Coder/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/master' into feat/mcp-client

Browse Source
This commit is contained in:
lintianle
2026-07-13 23:48:11 +08:00
parent af3152fefe 62eba83f4f
commit 2850b42654
611 changed files with 54555 additions and 10524 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
docs/AGENTS.md
View File

@@ -1,6 +1,6 @@
# AGENTS.md — The documentation standard
This file is the contract for every Markdown surface in the repo: each tier's job, the writing rules, and the word budgets the `verify-doc-budgets` gate enforces. The audit/apply workflow is the [dsh-doc-standards](../.agents/skills/dsh-doc-standards/SKILL.md) skill; the decision record is [the doc-tiers-and-budgets RFC](rfc/implemented/process/2026-07-04-doc-tiers-and-budgets.md).
This file is the contract for every Markdown files in the repo: each tier's job, the writing rules, and the word budgets that `verify-doc-budgets` enforces. The audit/apply workflow is the [dsh-doc-standards](../.agents/skills/dsh-doc-standards/SKILL.md) skill; the decision record is [the doc-tiers-and-budgets RFC](rfc/implemented/process/2026-07-04-doc-tiers-and-budgets.md).
## The tier taxonomy: one home per fact
@@ -31,14 +31,19 @@ Placement test: a story about a bug → postmortem. Why we chose X → RFC. How
- **Every new event's JSDoc carries an `@mode` tag** (emit | waterfall | parallel | serial); the catalog generator hard-errors without it. Write the JSDoc to stand alone — it becomes the catalog entry ([catalog RFC](rfc/implemented/process/2026-06-20-generated-cordis-catalog.md)).
- **The [core-data-structures catalog](core-data-structures/core.md) updates in the same change** that reshapes a documented type. `verify-type-equiv` catches drifted pastes, not never-documented new types ([what counts as core](core-data-structures/core.md#what-counts-as-core)).
- **Bilingual pairs update together**: editing either side obligates the counterpart and a re-record in the same change ([i18n contract](i18n/README.md)).
- Your audience is professional programmers. Prefer concise and straight-forward English over metaphor. Do not overuse words like "gate", "vocabulary", "surface", "seams".
## Budgets and the ceiling gate
## Wordcount Budgets
Standing docs accrete: every PR has a lesson it wants to append, and without displacement pressure nothing ever leaves. The gate is that pressure. [scripts/doc-budgets.manifest.json](../scripts/doc-budgets.manifest.json) lists the accretion-prone standing docs with a word ceiling each; `pnpm run verify-doc-budgets` (part of `doc-sync`, so CI and pre-push run it) fails when a doc exceeds its ceiling, and fails when a budgeted file is missing so a rename cannot orphan its budget.
Every PR has a lesson it wants to append, and without pressure nothing leaves. [scripts/doc-budgets.manifest.json](../scripts/doc-budgets.manifest.json) stores the allowed word-count ceiling for each budgeted standing doc; `pnpm run verify-doc-budgets` fails when a doc exceeds its ceiling or a budgeted file is missing.
- Ceilings are an enforcement frontier with working headroom: a ceiling sits at least 5% above the doc's current size — routine edits pass, real growth trips the gate — and ratchets down, keeping the margin, as the doc reaches target. Target budgets: root `AGENTS.md` ≤ 1,500 words; `architecture.md` ≤ 1,800; each subtree `AGENTS.md` ≤ 600, except this file (which carries the standard) ≤ 1,250; `packages/README.md` ≤ 600.
- When the gate goes red, first ask whether the added words belong in this tier and whether the existing wording can be condensed. If the words do not belong, relocate per the taxonomy above; if they belong but can be shorter, condense. If they truly need the space, raise the ceiling and justify the manifest diff in the PR. A ceiling set too low is a budget bug, and correcting it is the fix.
- Unbudgeted tiers (package READMEs, RFCs, reference matrices) have no ceiling — length is legitimate there when every row is a fact. Review and the slop checklist govern them instead.
When the gate goes red:
1. **Relocate** content that belongs in another tier; leave a one-line link if needed.
2. **Condense** content that belongs here but can be shorter.
3. **Raise** the ceiling only when the words truly need the space; justify the manifest diff in the PR. A too-low ceiling is a budget bug.
Ceilings keep working headroom: at least 5% above the current size, ratcheted down after trims. Target budgets: root `AGENTS.md` ≤ 1,500 words; `architecture.md` ≤ 1,800; each subtree `AGENTS.md` ≤ 600, except this file ≤ 1,250; `packages/README.md` ≤ 600. Unbudgeted tiers (package READMEs, RFCs, reference matrices) have no ceiling; review and the slop checklist govern them.
## The slop checklist

1
docs/agent-lifecycle.md
View File

@@ -39,6 +39,7 @@ sequenceDiagram
Tools-->>Session: tool-owned events when applicable
Driver->>Session: <code>tool/result</code> and <code>step/end</code>
Driver->>Hooks: <code>agent/turn-continuation</code> waterfall
Driver->>Hooks: <code>agent/turn-stop</code> serial terminal checkpoint
Driver->>Session: <code>turn/end</code>
Driver->>Persistence: <code>session/flush</code> parallel checkpoint
Driver-->>SDK: <code>agent/status</code> idle

79
docs/architecture.md
View File

@@ -1,23 +1,22 @@
# DeepSeek Harness Architecture
The **DeepSeek Harness SDK** is an SDK for building agent harnesses using the Cordis framework. The governing principle is simple: **everything is a plugin**. For example, the shipped agent loop is just one plugin in the default bundle, not a privileged kernel.
The **DeepSeek Harness SDK** builds agent harnesses on Cordis. The principle is simple: **everything is a plugin**. The shipped loop is one plugin, not a privileged kernel.
Read this page as the system map before changing `packages/`. It explains how the runtime is shaped, how the default loop moves work, where state lives, and where extensions attach. Type shapes live in [core-data-structures/](core-data-structures/core.md); exact event and service signatures live in the generated [events](cordis-catalog/events.md) and [services](cordis-catalog/services.md) catalogs; package contracts live in the [package map](../packages/README.md); rationale lives in the [RFCs](rfc/README.md). If Cordis itself is new to you, start with the [Cordis primer](cordis-primer.md).
## Overview
## System Shape
A harness is one [Cordis](cordis-primer.md) context. Packages contribute service keys, typed events, and disposable registrations: services expose stable calls (`ctx.llm`, `ctx.tools`, `ctx.sessions`), events provide interception and notifications (`agent/request`, `tools/pre-execute`, `session/event`), and registrations install prompt sections, tools, providers, adapters, or listeners.
A running harness is one Cordis context. Packages contribute service keys, typed events, and disposable registrations to that context. Services are the stable call surfaces (`ctx.llm`, `ctx.tools`, `ctx.sessions`); events are interception and notification points (`agent/request`, `tools/pre-execute`, `session/event`); registrations install prompt sections, tool schemas, providers, adapters, and listeners.
`packages/core/` groups the default agent flow; surrounding capabilities are equally first-class Cordis plugins.
The default distribution is a composition, not a hierarchy. `packages/core/` is a repository grouping for the default agent spine; capability seams around it are equally first-class plugins from a Cordis perspective.
### Default Service Spine
### Default Services
| ctx key | Package | Role |
|---|---|---|
| — | [`dsh-scope`](../packages/core/scope/README.md) | scoped-context registration primitive (library) |
| `ctx.sessions` | `dsh-session` | in-memory event-sourced sessions |
| `ctx.systemPrompt` | `dsh-system-prompt` | ordered prompt sections, tool schemas, and prompt variables |
| `ctx.tools` | `dsh-tools` | tool registry and [execution pipeline](tool-execution-pipeline.md) |
| `ctx.agents` | `dsh-agent` | live agent registry, public `Agent` handle, `agent/*` vocabulary |
| `ctx.agents` | `dsh-agent` | live agent registry, public `Agent` handle, `agent/*` events |
| `ctx.agentLoop` | `dsh-agent-loop` | shipped `ReactLoopAgent` driver |
### Capability Services
@@ -26,39 +25,42 @@ The default distribution is a composition, not a hierarchy. `packages/core/` is
|---|---|---|
| `ctx.llm` | [`llm/`](../packages/llm/README.md) | adapter registry and streaming model calls |
| `ctx.bash` | [`bash/`](../packages/bash/README.md) | foreground/background command execution |
| `ctx.sandbox` | [`sandbox/`](../packages/sandbox/README.md) | same-world process confinement (argv wrapping, per-call policy) |
| `ctx.codeRuntime` | [`code-runtime/`](../packages/code-runtime/README.md) | model-written program execution |
| `ctx.fs` | [`fs/`](../packages/fs/README.md) | filesystem provider primitives and policy events |
| `ctx.skills` | [`skill/`](../packages/skill/README.md) | skill provider registry and progressive disclosure |
| `ctx.web` | [`web/`](../packages/web/README.md) | search/fetch provider registries |
| `ctx.compact` | [`compact/`](../packages/compact/README.md) | session-surface compaction |
| `ctx.compact` | [`compact/`](../packages/compact/README.md) | session-log compaction |
| `ctx.subagents` | [`subagent/`](../packages/subagent/README.md) | named delegation providers |
| `ctx.workflows` | [`workflow/`](../packages/workflow/README.md) | script-driven multi-agent orchestration |
| `ctx.sessionPersistence` | [`session-persistence/`](../packages/session-persistence/README.md) | durable storage for session logs |
## Event Surface
## Event
Events are the harness extension API. Each service owns the vocabulary for the behavior it controls, and the generated [events catalog](cordis-catalog/events.md) is the exhaustive reference. The [producer/consumer map](event-producer-consumer.md) shows which packages emit or listen to each event.
Events form the service extension API; see the exhaustive [events catalog](cordis-catalog/events.md) and [producer/consumer map](event-producer-consumer.md).
### Event Domains
Use the event domain to decide where new behavior belongs:
- **Session events** are durable, replayable facts. Turn and step boundaries, user input, assistant output, tool calls, tool results, steering, compaction records, and tool-owned durable facts append to the session log and flow through `session/event`.
- **Agent events** are live runtime surfaces. They carry the live `Agent` handle for status, diagnostics, prompt admission, call-config shaping, result validation, and continuation policy.
- **Agent events** carry the live `Agent` handle for status, diagnostics, prompt admission, call-config shaping, result validation, and continuation policy.
- **Capability events** belong to the seam that owns the action. `tools/*`, `llm/*`, `system-prompt/*`, `fs/*`, and `subagent/*` let policy and adapters attach without importing the loop.
### Interception Semantics
Waterfall events behave like around-middleware: a listener delegates by calling `next()` and vetoes or takes over by returning without it. The full rule lives in [Cordis waterfall semantics](cordis-primer.md#cordis-waterfall-semantics).
Waterfall events behave like around-middleware: a listener delegates by calling `next()`; returning without it vetoes or takes over. Full rule: [Cordis waterfall semantics](cordis-primer.md#cordis-waterfall-semantics).
## Default Loop Lifecycle
The shipped loop drains queued work, assembles a request, streams a model answer, executes tools, decides whether to continue, and checkpoints durable state. The important architecture is where it pauses: each pause is a documented service call or event seam that another plugin can program against.
The shipped loop drains work, assembles requests, streams model answers, executes tools, applies continuation policy, and checkpoints state. Every pause is a service call or event available to plugins.
A **session** is one agent's append-only event log. A **turn** drains one queued batch and runs until the model stops asking for tools and no plugin requests continuation. A **step** is one model request plus the tool executions caused by that response. In the flow below ([sequence companion](agent-lifecycle.md)), quoted names are durable session events and event names are extension seams.
A **session** is one agent's append-only event log. A **turn** drains one queued batch and runs until the model stops asking for tools and no plugin requests continuation. A **step** is one model request plus the tool executions caused by that response. In the flow below ([sequence companion](agent-lifecycle.md)), quoted names are durable session events and event names are extension points.
### Turn Flow
```text
create agent -> emit agent/session-start(source)
prepare private session + agent.ctx -> await unpublished setup
-> enter session + agent -> session/created -> agent/created
-> enable driving -> agent/session-start(source) -> start driver
forever:
wait for queued messages
emit agent/status(running)
@@ -70,6 +72,7 @@ forever:
STEP loop:
drain steering
assemble system prompt and tool schemas
agent/session-prefix (first step)
agent/pre-step
'step/start'
snapshot the derived messages (the reconstruction boundary)
@@ -79,19 +82,20 @@ forever:
'assistant/message'
each tool call:
'tool/call'
tools/pre-execute -> dispatch -> tools/post-execute
tools/pre-execute -> monotonic guards -> tools/execute -> tools/post-execute -> tools/result
'tool/result'
append post-tool context and steering
'step/end'
agent/turn-continuation
agent/turn-stop (terminal policy)
stop unless tools or continuation policy ask for another step
'turn/end'
checkpoint persistence and notify idle/running status
```
Prompt assembly is single-path: `renderPrompt(assemble({ agent }))` IS the system prompt sent to the model. Plugins contribute ordered sections (static or computed from the per-call `AssembleContext`), tool schemas, and named variables interpolated as `{{name}}` at render — strictly, so an unknown or valueless reference fails the turn instead of shipping a hole. `dsh-system-prompt` itself owns the openers — the static `harness:identity` section (order −100) and the deployment's persona (order 0, from its `persona` config, shared by every agent in the context) — while the shipped loop registers the `model`/`cwd` variables; prompt-fact ownership is pinned by the [prompt-variables RFC](rfc/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md).
Prompt assembly is single-path: the loop sends `renderPrompt(await assemble(assembleContextFor(agent)))`; the helper couples the explicit agent and scope. Plugins contribute ordered sections, tool schemas, and named variables interpolated as `{{name}}` at render — strictly, so an unknown or valueless reference fails the turn instead of shipping a hole. `dsh-system-prompt` owns the openers — the static `harness:identity` section (order −100) and the deployment's global default persona (order 0, shadowable by a same-named agent-scoped section) — while the loop registers the `model`/`cwd` variables; prompt-fact ownership is pinned by the [prompt-variables RFC](rfc/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md).
Post-tool context lands after all tool results so tool-call/result adjacency stays stable. Steering drains between steps; leftover steering after a turn is re-queued as ordinary input.
Post-tool context lands after all tool results so tool-call/result adjacency stays stable. Steering drains between steps; ordinary leftover steering after a turn is re-queued as input. A terminal `agent/turn-stop` is the explicit exception: it runs after ordinary continuation and steering folding, then remains authoritative through turn close and flush so steering from those later listeners is discarded rather than becoming another step or turn; ordinary queued prompts are preserved.
### Failure Boundaries
@@ -101,21 +105,25 @@ Every session event is turn-enclosed. Reloading a crashed session preserves the
### Agent Handles
`ctx.agents` owns live agents and returns an `AgentHandle { agent, dispose() }`. `Agent` is the surface other plugins drive: `send()` queues work, `steer()` injects mid-turn content, `inject()` appends context and opens a one-shot injection turn when idle, `cancel()` is the public stop primitive, and `whenIdle()` observes quiescence. Lifecycle owners tear down with `await dispose()`.
`ctx.agents` owns live agents and returns an `AgentHandle { agent, dispose() }`. `Agent` is the API other plugins drive: `send()` queues work, `steer()` injects mid-turn content, `inject()` appends context and opens a one-shot injection turn when idle, `cancel()` is the public stop primitive, and `whenIdle()` observes quiescence. The caller fiber and concrete factory provider structurally co-own programmatic lifecycles; a consumer handle is the only non-structural teardown capability, and every owner reaches the same awaited disposer.
## State And Model Surface
### Agent Scope
Every live agent owns `agent.ctx` ([`dsh-scope`](../packages/core/scope/README.md), keyed by the agent). Its registrations are visible only to that agent, shadow same-named globals, and unwind with it. Its listeners hear only that agent's dispatches; an opaque carrier routes while the real subject stays explicit. `CreateAgentOptions.setup(agentCtx)` composes this world before publication and does not drive. Dev invariants and `verify-scoped-dispatch` keep carrier/subject identity aligned with event declarations. Rationale: [agent-scope RFC](rfc/implemented/architecture/2026-07-08-agent-scope-contexts.md); subagent `persona`, `toolFilter`, and `maxDepth` are the separate [composition-controls feature](rfc/implemented/feature/2026-07-12-subagent-persona-tool-filter-and-depth.md).
## State
### Session Log
The session log is the source of truth. `deriveMessages()` projects session events into the `Message[]` sent to the model; raw `assistant/chunk` events stay in the log for replay and UI fidelity. Replay, fork, resume, transcript rendering, telemetry, and persistence all derive from the same event stream.
**Model-visible ⟺ logged**: the log reconstructs every request — messages at `step/start`, headers by folding `request/header` — and dev invariants assert this ([reconstructability RFC](rfc/implemented/architecture/2026-07-05-reconstructable-requests.md)).
**Model-visible ⟺ logged**: the log reconstructs every request — messages at `step/start` fronted by the header's session prefix, headers by folding `request/header` — and dev invariants assert this ([reconstructability RFC](rfc/implemented/architecture/2026-07-05-reconstructable-requests.md)).
Durability is a plugin concern. Persistence backends buffer synchronous `session/event` notifications and the loop awaits a turn-end checkpoint before moving on. The `SessionPersistence` seam stores `SessionEvent` directly, with metadata in `SessionHeader`; JSONL and SQLite share one contract suite.
### Model Content
Messages are arrays of typed content blocks (`text`, `reasoning`, `tool-call`, `tool-result`). The union derives from the merge-extensible `ContentBlockMap`; the same pattern types `MessageSource`, `FinishReason`, `TurnTrigger`, and `TurnEndReason`. New block types are coordinated across adapters, UI bridges, compaction pricing, and persistence, so block vocabulary remains a repo-wide contract.
Messages are arrays of typed content blocks (`text`, `reasoning`, `tool-call`, `tool-result`). The union derives from the merge-extensible `ContentBlockMap`; the same pattern types `MessageSource`, `FinishReason`, `TurnTrigger`, and `TurnEndReason`. New block types are coordinated across adapters, UI bridges, compaction pricing, and persistence, so block types remain a repo-wide contract.
Streaming is a raw chunk protocol (`block-start` through `finish`) with `BlockAssembler` as the shared chunk-to-block assembler. The loop logs raw chunks while assembling them for dispatch. `LlmAdapter` is the provider seam: subclass, implement `stream()`, and register with `ctx.llm.registerAdapter(models, adapter)`. StreamChunk conventions live in [llm-streaming.md](core-data-structures/llm-streaming.md).
@@ -123,17 +131,17 @@ Streaming is a raw chunk protocol (`block-start` through `finish`) with `BlockAs
### Capability Pattern
A swappable capability usually splits into **interface / implementation / consumer**: the interface owns the `ctx` key and vocabulary; an implementation registers a backend; a consumer exposes model-facing behavior through `ctx.tools` or prompt assembly. The bash trio is the reference shape, and the [capability seam graph](capability-seams.md) shows the current package families.
A swappable capability usually splits into **interface / implementation / consumer**: the interface owns its `ctx` key and events, an implementation registers a backend, and a consumer exposes model behavior through tools or prompts. Bash is the reference; the [capability graph](capability-seams.md) shows every family.
Some seams bend the template deliberately. LLM keeps interface and consumer vocabulary together because adapters are the implementations. Filesystem adds policy as event gates around provider primitives. Web is one service with search and fetch provider registries, so provider swaps do not rename model tools. Subagents use a named provider registry because multiple delegation backends can coexist; `spawn` starts fresh, `fork` seeds from the parent's completed-turn prefix, and ACP can drive an out-of-process child ([subagent.md](core-data-structures/subagent.md)).
Some seams bend the template deliberately. LLM keeps interface and consumer vocabulary together because adapters are the implementations. Filesystem adds policy gates around provider primitives. Web is one service with search and fetch provider registries, so provider swaps do not rename model tools. Skills and subagents use named provider registries; local skills scan project/user roots, and other providers can add embedded or remote catalogs without registry/tool changes. Subagents spawn fresh, fork from the parent's completed-turn prefix, or use ACP children ([subagent.md](core-data-structures/subagent.md)).
### Bundles And Apps
`dsh-agent-core` is the default composition bundle: one plugin loading the providerless spine as code ([README](../packages/core/agent-core/README.md)). App packages compose it with a front door and own the boot `bin`: `dsh-stdio-agent` for the terminal REPL, and `dsh-acp-agent` for ACP over JSON-RPC stdio with no stdout logger ([ui/](../packages/ui/README.md)). A deployment is a thin `cordis.yml` leaf: swappable backends, one app entry, and optional product tools ([examples/](../examples/AGENTS.md), [runnable wirings](cookbook/extension-cookbook.md#runnable-wirings), [graph atlas](graph-atlas.md)).
`dsh-agent-core` is the default composition bundle: one plugin loading the shared spine ([README](../packages/core/agent-core/README.md)). App packages compose it with a front door and boot `bin`: `dsh-stdio-agent` for terminal REPL, and `dsh-acp-agent` for ACP over JSON-RPC stdio with no stdout logger ([ui/](../packages/ui/README.md)). A deployment is a thin `cordis.yml` leaf: swappable backends, one app entry, and optional product tools ([examples/](../examples/AGENTS.md), [runnable wirings](cookbook/extension-cookbook.md#runnable-wirings), [graph atlas](graph-atlas.md)).
### Where New Behavior Goes
New behavior should attach to a documented seam; changing the shipped loop requires updating this map.
New behavior should attach to a documented extension point; changing the shipped loop requires updating this map.
| Goal | Mechanism |
|---|---|
@@ -141,9 +149,20 @@ New behavior should attach to a documented seam; changing the shipped loop requi
| Add a model-facing capability | register a tool on `ctx.tools`; schemas flow into prompt assembly |
| Add command execution | implement and register a `ctx.bash` backend |
| Add filesystem access or policy | implement a `ctx.fs` provider or listen on `fs/*` policy events |
| Intercept prompts, requests, tool use, or continuation | listen on the relevant `agent/*` or `tools/*` waterfall |
| Confine spawned processes | a `ctx.sandbox` backend; consumers wrap their argv before spawning |
| Intercept prompts, requests, tool use, or continuation | listen on the relevant `agent/*` or `tools/*` waterfall; use serial `agent/turn-stop` for a monotonic terminal stop |
| Add a session-stable request prefix outside history | compose it on `agent/session-prefix`, once per loop instance; logged on the request header |
| Add UI or editor integration | drive `ctx.agents` and render from `session/event` |
| Add durable session state | add a `SessionEventMap` member and render/replay from the log |
| Fork a live session | use `ctx.sessions.fork(source, boundary?, childSessionId?)` |
| Scope a tool, prompt section, or listener to ONE agent | register it through that agent's `agent.ctx` (see Agent Scope) |
The [extension cookbook](cookbook/extension-cookbook.md) carries plugin skeletons and the feature-to-seam map; step-by-step guides cover [packages](cookbook/adding-a-package.md), [tools](cookbook/adding-a-tool.md), [LLM adapters](cookbook/adding-an-llm-adapter.md), and [vendored packages](cookbook/adding-a-vendored-package.md).
## Quick Reference
- Domain terms in the [glossary](glossary.md)
- Type definitions in [core-data-structures/](core-data-structures/core.md)
- Exact event and service signatures in [events](cordis-catalog/events.md)
- [services](cordis-catalog/services.md) catalogs
- package contracts in the [package map](../packages/README.md)
- [RFCs](rfc/README.md)

59
docs/capability-seams.md
View File

@@ -29,21 +29,36 @@ flowchart LR
pkg_tools["tools"]
pkg_tool_fs["tool-fs"]
pkg_tool_web["tool-web"]
svc_tools["ctx.tools<br/>Tool registry and execution waterfall"]
svc_tools["ctx.tools<br/>Tool registry and guarded execution pipeline"]
pkg_tool_ask_user["tool-ask-user"]
pkg_tool_bash["tool-bash"]
pkg_tool_cordis["tool-cordis"]
pkg_tool_skill["tool-skill"]
pkg_tool_subagent["tool-subagent"]
pkg_tool_todo["tool-todo"]
svc_agents["ctx.agents<br/>Agent registry"]
pkg_user_interaction["user-interaction"]
svc_userInteraction["ctx.userInteraction<br/>Human question/answer seam"]
pkg_stdio_agent["stdio-agent"]
pkg_skill["skill"]
svc_skills["ctx.skills<br/>Skill provider registry"]
pkg_skill_local["skill-local"]
svc_agents["ctx.agents<br/>Agent registry"]
svc_agentLoop["ctx.agentLoop<br/>Concrete loop driver"]
pkg_agent_core["agent-core"]
pkg_bash["bash"]
svc_bash["ctx.bash<br/>Bash executor seam"]
pkg_bash_local["bash-local"]
pkg_bash_sandbox["bash-sandbox"]
pkg_hooks_claude["hooks-claude"]
pkg_hooks_codex["hooks-codex"]
pkg_sandbox["sandbox"]
svc_sandbox["ctx.sandbox<br/>Process-sandbox seam"]
pkg_sandbox_local["sandbox-local"]
pkg_approval["approval"]
svc_approval["ctx.approval<br/>Approval seam"]
pkg_code_runtime["code-runtime"]
svc_codeRuntime["ctx.codeRuntime<br/>Code-execution seam"]
pkg_code_runtime_worker["code-runtime-worker"]
pkg_fs["fs"]
svc_fs["ctx.fs<br/>Filesystem provider seam"]
pkg_fs_local["fs-local"]
@@ -62,11 +77,20 @@ flowchart LR
pkg_web_search_perplexity["web-search-perplexity"]
pkg_web_search_deepseek["web-search-deepseek"]
pkg_web_fetch_local["web-fetch-local"]
pkg_workflow["workflow"]
svc_workflows["ctx.workflows<br/>Workflow script engine"]
pkg_workflow_workerthread["workflow-workerthread"]
pkg_tool_workflow["tool-workflow"]
pkg_acp --> svc_approval
pkg_acp --> svc_userInteraction
pkg_agent --> svc_agents
pkg_agent_loop --> svc_agentLoop
pkg_approval --> svc_approval
pkg_bash --> svc_bash
pkg_bash_local --> svc_bash
pkg_bash_sandbox --> svc_bash
pkg_code_runtime --> svc_codeRuntime
pkg_code_runtime_worker --> svc_codeRuntime
pkg_compact --> svc_compact
pkg_compact_basic --> svc_compact
pkg_fs --> svc_fs
@@ -75,10 +99,15 @@ flowchart LR
pkg_llm_deepseek --> svc_llm
pkg_llm_pi_ai --> svc_llm
pkg_llm_replay --> svc_llm
pkg_sandbox --> svc_sandbox
pkg_sandbox_local --> svc_sandbox
pkg_session --> svc_sessions
pkg_session_persistence --> svc_sessionPersistence
pkg_session_persistence_jsonl --> svc_sessionPersistence
pkg_session_persistence_sqlite --> svc_sessionPersistence
pkg_skill --> svc_skills
pkg_skill_local --> svc_skills
pkg_stdio_agent --> svc_userInteraction
pkg_subagent --> svc_subagents
pkg_subagent_acp --> svc_subagents
pkg_subagent_fork --> svc_subagents
@@ -86,24 +115,31 @@ flowchart LR
pkg_subagent_spawn --> svc_subagents
pkg_system_prompt --> svc_systemPrompt
pkg_tools --> svc_tools
pkg_user_interaction --> svc_userInteraction
pkg_web --> svc_web
pkg_web_fetch_local --> svc_web
pkg_web_search_deepseek --> svc_web
pkg_web_search_exa --> svc_web
pkg_web_search_perplexity --> svc_web
pkg_workflow --> svc_workflows
pkg_workflow_workerthread --> svc_workflows
svc_agentLoop --> pkg_agent_core
svc_agents --> pkg_acp
svc_agents --> pkg_agent_loop
svc_agents --> pkg_invariants
svc_agents --> pkg_stdio_agent
svc_agents --> pkg_subagent_inprocess
svc_approval --> pkg_tool_bash
svc_approval --> pkg_tools
svc_bash --> pkg_hooks_claude
svc_bash --> pkg_hooks_codex
svc_bash --> pkg_tool_bash
svc_codeRuntime --> pkg_tools
svc_compact --> pkg_compact_basic
svc_fs --> pkg_tool_fs
svc_llm --> pkg_agent_loop
svc_llm --> pkg_compact_basic
svc_sandbox --> pkg_bash_sandbox
svc_sessionPersistence --> pkg_acp
svc_sessionPersistence --> pkg_agent_loop
svc_sessions --> pkg_agent
@@ -111,6 +147,7 @@ flowchart LR
svc_sessions --> pkg_invariants
svc_sessions --> pkg_session_persistence
svc_sessions --> pkg_subagent_inprocess
svc_skills --> pkg_tool_skill
svc_subagents --> pkg_tool_subagent
svc_systemPrompt --> pkg_agent_loop
svc_systemPrompt --> pkg_tool_fs
@@ -118,12 +155,19 @@ flowchart LR
svc_systemPrompt --> pkg_tools
svc_tools --> pkg_acp
svc_tools --> pkg_agent_loop
svc_tools --> pkg_tool_ask_user
svc_tools --> pkg_tool_bash
svc_tools --> pkg_tool_cordis
svc_tools --> pkg_tool_fs
svc_tools --> pkg_tool_skill
svc_tools --> pkg_tool_subagent
svc_tools --> pkg_tool_todo
svc_tools --> pkg_tool_web
svc_userInteraction --> pkg_acp
svc_userInteraction --> pkg_stdio_agent
svc_userInteraction --> pkg_tool_ask_user
svc_web --> pkg_tool_web
svc_workflows --> pkg_tool_workflow
svc_fs -. event gate .-> pkg_fs_policy
```
@@ -133,14 +177,19 @@ flowchart LR
| `ctx.sessions` | `core` | [`session`](../packages/core/session) | - | [`agent-loop`](../packages/core/agent-loop), [`agent`](../packages/core/agent), [`session-persistence`](../packages/session-persistence/session-persistence), [`subagent-inprocess`](../packages/subagent/subagent-inprocess), [`invariants`](../packages/support/invariants) | - | Owns append-only Session instances and emits the durable session event feed. |
| `ctx.sessionPersistence` | `seam` | [`session-persistence`](../packages/session-persistence/session-persistence) | [`session-persistence-jsonl`](../packages/session-persistence/session-persistence-jsonl), [`session-persistence-sqlite`](../packages/session-persistence/session-persistence-sqlite) | [`agent-loop`](../packages/core/agent-loop), [`acp`](../packages/ui/acp) | - | Backends persist the same SessionEvent vocabulary; apps choose a backend at composition time. |
| `ctx.systemPrompt` | `core` | [`system-prompt`](../packages/core/system-prompt) | - | [`agent-loop`](../packages/core/agent-loop), [`tools`](../packages/core/tools), [`tool-fs`](../packages/fs/tool-fs), [`tool-web`](../packages/web/tool-web) | - | Collects prompt sections and model-facing tool schemas for each step. |
| `ctx.tools` | `core` | [`tools`](../packages/core/tools) | - | [`agent-loop`](../packages/core/agent-loop), [`tool-bash`](../packages/bash/tool-bash), [`tool-fs`](../packages/fs/tool-fs), [`tool-subagent`](../packages/subagent/tool-subagent), [`tool-todo`](../packages/todo/tool-todo), [`tool-web`](../packages/web/tool-web), [`acp`](../packages/ui/acp) | - | Registers tool definitions, exposes schemas to the prompt, and routes calls through tools/pre-execute and tools/post-execute. |
| `ctx.tools` | `core` | [`tools`](../packages/core/tools) | - | [`agent-loop`](../packages/core/agent-loop), [`tool-ask-user`](../packages/ui/tool-ask-user), [`tool-bash`](../packages/bash/tool-bash), [`tool-cordis`](../packages/cordis/tool-cordis), [`tool-fs`](../packages/fs/tool-fs), [`tool-skill`](../packages/skill/tool-skill), [`tool-subagent`](../packages/subagent/tool-subagent), [`tool-todo`](../packages/todo/tool-todo), [`tool-web`](../packages/web/tool-web), [`acp`](../packages/ui/acp) | - | Registers capabilities, owns Code Mode transport, and routes calls through pre-policy, monotonic guards, around dispatch, post-policy, and final-result observation. |
| `ctx.userInteraction` | `seam` | [`user-interaction`](../packages/ui/user-interaction) | [`stdio-agent`](../packages/ui/stdio-agent), [`acp`](../packages/ui/acp) | [`tool-ask-user`](../packages/ui/tool-ask-user), [`stdio-agent`](../packages/ui/stdio-agent), [`acp`](../packages/ui/acp) | - | UI front doors provide the active human-answer provider; tool-ask-user pauses a tool call on the provider-neutral ask() promise. |
| `ctx.skills` | `seam` | [`skill`](../packages/skill/skill) | [`skill-local`](../packages/skill/skill-local) | [`tool-skill`](../packages/skill/tool-skill) | - | Merges provider skill catalogs; tool-skill renders the session-prefix catalog and loads complete skill bodies. |
| `ctx.agents` | `core` | [`agent`](../packages/core/agent) | - | [`agent-loop`](../packages/core/agent-loop), [`acp`](../packages/ui/acp), [`subagent-inprocess`](../packages/subagent/subagent-inprocess), [`stdio-agent`](../packages/ui/stdio-agent), [`invariants`](../packages/support/invariants) | - | Owns live Agent handles and the create/resume factory seam. |
| `ctx.agentLoop` | `bundle` | [`agent-loop`](../packages/core/agent-loop) | - | [`agent-core`](../packages/core/agent-core) | - | The one concrete loop plugin; extension packages depend on dsh-agent events and services, not on this package. |
| `ctx.bash` | `seam` | [`bash`](../packages/bash/bash) | [`bash-local`](../packages/bash/bash-local) | [`tool-bash`](../packages/bash/tool-bash), [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex) | - | The model-facing bash tools and hook bridges consume this seam; sandboxed or remote executors can replace bash-local. |
| `ctx.codeRuntime` | `seam` | [`code-runtime`](../packages/code-runtime/code-runtime) | - | - | - | Runs one model-written program against host-provided async bindings; backends differ by substrate and language (the Code Mode RFC specifies the worker-thread backend and the tool-registry consumer). |
| `ctx.bash` | `seam` | [`bash`](../packages/bash/bash) | [`bash-local`](../packages/bash/bash-local), [`bash-sandbox`](../packages/bash/bash-sandbox) | [`tool-bash`](../packages/bash/tool-bash), [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex) | - | The model-facing bash tools and hook bridges consume this seam; sandboxed or remote executors replace bash-local without touching them. |
| `ctx.sandbox` | `seam` | [`sandbox`](../packages/sandbox/sandbox) | [`sandbox-local`](../packages/sandbox/sandbox-local) | [`bash-sandbox`](../packages/bash/bash-sandbox) | - | Consumers hand over the exact argv they are about to spawn; same-world backends wrap it under a per-call policy and report enforcement. |
| `ctx.approval` | `seam` | `approval` | [`acp`](../packages/ui/acp) | [`tools`](../packages/core/tools), [`tool-bash`](../packages/bash/tool-bash) | - | One-shot permission decisions dispatched over the `approval/request` waterfall; answerers are listeners (the ACP bridge for its own agents), absence fails closed to `unavailable`. |
| `ctx.codeRuntime` | `seam` | [`code-runtime`](../packages/code-runtime/code-runtime) | [`code-runtime-worker`](../packages/code-runtime/code-runtime-worker) | [`tools`](../packages/core/tools) | - | Runs one model-written program against host-provided async bindings; backends differ by substrate and language (the tool registry consumes it for Code Mode). |
| `ctx.fs` | `seam` | [`fs`](../packages/fs/fs) | [`fs-local`](../packages/fs/fs-local) | [`tool-fs`](../packages/fs/tool-fs) | [`fs-policy`](../packages/fs/fs-policy) | tool-fs executes read/write/edit through ctx.fs; fs-policy contributes observed-state checks through the fs/* event gate. |
| `ctx.compact` | `seam` | [`compact`](../packages/compact/compact) | [`compact-basic`](../packages/compact/compact-basic) | [`compact-basic`](../packages/compact/compact-basic) | - | The basic backend currently consumes the pre-step event directly; a model-facing compact tool remains deferred. |
| `ctx.subagents` | `seam` | [`subagent`](../packages/subagent/subagent) | [`subagent-spawn`](../packages/subagent/subagent-spawn), [`subagent-fork`](../packages/subagent/subagent-fork), [`subagent-acp`](../packages/subagent/subagent-acp), [`subagent-mock`](../packages/support/subagent-mock) | [`tool-subagent`](../packages/subagent/tool-subagent) | - | Providers implement transports; tool-subagent exposes one configured provider as a model-facing tool name. |
| `ctx.web` | `seam` | [`web`](../packages/web/web) | [`web-search-exa`](../packages/web/web-search-exa), [`web-search-perplexity`](../packages/web/web-search-perplexity), [`web-search-deepseek`](../packages/web/web-search-deepseek), [`web-fetch-local`](../packages/web/web-fetch-local) | [`tool-web`](../packages/web/tool-web) | - | Search and fetch providers register into one ctx.web seam; tool-web owns the stable model-facing names. |
| `ctx.workflows` | `seam` | [`workflow`](../packages/workflow/workflow) | [`workflow-workerthread`](../packages/workflow/workflow-workerthread) | [`tool-workflow`](../packages/workflow/tool-workflow) | - | One engine per context (bash shape, no named-provider registry); the worker-thread engine fans agent() calls out through ctx.subagents. |
Maintenance mode: hybrid: services are discovered from Cordis declarations; interface/implementation/consumer roles are classified in `scripts/gen-doc-graphs.ts` with a completeness guard.

490
docs/config-catalog.md
View File

@@ -11,7 +11,7 @@ A `Requires:` line lists the service keys the plugin `inject`s: its `cordis.yml`
## `@deepseek-ai/dsh-acp`
Requires: `agents` · `sessions` · `sessionPersistence` · `tools`
Requires: `agents` · `sessions` · `sessionPersistence` · `tools` · `userInteraction`
```ts config-catalog
/** Plugin config: the agent template ACP sessions are created from. */
@@ -31,7 +31,7 @@ export interface AcpConfig {
Depends on: `Stream` (`@agentclientprotocol/sdk`)
Source: [`packages/ui/acp/src/index.ts:115`](../packages/ui/acp/src/index.ts)
Source: [`packages/ui/acp/src/index.ts:250`](../packages/ui/acp/src/index.ts)
## `@deepseek-ai/dsh-acp-agent`
@@ -42,7 +42,8 @@ Source: [`packages/ui/acp/src/index.ts:115`](../packages/ui/acp/src/index.ts)
* pre-created agent — ACP creates agents at `session/new`); `persona` is the
* deployment persona (forwarded to the system-prompt plugin); `toolOrder` is
* the explicit model-facing tool order (forwarded to the system-prompt plugin);
* `persistenceRoot` is the JSONL backend's directory.
* `tools` is the tool registry's config (its presentation `mode`, forwarded
* through agent-core); `persistenceRoot` is the JSONL backend's directory.
*/
export interface Config {
/** Model name for ACP-created agents (must have a registered adapter). */
@@ -51,12 +52,18 @@ export interface Config {
persona?: string
/** Explicit model-facing tool order (the system-prompt plugin's `toolOrder` config; see dsh-system-prompt). */
toolOrder?: string[]
/** Tool-registry config — its presentation `mode` (forwarded through agent-core; see dsh-tools). */
tools?: ToolsConfig
/** Directory the JSONL session backend writes under. Defaults to `./.sessions`. */
persistenceRoot?: string
/** Skill registry, local-provider, and model-facing consumer config forwarded to agent-core. */
skills?: agentCore.SkillConfig
}
```
Source: [`packages/ui/acp-agent/src/index.ts:49`](../packages/ui/acp-agent/src/index.ts)
Depends on: [`agentCore`](../packages/core/agent-core/src/index.ts) · [`ToolsConfig`](#deepseek-aidsh-tools)
Source: [`packages/ui/acp-agent/src/index.ts:52`](../packages/ui/acp-agent/src/index.ts)
## `@deepseek-ai/dsh-agent-core`
@@ -66,10 +73,12 @@ Source: [`packages/ui/acp-agent/src/index.ts:49`](../packages/ui/acp-agent/src/i
* `agents` to the agent loop (an app that pre-creates no agents, like the ACP
* bridge, simply omits it), `persona` and `toolOrder` to the system-prompt
* plugin (the deployment's persona section and the explicit model-facing tool
* order). Every field is optional INPUT here because each owner's schema
* supplies the default (`[]` / `''` / absent — lexicographic); the schema is
* the INTERSECTION of the owners' own schemas, so validation and defaulting
* can never drift from them.
* order), the `tools` object to the tool registry (its presentation `mode`),
* and `skills` to the skill registry/local provider/tool consumer. Every field
* is optional INPUT here because each owner's schema supplies the default;
* the schema is the INTERSECTION of the owners' own schemas (with registry
* schemas nested under their bundle keys), so validation and defaulting can
* never drift from them.
*/
export interface Config {
/** The agent-loop `agents` list (see dsh-agent-loop's `Config`). */
@@ -78,40 +87,41 @@ export interface Config {
persona?: SystemPromptConfig['persona']
/** The explicit model-facing tool order (see dsh-system-prompt's `Config`). */
toolOrder?: SystemPromptConfig['toolOrder']
/** The tool registry's config — its presentation `mode` (see dsh-tools' `Config`). */
tools?: ToolsConfig
/** Skill registry, local provider, and model-facing consumer config. */
skills?: SkillConfig
}
/** Skill bundle config forwarded to the registry, local provider, and model-facing consumer. */
export interface SkillConfig {
/** Registry-level discovery cache settings. */
registry?: SkillRegistryConfig
/** Local filesystem skill provider settings. */
local?: SkillLocal.Config
/** Model-facing skill catalog and tool settings. */
tool?: toolSkill.Config
}
```
Depends on: [`AgentLoopConfig`](#deepseek-aidsh-agent-loop) · [`SystemPromptConfig`](#deepseek-aidsh-system-prompt)
Depends on: [`AgentLoopConfig`](#deepseek-aidsh-agent-loop) · [`SkillLocal`](../packages/skill/skill-local/src/index.ts) · [`SkillRegistryConfig`](#deepseek-aidsh-skill) · [`SystemPromptConfig`](#deepseek-aidsh-system-prompt) · [`ToolsConfig`](#deepseek-aidsh-tools) · [`toolSkill`](../packages/skill/tool-skill/src/index.ts)
Source: [`packages/core/agent-core/src/index.ts:69`](../packages/core/agent-core/src/index.ts)
Source: [`packages/core/agent-core/src/index.ts:87`](../packages/core/agent-core/src/index.ts)
## `@deepseek-ai/dsh-agent-loop`
Requires: `agents` · `sessions` · `llm` · `tools` · `systemPrompt`
```ts config-catalog
/**
* Plugin config: the agents to create — or resume, via `resumeSessionId` —
* declaratively at startup, so a cordis.yml deployment needs no code.
*/
/** Plugin configuration for declarative startup agents. */
export interface Config {
/** Agents created from configuration at startup. */
/** Agents created or resumed at plugin startup. */
agents: (AgentOptions & {
/** Agent id to register under; also seeds the fresh per-run session id (`${id}-session-<uuid>`). */
/** Registry identity for the live agent. */
id: AgentId
/**
* If set, the config agent RESUMES this persisted session id instead of
* starting a fresh `${id}-session-<uuid>`. Sourced from an env var in
* cordis.yml (`resumeSessionId: !!js process.env.RESUME_SESSION_ID`), so a
* demo can continue a prior conversation without code changes. Requires a
* `dsh-session-persistence` backend; the resume is deferred until that
* service is available (via `ctx.inject`) and the loaded session's events
* seed the live session so history continues.
*
* The schema accepts a plain string at runtime (cordis.yml values are
* untyped); the brand is compile-time only — the config format is the
* boundary where an id enters, so the TYPE declares the brand here.
*/
/** Optional workspace for a fresh session. */
cwd?: string
/** Persisted session to resume instead of creating a fresh session. */
resumeSessionId?: SessionId
})[]
}
@@ -119,7 +129,7 @@ export interface Config {
Depends on: [`AgentId`](../packages/core/agent/src/index.ts) · [`AgentOptions`](../packages/core/agent/src/index.ts) · [`SessionId`](../packages/core/session/src/index.ts)
Source: [`packages/core/agent-loop/src/index.ts:36`](../packages/core/agent-loop/src/index.ts)
Source: [`packages/core/agent-loop/src/index.ts:325`](../packages/core/agent-loop/src/index.ts)
## `@deepseek-ai/dsh-bash-local`
@@ -139,7 +149,70 @@ export interface Config {
}
```
Source: [`packages/bash/bash-local/src/index.ts:28`](../packages/bash/bash-local/src/index.ts)
Source: [`packages/bash/bash-local/src/index.ts:29`](../packages/bash/bash-local/src/index.ts)
## `@deepseek-ai/dsh-bash-sandbox`
Requires: `sandbox`
```ts config-catalog
/**
* Plugin config: the local executor's knobs plus the sandbox policy. All
* optional — `static Config` supplies the defaults (`mode: 'read-only'` is the
* fail-safe default; an example that wants a workspace-writable agent opts in
* explicitly). The runner choice is NOT configured here: which platform
* backend confines the command is the `ctx.sandbox` provider's config.
*/
export interface Config extends LocalConfig {
/** File-sandbox mode commands run under (default: `read-only`). */
mode?: SandboxMode
/**
* Root directory `workspace-write` mode may write under (default: the
* executor's default working directory — `cwd`, else `process.cwd()`).
*/
workspaceRoot?: string
}
```
Depends on: [`LocalConfig`](#deepseek-aidsh-bash-local) · [`SandboxMode`](core-data-structures/sandbox.md)
Source: [`packages/bash/bash-sandbox/src/index.ts:60`](../packages/bash/bash-sandbox/src/index.ts)
## `@deepseek-ai/dsh-code-runtime-worker`
```ts config-catalog
/** Plugin config: every execution cap, changeable from `cordis.yml` (no hardcoded tunables). */
export interface Config {
/**
* Busy-time budget in milliseconds: the run fails with kind `'timeout'`
* once the worker's MEASURED event-loop active time
* (`worker.performance.eventLoopUtilization()`) exceeds this. Metering
* measured busy time — not wall time, not host-side pending-call
* bookkeeping — is what makes the budget both fair (a program awaiting a
* slow tool accrues nothing) and ungameable (a hot loop accrues whether
* or not a decoy dispatch is in flight).
*/
computeMs?: number
/**
* Wall-clock ceiling in milliseconds; never pauses for anything. The
* backstop for what busy-time cannot see (a program awaiting a promise
* nobody will resolve).
*/
maxWallMs?: number
/** Shared byte budget for captured log text (console + raw stream writes), truncation marked in-band. */
maxLogBytes?: number
/**
* Byte cap for the completion value, measured by its real cross-boundary
* size (string bytes, or structured-clone wire size); an oversized or
* non-cloneable value crosses as a capped string rendering.
*/
maxValueBytes?: number
/** The worker's max old-generation heap in MiB (`resourceLimits`); overflow kills the worker, surfacing as kind `'worker-exit'`. */
maxOldGenerationSizeMb?: number
}
```
Source: [`packages/code-runtime/code-runtime-worker/src/index.ts:29`](../packages/code-runtime/code-runtime-worker/src/index.ts)
## `@deepseek-ai/dsh-compact-basic`
@@ -255,24 +328,6 @@ export interface Config {
Source: [`packages/hooks/hooks-codex/src/index.ts:43`](../packages/hooks/hooks-codex/src/index.ts)
## `@deepseek-ai/dsh-invariants`
Requires: `sessions`
```ts config-catalog
/** Plugin config. */
export interface Config {
/**
* Deep-freeze logged session-event data so mutating a logged event throws.
* Default true — this plugin only runs in dev/test, where freezing is the
* point. Set false to assert the event contract without freezing.
*/
freeze?: boolean
}
```
Source: [`packages/support/invariants/src/index.ts:45`](../packages/support/invariants/src/index.ts)
## `@deepseek-ai/dsh-llm-deepseek`
Requires: `llm`
@@ -404,6 +459,85 @@ export interface StreamableHttpConfig {
Source: [`packages/mcp/mcp-client/src/index.ts:91`](../packages/mcp/mcp-client/src/index.ts)
## `@deepseek-ai/dsh-repeat-tool-guard`
```ts config-catalog
/**
* Plugin config, validated by the same-named schemastery schema plus the
* load-time checks in `apply` (misconfiguration fails loud: an empty
* `thresholds` list, a non-integer, a value below 2, or a duplicate throws at
* plugin load, never a silent fall-back). `include`/`exclude` entries are
* `*`-wildcard predicates over tool names at call time, not references to
* registry entries — a pattern matching no currently registered tool is valid
* (`exclude: [mcp_*]` must stay legal in a deployment that loads no MCP tools).
*/
export interface Config {
/** Consecutive-repeat counts that trigger a reminder (default `[3, 5, 8]`). */
thresholds?: number[]
/** Tool-name patterns to track; empty means every tool is tracked. */
include?: string[]
/** Tool-name patterns transparent to the chain (neither count nor reset). */
exclude?: string[]
/**
* Maximum characters of canonical arguments quoted in the DETAILED reminder
* (default 500). Large payloads (a `write` body, a long command) would
* otherwise ride into the next request unbounded — precisely in a loop
* scenario; the cap bounds the reminder, never the detection (the chain key
* always compares the FULL canonical string).
*/
argumentsPreviewChars?: number
}
```
Source: [`packages/guard/repeat-tool-guard/src/index.ts:55`](../packages/guard/repeat-tool-guard/src/index.ts)
## `@deepseek-ai/dsh-sandbox-local`
```ts config-catalog
/** Plugin config. All optional — `static Config` supplies the defaults. */
export interface Config {
/**
* Override the sandbox runner argv (the bwrap-shaped profile arguments are
* appended). A NON-EMPTY argv is the operator's assertion that this runner
* exists and FULLY enforces the profile (confinement reports
* `enforcement: 'full'`, and — the runner's kernel mechanism being unknown
* — carries both Linux file-denial dialects as its denial signatures) —
* the runner chain and its probes are skipped,
* and a broken runner fails loudly at execution time. The operator also
* supplies {@link runnerFailureSignatures}, which distinguish the runner
* refusing its profile from the wrapped command failing normally.
* Absent (or empty — the schema normalizes an omitted array to `[]`): the
* built-in platform chains — Linux `bwrap` then the Landlock launcher
* (probed in that order), darwin `sandbox-exec` (the sole candidate,
* selected without a probe). Used for custom/alternative runners and
* for deterministic fake runners in keyless test tiers.
*/
runnerCommand?: string[]
/**
* Case-insensitive stderr substrings emitted when a configured
* {@link runnerCommand} refuses its profile before executing the wrapped
* command. Required and non-empty with `runnerCommand`; rejected without
* it. Missing/unexecutable runner errors are added automatically from
* `runnerCommand[0]`, while these signatures cover an executable runner's
* own failure dialect.
*/
runnerFailureSignatures?: string[]
/**
* Per-probe timeout in milliseconds for the chain's functional probes
* (default: 5000; must be a positive finite number — Node treats a 0
* `spawnSync` timeout as UNBOUNDED, so 0 is rejected at construction). A
* probe that exceeds it reads as an unusable rung, so a
* host slow enough to trip the default — cold NFS mounts, heavily loaded
* CI — would otherwise be misclassified `SANDBOX_UNAVAILABLE` with no
* config escape. Bounds ONE probe, and the chain walk runs each at most once
* per provider lifetime.
*/
probeTimeoutMs?: number
}
```
Source: [`packages/sandbox/sandbox-local/src/index.ts:36`](../packages/sandbox/sandbox-local/src/index.ts)
## `@deepseek-ai/dsh-session-persistence-jsonl`
Requires: `sessions`
@@ -457,6 +591,36 @@ export type JournalMode = 'wal' | 'delete' | 'truncate' | 'persist'
Source: [`packages/session-persistence/session-persistence-sqlite/src/index.ts:50`](../packages/session-persistence/session-persistence-sqlite/src/index.ts)
## `@deepseek-ai/dsh-skill`
```ts config-catalog
/** Skill registry configuration. */
export interface Config {
/** Maximum number of completed cwd/provider catalogs kept in memory. */
readonly collectCacheMaxEntries?: number
}
```
Source: [`packages/skill/skill/src/index.ts:113`](../packages/skill/skill/src/index.ts)
## `@deepseek-ai/dsh-skill-local`
Requires: `skills`
```ts config-catalog
/** Local filesystem skill provider configuration. */
export interface Config {
/** DeepSeek Harness config root. Defaults to `$DSH_HOME` or `~/.dsh`. */
dshHome?: string
/** Shared agent config root. Defaults to `$DSH_AGENTS_HOME` or `~/.agents`. */
agentsHome?: string
/** Additional skill roots scanned after project roots and before user roots. */
customSkillDirs?: string[]
}
```
Source: [`packages/skill/skill-local/src/index.ts:39`](../packages/skill/skill-local/src/index.ts)
## `@deepseek-ai/dsh-stdio-agent`
```ts config-catalog
@@ -466,7 +630,9 @@ Source: [`packages/session-persistence/session-persistence-sqlite/src/index.ts:5
* {@link @deepseek-ai/dsh-agent-core}'s forwarded `agents` list); `persona` is
* the deployment persona (forwarded to the system-prompt plugin); `toolOrder`
* is the explicit model-facing tool order (forwarded to the system-prompt plugin);
* `persistenceRoot` is the JSONL backend's directory; `welcome` is the UI banner.
* fresh sessions use `process.cwd()` as their workspace cwd; resumed sessions
* keep their persisted cwd. `persistenceRoot` is the JSONL backend's directory;
* `welcome` is the UI banner.
*/
export interface Config {
/** Model name for the `main` agent (must have a registered adapter). */
@@ -475,10 +641,14 @@ export interface Config {
persona?: string
/** Explicit model-facing tool order (the system-prompt plugin's `toolOrder` config; see dsh-system-prompt). */
toolOrder?: string[]
/** Tool-registry config — its presentation `mode` (forwarded through agent-core; see dsh-tools). */
tools?: ToolsConfig
/** Directory the JSONL session backend writes under. Defaults to `./.sessions`. */
persistenceRoot?: string
/** stdin-chat banner printed once on start. Defaults to `'ready.'`. */
welcome?: string
/** Skill registry, local-provider, and model-facing consumer config forwarded to agent-core. */
skills?: agentCore.SkillConfig
/**
* If set, the `main` agent RESUMES this persisted session id instead of
* starting fresh. Sourced from an env var in the leaf `cordis.yml`
@@ -488,7 +658,9 @@ export interface Config {
}
```
Source: [`packages/ui/stdio-agent/src/index.ts:60`](../packages/ui/stdio-agent/src/index.ts)
Depends on: [`agentCore`](../packages/core/agent-core/src/index.ts) · [`ToolsConfig`](#deepseek-aidsh-tools)
Source: [`packages/ui/stdio-agent/src/index.ts:65`](../packages/ui/stdio-agent/src/index.ts)
## `@deepseek-ai/dsh-subagent-acp`
@@ -548,7 +720,7 @@ Source: [`packages/subagent/subagent-acp/src/index.ts:30`](../packages/subagent/
## `@deepseek-ai/dsh-subagent-fork`
Requires: `subagents` · `agents`
Requires: `subagents`
```ts config-catalog
/** Config: the registry name to register the provider under. */
@@ -576,9 +748,11 @@ export interface Config {
/** Which start-time capabilities to advertise (default: all `true`). */
capabilities?: Partial<SubagentCapabilities>
/**
* The context contract to declare ({@link SubagentProvider.inheritsParentContext});
* default `false` (spawn-like). Set `true` to exercise the fork-shaped tool
* wording in consumer tests.
* The conversation-history descriptor to declare
* ({@link SubagentProvider.inheritsParentContext}); default `false` (fresh
* conversation). Set `true` to exercise seeded/fork wording in consumer
* tests. This flag says nothing about tool, service, scope, or authority
* inheritance.
*/
inheritsParentContext?: boolean
/**
@@ -591,11 +765,11 @@ export interface Config {
Depends on: [`SubagentCapabilities`](../packages/subagent/subagent/src/index.ts) · [`SubagentStopReason`](../packages/subagent/subagent/src/index.ts)
Source: [`packages/support/subagent-mock/src/index.ts:84`](../packages/support/subagent-mock/src/index.ts)
Source: [`packages/support/subagent-mock/src/index.ts:97`](../packages/support/subagent-mock/src/index.ts)
## `@deepseek-ai/dsh-subagent-spawn`
Requires: `subagents` · `agents`
Requires: `subagents`
```ts config-catalog
/** Config: the registry name to register the provider under. */
@@ -605,7 +779,7 @@ export interface Config {
}
```
Source: [`packages/subagent/subagent-spawn/src/index.ts:36`](../packages/subagent/subagent-spawn/src/index.ts)
Source: [`packages/subagent/subagent-spawn/src/index.ts:35`](../packages/subagent/subagent-spawn/src/index.ts)
## `@deepseek-ai/dsh-system-prompt`
@@ -616,7 +790,10 @@ export interface Config {
* The deployment's persona — the ONE deployment-authored fragment of the
* system prompt, rendered as the order-0 `deployment:persona` section
* (after the harness identity, before all tool guidance). Every agent in
* the context shares it, subagents included. Template, not free-form text:
* the context shares it by default; a per-agent persona is a SCOPED section
* of the same name registered through that agent's `agent.ctx` (it shadows
* this one for that agent — the subagent seam's `persona` request field does
* exactly that). Template, not free-form text:
* every complete `{{…}}` group is interpreted strictly against the
* registered prompt variables (the shipped agent loop registers `{{model}}`
* and `{{cwd}}`), and there is no escape syntax for literal `{{…}}` prose
@@ -651,7 +828,25 @@ export interface Config {
}
```
Source: [`packages/core/system-prompt/src/index.ts:179`](../packages/core/system-prompt/src/index.ts)
Source: [`packages/core/system-prompt/src/index.ts:227`](../packages/core/system-prompt/src/index.ts)
## `@deepseek-ai/dsh-tool-cordis`
Requires: `tools`
```ts config-catalog
/** Config for the tool-cordis plugin: the sandbox evaluation bound. */
export interface Config {
/**
* Milliseconds the SYNCHRONOUS portion of mount code may run in the vm
* before evaluation is aborted (default 5000). An async body escapes this
* bound — see docs/rfc/implemented/feature/2026-07-08-self-referential-cordis-toolset.md for the trust stance.
*/
vmTimeoutMs?: number
}
```
Source: [`packages/cordis/tool-cordis/src/index.ts:53`](../packages/cordis/tool-cordis/src/index.ts)
## `@deepseek-ai/dsh-tool-fs`
@@ -673,6 +868,20 @@ export interface Config {
Source: [`packages/fs/tool-fs/src/index.ts:48`](../packages/fs/tool-fs/src/index.ts)
## `@deepseek-ai/dsh-tool-skill`
Requires: `tools` · `skills`
```ts config-catalog
/** Model-facing skill catalog configuration. */
export interface Config {
/** Maximum normalized description length rendered in the session catalog; minimum 3. */
catalogDescriptionMaxLength?: number
}
```
Source: [`packages/skill/tool-skill/src/index.ts:19`](../packages/skill/tool-skill/src/index.ts)
## `@deepseek-ai/dsh-tool-subagent`
Requires: `tools` · `subagents`
@@ -692,24 +901,54 @@ export interface Config {
toolName?: string
/**
* Default per-child agent options (model) applied to every spawned child.
* Omitted fields fall back to the child loop's own defaults. There is no
* per-child persona: the deployment persona (the system-prompt plugin's
* `persona` config) is a context-wide section every agent shares.
* Omitted fields fall back to the child loop's own defaults.
*/
agentOptions?: AgentOptions
/**
* Per-child persona applied to every child this tool spawns: a scoped
* `deployment:persona` section shadowing the deployment's persona for the
* child alone. Requires the bound provider's `persona` capability
* (in-process backends support it; a request against one that doesn't is
* rejected at start). Omitted ⇒ the child renders the deployment persona.
*/
persona?: string
/**
* Tool scoping applied to every child this tool spawns (see
* `SubagentStartRequest.toolFilter`): the named global tools vanish from
* the child's prompt AND refuse to execute. Requires the provider's
* `toolFilter` capability. Unknown names fail the spawn loudly. Note the
* child otherwise sees every global tool — including this delegation tool
* itself; `deny`-listing it (or setting `maxDepth`) is how a deployment
* bounds recursion.
*/
toolFilter?: {
/** Global tool names the child keeps; everything else is removed. */
allow?: string[]
/** Global tool names removed from the child. */
deny?: string[]
}
/**
* Recursion cap applied to every child this tool spawns (see
* `SubagentStartRequest.maxDepth`): a spawn whose child would sit deeper
* than this in the delegation tree is rejected. Requires the provider's
* `depthLimit` capability. Must be a non-negative safe integer and is
* validated when the plugin loads. Omitted ⇒ unbounded (bound it in
* deployments that expose this tool to children).
*/
maxDepth?: number
}
```
Depends on: [`AgentOptions`](../packages/core/agent/src/index.ts)
Source: [`packages/subagent/tool-subagent/src/index.ts:44`](../packages/subagent/tool-subagent/src/index.ts)
Source: [`packages/subagent/tool-subagent/src/index.ts:47`](../packages/subagent/tool-subagent/src/index.ts)
## `@deepseek-ai/dsh-tool-web`
Requires: `tools` · `web` · `systemPrompt`
```ts config-catalog
/** Plugin config: which web tools to register, and the `web_search` source cap. */
/** Plugin config: which web tools to register, the source cap, and per-tool budgets. */
export interface Config {
/** Register `web_search`. Defaults to true. */
search?: boolean
@@ -717,10 +956,91 @@ export interface Config {
fetch?: boolean
/** Upper bound on sources returned by one `web_search` call. */
searchMaxResults?: number
/** Cooperative timeout budget (ms) for `web_fetch`. Defaults to 30000. */
fetchTimeoutMs?: number
/** Cooperative timeout budget (ms) for `web_search`. Defaults to 30000. */
searchTimeoutMs?: number
}
```
Source: [`packages/web/tool-web/src/index.ts:37`](../packages/web/tool-web/src/index.ts)
Source: [`packages/web/tool-web/src/index.ts:40`](../packages/web/tool-web/src/index.ts)
## `@deepseek-ai/dsh-tool-workflow`
Requires: `tools` · `workflows` · `systemPrompt`
```ts config-catalog
/** Config: the model-facing tool name plus result rendering caps. */
export interface Config {
/** The model-facing tool name to register (default `workflow`). */
toolName?: string
/** Rendered-result ceiling, in characters: a longer JSON value is truncated with a notice (default 50000). */
maxResultChars?: number
}
```
Source: [`packages/workflow/tool-workflow/src/index.ts:39`](../packages/workflow/tool-workflow/src/index.ts)
## `@deepseek-ai/dsh-tools`
Requires: `systemPrompt`
```ts config-catalog
/** Plugin config: how the registered tools are presented to the model. */
export interface Config {
/**
* The presentation mode. `'native'` (the default) contributes every
* visible end capability as a native wire function definition. Under
* `'code'` this registry contributes exactly ONE wire tool,
* `run_code`, plus the generated `tools:sdk` prompt section declaring every other tool as a
* TypeScript API the program calls. `'both'` contributes every native
* definition AND `run_code` + the SDK section. Non-native modes require a
* loaded `ctx.codeRuntime` whose `language` is `'typescript'` — a missing
* or mismatched runtime rejects every prompt assembly with an actionable
* error (misconfiguration fails loud, before any model request). A
* configured `systemPrompt.toolOrder` naming native tools likewise rejects
* every assembly under `'code'` (those names are no longer contributed) —
* a deployment switching modes updates its order config or drops it.
*/
mode?: ToolPresentationMode
}
/** How the registry presents its tools to the model (see {@link Config.mode}). */
export type ToolPresentationMode = 'native' | 'code' | 'both'
```
Source: [`packages/core/tools/src/index.ts:401`](../packages/core/tools/src/index.ts)
## `@deepseek-ai/dsh-user-approval`
```ts config-catalog
/** Plugin config. All optional — `static Config` supplies the defaults. */
export interface Config {
/**
* The deployment's default {@link ApprovalPolicy} for sessions without an
* `approval/policy` override — `'ask'` delegates to the composed answerers
* (fail-closed with none); `'never'` auto-rejects every ask without
* prompting (the deterministic CI/unattended stance).
*/
readonly policy?: ApprovalPolicy
}
/**
* A session's approval policy — what happens to an {@link ApprovalService}
* ask BEFORE any interactive answerer sees it:
*
* - `'ask'` (the default) — delegate to the composed answerers; with none
* composed the chain falls through to the fail-closed `'unavailable'`
* (exactly today's behavior).
* - `'never'` — never prompt anyone: every ask resolves `'rejected'`
* deterministically. The strict headless stance (CI, unattended runs) and
* the only policy value stated in the system prompt — unlike `'ask'`, its
* outcome is knowable without asking, so stating it cannot overclaim.
*/
export type ApprovalPolicy = 'ask' | 'never'
```
Source: [`packages/ui/user-approval/src/index.ts:270`](../packages/ui/user-approval/src/index.ts)
## `@deepseek-ai/dsh-web`
@@ -835,18 +1155,49 @@ export interface Config {
Source: [`packages/web/web-search-perplexity/src/index.ts:33`](../packages/web/web-search-perplexity/src/index.ts)
## `@deepseek-ai/dsh-workflow-workerthread`
Requires: `subagents`
```ts config-catalog
/** Plugin config (all optional — `static Config` supplies the defaults). */
export interface Config {
/** The `ctx.subagents` provider children run on (default `spawn`). */
provider?: string
/** Concurrent `agent()` ceiling; `0` (the default) auto-resolves to `min(16, max(1, cores - 2))`. */
maxConcurrentAgents?: number
/** Total `agent()` calls one run may start — the runaway-loop backstop (default 1000). */
maxTotalAgents?: number
/** Items accepted by a single `parallel()`/`pipeline()` call (default 4096). */
maxItemsPerCall?: number
/** vm timeout for the script's initial synchronous slice, inside the worker (default 5000 ms). */
syncTimeoutMs?: number
/**
* How long after a cancellation an unsettled script may keep running before
* the run force-settles `cancelled` and its worker is TERMINATED (default
* 5000 ms); also bounds `dispose()`.
*/
disposeGraceMs?: number
}
```
Source: [`packages/workflow/workflow-workerthread/src/index.ts:69`](../packages/workflow/workflow-workerthread/src/index.ts)
## Loadable plugins with no config
These load from a `cordis.yml` entry with no `config:` block; they declare no config surface.
- `@deepseek-ai/dsh-agent` ([`packages/core/agent/src/index.ts`](../packages/core/agent/src/index.ts))
- `@deepseek-ai/dsh-fs-policy` ([`packages/fs/fs-policy/src/index.ts`](../packages/fs/fs-policy/src/index.ts))
- `@deepseek-ai/dsh-invariants` — requires `sessions` ([`packages/support/invariants/src/index.ts`](../packages/support/invariants/src/index.ts))
- `@deepseek-ai/dsh-llm` ([`packages/llm/llm/src/index.ts`](../packages/llm/llm/src/index.ts))
- `@deepseek-ai/dsh-session` ([`packages/core/session/src/index.ts`](../packages/core/session/src/index.ts))
- `@deepseek-ai/dsh-subagent` ([`packages/subagent/subagent/src/index.ts`](../packages/subagent/subagent/src/index.ts))
- `@deepseek-ai/dsh-timeout-policy` — requires `tools` ([`packages/timeout/timeout-policy/src/index.ts`](../packages/timeout/timeout-policy/src/index.ts))
- `@deepseek-ai/dsh-tool-ask-user` — requires `tools` · `userInteraction` ([`packages/ui/tool-ask-user/src/index.ts`](../packages/ui/tool-ask-user/src/index.ts))
- `@deepseek-ai/dsh-tool-bash` — requires `tools` · `bash` · `systemPrompt` ([`packages/bash/tool-bash/src/index.ts`](../packages/bash/tool-bash/src/index.ts))
- `@deepseek-ai/dsh-tool-todo` — requires `tools` ([`packages/todo/tool-todo/src/index.ts`](../packages/todo/tool-todo/src/index.ts))
- `@deepseek-ai/dsh-tools` — requires `systemPrompt` ([`packages/core/tools/src/index.ts`](../packages/core/tools/src/index.ts))
- `@deepseek-ai/dsh-user-interaction` ([`packages/ui/user-interaction/src/index.ts`](../packages/ui/user-interaction/src/index.ts))
## Seam packages (not directly loadable)
@@ -856,7 +1207,9 @@ Abstract service classes — a deployment loads a concrete implementation packag
- `@deepseek-ai/dsh-code-runtime` — abstract `CodeRuntime` ([`packages/code-runtime/code-runtime/src/index.ts`](../packages/code-runtime/code-runtime/src/index.ts))
- `@deepseek-ai/dsh-compact` — abstract `CompactService` ([`packages/compact/compact/src/index.ts`](../packages/compact/compact/src/index.ts))
- `@deepseek-ai/dsh-fs` — abstract `FileSystem` ([`packages/fs/fs/src/index.ts`](../packages/fs/fs/src/index.ts))
- `@deepseek-ai/dsh-sandbox` — abstract `SandboxProvider` ([`packages/sandbox/sandbox/src/index.ts`](../packages/sandbox/sandbox/src/index.ts))
- `@deepseek-ai/dsh-session-persistence` — abstract `SessionPersistence` ([`packages/session-persistence/session-persistence/src/index.ts`](../packages/session-persistence/session-persistence/src/index.ts))
- `@deepseek-ai/dsh-workflow` — abstract `WorkflowService` ([`packages/workflow/workflow/src/index.ts`](../packages/workflow/workflow/src/index.ts))
## Library packages (no plugin entry)
@@ -866,4 +1219,7 @@ Imported as libraries by other packages; a `cordis.yml` cannot load them.
- `@deepseek-ai/dsh-app-boot` ([`packages/ui/app-boot/src/index.ts`](../packages/ui/app-boot/src/index.ts))
- `@deepseek-ai/dsh-brand` ([`packages/util/brand/src/index.ts`](../packages/util/brand/src/index.ts))
- `@deepseek-ai/dsh-hook-protocol` ([`packages/hooks/hook-protocol/src/index.ts`](../packages/hooks/hook-protocol/src/index.ts))
- `@deepseek-ai/dsh-scope` ([`packages/core/scope/src/index.ts`](../packages/core/scope/src/index.ts))
- `@deepseek-ai/dsh-subagent-inprocess` ([`packages/subagent/subagent-inprocess/src/index.ts`](../packages/subagent/subagent-inprocess/src/index.ts))
- `@deepseek-ai/dsh-subagent-subprocess` ([`packages/subagent/subagent-subprocess/src/index.ts`](../packages/subagent/subagent-subprocess/src/index.ts))
- `@deepseek-ai/dsh-timeout` ([`packages/util/timeout/src/index.ts`](../packages/util/timeout/src/index.ts))

14
docs/cookbook/adding-a-tool.md
View File

@@ -22,7 +22,7 @@ export function apply(ctx: Context) {
},
async execute(args, exec) {
// args is TYPED from the schema: { path: string; limit?: number }
// exec carries { callId, name, arguments, agent?, signal? }
// exec carries immutable identity + token; signal is the operational field
return [{ type: 'text', text: await readFile(args.path, 'utf8') }]
},
}))
@@ -34,7 +34,9 @@ Registration is effect-based: disposing the plugin fiber unregisters the tool (w
## Rules of the execute() contract
- **Args are validated for you.** `defineTool` validates the model-generated `arguments` against the `SchemaSpec` before `execute` runs (type, required keys, enum membership, nested objects/arrays — [runtime arg validation](../rfc/implemented/architecture/2026-06-11-runtime-arg-validation.md)), so inside `execute` the args already match `InferArgs`. You still hand-check value constraints the DSL can't express (non-empty strings, positive numbers, cross-field rules); throw a descriptive Error for those. Raw JSON-Schema tools registered directly (MCP) are NOT validated by the harness — they validate their own input.
- **Throwing means isError.** The registry catches anything `execute()` throws and returns `{isError: true}` to the model. Use that for infrastructure failures (bad input, spawn errors, aborts) — but REPORT domain failures in the result text instead (e.g. tool-bash returns `[exit code: 9]` with `isError: false`: the model decides what a failing command means).
- **Registration borrows your readonly definition.** A typed same-process contribution is not a serialization boundary; do not mutate its schema or replace callbacks after registration. `schemas()` materializes only the explicit model-facing projection. To hot-swap a tool, dispose its owning effect and register the replacement; mutable state inside the callback's closure remains ordinary plugin state.
- **Execution identity is protected.** The registry materializes `arguments` as detached lossless JSON in one recursive pass, freezes that value before policy starts, and assigns an opaque `exec.token`; `callId`, `name`, `arguments`, `agent`, `token`, and an optional enclosing-transport `parent` token stay immutable through dispatch. `parent` is identity-only and exposes no live outer execution. Treat `args` as readonly input. An around-dispatch wrapper may add, replace, or remove only `exec.signal` to impose cancellation or a deadline.
- **Throwing or returning non-JSON data means isError.** The registry catches anything `execute()` throws and materializes the complete post-policy result as lossless JSON before final observers run. A throw, malformed result, or non-JSON content/context/meta becomes `{isError: true}` so the live outcome cannot succeed and then fail at the durable log. Use errors for infrastructure failures (bad input, spawn errors, aborts), but report domain failures in the result text instead (for example, tool-bash returns `[exit code: 9]` with `isError: false` because the model decides what a failing command means).
- **Honor `exec.signal`.** Cancel in-flight work when it fires.
- **Attach durable card data with `meta` (optional).** `execute` may return `{ content, meta }` instead of a bare `ContentBlock[]` — `meta` is a JSON-serializable payload the core treats as opaque, persisted on the `tool/result` event and handed back to your `presentResult` (so a card that needs more than `args`, like `write`/`edit`'s applied-hunk diff, survives a session replay). Keep UI-only data here, never in the model-facing `content`.
- **Use `exec.agent` for async notifications.** `agent.inject(content, {source: {kind: 'plugin', plugin: '<name>'}})` appends durable context the NEXT model request sees — it is not a wake-up (an idle agent stays idle). Guard against disposed agents (try/catch).
@@ -45,9 +47,13 @@ Follow tool-bash's background pattern: a `run_in_background` flag returns a task
> TODO: each tool reimplements this background pattern by hand today. At some point we need a generic long-running-tool layer that handles task ids, incremental polling, kill, and completion notices uniformly.
## Permissions / sandboxing
## Execution policy and observation
Prefer not to build policy into the tool. The seam is the `tools/pre-execute` gate (deny/ask — see the permission-gate example in [extension-cookbook.md](./extension-cookbook.md)) and the `tools/post-execute` inspect/transform seam, or a sandboxing implementation behind the tool's executor seam.
Prefer not to build deployment policy into the tool. Use `tools/pre-execute` for extensible allow/deny/ask policy (the [permission-gate example](./extension-cookbook.md#a-hook-plugin-permission-gate)), `ctx.tools.guard()` for a final monotonic deny that later listeners cannot undo, `tools/execute` to wrap core dispatch with a deadline/retry/metrics scope, `tools/post-execute` to transform or attach model-facing context, and `tools/result` to observe the immutable normalized outcome without changing it. A sandboxing implementation can also sit behind the tool's executor capability seam; the exact contracts are in the [`dsh-tools` README](../../packages/core/tools/README.md#extension-points).
## Code Mode reaches your tool for free
Under the registry's non-native `mode` ([Code Mode](../../packages/core/tools/README.md)), each visible registered capability is callable from a `run_code` program as `await tools.<name>(args)` — nothing to add. The registry keeps `run_code` itself as reserved, unfilterable presentation infrastructure while restrictions still control which end capabilities appear in the scoped SDK and bindings. The generated SDK declares parameters from the same JSON Schema `defineTool` emits (constructs outside that subset degrade to `unknown`); each program call receives its own immutable execution whose `parent` is the enclosing `run_code` token, then re-enters the complete pre/guard/around/post/result pipeline. A failed call rejects the program-side promise with your error text. Design `description` and parameter `description`s as JSDoc a model reads while writing code, and remember that non-text result blocks reach programs as placeholders (text is the bridge's lingua franca).
## How your tool renders in an editor (ACP presentation)

21
docs/cookbook/extension-cookbook.md
View File

@@ -28,6 +28,8 @@ export function apply(ctx: Context) {
}
```
This waterfall is the reorderable policy layer. Use `ctx.tools.guard()` when an invariant needs a monotonic final denial, `tools/execute` when a plugin must wrap the actual dispatch lifetime (timeouts/retries/metrics; only `exec.signal` is replaceable), `tools/post-execute` for explicit result transformation, and `tools/result` for contained observation of the immutable final outcome. The [adding-a-tool guide](./adding-a-tool.md#execution-policy-and-observation) gives the selection rule.
## A UI plugin
A UI plugin renders from the `session/event` feed (the assistant token stream as `assistant/chunk`, plus turn/step boundaries and tool activity), and drives input back in via `agent.send()` / `agent.steer()`.
@@ -54,9 +56,9 @@ export function apply(ctx: Context) {
## A client-driver plugin (external protocol bridge)
A *client driver* is a UI plugin whose "user" is another program speaking a wire protocol rather than a human at a terminal. It owns the process's stdio (so it must run with **no stdout logger** — every non-protocol byte corrupts the stream), creates/resumes agents on demand through the `dsh-agent` factory seam, translates harness events (`session/event`, `agent/*`) into outbound protocol messages, and translates inbound requests back into `agent.send()` / `agent.cancel()`. Two harness-specific contracts make it correct: resolve each request exactly once off a settle signal (settle from the durable `turn/end` session event — the boundary is a session event, not an `agent/*` mirror — with `agent/status` as the fallback if a peer listener starved yours), and tear each agent down through its `AgentHandle.dispose()` (which stops the loop, `await`s its exit, and unregisters), not just `cancel()` — disposal must *reach* quiescence, not merely request it.
A *client driver* is a UI plugin whose "user" is another program speaking a wire protocol rather than a human at a terminal. It owns the process's stdio (so it must run with **no stdout logger** — every non-protocol byte corrupts the stream), creates/resumes agents on demand through the `dsh-agent` factory seam, translates harness events (`session/event`, `agent/*`) into outbound protocol messages, and translates inbound requests back into `agent.send()` / `agent.cancel()`. Two harness-specific contracts make it correct: correlate and settle each request exactly once from the durable `turn/end` session event even if rendering fails, and tear each agent down through its `AgentHandle.dispose()` (which stops the loop, `await`s its exit, and unregisters), not just `cancel()` — disposal must *reach* quiescence, not merely request it.
`packages/ui/acp` is the worked example: it bridges the agent to the Agent Client Protocol (JSON-RPC over stdio) so Zed and other ACP editors can drive it. See its README for the full method surface and the deferred-permission-gate note.
`packages/ui/acp` is the worked example: it bridges the agent to the Agent Client Protocol (JSON-RPC over stdio) so Zed and other ACP editors can drive it. See its README for the full method surface and the permission-prompt answerer it registers on the approval seam.
```ts
import type { Context } from 'cordis'
@@ -87,21 +89,26 @@ Three complete examples load their plugin trees from `cordis.yml`: [`examples/ec
Every product feature maps to a listener on a documented extension seam — the microkernel claim made checkable ([microkernel RFC](../rfc/implemented/architecture/2026-06-11-microkernel-event-taxonomy.md)). No row modifies the loop.
`system-prompt/assemble` is an expert cooperative whole-assembly transform: its returned assembly is authoritative, so listener authors own preserving active Code Mode and structured-output protocol contributions. Prefer `ctx.tools.restrict()` for tool filtering that must stay aligned across presentation, lookup, and execution.
| Product feature | Plugin mechanism |
|---|---|
| Hook system (user + project level) | listeners on `agent/session-start`, `agent/prompt-submit`, `agent/request`, `agent/step-result`, `tools/pre-execute`, `tools/post-execute`, `agent/turn-continuation` — each interception waterfall returns a typed Decision; the `dsh-hooks-claude` / `dsh-hooks-codex` bridges map hook config files onto these seams |
| `/goal` | force-continue via `agent/turn-continuation` + `steer()` reminders |
| `/loop` | on the `turn/end` session event, `send()` the next iteration; or force-continue |
| Dynamic workflow | orchestrator plugin on `turn/end` (or `step/end`) driving `send`/`steer` + subagents |
| Dynamic workflow | `ctx.workflows` + the worker-thread engine + the `workflow` tool; structured in-process children enforce output with scoped prompt/tool registrations, a monotonic tool guard, final `tools/result` commit (including enclosing `run_code`), and terminal `agent/turn-stop` |
| Queued + steering messages | core `Agent.send()` / `Agent.steer()` |
| Context compaction (auto + manual) | the `ctx.compact` seam + a backend (`dsh-compact-basic`) on the serial `agent/pre-step` seam; auto = token-pressure check before each step; a manual trigger invokes the same `ctx.compact` routine ([compaction RFC](../rfc/implemented/feature/2026-06-18-compaction-capability-seam.md) — the model-facing `/compact` consumer tool is deferred) |
| System prompt configurability | `ctx.systemPrompt.section()` with ordering |
| System prompt configurability | `ctx.systemPrompt.section()` with ordering and scope-local shadowing |
| AGENTS.md (root) | a section provider reading the file |
| AGENTS.md (subdir, on-touch) + file-change notices | `agent.inject()` from a watcher / tool-result listener |
| Built-in tools | `ctx.tools.register()`; schemas flow into the assembly automatically — the `dsh-tool-*` families (bash, fs, web, subagent, todo) are the shipped examples |
| ToolSearch / progressive disclosure | filter tools at `system-prompt/assemble` (the assembly carries the schemas; the loop logs the result as the request header, so disclosure stays reconstructable) |
| Tool sandbox (landlock / sandbox-exec) | `tools/pre-execute` (deny), or a sandboxing `BashExecutor` on the `dsh-bash` seam |
| Permission system / AskUserQuestion | `tools/pre-execute` (deny/ask); register an ask tool |
| ToolSearch / progressive disclosure | replace a scoped `ctx.tools.restrict()` registration as the visible set changes; the registry keeps presentation, lookup, and execution aligned |
| Tool deadline / retry / metrics | wrap core dispatch with `tools/execute`; a wrapper may replace `exec.signal`, delegate, and inspect the normalized result in one lexical lifetime |
| Final tool-result metrics / audit / capture | observe immutable authoritative outcomes with `tools/result`; use `tools/post-execute` instead only when the plugin must transform the result or attach context |
| Monotonic terminal turn policy | return `{ action: 'stop' }` from serial `agent/turn-stop`, after continuation and steering have already been folded |
| Subprocess sandbox (landlock / sandbox-exec) | use a `ctx.sandbox` backend through `dsh-bash-sandbox`; use `tools/pre-execute` for capability-level denial |
| Permission system / AskUserQuestion | return `ask` from `tools/pre-execute` and answer through `ctx.approval`; register a separate model-facing ask tool for ordinary user questions |
| Plan mode | `tools/pre-execute` (deny writes) + a mode prompt section via `ctx.systemPrompt.section()` or `agent.inject()` (model-visible ⟺ logged: `agent/request` shapes call config only) |
| Sub-agent delegation | the `ctx.subagents` provider registry (`dsh-subagent-spawn`/`-fork`/`-acp`) + `dsh-tool-subagent` exposing one configured provider to the model |
| MCP | one plugin per server: discover tools → `ctx.tools.register()` |

288
docs/cordis-catalog/events.md
View File

@@ -15,137 +15,179 @@ Dispatch modes: **emit** (fire-and-forget), **waterfall** (each listener gets `n
### `agent/created` — emit
An agent was registered in the AgentRegistry and is ready to receive messages.
An agent's fully composed scoped world was published in the AgentRegistry. Its session is already live in the session store. Setup is composition-only by contract; the subsequent `agent/session-start` boundary is the first supported place to inject or queue startup work. A synchronous listener throw vetoes publication and rollback emits the matching disposal edges; returned-promise rejection is observed and logged but cannot retroactively veto this synchronous boundary. A synchronous listener that requests the advanced registry detach does not remove the entry immediately: removal and the paired `agent/disposed` edge wait until the creation dispatch unwinds, so no later creation listener observes a disposal that preceded its own creation callback.
```ts cordis-catalog
'agent/created'(agent: Agent): void
'agent/created'(this: Scoped<Agent>, agent: Agent): void
```
Types: [Agent](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:264`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:316`](../../packages/core/agent/src/types.ts)
### `agent/disposed` — emit
An agent was disposed and removed from the registry; its fiber and any in-flight turn have been torn down.
An agent was removed from the registry. The concrete AgentLoop lifecycle emits this only after its driver and any in-flight turn reach quiescence; a custom agent registered through the public registry owns its own driver contract, which the registry cannot infer. Ordered teardown may still be detaching the session and unwinding scoped registrations when this runs.
```ts cordis-catalog
'agent/disposed'(agent: Agent): void
'agent/disposed'(this: Scoped<Agent>, agent: Agent): void
```
Types: [Agent](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:271`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:331`](../../packages/core/agent/src/types.ts)
### `agent/error` — emit
A step or turn errored. The loop reports a failure here (plus the logger) even when the error has no in-turn position for a session `error` event.
```ts cordis-catalog
'agent/error'(agent: Agent, turn: number, step: number, error: Error): void
'agent/error'(this: Scoped<Agent>, agent: Agent, turn: number, step: number, error: Error): void
```
Types: [Agent](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:420`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:605`](../../packages/core/agent/src/types.ts)
### `agent/pre-step` — serial
Awaited pre-step surface-mutation checkpoint, fired once per step AFTER `turn/start` (and after the prior step closed) but BEFORE this step's `step/start` — so anything a listener appends lands OUTSIDE the step, between `turn/start`/`step/end` and the upcoming `step/start`. `step` is the number of the step about to start. The loop awaits `ctx.serial('agent/pre-step', …)` after assembling the system prompt, then opens the step and derives the request history ONCE from whatever the surface now holds. This is where compaction belongs: it mutates the session surface in place (shadowing an older range with a summary node) with its log-only `compact/*` records cleanly outside any step, and the single subsequent derive reflects the mutation — so there is no double-derive and no listener can see (or be expected to act on) an assembled `messages` array that does not exist yet.
Serial (awaited in registration order), not a waterfall: a listener mutates the surface as a side effect; there is nothing to transform, but the loop must wait for the mutation to complete before opening the step and deriving. Cordis `serial` bails early if a listener returns a bail value; this event is typed and documented as `void`, so listeners must not return a semantic veto value. `fullSystemPrompt` is the assembled prompt a listener needs to measure pressure (the system prompt counts toward the budget). `signal` cancels any in-flight work a listener starts (e.g. a summarization model call).
Serial (awaited in registration order), not a waterfall: a listener mutates the surface as a side effect; there is nothing to transform, but the loop must wait for the mutation to complete before opening the step and deriving. Cordis `serial` bails early if a listener returns a bail value; this event is typed and documented as `void`, so listeners must not return a semantic veto value. `fullSystemPrompt` is the assembled prompt a listener needs to measure pressure (the system prompt counts toward the budget), and `sessionPrefix` is the instance's composed agent/session-prefix product for the same reason — every request carries it in front of the derived history, and it is composed BEFORE this seam fires precisely so a pressure gate counts the prefix the request will actually send (never a stale logged one). `signal` cancels any in-flight work a listener starts (e.g. a summarization model call). Scope-filtered dispatch (`@deepseek-ai/dsh-scope`): a listener registered through `agent.ctx` fires only for that agent's dispatches; a listener on a plain plugin context fires for every agent. The dispatch `this` is the scope carrier (`Scoped<Agent>`), built by the emitting side via `scopeTarget`/`agentEvents`.
```ts cordis-catalog
'agent/pre-step'(agent: Agent, turn: number, step: number, fullSystemPrompt: string, signal: AbortSignal): Promise<void> | void
'agent/pre-step'(this: Scoped<Agent>, agent: Agent, turn: number, step: number, fullSystemPrompt: string, sessionPrefix: readonly Message[], signal: AbortSignal): Promise<void> | void
```
Types: [Agent](../core-data-structures/core.md)
Types: [Agent](../core-data-structures/core.md) · [Message](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:349`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:438`](../../packages/core/agent/src/types.ts)
### `agent/prompt-submit` — waterfall
Waterfall: decide what happens to ONE drained queued message before it becomes a `user/message` — allow (optionally rewriting the prompt bytes or attaching `additionalContext`) or block it. Fires inside the already-open turn, per drained message. Maps onto Claude Code's `UserPromptSubmit` hook. Call `next()` to delegate to the default (allow unchanged), or return a PromptDecision without calling `next()` to short-circuit.
```ts cordis-catalog
'agent/prompt-submit'(agent: Agent, content: ContentBlock[], source: MessageSource, next: () => Promise<PromptDecision>): Promise<PromptDecision>
'agent/prompt-submit'(this: Scoped<Agent>, agent: Agent, content: ContentBlock[], source: MessageSource, next: () => Promise<PromptDecision>): Promise<PromptDecision>
```
Types: [Agent](../core-data-structures/core.md) · [ContentBlock](../core-data-structures/core.md) · [MessageSource](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:362`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:456`](../../packages/core/agent/src/types.ts)
### `agent/queued` — emit
A message entered the agent's inbox (queued or steering). `source` is the resolved source (defaults applied), not the caller's raw options.
A message entered the agent's inbox (queued or steering). Content and the resolved source are the detached, deeply-frozen values retained by the inbox. `source` has defaults applied and is not the caller's raw options.
```ts cordis-catalog
'agent/queued'(agent: Agent, content: ContentBlock[], info: { source: MessageSource; steering: boolean }): void
'agent/queued'(this: Scoped<Agent>, agent: Agent, content: ContentBlock[], info: { source: MessageSource; steering: boolean }): void
```
Types: [Agent](../core-data-structures/core.md) · [ContentBlock](../core-data-structures/core.md) · [MessageSource](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:289`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:360`](../../packages/core/agent/src/types.ts)
### `agent/request` — waterfall
Waterfall: shape the step's call configuration — model switching, sampling overrides — by returning a replacement LlmCallConfig (the frozen seed is the config the loop would otherwise use). Config is ALL a listener shapes here: every request is a pure function of the session log (the reconstructability RFC), so model-visible content flows through the log channels — `inject()`, steering, prompt-submit `additionalContext`, prompt sections via `system-prompt/assemble` — never through request mutation, and the loop records whatever config the request actually uses as a `request/header*` event before dispatch. The step's messages are already snapshotted when this fires (the `step/start` boundary): an `inject()` from a listener here lands in the log but joins the NEXT request. For surface mutation that must precede the snapshot (compaction), use agent/pre-step. Call `next()` to delegate, or return an LlmCallConfig without it to short-circuit.
Waterfall: shape the step's call configuration — model switching, sampling overrides — by returning a replacement LlmCallConfig (the frozen seed is the config the loop would otherwise use). Config is ALL a listener shapes here: every request is a pure function of the session log (the reconstructability RFC), so model-visible content flows through the log channels — `inject()`, steering, prompt-submit `additionalContext`, prompt sections via `system-prompt/assemble`, or the header-logged session prefix via agent/session-prefix — never through request mutation, and the loop records whatever config the request actually uses as a `request/header*` event before dispatch. The step's messages are already snapshotted when this fires (the `step/start` boundary): an `inject()` from a listener here lands in the log but joins the NEXT request. For surface mutation that must precede the snapshot (compaction), use agent/pre-step. Call `next()` to delegate, or return an LlmCallConfig without it to short-circuit.
```ts cordis-catalog
'agent/request'(agent: Agent, turn: number, step: number, config: LlmCallConfig, next: () => Promise<LlmCallConfig>): Promise<LlmCallConfig>
'agent/request'(this: Scoped<Agent>, agent: Agent, turn: number, step: number, config: LlmCallConfig, next: () => Promise<LlmCallConfig>): Promise<LlmCallConfig>
```
Types: [Agent](../core-data-structures/core.md) · [LlmCallConfig](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:385`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:485`](../../packages/core/agent/src/types.ts)
### `agent/session-prefix` — waterfall
Waterfall: compose the SESSION PREFIX — request-only messages placed in front of the ENTIRE derived history (directly after the provider's system slot) on every request this loop instance sends. Fired ONCE per loop instance, lazily before its first step's agent/pre-step seam — BEFORE the pre-step so a token-pressure gate (compaction) counts the prefix this instance will actually send, never a previous instance's logged one. The composed result is deep-frozen, recorded as `EpochHeader.messagePrefix` on the instance's anchoring `'initial'`/`'resume'` header snapshot, and reused verbatim for every subsequent request — never recomputed mid-session, so the provider prefix cache holds by construction (a process restart or `ctx.agents.resume()` is a new instance: it recomposes, and any drift lands attributably on the `'resume'` snapshot). Composition runs outside the step, before the boundary snapshot: a composing listener's session append joins the CURRENT request's derived history. A composition interrupted by a cancel/dispose landing inside the waterfall is discarded — never cached, logged, or sent — and the next turn recomposes under a live signal, so an abort-aware listener's degraded fallback cannot leak into later requests.
This is the home for session-stable openers the model must always see but that must NOT become durable history — a skills catalog, an AGENTS.md digest, a workspace baseline: `Session.deriveMessages()` never returns the prefix, and the header events are its only durable record, so the request stays reconstructable from the log. Content that CHANGES mid-session belongs in the append-only history channels instead — `agent.inject()`, a `tools/post-execute` decision's `additionalContext`, prompt-submit `additionalContext` — each a durable `context/message` paid once and prefix-cached thereafter.
The seed is a frozen empty list; a contributing listener returns a NEW array — never an in-place push. The canonical contribution is a PREPEND, `[mine, ...await next()]`: the waterfall unwinds innermost-first (the LAST-registered listener's `next()` resolves first), so prepending yields registration order on the wire, and every plugin using it composes deterministically. The append form `[...await next(), mine]` is legal but places a contribution AFTER every later-registered plugin's — reverse registration order when all contributors append. Call `next()` to delegate, or return a list without it to short-circuit. Scope-filtered dispatch (`@deepseek-ai/dsh-scope`): a listener registered through `agent.ctx` fires only for that agent's dispatches; a listener on a plain plugin context fires for every agent. The dispatch `this` is the scope carrier (`Scoped<Agent>`), built by the emitting side via `scopeTarget`/`agentEvents`.
```ts cordis-catalog
'agent/session-prefix'(this: Scoped<Agent>, agent: Agent, prefix: Message[], signal: AbortSignal, next: () => Promise<Message[]>): Promise<Message[]>
```
Types: [Agent](../core-data-structures/core.md) · [Message](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:537`](../../packages/core/agent/src/types.ts)
### `agent/session-start` — emit
The agent's session lifecycle began, fired once before its first turn. `source` says why (SessionStartSource: fresh startup, a resumed persisted session, …). A pure NOTIFICATION (emit, not waterfall): it carries no veto — a session-start listener that wants to seed context does so via `agent.inject()` (a `context/message` the first request sees), not by returning a decision. Cannot block the session from starting; that gap is deliberate (a bridge logs/injects, it does not gate startup).
The agent's session lifecycle began, fired once before its first turn. `source` says why (SessionStartSource: fresh startup, a resumed persisted session, …). A pure NOTIFICATION (emit, not waterfall): a listener cannot veto by returning a decision or throwing. A listener that wants to seed context does so via `agent.inject()` (a `context/message` the first request sees). A lifecycle owner can still dispose its structural ownership edge during this notification; publication rechecks liveness and then aborts before the driver starts.
```ts cordis-catalog
'agent/session-start'(agent: Agent, source: SessionStartSource): void
'agent/session-start'(this: Scoped<Agent>, agent: Agent, source: SessionStartSource): void
```
Types: [Agent](../core-data-structures/core.md)
Types: [Agent](../core-data-structures/core.md) · [SessionStartSource](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:304`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:381`](../../packages/core/agent/src/types.ts)
### `agent/status` — emit
Agent status changed (`idle` ⇄ `running`, or → `disposed`). Drive lifecycle off this transition, never off a status you just requested — `send()` does not flip status to `running` before it returns.
```ts cordis-catalog
'agent/status'(agent: Agent, status: AgentStatus): void
'agent/status'(this: Scoped<Agent>, agent: Agent, status: AgentStatus): void
```
Types: [Agent](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:280`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:345`](../../packages/core/agent/src/types.ts)
### `agent/step-result` — waterfall
Waterfall: post-process the assembled assistant Message before tool dispatch (validation, content rewriting, …).
```ts cordis-catalog
'agent/step-result'(agent: Agent, turn: number, step: number, message: Message, next: () => Promise<Message>): Promise<Message>
'agent/step-result'(this: Scoped<Agent>, agent: Agent, turn: number, step: number, message: Message, next: () => Promise<Message>): Promise<Message>
```
Types: [Agent](../core-data-structures/core.md) · [Message](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:395`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:552`](../../packages/core/agent/src/types.ts)
### `agent/turn-continuation` — waterfall
Waterfall: override the turn-continuation decision via a typed ContinuationDecision. The loop's `defaultDecision` is `continue` when the step had tool calls or steering was injected, else `stop`. Listeners force-continue (`/goal`, `/loop` — optionally attaching a `reason` recorded as next-step steering) or force-stop (budget guards). Call `next()` to delegate to the default, or return a decision to override.
```ts cordis-catalog
'agent/turn-continuation'(agent: Agent, turn: number, defaultDecision: ContinuationDecision, next: () => Promise<ContinuationDecision>): Promise<ContinuationDecision>
'agent/turn-continuation'(this: Scoped<Agent>, agent: Agent, turn: number, defaultDecision: ContinuationDecision, next: () => Promise<ContinuationDecision>): Promise<ContinuationDecision>
```
Types: [Agent](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:408`](../../packages/core/agent/src/types.ts)
Source: [`packages/core/agent/src/types.ts:570`](../../packages/core/agent/src/types.ts)
### `agent/turn-stop` — serial
Serial terminal-stop checkpoint after the ordinary `agent/turn-continuation` waterfall, any `continue.reason`, and the pending-steering continuation override have been folded. A listener returns `{ action: 'stop' }` to make this turn terminal, or `undefined` to abstain. Terminal stop is monotonic: listener order and steering cannot resume the turn, and pending steering is discarded rather than becoming another step or turn.
```ts cordis-catalog
'agent/turn-stop'(this: Scoped<Agent>, agent: Agent, turn: number): ContinuationStop | undefined
```
Types: [Agent](../core-data-structures/core.md)
Source: [`packages/core/agent/src/types.ts:588`](../../packages/core/agent/src/types.ts)
## `approval/*`
### `approval/request` — waterfall
Waterfall asking the composed answerers to decide one approval request. Dispatched only from ApprovalService.request — callers go through the service (which owns cancellation and the audit events), never through `ctx.waterfall` directly. A listener that can answer for this request's agent returns an outcome WITHOUT calling `next()` (the decision slot is single-occupancy, first listener to answer wins); a listener that does not recognize the agent MUST call `next()` so another answerer — or the fail-closed default `'unavailable'` — gets the question. Throwing is contained by the service and yields `'unavailable'`. Scope-filtered dispatch (`@deepseek-ai/dsh-scope`) keys the carrier by `req.agent`: a listener registered through `agent.ctx` receives only that agent's questions, while a plain-context listener receives every agent's. `req` is a readonly same-process value borrowed from the caller.
```ts cordis-catalog
'approval/request'(this: Scoped<ApprovalService>, req: ApprovalRequest, next: () => Promise<ApprovalOutcome>): Promise<ApprovalOutcome>
```
Types: [ApprovalOutcome](../core-data-structures/approval.md) · [ApprovalRequest](../core-data-structures/approval.md)
Source: [`packages/ui/user-approval/src/index.ts:70`](../../packages/ui/user-approval/src/index.ts)
## `fs/*`
@@ -203,77 +245,109 @@ Source: [`packages/llm/llm/src/index.ts:39`](../../packages/llm/llm/src/index.ts
### `session/created` — emit
A session was created in the store.
A session was created in the store. A synchronous listener throw vetoes publication and rollback emits the matching `session/disposed` edge; returned-promise rejection is observed and logged but cannot retroactively veto this synchronous boundary. A synchronous listener that requests the advanced detach does not remove the entry immediately: removal and the paired `session/disposed` edge wait until the creation dispatch unwinds. Scope-filtered dispatch (`@deepseek-ai/dsh-scope`): the carrier is the session's owner scope, captured when the session was ENTERED (an agent's session is entered through `agent.ctx`, so its events dispatch in that agent's scope; a bare `sessions.create()` from a plain plugin dispatches subject-less). A listener registered through `agent.ctx` hears only that agent's sessions; a plain plugin listener hears every session.
```ts cordis-catalog
'session/created'(session: Session): void
'session/created'(this: Scoped<Session>, session: Session): void
```
Source: [`packages/core/session/src/index.ts:39`](../../packages/core/session/src/index.ts)
Source: [`packages/core/session/src/index.ts:52`](../../packages/core/session/src/index.ts)
### `session/disposed` — emit
A previously announced session left the store. Emitted exactly once on normal detach or publication rollback, and never for a prepared/entered session whose `session/created` announcement did not begin. Listener failures (including returned-promise rejections) are logged and contained per listener so teardown always reaches quiescence. Scope-filtered dispatch uses the same owner carrier captured at entry; agent-scoped listeners hear only their own session's teardown.
```ts cordis-catalog
'session/disposed'(this: Scoped<Session>, session: Session): void
```
Source: [`packages/core/session/src/index.ts:64`](../../packages/core/session/src/index.ts)
### `session/event` — emit
An event was appended to a session log (sync, fire-and-forget). This is the per-append feed a UI or invariant plugin tails.
An event was appended to a session log (sync, fire-and-forget). This is the per-append feed a UI or invariant plugin tails. The log push is the commit point; synchronous throws and returned-promise rejections from observers are logged and contained per listener, so they cannot make a committed append appear to fail or starve later listeners. The exact callback list and Cordis internal-dispatch checks resolve before the push; callbacks themselves run only after it. Scope-filtered dispatch (`@deepseek-ai/dsh-scope`): the carrier is the session's owner scope, captured when the session was ENTERED (an agent's session is entered through `agent.ctx`, so its events dispatch in that agent's scope; a bare `sessions.create()` from a plain plugin dispatches subject-less). A listener registered through `agent.ctx` hears only that agent's sessions; a plain plugin listener hears every session.
```ts cordis-catalog
'session/event'(session: Session, event: SessionEvent): void
'session/event'(this: Scoped<Session>, session: Session, event: SessionEvent): void
```
Types: [SessionEvent](../core-data-structures/core.md)
Source: [`packages/core/session/src/index.ts:47`](../../packages/core/session/src/index.ts)
Source: [`packages/core/session/src/index.ts:83`](../../packages/core/session/src/index.ts)
### `session/flush` — parallel
Awaited durability checkpoint. The agent loop awaits `ctx.parallel('session/flush', session)` at every turn end; persistence plugins (JSONL, SQLite) drain their write-behind buffers here and on fiber dispose. Awaited (parallel), not a waterfall: every listener runs and the loop waits for all of them, but none can veto.
Awaited durability checkpoint. The agent loop awaits `ctx.sessions.flush(session)` at every turn end; persistence plugins (JSONL, SQLite) drain their write-behind buffers here and on fiber dispose. Awaited (parallel), not a waterfall: every listener runs and the caller waits for all of them, but none can veto. Dispatch it through SessionStore.flush — the store owns the carrier — never via a raw `ctx.parallel`. Scope-filtered dispatch (`@deepseek-ai/dsh-scope`): the carrier is the session's owner scope, captured when the session was ENTERED (an agent's session is entered through `agent.ctx`, so its events dispatch in that agent's scope; a bare `sessions.create()` from a plain plugin dispatches subject-less). A listener registered through `agent.ctx` hears only that agent's sessions; a plain plugin listener hears every session.
```ts cordis-catalog
'session/flush'(session: Session): Promise<void> | void
'session/flush'(this: Scoped<Session>, session: Session): Promise<void> | void
```
Source: [`packages/core/session/src/index.ts:57`](../../packages/core/session/src/index.ts)
Source: [`packages/core/session/src/index.ts:101`](../../packages/core/session/src/index.ts)
## `skill/*`
### `skill/provider-added` — emit
A skill provider became resolvable in the `ctx.skills` registry. Consumers can observe this instead of depending on Cordis plugin load order, which is concurrent for sibling plugins.
```ts cordis-catalog
'skill/provider-added'(provider: SkillProvider): void
```
Source: [`packages/skill/skill/src/index.ts:131`](../../packages/skill/skill/src/index.ts)
### `skill/provider-removed` — emit
A skill provider left the registry because its plugin fiber was disposed.
```ts cordis-catalog
'skill/provider-removed'(name: string): void
```
Source: [`packages/skill/skill/src/index.ts:137`](../../packages/skill/skill/src/index.ts)
## `subagent/*`
### `subagent/end` — emit
A subagent run settled — emitted when SubagentRun.result resolves (any stop reason). Paired with Events['subagent/start'].
A ready child settled. Scope-filtered dispatch uses the same delegating parent carrier as `subagent/start`, so the lifecycle pair reaches the same scoped audience.
```ts cordis-catalog
'subagent/end'(info: SubagentRunEndInfo): void
'subagent/end'(this: Scoped<SubagentService>, info: SubagentRunEndInfo): void
```
Source: [`packages/subagent/subagent/src/index.ts:98`](../../packages/subagent/subagent/src/index.ts)
Source: [`packages/subagent/subagent/src/index.ts:90`](../../packages/subagent/subagent/src/index.ts)
### `subagent/provider-added` — emit
A provider became resolvable in the SubagentService registry. Consumers that derive state from a named provider (e.g. the model-facing tool wording in `dsh-tool-subagent`) react HERE instead of assuming load order — the cordis Loader starts sibling plugins concurrently, so "listed earlier in cordis.yml" does not mean "registered earlier".
A provider became resolvable in the registry.
```ts cordis-catalog
'subagent/provider-added'(provider: SubagentProvider): void
```
Source: [`packages/subagent/subagent/src/index.ts:72`](../../packages/subagent/subagent/src/index.ts)
Source: [`packages/subagent/subagent/src/index.ts:66`](../../packages/subagent/subagent/src/index.ts)
### `subagent/provider-removed` — emit
A provider left the registry (its plugin's fiber was disposed — an unload or an HMR reload). Consumers holding provider-derived state drop it here; a reload re-fires `subagent/provider-added` with the fresh provider. Delivered with per-listener containment: a throwing subscriber is logged, never starves later subscribers, and never disrupts the provider's teardown.
A provider left the registry. Accepted runs remain holder-owned.
```ts cordis-catalog
'subagent/provider-removed'(name: string): void
```
Source: [`packages/subagent/subagent/src/index.ts:83`](../../packages/subagent/subagent/src/index.ts)
Source: [`packages/subagent/subagent/src/index.ts:72`](../../packages/subagent/subagent/src/index.ts)
### `subagent/start` — emit
A subagent run started — emitted after the provider is resolved and its capabilities validated, as the child run begins. Paired with Events['subagent/end'].
A provider established a ready child. For in-process providers, `ctx.agents.get(info.id)` resolves during this notification. Scope-filtered dispatch keys the carrier by the delegating parent, so a parent-scoped listener observes only its own delegations. Paired with `subagent/end`.
```ts cordis-catalog
'subagent/start'(info: SubagentRunInfo): void
'subagent/start'(this: Scoped<SubagentService>, info: SubagentRunInfo): void
```
Source: [`packages/subagent/subagent/src/index.ts:91`](../../packages/subagent/subagent/src/index.ts)
Source: [`packages/subagent/subagent/src/index.ts:82`](../../packages/subagent/subagent/src/index.ts)
## `system-prompt/*`
@@ -281,57 +355,147 @@ Source: [`packages/subagent/subagent/src/index.ts:91`](../../packages/subagent/s
Waterfall around prompt assembly — mutate or extend the PromptAssembly (sections + tools + variables) before it is rendered. Bound to the SystemPrompt service; call `next()` to delegate.
Scope-filtered dispatch (`@deepseek-ai/dsh-scope`): the carrier is keyed by `context.scope` — a listener registered through `agent.ctx` fires only for that agent's assemblies; a plain plugin listener fires for every assembly (scope-less ones included, dispatched subject-less).
The returned assembly is authoritative. This is an expert composition seam: a listener that removes or replaces another plugin's protocol contribution owns preserving that protocol's invariants.
```ts cordis-catalog
'system-prompt/assemble'(this: SystemPrompt, assembly: PromptAssembly, context: AssembleContext, next: () => Promise<PromptAssembly>): Promise<PromptAssembly>
'system-prompt/assemble'(this: Scoped<SystemPrompt>, assembly: PromptAssembly, context: AssembleContext, next: () => Promise<PromptAssembly>): Promise<PromptAssembly>
```
Source: [`packages/core/system-prompt/src/index.ts:38`](../../packages/core/system-prompt/src/index.ts)
Source: [`packages/core/system-prompt/src/index.ts:49`](../../packages/core/system-prompt/src/index.ts)
### `system-prompt/change` — emit
A section, tool provider, or variable provider was registered or unregistered (the assembly inputs changed).
A section, tool provider, or variable provider was registered or unregistered (the assembly inputs changed — possibly for one scope only). An UNFILTERED registry-subject notification, deliberately not scope-filtered dispatch: a global change concerns every agent's next assembly, so a scoped listener subscribing here sees every change, not just its own scope's.
```ts cordis-catalog
'system-prompt/change'(): void
```
Source: [`packages/core/system-prompt/src/index.ts:44`](../../packages/core/system-prompt/src/index.ts)
Source: [`packages/core/system-prompt/src/index.ts:59`](../../packages/core/system-prompt/src/index.ts)
## `tools/*`
### `tools/change` — emit
A tool was registered or unregistered (the available tool set changed).
A tool was registered or unregistered, or a scoped restriction changed (the available tool set changed — possibly for one scope only). An UNFILTERED registry-subject notification, deliberately not scope-filtered dispatch: a global change concerns every agent's next assembly, so a scoped listener subscribing here sees every change, not just its own scope's.
```ts cordis-catalog
'tools/change'(): void
```
Source: [`packages/core/tools/src/index.ts:97`](../../packages/core/tools/src/index.ts)
Source: [`packages/core/tools/src/index.ts:173`](../../packages/core/tools/src/index.ts)
### `tools/post-execute` — waterfall
### `tools/execute` — waterfall
Waterfall AFTER a tool runs — where hook plugins inspect the result and accept it (optionally REPLACING the model-facing content, and/or attaching `additionalContext` for the next request) or block it with corrective `feedback` (Claude Code's `PostToolUse`). Listeners receive `(exec, result, next)`: call `next()` to delegate to the default (accept unchanged), or return a PostToolDecision to override. The core tool dispatch sits between the two waterfalls as plain code, all inside `execute`'s outer try/catch (and the tool body keeps its own inner try/catch, so a thrown tool still reaches `post-execute` as an `isError` result).
Around-dispatch waterfall wrapping the registry's core tool dispatch, between the `tools/pre-execute` gate and the `tools/post-execute` seam. A listener receives `(exec, next)`: call `next()` to delegate to dispatch (returning its ToolExecutionResult, optionally wrapped), or return a replacement result without calling `next()` to short-circuit dispatch. The base `next()` IS the dispatch-with-normalization thunk — a thrown tool (or unknown tool) is already normalized to an `isError` result by the time a listener's `await next()` returns, so a wrapper never sees a raw throw from the tool body. This is the seam a timeout/retry/metrics plugin wraps: it can set or replace the one mutable field, `exec.signal` (e.g. with a per-call deadline), BEFORE `next()`, restore/delete it afterward, and inspect the result AFTER. Call identity (`token`, `callId`, `name`, `arguments`, `agent`, and `parent`) is immutable throughout the pipeline so a wrapper cannot change which tool and scope the pipeline accepted. (Cordis `next()` ignores passed arguments and re-invokes downstream with the shared payload, so a wrapper changes `exec.signal` in place rather than passing a new object to `next()`.) Multiple listeners compose by registration order — an outer one wraps the inner ones plus dispatch. Scope-filtered dispatch (`@deepseek-ai/dsh-scope`): the carrier is keyed by `exec.agent` — a listener registered through `agent.ctx` wraps only that agent's calls; a plain plugin listener wraps every call (including agent-less ones, which dispatch subject-less).
```ts cordis-catalog
'tools/post-execute'(this: ToolRegistry, exec: ToolExecution, result: ToolExecutionResult, next: () => Promise<PostToolDecision>): Promise<PostToolDecision>
'tools/execute'(this: Scoped<ToolRegistry>, exec: ToolExecution, next: () => Promise<ToolExecutionResult>): Promise<ToolExecutionResult>
```
Types: [ToolExecution](../core-data-structures/tools.md) · [ToolExecutionResult](../core-data-structures/tools.md)
Source: [`packages/core/tools/src/index.ts:92`](../../packages/core/tools/src/index.ts)
Source: [`packages/core/tools/src/index.ts:128`](../../packages/core/tools/src/index.ts)
### `tools/post-execute` — waterfall
Waterfall AFTER a tool runs — where hook plugins inspect the result and accept it (optionally REPLACING the model-facing content, and/or attaching `additionalContext` for the next request) or block it with corrective `feedback` (Claude Code's `PostToolUse`). Listeners receive `(exec, result, next)`: call `next()` to delegate to the default (accept unchanged), or return a PostToolDecision to override. Core tool dispatch runs earlier as the base `next()` of the `tools/execute` waterfall, all inside `execute`'s outer try/catch (and the tool body keeps its own inner try/catch, so a thrown tool still reaches `post-execute` as an `isError` result). Scope-filtered dispatch (`@deepseek-ai/dsh-scope`): the carrier is keyed by `exec.agent` — a listener registered through `agent.ctx` fires only for that agent's calls; a plain plugin listener fires for every call (including agent-less ones, which dispatch subject-less).
```ts cordis-catalog
'tools/post-execute'(this: Scoped<ToolRegistry>, exec: ToolExecution, result: Readonly<ToolExecutionResult>, next: () => Promise<PostToolDecision>): Promise<PostToolDecision>
```
Types: [ToolExecution](../core-data-structures/tools.md) · [ToolExecutionResult](../core-data-structures/tools.md)
Source: [`packages/core/tools/src/index.ts:148`](../../packages/core/tools/src/index.ts)
### `tools/pre-execute` — waterfall
Waterfall BEFORE a tool runs — the gate where sandbox, permission, and hook plugins allow or deny a call (Claude Code's `PreToolUse`). Listeners receive `(exec, next)`: call `next()` to delegate to the default (allow), or return a PreToolDecision without calling `next()` to short-circuit. A `deny` skips dispatch and yields an `isError` result; the tool body never runs. Input rewrite is deliberately NOT offered here (see PreToolDecision); `ask` degrades to deny until the permission system lands (`FIXME(permissions)`).
Waterfall BEFORE a tool runs — the gate where sandbox, permission, and hook plugins allow or deny a call (Claude Code's `PreToolUse`). Listeners receive `(exec, next)`: call `next()` to delegate to the default (allow), or return a PreToolDecision without calling `next()` to short-circuit. A `deny` skips dispatch and yields an `isError` result; the tool body never runs. Input rewrite is deliberately NOT offered here (see PreToolDecision); `ask` is serviced by the `ctx.approval` seam when one is mounted, and degrades to deny otherwise. Scope-filtered dispatch (`@deepseek-ai/dsh-scope`) keys the carrier by `exec.agent`: a listener registered through `agent.ctx` fires only for that agent's calls, while a plain plugin listener fires for every call (including agent-less ones, which dispatch subject-less).
```ts cordis-catalog
'tools/pre-execute'(this: ToolRegistry, exec: ToolExecution, next: () => Promise<PreToolDecision>): Promise<PreToolDecision>
'tools/pre-execute'(this: Scoped<ToolRegistry>, exec: ToolExecution, next: () => Promise<PreToolDecision>): Promise<PreToolDecision>
```
Types: [ToolExecution](../core-data-structures/tools.md)
Source: [`packages/core/tools/src/index.ts:76`](../../packages/core/tools/src/index.ts)
Source: [`packages/core/tools/src/index.ts:101`](../../packages/core/tools/src/index.ts)
### `tools/result` — emit
Synchronous notification of the authoritative FINAL tool outcome, after the complete pre/execute/post pipeline, final lossless-JSON validation, and outer error normalization. Unlike the three waterfalls, this seam cannot transform the result: each listener receives the now-frozen execution object and a deep-frozen result snapshot; listener failures are contained and logged, and ToolRegistry.execute still returns the outcome. Scope-filtered dispatch (`@deepseek-ai/dsh-scope`): keyed by `exec.agent`, using the same carrier as the pipeline.
```ts cordis-catalog
'tools/result'(this: Scoped<ToolRegistry>, exec: Readonly<ToolExecution>, result: Readonly<ToolExecutionResult>): undefined
```
Types: [ToolExecution](../core-data-structures/tools.md) · [ToolExecutionResult](../core-data-structures/tools.md)
Source: [`packages/core/tools/src/index.ts:163`](../../packages/core/tools/src/index.ts)
## `workflow/*`
### `workflow/agent-end` — emit
One `agent()` call settled (clean result, child failure, or run cancellation). Paired with Events['workflow/agent-start'] by `agent.seq`, exactly once per started call on every stop path — on an engine termination path (a worker killed past its grace) the end is engine-synthesized with outcome `'cancelled'`.
```ts cordis-catalog
'workflow/agent-end'(info: WorkflowRunInfo, agent: WorkflowAgentEndInfo): void
```
Source: [`packages/workflow/workflow/src/index.ts:96`](../../packages/workflow/workflow/src/index.ts)
### `workflow/agent-start` — emit
One `agent()` call established a ready child run. Paired with Events['workflow/agent-end'] by `agent.seq`. A call that never receives a ready run from the provider emits neither event in this pair.
```ts cordis-catalog
'workflow/agent-start'(info: WorkflowRunInfo, agent: WorkflowAgentInfo): void
```
Source: [`packages/workflow/workflow/src/index.ts:85`](../../packages/workflow/workflow/src/index.ts)
### `workflow/end` — emit
A workflow run settled (any stop reason). Fired when WorkflowRun.result resolves. Paired with Events['workflow/start'].
```ts cordis-catalog
'workflow/end'(info: WorkflowRunInfo, result: WorkflowResultInfo): void
```
Source: [`packages/workflow/workflow/src/index.ts:106`](../../packages/workflow/workflow/src/index.ts)
### `workflow/log` — emit
The script emitted a narration line (a `log(message)` call).
```ts cordis-catalog
'workflow/log'(info: WorkflowRunInfo, message: string): void
```
Source: [`packages/workflow/workflow/src/index.ts:75`](../../packages/workflow/workflow/src/index.ts)
### `workflow/phase` — emit
The script entered a phase (a `phase(title)` call) — progress grouping for observers; no execution semantics.
```ts cordis-catalog
'workflow/phase'(info: WorkflowRunInfo, title: string): void
```
Source: [`packages/workflow/workflow/src/index.ts:68`](../../packages/workflow/workflow/src/index.ts)
### `workflow/start` — emit
A workflow run started — the script's meta block validated, the body about to execute. Paired with Events['workflow/end'].
```ts cordis-catalog
'workflow/start'(info: WorkflowRunInfo): void
```
Source: [`packages/workflow/workflow/src/index.ts:60`](../../packages/workflow/workflow/src/index.ts)
## Inherited events (cordis core + loader/hmr/timer)

132
docs/cordis-catalog/services.md
View File

@@ -11,34 +11,48 @@ The **harness tier** below (the `@deepseek-ai/dsh-*` packages) is the vocabulary
## `ctx.agentLoop` — `AgentLoop`
The agent-loop plugin (`ctx.agentLoop`): creates ReactLoopAgents, runs their loops, and registers them in `ctx.agents`. Also implements the AgentFactory seam, so plugins create/resume agents through `ctx.agents` (the interface) without depending on this concrete package.
The loop itself is deliberately thin — every behavior beyond "call the model, run the tools, repeat" belongs to plugins listening on the event taxonomy declared in @deepseek-ai/dsh-agent.
Concrete ReactLoopAgent factory and driver service.
```ts cordis-catalog
create(id: AgentId, options: AgentOptions = {}): ReactLoopAgent
createAgent(options: CreateAgentOptions): AgentHandle
async resume(options: ResumeAgentOptions): Promise<AgentHandle>
create(id: AgentId, options: AgentOptions = {}, meta: Pick<SessionHeader, 'cwd'> = {}): ReactLoopAgent
async createAgent(ownerCtx: Context, options: CreateAgentOptions): Promise<AgentHandle>
async resume(ownerCtx: Context, options: ResumeAgentOptions): Promise<AgentHandle>
```
Source: [`packages/core/agent-loop/src/index.ts:68`](../../packages/core/agent-loop/src/index.ts)
Source: [`packages/core/agent-loop/src/index.ts:338`](../../packages/core/agent-loop/src/index.ts)
## `ctx.agents` — `AgentRegistry`
Agent registry (`ctx.agents`): tracks live agents so UI, hook, and orchestrator plugins can find them without depending on the concrete loop package. Agent *creation* is provided by whichever plugin implements the AgentFactory (phase 1: `@deepseek-ai/dsh-agent-loop`), registered via setFactory.
Agent registry (`ctx.agents`): tracks live agents so UI, hook, and orchestrator plugins can find them without depending on the concrete loop package. Agent *creation* is provided by whichever plugin implements the AgentFactory (`@deepseek-ai/dsh-agent-loop`), registered via setFactory.
```ts cordis-catalog
setFactory(factory: AgentFactory): () => void
create(options: CreateAgentOptions): AgentHandle
async create(options: CreateAgentOptions): Promise<AgentHandle>
async resume(options: ResumeAgentOptions): Promise<AgentHandle>
register(agent: Agent): () => void
enter(agent: Agent): () => void
announce(agent: Agent): void
get(id: AgentId): Agent | undefined
list(): Agent[]
```
Types: [Agent](../core-data-structures/core.md)
Source: [`packages/core/agent/src/index.ts:117`](../../packages/core/agent/src/index.ts)
Source: [`packages/core/agent/src/index.ts:203`](../../packages/core/agent/src/index.ts)
## `ctx.approval` — `ApprovalService`
The `ctx.approval` service: dispatches ApprovalRequests to the `approval/request` waterfall and audits every ask/outcome pair to the requesting agent's session log. Stateless between requests — grants are returned to the caller, never stored here.
Owns the policy tier too (`effective = fold(the session's 'approval/policy' events) ?? config.policy`): `request()` resolves `'never'` to `'rejected'` before dispatching any interactive answerer, a per-agent prompt section states a `'never'` policy (and only that one in prose — an `'ask'` promise could overclaim an answerer that headless compositions do not have), and an `agent/pre-step` narrator injects at most one coalesced notice when a session's effective policy moved past what the model was last told.
```ts cordis-catalog
async request(req: ApprovalRequest): Promise<ApprovalOutcome>
```
Types: [ApprovalOutcome](../core-data-structures/approval.md) · [ApprovalRequest](../core-data-structures/approval.md)
Source: [`packages/ui/user-approval/src/index.ts:294`](../../packages/ui/user-approval/src/index.ts)
## `ctx.bash` — `BashExecutor` (abstract seam)
@@ -65,7 +79,7 @@ onTaskDone(listener: BashTaskListener): () => void
Types: [BashExecRequest](../core-data-structures/bash.md) · [BashExecSpec](../core-data-structures/bash.md) · [BashRunResult](../core-data-structures/bash.md) · [BashTask](../core-data-structures/bash.md) · [BashTaskRead](../core-data-structures/bash.md)
Source: [`packages/bash/bash/src/index.ts:59`](../../packages/bash/bash/src/index.ts)
Source: [`packages/bash/bash/src/index.ts:62`](../../packages/bash/bash/src/index.ts)
## `ctx.codeRuntime` — `CodeRuntime` (abstract seam)
@@ -98,11 +112,13 @@ Implementations MUST honor:
- **Blocking**: no compaction begins while another is in progress for the same session. The recommended mechanism is the log-recorded lock — append `compact/start` before the slow work and `compact/end` after (even on failure) — so the lock is visible to replay and crash recovery.
```ts cordis-catalog
abstract compactIfNeeded( agent: CompactAgentContext, fullSystemPrompt: string, signal: AbortSignal, ): Promise<CompactionResult | null>
abstract compactIfNeeded( agent: CompactAgentContext, fullSystemPrompt: string, sessionPrefix: readonly Message[], signal: AbortSignal, ): Promise<CompactionResult | null>
abstract compactRegion( session: Session, start: number, end: number, agent: CompactAgentContext, signal?: AbortSignal, ): Promise<CompactionResult>
```
Source: [`packages/compact/compact/src/index.ts:63`](../../packages/compact/compact/src/index.ts)
Types: [Message](../core-data-structures/core.md)
Source: [`packages/compact/compact/src/index.ts:65`](../../packages/compact/compact/src/index.ts)
## `ctx.fs` — `FileSystem` (abstract seam)
@@ -145,6 +161,24 @@ Types: [GenerateOptions](../core-data-structures/core.md) · [StreamChunk](../co
Source: [`packages/llm/llm/src/index.ts:88`](../../packages/llm/llm/src/index.ts)
## `ctx.sandbox` — `SandboxProvider` (abstract seam)
Abstract process-sandbox service. Subclass, implement confine, and load the subclass as a plugin — it registers as `ctx.sandbox` (one implementation per context; loading a second throws, cordis' standard duplicate-service behavior).
Semantics every implementation must honor:
- confine either returns an argv whose runner ENFORCES the policy or fails closed — at `confine` time with SandboxUnavailableError (no backend for this host), or at EXECUTION time by the runner itself refusing to run the command (exiting without exec'ing it, identified by ConfinedArgv.runnerFailureSignatures). A silent unconfined passthrough is never a legal outcome on either path.
- Probing exists to ARBITRATE between multiple candidate backends and may be skipped when a platform has exactly one: the sole candidate is selected directly and the runner's exec-time fail-closed refusal carries the safety property. When probing does run, it is functional (actually enforcing a profile, not a version check), at most once per provider lifetime; `confine` itself spawns nothing beyond that one-time probing.
- The returned ConfinedArgv.enforcement states the backend's actual completeness for THIS host; `partial` is reported, never silently upgraded to `full`.
```ts cordis-catalog
abstract confine(argv: readonly string[], policy: SandboxPolicy): ConfinedArgv
```
Types: [ConfinedArgv](../core-data-structures/sandbox.md) · [SandboxPolicy](../core-data-structures/sandbox.md)
Source: [`packages/sandbox/sandbox/src/index.ts:180`](../../packages/sandbox/sandbox/src/index.ts)
## `ctx.sessionPersistence` — `SessionPersistence` (abstract seam)
Abstract durable session-persistence service. Subclass, implement the abstract methods, and load the subclass as a plugin — it registers as `ctx.sessionPersistence` (one implementation per context; loading a second throws, cordis' standard duplicate-service behavior).
@@ -153,7 +187,7 @@ Contracts every implementation MUST honor (a DB backend asserts them inside a tr
- **Append-only; a crashed turn is closed, not truncated.** Committed events — those at or below a flushed `turn/end` — are never rewritten. A crash can leave an unclosed final turn whose events are real (and possibly large); load preserves them and closes the orphaned turn with synthetic boundary events (see load). Only a never-fully-written torn tail fragment is discarded.
- **Contiguous seq.** A persisted log is contiguous: `events[i].seq === i`. load rejects a parse error or a `seq` gap in the COMMITTED region (unloadable); append's first event `seq` MUST equal the backend's stored next-seq (after `load` has balanced any interrupted turn).
- **JSON-serializable data.** `SessionEventMap` is merge-extensible and `event.data` is typed only as `SessionEventMap[K]`, so append REJECTS non-JSON-serializable data with an error naming the offending event type. A backend snapshots (serializes/clones) each event when it buffers, since `session.events` hands out the live mutable object.
- **JSON-serializable events.** `SessionEventMap` is merge-extensible, so append materializes each complete batch through the shared lossless-JSON boundary before buffering it. The public `session.events` view is immutable, but persistence still snapshots direct/replay callers at this independent trust boundary.
- **Durability.** append returns only once the batch is durable (the file backend fsyncs; a DB commits). create MAY defer the physical write until the first append (lazy materialization).
```ts cordis-catalog
@@ -178,25 +212,39 @@ create(id?: SessionId, options?: CreateSessionOptions): Session
prepare(id?: SessionId, options?: CreateSessionOptions): Session
enter(session: Session): () => void
announce(session: Session): void
async flush(session: Session): Promise<void>
get(id: SessionId): Session | undefined
list(): Session[]
fork(source: SessionForkSource, boundary?: number, childSessionId?: SessionId): Session
```
Source: [`packages/core/session/src/index.ts:405`](../../packages/core/session/src/index.ts)
Source: [`packages/core/session/src/index.ts:590`](../../packages/core/session/src/index.ts)
## `ctx.skills` — `SkillService`
Registry of skill providers. It merges provider catalogs with stable first-wins duplicate handling, exposes sorted model-visible summaries, and loads full skill bodies on demand.
```ts cordis-catalog
registerProvider(provider: SkillProvider): () => void
register(skill: SkillRegistration): () => void
async list(options: SkillLookupOptions = {}): Promise<SkillSummary[]>
async get(name: string, options: SkillLookupOptions = {}): Promise<SkillDefinition | undefined>
```
Source: [`packages/skill/skill/src/index.ts:158`](../../packages/skill/skill/src/index.ts)
## `ctx.subagents` — `SubagentService`
The `subagents` service: a registry of named SubagentProviders and a capability-checked start surface.
Named provider registry and capability-checked start surface.
```ts cordis-catalog
registerProvider(provider: SubagentProvider): () => void
getProvider(name: string): SubagentProvider | undefined
list(): string[]
start(name: string, request: SubagentStartRequest): SubagentRun
async start(name: string, request: SubagentStartRequest): Promise<SubagentRun>
```
Source: [`packages/subagent/subagent/src/index.ts:144`](../../packages/subagent/subagent/src/index.ts)
Source: [`packages/subagent/subagent/src/index.ts:123`](../../packages/subagent/subagent/src/index.ts)
## `ctx.systemPrompt` — `SystemPrompt`
@@ -204,27 +252,42 @@ Registry service (`ctx.systemPrompt`): plugins contribute ordered text sections,
```ts cordis-catalog
section(section: PromptSection): () => void
tools(provider: () => ToolSchema[]): () => void
tools(provider: (context: AssembleContext) => ToolProviderResult): () => void
variable(name: string, provider: (context: AssembleContext) => string | undefined): () => void
async assemble(context: AssembleContext = {}): Promise<PromptAssembly>
```
Source: [`packages/core/system-prompt/src/index.ts:291`](../../packages/core/system-prompt/src/index.ts)
Source: [`packages/core/system-prompt/src/index.ts:342`](../../packages/core/system-prompt/src/index.ts)
## `ctx.tools` — `ToolRegistry`
Tool registry (`ctx.tools`): tool plugins register definitions; the agent loop executes calls through the `tools/pre-execute` → dispatch → `tools/post-execute` pipeline. The registry contributes its schemas into the system-prompt assembly.
Tool registry (`ctx.tools`): tool plugins register definitions; the agent loop executes calls through the `tools/pre-execute` → guards → `tools/execute` → `tools/post-execute` → `tools/result` pipeline. The registry contributes its schemas into the system-prompt assembly — WHICH schemas is governed by its `mode` config (see Config.mode); under a non-native mode it also owns the reserved `run_code` presentation transport and the `tools:sdk` prompt section.
Two registration layers (`@deepseek-ai/dsh-scope`): a registration through a plain plugin context is GLOBAL (visible to every agent); one through a scoped context (`agent.ctx`) is filed in that scope's layer — visible to that agent alone, disposed with the scope, and SHADOWING a global tool of the same name for that agent (most-specific-wins; within one layer a duplicate name still throws). restrict masks the global layer per scope. One private visibility resolver feeds the registry's prompt contribution, get, and execute — and, under a non-native mode, the SDK section and `run_code`'s bindings — so those registry-owned presentation and dispatch paths agree. An expert `system-prompt/assemble` listener may deliberately replace the final wire composition and owns any resulting divergence.
```ts cordis-catalog
register(definition: ToolDefinition): () => void
get(name: string): ToolDefinition | undefined
schemas(): ToolSchema[]
async execute(exec: ToolExecution): Promise<ToolExecutionResult>
restrict(filter: ToolRestriction): () => void
guard(guard: ToolGuard): () => void
get(name: string, scope?: ScopeKey): ToolDefinition | undefined
schemas(scope?: ScopeKey): ToolSchema[]
async execute(exec: ToolExecutionInput): Promise<ToolExecutionResult>
```
Types: [ToolDefinition](../core-data-structures/tools.md) · [ToolExecution](../core-data-structures/tools.md) · [ToolExecutionResult](../core-data-structures/tools.md)
Types: [ToolDefinition](../core-data-structures/tools.md) · [ToolExecutionInput](../core-data-structures/tools.md) · [ToolExecutionResult](../core-data-structures/tools.md)
Source: [`packages/core/tools/src/index.ts:278`](../../packages/core/tools/src/index.ts)
Source: [`packages/core/tools/src/index.ts:493`](../../packages/core/tools/src/index.ts)
## `ctx.userInteraction` — `UserInteractionService`
`ctx.userInteraction`: one active UI provider plus an `ask()` surface.
```ts cordis-catalog
registerProvider(provider: UserInteractionProvider): () => void
async ask(request: AskUserQuestionRequest): Promise<AskUserQuestionAnswer>
```
Source: [`packages/ui/user-interaction/src/index.ts:82`](../../packages/ui/user-interaction/src/index.ts)
## `ctx.web` — `WebService`
@@ -248,6 +311,23 @@ async fetch(request: WebFetchRequest, exec?: WebExecContext): Promise<WebFetchRe
Source: [`packages/web/web/src/index.ts:87`](../../packages/web/web/src/index.ts)
## `ctx.workflows` — `WorkflowService` (abstract seam)
Abstract workflow execution service. Subclass, implement start, and load the subclass as a plugin — it registers as `ctx.workflows` (one implementation per context; loading a second throws, cordis' standard duplicate-service behavior).
Semantics every implementation must honor:
- start throws synchronously for a request that cannot begin (an unparseable script, an invalid meta block). Once it returns a WorkflowRun, `result` NEVER rejects — every failure resolves with `stopReason: 'error'` (or `'cancelled'`) — and once the run is cancelled, `result` SETTLES within the implementation's bounded grace even if the script itself never settles (a consumer awaiting `result` must never be wedged past a cancellation).
- The `workflow/*` events fire through emitWorkflowEvent (borrowed immutable data, per-listener containment); `workflow/end` fires exactly once per started run, after `result` is settled or as it settles.
- `dispose()` reaches quiescence within a bounded grace: it cancels, waits for the script to settle AND its started children to finish disposing, and abandons whatever is left rather than hanging its caller (the engine documents what abandonment leaves behind).
- Runs are HOLDER-OWNED: the engine hands control (`cancel`/`dispose`) to the `start()` caller and does not track its live runs — disposing the engine's own fiber mid-run deliberately leaves those runs to their holders' teardown, so an engine reload cannot yank a run out from under the consumer awaiting it.
```ts cordis-catalog
abstract start(request: WorkflowStartRequest): WorkflowRun
```
Source: [`packages/workflow/workflow/src/index.ts:211`](../../packages/workflow/workflow/src/index.ts)
## Inherited `ctx` members (cordis core + loader/hmr/timer)
The framework `ctx` surface every plugin also sees, beyond the harness services above. This is pinned vendor source ([vendoring policy](../../vendor/README.md)); it is summarized here so the page is a complete picture of what `ctx` offers, without elevating framework internals to the harness tier's prominence.

30
docs/cordis-primer.md
View File

@@ -4,35 +4,35 @@ Cordis is the vendored plugin framework underneath the DeepSeek Harness SDK. Thi
## Cordis In Five Ideas
- **A plugin is a unit of behavior.** It can be a function with optional `inject` and `apply(ctx)` fields, or a `Service` subclass whose lifecycle Cordis mounts into the current context.
- **A context is the service container.** A service claims a stable `ctx.<key>` such as `ctx.tools`, `ctx.llm`, or `ctx.sessions`; other plugins program against that key instead of importing a concrete implementation.
- **`inject` is the dependency gate.** A plugin that names required services waits until those services exist, so load order is expressed through service requirements rather than manual boot sequencing.
- **Events are typed extension seams.** Services declare event names through TypeScript declaration merging, then dispatch them as `emit`, `waterfall`, `parallel`, or `serial` depending on whether listeners observe, wrap, fan out, or run in order.
- **Registrations are disposable effects.** Prompt sections, tool schemas, adapters, providers, and listeners are installed through `ctx.effect()` or `ctx.on()` so reload and teardown unwind them predictably.
- **A plugin is a object that implements Service.** It can be a function with optional `inject` and `apply(ctx)` fields, or a `Service` subclass whose lifecycle Cordis mounts into the current context.
- **A context is a repository of services.** A service claims a stable `ctx.<key>` such as `ctx.tools`, `ctx.llm`, or `ctx.sessions` from a context; other plugins find services via key instead of importing a concrete implementation.
- **Declare service dependency via `inject`.** A plugin that names required services waits until those services exist, so load order is expressed through service requirements rather than manual boot sequencing.
- **Typed Events for communication.** Services declare event names through TypeScript declaration merging, then dispatch them as `emit`, `waterfall`, `parallel`, or `serial` depending on whether listeners observe, wrap, fan out, or run in order.
- **Registrations are reversible effects.** Prompt sections, tool schemas, adapters, providers, and listeners are installed through `ctx.effect()` or `ctx.on()` so reload and teardown unwind them predictably.
## Dispatch Modes
Use the mode to understand what a listener can do:
Every event can have one of the following dispatch mode and can only be dispatched by these methods accordingly.
| Mode | Shape |
|---|---|
| `emit` | synchronous notification; listeners observe but do not shape the result |
| `waterfall` | around-middleware; each listener receives `next()` and may wrap, rewrite, or veto |
| `parallel` | awaited fan-out; all listeners run and the dispatcher waits for them |
| `serial` | awaited in registration order; a non-void bail value stops the chain |
| Mode | Awaited? | Dispatch Order | Has Return Value? |
|---|---|---|---|
| `emit` | No | listeners observe in registration order | No |
| `waterfall` | No | listeners observe in registration order | Yes |
| `parallel` | Yes | all listeners observe the event in parallel | No |
| `serial` | Yes | listeners observe in registration order | Yes |
The mode is part of the event's public contract. New harness events document it with an `@mode` tag so the generated catalog can check declarations against dispatch sites.
## Cordis Waterfall Semantics
`ctx.waterfall` is around-middleware, not a reducer. A listener receives `(...args, next)`. Call `next()` to delegate, optionally wrapping the result; return without `next()` to short-circuit. Values propagate through `next()`'s return value.
`ctx.waterfall` is around-middleware. A listener receives `(...args, next)`. Call `next()` to delegate the possibly wrapped result to the next service; return without `next()` to short-circuit. Values propagate through `next()`'s return value.
Cooperative listeners usually mutate a shared request or decision object and then delegate. Returning a replacement is a takeover: downstream listeners see the replacement, and earlier mutations on the original object do not carry forward. Use `prepend: true` only when the listener must run before ordinary registrations.
Cooperative listeners usually mutate a shared request or decision object and then delegate. A listener can also choose to replace the result entirely and downstream listeners will only see the result after replacement. Use `prepend: true` only when the listener must run before ordinary registrations.
For single-decision events, short-circuiting is the design. A policy listener can return without `next()` when it owns the decision, while a listener that only annotates or observes must delegate.
## Practical Rules
Own vocabulary where the behavior lives: a tool pipeline event belongs to `ctx.tools`, model streaming belongs to `ctx.llm`, and live agent coordination belongs to `ctx.agents`. Prefer events for interception and policy; prefer service methods for direct capability calls.
Encapsulate behavior into plugins: a tool pipeline event belongs to `ctx.tools`, model streaming belongs to `ctx.llm`, and live agent coordination belongs to `ctx.agents`. Prefer events for interception and policy; prefer service methods for direct capability calls.
Every registration should have a disposer, either by returning one from `ctx.effect()` or using a Cordis helper that does it for you. If teardown order matters, keep the related work in one effect so disposal unwinds in the intended sequence.

64
docs/core-data-structures/approval.md Normal file
View File

@@ -0,0 +1,64 @@
# User Approval
The user-approval seam of [dsh-user-approval](../../packages/ui/user-approval) answers one question: may this specific action proceed? It owns the shared request/outcome vocabulary, the `ctx.approval` dispatch service, the `approval/request` answerer waterfall, the log-only audit pair, and the per-session `ask`/`never` policy. UI channels such as [dsh-acp](../../packages/ui/acp) provide answerers; callers such as [dsh-tools](../../packages/core/tools) and [dsh-tool-bash](../../packages/bash/tool-bash) consume the closed outcome and fail closed unless it is `allowed-once`.
Source: [`packages/ui/user-approval/src/index.ts`](../../packages/ui/user-approval/src/index.ts)
## Identity and outcome
Every request receives a fresh `ApprovalRequestId`. The brand pairs the `approval/asked` and `approval/decided` audit events without making approval ids interchangeable with tool-call, session, or agent ids.
```ts type-equiv
type ApprovalRequestId = Branded<'ApprovalRequestId'>
```
`ApprovalOutcome` is closed and fail-closed. `allowed-once` grants only the asked-about action; callers deny on `rejected`, `cancelled`, and `unavailable`. A missing, non-owning, throwing, or non-conforming answerer becomes `unavailable` rather than opening the gate.
```ts type-equiv
type ApprovalOutcome = 'allowed-once' | 'rejected' | 'cancelled' | 'unavailable'
```
## Per-session policy
`ApprovalPolicy` determines what happens before interactive answerers run. `ask` delegates to the composed answerer chain, whose no-answer default is `unavailable`; `never` deterministically returns `rejected` without dispatching any answerer. The effective value is the last `approval/policy` event in the session log, falling back to the service config. `setApprovalPolicy(session, policy)` is the single write path, so replay reconstructs the override.
```ts type-equiv
type ApprovalPolicy = 'ask' | 'never'
```
The prompt section states the deterministic `never` behavior and records either policy with a source-owned marker. The pre-step narrator reads that marker from the logged request header after restart; it does not infer state from deployment persona prose. An idle ACP switch is held in the bridge until the next `turn/start`, because approval audit and policy events must remain turn-enclosed for durable replay.
## Approval request
`ApprovalRequest` identifies the agent and tool action closely enough to route and audit the question. It deliberately omits tool arguments: an answerer attaches the prompt to the already-streamed tool call through `callId` instead of rendering a second copy that could drift.
```ts type-equiv
interface ApprovalRequest {
/**
* The agent on whose behalf the question is asked. Routes the question (a
* UI answerer only answers for agents it owns) and receives the audit
* events on its session log.
*/
readonly agent: Agent
/** The tool the question is about (presentation and audit). */
readonly toolName: string
/**
* The exact tool call being decided, when the asker has one — lets a UI
* attach the prompt to the tool call it already streamed.
*/
readonly callId?: CallId
/** The asker's human-readable explanation of WHY it is asking. */
readonly reason?: string
/**
* Aborting withdraws the question: the request settles `'cancelled'`
* immediately and a late answer from a still-pending answerer is discarded.
*/
readonly signal?: AbortSignal
}
```
## Dispatch and audit
`ctx.approval.request(req)` requires the requesting session to be inside an open turn. It appends `approval/asked`, obtains one outcome, appends the matching `approval/decided`, and resolves with that outcome. The `never` policy is enforced inside the service before waterfall dispatch, so even an answerer registered later with `prepend` cannot bypass it. Answerers return an outcome when they own the request or call `next()` to delegate; the first answer occupies the single decision slot.
The audit events are log-only and do not enter the model transcript. Model-visible behavior is the caller's derived tool result, while the request header records the prompt policy that the model actually saw. Service disposal removes its prompt section and pre-step narrator together; answerer listeners are independently effect-bound to their owning plugins.

85
docs/core-data-structures/bash.md
View File

@@ -44,6 +44,20 @@ interface BashExecRequest {
* ownerless background start (a non-agent caller).
*/
owner?: OwnerToken | undefined
/**
* Explicit per-call sandbox-policy input, overriding the executor's
* configured default mode for THIS call. Never a silent default: a
* consumer sets it only from an explicit policy source — an
* `'allowed-once'` grant a human just issued through `ctx.approval` (the
* escalation flow in the sandbox RFC § Escalation, which outranks), or the
* session's standing override folded from its own `bash/sandbox-mode`
* events (the sandbox RFC § Per-session mode switching — the user's recorded per-session
* choice). A sandboxing executor confines THIS call under the given mode;
* a non-sandboxing executor carries the field and confines nothing (the
* tool layer stamps neither escalation nor overrides without a sandboxing
* executor — see {@link BashExecutor.sandboxMode}).
*/
sandboxMode?: SandboxMode | undefined
}
```
@@ -79,6 +93,16 @@ interface BashExecSpec {
* task. `start()` stores it; `run()` (foreground) ignores it.
*/
owner: OwnerToken | undefined
/**
* The sandbox mode this call executes under, REQUIRED-but-nullable for the
* same visibility reason as `owner`. A sandboxing executor's `resolve()`
* stamps the effective mode (the request's explicit override, else its
* configured default) so `run()`/`start()` read the spec, never the config;
* a non-sandboxing executor carries the request value through verbatim and
* ignores it (`undefined` under such an executor means what its README says:
* unconfined execution).
*/
sandboxMode: SandboxMode | undefined
}
```
@@ -106,6 +130,12 @@ interface BashRunResult {
timeoutMs: number
stdout: CollectedOutput
stderr: CollectedOutput
/**
* Sandbox facts, present iff a sandboxing executor ran the command — an
* unsandboxed executor (e.g. `dsh-bash-local`) never sets it. See
* {@link BashSandboxInfo} for the `denied` classification semantics.
*/
sandbox?: BashSandboxInfo
}
```
@@ -122,9 +152,52 @@ interface CollectedOutput {
}
```
## File sandbox: `BashSandboxInfo`
A sandbox-consuming executor (`dsh-bash-sandbox`) exposes its configured fallback through `BashExecutor.sandboxMode`. The tool layer folds each agent session's durable `bash/sandbox-mode` override, stamps the effective mode onto the request, states it in the per-agent prompt, and may replace it for one user-approved strictly wider call. The mode/enforcement vocabulary is owned and cataloged by the [`@deepseek-ai/dsh-sandbox` seam](sandbox.md), whose provider wraps the executor's argv; modes govern FILE effects only, not network or process visibility.
A sandboxed run always reports the facts it executed under on `BashRunResult.sandbox`: `denied` is the executor's conservative classification of a failure as sandbox-caused (a failed exit whose stderr carries a filesystem-permission signature — never a clean exit or a signal kill), read from the collected stderr tail; `enforcement` reports how completely the selected backend governs the mode's file effects (`SandboxEnforcement = 'full' | 'partial'` — `partial` when an older Landlock ABI governs only a subset of the requested accesses; absent under `danger-full-access`, where nothing is confined); `runnerFailed` marks the opposite of a denial — the sandbox RUNNER itself failed and the command never ran (stamped only on settled background tasks; a foreground run surfaces the same condition as the thrown `SANDBOX_UNAVAILABLE` error):
```ts type-equiv
interface BashSandboxInfo {
/** The mode the command actually ran under. */
mode: SandboxMode
/**
* True when the executor classifies this run's failure as the sandbox
* denying a file operation. The classification is CONSERVATIVE (a failed
* exit whose stderr carries a filesystem-permission signature) and reads
* the COLLECTED stderr — the bounded in-memory tail per
* {@link CollectedOutput} semantics, so a signature that survives only in a
* spill file is missed toward `denied: false`. A plain command failure
* keeps `denied: false` even under a sandboxed mode.
*/
denied: boolean
/**
* How completely the runner enforced `mode`'s file effects — see
* {@link SandboxEnforcement}. Absent exactly when `mode` is
* `danger-full-access`: nothing is confined, so there is no enforcement to
* report.
*/
enforcement?: SandboxEnforcement
/**
* True when the executor classifies this failure as the SANDBOX RUNNER
* itself failing (missing binary, refused profile, fail-closed refusal
* before exec) — the command NEVER RAN; this is a sandbox failure, not a
* task failure, and it outranks `denied` (a runner's own error text can
* contain denial words). Only ever stamped on settled BACKGROUND tasks: a
* foreground run surfaces the same condition as the thrown
* `SANDBOX_UNAVAILABLE` error instead (the foreground path has an error
* channel; a settled task's facts are its only channel).
*/
runnerFailed?: boolean
}
```
One more piece completes the vocabulary: the `SANDBOX_UNAVAILABLE` error code (owned by the [sandbox seam](sandbox.md)) is what the `ctx.sandbox` provider throws — and the executor propagates — when a confined mode has no usable backend. A selected runner refusing its profile reaches the same fail-closed foreground error; a settled background task records `runnerFailed`. The model sees the current effective mode in the prompt, receives denial/runner facts in results, and can request a one-shot strictly wider retry through `sandbox_permissions` plus `justification`; `ctx.approval` must grant that exact call before anything executes. The complete policy and switching design is the [sandbox RFC](../rfc/implemented/feature/2026-07-06-sandbox.md).
## Background tasks: `BashTask`
A long-running command started with `start()` is tracked as a `BashTask`. `BashTaskStatus` is `'running' | 'completed' | 'killed'`; `done` resolves when the underlying process closes and never rejects.
A long-running command started with `start()` is tracked as a `BashTask`. `BashTaskStatus` is `'running' | 'completed' | 'killed'`; `done` resolves when the underlying process closes and never rejects. A sandboxing executor stamps `sandbox` once the task settles — classification runs against the settled task's collected stderr — so the field is absent while running and under an unsandboxed executor.
```ts type-equiv
interface BashTask {
@@ -137,6 +210,16 @@ interface BashTask {
signal: NodeJS.Signals | null
/** Resolves when the underlying process closes (never rejects). */
readonly done: Promise<void>
/**
* Sandbox facts for this task's execution, stamped by a sandboxing executor
* once the task settles and BEFORE completion listeners are notified — an
* `onTaskDone` consumer and a `done` awaiter both see it. Denial
* classification runs against the settled task's collected stderr, so the
* field cannot exist earlier: absent while the task is running and under an
* executor that does not sandbox. See {@link BashSandboxInfo} for the
* `denied` semantics.
*/
sandbox?: BashSandboxInfo
}
```

2
docs/core-data-structures/code-runtime.md
View File

@@ -1,6 +1,6 @@
# Code Runtime
The code-execution seam — a [capability seam](../rfc/implemented/architecture/2026-06-13-capability-seams.md) whose interface ([dsh-code-runtime](../../packages/code-runtime/code-runtime), `ctx.codeRuntime`) runs one model-written program against host-provided async bindings and reports what it printed and returned. Code execution is **one optional capability**, not part of the agent-loop spine — so its vocabulary lives here, not in [core.md](core.md). Backends differ by execution substrate and source language, both readonly descriptors on the service; the worker-thread backend and the tool-registry consumer (Code Mode) are specified in the [Code Mode RFC](../rfc/proposed/feature/2026-06-15-code-mode.md).
The code-execution seam — a [capability seam](../rfc/implemented/architecture/2026-06-13-capability-seams.md) whose interface ([dsh-code-runtime](../../packages/code-runtime/code-runtime), `ctx.codeRuntime`) runs one model-written program against host-provided async bindings and reports what it printed and returned. Code execution is **one optional capability**, not part of the agent-loop spine — so its vocabulary lives here, not in [core.md](core.md). Backends differ by execution substrate and source language, both readonly descriptors on the service; the worker-thread backend and the tool-registry consumer (Code Mode) are specified in the [Code Mode RFC](../rfc/implemented/feature/2026-06-15-code-mode.md).
Source: [`packages/code-runtime/code-runtime/src/types.ts`](../../packages/code-runtime/code-runtime/src/types.ts)

2
docs/core-data-structures/compaction.md
View File

@@ -50,6 +50,6 @@ interface CompactionResult {
## The service
`CompactService` (`ctx.compact`, abstract — defined in [`packages/compact/compact/src/index.ts`](../../packages/compact/compact/src/index.ts)) declares two abstract methods: `compactIfNeeded(agent, fullSystemPrompt, signal)` checks token pressure and compacts an older range if the history is too large (returning `null` when nothing needs it), and `compactRegion(session, start, end, agent, signal?)` forcibly summarizes surface nodes `[start, end]` into a single replacement node. `compactIfNeeded`'s parameters are all required — the loop's `agent/pre-step` checkpoint supplies the agent, the assembled `fullSystemPrompt`, and the turn `signal`. A backend summarizing via `ctx.llm.stream()` must forward `signal` into the call's `GenerateOptions.signal`, so an abort or dispose tears down the in-flight summarization. The entire strategy — token estimation, retention policy, event sequencing, summarization — is a HOW decision owned by the implementation.
`CompactService` (`ctx.compact`, abstract — defined in [`packages/compact/compact/src/index.ts`](../../packages/compact/compact/src/index.ts)) declares two abstract methods: `compactIfNeeded(agent, fullSystemPrompt, sessionPrefix, signal)` checks token pressure and compacts an older range if the history is too large (returning `null` when nothing needs it), and `compactRegion(session, start, end, agent, signal?)` forcibly summarizes surface nodes `[start, end]` into a single replacement node. `compactIfNeeded`'s parameters are all required — the loop's `agent/pre-step` checkpoint supplies the agent, the assembled `fullSystemPrompt`, the instance's composed `sessionPrefix` (request-only messages the derived history omits, so the pressure estimate must count them), and the turn `signal`. A backend summarizing via `ctx.llm.stream()` must forward `signal` into the call's `GenerateOptions.signal`, so an abort or dispose tears down the in-flight summarization. The entire strategy — token estimation, retention policy, event sequencing, summarization — is a HOW decision owned by the implementation.
Auto-compaction runs on the serial `agent/pre-step` loop seam (fired once per step, after `turn/start` and BEFORE the step opens and its request history is derived), not the `agent/request` waterfall: compaction mutates the session surface in place — with its log-only `compact/*` records landing cleanly outside any step — and the loop derives the request from the already-compacted surface. Retention is turn-agnostic — the only structural guard is tool-pairing balance (a compacted region's edges are balanced cuts on the surface, so it never splits a step's tool-calls from their results), so a single runaway turn that alone exceeds the window compacts its own early closed steps rather than being retained verbatim. The backend that ships this (`dsh-compact-basic`) documents the retention walk, summary shrink validation, bounded re-compaction, and the crash/recoverable failure taxonomy.

50
docs/core-data-structures/core.md
View File

@@ -16,15 +16,22 @@ Everything else is documented on a **sub-page**, not here. The rule that draws t
| Sub-page | Owns |
|---|---|
| [llm-streaming.md](llm-streaming.md) | the `StreamChunk` wire protocol + adapter contract, `BlockAssembler`, the `LlmAdapter` seam |
| [scope.md](scope.md) | scoped registration identity, dispatch carriers, and the owned `Scope` context |
| [session.md](session.md) | the full `SessionEventMap` variant catalog, `TurnTrigger`/`TurnEndReason`, `deriveMessages()`, the turn-enclosure invariant |
| [persistence.md](persistence.md) | the durability seam: `SessionPersistence`, JSONL + SQLite backends, `session/flush`, crash recovery, `SessionHeader` |
| [tools.md](tools.md) | `ToolDefinition` full fields, the schema DSL, `ToolExecution`/`ToolResult`, tool-presentation UI types, the `tools/pre-execute`/`tools/post-execute` pipeline |
| [system-prompt.md](system-prompt.md) | per-assembly context, tool-provider results, prompt sections, and cooperative assembly |
| [tools.md](tools.md) | `ToolDefinition` full fields, the schema DSL, `ToolExecution`/`ToolResult`, tool-presentation UI types, and the guarded execution pipeline |
| [user-interaction.md](user-interaction.md) | the UI-backed human question/answer seam: `AskUserQuestionRequest`, answer/options vocabulary, provider API, error taxonomy |
| [approval.md](approval.md) | the one-shot user-approval seam: `ApprovalRequest`, `ApprovalOutcome`, per-session policy, audit and answerer contracts |
| [bash.md](bash.md) | the bash executor seam: `BashExecRequest`/`Spec`, `BashRunResult`, background `BashTask`s |
| [sandbox.md](sandbox.md) | the process-confinement seam: file-effect modes, `SandboxPolicy`, `ConfinedArgv`, enforcement and fail-closed errors |
| [code-runtime.md](code-runtime.md) | the code-execution seam: `CodeRunRequest`/`Result`, binding namespaces, captured logs, the `CodeRunFailure` taxonomy |
| [filesystem.md](filesystem.md) | the filesystem seam: `FsTarget`, read/write/edit outcomes, observed-file state, `FsErrorCode` |
| [skills.md](skills.md) | the skill service: discovery priority, `SkillSummary`/`SkillDefinition`, session-prefix catalog, model-facing `skill` loading |
| [compaction.md](compaction.md) | the compaction seam: the `compact/*` session events, `CompactionResult`, the `CompactService` interface |
| [subagent.md](subagent.md) | the subagent seam: the named-provider registry, `SubagentStartRequest`/`Result`/`Run`, the start-time-vs-runtime capability split |
| [web.md](web.md) | the web access seam: `WebSearchRequest`/`Result`, `WebFetchRequest`/`Result`, `WebFetchBody`, provider/capability status, `WebError` |
| [workflow.md](workflow.md) | the workflow seam: `WorkflowStartRequest`, `WorkflowMeta`, `WorkflowRun`/`Result`, the `workflow/*` event payloads, `WorkflowError` fatality |
> Type definitions on this page are pasted **verbatim** from source and drift-checked by `pnpm run verify-type-equiv` (see [development.md](../development.md#documenting-types-verbatim-ts-type-equiv)). Inline JSDoc is omitted for readability; follow the source link for the full contracts.
@@ -128,6 +135,12 @@ Source: [`packages/llm/llm/src/types.ts`](../../packages/llm/llm/src/types.ts)
```ts type-equiv
interface GenerateOptions {
model: string
/**
* Ordered conversation messages, exactly as the provider sees them (after
* the `system` slot). A loop-built request assembles them as
* `EpochHeader.messagePrefix` + the derived history (dsh-agent-loop); a
* hand-built one-shot passes any list.
*/
messages: Message[]
/** System prompt text (adapters map to the provider's system slot). */
system?: string
@@ -188,7 +201,9 @@ The model-facing `ToolSchema` is the wire shape; the registered `ToolDefinition`
### The request envelope: `LlmCallConfig` and the logged header
Requests are built by the loop, not shaped per call: the non-content half of a request — the `EpochHeader`: this call configuration plus the rendered system prompt and the tool schemas in the assembly's canonical order (dsh-system-prompt's `toolOrder` config, lexicographic when unset) — is logged session state (`request/header` snapshot and delta events, [session.md](session.md#the-request-header-events-requestheader-and-requestheader-delta)), so every conversation request is a pure function of the session log ([reconstructability RFC](../rfc/implemented/architecture/2026-07-05-reconstructable-requests.md)). The `agent/request` waterfall receives a frozen `LlmCallConfig` seed and a listener returns a replacement to switch model or sampling — the loop logs whatever the request actually uses. Loop-built requests arrive at `llm/stream` deep-frozen; mutation throws.
Requests are built by the loop, not shaped per call: the non-history half of a request — the `EpochHeader`: this call configuration plus the rendered system prompt, the tool schemas in the authoritative returned assembly order (initially canonicalized by dsh-system-prompt's `toolOrder` config, or lexicographically when unset), and the session prefix — is logged session state (`request/header` snapshot and delta events, [session.md](session.md#the-request-header-events-requestheader-and-requestheader-delta)), so every conversation request is a pure function of the session log ([reconstructability RFC](../rfc/implemented/architecture/2026-07-05-reconstructable-requests.md)). The `agent/request` waterfall receives a frozen `LlmCallConfig` seed and a listener returns a replacement to switch model or sampling; the `agent/session-prefix` waterfall — fired once per loop instance — composes the request-only messages fronting the derived history (recorded as the header's `messagePrefix`) — the loop logs whatever the request actually uses. Loop-built requests arrive at `llm/stream` deep-frozen; mutation throws.
On the wire, a loop-built request reads in this order: the `system` slot (the rendered prompt assembly) → `messagePrefix` (the frozen session prefix) → the derived history — the boundary snapshot, whose tail is the newest `user/message` on a turn's first step and the previous step's tool results on later steps. The prefix never enters the derived history; its durable record is the header events, and the dev invariant recomputes exactly this equation against every loop-built request.
FIXME(call-config-shape): revisit the exact definition of this type — which fields are genuinely epoch-level for cache purposes (`model` certainly; the sampling scalars sit here out of caution), and where provider-specific extras (reasoning options, extra body params) belong when an adapter needs them.
@@ -244,12 +259,29 @@ interface Agent {
readonly session: Session
readonly status: AgentStatus
/** Queue a user message. Starts a turn when idle; otherwise waits for the next turn. */
/**
* The agent's scope context (`@deepseek-ai/dsh-scope`, key = this agent):
* registrations through it — tools, prompt sections/variables, listeners,
* restrictions — are visible to this agent only and unwind when it is
* disposed; `agent.ctx.on('agent/…')` listeners fire only for this agent.
*/
readonly ctx: Context
/**
* Queue a user message. Starts a turn when idle; otherwise waits for the next
* turn. Content and the resolved source are accepted as one detached,
* deeply-frozen lossless-JSON record before notification or enqueue, so
* caller or `agent/queued` listener in-place mutation cannot change later
* log/model input. Throws synchronously when either value is not losslessly
* JSON-serializable; `agent/prompt-submit` may still return an explicit
* replacement.
*/
send(content: ContentBlock[], options?: SendOptions): void
/**
* Steer a running turn: content is injected between steps of the current
* turn. When idle, behaves like {@link send}.
* turn. Uses the same owned-value and synchronous-validation boundary as
* {@link send}; when idle, behaves exactly like that method.
*/
steer(content: ContentBlock[], options?: SendOptions): void
@@ -321,7 +353,7 @@ interface Agent {
}
```
`AgentStatus` is `'idle' | 'running' | 'disposed'`. `AgentId` is a branded string. `AgentOptions` (`model?`) is merge-extensible — plugins add creation options by declaration merging; the persona is NOT an agent option but the `dsh-system-prompt` plugin's `persona` config, shared context-wide. The `agent/*` event taxonomy (lifecycle emits incl. `agent/session-start`, the serial `agent/pre-step` surface-mutation seam, and the `agent/prompt-submit`/`agent/request`/`agent/step-result`/`agent/turn-continuation` waterfalls) is in [architecture.md § Event taxonomy](../architecture.md#event-taxonomy); turn/step boundaries are durable `session/event` records, not `agent/*` emits.
`AgentStatus` is `'idle' | 'running' | 'disposed'`. `AgentId` is a branded string. `AgentOptions` (`model?`) is merge-extensible — plugins add creation options by declaration merging. Persona is not an agent option: the `dsh-system-prompt` config supplies the global default, and an agent-scoped `deployment:persona` section may shadow it. The `agent/*` event taxonomy (lifecycle emits incl. `agent/session-start`, serial `agent/pre-step`/`agent/turn-stop` checkpoints, and the `agent/prompt-submit`/`agent/request`/`agent/session-prefix`/`agent/step-result`/`agent/turn-continuation` waterfalls) is in [architecture.md § Event taxonomy](../architecture.md#event-taxonomy); turn/step boundaries are durable `session/event` records, not `agent/*` emits.
## Interception decisions
@@ -352,12 +384,20 @@ type ContinuationDecision =
| { action: 'continue'; reason?: HookContext }
```
`agent/turn-stop` returns the stop-only `ContinuationStop` subset or `undefined`. The loop calls this serial checkpoint after folding the ordinary decision, its reason, and pending steering; a stop is terminal and discards pending steering.
```ts type-equiv
type ContinuationStop = Extract<ContinuationDecision, { action: 'stop' }>
```
`agent/session-start` carries a `SessionStartSource` (why the session lifecycle began; a bridge keys its SessionStart matcher on it):
```ts type-equiv
type SessionStartSource = 'startup' | 'resume' | 'clear' | 'compact'
```
`agent/session-prefix` composes the session prefix — a plain `Message[]`, no dedicated payload type. Fired ONCE per loop instance, lazily on its first request: the composed list is deep-frozen, recorded as the header's `messagePrefix` ([the request envelope](#the-request-envelope-llmcallconfig-and-the-logged-header)), and placed in front of the ENTIRE derived history on every request the instance sends — the home for session-stable openers like a skills catalog or an AGENTS.md digest, never returned by `deriveMessages()`. Reuse is structural, so the prefix cannot drift mid-session (resume = a new instance = a recompose); content that changes mid-session goes through the append-only history channels instead (`agent.inject()`, `tools/post-execute` / prompt-submit `additionalContext`). Not a Decision union: the seam contributes content instead of vetoing, so the shape is the contribution itself.
## `ToolDefinition`
The one pipeline-authoring type that is core: what every registered tool *is* — a model-facing `ToolSchema` plus an `execute` function and optional UI presenters. A tool author rarely constructs it by hand (the `defineTool` DSL builds it with typed args), but it is the contract the registry holds and the loop dispatches through.

21
docs/core-data-structures/persistence.md
View File

@@ -25,15 +25,15 @@ interface SessionHeader {
* session is created. A persistence backend rejects any other version on load
* (no migration — see the constant).
*/
version: number
readonly version: number
/** The session's id (mirrors the {@link Session}'s id). */
id: SessionId
readonly id: SessionId
/** Unix epoch milliseconds when the session was created. */
createdAt: number
readonly createdAt: number
/** Absolute working directory the session was created in (if any). */
cwd?: string
readonly cwd?: string
/** The session this one was forked from (seed lineage), if any. */
parentSession?: SessionId
readonly parentSession?: SessionId
/**
* How many leading events were INHERITED via a seed rather than produced by
* this session — the seed boundary. Set when a fork seeds a child with a
@@ -43,7 +43,7 @@ interface SessionHeader {
* harness can skip the inherited prefix when deriving the child's OWN script
* (the seeded events are the parent's, not this child's model calls).
*/
seedLength?: number
readonly seedLength?: number
}
```
@@ -54,7 +54,7 @@ Creating a `Session` through the store takes a `seed` (replay/fork an existing e
```ts type-equiv
interface CreateSessionOptions {
/** Events to seed the new session with (replay/fork). */
seed?: SessionEvent[]
readonly seed?: readonly SessionEvent[]
/**
* Creation metadata. The store fills in `version`/`id` and defaults
* `createdAt` to now; the caller supplies the storage-level fields (validated
@@ -67,7 +67,12 @@ interface CreateSessionOptions {
* length, not the original boundary — the caller must pass the persisted
* boundary back. A fresh fork passes its actual seeded-prefix length.
*/
meta?: { cwd?: string; parentSession?: SessionId; createdAt?: number; seedLength?: number }
readonly meta?: {
readonly cwd?: string
readonly parentSession?: SessionId
readonly createdAt?: number
readonly seedLength?: number
}
}
```

82
docs/core-data-structures/sandbox.md Normal file
View File

@@ -0,0 +1,82 @@
# Process Sandbox
The process-sandbox seam of [dsh-sandbox](../../packages/sandbox/sandbox) wraps a same-world subprocess argv in a file-effect policy without coupling consumers to a platform runner. [dsh-sandbox-local](../../packages/sandbox/sandbox-local) supplies the Linux bwrap/Landlock and macOS Seatbelt backends; [dsh-bash-sandbox](../../packages/bash/bash-sandbox) is the first consumer. Containers, microVMs, and remote execution are sibling implementations of whole capability seams, not providers of `ctx.sandbox`.
Source: [`packages/sandbox/sandbox/src/index.ts`](../../packages/sandbox/sandbox/src/index.ts)
## Modes and enforcement
`SandboxMode` governs filesystem effects only. `read-only` denies writes except the required `/dev/null` sink; `workspace-write` permits writes under the workspace root and the backend's promised temp area; `danger-full-access` bypasses confinement. Network and process visibility are outside this vocabulary.
```ts type-equiv
type SandboxMode = 'read-only' | 'workspace-write' | 'danger-full-access'
```
Only the first two modes can be sent to a provider. A `danger-full-access` consumer spawns its original argv and does not call `ctx.sandbox`.
```ts type-equiv
type ConfinedSandboxMode = Exclude<SandboxMode, 'danger-full-access'>
```
Enforcement is a reported fact. `full` means the backend governs every file effect promised by the mode; `partial` means an active backend or older kernel ABI governs only a subset, so consumers that require the absolute promise must reject or surface that distinction.
```ts type-equiv
type SandboxEnforcement = 'full' | 'partial'
```
## Per-call policy
The policy is fully resolved and carried per call. This permits concurrent consumers and one-shot escalated retries to ask the same provider for different boundaries without mutating provider state.
```ts type-equiv
interface SandboxPolicy {
/** The file-effect mode this execution runs under. */
mode: ConfinedSandboxMode
/** Absolute root directory `workspace-write` may write under. */
workspaceRoot: string
}
```
## Wrapped argv and classification dialects
`ConfinedArgv` is what the consumer spawns. Besides the replacement argv, it carries the backend's enforcement fact and two orthogonal stderr dialects. `denialSignatures` identify the confined command being blocked while the sandbox works correctly. `runnerFailureSignatures` identify the sandbox runner refusing or failing before it executes the command; consumers check these first and surface a sandbox infrastructure failure, never an ordinary task failure.
```ts type-equiv
interface ConfinedArgv {
/** The wrapped argv (runner, profile, separator, then the caller's argv). */
argv: string[]
/** How completely the selected backend enforces the policy's file effects. */
enforcement: SandboxEnforcement
/**
* The selected backend's denial DIALECT: the case-insensitive stderr
* substrings a file effect denied by THIS backend produces (EROFS text
* under bwrap's read-only binds, EACCES under Landlock, EPERM under
* Seatbelt). A consumer that infers denials from a failed run's stderr
* matches against exactly these rather than a cross-backend union — the
* union claims denials a given backend never produces.
*/
denialSignatures: readonly string[]
/**
* How the RUNNER ITSELF failing identifies itself: case-insensitive stderr
* substrings produced when the sandbox binary is missing, refuses its
* profile, or fails closed before exec'ing the command (`bwrap: `,
* `landlock-run: `, `sandbox-exec: ` — each covers both the runner's own
* error prefix and the shell's runner-not-found message). ORTHOGONAL to
* {@link denialSignatures}: a denial is the confined COMMAND being blocked
* (the sandbox working as designed); a runner failure means the command
* NEVER RAN and must surface as a sandbox failure, not a task failure —
* consumers check these signatures FIRST (a runner's own error text may
* contain denial words, e.g. an unopenable grant root reporting
* `Permission denied`).
*/
runnerFailureSignatures: readonly string[]
}
```
An operator-configured local runner must supply at least one `runnerFailureSignatures` entry for its own pre-exec refusal dialect; the provider adds outer-shell missing and unexecutable forms automatically. This makes an executable custom runner rejecting its profile distinguishable from the wrapped command exiting with the same status.
## Provider and fail-closed errors
`ctx.sandbox.confine(argv, policy)` returns a `ConfinedArgv` or throws `SandboxUnavailableError` with code `SANDBOX_UNAVAILABLE` when no usable backend exists. A selected runner can also fail closed at execution time, in which case its failure signature carries the same infrastructure meaning. Silent unconfined passthrough is never legal for a confined policy.
Provider probing arbitrates between multiple candidates and is cached for the provider lifetime. A platform with one candidate may select it directly; execution-time refusal retains the safety property. The local provider reports bwrap and Seatbelt as full and preserves the Landlock launcher's full/partial kernel verdict.

31
docs/core-data-structures/scope.md Normal file
View File

@@ -0,0 +1,31 @@
# Scoped Registration
The [scope package](../../packages/core/scope) supplies the identity and carrier vocabulary that makes one registration context mean both per-agent visibility and shared lifetime ownership. It is a library primitive rather than a Cordis service; the [agent-scope runtime-design RFC](../rfc/implemented/architecture/2026-07-12-agent-scope-runtime-design.md#scope-routing-one-opaque-key-selects-one-layer) owns the implementation rationale, while the package [README](../../packages/core/scope/README.md) owns the callable API and filtering semantics.
Source: [`packages/core/scope/src/index.ts`](../../packages/core/scope/src/index.ts).
## Identity and dispatch carrier
`ScopeKey` is an opaque object identity. The shipped loop uses the live `Agent` object as its own key, but the primitive never inspects the object.
```ts type-equiv
type ScopeKey = object
```
`Scoped<T>` is the compile-time brand on the opaque routing receiver returned by `scopeTarget(base, key)`. Scope-filtered event declarations require this carrier as their `this` type, while the real event subject remains an explicit argument.
```ts type-equiv
type Scoped<T extends object> = object & { readonly [ScopedBrand]: T }
```
## Owned registration context
`Scope` pairs the tagged registration context with two teardown surfaces. `rawDispose` preserves the exact Cordis disposer identity needed by an ordered composite effect; `dispose()` is the public shared quiescence boundary for direct and racing callers.
```ts type-equiv
interface Scope {
ctx: Context
rawDispose: () => Promise<void> | void
dispose(): Promise<void>
}
```

29
docs/core-data-structures/session.md
View File

@@ -74,16 +74,15 @@ interface SessionEventMap {
*/
'request/header': { header: EpochHeader; reason: RequestHeaderReason }
/**
* Amendment to the folded {@link EpochHeader}: at least one of a
* {@link SystemDelta}, a {@link ToolsDelta}, or a whole replacement
* {@link LlmCallConfig} (four scalars — not worth diffing). Appended by the
* loop inside the step, before dispatch, when the header for this request
* differs from the fold of the log so far; the writer verifies
* `applyHeaderDelta(previous, delta)` reproduces the new header exactly and
* falls back to a `'fallback'` `request/header` snapshot when it cannot, so
* a logged delta ALWAYS round-trips. NOT a {@link SurfaceEventType}.
* Amendment to the folded {@link EpochHeader}: system line-trim, name-keyed
* tools delta, whole replacement config, or whole replacement session
* prefix (an EMPTY array encodes the transition to "none"). The
* writer verifies `applyHeaderDelta(previous, delta)` reproduces the new
* header exactly and falls back to a `'fallback'` `request/header` snapshot
* when it cannot, so a logged delta ALWAYS round-trips. NOT a
* {@link SurfaceEventType}.
*/
'request/header-delta': { system?: SystemDelta; tools?: ToolsDelta; config?: LlmCallConfig }
'request/header-delta': { system?: SystemDelta; tools?: ToolsDelta; config?: LlmCallConfig; messagePrefix?: Message[] }
}
```
@@ -100,7 +99,7 @@ export interface TodoItem {
### The request header events: `request/header` and `request/header-delta`
The request envelope — the `EpochHeader` (call config + rendered system prompt + assembled tool schemas) — is logged session state, so every conversation request is a pure function of the log (the reconstructability RFC). A `request/header` snapshot (reason `'initial' | 'resume' | 'fallback'`) anchors the fold at conversation birth, process boundaries, and delta-encoding fallbacks; `request/header-delta` events amend it mid-run. `foldRequestHeader(events)` reconstructs the header any request was built under; the writer round-trip-verifies every delta before logging it, so a well-formed log always folds. Neither is a `SurfaceEventType` — they produce no LLM message.
The request envelope — the `EpochHeader` (call config + rendered system prompt + assembled tool schemas + the session prefix) — is logged session state, so every conversation request is a pure function of the log (the reconstructability RFC). A `request/header` snapshot (reason `'initial' | 'resume' | 'fallback'`) anchors the fold at conversation birth, process boundaries, and delta-encoding fallbacks; `request/header-delta` events amend it mid-run. `foldRequestHeader(events)` reconstructs the header any request was built under; the writer round-trip-verifies every delta before logging it, so a well-formed log always folds. Neither is a `SurfaceEventType` — they produce no LLM message.
```ts type-equiv
export interface EpochHeader {
@@ -110,10 +109,18 @@ export interface EpochHeader {
system?: string
/** Assembled tool schemas; absent for a tool-less request. */
tools?: ToolSchema[]
/**
* The session prefix: request-only messages sent BEFORE the entire derived
* history (the `agent/session-prefix` waterfall's product, composed once
* per loop instance and reused for every request it sends). Not session
* history — `deriveMessages()` never returns it — so the header is its
* only durable record; absent when the instance composed none.
*/
messagePrefix?: Message[]
}
```
Canonical form: an empty system prompt and an empty tool list are ABSENT fields, matching how requests are built. The delta payloads (`SystemDelta` — a common-prefix/suffix line trim; `ToolsDelta` — name-keyed added/removed/changed) live beside the events in [`packages/core/session/src/types.ts`](../../packages/core/session/src/types.ts).
Canonical form: an empty system prompt, an empty tool list, and an empty session prefix are ABSENT fields, matching how requests are built. `messagePrefix` is the durable record of the `agent/session-prefix` waterfall's product (the request is `messagePrefix + derived history`); composed once per loop instance and anchored by that instance's snapshot, so the loop never produces a prefix delta in practice — the delta arm (whole-array replacement, an empty array encoding the transition back to absence) exists for codec totality. The other delta payloads (`SystemDelta` — a common-prefix/suffix line trim; `ToolsDelta` — name-keyed added/removed/changed) live beside the events in [`packages/core/session/src/types.ts`](../../packages/core/session/src/types.ts).
## `SessionEvent<T>` — one log entry

116
docs/core-data-structures/skills.md Normal file
View File

@@ -0,0 +1,116 @@
# Skills
The [skill capability family](../../packages/skill) is split across three packages: the registry ([dsh-skill](../../packages/skill/skill), `ctx.skills`) merges provider catalogs; the local provider ([dsh-skill-local](../../packages/skill/skill-local)) scans project/custom/user directories; the consumer ([dsh-tool-skill](../../packages/skill/tool-skill)) owns the session-prefix catalog and model-facing `skill` tool. Skills are optional instructions, not session events, so their vocabulary lives here rather than in [core.md](core.md).
Source: [`packages/skill/skill/src/index.ts`](../../packages/skill/skill/src/index.ts), [`packages/skill/skill-local/src/index.ts`](../../packages/skill/skill-local/src/index.ts), and [`packages/skill/tool-skill/src/index.ts`](../../packages/skill/tool-skill/src/index.ts).
## Provider registry
`ctx.skills` is a multi-provider registry. Providers can represent local directories, embedded plugin data, HTTP catalogs, or another source. Provider plugins register synchronously during `apply()`; remote initialization, authentication, and discovery are awaited by `list()`. Provider objects, lookup options, and candidates are readonly same-process contracts, so the registry borrows them instead of manufacturing defensive snapshots. The registry still validates semantic fields, resolves duplicate skill names first-wins by rank/provider order/local order, and sorts the final summaries by `name` for deterministic consumers. A provider `list()` rejection is logged and skipped without caching the degraded catalog; malformed candidates still fail fast because they violate the provider contract.
```ts type-equiv
interface SkillProvider {
readonly name: string
readonly list: (options: SkillLookupOptions) => Promise<readonly SkillCandidate[]>
readonly get: (candidate: SkillCandidate, options: SkillLookupOptions) => Promise<SkillDefinition | undefined>
}
```
## Local discovery priority
The shipped local provider scans roots in rank order:
| Rank | Source | Root |
|---|---|---|
| 100 | `project-dsh` | `<projectRoot>/.dsh/skills` |
| 200 | `project-agents` | `<projectRoot>/.agents/skills` |
| 300 | `custom` | `Config.customSkillDirs` |
| 400 | `user-dsh` | `<dshHome>/skills` |
| 500 | `user-agents` | `<agentsHome>/skills` |
The project root is the nearest ancestor containing `.git`; without one, the current cwd is used. When `ctx.fs` is available, the git-root walk probes `.git` through the filesystem service so remote or sandboxed workspaces do not fall back to the host filesystem boundary. The user DSH root skips its `.system` child. The local provider does not ship built-in system skills; deployments supply built-ins through another provider.
## Skill identity
Skill names are kebab-case (`^[a-z0-9]+(?:-[a-z0-9]+)*$`). The local provider accepts directory bundles (`<name>/SKILL.md`) and flat Markdown files (`<name>.md`). Nested recursive `**/SKILL.md` discovery is intentionally outside v1.
```ts type-equiv
type SkillSource = 'project-dsh' | 'project-agents' | 'runtime' | 'user-dsh' | 'user-agents' | 'custom' | (string & {})
```
## Summaries, candidates, and complete definitions
`SkillSummary` is the registry's model-invocable summary shape. Consumers choose which fields to render; the session catalog uses only `name` and `description`, never the body or absolute file path. `disableModelInvocation` hides a skill from model listings while allowing trusted code to load it by name.
```ts type-equiv
interface SkillSummary {
readonly name: string
readonly description: string
readonly whenToUse?: string
readonly disableModelInvocation?: boolean
readonly source: SkillSource
readonly provider: string
readonly resourceBase?: SkillResourceBase
}
```
`SkillCandidate` is the provider-to-registry shape. `locator` is opaque provider state; the registry only stores it and gives it back to the winning provider's `get()`.
```ts type-equiv
interface SkillCandidate extends SkillSummary {
readonly rank: number
readonly locator: unknown
readonly path?: string
readonly metadata?: Readonly<Record<string, unknown>>
}
```
`SkillDefinition` is the complete parsed result returned by `ctx.skills.get()` and used by the `skill` tool. `resourceBase` tells the tool how to render relative-resource guidance for local, URL, or provider-managed skills.
```ts type-equiv
type SkillResourceBase =
| { readonly kind: 'directory'; readonly path: string }
| { readonly kind: 'url'; readonly url: string }
| { readonly kind: 'opaque'; readonly description: string }
```
```ts type-equiv
interface SkillDefinition extends SkillSummary {
readonly content: string
readonly path?: string
readonly metadata?: Readonly<Record<string, unknown>>
}
```
Runtime skills use the same complete shape and participate in the same first-wins collection order. The returned disposer removes the contribution and invalidates discovery caches.
```ts type-equiv
type SkillRegistration = Omit<SkillDefinition, 'provider'> & {
readonly provider?: string
}
```
## Lookup and configuration
Skill lookup is cwd-sensitive because providers may expose workspace-local skills, and its optional signal cancels provider work for the caller. Providers receive the same readonly options object used for cache identity and loading. Cancellation is checked before and after catalog selection, including cache hits, and races both discovery and full-definition loading. If no git root is found, the local provider treats the supplied cwd itself as the project root.
```ts type-equiv
interface SkillLookupOptions {
readonly cwd?: string | undefined
readonly signal?: AbortSignal | undefined
}
```
The registry owns only its discovery-cache bound. The local provider owns filesystem roots (`dshHome`, `agentsHome`, and `customSkillDirs`). The consumer owns its catalog description bound.
```ts type-equiv
interface Config {
readonly collectCacheMaxEntries?: number
}
```
## Session catalog and tool contract
`dsh-tool-skill` contributes a user-role `<system-reminder>` through `agent/session-prefix`. The catalog contains sorted skill `name` and normalized, XML-escaped `description` only; it omits bodies, paths, sources, providers, and routing hints. Prefix discovery forwards the caller's abort signal through `SkillLookupOptions`. `catalogDescriptionMaxLength` is the consumer config for the description bound, with default `500` and integer minimum `3`. Its request-only, header-logged lifecycle is defined by the [session-prefix RFC](../rfc/implemented/feature/2026-07-07-session-prefix.md).
The model-facing `skill({ name })` tool validates the kebab-case name, loads the complete definition for the calling agent cwd, reports an unresolved skill as unknown or no longer available, rejects `disableModelInvocation` skills, and returns a tool result containing `<skill_content name="...">`, `<skill_resources>`, and `<skill_instructions>`. `resourceBase` resolves explicitly referenced scripts, references, and assets only as needed; the loaded result does not enumerate a skill directory. The tool result is the model-visible path for complete instructions.

49
docs/core-data-structures/subagent.md
View File

@@ -12,37 +12,41 @@ A provider advertises its **start-time** features on a static descriptor the ser
```ts type-equiv
interface SubagentCapabilities {
outputSchema: boolean
depthLimit: boolean
toolFilter: boolean
readonly outputSchema: boolean
readonly depthLimit: boolean
readonly toolFilter: boolean
readonly persona: boolean
}
```
## The start request
What a caller asks for when starting a subagent. The tool layer builds this from the model's `{ description, prompt }` plus its own config; the service validates the start-time capabilities against the named provider, then passes it to `provider.start`. `parent` is REQUIRED — in-process backends read `parent.session.header` for the working directory, the `parentSession` lineage, and the delegation depth. The three optional fields (`outputSchema`, `maxDepth`, `toolFilter`) each gate on the matching `SubagentCapabilities` flag. `outputSchema` is an object-rooted JSON Schema within the subset `assertSupportedOutputSchema` (dsh-tools) enforces — a schema outside it is rejected loud at start; the in-process backends realize it with a forced `structured_output` capture tool (see the [driver README](../../packages/subagent/subagent-inprocess/README.md)).
What a caller asks for when starting a subagent. The tool layer builds this from the model's `{ description, prompt }` plus its own config; the service validates the start-time capabilities against the named provider, then passes it to `provider.start`. `parent` is REQUIRED — in-process backends read `parent.session.header` for the working directory, the `parentSession` lineage, and the delegation depth. The four optional fields (`outputSchema`, `maxDepth`, `toolFilter`, `persona`) each gate on the matching `SubagentCapabilities` flag — in-process backends realize `toolFilter` as a scoped `tools.restrict()` and `persona` as a scoped shadowing `deployment:persona` section, both composed in the child's creation window. `outputSchema` is an object-rooted JSON Schema within the subset `assertSupportedOutputSchema` (dsh-tools) enforces — a schema outside it is rejected loud at start; the in-process backends realize it with a forced `structured_output` capture tool (see the [driver README](../../packages/subagent/subagent-inprocess/README.md)).
```ts type-equiv
interface SubagentStartRequest {
prompt: ContentBlock[]
parent: Agent
signal?: AbortSignal
agentOptions?: AgentOptions
outputSchema?: StructuredOutputSchema
maxDepth?: number
toolFilter?: { allow?: string[]; deny?: string[] }
readonly prompt: ContentBlock[]
readonly parent: Agent
readonly signal: AbortSignal
readonly agentOptions?: AgentOptions
readonly outputSchema?: StructuredOutputSchema
readonly maxDepth?: number
readonly toolFilter?: ToolRestriction
readonly persona?: string
}
```
`signal` is the single cancellation channel before and after readiness. The [subagent composition-controls RFC](../rfc/implemented/feature/2026-07-12-subagent-persona-tool-filter-and-depth.md) owns the persona, live global-tool filter, absolute-depth, and visibility-not-authority rationale.
## The terminal result: `SubagentResult`
The outcome of a run, resolved by `SubagentRun.result`. `structured` is present iff the request carried an `outputSchema` AND the provider honored it. A non-`completed` `stopReason` means `output` may be partial — the consumer maps it to an `isError` tool result rather than reporting partial output as success.
The outcome of a run, resolved by `SubagentRun.result`. `structured` is present only after a requested `outputSchema` was successfully satisfied; requesting a schema does not guarantee it, and a provider may return `stopReason: 'error'` when the child fails or finishes without a valid capture. A non-`completed` `stopReason` means `output` may be partial — the consumer maps it to an `isError` tool result rather than reporting partial output as success.
```ts type-equiv
interface SubagentResult {
output: ContentBlock[]
structured?: unknown
stopReason: SubagentStopReason
readonly output: ContentBlock[]
readonly structured?: unknown
readonly stopReason: SubagentStopReason
}
```
@@ -60,37 +64,36 @@ interface SubagentStopReasonMap {
## A live run: `SubagentRun`
The handle the consumer holds while a child executes. The consumer awaits `result`, may `cancel` mid-flight, and MUST `dispose` on every path to reach child quiescence (no leaked idle child / session). `result` does NOT reject on a child-level failure — a model/transport failure resolves with `stopReason: 'error'` — so the consumer maps a non-`completed` reason to an `isError` result; it rejects only on an infrastructure fault the seam cannot represent. `sendMessage` and `resume` are OPTIONAL: a provider that supports the runtime capability defines the method; one that doesn't omits it.
The handle the consumer holds after a provider has established a ready child. The consumer awaits `result` and MUST `dispose` on every path to cancel remaining work and reach child quiescence. `result` does NOT reject on a child-level failure — a model/transport failure resolves with `stopReason: 'error'` — so the consumer maps a non-`completed` reason to an `isError` result; it rejects only on an infrastructure fault the seam cannot represent. `sendMessage` and `resume` are OPTIONAL: a provider that supports the runtime capability defines the method; one that doesn't omits it.
```ts type-equiv
interface SubagentRun {
readonly id: AgentId
readonly result: Promise<SubagentResult>
cancel(reason?: string): void
dispose(): Promise<void>
sendMessage?(content: ContentBlock[]): void
resume?(content: ContentBlock[]): SubagentRun
resume?(content: ContentBlock[]): Promise<SubagentRun>
}
```
## The provider seam: `SubagentProvider`
One transport for running a child agent. Implementations register under a unique name via `SubagentService.registerProvider`; multiple coexist in one context. The service validates every requested start-time capability before calling `start`, so an implementation may assume e.g. `request.maxDepth` is honorable when present. `inheritsParentContext` is a DESCRIPTIVE fact beside the capabilities (nothing validates against it): whether a child sees the parent conversation (`fork`: true, `spawn`/`acp`: false) — the model-facing consumer derives truthful tool wording from it.
One transport for running a child agent. Implementations register under a unique name via `SubagentService.registerProvider`; multiple coexist in one context. The service validates every requested start-time capability before calling `start`, so an implementation may assume e.g. `request.maxDepth` is honorable when present. `inheritsParentContext` is a DESCRIPTIVE fact beside the capabilities (nothing validates against it): whether a child sees the parent conversation (`fork`: true, `spawn`/`acp`: false) — the model-facing consumer derives truthful tool wording from it. It describes conversation history only, not tool registrations, injected services, or authority inheritance.
```ts type-equiv
interface SubagentProvider {
readonly name: string
readonly capabilities: SubagentCapabilities
readonly inheritsParentContext: boolean
start(request: SubagentStartRequest): SubagentRun
start(request: SubagentStartRequest): Promise<SubagentRun>
}
```
The service (`ctx.subagents`) emits `subagent/start` when a run begins and `subagent/end` when it settles (see the [events catalog](../cordis-catalog/events.md)). `subagent/end` carries `lastAssistantMessage` (the child's final `output`) on the settle path, so an observer sees WHAT the subagent produced without holding the run (absent when the run rejected at the infrastructure level — no result was produced). These are **observe-only** events: both are plain `emit`s (the `subagent/end` fires from a detached `.then` after the result settles and awaits no listener), so a subscriber observes but cannot change the run. Both emits contain a thrown listener **per listener** (logged, never propagated): one bad subscriber can neither strand a live run, surface as an unhandled rejection on the detached settle hook, nor starve the listeners registered after it.
`SubagentProvider.start()` and `ctx.subagents.start()` are the publication boundary: their promises fulfill only with a ready run. The service attaches result observation, emits `subagent/start`, and returns the same holder-owned run; a rejected start has already cleaned provider-owned partial resources and emits neither lifecycle event. For an in-process provider, a start listener can resolve the live child with `ctx.agents.get(info.id)`; a remote provider need not publish into the local registry. `subagent/end` carries `lastAssistantMessage` (the child's final `output`) on the settle path and reports `error` on infrastructure rejection. Both lifecycle events are observe-only emits with per-listener exception containment.
## In-process backends: depth and seed
The two in-process backends ([dsh-subagent-spawn](../../packages/subagent/subagent-spawn) fresh, [dsh-subagent-fork](../../packages/subagent/subagent-fork) seeded) run the child as a child `Agent` on the same context via `ctx.agents.create`. Two pieces of vocabulary ride on the existing agent/session types rather than new core types:
The two in-process backends ([dsh-subagent-spawn](../../packages/subagent/subagent-spawn) fresh, [dsh-subagent-fork](../../packages/subagent/subagent-fork) seeded) run the child as an ordinary `Agent` in the same application. The provider creates it directly through `parent.ctx`, passes the required signal into the core creation transaction, and delegates quiescent disposal to the returned `AgentHandle`. Provider removal prevents new starts but does not revoke an accepted run. The child receives a flat new scope rather than inheriting the parent's registrations. Two pieces of vocabulary ride on the existing agent/session types rather than new core types:
- **Delegation depth** is a merge-extensible `AgentOptions.subagentDepth` field (`0` for a top-level agent, parent + 1 for a child). The seam owns it — the loop neither sets nor reads it — so a nested spawn reads its parent's depth from `parent.options.subagentDepth` and the `depthLimit` capability caps the tree by refusing a child whose depth would exceed `request.maxDepth`.
- **Delegation depth** is a merge-extensible `AgentOptions.subagentDepth` field (`0` for a top-level agent, parent + 1 for a child). Only `undefined` means top level; every stored present value must be a non-negative safe integer. The seam owns it — the loop neither sets nor reads it — so a nested spawn validates its parent's stored depth, rejects a derived child depth outside the safe-integer domain, and applies a defined absolute `request.maxDepth` cap to that child.
- **Fork seeding** uses `CreateAgentOptions.seed` (a `SessionEvent[]` prefix threaded through `AgentLoop.createAgent` → `ctx.sessions.prepare({ seed })`, the same primitive `resume` uses). The fork backend passes a *balanced completed-turn prefix* of the parent's log — the parent's events up to and including its last `turn/end` — so the seed is contiguous-from-0 and the [invariants](../../packages/support/invariants) replay accepts it (the in-flight, unbalanced turn is excluded).

38
docs/core-data-structures/system-prompt.md Normal file
View File

@@ -0,0 +1,38 @@
# System Prompt Assembly
The [system-prompt package](../../packages/core/system-prompt) owns the data exchanged between prompt contributors and one assembly call. The package [README](../../packages/core/system-prompt/README.md) documents registration, ordering, scoping, and rendering behavior; this page pins the literal cross-package shapes that plugins implement or pass.
Source: [`packages/core/system-prompt/src/index.ts`](../../packages/core/system-prompt/src/index.ts).
## Assembly context
`AssembleContext` identifies the scope layer one assembly resolves. It is merge-extensible: `dsh-agent` adds the optional live `agent` field, and `assembleContextFor(agent)` sets that field and `scope` together.
```ts type-equiv
interface AssembleContext {
scope?: ScopeKey
}
```
## Tool-provider result
`ToolProviderResult.schemas` is the model-visible set for the current assembly. `knownNames` is the provider's pre-restriction name universe used to distinguish a configured-name typo from a known tool that is deliberately hidden in this scope.
```ts type-equiv
interface ToolProviderResult {
readonly schemas: readonly ToolSchema[]
readonly knownNames?: readonly string[]
}
```
## Prompt sections
`PromptSection` is a readonly same-process registration contract. Its text may be static or resolved from the current assembly context.
```ts type-equiv
interface PromptSection {
readonly name: string
readonly order: number
readonly text: string | ((context: AssembleContext) => string)
}
```

68
docs/core-data-structures/tools.md
View File

@@ -1,6 +1,6 @@
# Tools
The tool pipeline of [dsh-tools](../../packages/core/tools). [core.md](core.md) introduces `ToolDefinition` as the one pipeline-authoring type promoted to the spine and `ToolSchema` as the model-facing wire shape. This page owns the full `ToolDefinition`, the typed schema DSL that builds it, the waterfall execution shapes, and the UI-presentation vocabulary.
The tool pipeline of [dsh-tools](../../packages/core/tools). [core.md](core.md) introduces `ToolDefinition` as the one pipeline-authoring type promoted to the spine and `ToolSchema` as the model-facing wire shape. This page owns the full `ToolDefinition`, the typed schema DSL that builds it, the guarded execution shapes, and the UI-presentation vocabulary.
Source: [`packages/core/tools/src/index.ts`](../../packages/core/tools/src/index.ts) · [`packages/core/tools/src/schema.ts`](../../packages/core/tools/src/schema.ts) · [`packages/core/tools/src/presentation.ts`](../../packages/core/tools/src/presentation.ts)
@@ -11,6 +11,14 @@ A `ToolSchema` (the model-facing fields) plus the `execute` function and optiona
```ts type-equiv
interface ToolDefinition extends ToolSchema {
execute(args: unknown, exec: ToolExecution): Promise<ToolExecuteReturn>
/**
* Cooperative tool-call timeout budget in milliseconds. Omit for no deadline.
* Enforced by `@deepseek-ai/dsh-timeout-policy` (a `tools/execute` wrapper); it
* is NEVER sent to the model — `schemas()` whitelists only name/description/
* parameters. Declaring it asserts this tool forwards `exec.signal` to a
* cooperative implementation that can reach quiescence when the signal aborts.
*/
timeoutMs?: number
/**
* Optional: how to present the PENDING state of one call in a UI, derived from
* the call's `args` (parsed arguments, `unknown` — the tool validates/narrows
@@ -73,22 +81,60 @@ type InferArgs<S extends SchemaSpec> = Simplify<
`defineTool({ name, description, parameters, execute, … })` ties it together: `parameters` is a `SchemaSpec`, `execute(args, exec)` gets `args: InferArgs<typeof parameters>`, and the helper converts the spec to JSON Schema (`schemaSpecToJsonSchema`) for the wire and validates model-generated args (`validateArgs`) before the typed body runs. A mismatch throws `ToolArgsError` (`code: 'INVALID_ARGS'`), which the registry turns into an `isError` result so the model can self-correct. Why a custom DSL and not schemastery: tool parameters need JSON Schema (the LLM wire format), not validation/transformation — the lightweight DSL gives the best authoring DX with the smallest surface.
## Execution: the `tools/pre-execute` / `tools/post-execute` pipeline shapes
Registration is a trusted same-process contract. The registry borrows the typed definition as readonly input and validates only semantic requirements such as a positive finite `timeoutMs`; `schemas()` materializes the explicit model-facing projection at the model boundary so execution and presentation share one resolved definition without leaking callbacks onto the wire.
`ctx.tools.execute()` runs each call through a two-waterfall pipeline — `tools/pre-execute` (the allow/deny/ask gate) → core dispatch → `tools/post-execute` (inspect/replace the result, attach context) — the seams where sandbox, permission, hook, and plan-mode plugins gate or transform a call. The pending call is a `ToolExecution`; the outcome is a `ToolExecutionResult`.
## `ToolRestriction` — one scope's live global filter
`ToolRestriction` applies only to the live deployment-global tool layer. The registry compiles readonly names into private sets, intersects multiple restrictions, then overlays scope-local tools. A deny-only filter admits later unlisted globals, while an allow-list excludes them.
```ts type-equiv
interface ToolExecution {
callId: CallId
name: string
interface ToolRestriction {
readonly allow?: readonly string[]
readonly deny?: readonly string[]
}
```
## Execution: extensible waterfalls plus monotonic policy
`ctx.tools.execute()` accepts a caller-owned `ToolExecutionInput`, materializes its parsed JSON arguments once into a pipeline-owned `ToolExecution`, and runs that call through `tools/pre-execute` (the reorderable allow/deny/ask waterfall) → registered monotonic guards → `tools/execute` (around-dispatch wrappers) → `tools/post-execute` (inspect/replace the result) → `tools/result` (the immutable authoritative outcome). The outcome is a `ToolExecutionResult`.
```ts type-equiv
type ToolExecutionToken = symbol & { readonly [toolExecutionTokenBrand]: true }
```
```ts type-equiv
interface ToolExecutionInput {
readonly callId: CallId
readonly name: string
/** Parsed JSON arguments (unknown — tools validate their own input). */
arguments: unknown
readonly arguments: unknown
/** The agent on whose behalf the call runs (set by the agent loop). */
agent?: Agent
readonly agent?: Agent
/**
* Opaque token of the enclosing transport execution, when one exists. Code
* Mode sets this on SDK sub-dispatches so commit-style observers can wait for
* the outer `run_code` outcome without receiving its live mutable execution.
*/
readonly parent?: ToolExecutionToken
signal?: AbortSignal
}
```
```ts type-equiv
interface ToolExecution extends ToolExecutionInput {
/** Registry-assigned identity shared with nested calls only as their opaque `parent` token. */
readonly token: ToolExecutionToken
}
```
`ToolExecutionToken` is a compile-time opaque fresh `Symbol` at runtime; identity comparison is its only operation. Before policy runs, `ctx.tools.execute()` materializes `arguments` as detached lossless JSON, assigns the token, and deep-freezes the accepted arguments. A non-JSON value is normalized to an error before policy. `token`, `callId`, `name`, `arguments`, `agent`, and the optional `parent` token are readonly throughout the waterfalls, while an around-dispatch wrapper may add, replace, or remove only optional `signal`. After the complete pipeline the registry freezes the execution and exposes its stable identity to `tools/result` observers.
A `ToolGuard` is scope-aware final pre-dispatch policy. Its shape deliberately has no allow result: `undefined` preserves the waterfall decision, while a returned reason can only reduce permission, so a later listener cannot undo it.
```ts type-equiv
type ToolGuard = (execution: Readonly<ToolExecution>) => string | undefined
```
```ts type-equiv
interface ToolExecutionResult {
callId: CallId
@@ -121,7 +167,9 @@ interface ToolExecutionResult {
}
```
Each interception waterfall returns a typed **Decision** (the idiom shared with the `agent/*` seams). `tools/pre-execute` listeners receive `(exec, next)` and return a `PreToolDecision`; `tools/post-execute` listeners receive `(exec, result, next)` and return a `PostToolDecision`:
The registry materializes and freezes the final accepted result immediately before `tools/result`. Its content, structured error, additional context, and presentation metadata must round-trip losslessly through JSON; an invalid outcome becomes a JSON-safe `isError` result, so the observed live outcome is safe for the later durable `tool/result` append.
Each interception waterfall returns a typed **Decision** (the idiom shared with the `agent/*` seams). `tools/pre-execute` listeners receive `(exec, next)` and return a `PreToolDecision`; `tools/execute` wrappers return a `ToolExecutionResult`; `tools/post-execute` listeners receive `(exec, result, next)` and return a `PostToolDecision`:
```ts type-equiv
type PreToolDecision =
@@ -136,7 +184,7 @@ type PostToolDecision =
| { kind: 'block'; feedback: ContentBlock[]; additionalContext?: HookContext }
```
Call `next()` to delegate to the default (allow / accept-unchanged), or return a decision to short-circuit. A `pre-execute` `deny` (or `ask`, which degrades to deny until the permission system lands) skips dispatch and yields an `isError` result; input rewrite is deliberately NOT offered on `PreToolDecision` (it would desync the pre-execution audit/history/UI from what ran — its own proposed RFC). A `post-execute` `accept` may replace the model-facing `content` (clean, because `tool/result` is logged after `execute()` returns); a `block` turns the call into an `isError` whose content is the corrective `feedback`. Core dispatch sits between the waterfalls as plain code; the tool body keeps its own try/catch so a thrown tool still reaches `post-execute` as an `isError`. An unregistered tool routes through the same catch as a tool-thrown error, so both failure classes get a structured `{ name, code }` (`ToolNotFoundError` → `UNKNOWN_TOOL`) — the loop records a failed tool call instead of failing the whole turn.
Call `next()` to delegate to the default (allow / dispatch / accept-unchanged), or return a decision/result to short-circuit. A `pre-execute` `deny` skips dispatch and yields an `isError` result. An `ask` resolves through the optional approval seam: only `allowed-once` proceeds, while every non-grant, missing channel/service, or agent-less request becomes a normalized denial. A registered `ToolGuard` then runs and can still impose a final denial. Input rewrite is deliberately NOT offered on `PreToolDecision` because it would desync the pre-execution audit/history/UI from what ran. A `post-execute` `accept` may replace the model-facing `content`; a `block` turns the call into an `isError` whose content is the corrective `feedback`. The synchronous `tools/result` notification then receives the frozen execution identity and a deep-frozen result snapshot after every wrapper, post decision, and outer error catch; observers cannot transform the outcome or race each other through payload mutation, and one observer failure neither changes the result nor starves peers. An unregistered tool routes through the same catch as a tool-thrown error, so both failure classes get a structured `{ name, code }` (`ToolNotFoundError` → `UNKNOWN_TOOL`) — the loop records a failed tool call instead of failing the whole turn.
## The structured-output schema subset

97
docs/core-data-structures/user-interaction.md Normal file
View File

@@ -0,0 +1,97 @@
# User Interaction
The user-interaction seam of [dsh-user-interaction](../../packages/ui/user-interaction). It is the provider-neutral vocabulary a tool or permission plugin uses when it needs the human to answer before the agent can continue. UI surfaces provide the active `UserInteractionProvider`: `dsh-stdio-agent` renders questions in readline, and `dsh-acp` maps them to ACP form elicitations.
Source: [`packages/ui/user-interaction/src/index.ts`](../../packages/ui/user-interaction/src/index.ts)
## Question options
`AskUserQuestionOption` is the selectable-choice shape. `label` is the user-facing option text and also the model-facing selected value; `description` is optional UI help text.
```ts type-equiv
interface AskUserQuestionOption {
/** User-facing label. */
label: string
/** Optional extra context rendered by capable UIs. */
description?: string
}
```
## Question item
`AskUserQuestionItem` is one question in a request. The model supplies a stable `id`, which is echoed back with the answer so batched questions remain routable.
```ts type-equiv
interface AskUserQuestionItem {
/** Stable model-provided question id, echoed in the answer. */
id: string
/** The question to display. */
question: string
/** Optional short heading/group label. */
header?: string
/** Optional choices the UI can render as a menu. */
options?: AskUserQuestionOption[]
/** Whether more than one option may be selected. Defaults to single-select. */
multiSelect?: boolean
}
```
## Ask request
`AskUserQuestionRequest` is the cross-package request. `questions` is an array so a UI can present related prompts in one flow while preserving a stable id per answer.
```ts type-equiv
interface AskUserQuestionRequest {
/** Questions to display. */
questions: AskUserQuestionItem[]
/** Calling agent, when the request came from an agent tool call. */
agent?: Agent
/** Abort signal for the owning tool/step. */
signal?: AbortSignal
}
```
## Answer
Providers return one answer per answered question id. `selected` contains selected option labels, and `custom` carries a free-form "Other" answer when the user typed one. When `custom` is present, `selected` is empty; custom text is an answer override, not a supplement to selected choices.
```ts type-equiv
interface AskUserQuestionAnswerItem {
/** The answered question id. */
id: string
/** Selected option labels. Empty when the answer is purely custom text. */
selected: string[]
/** Optional free-text "Other" answer. */
custom?: string
}
```
```ts type-equiv
interface AskUserQuestionAnswer {
/** Structured answers keyed by question id. */
answers: AskUserQuestionAnswerItem[]
}
```
## Provider
Only one provider may be active in a context. Provider registration is effect-bound so HMR/disposal removes the active UI.
```ts type-equiv
interface UserInteractionProvider {
ask(request: AskUserQuestionRequest): Promise<AskUserQuestionAnswer>
}
```
## Errors
`UserInteractionError` extends `HarnessError`, so `ctx.tools.execute()` preserves `{ name, code }` for model-facing tool failures such as `EMPTY_QUESTIONS`, `NO_PROVIDER`, `ASK_ABORTED`, or ACP-side cancellation.
```ts type-equiv
class UserInteractionError extends HarnessError {
constructor(message: string, code: string, options?: ErrorOptions) {
super(message, code, options)
this.name = 'UserInteractionError'
}
}
```

69
docs/core-data-structures/workflow.md Normal file
View File

@@ -0,0 +1,69 @@
# Workflow
The workflow seam — an agent running a model-written orchestration SCRIPT that fans out subagents. Like [subagent](subagent.md) it is **one optional capability**, not part of the agent-loop spine, so its vocabulary lives here rather than in [core.md](core.md). Unlike the subagent registry it takes the bash shape: ONE engine implementation per context provides `ctx.workflows`; there is no named-provider registry (a second engine is a plugin swap, not a co-resident).
Interface: [dsh-workflow](../../packages/workflow/workflow) (`ctx.workflows` + the vocabulary below). The implementation is [dsh-workflow-workerthread](../../packages/workflow/workflow-workerthread) (a `node:worker_threads` engine — one worker per run, the script's vm context inside it); the model-facing consumer is [dsh-tool-workflow](../../packages/workflow/tool-workflow). The proposal and rationale: [the dynamic-workflows RFC](../rfc/implemented/feature/2026-07-05-dynamic-workflows.md).
Source: [`packages/workflow/workflow/src/types.ts`](../../packages/workflow/workflow/src/types.ts)
## The start request
What a caller asks for when starting a run. The tool layer builds this from the model's `{ script, meta, args }` call plus the calling agent; `meta` and `args` are plain JSON DATA (the engine shape-validates `meta` and rejects loud BEFORE anything runs — no script text is ever evaluated to obtain it). `parent` is REQUIRED — every child the script spawns is attributed to it (cwd, lineage, and depth flow through the [subagent seam](subagent.md)).
```ts type-equiv
interface WorkflowStartRequest {
script: string
meta: WorkflowMeta
args?: unknown
parent: Agent
signal?: AbortSignal
}
```
## The workflow's identity: `WorkflowMeta`
The identity block carried as data on the start request (the tool's `meta` parameter; the field vocabulary matches the Claude Code dynamic-workflows meta block). `phases` is progress vocabulary only: `phase()` calls match titles for observers; no execution structure is implied.
```ts type-equiv
interface WorkflowMeta {
name: string
description: string
whenToUse?: string
phases?: WorkflowPhase[]
}
```
## The terminal result: `WorkflowResult`
The outcome of one run, resolved by `WorkflowRun.result`. `value` is the script's materialized return value — plain host-realm JSON data (`null` when the script returned nothing) — meaningful only for `completed`. `stopReason` is a CLOSED union (engine-owned; consumers may exhaust it): `completed` | `cancelled` | `error`. A non-`completed` reason carries the failure in `error`, and the consumer maps it to an `isError` tool result rather than reporting partial output as success.
```ts type-equiv
interface WorkflowResult {
value: unknown
stopReason: WorkflowStopReason
error?: string
agentsStarted: number
}
```
## A live run: `WorkflowRun`
The handle the consumer holds while a script executes. The consumer awaits `result`, may `cancel` mid-flight, and MUST `dispose` on every path. `result` does NOT reject — a script failure resolves with `stopReason: 'error'` — and once the run is cancelled it SETTLES within the engine's bounded grace even if the script itself never settles (the engine force-settles `cancelled`; the worker-thread engine then terminates the script's worker), so a consumer awaiting `result` is never wedged past a cancellation. `dispose()` = cancel + that bounded settle + child quiescence; it never hangs on a stuck script.
```ts type-equiv
interface WorkflowRun {
readonly id: WorkflowRunId
readonly meta: WorkflowMeta
readonly result: Promise<WorkflowResult>
cancel(reason?: string): void
dispose(): Promise<void>
}
```
## Failure discipline: `WorkflowError.fatal`
Hook misuse inside a script — bad arguments, unknown/deferred `agent()` options, a schema outside the [structured-output subset](../../packages/core/tools/README.md), a tripped cap, a seam start failure, cancellation — throws a `WorkflowError` with `fatal: true`. The `parallel()`/`pipeline()` combinators RE-THROW fatal errors instead of mapping the item to `null`: a typo'd option must kill the script loudly, never dissolve into something that reads as an ordinary child failure. The per-item `null` is reserved for child-run failures (a non-`completed` stop reason) and ordinary in-stage script errors.
## Events
The `workflow/*` events (`workflow/start`, `workflow/phase`, `workflow/log`, `workflow/agent-start`, `workflow/agent-end`, `workflow/end` — see the [events catalog](../cordis-catalog/events.md)) are **observe-only** emits carrying DATA SNAPSHOTS: every payload starts with `WorkflowRunInfo` (id + meta), never the live `WorkflowRun`, so a subscriber cannot gain `cancel`/`dispose`, and `workflow/end` deliberately omits the result value (a listener observing outcomes must not receive a mutable alias of the caller's result). Every emit is per-listener contained — a throwing subscriber is logged, never propagated, and cannot starve the listeners registered after it — and every listener receives its own payload clone, so mutating it corrupts neither the engine nor other listeners; the containment mirrors `subagent/start`/`subagent/end`.

4
docs/development.i18n.yaml
View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write
development.md: bd6f6b561480419abea7a42a44b4078e2c59b1cb
development.zh.md: 54bf19765d2b4dc419e6b71684dbcfcd28230541
development.md: ea5f2e5d08acbaf1dfce4661530218dbf1a051b9
development.zh.md: 50877796f2ff47ad46cc67b35f3cd7b5704c8315

2
docs/development.md
View File

@@ -67,7 +67,7 @@ These hooks do not exactly mirror CI. Notably, `pre-push` runs unit tests withou
## CI gates
The keyless GitHub workflow has eight jobs: five Node 24 lanes run static gates, lint, coverage, snapshot replay, and artifact gates separately, and three compatibility jobs run `pnpm run check:node-compat` on Node 22.19, 24, and 26. The lane schedulers fan out independent gates from `package.json`: constraints, typecheck, lint, coverage, snapshot replay, `doc-sync` members, module-graph freshness, `knip`, and the echo-agent smoke test.
The keyless GitHub workflow has eight jobs: five Node 24 lanes run static gates, lint, coverage, snapshot replay, and artifact gates separately, and three compatibility jobs run `pnpm run check:node-compat` on Node 22.19, 24, and 26. The compatibility command runs the TypeScript typecheck and a keyless workflow-workerthread source-launch smoke on every runtime, so the matrix proves that the source graph typechecks and that a real unbuilt Worker loader path executes; the other lane schedulers fan out independent gates from `package.json`: constraints, lint, coverage, snapshot replay, `doc-sync` members, module-graph freshness, `knip`, and the echo-agent smoke test.
`pnpm run build` feeds the artifact lane, and `publint`, `verify-node-next-types`, and built-bin smoke tests wait for build output. The separate real-API workflow runs `pnpm run test:e2e` with a secret and `DSH_E2E_MAX_WORKERS=14`.

2
docs/development.zh.md
View File

@@ -67,7 +67,7 @@ vendor manifest 守卫检查 `vendor/*/src` 下的改动是否连同对应的 `v
## CI 门禁
keyless GitHub 工作流有八个 job：五个 Node 24 lane 分别运行 static gates、lint、coverage、snapshot replay 和 artifact gates，三个兼容性 job 在 Node 22.19、24 和 26 上运行 `pnpm run check:node-compat`。各 lane 调度器并发运行来自 `package.json` 的独立门禁：constraints、typecheck、lint、coverage、snapshot replay、`doc-sync` 成员、module graph 新鲜度、`knip` 和 echo-agent 冒烟测试。
keyless GitHub 工作流有八个 job：五个 Node 24 lane 分别运行 static gates、lint、coverage、snapshot replay 和 artifact gates，三个兼容性 job 在 Node 22.19、24 和 26 上运行 `pnpm run check:node-compat`。兼容性命令会在每个运行时上运行 TypeScript 类型检查和 keyless 的 workflow-workerthread 源码启动冒烟测试，因此该矩阵既证明源码图能通过类型检查，也会实际执行一条未构建的 Worker loader 路径；其他 lane 调度器并发运行来自 `package.json` 的独立门禁：constraints、lint、coverage、snapshot replay、`doc-sync` 成员、module graph 新鲜度、`knip` 和 echo-agent 冒烟测试。
`pnpm run build` 供给 artifact lane，`publint`、`verify-node-next-types` 和 built-bin 冒烟测试等待 build 输出。单独的真实 API 工作流带密钥运行 `pnpm run test:e2e`，并设置 `DSH_E2E_MAX_WORKERS=14`。

66
docs/event-producer-consumer.md
View File

@@ -7,32 +7,52 @@ This matrix shows which packages dispatch each harness-owned event and which pac
| Event | Mode | Declared in | Dispatchers | Listeners |
| --- | --- | --- | --- | --- |
| `agent/created` | `emit` | [`packages/core/agent/src/types.ts:264`](../packages/core/agent/src/types.ts) | [`agent`](../packages/core/agent) (`emit`) | [`stdio-agent`](../packages/ui/stdio-agent) |
| `agent/disposed` | `emit` | [`packages/core/agent/src/types.ts:271`](../packages/core/agent/src/types.ts) | [`agent`](../packages/core/agent) (`emit`) | [`stdio-agent`](../packages/ui/stdio-agent) |
| `agent/error` | `emit` | [`packages/core/agent/src/types.ts:420`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`emit`) | - |
| `agent/pre-step` | `serial` | [`packages/core/agent/src/types.ts:349`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`serial`) | [`compact-basic`](../packages/compact/compact-basic) |
| `agent/prompt-submit` | `waterfall` | [`packages/core/agent/src/types.ts:362`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`waterfall`) | [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex) |
| `agent/queued` | `emit` | [`packages/core/agent/src/types.ts:289`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`emit`) | - |
| `agent/request` | `waterfall` | [`packages/core/agent/src/types.ts:385`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`waterfall`) | - |
| `agent/session-start` | `emit` | [`packages/core/agent/src/types.ts:304`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`emit`) | [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex) |
| `agent/status` | `emit` | [`packages/core/agent/src/types.ts:280`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`emit`) | [`acp`](../packages/ui/acp), [`invariants`](../packages/support/invariants), [`stdio-agent`](../packages/ui/stdio-agent) |
| `agent/step-result` | `waterfall` | [`packages/core/agent/src/types.ts:395`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`waterfall`) | - |
| `agent/turn-continuation` | `waterfall` | [`packages/core/agent/src/types.ts:408`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`waterfall`) | [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex) |
| `agent/created` | `emit` | [`packages/core/agent/src/types.ts:316`](../packages/core/agent/src/types.ts) | [`agent`](../packages/core/agent) (`events.dispatch`) | [`stdio-agent`](../packages/ui/stdio-agent) |
| `agent/disposed` | `emit` | [`packages/core/agent/src/types.ts:331`](../packages/core/agent/src/types.ts) | [`agent`](../packages/core/agent) (`events.dispatch`) | [`stdio-agent`](../packages/ui/stdio-agent) |
| `agent/error` | `emit` | [`packages/core/agent/src/types.ts:605`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`emit`) | - |
| `agent/pre-step` | `serial` | [`packages/core/agent/src/types.ts:438`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`serial`) | [`compact-basic`](../packages/compact/compact-basic), [`user-approval`](../packages/ui/user-approval) |
| `agent/prompt-submit` | `waterfall` | [`packages/core/agent/src/types.ts:456`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`waterfall`) | [`acp`](../packages/ui/acp), [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex), [`repeat-tool-guard`](../packages/guard/repeat-tool-guard) |
| `agent/queued` | `emit` | [`packages/core/agent/src/types.ts:360`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`emit`) | - |
| `agent/request` | `waterfall` | [`packages/core/agent/src/types.ts:485`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`waterfall`) | - |
| `agent/session-prefix` | `waterfall` | [`packages/core/agent/src/types.ts:537`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`waterfall`) | [`tool-skill`](../packages/skill/tool-skill) |
| `agent/session-start` | `emit` | [`packages/core/agent/src/types.ts:381`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`emit`) | [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex), [`invariants`](../packages/support/invariants) |
| `agent/status` | `emit` | [`packages/core/agent/src/types.ts:345`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`emit`) | [`invariants`](../packages/support/invariants), [`repeat-tool-guard`](../packages/guard/repeat-tool-guard), [`stdio-agent`](../packages/ui/stdio-agent) |
| `agent/step-result` | `waterfall` | [`packages/core/agent/src/types.ts:552`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`waterfall`) | - |
| `agent/turn-continuation` | `waterfall` | [`packages/core/agent/src/types.ts:570`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`waterfall`) | [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex) |
| `agent/turn-stop` | `serial` | [`packages/core/agent/src/types.ts:588`](../packages/core/agent/src/types.ts) | [`agent-loop`](../packages/core/agent-loop) (`serial`) | [`subagent-inprocess`](../packages/subagent/subagent-inprocess) |
| `approval/request` | `waterfall` | [`packages/ui/user-approval/src/index.ts:70`](../packages/ui/user-approval/src/index.ts) | [`user-approval`](../packages/ui/user-approval) (`waterfall`) | [`acp`](../packages/ui/acp) |
| `fs/edit-intent` | `waterfall` | [`packages/fs/fs/src/index.ts:123`](../packages/fs/fs/src/index.ts) | [`tool-fs`](../packages/fs/tool-fs) (`waterfall`) | [`fs-policy`](../packages/fs/fs-policy) |
| `fs/observed` | `emit` | [`packages/fs/fs/src/index.ts:138`](../packages/fs/fs/src/index.ts) | [`tool-fs`](../packages/fs/tool-fs) (`emit`) | [`fs-policy`](../packages/fs/fs-policy) |
| `fs/write-intent` | `waterfall` | [`packages/fs/fs/src/index.ts:109`](../packages/fs/fs/src/index.ts) | [`tool-fs`](../packages/fs/tool-fs) (`waterfall`) | [`fs-policy`](../packages/fs/fs-policy) |
| `llm/stream` | `waterfall` | [`packages/llm/llm/src/index.ts:39`](../packages/llm/llm/src/index.ts) | [`llm`](../packages/llm/llm) (`waterfall`) | [`invariants`](../packages/support/invariants), [`llm-replay`](../packages/support/llm-replay) |
| `session/created` | `emit` | [`packages/core/session/src/index.ts:39`](../packages/core/session/src/index.ts) | [`session`](../packages/core/session) (`emit`) | [`invariants`](../packages/support/invariants), [`session-persistence`](../packages/session-persistence/session-persistence) |
| `session/event` | `emit` | [`packages/core/session/src/index.ts:47`](../packages/core/session/src/index.ts) | [`session`](../packages/core/session) (`emit`) | [`acp`](../packages/ui/acp), [`invariants`](../packages/support/invariants), [`session-persistence`](../packages/session-persistence/session-persistence), [`stdio-agent`](../packages/ui/stdio-agent) |
| `session/flush` | `parallel` | [`packages/core/session/src/index.ts:57`](../packages/core/session/src/index.ts) | [`agent-loop`](../packages/core/agent-loop) (`parallel`) | [`session-persistence`](../packages/session-persistence/session-persistence) |
| `subagent/end` | `emit` | [`packages/subagent/subagent/src/index.ts:98`](../packages/subagent/subagent/src/index.ts) | [`subagent`](../packages/subagent/subagent) (`events.dispatch`) | [`hooks-claude`](../packages/hooks/hooks-claude) |
| `subagent/provider-added` | `emit` | [`packages/subagent/subagent/src/index.ts:72`](../packages/subagent/subagent/src/index.ts) | [`subagent`](../packages/subagent/subagent) (`emit`) | [`tool-subagent`](../packages/subagent/tool-subagent) |
| `subagent/provider-removed` | `emit` | [`packages/subagent/subagent/src/index.ts:83`](../packages/subagent/subagent/src/index.ts) | - | [`tool-subagent`](../packages/subagent/tool-subagent) |
| `subagent/start` | `emit` | [`packages/subagent/subagent/src/index.ts:91`](../packages/subagent/subagent/src/index.ts) | [`subagent`](../packages/subagent/subagent) (`events.dispatch`) | [`hooks-claude`](../packages/hooks/hooks-claude) |
| `system-prompt/assemble` | `waterfall` | [`packages/core/system-prompt/src/index.ts:38`](../packages/core/system-prompt/src/index.ts) | [`system-prompt`](../packages/core/system-prompt) (`waterfall`) | - |
| `system-prompt/change` | `emit` | [`packages/core/system-prompt/src/index.ts:44`](../packages/core/system-prompt/src/index.ts) | [`system-prompt`](../packages/core/system-prompt) (`emit`) | - |
| `tools/change` | `emit` | [`packages/core/tools/src/index.ts:97`](../packages/core/tools/src/index.ts) | [`tools`](../packages/core/tools) (`emit`) | - |
| `tools/post-execute` | `waterfall` | [`packages/core/tools/src/index.ts:92`](../packages/core/tools/src/index.ts) | [`tools`](../packages/core/tools) (`waterfall`) | [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex) |
| `tools/pre-execute` | `waterfall` | [`packages/core/tools/src/index.ts:76`](../packages/core/tools/src/index.ts) | [`tools`](../packages/core/tools) (`waterfall`) | [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex) |
| `session/created` | `emit` | [`packages/core/session/src/index.ts:52`](../packages/core/session/src/index.ts) | [`session`](../packages/core/session) (`events.dispatch`) | [`invariants`](../packages/support/invariants), [`session-persistence`](../packages/session-persistence/session-persistence) |
| `session/disposed` | `emit` | [`packages/core/session/src/index.ts:64`](../packages/core/session/src/index.ts) | [`session`](../packages/core/session) (`events.dispatch`) | - |
| `session/event` | `emit` | [`packages/core/session/src/index.ts:83`](../packages/core/session/src/index.ts) | [`session`](../packages/core/session) (`events.dispatch`) | [`acp`](../packages/ui/acp), [`invariants`](../packages/support/invariants), [`session-persistence`](../packages/session-persistence/session-persistence), [`stdio-agent`](../packages/ui/stdio-agent) |
| `session/flush` | `parallel` | [`packages/core/session/src/index.ts:101`](../packages/core/session/src/index.ts) | [`session`](../packages/core/session) (`events.dispatch`) | [`session-persistence`](../packages/session-persistence/session-persistence) |
| `skill/provider-added` | `emit` | [`packages/skill/skill/src/index.ts:131`](../packages/skill/skill/src/index.ts) | [`skill`](../packages/skill/skill) (`emit`) | - |
| `skill/provider-removed` | `emit` | [`packages/skill/skill/src/index.ts:137`](../packages/skill/skill/src/index.ts) | [`skill`](../packages/skill/skill) (`emit`) | - |
| `subagent/end` | `emit` | [`packages/subagent/subagent/src/index.ts:90`](../packages/subagent/subagent/src/index.ts) | [`subagent`](../packages/subagent/subagent) (`events.dispatch`) | [`hooks-claude`](../packages/hooks/hooks-claude) |
| `subagent/provider-added` | `emit` | [`packages/subagent/subagent/src/index.ts:66`](../packages/subagent/subagent/src/index.ts) | [`subagent`](../packages/subagent/subagent) (`emit`) | [`tool-subagent`](../packages/subagent/tool-subagent) |
| `subagent/provider-removed` | `emit` | [`packages/subagent/subagent/src/index.ts:72`](../packages/subagent/subagent/src/index.ts) | [`subagent`](../packages/subagent/subagent) (`events.dispatch`) | [`tool-subagent`](../packages/subagent/tool-subagent) |
| `subagent/start` | `emit` | [`packages/subagent/subagent/src/index.ts:82`](../packages/subagent/subagent/src/index.ts) | [`subagent`](../packages/subagent/subagent) (`events.dispatch`) | [`hooks-claude`](../packages/hooks/hooks-claude) |
| `system-prompt/assemble` | `waterfall` | [`packages/core/system-prompt/src/index.ts:49`](../packages/core/system-prompt/src/index.ts) | [`system-prompt`](../packages/core/system-prompt) (`waterfall`) | - |
| `system-prompt/change` | `emit` | [`packages/core/system-prompt/src/index.ts:59`](../packages/core/system-prompt/src/index.ts) | [`system-prompt`](../packages/core/system-prompt) (`emit`) | - |
| `tools/change` | `emit` | [`packages/core/tools/src/index.ts:173`](../packages/core/tools/src/index.ts) | [`tools`](../packages/core/tools) (`emit`) | - |
| `tools/execute` | `waterfall` | [`packages/core/tools/src/index.ts:128`](../packages/core/tools/src/index.ts) | [`tools`](../packages/core/tools) (`waterfall`) | [`timeout-policy`](../packages/timeout/timeout-policy) |
| `tools/post-execute` | `waterfall` | [`packages/core/tools/src/index.ts:148`](../packages/core/tools/src/index.ts) | [`tools`](../packages/core/tools) (`waterfall`) | [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex), [`repeat-tool-guard`](../packages/guard/repeat-tool-guard) |
| `tools/pre-execute` | `waterfall` | [`packages/core/tools/src/index.ts:101`](../packages/core/tools/src/index.ts) | [`tools`](../packages/core/tools) (`waterfall`) | [`hooks-claude`](../packages/hooks/hooks-claude), [`hooks-codex`](../packages/hooks/hooks-codex) |
| `tools/result` | `emit` | [`packages/core/tools/src/index.ts:163`](../packages/core/tools/src/index.ts) | [`tools`](../packages/core/tools) (`events.dispatch`) | [`subagent-inprocess`](../packages/subagent/subagent-inprocess) |
| `workflow/agent-end` | `emit` | [`packages/workflow/workflow/src/index.ts:96`](../packages/workflow/workflow/src/index.ts) | [`workflow`](../packages/workflow/workflow) (`events.dispatch`) | - |
| `workflow/agent-start` | `emit` | [`packages/workflow/workflow/src/index.ts:85`](../packages/workflow/workflow/src/index.ts) | [`workflow`](../packages/workflow/workflow) (`events.dispatch`) | - |
| `workflow/end` | `emit` | [`packages/workflow/workflow/src/index.ts:106`](../packages/workflow/workflow/src/index.ts) | [`workflow`](../packages/workflow/workflow) (`events.dispatch`) | - |
| `workflow/log` | `emit` | [`packages/workflow/workflow/src/index.ts:75`](../packages/workflow/workflow/src/index.ts) | [`workflow`](../packages/workflow/workflow) (`events.dispatch`) | - |
| `workflow/phase` | `emit` | [`packages/workflow/workflow/src/index.ts:68`](../packages/workflow/workflow/src/index.ts) | [`workflow`](../packages/workflow/workflow) (`events.dispatch`) | - |
| `workflow/start` | `emit` | [`packages/workflow/workflow/src/index.ts:60`](../packages/workflow/workflow/src/index.ts) | [`workflow`](../packages/workflow/workflow) (`events.dispatch`) | - |
## Non-harness or undeclared event strings seen in package source
| Event string | Dispatchers | Listeners |
| --- | --- | --- |
| `internal/dispatch` | - | [`invariants`](../packages/support/invariants) |
Maintenance mode: hybrid generated: Cordis event declarations and most producer/listener edges are AST-scanned; dynamic dispatch sites are classified in `scripts/gen-doc-graphs.ts`.

17
docs/glossary.md Normal file
View File

@@ -0,0 +1,17 @@
# Glossary
Domain vocabulary for the DeepSeek Harness SDK uses one canonical term per concept. Terms link to their entries with standard Markdown anchors; implementation detail stays in package READMEs and RFCs.
FIXME(glossary-completeness): Expand this glossary before the first release so it covers the SDK's other core and capability subsystems, not only agent scope.
## agent-scope
- **scope** — the unit of per-agent registration: a contribution (tool, prompt section, variable, restriction, listener) is either *global* (visible to every agent) or *scoped* (owned by exactly one [scope key](#scope-key)). Two levels, flat: scoped registrations do not inherit down to subagents; subtree behavior is expressed with [lineage](#lineage) data, never scope structure.
- **scope key** — the opaque identity a scope is keyed by, compared by object identity. The harness convention: a live agent is the key of its own scope. <a id="scope-key"></a>
- **agent context (`agent.ctx`)** — the agent's scoped context; registrations through it are scope-visible AND scope-lifetime (one fact drives both), and listeners on it participate in that agent's scope-filtered dispatches. Registry-subject events may remain deliberately unfiltered under their own event contracts.
- **scope carrier** — the `thisArg` a scope-filtered dispatch carries (built by `scopeTarget`); its filter admits untagged listeners plus the subject's own. A *subject-less* carrier (no key) admits untagged listeners only.
- **scoped dispatch** — the rule: an event about one agent's activity dispatches with that agent's carrier. Events about a registry itself (a tool was added) are *registry-subject* and stay unfiltered.
- **shadowing** — most-specific-wins name resolution: a scoped tool/section/variable replaces its same-named global twin for that scope alone. The per-agent persona and per-agent tool-variant mechanism.
- **restriction / scope-local registration** — a restriction (`tools.restrict`) filters the GLOBAL tool surface for one scope (compose by intersection); scope-local registrations are merged after that filter. A filtered-away global tool is absent from the prompt AND refuses execution, indistinguishably from a nonexistent one.
- **setup window** — the creation slot where a creator composes an agent's scoped world (`CreateAgentOptions.setup`): after the scope and agent object exist but before the agent or session is published, `agent/session-start` fires, or the first prompt is assembled. Setup registers; it never drives the agent.
- **lineage** — parent/child facts carried as data (`parentSession`, `subagentDepth`); never affects visibility. <a id="lineage"></a>

1
docs/graph-atlas.md
View File

@@ -14,6 +14,7 @@ The process decision behind this index is recorded in [the documentation graph R
| [capability seams and core services](capability-seams.md) | `hybrid generated` |
| [echo-agent app composition](../examples/echo-agent/composition.md) | `hybrid generated` |
| [coding-agent app composition](../examples/coding-agent/composition.md) | `hybrid generated` |
| [cordis-agent app composition](../examples/cordis-agent/composition.md) | `hybrid generated` |
| [acp-agent app composition](../examples/acp-agent/composition.md) | `hybrid generated` |
| [event producer/consumer matrix](event-producer-consumer.md) | `hybrid generated` |
| [agent turn and step lifecycle](agent-lifecycle.md) | `curated` |

173
docs/module-graph.md
View File

@@ -9,6 +9,7 @@ Inter-package dependencies among the `@deepseek-ai/dsh-*` harness packages, deri
flowchart TD
subgraph group_util["packages/util"]
pkg_brand["brand"]
pkg_timeout["timeout"]
end
subgraph group_llm["packages/llm"]
pkg_llm["llm"]
@@ -19,6 +20,7 @@ flowchart TD
pkg_agent["agent"]
pkg_agent_core["agent-core"]
pkg_agent_loop["agent-loop"]
pkg_scope["scope"]
pkg_session["session"]
pkg_system_prompt["system-prompt"]
pkg_tools["tools"]
@@ -26,6 +28,7 @@ flowchart TD
subgraph group_bash["packages/bash"]
pkg_bash["bash"]
pkg_bash_local["bash-local"]
pkg_bash_sandbox["bash-sandbox"]
pkg_tool_bash["tool-bash"]
end
subgraph group_fs["packages/fs"]
@@ -34,6 +37,11 @@ flowchart TD
pkg_fs_policy["fs-policy"]
pkg_tool_fs["tool-fs"]
end
subgraph group_skill["packages/skill"]
pkg_skill["skill"]
pkg_skill_local["skill-local"]
pkg_tool_skill["tool-skill"]
end
subgraph group_compact["packages/compact"]
pkg_compact["compact"]
pkg_compact_basic["compact-basic"]
@@ -44,6 +52,7 @@ flowchart TD
pkg_subagent_fork["subagent-fork"]
pkg_subagent_inprocess["subagent-inprocess"]
pkg_subagent_spawn["subagent-spawn"]
pkg_subagent_subprocess["subagent-subprocess"]
pkg_tool_subagent["tool-subagent"]
end
subgraph group_web["packages/web"]
@@ -54,9 +63,15 @@ flowchart TD
pkg_web_search_exa["web-search-exa"]
pkg_web_search_perplexity["web-search-perplexity"]
end
subgraph group_timeout["packages/timeout"]
pkg_timeout_policy["timeout-policy"]
end
subgraph group_todo["packages/todo"]
pkg_tool_todo["tool-todo"]
end
subgraph group_cordis["packages/cordis"]
pkg_tool_cordis["tool-cordis"]
end
subgraph group_hooks["packages/hooks"]
pkg_hook_protocol["hook-protocol"]
pkg_hooks_claude["hooks-claude"]
@@ -78,57 +93,106 @@ flowchart TD
pkg_acp_agent["acp-agent"]
pkg_app_boot["app-boot"]
pkg_stdio_agent["stdio-agent"]
pkg_tool_ask_user["tool-ask-user"]
pkg_user_approval["user-approval"]
pkg_user_interaction["user-interaction"]
end
subgraph group_code_runtime["packages/code-runtime"]
pkg_code_runtime["code-runtime"]
pkg_code_runtime_worker["code-runtime-worker"]
end
subgraph group_guard["packages/guard"]
pkg_repeat_tool_guard["repeat-tool-guard"]
end
subgraph group_mcp["packages/mcp"]
pkg_mcp_client["mcp-client"]
end
subgraph group_sandbox["packages/sandbox"]
pkg_sandbox["sandbox"]
pkg_sandbox_local["sandbox-local"]
end
subgraph group_workflow["packages/workflow"]
pkg_tool_workflow["tool-workflow"]
pkg_workflow["workflow"]
pkg_workflow_workerthread["workflow-workerthread"]
end
pkg_llm --> pkg_brand
pkg_bash --> pkg_brand
pkg_code_runtime_worker --> pkg_code_runtime
pkg_llm_deepseek --> pkg_llm
pkg_llm_pi_ai --> pkg_llm
pkg_session --> pkg_brand
pkg_session --> pkg_llm
pkg_session --> pkg_scope
pkg_system_prompt --> pkg_llm
pkg_bash_local --> pkg_bash
pkg_system_prompt --> pkg_scope
pkg_fs --> pkg_brand
pkg_fs --> pkg_llm
pkg_web --> pkg_llm
pkg_sandbox --> pkg_llm
pkg_agent --> pkg_brand
pkg_agent --> pkg_llm
pkg_agent --> pkg_scope
pkg_agent --> pkg_session
pkg_agent --> pkg_system_prompt
pkg_bash --> pkg_brand
pkg_bash --> pkg_sandbox
pkg_bash --> pkg_session
pkg_fs_local --> pkg_fs
pkg_fs_policy --> pkg_fs
pkg_skill_local --> pkg_fs
pkg_skill_local --> pkg_skill
pkg_compact --> pkg_llm
pkg_compact --> pkg_session
pkg_web_fetch_local --> pkg_timeout
pkg_web_fetch_local --> pkg_web
pkg_web_search_deepseek --> pkg_web
pkg_web_search_exa --> pkg_web
pkg_web_search_perplexity --> pkg_web
pkg_hook_protocol --> pkg_bash
pkg_hook_protocol --> pkg_session
pkg_session_persistence --> pkg_session
pkg_llm_replay --> pkg_llm
pkg_llm_replay --> pkg_session
pkg_tools --> pkg_agent
pkg_tools --> pkg_llm
pkg_tools --> pkg_system_prompt
pkg_sandbox_local --> pkg_llm
pkg_sandbox_local --> pkg_sandbox
pkg_bash_local --> pkg_bash
pkg_bash_local --> pkg_timeout
pkg_compact_basic --> pkg_agent
pkg_compact_basic --> pkg_compact
pkg_compact_basic --> pkg_llm
pkg_compact_basic --> pkg_session
pkg_hook_protocol --> pkg_bash
pkg_hook_protocol --> pkg_session
pkg_session_persistence_jsonl --> pkg_session
pkg_session_persistence_jsonl --> pkg_session_persistence
pkg_session_persistence_sqlite --> pkg_session
pkg_session_persistence_sqlite --> pkg_session_persistence
pkg_invariants --> pkg_agent
pkg_invariants --> pkg_llm
pkg_invariants --> pkg_scope
pkg_invariants --> pkg_session
pkg_user_approval --> pkg_agent
pkg_user_approval --> pkg_brand
pkg_user_approval --> pkg_llm
pkg_user_approval --> pkg_scope
pkg_user_approval --> pkg_session
pkg_user_approval --> pkg_system_prompt
pkg_user_interaction --> pkg_agent
pkg_user_interaction --> pkg_llm
pkg_workflow --> pkg_agent
pkg_workflow --> pkg_brand
pkg_workflow --> pkg_llm
pkg_tools --> pkg_agent
pkg_tools --> pkg_code_runtime
pkg_tools --> pkg_llm
pkg_tools --> pkg_scope
pkg_tools --> pkg_session
pkg_tools --> pkg_system_prompt
pkg_tools --> pkg_user_approval
pkg_bash_sandbox --> pkg_bash
pkg_bash_sandbox --> pkg_bash_local
pkg_bash_sandbox --> pkg_sandbox
pkg_agent_loop --> pkg_agent
pkg_agent_loop --> pkg_llm
pkg_agent_loop --> pkg_scope
pkg_agent_loop --> pkg_session
pkg_agent_loop --> pkg_session_persistence
pkg_agent_loop --> pkg_system_prompt
@@ -136,46 +200,76 @@ flowchart TD
pkg_tool_bash --> pkg_agent
pkg_tool_bash --> pkg_bash
pkg_tool_bash --> pkg_llm
pkg_tool_bash --> pkg_sandbox
pkg_tool_bash --> pkg_system_prompt
pkg_tool_bash --> pkg_tools
pkg_tool_bash --> pkg_user_approval
pkg_tool_fs --> pkg_fs
pkg_tool_fs --> pkg_llm
pkg_tool_fs --> pkg_session
pkg_tool_fs --> pkg_system_prompt
pkg_tool_fs --> pkg_tools
pkg_tool_skill --> pkg_agent
pkg_tool_skill --> pkg_llm
pkg_tool_skill --> pkg_skill
pkg_tool_skill --> pkg_tools
pkg_subagent --> pkg_agent
pkg_subagent --> pkg_llm
pkg_subagent --> pkg_scope
pkg_subagent --> pkg_tools
pkg_tool_web --> pkg_llm
pkg_tool_web --> pkg_system_prompt
pkg_tool_web --> pkg_tools
pkg_tool_web --> pkg_web
pkg_timeout_policy --> pkg_llm
pkg_timeout_policy --> pkg_timeout
pkg_timeout_policy --> pkg_tools
pkg_tool_todo --> pkg_agent
pkg_tool_todo --> pkg_session
pkg_tool_todo --> pkg_tools
pkg_tool_cordis --> pkg_scope
pkg_tool_cordis --> pkg_tools
pkg_hooks_codex --> pkg_agent
pkg_hooks_codex --> pkg_hook_protocol
pkg_hooks_codex --> pkg_llm
pkg_hooks_codex --> pkg_session
pkg_hooks_codex --> pkg_tools
pkg_acp --> pkg_agent
pkg_acp --> pkg_bash
pkg_acp --> pkg_llm
pkg_acp --> pkg_sandbox
pkg_acp --> pkg_session
pkg_acp --> pkg_session_persistence
pkg_acp --> pkg_tools
pkg_acp --> pkg_user_approval
pkg_acp --> pkg_user_interaction
pkg_tool_ask_user --> pkg_agent
pkg_tool_ask_user --> pkg_tools
pkg_tool_ask_user --> pkg_user_interaction
pkg_repeat_tool_guard --> pkg_agent
pkg_repeat_tool_guard --> pkg_tools
pkg_mcp_client --> pkg_llm
pkg_mcp_client --> pkg_tools
pkg_tool_workflow --> pkg_agent
pkg_tool_workflow --> pkg_llm
pkg_tool_workflow --> pkg_system_prompt
pkg_tool_workflow --> pkg_tools
pkg_tool_workflow --> pkg_workflow
pkg_agent_core --> pkg_agent
pkg_agent_core --> pkg_agent_loop
pkg_agent_core --> pkg_invariants
pkg_agent_core --> pkg_llm
pkg_agent_core --> pkg_session
pkg_agent_core --> pkg_skill
pkg_agent_core --> pkg_skill_local
pkg_agent_core --> pkg_system_prompt
pkg_agent_core --> pkg_tool_bash
pkg_agent_core --> pkg_tool_skill
pkg_agent_core --> pkg_tools
pkg_subagent_acp --> pkg_agent
pkg_subagent_acp --> pkg_llm
pkg_subagent_acp --> pkg_subagent
pkg_subagent_acp --> pkg_subagent_subprocess
pkg_subagent_inprocess --> pkg_agent
pkg_subagent_inprocess --> pkg_llm
pkg_subagent_inprocess --> pkg_session
@@ -195,6 +289,13 @@ flowchart TD
pkg_subagent_mock --> pkg_agent
pkg_subagent_mock --> pkg_llm
pkg_subagent_mock --> pkg_subagent
pkg_workflow_workerthread --> pkg_agent
pkg_workflow_workerthread --> pkg_brand
pkg_workflow_workerthread --> pkg_llm
pkg_workflow_workerthread --> pkg_session
pkg_workflow_workerthread --> pkg_subagent
pkg_workflow_workerthread --> pkg_tools
pkg_workflow_workerthread --> pkg_workflow
pkg_subagent_fork --> pkg_agent
pkg_subagent_fork --> pkg_session
pkg_subagent_fork --> pkg_subagent
@@ -205,61 +306,85 @@ flowchart TD
pkg_acp_agent --> pkg_agent_core
pkg_acp_agent --> pkg_app_boot
pkg_acp_agent --> pkg_session_persistence_jsonl
pkg_acp_agent --> pkg_tools
pkg_acp_agent --> pkg_user_interaction
pkg_stdio_agent --> pkg_agent
pkg_stdio_agent --> pkg_agent_core
pkg_stdio_agent --> pkg_app_boot
pkg_stdio_agent --> pkg_llm
pkg_stdio_agent --> pkg_session
pkg_stdio_agent --> pkg_session_persistence_jsonl
pkg_stdio_agent --> pkg_tool_ask_user
pkg_stdio_agent --> pkg_tools
pkg_stdio_agent --> pkg_user_interaction
```
| Package | Group | Depends on |
| --- | --- | --- |
| [`brand`](../packages/util/brand) | `util` | — |
| [`timeout`](../packages/util/timeout) | `util` | — |
| [`scope`](../packages/core/scope) | `core` | — |
| [`skill`](../packages/skill/skill) | `skill` | — |
| [`subagent-subprocess`](../packages/subagent/subagent-subprocess) | `subagent` | — |
| [`acp-snapshot`](../packages/support/acp-snapshot) | `support` | — |
| [`app-boot`](../packages/ui/app-boot) | `ui` | — |
| [`code-runtime`](../packages/code-runtime/code-runtime) | `code-runtime` | — |
| [`llm`](../packages/llm/llm) | `llm` | [`brand`](../packages/util/brand) |
| [`bash`](../packages/bash/bash) | `bash` | [`brand`](../packages/util/brand) |
| [`code-runtime-worker`](../packages/code-runtime/code-runtime-worker) | `code-runtime` | [`code-runtime`](../packages/code-runtime/code-runtime) |
| [`llm-deepseek`](../packages/llm/llm-deepseek) | `llm` | [`llm`](../packages/llm/llm) |
| [`llm-pi-ai`](../packages/llm/llm-pi-ai) | `llm` | [`llm`](../packages/llm/llm) |
| [`session`](../packages/core/session) | `core` | [`brand`](../packages/util/brand), [`llm`](../packages/llm/llm) |
| [`system-prompt`](../packages/core/system-prompt) | `core` | [`llm`](../packages/llm/llm) |
| [`bash-local`](../packages/bash/bash-local) | `bash` | [`bash`](../packages/bash/bash) |
| [`session`](../packages/core/session) | `core` | [`brand`](../packages/util/brand), [`llm`](../packages/llm/llm), [`scope`](../packages/core/scope) |
| [`system-prompt`](../packages/core/system-prompt) | `core` | [`llm`](../packages/llm/llm), [`scope`](../packages/core/scope) |
| [`fs`](../packages/fs/fs) | `fs` | [`brand`](../packages/util/brand), [`llm`](../packages/llm/llm) |
| [`web`](../packages/web/web) | `web` | [`llm`](../packages/llm/llm) |
| [`agent`](../packages/core/agent) | `core` | [`brand`](../packages/util/brand), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`system-prompt`](../packages/core/system-prompt) |
| [`sandbox`](../packages/sandbox/sandbox) | `sandbox` | [`llm`](../packages/llm/llm) |
| [`agent`](../packages/core/agent) | `core` | [`brand`](../packages/util/brand), [`llm`](../packages/llm/llm), [`scope`](../packages/core/scope), [`session`](../packages/core/session), [`system-prompt`](../packages/core/system-prompt) |
| [`bash`](../packages/bash/bash) | `bash` | [`brand`](../packages/util/brand), [`sandbox`](../packages/sandbox/sandbox), [`session`](../packages/core/session) |
| [`fs-local`](../packages/fs/fs-local) | `fs` | [`fs`](../packages/fs/fs) |
| [`fs-policy`](../packages/fs/fs-policy) | `fs` | [`fs`](../packages/fs/fs) |
| [`skill-local`](../packages/skill/skill-local) | `skill` | [`fs`](../packages/fs/fs), [`skill`](../packages/skill/skill) |
| [`compact`](../packages/compact/compact) | `compact` | [`llm`](../packages/llm/llm), [`session`](../packages/core/session) |
| [`web-fetch-local`](../packages/web/web-fetch-local) | `web` | [`web`](../packages/web/web) |
| [`web-fetch-local`](../packages/web/web-fetch-local) | `web` | [`timeout`](../packages/util/timeout), [`web`](../packages/web/web) |
| [`web-search-deepseek`](../packages/web/web-search-deepseek) | `web` | [`web`](../packages/web/web) |
| [`web-search-exa`](../packages/web/web-search-exa) | `web` | [`web`](../packages/web/web) |
| [`web-search-perplexity`](../packages/web/web-search-perplexity) | `web` | [`web`](../packages/web/web) |
| [`hook-protocol`](../packages/hooks/hook-protocol) | `hooks` | [`bash`](../packages/bash/bash), [`session`](../packages/core/session) |
| [`session-persistence`](../packages/session-persistence/session-persistence) | `session-persistence` | [`session`](../packages/core/session) |
| [`llm-replay`](../packages/support/llm-replay) | `support` | [`llm`](../packages/llm/llm), [`session`](../packages/core/session) |
| [`tools`](../packages/core/tools) | `core` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`system-prompt`](../packages/core/system-prompt) |
| [`sandbox-local`](../packages/sandbox/sandbox-local) | `sandbox` | [`llm`](../packages/llm/llm), [`sandbox`](../packages/sandbox/sandbox) |
| [`bash-local`](../packages/bash/bash-local) | `bash` | [`bash`](../packages/bash/bash), [`timeout`](../packages/util/timeout) |
| [`compact-basic`](../packages/compact/compact-basic) | `compact` | [`agent`](../packages/core/agent), [`compact`](../packages/compact/compact), [`llm`](../packages/llm/llm), [`session`](../packages/core/session) |
| [`hook-protocol`](../packages/hooks/hook-protocol) | `hooks` | [`bash`](../packages/bash/bash), [`session`](../packages/core/session) |
| [`session-persistence-jsonl`](../packages/session-persistence/session-persistence-jsonl) | `session-persistence` | [`session`](../packages/core/session), [`session-persistence`](../packages/session-persistence/session-persistence) |
| [`session-persistence-sqlite`](../packages/session-persistence/session-persistence-sqlite) | `session-persistence` | [`session`](../packages/core/session), [`session-persistence`](../packages/session-persistence/session-persistence) |
| [`invariants`](../packages/support/invariants) | `support` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`session`](../packages/core/session) |
| [`agent-loop`](../packages/core/agent-loop) | `core` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`session-persistence`](../packages/session-persistence/session-persistence), [`system-prompt`](../packages/core/system-prompt), [`tools`](../packages/core/tools) |
| [`tool-bash`](../packages/bash/tool-bash) | `bash` | [`agent`](../packages/core/agent), [`bash`](../packages/bash/bash), [`llm`](../packages/llm/llm), [`system-prompt`](../packages/core/system-prompt), [`tools`](../packages/core/tools) |
| [`invariants`](../packages/support/invariants) | `support` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`scope`](../packages/core/scope), [`session`](../packages/core/session) |
| [`user-approval`](../packages/ui/user-approval) | `ui` | [`agent`](../packages/core/agent), [`brand`](../packages/util/brand), [`llm`](../packages/llm/llm), [`scope`](../packages/core/scope), [`session`](../packages/core/session), [`system-prompt`](../packages/core/system-prompt) |
| [`user-interaction`](../packages/ui/user-interaction) | `ui` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm) |
| [`workflow`](../packages/workflow/workflow) | `workflow` | [`agent`](../packages/core/agent), [`brand`](../packages/util/brand), [`llm`](../packages/llm/llm) |
| [`tools`](../packages/core/tools) | `core` | [`agent`](../packages/core/agent), [`code-runtime`](../packages/code-runtime/code-runtime), [`llm`](../packages/llm/llm), [`scope`](../packages/core/scope), [`session`](../packages/core/session), [`system-prompt`](../packages/core/system-prompt), [`user-approval`](../packages/ui/user-approval) |
| [`bash-sandbox`](../packages/bash/bash-sandbox) | `bash` | [`bash`](../packages/bash/bash), [`bash-local`](../packages/bash/bash-local), [`sandbox`](../packages/sandbox/sandbox) |
| [`agent-loop`](../packages/core/agent-loop) | `core` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`scope`](../packages/core/scope), [`session`](../packages/core/session), [`session-persistence`](../packages/session-persistence/session-persistence), [`system-prompt`](../packages/core/system-prompt), [`tools`](../packages/core/tools) |
| [`tool-bash`](../packages/bash/tool-bash) | `bash` | [`agent`](../packages/core/agent), [`bash`](../packages/bash/bash), [`llm`](../packages/llm/llm), [`sandbox`](../packages/sandbox/sandbox), [`system-prompt`](../packages/core/system-prompt), [`tools`](../packages/core/tools), [`user-approval`](../packages/ui/user-approval) |
| [`tool-fs`](../packages/fs/tool-fs) | `fs` | [`fs`](../packages/fs/fs), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`system-prompt`](../packages/core/system-prompt), [`tools`](../packages/core/tools) |
| [`subagent`](../packages/subagent/subagent) | `subagent` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`tools`](../packages/core/tools) |
| [`tool-skill`](../packages/skill/tool-skill) | `skill` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`skill`](../packages/skill/skill), [`tools`](../packages/core/tools) |
| [`subagent`](../packages/subagent/subagent) | `subagent` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`scope`](../packages/core/scope), [`tools`](../packages/core/tools) |
| [`tool-web`](../packages/web/tool-web) | `web` | [`llm`](../packages/llm/llm), [`system-prompt`](../packages/core/system-prompt), [`tools`](../packages/core/tools), [`web`](../packages/web/web) |
| [`timeout-policy`](../packages/timeout/timeout-policy) | `timeout` | [`llm`](../packages/llm/llm), [`timeout`](../packages/util/timeout), [`tools`](../packages/core/tools) |
| [`tool-todo`](../packages/todo/tool-todo) | `todo` | [`agent`](../packages/core/agent), [`session`](../packages/core/session), [`tools`](../packages/core/tools) |
| [`tool-cordis`](../packages/cordis/tool-cordis) | `cordis` | [`scope`](../packages/core/scope), [`tools`](../packages/core/tools) |
| [`hooks-codex`](../packages/hooks/hooks-codex) | `hooks` | [`agent`](../packages/core/agent), [`hook-protocol`](../packages/hooks/hook-protocol), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`tools`](../packages/core/tools) |
| [`acp`](../packages/ui/acp) | `ui` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`session-persistence`](../packages/session-persistence/session-persistence), [`tools`](../packages/core/tools) |
| [`acp`](../packages/ui/acp) | `ui` | [`agent`](../packages/core/agent), [`bash`](../packages/bash/bash), [`llm`](../packages/llm/llm), [`sandbox`](../packages/sandbox/sandbox), [`session`](../packages/core/session), [`session-persistence`](../packages/session-persistence/session-persistence), [`tools`](../packages/core/tools), [`user-approval`](../packages/ui/user-approval), [`user-interaction`](../packages/ui/user-interaction) |
| [`tool-ask-user`](../packages/ui/tool-ask-user) | `ui` | [`agent`](../packages/core/agent), [`tools`](../packages/core/tools), [`user-interaction`](../packages/ui/user-interaction) |
| [`repeat-tool-guard`](../packages/guard/repeat-tool-guard) | `guard` | [`agent`](../packages/core/agent), [`tools`](../packages/core/tools) |
| [`mcp-client`](../packages/mcp/mcp-client) | `mcp` | [`llm`](../packages/llm/llm), [`tools`](../packages/core/tools) |
| [`agent-core`](../packages/core/agent-core) | `core` | [`agent`](../packages/core/agent), [`agent-loop`](../packages/core/agent-loop), [`invariants`](../packages/support/invariants), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`system-prompt`](../packages/core/system-prompt), [`tool-bash`](../packages/bash/tool-bash), [`tools`](../packages/core/tools) |
| [`subagent-acp`](../packages/subagent/subagent-acp) | `subagent` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`subagent`](../packages/subagent/subagent) |
| [`tool-workflow`](../packages/workflow/tool-workflow) | `workflow` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`system-prompt`](../packages/core/system-prompt), [`tools`](../packages/core/tools), [`workflow`](../packages/workflow/workflow) |
| [`agent-core`](../packages/core/agent-core) | `core` | [`agent`](../packages/core/agent), [`agent-loop`](../packages/core/agent-loop), [`invariants`](../packages/support/invariants), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`skill`](../packages/skill/skill), [`skill-local`](../packages/skill/skill-local), [`system-prompt`](../packages/core/system-prompt), [`tool-bash`](../packages/bash/tool-bash), [`tool-skill`](../packages/skill/tool-skill), [`tools`](../packages/core/tools) |
| [`subagent-acp`](../packages/subagent/subagent-acp) | `subagent` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`subagent`](../packages/subagent/subagent), [`subagent-subprocess`](../packages/subagent/subagent-subprocess) |
| [`subagent-inprocess`](../packages/subagent/subagent-inprocess) | `subagent` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`subagent`](../packages/subagent/subagent), [`system-prompt`](../packages/core/system-prompt), [`tools`](../packages/core/tools) |
| [`tool-subagent`](../packages/subagent/tool-subagent) | `subagent` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`subagent`](../packages/subagent/subagent), [`tools`](../packages/core/tools) |
| [`hooks-claude`](../packages/hooks/hooks-claude) | `hooks` | [`agent`](../packages/core/agent), [`hook-protocol`](../packages/hooks/hook-protocol), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`subagent`](../packages/subagent/subagent), [`tools`](../packages/core/tools) |
| [`subagent-mock`](../packages/support/subagent-mock) | `support` | [`agent`](../packages/core/agent), [`llm`](../packages/llm/llm), [`subagent`](../packages/subagent/subagent) |
| [`workflow-workerthread`](../packages/workflow/workflow-workerthread) | `workflow` | [`agent`](../packages/core/agent), [`brand`](../packages/util/brand), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`subagent`](../packages/subagent/subagent), [`tools`](../packages/core/tools), [`workflow`](../packages/workflow/workflow) |
| [`subagent-fork`](../packages/subagent/subagent-fork) | `subagent` | [`agent`](../packages/core/agent), [`session`](../packages/core/session), [`subagent`](../packages/subagent/subagent), [`subagent-inprocess`](../packages/subagent/subagent-inprocess) |
| [`subagent-spawn`](../packages/subagent/subagent-spawn) | `subagent` | [`subagent`](../packages/subagent/subagent), [`subagent-inprocess`](../packages/subagent/subagent-inprocess) |
| [`acp-agent`](../packages/ui/acp-agent) | `ui` | [`acp`](../packages/ui/acp), [`agent-core`](../packages/core/agent-core), [`app-boot`](../packages/ui/app-boot), [`session-persistence-jsonl`](../packages/session-persistence/session-persistence-jsonl) |
| [`stdio-agent`](../packages/ui/stdio-agent) | `ui` | [`agent`](../packages/core/agent), [`agent-core`](../packages/core/agent-core), [`app-boot`](../packages/ui/app-boot), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`session-persistence-jsonl`](../packages/session-persistence/session-persistence-jsonl) |
| [`acp-agent`](../packages/ui/acp-agent) | `ui` | [`acp`](../packages/ui/acp), [`agent-core`](../packages/core/agent-core), [`app-boot`](../packages/ui/app-boot), [`session-persistence-jsonl`](../packages/session-persistence/session-persistence-jsonl), [`tools`](../packages/core/tools), [`user-interaction`](../packages/ui/user-interaction) |
| [`stdio-agent`](../packages/ui/stdio-agent) | `ui` | [`agent`](../packages/core/agent), [`agent-core`](../packages/core/agent-core), [`app-boot`](../packages/ui/app-boot), [`llm`](../packages/llm/llm), [`session`](../packages/core/session), [`session-persistence-jsonl`](../packages/session-persistence/session-persistence-jsonl), [`tool-ask-user`](../packages/ui/tool-ask-user), [`tools`](../packages/core/tools), [`user-interaction`](../packages/ui/user-interaction) |

92
docs/persistence-catalog.md
View File

@@ -11,6 +11,40 @@ The on-disk envelope around every payload is `SessionEvent` — `type`, monotoni
## Events
### `approval/*`
#### `approval/asked` — log-only
An approval question was put to the answerer chain — log-only audit (like `hook/*`; NOT a surface event, carries no `surfaceOp`). `id` pairs it with the `approval/decided` that always follows; `toolName` is the tool the question is about, `callId` the exact tool call when the asker had one, `reason` the asker's human-readable explanation (e.g. a hook's permission-decision reason).
```ts persistence-catalog
'approval/asked': { id: ApprovalRequestId; toolName: string; callId?: CallId; reason?: string }
```
Types: [CallId](core-data-structures/core.md)
Source: [`packages/ui/user-approval/src/index.ts:84`](../packages/ui/user-approval/src/index.ts)
#### `approval/decided` — log-only
The outcome of a prior `approval/asked` (same `id`) — log-only audit. Exactly one per ask, appended when the outcome is known: a decision, a cancellation, or the fail-closed `'unavailable'`.
```ts persistence-catalog
'approval/decided': { id: ApprovalRequestId; outcome: ApprovalOutcome }
```
Source: [`packages/ui/user-approval/src/index.ts:95`](../packages/ui/user-approval/src/index.ts)
#### `approval/policy` — log-only
The session's approval policy was switched — log-only, durable, replayable, never in the model transcript (the model learns the policy from the prompt section and the narrator's notices). The LAST such event is the session's override (effectiveApprovalPolicy); who asked for it is derivable from position (an event after the log's last `request/header*` was a runtime switch by the user).
```ts persistence-catalog
'approval/policy': { policy: ApprovalPolicy }
```
Source: [`packages/ui/user-approval/src/index.ts:107`](../packages/ui/user-approval/src/index.ts)
### `assistant/*`
#### `assistant/chunk` — log-only
@@ -23,7 +57,7 @@ Raw stream chunk — token-level replay fidelity.
Types: [StreamChunk](core-data-structures/llm-streaming.md)
Source: [`packages/core/session/src/types.ts:304`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:322`](../packages/core/session/src/types.ts)
#### `assistant/message` — surface
@@ -35,7 +69,19 @@ Assembled assistant message for one step (derived history uses this). Carries th
Types: [ContentBlock](core-data-structures/core.md) · [TokenUsage](core-data-structures/llm-streaming.md)
Source: [`packages/core/session/src/types.ts:311`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:329`](../packages/core/session/src/types.ts)
### `bash/*`
#### `bash/sandbox-mode` — log-only
The session's sandbox mode was switched — log-only (like `approval/*`; NOT a surface event, carries no `surfaceOp`): durable and replayable, never in the model transcript. The LAST such event is the session's override (effectiveSandboxMode); who asked for it is derivable from position (an event after the log's last `request/header*` was a runtime switch by the user; see the tool layer's narrator).
```ts persistence-catalog
'bash/sandbox-mode': { mode: SandboxMode }
```
Source: [`packages/bash/bash/src/session-mode.ts:31`](../packages/bash/bash/src/session-mode.ts)
### `compact/*`
@@ -83,7 +129,7 @@ In-session context injection (file-change notices, subdir AGENTS.md, skill conte
Types: [ContentBlock](core-data-structures/core.md) · [MessageSource](core-data-structures/core.md)
Source: [`packages/core/session/src/types.ts:302`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:320`](../packages/core/session/src/types.ts)
### `hook/*`
@@ -119,7 +165,7 @@ A queued prompt an `agent/prompt-submit` listener VETOED — the durable record
Types: [ContentBlock](core-data-structures/core.md) · [MessageSource](core-data-structures/core.md)
Source: [`packages/core/session/src/types.ts:296`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:314`](../packages/core/session/src/types.ts)
### `request/*`
@@ -131,17 +177,17 @@ Full snapshot of the EpochHeader the NEXT request is built under, with the Reque
'request/header': { header: EpochHeader; reason: RequestHeaderReason }
```
Source: [`packages/core/session/src/types.ts:356`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:374`](../packages/core/session/src/types.ts)
#### `request/header-delta` — log-only
Amendment to the folded EpochHeader: at least one of a SystemDelta, a ToolsDelta, or a whole replacement LlmCallConfig (four scalars — not worth diffing). Appended by the loop inside the step, before dispatch, when the header for this request differs from the fold of the log so far; the writer verifies `applyHeaderDelta(previous, delta)` reproduces the new header exactly and falls back to a `'fallback'` `request/header` snapshot when it cannot, so a logged delta ALWAYS round-trips. NOT a SurfaceEventType.
Amendment to the folded EpochHeader: at least one of a SystemDelta, a ToolsDelta, a whole replacement LlmCallConfig (four scalars — not worth diffing), or a whole replacement session prefix (`messagePrefix` — small advisory content, replaced whole; an EMPTY array encodes the transition to "none", mirroring the canonical form's absent field — the loop never produces one in practice: the prefix is composed once per instance and anchored by that instance's snapshot, so this arm exists for codec totality). Appended by the loop inside the step, before dispatch, when the header for this request differs from the fold of the log so far; the writer verifies `applyHeaderDelta(previous, delta)` reproduces the new header exactly and falls back to a `'fallback'` `request/header` snapshot when it cannot, so a logged delta ALWAYS round-trips. NOT a SurfaceEventType.
```ts persistence-catalog
'request/header-delta': { system?: SystemDelta; tools?: ToolsDelta; config?: LlmCallConfig }
'request/header-delta': { system?: SystemDelta; tools?: ToolsDelta; config?: LlmCallConfig; messagePrefix?: Message[] }
```
Source: [`packages/core/session/src/types.ts:367`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:391`](../packages/core/session/src/types.ts)
### `steering/*`
@@ -155,7 +201,7 @@ Steering content injected between steps of a running turn.
Types: [ContentBlock](core-data-structures/core.md) · [MessageSource](core-data-structures/core.md)
Source: [`packages/core/session/src/types.ts:329`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:347`](../packages/core/session/src/types.ts)
### `step/*`
@@ -167,7 +213,7 @@ Closes step `step` of turn `turn`.
'step/end': { turn: number; step: number }
```
Source: [`packages/core/session/src/types.ts:283`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:301`](../packages/core/session/src/types.ts)
#### `step/start` — log-only
@@ -177,7 +223,7 @@ Opens step `step` of turn `turn` — one model call plus the tool executions it
'step/start': { turn: number; step: number }
```
Source: [`packages/core/session/src/types.ts:281`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:299`](../packages/core/session/src/types.ts)
### `todo/*`
@@ -193,7 +239,7 @@ NOT a SurfaceEventType: it produces no LLM message and never reaches `deriveMess
Types: [TodoItem](core-data-structures/session.md)
Source: [`packages/core/session/src/types.ts:343`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:361`](../packages/core/session/src/types.ts)
### `tool/*`
@@ -207,7 +253,19 @@ The model requested one tool invocation: `name` with the raw `arguments` JSON st
Types: [CallId](core-data-structures/core.md)
Source: [`packages/core/session/src/types.ts:317`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:335`](../packages/core/session/src/types.ts)
#### `tool/code-dispatch` — log-only
One bridged sub-dispatch from a `run_code` program: the parent `run_code` call id, the deterministic sub-call id (`<parent>:code:<n>`), the tool `name` with its JSON-normalized `arguments` — the exact value dispatched, normalized BEFORE dispatch, so this append can never fail on payload shape — whether the sub-call errored, and a bounded `resultSummary` of its model-facing text. Log-only: `deriveMessages()` ignores it, so sub-calls never re-enter model context; persistence and UIs get every call. Appended inside the parent `run_code`'s execution (the bridge drains its queue before returning), so the turn-enclosure invariant holds by construction.
```ts persistence-catalog
'tool/code-dispatch': { parentCallId: CallId; subCallId: CallId; name: string; arguments: unknown; isError: boolean; resultSummary: string }
```
Types: [CallId](core-data-structures/core.md)
Source: [`packages/core/tools/src/code-mode.ts:38`](../packages/core/tools/src/code-mode.ts)
#### `tool/result` — surface
@@ -219,7 +277,7 @@ A completed tool call's model-facing result, plus an optional tool-private `meta
Types: [CallId](core-data-structures/core.md) · [ContentBlock](core-data-structures/core.md)
Source: [`packages/core/session/src/types.ts:327`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:345`](../packages/core/session/src/types.ts)
### `turn/*`
@@ -233,7 +291,7 @@ Closes turn `turn` with the TurnEndReason that ended it. The loop fires the awai
Types: [TurnEndReason](core-data-structures/session.md)
Source: [`packages/core/session/src/types.ts:279`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:297`](../packages/core/session/src/types.ts)
#### `turn/start` — log-only
@@ -245,7 +303,7 @@ Opens turn `turn`. `trigger` records what started it — a drained message batch
Types: [TurnTrigger](core-data-structures/session.md)
Source: [`packages/core/session/src/types.ts:273`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:291`](../packages/core/session/src/types.ts)
### `user/*`
@@ -259,4 +317,4 @@ A user-visible prompt (queued message drained at turn start).
Types: [ContentBlock](core-data-structures/core.md) · [MessageSource](core-data-structures/core.md)
Source: [`packages/core/session/src/types.ts:285`](../packages/core/session/src/types.ts)
Source: [`packages/core/session/src/types.ts:303`](../packages/core/session/src/types.ts)

31
docs/rfc/INDEX.md
View File

@@ -8,19 +8,20 @@ Generated by `pnpm run gen-rfc-index` from the RFC tree — never edit by hand;
| Title | First proposed |
|---|---|
| [Agent Client Protocol (ACP) support — drive the coding agent from external editors](proposed/feature/2026-06-14-acp-agent-client-protocol.md) | 2026-06-14 |
| [Multiplex concurrent ACP sessions over one connection](proposed/feature/2026-06-14-acp-multi-session.md) | 2026-06-14 |
| [Code Mode — the model writes TypeScript against the tool registry](proposed/feature/2026-06-15-code-mode.md) | 2026-06-15 |
| [Pre-tool input rewrite — a consistent design](proposed/feature/2026-06-30-pre-tool-input-rewrite.md) | 2026-06-30 |
| [Claude Code and Codex subagent backends (out-of-process delegation to external coding agents)](proposed/feature/2026-07-07-claude-code-and-codex-subagent-backends.md) | 2026-07-07 |
| [Repeat-tool-call guard plugin](proposed/feature/2026-07-08-repeat-tool-guard.md) | 2026-07-08 |
| [Interactive side sessions and merge-back](proposed/feature/2026-07-08-interactive-side-sessions.md) | 2026-07-08 |
| [Stream workflow progress through tool calls](proposed/feature/2026-07-13-stream-workflow-progress-through-tool-calls.md) | 2026-07-13 |
### Simplification
| Title | First proposed |
|---|---|
| [Unify the agent id and the session id](proposed/simplification/2026-06-20-unify-agent-and-session-id.md) | 2026-06-20 |
| [Prune dead core-spine surface — `SurfaceManager.invalidate()`, the loop-internal exports, `ToolExecutionResult.callId`](proposed/simplification/2026-07-04-prune-dead-core-spine-surface.md) | 2026-07-04 |
| [Prune dead public and result surface](proposed/simplification/2026-07-04-prune-dead-core-spine-surface.md) | 2026-07-04 |
| [Drop unconsumed skill provider events](proposed/simplification/2026-07-12-drop-unconsumed-skill-provider-events.md) | 2026-07-12 |
| [Prune unused web seam fields](proposed/simplification/2026-07-12-prune-unused-web-seam-fields.md) | 2026-07-12 |
| [Simplify session-log representation](proposed/simplification/2026-07-12-simplify-session-log-representation.md) | 2026-07-12 |
### Architecture
@@ -51,19 +52,31 @@ Generated by `pnpm run gen-rfc-index` from the RFC tree — never edit by hand;
| Title | First proposed |
|---|---|
| [Agent Client Protocol (ACP) support — drive the coding agent from external editors](implemented/feature/2026-06-14-acp-agent-client-protocol.md) | 2026-06-14 |
| [Multiplex concurrent ACP sessions over one connection](implemented/feature/2026-06-14-acp-multi-session.md) | 2026-06-14 |
| [Code Mode — the model writes TypeScript against the tool registry](implemented/feature/2026-06-15-code-mode.md) | 2026-06-15 |
| [Filesystem tool schemas — model-facing read/write/edit shapes](implemented/feature/2026-06-17-filesystem-tool-schemas.md) | 2026-06-17 |
| [Rich ACP bash rendering — the terminal card via the `_meta` convention](implemented/feature/2026-06-18-acp-terminal-and-tool-rendering.md) | 2026-06-18 |
| [Compaction as a capability seam (abstract contract + basic backend)](implemented/feature/2026-06-18-compaction-capability-seam.md) | 2026-06-18 |
| [Subagent capability seam](implemented/feature/2026-06-21-subagent-capability-seam.md) | 2026-06-21 |
| [ACP subagent backend (out-of-process delegation)](implemented/feature/2026-06-22-acp-subagent-backend.md) | 2026-06-22 |
| [Ask-user question capability](implemented/feature/2026-06-25-ask-user-question.md) | 2026-06-25 |
| [The `todo_write` tool — model task list as event-sourced session state](implemented/feature/2026-06-29-todo-write-tool.md) | 2026-06-29 |
| [dsh-hooks-claude + dsh-hooks-codex — the Claude Code / Codex hook bridges](implemented/feature/2026-06-30-hook-bridges.md) | 2026-06-30 |
| [dsh-hook-protocol — the shared Claude Code / Codex hook wire-protocol core](implemented/feature/2026-06-30-hook-protocol-lib.md) | 2026-06-30 |
| [Interception seams — the typed-Decision surface a hook programs against](implemented/feature/2026-06-30-interception-seams.md) | 2026-06-30 |
| [SessionStore fork API](implemented/feature/2026-06-30-session-store-fork-api.md) | 2026-06-30 |
| [Subagent lifecycle enrichment — lastAssistantMessage (observe-only)](implemented/feature/2026-06-30-subagent-observe-enrich.md) | 2026-06-30 |
| [Dynamic workflows — a script-driven multi-agent orchestration seam](implemented/feature/2026-07-05-dynamic-workflows.md) | 2026-07-05 |
| [Skill system — progressive disclosure instructions for agents](implemented/feature/2026-07-05-skill-system.md) | 2026-07-05 |
| [The approval seam — one-shot permission decisions over a waterfall of answerers](implemented/feature/2026-07-06-approval-seam.md) | 2026-07-06 |
| [Explicit model-facing tool order](implemented/feature/2026-07-06-explicit-tool-order.md) | 2026-07-06 |
| [The subprocess sandbox — confinement seam, native runners, escalation, and per-session modes](implemented/feature/2026-07-06-sandbox.md) | 2026-07-06 |
| [MCP client plugin — connect to external MCP servers and bridge their tools](implemented/feature/2026-07-07-mcp-client-plugin.md) | 2026-07-07 |
| [The session prefix — request-only messages in front of the derived history](implemented/feature/2026-07-07-session-prefix.md) | 2026-07-07 |
| [Repeat-tool-call guard plugin](implemented/feature/2026-07-08-repeat-tool-guard.md) | 2026-07-08 |
| [The self-referential cordis toolset](implemented/feature/2026-07-08-self-referential-cordis-toolset.md) | 2026-07-08 |
| [Configure subagent persona, tool visibility, and depth](implemented/feature/2026-07-12-subagent-persona-tool-filter-and-depth.md) | 2026-07-12 |
### Simplification
@@ -95,7 +108,7 @@ Generated by `pnpm run gen-rfc-index` from the RFC tree — never edit by hand;
|---|---|
| [Provider-neutral content-block vocabulary owned by dsh-llm](implemented/architecture/2026-06-11-content-block-vocabulary.md) | 2026-06-11 |
| [Custom typed tool-schema DSL instead of schemastery](implemented/architecture/2026-06-11-custom-schema-dsl.md) | 2026-06-11 |
| [Dev-mode invariants over compile-time deep-readonly](implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md) | 2026-06-11 |
| [Source-owned session immutability and dev-mode invariants](implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md) | 2026-06-11 |
| [Event-sourced sessions with derived message history](implemented/architecture/2026-06-11-event-sourced-sessions.md) | 2026-06-11 |
| [Microkernel — extension via Cordis event taxonomy, one concrete loop](implemented/architecture/2026-06-11-microkernel-event-taxonomy.md) | 2026-06-11 |
| [Runtime arg validation at the model boundary](implemented/architecture/2026-06-11-runtime-arg-validation.md) | 2026-06-11 |
@@ -124,6 +137,10 @@ Generated by `pnpm run gen-rfc-index` from the RFC tree — never edit by hand;
| [Prompt variables and tool-guidance ownership](implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md) | 2026-07-05 |
| [Every LLM request is reconstructable from the session log](implemented/architecture/2026-07-05-reconstructable-requests.md) | 2026-07-05 |
| [Subagent provider-lifecycle events — `subagent/provider-added` / `subagent/provider-removed`](implemented/architecture/2026-07-05-subagent-provider-lifecycle-events.md) | 2026-07-05 |
| [A shared timeout/deadline primitive, with hard-kill left to each capability](implemented/architecture/2026-07-06-timeout-deadline-library.md) | 2026-07-06 |
| [Tool-call timeout policy as a plugin](implemented/architecture/2026-07-07-tool-call-timeout-policy.md) | 2026-07-07 |
| [The agent is a registration scope](implemented/architecture/2026-07-08-agent-scope-contexts.md) | 2026-07-08 |
| [Agent-scope runtime design and correctness](implemented/architecture/2026-07-12-agent-scope-runtime-design.md) | 2026-07-12 |
### Process
@@ -187,6 +204,8 @@ Generated by `pnpm run gen-rfc-index` from the RFC tree — never edit by hand;
| [Return the ACP bridge to one live session per connection](rejected/simplification/2026-06-20-single-session-acp-bridge.md) | 2026-06-20 |
| [Truncate interrupted final turns on load](rejected/simplification/2026-06-20-truncate-interrupted-turns.md) | 2026-06-20 |
| [Prune the unimplemented subagent seam vocabulary](rejected/simplification/2026-07-04-prune-unimplemented-subagent-vocabulary.md) | 2026-07-04 |
| [Collapse workflows to the exercised foreground core](rejected/simplification/2026-07-12-collapse-workflow-to-foreground-core.md) | 2026-07-12 |
| [Prune unused skill registry surface](rejected/simplification/2026-07-12-prune-unused-skill-registry-surface.md) | 2026-07-12 |
### Architecture

53
docs/rfc/implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md
View File

@@ -1,29 +1,58 @@
# RFC: Dev-mode invariants over compile-time deep-readonly
# RFC: Source-owned session immutability and dev-mode invariants
Status: implemented
## Problem
The session log is append-only by contract, but the types don't enforce it: `session.events` returns `readonly SessionEvent[]` whose *elements* are mutable, and `deriveMessages()` handed the logged `content` arrays/blocks out by reference. The loop then passes those derived messages into the `agent/request` waterfall and on to adapters, where mutating the request is sanctioned — so a request middleware could reach back and rewrite history, silently breaking replay equivalence and the derived-history guarantee. Separately, the event taxonomy (turn/step nesting, seq monotonicity, tool-call/result pairing, legal status transitions) was asserted only where individual tests happened to look.
The session log needs two different protections: immutable ownership of each stored fact, and checks for relationships among facts across time and service seams. Conflating them in an optional development plugin would leave production history vulnerable; trying to express both through TypeScript readonly types would not create a runtime boundary or describe relational rules.
Two ways to defend the log: make immutability part of the type (`DeepReadonly<SessionEvent>` on the way out), or catch corruption at runtime in dev. The runtime-validation proposal took the runtime route; [the deep-readonly proposal](../../rejected/architecture/2026-06-11-immutable-public-surfaces.md) took the type route.
The session log is the durable source of truth for replay, request reconstruction, persistence, and user-visible history. Code outside the session package must be able to inspect that history without retaining a reference that can rewrite it later, and inputs accepted from callers must not remain connected to caller-owned mutable objects.
Immutability of individual values is only half of the contract. A log can contain perfectly immutable records whose sequence, turn/step nesting, tool-call pairing, scoped delivery, or reconstructed model request is wrong. Those rules relate multiple records or services and cannot be established by freezing one object.
TypeScript readonly types are not a sufficient runtime boundary. They disappear when the program runs, a cast can bypass them, and a recursive `DeepReadonly<T>` would spread through every log and message consumer even though some downstream request-processing APIs intentionally work with mutable values.
## Decision
Reject the pervasive `DeepReadonly<T>` type flip. Instead:
Responsibility is split between an always-on storage boundary and optional development assertions.
1. **Always-on:** `deriveMessages()` deep-clones the content it emits (one `structuredClone` per derived message). In-flight mutation of a request can no longer reach the log — this is the real fix, and it costs nothing meaningful next to a model call.
2. **Dev-mode:** a new `dsh-invariants` plugin (pure listeners, off in production, on in tests and demos) asserts the event contract and `Object.freeze`s logged event data so any *other* code that mutates a logged event throws instead of corrupting silently. Seeded sessions are frozen and checked on `session/created` (the constructor copies the seed without emitting `session/event`).
### Session owns immutable history
The invariants encode the *real* contract, not an idealized one: a `tool/call` may have no `tool/result` (a thrown tool-execution pipeline step ends the turn), and both `idle→disposed` and `running→disposed` are legal.
`Session` accepts an event only after one recursive pass has materialized a lossless JSON snapshot. That pass rejects unsupported values and produces the exact detached record that enters the log, so validation and storage cannot observe different values from a stateful getter or retain caller-owned nested references.
The accepted event and all of its descendants are deep-frozen before publication. `append()` returns that owned frozen event, `session/event` observers receive the same record, and `session.events` returns a frozen array snapshot. A previously returned array does not grow after a later append. Seed records pass through the same validation, snapshot, and freeze boundary before construction succeeds.
This guarantee belongs in `Session`, not in an optional listener, because every composition relies on trustworthy history. A production deployment, a focused test, or a custom embedding receives the same storage semantics whether or not development support plugins are registered.
### Derived requests remain detached
`deriveMessages()` projects logged surface events into detached, deep-frozen `Message` objects and returns a fresh array snapshot. Request assembly can therefore combine derived history with other inputs without exposing a path back into the log. The cache reuses safe immutable projections rather than recloning the complete history for each model call.
### The invariants plugin checks relationships
`dsh-invariants` is a pure-listener development plugin. It does not freeze records and has no configuration; disposal removes only its assertions. It checks rules that require trace state or observation of another seam, including monotonic sequence numbers, turn and step nesting, tool-call/result pairing, legal agent-status transitions, subject-correct scoped dispatch, and equality between a loop-built request and the request reconstructed from its session-log prefix.
When the plugin attaches to an existing or seeded session, it replays the immutable log to rebuild trace state. This makes hot reload safe in the middle of a turn without giving the plugin ownership of session storage.
## Alternatives considered
**The pervasive `DeepReadonly<T>` type flip** ([the rejected proposal](../../rejected/architecture/2026-06-11-immutable-public-surfaces.md)) — compile-time only (a plugin casts straight through it), high type-noise across every log/message consumer and adapter, and it would force readonly types through code where mutation is the sanctioned API. The clone draws the mutable/immutable boundary exactly at "logged vs in-flight" without any of that noise.
### Pervasive deep-readonly types
[The rejected immutable-public-surfaces proposal](../../rejected/architecture/2026-06-11-immutable-public-surfaces.md) would apply a recursive readonly type across public log and message surfaces. That provides editor feedback but not a runtime guarantee: TypeScript types are erased and plugin code can cast through them. It also pushes readonly types into consumers where mutation is intentional. Runtime ownership at the `Session` boundary protects every caller without that type propagation.
### Development-only freezing
Freezing history only when an invariants plugin is installed would make the core guarantee composition-dependent. Code could pass development tests and still corrupt history in production or in a focused composition that omits the plugin. Storage immutability is therefore always on, while the more expensive relational checks remain opt-in development support.
### Clone only when deriving messages
Detaching `deriveMessages()` would protect the most common request path but leave other readers of `session.events`, append return values, and session-event observers able to mutate durable history. The log must protect its own boundary; derived projections are an additional isolation boundary, not a substitute.
## Consequences
- History corruption is caught loudly in tests and demos, at zero production cost and zero type noise. The trade-off is that the guarantee is dynamic (a dev-mode tripwire) rather than static.
- The invariants plugin doubles as executable documentation of the event taxonomy — the assertions are the contract.
- `Session.events` keeps its `readonly SessionEvent[]` type; no consumer churn.
- This folds in [the deep-readonly proposal](../../rejected/architecture/2026-06-11-immutable-public-surfaces.md) — there is no separate deep-readonly record; this records the decision to *not* pursue that approach. `InvariantError` is a plain `Error` with a `code` for now; a later taxonomy change can promote it.
- Every accepted live or seeded session event is detached from caller-owned inputs and deeply immutable before any observer can receive it.
- `session.events` exposes stable immutable snapshots instead of the private growing array.
- Request-side mutation cannot reach stored history through derived messages.
- Development builds can enable relational assertions without changing storage behavior, and disposing or omitting the plugin does not weaken log immutability.
- `dsh-invariants` has no `Config` surface because it has no behavior to tune.
- The runtime boundary carries a recursive snapshot-and-freeze cost once per accepted event; later readers and cached projections reuse the owned immutable records.

7
docs/rfc/implemented/architecture/2026-06-11-microkernel-event-taxonomy.md
View File

@@ -10,9 +10,10 @@ The product principle is "everything is a plugin": hooks, /goal, /loop, dynamic
Pure Cordis event taxonomy. The loop's extension seams are typed events with deliberate dispatch modes:
- **waterfall** (around-middleware) where plugins mutate or veto: `agent/prompt-submit`, `agent/request`, `agent/step-result`, `agent/turn-continuation`, `tools/pre-execute`, `tools/post-execute`, `llm/stream`, `system-prompt/assemble`.
- **emit** (sync fire-and-forget) for notifications: turn/step boundaries, stream chunks, lifecycle, errors.
- **parallel** (awaited) for the one durability checkpoint: `session/flush`.
- **waterfall** (around-middleware) where plugins transform, veto, or wrap: `agent/prompt-submit`, `agent/request`, `agent/step-result`, `agent/turn-continuation`, `tools/pre-execute`, `tools/execute`, `tools/post-execute`, `llm/stream`, `system-prompt/assemble`.
- **serial** (awaited in listener order; a bail value stops later listeners) for ordered checkpoints: every `agent/pre-step` listener runs when all abstain, while the first stop returned from `agent/turn-stop` makes the terminal decision final.
- **parallel** (awaited fan-out) where every listener must get an independent chance: the `session/flush` durability checkpoint.
- **emit** (synchronous fire-and-forget) for notifications: turn/step boundaries, stream chunks, lifecycle, errors, and the contained immutable `tools/result` observation.
The event vocabulary lives in interface packages (dsh-agent declares the agent/* events); `@deepseek-ai/dsh-agent-loop` is the only concrete loop plugin and is itself swappable — nothing outside it may depend on it.

6
docs/rfc/implemented/architecture/2026-06-14-session-persistence.md
View File

@@ -6,7 +6,7 @@ Status: implemented
## Problem
Sessions lived only in memory. The example `session-jsonl.ts` plugin (duplicated byte-for-byte in both examples) was write-only telemetry: it buffered `session/event` and appended JSON lines, with no read/replay path, no crash-safety (no fsync, no atomic write, a fire-and-forget dispose drain), no listing, and no format versioning. Nothing could rehydrate a past session from disk into a live agent, so durable resume ("continue yesterday's task"), durable forking, and the ACP `session/load` method ([ACP support](../../proposed/feature/2026-06-14-acp-agent-client-protocol.md)) were all impossible.
Sessions lived only in memory. The example `session-jsonl.ts` plugin (duplicated byte-for-byte in both examples) was write-only telemetry: it buffered `session/event` and appended JSON lines, with no read/replay path, no crash-safety (no fsync, no atomic write, a fire-and-forget dispose drain), no listing, and no format versioning. Nothing could rehydrate a past session from disk into a live agent, so durable resume ("continue yesterday's task"), durable forking, and the ACP `session/load` method ([ACP support](../../implemented/feature/2026-06-14-acp-agent-client-protocol.md)) were all impossible.
The [event-sourced model](2026-06-11-event-sourced-sessions.md) makes the append-only log the single source of truth and derives LLM history from it. Persistence had to stay faithful to that: persist the existing `SessionEvent` directly, with no parallel "persisted message" type that the log is converted to and from. The backend also had to be swappable — a file store now, a database store later — behind one interface.
@@ -23,7 +23,7 @@ Key choices recorded here because they are durable, contested, and surprising:
- **Append-only; a crashed turn is closed, never truncated.** Committed events — those at or below a flushed `turn/end` — are never rewritten. The loop only flushes at `turn/end`, so a crash can leave a durable log whose final turn never closed: real, fully-written events sit after the last `turn/end`. **A single turn can be huge in a long-horizon task** (many steps, large tool output spanning a long autonomous run), so discarding the interrupted turn would silently destroy a large amount of real work — truncating a turn is wrong. Instead, on reload `load` PRESERVES those events and CLOSES the orphaned turn by durably appending the minimal synthetic boundary events: an error `tool/result` for every `tool-call` the crash left unanswered, then a `step/end` if a step was still open, then a `turn/end` carrying the merge-extensible `{ kind: 'interrupted' }` reason (a marker that records the turn was cut short by a crash, not completed by the model — no loop ever emits it). The synthetic tool results matter for resume correctness: the loop logs the `assistant/message` (carrying the `tool-call` blocks) BEFORE running the tools, so a crash mid-tool leaves calls without results; `deriveMessages()` would then replay a dangling assistant tool-call, which every provider rejects as an invalid transcript on the next request. Answering each orphaned call with an error result keeps the rehydrated history valid. `load` returns the balanced log, so a resumed session is immediately usable. The ONLY thing discarded is a never-fully-written **torn tail fragment** — a final record whose bytes (JSONL) or row were never completely flushed; that fragment is not a valid event and is dropped before the synthetic closers are written. A parse error or `seq` gap in the COMMITTED region (at or before the last real `turn/end`) is genuine corruption and makes the session unloadable.
- **File backend canonical, DB backend a proven drop-in.** `SessionEvent` maps 1:1 onto a row `(session_id, seq, type, time, data)` — `append` is INSERT (in a transaction asserting the contiguous-seq contract), `load` is SELECT … ORDER BY seq. `dsh-session-persistence-sqlite` is exactly this: a `SessionPersistence` subclass with no interface change (opencode runs this exact shape on SQLite/WAL), and it passes the same `runPersistenceContract` suite as the JSONL backend — so the contract holds both backends to identical semantics (lazy materialization, interrupted-turn close on load, contiguous-seq), expressed once over file bytes and once over rows.
- **Metadata is out-of-log.** Format version, cwd, and lineage are storage concerns, not replayable conversation state, so they live in a `SessionHeader` owned by `dsh-session` and attached to a `Session` via a new readonly `session.header` — never in `SessionEventMap`, never reaching `deriveMessages()`. The alternative (a merge-extensible `session/meta` event as log line 0) was rejected: an in-log event would ride along with a seeded/forked session for free, but metadata is not replayable state, so the explicit out-of-log header seam is the cleaner cost. (The header was originally split into an immutable `SessionHeader` plus a mutable `SessionSummary` whose union was `SessionMeta`; the mutable summary was later removed as dead state — see [Drop the mutable session summary](../simplification/2026-06-19-drop-mutable-session-summary.md).)
- **Resume is an async factory, not a change to synchronous create.** `ctx.agents.resume({ resumeSessionId })` awaits `ctx.sessionPersistence.load`, recreates the live session with the loaded events (so `lastTurnNumber`/`deriveMessages` continue), and starts a fresh agent on the resumed id (NOT `${agentId}-session`). The agent-loop does NOT hard-inject `sessionPersistence` (that would pend non-persistent demos forever); `resume` rejects with a clear error when it is absent.
- **`ctx.agents.create()` and `ctx.agents.resume()` are async factories; resume additionally crosses the persistence boundary.** `ctx.agents.resume({ resumeSessionId })` awaits `ctx.sessionPersistence.load`, recreates the live session with the loaded events (so `lastTurnNumber`/`deriveMessages` continue), and starts a fresh agent on the resumed id (NOT `${agentId}-session`). The agent-loop does NOT hard-inject `sessionPersistence` (that would pend non-persistent demos forever); `resume` rejects with a clear error when it is absent.
## Alternatives considered
@@ -33,4 +33,4 @@ Format versioning: the header carries a `version`; `load` rejects any non-curren
## Consequences
Two new packages and the metadata seam in `dsh-session` (`session.header`, the `create(id?, options?)` signature). Bought: durable resume/fork, a read/replay path, crash tolerance, and the foundation the ACP `session/load` ([ACP support](../../proposed/feature/2026-06-14-acp-agent-client-protocol.md)) needs — all over the existing event-sourced log, with the backend swappable behind one interface. The reusable `runPersistenceContract` suite holds every backend to the same append-only / contiguous-seq / lazy-materialization / serializability semantics. This completes [event-sourced sessions](2026-06-11-event-sourced-sessions.md)'s deferred "real persistence backend" and resolves its `TODO(review)` on the event vocabulary: persisting the log freezes its shape, and the `assistant/chunk` fidelity question is answered above (persist verbatim).
Two new packages and the metadata seam in `dsh-session` (`session.header`, the `create(id?, options?)` signature). Bought: durable resume/fork, a read/replay path, crash tolerance, and the foundation the ACP `session/load` ([ACP support](../../implemented/feature/2026-06-14-acp-agent-client-protocol.md)) needs — all over the existing event-sourced log, with the backend swappable behind one interface. The reusable `runPersistenceContract` suite holds every backend to the same append-only / contiguous-seq / lazy-materialization / serializability semantics. This completes [event-sourced sessions](2026-06-11-event-sourced-sessions.md)'s deferred "real persistence backend" and resolves its `TODO(review)` on the event vocabulary: persisting the log freezes its shape, and the `assistant/chunk` fidelity question is answered above (persist verbatim).

12
docs/rfc/implemented/architecture/2026-06-18-agent-lifecycle-and-ownership-seams.md
View File

@@ -16,9 +16,9 @@ A new `cancel()` verb on the `Agent` interface — the single public stop primit
### 2. `AgentHandle` async disposer
`ctx.agents.create`/`resume` (and the `AgentFactory` interface) return `AgentHandle = { agent: Agent; dispose(): Promise<void> }`. The disposer is a **capability** — only the holder can tear down exactly this agent: stop its loop, `await` the loop's exit (true quiescence, not just the `disposed` status flip), unregister it, and remove its session from the store. `ctx.agents.get(id)` still returns a bare `Agent`. Config-created agents stay owned by the `AgentLoop` fiber (the handle is discarded). ACP holds each session's disposer in its `SessionRecord` and runs it on disconnect/teardown, so a bare client disconnect leaves no registered agent and no session-store entry — even when `session/load` races teardown (the just-resumed handle is disposed before the closed-guard throw).
`ctx.agents.create`/`resume` (and the `AgentFactory` interface) return `AgentHandle = { agent: Agent; dispose(): Promise<void> }`. The disposer is a **consumer capability** — a registry observer holding only the bare `Agent` cannot tear it down. The caller fiber and registered factory provider are structural co-owners: caller unload enforces structured ownership, while provider unload must stop old instances whose scoped dependency surface resolves through that provider. All three paths reach the same memoized teardown: stop the loop, await its exit and idle flushes (true quiescence, not just the `disposed` status flip), detach the agent, detach its session, and unwind its scope. Each public ID becomes reusable when its exact registry entry detaches; there is no separate reservation-release phase. Config-created agents are already owned by the `AgentLoop` fiber (the handle is discarded). ACP holds each session's disposer in its `SessionRecord` and runs it on disconnect/teardown, so a bare client disconnect leaves no registered agent and no session-store entry — even when `session/load` races teardown (the just-resumed handle is disposed before the closed-guard throw).
**Teardown ORDER is load-bearing for durability**, and the implementation folds the session lifecycle into the agent's SINGLE composite cordis effect (`SessionStore.prepare`/`enter`/`announce`, replacing a sibling-effect split). A fiber unload disposes sibling effects concurrently (`Promise.all`), which would race the session's `onAppend` detach against the loop's closing `session/flush` and drop the closing `turn/end`; inside one effect the disposers run as an ordered LIFO chain (loop stopped + `await agent.done` BEFORE the session detaches), so the loop's final flush is captured on BOTH the handle's `dispose()` and a fiber unload. The register disposer's `agent/disposed` emit is contained (a throwing listener must not reject the chain and skip the later session detach).
**Teardown ORDER is load-bearing for durability**, and the implementation folds the session lifecycle into the agent's SINGLE composite cordis effect (`SessionStore.prepare`/`enter`/`announce`, replacing a sibling-effect split). A fiber unload disposes sibling effects concurrently (`Promise.all`), which would race removing the session store's append publication hooks against the loop's closing `session/flush` and drop the closing `turn/end`; inside one effect the disposers run as an ordered LIFO chain (loop stopped + `await agent.done` BEFORE the session detaches), so the loop's final flush is captured on BOTH the handle's `dispose()` and a fiber unload. The contained `agent/disposed` and `session/disposed` notifications cannot reject the chain or skip later teardown.
### 3. Bash owner token in the seam
@@ -33,16 +33,14 @@ These invariants hold and are pinned by tests:
- A `tool-bash` HMR reload does NOT make an existing background task readable or killable by a different session (ownership survives on the executor).
- Existing non-ACP demos still work without managing handles explicitly; config-created agents remain owned by the `AgentLoop` plugin fiber.
## Seam precondition (recorded)
## Session owner tokens are unique among live agents
The bash owner-token comparison relies on `session.header.id` being unique among live agents. The agent registry does NOT enforce this — it rejects a duplicate *agentId*, not a duplicate session id, and `createAgent` accepts an arbitrary `sessionId`. This is NOT reachable via ACP (UUID sessionId, `agentId === sessionId`, duplicate-load rejected), so it is not a live product hole, but a programmatic caller that registers two agents with the same session id would break bash isolation and mis-route the completion notice. The access *policy* (token comparison) stays in `tool-bash` (the consumer); the bash seam stores only an opaque `owner` string and never interprets it — the correct interface/impl/consumer split.
The planned resolution is to remove the precondition by construction — see [unify the agent id and the session id](../../proposed/simplification/2026-06-20-unify-agent-and-session-id.md): once an agent IS its session (one id), the registry's existing unique-`agentId` check is a unique-session-id guarantee and no two live agents can share a session token.
The bash owner-token comparison relies on `session.header.id` being unique among live agents. Concurrent same-ID operations may both prepare privately, but publication enters the session and agent in order; `SessionStore.enter()` rejects a duplicate live session id, and every losing transaction rolls its private state back. A programmatic caller therefore cannot publish two live agents with one session token. The access *policy* (token comparison) stays in `tool-bash` (the consumer); the bash seam stores only an opaque `owner` string and never interprets it — the correct interface/implementation/consumer split.
## Alternatives considered
- **A public `BashTask.owner` field** instead of the `BashExecutor.ownerOf(id)` seam — rejected: one read path, no redundant API.
- **Sibling cordis effects for the agent's session lifecycle** — rejected: a fiber unload disposes sibling effects concurrently (`Promise.all`), racing the session's `onAppend` detach against the loop's closing `session/flush`; the single composite effect's ordered LIFO chain is what captures the closing `turn/end` on both disposal paths.
- **Sibling cordis effects for the agent's session lifecycle** — rejected: a fiber unload disposes sibling effects concurrently (`Promise.all`), racing removal of the store-owned append publication hooks against the loop's closing `session/flush`; the single composite effect's ordered LIFO chain is what captures the closing `turn/end` on both disposal paths.
- **A separate step-only `abort()` beside `cancel()`** — shipped originally, then removed as unused; `cancel()` is the single public stop primitive ([the public-stop-surface RFC](../simplification/2026-06-20-public-agent-stop-surface.md)).
## Consequences

4
docs/rfc/implemented/architecture/2026-06-30-event-domain-semantics.md
View File

@@ -28,9 +28,9 @@ This is the foundational change in a stack that adds a Hooks subsystem; it estab
## Consequences
- The loop no longer emits any boundary mirror; `closeStep` appends `step/end` only and `closeTurn` appends `turn/end` only. A throwing `step/end`/`turn/end` session-event listener is the surviving boundary-listener failure path (contained inside `closeStep`/`closeTurn` — `Session.append` pushes the event before notifying listeners, so the boundary is durable and the turn closes balanced regardless).
- The loop no longer emits any boundary mirror; `closeStep` appends `step/end` only and `closeTurn` appends `turn/end` only. `Session.append` owns post-commit observer containment, so a throwing boundary observer cannot change the turn outcome or starve later consumers; an acceptance or internal validation failure still escapes before the boundary enters the log.
- Tests that observed boundaries via the removed emits now observe the durable `turn/start`/`turn/end`/`step/start`/`step/end` session events — the behavior they pin (boundary ordering, step counting) is unchanged; only the feed they read moved to the canonical one. The tests that exercised a *throwing turn-boundary emit listener* were deleted, because that code path no longer exists (there is no emit to throw from). Per [AGENTS.md "tests document behavior, not golden truth"](../../../../AGENTS.md), the behavior and its test moved (or died) together.
- The loop marks the step open (`stepOpen = true`) BEFORE appending `step/start`, because `Session.append` pushes the event to the log before notifying `session/event` listeners (validation throws happen earlier, before the push — see [the session append contract](../../../core-data-structures/session.md)). So a throwing `step/start` session-event listener runs with the step already open and the event already in the log: the loop's outer catch then calls `closeStep()`, which appends the balancing `step/end`, and the turn closes balanced with an error (`turn/start → step/start → step/end → turn/end` — verified by the invariants oracle in the regression test). Closing the open step is owed precisely because the marker is set first.
- The loop marks the step open (`stepOpen = true`) only after `append('step/start')` returns. Internal dispatch validation runs before the log push and may reject without opening a step; post-commit `session/event` observer failures are contained inside `Session.append`. The marker therefore represents exactly the committed boundary that owes a later `step/end`.
- The full realization of this is [the simplification RFC "Stop mirroring durable boundaries as agent events"](../simplification/2026-06-20-remove-agent-boundary-mirror-events.md): all four boundary mirrors are removed and every consumer reads boundaries off `session/event`. `agent/steering` (not a boundary mirror) stayed outside that RFC's scope and was removed by its own follow-up, [Remove the `agent/steering` mirror emit](../simplification/2026-07-04-remove-agent-steering-mirror.md) — it mirrored the durable `steering/message`.
- The cordis events catalog (`docs/cordis-catalog/events.md`) is regenerated to drop the mirror events.

12
docs/rfc/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md
View File

@@ -16,11 +16,11 @@ The assembled system prompt had four defects, all of one family: facts the harne
## Decision
**One principle: every fact in the prompt has exactly one owner.** The model name and workspace are config/session facts → the harness exposes them as variables and the persona references them. Per-tool semantics and when-to-use → the tool's `description`. Cross-call habits a description cannot carry → the tool package's prompt section. Identity and behavior → the deployment's persona, and nothing else.
**One principle: every fact in the prompt has exactly one owner.** The model name and workspace are config/session facts → the harness exposes them as variables and the persona references them. Per-tool semantics and when-to-use → the tool's `description`. Cross-call habits a description cannot carry → the tool package's prompt section. Harness provenance → the static `harness:identity` section. Deployment role and behavior → the deployment's persona.
### Assemble context
`SystemPrompt.assemble(context)` takes an `AssembleContext` — declared EMPTY and merge-extensible in `dsh-system-prompt` (the package stays agnostic of who assembles); `dsh-agent` declaration-merges `agent?: Agent` onto it (a new type-level edge `agent → system-prompt`, no cycle — `tools` already depends on both). The loop passes `{ agent }` each step; section text providers become `string | ((context) => string)` (zero-arg providers stay valid), and the `system-prompt/assemble` waterfall gains the context parameter so a listener can filter or extend per agent.
`SystemPrompt.assemble(context)` takes a merge-extensible `AssembleContext`. `dsh-system-prompt` declares the optional `scope` selector used for scoped routing, while `dsh-agent` declaration-merges the optional typed `agent` field onto it (a type-level edge `agent → system-prompt`, with no runtime dependency cycle). The loop calls `assembleContextFor(agent)` each step so both fields identify the same agent; section text providers may read that context, and the `system-prompt/assemble` waterfall receives it so a listener can filter or extend per agent.
### Prompt variables
@@ -30,15 +30,15 @@ Plugins contribute named values via `ctx.systemPrompt.variable(name, provider)`;
### Persona as the order-0 section
`dsh-system-prompt` itself registers the two harness-owned sections (they must survive a swapped loop plugin, so they do NOT live on `dsh-agent-loop`): the static `harness:identity` at order `-100` — every prompt opens by stating the agent is powered by the DeepSeek Harness SDK — and `deployment:persona` at order 0, whose text is the plugin's own `persona` config. The persona is per-DEPLOYMENT, not per-agent: every agent in the context (subagents included) renders the same one, `AgentOptions.systemPrompt` is deleted along with the per-agent forwarding plumbing (the app configs' `systemPrompt` keys become a `persona` key routed to this plugin through `dsh-agent-core`), and the ACP bridge and `dsh-tool-subagent` stop carrying persona configuration entirely. The loop's special-case join is deleted: `fullSystemPrompt ≡ renderPrompt(assembly)`, one ordered pipeline for everything the model sees, and `agent/pre-step` (compaction's token-pressure input) measures exactly the real prompt. Order bands are now convention: harness identity `-100`, persona `0`, tool guidance `100–199`; other negative orders also render before the persona.
`dsh-system-prompt` itself registers the two harness-owned sections (they must survive a swapped loop plugin, so they do NOT live on `dsh-agent-loop`): the static `harness:identity` at order `-100` — every prompt opens by stating the agent is powered by the DeepSeek Harness SDK — and the global default `deployment:persona` at order 0, whose text is the plugin's own `persona` config. `AgentOptions.systemPrompt` and the loop's special-case join are gone: `fullSystemPrompt ≡ renderPrompt(assembly)`, one ordered pipeline for everything the model sees, and `agent/pre-step` (compaction's token-pressure input) measures exactly the real prompt. An agent-scoped section with the same `deployment:persona` name shadows the default for that agent; programmatic setup may register one directly, and the subagent persona feature installs one before publishing an in-process child when the selected provider supports it. Order bands are convention: harness identity `-100`, persona `0`, tool guidance `100–199`; other negative orders also render before the persona.
### Tool guidance ownership
Per-tool semantics and when-to-use live in tool DESCRIPTIONS, which already ship in every request — the YAML prose was ~fully redundant with them. Sections carry only the cross-call habits a single call's description cannot: `dsh-tool-bash` contributes `tool:bash` (order 105) — check the `[exit code: N]` marker on every result; `dsh-tool-fs`'s read section gains the "not shell commands like cat" contrast. `todo_write` and the subagent tools need NO section — their descriptions already carry the whole contract. The leaf personas shrink to identity + behavior (verify your work; keep answers brief), and the welcome banner stops enumerating tools.
### The subagent context contract
### The subagent conversation-history descriptor
`SubagentProvider` gains `readonly inheritsParentContext: boolean` — a DESCRIPTIVE fact beside `capabilities`, not in it (capabilities are start-time validation; nothing validates against this flag). Spawn and ACP declare `false`, fork declares `true`. `dsh-tool-subagent` derives both the tool description and the `prompt` parameter description from the flag (`providerWording`): the fork instance now tells the model the child inherits the conversation's completed turns (not the in-flight turn) and that its prompt should state only what is new. Deriving the description from a provider that arrives on its own fiber is what forced the provider-lifecycle events and the tool's reactive registration — that mechanism, its Loader-concurrency rationale, and its rejected alternatives are recorded in [the provider-lifecycle-events RFC](2026-07-05-subagent-provider-lifecycle-events.md).
`SubagentProvider` gains `readonly inheritsParentContext: boolean` — a DESCRIPTIVE conversation-history fact beside `capabilities`, not in it (capabilities are start-time validation; nothing validates against this flag). Spawn and ACP declare `false`, fork declares `true`. The name refers only to conversation seeding, not Cordis scope, services, tools, or authority. `dsh-tool-subagent` derives both the tool description and the `prompt` parameter description from the flag (`providerWording`): the fork instance now tells the model the child is seeded with the conversation's completed turns (not the in-flight turn) and that its prompt should state only what is new. Deriving the description from a provider that arrives on its own fiber is what forced the provider-lifecycle events and the tool's reactive registration — that mechanism, its Loader-concurrency rationale, and its rejected alternatives are recorded in [the provider-lifecycle-events RFC](2026-07-05-subagent-provider-lifecycle-events.md).
## Alternatives considered
@@ -56,7 +56,7 @@ Per-tool semantics and when-to-use live in tool DESCRIPTIONS, which already ship
## Shipped invariants
- `renderPrompt(assemble({ agent }))` for the coding-agent example renders the persona FIRST (with the agent's model name interpolated), then the fs/bash/web guidance sections; the loop has no other prompt-composition path.
- `renderPrompt(await assemble(assembleContextFor(agent)))` for the coding-agent example renders the harness identity, then the persona (with the agent's model name interpolated), then the fs/bash/web guidance sections; the loop has no other prompt-composition path.
- The `subagent_fork` schema description says the child inherits the conversation; the `subagent` one says it does not. The tool follows its provider: absent before the backend activates, present after, gone when the backend unloads, re-worded from the fresh provider on reload.
- Unknown/valueless/malformed/unbalanced `{{…}}` references throw with the section name in the message; duplicate section, variable, and tool-name registrations all throw.
- Snapshot goldens are prompt-independent by construction: llm-replay keys replay on (turn, step) chunk streams and never re-verifies the outgoing request.

9
docs/rfc/implemented/architecture/2026-07-05-reconstructable-requests.md
View File

@@ -20,13 +20,13 @@ Prefix-cache stability is corollary #1, not the headline: an append-only log pro
**Messages.** `Session.deriveMessages()` is cached: each surface node is projected exactly once, when first seen, through the public per-node function `deriveEventMessage(event)`; a surface rewrite (a compaction `replace` — `SurfaceManager.replaceGeneration`) rebuilds. Callers get a fresh array per call over shared, deep-frozen messages: mutating logged history through a projection is unrepresentable (it throws), replacing the old clone-per-call isolation. External reconstructors fold the same public function over a log prefix, so no two paths can disagree.
**The header.** The request's non-content half — `EpochHeader`: call config (`LlmCallConfig`: model + sampling scalars), rendered system prompt, assembled tool schemas — is logged session state, in canonical form (empty system/tools ≡ absent). Two log-only, turn-enclosed events in dsh-session carry it: `request/header`, a full snapshot with reason `'initial' | 'resume' | 'fallback'`, and `request/header-delta`, an amendment (`SystemDelta`: a common-prefix/suffix line trim; `ToolsDelta`: name-keyed added/removed/changed; `config`: replaced whole). The pure trio `foldRequestHeader` / `diffHeader` / `applyHeaderDelta` reconstructs; the live session tracks the fold with the same lazy cursor as the message cache. Snapshots anchor the fold where a fold needs anchors — conversation birth and process boundaries — and each loop instance appends one on its first request (`'initial'` when the log has none, `'resume'` otherwise, even when nothing changed: the boundary itself is a recorded fact, and cross-restart drift becomes attributable while an unchanged header resumes byte-identical). Deltas are an encoding optimization with a safety valve, never a correctness dependency: the writer verifies `applyHeaderDelta(prev, delta)` reproduces the new header exactly and records a `'fallback'` snapshot when the encoding cannot express a change (a pure tool reordering), so a well-formed log always folds.
**The header.** The request's non-history half — `EpochHeader`: call config (`LlmCallConfig`: model + sampling scalars), rendered system prompt, assembled tool schemas, and the session prefix (`messagePrefix`, below) — is logged session state, in canonical form (empty system/tools/prefix ≡ absent). Two log-only, turn-enclosed events in dsh-session carry it: `request/header`, a full snapshot with reason `'initial' | 'resume' | 'fallback'`, and `request/header-delta`, an amendment (`SystemDelta`: a common-prefix/suffix line trim; `ToolsDelta`: name-keyed added/removed/changed; `config`: replaced whole; `messagePrefix`: replaced whole, an empty array encoding the transition to absence — an arm the loop never exercises in practice, kept for codec totality). The pure trio `foldRequestHeader` / `diffHeader` / `applyHeaderDelta` reconstructs; the live session tracks the fold with the same lazy cursor as the message cache. Snapshots anchor the fold where a fold needs anchors — conversation birth and process boundaries — and each loop instance appends one on its first request (`'initial'` when the log has none, `'resume'` otherwise, even when nothing changed: the boundary itself is a recorded fact, and cross-restart drift becomes attributable while an unchanged header resumes byte-identical). Deltas are an encoding optimization with a safety valve, never a correctness dependency: the writer verifies `applyHeaderDelta(prev, delta)` reproduces the new header exactly and records a `'fallback'` snapshot when the encoding cannot express a change (a pure tool reordering), so a well-formed log always folds.
**The loop, transmission-stateless.** Per step: render assembly (every step — value comparison needs no change-signal discipline, and a section that varies per step surfaces as a *logged* header event per step instead of a silent bust) → `agent/pre-step` (compaction's surface mutations land before derivation) → **messages snapshot, then `step/start` appended as the next operation in the same synchronous frame** → seed the call config (first request of the instance: from `AgentOptions`, so explicit options always beat the logged baseline — fork model-overrides and resume reconfiguration stay correct; afterwards: from the folded header) → the `agent/request` waterfall, re-typed `(agent, turn, step, config: LlmCallConfig, next) → LlmCallConfig` — a frozen seed and a returned replacement are ALL a listener shapes; content flows through the log channels (`inject()`, steering, prompt-submit `additionalContext`, sections via `system-prompt/assemble`) — → the header event the request owes the log → build `GenerateOptions` from the snapshot + header, deep-freeze (`deepFreeze` exempts the `AbortSignal`, the one live control channel — freezing one breaks `AbortController.abort()`), dispatch. The loop's only in-process bookkeeping is one boolean: whether this instance has logged its anchoring snapshot.
**The loop, transmission-stateless.** Per step: render assembly (every step — value comparison needs no change-signal discipline, and a section that varies per step surfaces as a *logged* header event per step instead of a silent bust) → on the instance's FIRST step only, the `agent/session-prefix` waterfall — request-ONLY messages fronting the entire derived history (a frozen empty seed, contributions returned as an extension of `next()`; the home for session-stable openers that must NOT become history — a skills catalog, an AGENTS.md digest), deep-frozen and cached on the instance so reuse is structural and the prefix cannot drift mid-session — → `agent/pre-step`, carrying the composed prefix (compaction's surface mutations land before derivation, and its pressure gate counts the prefix this instance will actually send — never a previous instance's logged one, which could under-gate a resumed/forked instance whose contributor grew) → **messages snapshot, then `step/start` appended as the next operation in the same synchronous frame** → seed the call config (first request of the instance: from `AgentOptions`, so explicit options always beat the logged baseline — fork model-overrides and resume reconfiguration stay correct; afterwards: from the folded header) → the `agent/request` waterfall, re-typed `(agent, turn, step, config: LlmCallConfig, next) → LlmCallConfig` — a frozen seed and a returned replacement are ALL a listener shapes; durable content flows through the log channels (`inject()`, steering, prompt-submit `additionalContext`, sections via `system-prompt/assemble`) — → the header event the request owes the log, carrying the prefix as `messagePrefix` (no session event carries it, so the header is its only durable record; resume = a new instance = a recompose, anchored by its `'resume'` snapshot) → build `GenerateOptions` from `messagePrefix + snapshot` + header, deep-freeze (`deepFreeze` exempts the `AbortSignal`, the one live control channel — freezing one breaks `AbortController.abort()`), dispatch. The loop's per-instance bookkeeping is one boolean plus the cached prefix: whether this instance has logged its anchoring snapshot, and what it composed.
**The reconstruction boundary is `step/start`, unconditionally.** A step's messages are the derivation over `events[0..stepStartSeq)`. Because the snapshot precedes the `step/start` append in the same synchronous frame, nothing can enter this request past the boundary: an `agent.inject()` from an `agent/request` listener (or any concurrent task, or a `session/event` listener firing on `step/start` itself) lands in the log after the boundary and joins the NEXT request. For waterfall-window appends this matches the prior loop (it also derived before its waterfall); for a synchronous `step/start` listener it is a deliberate change — such a listener could previously reach the current request — and `agent/pre-step` is the sanctioned seam for content that must affect the CURRENT request. A step's header for reconstruction is the fold after its own `request/header*` event (which sits between its `step/start` and first response event) or the fold carried forward.
**The reconstruction boundary is `step/start`, unconditionally.** A step's messages are the derivation over `events[0..stepStartSeq)`. Because the snapshot precedes the `step/start` append in the same synchronous frame, an `agent.inject()` from an `agent/request` listener or any concurrent task lands after the boundary and joins the NEXT request. `session/event` is observe-only during publication: a reentrant append is rejected until the current callback list drains, preventing nested event delivery from overtaking the event being observed. `agent/pre-step` is the sanctioned seam for content that must affect the CURRENT request. A step's header for reconstruction is the fold after its own `request/header*` event (which sits between its `step/start` and first response event) or the fold carried forward.
**Enforcement.** Dev-mode ([dsh-invariants](../../../../packages/support/invariants/src/index.ts)), on `llm/stream`: a frozen request with a live `sessionId` — the loop-built marker; hand-built one-shots are unfrozen and skipped — must carry messages deep-equal to the boundary derivation, rebuilt through a FRESH `Session` over `events[0..stepStartSeq)` so the live cache cannot vouch for itself, and header fields equal to `foldRequestHeader` over the log. There is no divergence allowance and nothing to allow: no seam can put unlogged content into a request. `prepend: true` only defends against the replay adapter's short-circuit (an append-registered listener); two prepended listeners have no defined mutual order in cordis, so correctness rests on the seq-bounded fold, never on listener timing. Measurement stays lean: the with-key e2e ([request-cache.e2e.ts](../../../../packages/core/agent-loop/tests/request-cache.e2e.ts)) proves `usage.cacheReadTokens > 0` on every request after the first against the live API, and per-step usage in the log is the production observable — a header event or compaction shows up as a cache-read collapse on the next step.
**Enforcement.** Dev-mode ([dsh-invariants](../../../../packages/support/invariants/src/index.ts)), on `llm/stream`: a frozen request with a live `sessionId` — the loop-built marker; hand-built one-shots are unfrozen and skipped — must carry messages deep-equal to the folded header's `messagePrefix` followed by the boundary derivation — the derivation rebuilt through a FRESH `Session` over `events[0..stepStartSeq)` so the live cache cannot vouch for itself — and header fields equal to `foldRequestHeader` over the log. There is no divergence allowance and nothing to allow: no seam can put unlogged content into a request — the `agent/session-prefix` seam's product enters only because the header event records it first. `prepend: true` only defends against the replay adapter's short-circuit (an append-registered listener); two prepended listeners have no defined mutual order in cordis, so correctness rests on the seq-bounded fold, never on listener timing. Measurement stays lean: the with-key e2e ([request-cache.e2e.ts](../../../../packages/core/agent-loop/tests/request-cache.e2e.ts)) proves `usage.cacheReadTokens > 0` on every request after the first against the live API, and per-step usage in the log is the production observable — a header event or compaction shows up as a cache-read collapse on the next step.
### The MiniCode shape: adopted, with the provenance arrow inverted
@@ -44,6 +44,7 @@ What survives from `LLMClient`: the conversation is maintained, not rebuilt —
## Consequences
- A request that is not explained by the log cannot be constructed by accident — not by the loop, not by a listener; mutating a built request throws; every header change is a durable, diffable log event.
- Choosing between the advisory channels is a change-frequency decision, and the design makes the stable one structural: an `agent/session-prefix` contribution is composed once per loop instance and reused verbatim, so it extends the cacheable prefix at zero marginal cost and CANNOT bust the provider cache mid-session; content that changes mid-session flows through the append-only history channels — `agent.inject()`, a `tools/post-execute` decision's `additionalContext`, prompt-submit `additionalContext` — each a durable `context/message` paid once and prefix-cached thereafter, at the price of accumulating in history and the log. Route session-frozen openers to the prefix and change notices to the history channels; a per-step request-only tail slot was deliberately dropped (no consumer, and a durable append covers every current update pattern).
- What still costs full price at the provider is inherent and logged: compaction (its `compact/*` events and replace node), a real prompt/tool change (`request/header-delta`), a config switch (ditto), a process boundary with drift (`'resume'` snapshot differing from its predecessor). The provider's own reasoning-content exclusion is managed server-side.
- The `step/start`-listener behavior change (above) is the one observable semantics change for plugins; `agent/pre-step` is the current-request seam.
- Tool-result trimming (planned) needs no new mechanism: a logged single-node surface replace (`start === end`) carrying a trimmed `tool/result` under the same `callId` — compaction-family, replay-correct, cache-bust batched by the same pressure logic.

98
docs/rfc/implemented/architecture/2026-07-06-timeout-deadline-library.md Normal file
View File

@@ -0,0 +1,98 @@
# RFC: A shared timeout/deadline primitive, with hard-kill left to each capability
Status: implemented
## Problem
Timeout handling was drifting apart across the tool-bearing capabilities, and the divergence was not superficial — it was the same logic re-implemented three ways, each with its own subtle correctness burden.
- **bash** ([packages/bash/bash-local/src/run.ts](../../../../packages/bash/bash-local/src/run.ts)) had a full, correct timeout inside the process plumbing: a config-clamped `timeoutMs`, two independent triggers — a `killTimer` for the timeout and an `onAbort` listener for upstream cancellation — each calling one `kill()` closure that escalates SIGTERM→grace→SIGKILL on the process group, and two orthogonal outcome booleans (`timedOut`, `aborted`) latched independently.
- **web_fetch** ([packages/web/web-fetch-local/src/provider.ts](../../../../packages/web/web-fetch-local/src/provider.ts)) had a correct but *hand-rolled* timeout: it constructed an `AbortController`, wired `setTimeout(() => controller.abort(new WebError(…, 'WEB_FETCH_TIMEOUT')))`, manually added and removed the upstream-signal listener, cleared the timer in a `finally`, and recovered the timeout reason from `signal.reason` in a `translateAbortOrNetwork` helper because the reader surfaces a bare `AbortError`.
- **web_search** ([packages/web/tool-web/src/search.ts](../../../../packages/web/tool-web/src/search.ts)) had **no timeout at all**: `WebSearchRequest` ([packages/web/web/src/types.ts](../../../../packages/web/web/src/types.ts)) carries no `timeoutMs` field, and each provider's `search()` only forwards `exec.signal`. (web_search stays untimed here — see Consequences.)
Each new external-process or network tool re-derived the same four things — clamp the requested value, start a timer, fuse the timeout with upstream cancellation, and distinguish "timed out" from "cancelled" on the way out — and the fusion and reason-recovery are exactly the parts that are easy to get subtly wrong (web_fetch's `signal.reason` dance is evidence). At the same time, the *termination* each performs is irreducibly different: bash kills an OS process group (work runs in a child process, outside this runtime, reachable only by signal), while web aborts an in-process `fetch` (undici tears down the socket). There is no single mechanism that can stop all of them.
The two reference agents surveyed converged on the same split. Codex models "what will end this exec early" as one value (`ExecExpiration`, an enum fusing timeout and a cancellation token) whose `wait_with_outcome()` returns `TimedOut | Cancelled`, while the actual `kill_process_group` lives outside it — and that abstraction is reused *only* across the exec family, with MCP, model-stream, and guardian each keeping their own bespoke `tokio::time::timeout`. Claude Code shares nothing: bash and ripgrep each own a private SIGTERM→SIGKILL kill and distinguish timeout from cancellation by throwing distinct error types, while file I/O has no timeout. Both confirm the boundary drawn here: the timing-and-classification half is worth sharing within a family of like-terminated operations; the termination half is not shareable and stays in each capability.
## Decision
`@deepseek-ai/dsh-timeout` lives under `packages/util/` (peer to `dsh-brand`) and owns the *timing and classification* half of timeout; the *termination* half — the hard kill — stays in each capability's implementation. It is a library of pure functions, **not** a cordis service or plugin: it takes no `ctx`, registers nothing, holds no cross-call state, and emits no events. There is deliberately no central "timeout service" that would have to know how to stop every capability's work — that knowledge is exactly what a microkernel keeps out of shared layers, and what Codex's exec-only `ExecExpiration` scope demonstrates.
### The library surface
Three functions plus one reason type:
```ts ignore-check
/** The internal reason attached to a timeout abort, so consumers can classify it after the fact. */
export class TimeoutReason extends Error {
override name = 'TimeoutReason'
constructor(readonly code: string, readonly timeoutMs: number) {
super(`${code} after ${timeoutMs}ms`)
}
}
/** Validate/fill a caller's optional positive hint from the backend's default, then cap at its max. */
export function clampTimeout(
requested: number | undefined,
def: number,
max: number,
name = 'timeoutMs',
): number
/**
* Build a deadline signal that aborts on upstream cancellation OR on timeout,
* with the timeout carrying a `TimeoutReason`. `timeoutMs <= 0` means "no
* timeout" (background tasks): forward only the upstream signal, arm no timer.
* The returned object's `[Symbol.dispose]` clears the timer — `using` for a
* scope-lifetime consumer, a manual call for an event-lifetime one.
*/
export function deadline(
upstream: AbortSignal | undefined,
timeoutMs: number,
code: string,
): { signal: AbortSignal; [Symbol.dispose](): void }
/** Recover the TimeoutReason from an aborted signal (or error); `code` scopes the match to this deadline's timer. */
export function timeoutOf(x: AbortSignal | { reason?: unknown }, code?: string): TimeoutReason | undefined
```
`deadline` is `AbortSignal.any([upstream, <timeout controller>])` with three things the standard library does not give: a typed, identifiable `TimeoutReason` on the timeout abort (native `AbortSignal.timeout()` yields a fixed `TimeoutError`, indistinguishable across timeout kinds), an internal `timeoutMs <= 0` "no timeout" sentinel for backend-owned background work, and a `Symbol.dispose` cleanup that works with both `using` and manual disposal. `AbortSignal.any` is a Node ≥ 20 primitive; it is the single mechanism that fuses two abort sources into one, adopting the reason of whichever fires first. External request hints validate as positive finite numbers via `clampTimeout` before they reach `deadline`; `0` is not a model-/plugin-facing "disable timeout" value. When `timeoutMs <= 0` and no upstream signal is present, `deadline()` returns a never-aborting signal plus a no-op disposer so callers keep one call shape. `TimeoutReason` is an internal classification reason: providers translate it into seam-specific public errors or result fields before returning to callers. `timeoutOf`'s optional `code` scopes classification to the caller's own deadline: when the `upstream` is itself a deadline (a future `tools/execute` middleware arming a per-call deadline), `AbortSignal.any` preserves the outer `TimeoutReason` if it fires first, and an unscoped match would misreport the outer timeout as the inner capability's own; scoping to `code` reads a foreign timeout as an ordinary upstream cancel.
### The division of labor
| Concern | Owner |
|---|---|
| Validate request hint and clamp default/max | `dsh-timeout` (`clampTimeout`) — pure arithmetic plus the shared positive-finite request contract |
| Arm timer, abort on deadline, carry reason, fuse with upstream cancel | `dsh-timeout` (`deadline`) |
| Clear the timer | `dsh-timeout` (`[Symbol.dispose]`) |
| Classify the first abort reason after abort | `dsh-timeout` (`timeoutOf`) |
| **Actually terminate the work** | the capability's implementation |
| The default/max *values* | the capability's config |
| The timeout `code` string | the capability (`WEB_FETCH_TIMEOUT` ≠ `BASH_TIMEOUT`) |
The signal only *notifies*; termination is always the listener's job, and the listener differs by capability. bash writes its own `addEventListener('abort', kill)` because the OS process lives outside this runtime and nothing else will kill it; web hands `d.signal` to `fetch` and undici tears down the socket. This is why file read/write/edit take **no** `timeoutMs`: a local syscall is best-effort-abortable at most, a timeout could not force `fsync`/`rename` to stop, and adding one would be an implicit default that violates explicit-over-implicit. Both reference agents leave file I/O untimed for the same reason.
### How each capability consumes it
- **web_fetch** — the tool stays validate-and-forward; the provider's hand-rolled controller + `setTimeout` + manual listener + `finally` + `signal.reason` recovery is replaced by provider-owned `deadline`/`timeoutOf`. A pre-aborted upstream signal still throws `WEB_ABORTED` up front; otherwise `fetch` runs against the fused `d.signal`, and `translateAbortOrNetwork` classifies a thrown error by the signal (`timeoutOf` → `WEB_FETCH_TIMEOUT`, else aborted → `WEB_ABORTED`, else network → `WEB_PROVIDER_ERROR`). The public error-code contract is unchanged, and `TimeoutReason` never crosses the web seam as the public error.
- **bash** — `resolve()` stays a pure request-to-spec step: it clamps with `clampTimeout(request.timeoutMs, config.timeoutMs, config.maxTimeoutMs, 'bash-local: request.timeoutMs')` and carries `request.signal` through unchanged. Foreground `run()` owns the timeout: `using d = deadline(spec.signal, spec.timeoutMs, 'BASH_TIMEOUT')`, then `runBash` receives only `d.signal`. `runBash` no longer owns any timer — it listens for abort and runs its existing SIGTERM→grace→SIGKILL process-group kill, and its `SpawnSpec`/`SpawnOutcome` no longer carry `timeoutMs`/`timedOut`/`aborted` (the executor classifies from the deadline signal instead). `run()` computes `timedOut = timeoutOf(d.signal, 'BASH_TIMEOUT') !== undefined` and `aborted = d.signal.aborted && !timedOut`, so the public seam booleans (`BashRunResult.timedOut`/`aborted`) are mutually exclusive — the shared deadline reports the cause that first cut the command short, and the `code` scope keeps a nested outer deadline from being misread as bash's own timeout. Background `start()` creates no deadline and forwards only the upstream signal, so background tasks stay timeout-free; a task's killed-vs-completed status reads its own `spec.signal.aborted`.
## Consequences
- `runBash`'s outcome no longer independently latches `timedOut` and `aborted`; a timeout and a user abort racing before process close now report a single first-abort cause instead of both being true. The uniform SIGTERM→grace→SIGKILL kill is unchanged, and the seam type `BashRunResult` keeps both booleans (now mutually exclusive), so `dsh-tool-bash`'s result rendering is untouched.
- `SpawnSpec.timeoutMs` and `SpawnOutcome.timedOut`/`aborted` were removed rather than kept as always-zero/always-false vestiges: with `runBash` owning no timer and the executor owning classification, they were read nowhere. This is the one deviation from the literal proposal shape (which passed `timeoutMs: 0` into `runBash`); an always-0 field read by nothing is dead weight under the per-file coverage gate.
- web_fetch shed its bespoke controller/timer/listener/reason-recovery; the classifier now keys off the deadline signal (`timeoutOf` + `aborted`) rather than the thrown error's shape, which is robust across both the request-phase reject-with-reason and the read-phase bare-`AbortError`.
- `AbortSignal.any` and `using`/`Symbol.dispose` enter the repo for the first time here (Node ≥ 24 baseline, already met).
Out of scope, named to mark the boundary: `web_search` can gain an optional model-facing `timeout_ms` once its tool-schema/snapshot coverage is planned; future ripgrep-backed fs discovery tools can consume the same provider-owned deadline shape once they exist; a `tools/execute` waterfall middleware could arm a default deadline for every tool call by driving `exec.signal` — that would be a plugin that *consumes* this library and still only notifies, the hard kill remaining each capability's job.
## Alternatives considered
**A unified timeout *plugin* / `ctx.timeout` service.** Rejected on microkernel grounds. A service that could stop any tool's work would have to understand every capability's termination mechanism (process-group SIGKILL, socket teardown, syscall-boundary checks) — the "kernel knows too much" the architecture forbids. Codex's `ExecExpiration` is scoped to the exec family precisely because the kill it drives (`killpg`) is process-family-specific; MCP and model-stream keep their own. There is no coherent middle layer that owns termination for everything, so the shared piece can only be the pure timing/classification half — a library, not a service.
**Per-tool ad-hoc timeout, no shared code (the prior status quo, and Claude Code's choice).** Rejected because it was already producing divergence and duplicated correctness burden: web_fetch hand-rolled the exact controller/reason logic that future network/process-backed tools would each have to re-derive, and the fusion + `signal.reason` recovery are the error-prone parts. Claude Code tolerates full duplication; this repo has a single shared abort channel (`exec.signal` on every `execute`) that makes a small shared primitive strictly cleaner, so the cost/benefit differs.
**A `withTimeout(promise, ms)` wrapper instead of a signal factory.** Rejected because racing a promise against a timer resolves the *tool-call* promise on deadline without stopping the underlying work — the child process or fetch socket leaks on. Handing out a signal and requiring the capability to listen is what forces a real termination path to exist. This mirrors the "dispose must reach quiescence, not just request it" defensive rule.
**Keep bash's two independent triggers (`killTimer` + `onAbort`) rather than fusing.** Rejected for the convergence goal: fusing into one `deadline` signal removes bash's bespoke timer and gives every capability one shape. The trade-off is that bash's `timedOut`/`aborted` booleans become first-abort classifications rather than independent facts that can both be true when timeout and user abort race before process close. That is acceptable because the result reports the cause that first cut the command short; the termination action stays the same uniform SIGTERM→grace→SIGKILL kill. Note the deliberate non-alignment with Codex: Codex forks its kill by outcome (timeout → immediate SIGKILL; cancel → SIGTERM + 50 ms grace → SIGKILL), whereas the fused signal drives one uniform `kill()` for both, matching Claude Code's unified bash kill. Splitting the kill by `timeoutOf` is possible later if a need appears; there is none now.

110
docs/rfc/implemented/architecture/2026-07-07-tool-call-timeout-policy.md Normal file
View File

@@ -0,0 +1,110 @@
# RFC: Tool-call timeout policy as a plugin
Status: implemented
## Problem
The [timeout/deadline RFC](2026-07-06-timeout-deadline-library.md) extracted the timing-and-classification primitive into `@deepseek-ai/dsh-timeout`, but timeout policy was still attached to individual capabilities and model-facing schemas. `bash` exposed `timeoutMs`; `web_fetch` exposed `timeout_ms`; `web_search` had no model-facing timeout even though providers already honor `exec.signal`; a future grep/glob tool would either import the timeout library directly or invent its own timeout policy. That is the wrong authoring shape for a plugin SDK: a tool author should normally forward `exec.signal` to the implementation it calls, and deployment policy should decide the budget.
At the same time, not every timeout in the repo is a model-facing tool-call budget. Hooks execute command hooks by calling `ctx.bash` directly, not through `ctx.tools.execute()`, and the `bash` model tool multiplexes foreground execution, background start, background polling, and hook reuse through the same backend. Moving every timeout into a tool plugin in one step would conflate those paths and risk breaking hook timeout semantics.
## Decision
Tool-call timeout is a policy that applies only to model-facing tool execution, in three parts:
- `@deepseek-ai/dsh-timeout` remains the shared library that owns `deadline()` and `timeoutOf()`.
- `@deepseek-ai/dsh-tools` has an around-dispatch waterfall, `tools/execute`, between `tools/pre-execute` and `tools/post-execute`.
- `@deepseek-ai/dsh-timeout-policy` reads each tool's declared `timeoutMs` from the registry and wraps a call that has one by deriving a new `exec.signal`.
The execution pipeline is:
```text
ctx.tools.execute(exec)
-> tools/pre-execute
-> tools/execute
-> registry dispatch (the base next())
-> tool.execute(args, exec)
-> thrown tool errors normalize to ToolExecutionResult
-> tools/post-execute
```
The default behavior is conservative: a tool that declares no `timeoutMs` receives no `TOOL_TIMEOUT` deadline from the plugin.
### The `tools/execute` around seam
`@deepseek-ai/dsh-tools` declares a `tools/execute` waterfall whose base `next()` is the dispatch-with-normalization thunk — the same inner `try`/`catch` that turns a thrown tool (or unknown tool) into an `isError` `ToolExecutionResult`. A listener receives `(exec, next)`: it calls `next()` to delegate to dispatch (returning its result, optionally wrapped) or returns a replacement result to short-circuit dispatch. The whole pipeline still sits inside `execute`'s outer try/catch, so a throwing listener becomes an `isError` result, never a turn failure.
That the catch is the base `next` — not something outside the waterfall — is load-bearing: when a provider sees the timeout signal and throws its own upstream-abort error, registry dispatch first converts it to a normal error result, and only then can `timeout-policy` replace the final result with `TOOL_TIMEOUT`.
### The `timeout-policy` plugin
The plugin is `@deepseek-ai/dsh-timeout-policy`, a zero-config function/namespace plugin (`name` / `inject` / `apply`) in the `packages/timeout/` group. The per-tool budget is DECLARED on the tool, not on this plugin: a `ToolDefinition` carries an optional `timeoutMs`, which the owning tool plugin sets from its own config. `dsh-tool-web`, for example, resolves `fetchTimeoutMs` / `searchTimeoutMs` (default 30000) onto the `web_fetch` / `web_search` definitions:
```yaml
- id: timeout-policy
name: '@deepseek-ai/dsh-timeout-policy'
- id: tool-web
name: '@deepseek-ai/dsh-tool-web'
config:
fetchTimeoutMs: 30000
searchTimeoutMs: 30000
```
Keeping the tool name out of this plugin's config is deliberate: a budget keyed by a free-text tool name could be mistyped (`web_fech`) and then silently apply to nothing. Declaring `timeoutMs` on the tool makes that failure class structurally impossible — the enforcer reads `ctx.tools.get(exec.name)?.timeoutMs`, and `exec.name` is the tool being dispatched, so the lookup always resolves and there is no unknown-name path to warn or throw about. `timeoutMs` is validated positive-finite by `defineTool` at definition time. For a tool that declares a budget the listener arms `deadline(exec.signal, timeoutMs, 'TOOL_TIMEOUT')`, swaps the derived signal onto `exec` for the downstream dispatch, restores the caller's own signal afterward, and returns a structured `TOOL_TIMEOUT` result when `timeoutOf(d.signal, 'TOOL_TIMEOUT')` matches. A tool with no declared budget delegates unchanged.
Signal replacement is by **in-place mutation of `exec.signal`**, not by passing a new object to `next()`. Cordis's waterfall `next()` ignores any arguments handed to it and re-invokes downstream listeners with the shared payload array (`vendor/cordis/src/events.ts`), so the documented cordis idiom — mutate the shared object, then delegate — is the only mechanism that reaches dispatch. The plugin restores `exec.signal` to the caller's original in a `finally` so `tools/post-execute` never sees this plugin's (possibly already-aborted) deadline signal.
`timeout-policy` owns both uses of the `TOOL_TIMEOUT` code: the internal deadline code passed to `deadline()`/`timeoutOf()` (scoped so a nested outer deadline reads as an ordinary cancel) and the structured tool-result error code. Its replacement result is:
```ts ignore-check
function toolTimeoutResult(callId: CallId, timeoutMs: number): ToolExecutionResult {
return {
callId,
content: [{ type: 'text', text: `Error: tool call timed out after ${timeoutMs}ms` }],
isError: true,
error: { name: 'ToolTimeoutError', code: 'TOOL_TIMEOUT' },
}
}
```
This is a cooperative deadline. It does not kill arbitrary work by racing the tool promise; the tool or the capability it calls must honor `exec.signal` and reach quiescence. Declaring `timeoutMs` therefore MEANS "this tool is cooperative with `exec.signal`", which the plugin README states as its contract.
No new session event is needed for reconstructability: `TOOL_TIMEOUT` is the final model-facing `tool/result` for that call, so the existing session log already records the content and structured `{ name, code }` error the next model request sees.
### Existing tool adaptation
`web_fetch` and `web_search` are migrated. `dsh-tool-web` keeps ownership of their model-facing schemas, and those schemas expose no timeout knob: `web_fetch` dropped its `timeout_ms` parameter to match the reference-agent shape, and `web_search` stays query-only. The tool bodies do not import `@deepseek-ai/dsh-timeout`; they forward `exec.signal` to `ctx.web`.
`dsh-web-fetch-local` keeps a provider-level timeout (`timeoutMs`/`maxTimeoutMs`) as a large resource backstop for direct `ctx.web.fetch()` callers and misconfigured deployments; it owns no model-facing timeout. When a `TOOL_TIMEOUT` signal reaches the fetch provider first, provider-scoped classification treats it as upstream `WEB_ABORTED`, and the outer `tools/execute` wrapper replaces the final tool result with `TOOL_TIMEOUT`. A shipped web-tool deployment configures the provider backstop above the `timeout-policy` budget so the tool-call policy normally wins for model calls.
`bash` stays on the current backend timeout path. `dsh-tool-bash` continues to expose `timeoutMs` and `run_in_background`; `dsh-bash-local` continues to use `@deepseek-ai/dsh-timeout` for `BASH_TIMEOUT`; hook bridges continue to call `runHook()` and pass `timeoutMs` through `ctx.bash`. This keeps foreground/background/hook behavior stable.
`read`, `write`, `edit`, `todo_write`, `bash_output`, and `bash_kill` do not opt into tool-call timeout: they are local filesystem or short registry/session operations where a deadline would be best-effort only or unnecessary.
A future model-facing grep/glob tool can be implemented on top of `ctx.bash` without importing `@deepseek-ai/dsh-timeout`: it forwards `exec.signal` to `ctx.bash`, and declares its own `timeoutMs` (from its plugin's config) for the enforcer to apply. If bash-local's backend timeout becomes a problem for such a tool, the bash seam can later add a caller-owned-deadline mode; that is outside this cut.
## Alternatives considered
**Name the plugin `tool-timeout`.** The literal RFC name matched the `gen-tool-catalog` completeness guard's `packages/*/tool-*` glob, which requires every match to register a model-facing tool. This plugin registers none — it is a `tools/execute` wrapper — so a `tool-*` name would either fail `verify-tool-catalog` or force a misleading boot entry. The package is `@deepseek-ai/dsh-timeout-policy` in a new `packages/timeout/` group; the cordis.yml `id` can still be `timeout-policy`.
**Keep per-tool timeout handling only.** This was the shape for `bash` and `web_fetch`, and it matches Claude Code and Codex for shell commands. It loses for web-style tools because every new timeout-capable tool must choose validation, cap semantics, docs, snapshots, and classification. The plugin centralizes policy and classification while leaving each tool's schema focused on business input.
**Move all timeout policy out of bash-local immediately.** Cleaner long-term — bash-local would become a pure subprocess executor and all callers would own their deadlines. It loses as the first step because hooks call `ctx.bash` directly and the bash model tool has foreground/background semantics that are not the same tool-call lifetime. Keeping `BASH_TIMEOUT` preserves those paths while tool-call timeout proves itself on simpler tools.
**Use a global default budget for every tool.** Convenient, but it surprises tool authors: any tool that accidentally runs longer than the global budget would start failing once the plugin loads. A per-tool declared budget makes adoption deliberate.
**Expose a model-facing `timeout_ms` override.** Claude Code's `WebFetch`/`WebSearch` and Codex's web tools keep timeout out of the model-call shape. A model override would make timeout part of prompt semantics and force schema/argument-stripping rules into `timeout-policy`. Web timeout stays deployment policy only.
**Let `timeout-policy` match tool arguments itself.** A rule engine such as "disable timeout when `bash.run_in_background` is true" would make the policy plugin know tool-specific argument semantics. Avoided by not migrating bash to tool-call timeout.
**Use `tools/pre-execute` plus `tools/post-execute` instead of a new around seam.** A pre listener could arm a deadline and mutate `exec.signal`; a post listener could classify and replace. That loses because the deadline lifetime would cross two independent waterfalls: a call-id map, cleanup on every pre-deny/tool-throw/post-throw/dispose path, and ordering rules with every other listener. `tools/pre-execute` is also the allow/deny gate, not an execution wrapper. `tools/execute` gives the timeout one lexical scope: arm, delegate, classify, dispose.
**Use `Promise.race` to enforce timeouts for non-cooperative tools.** Rejected for the same reason as the timeout-library RFC: it returns control to the caller while the underlying process, fetch, or provider operation may still be running. The plugin only sends a signal; termination remains the implementation's responsibility.
## Consequences
- `@deepseek-ai/dsh-tools` gains an around-dispatch surface after the interception seams deliberately split pre/post tool hooks. Its contract is narrow — wrap registry dispatch, not replace the pre-gate or post-result policy — and the base `next()` is dispatch-with-normalization so a wrapper never sees a raw tool throw.
- Multiple `tools/execute` listeners compose by ordinary Cordis waterfall order: a listener that calls `next()` wraps downstream listeners plus dispatch; one that returns without `next()` short-circuits them. A deployment combining timeout with a future retry/sandbox/metrics wrapper chooses semantics by registration order ("timeout covers the whole retry" vs "timeout covers each attempt").
- Opt-in by declaration is a deliberate misconfiguration risk: a tool can declare a `timeoutMs` without honoring `exec.signal`, and that tool will not stop on timeout. The plugin contract states that declaring a budget means cooperative; the web tools prove the pattern on tools that already forward the signal.
- During the transition `bash` and the migrated web tools use different timeout paths on purpose: `TOOL_TIMEOUT` is the model-facing tool-call budget, while `BASH_TIMEOUT` remains the bash backend timeout used by bash and hooks.
- Deviation from the literal proposal, recorded per the implemented-RFC rule: the plugin package is `@deepseek-ai/dsh-timeout-policy` (not `tool-timeout`), signal replacement is in-place `exec.signal` mutation before `next()` (not `next({ ...exec, signal })`, which cordis ignores), and the per-tool budget is declared on the `ToolDefinition` (`timeoutMs`, set by the owning tool plugin from its config) rather than mapped by tool name in this plugin's config — so the enforcer is zero-config and a mistyped tool name is impossible. All three are described in `## Decision` above.

170
docs/rfc/implemented/architecture/2026-07-08-agent-scope-contexts.md Normal file
View File

@@ -0,0 +1,170 @@
# RFC: The agent is a registration scope
Status: implemented
## Problem
One application needs to share infrastructure across many agents while letting each agent have its own tools, prompt contributions, policies, and listeners. Shared adapters, persistence, and user interfaces belong to the deployment; a persona, tool variant, or listener often belongs to one agent.
A separate service graph per agent duplicates shared infrastructure. One global registration graph has the opposite failure: an agent-specific contribution can leak into unrelated agents. Contributors need one ordinary registration mechanism that determines both who can see a contribution and when it is cleaned up.
The mechanism also needs a publication boundary. An agent must not become visible before its local world is complete, and teardown must retain that world until final work has stopped.
## Decision
Every live agent owns one flat registration layer exposed as `agent.ctx`. Code registers through the context that owns a contribution; scope-aware services combine deployment-global registrations with exactly one matching agent layer; operations choose that layer from their real agent; and the layer exists for the agent's complete published lifetime.
Cordis is the plugin framework underneath the SDK. A Cordis **context** is the object plugins use to access services and register effects whose cleanup follows that context. The [Cordis primer](../../../cordis-primer.md) explains the framework in more detail.
For most contributors, the complete contract is four rules:
| Question | Rule |
|---|---|
| Where do I register behavior for one agent? | Call the ordinary registration API through `agent.ctx` |
| What does an operation for an agent see? | Deployment globals plus that agent's layer, using the owning service's merge rules |
| Which scoped listeners run? | Unscoped listeners plus listeners registered for the operation's agent |
| How long does the layer exist? | Setup completes before publication; disposal keeps it until work reaches quiescence |
The scope is flat. Resolution never walks parent or sibling scopes, and lifetime ownership does not imply registration inheritance.
```mermaid
flowchart LR
plain["Plain plugin context<br/>cleanup follows the plugin"] -->|"registers into"| globalLayer["Deployment-global layer"]
agentAContext["agentA.ctx<br/>cleanup follows Agent A"] -->|"registers into"| agentALayer["Agent A layer"]
agentBContext["agentB.ctx<br/>cleanup follows Agent B"] -->|"registers into"| agentBLayer["Agent B layer"]
operationA["Operation for Agent A"] -->|"selects"| agentAView["Agent A view<br/>globals plus A local"]
globalLayer --> agentAView
agentALayer --> agentAView
operationB["Operation for Agent B"] -->|"selects"| agentBView["Agent B view<br/>globals plus B local"]
globalLayer --> agentBView
agentBLayer --> agentBView
```
The missing cross-edges are the isolation rule: Agent A's local registrations do not enter Agent B's view, and a parent's registrations do not enter a child merely because the parent owns the child's lifetime.
The companion [runtime-design RFC](2026-07-12-agent-scope-runtime-design.md) explains the implementation and correctness reasoning. The [subagent composition-controls RFC](../feature/2026-07-12-subagent-persona-tool-filter-and-depth.md) owns the separate `persona`, `toolFilter`, and `maxDepth` feature.
### Registration origin chooses visibility and cleanup
A registration made through a plain plugin context is deployment-global and is disposed with that plugin. The same method called through `agent.ctx` contributes to one agent and is disposed with that agent's scope.
| Registration origin | Default visibility | Disposed with |
|---|---|---|
| Plain plugin context | Every eligible agent view | Registering plugin |
| `agent.ctx` | Exactly that agent's view | Agent scope |
Tools, prompt sections and variables, tool restrictions, guards, and scoped event listeners adopt this contract. Named local values ordinarily shadow a same-named global value for that agent; each owning service documents exceptions and merge behavior.
The ordinary contributor pattern is to register the complete local world during agent setup:
```js
const handle = await ctx.agents.create({
agentId: AgentId('reviewer'),
sessionId: SessionId('reviewer-session'),
agentOptions: { model: 'model-name' },
setup(agentCtx) {
agentCtx.systemPrompt.section({
name: 'deployment:persona',
order: 0,
text: 'Review code, but do not modify files.',
})
agentCtx.tools.register({
name: 'review_summary',
description: 'Return the review summary.',
parameters: { type: 'object', properties: {} },
async execute() {
return [{ type: 'text', text: 'review complete' }]
},
})
},
})
ctx.tools.get('review_summary') // undefined: not global
ctx.tools.get('review_summary', handle.agent) // the reviewer-local tool
await handle.dispose()
ctx.tools.get('review_summary', handle.agent) // undefined: scope is gone
```
Setup receives a full trusted Cordis context so it can compose ordinary plugins and services. Its contract is composition-only: driving or publishing the in-flight agent through casts or internal registry calls is unsupported.
### The operation chooses the view
Registration origin and operation subject are separate facts. Calling a service through `agent.ctx` selects where a new registration belongs; it does not bind later reads to that agent.
Tool lookup and execution receive the agent they act for. Prompt assembly receives an assembly context for the agent whose request is being built. Event dispatch receives its domain subject. This keeps shared service instances reusable across agents while making each operation's view explicit.
Only services that adopt the scope contract resolve an agent layer. `agent.ctx` does not automatically change arbitrary Cordis service calls.
### Scoped events keep routing separate from event data
An event about Agent A normally reaches unscoped listeners and A-scoped listeners, not B-scoped listeners. An event without an agent subject reaches only unscoped listeners.
At the Cordis level, `Scoped<T>` is an opaque routing receiver. It carries the filter used to choose listeners but is not the domain object. Event signatures therefore keep the real `Agent`, tool execution, approval request, or other subject as an explicit argument that listeners can inspect.
A listener registered with `{ global: true }` deliberately bypasses contextual audience filtering while its cleanup still follows the registering context. Registry-membership notifications remain unfiltered because they describe shared registry state rather than one agent's operation. The generated [event catalog](../../../cordis-catalog/events.md) is the exhaustive event reference.
### Creation publishes last and disposal revokes last
`ctx.agents.create()` and `resume()` build an unpublished session, scope, agent, and driver. They await `setup`, admit the final session and agent entries, announce them in order, start the loop, and only then return a handle.
An optional creation signal cancels work only while create or resume is pending. After the promise resolves, the returned `AgentHandle` owns explicit disposal.
If loading, setup, admission, or publication fails, the private transaction rolls back everything it prepared. Concurrent operations using the same caller-supplied live ID may both reach setup, but final registry entry admits only one; every loser rejects and cleans its private resources. Sequential reuse after awaited disposal remains valid.
`AgentHandle.dispose()` reverses the boundary. It deactivates creation or driving, waits for synchronous publication to unwind, stops and drains the driver and final session flushes, detaches the agent and session, and finally disposes the scope. Repeated or racing disposal requests join one completion promise.
The calling Cordis context and the concrete AgentLoop factory are structural co-owners. Unloading either disposes the transaction or live agent.
```mermaid
flowchart TB
request["Create or resume"] --> privateWorld["Build private session, scope, agent, and driver"]
privateWorld --> setup["Await composition through agent.ctx"]
setup --> admission["Admit final session and agent entries"]
admission --> publish["Announce lifecycle and start the driver"]
publish --> live["Return AgentHandle"]
privateWorld -->|"failure, cancellation, or owner loss"| rollback["Rollback private work"]
setup -->|"failure, cancellation, or owner loss"| rollback
admission -->|"duplicate or owner loss"| rollback
publish -->|"listener failure or owner loss"| rollback
live -->|"handle or owner disposal"| quiesce["Stop and drain work"]
rollback --> quiesce
quiesce --> detach["Detach agent, then session"]
detach --> revoke["Dispose the agent scope"]
```
## Security and authority are non-goals
Agent scopes compose trusted same-process registrations. They do not sandbox plugins, define a parent-to-child authority lattice, freeze grants at creation, or guarantee that a child can do no more than its parent.
A parent may own a child whose visible tools are wider than its own because lifetime ownership does not donate or cap registrations. A plugin holding a Cordis context also runs in the same process and can call available services directly.
Deployments that need non-escalation require a separate authority representation, propagation rule, and execution check. Parent-subset grants, creation-time authorization snapshots, explicit future-grant APIs, and generic capability/output/termination tags are outside this decision.
## Alternatives considered
The rejected designs either separate visibility from cleanup, cover only one registration family, duplicate shared infrastructure, or conflate lifetime ownership with inheritance.
### Pass an agent option to every registration
An API such as `tools.register(definition, { agent })` repeats scope plumbing in every registry and permits visibility ownership to drift from cleanup ownership. Registering through `agent.ctx` makes both facts follow one Cordis effect owner.
### Filter events while keeping registries global
Listener filtering prevents the wrong hook from running but does not scope tool schemas, executable lookup, prompt sections, variables, or other registered data. Agent-local composition would still require temporary global mutation.
### Create one service graph per agent
The required view is shared deployment services plus one local registration layer. Per-agent graphs duplicate adapters and complicate shared persistence, provider registries, and application boot.
### Inherit parent registration scopes
Parentage describes lifetime and conversation lineage, not a universal merge policy. Hierarchical lookup makes unrelated services inherit accidentally and cannot define security without a separate authority model.
## Consequences
Contributors use one familiar pattern: register shared behavior through a plugin context, register local behavior through `agent.ctx`, select the real agent on operations, and dispose the returned handle. Setup is atomic from an observer's perspective, and teardown preserves local behavior until work stops.
The cost is explicit subject selection, asynchronous programmatic creation, and service-specific scope adoption. Flat registration scope is intentionally not authority, and subagent composition controls remain a separate feature rather than hidden scope semantics.

390
docs/rfc/implemented/architecture/2026-07-12-agent-scope-runtime-design.md Normal file
View File

@@ -0,0 +1,390 @@
# RFC: Agent-scope runtime design and correctness
Status: implemented
## Problem
The [agent-scope contract](2026-07-08-agent-scope-contexts.md) is simple for contributors: register through `agent.ctx`, resolve one global-plus-agent view, publish only after setup, and retain the scope until work stops. The runtime must preserve that contract across a cooperative plugin framework, asynchronous creation, reentrant listeners, durable session commits, and worker or process failure.
The main design risk is adding a second mechanism for every race. Separate reservations, readiness sentinels, cancellation relays, snapshot layers, and protection registries can mirror the same fact until no reader can tell which one is authoritative. That machinery also encourages the runtime to treat trusted typed calls as hostile serialization boundaries.
The implementation needs enough state to preserve real ownership and settlement boundaries, but no more. A correctness reviewer must be able to follow one fact from acceptance through publication and teardown without reconciling parallel representations.
## Decision
The runtime uses one mechanism per independent fact. Scope routing has an opaque carrier; each live registry object has one entry record; each create or resume operation has one transaction; typed same-process calls borrow readonly values; real data boundaries materialize once; the cooperative prompt-assembly result is authoritative; and worker/process code retains separate terminal and quiescence state only where different owners can genuinely race.
The design can be skimmed as seven choices:
| Problem | Authoritative mechanism |
|---|---|
| Select global plus one agent's registrations | Opaque scope key and routing carrier |
| Own one live agent or session | One registry entry captured by its disposer |
| Coordinate create/resume | One `AgentCreationTransaction` |
| Protect durable, queued, model, or wire data | Materialize once at that boundary |
| Pass typed values inside one process | Readonly borrowed contract |
| Compose the model-visible prompt and tool surface | One shared tool view plus the authoritative assembly-waterfall result |
| Coordinate subagent, worker, and process shutdown | One cancellation signal plus the independent terminal/quiescence facts of that boundary |
The rest of this RFC expands those choices in dependency order. It first explains the Cordis mechanics, then scope routing, creation and session commit, tools and prompts, subagents and workflows, and finally the checks that make the reasoning executable.
The [July 8 RFC](2026-07-08-agent-scope-contexts.md) remains the contributor contract. The separate [subagent composition-controls RFC](../feature/2026-07-12-subagent-persona-tool-filter-and-depth.md) owns `persona`, `toolFilter`, and `maxDepth`; this document discusses only how their setup fits the lifecycle.
## Cordis model: context, fiber, effect, receiver, and waterfall
Five Cordis ideas are required to understand the implementation. A context selects services and registration ownership; a fiber is one live plugin or child lifecycle; an effect attaches cleanup to a fiber; an event receiver selects listeners; and a waterfall lets listeners transform or veto an operation in sequence.
### A context is an ownership path through one service graph
All agents share one Cordis service graph. A derived context does not clone `ToolRegistry`, `SystemPrompt`, persistence, or model adapters; it changes how registrations made through that context are tagged and which effects own their cleanup.
`agent.ctx` is such a derived context. Service calls still reach the shared instances, while a registration can inspect its calling context and store a contribution under the nearest scope key. Ordinary plugin contexts carry no scope key and therefore register globally.
### Fibers and effects make cleanup structural
A Cordis fiber is the live instance created when a plugin or child context is activated. Its state records whether that lifecycle is active, unloading, failed, or disposed. `ctx.effect()` and `ctx.on()` return disposers and also attach those disposers to the registering fiber, so unloading a plugin or agent scope removes everything registered through that context without a separate inventory.
The vendored Cordis fiber implementation establishes ownership before arbitrary setup or `internal/plugin` observers run. A reentrant unload can see the child fiber or effect that has started, reject effects added after unload begins, and join cleanup already started through a public single-shot disposer. Teardown observers are contained individually so one callback cannot prevent structural cleanup.
These are framework lifecycle guarantees rather than agent-specific policy. Agent creation depends on them because setup can activate arbitrary plugins and synchronously reenter owner disposal.
### Receivers route listeners; waterfalls compose decisions
Cordis filters listeners using the dispatch receiver (`this`), while harness listeners need an explicit agent, execution, request, or other subject. `Scoped<T>` marks the receiver expected by a scoped event declaration, but the runtime carrier deliberately exposes no subject API.
Product helpers therefore construct the carrier and pass the domain subject separately. This prevents listener routing from becoming an alternate object model and keeps event signatures understandable without knowledge of carrier internals.
A Cordis waterfall is middleware-style dispatch. Each listener receives `next()`: calling it delegates to the remaining listeners and base operation, while returning without it vetoes or replaces the downstream result. Waterfalls power prompt assembly and tool policy; ordinary emit events notify synchronously, and parallel events await all listeners without a veto result.
## Scope routing: one opaque key selects one layer
The scope package implements the smallest object needed for Cordis routing. Its carrier holds only a composed service filter and scope predicate, while the package records the opaque key privately and exposes the scope fiber's quiescent disposer separately.
### Scope identity uses object identity
A `ScopeKey` is an opaque object compared by identity. The harness uses the live `Agent` as its own key, but the primitive is domain-neutral and supports other scoped owners.
`createScope(parent, key)` returns a scope whose `ctx` shares the parent's services and whose effects are tagged with that key. `scopeOf(ctx)` reads the nearest registration key. `scopeTarget(base, key)` creates the event receiver whose filter preserves the base receiver's Cordis service filter, then admits unscoped listeners and listeners with that exact key.
The receiver is a small carrier rather than a transparent proxy for the domain object. Code that needs the agent receives the explicit event argument; code that needs registration ownership receives `agent.ctx`.
### Registry reads overlay one exact map
Scope-aware registries store global contributions separately from identity-keyed local contributions. A read resolves the global layer and at most one local layer; it never traverses parentage.
Each service retains its domain rule. Named prompt values and tools use local shadowing, tool restrictions filter globals before local tools are added, and events select listener audiences rather than registered data. Scope supplies identity and ownership, not a universal merge algorithm.
### Fused dispatch helpers prevent subject drift
`agentEvents(context, agent)` constructs the agent's carrier and injects the same agent as the event subject. Session, tool, approval, prompt, and subagent services likewise derive routing from the object they already own instead of accepting an unrelated key.
The type marker rejects ordinary bare-receiver mistakes, and development invariants cover direct JavaScript or casted dispatch. The subject remains explicit because routing correctness and useful event data are different concerns.
## Agent creation: one transaction owns the complete operation
Create and resume are one asynchronous lifecycle with several phases, not several lifecycles. `AgentCreationTransaction` owns caller and factory liveness, optional cancellation, private resources, publication, rollback, and the memoized teardown observed by every owner.
### Registry entries are the only live identity records
AgentRegistry and SessionStore each keep one entry per live object. The entry holds the stable ID, object, scoped carrier, and the small amount of publication or append state that belongs to that object.
A detach closure captures its exact entry. It deletes only when the map still points to that entry, so an old disposer cannot delete a later object that reuses the same ID. No registry rereads a mutable caller object to decide identity.
There is no reservation API. Caller-supplied IDs are admitted at final entry. Concurrent same-ID operations may both complete private setup; exactly one final `enter()` succeeds, and every loser rolls its private resources back. Sequential reuse is valid after the earlier disposer reaches quiescence.
### The transaction owns preparation before awaiting it
The transaction is installed under both the calling Cordis context and the concrete AgentLoop factory before persistence load or setup can suspend. It also observes an optional create/resume signal until the public operation settles.
Create prepares a new Session. Resume loads and validates the persisted Session before preparing the same live session identity. Both paths then build the scope, agent, and driver and invoke the same setup/publication algorithm.
The factory stores concrete trace targets but invokes them through a caller-bound Cordis trace. This preserves dependency origin and caller ownership without stacking trace proxies.
### Setup is trusted composition inside a private world
Setup receives the full child context and may await plugin activation. It can register tools, prompt sections, restrictions, listeners, and other effects, but the public contract does not support driving or publishing the in-flight agent through casts or internal registry calls.
The transaction races asynchronous load and setup against deactivation rather than waiting forever for a promise owned by external code. If cancellation or owner unload wins, public creation rejects after transaction-owned cleanup even when the external promise never settles.
### Publication has one ordered commit path
Publication admits and announces resources in the order required by observers:
1. Enter the session.
2. Enter the agent.
3. Announce `session/created`.
4. Announce `agent/created`.
5. Enable public driving.
6. Emit `agent/session-start`.
7. Start the driver.
The agent never drives before both registries and creation notifications agree. A synchronous listener may veto or dispose an owner; the transaction records publication in progress and waits for that callback stack to unwind before teardown continues. Every creation announcement that begins has a matching disposal announcement during rollback.
The sequence diagram isolates the non-obvious race: a synchronous creation listener can request disposal while the publication call stack still owns both registry entries. Teardown must deactivate immediately but wait for that stack to unwind before stopping and detaching anything.
```mermaid
sequenceDiagram
participant Tx as AgentCreationTransaction
participant Registries
participant Listener as Synchronous listener
participant Driver
Tx->>Tx: mark publication in progress
Tx->>Registries: announce agent/created
Registries->>Listener: invoke inside the same call stack
Listener->>Tx: dispose reentrantly
Tx->>Tx: deactivate, teardown waits for publication
Tx-->>Listener: disposal request accepted
Listener-->>Registries: return
Registries-->>Tx: announcement unwound
Tx->>Tx: resolve publication settlement
Tx->>Driver: stop and drain
Tx->>Registries: detach agent, then session
Tx->>Tx: dispose scope and resolve teardown
```
### Teardown preserves work before revoking registrations
Every teardown request joins one memoized path. The order is:
1. Deactivate creation or driving and let synchronous publication finish.
2. Stop and drain the driver, including idle injection flushes.
3. Detach the agent.
4. Detach the session.
5. Dispose the agent scope.
6. Retire transaction ownership tracking.
This order lets final agent and session events use the matching scoped listeners and keeps persistence observers attached through the final flush. Scope disposal comes last because registration revocation is the externally visible lifetime boundary.
## Session append: materialize, validate, commit, notify
Session events cross a durable boundary, so append owns their data. The rest of the algorithm uses one attached entry and one commit point.
### Durable data is materialized once
Session headers, seeds, and appended events are lossless JSON data. The Session constructor or append path materializes and validates them before storage and exposes frozen snapshots, so later caller mutation cannot change persistence, replay, or model reconstruction.
This is a real ownership boundary: the values leave the caller, may be persisted, and must reconstruct the same request later. It is intentionally stricter than a typed same-process callback or registry definition.
### Pre-commit listeners can veto; post-commit observers cannot
Append follows one sequence:
1. Materialize the durable event and surface intent.
2. Claim the SessionEntry and reject reentrant append on that entry.
3. Resolve scoped callbacks and run internal invariant validation.
4. Push exactly once; this is the commit point.
5. Notify each observer independently, containing synchronous and asynchronous failures.
6. Release append state and honor a detach requested during publication.
No observer error makes a committed event look uncommitted, and one bad listener cannot starve later listeners. Session invariants stage their transition before commit and apply it only when the same event reaches the contained post-commit observer.
`flush()` starts every persistence listener and awaits every result before reporting failure. This deliberate all-settled behavior prevents a synchronous failure from starving another backend or final flush.
## Trust boundaries: copy only when ownership actually changes
The runtime distinguishes typed in-process contracts from serialization and durability boundaries. This is the main simplification rule for values and callbacks.
| Boundary | Ownership rule |
|---|---|
| Typed service/plugin call in the same process | Borrow readonly values and callbacks |
| Parsed plugin configuration or external file | Validate semantic and structural input |
| Queued inbox message | Materialize before asynchronous consumption |
| Model/tool JSON input or output | Materialize at the model/tool boundary |
| Durable session or persistence data | Materialize and validate before commit |
| Worker, process, or wire message | Serialize, validate, and own the decoded value |
Tests that fabricate hostile getters, replace typed callbacks after handoff, or cast fake service objects do not define a production contract by themselves. The runtime keeps checks where data crosses a parser, queue, model, durable, file, worker, process, or wire boundary and relies on readonly types plus plugin discipline inside the trusted process.
Callback containment is separate from data ownership. Listeners are arbitrary extension code and can throw even when their arguments are trusted; publication and post-commit paths still contain failures according to their event contract.
## Tools and prompts: one view, authoritative assembly, committed outcomes
Tool presentation and execution share one private resolver. Prompt assembly remains trusted cooperative composition: registries supply the ordered input, and the assembly waterfall's returned value is exactly what the loop logs and sends. Execution uses separate one-way boundaries only where policy or outcome settlement must be monotonic.
### One resolver defines the tool view
The private resolver applies the current presentation mode, live global restrictions, exact local overlay, and local shadowing. Schemas, lookup, execution, Code Mode SDK generation, and restriction validation all use that resolver or its pre-restriction global-name view.
The [subagent composition-controls RFC](../feature/2026-07-12-subagent-persona-tool-filter-and-depth.md#tool-filtering-is-one-live-global-view-rule) owns the user-visible allow/deny semantics. The implementation requirement is agreement: a filtered-away global cannot remain executable through a different lookup path, and a locally shadowed definition is the same definition presented and executed.
`ToolRestriction` accepts readonly allow/deny names and compiles them into internal sets. Multiple restrictions intersect. Public `visible()` and `knownNames()` methods are unnecessary because only the registry needs the intermediate views.
### Tool execution owns identity and boundary materialization
The registry assigns every execution a fresh branded `Symbol` token. Nested Code Mode calls carry the outer token as `parent`, so structured output can correlate an inner capture with its enclosing `run_code` result by identity.
A fresh registry-assigned Symbol provides collision-free execution identity without a WeakSet membership registry. Callers cannot supply the execution's own token through `ToolExecutionInput`; they only receive the pipeline-owned `ToolExecution` after the registry creates it. This is a trusted typed contract, not a runtime defense against arbitrary casts or JavaScript callers.
Arguments are materialized once where model/tool JSON enters the pipeline. Pre-, around-, and post-execute listeners operate on the typed execution and decisions. Call ID correlation, approval, monotonic guards, and Code Mode nesting remain explicit relational checks.
After the last post-execute listener, the registry materializes and freezes the accepted final result once. Every synchronous `tools/result` observer receives that exact committed object, and observer failures are contained individually. An outer pipeline failure is normalized into a committed error result, so observers can discard staged work against the same authoritative boundary.
### The assembly waterfall owns the final model-visible composition
SystemPrompt first resolves the global-plus-agent sections, variables, and tool providers into a deterministic registry contribution. The scope-filtered `system-prompt/assemble` waterfall may then reorder, replace, add, or remove any section, variable, or schema. Its returned assembly is authoritative; there is no later restoration pass and no finality metadata on ordinary prompt sections, tool definitions, or provider results.
This is a trusted same-process extension seam, not an authority boundary. A listener that changes Code Mode's `run_code` schema or `tools:sdk` instructions, or a structured child's capture schema or instruction, owns preserving a coherent protocol in the assembly it returns. ToolRegistry still reserves `run_code` against ordinary tool registration and restriction because those are registry invariants, but assembly middleware remains free to transform the final model-visible surface.
Scope solves the real isolation problem directly. Structured-output contributions register in the child's exact scope, while Code Mode derives its transport and SDK from the same resolved tool view. A second named-protection system would need another ownership and collision rule across arbitrary schema providers—including providers that intentionally contribute duplicate names—without creating a new trust boundary.
### Structured output commits only authoritative outcomes
Structured output combines child-scoped composition with a two-phase execution commit. The child registers its `structured_output` tool and instruction before publication; a trusted assembly listener may transform those ordinary contributions and is responsible for preserving the protocol if the child is expected to complete. The tool body validates a candidate and stages it by the current `ToolExecution`, but successful capture is decided only by immutable `tools/result` observations.
For a native call, the observer deletes the stage and commits its value only when that exact execution's final result succeeds. A post-execute block or outer pipeline failure therefore cannot leave a captured value behind.
For a Code Mode SDK call, the inner successful result records `{ parentToken, value }` rather than committing. The observer waits for the `run_code` execution whose token matches `parentToken` and commits only if that outer final result also succeeds. Program failure, runtime abort, or outer post-policy denial discards the pending value.
Once a value is pending or committed, a scoped monotonic guard denies later tool calls. After commit, the ordinary serial `agent/turn-stop` listener returns a stop decision after continuation and steering have already folded. A schema-validation failure remains an ordinary `INVALID_ARGS` tool error and leaves the child able to retry within the same turn.
Pure Code Mode's registry contribution omits `structured_output` from native wire schemas and exposes it through the generated SDK. The assembly waterfall may deliberately change that presentation; execution still validates against the child-scoped definition, and the listener owns the consistency of any alternate model-visible route it creates.
### Three execution boundaries are deliberately one-way
Prompt assembly is intentionally cooperative, but three execution facts need one-way settlement after their extensible stages:
| Boundary | Final power | Why ordinary listener order is insufficient |
|---|---|---|
| Tool pre-policy | Deny monotonically | A later listener must not re-allow an already denied call |
| Tool result | Observe the immutable committed outcome | Structured output must commit only the result that actually escaped the pipeline |
| Turn continuation | Stop after ordinary continuation folding | A committed terminal output must end the turn |
`ToolGuard` is the monotonic policy registry. Committed tool observation is the contained `tools/result` point described above. Terminal structured output listens on the ordinary serial `agent/turn-stop` fold after normal continuation and steering decisions; no public `strictSerial()` dispatcher is needed for the typed listener contract.
### Skill and approval services trust typed callers
Skill registry definitions and approval policies are readonly same-process contracts. Their services do not clone callback objects or defend against post-handoff callback replacement.
Skill still validates external skill files and parsed provider output, routes catalogs through the calling agent's tool view, and disposes registrations exactly. Approval still resolves policy, observes cancellation, routes `approval/request` by `request.agent`, records the durable audit pair, and contains answerer and post-commit observer failures.
## Subagents: readiness is the start promise
Subagent startup has one ownership transfer. The provider owns partial resources until its start promise fulfills with a ready published run; the caller owns the returned run and must dispose it.
### The service contract has one cancellation channel
`SubagentProvider.start()` and `SubagentService.start()` return `Promise<SubagentRun>`. The promise fulfills only after the backend has established the child it promises, so callers and `subagent/start` observers never need a second `run.started` readiness promise.
`SubagentStartRequest.signal` is required. Aborting it requests cancellation during startup and after readiness. `SubagentRun.dispose()` also requests cancellation and awaits quiescence. There is no separate public `run.cancel()` channel.
Optional `sendMessage()` supports a live backend that can accept steering. Optional `resume()` returns `Promise<SubagentRun>` because the resumed child has the same asynchronous readiness boundary.
The service validates provider capabilities and request semantics before calling the provider. A provider rejection cleans any partial resources before the rejection escapes and emits no `subagent/start`/`subagent/end` pair. After fulfillment, the service attaches result observation, emits scoped start, and returns the run. Provider removal prevents later starts but does not revoke a run already accepted by the provider.
### In-process providers reuse the core transaction
Spawn and fork share one in-process driver. It creates the child through `parent.ctx`, passes the required signal into the core creation transaction, and installs persona, tool restriction, and structured-output contributions during unpublished setup.
The provider awaits creation and returns only the published run. At the handoff, core creation detaches its creation-only abort listener; the provider immediately rechecks the signal before installing the live-run listener, so an abort in that narrow interval disposes the new handle instead of escaping cancellation. Parent teardown follows the child because the operation belongs to `parent.ctx`; provider unload blocks new starts but does not become a second revocation owner for accepted runs. The run disposer cancels the child and awaits the AgentHandle's ordered teardown.
Spawn uses an empty session seed. Fork uses a validated completed-turn prefix. Conversation seeding changes history only and does not import scope, tools, services, or authority.
### ACP providers own the process until readiness or cleanup
An ACP provider crosses a real process and wire boundary, so it retains validation, environment scrubbing, message serialization, abort/process races, and kill-to-exit quiescence.
Start resolves only after `initialize` and `newSession` succeed. Abort, spawn failure, RPC failure, or invalid startup response reaps the process before rejection. After readiness, result maps the ACP prompt outcome and streamed output; dispose requests cancellation, closes the connection, and awaits process exit through one memoized path.
## Workflows and ACP UI: retain only independent async facts
Worker and editor bridges need more state than same-process registries because messages, process death, and rendering can settle independently. Their state is organized around those real facts rather than duplicate cancellation protocols.
### Workflow children are pending starts or published records
The workflow host keeps pending provider-start promises and published child records. A child moves from pending to published only when async `SubagentService.start()` fulfills; rejected starts clean their partial provider work and produce no child lifecycle pair.
One host-owned AbortController supplies the required signal to pending and live children. Closing workflow admission aborts that signal, so there is no duplicate `ChildCancel` worker RPC or explicit host-side `run.cancel()` fanout. Quiescence waits for both pending starts and published child disposal.
The worker boundary still serializes requests and outcomes. The host retains first-terminal-outcome arbitration, exact child accounting, worker-death handling, grace termination, late/duplicate message rejection, and bounded cleanup because result receipt, worker exit, and child quiescence are genuinely independent facts.
### Terminal result and physical cleanup remain separate
The workflow result records the first accepted terminal outcome according to the public precedence rules. Cleanup can continue after that result is chosen: live children still need disposal, a worker still needs termination, and a slow external backend may outlive the configured grace bound.
Public disposal claims its memoized promise before invoking callbacks. Worker death closes admission before processing any queued late child request, synthesizes missing lifecycle ends, and starts child/process cleanup without rewriting an outcome already claimed.
### ACP prompt settlement does not depend on rendering success
The ACP UI correlates a prompt with its observed turn directly. It does not scan from a `logWatermark` or use session status as a second reconciliation oracle.
Prompt handling settles correlation in a `finally` around transcript rendering. A rendering failure can fail presentation, but it cannot skip prompt settlement or leave the session permanently in flight. Concurrent loads of the same persisted caller-supplied session ID remain excluded because that is a real persistence identity race, not a UUID collision concern.
## Correctness enforcement
The design is enforced at types, runtime escape points, generated contracts, and behavioral tests. No one layer is asked to prove what it cannot observe.
### Types make the ordinary path hard to misuse
Readonly contracts describe borrowed same-process values. `Scoped<T>` marks event receivers, `agentEvents()` fuses carrier and subject, tool inputs omit registry-owned tokens, and subagent async return types expose readiness directly.
TypeScript cannot govern JavaScript casts, direct Cordis dispatch, process messages, or durable files, so runtime enforcement remains at those escape points.
### Runtime invariants cover cross-service facts
The invariants plugin verifies that every declared scoped event uses a marked carrier and that event families exposing a subject use the matching key. Session trace validation stages before append commit and advances after the same event commits.
The plugin does not police trusted setup by scanning registries or reject prompt assembly objects fabricated through casts. Those checks would turn composition contracts into speculative runtime machinery without protecting a real external boundary.
### Generated artifacts keep public contracts aligned
The event catalog, service catalog, producer/consumer matrix, configuration catalog, module graph, tool catalog, and type-equivalence blocks are generated or freshness-gated from source. `verify-scoped-dispatch` keeps the declared scoped-event set aligned with runtime invariant coverage.
Behavioral tests pin scoped routing and disposal, final-entry collision cleanup, publication rollback, ordered quiescence, durable pre/post-commit behavior, live tool filtering across presentation and execution, cooperative prompt assembly, structured-output commit in native and Code Mode, async subagent startup and signal cancellation, worker terminal arbitration, ACP settlement, and process teardown.
## Alternatives considered
The [July 8 RFC](2026-07-08-agent-scope-contexts.md#alternatives-considered) owns alternatives to the public flat-scope contract. The alternatives here concern implementation shape.
### Use a transparent proxy as the scope carrier
A proxy that impersonates the subject must preserve property, callable, constructable, private-field, descriptor, and proxy-invariant behavior that listener routing never needs. A small opaque carrier keeps the filter and key while the explicit event argument carries the subject.
### Reserve agent and session IDs before setup
Reservations prevent duplicate private setup work but require cross-service capabilities, release ordering, abandoned-reservation cleanup, and prepared-object binding. IDs are caller-supplied and concurrent reuse is caller error; final entry can choose the winner while the losing transaction rolls back cleanly.
### Snapshot every typed same-process argument
Universal copying defends against stateful getters and callers that violate readonly contracts, but it adds allocation, duplicated validators, and paths that can forget to copy. Materialization belongs at parser, queue, model, durable, worker, process, and wire boundaries where ownership actually changes.
### Give readiness, cancellation, and disposal separate controllers
Parallel sentinels can all mirror whether one operation is live. One transaction or start promise owns the operation; separate promises remain only where publication unwind, external work, terminal result, and physical quiescence can settle independently.
### Keep synchronous subagent start plus `run.started`
This splits provider acceptance from readiness and forces every consumer to register a partial run, attach result observation, await readiness, and clean up readiness failure. An async start promise makes provider-to-caller ownership transfer the readiness boundary itself.
### Restore selected prompt or tool contributions after assembly
A post-waterfall restoration pass would create a second composition rule after the documented cooperative seam. Correctly assigning canonical presence or absence would also require provider ownership and collision rules for arbitrary tool-schema providers, whose ordinary output may contain duplicate names. Scoped registration already supplies the required per-agent isolation, and trusted assembly listeners own the protocol consistency of what they return, so named restoration adds machinery without establishing an independent boundary.
### Remove worker/process lifecycle guards with same-process hardening
Worker messages, process death, and durable input do cross ownership and serialization boundaries. First-outcome arbitration, validation, environment scrubbing, and quiescent process cleanup remain necessary even though hostile same-process callback machinery does not.
## Consequences
The implementation is smaller and its proof follows the same shape as its ownership graph. One key selects a layer, one entry owns a live registry object, one transaction owns creation, one resolver owns a tool view, and one async promise transfers subagent ownership.
### What the design guarantees
- A scoped contribution is visible only in its exact agent view and is disposed with that scope.
- Create and resume expose no partially configured handle; final-entry losers and publication failures clean every prepared resource.
- Disposal retains scoped listeners and persistence through driver drain and final session work, then revokes the scope.
- Durable, queued, model, worker, process, and wire values are owned at their real boundary; typed same-process values follow readonly contracts.
- ToolRegistry's presentation, lookup, and execution resolve the same live view before expert assembly transforms, and committed results have one immutable observation point.
- Registry contributions are deterministic inputs, while the trusted assembly waterfall owns the final model-visible composition.
- Subagent start returns only a ready run, required signals cancel pending or live work, and disposal reaches the backend's quiescence contract.
- Worker/process result precedence and cleanup remain correct under death, late messages, and bounded teardown.
### Costs and limits
Scope-aware services still maintain global and identity-keyed maps, and operations must carry their real agent explicitly. Async create/resume and subagent start require callers to await ownership transfer and dispose returned handles.
A trusted `system-prompt/assemble` listener can remove or replace Code Mode and structured-output protocol pieces. This is deliberate: the listener owns final composition and must preserve any protocol the deployment expects to remain usable.
The design trusts typed plugins in the same process. It does not defend against arbitrary casts, stateful getters, mutation that violates readonly contracts, or a plugin deliberately using ambient service access outside the supported composition API.
The [security and authority non-goal](2026-07-08-agent-scope-contexts.md#security-and-authority-are-non-goals) remains fundamental. These mechanisms prove registration composition, publication, and lifetime ownership; they do not prove confinement or parent-to-child non-escalation.

57
docs/rfc/implemented/feature/2026-06-14-acp-agent-client-protocol.md Normal file
View File

@@ -0,0 +1,57 @@
# RFC: Agent Client Protocol (ACP) support — drive the coding agent from external editors
Status: implemented
## Problem
The harness originally exposed agents only through a readline loop. That surface could carry text, but it gave an editor no structured way to create or resume sessions, correlate prompt completion, stream reasoning and tool activity, render tool-specific UI, ask for permission, or cancel one conversation without disturbing another. ACP defines those interactions as JSON-RPC over stdio, and Zed is the target client used to make concrete compatibility decisions.
The bridge must preserve the harness's existing ownership boundaries. It cannot depend on the concrete agent loop, bypass the tool registry, execute shell commands in the editor, or invent a second source of session truth. stdout is also the protocol transport, so any accidental log output corrupts the connection.
## Decision
`@deepseek-ai/dsh-acp` is a UI/client-driver plugin under `packages/ui/acp`. It uses `@agentclientprotocol/sdk`'s `AgentSideConnection` over stdin/stdout and programs only interface services: the agent create/resume factory, session persistence, tool registry, user interaction, and optional approval/bash capabilities. It does not change the agent loop and is not a capability-seam implementation.
The bridge implements the following stable session path:
- `initialize` negotiates the protocol version, advertises text plus `resource_link` prompts, and advertises `loadSession`.
- `session/new` validates an absolute `cwd`, stores it in `SessionHeader`, creates an agent through `ctx.agents`, and returns any composition-backed config options.
- `session/load` validates the requested cwd against persisted metadata before constructing an agent, reserves the id across the asynchronous resume, replays user/assistant/tool events as ACP updates, and reports the resumed config-option fold.
- `session/prompt` accepts text and resource links, rejects unsupported or empty content, allows one in-flight prompt per session, and settles against that prompt's owning `turn/end`. An error turn rejects the RPC; other closed turn reasons map through a total ACP stop-reason codec.
- `session/cancel` calls the queue-aware agent cancel path and settles only the addressed session's prompt.
Tool-call presentation remains tool-owned. A tool's `presentCall` and `presentResult` return the `generic`, `terminal`, or `diff` render-intent variants; the bridge switches on that union and maps it to ACP. Presenter-less tools receive a generic fallback. Bash terminal cards use Zed's capability-gated `_meta.terminal_info`, `_meta.terminal_output`, and `_meta.terminal_exit` convention; the harness still executes the command through `ctx.bash`, preserving sandbox, environment scrub, ownership, and cwd. Clients without that extension receive ordinary text content. Filesystem tools provide diff cards and file locations without hard-coded tool-name branches in the bridge.
Permission handling is an answerer on the [user-approval seam](2026-07-06-approval-seam.md), not an ask-every-tool policy in ACP. An `approval/request` for a bridge-owned agent with a call id becomes `session/request_permission` on that agent's editor session, with one-shot allow/reject choices. Foreign or call-less requests delegate; a missing or failed answerer remains fail-closed. The plugin that asks—such as a pre-execute policy or bash escalation—owns the decision to ask.
The bridge advertises ACP config options instead of session modes. `sandbox-mode` exists only when the mounted bash executor reports sandbox capability, and `approval-policy` exists only when `ctx.approval` is composed. Each option is an independent select whose current value is the session event fold over the composition default. `session/set_config_option` validates against the owning domain vocabulary and writes through `setSandboxMode` or `setApprovalPolicy`. An open-turn switch appends immediately; an idle switch is overlaid in the response and anchored at the next turn start. Until that anchor it is memory-only and a crash reverts to the durable fold. ACP session modes are deliberately not modeled because one mode list cannot represent these orthogonal knobs and config options are the forward protocol surface. Runtime model selection remains outside this decision; `AcpConfig.model` is connection-wide.
The bridge also provides the ACP-backed `UserInteractionProvider`: `ask_user_question` requests become form elicitations on the owning session. Select, multi-select, option descriptions, and custom-answer override semantics are preserved.
Lifecycle ownership is explicit. The bridge holds an `AgentHandle` per live session. Disconnect and Cordis disposal cancel pending prompts, dispose every handle in parallel, await loop quiescence and persistence flush, and then remove the records. Stream notification failures are contained so a vanished client cannot corrupt an agent turn. The ACP app composition loads no stdout logger; a test guards stdout as framed JSON-RPC only.
The precise supported and deferred protocol rows live in [`packages/ui/acp/acp-feature-support.md`](../../../../packages/ui/acp/acp-feature-support.md); the package README is the operational contract.
## Alternatives considered
**A prepended `tools/execute` listener that asks on every ACP-owned call** — rejected. It would hard-code permission policy into the UI bridge, ask even when no policy requires it, and could not serve approval requests that arise after execution begins. The shared user-approval seam keeps mechanism, asking policy, and UI answerer separate.
**Inject the concrete `agentLoop`** — rejected. Agent creation, resume, idle observation, and disposal are interface-level ownership operations on `dsh-agent`; a UI plugin does not need a dependency-rule exception.
**Execute bash through ACP `terminal/*`** — rejected. That would move execution outside the harness and bypass its sandbox, credential scrub, task ownership, cwd resolution, and session log. Terminal metadata is presentation only.
**Represent sandbox and approval as ACP session modes** — rejected. They are independent composable settings, while a single current mode is mutually exclusive. ACP config options represent both without a cross-product and match the protocol's forward direction.
**Hijack stdout defensively** — rejected. Process-wide monkey-patching is outside Cordis effect ownership and races the protocol transport. The app composition owns stdout purity.
## Consequences
Editors can create, load, prompt, cancel, render, ask, and reconfigure multiple harness sessions over one ACP connection without a loop-specific dependency. The session event log remains the durable source for replay, prompt settlement, cwd, and per-session configuration. Tool presentation and human-answer channels remain extensible plugin contracts instead of ACP-specific behavior.
The bridge deliberately does not implement session list/delete/resume/close capabilities, MCP passthrough, additional directories, image/audio/embedded-resource prompts, runtime model selection, plans, slash commands, usage updates, editor filesystem delegation, or the ACP terminal execution sub-protocol. The feature checklist records these as unsupported rather than silently accepting them.
An idle config selection is truthful in the live response but not durable until the next turn anchors it. Crashing before that boundary loses the pending selection; this is the cost of keeping session events turn-enclosed and replay-safe.
## Verification
The ACP suites cover the in-memory protocol codec, create/load replay, exact prompt settlement, cancellation races, unsupported content, tool presentation, terminal capability fallback, permission outcome mapping, config-option validation and persistence, multi-session isolation, disconnect/disposal quiescence, and HMR cleanup. Snapshot and built-bin tests exercise the app composition, while the real-API e2e self-skips without a key.

37
docs/rfc/implemented/feature/2026-06-14-acp-multi-session.md Normal file
View File

@@ -0,0 +1,37 @@
# RFC: Multiplex concurrent ACP sessions over one connection
Status: implemented
## Problem
An ACP editor can keep several conversations alive over one agent subprocess. A single-active-session bridge would force extra processes and would not match Zed's client model, which tracks multiple session ids and concurrent loads. Multiplexing introduces isolation risks: events, prompt completion, cancellation, permission prompts, config selections, and predictable background-task ids must never cross session boundaries.
## Decision
The ACP bridge stores live sessions in `Map<SessionId, SessionRecord>` and keeps a `WeakMap<Agent, SessionId>` reverse index for agent-scoped callbacks. A record owns its agent handle, in-flight prompt, live tool-call presentation state, pending idle config switches, session cwd, and client capability snapshot. A separate loading-id set reserves each id before asynchronous resume so two pipelined loads cannot construct duplicate agents; distinct ids may load concurrently.
Every `session/event` and `agent/status` callback resolves the owning record before sending or settling anything. Each session permits one in-flight prompt independently. The prompt records a log watermark, captures its own `turn/start`, and settles only on the matching `turn/end`; a late end from a cancelled prior turn cannot resolve a newer prompt. `session/cancel` addresses one record and calls only that agent's queue-aware cancel path.
Permission ownership uses the same reverse index. The ACP `approval/request` answerer prompts only the editor session that owns the requesting agent and delegates foreign requests. User-interaction elicitations likewise route by agent ownership. Per-session sandbox and approval config values fold only that session's events, with pending idle switches stored on that record until the next turn anchors them.
Background bash tasks carry an opaque owner token equal to the owning session id. `bash_output` and `bash_kill` compare the caller's token with the executor's task ownership before reading or killing; a predictable task id alone grants no access. Ownership is stored with the executor task, so a tool plugin reload does not erase it.
Connection teardown clears the live map, settles each pending prompt as cancelled, and disposes all `AgentHandle`s in parallel. Each handle stops and awaits its loop, flushes the session while attached, unregisters the agent, and removes the session. Teardown is memoized and shared by client disconnect and plugin disposal.
## Alternatives considered
**One live session per connection** — rejected. It adds process overhead and contradicts the target client's multi-session shape without removing multiplexing needs from the editor.
**A per-session `ctx.extend()`** — rejected. A child context does not by itself create a child plugin fiber, so listeners would still belong to the bridge fiber. The implemented bridge instead uses global listeners with explicit O(1) demultiplexing and per-session owned records; agent lifecycle is owned by `AgentHandle`.
**Agent object identity as bash-task ownership** — rejected. A resumed or replaced agent object may legitimately represent the same durable session. The opaque session token is the cross-boundary identity that should survive plugin reloads.
## Consequences
N sessions can stream, prompt, request permission, switch config, and run background tasks concurrently without interleaving or cross-settling. A cancel or dispose in one session does not affect its neighbors. The bridge pays for explicit maps and isolation tests, but it does not add one listener set per session and therefore avoids listener fan-out during long-lived connections.
The bridge still exposes no protocol method to close one live session independently. Today records leave together on connection teardown; session close/resume lifecycle capabilities remain deferred in the ACP feature checklist.
## Verification
The multi-session suite drives concurrent sessions through interleaved updates, independent in-flight prompts, targeted cancellation, same-id and distinct-id load races, permission routing, config isolation, and teardown. Tool-bash tests prove one session cannot read or kill another session's background task.

139
docs/rfc/implemented/feature/2026-06-15-code-mode.md Normal file
View File

@@ -0,0 +1,139 @@
# RFC: Code Mode — the model writes TypeScript against the tool registry
Status: implemented
## Problem
In the registry's native presentation, the agent loop advertises every visible capability as a JSON-schema function definition. `ToolRegistry` contributes its schemas to the system-prompt assembly, the assembly's `tools` land on the wire (and in the logged request header), the model invokes one `tool-call` block per step, and the loop dispatches each call through `ctx.tools.execute()` **sequentially** (parallel tool execution is an explicit open TODO in `dsh-tools` and [docs/architecture.md](../../../architecture.md)), with **every** intermediate `tool-result` re-entering the model's context on the next request.
For multi-step tool work this is token-heavy and serial. The model cannot compose tools — loop over a result set, branch on an intermediate value, fan out, post-process — without a full model round-trip per call, and each round-trip drags the entire intermediate result back into context whether the model needs it or not.
Cloudflare's [Code Mode](https://blog.cloudflare.com/code-mode/) proposes an alternative grounded in a simple observation: LLMs are better at writing code than at emitting tool calls, because they have seen millions of lines of real code and comparatively few contrived tool-calling traces. Instead of one tool call per step, the model writes a TypeScript program against a generated API over the tools, the program executes in a sandboxed runtime, and the model curates what comes back — only what it prints or returns — instead of every intermediate result.
Tool presentation belongs to the registry that owns tool visibility: implementing a second presentation as an after-the-fact waterfall transform would make correctness depend on listener order and fight [reconstructable requests](../../implemented/architecture/2026-07-05-reconstructable-requests.md). The execution substrate is also part of the foundation rather than a placeholder: Node `worker_threads` provides a separate isolate, an empty environment, heap caps, and termination of a hot synchronous loop, while fitting the harness's existing trust model (§Trust posture).
## Decision
Three decisions, each elaborated in its own section below:
1. **Code Mode is a first-class presentation mode of `ToolRegistry`** (`dsh-tools`), selected by a validated `mode` config: `'native'` (the default, contributing the visible capability schemas), `'code'` (the registry contributes only its reserved `run_code` transport plus a generated SDK `.d.ts` in the system prompt), or `'both'` (native schemas and the transport + SDK). The registry shapes its canonical contribution at the source; the cooperative prompt-assembly result remains authoritative, and the logged request header records exactly that returned presentation.
2. **Code execution is a capability seam** — `packages/code-runtime/` contains the interface package `@deepseek-ai/dsh-code-runtime`, which owns `ctx.codeRuntime` ([capability seams](../../implemented/architecture/2026-06-13-capability-seams.md); consumer = `dsh-tools`, with core-consumes-a-seam precedent in `agent-loop` → `dsh-llm`). The runtime knows nothing about tools: it is handed a program and named async bindings, runs the program, and reports `{ value, logs, error? }`. Language and substrate are backend properties, so a future Python or container backend is another implementation package, not a redesign.
3. **The shipped implementation is `@deepseek-ai/dsh-code-runtime-worker`**: one fresh Node worker thread per run, executing the model's TypeScript after type-strip, with bindings bridged over the message port, an empty environment, configurable heap/output/time caps, and hard termination. Its trust posture is bash-equivalent by design — no unsafe-acknowledgement flags — because the harness already ships `dsh-bash-local`, which executes arbitrary model-written shell commands with strictly *more* ambient authority.
### The registry owns the mode
`ToolRegistry` gains a schemastery-validated config (`static Config`), its first: `mode: 'native' | 'code' | 'both'`, default `'native'`. A deployment flips it from `cordis.yml` (`tools: { mode: code }`) — no code edit, per the no-hardcoded-tunables convention.
**Wire tool list = the registry's contribution before cooperative assembly.** The registry feeds assembly through a mode-aware provider: `'native'` contributes every capability visible to that assembly scope, `'code'` contributes only `run_code`, and `'both'` contributes both. Because [`PromptAssembly.tools` is the single source the loop's request header snapshots](../../../../packages/core/system-prompt/src/index.ts), the final presentation is logged and reconstructable. The reserved transport is not a capability: it lives outside global/scoped registration and restriction layers, cannot be registered or shadowed there, and cannot be named by `ctx.tools.restrict()`. The mode governs this provider's input to assembly; other direct `systemPrompt.tools()` providers own their schemas, and the trusted assembly waterfall owns the returned wire list.
**Interaction with `toolOrder`, stated up front:** a configured `systemPrompt.toolOrder` naming native capabilities rejects every assembly under `mode: 'code'`, because those names are outside that mode's wire-validation universe. This is correct behavior, not a bug: a deployment using Code Mode updates its order config or drops it.
**The SDK prompt section.** Under `'code'` and `'both'` the registry registers one lazy prompt section (`tools:sdk`, in the 100–199 tool-guidance order band) whose thunk regenerates, for each assembly scope, a TypeScript declaration of every visible end-capability tool plus fixed usage instructions. It uses the same visibility resolver as lookup and execution, so scoped grants and shadows appear while restricted globals disappear; the reserved `run_code` transport itself is excluded. The thunk emits tools in lexicographic name order, so an unchanged visible set produces byte-identical text.
**Assembly ownership.** `run_code` and `tools:sdk` enter the trusted `system-prompt/assemble` waterfall as normal assembly inputs. A scoped `tools:sdk` section may shadow the global default before dispatch, and a listener may remove or replace either contribution. The waterfall's returned assembly is final, so whoever changes these inputs owns preserving a viable Code Mode protocol when the deployment expects Code Mode to remain usable; no restoration pass overrides deliberate composition.
**Codegen.** A pure `jsonSchemaToTs(schema)` module inside `dsh-tools` (sibling of `json-schema.ts` — `schemas()` and the SDK are two projections of the same store) maps the JSON-Schema subset the `defineTool` DSL emits (object/string/number/boolean/array, `properties`, `required`, string `enum` → literal union, nested objects, array `items`, `description` → JSDoc) to a TS type literal. It is **total**: any construct outside that subset (`$ref`, `oneOf`/`anyOf`, `integer`, future MCP shapes, …) degrades to `unknown` without throwing. Because `ToolSchema.name` is an arbitrary string, the SDK is declared as one object constant — `declare const tools: { "some-mcp-tool"(args: …): Promise<string>; bash(args: …): Promise<string>; … }` — quoted keys make every name reachable with no sanitization or alias-collision logic. Typing is advisory (the runtime executes type-stripped JS); the instructions say so.
### The run_code tool and the dispatch bridge
Under `'code'` and `'both'` the registry owns `run_code` as a reserved presentation transport with one required parameter, `{ code: string }`. It is represented by a normal `ToolDefinition` for dispatch but stays outside the filterable capability layers, so restrictions cannot accidentally remove Code Mode's only entry point. Calls traverse the complete tool pipeline — `tools/pre-execute` → monotonic guards → `tools/execute` around dispatch → `tools/post-execute` → immutable `tools/result` notification — exactly like native calls; a permission plugin can inspect the program text before it runs, and final-result observers see the normalized outer outcome. Its `execute(args, exec)`:
1. **Builds the bindings**: the bridge owns a **run-scoped `AbortController`** whose signal follows `exec.signal` (an outer cancel propagates in) and which the bridge itself fires the moment the run settles for any reason — completion, program exception, `computeMs`/`maxWallMs` expiry, worker exit. For every visible capability tool, the binding is an async function that (a) checks the run signal before and after, (b) **JSON-normalizes the argument** — a `JSON.parse(JSON.stringify(args))` round-trip, rejecting that one call with a descriptive `Error` when the value does not survive (`BigInt`, circular structures) — because the seam's structured-clone boundary is wider than JSON while the session log accepts only JSON, (c) awaits its turn on the **per-run serialization queue** (below), (d) calls `this.execute({ callId, name, arguments, agent: exec.agent, parent: exec.token, signal: runSignal })` with a deterministic sub-id `` CallId(`${exec.callId}:code:${n}`) ``, (e) appends a `tool/code-dispatch` session event, and (f) maps the result: success → the text-block contents joined as a `string` (non-text blocks become placeholders), `isError` → **the binding rejects** with an `Error` carrying the result text. The child's readonly `parent` is only the outer execution's frozen, property-free token, so commit-style observers can correlate outcomes without receiving a mutation path into the live `run_code` wrapper. Every sub-call still traverses the full pipeline under its own immutable identity and registry-assigned token. The run signal, rather than the bare outer one, lets budget expiry abort an in-flight sub-tool instead of orphaning it. Rejection gives programs ordinary `try/catch` and `Promise.all` failure semantics rather than a bespoke result envelope.
2. **Runs the program**: `ctx.codeRuntime.run({ program: args.code, bindings: [{ global: 'tools', functions }], signal: runController.signal })`. The runtime receives the run-scoped signal, not only the caller's outer signal, so any way the outer run settles also aborts work inside the runtime.
3. **Surfaces the outcome — after reaching quiescence.** When `ctx.codeRuntime.run()` settles, whether by fulfillment or rejection, the bridge fires the run-scoped abort (cancelling any in-flight sub-dispatch and abandoning queued-unstarted ones), then **awaits the dispatch queue's drain before returning or propagating**, per the dispose-to-quiescence rule in [defensive patterns](../../../defensive-patterns.md): an aborted in-flight sub-call still settles and logs its `isError` `tool/code-dispatch` event *inside* the open turn, and nothing can append after `run_code` settles. A successful result then returns one text block — the captured console/stdout output followed by the rendered return value (if any) — plus a `meta` payload (capped logs, dispatch count) for presentation. A fulfilled run with `result.error` throws a `CodeRunFailedError extends HarnessError` (`code: 'CODE_RUN_FAILED'`, message = the error kind and text plus captured logs so the model can self-correct); a backend rejection propagates through the same registry error boundary. Both become structured `isError` tool results.
**Sub-call `additionalContext` is suppressed, deliberately.** A `tools/post-execute` hook may attach `additionalContext` to a call; for loop-dispatched calls the loop buffers those and appends each as a `context/message` only after the step's `tool/result`s, preserving call/result adjacency. A sub-dispatch result's `additionalContext` has no such safe outlet from inside a running `run_code`: injecting immediately would land a `context/message` between the parent's `tool/call` and its `tool/result` (breaking the adjacency the buffering exists to protect), and `PostToolDecision.additionalContext` is singular where a program may produce many. The MVP therefore drops sub-call `additionalContext`, pinned by a test and stated in the hooks bridge's docs; the follow-up (a plural context channel or loop-level sub-dispatch buffering) is deferred until a real hook needs it through Code Mode.
**Concurrency: serialized, enforced by the binding.** The bindings are async, so a model writing `Promise.all([tools.a(…), tools.b(…)])` starts both immediately — concurrent dispatch would be the default, while the tool contract carries no concurrency-safety metadata (the open parallel-execution TODO). Each `run_code` invocation therefore owns a dispatch queue and every binding call chains onto it, so even `Promise.all` executes the underlying `ctx.tools.execute()` calls one at a time in submission order; when the run settles, queued-but-unstarted dispatches are abandoned. Lifting this per tool remains tied to tools declaring themselves concurrency-safe.
**Presentation.** `run_code`'s render intent is decided here per the [render-intent RFC](../../implemented/architecture/2026-07-02-tool-render-intent-union.md): `presentCall` → a `generic` card, `kind: 'execute'`, title = the program text, `rawInput` = the same program text; `presentResult` → a `generic` card whose content is the captured output (from `meta`). The program is the title because ACP execute cards reliably render that field while some clients omit body and raw-input content. This is not a `terminal` card: that card's semantics are "a shell command in a working directory", which a program is not.
### Observability: `tool/code-dispatch`
Each sub-dispatch appends one session event, declared by `dsh-tools` via `SessionEventMap` declaration merging (the map is merge-extensible for exactly this; `todo/write` is the log-only precedent): `tool/code-dispatch` with `{ parentCallId, subCallId, name, arguments, isError, resultSummary }` — `arguments` being the bridge's JSON-normalized value, the very one dispatched, so the append cannot fail on payload shape. It is log-only — `deriveEventMessage()` ignores unknown event types by design, so sub-calls never re-enter model context — but persistence and UIs get every call. As a log event it carries JSDoc prose but **no `@mode` tag** (that vocabulary belongs to cordis bus events; the persistence-catalog generator hard-errors on one) and lands in the regenerated `docs/persistence-catalog.md`; appends happen inside `run_code`'s execution, so the turn-enclosure invariant is satisfied by construction. A `run_code` execution arriving without `exec.agent` (the loop always supplies it; direct programmatic calls may not) still runs and simply skips event logging, exactly as the `ToolExecution` contract allows.
### The code-runtime seam
`packages/code-runtime/code-runtime/` — `@deepseek-ai/dsh-code-runtime`, depending only on `cordis`. An abstract `CodeRuntime extends Service` (`super(ctx, 'codeRuntime')`) plus the vocabulary:
- `CodeRunRequest = { program: string; bindings: CodeBindingNamespace[]; signal?: AbortSignal }`
- `CodeBindingNamespace = { global: string; functions: Record<string, (args: unknown) => Promise<unknown>> }` — the runtime exposes each namespace as a global object of async functions inside the program; binding arguments and resolutions must be structured-cloneable (a runtime may cross a serialization boundary; ours does).
- `CodeRunResult = { value?: unknown; logs: CodeLogEntry[]; error?: CodeRunFailure }` — program execution outcomes, including exception, timeout, abort, and worker exit, resolve as the `error` field. `run()` may reject only for caller/seam misuse (for example a duplicate binding namespace); consumers still contain a non-conforming backend rejection at their own error boundary.
- `CodeLogEntry = { source: 'console' | 'stdout' | 'stderr'; level?: 'log' | 'info' | 'warn' | 'error' | 'debug'; text: string }`
- `CodeRunFailure = { kind: 'exception' | 'timeout' | 'abort' | 'worker-exit'; message: string }` — orthogonal outcomes reported independently per [defensive patterns](../../../defensive-patterns.md); a timed-out run is not an exception, an abort is not a timeout.
- Two readonly backend descriptors, informational not gating: `language` (what the program must be written in — `'typescript'` for the shipped backend; a Python backend would say so, and pair with its own SDK generator on the presentation side) and `isolation` (`'worker-thread'` for the shipped backend; `'process'`, `'container'`, … for future ones). `dsh-tools` requires `language === 'typescript'` in the MVP — its codegen emits TS — and fails the assembly loudly otherwise, the same misconfiguration idiom as `toolOrder` violations (as when `mode` is non-native with no `ctx.codeRuntime` loaded at all).
Per explicit-over-implicit at seams, the request spells out everything the runtime acts on; defaulting (timeouts, caps) is the implementation's validated config, never a hidden `??` inside `run()`. Consumption uses the loop's optional-backend idiom: Cordis has no optional injection — every `inject` entry gates activation — so a static `inject` on the registry would hold `ctx.tools` (and every tool plugin behind it) hostage to a code runtime existing even under `mode: 'native'`; instead the registry reads `ctx.get('codeRuntime')` at use time, exactly as `agent-loop` consumes `sessionPersistence`, with absence failing loud in the provider thunk. The seam has concrete divergence on both axes: the worker-thread substrate can be replaced by a container or microVM implementation, and the TypeScript language contract can be paired with a language-specific SDK and runtime. `dsh-tools` consumes only the interface and tests against a trivial in-repo fake, exactly the interface/implementation/consumer shape of the bash template.
### The worker-thread runtime
`@deepseek-ai/dsh-code-runtime-worker`, the second package of the `packages/code-runtime/` group. Per `run()`:
1. **Type-strip host-side** with Node's built-in `stripTypeScriptTypes` (`node:module`; present across the repo's whole engines range, `^22.19.0 || >=24.0.0`, and position-preserving, so runtime error line numbers match the model's source). Strip-only mode rejects non-erasable syntax (`enum`, namespaces) — that rejection returns as `error.kind: 'exception'` with Node's message, the SDK instructions say "erasable TypeScript only", and the model self-corrects like any other program error. A syntax-level failure never spawns a worker.
2. **Spawn one fresh `Worker` per run** from the package's own bootstrap module: `env: {}` (truly empty — stronger than the scrubbed-env rule for spawned commands), `resourceLimits` from config, `stdout`/`stderr` captured into `logs` rather than inherited. No pooling and no cross-run state: a program's world dies with its worker, which keeps runs reconstructable from the log alone and makes state bleed unrepresentable.
3. **Execute** in the bootstrap: the stripped program becomes the body of an `AsyncFunction` whose parameters are the binding globals and a capturing `console` shim, so top-level `await` and `return` work and the program's completion value is the run's `value` (structured-cloneable values cross as-is; anything else is replaced by its `util.inspect` rendering, documented).
4. **Bridge bindings over the message port**: each binding function in the worker posts `{ id, global, name, args }` and awaits the reply; the host validates the name against the request's bindings, invokes, and replies `{ id, ok, value }` or `{ id, ok: false, message }` (a host-side binding rejection becomes a program-side rejection). The worker-side namespace objects are built null-prototype via `defineProperty`, so a binding named `__proto__`, `constructor`, or `toString` is an ordinary own property, not a prototype collision. Unknown names, duplicate ids, and post-settlement messages are rejected or ignored — the port protocol assumes a hostile peer, because the peer runs model code.
5. **Enforce caps — two independent budgets, because the peer is hostile.** The compute budget (`computeMs`) meters the worker's *measured busy time* via `worker.performance.eventLoopUtilization()` polling — not host-side "is an RPC pending" bookkeeping, which a program defeats by firing an un-awaited call at a slow tool and then spinning hot while the host thinks it is waiting. Measured busy time cannot be gamed: a hot loop accrues it whether or not a dispatch is in flight, and a program genuinely awaiting a slow tool accrues none, so a long-running `bash` sub-call still does not kill an innocent run. The wall ceiling (`maxWallMs`) never pauses for anything and backstops what busy-time cannot see (a program awaiting a promise nobody will resolve). Budget expiry, `signal` abort, and run completion all funnel into `worker.terminate()`, which ends hot synchronous loops too (measured; this was `node:vm`'s unfixable gap); the failure reports which budget fired. Heap overflow surfaces as the worker's OOM exit → `error.kind: 'worker-exit'`. Log and value sizes are capped by config, truncation marked in-band. All caps are validated config fields with defaults (`computeMs: 60_000`, `maxWallMs: 600_000`, `maxLogBytes: 65_536`, `maxValueBytes: 32_768`, `maxOldGenerationSizeMb: 512`), changeable from `cordis.yml`.
6. **Dispose to quiescence**: the service's own disposal terminates in-flight workers and *awaits* their exits before resolving, per [defensive patterns](../../../defensive-patterns.md).
### Trust posture
The worker runtime is **containment, not a security boundary**. Model code in the worker can reach Node globals — `fetch`, `process` (with an empty env), dynamic `import()` of built-ins — so a deliberately adversarial program has ambient authority comparable to what the harness's own `bash` tool grants every model turn: `dsh-bash-local` runs arbitrary model-written commands with the host filesystem, network, and a scrubbed-but-populated environment. One asymmetry runs the other way: `worker.terminate()` ends the thread, not OS processes a program may have spawned via `node:child_process` — weaker than `bash-local`'s process-group kill for direct children (equivalent for double-forked daemons, which survive both); the wall-clock ceiling bounds the worker itself, and orphan cleanup is the same deployment-level concern it is for bash. Code Mode is gated where bash is gated — `tools/pre-execute`, where permission/sandbox plugins veto or approve the program — and adds containment bash does not have: empty env, heap caps, hard termination of the program itself, and a separate isolate. A `node:vm` executor with no containment would need explicit unsafe acknowledgement; imposing that ceremony on the better-contained worker while bash needs none would be posture theater. A deployment that needs a hard boundary (untrusted multi-tenant input) needs it for bash too; that is a future `isolation: 'container'` backend, and the `isolation` descriptor lets deployments distinguish backends.
### What the model sees
The `tools:sdk` section carries the `.d.ts` plus fixed instructions: the program is the body of an async TypeScript function (erasable syntax only — no `enum`/namespaces; type annotations are advisory); call tools as `await tools.name(args)` (quoted access for exotic names); a failed tool call **rejects** with an `Error` carrying the tool's error text — catch it to handle and continue; calls run **sequentially** even under `Promise.all`; emit results via `return` and/or `console.log`, and only that curated output returns to the context — intermediate tool results never do. That last line is the payoff the whole design serves: output-side context cost becomes the model's own editorial decision. On the input side the `.d.ts` is not free — for a large tool surface it can rival the native JSON schemas it replaces (and `'both'` pays for the two side by side) — but it is prefix-stable, so provider prefix caching amortizes it; the win is workload-dependent and the RFC claims no more.
## Consequences
The design consists of the `dsh-code-runtime` interface package, the `dsh-code-runtime-worker` backend, and the `dsh-tools` presentation and dispatch integration.
Shipped surface:
- **The seam**: `packages/code-runtime/` — `@deepseek-ai/dsh-code-runtime` (abstract `CodeRuntime`, the vocabulary above, `ctx.codeRuntime`) and `@deepseek-ai/dsh-code-runtime-worker` (the worker-thread backend, every cap a validated config field). Rows in the service map, capability-seams graph, config catalog, and cordis catalog.
- **The registry surface**: `ToolRegistry`'s `mode` config, mode-aware wire contribution, lazy `tools:sdk` section and reserved `run_code` transport, `jsonSchemaToTs`/`renderToolsSdk` (exported), the dispatch bridge and `CodeRunFailedError`, and the `tool/code-dispatch` log event (declaration-merged into `SessionEventMap`, regenerated into the persistence catalog; `run_code` in the tool catalog).
- **The composed surface**: the `tools` config forwards through `agent-core` and both app packages (`stdio-agent`, `acp-agent`); `demo:code-mode` boots each UI example's `code-mode.cordis.yml` overlay (the worker runtime + `mode: 'code'` over the base tree); every program sub-dispatch resolves the same scoped capability view and re-enters the complete tool pipeline with an immutable link to its enclosing transport execution.
- **Interactions inherited by deployments**: a `toolOrder` naming native tools rejects every assembly under `'code'` (update or drop the order config when switching modes); restrictions can hide end capabilities but cannot remove the registry-owned presentation transport, while assembly listeners may rewrite the final model-visible surface and own its protocol integrity; sub-call `additionalContext` is dropped by the bridge (a plural context channel is deferred until a real hook needs it through Code Mode); sub-dispatch stays serialized until tools can declare concurrency safety — the same metadata the native parallel-dispatch TODO waits on.
## Testing
What the suites pin, per tier:
- **Unit — worker runtime** (real workers, no mocks): output/value capture and log-source attribution; error kinds (exception incl. non-erasable syntax, abort, worker-exit under OOM); the two budgets from both sides (a hot loop behind an un-awaited pending dispatch dies at `computeMs` busy time; a program idling on a slow binding outlives `computeMs` and dies only at `maxWallMs`); binding-bridge hostility (junk/forged port traffic incl. non-object messages and forged `log`/`done` cap bypass attempts, unknown names, duplicate ids, post-settlement replies, `__proto__`/`constructor`/`toString` binding names); structured-clone fallback and cap truncation; `env` emptiness verified from inside a program; dispose-awaits-exit. A real-load-path e2e runs the BUILT package under plain `node` so the worker entry resolves both unbuilt (tsx) and built — the published-artifact guard from [docs/testing.md](../../../testing.md).
- **Unit — registry integration**: the codegen table (DSL subset, quoted names, `unknown` degradation, byte-identical determinism); provider contribution per mode (`'native'` capabilities, `'code'` exactly `[run_code]`, `'both'` capabilities + `run_code`); reserved-name, restriction, scoped shadowing, authoritative assembly transformation, and `toolOrder × mode` invariants; missing-runtime / wrong-language loud failures; full-pipeline and opaque parent-token behavior for sub-dispatches; serialization non-overlap (a probe tool records enter/exit under `Promise.all`); abort aborting the in-flight sub-dispatch and abandoning queued ones; binding rejection on `isError` and on JSON-unrepresentable arguments; `CodeRunFailedError` → structured `isError` carrying kind + logs; `tool/code-dispatch` payloads (JSON-normalized arguments identical to what dispatched); `deriveMessages()` ignoring the event; sub-call `additionalContext` suppression; HMR safety.
- **e2e (with-key, self-skips)**: a real model under `mode: 'code'` composes two bash calls in one program (`examples/coding-agent/tests/code-mode.e2e.ts`) — every logged `request/header` carries exactly `[run_code]`, the dispatch events land under the parent call, the file the program wrote exists, and the final answer is the curated output.
- **Snapshot (keyless replay)**: goldens for a `run_code` turn under `'code'` and `'both'` (`code-mode-turn`, `both-mode-turn`), each its own header-pinning class — the SDK section text, the collapsed header tool list, the dispatch events, and the result card are committed and replayed.
## Alternatives considered
**An add-on consumer plugin with zero core changes.** Rejected because `agent/request` is call-config-only under [reconstructable requests](../../implemented/architecture/2026-07-05-reconstructable-requests.md), while transforming an assembled tool list would have to undo `toolOrder` canonicalization without owning its config and would depend on listener order. Which tools the model is offered, and in which representation, is the registry's single concern: native schemas and the SDK are two projections of one visible store.
**`node:vm` as the reference runtime, with hardening deferred.** Rejected: `node:vm` is not isolation (prototype-chain escapes reach the host realm) and cannot interrupt a hot loop. A worker thread provides a separate isolate, empty environment, `resourceLimits`, and reliable `terminate()` at bash-equivalent trust, so the reference and production implementation are one package without an unsafe-acknowledgement ceremony.
**Result elision / summarization over native tool-calling.** Addresses only the context-bloat half of the problem: trimming old `tool-result`s is cheap to add as a logged surface replacement under reconstructable requests, but still pays one model round-trip per call and cannot express loops, branches, or joins. Complementary, not competing; it can layer under Code Mode for residual native calls.
**Parallel native dispatch in the loop.** The other answer to round-trip cost; still valid future work (the open TODO), still blocked on concurrency-safety metadata, and still no composition — it parallelizes calls the model already decided on in one step. Code Mode's serialized-queue decision keeps the two compatible: when the metadata lands, both native parallel dispatch and per-tool binding parallelism unlock together.
**Always-exclusive (Cloudflare-faithful, no mode).** Rejected for this SDK's primary consumer: a coding agent's bread-and-butter single calls (`bash`, `read`, `edit`) are already ideal as native calls, and forcing every edit through a program taxes the common case. The mode config keeps the faithful form (`'code'`) one line away without imposing it.
**Per-tool visibility tiers (this tool native, that tool code-only).** Deferred: it needs per-tool metadata and a presentation split that `'native' | 'code' | 'both'` does not, and its design depends on evidence about how models split usage under `'both'`.
**Sanitized identifier aliases in the SDK** (`my-tool` → `my_tool`, Cloudflare's approach). Rejected: quoted keys on a `declare const` make every name reachable with zero alias-collision logic; models handle `tools["my-tool"](…)` fine.
**A REPL-style persistent kernel** (state survives across `run_code` calls). Rejected for the MVP: cross-call state would be invisible to the session log, breaking the reconstructability guarantee that every request is a pure function of the log; fresh-per-run keeps it. A kernel-style backend remains expressible behind the seam later, with its own logging story.
## Risks
**The worker is not a hard security boundary.** Deliberate and documented (§Trust posture): posture equals the existing bash tool, containment exceeds it, gating uses the same seams. Deployments needing more need a future `isolation: 'container'` backend — tracked as the seam's designed extension, not a TODO on this design.
**`stripTypeScriptTypes` is marked experimental.** It is the same engine (amaro/swc) behind Node's own native `.ts` execution, exposed as an API across this repo's whole engines range. Mitigations: the runtime's unit suite pins the behaviors relied on (position preservation, erasable-only rejection message shape loosely), the call sits behind one private function, and `amaro`/`sucrase` are drop-in replacements if the API shifts. The erasable-only subset is a model-facing contract line, and the error path is a working feedback loop, not a dead end.
**Prompt cost of the SDK, especially under `'both'`.** The `.d.ts` can rival the native schemas it complements; `'both'` carries two representations. Prefix stability + provider caching amortize per-session cost; the mode is per-deployment; the RFC makes no unconditional-savings claim. Measured guidance (when to prefer which mode) is explicitly post-ship learning.
**Registry scope growth.** `dsh-tools` absorbs codegen, a tool, a bridge, and an event. Contained by module boundaries inside the package (`ts-types.ts`, `code-mode.ts` beside `schema.ts`/`json-schema.ts`/`presentation.ts`) and by the seam: everything substrate-shaped lives behind `ctx.codeRuntime`.
**Structured-clone limits at the binding boundary.** The seam's clone boundary admits values JSON does not (`Date`, `Map`, `BigInt`), and the session log accepts only JSON — left unhandled, a sub-call could execute and then fail at `tool/code-dispatch` append time. Closed by the bridge's JSON-normalization step (§ the dispatch bridge): what does not survive the round-trip rejects that binding call before dispatch, so every executed sub-call is loggable by construction. The seam itself keeps the wider structured-clone contract (it is about the port, and stated so a future binding producer cannot discover it in production); consumers with stricter payload needs enforce them at their own boundary, as the bridge does. Non-text sub-result content is reduced to placeholders — a known MVP limitation, recorded in the SDK instructions.
**Serialized-only sub-dispatch.** `Promise.all` gains no wall-clock parallelism yet, only fewer round-trips; models may over-expect. The instructions state it; lifting it is tied to the same concurrency-safety metadata the native parallel-dispatch TODO needs.
**Budget metering reads the event loop, not a flag.** Busy-time polling (`eventLoopUtilization()`) is coarser than an exact CPU meter — a budget expires up to one poll interval late — and its correctness claim ("a pending dispatch cannot pause it") is load-bearing against a hostile program. Both sides are unit-tested (hot loop with a pending decoy dispatch dies at `computeMs`; idle-on-slow-binding survives to `maxWallMs`), and the poll interval is an internal constant, not config — nothing a deployment could mis-tune into a bypass.

2
docs/rfc/implemented/feature/2026-06-18-acp-terminal-and-tool-rendering.md
View File

@@ -4,7 +4,7 @@ Status: implemented
## Problem
The ACP bridge lets each tool own its call rendering via `presentCall`/`presentResult` (see [tool-call UI presentation](../../proposed/feature/2026-06-14-acp-agent-client-protocol.md) and `packages/core/tools`). For `bash` we surface the exact command as the `tool_call` title, the model's `description` as a content text block, `kind: 'execute'`, and the completed output wrapped in a fenced ` ```console ` text block.
The ACP bridge lets each tool own its call rendering via `presentCall`/`presentResult` (see [tool-call UI presentation](../../implemented/feature/2026-06-14-acp-agent-client-protocol.md) and `packages/core/tools`). For `bash` we surface the exact command as the `tool_call` title, the model's `description` as a content text block, `kind: 'execute'`, and the completed output wrapped in a fenced ` ```console ` text block.
That is a correct, capability-free baseline, but not how the reference editors render a *terminal* tool at its best. An editor like Zed has a dedicated terminal tool-call card — a header showing the working directory, the command as the label, the command output rendered as a terminal, and an exit-status pill — but it only builds that card when the `tool_call` carries terminal metadata (below). With a plain text block the output appears as static markdown and there is no cwd header. (Zed also HIDES `rawInput` for `kind: 'execute'`, which is why the command IS the title — both reference adapters do the same. The human-readable description rides as a separate content block above the card; note this is a DELIBERATE divergence — claude-agent-acp DROPS the description in terminal mode and renders only the card — we keep the summary visible alongside.)

12
docs/rfc/implemented/feature/2026-06-21-subagent-capability-seam.md
View File

@@ -35,18 +35,18 @@ A new package group `packages/subagent/`:
| `@deepseek-ai/dsh-subagent-mock` | support: a scripted provider for testing the seam through the real load path |
| `@deepseek-ai/dsh-tool-subagent` | consumer: the model-facing `subagent` tool over `ctx.subagents` |
### The primitive: `start → SubagentRun`
### The primitive: async `start → SubagentRun`
A provider exposes `start(request) → SubagentRun`. The run carries a `result` promise (the terminal `SubagentResult`), `cancel()`, and `dispose()`. The transport-neutral verb is **`start`**; "spawn" is reserved for the in-process `dsh-subagent-spawn` backend's identity, not the service verb. The service's `start(name, request)` resolves the named provider, validates capabilities, delegates, and emits `subagent/start` / `subagent/end` around the run.
A provider exposes `start(request) → Promise<SubagentRun>`. Promise fulfillment is the publication/readiness and provider-to-caller ownership boundary: for an in-process backend the child is already published in `ctx.agents`, and for ACP the remote session already exists. `SubagentStartRequest.signal` is the single cancellation channel before and after readiness; `SubagentRun` carries the terminal `result` and a `dispose()` method that cancels remaining work and awaits quiescence. The transport-neutral verb is **`start`**; "spawn" is reserved for the in-process `dsh-subagent-spawn` backend's identity, not the service verb. A rejected start cleans provider-owned partial resources and emits neither subagent lifecycle event.
### Two kinds of optional capability, discovered two ways
- **Start-time features** (`outputSchema`, `depthLimit`, `toolFilter`) ride on a static `provider.capabilities` descriptor. The service checks every requested one BEFORE delegating and **rejects loud** (`SubagentError('UNSUPPORTED_CAPABILITY')`) if the provider lacks it — never accepted-then-ignored. They must be checked before a run exists, which is why they cannot be runtime methods.
- **Start-time features** (`outputSchema`, `depthLimit`, `toolFilter`, `persona`) ride on a static `provider.capabilities` descriptor. The service checks every requested one BEFORE delegating and **rejects loud** (`SubagentError('UNSUPPORTED_CAPABILITY')`) if the provider lacks it — never accepted-then-ignored. They must be checked before a run exists, which is why they cannot be runtime methods.
- **Runtime features** (steering via `sendMessage`, follow-up via `resume`) are **optional methods** on `SubagentRun`. The method's presence IS the capability, and TypeScript narrowing is the discovery mechanism: a consumer cannot call an absent method without narrowing first, so there is no silent-degradation path and no separate flags object to keep in sync.
### Fork vs. fresh are separate backends, not a flag
Rather than a `context: 'fresh' | 'fork'` request field, the distinction is the provider's identity: `dsh-subagent-spawn` (fresh, isolated, own system prompt) and `dsh-subagent-fork` (seeded from the parent's log) are two registered providers. You pick behavior by picking a provider — consistent with the registry being the selection mechanism. The fork backend seeds only a **balanced, completed-turn prefix** of the parent log: at tool-execute time the parent's turn is open (it holds the `assistant/message` and the dangling spawn `tool/call` with no `tool/result`), and seeding that raw prefix would give the child an unbalanced turn the [invariants](../../../../packages/support/invariants/src/index.ts) freeze-check rejects.
Rather than a `context: 'fresh' | 'fork'` request field, the distinction is the provider's identity: `dsh-subagent-spawn` (fresh, isolated, own system prompt) and `dsh-subagent-fork` (seeded from the parent's log) are two registered providers. You pick behavior by picking a provider — consistent with the registry being the selection mechanism. The fork backend seeds only a **balanced, completed-turn prefix** of the parent log: at tool-execute time the parent's turn is open (it holds the `assistant/message` and the dangling spawn `tool/call` with no `tool/result`), and seeding that raw prefix would give the child an unbalanced turn that the [invariants](../../../../packages/support/invariants/src/index.ts) trace replay rejects.
### Child isolation and the parent log
@@ -54,7 +54,7 @@ Each subagent runs in its **own `Session`** (own id, `parentSession` lineage), p
### Synchronous collect (first cut)
The `dsh-tool-subagent` consumer awaits `run.result` and returns the child's final output as the tool result, blocking the parent's turn until the child finishes. It does so inside a `try/finally` that always `dispose()`s the run (no leaked idle child/session on any path), bridges `exec.signal` to `run.cancel()`, and maps a non-`completed` stop reason to an `isError` result rather than returning partial output as success. Steering (`sendMessage`) is part of the contract but **intentionally unused** this cut.
The `dsh-tool-subagent` consumer passes its execution signal into the start request, awaits the ready run's `result`, and returns the child's final output as the tool result, blocking the parent's turn until the child finishes. A `try/finally` always `dispose()`s the run, so no success, failure, or cancellation path leaks an idle child/session. A non-`completed` stop reason maps to an `isError` result rather than returning partial output as success. Steering (`sendMessage`) is part of the contract but intentionally unused in this consumer.
### Provider selection is config, not model-facing
@@ -66,7 +66,7 @@ The seam is tested through the real cordis Loader / export path, not a hand-buil
## Consequences
- **Recursion.** Without a guard, an in-process child inherits the spawn tool and can spawn unboundedly. Depth-limit is an optional capability (the in-process backends enforce it; ACP advertises it off and rejects a `maxDepth` request); tool-filtering is likewise optional. Tool-filtering, when implemented, needs a `tools/pre-execute` deny in the child context — schema filtering alone is insufficient because a model can hallucinate a denied tool name.
- **Recursion.** Without a bound, an in-process child can see the delegation tool and recurse. The in-process backends implement the optional absolute depth limit and scoped live-global `toolFilter`; ACP advertises both capabilities off and rejects such a request. The [subagent composition-controls RFC](2026-07-12-subagent-persona-tool-filter-and-depth.md) owns their exact semantics and security limits.
- **Blocking the parent turn.** Synchronous collect holds the parent's `runStep` open for the child's full duration. This is acceptable for the first cut; **background / poll / spill semantics are deferred to a future redesign that unifies long-running-tool handling across subagents AND bash** (a sub-agent and a long `bash` background task pose the same "the model started something slow, how does it collect later" problem, and should share one mechanism rather than each inventing its own).
- **Live progress.** This cut surfaces only lifecycle + final result; a per-chunk child→parent update stream is deferred with the background redesign.
- **ACP client surface.** Proxying `fs`/`terminal` from the ACP child back to the parent (a shared-workspace mode) is future work; the first cut advertises neither, so the child self-serves in its own process.

2
docs/rfc/implemented/feature/2026-06-22-acp-subagent-backend.md
View File

@@ -34,7 +34,7 @@ The child is a separate process, so it inherits an environment. Credential-shape
Designed at every tier the backend touches, per the root AGENTS.md rule that a new capability shape names its coverage at every tier at plan time:
- **Keyless unit/integration** (`subagent-acp.spec.ts`): spawns a scripted mock ACP server subprocess (`tests/mock-acp-server.ts`) and drives it through the real backend over real ACP stdio. Covers: the prompt round-trip + output accumulation; every StopReason mapping; cancellation via `run.cancel()` and via the request signal; the already-aborted-before-start case; the cancel-races-ahead-of-newSession case; a torn-pipe-after-cancel (child crashes on cancel) settling `aborted`; permission auto-answer under both policies (including the allow-policy-no-allow-option fallback); a non-message update consumed but not accumulated; a nonexistent-command spawn failure settling `error`; HMR provider cleanup; and the namespace export shape. 100% per-file coverage.
- **Keyless unit/integration** (`subagent-acp.spec.ts`): spawns a scripted mock ACP server subprocess (`tests/mock-acp-server.ts`) and drives it through the real backend over real ACP stdio. Coverage includes the prompt round-trip and output accumulation; every StopReason mapping; cancellation through the required request signal and through disposal; already-aborted and cancel-races-ahead-of-newSession starts; a torn pipe after cancellation settling `aborted`; permission auto-answer under both policies; non-message updates; nonexistent-command startup failure with process reaping; provider HMR; and the namespace export shape.
- **With-key e2e** (`subagent-acp.e2e.ts`): the harness drives ITSELF — the backend spawns the real `acp-agent` example process and a real model in that child answers a prompt (PONG) and does real file work (writes `proof.txt`, verified on disk). Self-skips without `DEEPSEEK_API_KEY`. This is the "talk to our own process" smoke and the out-of-process analogue of the in-process spawn e2e.
- **Snapshot**: deferred as `TODO(acp-subagent-replay)`. An ACP child is a distinct replay shape — each child is its own PROCESS with its own single-agent replay (booted under `DSH_SNAPSHOT=replay` with its own sessions-root + fixture), unlike the in-process per-session keying that [the per-session replay RFC](../testing/2026-06-22-subagent-snapshot-replay.md) added. The keyless mock-server tests give deterministic coverage of the backend in the meantime; the snapshot follow-up would record the parent driving a real-but-replayed ACP child.

49
docs/rfc/implemented/feature/2026-06-25-ask-user-question.md Normal file
View File

@@ -0,0 +1,49 @@
# RFC: Ask-user question capability
Status: implemented
## Problem
The agent sometimes cannot proceed safely from model inference alone: it needs the human to choose a path, confirm a risky/default action, or provide missing information. Before this change, the only way to get that answer was for the model to ask in assistant text and then stop, which broke the normal tool-call loop: the agent had no structured way to pause, no option metadata for UIs, no abort/error taxonomy, and no way for non-stdio front doors to present the question consistently.
This is a user-facing capability, but it also crosses package boundaries. A model-facing tool needs a provider-neutral request vocabulary; each UI surface needs to decide how to show and collect the answer; the agent loop should remain unchanged because a tool call already has the right async shape.
## Decision
Introduce `dsh-user-interaction` as the provider-neutral interface package for `ctx.userInteraction`, colocated with the model-facing consumer `dsh-tool-ask-user` under `packages/ui`. The grouping is intentional: asking a human is a UI-backed product affordance, not part of the providerless core spine. The seam still owns the stable request/answer/error vocabulary, while UI product surfaces provide the concrete provider that collects the answer. The tool registers `ask_user_question`, forwards `{ questions, agent, signal }`, and returns the provider-computed structured answers as the tool result.
The model-facing request vocabulary is deliberately aligned with the product-research schema: `ask_user_question({ questions: [{ id, question, header?, options?: [{ label, description? }], multi_select? }] })`. `id` is supplied per question and echoed in the result so a batch can be routed without relying on question text. `label` is both user-facing display text and the selected value returned to the model; there is no separate `value`, no `recommended`, no `allow_custom`, and no `desc` alias.
Providers return `{ answers: [{ id, selected, custom? }] }`. `selected` is always an array of selected option labels, so single-select and `multi_select` answers share one result shape. `custom` carries a free-text "Other" answer; optionless questions collect `custom` directly. When `custom` is present, it overrides any selected choices and `selected` is empty.
`UserInteractionError` extends `HarnessError`, so failures such as `NO_PROVIDER`, `ASK_ABORTED`, ACP cancellation, or missing session routing survive `ctx.tools.execute()` as machine-routable `{ name, code }` tool errors. This matches the structured-error taxonomy and lets the model or a wrapping plugin distinguish "user cancelled" from a generic thrown exception.
## UI mappings
`dsh-stdio-agent`'s in-package readline module renders each question, shows each option's `description` on the next line, supports comma/space-separated numeric choices for `multi_select`, accepts free-form custom answers, and rejects pending questions on abort, provider disposal, or stdin EOF. A batched request is asked in order and resolved as one answer object. The stdio provider serializes simultaneous requests with an internal queue so only one prompt owns stdin at a time.
`dsh-acp` provides the same seam for ACP sessions. It routes an ask request from the calling `Agent` through the bridge's `agent→sessionId` reverse map and calls ACP `unstable_createElicitation` with a session-scoped form for each question. Single-select options become a `choice` string enum; `multi_select` options become a `choice` array enum; optionless questions use a required `custom` text field. If the client returns both `choice` and non-empty `custom`, the custom answer wins. ACP `decline`/`cancel`, a missing answer, a missing session, and a client without elicitation support all become structured `UserInteractionError`s.
The ACP mapping deliberately uses elicitation, not `session/request_permission`. `request_permission` is still reserved for the separate permission gate: it is a yes/no-or-policy authorization protocol around tool execution. `ask_user_question` is a general information-gathering tool with optional free-form answers, so ACP form elicitation is the closer protocol fit. The bridge's session routing is shared with the future permission gate, but the user intent is different.
## Alternatives considered
**Assistant text followed by a stopped turn.** The model could ask the user in plain assistant text and then stop. That loses the structured option metadata, gives UIs no provider-neutral way to render a choice, and forces the next human answer to arrive as a new user prompt rather than as the result of the operation that needed the answer.
**Core-owned ask-user packages.** The first implementation split the seam and the model-facing tool across `packages/core` and `packages/ui`, but both names describe one UI-backed human-interaction affordance. The seam remains provider-neutral, but it is not providerless core infrastructure like sessions, tools, or the agent registry. Keeping `dsh-user-interaction` and `dsh-tool-ask-user` together under `packages/ui` makes the package map match the product boundary: apps and bridges provide the human-answer provider, and the stdio app opts into the model-facing tool.
**ACP `session/request_permission`.** Permission requests are authorization around tool execution; `ask_user_question` is information gathering with optional free-form answers. Using permission for general questions would collapse two different product concepts and make the future permission gate harder to reason about.
**A loop-level pause primitive.** The agent loop already knows how to await a tool call and resume from a tool result. Adding a new loop special case would duplicate that async shape and make every loop implementation learn about a UI concern.
## Consequences
ACP elicitation is currently marked unstable in the SDK. The fallback is still structured: if a client does not implement it, the tool returns `ASK_FAILED` rather than hanging. A later ACP stabilization may rename or reshape the method; that migration should stay inside `dsh-acp` because the core `ctx.userInteraction` vocabulary is provider-neutral.
The feature gives the model a powerful pause primitive, so prompt guidance matters. The tool description tells the model to ask concise questions and use options when possible. Product policy can later wrap `tools/execute` to restrict when the tool is allowed, but the loop should not special-case it.
`dsh-user-interaction` and `dsh-tool-ask-user` both live in `packages/ui` because they form one product-facing human-interaction capability. `agent-core` does not load either the tool or a provider. `stdio-agent` opts into the seam, its readline provider, and the model-facing tool. `acp-agent` keeps only the `userInteraction` seam/provider by default: ACP elicitation support is still client-dependent, so an ACP leaf must opt into the model-facing tool deliberately once its client can complete elicitation requests.
## Testing
Unit coverage pins provider registration/disposal, duplicate-provider rejection, abort-before-provider, empty-question rejection, structured tool errors through `ctx.tools.execute()`, batched answers, multi-select answers, custom answers, and the model schema including the removal of `value`, `recommended`, `allow_custom`, and `desc`. `dsh-stdio-agent` tests cover option descriptions, queued requests, EOF/abort cleanup, optionless free-form input, invalid option reprompts, duplicate multi-select numbers, and batched question flows. ACP bridge tests drive a real in-memory ACP connection with the real `ask_user_question` tool and verify selected-option, custom-overrides-choice, multi-select, and optionless free-form elicitation paths continue the agent loop.

9
docs/rfc/implemented/feature/2026-06-30-hook-bridges.md
View File

@@ -26,9 +26,11 @@ Each bridge maps the neutral `MergedHookOutcome` from the shared lib onto the se
| `tools/pre-execute` | `deny`→`deny`; `ask`→`ask` | `block`→`deny` (no allow/ask) |
| `tools/post-execute` | `deny`→`block`+feedback; context-only→delegate+fold | same |
| `agent/turn-continuation` | blocking Stop → `continue` (reason = next-step steering) | same |
| `subagent/start` (emit) | additionalContext → inject into the live child | — (not a Codex event) |
| `subagent/start` (emit) | additionalContext → inject into a live in-process child; a remote child has no local injection target | — (not a Codex event) |
| `subagent/end` (emit) | observe-only | — |
The CC bridge's `ask` result is a real permission path, not a terminal bridge decision: `dsh-tools` resolves it through the optional [approval seam](2026-07-06-approval-seam.md). A composed ACP answerer prompts the owning editor session and `allowed-once` proceeds; without an ApprovalService or answerer, the call fails closed to `deny`.
### Context source is always the plugin (the mislabel guard)
`agent.inject()` defaults a missing `MessageSource` to `{ kind: 'user' }` — which would record plugin-injected context as if the user had typed it. So every bridge `inject()` and every `HookContext` passes an explicit `{ kind: 'plugin', plugin: 'hooks-claude' | 'hooks-codex' }` source. A test asserts the resulting `context/message.source` is the plugin, never `user`.
@@ -52,11 +54,10 @@ Two different cwds, kept distinct on purpose. The hooks **themselves** run in th
## Deferred (faithful-but-degraded)
- **Tool-input rewrite.** A CC/Codex `updatedInput` is logged + warned, not honored — input rewrite is a deferred consistency-design problem ([the pre-tool-input-rewrite RFC](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md)), because the pre-execution args are read by `tool/call` audit + `assistant/message` history + ACP/tool-bash presentation, so an honest rewrite is a design unit, not a field.
- **Stop loop-guard** (`TODO(stop-loop-guard)`). CC/Codex break an infinite force-continue with `stop_hook_active` (true once a Stop hook fired this run) plus a max-consecutive cap; both are deferred. Today `stop_hook_active` is always `false`, so a Stop hook that unconditionally blocks would force-continue every step — a hook author must self-limit until the guard lands.
- **Permission `ask`** degrades to `deny` at the `tools/pre-execute` seam (`FIXME(permissions)` in the interception-seams RFC) — there is no interactive permission prompt yet.
- **Stop loop-guard** (`TODO(stop-loop-guard)`). CC/Codex break an infinite force-continue with `stop_hook_active` (true once a Stop hook fired this run) plus a max-consecutive cap; both are deferred. `stop_hook_active` is always `false`, so a Stop hook that unconditionally blocks would force-continue every step — a hook author must self-limit until the guard lands.
- **Hook `continue:false` (hard halt).** A hook can ask to halt the whole run (CC/Codex `continue:false`); the shared merge folds it into `MergedHookOutcome.stop`/`stopReason`, but no bridge acts on it (`TODO(hook-continue-false)`) — the interception seams have no "hard-halt the agent" primitive yet (a Decision blocks/steers a single point, not the run). Deferred with the loop-guard work; the halt request is recorded in the `hook/result` log, and the hook keeps its per-point effect (decision/context) meanwhile.
- **Config discovery.** The path is explicit in `cordis.yml` and process-level (see above); the full multi-layer CC/Codex precedence walk, per-session project-local discovery, and the trust/hash model are not reimplemented (`TODO(per-session-hook-config)`).
- **Session-start / subagent-start context is best-effort, not gated (`TODO(session-start-gating)`).** `agent/session-start` is a synchronous emit and the bridge runs its hook on a detached `.then`, so the injected `additionalContext` is not guaranteed to land before the first turn reaches the model — a slow hook can miss the first request (the context then arrives as a later injection). `subagent/start` is sharper: an in-process provider may have already queued the child's prompt before the listener runs, and a short-lived child can finish before the detached inject fires. Making startup context a gated/awaited primitive is a loop-level change deferred to the interception seams; today the contract is "injected as soon as the hook resolves", not "before the first request". The bridge tests do NOT wait on the injection where they assert the guaranteed-timing behavior, so they document the real (best-effort) timing rather than masking it.
- **Session-start / subagent-start context is best-effort, not gated (`TODO(session-start-gating)`).** `agent/session-start` is a synchronous emit and the bridge runs its hook on a detached `.then`, so the injected `additionalContext` is not guaranteed to land before the first turn reaches the model — a slow hook can miss the first request (the context then arrives as a later injection). `subagent/start` is emitted only after child publication, so the bridge can capture the live in-process child synchronously, but the result driver may queue the prompt as that same readiness boundary resolves and a short-lived child can finish before the detached hook injects. Making startup context a gated/awaited primitive is a loop-level change deferred to the interception seams; the contract is "injected as soon as the hook resolves", not "before the first request". The bridge tests do NOT wait on the injection where they assert the guaranteed-timing behavior, so they document the real (best-effort) timing rather than masking it.
## Alternatives considered

32
docs/rfc/implemented/feature/2026-06-30-interception-seams.md
View File

@@ -6,21 +6,31 @@ Status: implemented
The harness needs a hooks subsystem: users extend or gate the agent at lifecycle points the way Claude Code (CC) and Codex do. The key reframe driving this design is that **"native hooks" are not a package** — a native hook is just an ordinary Cordis plugin subscribing to the canonical lifecycle events. So the real product is a *powerful, well-typed canonical event surface*; the CC/Codex bridges (the `dsh-hooks-claude` / `dsh-hooks-codex` packages) are merely translators that map an external shell-hook protocol onto that same surface. Anything a bridge can do, a plain plugin can do directly — more powerfully (no serialization boundary, full `ctx`, typed returns).
Before this change the interception surface was incomplete and inconsistent for that goal: there was no per-prompt seam (CC's `UserPromptSubmit`), no session-start signal (CC's `SessionStart`), the single `tools/execute` waterfall conflated the pre-gate and post-inspect phases (CC splits `PreToolUse`/`PostToolUse`), and `agent/turn-continuation` returned a bare `boolean` with no room for a force-continue *reason*. The [event-domain-semantics RFC](../architecture/2026-06-30-event-domain-semantics.md) pinned down the three-domain rule and the typed-Decision idiom as the interception convention; this RFC builds the actual seams on top of it.
The surface needs distinct contracts for per-prompt policy (CC's `UserPromptSubmit`), session-start observation (CC's `SessionStart`), pre-tool policy, around-dispatch control, post-tool transformation, final-result observation, and continuation with a model-facing reason. Conflating those phases gives plugins mutation channels they do not need and makes finality depend on listener ordering. The [event-domain-semantics RFC](../architecture/2026-06-30-event-domain-semantics.md) supplies the three-domain rule and the typed-Decision idiom; this RFC applies them to the lifecycle seams.
## Decision
Add/​reshape the interception seams so every one returns a small, seam-specific **typed Decision union**, and the set covers the hook points in scope (`session-start`, `prompt-submit`, `pre-tool`, `post-tool`, `stop`-via-continuation).
The canonical surface separates transformable policy, around-dispatch control, and observe-only notification. Policy waterfalls return small seam-specific **typed Decision unions**; wrappers return normalized results; notifications receive immutable snapshots and cannot affect the outcome. The set covers the hook points in scope (`session-start`, `prompt-submit`, `pre-tool`, `post-tool`, `stop`-via-continuation) while leaving non-hook execution policy independently composable.
**New `agent/*` events** (`dsh-agent`):
**Agent events** (`dsh-agent`):
- `agent/session-start(agent, source)` — emit, once before turn 1, carrying a `SessionStartSource` (`startup` for a fresh/forked create, `resume` for a reloaded persisted session; `clear`/`compact` reserved). A pure notification — it CANNOT block startup (a deliberate gap: a bridge logs/injects, it does not gate startup). A listener seeds context via `agent.inject()`.
- `agent/prompt-submit(agent, content, source, next) → PromptDecision` — waterfall, fired per drained queued message inside the open turn, before the `user/message` append. `allow` (optionally rewriting the prompt `content` or attaching `additionalContext`) or `block` (dropping the prompt; the loop appends a durable `prompt/blocked` in its place — see the dispatch note below).
**Reshaped** `agent/turn-continuation` from `(…, defaultDecision: boolean) → boolean` to `(…, defaultDecision: ContinuationDecision) → ContinuationDecision`. A `{action:'continue', reason?}` may carry model-facing context recorded as next-step steering in the same turn — the typed twin of the existing `/goal` step-end-steer pattern.
**`agent/turn-continuation`** receives and returns a `ContinuationDecision`. A `{action:'continue', reason?}` may carry model-facing context recorded as next-step steering in the same turn — the typed twin of the `/goal` step-end-steer pattern.
**Split** the single `tools/execute` waterfall into `tools/pre-execute` (→ `PreToolDecision` allow/deny/ask gate) and `tools/post-execute` (→ `PostToolDecision` accept/block, optionally replacing content or attaching `additionalContext`). Core dispatch sits between them as plain code inside `ToolRegistry.execute`'s outer try/catch, and the tool body keeps its own inner try/catch so a thrown tool still becomes an `isError` result that `post-execute` listeners can inspect.
### The tool pipeline gives each phase one kind of authority
**New `TurnEndReason` variant** `rejected` (`dsh-session`): a turn whose entire prompt batch was blocked by `prompt-submit`.
Every call follows one ordered pipeline: `tools/pre-execute` → monotonic guards → `tools/execute` → core dispatch → `tools/post-execute` → `tools/result`. The registry reads each caller-owned input field once, materializes `arguments` as detached lossless JSON in one recursive pass, and snapshots `ToolExecutionInput` into a pipeline execution with its own opaque token. Identity fields and deeply frozen arguments are immutable for the whole pipeline, and a nested call's `parent` contains only the enclosing execution's token rather than its live object. Optional `signal` is the only operational field an around-dispatch wrapper may add, replace, or remove, and the complete object freezes before final observers run. This identity contract prevents a policy listener from silently changing what the log, UI, and tool body believe ran.
- **`tools/pre-execute`** is the extensible waterfall gate. Its `PreToolDecision` allows, denies, or asks. Deny skips `tools/execute` and core dispatch. Ask resolves through the optional approval seam: only `allowed-once` continues through guards and dispatch; rejection, cancellation, an unavailable channel, a missing approval service, or an agent-less call becomes a normalized denial. Every outcome still reaches post-policy and final observers.
- **`ctx.tools.guard()`** installs synchronous scope-aware policy after the whole pre-execute waterfall. A guard may deny or abstain, never force-allow, so listener ordering cannot resurrect an operation that a final invariant forbids.
- **`tools/execute`** is the around-dispatch waterfall for timeout, retry, and metrics plugins. A wrapper delegates to core dispatch with `next()`, may add, replace, or remove only `exec.signal` before doing so, and receives the already-normalized result of a thrown or unknown tool; returning its own valid result short-circuits dispatch.
- **`tools/post-execute`** is the inspect/transform waterfall. Its `PostToolDecision` accepts, blocks with feedback, optionally replaces content, or attaches `additionalContext`; in-place mutation of the result is not a transform channel, because the registry rebuilds the outcome from a protected snapshot plus the returned decision.
- **`tools/result`** is the synchronous contained notification after every transform, lossless-JSON materialization, and the outer error boundary. It receives the same frozen execution identity and an immutable snapshot of the authoritative result; observer failures are contained per listener and cannot change or reject `ToolRegistry.execute()`'s returned outcome.
Core dispatch and the tool body sit inside normalization boundaries, so tool, listener, malformed-result, non-JSON result, and identity-shape failures resolve as JSON-safe `isError` results rather than escaping the turn. A post-execute listener can therefore inspect a thrown tool, and a final observer sees exactly what the caller receives and the session log can persist.
**`TurnEndReason.rejected`** (`dsh-session`): a turn whose entire prompt batch was blocked by `prompt-submit`.
### Three load-bearing loop decisions
@@ -30,13 +40,13 @@ Add/​reshape the interception seams so every one returns a small, seam-specifi
3. **A forced `continue` `reason` is enqueued through the steering channel**, so the next step's top-of-loop drain records it as steering for the continued turn — next-*step* steering within the SAME turn, not a next-*turn* prompt (matching the existing `hasSteering` force-continue override).
### Pre-tool INPUT rewrite is DEFERRED (the over-reach signal)
### Pre-tool input rewrite is a separate consistency decision
`PreToolDecision` is allow/deny/ask only — **no `arguments` rewrite**. Output replacement (`PostToolDecision.accept.content`) is safe because `tool/result` is logged AFTER execution (one source of truth). Input rewrite is NOT safe today: `assistant/message` (the model-history source) and `tool/call` (the audit record) are both logged BEFORE execution, and live consumers READ `tool/call.arguments` for presentation (the ACP bridge remembers them for `presentResult`; `dsh-tool-bash` derives the title/cwd/terminal-vs-background from them). A rewrite that changed only execution would make the UI show one command while another RAN. Designing that consistently (rewriting the audit + history + presentation as one unit) is a real consistency-design problem CC itself warns is racy — so it gets its own [proposed RFC](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md), and `TODO(pre-tool-input-rewrite)` anchors it at the loop's pre-execute call site. This does not regress any production consumer (no production `tools/execute` listener mutated `exec.arguments`). The low-level capability to mutate `exec` in a `pre-execute` listener still exists (unadvertised — a test shim uses it to thread a generated id), but it is not a first-class advertised contract.
`PreToolDecision` is allow/deny/ask only — **no `arguments` rewrite**. Output replacement is safe because `tool/result` is logged after execution from the final result. Input rewrite is different: `assistant/message` (model history) and `tool/call` (the audit record) are logged before `ToolRegistry.execute()`, while ACP and tool presentation read those arguments. The registry therefore seals the materialized arguments before `tools/pre-execute`; no listener or test shim can mutate them in place. An honest rewrite must update history, audit, presentation, and execution as one unit before that identity is created, which belongs to the separate [pre-tool input-rewrite proposal](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md) and its loop-side `TODO(pre-tool-input-rewrite)`.
### What this PR does NOT do
### Boundaries
It does **not** declare `hook/*` SessionEvents (the durable hook-invocation log) — those belong to the `dsh-hook-protocol` library, because a native plugin can already use the typed Decisions without a durable hook log. A worked native-plugin example/test in this PR (`packages/core/agent-loop/tests/interception.spec.ts`) proves all the seams compose end-to-end through the REAL loop with NO `hook/*` involved — the concrete proof that "native hooks are just a plugin". Compaction (`PreCompact`/`PostCompact`), the Notification hook, Codex `PermissionRequest`, the permission/`ask` system, and the Stop loop-guard remain deferred (`FIXME(permissions)` marks the `ask`→deny degrade).
The seam package does **not** declare `hook/*` session events (the durable hook-invocation log); those belong to `dsh-hook-protocol`, because a native plugin uses typed decisions without an external hook log. The native-plugin integration test (`packages/core/agent-loop/tests/interception.spec.ts`) composes the seams through the real loop with no `hook/*` protocol. Compaction (`PreCompact`/`PostCompact`), Notification, and Codex `PermissionRequest` remain outside this decision. The [approval seam](2026-07-06-approval-seam.md) resolves `ask` decisions through `ctx.approval`, while terminal monotonic stopping is owned separately by `agent/turn-stop`.
## Alternatives considered
@@ -45,4 +55,4 @@ It does **not** declare `hook/*` SessionEvents (the durable hook-invocation log)
## Consequences
The canonical interception surface is now complete and uniformly typed: a native plugin returns typed decisions directly, and a CC/Codex bridge maps its protocol fields onto the same unions. The loop gained four firing points (session-start emit, prompt-submit waterfall, the post-tool context buffer, the continuation reshape) and the `dsh-tools` registry runs a two-waterfall pipeline; both are documented in [architecture.md](../../../architecture.md) and the package READMEs, and the decision types in [core-data-structures](../../../core-data-structures/core.md#interception-decisions) + [tools.md](../../../core-data-structures/tools.md). All existing `tools/execute` and `turn-continuation` listeners (tests, docs) migrated to the new seams. The ACP bridge maps the new `rejected` reason to `cancelled` (its codec). A pure internal change with no editor-visible transcript shift for the existing scenarios — the new behavior only fires when a hook is registered — so the snapshot goldens are unchanged; a hook-driven snapshot scenario lands with the `dsh-hooks-claude` bridge, which is what makes a hook observable end-to-end through ACP.
The canonical interception surface is uniformly typed without giving every extension the same power: hooks return decisions, execution wrappers wrap, terminal guards only deny, and final observers only observe. The loop owns session-start, prompt-submit, post-tool context buffering, and continuation; `dsh-tools` owns identity sealing and the five-phase execution pipeline. Their contracts are documented in [architecture.md](../../../architecture.md), package READMEs, [core interception decisions](../../../core-data-structures/core.md#interception-decisions), and [tool structures](../../../core-data-structures/tools.md). The ACP bridge maps `rejected` turns to its `cancelled` codec value, while hook-driven snapshots verify the observable bridge behavior end to end.

6
docs/rfc/implemented/feature/2026-06-30-subagent-observe-enrich.md
View File

@@ -6,13 +6,13 @@ Status: implemented
The hooks subsystem ([interception seams RFC](2026-06-30-interception-seams.md)) lets a plugin observe and gate the agent at lifecycle points. Claude Code and Codex both expose **SubagentStart / SubagentStop** hooks, and CC's carry the subagent's final message. The harness already emits `subagent/start` and `subagent/end` lifecycle events ([the subagent capability-seam](2026-06-21-subagent-capability-seam.md)), but their payloads were minimal (`provider`, `id`, and on end `stopReason`) — not enough for a hooks bridge to report WHAT a subagent produced without separately reaching for the live run.
This RFC enriches the end payload. It is deliberately **observe-only**: no control-flow change, no waterfall, no `start()` restructure. A run-affecting subagent-stop decision (continuation, injection that changes the run) is a separate, larger redesign and stays out of scope.
This RFC enriches the end payload. It is deliberately **observe-only**: no control-flow change and no waterfall. A run-affecting subagent-stop decision (continuation, injection that changes the run) is a separate, larger redesign and stays out of scope.
## Decision
**Add `lastAssistantMessage` — the child's final output — to `SubagentRunEndInfo`.** On the settle path it is a DEEP CLONE of `SubagentResult.output` (so an observer sees WHAT the subagent produced without holding the run). On the REJECT path (an infrastructure fault where no `SubagentResult` was produced — the seam only knows `stopReason: 'error'`) it is absent. The clone is load-bearing for observe-only: the `subagent/end` emit fires from a detached `.then` registered *before* `start()` returns, i.e. before the caller's own `await run.result` continuation — handing listeners the same array reference would let a mutating listener corrupt the caller's `SubagentResult.output`. `structuredClone` makes the event a read-only view (a regression test mutates the event's array and asserts the caller's result is untouched); a clone failure is contained (logged, the event still fires without `lastAssistantMessage`) rather than becoming an unhandled rejection on the detached `.then`.
**Add `lastAssistantMessage` — the child's final output — to `SubagentRunEndInfo`.** On the settle path it is the readonly typed `SubagentResult.output`, so an observer sees what the child produced without holding the run. On an infrastructure rejection where no `SubagentResult` exists, it is absent and the event reports `stopReason: 'error'`. Providers and listeners are trusted same-process collaborators and honor the borrowed immutable payload contract.
Both events stay plain **`emit`s**. `subagent/end` fires from a detached `.then` on `run.result` and awaits no listener, so it is genuinely observe-only by construction — a `subagent/start` listener can still reach the live child via `ctx.agents.get(info.id)` and `inject()` into it; a `subagent/end` listener can only observe (the run has settled). Per-listener containment (already in place) keeps one bad subscriber from stranding a live run or surfacing as an unhandled rejection on the detached settle hook.
Both events stay plain **`emit`s**. Async `SubagentService.start()` attaches result observation to the ready provider run, emits `subagent/start`, and then returns the run; an in-process listener can therefore reach the published child via `ctx.agents.get(info.id)`, while a remote provider need not have a local registry entry. A rejected provider start emits neither event. The callbacks remain observe-only and per-listener containment keeps one bad subscriber from stranding a live run or starving later listeners.
## Alternatives considered

72
docs/rfc/implemented/feature/2026-07-05-dynamic-workflows.md Normal file
View File

@@ -0,0 +1,72 @@
# RFC: Dynamic workflows — a script-driven multi-agent orchestration seam
Status: implemented
## Problem
The harness can delegate ONE task to ONE child (`dsh-tool-subagent`), but work that fans out across many independent pieces — an audit over many files, a migration, multi-angle research, adversarial verification of findings — forces the model to orchestrate turn by turn: every intermediate result lands in the parent context, the plan lives nowhere durable, and coordination costs a model round-trip per step. Claude Code ships this capability as [dynamic workflows](https://code.claude.com/docs/en/workflows): the model writes a JavaScript orchestration script, a runtime executes it, and the script — not the conversation — holds the loop, the branching, and the intermediate results.
## Decision
A workflow capability family at `packages/workflow/` in the bash seam shape (interface / implementation / consumer), plus the structured-output foundation it needs on the subagent seam.
### The script contract (Claude Code-compatible)
A workflow call is two parts: a `meta` JSON parameter (the identity block — `name`, `description`, optional `whenToUse`/`phases`; the field vocabulary matches Claude Code's meta block) and a `script` — a plain-JS body with top-level `await`, ending in `return <json-value>`. Meta is DATA, never code: the engine shape-validates it and evaluates no script text to obtain it (a body still opening with a CC-style `export const meta` statement is rejected with a pointed message). The body sees exactly: `agent(prompt, {label, phase, schema, model})`, `parallel(thunks)`, `pipeline(items, ...stages)` (NO cross-stage barrier; `(prev, item, index)` callbacks), `phase(title)`, `log(message)`, and `args`. CC semantics are preserved where they matter to script authors: a failed child resolves `null` (scripts `.filter(Boolean)`); an ordinary stage throw nulls the ITEM and skips its remaining stages. CC's determinism bans (`Date.now()`/`Math.random()`/argless `new Date()` throwing) are NOT enforced — they exist for CC's journaling/resume, which this cut defers — so a CC-authored BODY runs unchanged (its meta header moves into the parameter) while scripts written here may freely read the clock.
One deliberate strictness DIVERGENCE from CC: hook misuse — unknown or deferred options (`effort`/`isolation`/`agentType`), malformed arguments, schemas outside the supported subset, tripped caps, seam start failures — throws a `WorkflowError` with `fatal: true`, and the combinators RE-THROW fatal errors instead of nulling the item. Without this, a typo'd option dissolves into a `null` indistinguishable from a child failure — the accepted-then-ignored failure mode this repo bans. One addition: the tool's `args` parameter is a JSON OBJECT (a bare list is wrapped as a field) so the wire schema stays honest.
### The seam (dsh-workflow)
`ctx.workflows` is an abstract `WorkflowService` in the bash shape — one engine per context, no named-provider registry (engines are deployment swaps, not co-residents). `start(request)` throws synchronously for a script that cannot begin; a returned `WorkflowRun`'s `result` NEVER rejects (failures resolve as `stopReason: 'error' | 'cancelled'`). The `workflow/*` events are observe-only emits carrying DATA SNAPSHOTS (id + meta; `workflow/end` omits the result value), per-listener contained, mirroring `subagent/start`/`subagent/end` — control stays with the run's holder. Vocabulary details: [core-data-structures/workflow.md](../../../core-data-structures/workflow.md).
### The engine (dsh-workflow-workerthread): one worker thread per run
**Trust premise (governs every engine decision below)**: workflow scripts are MODEL-WRITTEN — the same trust level as the model's existing bash access — so the engine defends against BUGGY scripts, never hostile ones. In scope: `result` never rejects, no unhandled rejections from dropped hook promises, loud rejection of values JSON cannot carry, fatal-vs-null hook discipline, cancellation that always frees the caller. Out of scope, deliberately: adversarial values (throwing/spinning accessors, proxies with hostile traps, prototype forgery, `prepareStackTrace` hijack) AND Node-API escape from the script's context — the vm context shares object machinery with its surrounding realm, so a script can reach the `Function` constructor (`globalThis.constructor.constructor`) and from it `process` and every Node builtin; the absent globals are API surface, not containment, and a worker thread is NOT a security boundary (an escapee holds process-wide privileges — Node's permission model is per-process). Worker-side code MAY run script code while reading script values, and that is accepted: a synchronous spin costs the script its OWN thread (terminated at the post-cancel grace), never the host loop, so containing error VALUES would be cost without a threat model. Genuine sandboxing (isolated-vm, a separate process) remains an engine swap behind the seam, not incremental defenses here.
**Why node:worker_threads**: one run uses one unpooled worker because a workflow run is already heavyweight relative to thread startup. The script runs in a vm context inside the worker, keeping the script-visible surface to the hook contract instead of exposing a bare worker realm, while `agent()` bridges by message-port RPC to I/O-bound child loops on the host. This keeps `start()` from blocking the host on the script's synchronous slice, makes the post-cancel deadline end in a real `worker.terminate()`, and gives cross-thread values a serialization boundary by construction. isolated-vm was rejected for its maintenance state, required `--no-node-snapshot` consumer flag on Node ≥ 20, and node-gyp fallback.
Host-side meta validation and body pre-parsing preserve the seam's synchronous errors, and private enum-keyed payload maps define the wire protocol. Pending async starts, published child records, one host cancellation signal, worker-death reaping, result precedence, and disposal quiescence preserve the subagent run contract across that wire; the [agent-scope runtime-design RFC](../architecture/2026-07-12-agent-scope-runtime-design.md#workflow-children-are-pending-starts-or-published-records) owns those race algorithms. Coverage uses an in-process `MessageChannel` for worker-side logic that main-process V8 coverage cannot see and separately proves the built `lib/worker.js`—a second tsdown entry sanctioned by the `"./worker"` subpath export—under plain Node in the built-bin smoke gate.
**Meta as data, never evaluated**: the meta block reaches the seam as a plain JSON request field (the tool's schema-validated `meta` parameter) and the engine only shape-validates it, every violation named. This is a host-isolation invariant, not a convenience: evaluating a meta literal host-side — even one contractually "pure", in an empty timed vm context — hands script-controlled getters a host stack with no timeout the moment the result is READ, defeating the exact spin isolation the worker thread buys.
**Value boundary**: values leaving the script (meta, hook options, schemas, the return value) go through `materializeFromRealm` — a plain recursive walk that rejects loud everything JSON cannot carry (exotic prototypes, functions, symbols, cycles, sparse arrays, non-finite numbers, nested `undefined`), copying via `Object.defineProperty` so a `"__proto__"` key becomes a data property, never a prototype mutation; getters are read ordinarily and their RESULT crosses (a throwing read fails loud) — which is also what makes every later postMessage hop total. Values entering the realm (`args`, `agent()` results, hook promises and failures, combinator arrays) are handed over directly as worker-realm values — the script is trusted, so outer prototypes are not a leak; `args` rides the `workerData` structured clone (the caller-isolation copy) and is cloned once more so a script scribbling on it cannot mutate the session's init object. Hook failures are `WorkflowError`s built OUTSIDE the script's context: the combinators recognize fatality by `instanceof` against the engine's own class (unforgeable from the script), and the script-visible consequence — in-script `instanceof Error` is `false` for hook errors; branch on `e.name`/`e.code` — is documented in the engine README. Realm functions (stages, thunks) are called, never materialized. Thrown script values are rendered by a total renderer (stack → message → `String()`, fixed label if rendering throws), so `result` cannot reject. Caps (`maxConcurrentAgents` auto = `min(16, max(1, availableParallelism() - 2))`, `maxTotalAgents` 1000, `maxItemsPerCall` 4096) and timeouts are validated Config, not literals.
### The consumer (dsh-tool-workflow)
A `workflow` tool mirroring `dsh-tool-subagent`'s synchronous shape: start, await, `try/finally` dispose, abort-bridge `exec.signal`, non-`completed` → `isError`. Render intent: a `generic` card titled by the call's `meta.name` parameter (presentation is a pure function of args). The tool description IS the model-facing authoring spec. The usage policy ships with the tool as its own `tool:<toolName>` prompt section (explicit-ask-only guidance — tool guidance lives in tool plugins, never in the deployment persona); the harness has no ultracode-style effort gate.
### The foundation: structured output on the subagent seam
`SubagentStartRequest.outputSchema` is implemented by `dsh-subagent-inprocess` for both in-process backends. Each structured child receives its own scoped capture tool, instruction, and enforcement registrations on `child.ctx`; concurrent children can use different schemas without sharing mutable policy, and disposing the child removes the entire attachment.
An output schema makes a schema-valid committed capture mandatory for successful child completion. The scoped runtime presents the capture tool and instruction, commits only a successful final outcome—including the enclosing `run_code` outcome for an SDK call—denies later side effects after capture becomes pending, and stops the child without another model step after commit. A validation failure remains a retryable tool error; clean completion without a committed capture settles as an error.
`StructuredOutputSchema` is the raw enforceable JSON-Schema subset in `dsh-tools` (single-string `type`, `properties`/`required`/`additionalProperties`, `items`, scalar `enum`/`const`), and unsupported keywords fail loudly because that wire data becomes the capture tool's parameters verbatim. The [agent-scope runtime-design RFC](../architecture/2026-07-12-agent-scope-runtime-design.md#structured-output-commits-only-authoritative-outcomes) owns the assembly, commit, guard, and terminal-stop correctness algorithms.
## Deferred (documented non-goals of this cut)
- **Background collection** (start tool → run id → completion notice → collect), designed alongside bash/subagent background unification.
- **Journaling + resume** (`resumeFromRunId`, cached agent() prefixes) — implementing it reintroduces CC's determinism bans as a script-contract tightening (scripts may read the clock today).
- **Saved/bundled workflows** (a `.deepseek/workflows/` registry, slash-command surface) and **script persistence to a run directory** (the tool-call event already records the script durably).
- **Nested `workflow()`**, **token `budget`**, and the `effort`/`isolation`/`agentType` agent options (each rejects loud with a message naming it deferred).
- **An overall run wall-clock timeout** — cancellation always frees the caller (result settles within the grace), so a cap on total run time is a policy knob for the background redesign, not a correctness need here.
- **Engine hardening beyond worker threads**: an isolated-vm or separate-process engine behind the same seam (actual sandboxing; memory limits).
- **ACP progress UI** over the `workflow/*` events (a `/workflows`-style view); the events exist for it.
- **ACP-backend structured output** and **`toolFilter`** (both still capability-gated `false`).
## Alternatives considered
- **Hostile-value containment in the host** (trap-free proxy rejection, accessor-never-invoked descriptor walks, realm-side pre-rendering of thrown values, realm-built promises/arrays/error clones with structural fatal recognition): rejected because every defense targets an author the trust premise accepts, while the thread's serialization boundary already makes cross-realm values total by construction.
- **In-process `node:vm` execution**: mechanically simplest — no RPC, no thread — but `start()` blocks the caller for the script's initial synchronous slice, a synchronous spin past the first await cannot be killed in-process (the vm `timeout` covers only that first slice), and `dispose()` could only abandon an unsettling script on the host loop. The worker-thread engine keeps the same vm-context script surface while unblocking the host and making termination real.
- **Background execution as the default** (CC's shape): deferred; foreground-synchronous matches `dsh-tool-subagent`'s cut, and background semantics should be designed ONCE across bash/subagent/workflow rather than per-tool.
- **Workflow-layer JSON parsing for `agent({schema})`**: duplicating a seam concern at one consumer while the seam's capability flag stayed dishonestly `false`.
- **Meta embedded in the script as `export const meta = {...}`** (CC's exact format): keeps scripts self-contained and CC scripts drop-in, but obtaining meta requires evaluating model-written text on the host. Even an empty timed vm context cannot bound script-controlled getters when the host reads the resulting object. A JSON parameter removes the scanner, evaluation, and host-spin hole; the cost is that a CC script's meta header must move into the parameter (the body stays drop-in).
- **`SchemaSpec` as the outputSchema type**: the author-facing DSL cannot express what arrives as data and cannot be validated against without conversion loss.
- **A schema-object library (zod, or the repo's schemastery) for the structured-output subset**: the schema is wire data — plain JSON that crosses the vm realm boundary in `agent({schema})` and lands verbatim in the forced tool's parameters — exactly where live schema objects cannot sit; consuming raw JSON Schema at runtime would need a third-party converter on top (zod core only emits JSON Schema, not the reverse), and it would put a second schema language beside schemastery's config role.
- **ajv for value validation**: it validates FULL JSON Schema, so the subset gate — the module's actual point, since every accepted keyword must be one the harness enforces — would remain hand-written regardless; it compiles validators through `new Function`; and it would be dsh-tools' first runtime dependency, all to replace the ~70-line value walker while the path-qualified, every-violation error reporting stays custom either way.
- **Provider JSON mode (`response_format: {type: json_object}`) instead of the forced capture tool**: the official API guarantees valid JSON, not schema-conforming JSON (no `json_schema` type; the docs' own guidance is to validate client-side, with the schema riding in the prompt), so both walkers survive untouched and only the capture-tool mechanics could go — at the cost of tools during a structured child's run (whether `response_format` composes with tool calling is undocumented), the in-turn validation retry (`ToolArgsError` keeps recovery inside the turn; a JSON-mode empty body — a documented failure mode — ends the turn, and the only recovery is the re-prompt loop this design rejects), and a new per-adapter `LlmCallConfig` surface. The accepted upgrade path is strict TOOL schemas (provider-side constrained decoding on tool parameters) when available: the same forced tool and subset gate, with the gate narrowed to the provider's strict subset.
## Consequences
The harness gains CC-compatible script orchestration: fan-out plans live in a rerunnable artifact instead of the parent context, and `outputSchema` yields an authoritative structured child result across native and Code Mode presentation. The cost, bounded by the trust premise, is a worker thread per run (~tens-of-ms spin-up), every hook crossing a message port as RPC, and a termination-path `agentsStarted` that degrades to the host-observed count; in exchange `start()` never blocks the host, a post-cancel grace ends in a real `worker.terminate()`, and the value boundary is serialization by construction. A worker thread is still not a security boundary — scripts share the model's trust level, and actual sandboxing requires an isolated-vm/separate-process engine behind the seam. The fatal-vs-null strictness divergence from CC means a CC-authored script that relies on option typos dissolving to `null` behaves differently, preserving the repo's no-accepted-then-ignored rule. Consumers must hold the run handle for control (`cancel`/`dispose`); observers get data snapshots only, so no listener can extend a run's lifetime or corrupt another's view.

53
docs/rfc/implemented/feature/2026-07-05-skill-system.md Normal file
View File

@@ -0,0 +1,53 @@
# RFC: Skill system — progressive disclosure instructions for agents
Status: implemented
## Problem
Agent products have converged on a skill pattern: keep the request prompt small by listing only available instruction bundles, then load the full body when the model decides a task matches. Codex, Claude Code, OpenCode, and Kimi Code differ in details, but all separate discovery metadata from complete instructions so a workspace can carry reusable behavior without paying the full prompt cost on every turn.
DeepSeek Harness uses the same primitive so project-specific review, plugin-authoring, and tool-usage guidance lives next to the workspace or the user's agent configuration instead of being hard-coded into the loop.
## Decision
`@deepseek-ai/dsh-skill` is the pure provider registry (`ctx.skills`), `@deepseek-ai/dsh-skill-local` is the shipped local filesystem provider, and `@deepseek-ai/dsh-tool-skill` owns the session-prefix catalog and model-facing loader tool. `dsh-agent-core` loads the registry, local provider, and consumer by default so stdio and ACP apps get the same behavior while embedded or remote providers contribute skills without changing the registry or consumer. Its `skills` config forwards `registry`, `local`, and `tool` branches to those owners.
Provider plugins register synchronously during `apply()`. Provider catalogs return ranked candidates from awaited `list()` calls, where remote providers perform initialization, authentication, and discovery while honoring the lookup abort signal. The registry validates each candidate, resolves same-name skills first-wins by rank, provider registration order, and provider-local order, then sorts summaries by skill name for deterministic consumers. It caches only completed catalog snapshots and retries when a provider/runtime revision changes during discovery, so an unload cannot freeze a stale, unresolvable skill into a session prefix. Runtime `ctx.skills.register(...)` remains a convenience for embedded in-process skills and uses project-over-user priority; `runtime` is reserved as the registry-owned provider name.
The local provider scans cwd-sensitive project roots, custom roots, and user roots in first-wins rank order: project `.dsh`, project `.agents`, `customSkillDirs`, user `.dsh`, then user `.agents`. The user `.dsh/skills` scan skips `.system` so a system-owned directory is not treated as normal user content. DeepSeek Harness does not ship built-in system skills; embedded or remote providers supply additional skills when configured.
Each skill is either `<name>/SKILL.md` or `<name>.md` with YAML frontmatter. `name` and `description` are required; `whenToUse`, `disableModelInvocation`, and `metadata` are optional. Names are kebab-case. YAML frontmatter is parsed with the `yaml` package instead of `js-yaml` or a hand-written parser: `yaml` is the already-declared modern parser for this package's limited frontmatter needs, and a narrow parser would either reject valid YAML users expect to work or grow into an unreviewed YAML subset.
Local skill filesystem I/O goes through `ctx.fs` when a filesystem service is loaded: project-root lookup probes `.git` with `resolve` and `stat`, root discovery uses `listDir`, and skill reads use `readText`. The Node filesystem remains a fallback for minimal contexts that mount `dsh-skill-local` without the fs seam. Missing roots, unreadable or malformed skill files, and transient provider `list()` failures degrade to warn-and-skip so one bad source does not make every agent request fail; malformed candidates still fail fast because they are provider contract violations.
`dsh-tool-skill` contributes one user-role `<system-reminder>` catalog through [`agent/session-prefix`](2026-07-07-session-prefix.md). The catalog contains sorted skill name and description only; it excludes bodies, paths, sources, providers, and routing hints. Descriptions are whitespace-normalized, XML-escaped, and capped by `catalogDescriptionMaxLength`, whose default is `500` and minimum is `3`. The session-prefix seam freezes the request-only catalog per loop instance and records it in the request header, preserving reconstructability without adding it to durable history. Full skill bodies are never included in the catalog.
The `skill({ name })` tool loads one full skill for the current agent cwd and returns a tool result containing `<skill_content name="...">`, `<skill_resources>`, and `<skill_instructions>`. `resourceBase` supplies a directory, URL, or opaque provider-managed base for explicitly referenced scripts, references, and assets; resources load only as needed, without directory enumeration. An unresolved name reports that the skill is unknown or no longer available; invalid names and skills marked `disableModelInvocation` retain distinct tool errors. The tool result is the model-visible disclosure path.
The data structures and catalog/tool contract are documented in [skills.md](../../../core-data-structures/skills.md), with service signatures in the generated [services catalog](../../../cordis-catalog/services.md).
## Alternatives considered
**Inject full skill bodies into every system prompt.** Rejected because it destroys progressive disclosure and makes every request pay for instructions that may not apply.
**Expose skills only as slash commands.** Rejected because model-initiated loading is the core capability; slash/ACP command advertisement does not change discovery.
**Put local filesystem scanning directly inside `ctx.skills`.** Rejected because coding agents, web agents, and future plugin ecosystems need different skill sources. A provider registry mirrors the subagent seam: the registry owns conflict resolution and consumers, while implementations own loading.
**Use a system-prompt section.** Rejected because the rendered system prompt is a single string, while the catalog is a user-role `<system-reminder>` message with request-only lifecycle requirements. [`agent/session-prefix`](2026-07-07-session-prefix.md) is the selected mechanism: it places the catalog ahead of derived history and records the composed message in the request header.
**Materialize built-in DSH authoring skills under `~/.dsh/skills/.system`.** Rejected because bundled skills do not write user home on startup, and embedded or remote providers supply configured skills.
**Recursively discover nested `**/SKILL.md`.** Rejected. Flat files and one-level directory bundles cover the configured roots while keeping duplicate handling and catalog order easy to reason about.
**Hand-parse frontmatter.** Rejected because the accepted schema includes an open `metadata` object. A narrow parser would either reject valid YAML users expect to work or grow into an unreviewed YAML subset.
## Consequences
The agent-core spine includes one session-prefix contributor, one local provider, and one model-facing tool. Skill discovery is cwd-sensitive, so callers that create agents with different session cwd values can observe different project skill overrides by design.
The catalog is deterministic for a fixed root set and runtime registration revision, but disk changes are not watched; discovery is memoized until runtime registration invalidates the cache or the process restarts.
## Deferred
Forked skill contexts (`context: fork`), direct user/slash invocation (`user-invocable`), parameter declarations and hints (`arguments` and `argument-hint`), and per-skill tool constraints (`allowed-tools` and `disallowed-tools`) are outside the shipped contract. The registry, local provider, and model-facing tool do not parse, advertise, or enforce these fields.

141
docs/rfc/implemented/feature/2026-07-06-approval-seam.md Normal file
View File

@@ -0,0 +1,141 @@
# RFC: The approval seam — one-shot permission decisions over a waterfall of answerers
Status: implemented
## Problem
Two callers need to put one question — "may this specific action proceed?" — to a human: `tools/pre-execute`'s `ask` decision (including the Claude-Code hook bridge's `permissionDecision: ask`) and the [sandbox RFC](2026-07-06-sandbox.md)'s post-denial one-shot escalation retry. A shared seam keeps them from inventing separate outcome vocabularies, UI routing, cancellation, and audit trails, while guaranteeing that a deployment with no UI can never grant an unanswerable request.
The routing problem is ownership: an approval prompt must reach the editor session that owns the asking agent (the ACP bridge multiplexes N sessions over one connection), fail closed for agents nobody owns (in-process subagents, tests), and stay out of deployments that compose no UI (headless, CI).
## Decision
One package, `dsh-user-approval` (`packages/ui/user-approval`), owning the vocabulary and the `ctx.approval` service — the MECHANISM. The POLICY — who answers, and whether a session is asked at all — lives outside it: answerers are `approval/request` waterfall listeners registered by the plugins that own the channel (the ACP bridge; future terminal UIs; test scripts), and a per-session policy tier can decide before any human is involved. Consumers (`dsh-tools`' ask routing, the sandbox escalation gate) resolve a question to a closed outcome and derive their own tool results from it. Deliberately ONE package, not the capability-seam three (see Alternatives).
### How a deployment uses it
One `cordis.yml` entry mounts the seam. Not loading it is the fail-closed opt-out: consumers deny unanswerable requests with zero approval code registered.
```yaml
- id: approval
name: '@deepseek-ai/dsh-user-approval'
# config:
# policy: never # deployment default for sessions without an override; 'ask' when omitted
```
The entry alone provides mechanism, not a channel: with no answerer composed, every ask resolves `unavailable` and the asking tool call denies — fail-closed needs no configuration. Composing the ACP app (`@deepseek-ai/dsh-acp-agent`, as in [the sandbox example](../../../../examples/sandbox-acp-agent/README.md)) completes the loop: its bridge registers an answerer that prompts the owning editor session via `session/request_permission`, so a hook's `ask` or an escalation request surfaces as a one-shot Allow/Reject prompt attached to the already-streamed tool call. `policy: never` is the unattended stance — every ask auto-rejects deterministically, stated in the system prompt, no human in the loop. `policy` is validated against the closed list at plugin load; anything else throws.
What a composed deployment observes: `allowed-once` lets exactly that call proceed; rejection, dismissal, and channel absence deny with three distinct reasons the model can tell apart; every ask lands a durable `approval/asked`/`approval/decided` pair on the asking agent's session log; nothing about a grant persists past the call that asked.
One ask under this composition, verbatim from the sandbox example's recorded `escalation-approved` scenario — the model requests a sandbox escalation, the gate asks, the bridge prompts the owning editor, the user clicks Allow once:
```
tool/call bash {"command": "printf 'escalated\n' > escalated.txt && cat escalated.txt",
"sandbox_permissions": "workspace-write",
"justification": "the user asked to write escalated.txt in the workspace"}
approval/asked {"toolName": "bash", "callId": "call_00_…",
"reason": "escalate sandbox to workspace-write: the user asked to write escalated.txt in the workspace"}
→ session/request_permission {"toolCall": {"toolCallId": "call_00_…"},
"options": [{"optionId": "allow-once", "name": "Allow once", "kind": "allow_once"},
{"optionId": "reject-once", "name": "Reject", "kind": "reject_once"}]}
← the user picks "Allow once" on the prompt the editor attaches to the streamed bash call
approval/decided {"outcome": "allowed-once"}
tool/result "escalated" — this one call ran under the wider mode; the grant died with it
```
The `escalation-rejected` twin ends in `{"outcome": "rejected"}` instead: nothing executes, and the model's result carries the asker's verbatim fail-closed text (`the user rejected escalating this command to "workspace-write"`). A hook's `permissionDecision: ask` rides the identical wire; only the asker and its deny texts differ (§ Ask routing in dsh-tools). Headless, the same request skips the prompt entirely and settles `unavailable`.
### Design detail
#### The seam: mechanism and policy split
After request validation and a successful `approval/asked` append, the answerer phase always resolves to a closed `ApprovalOutcome` — `allowed-once` / `rejected` / `cancelled` / `unavailable`. `ApprovalRequest` is a readonly same-process contract, so the service borrows its routing identity and cancellation signal instead of copying the record or capturing a parallel callback bundle. It dispatches the `approval/request` waterfall, races the request signal (abort settles `cancelled`; a late answer is discarded, never double-audited), contains a throwing answerer as `unavailable`, normalizes a rogue non-vocabulary return to `unavailable`, and lands the log-only audit pair `approval/asked`/`approval/decided` (paired by the branded `ApprovalRequestId`) on the request agent's session log. Request acceptance and either pre-commit audit append may still reject; returning a decision that could not be logged would violate the pair. Session owns post-commit observer containment, so a callback failure cannot turn an authoritative audit append into a rejected request or suppress the matching event. Grants are one-shot by definition: `allowed-once` authorizes the single asked-about action, never a class of future ones, and the service stores nothing between requests. `request()` also throws before appending anything when the agent's session has no open turn — the audit pair must be turn-enclosed, the turn being the durable log's commit/replay boundary (a bare event between turns is dropped as crash tail on reload); every ask path runs mid-turn already, and idle asks are a deferred design.
Answerers are the policy, and they are `approval/request` waterfall listeners. The waterfall buys exactly what the seam needs: with zero listeners the dispatch falls through to the caller-supplied default — `unavailable`, so fail-closed needs no configuration and no code in any deployment; a listener that recognizes the request's agent answers by returning an outcome without calling `next()` (the decision slot is single-occupancy, first answer wins — the same documented semantics as the `fs/write-intent` gate); a listener that does not recognize the agent MUST delegate via `next()` so another answerer or the default gets the question; and listeners dispose with their owning fiber, so an unloaded UI plugin degrades the next ask to `unavailable` instead of leaving a dangling channel. Registration order across sibling plugins is not load-order deterministic (the loader starts siblings concurrently), so a deployment composes ONE terminal answerer and reserves `prepend` listeners for decide-or-delegate gates.
`ApprovalRequest` carries the asking `agent` (routes the question; receives the audit events), the `toolName`, the optional exact `callId`, the asker's human-readable `reason`, and the optional `signal`. The caller retains ownership and honors the readonly contract for the duration of `request()`. The vocabulary is deliberately self-contained — it names the tool-call by the `CallId` brand from `dsh-llm` and never imports `dsh-tools` — because `dsh-tools` depends on `dsh-user-approval` (the ask routing) and a `ToolCallView` import would close a package cycle. It deliberately does NOT carry tool arguments: a UI answerer attaches the prompt to the already-streamed tool call via `callId` instead of re-rendering the call.
#### Ask routing in dsh-tools
`ToolRegistry.execute()` resolves an `ask` decision through the seam before the shared deny path: `allowed-once` proceeds to guards and dispatch, and the three non-grants deny with distinct reasons — "the user rejected…", "…was cancelled", "…no approval channel is available" — so the model can tell a human "no" from an absent channel. The seam is consumed opportunistically (`ctx.get('approval')`, the `tool-bash`/`agent-loop` pattern), not statically injected: with no ApprovalService, or after one unmounts, the next ask fails closed without gating the registry's fiber. An agent-less execution also fails closed — without an agent there is no session to audit to and no UI to route to.
#### The per-session policy tier
The seam also owns the session-scoped approval policy — the approval knob of the two-knob per-session switching design ([the sandbox RFC](2026-07-06-sandbox.md) § Per-session modes is the pattern's home: one log-only event per knob, a pure fold, THE write path, ACP config-option advertisement, and turn-anchoring). `ApprovalPolicy` is `'ask' | 'never'`, and `effectiveApprovalPolicy(events) ?? Config.policy` (default `'ask'`) decides every request BEFORE any interactive answerer: the service resolves a `'never'` session to `'rejected'` INSIDE `request()`, before dispatching the waterfall at all — no listener registration, including a later `prepend`, can sit ahead of it — while `'ask'` dispatches unchanged and falls through to fail-closed `'unavailable'` when nobody answers. Visibility follows the switching design's two layers with one asymmetry: the prompt section states ONLY `'never'` (deterministic, availability-independent — "you will be prompted" would overclaim in a composition with no answerer, and absence under a logged header is exactly how the narrator reads `'ask'` back), the narrator injects at most one coalesced notice per switch, and the audit pair still lands on every ask, including the policy's auto-rejections.
#### The ACP answerer
The bridge registers the first real answerer: it resolves the owning session through its existing `WeakMap<Agent, sessionId>` reverse map, issues `session/request_permission` with the request's `callId` as the `toolCall` reference and the one-shot options `allow_once`/`reject_once`, and maps the response — selected `allow-once` → `allowed-once`, any other selection → `rejected` (an unknown optionId from a non-conforming client never grants), client `cancelled` → `cancelled`. A request for a foreign agent — or one without a `callId`, since the protocol prompt must attach to a tool call — delegates via `next()`. A rejected RPC (client gone mid-prompt) propagates to the service, which contains it as `unavailable`. Whether a call ASKS at all is policy — a hook or `tools/pre-execute` plugin returning `ask` — never the bridge's own judgment.
The answerer routes through the bridge's reverse-map ownership seam described by [the ACP support RFC](../../implemented/feature/2026-06-14-acp-agent-client-protocol.md), implementing the per-session permission ownership required by [the multi-session RFC](../../implemented/feature/2026-06-14-acp-multi-session.md).
#### Audit, and what the model sees
`approval/asked` / `approval/decided` are log-only session events (the `hook/invoked`/`hook/result` precedent): durable, replayable, never in the model transcript. The model's entire view of an approval is the tool result the asker derives from the outcome — reconstructability holds because that result is an ordinary logged `tool/result`. One `decided` lands per `asked`, whatever the outcome, including an already-aborted signal (settled `cancelled` without dispatching), a contained answerer failure, or a session observer that throws after either event is already appended.
#### Entities and dependencies
One package, no cycles: `dsh-user-approval` peers on `cordis`, `dsh-session` (event-map merge + append), `dsh-agent` (the `Agent` type), `dsh-llm` (`CallId`, via `dsh-brand`). `dsh-tools` and `dsh-acp` each peer on it; the escalation phase's asker lives in `dsh-tool-bash` (see [the sandbox RFC](2026-07-06-sandbox.md) § Escalation), so the sandbox family keeps its ZERO-edge relation (the executor contributes the per-call override mechanism, and transport seams never ask humans questions). The seam is one package, not the capability-seam three: the service body (dispatch + audit) has no replaceable implementation — the replaceable part is the answerer listeners, and those live with their owners (the bridge; future terminal UIs; test scripts). `@cordisjs/plugin-capability` stays orthogonal (a static grant registry answers "is this already authorized", not "ask the user now"), and `subagent-acp`'s child-side `permission` auto-answer is untouched — routing a child's approvals to the parent session is deferred (§ Deferred).
### Testing
Unit tier: the service's outcome branches (fail-closed default, first-wins slot, delegation, containment, rogue-value normalization, abort-before and abort-during with late-answer discard, fresh ids, fiber-disposal degradation), scoped routing, post-append observer throws on both audit events, and the policy tier (both values × dispatch/decide, a `'never'` decision unbypassable even by an answerer prepended AFTER the service, audit pair intact) in `dsh-user-approval`; the ask routing matrix (grant dispatches; three non-grant reasons pinned verbatim; unmounted and agent-less degrades; the registry's own exhaustiveness backstop against a non-conforming stand-in) in `dsh-tools`; the answerer (wire shape of the prompt, outcome mapping, unknown-option conservatism, foreign-agent and call-less delegation) driven through a real bridge + scripted client in `dsh-acp`.
Snapshot tier: the harness accepts scripted permission answers (`permissionAnswers` in a scenario's `input.json`, consumed FIFO; an unscripted prompt answers `cancelled`, fail closed). The seam's wire is recorded end to end in the sandbox example's suite: both escalation branches drive `session/request_permission` through this seam over scripted answers (grant and rejection), and the recorded `mode-switching` scenario pins the `'never'` prompt sentence and the policy-switch notice ([the sandbox RFC](2026-07-06-sandbox.md) § Testing).
## Deferred
- **`allow_always` grant storage** — honoring a persistent grant means designing storage, scope identity (call? path? prefix? session? time window?), and revocation; until designed, only the one-shot options are advertised ([the sandbox RFC](2026-07-06-sandbox.md) § Escalation records the open scope question).
- **A recorded hook-driven `ask` through a composed answerer** — the human-prompt wire is recorded through the sandbox example's escalation branches. The hook matrix's `hook-cc-pretool-ask` pins the no-ApprovalService fallback denial, while the hook-producer-plus-answerer composition remains on the unit tier.
- **Routing a child agent's approvals to the parent session** — `subagent-acp`'s child auto-answers its own `permission` requests; surfacing them to the parent's editor is its own design.
## Alternatives considered
- **A single registered provider instead of waterfall listeners** — rejected: a `registerProvider()` surface forces every composition question — allowlist pre-filters, external hook deciders, scripted test answers, a policy gate in front of a human — inside one provider implementation. The waterfall gets composition, fail-closed absence, and HMR disposal from machinery the runtime already has; the seam's JSDoc pins the single-decision-slot convention instead of inventing a provider registry.
- **An inline `tools/pre-execute` permission gate in the ACP bridge** — rejected: prompting for every bridge-owned call hardwires the asking POLICY into the UI plugin, cannot serve a second asker (sandbox escalation happens after execution starts, with no pre-execute moment), and leaves hook-produced `ask` decisions without a shared mechanism.
- **The generic user-interaction seam (`ctx.userInteraction`)** — rejected as the approval mechanism: the two share a skeleton (route by agent, block for a human, handle absence), but approval's contract is narrower in every dimension that matters: a closed outcome vocabulary instead of free text, a protocol-native prompt attached to a tool call instead of a generic form, mandatory fail-closed absence, and audit events. Approval therefore does not ride the shipped `packages/ui/user-interaction` / `ask_user_question` elicitation path — an elicitation form is not a permission prompt, and a free-text answer is not a closed outcome; sharing provider plumbing stays open if the two ever converge.
- **Static optional injection in `dsh-tools`** — rejected: the vendored cordis `Inject` type has no optional flag — the object form maps service names to intercept config, and a declared inject gates the fiber. `ctx.get('approval')` is the documented opportunistic-consumption pattern (the `tool-bash` owner-token lookup, the loop's persistence probe), reads presence per call, and degrades correctly across HMR without extra machinery.
- **The capability-seam three-package split** — rejected: interface/implementation/consumer fits a seam whose implementation is swappable (bash-local vs bash-sandbox). Here the service body is fixed mechanism and the variable part is listeners that live with their owners — splitting would manufacture an implementation package with nothing in it ("don't split preemptively").
- **Offering `allow_always` now** — rejected: the protocol can express it, but honoring it means designing grant storage, scope identity, and revocation (§ Deferred). Advertising an option the harness cannot honor manufactures doomed grants.
## Consequences
The implemented contract is pinned by the suites in Testing:
- With an ApprovalService and an answerer composed, a hook's `ask` reaches a human and `allowed-once` dispatches the tool; every other outcome denies with its distinct reason.
- A `'never'` session auto-rejects every ask without prompting anyone, states the policy in its prompt, and narrates switches (the shared switching mechanics are pinned in [the sandbox RFC](2026-07-06-sandbox.md)).
- Every unanswerable path fails closed to `unavailable`: no service, no listener, a foreign or agent-less request, a throwing answerer, a rogue return value, or a dead client connection.
- Every `request()` routes through its readonly agent identity and lands exactly one `approval/asked`/`approval/decided` pair on that agent's log, replayable and invisible to the model transcript; post-append observer failures cannot split the pair.
- Prompts route per-session through the bridge's ownership map; one session's prompt can never reach another session's editor.
- A deployment with no ApprovalService emits no approval prompt or approval audit events and denies every `ask` request.
Costs and accepted limits:
- **Two decide-eager answerers race for the slot.** Sibling-plugin listener order is not deterministic, so the seam cannot referee competing terminal answerers — mitigated by convention (one terminal answerer per deployment; `prepend` only for decide-or-delegate gates) rather than a priority mechanism the event bus does not have.
- **Production exercise rests on one composition.** `ask` has two producer families — the hook bridges through `tools/pre-execute`, and sandbox escalation through its own gate — with the wire recorded in the sandbox example's snapshot suite, so the seam's real-world coverage is that one composition until more deployments compose it.
- **Ownership keys on `Agent` object identity.** The answerer resolves sessions through the bridge's existing WeakMap; every current path hands the same object through the loop and the seams, but a future boundary that clones or proxies agents would make the bridge delegate and fail closed — safe, but silently UI-less — and would need session-id matching instead.
## FAQ
Behavioral and usage questions only — every "why not X?" design question lives in [Alternatives considered](#alternatives-considered), whose job is exactly that.
- **What happens in a deployment with no answerer at all (headless, CI)?** Every ask falls through the empty waterfall to `unavailable` and the tool call denies with the "no approval channel is available" reason. Fail-closed is the zero-listener default, not a configuration.
- **Can a grant persist — "always allow this"?** No. `allowed-once` authorizes the single asked-about action and the service stores nothing between requests; `allow_always` is deliberately not advertised until grant storage is designed (§ Deferred).
- **What does the model see of an approval?** Only the tool result the asker derives from the outcome — the audit pair never enters the transcript. The three non-grant reasons are distinct, so the model can tell a human "no" from a dismissed prompt from a missing channel.
- **Who decides whether a call asks in the first place?** Policy producers: a hook returning `permissionDecision: ask`, any `tools/pre-execute` listener, or the sandbox escalation gate. The seam and the bridge only route and answer; neither injects its own judgment about what deserves a prompt.
- **What happens when the user dismisses the prompt, or the turn aborts mid-ask?** Dismissal maps to `cancelled` with its own deny text. An already-aborted signal settles `cancelled` without dispatching; an abort during the ask discards the late answer — one audit pair either way, never two.
- **What if the client answers with an option the harness never offered?** Any selection other than the offered `allow_once` maps to `rejected` — an unknown optionId from a non-conforming client can never grant.
- **How do subagents' approvals route?** An agent no answerer owns delegates through the whole waterfall and fails closed — in-process subagents are deliberately unanswerable. `subagent-acp`'s child-side auto-answer is separate; routing a child's asks to the parent's editor is deferred (§ Deferred).
- **What does `policy: 'never'` actually change at runtime?** The service resolves every ask for that session to `rejected` before dispatching any answerer (in-service, so no registration order can bypass it); the system prompt states the policy; switches are narrated at boundaries; the audit pair still lands for every auto-rejection.
- **What happens across a hot reload, or when the UI plugin unloads mid-session?** Answerers dispose with their owning fiber, so the next ask degrades to `unavailable` instead of hanging on a dead channel; remounting re-registers the answerer with no catch-up state.
- **Where does the user see what they are approving?** On the tool call itself: the prompt attaches to the already-streamed call via `callId` — arguments included — and adds the asker's human-readable `reason`; the request carries no argument copy of its own.
## Prior art
In-repo precedents this design copies or contrasts with:
- The `fs/write-intent` gate (`packages/fs/fs/`) — the documented single-occupancy decision-slot waterfall semantics (first answer wins, delegate via `next()`) the answerer contract reuses.
- `hook/invoked`/`hook/result` — the log-only audit-pair precedent `approval/asked`/`approval/decided` follows; [the hook-bridges RFC](2026-06-30-hook-bridges.md) ships `permissionDecision: ask`, the first producer.
- [The interception-seams RFC](2026-06-30-interception-seams.md) — the `tools/pre-execute` `allow`/`deny`/`ask` vocabulary whose `ask` this seam services.
- [The ACP support RFC](../../implemented/feature/2026-06-14-acp-agent-client-protocol.md) — the `WeakMap<Agent, sessionId>` ownership seam the answerer routes through; [the multi-session RFC](../../implemented/feature/2026-06-14-acp-multi-session.md) — the per-session permission-ownership blocker this implements.
- The opportunistic `ctx.get()` consumption pattern (`tool-bash`'s owner-token lookup, the loop's persistence probe) — how `dsh-tools` consumes the seam without gating its fiber on it.

6
docs/rfc/implemented/feature/2026-07-06-explicit-tool-order.md
View File

@@ -17,7 +17,7 @@ The system-prompt assembly owns the canonical model-facing tool order, exactly w
- The list must contain the rest entry exactly once and no duplicate names.
- When `toolOrder` is unset, the canonical order is plain lexicographic name order (code-unit comparison, locale-independent), so determinism requires no configuration.
The policy is applied where the list is born: `assemble()`, before the `system-prompt/assemble` waterfall. The assembly canonicalizes the tools it collects from providers the same way it sorts sections by their `order` field — on the initial assembly, killing the registration-order entropy at its source. Everything downstream inherits the order untouched: the waterfall, the loop's `EpochHeader`, the `request/header` event, the deep-frozen request, and the dev invariant's cross-check all see one deterministic list, with no new loop change.
The policy is applied where the list is born: `assemble()`, before the `system-prompt/assemble` waterfall. The assembly canonicalizes the tools it collects from providers the same way it sorts sections by their `order` field — on the initial assembly, killing the registration-order entropy at its source. The waterfall therefore starts from one deterministic list; when a listener leaves that order intact, the loop's `EpochHeader`, the `request/header` event, the deep-frozen request, and the dev invariant's cross-check inherit it with no new loop change.
Scope is deliberately narrow: this fixes the REGISTRATION-ORDER race, not plugin behavior. A `system-prompt/assemble` listener may still add, remove, or rearrange tools — same as it may edit sections after their sort — and owns the determinism of what it emits; the waterfall contract already demands deterministic listeners (the reconstructability invariant would catch a listener that diverges between build and replay).
@@ -36,8 +36,8 @@ Config plumbing follows the `persona` precedent, and `toolOrder` sits beside it:
## Consequences
- Every assembly — and therefore every `request/header` event and model request — has a deterministic tool order on every host; the CI-vs-local golden flip is structurally gone. The default order is lexicographic, no longer registration order.
- `PromptAssembly.tools` itself is canonical, so every assembly consumer (the loop, waterfall listeners, any future prompt inspector) sees the model-facing order; provider registration order is observable nowhere downstream of the registry.
- Every registry-built assembly starts with a deterministic tool order on every host; absent an expert listener that deliberately changes it, every `request/header` event and model request inherits that order. The CI-vs-local registration-order flip is structurally gone, and the default is lexicographic.
- The initial `PromptAssembly.tools` is canonical, so waterfall listeners start from the model-facing order; provider registration order is observable nowhere before that cooperative seam.
- The snapshot suite's single pinned request-header fixture (`text-turn`) carries the new canonical tool order; every other ACP snapshot keeps the header bulk scrubbed as `{{system}}`/`{{tools}}`, per the pinned-header design.
- A pure tool reordering between steps is representable only as a `request/header` `'fallback'` snapshot (the name-keyed `ToolsDelta` cannot express it); with a stable canonical order such reorders no longer occur in practice, so the fallback path stays a safety valve.
- The `toolOrder` key rides the app → `agent-core` → `SystemPrompt` forwarding chain, so deployments set it next to `persona` in the app config; `dsh-llm` and the agent loop are untouched.

225
docs/rfc/implemented/feature/2026-07-06-sandbox.md Normal file
View File

@@ -0,0 +1,225 @@
# RFC: The subprocess sandbox — confinement seam, native runners, escalation, and per-session modes
Status: implemented
## Problem
A coding agent needs this product path: bash subprocesses — and the hook commands that ride them — execute under a restricted file sandbox by default; if and only if the sandbox actually denies an operation, the model may request one user approval for that same operation and, once granted, retry it once with wider permissions. An every-tool boundary is deliberately NOT the claim: fs/web/todo execute in-process where an `execve` wrapper is meaningless (§ In-process tools), and the cross-family boundary is staged follow-up work (§ Deferred phases). Without a shared vocabulary, every tool reinvents approval fields, denial parsing, retry matching, and permission-state hints.
The harness is an SDK, so confinement must be a capability developers COMPOSE: whether to sandbox, and which backend per platform, belongs in the leaf `cordis.yml` as a first-class entry — not inside one executor's private machinery. And the first-choice runner, `bwrap`, is unusable on exactly the hosts a sandbox matters most (minimal containers, disabled unprivileged userns, LSMs that deny `mount`), so a fallback runner has to ship with the SDK rather than be assumed on the host.
Confinement alone leaves two gaps. A denial with no escalation path is terminal — the model can only give up, which pressure-cooks operators into configuring `workspace-write` or `danger-full-access` globally and defeats the sandbox. And the model-visible knobs (the sandbox mode, the approval policy) change over an agent's lifetime — an ACP user flips a per-session setting, an operator edits `cordis.yml` while the process is down — while the model must never act on a stale belief about them: what IS the state on every request, what changed while the agent lives, and what changed while nobody was watching all need answers.
## Decision
One seam, one per-platform chain of local backends, one consumer, and two levers on top: a per-call escalation path and per-session runtime modes. Everything below composes from the leaf `cordis.yml`; nothing touches `agent-loop`. The scope is deliberately bounded: the phases this RFC names but does not design — per-session workspace root, cross-family fs enforcement, the `subagent-acp` consumer, more environments, a Windows chain — are listed under § Deferred phases, each a follow-up design, not a config knob.
### How a deployment uses it
Three `cordis.yml` entries turn an unconfined coding agent into the sandboxed product path; [`examples/sandbox-acp-agent`](../../../../examples/sandbox-acp-agent/README.md) is this composition, live:
```yaml
- id: sandbox
name: '@deepseek-ai/dsh-sandbox-local' # the per-platform runner provider (ctx.sandbox)
- id: bash
name: '@deepseek-ai/dsh-bash-sandbox' # the confined executor, replacing dsh-bash-local behind ctx.bash
config:
mode: read-only # the deployment default every session starts from
workspaceRoot: !!js process.cwd() # the boundary workspace-write may write under
- id: approval
name: '@deepseek-ai/dsh-user-approval' # the escalation gate's channel (the approval RFC)
```
The swap is invisible to every consumer of `ctx.bash`: the bash tools, hook commands, and background tasks run exactly as before, spawned through the wrapped argv the provider returns. Deleting the `sandbox` and `bash` entries and loading `@deepseek-ai/dsh-bash-local` instead is the opt-out — execution is unconfined again and the escalation fields vanish from the tool schema, because they are capability-gated on the mounted executor, not on configuration. Omitting only `approval` keeps confinement but fails every escalation closed with its own error text.
Misconfiguration fails loud: `mode` outside the closed vocabulary is rejected at plugin load, and a host with no usable backend throws the structured `SANDBOX_UNAVAILABLE` — at `confine()` before the command ever spawns — rather than degrading to unconfined execution. `runnerCommand` on `dsh-sandbox-local` is the operator's explicit assertion of a bwrap-compatible runner (chain and probes skipped); it doubles as the deterministic fake-runner seam for keyless tests.
What the model then experiences: denied file effects come back as result facts with a `[sandbox: file access denied under <mode> mode]` marker plus standing instructions not to retry around them; under a confining executor the schema offers `sandbox_permissions` + `justification` for the one-approval escalated retry (validated strictly wider than the session's effective mode at execution); the system prompt deliberately does NOT state the sandbox mode — the model learns the boundary from the marker (which names the mode) when it hits it, instead of preemptively refusing work a standing declaration discourages. What an ACP editor experiences: a `sandbox-mode` and an `approval-policy` config-option select per session (each advertised only when its knob is composable), switchable at runtime; a sandbox switch simply changes what subsequent commands may do, while an approval-policy switch to `'never'` is stated in the prompt and narrated.
The product path, concretely (the escalation arc is verbatim from the recorded `escalation-approved` scenario; the denial leg is pinned on the real-kernel e2e tier):
```
tool/result … [sandbox: file access denied under read-only mode] ← the write RAN; the kernel refused it
tool/call bash {"command": "printf 'escalated\n' > escalated.txt && cat escalated.txt",
"sandbox_permissions": "workspace-write",
"justification": "the user asked to write escalated.txt in the workspace"}
→ the editor is prompted on this very call (session/request_permission through the approval seam); Allow once
tool/result "escalated" — THIS call ran under workspace-write and its result facts say so; the session stays read-only
```
Reject instead and nothing executes: the result is the verbatim `the user rejected escalating this command to "workspace-write"`, and the teaching makes that final — no re-ask.
### Design detail
#### Grounding — verified against the code
- Runtime OS subprocesses exist at exactly two sites: the `ctx.bash` seam's single spawn (`packages/bash/bash-local/src/run.ts`; hook commands flow through `ctx.bash`, so bash confinement covers them transitively) and `subagent-acp`'s child agents (`packages/subagent/subagent-acp/src/run.ts`) — the second consumer that makes a shared seam due rather than preemptive under the [capability seams RFC](../architecture/2026-06-13-capability-seams.md)'s "don't split preemptively" rule.
- Everything else executes inside the harness process (fs is in-process `node:fs`, web is in-process `fetch`, every `ToolDefinition.execute()` closes over `ctx`): an OS sandbox wraps `execve` and cannot wrap an in-process function call, so "sandbox any tool" is policy at each tool's seam, never a mechanical transport change.
- `tools/pre-execute` (`allow`/`deny`/`ask`) exists, with `ask` serviced by [the approval seam](2026-07-06-approval-seam.md); the fs intent gates are version guards with no mode input yet.
- `dsh-bash`'s request/spec split (`BashExecRequest` → `resolve()` → `BashExecSpec`) carries per-call fields the way escalation needs — `owner` is the template: request-optional, spec required-but-nullable, carried verbatim — and the result types already speak `SandboxMode`, so a per-call policy field adds no dependency edge.
- The pinned-header snapshot design means a schema/description change churns at most one pinning fixture per suite, and the escalation fields are advertised only under a sandboxing executor — so they live in exactly one pinned header, the sandbox example suite's `mode-switching` fixture.
#### The seam: `ctx.sandbox`
`dsh-sandbox` owns the vocabulary and the `SandboxProvider` contract: `confine(argv, policy)` returns the argv to spawn INSTEAD of the caller's own — wrapped so the process and everything it spawns run confined — plus the `enforcement` completeness the selected backend achieves, its denial dialect (`denialSignatures`, the stderr substrings that backend's kernel prints on a denied file effect), and its runner-failure dialect (`runnerFailureSignatures`, how the runner ITSELF failing — and therefore the command never running — identifies itself); with no usable backend it throws the fail-closed `SANDBOX_UNAVAILABLE` error, never a silent unconfined passthrough. The vocabulary: `SandboxMode` (`read-only` / `workspace-write` / `danger-full-access`, FILE effects only — network and process visibility are not claimed), `SandboxEnforcement` (`full` / `partial`), `SandboxPolicy` (mode + workspace root).
Policy rides each CALL, not the provider: two consumers may confine under different policies at the same instant (bash under `read-only` while a confined child agent keeps its state directory writable), and an approved escalated retry is a new call with a wider policy — inexpressible under a config-fixed provider mode.
The seam confines SAME-WORLD subprocesses only: a backend shares the host's filesystem and kernel. Containers, microVMs, and remote executors are NOT backends of this seam — they replace whole capability implementations (`ctx.bash`, `ctx.fs`) as environment-coherent groups, because an agent whose bash runs in a container while its fs tools write the host lives in two split worlds.
Left open, for the phase that needs them: whether network restriction arrives as a separate `network_mode` or merges into `sandbox_mode` once a runner enforces both, and whether `SandboxPolicy` grows extra writable-root grants now (the launcher already speaks `--rw <path>`) or only when escalation needs them.
#### Local backends and the shipped launcher
`dsh-sandbox-local` selects BY PLATFORM, once per lifetime, and caches the verdict: each platform names its runner chain, a chain of one is selected directly — probing arbitrates between candidates, and a sole candidate leaves nothing to arbitrate — and a chain of several is probed FUNCTIONALLY in preference order (build and enforce a real profile, never `--version` — a present-but-unusable `bwrap` must fail its probe). Linux: `bwrap` first (its mount profile is closest to the mode vocabulary: whole tree read-only, fresh `/dev`+`/proc`, `workspace-write` adds an ephemeral `/tmp` and rebinds the workspace root; deliberately no `--unshare-pid` and no network claim), else the npm-distributed `landlock-run` Landlock launcher. darwin: `sandbox-exec` speaking a Seatbelt (SBPL) profile — allow-default with `(deny file-write*)` plus write allow-lists, every granted root canonicalized because Seatbelt matches resolved paths (`/tmp` IS `/private/tmp`) — unprobed, the sole candidate. A platform with no chain fails closed at `confine()`; an unprobed runner that turns out unusable fails closed at EXECUTION instead — it refuses to run the command, and every wrap carries `runnerFailureSignatures` (the runner's own error prefix, which also matches the shell's runner-not-found message) so the consumer classifies that as a SANDBOX failure, never a task failure: on either path the command neither runs unconfined nor slips through as a plain failure. A non-empty `runnerCommand` config is the operator's assertion of a runner that fully enforces the bwrap-shaped profile — chain and probes skipped; it doubles as the deterministic fake-runner seam for keyless tests. It is not exempt from fail-closed execution: its wrap carries argv0-scoped outer-shell failure shapes (`exec: <argv0>: not found`, `<argv0>: No such file or directory`, `<argv0>: Permission denied`) as its runner-failure dialect, so a missing or unexecutable configured runner classifies as a sandbox failure like every other rung — never as a failing command, and never as a denial.
The launcher is a ~300-line C program (plain C11 over the raw Landlock UAPI — no libraries beyond a statically linked musl, so the audit surface is that one file plus the kernel's stable syscall contract): `--ro <path>` / `--rw <path>` grants, `--`, the wrapped argv; it installs the ruleset on itself and `exec`s (rulesets are inherited across `execve`, and it sets `no_new_privs` before restricting); `--probe` enforces a maximal ruleset in a short-lived child and exits 0 only when the kernel actually enforces; launcher failures exit 125 without exec'ing.
The launcher lives in its own repository and reaches the harness as the npm package family [`node-addon-landlock-run`](https://www.npmjs.com/package/node-addon-landlock-run) (the per-platform-package pattern of `node-addon-require-builtin` and esbuild): an entry package — `dsh-sandbox-local`'s one runtime dependency — plus per-platform binary packages selected at install time by npm's `os`/`cpu` fields. The entry package owns the launcher's CLI contract end to end (`launcherPath()` resolution with a never-existing fallback, the functional `probe()`, `grantArgs()` flag spelling), versioned together with the binary so probe-report parsing can never drift against it; the harness keeps only the policy side, `landlockProfileArgs()` mapping the mode vocabulary to grants. Native-only per-architecture builds, pack gates (binary presence, executability, ELF architecture), and the byte-pinned publish rehearsal are that repository's release pipeline; this repo's Landlock CI legs install the published family from the registry — the true consumer path — and prove real-kernel confinement through it.
FIXME: Revisit the separate-repository boundary and try to maintain the launcher source and its platform package family inside this monorepo, so the native release surface and harness contract evolve together.
Profile parity is honest rather than identical: under Landlock, `read-only` grants `--ro /` plus `--rw /dev/null` (the node, not `/dev` — the host's `/dev/shm` is a persistent shared tmpfs), and `workspace-write` grants the HOST `/tmp` where bwrap's is ephemeral; under Seatbelt, `read-only` likewise grants only the `/dev/null` literal, and `workspace-write` grants the host `/tmp` plus the per-user darwin temp dir (`os.tmpdir()` — the platform's real temp area for mkstemp-family tools; omitting it would deny what the mode promises). Every wrap carries the rung's denial dialect (`denialSignatures`: EROFS text under bwrap, EACCES under Landlock, EPERM under Seatbelt) so consumers match the active backend rather than a cross-runner union. Enforcement is honest per ABI level: an older kernel enforces the subset its ABI governs (path truncate is ungoverned before ABI v3), the probe's report line distinguishes the cases, and every confined result carries the structured `enforcement: 'full' | 'partial'` fact — refusing partial enforcement would deny the fallback to precisely the older-kernel hosts that need it. The bwrap and Seatbelt profiles govern every promised file effect by construction, so their passing probes always report `full`.
#### The bash consumer
`dsh-bash-sandbox` extends `LocalBashExecutor` (spawn mechanics, process-group kills, spill files, background tasks, credential scrub inherited verbatim) and hands `ctx.sandbox` the exact `['bash', '-c', command]` argv it is about to spawn. A sandbox denial is a RESULT FACT, not an error: the command RAN and the kernel refused a file operation, so `result.sandbox.denied` is orthogonal to `exitCode`/`signal`. Classification is conservative text inference over the collected stderr tail against the WRAP's own dialect, so a backend is never credited with a denial text its kernel does not speak (bare EPERM under a Linux runner names non-file boundaries the mode vocabulary does not govern); the known residual false positive is non-sandbox text in the active dialect (an ssh auth failure under Landlock, a refused `kill` under Seatbelt), and a structured runner signal wins once one exists. A RUNNER failure is the opposite of a denial and outranks it in classification (a runner's error text can itself contain denial words): the wrap's `runnerFailureSignatures` matching a failed run means the sandbox broke and the command NEVER RAN — the foreground path re-throws it as the structured `SANDBOX_UNAVAILABLE` error (the late twin of the confine-time throw, carrying the runner's first stderr line), a settled background task stamps `sandbox.runnerFailed` and `bash_output` renders its own marker — so a broken sandbox can never read as a failing command.
The model's view is result facts only: the static tool description explains the denial marker (`[sandbox: file access denied under <mode> mode]`), encourages attempting commands that may be denied, and forbids retrying around a denial; when the escalation fields are advertised, a denied result additionally carries the escalation hint itself, so the sanctioned same-turn retry is prompted at the decision point rather than depending on the model recalling the description (§ Escalation). No prompt section states the sandbox mode (§ Per-session modes).
#### Escalation: one approved wider retry after a denial
The seam level is mechanism only. `BashExecRequest` carries `sandboxMode?: SandboxMode`, an explicit per-call policy input; `BashExecSpec` carries it required-but-nullable (the `owner` pattern: a forgotten field is a visible `undefined`, and `resolve()` is the one explicit defaulting step); `BashExecutor` exposes the capability fact `get sandboxMode(): SandboxMode | undefined` — `undefined` in the base class, the configured mode in `SandboxBashExecutor` — so the tool layer can advertise only what the mounted executor honors: composition truth, not configuration. The seam honors ANY explicit mode, including a narrower one; the wider-only ladder is escalation policy and lives in the tool. A non-sandboxing executor (`dsh-bash-local`) carries the field verbatim and confines nothing — the field reaching it means the caller bypassed the tool's gate, and its honest behavior stays unconfined execution, not a guess at enforcement it does not have.
`SandboxBashExecutor.resolve()` stamps the effective mode — escalation grant > session override > configured default — so `run()`/`start()` read the spec, never the config. The `danger-full-access` branch, the confine call, and the result facts all key off the spec's mode, and the per-task facts map carries each task's mode alongside its wrap facts (`notifyTaskDone()` stamps from the map entry): one escalated call — foreground or background — reports the mode it ACTUALLY ran under while every neighbor keeps its own.
The tool gate advertises two extra parameters exactly when `ctx.bash.sandboxMode` reports a confining mode at registration: `sandbox_permissions`, an enum of the closed escalation-target vocabulary — `workspace-write`/`danger-full-access`, every mode a session could ever escalate TO — and `justification`, required together with it. The enum is deliberately NOT cut down to the modes wider than the executor's DEFAULT: schemas are registry-global while the effective mode is per-session and switchable, so a default-relative ladder strands a session overridden NARROWER than the default (with a `danger-full-access` default and a `read-only` override it would advertise nothing at all — confined, but with no lever). Strict widening is instead enforced at EXECUTION against the call's effective mode (session override ?? executor default): a request that is not strictly wider fails closed with its own text and prompts no one. An escalating call resolves approval BEFORE anything executes — no `ctx.approval` composed, or no agent on the execution, fails closed with its own text; otherwise `ctx.approval.request({ agent, toolName: 'bash', callId, reason, signal })` with the audit-self-contained reason `escalate sandbox to ${mode}: ${justification}`, while the UI attaches the prompt to the already-streamed call (the command is visible there; the approval RFC's no-arguments rule holds). The four outcomes map to distinct results: `allowed-once` stamps `sandboxMode` onto the bash request and proceeds; `rejected`, `cancelled`, and `unavailable` each produce their own error text, so the model can tell a human "no" from a dismissed prompt from a missing channel. The grant is consumed by the very call that asked; nothing is stored.
The tool description teaches — and a denied result itself prompts — the SAME-TURN flow when the fields exist: on a denial a wider mode would cure, escalate immediately in that turn by retrying the exact command once with `sandbox_permissions` (the narrowest mode that suffices) + `justification`, without detouring through chat to ask first — the approval prompt raised by the retry IS how the user consents. Never speculatively: an escalation is grounded in a real denial — normally the one the command just hit, up front only when the session already denied the same access — and a prompt stating approvals are disabled turns the exception off entirely; a rejected escalation is final for that command. Denial-grounding is deliberately model discipline plus human judgment, not harness bookkeeping — the human sees the exact command and justification on the prompt (see Alternatives for why hard-matching is rejected). No new session events anywhere: the attempt is an ordinary `tool/call` whose logged arguments carry the two fields, the decision is the approval seam's `approval/asked`/`approval/decided` pair, the outcome is an ordinary `tool/result` whose sandbox facts name the mode it ran under. The asker lives in `dsh-tool-bash`, NOT the executor: a transport seam has no `agent`, no `callId`, and no business asking humans questions.
Left open, recorded for the phase that picks them up: what a grant's scope identity is beyond the sandbox mode — the exact call, a path, a command prefix, the session, a time window — the question `allow_always` grant storage must answer before that option can be advertised; how cancellation behaves while an approval prompt is pending; and how escalation is defined for `run_in_background` denials that arrive via `bash_output`.
#### Per-session modes: the session log as the store
```
effective(session) = findLast(the session's own knob events)?.value ?? the composition-config default
```
The default is composition config (`cordis.yml`) — operator-owned, process-wide. A runtime switch is a SESSION-SCOPED override recorded as one log-only event in that session's own log. Restart immunity (resuming a session replays its log, so overrides come back with zero catch-up machinery) and multi-session isolation (one editor tab's `workspace-write` cannot disturb another's `read-only`) both fall out by construction, and no external config store exists anywhere.
**One event per knob, owned by its domain** — the merge-extensible `SessionEventMap` idiom every existing event family already follows (`approval/*` in `dsh-user-approval`, `hook/*` in the hooks packages):
```ts
interface SessionEventMap {
'bash/sandbox-mode': { mode: 'read-only' | 'workspace-write' | 'danger-full-access' }
'approval/policy': { policy: 'ask' | 'never' }
}
```
Each owner exports the same three-piece kit: the event declaration, a pure fold (`effectiveSandboxMode(events)` / `effectiveApprovalPolicy(events)` — a `findLast`, typed to the domain's closed union), and THE write path (`setSandboxMode(session, mode)` / `setApprovalPolicy(session, policy)` — a switch IS its event; nothing mutates state out of band). No shared owner service, no generic facts map, no registry: a third knob copies the ~40-line pattern into its own package. Execution follows the fold on both sides — the bash tool's per-call stamp reads it as the middle rung of the § Escalation precedence chain, and the approval seam's `'never'` gate is [the approval RFC](2026-07-06-approval-seam.md)'s side of the same pattern.
**Visibility is deliberately asymmetric between the knobs.** The SANDBOX mode is stated nowhere and its switches are not narrated: a standing "you are read-only" declaration teaches the model to refuse preemptively (observed live: sessions where the model would not even attempt a write it could have escalated), while the denial marker already names the mode the command ran under at exactly the moment the boundary matters — behavior, not belief, carries the state, and a switch simply changes what the next command does. The APPROVAL policy keeps both layers, because its failure mode is the opposite: an auto-rejected ask under `'never'` returns "the user rejected …" wording no behavior can disambiguate, so the prompt states `'never'` (and ONLY `'never'` — an `'ask'` promise is unknowable without asking, and absence under a logged header is how the narrator reads `'ask'` back), and an `agent/pre-step` narrator injects at most one coalesced notice per policy switch: idle flip-flops collapse to one notice at the next turn's first step, a net-zero round trip to none, and a mid-turn change is narrated no later than the next step. Its "last told" is in-memory with a log-derived fallback (the folded header's system text parsed against the closed candidate sentence; LAST occurrence wins, so a persona quoting it cannot shadow the real section), so restarts lose nothing; attribution is positional (a knob event after the log's last `request/header*` reads `changed by the user`, a drift with no such event reads `changed by the operator/config`).
**The editor surface** is protocol-native [Session Config Options](https://agentclientprotocol.com/protocol/session-config-options) — the spec's replacement for session modes (slated for removal in ACP v2), already SDK-typed. The bridge advertises one independent `select` per composable knob — `sandbox-mode` (category `mode`) iff the mounted executor confines, `approval-policy` iff the approval seam is composed — with `currentValue` folded from each session's own log, in `session/new` and `session/load` responses. `session/set_config_option` validates against the same closed lists, routes to the domain setter, and returns the complete refreshed state (the spec contract).
**Anchoring: turn-enclosure is the commit boundary.** The turn-enclosure contract makes a bare between-turns append invalid (the JSONL backend treats a post-`turn/end` tail as crash garbage; dev invariants throw). A switch while a turn is open appends immediately — openness read from the LOG (last boundary event is `turn/start`), not `agent.status`, which stays `running` between queued turns. An idle switch is held on the bridge's session record and anchored at the next turn's `agent/prompt-submit` — inside the turn, before anything in it assembles or executes, last write per knob, and OUTSIDE any `session/event` emit (appending from inside that feed reorders events for later-registered listeners — a bug the dev invariants caught live). Until anchored, the switch exists only in bridge memory: responses overlay it truthfully, and a crash before the next turn reverts it — `session/load` then reports the fold's truth, so the editor UI self-corrects rather than lies.
#### In-process tools
fs/web/todo execute in-process, so their sandbox semantics are policy at their seams: the fs intent gates deciding by the shared mode vocabulary (§ Deferred phases, cross-family) make `read-only` a real boundary instead of a bash-only approximation — until then the contract says so honestly. No generic per-tool sandbox runtime: a host-mediated tool leaves the process only by returning declarative effects the host validates, which is a rewrite, not a wrapper.
FIXME: Revisit this tool-local boundary. The follow-up design needs to determine whether sandboxing becomes a global harness capability that applies uniformly to every tool, instead of expressing in-process enforcement independently at each tool seam.
### Testing
- Unit tier (no real runner anywhere): profile dialects, per-platform chain selection (sole candidate unprobed, no chain fails closed, multi-candidate probe order), verdict caching, the fail-closed end, probe-report parsing, and the launcher/`sandbox-exec` CLI contracts via fake runner scripts in `dsh-sandbox-local`; wrapping, policy hand-off, fact stamping, and runner-failure-outranks-denial classification (foreground throw, background `runnerFailed` fact) against a fake provider in `dsh-bash-sandbox`; the error's structured identity in `dsh-sandbox`. The escalation matrix spans the three bash packages: verbatim carry-through in `dsh-bash-local`, stamp/branch/per-task-facts in `dsh-bash-sandbox`, and the capability gate, `justification` pairing, fail-closed texts (pinned verbatim), and grant stamping in `dsh-tool-bash`. The switching surface pins the folds, the stamping precedence, the `'never'` gate, per-session section rendering, the full narrator matrix (cold start, coalescing, net-zero, resume drift with operator wording, positional attribution, persona-shadow hardening), and the bridge's advertisement gating, validation rejections, idle-vs-mid-turn anchoring (dev invariants mounted), and `session/load` reporting over a real two-process JSONL round trip.
- Keyless real-runner e2e, split along the seam and per rung: CI's `sandbox-e2e` matrix runs bwrap and Landlock on Linux (the Landlock leg once per architecture, each confining through the registry-installed launcher) and Seatbelt on macOS against real kernels, failing on a silent all-skip. World-proofs live in `dsh-sandbox-local` (denied writes absent on disk, workspace writes landing, temp-area grants pinned, kernel denial text matching the advertised dialect) and `dsh-bash-sandbox` (the through-`ctx.bash` consumer proofs, including denied-then-overridden-write-lands). This package's own publish path is rehearsed without publishing (`packed-install.e2e.ts`): `pnpm pack`, tarballs installed into a throwaway consumer with the launcher family resolving from the registry, plain `node` confining through the INSTALLED launcher — asserted executable apart, so a mode-stripped binary can never masquerade as a non-enforcing kernel. The switching surface has its own keyless e2e (`examples/sandbox-acp-agent`): the real `cordis.yml` tree advertises both options, honors switches end to end, and rejects out-of-vocabulary values.
- With-key e2e (`examples/sandbox-acp-agent/tests/escalation.e2e.ts`): real model + real runner + the REAL bridge answerer, world-verified — denied under `read-only`, escalates with justification, the scripted editor grants and the retried write lands on disk, while a rejected escalation leaves no write. Self-skips without `DEEPSEEK_API_KEY` or a usable runner (e2e.yml installs bubblewrap so it actually executes in CI).
- Snapshot tier (`examples/sandbox-acp-agent/tests/acp.snapshot.ts`): the keyless config-option wire; the recorded mode-switching arc as the suite's pinned header — necessarily, since mid-session switches emit the `request/header-delta`s the uniformity guard licenses only in the pin — committing both switches, the prompt-section delta and one "changed by the user" notice per knob, and a confined write landing under the switched mode; and both recorded escalation branches over scripted `permissionAnswers` (grant runs confined under `workspace-write`; rejection executes nothing and pins the fail-closed text). Replay re-executes every fixture's bash calls under the host's real runner (ci.yml's snapshot lane installs bubblewrap). Deliberately absent: a fixture carrying a real DENIAL — denial stderr is the backend's dialect and would pin a fixture to its recording platform; the escalation prompts assert the prior denial instead, and the denial→marker path stays on the tiers above.
## Deferred phases
Each phase gets its full design when picked up, validated against the code at that time, and lands with unit, real-API e2e, and snapshot coverage at the tiers it touches.
- **Per-session workspace root** — the executor's write boundary stays config-fixed for its lifetime while each ACP session has its own cwd; a per-session root rides the same per-call policy carrier once designed.
- **Cross-family boundary** — the fs intent gates decide by the shared mode, making `read-only`/`workspace-write` real boundaries beyond bash.
- **Second consumer** — `subagent-acp` optionally confines child agents (per-call policy; unconfined default — a child agent must write its own persistence).
- **More environments** — an environment-coherent capability group example (e.g. bash+fs against one container).
- **Windows chain** — `PLATFORM_CHAINS.win32` is reserved and empty (fail-closed); filling it means a confinement runner from the AppContainer/restricted-token family, shipped from its own repository on the `node-addon-landlock-run` template, plus its profile dialect and denial/runner-failure signatures.
## Alternatives considered
- **Command-string heuristic preflight** — rejected: cannot understand expansion/subprocesses/symlinks; the strict attempt (run it, let the kernel decide) is the only trustworthy denial signal.
- **Functionally probe even a platform's sole backend** — rejected: probing arbitrates between candidates; with one there is nothing to decide, and probe cost taxes the first confined command of every session (prohibitive for heavy future backends). The runner's own exec-time fail-closed refusal plus `runnerFailureSignatures` classification carries the safety property instead.
- **Commit the built launcher binaries** — rejected: a binary in a diff is unreviewable and churns history; reviewed source + native CI builds + the launcher repo's byte-pinned publish rehearsal keep bytes out of every tree.
- **Compile the launcher on install** — rejected: pushes a C toolchain onto every consumer; a fallback that exists only where a compiler happens to be is not a fallback.
- **Cross-compile both architectures from one builder** — rejected: requires carrying a pinned cross toolchain (rustup targets, zig, or a container image) solely to rebuild two ~70 KB binaries; per-architecture native runners already exist and each builds its own platform package (the `node-addon-require-builtin` model, the launcher repo's own pipeline).
- **No fallback (bwrap or fail closed)** — rejected: concentrates failure on the hosts a sandbox matters most, degrading to `danger-full-access` by resignation.
- **Keep the mechanism inside `dsh-bash-sandbox`** — rejected: blocks the existing second consumer, makes future phases read mode out of a bash plugin's config, and cannot express escalation.
- **Config-fixed mode on the provider** — rejected: one mode per process; cannot serve concurrent consumers with different policies nor the one-shot widened retry.
- **One interface spanning containers/VMs too** — rejected: `confine(argv)` presupposes a shared filesystem; environment isolation is capability-sibling backends deployed as coherent groups.
- **Generic ToolRuntime wrapping any tool** — rejected: mechanically false for in-process tools (closures over `ctx`); the declarative-effects rewrite is unjustified for fs/web/todo.
- **Ask inside the executor (`dsh-bash-sandbox`)** — rejected: no `agent` to route through, no `callId` to attach the prompt to; adding them teaches a transport seam about sessions and UIs — the tool layer holds both and owns the model-facing vocabulary.
- **Auto-retry inside the same tool call** — rejected: a hidden re-entry the log cannot reconstruct: one `tool/call` would have produced two executions with different policies — the retry is a NEW logged call with its own arguments and result facts.
- **Advertise the escalation fields unconditionally** — rejected: under `dsh-bash-local` they are a dead lever — advertising an option the harness cannot honor manufactures doomed grants; capability-gating costs one registration-time read.
- **A default-relative escalation ladder (advertise only the modes wider than the executor's registration-time default)** — rejected: per-session overrides make the default the wrong baseline — a session switched narrower than the default loses exactly the lever it needs, and under a `danger-full-access` default the fields vanish entirely while a `read-only`-overridden session stays confined with no escalation path. The enum pins the closed target vocabulary; strict widening is a per-call execution check against the session's effective mode.
- **Per-session dynamic tool schemas** — rejected: schemas are registry-global by design (one assembly vocabulary, the pinned-header snapshot contract), and re-registering per session would buy only what the execution-time strict-wider check already guarantees, at the cost of a per-session schema surface and header churn on every switch.
- **Hard-match the retry to a prior denial** — rejected: command-string identity is fragile (quoting, `workdir`, env prefixes, a pipeline retried as its failing stage) — false-rejects honest retries or is trivially satisfied; the real boundary is the human seeing command + justification. Revisit only if `allow_always` grant storage ever needs machine-checkable scopes.
- **A generic `env/state` facts map with an owner service** — rejected: approval and sandbox compose independently, so neither's state may drag in a third package; single-key folds are one `findLast` each, dissolving the owner service; no invariant spans the knobs, so atomic multi-key patches bought nothing.
- **Narrate via `agent/user-message` + a bus event** — rejected: it presupposes a turn-entry seam that does not exist (the real seam is `agent/prompt-submit`), and pre-step's position serves both the coalesced turn-entry notice and the mid-turn immediacy bound with one listener.
- **A standing prompt statement of the sandbox mode (+ a switch narrator)** — shipped first, then removed on live evidence: with `Bash commands run under the "read-only" file sandbox.` in every request, the model refused to ATTEMPT denied-then-escalatable work (five of twelve turns in the first manual session ended with zero tool calls), turning the sandbox into a soft lockout. The denial marker names the mode at the moment it matters and the escalation fields carry the recovery; the approval knob keeps its statement because an auto-rejection is behaviorally indistinguishable from a human "no".
- **Track "last told" with its own bookkeeping events** — rejected: the `request/header*` fold already records the exact prompt the model saw; parsing the closed candidate sentences back replaces a second bookkeeping stream — events are needed only where they ARE the store.
- **ACP session modes instead of config options** — rejected: one mode list cannot carry two orthogonal knobs; config options are the spec's designed surface and modes are slated for removal in ACP v2.
## Consequences
What shipped pins — the tiers in Testing hold each:
- A denied command retried with `sandbox_permissions` + `justification` prompts the user through the composed answerer chain; a grant runs THAT call under the wider mode (result facts say so) while every other call keeps its own effective mode; every non-grant outcome produces its distinct error text and executes nothing.
- The escalation fields exist exactly when the mounted executor confines; a request that is not strictly wider than the call's effective mode fails closed with its own text and prompts no one; a deployment with no ApprovalService fails escalating calls closed and leaves plain calls untouched.
- The system prompt never states the sandbox mode (an approval `'never'` policy is the one stated knob), and the whole exchange — headers, knob events, notices, approvals, results — reconstructs from the session log alone, with no event types beyond the two knob events.
- N idle-time flips produce at most one anchored event per knob (a net-zero sequence anchors none — a no-op push from a client echoing current selections records nothing); an approval-policy switch is narrated in at most one coalesced notice; a mid-turn sandbox switch is honored by the next call's stamp.
- A resumed session's overrides apply and are reported to the editor with no special-casing; a default changed while the process was down is narrated before the session's first new request, attributed to the operator.
- Two concurrent sessions never see each other's state, notices, or config options.
- `agent-loop` is untouched — everything rides `systemPrompt.section`, `SessionEventMap` merging, `agent.inject()`, `agent/pre-step`, `agent/prompt-submit`, and the ACP handler surface.
Costs and accepted limits:
- **The one-wrapper illusion is given up knowingly.** A `tools/pre-execute` wrapper plus prompt conventions does not solve sandbox approval — the correct design costs structured denials, native runner probes, per-call policy carriage, and consistent cross-family enforcement, and this design pays it.
- **`read-only` is not yet a cross-family boundary.** Until the fs intent gates decide by the shared mode, the claim holds for bash only; the contract says so honestly (§ In-process tools).
- **Windows has no backend.** Its chain slot is reserved empty — fail-closed, never a fallthrough; filling it is a deferred phase.
- **The Seatbelt rung leans on Apple's deprecated-but-shipped `sandbox-exec` CLI.** As darwin's sole candidate it is selected without probing, so a future removal surfaces at execution as the runner-failure classification — re-thrown `SANDBOX_UNAVAILABLE`, the command never runs; fail closed, never open.
- **Landlock confinement is only as complete as the running kernel's ABI.** Reported as `enforcement: 'partial'` rather than refused — the deliberate trade that keeps the fallback available on older-kernel hosts.
- **The launcher arrives as a registry dependency.** Trusted through its own repository's release pipeline (reviewed C source, native CI builders, byte-pinned publish rehearsal) plus this repo's version pin — the real-kernel e2e legs are what vouch for behavior through the installed bytes.
- **The model may over-ask.** Escalating without denial grounding, or picking `danger-full-access` where `workspace-write` suffices: the description steers and the enum forces the ladder, but the human prompt is the actual gate; the `approval/asked` reasons make over-asking auditable, and a `prepend` policy answerer can auto-reject patterns a deployment never wants.
- **The advertised target set is static while the effective mode is per-session** (schemas are registry-global) — a session already at the widest mode is still offered the fields. Harmless by construction: the strict-wider check at execution, not the enum, is the safety boundary — a non-widening request fails with its own text and never prompts anyone.
- **A granted escalation is not a working sandbox.** An unavailable backend still fails closed even for a granted escalation to a confining mode — at `confine()` when the platform has no chain or every probe fails, at execution when an unprobed sole runner refuses (classified as a sandbox failure, not a command failure) — while a granted `danger-full-access` run never touches the provider at all: there the grant, not the probe, is the authority.
- **An idle switch lives in bridge memory until the next turn anchors it.** A crash in that window reverts it (reported honestly on `session/load`), and a session that never runs another turn never persists it — accepted, with the loop-owned idle commit turn named as future work if durability becomes a requirement.
- **The approval narrator's restart baseline parses prompt prose.** The closed candidate sentence is owned by the writing module itself, so a wording change is a coordinated writer+parser edit in one file; a session whose headers predate the section silently adopts the current policy without a notice.
- **The approval section is still a dynamic prompt surface** (a `'never'` switch breaks provider prompt-prefix caching for that session). Accepted: policy switches are rare, and a model acting on a stale `'never'` is worse. The sandbox knob no longer touches the prompt at all.
- **The model may hold a stale belief about the sandbox mode** (nothing announces a switch). Accepted deliberately: the next attempt's marker or success corrects it, and the observed failure mode of announcing — preemptive refusal — is worse than one wasted retry.
## FAQ
Behavioral and usage questions only — every "why not X?" design question lives in [Alternatives considered](#alternatives-considered), whose job is exactly that.
- **A command came back with `[sandbox: file access denied under read-only mode]` — did it fail?** It RAN, and the kernel refused a file effect: the denial is a result fact orthogonal to exit code. The teaching forbids retrying around it; the one sanctioned move is the same command retried once with an escalation request.
- **How is a BROKEN sandbox told apart from a failing command?** Runner failure outranks denial in classification: a failed run matching the wrap's `runnerFailureSignatures` means the command NEVER ran — foreground re-throws the structured `SANDBOX_UNAVAILABLE` with the runner's stderr line, a background task stamps `sandbox.runnerFailed` and renders its own marker. A broken sandbox can never read as a failing command, and the command never runs unconfined.
- **What happens on a platform with no backend — Windows today?** `confine()` throws the fail-closed `SANDBOX_UNAVAILABLE` and the command never spawns; `win32` is a reserved EMPTY chain, pinned by test to fail closed identically until a Windows runner fills it (§ Deferred phases).
- **`bwrap` is installed on my host but unusable (disabled unprivileged userns, an LSM denying `mount`) — what happens?** The chain probe is functional — it builds and enforces a real profile rather than checking `--version` — so a present-but-unusable `bwrap` fails its probe, selection falls to the registry-installed Landlock launcher, and the verdict is cached for the provider's lifetime.
- **Does the sandbox restrict network or process visibility?** No — `SandboxMode` claims FILE effects only; the bwrap profile deliberately does not unshare pid, and no backend claims network. Whether network restriction becomes its own knob is left open in § The seam.
- **Which tools actually run confined?** OS subprocesses through `ctx.bash` — the bash tools, and hook commands transitively. fs/web/todo execute in-process, where an `execve` wrapper is mechanically meaningless; their `read-only` semantics arrive with the cross-family deferred phase, and until then the contract says bash-only honestly.
- **Does a granted escalation persist, or cover background tasks?** Neither: the grant is consumed by the very call that asked (foreground or background), that one call reports the mode it actually ran under, and every neighbor keeps its own. How escalation should be DEFINED for a background denial that only surfaces later via `bash_output` is left open in § Escalation.
- **When does an editor's mode switch take effect?** Mid-turn: appended immediately, honored by the very next call's stamp. Idle: held on the bridge's session record, anchored at the next turn's `agent/prompt-submit`, with N flips coalescing to at most one event (none if net-zero); a crash before anchoring reverts it and `session/load` reports the truth. The model is not told — its next command simply behaves under the new mode.
- **What survives a restart — and what if the operator changed the config default while the process was down?** Overrides replay from the session log (`effective = fold ?? config`), so a resumed session keeps its modes with zero catch-up machinery; a default that drifted offline changes behavior the same way a switch does (the approval policy, being stated, is additionally narrated with operator/config attribution).
- **What does `enforcement: 'partial'` on a result mean?** The selected backend enforces the subset its kernel ABI governs — e.g. Landlock before ABI v3 does not govern path truncate — and says so structurally instead of refusing the host; the probe's report line distinguishes the cases. The bwrap and Seatbelt profiles govern every promised file effect by construction, so they always report `full`.
## Prior art
In-repo precedents this design copies or contrasts with:
- [The capability-seams RFC](../architecture/2026-06-13-capability-seams.md) — the interface/implementation/consumer split and the "don't split preemptively" timing rule the second consumer satisfied.
- The `dsh-bash` request/spec split and its `owner` field ([the bash vocabulary catalog](../../../core-data-structures/bash.md)) — the per-call carrier template `sandboxMode` rides, and the explicit-`resolve()` defaulting convention.
- [The approval seam RFC](2026-07-06-approval-seam.md) — the channel escalation asks through; its answerer waterfall, audit pair, and one-package rationale are recorded there.
- [Event-sourced sessions](../architecture/2026-06-11-event-sourced-sessions.md) and [the turn-enclosure invariant](../architecture/2026-06-15-turn-enclosure-invariant.md) — the log-as-store foundation the per-session modes fold over, and the commit boundary the anchoring design obeys.
- [The interception-seams RFC](2026-06-30-interception-seams.md) — the `tools/pre-execute` vocabulary the escalation gate deliberately does not reuse (an escalating call has no pre-execute moment of its own).

42
docs/rfc/implemented/feature/2026-07-07-session-prefix.md Normal file
View File

@@ -0,0 +1,42 @@
# RFC: The session prefix — request-only messages in front of the derived history
Status: implemented
## Problem
A plugin often owns a session-stable opener the model must always see — a skills catalog, an AGENTS.md digest, a workspace baseline. Before this seam the harness offered two homes, and both are wrong for that content. The system prompt is one rendered string: message-shaped content (a user-role `<system-reminder>` envelope, a multi-message primer) does not fit it, and providers weight conversation messages differently from system text. Durable history (`agent.inject()`, a `context/message` at session start) makes the opener permanent: every `deriveMessages()` consumer replays it, the compaction retention walk owns it, forks bake it in stale, and a resume cannot refresh it — a catalog captured at session birth outlives the world it described.
The obvious third option — let a plugin edit the request's `messages` on the way out — is banned by [the reconstructable-requests RFC](../architecture/2026-07-05-reconstructable-requests.md): every loop-built request is a pure function of the session log, so whatever channel carries the opener must log exactly what it sends. What was missing was a request-only message channel with a durable record.
## Decision
`agent/session-prefix` is a waterfall on the agent event map ([`packages/core/agent/src/types.ts`](../../../../packages/core/agent/src/types.ts)): listeners receive a frozen empty seed and return an extension (the canonical contribution is a prepend, `[mine, ...await next()]`, which yields registration order on the wire). The loop ([`packages/core/agent-loop/src/loop.ts`](../../../../packages/core/agent-loop/src/loop.ts)) fires it once per loop instance, lazily before the instance's first `agent/pre-step`; the composed list is deep-cloned, deep-frozen, cached on the instance, and placed in front of the ENTIRE derived history — directly after the provider's system slot — on every request the instance sends ([wire order](../../../core-data-structures/core.md#the-request-envelope-llmcallconfig-and-the-logged-header)).
Three properties carry the design:
- **Request-only, header-logged.** `deriveMessages()` never returns the prefix; its one durable record is `EpochHeader.messagePrefix` on the instance's anchoring `request/header` snapshot — the channel the reconstructable-requests RFC already owns for the request's non-history half, so no new session event exists. The dev invariant ([dsh-invariants](../../../../packages/support/invariants/src/index.ts)) recomputes `messagePrefix + boundary derivation` against every loop-built request; an unlogged prefix cannot reach the wire.
- **Frozen per instance.** Reuse is structural, not disciplined: the cached product cannot change mid-session, so the provider's prompt cache holds by construction and the prefix extends the cacheable region at zero marginal cost per step. A process restart or `ctx.agents.resume()` is a new instance: it recomposes, and any drift lands attributably on the `'resume'` header snapshot. This is the routing rule the seam creates: session-frozen openers ride the prefix; content that changes mid-session rides the append-only history channels (`agent.inject()`, a `tools/post-execute` decision's `additionalContext`, prompt-submit `additionalContext` — [the interception-seams RFC](2026-06-30-interception-seams.md)), each a durable `context/message` paid once and prefix-cached thereafter.
- **Composed before the pressure gate.** Composition precedes the instance's first `agent/pre-step`, and the seam hands the composed value through: `agent/pre-step` carries a `sessionPrefix` parameter and `CompactService.compactIfNeeded(agent, fullSystemPrompt, sessionPrefix, signal)` counts it in its token-pressure estimate — a gate reading the previous instance's folded prefix instead would under-gate a resumed or forked instance whose contributor grew, skipping compaction and shipping an over-window first request. A composition interrupted by a cancel/dispose landing inside the waterfall is discarded, never cached: an abort-aware listener's degraded fallback cannot leak into later requests, and the next turn recomposes under a live signal.
Because composition runs before the boundary snapshot, a composing listener's session append joins the CURRENT request's derived history. Compaction structurally cannot touch the prefix (or the system prompt): it rewrites surface nodes, and header state never enters the surface.
## Testing
**Unit** — [interception.spec.ts](../../../../packages/core/agent-loop/tests/interception.spec.ts) pins compose-once across turns and steps (one composition, zero `request/header-delta`s), canonical prepend ordering, empty-prefix omission from the header, the frozen seed (in-place push throws), held-reference mutation immunity, and composition-precedes-pre-step with the seam receiving the composed value; [cancel.spec.ts](../../../../packages/core/agent-loop/tests/cancel.spec.ts) pins cancel/dispose landing inside the composition window and the discard-and-recompose stale-cache guard; dsh-session codec tests cover the `messagePrefix` fold/diff/apply arms (empty ≡ absent); dsh-invariants tests pin the `messagePrefix + derivation` equation; dsh-compact-basic tests pin that the pressure estimate counts the handed prefix. **Snapshot** — the acp-snapshot normalizer scrubs header prefixes to count-preserving `{{messagePrefix}}` tokens (unit-covered in dsh-acp-snapshot); header content itself is pinned per [the pinned-header scenario RFC](../testing/2026-07-06-pin-request-header-content-in-one-scenario.md), and the example tree loads no prefix contributor, so live goldens stay prefix-free. **e2e** — none prefix-specific: the seam is provider-independent and deterministic; the with-key cache measurement in [request-cache.e2e.ts](../../../../packages/core/agent-loop/tests/request-cache.e2e.ts) already proves the cacheable-prefix economics the design rests on.
## Alternatives considered
- **Per-request `before`/`after` slots recomputed every step** (the shape first proposed: a waterfall firing on every request, contributing frozen `before` messages ahead of the history and fresh `after` messages behind it) — rejected. A per-step `before` recompose invites silent drift — nothing anchors it to the log short of logging a header delta per step — and an `after` slot sits behind the growing history, so its tokens re-pay on every request and everything after it is uncacheable. Measured against the alternatives, every current update pattern is served cheaper by a durable append (paid once, cache-read thereafter), and the only content with no home was the session-stable opener — which wants freezing, not recomputation.
- **A system-prompt section** (`system-prompt/assemble`) — rejected for this content: the assembly renders to the single `system` string, so message-shaped openers do not fit, and the system prompt is deliberately re-assembled per step (with header deltas when it changes) while the opener wants instance-frozen semantics.
- **A durable history opener** (`inject()` at session start) — rejected: permanent history is the failure mode in the problem statement — replayed everywhere, compactable, stale across resumes.
- **Compose per turn instead of per instance** — rejected: a turn-boundary recompose either desyncs silently from the log or forces a header delta per change, and it busts the provider cache exactly as often as it fires; the legitimate refresh point is the instance boundary, where the `'resume'` snapshot already records drift attributably.
- **Compose lazily at the first request and let compaction read the folded header** (the shape as first merged) — superseded in review: the fold matches the live prefix only from the instance's second request on, so on a resumed/forked instance's first step the pressure gate read the PREVIOUS instance's prefix and could under-gate. Composing before the first pre-step and handing the live value through the seam makes the estimate exact at every step.
- **A dedicated session event carrying the prefix** — rejected: the header events are the request's non-history record by design; a second event would be a second home for the same fact and another codec to keep total.
## Consequences
- `agent/pre-step` and `CompactService.compactIfNeeded` carry a `sessionPrefix` parameter: every pre-step listener and compaction backend sees the real per-instance value (all in-repo implementations updated in the same change, per the pre-release stance).
- A contributor whose content changes mid-session is not re-read until the next instance — by design. A deployment needing mid-session catalog updates routes the change notice through the append-only history channels and pays one durable `context/message`.
- The dropped `after` slot leaves no request-only channel near the request tail; nothing in the repo needs one, and adding it back would re-open the every-step re-pay cost the design exists to avoid.
- The `request/header-delta` `messagePrefix` arm (whole-array replacement, empty array encoding transition to absence) exists for codec totality; the loop never exercises it, because the cached prefix cannot change within an instance.
- An empty composition is canonical absence: no-contributor deployments log no extra header bytes and their requests are the bare derivation.

73
docs/rfc/implemented/feature/2026-07-08-repeat-tool-guard.md Normal file
View File

@@ -0,0 +1,73 @@
# RFC: Repeat-tool-call guard plugin
Status: implemented
## Problem
A model stuck in a loop re-issues the same tool call with byte-identical arguments — re-running a failing grep, re-reading an unchanged file, polling a command that already gave its answer — and each round trip burns tokens, wall-clock, and (for paid APIs) money without adding information. The harness has nothing that notices: the loop has no step budget, no plugin tracks call repetition, and the model only escapes when it happens to vary its own behavior. The failure mode is real and cheap to detect — [pi-repeat-tool-guard](https://github.com/Kingwl/pi-repeat-tool-guard) ships exactly this as a pi coding-agent extension: count consecutive identical calls and, past a threshold, append a `<system-reminder>` telling the model to stop repeating itself and change course.
The harness already has every seam the pi extension uses, and better ones: [the interception-seams RFC](2026-06-30-interception-seams.md) gives `tools/post-execute` a sanctioned way to attach model-facing context to a finished call, the loop buffers and injects that context with call/result adjacency preserved, and injected context is a logged `context/message` — so a native guard satisfies the model-visible ⟺ logged rule with no new session event. What was missing was only the plugin itself.
## Decision
The guard is a loop-hygiene plugin, not a model-facing tool: it never appears in the tool list, never vetoes or rewrites a call, and adds exactly one behavior — it watches each agent's stream of tool calls, counts runs of consecutive calls to the same tool with identical canonicalized arguments, and at configured run lengths injects an escalating advisory reminder telling the model to stop repeating itself, re-read the last result, and either change approach or conclude. The purpose is to break unproductive loops within a few wasted steps instead of letting them run to the turn's natural end — while leaving the decision (retry differently, gather more evidence, or finish) entirely with the model, so a legitimately repeated call is delayed by nothing and blocked by nothing.
The plugin is `@deepseek-ai/dsh-repeat-tool-guard` at `packages/guard/repeat-tool-guard/`, opening the `guard/` group for loop-hygiene plugins (single-package groups have precedent: [the todo-write RFC](2026-06-29-todo-write-tool.md) shipped `todo/tool-todo`). It registers three listeners and holds all state in plugin-local maps keyed by `AgentId` — the tool registry is a context-level singleton whose waterfalls interleave every agent's calls (subagents run on the same context), so per-agent keying is correctness, not polish.
- **`tools/post-execute` (waterfall)** — the one detection point. The listener receives `(exec, result)` together, so counting and reminder delivery need no cross-event pending map (the pi extension needs one only because its `tool_call`/`tool_result` hooks are separate events). It always delegates via `next()` and, when a threshold is hit, folds a reminder onto the downstream decision's `additionalContext` — the observe-and-enrich posture [the hooks bridges](2026-06-30-hook-bridges.md) already use, honoring the waterfall contract. Counting happens here rather than in `tools/pre-execute` because post-execute also runs for denied calls (`ToolRegistry.execute` routes a deny through the same pipeline), and a model hammering a denied call is exactly the loop worth breaking.
- **`agent/prompt-submit` (waterfall)** — pure reset hook: delegate via `next()`, clear the submitting agent's chain. A user interjection changes the context; repetition across it is not a loop.
- **`agent/status` (emit)** — on `disposed`, drop the agent's state, bounding the maps over harness lifetime.
### Detection semantics
The chain key is `(tool name, canonical arguments)`; a call identical to the previous tracked call increments the agent's consecutive counter, a different tracked call resets it to 1. Canonicalization is a deep key-sort plus `JSON.stringify`: `ToolExecution.arguments` is by construction the loop's `JSON.parse` output (or the raw string fallback for malformed argument JSON, which is itself a comparable value), so the pi original's bigint/circular/`undefined` handling has no inputs here and is deliberately dropped.
Two deliberate rules, both documented in [the package README](../../../../packages/guard/repeat-tool-guard/README.md) because they are behavior a reader would otherwise guess at:
- **Untracked calls are transparent to the chain.** A call excluded by `include`/`exclude` neither increments nor resets the counter, so `grep X → todo_write → grep X` still counts as two consecutive `grep X` when `todo_write` is excluded. This is what makes exclusion useful — bookkeeping tools interleaved into a loop must not launder it — and it is the pi extension's (undocumented) semantics, kept on purpose and written down.
- **Calls without an agent are ignored.** A direct `ctx.tools.execute()` caller (tests, non-loop consumers) has no model to remind and no `AgentId` to key on.
### Reminder delivery
Reminders ride `additionalContext` (source `{kind: 'plugin', plugin: 'repeat-tool-guard'}` — the label is load-bearing per `HookContext`), never a `content` replacement: the `tool/result` event stays the tool's own output for audit, and the loop appends buffered context as `context/message`(s) after the step's results, which the session renders as the tagged synthetic-user envelope and derived history replays. Thresholds escalate: the first configured threshold gets a short "you are repeating yourself, analyze the previous result" nudge; each later threshold gets the detailed form naming the tool, the repeat count, and the canonical arguments (head-truncated at `argumentsPreviewChars`, default 500 — a looping `write`-sized payload must not ride into the next request unbounded; the chain key always compares the full canonical string), and stating that the calls made no progress. The pi original hardcodes the gentle text to the literal count 3; the guard keys it to `thresholds[0]`, fixing that bug in the port. When the downstream decision already carries `additionalContext` (a hook bridge on the same call), the guard concatenates content under its own `source` — a `HookContext` holds one `MessageSource`, and `source.kind` is what framing depends on.
### Config
```yaml
- id: repeat-tool-guard
name: '@deepseek-ai/dsh-repeat-tool-guard'
config:
thresholds: [3, 5, 8] # default; consecutive counts that trigger a reminder
include: [] # tool-name patterns to track; empty ⇒ all tools
exclude: [todo_write] # tool-name patterns transparent to the chain
argumentsPreviewChars: 500 # default; cap on arguments quoted in the detailed reminder
```
`thresholds` is validated at load and throws on an empty list, a non-integer, a value below 2, or a duplicate — misconfiguration fails loud, replacing the pi original's silent fall-back to defaults. `include`/`exclude` entries support `*` wildcards. Patterns are predicates over whatever tools exist at call time, not references to a registry entry, so an entry matching no currently registered tool is NOT an error — unlike `toolOrder`'s referent check, `exclude: [mcp_*]` must stay valid in a deployment that loads no MCP tools.
## Testing
**Unit** — the suite drives a real agent loop against a scripted mock adapter (no network) and covers, at per-file 100%: counting/reset semantics (identical, different-tracked, untracked-transparent, prompt-submit reset, disposal cleanup, per-agent isolation), canonicalization (deep key-order insensitivity), threshold escalation including the `thresholds[0]` gentle-text rule, denied-call counting, no-agent transparency, wildcard escaping, config fail-loud cases, and both fold-onto-downstream paths (block and accept-with-replacement). **Snapshot** — the `repeat-tool-guard` scenario in the acp-agent example suite scripts five identical `todo_write` calls and pins both reminder tiers (gentle at the third, detailed at the fifth) as `context/message`s in the ACP transcript and the session log; the guard is loaded in the example's live tree (`cordis.yml`), inert for every other scenario (none repeats a call three times). The scenario is authored keyless (like `error-finish`/`cancel`): deterministically forcing a live model to repeat one call three times is not a stable recording. **e2e** — none: the plugin is provider-independent and deterministic, and the seam contracts it relies on are e2e-covered by their owners.
## Alternatives considered
- **Append the reminder into the tool result** (`accept` with replaced `content` — the pi extension's mechanism, which patches result content because that is the only channel its API offers) — rejected: it makes the logged `tool/result` lie about what the tool returned, and `additionalContext` exists precisely as the separate sanctioned channel for post-execute commentary, with loop-level buffering that preserves call/result adjacency.
- **Count in `tools/pre-execute` with a pending-reminder map** (the pi two-phase shape) — rejected: post-execute alone sees `(exec, result)` together and also fires for denied calls, so one listener with no cross-event state covers strictly more attempts with less machinery.
- **Escalate to `block` at the highest threshold** — rejected for the initial scope: a blocked call punishes legitimate identical repeats (polling a long-running terminal, re-checking a file the agent expects to change), and an advisory reminder keeps the model in control. Revisit with evidence; the decision shape (`PostToolDecision`) already supports it.
- **A per-deployment external hook via the CC/Codex bridges** (a `PostToolUse` script) — rejected as the answer: it works for one deployment, but a shipped, unit-tested, `cordis.yml`-configurable plugin is the harness-native form, without per-call subprocess cost.
- **A loop-level step or repetition budget in `agent-loop`** — rejected: "plugins, not loop changes"; a hard step budget is a blunter, orthogonal control that would need its own proposal.
- **Fuzzy/near-identical detection** (normalized paths, similar-but-not-equal arguments) — rejected: exact match after canonicalization is cheap, deterministic, and explainable to the model; similarity thresholds invite false positives and need evidence before they earn complexity.
- **Placing the package in `core/`** — rejected: core is the product spine; a behavioral guard is an optional leaf plugin, and the `todo/` precedent is a small dedicated group per plugin family.
## Consequences
- The reminder is advisory by design: idempotent polling patterns that repeat identical calls on purpose still receive nudges past the thresholds, and the pressure valves are config (`thresholds`, `exclude`) plus reminder text that explicitly allows finishing when enough evidence has been gathered. Each trigger costs reminder tokens on the next request; thresholds bound the frequency.
- Chain state is in-memory only: a session resumed from persistence starts with a fresh chain, so a loop spanning a resume draws its reminders later than a live one — accepted, the guard is a heuristic nudge, not a logged invariant, and persisting counter state would buy little for real complexity.
- When multiple post-execute producers attach context on one call, the fold concatenates under the guard's `source`; ordering between plugins follows listener registration order. The seam cannot represent mixed provenance — a limit inherited from `HookContext`, not owned by this plugin.
- Implementing the snapshot tier surfaced a hidden assumption in the suite kit: the fixture guard equated "authored model scenario" with "override-driven". The `Scenario` table now carries an explicit `overridden` flag, and the sidecar's presence is checked BOTH ways against it (an unregistered stray sidecar would silently replace the derived script) — the suite kit is stricter than it was before this plugin existed.
## Deferred
- Compaction does not reset chains: a compacted history changes what the model sees, but the repetition risk usually survives compaction.
- Escalating to `block` at a high threshold is not implemented; `PostToolDecision` already supports it if evidence arrives.
- Subagent chains stay isolated per agent; no sharing mechanism exists until a concrete case appears.

84
docs/rfc/implemented/feature/2026-07-08-self-referential-cordis-toolset.md Normal file
View File

@@ -0,0 +1,84 @@
# RFC: The self-referential cordis toolset
Status: implemented
## Problem
Everything in this harness is a cordis plugin, but the agent running inside that plugin runtime cannot see or touch it: it cannot enumerate the services and events around it, cannot extend itself with a new tool mid-session, and cannot compose capabilities it invents. Handing the model that power is worth exploring — a self-referential agent that inspects and modifies its own runtime — but it raises three correctness problems at once, and the design is about answering them rather than the raw "let the model run code" mechanic.
First, model-written registration must be validated where it happens: a malformed tool schema has to fail at registration, not when a later request tries to assemble it into a prompt. Second, model-written code has to call service APIs whose source it has never seen — guessed method signatures and, worse, guessed return-value shapes cost many steps of blind probing. Third, everything the model mounts must be fully disposable, by the model on demand and by the ordinary plugin lifecycle when the host plugin reloads, or a long session accretes orphaned listeners and tools.
## Decision
The toolset ships as [`@deepseek-ai/dsh-tool-cordis`](../../../../packages/cordis/tool-cordis/README.md) — a new top-level `packages/cordis/` group — and is demoed by [`examples/cordis-agent`](../../../../examples/cordis-agent/README.md). It gives the model three tools over the live cordis runtime it is running inside: inspect it, mount model-written plugins into it, dispose them again.
The trust stance, stated once and threaded through the rest: the `node:vm` sandbox isolates the global context only — it prevents accidental global pollution, not malice — and the `ctx` a mounted plugin's `apply` receives is a whitelist façade that narrows the *surface* (framework internals withheld) but not the *privilege* of what it exposes. The verbs the façade does expose reach the real runtime: a mounted tool can shell out through `ctx.bash`, read the filesystem through `ctx.fs`, reach the network through `ctx.web`. Neither the sandbox nor the façade is a security boundary; handing the model this power is the point of the toolset. A deployment loads this plugin exactly as deliberately as it grants a bash tool — an opt-in capability in the app's `cordis.yml`, never a product default.
### The three tools
| Tool | Contract |
|---|---|
| `cordis_inspect` | Read-only report over the live runtime, one Markdown section per `what` value (omit `what` for all sections). Never mutates. |
| `cordis_mount` | Evaluates `code` (the body of an async JavaScript function) in a `node:vm` sandbox; the code must `return` a cordis plugin, which is mounted as a child of the `cordis-dynamic` group fiber and tracked under a fresh id (`dyn-1`, `dyn-2`, …). |
| `cordis_unmount` | Disposes one dynamic mount by id and returns only after disposal reaches quiescence — every registration the plugin made is unwound, not merely requested to stop. |
`cordis_inspect` sections: `services` (every provided ctx service and the owning fiber, non-active owners flagged), `plugins` (a flat list of every loaded plugin with its lifecycle state, from `ctx.registry` — what capabilities are loaded, deliberately not the tree shape), `tools` (what the model can call), `dynamic` (the mount table: id, name, state, provided services, awaited services), `api` (live service signatures + the type shapes they reference, from the generated catalog), and `events` (harness events with dispatch mode and signature). The model-facing tool descriptions carry the operational rules the model needs at call time; [the generated tool catalog](../../../tool-catalog.md) is their exhaustive rendering.
### Sandbox semantics
Mount code runs via `vm.createContext` + `runInContext`, wrapped as the body of an async function under a per-mount filename (`cordis-mount-<id>.js`). The vm gives the code a fresh realm: writes to `globalThis` stay inside the sandbox, and no Node API is handed in — capability access is *steered* toward the cordis services (`ctx.fs` for files, `ctx.web` for HTTP, `ctx.bash` for processes, the `ctx.timer` helpers for timing) rather than Node built-ins, so a well-behaved mount stays inspectable through `cordis_inspect` and disposable with its fiber. This is steering, not containment: consistent with the trust stance above, the small global surface keeps *honest* code on the cordis services but is not a security boundary — the host-realm helpers it exposes (`harness`, `console`, `btoa`) are reachable functions, so mount code that goes looking (through such a helper's `.constructor`, say) can still reach the host realm and Node itself, which is accepted because the `ctx` a mount ultimately receives is fully privileged anyway. The `vmTimeoutMs` config bounds only the synchronous portion of evaluation; an async body escapes the bound (also acceptable under the trust stance).
Sandbox globals are deliberately small: a tagged write-through `console` (`[cordis:<id>] …` on the host stdout/stderr, so a listener that fires long after the mount call still lands somewhere the user sees), the `harness.defineTool` / `harness.registerTool` registration pair, the encoding primitives fresh vm contexts lack (`btoa`/`atob` as host closures over `Buffer` — a sanctioned exception, `Buffer` itself is never exposed — plus `TextEncoder`/`TextDecoder`), and callable traps over the withheld Node APIs (`require`, `setTimeout`/`setInterval`/`setImmediate`/`clearTimeout`/`clearInterval`, `fetch`) that throw a redirect naming the cordis alternative. Only function-shaped globals are trapped; `process` and `Buffer` stay `undefined` so a `typeof` feature probe stays inert rather than detonating a throwing accessor.
Three boundary mechanisms make model-written code behave correctly across the realm seam. **Dual-realm `instanceof`**: most objects sandbox code touches are host-realm (tool `args`, event payloads, service returns), so a plain `x instanceof Array` in the vm would silently be false — a per-sandbox prelude gives the vm realm's own constructors a `Symbol.hasInstance` that checks both the vm constructor and its host counterpart, patching only vm-realm globals. **Realm normalization of tool results**: objects built inside the vm carry the vm realm's `Object.prototype`, which the session log's append-time plainness check (`isJsonValue` in `dsh-session`, a prototype-identity comparison) rejects, so the sandbox's `harness.defineTool` JSON round-trips every `execute` return into the host realm — which also projects it onto exactly what the log durably stores — and then shape-checks it against the two `ToolExecuteReturn` forms, so a JSON-valid but wrong-shape return (a bare string, `{ content: 'ok' }`) fails that one call with a teaching error instead of entering the log as corrupt tool-result content. **A whitelist context façade**: the `ctx` a mounted plugin's `apply` receives is NOT the real context nor a pass-through proxy over it — it is a façade exposing only what a mount legitimately needs (`tools.register` marker-guarded, a read-only `tools.get`/`schemas`, `on`/`once`, `provide`, the timer helpers, and the services the plugin DECLARED in `inject`), with every framework-plumbing member (`root`, `parent`, `fiber`, `reflect`, `registry`, `extend`, `isolate`, `intercept`, `plugin`, `set`, `mixin`, …) denied with a teaching error. This closes an escape *class* rather than a single hole: a proxy that merely special-cased `ctx.tools` still handed back the raw context through `ctx.root`, `ctx.extend()`, or a service instance's `.ctx`, and mount code could then `ctx.root.tools.register({…})` to bypass the marker check and realm normalization — a raw vm-realm result then errors a real agent turn at the plainness check. The façade has no context-valued member to reach, and the one indirect leak (an injected-service method returning a `Context`) is rejected on the way back to sandbox code. Two narrower rules complete the surface. First, **service access requires an `inject` declaration**: reaching a service the mount did not declare is refused even when a global provider is live — otherwise a mount could depend on a provider cordis never sees, and unmounting that provider would neither park the consumer nor unwind the tools it registered, leaving a model-visible tool that fails only at execution time. Because the read is gated on the declaration, cross-mount `provide`/`inject` keeps its lifecycle guarantees (the plugin's own `inject` and the fiber's pending/active gating drive activation and unload); only the `apply`-time `ctx` surface is narrowed. Second, **`ctx.tools.get` returns a read-only schema view** (name/description/parameters), never the live `ToolDefinition` — handing back the definition would expose its `execute`, letting mount code call another tool directly and bypass `ToolRegistry.execute` and its pre/post-execute hooks and accounting; a mount that wants to invoke a tool must go through the registry, and one that wants to introspect gets the same view `schemas()` returns.
Boundary errors are written around the mistakes models actually make (see [Consequences](#consequences) for how each was found), and the boundary normalizes rather than lectures wherever the input has exactly one meaning: schema `parameters` accept the JSON-Schema dialect models write by strong prior — the `{ type: 'object', properties, required: […] }` wrapper unwraps to the SchemaSpec DSL (the `required` array becoming per-property flags, at any nesting level), `type: 'integer'` maps to `number`, and `required: false` reads as optional — while genuinely meaningless input is rejected with the vocabulary enumerated (an unknown type lists the five valid ones; a non-boolean `required` names the rule). The remaining teaching errors: an unbalanced `});` closing gets the vm's offending source line plus a "code is a function body" reminder; TypeScript syntax gets the remove-annotations fix (detected on the failing line only, so an ` as ` inside a description string does not misfire); a forgotten `return` gets the two valid plugin forms; a Node built-in call gets the redirect to its cordis service; a tool-name collision on re-mount gets the unmount-first-then-remount recipe.
### The dynamic group and mount lifecycle
Every dynamic mount is a child of a single `cordis-dynamic` group fiber, itself a child of the `tool-cordis` plugin's fiber. The group exists so the mounts form one subtree: they are disposed as a unit, and disposing `tool-cordis` (HMR reload, config unload) cascades over every mount through the ordinary parent→child fiber lifecycle — no bespoke cleanup. Mounting settles before it reports: the returned fiber is `await()`ed, and a startup error (a throwing `apply`, a duplicate tool name, a duplicate service) disposes the fiber and surfaces as the tool error, so a failed mount never lingers. A settled fiber that is not active is a legal pending mount — cordis semantics for unsatisfied `inject` — kept mounted and reported with what it waits for. Everything the plugin registers is an effect on its fiber, so `cordis_unmount` is nothing but an awaited `fiber.dispose()`.
### Cross-mount composition via provide/inject
Mounts relate to each other through ordinary cordis service semantics, with their ids as the lifecycle handles: mount A calls `ctx.provide('foo', value)`, mount B declares `inject: ['foo']` and activates the moment `foo` exists; mounted first, B stays pending and names the missing service; unmounting A sends B back to pending (its registrations unwound) and a later re-provide re-runs B's `apply` through a fresh sandbox façade; a duplicate provide fails loud with the owning fiber named. One realm caveat: a service value provided by a mount is a vm-realm object — method calls on it work from anywhere, but consumers must not assume host prototypes on it.
### The generated API catalog
`cordis_inspect what:"api"` and `what:"events"` answer from a machine-readable catalog generated at build time, never a hand-maintained table that would drift from the JSDoc it paraphrases. [`scripts/gen-cordis-api.ts`](../../../../scripts/gen-cordis-api.ts) reuses `collectServices` / `collectEvents` from [`scripts/gen-cordis-catalog.ts`](../../../../scripts/gen-cordis-catalog.ts) — the same AST walk that generates [the cordis service catalog](../../../cordis-catalog/services.md) and [events catalog](../../../cordis-catalog/events.md) — and emits `packages/cordis/tool-cordis/src/api-catalog.ts`, a committed, banner-commented data module. The artifact carries, per service, its key + one-line summary + raw method signatures; per event, name + `@mode` + signature + summary; the comment-stripped declarations of every exported type the service signatures reference (transitive closure — so a consumer sees that a bash run's `stdout` is `{ text, truncated }`, not a string); plus the curated inherited `ctx` surface shared with the cordis catalog generator. A type name declared in more than one package (each plugin's `Config`) is dropped as ambiguous, and an oversized declaration is truncated with a marker.
Freshness is gated like every generated artifact: `pnpm run verify-cordis-api` (in `doc-sync`) regenerates in memory and fails on any diff, so a JSDoc edit that changes a public signature cannot ship without regenerating the catalog the model reads. At runtime the inspect tool intersects the catalog with the live runtime rather than dumping it: live catalogued services render summary + signatures, live services without a catalog entry (mount-provided ones) render name + owning fiber, catalogued services with no live provider are listed tersely, and the referenced type shapes follow.
### Configuration, rendering, and observability
The plugin exposes one config field, validated by schemastery and documented in [the config catalog](../../../config-catalog.md): `vmTimeoutMs` (default 5000), the millisecond bound on the synchronous portion of mount-code evaluation. Tool names, the `cordis-dynamic` group name, and the `dyn-` id prefix are structural vocabulary and stay fixed. All three tools render as `generic` cards per [the tool cookbook](../../../cookbook/adding-a-tool.md) (`cordis_inspect` a `read`, `cordis_mount` an `execute` carrying the code as `rawInput`, `cordis_unmount` a `delete`), with no `presentResult` overrides.
Model-visible ⟺ logged holds with no new session event type: a mount or unmount is visible only through its own `tool/call` / `tool/result` pair, which the loop logs, and the changed tool set a mount induces is logged by the request-header delta the loop already emits when schemas change between steps. There is deliberately no `cordis/mount` provenance event — it would duplicate what the tool-call pair records. Dynamic mounts are process-lifetime, not session state: resuming a persisted session rehydrates the conversation but does not re-mount plugins.
## Alternatives considered
**A structured per-capability registration tool instead of `cordis_mount`.** The most tempting alternative is a `cordis_register_tool` with explicit `name` / `description` / `parameters` / `code` fields (and siblings `cordis_register_listener`, `cordis_register_service`, …) rather than a single "mount a plugin" primitive. It was rejected because its one real win — no plugin boilerplate for the single commonest case — does not pay for its costs, while a single mount primitive answers every capability at once.
| Dimension | Structured per-capability tools | Single `cordis_mount` |
|---|---|---|
| Schema correctness | `parameters` is still a model-written JSON object needing SchemaSpec validation, merely one step earlier | The same validation runs at the sandbox boundary, with the same instructive errors |
| The code field | An `execute` body is still model-written JS in a vm; the realm and service-call correctness problems are unchanged | One sandbox, one normalization path, one guarded registration |
| Capability coverage | Tools only; listeners, services, `inject` relations each need another structured tool — a surface that grows without bound | One vocabulary (a cordis plugin) covers every effect, present and future |
| Cross-mount composition | Not expressible in a tool-registration payload | Native `provide`/`inject`, ordinary cordis semantics |
| Inspectability | Registers something the plugin list cannot show as a plugin | What the model mounts is exactly what `cordis_inspect` renders |
| Model ergonomics | Wins for the single most common case (no plugin boilerplate) | Mitigated by the canonical recipe in the mount description plus boundary errors that teach the fix |
The correctness investment therefore goes where it pays for every capability at once: the generated API catalog surfaced through `cordis_inspect`, and sandbox-boundary validation whose error messages teach the correct call. A structured registration tool remains addable later as sugar that synthesizes mount code; nothing here forecloses it.
**A hand-maintained service/event reference in the tool.** The first cut of the inspect tool carried a hand-written table of service method signatures. It was replaced by the generated `api-catalog.ts` because a hand table drifts from the JSDoc the moment a signature changes and nothing gates the drift, whereas the generated artifact is freshness-checked against the same AST the docs use.
**A new `cordis/mount` session event.** A durable provenance event recording each mount (source, name) has clear precedent (`hook/invoked`, `compact/start`). It was declined for v1: mount and unmount are already visible as `tool/call` / `tool/result` pairs and the tool-set change is already logged as a request-header delta, so a dedicated event would only duplicate the record. It remains addable if an audit use case needs mount provenance separable from the tool call.
**A hardened / capability-restricted sandbox.** Trapping Node built-ins and handing mount code a whitelist façade rather than the raw context might suggest an intent to sandbox for safety. It is explicitly not that: the traps and the façade narrow the *surface* mount code sees — steering it onto cordis services and away from leak-prone Node built-ins and framework internals — for correctness and to close the unguarded-context escape, but the capabilities the façade exposes (`ctx.bash`, `ctx.fs`, `ctx.web`) reach the real runtime, so it is not a security boundary. A real one (separate process, permission prompts) was out of scope for a dev/opt-in toolset and would fight the entire point — handing the model the live runtime.
## Consequences
The toolset is a deliberate opt-in with a fully-privileged `ctx`, so a deployment adopts it as consciously as a bash tool. Several facts follow that the tool descriptions warn the model about directly: a waterfall listener (e.g. `tools/pre-execute`) that returns without calling `next()` vetoes the chain, so a mounted listener can lobotomize the agent's own tool dispatch ([waterfall semantics](../../../cordis-primer.md#cordis-waterfall-semantics)); mount code runs inside a tool call of the current turn, so awaiting anything that resolves only after the turn deadlocks; `vmTimeoutMs` bounds synchronous evaluation only; and mounts do not survive session resume.
The instructive boundary errors were not guessed — they were written against live self-design sessions in which a real model was asked to build itself coding tools. Those sessions surfaced the failure modes now mitigated: the model closed a returned plugin object with `});` and got only a bare `Unexpected token ')'` it retried blind; it hit a false-positive "this is TypeScript" hint because a description string contained the word "as"; it guessed a bash run's `stdout` was a string and burned six steps building throwaway debug tools to discover it is `{ text, truncated }`; and it wrote tool schemas in the JSON-Schema dialect (`type: 'integer'`, `required: false`, then the full wrapper) three rejections in a row — the rejection text itself pushing it from a nearly-correct DSL attempt back to raw JSON Schema. The fixes — source-line-plus-caret parse errors, line-scoped TypeScript detection, the type-shape closure in the API catalog, the redirect traps, and schema-dialect normalization in place of rejection — cut later sessions from dozens of tool calls with repeated errors to a first-try success on every capability, including a model that hit a Node-`setTimeout` trap and self-corrected to `inject: ['timer']` in one step.
Coverage is named per tier: package unit specs drive the three tools through a real `ToolRegistry` on a real fiber tree (the mount success/failure family, vm isolation, dual-realm `instanceof`, realm normalization against the real `isJsonValue`, the SchemaSpec and raw-registration rejections, the Node-API traps, the cross-mount provide/inject matrix, catalog-backed `api`/`events` rendering, config validation, presenters, quiescent unmount, and the HMR cascade), a `MockAdapter` loop test proves a tool mounted in one step is dispatchable in the next, and the example carries a keyless Loader smoke plus a with-key smoke that world-verifies a live model mounting a listener, building its own tool, and composing two mounts. No snapshot scenario is added: the toolset ships in no ACP-served app, so it changes no editor-facing transcript, and its presenters are unit-tested pure functions — adding it to the ACP example solely for a golden would rewrite the pinned request-header tool set of every recorded scenario.

92
docs/rfc/implemented/feature/2026-07-12-subagent-persona-tool-filter-and-depth.md Normal file
View File

@@ -0,0 +1,92 @@
# RFC: Configure subagent persona, tool visibility, and depth
Status: implemented
## Problem
A reusable subagent provider answers how to run a child, but different delegation tools need different child behavior. One deployment may want a reviewer persona, a research-only tool set, or a hard recursion bound without creating a new provider for every combination.
These controls affect the child's first model request and therefore cannot be installed after the child is visible. They also need honest provider support: an ACP backend cannot silently accept an in-process-only tool filter, and a filter must not be described as a security boundary when every plugin runs in the same trusted process.
## Decision
Subagent starts have three independent composition controls: `persona`, `toolFilter`, and `maxDepth`. A provider advertises support for each control, the service rejects unsupported requests before starting a run, and an in-process provider installs the requested composition while the child is still unpublished.
The controls answer different questions:
| Control | Question | Result |
|---|---|---|
| `persona` | What role instructions replace the deployment persona for this child? | A child-local prompt section shadows `deployment:persona` |
| `toolFilter` | Which deployment-global tools enter this child's visible tool view? | A scoped restriction filters globals before child-local tools are added |
| `maxDepth` | How deep may this delegation tree grow? | A start whose child depth exceeds the absolute cap is rejected |
`dsh-tool-subagent` exposes the controls as plugin configuration and copies them into each request it creates. Direct `SubagentService` callers may choose them per request. The provider capability descriptor remains the source of truth for whether a backend can honor each field.
### Persona is a scoped shadow
The persona control changes one child without changing deployment-wide prompt assembly. During unpublished setup, an in-process provider registers a child-scoped section named `deployment:persona`; ordinary most-specific-wins resolution replaces the global section only in that child's assemblies.
The value has the same strict template semantics as the deployment persona. Omitting it inherits the deployment section through the global layer; an explicit empty string shadows the global persona with an empty section. Parent and sibling personas never enter the child's flat scope.
This uses the normal system-prompt registration mechanism rather than a second persona channel. The first prompt therefore sees the same named contribution that later prompts and prompt-inspection tools see.
### Tool filtering is one live global-view rule
The tool filter controls visibility and executable lookup together. An in-process provider installs `ToolRegistry.restrict()` in the child's scope before publication, and the registry's single resolver applies the same result to prompt schemas, lookup, execution, and Code Mode SDK generation.
Resolution follows these rules:
1. Each restriction applies `allow` before `deny` to the live deployment-global tool registry.
2. Multiple restrictions intersect, so every installed restriction must admit a global tool.
3. Child-scoped tools are added after global filtering and may shadow an admitted global tool.
4. Reserved `run_code` presentation and other scope-local protocol contributions are outside the global filter.
Configuration fails loudly when a filter supplies neither `allow` nor `deny`, or names something outside the current global restrictable set, including a scope-local-only or reserved name. `allow: []` is valid and deliberately hides every global tool. These checks catch misspellings and prevent configuration from appearing effective when it cannot affect the named entry.
The global registry remains live. A deny-only filter admits a later global name unless it explicitly denies that name; an allow-list excludes a later global name unless it explicitly allows that name. Removing a global tool removes it from every resolved view. These semantics preserve hot registration while making the difference between allow and deny explicit.
### Depth is an absolute tree cap
The depth limit bounds recursive delegation independently of tool visibility. A top-level agent has depth zero; an in-process child has its parent's validated depth plus one. `maxDepth` is an absolute non-negative safe integer, and a start rejects before child ownership begins when the derived child depth is greater than the cap.
Every public entry validates the domain rather than relying on one model-facing configuration path. Negative values, fractions, negative zero, non-finite values, unsafe integers, malformed stored parent depth, and derived overflow all reject. Omitting the cap leaves depth unbounded by this mechanism.
A deployment can combine depth and filtering. For example, it may keep the delegation tool visible at depth one but set `maxDepth: 1`, or deny the delegation tool entirely in children. Neither choice changes the provider's conversation-history behavior.
### Capability gating keeps providers honest
Capabilities separate a requested feature from a provider implementation. `SubagentCapabilities` advertises `persona`, `toolFilter`, and `depthLimit`; `SubagentService.start()` checks every present request field against those flags before calling the provider.
This lets spawn and fork providers share the in-process implementation while external providers advertise only what they can enforce. A request never degrades silently: selecting an unsupported control produces `UNSUPPORTED_CAPABILITY`, and no run or lifecycle event exists.
### Unpublished setup makes the first request correct
All child-local composition is complete before the child becomes observable. The in-process provider supplies one setup callback to agent creation; that callback installs persona, tool restriction, and structured-output contributions in the child's scope. Only after setup succeeds does creation publish the session and agent and allow the driver to start.
A setup failure rolls back the private child. No observer can acquire a child whose first prompt used the deployment persona or unfiltered tool set and whose later prompts use the requested configuration.
## Visibility is not authority
These controls compose trusted same-process behavior; they do not authorize it. `toolFilter` changes the child view resolved by the tool registry, but it does not create a parent-to-child grant lattice, require a child to be a subset of its parent, sandbox plugins, or prevent code with another Cordis context from calling services directly.
In particular, a child-local tool is added after the global filter and may be absent from the parent's view. A deny-only child also sees later global tools not named by the deny-list. Those are deliberate live-composition semantics, not non-escalation guarantees.
A security design would need a separate authority representation, propagation rule, and execution-time enforcement point. Creation-time grant snapshots, parent-subset grants, explicit future-grant APIs, and generic capability/output/termination tags are outside this feature.
## Alternatives considered
**Create one provider per persona or tool set.** This multiplies providers that share the same transport and lifecycle implementation, makes dynamic deployment configuration awkward, and still needs a recursion mechanism. Providers remain about execution transport; requests carry per-child composition.
**Copy the parent's complete tool view.** Registration scope is flat by design, and lifetime ownership does not imply visibility inheritance. Copying a resolved view would also freeze dynamic global registrations and conflate composition with authority without defining either contract fully.
**Snapshot allowed global tools at child creation.** A frozen allow-set makes future registration uniformly unavailable, but it changes hot-registration semantics and starts an authorization design. The implemented filter stays a live registry predicate and documents allow-versus-deny behavior directly.
**Hide only tool schemas.** Presentation-only filtering lets the model execute a tool that the prompt says does not exist through Code Mode or a forged call. One resolver governs both presentation and execution instead.
**Use only tool filtering to stop recursion.** Removing the delegation tool is useful but provider-specific and does not protect direct service callers or alternate delegation tools. Absolute depth is an independent structural bound.
## Consequences
Contributors can configure child role, visible global tools, and recursion without defining new providers. Capability checks fail before ownership starts, unpublished setup makes the first request consistent, and one tool resolver prevents presentation/execution drift.
The cost is that deployments must understand live allow/deny behavior and the distinction between visibility and authority. Provider authors must advertise each supported control accurately, and in-process providers must install every requested contribution before publication. The controls deliberately do not solve security confinement or parent-to-child non-escalation.

4
docs/rfc/implemented/process/2026-07-06-node-engine-floor.md
View File

@@ -8,7 +8,7 @@ The Node 22 branch of the root `engines.node` range is a contract for the instal
## Decision
Set `engines.node` to `^22.19.0 || >=24.0.0` and test the keyless CI compatibility matrix on `['22.19', 24, 26]`. The real-API e2e workflow stays on Node 24 because it exercises API integration rather than the runtime floor.
Set `engines.node` to `^22.19.0 || >=24.0.0` and test the keyless CI compatibility matrix on `['22.19', 24, 26]`. Every matrix leg runs the TypeScript typecheck plus a keyless source-mode worker smoke, so the floor is exercised through both a complete source typecheck and a real unbuilt runtime path. The real-API e2e workflow stays on Node 24 because it exercises API integration rather than the runtime floor.
Two Node features gate the source runtime:
@@ -22,7 +22,7 @@ Those source features clear on the 22.x line at **22.18**, but the installed Pi
## Consequences
- The advertised LTS branch no longer undercuts the Pi adapter dependency floor.
- CI proves the Node 22 LTS floor directly with Node 22.19, keeps the Node 24 branch on `node: 24`, and keeps Node 26 for the next even line.
- CI proves the Node 22 LTS floor directly with Node 22.19, keeps the Node 24 branch on `node: 24`, and keeps Node 26 for the next even line; each leg typechecks the source graph and launches the unbuilt workflow worker for real.
- The built-bin smoke needs no version-conditional flag: at 22.19 type-stripping is already the default, so the test stays the plain `node lib/bin.js` path it documents.
- A future dependency or source API that raises the runtime floor must move `engines.node`, the compatibility matrix, and this RFC in the same change.

6
docs/rfc/implemented/process/2026-07-06-parallel-github-ci-gates.md
View File

@@ -10,9 +10,9 @@ The hard part is the artifact boundary. `publint`, `verify-node-next-types`, and
## Decision
[CI](../../../../.github/workflows/ci.yml) keeps the keyless workflow to a few broad jobs instead of one job per gate. The Node 24 matrix has five lanes: static gates (`pnpm run check:ci:static`), lint (`pnpm run check:ci:lint`), coverage (`pnpm run check:ci:coverage`), snapshot replay (`pnpm run check:ci:snapshot`), and artifact gates (`pnpm run check:ci:artifacts`). The Node 26 compatibility job installs once and runs `pnpm run check:node-compat`.
[CI](../../../../.github/workflows/ci.yml) keeps the keyless workflow to a few broad jobs instead of one job per gate. The Node 24 matrix has five lanes: static gates (`pnpm run check:ci:static`), lint (`pnpm run check:ci:lint`), coverage (`pnpm run check:ci:coverage`), snapshot replay (`pnpm run check:ci:snapshot`), and artifact gates (`pnpm run check:ci:artifacts`). The compatibility matrix has Node 22.19, 24, and 26 jobs; each installs once and runs `pnpm run check:node-compat`.
Each lane delegates to [scripts/run-gates.ts](../../../../scripts/run-gates.ts), an in-process scheduler with bounded concurrency (`DSH_GATE_CONCURRENCY`). The static lane fans out constraints, the echo-agent demo smoke, `doc-sync` leaf gates, module-graph freshness, and `knip`; the lint lane runs ESLint with its own Node heap cap and a content-strategy ESLint cache; the coverage lane runs Vitest coverage with bounded file workers (`DSH_COVERAGE_MAX_WORKERS`); the snapshot lane isolates replay; the artifact lane builds once and then fans out the artifact consumers; the Node 26 compatibility job owns the TypeScript typecheck. The scheduler buffers each gate's output and prints a named result block with duration, so independent failures stay attributable inside each broad job log.
Each lane delegates to [scripts/run-gates.ts](../../../../scripts/run-gates.ts), an in-process scheduler with bounded concurrency (`DSH_GATE_CONCURRENCY`). The static lane fans out constraints, the echo-agent demo smoke, `doc-sync` leaf gates, module-graph freshness, and `knip`; the lint lane runs ESLint with its own Node heap cap and a content-strategy ESLint cache; the coverage lane runs Vitest coverage with bounded file workers (`DSH_COVERAGE_MAX_WORKERS`); the snapshot lane isolates replay; the artifact lane builds once and then fans out the artifact consumers. Every compatibility job runs the TypeScript typecheck and a keyless workflow-workerthread source-launch smoke, which starts a real unbuilt worker and therefore catches Node-version-specific loader/runtime failures that typechecking cannot. The scheduler buffers each gate's output and prints a named result block with duration, so independent failures stay attributable inside each broad job log.
Generated `.sessions/` logs and `.doc-typecheck-*` temp directories are ignored by lint. The aggregate local CI mode still runs demo smoke after lint, while the split GitHub static lane can run demo smoke directly because lint is isolated in its own lane.
@@ -36,4 +36,4 @@ The broad-lane split repeats checkout, setup, and install more often than a sing
The split introduces a maintenance obligation: when `package.json` adds or removes a gate that belongs in CI, [scripts/run-gates.ts](../../../../scripts/run-gates.ts) needs the matching leaf. That obligation is intentional because the runner is the parallel execution plan for the same gate vocabulary, not a separate quality policy.
The Node 26 signal is narrower than the primary Node 24 signal. It proves the source graph on the newer runtime without doubling documentation, coverage, publication, snapshot, and smoke checks whose failures are not expected to vary by Node minor version.
The compatibility signal is narrower than the primary Node 24 signal. It proves that the source graph typechecks and that the real unbuilt workflow-worker launch path executes on every advertised runtime line without doubling documentation, coverage, publication, snapshot replay, and unrelated smoke checks whose failures are not expected to vary by Node version.

2
docs/rfc/implemented/simplification/2026-06-20-remove-agent-boundary-mirror-events.md
View File

@@ -13,7 +13,7 @@ Status: implemented
## Problem
The loop records the canonical transcript in `SessionEvent` and also emitted a parallel set of live `agent/*` boundary mirror events: `agent/turn-start`, `agent/turn-end`, `agent/step-start`, and `agent/step-end`. The mirrors made consumers choose between two sources of truth for the SAME durable fact. ACP already chose the session log for the editor-facing transcript because a throwing peer listener can prevent later `agent/*` listeners from observing a boundary, while the session event was already appended. The stdio UI was the only production consumer that still rendered turn boundaries from the mirror events; it already rendered tool calls and results from `session/event`.
The loop records the canonical transcript in `SessionEvent` and also emitted a parallel set of live `agent/*` boundary mirror events: `agent/turn-start`, `agent/turn-end`, `agent/step-start`, and `agent/step-end`. The mirrors made consumers choose between two sources of truth for the SAME durable fact. ACP already chose the session log for the editor-facing transcript because it is the one durable, replayable record; consuming a live mirror would require reconciling its timing with the boundary already stored in that log. The stdio UI was the only production consumer that still rendered turn boundaries from the mirror events; it already rendered tool calls and results from `session/event`.
This duplication is not free. Every lifecycle change had to update the session event, the mirror event, docs, invariants, tests, and snapshot expectations. The duplicate boundary events also made failure ordering subtle: a turn can be durably closed before a live `agent/turn-end` listener runs, so a post-boundary listener failure has no valid in-log position left and must be reported out of band.

4
docs/rfc/implemented/testing/2026-06-19-acp-snapshot-tests.md
View File

@@ -51,7 +51,7 @@ The ACP server app loads `@deepseek-ai/dsh-llm-deepseek`, whose `apply` throws w
A snapshot run asserts **two** normalized surfaces, because the harness's external surfaces are distinct:
1. The **stdout transcript** — the framed `session/update` JSON-RPC the editor sees. Catches regressions in the ACP bridge's event→update translation (`streamSessionEventUpdate`). Compared against a committed `stdout.golden.jsonl`.
2. The **re-persisted session JSONL** — the log the replay run itself persists, compared against the scenario's `session.jsonl`. Catches regressions in the loop, tool dispatch, and turn/step structure that never surface on stdout. There is no separate session golden: `session.jsonl` is BOTH the replay source (recorded scenarios) and the expected produced log. Both sides pass through `normalizeSessionLog` before comparing — the fixture is raw-harvested (its own real session id / cwd / timestamps) and the replay output has fresh ones, so each is scrubbed against ITS OWN volatile values (the fixture's read from its header line) and the comparison is on normalized form. Request-header CONTENT (the composed system prompt + tool schemas) is additionally scrubbed to `{{system}}`/`{{tools}}` tokens on both sides — in the stored fixtures too — for every scenario except the one that pins it ([pinned-header RFC](2026-07-06-pin-request-header-content-in-one-scenario.md)). For an authored override scenario the same `session.jsonl` holds the expected produced log; `replay.override.json` drives the model, and `llm-replay` ignores the fixture for model chunks when an override exists, so committing the expected log there does not affect replay.
2. The **re-persisted session JSONL** — the log the replay run itself persists, compared against the scenario's `session.jsonl`. Catches regressions in the loop, tool dispatch, and turn/step structure that never surface on stdout. There is no separate session golden: `session.jsonl` is both the replay source (recorded scenarios) and the expected produced log. Both sides pass through `normalizeSessionLog` before comparing — the fixture is raw-harvested (its own real session id / cwd / timestamps) and the replay output has fresh ones, so each is scrubbed against its own volatile values (the fixture's read from its header line) and the comparison is on normalized form. Every stored JSONL additionally scrubs composed prompt text to `{{system}}`; each header class's pinning scenario stores that prompt readably in `system-prompt.golden.md` and keeps the complete tool schemas in its JSONL, while other scenarios scrub schemas to `{{tools}}` ([pinned-header RFC](2026-07-06-pin-request-header-content-in-one-scenario.md)). For an authored override scenario the same `session.jsonl` holds the expected produced log; `replay.override.json` drives the model, and `llm-replay` ignores the fixture for model chunks when an override exists, so committing the expected log there does not affect replay.
The two are genuinely additive: stdout is the bridge's *lossy projection* of the log (it drops `assistant/message.usage`, `step/*`, exact `seq`/`time`, and renders tool I/O differently), so a loop/tool/turn-structure regression can change the JSONL while leaving the stdout projection identical, and a bridge-translation regression can change stdout while the JSONL is untouched. Asserting the JSONL equality also echoes the proposed [universal replay fixture](../../proposed/testing/2026-06-11-deterministic-and-stress-testing.md) idea.
@@ -68,7 +68,7 @@ The replay plugin lives in its own package, `@deepseek-ai/dsh-llm-replay` (`pack
### Two subcommands, replay in the default gate
`pnpm run test:snapshot` runs replay (keyless) and is composed into the default `pnpm run test` gate so every PR gets the regression check (the main `vitest.config.ts` include stays narrow; the gate is `test && test:snapshot`). `pnpm run test:snapshot:record` requires `DEEPSEEK_API_KEY` (loaded from repo `.env` first), hits the real API, harvests the produced `session.jsonl` (the replay source AND the expected-log artifact), and `--update`s the stdout golden in one pass. Both forward a scenario filter. A missing fixture in replay **fails loud** with a "record first" message rather than self-skipping (the e2e self-skip rule is a CI-secret accommodation, not appropriate here — a committed-fixture test that silently vanishes is a coverage hole). A no-model scenario's `session.jsonl` simply has no `assistant/chunk` events (empty derived script); fail-loud still applies if a model call happens with no entry. An orphan-fixture guard test fails on a golden/fixture not referenced by any scenario (Vitest does not prune orphaned raw goldens), and a per-kind required-fixture guard asserts each scenario ships exactly the files its kind needs (`input.json` + `stdout.golden.jsonl` + `session.jsonl` for ALL scenarios — the harness passes `<dir>/session.jsonl` to `llm-replay` unconditionally, so even a no-model scenario needs its header-only fixture or `loadReplayScript()` fails; `replay.override.json` additionally for authored model scenarios).
`pnpm run test:snapshot` runs replay (keyless) and is composed into the default `pnpm run test` gate so every PR gets the regression check (the main `vitest.config.ts` include stays narrow; the gate is `test && test:snapshot`). `pnpm run test:snapshot:record` requires `DEEPSEEK_API_KEY` (loaded from repo `.env` first), hits the real API, harvests the produced `session.jsonl` (the replay source AND the expected-log artifact), and `--update`s the stdout golden in one pass. Both forward a scenario filter. A missing fixture in replay **fails loud** with a "record first" message rather than self-skipping (the e2e self-skip rule is a CI-secret accommodation, not appropriate here — a committed-fixture test that silently vanishes is a coverage hole). A no-model scenario's `session.jsonl` simply has no `assistant/chunk` events (empty derived script); fail-loud still applies if a model call happens with no entry. An orphan-fixture guard test fails on a golden/fixture not referenced by any scenario (Vitest does not prune orphaned raw goldens), and a per-kind required-fixture guard asserts each scenario ships exactly the files its kind needs (`input.json` + `stdout.golden.jsonl` + `session.jsonl` for ALL scenarios — the harness passes `<dir>/session.jsonl` to `llm-replay` unconditionally, so even a no-model scenario needs its header-only fixture or `loadReplayScript()` fails; `replay.override.json` exactly for the scenarios whose table entry sets `overridden` — required with the flag, forbidden without it, because the harness forwards the sidecar purely on file existence and an unregistered stray would silently replace the derived script).
## Alternatives considered

17
docs/rfc/implemented/testing/2026-07-06-pin-request-header-content-in-one-scenario.md
View File

@@ -4,29 +4,30 @@ Status: implemented
## Problem
Every model-driving ACP snapshot fixture (`session.jsonl`) embedded the full composed system prompt and the complete tool-schema list in its `request/header` event — roughly 8 KB on one line, per fixture. That content is identical across the suite (byte-identical tool list everywhere, including subagent children; identical prompt modulo each recording's temp cwd), so any change touching a tool description or a system-prompt line had to update every fixture: re-record everything against the live API (churning model responses and stdout goldens along the way) or hand-edit ~35 giant header lines. Introducing the dynamic-workflows feature — one new tool plus one prompt paragraph — rewrote every snapshot fixture in the repo, burying the behavioral diff a reviewer should be reading.
An ACP snapshot suite needs to prove the exact composed system prompt and tool-schema list sent in each `request/header`, but duplicating that content inside every `session.jsonl` makes a prompt or schema edit rewrite dozens of giant one-line JSON records. Keeping one raw header avoids the duplication but still makes prompt review poor: prose is JSON-escaped onto one line and mixed with thousands of characters of tool schemas.
## Decision
Exactly one scenario — `text-turn`, flagged `pinsHeader` in the `acp.snapshot.ts` scenario table — commits and compares the full request-header content; the pin mechanics live in [`dsh-acp-snapshot`](../../../../packages/support/acp-snapshot/README.md), whose suite factory enforces one pin per consuming suite. Every other fixture stores and compares that content as stable tokens via the pure normalizer `scrubRequestHeaders` in that package's `normalize.ts`: a `request/header` event's `header.system` becomes `"{{system}}"` and `header.tools` becomes `"{{tools}}"`; a `request/header-delta` keeps its structural facts — the system delta's `keepStart`/`keepEnd` line positions with one `{{system}}` token per inserted line, the tools delta's added/removed/changed tool names — and tokenizes only the bulk (prompt text, schema bodies), so two different deltas still compare different. The scrub is composed in front of `normalizeSessionLog` on BOTH sides of a non-pinning scenario's log compare and applied to the harvested logs record mode writes, so a re-record cannot smuggle the content back. Absent fields stay absent — WHETHER a header carried a prompt or tools is behavior and stays visible — and `config`/`reason` stay verbatim: a model swap churns every fixture by design (it invalidates the recorded responses), while a prompt or schema edit churns none of them (replay derives model behavior exclusively from `assistant/chunk` events and never reads header content — see `dsh-llm-replay`).
Exactly one scenario per header-composition class is flagged `pinsHeader`. Its directory splits the pin by review format: `system-prompt.golden.md` contains the normalized composed prompt as ordinary Markdown, while `session.jsonl` keeps the full tool-schema list, config, and reason but stores `header.system` as `"{{system}}"`. Every other JSONL stores both the system prompt and tool list as `"{{system}}"` / `"{{tools}}"`. The pin mechanics live in [`dsh-acp-snapshot`](../../../../packages/support/acp-snapshot/README.md), whose suite factory enforces one pin per class.
A system-prompt or tool-schema change therefore lands as exactly one committed-fixture diff — the pinned `text-turn` header line — updated by hand or by re-recording that one scenario (`pnpm run test:snapshot:record` with `-t text-turn`).
The pure `scrubSystemPrompts` normalizer applies to every stored session fixture and tokenizes both an initial header's prompt and a header delta's inserted prompt lines. `scrubRequestHeaders` additionally tokenizes tool schemas and session-prefix content for non-pinning scenarios while retaining structural facts: system-delta positions and arity, added/removed/changed tool names, prefix message count, field presence, config, and reason. Record and refresh write-back apply the appropriate scrub before writing JSONL and regenerate the Markdown prompt from the normalized live header, so neither path can reintroduce prompt text into JSONL or leave the readable snapshot stale.
Guards make the split self-enforcing. On disk (fixture meta-tests): every non-pinning `session*.jsonl` must be a fixed point of `scrubRequestHeaders` (unscrubbed content crept in — apply the scrub), the pinning scenario's fixture must NOT be one (the pin lost its content), and exactly one scenario must pin. Live (every non-pinning scenario run): each `request/header` the run produces — parent, spawn child, fork child, initial or resume — must equal the pinned fixture's header after both sides normalize their own volatile values, and no `request/header-delta` may appear at all (a mid-run header change diverges from the pin by construction, and its content would be invisible under the scrub), so the single-pin premise is asserted rather than assumed.
Guards make the split self-enforcing. On disk, every `session*.jsonl` is a fixed point of `scrubSystemPrompts`, only non-pinning fixtures are fixed points of the full header scrub, `system-prompt.golden.md` exists exactly beside pinning fixtures, and each class has one pin. Live, every `request/header` produced by a parent, spawn child, fork child, initial request, or resume must match both halves of its class's pin after volatile-value normalization. A header without a string prompt or any `request/header-delta` fails loud because the two static pin artifacts cannot represent it.
One pin covers the whole suite because every session — parent, spawn child, fork child — composes the identical tool list and the identical prompt modulo cwd, and the uniformity guard fails the suite the moment that stops holding. If header composition ever becomes session-dependent by design (a restricted subagent toolset, say), the divergent shape gets its own pinning scenario.
## Alternatives considered
- **Re-record or hand-edit every fixture per change** — the status quo; the churn this RFC removes.
- **Scrub at compare time only, keeping fixtures raw** — the compares go green without fixture edits, but every committed fixture then carries a permanently stale copy of the prompt and schemas: dead weight that misleads readers and still rewrites wholesale on the next re-record. Storing the tokens keeps the fixture honest about what it does and does not pin.
- **Re-record or hand-edit every fixture per change** — preserves exact headers but buries behavioral diffs under duplicated prompt and schema content.
- **Scrub at compare time only, keeping fixtures raw** — lets compares pass while committed fixtures retain stale duplicate content and rewrite wholesale on the next recording. Stored tokens state honestly what each JSONL does not pin.
- **Scrub everywhere, pin nowhere** — loses the only end-to-end record of the composed header as actually sent (prompt assembly, registered-tool order, full schemas). The generated tool catalog documents each tool in isolation; only a real fixture pins the composed set.
- **Keep the one full pin entirely in JSONL** — removes suite-wide duplication but leaves system-prompt changes as an escaped one-line diff entangled with the tool list. Markdown gives prompt prose its natural review format without weakening the header assertion.
- **Slim the session log itself (log a content digest, store the header elsewhere)** — violates the reconstructability contract: the product log must reproduce each request bit-for-bit ([reconstructable-requests RFC](../architecture/2026-07-05-reconstructable-requests.md)). Header bulk is a test-artifact concern, solved in test normalization; the live log is untouched.
## Verification
All 37 snapshot scenarios replay green with the scrubbed fixtures (the committed fixtures were rewritten once through `scrubRequestHeaders` itself; `text-turn` untouched). The fixed-point, pin-retains-content, exactly-one-pin, live header-uniformity, and no-unpinned-delta guards run inside the suite, and `scrubRequestHeaders` has unit coverage for both header event types, delta structure preservation (line positions, insert arity, tool names), absent-field preservation, config/reason retention, byte-for-byte pass-through of other lines, and idempotence.
The suite replays every scenario against the split pins. Unit coverage exercises both scrub levels, Markdown formatting, record/refresh regeneration, normalized prompt extraction, fixed-point enforcement, required-file symmetry, header uniformity, and delta rejection.
## Consequences
A tool-description or system-prompt change churns one committed fixture line instead of every fixture in the suite, so snapshot diffs read as behavior again, and ~270 KB of duplicated header bytes leave the repo. The cost: non-pinning fixtures no longer display header content, so reading one shows tokens where the prompt and schemas were — the pinned `text-turn` fixture is the place to look, and the live uniformity guard guarantees it speaks for every session in the suite. A header change surfaces as a suite-wide test failure whose fix is the one pinned line, rather than as ~35 fixture rewrites.
A system-prompt change produces a normal line-oriented Markdown diff in one file per affected composition class; a tool-description change produces one pinned JSONL line per class; ordinary behavioral fixtures remain untouched. Session fixtures display tokens for omitted content, and the live uniformity guard makes each split pin authoritative for every session in its class. The pinning scenario carries one extra generated artifact whose terminal newline is canonicalized for repository hygiene.

2
docs/rfc/implemented/testing/2026-07-08-shared-acp-snapshot-package.md
View File

@@ -16,7 +16,7 @@ The machinery lives in [`packages/support/acp-snapshot`](../../../../packages/su
**`src/normalize.ts`** — the pure normalizers, hook-free by policy: when a future event carries a new volatile field (an approval duration, say), the shared normalizer learns it in the same change, keeping one home for what "normalized" means rather than per-suite scrub extensions.
**`src/suite.ts`** — the `Scenario` type and `defineAcpSnapshotSuite(options)`, registering the per-scenario compares, record-mode fixture write-back, the header pin with its live uniformity guard, and the fixture guard block (no orphan scenario dirs, required files present, exactly one pin, non-pinning fixtures are `scrubRequestHeaders` fixed points). The pinned-header contract ([pinned-header RFC](2026-07-06-pin-request-header-content-in-one-scenario.md)) is per-suite: each suite flags exactly one `pinsHeader` scenario (the factory throws on zero, a meta-test rejects more than one; WHICH scenario pins is the table's reviewable choice), and the uniformity guard compares only that suite's sessions. The pure helpers (`childFixturePaths`, `fixtureContext`, `normalizedHeaders`, `headerDeltaCount`) are exported for direct unit coverage.
**`src/suite.ts`** — the `Scenario` type and `defineAcpSnapshotSuite(options)`, registering the per-scenario compares, record/refresh fixture write-back, the header pin with its live uniformity guard, and the fixture guard block (no orphan scenario dirs, required files present, exactly one pin per class, every JSONL a `scrubSystemPrompts` fixed point, non-pinning fixtures also `scrubRequestHeaders` fixed points). The pinned-header contract ([pinned-header RFC](2026-07-06-pin-request-header-content-in-one-scenario.md)) is per-suite: each header class flags exactly one `pinsHeader` scenario, whose `system-prompt.golden.md` and JSONL tool list split the composed header into reviewable artifacts; the uniformity guard compares both against every live header in that class. The pure helpers (`childFixturePaths`, `fixtureContext`, `normalizedHeaders`, `normalizedSystemPrompts`, `formatSystemPromptSnapshot`, `headerDeltaCount`) are exported from the module for direct unit coverage.
## Alternatives considered

84
docs/rfc/proposed/feature/2026-06-14-acp-agent-client-protocol.md
View File

@@ -1,84 +0,0 @@
# RFC: Agent Client Protocol (ACP) support — drive the coding agent from external editors
Status: proposed
> **Implementation status (MVP landed):** steps 1, 2, 3, 4, 6, 7, 8 are implemented in `packages/ui/acp` + `examples/acp-agent`. **Step 5 (the `session/request_permission` permission gate) is deferred** — the bridge ships a pass-through (tools run with the executor's full authority) marked `TODO(rfc010-permission-gate)`, and lays down only the `WeakMap<Agent, sessionId>` ownership seam the gate will build on. Status stays `proposed` until the gate lands. `session/cancel` is the queue-aware `agent.cancel()`: it aborts a running step, clears queued + steering work, and drops a turn that is about to start, so a queued-but-not-yet-started prompt never runs and a later prompt cannot be batched into the cancelled turn. **Per-session `cwd` is now honored** (lifting the original "launch the server in the workspace root" restriction — see § Deferred): `session/new` accepts any absolute `cwd`, and `session/load` requires the request `cwd` to match the persisted session `cwd` so the editor and bash executor agree on the workspace.
## Problem
The coding agent is reachable only through the readline `stdio-chat` plugin: it reads lines from stdin, calls `agent.send()`, and prints the assistant token stream (`session/event` `assistant/chunk`) to stdout. There is no structured protocol, so the agent cannot be embedded in an editor — no streaming render, no tool-call display, no permission UI, no resumable sessions.
Editors are converging on the Agent Client Protocol (ACP), which Zed and others speak: JSON-RPC 2.0 over newline-delimited stdio, modeled on the Language Server Protocol. An editor boots the agent as a subprocess and exchanges `initialize` / `session/new` / `session/prompt`, rendering streamed `session/update` notifications and `session/request_permission` prompts. The goal is for the agent to be a drop-in ACP server — implement the protocol once and run in any ACP client, with no per-editor glue.
This RFC has a hard prerequisite on [session persistence](../../implemented/architecture/2026-06-14-session-persistence.md): it assumes durable session persistence (the `SessionPersistence` service and the async `AgentLoop.resume` seam) is implemented, so resuming a session via `session/load` is in scope. None of those APIs exist yet — `AgentLoop` currently exposes only the synchronous `create` — so ACP must land after, or in the same change as, [session persistence](../../implemented/architecture/2026-06-14-session-persistence.md), and pins to its `resume(agentId, resumeSessionId)` contract. Session persistence persists every `SessionEvent` verbatim (including `assistant/chunk`), so a loaded session has the stream chunks needed to replay turns to the client.
## Proposal
A new plugin package `@deepseek-ai/dsh-acp` — a client-driver / UI plugin, the structured analogue of `stdio-chat`. It is NOT a change to the loop and NOT an [capability seams](../../implemented/architecture/2026-06-13-capability-seams.md) interface/implementation/consumer capability split; it consumes the existing `agent/*` event taxonomy and the `tools/pre-execute`/`tools/post-execute` waterfalls.
It depends on the official `@agentclientprotocol/sdk` (the `AgentSideConnection` class) — Apache-2.0, actively versioned. The SDK declares a `zod` peer dependency and imports `zod/v4` at runtime, so `packages/ui/acp` must declare `zod` itself (per the workspace dependency constraints). This is the renamed successor to `@zed-industries/agent-client-protocol`, which is now deprecated on npm.
The mapping between ACP and existing harness seams — each row names the seam and any required extension:
| ACP (client ⇄ agent) | Harness seam | Notes |
|---|---|---|
| `initialize` | static handler | negotiate `protocolVersion` (echo the supported version, else error); advertise text-only `promptCapabilities` and `loadSession: true`; report agent name/version |
| `session/new {cwd, mcpServers, additionalDirectories}` → `{sessionId}` | the `dsh-agent` create factory (see Dependency note + Plan) | the seam must accept `{ sessionId, meta }` so the ACP-generated `sessionId` becomes the live/persisted session id and the validated `cwd` is attached as the `SessionHeader` (today `AgentLoop.create(id)` hardcodes `${id}-session` and takes no metadata); reject a 2nd session (single-session MVP, see [ACP multi-session](2026-06-14-acp-multi-session.md)); `cwd` validated (require absolute) — any absolute cwd is honored: it becomes the session's `SessionHeader.cwd` and the default bash workdir (per-session cwd, see § Deferred → RESOLVED), so the server need not launch in the workspace; non-empty `mcpServers` and `additionalDirectories` are rejected for the MVP because silently ignoring requested servers/roots would desync the client's tool and filesystem-scope UI |
| `session/load {sessionId, cwd, mcpServers, additionalDirectories}` | the `dsh-agent` resume factory ([session persistence](../../implemented/architecture/2026-06-14-session-persistence.md) + Dependency note) | load `{ meta, events }`, seed the session, re-derive history via `deriveMessages()`, replay prior turns to the client as `session/update` per the ACP load contract; `mcpServers` and `additionalDirectories` rejected as in `session/new` |
| `session/prompt {prompt}` | `agent.send()` (idle) | text blocks → `TextBlock`; reject image/audio per advertised capabilities; one in-flight prompt per session |
| resolve `session/prompt` → `{stopReason}` | the `turn/end` `session/event` (its `reason`) | map the harness kebab `TurnEndReason` to the ACP snake_case `StopReason` wire enum: `completed`→`end_turn`, `max-tokens`→`max_tokens`, `aborted`(cancel)→`cancelled`, plus `refusal`/`max_turn_requests` when applicable; honor the batch-into-one-turn and send-not-synchronously-running settle semantics |
| `session/update: agent_message_chunk` | `session/event` `assistant/chunk` `text-delta` only | do NOT also emit on `block-end(TextBlock)` — it carries the fully-assembled block and would duplicate the streamed text |
| `session/update: agent_thought_chunk` | `session/event` `assistant/chunk` `reasoning-delta` | |
| `session/update: tool_call` (pending→in_progress) | `session/event` `tool/call` | demux via a Session→sessionId map; `kind` inferred from the tool name |
| `session/update: tool_call_update` (completed/failed) | `session/event` `tool/result` | a throwing `tools/execute` yields NO `tool/result` → fail the pending tool UI from `agent/error`/turn-end |
| `session/request_permission {sessionId, toolCall, options}` | prepended `tools/execute` listener | no-op unless `exec.agent` is ACP-owned; await the outcome; `selected/allow_*` → `next()`; `reject_*`/`cancelled` → veto `ToolExecutionResult{isError}` |
| `session/cancel` (notification) | `agent.cancel(reason)` | the queue-aware cancel (abort running step, clear queued + steering, drop an about-to-start turn); settle the in-flight prompt as `cancelled`; resolve any pending permission as `cancelled` exactly once |
The permission gate is the first real consumer of the `tools/execute` veto seam (the documented "single veto/sandbox/permission seam" plus the deferred "Permission system" TODO in [docs/architecture.md](../../../architecture.md)). It is a single global listener registered with `prepend: true` so it runs before any other tool wrapper. `ToolExecution.agent` is optional and the `Agent` interface carries no origin marker, so the bridge tracks ownership itself: it records each agent it creates in a `WeakMap<Agent, sessionId>` and the gate no-ops (calls `next()` immediately) for any `exec.agent` it does not own — non-ACP agents and the no-agent case pass straight through. For an owned agent it resolves the session, issues `session/request_permission`, and stores the pending resolver on that session's record so the outcome — or a `session/cancel`/connection-close — settles it exactly once.
Lifecycle and disposal: the connection, listeners, and in-flight permission promises register via `ctx.effect`/`ctx.on`; teardown is async and must *reach* quiescence, not just request it — close the connection, settle/reject pending permissions, and dispose each owned agent through its `AgentHandle.dispose()` (which stops the loop, `await`s its exit, and unregisters). Owner teardown goes through that handle seam, not the loop's concrete `agent.done` (which exists only on `ReactLoopAgent`); a non-owner that merely wants to *observe* the current work settling without tearing the agent down awaits the interface-level `agent.whenIdle()`. Every listener contains its `send()` exceptions (log, never reject the turn) because stream chunks are emitted inside the model step, so a throwing listener would corrupt the turn.
**Dependency note (architecture rule).** [docs/architecture.md](../../../architecture.md) states "plugins depend on interface packages, never on `dsh-agent-loop`." Creating and resuming agents is currently only on the concrete `AgentLoop` (`ctx.agentLoop`), so this RFC proposes adding an **abstract create/resume factory** to the `dsh-agent` interface (registry-level `create({ sessionId, meta })` / `resume(...)`), implemented by the loop, so `dsh-acp` injects only `agents` (the interface) and the dependency rule holds. The alternative — injecting the concrete `agentLoop` and recording a documented exception in the architecture doc — is explicitly the non-preferred fallback.
## Plan
1. Package scaffold `packages/ui/acp/` per [the cookbook](../../../cookbook/adding-a-package.md); add `@agentclientprotocol/sdk` and `zod`. Add the abstract create/resume factory to `dsh-agent` (the interface) so the bridge can `inject: ['agents', 'sessions', 'tools', 'sessionPersistence']` without depending on the concrete loop; `sessionPersistence` is required because `session/load` advertises `loadSession: true`. (Fallback only if the factory is judged not worth it: inject `agentLoop` directly and record the architecture-rule exception in `docs/architecture.md`.)
2. Connection plus `initialize`/`session/new`: wire `AgentSideConnection` to stdin/stdout; protocolVersion negotiation; the single-session guard; create the live session through the new `{ sessionId, meta }` factory seam (so the ACP `sessionId` and validated `cwd` become the session's id and header); the `sessionId↔agent` and `Session↔sessionId` maps.
3. Internal edit — turn-end reason fidelity (sanctioned: edit internals to fit ACP). Extend `TurnEndReasonMap` in the proper places: (a) declaration-merge a `max-tokens` variant in the owning package (`packages/core/session/src/types.ts`, alongside `completed|aborted|error|disposed`) — add `max-tokens` because `FinishReasonMap` produces it (DeepSeek maps `length` → `max-tokens`); do not add `refusal`, since no current adapter produces it (unknown DeepSeek finish reasons collapse to `error`), but leave a comment in `TurnEndReasonMap` noting `refusal` should be added when an adapter first emits it (`FinishReasonMap` is merge-extensible); (b) make `agent-loop`'s `loop.ts` populate the reason from the model `finish` chunk — `assembler.finish` lives inside `runStep`, so `runStep` must return it up to `runTurn`, and the rule is "the last step's finish reason wins, but any `max-tokens` in the turn surfaces as `max-tokens`"; (c) no consumer exhaustively switches over `TurnEndReason` today (the invariants plugin switches on `SessionEventType`, and `deriveMessages` ignores `turn/end`), so adding `max-tokens` is a non-breaking extension — but recheck before landing; (d) update [docs/architecture.md](../../../architecture.md) (the CI-verified loop-lifecycle/event-taxonomy doc) and the affected package READMEs/JSDoc (`dsh-session`, `dsh-agent`, `dsh-agent-loop`) per the repo doc-sync policy. This replaces a fragile "observe the finish chunk in the bridge" hack with a real, documented contract.
4. Prompt-turn streaming plus load: translate `session/event` (the `assistant/chunk` token stream plus boundaries and tool activity) into `session/update`; resolve `session/prompt` on settle, mapping the harness `TurnEndReason` to the ACP `StopReason` wire enum (`completed`→`end_turn`, `max-tokens`→`max_tokens`, `aborted`→`cancelled`) — a small total function with a test asserting the exact wire strings, since the SDK rejects an unknown `stopReason`. Concrete correlation, since the loop batches queued messages into one turn and `send()` does not synchronously flip to running: install the `session/event` listener before `send()`; capture the prompt's owning turn from its `turn/start` record, then resolve on that turn's `turn/end` (with `agent/status` idle/disposed as a fallback); reject an empty/whitespace prompt up front rather than calling `send()` (no turn would ever start, so the RPC would hang). Implement `session/load` on the session-persistence resume seam.
5. Permission gate: a single `tools/execute` listener registered with `prepend: true`, owning a `WeakMap<Agent, sessionId>` of bridge-created agents; no-op (`next()`) for unowned/no-agent calls; for owned calls → `session/request_permission` → allow (`next()`) / veto; settle the stored resolver exactly once on outcome, cancel, or connection close.
6. Example wiring (extract a shared base). `@cordisjs/plugin-include` is itself a plugin entry that resets `ctx.baseUrl` and loads a path, so a child `cordis.yml` can nest-include a shared base; the extraction is safe because every dependent plugin declares `inject` (loader groups initialize via `Promise.all`, so YAML order is NOT the dependency mechanism — never rely on it). Extract the provider/tool core (`llm, sessions, system-prompt, tools, agents, invariants, llm-deepseek, bash-local, tool-bash`) into `examples/base.yml`; have both `coding-agent` and a new `examples/acp-agent/` include it and add their own UI plugin plus logger. Keep `agent-loop` per-example (NOT in the base): `AgentLoop` creates its configured agents in its constructor, and the two examples disagree — `coding-agent` needs a pre-created `main` (its `stdio-chat` calls `ctx.agents.get('main')`), while `acp-agent` must pre-create none (ACP `session/new` creates agents). So `coding-agent` declares `agent-loop` with `agents: [{ id: main, … }]` and `acp-agent` with `agents: []`. `acp-agent` loads `dsh-session-persistence-jsonl` (from [session persistence](../../implemented/architecture/2026-06-14-session-persistence.md) — required for `session/load`), omits the stdout logger (see Risks), and adds `pnpm run demo:acp` plus the Zed `agent_servers` snippet.
7. Tests (the repo cares a lot here): a property-based test for the protocol shape (precedent: [property-based testing](../../implemented/testing/2026-06-11-property-based-testing.md)) — fuzz arbitrary harness event sequences and assert ACP-stream invariants (never a `tool_call_update` before its `tool_call`; exactly one `session/prompt` resolution per prompt; monotonic, well-formed ordering; `stopReason` in the legal set); codec unit tests over an in-memory `Duplex` pair (drive `AgentSideConnection` without a subprocess; assert exact frames for `initialize`, `session/new`, a full prompt turn); the mandatory HMR-safety test (dispose the fiber; assert the connection closed, all `ctx.on` listeners gone, any in-flight `request_permission` settled); failure-path tests (connection closes mid-stream; closes with a permission pending; a notification `send()` rejects but the turn survives; `finish{kind:'error'|'aborted'}`; a `tools/execute` throw with no `tool/result`; a second `session/new` rejected; a `session/prompt` while one is in flight; an empty prompt rejected without hanging; a `session/load` re-derives identical history and replays it); and an e2e (`*.e2e.ts`, self-skips without `DEEPSEEK_API_KEY`) that boots `examples/acp-agent`, connects a `ClientSideConnection`, sends a real prompt, owns and disposes the harness in `afterEach`, and verifies the world (files on disk), not the agent's self-report.
8. Docs: module/JSDoc plus a package README; extend [the extension cookbook](../../../cookbook/extension-cookbook.md) with the client-driver pattern. Flip Status to `implemented` on landing; record a decision in this RFC only if it proves durable, contested, and surprising (candidates: the `tools/execute` permission-ownership rule, the npm-dependency choice) — not auto-required.
Deferred (each names its owning future work):
- Multiplexing concurrent sessions → [ACP multi-session](2026-06-14-acp-multi-session.md).
- ~~`cwd` honoring.~~ **RESOLVED.** Originally there was no path from `session/new.cwd` to the bash workdir (`tool-bash` forwarded only an explicit `args.workdir`; `LocalBashExecutor.resolve` defaulted to its own config or `process.cwd()`), so the MVP validated `cwd` (require absolute) AND required the server to launch in the workspace root, erroring on a mismatch. This is now lifted: the validated `cwd` is stored as `SessionHeader.cwd`, and `dsh-tool-bash` defaults the bash workdir to the calling agent's `session.header.cwd` (an explicit model `workdir` still wins; a relative one resolves against it). Any absolute `cwd` is honored — the server need not launch in the workspace, and N sessions can each target a different directory. Widening scope beyond the single cwd (`additionalDirectories`) remains deferred.
- Client `terminal/*` proxying (a live editor terminal) and `fs/*` (editor-rendered diffs) — a future `BashExecutor` over the [capability seams](../../implemented/architecture/2026-06-13-capability-seams.md) bash seam, gated on `clientCapabilities.terminal`.
- Image/audio prompts (blocked on the DeepSeek adapter, which skips `image` blocks today), modes, auth, `available_commands`/slash-commands, `plan`, and `usage_update`.
## Alternatives considered
- **A process-wide stdout hijack inside `dsh-acp`** (defensively monkey-patching `console.log` / `process.stdout.write`) — rejected: it lives outside Cordis' effect-scoped, HMR-friendly plugin model, races the connection's own stdout handoff, and fights the logger. The stdout guarantee is config-only.
- **Injecting `agentLoop` directly instead of the abstract create/resume factory** — the recorded fallback, taken only if the factory seam is judged not worth it, with the architecture-rule exception recorded in `docs/architecture.md`.
## Acceptance criteria
- The `acp-agent` example speaks ACP over stdio end-to-end: `initialize`, `session/new` with a validated absolute `cwd` honored as the session workspace, streamed `session/update` frames per prompt turn, `session/load` re-deriving identical history, and `session/prompt` resolving with the correct wire `stopReason`.
- stdout carries only framed JSON-RPC (asserted by test); the permission gate settles every `session/request_permission` exactly once — on outcome, cancel, or connection close.
- The plan's test set runs green: the property-based protocol invariants, the codec unit tests over an in-memory duplex pair, the HMR-safety test, the failure-path matrix, and the self-skipping real-API e2e that verifies the world.
## Risks
stdout is the protocol — guaranteed by config, not by monkey-patching. The console logger writes through `console.log` to stdout, so any stdout UI/logger plugin corrupts JSON-RPC. The guarantee is config-only: the `acp-agent` example loads no stdout plugin (no console logger, no `stdio-chat`) and, if logging is wanted, uses a stderr exporter. A defensive process-wide `process.stdout.write`/`console.log` hijack inside `dsh-acp` is explicitly rejected — it lives outside Cordis' effect-scoped, HMR-friendly plugin model, races the connection's own stdout handoff, and fights the logger. A test asserts the example emits only framed JSON-RPC on stdout.
New third-party runtime dependency plus protocol drift: `@agentclientprotocol/sdk` is young (0.25.x, recently renamed) and evolving. Pin the version and isolate churn to the one bridge package. This is not a vendoring-policy violation — [vendoring Cordis as source](../../implemented/process/2026-06-11-vendor-cordis-as-source.md) vendors the framework; genuine third-party deps already live on npm (`@earendil-works/pi-ai`).
Turn-settle and prompt-correlation hazards: honor "queued messages batch into one turn" and "`send()` does not synchronously flip to running" (see `stdio-chat.ts` and the defensive-patterns section of [docs/architecture.md](../../../architecture.md)); gate resolution on an observed running→idle transition and handle the empty-prompt / no-work branch so an RPC can't hang.
Permission-await and disposal hangs: a pending `request_permission` whose connection closes or whose turn aborts must settle exactly once; disposal must reach quiescence — tear each owned agent down through `AgentHandle.dispose()` (which stops the loop and awaits its exit), rather than orphaning awaits on a closed pipe.
The 100% per-file coverage gate (repo policy) makes a branch-heavy protocol bridge real work. Accepted deliberately, surfaced so it isn't a surprise at PR time.
ACP protocol-shape details (exact method names, `session/update` variants, permission option kinds, stop reasons) are taken from the ACP spec and the `@agentclientprotocol/sdk` types; they are not independently verifiable until the dependency is added, so the implementation pins the SDK version and conforms to its types rather than to this RFC's prose where they differ.

48
docs/rfc/proposed/feature/2026-06-14-acp-multi-session.md
View File

@@ -1,48 +0,0 @@
# RFC: Multiplex concurrent ACP sessions over one connection
Status: proposed
> **Implementation status:** the multi-session bridge (steps 1, 3, 4) and the bash task-ownership isolation are implemented in `packages/ui/acp` + `packages/bash/tool-bash`. **Per-session *permission* ownership is deferred** — it depends on [the ACP support permission gate](2026-06-14-acp-agent-client-protocol.md) (`TODO(rfc010-permission-gate)`), which is itself deferred; the `agent→sessionId` reverse map the gate will route through is in place. Step 2's per-session disposer scope is now implemented (see [agent lifecycle & ownership seams](../../implemented/architecture/2026-06-18-agent-lifecycle-and-ownership-seams.md)): the factory returns a per-agent `AgentHandle` whose `dispose()` stops the loop, awaits quiescence, unregisters the agent, and removes its session, so a bare client disconnect leaves no registered agent or session-store entry. Status stays `proposed` until per-session permission ownership lands.
> **Target-client note:** Zed is the current target ACP client, and its ACP client maintains a `HashMap<SessionId, AcpSession>` plus `pending_sessions` for concurrent `session/load` calls. The competing simplification to return to one live session per connection was rejected after checking that target-client shape; this RFC remains the path for finishing multiplexing and per-session permission ownership. See [the rejected simplification](../../rejected/simplification/2026-06-20-single-session-acp-bridge.md).
## Problem
[ACP support](2026-06-14-acp-agent-client-protocol.md) ships with a single active session per connection: a second `session/new` is rejected. Editors expect to run several conversations over one agent subprocess — a user opens multiple threads, or a client pre-warms sessions. The single-session guard is a deliberate MVP scope cut, not an architectural limit; this RFC lifts it.
This paragraph is historical: the multi-session bridge has landed. The remaining proposed work is per-session permission ownership plus the lifecycle seams now tracked in [agent lifecycle and ownership seams](../../implemented/architecture/2026-06-18-agent-lifecycle-and-ownership-seams.md).
## Proposal
The harness core already supports many agents (`AgentRegistry.list()` and `AgentLoop.create` impose no count limit), so multiplexing is a bridge-layer change in `@deepseek-ai/dsh-acp`, not a loop or core change.
- Lift the single-session guard in `session/new`; allow N live sessions, each mapped to its own `ReactLoopAgent`.
- The bridge's `sessionId→agent` and `Session→sessionId` maps (introduced single-entry by [the ACP support RFC](2026-06-14-acp-agent-client-protocol.md)) become true multi-entry, plus a third `agent→sessionId` reverse map: the `tools/execute` permission gate receives only `exec.agent` (no sessionId), so it needs an O(1) reverse lookup to find the owning session. Every `agent/*` event and every `session/event` is demuxed strictly by id, so two sessions streaming at once never interleave their `session/update` notifications.
- Per-session prompt queues: [the ACP support RFC](2026-06-14-acp-agent-client-protocol.md)'s single-entry in-flight-prompt state becomes multi-entry — one in-flight prompt *per session*, tracked per `sessionId`.
- Per-session cancel routing: `session/cancel` cancels only its own session's agent (via the queue-aware `agent.cancel()`) and settles only that session's in-flight prompt. The cancel is scoped to that one agent — a per-agent `AbortController` for the running step plus the agent's own queued/steering FIFOs — so it never touches another session's stream or pending prompt.
- Per-session permission ownership: a `session/request_permission` and its outcome are bound to the originating session via the reverse map, so a permission prompt or a cancel in one session can never resolve another session's pending permission.
## Plan
1. Generalize the two id maps to multi-entry and add the `agent→sessionId` reverse map; add a per-session record holding the agent, the in-flight-prompt state, the pending-permission registry, and the session's disposer scope (see step 2).
2. Give each session a real per-session disposer scope, NOT `ctx.extend()` — in Cordis `ctx.extend()` only creates a child context/prototype, but `ctx.on()` registered on it is still owned by the current plugin fiber, so disposing it would not remove that session's listeners. Use a genuine child fiber (load a per-session sub-plugin, e.g. `ctx.plugin(...)` returning a fork, or collect each session's `ctx.on` disposers in its session record and call them on teardown). Demux every `agent/*` and `session/event` by id into the right session record. Note the single global `tools/execute` listener stays on the bridge root (it must see all agents) and routes via the reverse map.
3. Lift the `session/new` guard; keep `session/load` ([from ACP support](2026-06-14-acp-agent-client-protocol.md)) working per session.
4. Tests for cross-session isolation: two sessions streaming and permission-prompting concurrently never interleave; a cancel/abort in one session leaves the other's stream and pending permission untouched; per-session in-flight-prompt enforcement holds independently; disposing one session leaves the others running.
## Alternatives considered
**A per-session `ctx.extend()` scope** — rejected: in Cordis, `ctx.extend()` only creates a child context/prototype, and `ctx.on()` registered on it is still owned by the current plugin fiber, so disposing it would not remove that session's listeners. A genuine child fiber (or a per-session collection of disposers) is required.
## Acceptance criteria
- N concurrent sessions stream and permission-prompt without interleaving their `session/update` notifications; a cancel in one session leaves every other session's stream, queued prompts, and pending permissions untouched.
- Disposing one session removes exactly its own listeners; connection teardown reaches quiescence across all sessions.
- One session's agent cannot read or kill another session's background bash task.
## Risks
Listener fan-out cost: each session adds listeners; ensure disposal of one session removes exactly its own and the connection teardown ([from ACP support](2026-06-14-acp-agent-client-protocol.md)) still reaches quiescence across all sessions.
The subtle correctness trap is cross-session leakage — a cancel or abort on one session settling another session's pending permission. The per-session permission ownership rule (routed via the `agent→sessionId` reverse map) and its isolation test are the guard.
Shared background-task state: the bash executor's task ids are global and predictable (`bash-1`, `bash-2`, …), and `bash_output`/`bash_kill` look up by id without checking the caller. Under one session this is benign; under N sessions one session's agent could read or kill another's background task. This is a pre-existing `tool-bash` gap that multi-session turns into a real isolation hole — fixing it (validate the caller against the task owner) belongs with this RFC or a companion `tool-bash` change.

139
docs/rfc/proposed/feature/2026-06-15-code-mode.md
View File

@@ -1,139 +0,0 @@
# RFC: Code Mode — the model writes TypeScript against the tool registry
Status: proposed
## Problem
Today the agent loop advertises every registered tool to the model as a native JSON-schema function definition. `ToolRegistry` contributes its schemas to the system-prompt assembly, the assembly's `tools` land on the wire (and in the logged request header), the model invokes one `tool-call` block per step, and the loop dispatches each call through `ctx.tools.execute()` **sequentially** (parallel tool execution is an explicit open TODO in `dsh-tools` and [docs/architecture.md](../../../architecture.md)), with **every** intermediate `tool-result` re-entering the model's context on the next request.
For multi-step tool work this is token-heavy and serial. The model cannot compose tools — loop over a result set, branch on an intermediate value, fan out, post-process — without a full model round-trip per call, and each round-trip drags the entire intermediate result back into context whether the model needs it or not.
Cloudflare's [Code Mode](https://blog.cloudflare.com/code-mode/) proposes an alternative grounded in a simple observation: LLMs are better at writing code than at emitting tool calls, because they have seen millions of lines of real code and comparatively few contrived tool-calling traces. Instead of one tool call per step, the model writes a TypeScript program against a generated API over the tools, the program executes in a sandboxed runtime, and the model curates what comes back — only what it prints or returns — instead of every intermediate result.
An earlier draft of this RFC designed Code Mode as an add-on consumer plugin with zero core changes, deferring the execution substrate to a follow-up. Both constraints are dropped here, deliberately. First, the harness is pre-release and optimizes for the correct foundation over blast radius: tool presentation is the registry's own concern, and bolting a second presentation onto it from outside means transforming the registry's contribution after the fact — a waterfall listener whose correctness depends on listener ordering, which fights the [reconstructable-requests](../../implemented/architecture/2026-07-05-reconstructable-requests.md) design instead of riding it (that refactor removed request mutation from `agent/request`, the seam the old draft relied on). Second, the substrate question is answerable now: a Node `worker_threads` runtime gives real containment — separate isolate, empty environment, heap caps, and a `terminate()` that reliably stops a hot synchronous loop — where the old draft's `node:vm` stub had none of those, and it fits the harness's existing trust model (§Trust posture) without a hardening follow-up.
## Proposal
Three decisions, each elaborated in its own section below:
1. **Code Mode is a first-class presentation mode of `ToolRegistry`** (`dsh-tools`), selected by a validated `mode` config: `'native'` (today's behavior, the default), `'code'` (the wire carries exactly one tool, `run_code`, plus a generated SDK `.d.ts` in the system prompt), or `'both'` (native schemas *and* `run_code` + SDK). The registry's existing tool-schema provider contributes whatever the mode dictates, so the wire tool list is shaped at its source — no interception, no listener-ordering caveats — and the logged request header records it for free.
2. **Code execution is a capability seam** — a new group `packages/code-runtime/` with the interface package `@deepseek-ai/dsh-code-runtime` owning `ctx.codeRuntime` ([capability seams](../../implemented/architecture/2026-06-13-capability-seams.md); consumer = `dsh-tools`, with core-consumes-a-seam precedent in `agent-loop` → `dsh-llm`). The runtime knows nothing about tools: it is handed a program and named async bindings, runs the program, and reports `{ value, logs, error? }`. Language and substrate are backend properties, so a future Python or container backend is a new implementation package, not a redesign.
3. **The shipped implementation is `@deepseek-ai/dsh-code-runtime-worker`**: one fresh Node worker thread per run, executing the model's TypeScript after type-strip, with bindings bridged over the message port, an empty environment, configurable heap/output/time caps, and hard termination. Its trust posture is bash-equivalent by design — no unsafe-acknowledgement flags — because the harness already ships `dsh-bash-local`, which executes arbitrary model-written shell commands with strictly *more* ambient authority.
### The registry owns the mode
`ToolRegistry` gains a schemastery-validated config (`static Config`), its first: `mode: 'native' | 'code' | 'both'`, default `'native'`. A deployment flips it from `cordis.yml` (`tools: { mode: code }`) — no code edit, per the no-hardcoded-tunables convention.
**Wire tool list = the registry's contribution.** The registry already feeds the assembly through `ctx.systemPrompt.tools(() => this.schemas())`; the provider becomes mode-aware: `'native'` contributes all schemas (unchanged), `'code'` contributes only `run_code`'s schema, `'both'` contributes all schemas plus `run_code`. Because [`PromptAssembly.tools` is the single source the loop's request header snapshots](../../../../packages/core/system-prompt/src/index.ts), the collapse is automatically logged and reconstructable — model-visible ⟺ logged holds with zero new mechanism. Scope of the guarantee, stated honestly: the mode governs the **registry's** contribution, and the registry is the only shipped schema source — but `systemPrompt.tools()` is a public multi-provider API and the `system-prompt/assemble` waterfall may transform the assembly, so a deployment that wires a second direct provider (or a mutating listener) owns what it adds, exactly as in native mode. Those are deliberate acts; what the design eliminates is the *accidental* leak the old draft worried about — a listener-ordering race around an after-the-fact collapse — and the shipped-configuration invariant (`'code'` ⇒ assembled tools exactly `[run_code]`) is pinned by tests and, like every request, by the logged header.
**Interaction with `toolOrder`, stated up front:** a configured `systemPrompt.toolOrder` naming native tools rejects every assembly under `mode: 'code'` (those names are no longer contributed), by the existing fail-loud rule for unlisted names. This is correct behavior, not a bug: a deployment switching modes updates its order config or drops it.
**The SDK prompt section.** Under `'code'` and `'both'` the registry registers one lazy prompt section (`tools:sdk`, in the 100–199 tool-guidance order band) whose thunk regenerates, at each assembly, a TypeScript declaration of every registered tool except `run_code` itself, plus fixed usage instructions. The thunk reads the live store and emits tools in lexicographic name order, so its output is deterministic and stable across steps — an unchanged tool set produces byte-identical text (prefix-cache-friendly; a mid-session registration surfaces as one logged header delta, exactly like a native-mode tool change).
**Codegen.** A pure `jsonSchemaToTs(schema)` module inside `dsh-tools` (sibling of `json-schema.ts` — `schemas()` and the SDK are two projections of the same store) maps the JSON-Schema subset the `defineTool` DSL emits (object/string/number/boolean/array, `properties`, `required`, string `enum` → literal union, nested objects, array `items`, `description` → JSDoc) to a TS type literal. It is **total**: any construct outside that subset (`$ref`, `oneOf`/`anyOf`, `integer`, future MCP shapes, …) degrades to `unknown` without throwing. Because `ToolSchema.name` is an arbitrary string, the SDK is declared as one object constant — `declare const tools: { "some-mcp-tool"(args: …): Promise<string>; bash(args: …): Promise<string>; … }` — quoted keys make every name reachable with no sanitization or alias-collision logic. Typing is advisory (the runtime executes type-stripped JS); the instructions say so.
### The run_code tool and the dispatch bridge
Under `'code'` and `'both'` the registry registers `run_code` in itself as an ordinary tool — one required parameter `{ code: string }` — so the unchanged loop dispatches it through the normal pipeline and `tools/pre-execute` / `tools/post-execute` gate it like any other call (a permission plugin can inspect the program text before it runs). Its `execute(args, exec)`:
1. **Builds the bindings**: the bridge owns a **run-scoped `AbortController`** whose signal follows `exec.signal` (an outer cancel propagates in) and which the bridge itself fires the moment the run settles for any reason — completion, program exception, `computeMs`/`maxWallMs` expiry, worker exit. For every registered tool except `run_code`, the binding is an async function that (a) checks the run signal before and after (throwing stops the program — necessary because `ctx.tools.execute()` converts errors to `isError` data), (b) **JSON-normalizes the argument** — a `JSON.parse(JSON.stringify(args))` round-trip, rejecting that one call with a descriptive `Error` when the value does not survive (`BigInt`, circular structures) — because the seam's structured-clone boundary is wider than JSON while the session log accepts only JSON: normalizing BEFORE dispatch makes the dispatched form and the logged form the same JSON value by construction, so an executed sub-call can never fail at logging time, (c) awaits its turn on the **per-run serialization queue** (below), (d) calls `this.execute({ callId, name, arguments, agent: exec.agent, signal: runSignal })` with a deterministic sub-id `` CallId(`${exec.callId}:code:${n}`) `` — the run signal, not the bare outer one, so a budget expiry aborts an in-flight sub-tool (`bash-local` kills on its spec signal) instead of orphaning it, (e) appends a `tool/code-dispatch` session event, and (f) maps the result: success → the text-block contents joined as a `string` (non-text blocks become placeholders, an MVP limitation), `isError` → **the binding rejects** with an `Error` carrying the result text. Rejection is the deliberate model-facing contract — real code signals failure by throwing, `try/catch` and `Promise.all` short-circuiting behave as every model has seen them behave — where the old draft's `{ output, isError }` envelope made error handling a bespoke convention.
2. **Runs the program**: `ctx.codeRuntime.run({ program: args.code, bindings: [{ global: 'tools', functions }], signal: exec.signal })`.
3. **Surfaces the outcome — after reaching quiescence.** When `ctx.codeRuntime.run()` resolves, the bridge fires the run-scoped abort (cancelling any in-flight sub-dispatch and abandoning queued-unstarted ones), then **awaits the dispatch queue's drain before returning**, per the dispose-to-quiescence rule in [defensive patterns](../../../defensive-patterns.md): an aborted in-flight sub-call still settles and logs its `isError` `tool/code-dispatch` event *inside* the open turn, and nothing can append after `run_code` returns. A successful run then returns one text block — the captured console/stdout output followed by the rendered return value (if any) — plus a `meta` payload (capped logs, dispatch count) for presentation. A run with `result.error` throws a `CodeRunFailedError extends HarnessError` (`code: 'CODE_RUN_FAILED'`, message = the error kind and text plus captured logs so the model can self-correct); the registry's existing catch turns it into a structured `isError` result.
**Sub-call `additionalContext` is suppressed, deliberately.** A `tools/post-execute` hook may attach `additionalContext` to a call; for loop-dispatched calls the loop buffers those and appends each as a `context/message` only after the step's `tool/result`s, preserving call/result adjacency. A sub-dispatch result's `additionalContext` has no such safe outlet from inside a running `run_code`: injecting immediately would land a `context/message` between the parent's `tool/call` and its `tool/result` (breaking the adjacency the buffering exists to protect), and `PostToolDecision.additionalContext` is singular where a program may produce many. The MVP therefore drops sub-call `additionalContext`, pinned by a test and stated in the hooks bridge's docs; the follow-up (a plural context channel or loop-level sub-dispatch buffering) is deferred until a real hook needs it through Code Mode.
**Concurrency: serialized, enforced by the binding.** The bindings are async, so a model writing `Promise.all([tools.a(…), tools.b(…)])` starts both immediately — concurrent dispatch would be the *default*, while the tool contract still carries no concurrency-safety metadata (the open parallel-execution TODO). Each `run_code` invocation therefore owns a dispatch queue and every binding call chains onto it, so even `Promise.all` executes the underlying `ctx.tools.execute()` calls one at a time in submission order; when the run settles, queued-but-unstarted dispatches are abandoned. Lifting this per-tool once tools can declare themselves concurrency-safe is deferred work, same as before.
**Presentation.** `run_code`'s render intent is decided here per the [render-intent RFC](../../implemented/architecture/2026-07-02-tool-render-intent-union.md): `presentCall` → a `generic` card, `kind: 'execute'`, title `Run code`, `rawInput` = the program text; `presentResult` → a `generic` card whose content is the captured output (from `meta`). Not a `terminal` card: that card's semantics are "a shell command in a working directory", which a program is not.
### Observability: `tool/code-dispatch`
Each sub-dispatch appends one session event, declared by `dsh-tools` via `SessionEventMap` declaration merging (the map is merge-extensible for exactly this; `todo/write` is the log-only precedent): `tool/code-dispatch` with `{ parentCallId, subCallId, name, arguments, isError, resultSummary }` — `arguments` being the bridge's JSON-normalized value, the very one dispatched, so the append cannot fail on payload shape. It is log-only — `deriveEventMessage()` ignores unknown event types by design, so sub-calls never re-enter model context — but persistence and UIs get every call. As a log event it carries JSDoc prose but **no `@mode` tag** (that vocabulary belongs to cordis bus events; the persistence-catalog generator hard-errors on one) and lands in the regenerated `docs/persistence-catalog.md`; appends happen inside `run_code`'s execution, so the turn-enclosure invariant is satisfied by construction. A `run_code` execution arriving without `exec.agent` (the loop always supplies it; direct programmatic calls may not) still runs and simply skips event logging, exactly as the `ToolExecution` contract allows.
### The code-runtime seam
`packages/code-runtime/code-runtime/` — `@deepseek-ai/dsh-code-runtime`, depending only on `cordis`. An abstract `CodeRuntime extends Service` (`super(ctx, 'codeRuntime')`) plus the vocabulary:
- `CodeRunRequest = { program: string; bindings: CodeBindingNamespace[]; signal?: AbortSignal }`
- `CodeBindingNamespace = { global: string; functions: Record<string, (args: unknown) => Promise<unknown>> }` — the runtime exposes each namespace as a global object of async functions inside the program; binding arguments and resolutions must be structured-cloneable (a runtime may cross a serialization boundary; ours does).
- `CodeRunResult = { value?: unknown; logs: CodeLogEntry[]; error?: CodeRunFailure }` — an error is a field on a resolved result, never a rejection of `run()`.
- `CodeLogEntry = { source: 'console' | 'stdout' | 'stderr'; level?: 'log' | 'info' | 'warn' | 'error' | 'debug'; text: string }`
- `CodeRunFailure = { kind: 'exception' | 'timeout' | 'abort' | 'worker-exit'; message: string }` — orthogonal outcomes reported independently per [defensive patterns](../../../defensive-patterns.md); a timed-out run is not an exception, an abort is not a timeout.
- Two readonly backend descriptors, informational not gating: `language` (what the program must be written in — `'typescript'` for the shipped backend; a Python backend would say so, and pair with its own SDK generator on the presentation side) and `isolation` (`'worker-thread'` for the shipped backend; `'process'`, `'container'`, … for future ones). `dsh-tools` requires `language === 'typescript'` in the MVP — its codegen emits TS — and fails the assembly loudly otherwise, the same misconfiguration idiom as `toolOrder` violations (as when `mode` is non-native with no `ctx.codeRuntime` loaded at all).
Per explicit-over-implicit at seams, the request spells out everything the runtime acts on; defaulting (timeouts, caps) is the implementation's validated config, never a hidden `??` inside `run()`. Consumption uses the loop's established optional-backend idiom: cordis has no optional injection — every `inject` entry gates activation — so a static `inject` on the registry would hold `ctx.tools` (and every tool plugin behind it) hostage to a code runtime existing even under `mode: 'native'`; instead the registry reads `ctx.get('codeRuntime')` at use time, exactly as `agent-loop` consumes `sessionPersistence`, with absence failing loud in the provider thunk as above. The seam split is justified by real planned divergence on both axes — substrate (worker now; container/microVM later) and language (the Python/AssemblyScript direction sketched in the earlier draft survives as future work) — not by speculation: `dsh-tools` consumes the interface today and tests against a trivial in-repo fake, exactly the interface/implementation/consumer shape of the bash template.
### The worker-thread runtime
`@deepseek-ai/dsh-code-runtime-worker`, the second package of the `packages/code-runtime/` group. Per `run()`:
1. **Type-strip host-side** with Node's built-in `stripTypeScriptTypes` (`node:module`; present across the repo's whole engines range, `^22.19.0 || >=24.0.0`, and position-preserving, so runtime error line numbers match the model's source). Strip-only mode rejects non-erasable syntax (`enum`, namespaces) — that rejection returns as `error.kind: 'exception'` with Node's message, the SDK instructions say "erasable TypeScript only", and the model self-corrects like any other program error. A syntax-level failure never spawns a worker.
2. **Spawn one fresh `Worker` per run** from the package's own bootstrap module: `env: {}` (truly empty — stronger than the scrubbed-env rule for spawned commands), `resourceLimits` from config, `stdout`/`stderr` captured into `logs` rather than inherited. No pooling and no cross-run state: a program's world dies with its worker, which keeps runs reconstructable from the log alone and makes state bleed unrepresentable.
3. **Execute** in the bootstrap: the stripped program becomes the body of an `AsyncFunction` whose parameters are the binding globals and a capturing `console` shim, so top-level `await` and `return` work and the program's completion value is the run's `value` (structured-cloneable values cross as-is; anything else is replaced by its `util.inspect` rendering, documented).
4. **Bridge bindings over the message port**: each binding function in the worker posts `{ id, global, name, args }` and awaits the reply; the host validates the name against the request's bindings, invokes, and replies `{ id, ok, value }` or `{ id, ok: false, message }` (a host-side binding rejection becomes a program-side rejection). The worker-side namespace objects are built null-prototype via `defineProperty`, so a binding named `__proto__`, `constructor`, or `toString` is an ordinary own property, not a prototype collision. Unknown names, duplicate ids, and post-settlement messages are rejected or ignored — the port protocol assumes a hostile peer, because the peer runs model code.
5. **Enforce caps — two independent budgets, because the peer is hostile.** The compute budget (`computeMs`) meters the worker's *measured busy time* via `worker.performance.eventLoopUtilization()` polling — not host-side "is an RPC pending" bookkeeping, which a program defeats by firing an un-awaited call at a slow tool and then spinning hot while the host thinks it is waiting. Measured busy time cannot be gamed: a hot loop accrues it whether or not a dispatch is in flight, and a program genuinely awaiting a slow tool accrues none, so a long-running `bash` sub-call still does not kill an innocent run. The wall ceiling (`maxWallMs`) never pauses for anything and backstops what busy-time cannot see (a program awaiting a promise nobody will resolve). Budget expiry, `signal` abort, and run completion all funnel into `worker.terminate()`, which ends hot synchronous loops too (measured; this was `node:vm`'s unfixable gap); the failure reports which budget fired. Heap overflow surfaces as the worker's OOM exit → `error.kind: 'worker-exit'`. Log and value sizes are capped by config, truncation marked in-band. All caps are validated config fields with defaults (`computeMs: 60_000`, `maxWallMs: 600_000`, `maxLogBytes: 65_536`, `maxValueBytes: 32_768`, `maxOldGenerationSizeMb: 512`), changeable from `cordis.yml`.
6. **Dispose to quiescence**: the service's own disposal terminates in-flight workers and *awaits* their exits before resolving, per [defensive patterns](../../../defensive-patterns.md).
### Trust posture
The worker runtime is **containment, not a security boundary**, and the RFC says so without ceremony. Model code in the worker can reach Node globals — `fetch`, `process` (with an empty env), dynamic `import()` of built-ins — so a deliberately adversarial program has ambient authority comparable to what the harness's own `bash` tool already grants every model turn: `dsh-bash-local` runs arbitrary model-written commands with the host filesystem, network, and a scrubbed-but-populated environment. One asymmetry runs the other way and is stated plainly: `worker.terminate()` ends the thread, not OS processes a program may have spawned via `node:child_process` — weaker than `bash-local`'s process-group kill for direct children (equivalent for double-forked daemons, which survive both); the wall-clock ceiling bounds the worker itself, and orphan cleanup is the same deployment-level concern it already is for bash. Code Mode is gated where bash is gated — `tools/pre-execute`, where permission/sandbox plugins veto or approve the program before it runs — and adds containment bash does not have: empty env, heap caps, hard termination of the program itself, a separate isolate. The earlier draft's two-flag unsafe ceremony (`{ unsafe: true }` constructor + `allowUnsafeRuntime`) existed for a `node:vm` stub with *no* containment and is dropped with it; demanding scarier flags for the better-contained executor than for bash would be posture theater. A deployment that needs a hard boundary (untrusted multi-tenant input) needs it for bash too; that is a future `isolation: 'container'` backend, and the `isolation` descriptor exists so such a deployment can tell backends apart.
### What the model sees
The `tools:sdk` section carries the `.d.ts` plus fixed instructions: the program is the body of an async TypeScript function (erasable syntax only — no `enum`/namespaces; type annotations are advisory); call tools as `await tools.name(args)` (quoted access for exotic names); a failed tool call **rejects** with an `Error` carrying the tool's error text — catch it to handle and continue; calls run **sequentially** even under `Promise.all`; emit results via `return` and/or `console.log`, and only that curated output returns to the context — intermediate tool results never do. That last line is the payoff the whole design serves: output-side context cost becomes the model's own editorial decision. On the input side the `.d.ts` is not free — for a large tool surface it can rival the native JSON schemas it replaces (and `'both'` pays for the two side by side) — but it is prefix-stable, so provider prefix caching amortizes it; the win is workload-dependent and the RFC claims no more.
## Plan
Four stacked PRs, each gates-green (`typecheck`, `lint`, `test:coverage`, `test:snapshot`, `doc-sync`, `verify-module-graph`, `build`, `hygiene`) with docs updated in the same change:
1. **This RFC revision** (docs-only): the file rewritten as above (renamed `2026-06-15-code-mode.md`, same first-proposed date), regenerated RFC index.
2. **`dsh-code-runtime`** (interface package): the group `packages/code-runtime/`, abstract `CodeRuntime`, vocabulary types, ctx-key declaration; docs in the same change — group README + package README, the `packages/README.md` group table row, the `ctx.codeRuntime` row in [docs/architecture.md](../../../architecture.md)'s service map, and the regenerated cordis catalog (the new service class). Unit tier: HMR safety (dispose removes `ctx.codeRuntime`), contract docs. This package has no behavior to snapshot or e2e; its coverage story is unit-only by design.
3. **`dsh-code-runtime-worker`**: the implementation above, plus the regenerated config catalog (its `Config`). Unit tier (real workers, no mocks — they are cheap and local): output/value capture, log source attribution, error kinds (exception incl. non-erasable syntax, abort, worker-exit under OOM), the two budgets from both sides (a hot loop with an un-awaited pending dispatch still dies at `computeMs` busy time; a program idling on a slow binding outlives `computeMs` and dies only at `maxWallMs`), binding bridge hostility cases (unknown name, duplicate id, post-settlement message, `__proto__`/`constructor`/`toString` binding names), structured-clone fallback, cap truncation, `env` emptiness verified from inside the program, dispose-awaits-exit. A real-load-path test runs the built package (`lib/`) so the worker entry resolves both unbuilt (tsx) and built — the published-bin guard from [docs/testing.md](../../../testing.md).
4. **Native code mode in `dsh-tools`** + the end-to-end surface: mode config, provider switch, `tools:sdk` section, `jsonSchemaToTs`, `run_code` + dispatch bridge + serialization queue, `tool/code-dispatch` event; regenerated tool, config, and persistence catalogs; [docs/architecture.md](../../../architecture.md) (tool-pipeline prose) and the [adding-a-tool cookbook](../../../cookbook/adding-a-tool.md) cross-reference updated in the same change; an `examples/` leaf + `demo:code` script wiring the worker runtime with `mode: 'code'`; move this RFC to `implemented/`. Coverage named per tier now, per the plan-time rule: **unit** — codegen table (DSL subset, quoted names, `unknown` degradation, determinism), provider contribution per mode, `toolOrder × mode` rejection, missing-runtime/wrong-language loud failures, serialization non-overlap (a probe tool records enter/exit under `Promise.all`), abort stops the queue, binding rejection on `isError`, `CodeRunFailedError` → structured `isError`, event payloads, `deriveMessages()` ignores the event, HMR safety (mode flip via config reload removes tool + section); **e2e (with-key, self-skips)** — a real model, `mode: 'code'`, a task requiring two tool calls and curation, asserting the wire tool list was exactly `run_code` and the transcript's dispatch events; **snapshot (keyless replay)** — goldens for a `run_code` turn in `'code'` and `'both'`, pinning the SDK section text, the collapsed header tools, dispatch events, and the result card.
The four PRs land in order (each on the previous); per stacked-review practice, review fixes land on the PR that introduced them and merge down.
## Alternatives considered
**An add-on consumer plugin, zero core changes (the previous draft of this RFC).** Rejected on both halves. The wire-collapse half aged out from under it: it targeted the `agent/request` waterfall, which [reconstructable requests](../../implemented/architecture/2026-07-05-reconstructable-requests.md) has since re-typed to call-config-only, and the surviving alternative — transforming the assembly a waterfall listener receives — is strictly worse than contributing the right list in the first place (transformation must undo `toolOrder` canonicalization it cannot see the config for, and its correctness depends on where it sits in a listener chain). The deeper reason is ownership: which tools the model is offered, in which representation, is the registry's single concern — `schemas()` for function calling and the SDK for Code Mode are two projections of one store, and splitting the second projection into a satellite package would preserve a boundary the domain does not have.
**`node:vm` as the reference runtime, hardening deferred (also the previous draft).** Rejected: `node:vm` is not isolation (prototype-chain escapes reach the host realm), cannot interrupt a hot loop, and forced the draft into a two-flag unsafe ceremony plus a mandatory follow-up RFC. The worker thread delivers the missing properties now — separate isolate, empty env, `resourceLimits`, reliable `terminate()` (all verified by probe before this revision) — at bash-equivalent trust, so the reference implementation and the production one are the same package and the ceremony dissolves.
**Result elision / summarization over native tool-calling.** Addresses only the context-bloat half of the problem: trimming old `tool-result`s (now cheap to add as a logged surface replace, per the reconstructable-requests consequences) still pays one model round-trip per call and cannot express loops, branches, or joins. Complementary, not competing; it can layer under Code Mode for residual native calls.
**Parallel native dispatch in the loop.** The other answer to round-trip cost; still valid future work (the open TODO), still blocked on concurrency-safety metadata, and still no composition — it parallelizes calls the model already decided on in one step. Code Mode's serialized-queue decision keeps the two compatible: when the metadata lands, both native parallel dispatch and per-tool binding parallelism unlock together.
**Always-exclusive (Cloudflare-faithful, no mode).** Rejected for this SDK's primary consumer: a coding agent's bread-and-butter single calls (`bash`, `read`, `edit`) are already ideal as native calls, and forcing every edit through a program taxes the common case. The mode config keeps the faithful form (`'code'`) one line away without imposing it.
**Per-tool visibility tiers (this tool native, that tool code-only).** Deferred again, knowingly: it needs per-tool metadata and a presentation split that `'native' | 'code' | 'both'` does not, and every learning it depends on (how models actually split usage under `'both'`) arrives only after this ships.
**Sanitized identifier aliases in the SDK** (`my-tool` → `my_tool`, Cloudflare's approach). Rejected: quoted keys on a `declare const` make every name reachable with zero alias-collision logic; models handle `tools["my-tool"](…)` fine.
**A REPL-style persistent kernel** (state survives across `run_code` calls). Rejected for the MVP: cross-call state would be invisible to the session log, breaking the reconstructability guarantee that every request is a pure function of the log; fresh-per-run keeps it. A kernel-style backend remains expressible behind the seam later, with its own logging story.
## Acceptance criteria
- `mode: 'native'` (and unset) is byte-for-byte today's behavior: same assemblies, same headers, same snapshots.
- Under `mode: 'code'`, the assembled tool list (and thus the logged `request/header`) is exactly `[run_code]`; under `'both'`, every native schema plus `run_code`; the `tools:sdk` section is present in both, absent under `'native'`, and its text is deterministic for a fixed tool set (byte-identical across consecutive assemblies).
- The generated `.d.ts` covers every registered tool except `run_code`, non-identifier names via quoted keys, unsupported schema constructs as `unknown`, without codegen ever throwing.
- A program calling two tools returns only its curated output; each sub-call appears as a `tool/code-dispatch` event ordered by log `seq`, flows through `tools/pre-execute`/`post-execute` (a deny reaches the program as a binding rejection), and never enters derived messages; a binding argument that does not survive JSON normalization (`BigInt`, a circular structure) rejects before dispatch — nothing executes unlogged.
- `Promise.all` over three bindings produces non-overlapping `ctx.tools.execute()` intervals (probe-tool assertion); aborting mid-program stops the worker and dispatches nothing further; a budget expiry during a slow sub-dispatch aborts that dispatch (the probe tool observes its signal fire), `run_code` returns only after the queue drains, and no `tool/code-dispatch` event lands after `run_code`'s own `tool/result` in the log.
- Worker runtime: a hot `for(;;){}` run ends at the `computeMs` busy-time budget with `error.kind: 'timeout'` — including when the program fired an un-awaited binding call first (the pending-RPC decoy); a program idling on a slow binding does not consume `computeMs` and is bounded only by `maxWallMs`; OOM under `resourceLimits` yields `kind: 'worker-exit'` with the host process healthy; `process.env` inside a program is empty; non-erasable syntax yields `kind: 'exception'` without a worker spawn; disposal awaits worker exit.
- Misconfiguration is loud before any model request: non-native mode with no `ctx.codeRuntime`, a runtime whose `language ≠ 'typescript'`, and `toolOrder` naming a non-contributed tool all reject the assembly with actionable messages.
- The demo runs against the real API via `demo:code`; the snapshot goldens replay keylessly; all repo gates pass on every PR of the stack.
## Risks
**The worker is not a hard security boundary.** Deliberate and documented (§Trust posture): posture equals the existing bash tool, containment exceeds it, gating uses the same seams. Deployments needing more need a future `isolation: 'container'` backend — tracked as the seam's designed extension, not a TODO on this design.
**`stripTypeScriptTypes` is marked experimental.** It is the same engine (amaro/swc) behind Node's own native `.ts` execution, exposed as an API across this repo's whole engines range. Mitigations: the runtime's unit suite pins the behaviors relied on (position preservation, erasable-only rejection message shape loosely), the call sits behind one private function, and `amaro`/`sucrase` are drop-in replacements if the API shifts. The erasable-only subset is a model-facing contract line, and the error path is a working feedback loop, not a dead end.
**Prompt cost of the SDK, especially under `'both'`.** The `.d.ts` can rival the native schemas it complements; `'both'` carries two representations. Prefix stability + provider caching amortize per-session cost; the mode is per-deployment; the RFC makes no unconditional-savings claim. Measured guidance (when to prefer which mode) is explicitly post-ship learning.
**Registry scope growth.** `dsh-tools` absorbs codegen, a tool, a bridge, and an event. Contained by module boundaries inside the package (`ts-types.ts`, `code-mode.ts` beside `schema.ts`/`json-schema.ts`/`presentation.ts`) and by the seam: everything substrate-shaped lives behind `ctx.codeRuntime`.
**Structured-clone limits at the binding boundary.** The seam's clone boundary admits values JSON does not (`Date`, `Map`, `BigInt`), and the session log accepts only JSON — left unhandled, a sub-call could execute and then fail at `tool/code-dispatch` append time. Closed by the bridge's JSON-normalization step (§ the dispatch bridge): what does not survive the round-trip rejects that binding call before dispatch, so every executed sub-call is loggable by construction. The seam itself keeps the wider structured-clone contract (it is about the port, and stated so a future binding producer cannot discover it in production); consumers with stricter payload needs enforce them at their own boundary, as the bridge does. Non-text sub-result content is reduced to placeholders — a known MVP limitation, recorded in the SDK instructions.
**Serialized-only sub-dispatch.** `Promise.all` gains no wall-clock parallelism yet, only fewer round-trips; models may over-expect. The instructions state it; lifting it is tied to the same concurrency-safety metadata the native parallel-dispatch TODO needs.
**Budget metering reads the event loop, not a flag.** Busy-time polling (`eventLoopUtilization()`) is coarser than an exact CPU meter — a budget expires up to one poll interval late — and its correctness claim ("a pending dispatch cannot pause it") is load-bearing against a hostile program. Both sides are unit-tested (hot loop with a pending decoy dispatch dies at `computeMs`; idle-on-slow-binding survives to `maxWallMs`), and the poll interval is an internal constant, not config — nothing a deployment could mis-tune into a bypass.

19
docs/rfc/proposed/feature/2026-06-30-pre-tool-input-rewrite.md
View File

@@ -4,7 +4,7 @@ Status: proposed
## Problem
The [interception-seams RFC](../../implemented/feature/2026-06-30-interception-seams.md) added `tools/pre-execute` returning a `PreToolDecision` (allow/deny/ask) — but deliberately NOT input rewrite (a hook changing a tool call's `arguments` before it runs). Claude Code's `PreToolUse` hook offers an `updatedInput`, so a faithful CC bridge wants the same. This RFC designs that, separately, because doing it consistently is a real problem — not a field to bolt onto the allow decision.
The [interception-seams RFC](../../implemented/feature/2026-06-30-interception-seams.md) defines `tools/pre-execute` as an allow/deny/ask gate over an execution whose identity is already protected and whose arguments are deeply frozen. Claude Code's `PreToolUse` hook also offers `updatedInput`, so a faithful bridge needs an explicit rewrite mechanism. A rewrite cannot be a mutation escape hatch on the existing execution object: it must keep the durable history, audit record, presentation, and executed value consistent.
## The problem: three readers of pre-execution arguments
@@ -14,37 +14,38 @@ In the loop, a tool call's arguments are committed to the log and read by live c
2. **`tool/call`** is the durable AUDIT record, appended before `ctx.tools.execute()`.
3. **Live presentation reads `tool/call.arguments`**: the ACP bridge remembers them and passes them to `presentResult`; `dsh-tool-bash` derives the card title, the rawInput, the cwd, and the terminal-vs-background treatment from them.
So an "input rewrite" that changes ONLY what executes would make the UI show one command while another RAN, and render result state against the wrong arguments — a real inconsistency, not a documentable gap. (The existing low-level capability to mutate `exec.arguments` in a listener has exactly this latent inconsistency; it is unadvertised precisely because of this — yet not unused: a tool-bash integration test rewrites a scripted call's arguments through it (`packages/bash/tool-bash/tests/integration.spec.ts`), so this design must either sanction that path with the consistency unit below or seal it — `readonly` arguments at the seam, with the test shim moved onto a behavior-level helper.)
An execution-only rewrite would make the UI show one command while another ran and render the result against the wrong arguments. The registry prevents that failure mode today: it structured-clones and deep-freezes `arguments`, makes the execution identity properties non-writable, and exposes no test shim or listener path that can replace them. The rewrite design must preserve that protected-identity boundary rather than weaken it.
## Proposal
A sketch, to validate against the code when built. Treat input rewrite as a consistency unit: when a `pre-execute` hook supplies `updatedInput`, the rewrite must be reflected in ALL three readers, atomically, before execution:
A rewrite is a pre-identity consistency transaction. When a hook supplies `updatedInput`, the effective value must be chosen before the registry constructs its immutable `ToolExecution`, and it must be reflected in all three readers atomically:
- The `tool/call` audit event records the REWRITTEN arguments (with the original retained in a sidecar field for the audit trail — a hook changed the call, and both the original and the effective arguments are facts worth keeping).
- The `assistant/message` in derived history must agree with what executed — options to evaluate: rewrite the assistant message's tool-call block in place (changes what the model "sees it said"), or record a separate correction the next request carries. The CC model is that the model sees the rewrite took effect.
- Presentation (`presentCall`/`presentResult`) reads the rewritten arguments, so the UI shows what actually ran.
The shape would extend `PreToolDecision` with an allow-variant `arguments` (or a dedicated `{kind:'rewrite', arguments}`), and the loop would thread the rewrite through the three readers above rather than only into `ctx.tools.execute()`.
Extending `PreToolDecision` at its current firing point is insufficient: both durable records already exist by then, and the execution identity is protected. The implementation must either move the relevant decision before the log commit or add a dedicated earlier rewrite decision over the pending model call. After the loop commits the effective arguments to history and audit, it constructs the ordinary immutable execution and runs the existing allow/deny/ask and tool pipeline unchanged.
## Alternatives considered
### Why not now
### Why not mutate the execution object?
The interception-seams RFC notes input rewrite "fought the code across two review rounds" — the signal AGENTS.md names for an over-reaching change. Shipping allow/deny/ask first keeps the seam honest (no advertised contract that silently desyncs the UI), and a CC/Codex bridge that receives an `updatedInput` logs it and surfaces a faithful-but-degraded warning (like `ask`→deny) until this lands. This RFC is the home for the consistent design; `TODO(pre-tool-input-rewrite)` in the loop's pre-execute call site anchors it.
Allowing a pre-execute listener to assign `exec.arguments` would provide only an execution rewrite, leaving model history, audit, and presentation unchanged. Keeping the identity protected makes such partial behavior unrepresentable. Until the consistency transaction exists, a CC/Codex bridge logs and warns about `updatedInput` rather than claiming it was honored; `TODO(pre-tool-input-rewrite)` at the loop dispatch site anchors the missing earlier phase.
## Acceptance criteria
- A `pre-execute` rewrite is reflected in all three readers atomically before execution: the `tool/call` audit records the rewritten arguments (the original retained in a sidecar field), derived history agrees with what executed, and presentation renders the rewritten arguments.
- The unadvertised `exec.arguments` mutation path is either sanctioned by this consistency unit or sealed (`readonly` arguments at the seam, the test shim moved onto a behavior-level helper).
- A requested rewrite is resolved before `ToolExecution` identity is created and reflected in all three readers atomically: the `tool/call` audit records the rewritten arguments (the original retained in a sidecar field), derived history agrees with what executed, and presentation renders the rewritten arguments.
- The effective `ToolExecution.arguments` remains deeply frozen and non-writable throughout pre-policy, guards, dispatch, post-policy, and final observation; no mutation shim is introduced.
- The CC/Codex bridges honor `updatedInput` instead of logging the faithful-but-degraded warning.
## Risks
- Rewriting the `assistant/message` tool-call block changes what the model "sees it said"; whether any provider rejects that on replay is the open question that must be settled empirically before the decision shape freezes.
- Until this lands, the unadvertised mutation path keeps its latent UI-desync inconsistency.
- An earlier rewrite phase changes the ordering relationship among `assistant/message`, `tool/call`, hook audit events, and execution; the design must pin that ordering without weakening turn enclosure or call/result adjacency.
## Open questions
- Does rewriting the `assistant/message` tool-call block corrupt any provider's expectation on replay, or is a separate correction safer?
- Should the original arguments be preserved on the `tool/call` event (audit) and, if so, under what field?
- Does the rewrite decision move before the log commit or become a dedicated earlier seam, and how do existing pre-tool allow/deny hooks avoid running twice?
- How does this interact with a future permission `ask` flow (a user approving a rewritten call)?

48
docs/rfc/proposed/feature/2026-07-08-interactive-side-sessions.md Normal file
View File

@@ -0,0 +1,48 @@
# RFC: Interactive side sessions and merge-back
Status: proposed
## Problem
A user deep in a live session often wants to fork the conversation — ask "why did we structure it this way", explore an alternative, get an explanation — without polluting the main context and without abandoning the surface they are in. The harness has every primitive this needs and no product face for it: [the session-store fork API](../../implemented/feature/2026-06-30-session-store-fork-api.md) produces a `Session` with no attached agent and no way for a client to reach it, and [the fork subagent](../../implemented/feature/2026-06-21-subagent-capability-seam.md) seeds a child with the parent's prefix but runs it as a model-driven task whose whole transcript collapses into one tool result — neither yields a forked conversation the USER can talk to.
The return path is missing entirely: nothing carries a conclusion from one branch back into another. Whatever the user learns in an exploration branch is copy-pasted by hand or lost, with no provenance and no replayable record.
Terminal-first competitors ship half of each: single-shot side questions with inherited context, and user-switched branch copies that force a client restart. None ship the return verb. A harness that owns its session store and context assembly can do both cheaply — and can do so without breaking the provider prefix cache, which third-party wrappers structurally cannot.
## Proposal
A **side session** is an ordinary live session forked from a source session at its last completed turn, attached to its own agent, framed as a read-only advisor, with one new verb — **merge-back** — that hands a condensed note to the parent.
- **Fork + attach composes existing primitives.** The child is created via `ctx.agents.create({ seed, meta })` with the parent's balanced completed-turn prefix (the same slice the fork subagent takes) and `parentSession`/`seedLength` lineage stamped in `meta`. No new core service and no session-store change, for the same reasons [the fork API RFC](../../implemented/feature/2026-06-30-session-store-fork-api.md) rejected a standalone fork service.
- **Advisor framing rides the log, not the system prompt.** The rules ("you are a read-only side advisor; explain, do not mutate; refuse task continuation") are `inject()`ed as a `context/message` with source `{ kind: 'plugin', plugin: 'sidechat' }` immediately after creation. The child's system prompt stays byte-identical to the parent's, so the provider's prefix cache covers the inherited history.
- **Merge-back is one condense turn plus one injection.** The child is prompted for a bounded handback note (a hard length cap), which is then `inject()`ed into the PARENT as a `context/message` with the same plugin source. The parent's next request sees it at its chronological position; replay and [reconstructability](../../implemented/architecture/2026-07-05-reconstructable-requests.md) hold by construction; no new event type enters [the session vocabulary](../../../core-data-structures/session.md).
- **The surface binding is deliberately unspecified.** How a user invokes the fork and the merge, and how a handback note is presented, are client concerns; while the harness speaks through protocols whose UI it does not control, this RFC pins only the surface-agnostic mechanics above and leaves presentation to the first surface the project owns.
Scope also excludes: rewind productization (mechanically the same fork-to-an-earlier-boundary, an entirely different product question), session tree views, a model-facing side-session tool, and `forkName`/`mergedInto` metadata.
## Prototype
A spike (branch `spike/side-sessions-b`) validates the mechanics against the live adapter: a fork that leaves the source log untouched, a child seeded with the full inherited prefix, a multi-turn advisor exchange that demonstrably uses parent history, and a merge-back the parent's next turn quotes correctly.
## Alternatives considered
- **Carry side conversations through the subagent seam.** Rejected: a subagent child is a model-driven run — the parent's model spawns it, drives it, and consumes its result as a tool result. A side session is user-driven, needs its own client-visible session and lifecycle, and must outlive any single parent turn.
- **Persona via `system-prompt/assemble` section filtering.** Rejected as the default path: any system-prompt byte change invalidates the provider prefix cache from token 0, forfeiting the cheap fork that makes side sessions attractive on long histories. The filter seam remains available for deployments that prefer hard prompt separation over cache reuse.
- **A dedicated `sidechat/*` event family for the handback.** Deferred: `context/message` with a mandatory plugin source already satisfies durability, provenance, and replay. A first-class event earns its catalog, persistence, and snapshot costs only if a UI needs to render handbacks as dedicated cards.
- **Binding the proposal to a protocol surface now.** Rejected in review: the harness currently speaks through client-owned UIs it does not control, so any presentation contract written today would be speculative. The RFC pins the surface-agnostic mechanics and defers presentation to the first surface the project owns.
- **Surface-level mirroring of the handback.** Rejected in review: a visible record emitted outside the log vanishes on replay while the model still sees it. Whatever surface eventually renders the handback must derive its presentation from the durable `context/message`, so live and replayed views come from the same event.
## Acceptance criteria
- Forking a live session yields a child agent seeded with the source's balanced completed-turn prefix, with `parentSession` and `seedLength` in its header and a system prompt byte-identical to the parent's; the source log is untouched by the fork.
- The advisor framing is exactly one plugin-sourced `context/message` at the head of the child's appended history — never a system-prompt change.
- Merge-back appends exactly one length-capped `context/message` to the parent with source `plugin: sidechat`; the parent's next request sees it, and replay reproduces it at the same position.
- Parent and child run concurrently without cross-talk between their logs or streams.
- Coverage: unit tests for the fork/attach and merge-back mechanics; surface-level snapshot coverage lands with whichever surface first binds the feature.
## Risks
- **Read-only is advisory in v1.** The rules are injected context, not enforcement; a determined prompt can still drive mutating tools. The hard gate is a `tools/pre-execute` deny via [the interception seams](../../implemented/feature/2026-06-30-interception-seams.md), and the RFC's advisor framing is written so that gate can be added without changing the mechanics.
- **A compacted source forks its compacted view.** The child inherits the summary, not the original turns; whichever surface binds the feature should disclose this once [compaction](../../implemented/feature/2026-06-18-compaction-capability-seam.md) ships in this path.
- **Handback notes spend parent tokens.** The length cap and one-note-per-merge bound the cost, but a user who merges repeatedly accumulates notes; a future consolidation pass belongs to the compaction work, not here.

79
docs/rfc/proposed/feature/2026-07-08-repeat-tool-guard.md
View File

@@ -1,79 +0,0 @@
# RFC: Repeat-tool-call guard plugin
Status: proposed
## Problem
A model stuck in a loop re-issues the same tool call with byte-identical arguments — re-running a failing grep, re-reading an unchanged file, polling a command that already gave its answer — and each round trip burns tokens, wall-clock, and (for paid APIs) money without adding information. The harness has nothing that notices: the loop has no step budget, no plugin tracks call repetition, and the model only escapes when it happens to vary its own behavior. The failure mode is real and cheap to detect — [pi-repeat-tool-guard](https://github.com/Kingwl/pi-repeat-tool-guard) ships exactly this as a pi coding-agent extension: count consecutive identical calls and, past a threshold, append a `<system-reminder>` telling the model to stop repeating itself and change course.
The harness already has every seam the pi extension uses, and better ones: [the interception-seams RFC](../../implemented/feature/2026-06-30-interception-seams.md) gives `tools/post-execute` a sanctioned way to attach model-facing context to a finished call, the loop buffers and injects that context with call/result adjacency preserved, and injected context is a logged `context/message` — so a native guard satisfies the model-visible ⟺ logged rule with no new session event. What is missing is only the plugin itself.
## Proposal
The guard is a loop-hygiene plugin, not a model-facing tool: it never appears in the tool list, never vetoes or rewrites a call, and adds exactly one behavior — it watches each agent's stream of tool calls, counts runs of consecutive calls to the same tool with identical canonicalized arguments, and at configured run lengths injects an escalating advisory reminder telling the model to stop repeating itself, re-read the last result, and either change approach or conclude. The purpose is to break unproductive loops within a few wasted steps instead of letting them run to the turn's natural end — while leaving the decision (retry differently, gather more evidence, or finish) entirely with the model, so a legitimately repeated call is delayed by nothing and blocked by nothing.
The shape: one new leaf plugin package, `@deepseek-ai/dsh-repeat-tool-guard` at `packages/guard/repeat-tool-guard/`, opening a `guard/` group for loop-hygiene plugins (single-package groups have precedent: [the todo-write RFC](../../implemented/feature/2026-06-29-todo-write-tool.md) shipped `todo/tool-todo`). The plugin registers three listeners via `ctx.effect()` and holds all state in plugin-local maps keyed by `AgentId` — the tool registry is a context-level singleton whose waterfalls interleave every agent's calls (subagents run on the same context), so per-agent keying is correctness, not polish.
- **`tools/post-execute` (waterfall)** — the one detection point. The listener receives `(exec, result)` together, so counting and reminder delivery need no cross-event pending map (the pi extension needs one only because its `tool_call`/`tool_result` hooks are separate events). It always delegates via `next()` and, when a threshold is hit, folds a reminder onto the downstream decision's `additionalContext` — the observe-and-enrich posture [the hooks bridges](../../implemented/feature/2026-06-30-hook-bridges.md) already use, honoring the waterfall contract. Counting happens here rather than in `tools/pre-execute` because post-execute also runs for denied calls (`ToolRegistry.execute` routes a deny through the same pipeline), and a model hammering a denied call is exactly the loop worth breaking.
- **`agent/prompt-submit` (waterfall)** — pure reset hook: delegate via `next()`, clear the submitting agent's chain. A user interjection changes the context; repetition across it is not a loop.
- **`agent/status` (emit)** — on `disposed`, drop the agent's state, bounding the maps over harness lifetime.
### Detection semantics
The chain key is `(tool name, canonical arguments)`; a call identical to the previous tracked call increments the agent's consecutive counter, a different tracked call resets it to 1. Canonicalization is a deep key-sort plus `JSON.stringify`: `ToolExecution.arguments` is by construction the loop's `JSON.parse` output (or the raw string fallback for malformed argument JSON, which is itself a comparable value), so the pi original's bigint/circular/`undefined` handling has no inputs here and is deliberately dropped.
Two deliberate rules, both documented in the package README because they are behavior a reader would otherwise guess at:
- **Untracked calls are transparent to the chain.** A call excluded by `include`/`exclude` neither increments nor resets the counter, so `grep X → todo_write → grep X` still counts as two consecutive `grep X` when `todo_write` is excluded. This is what makes exclusion useful — bookkeeping tools interleaved into a loop must not launder it — and it is the pi extension's (undocumented) semantics, kept on purpose and written down.
- **Calls without an agent are ignored.** A direct `ctx.tools.execute()` caller (tests, future non-loop consumers) has no model to remind and no `AgentId` to key on.
### Reminder delivery
Reminders ride `additionalContext` (source `{kind: 'plugin', plugin: 'repeat-tool-guard'}` — the label is load-bearing per `HookContext`), never a `content` replacement: the `tool/result` event stays the tool's own output for audit, and the loop already appends buffered context as `context/message`(s) after the step's results, which the session renders as the tagged synthetic-user envelope and derived history replays. Thresholds escalate: the first configured threshold gets a short "you are repeating yourself, analyze the previous result" nudge; each later threshold gets the detailed form naming the tool, the repeat count, and the canonical arguments, and stating that the calls made no progress. The pi original hardcodes the gentle text to the literal count 3; the guard keys it to `thresholds[0]`, fixing that bug in the port. When the downstream decision already carries `additionalContext` (a hook bridge on the same call), the guard folds content following the shared-merge precedent in `dsh-hook-protocol`.
### Config
```yaml
- id: repeat-tool-guard
name: '@deepseek-ai/dsh-repeat-tool-guard'
config:
thresholds: [3, 5, 8] # default; consecutive counts that trigger a reminder
include: [] # tool-name patterns to track; empty ⇒ all tools
exclude: [todo_write] # tool-name patterns transparent to the chain
```
`thresholds` is validated at load and throws on an empty list, a non-integer, a value below 2, or a duplicate — misconfiguration fails loud, replacing the pi original's silent fall-back to defaults. `include`/`exclude` entries support `*` wildcards. Patterns are predicates over whatever tools exist at call time, not references to a registry entry, so an entry matching no currently registered tool is NOT an error — unlike `toolOrder`'s referent check, `exclude: [mcp_*]` must stay valid in a deployment that loads no MCP tools.
### Testing
Coverage named at plan time, per tier: **unit** — counting/reset semantics (identical, different-tracked, untracked-transparent, prompt-submit reset, disposal cleanup, per-agent isolation), canonicalization, threshold escalation including the `thresholds[0]` gentle-text rule, config fail-loud cases, and the fold-onto-downstream-decision path, to per-file 100% like every `packages/*/*/src` file. **Snapshot** — one scripted-replay scenario where the model repeats a call to threshold and the reminder `context/message` appears in the transcript, pinning the model-visible text and its envelope (this is a transcript-surface change; the ACP snapshot suite is the tier that owns it). **e2e** — none: the plugin is provider-independent and deterministic, and forcing a live model to repeat a call three times is not a stable test; the seam contracts it relies on are already e2e-covered by their owners.
## Alternatives considered
- **Append the reminder into the tool result** (`accept` with replaced `content` — the pi extension's mechanism, which patches result content because that is the only channel its API offers) — rejected: it makes the logged `tool/result` lie about what the tool returned, and `additionalContext` exists precisely as the separate sanctioned channel for post-execute commentary, with loop-level buffering that preserves call/result adjacency.
- **Count in `tools/pre-execute` with a pending-reminder map** (the pi two-phase shape) — rejected: post-execute alone sees `(exec, result)` together and also fires for denied calls, so one listener with no cross-event state covers strictly more attempts with less machinery.
- **Escalate to `block` at the highest threshold** — rejected for the initial scope: a blocked call punishes legitimate identical repeats (polling a long-running terminal, re-checking a file the agent expects to change), and an advisory reminder keeps the model in control. Revisit with evidence; the decision shape (`PostToolDecision`) already supports it.
- **A per-deployment external hook via the CC/Codex bridges** (a `PostToolUse` script) — rejected as the answer: it works today for one deployment, but a shipped, unit-tested, `cordis.yml`-configurable plugin is the harness-native form, without per-call subprocess cost.
- **A loop-level step or repetition budget in `agent-loop`** — rejected: "plugins, not loop changes"; a hard step budget is a blunter, orthogonal control that would need its own proposal.
- **Fuzzy/near-identical detection** (normalized paths, similar-but-not-equal arguments) — rejected: exact match after canonicalization is cheap, deterministic, and explainable to the model; similarity thresholds invite false positives and need evidence before they earn complexity.
- **Placing the package in `core/`** — rejected: core is the product spine; a behavioral guard is an optional leaf plugin, and the `todo/` precedent is a small dedicated group per plugin family.
## Acceptance criteria
- `packages/guard/repeat-tool-guard/` exists, registers all listeners through `ctx.effect()`, and is loadable from a `cordis.yml` with the config above; the config catalog regenerates with its entry.
- Invalid `thresholds` (empty, non-integer, `< 2`, duplicate) throw at plugin load.
- Unit suite covers the semantics list above at per-file 100%; a snapshot scenario replays a threshold-crossing repetition and pins the reminder `context/message` in the transcript on macOS and Linux.
- The reminder is reconstructable from the session log alone (it is an ordinary `context/message` with a plugin source — no new session event).
- The package README opens with the plugin's purpose — an advisory loop-breaker that is not a model-facing tool, never blocks or rewrites a call, and only injects reminders — then documents the transparency rule, the per-agent keying, and the in-memory-only state; `doc-sync` is green.
## Risks
- **False positives on legitimately repeated calls.** Idempotent polling patterns repeat identical calls on purpose; the reminder is advisory and thresholds/`exclude` are the pressure valves, but a badly tuned deployment adds noise to the transcript. Mitigation: conservative defaults and the reminder text explicitly allowing "finish the task if enough evidence has been gathered".
- **Reminder tokens are model-visible cost.** Each trigger appends a paragraph to the next request; thresholds bound the frequency, but a pathological agent can hit 3/5/8 repeatedly across different keys.
- **State is in-memory only.** A session resumed from persistence starts with a fresh chain, so a loop spanning a resume gets its reminders later than a live one — accepted: the guard is a heuristic nudge, not a logged invariant, and persisting counter state would buy little for real complexity.
- **Multiple context producers on one call.** When a hook bridge and the guard both attach `additionalContext`, ordering follows listener registration order; the fold keeps both, but the combined envelope's readability depends on merge behavior that this RFC inherits rather than owns.
## Open questions
- Should compaction reset chains? A compacted history changes what the model sees, but the repetition risk usually survives compaction; the initial answer is no.
- Should subagents inherit the parent's thresholds via config only, or ever share chain state? Per-agent isolation is the proposed default; sharing looks like a smell until a concrete case appears.

41
docs/rfc/proposed/feature/2026-07-13-stream-workflow-progress-through-tool-calls.md Normal file
View File

@@ -0,0 +1,41 @@
# RFC: Stream workflow progress through tool calls
Status: proposed
## Problem
The workflow engine intentionally emits balanced `workflow/*` observation events for run, phase, narration, and child-agent progress, but no production consumer presents them. Editors therefore show one pending workflow tool card until the final result even while the engine already reports which phase is active, what the script logged, and which children started or settled. The [dynamic-workflows decision](../../implemented/feature/2026-07-05-dynamic-workflows.md) explicitly reserves ACP progress UI for this event stream.
Making `dsh-acp` listen to workflow events directly would invert the capability boundary: the generic UI bridge would depend on an optional workflow package and special-case one tool name. The tool pipeline already owns the routing facts a live update needs—agent and call id—but exposes only pure pending/final presenters, so a long-running tool has no provider-neutral way to report transient UI state between them.
## Proposal
Add a live progress channel to `dsh-tools`. The registry-owned `ToolExecution` gains `reportProgress(view): boolean`, where `view` is a detached provider-neutral generic progress snapshot containing an optional replacement title and UI-facing content blocks. Progress cannot change the call's args-derived card tag, kind, raw input, locations, terminal intent, or diff intent; it updates only the live title/content within the presentation chosen up front. While the execution is active, the method validates and snapshots the view, then dispatches a contained, agent-scoped `tools/progress` observation carrying the authoritative execution identity and snapshot. Once final-result processing begins it returns `false` and emits nothing, so a late asynchronous reporter cannot overwrite a terminal card. Observer exceptions are logged and cannot fail the tool.
`dsh-acp` consumes `tools/progress` generically. It resolves the execution's agent through its existing agent-to-session map and emits an in-progress `tool_call_update` for the same call id. Because reporting is available only inside the tool execution pipeline, the durable `tool/call` and its ACP `tool_call` always precede the first update; closing the reporter before `tools/result` ensures no progress update follows the completed/failed card. Progress is live UI state rather than model input or durable history: session replay continues to reconstruct the pending and final cards from `tool/call` and `tool/result` without replaying transient updates.
`dsh-tool-workflow` becomes the first producer. Each tool execution installs a compact event capture before calling `ctx.workflows.start()`, because a valid engine may emit progress synchronously inside `start()`. Until the call returns, the capture reduces observed events into candidate states keyed by `WorkflowRunInfo.id`; it then selects the returned `WorkflowRun.id`, discards other candidates, reports the accumulated snapshot, and routes later matching events directly. If `start()` throws, the capture is disposed and its candidates are dropped. This preserves engine swappability without adding observer correlation to `WorkflowStartRequest` or requiring progress to wait until `start()` returns.
The reducer consumes the existing start, phase, log, agent-start, agent-end, and end events, reporting a replacement snapshot with the current phase, latest log line, active child labels, and completed/failed/cancelled counts. It does not accumulate a narration transcript; settled children leave the active set and become counters. `workflow/end`, tool settlement, or plugin disposal removes the reducer entry and event capture. The six workflow events, their metadata, paired child lifecycle, run handle, cancellation channels, and observer containment remain unchanged; third-party observers can continue consuming them directly.
Update the tool execution/presentation docs, generated event and API catalogs, workflow package docs, and the workflow data-structure catalog. ACP integration coverage must exercise the real workflow tool and worker seam with a scripted model boundary; the primary ACP snapshot suite adds one workflow-progress scenario because this changes the editor-facing transcript.
## Alternatives considered
**Delete the workflow observation surface.** Rejected in [the collapse-workflow simplification](../../rejected/simplification/2026-07-12-collapse-workflow-to-foreground-core.md): the events and their balanced lifecycle are intentional, and the missing piece is a consumer.
**Teach ACP about workflows directly.** This could map `WorkflowRunInfo` to a session and card, but it would make the generic bridge depend on an optional capability and bypass the rule that tools own presentation intent. A tool-progress channel solves the same routing problem for every long-running tool.
**Persist every progress update as a session event.** That would make live narration replayable, but it would permanently enlarge logs with state whose authoritative durable outcome is already the tool call/result pair. If resumable workflow progress becomes a product requirement, it needs a workflow-journaling design rather than UI snapshots disguised as durable facts.
## Acceptance criteria
- `ToolExecution.reportProgress()` is registry-owned, agent-scoped, snapshotting, observer-contained, and returns `false` without dispatch after terminal processing starts.
- ACP routes progress to the correct call in the correct live session; concurrent workflows in different sessions cannot cross-talk, and no `tool_call_update` appears before its `tool_call` or after its terminal update.
- Workflow progress shows the current phase, latest log line, active children, and outcome counts while preserving all existing `workflow/*` events and run semantics; a seam test engine that emits start, phase, log, child, and end events synchronously inside `start()` loses none of that reducer state.
- Cancellation, worker death, tool failure, session close, and plugin disposal release reducer state; replay emits only the durable pending/final card pair.
- Unit, workflow integration, ACP integration, snapshot, typecheck, coverage, doc-sync, module-graph, build, and hygiene gates pass.
## Risks
This adds a public live-progress method and event to the tool seam, so implementations must keep the active/terminal boundary exact and detach snapshots before observers see them. The pre-start capture can briefly observe unrelated workflow runs, so it holds only compact candidate state keyed by run id and drops every non-matching candidate as soon as `start()` returns. A workflow can emit many progress changes; the bounded reducer avoids transcript growth but still sends one UI update per meaningful event after correlation. If measured clients need coalescing, it must be a defaulted validated bridge configuration rather than a hardcoded throttle. Transient progress intentionally disappears on replay, so the final tool result remains the only durable workflow card content.

54
docs/rfc/proposed/simplification/2026-06-20-unify-agent-and-session-id.md
View File

@@ -4,57 +4,35 @@ Status: proposed
## Problem
The agent factory carries TWO ids for what is, in every live consumer, one thing:
The agent factory carries two ids for each live agent/session pair: `agentId`, the `AgentRegistry` routing handle, and `sessionId`, the event-sourced/persisted-log identity. `CreateAgentOptions` takes both; `ResumeAgentOptions` takes `agentId` plus `resumeSessionId`; in-process subagents mint two independent UUIDs despite recording lineage separately.
- `agentId` — the `AgentRegistry` handle (the actor identity; the registry rejects a duplicate).
- `sessionId` — the event-sourced session / persisted-log identity (`session.header.id`).
ACP already uses the same value for both identities. Where they diverge, stdio keeps `labelBySession` solely to recover an agent label from session events, and hooks expose both values for authors to reconcile. No production path reattaches one live agent object to several sessions or drives one session through several agent ids.
`CreateAgentOptions` takes both separately; `ResumeAgentOptions` takes an `agentId` plus a `resumeSessionId`. They diverge in exactly three places:
The [agent-scope runtime](../../implemented/architecture/2026-07-12-agent-scope-runtime-design.md) has no reservation side tables: create and resume use one `AgentCreationTransaction`, and agent/session entries use the same final-entry collision rule. Separate ids therefore do not duplicate asynchronous liveness, rollback, or quiescence machinery. Identity unification is only an API and representation simplification: it deletes one caller-supplied id, one UUID per in-process child, and the remaining translation paths without changing the transaction lifecycle.
- **Config-driven create** (`AgentLoop.create`): a stable `agentId` (e.g. `"echo"`) with a fresh per-run `sessionId` (`${id}-session-<uuid>`).
- **Resume**: a caller-supplied `agentId` (e.g. `"main"`) on a persisted `resumeSessionId`.
- **In-process subagent children**: the backend mints the child's `agentId` and `sessionId` as two independent UUIDs (`packages/subagent/subagent-inprocess/src/index.ts`) that nothing distinguishes — `parentSession` records lineage independently.
Where a live consumer looks an agent up, no lookup needs an id translation: the ACP bridge — the primary production path — already unifies the two (`agentId === sessionId === <uuid>`; both factory call sites brand `AgentId(sessionId)` directly, and its reverse lookup keys on the `Agent` object itself), and the CC hooks bridge resolves subagent children directly by the `agentId` its lifecycle event carries. The one production population whose two ids actually DIVERGE is the in-process subagent children — the same cosmetic separation as the config path, and the same one-field simplification under unification. One consumer already pays the two-id tax: ui-stdio keeps a `labelBySession` map (seeded from the registry, maintained by `agent/created`/`agent/disposed` listeners) solely to translate `session.header.id` back to an agent id for its turn labels — machinery that deletes outright when the ids unify. And the CC hooks bridge stamps `session_id: agent.session.header.id` into every hook payload, so under unification a subagent hook's `session_id` and `agent_id` become the same string — one less identity for a hook author to reconcile.
The separation is **latent generality no consumer exercises**: nothing reads a *stable* `agentId` back across runs (each process starts fresh, and persistence keys off the session id, never the agent id). The config path's "stable agentId, fresh sessionId" buys nothing concrete — it is cosmetic. And the `agentId !== sessionId` case is precisely what opens the bash owner-token alias hole: the bash completion-notice routes by `session.header.id`, but the registry enforces uniqueness only on `agentId`, so a programmatic caller registering two agents with different agent ids but the SAME session id can mis-route a notice (see [agent lifecycle and ownership seams](../../implemented/architecture/2026-06-18-agent-lifecycle-and-ownership-seams.md) § Seam precondition). The current code documents this as a precondition rather than guaranteeing it.
Session itself repeats the same fact as `Session.id` and `Session.header.id`. Construction rejects a header whose id differs, so the aliases are constrained equal; the durable boundary must nevertheless validate the duplicate, and production consumers choose between its two homes.
## Proposal
Make an agent BE its session: one id. An agent's registry handle IS its `session.header.id`.
Make an agent's registry id equal its session id. `CreateAgentOptions` accepts one id used for both final registry entries; resume registers the agent under the resumed session id; subagent creation mints one combined id; Session keeps one identity home by deriving `id` from `header.id` or removing the alias. Keep the existing creation transaction, final-entry collision checks, and exact-entry detach semantics; remove only maps and fields whose sole job is translating between the ids.
- `CreateAgentOptions` drops the separate `sessionId` — the single `id` is both the registry handle and the live/persisted session id. (ACP already passes the same UUID for both, so its call site simplifies to one field.)
- `ResumeAgentOptions` drops the separate `agentId` — resuming `sessionId` X registers the agent under id X. (ACP already does this.)
- The config path (`AgentLoop.create`) uses its configured `id` directly as the session id, applying whatever resume-or-create policy it adopts (today it appends a per-run uuid to avoid colliding with an on-disk log; that policy moves onto the single id, e.g. the config id IS the session and a durable backend resumes it — to be settled in the implementing PR).
- The registry's existing unique-`agentId` check becomes, by construction, a unique-session-id guarantee — the bash alias hole is closed with NO new defensive invariant: two agents cannot share a session id because the session id is the agent id.
The config-driven path must first settle its currently hidden resume-or-create policy. Today it uses a stable agent label and fresh UUID-suffixed session id to avoid colliding with a durable log on the next run. Under unification it must deliberately resume the fixed id, mint a fresh combined id, or expose an explicit policy; implementation must not pick silently.
`agent/created` and `agent/disposed` remain outside this proposal. They are paired publication lifecycle events, not identity aliases; any later consumer-free removal needs its own proposal after a fresh search.
## Alternatives considered
### Why not just enforce session-id uniqueness in `AgentRegistry.register()`?
That was the review's first suggestion. It would couple the generic registry to a session-uniqueness assumption (the registry tracks *agents*, not sessions) and entrench the very separation this RFC removes. Unifying the ids closes the hole more cleanly — there is nothing left to enforce.
**Keep separate routing and log identities.** The config-driven loop uses a stable configured agent id with a fresh UUID session on each fresh process start. That is a real use of the distinction: a stable routing/display label plus a new durable conversation. Unification can proceed only after choosing whether this path resumes a fixed identity, mints a combined per-run identity, or exposes the policy explicitly. If the stable label is a required product contract, reject this proposal rather than hiding it in another map.
## Acceptance criteria
- `ctx.agents.create`/`resume` take a single id; the ACP bridge passes one id.
- The config-driven agent path has a deliberate, documented session-id policy (no silent per-run id divergence that no consumer reads).
- The bash owner-token alias hole is gone by construction (no two live agents can share a session id).
- All existing behavior the tests pin (ACP create/resume/load, config startup, durability) still holds — or the tests change WITH the behavior where the divergence was an artifact (per AGENTS.md "tests document behavior, not golden truth").
- Agent create/resume and subagent creation carry one identity; `Session` stores it in one place.
- The existing creation transaction keeps final-entry collision, exact-entry detach, rollback, and quiescence guarantees without adding identity-specific lifecycle state.
- ACP, stdio, hooks, bash ownership, persistence, and lineage need no agent/session id translation.
- The config-driven resume-or-create policy is explicit and covered across a durable restart.
- `agent/created`/`agent/disposed` are removed only if a post-change production search finds no listener; otherwise they and their publication semantics stay.
- Typecheck, coverage, snapshots, doc-sync, module-graph verification, build, and hygiene pass.
## Risks
This touches public factory interfaces (`CreateAgentOptions`, `ResumeAgentOptions`, `AgentFactory`) and the config-agent id scheme, so it is a deliberate cross-package change, not a local patch — it ships as its own PR (converged with Codex); the bash owner-token precondition it closes is documented in [agent lifecycle and ownership seams](../../implemented/architecture/2026-06-18-agent-lifecycle-and-ownership-seams.md).
The genuine risks of collapsing the two ids into one (the case AGAINST this proposal — to be weighed honestly before implementing):
- **It forecloses a one-agent-resumes-many-sessions / one-session-driven-by-many-agents future.** Today the separate ids leave room for an agent (a stable actor) to detach from one session and attach to another, or for a handoff where a new agent process adopts an existing session under a new actor handle. Unifying makes "agent" and "session" the same lifetime, so any such future needs a NEW seam (e.g. an explicit `actorId` distinct from the session) — re-introducing the very separation we removed. We judge this generality currently unused, but it is a door this change closes.
- **Subagents / fork / spawn may WANT a stable actor id across forked sessions.** The [subagent seam](../../implemented/feature/2026-06-21-subagent-capability-seam.md) runs a child agent seeded from a parent's event log (fork). If a future design wants "the same agent identity across a fork" (parent and child share an actor but have distinct session logs), a unified id blocks it. The implementing PR must check the intended fork/spawn model BEFORE unifying, or accept that fork always mints a fresh combined id. (As shipped, each subagent child mints its own distinct agent id — `parentSession` records lineage — so the seam does not currently rely on a shared actor id, but unifying would foreclose adding one.)
- **The config-driven resume-or-create policy becomes load-bearing, not cosmetic.** Today the per-run-uuid session id quietly sidesteps the "a fixed id collides with its own on-disk log on the second run" problem. Once the id is unified and stable, a config agent restarting MUST decide resume-vs-fresh deliberately — there is no longer a throwaway session id to hide behind. Getting this wrong reintroduces the create-collision the uuid was avoiding (a durable backend refuses to re-create an id whose log exists). This is the one real design decision the implementing PR owns, and it is easy to get subtly wrong.
- **Persisted/on-disk identity becomes the agent identity.** Unifying means the registry handle is now a persisted, externally-meaningful string (a session id a client chose), not an internal label. A caller that previously used a short human label (`"main"`) as the agent id now must use the session id. This is fine for ACP (already a UUID) but is a semantic narrowing for any programmatic embedder that relied on naming its agents independently of session storage.
- **Migration churn touches every create/resume call site and its tests.** `CreateAgentOptions`/`ResumeAgentOptions` shape changes ripple to ACP, the config path, the agent-loop factory, and ~dozens of test fixtures that currently pass distinct `agentId`/`sessionId` (some deliberately distinct to exercise the divergence — those tests change WITH the behavior, per AGENTS.md "tests document behavior, not golden truth"). The risk is mechanical but broad; a missed call site is a type error, but a missed *test* could silently lose coverage of a path.
The one real design question the implementing PR must settle first is the config-driven resume-or-create policy once the id is unified (today's per-run-uuid behavior is a demo simplification already flagged `TODO(demo)`). If, on closer look, the fork/spawn or multi-session-actor futures turn out to be wanted, this RFC should be REJECTED in favor of the lighter "enforce session-id uniqueness in the registry" guard — the alias hole is not reachable via ACP, so keeping the ids separate and merely documenting (or mechanically enforcing) the precondition remains a valid alternative.
This forecloses latent multi-session-actor and session-handoff designs, makes persisted client-chosen session identity the registry identity, and touches every factory fixture. The config restart decision is blocking, not mechanical. If separate routing identity is a real requirement, reject this RFC and retain the current caller-supplied pair plus final-entry arbitration.

54
docs/rfc/proposed/simplification/2026-07-04-prune-dead-core-spine-surface.md
View File

@@ -1,32 +1,60 @@
# RFC: Prune dead core-spine surface — `SurfaceManager.invalidate()`, the loop-internal exports, `ToolExecutionResult.callId`
# RFC: Prune dead public and result surface
Status: proposed
## Problem
Three pieces of public spine surface share one defect class: their only possible role is to be ignored, or their trigger is unreachable.
Several package-root exports, result fields, and convenience methods have no production consumer. They survive because tests import internals through public entry points or because a type anticipated a caller that never arrived. Each item is small in isolation, but together they enlarge the SDK contract, generated catalogs, documentation, and regression matrix without enabling a shipped path.
1. **`SurfaceManager.invalidate()`** (`packages/core/session/src/surface.ts`). Its documented trigger — "the log has been replaced wholesale (e.g. after Session seed)" — is structurally unreachable: seeding happens inside the `Session` constructor, `_surface` is created lazily on first access, and the log reference is never reassigned afterward, so no constructed `SurfaceManager` ever observes a wholesale replacement. Sole caller: its own unit test. A rollback primitive protecting a scenario the implementation cannot produce.
2. **The `runLoop`, `Inbox`, and `InboxMessage` exports** (`packages/core/agent-loop/src/index.ts`). `runLoop` has no importer outside the package — the only callers are the package's own internals (the agent constructs its loop with it), so the public re-export has zero consumers; `Inbox`/`InboxMessage` likewise reach outside code only through the package's own inbox spec (switchable to the source module). The exports contradict the package's own docs — the inbox module doc says the public surface is `Agent.send()`/`Agent.steer()` — and the [architecture dependency rule](../../../architecture.md): nothing programs against `dsh-agent-loop`; a replacement loop is a different bundle built on `dsh-agent`, not a consumer of this package's internals. `ReactLoopAgent` stays exported (cross-package tests construct it by package name).
3. **`ToolExecutionResult.callId`** (`packages/core/tools/src/index.ts`; the *input* `ToolExecution.callId` stays). Zero readers — and no listener can even construct a result: `tools/pre-execute`/`tools/post-execute` listeners return Decisions, the registry builds every result itself and always sets `callId` to the input `exec.callId`, and the post-execute dispatch snapshots the outcome before the waterfall precisely so a listener mutating the shared result reference cannot corrupt the id. The loop independently ignores `result.callId` in favor of its own `call.id`, and two regression tests exist solely to prove the field cannot matter (the loop's ignores-result-callId test and the registry's mutation guard). A field that is by construction a copy of its input, defended by snapshot machinery, and pinned by tests proving it is ignored is pure liability surface; the ACP bridge correlates via the session event's `data.callId`, never via the execution result.
The production corpus is `packages/*/*/src`, example sources/config, and runtime scripts. Tests, package READMEs, and RFC prose are evidence of publication but not fixed callers. `cordis_inspect` makes `packages/cordis/tool-cordis/src/api-catalog.ts` model-visible, and `cordis_mount` can invoke injected services through guarded real-service proxies, so catalogued service methods and returned shapes are a genuine dynamic product surface. The table therefore distinguishes absence of a fixed repository caller from unreachability: rows touching catalogued vocabulary intentionally contract what model-written mounts can discover and call, while package-root implementation helpers are not reached through that service façade. Exact-symbol searches produce the following inventory:
| Surface | Production evidence | Simplification |
| --- | --- | --- |
| `SurfaceManager.invalidate()` | Only its unit test calls it; seeding completes before the lazily-created manager exists and the session never replaces its log reference. | Delete it and its impossible wholesale-replacement contract. |
| `ToolExecutionResult.callId` | Every hook already receives the immutable `ToolExecution`; the loop and ACP correlate through the call/session event. No consumer reads the duplicate result field. | Remove the field, copy/mismatch guards, and tests that prove the duplicate cannot disagree. |
| `ReactLoopAgent` root export | Outside-package named imports are tests; production programs against `Agent` and creates/resumes through `ctx.agents`. | Return/interface-type `Agent` and make the concrete loop class package-internal; keep the deliberate synchronous config-only `AgentLoop.create()` path. |
| `workflow-workerthread` protocol/runtime/session re-exports and named `WorkerWorkflowEngine` | Every package-name consumer uses the default engine; the workflow RFC already defines the worker wire protocol as private. | Keep the default plugin class/config contract; drop the duplicate named class export and keep protocol modules source-private. |
| `code-runtime-worker` protocol/bootstrap re-exports | Outside-package production/e2e consumers use `WorkerCodeRuntime` and config, not `BootstrapPort`, `PatchableStream`, or worker message/boot types. | Keep the runtime class/config contract and make its wire/bootstrap vocabulary source-private. |
| ACP translation/presenter root exports | `agentOptions`, `streamSessionEventUpdate`, `todosToPlan`, `ToolPresenter`, `nullToolPresenter`, and `TerminalRendering` have only same-file or ACP-test consumers; the sole outside-package production consumer mounts the plugin namespace. | Keep `name`, `inject`, `Config`, `AcpConfig`, and `apply`; make translation/presentation helpers source-private and test them in-package. |
| `providerWording` and `completedTurnPrefix` root exports | Each has one same-package production caller; only the balanced-prefix helper has a same-package white-box test. | Make them source-private and test provider behavior. |
| `depthOf`, `SubagentDepthError`, `SENSITIVE_ENV_PATTERN`, `waitForExit`, and `exitsWithin` root exports | Production subagent backends consume the in-process runner and subprocess construction/disposal helpers, not these enforcement/test internals. | Keep depth/environment/exit behavior but make the helpers and error/regex source-private; test through spawn and disposal. |
| `PersistenceCoordinator.inits`, backend `inits` accessors, `seedCoversPrefix`, and `assertSerializable` | The accessors exist for white-box tests; `seedCoversPrefix` has no outside production importer; `assertSerializable` has no production caller and duplicates the coordinator append boundary's lossless snapshot. | Observe initialization through `session/flush`, make `seedCoversPrefix` source-private, and delete `assertSerializable`. Keep both backends, `SessionHeader`, and SQLite's version contract. |
| `LlmError.status` and replay status | Adapters/replay populate it, but production branches on stable error code/message and never reads raw status. | Remove the unread field and replay plumbing while preserving error classification. |
| `BlockAssembler.push()` return value | Both production callers ignore the returned completed block. | Return `void`; keep the deliberately public `blocks()`/`message()` contract. |
| `compactRegion`'s separate `session` argument | The fixed caller passes the same object already present as `agent.session`; the model-visible mount API can also call the method, but accepting two identities permits a mounted plugin to provide an incoherent pair. | Keep the manual-region seam while deliberately narrowing it to `agent.session` as the one source of truth. |
| `CompactionResult.startSeq`, `summarySeq`, `endSeq`, and `summary` | The production consumer reads only shadowed range/seq/token accounting; the durable log owns summary and event identity. | Remove the four result echoes while keeping both shared transcript renderers. |
| `BasicCompactService` estimation/summarization visibility | No outside production caller invokes the five methods; the implemented RFC names only `estimateContentTokens()` and `summarize()` as subclass hooks. | Make those two `protected` and the three orchestration-only estimators private. |
| `CodeLogEntry.source`/`level` and `RunCodeMeta.dispatches` | Every production consumer maps logs to text; no presenter/model path reads the other fields or the persisted dispatch count. | Make code-runtime logs strings (or text-only entries) and remove result-meta dispatch plumbing; keep the local counter that mints deterministic dispatch ids. |
| `ToolNotFoundError.toolName`, `SystemPrompt.config`, and `BashTask.command` | Each stored public value has no production reader. | Drop the unread field while retaining error messages, resolved configuration behavior, and task lifecycle. |
| Backend package-root implementation helpers | The exact inventory below is called only through relative same-package imports. Production namespace imports mount the retained plugin contract without reading these properties; named root consumers are tests. | Retain each adapter/provider/service and its config/error contract; stop exporting the listed helper functions/constants at package roots. |
| Consumer package-root implementation helpers | The exact inventory below has only same-package production callers. Production namespace imports mount plugin contracts without reading helper properties; named root consumers are tests. | Retain plugin contracts and stable error codes; move tests to package-local modules or public behavior and stop exporting the listed helpers at package roots. |
### Grouped helper-export inventory
- `dsh-llm-deepseek`: `httpErrorCode`, `serializeMessages`, `serializeRequest`, `DONE`, `parseSse`, `mapFinishReason`, `mapUsage`, and `translate`; `dsh-llm-pi-ai`: `buildModel`, `mapStopReason`, `mapUsage`, `toPiContext`, and `toStreamChunks`.
- `dsh-bash-local`: `DEFAULT_GRACE_MS`, `ENV_OVERRIDES`, `killGroup`, `OutputCollector`, and `runBash`; `dsh-bash-sandbox`: `shellQuote`, `classifyDenial`, and `classifyRunnerFailure`; `dsh-sandbox-local`: `bwrapProfileArgs`, `landlockProfileArgs`, and `seatbeltProfileArgs`. The public mutable test-injection fields and their types are outside this proposal.
- `dsh-fs-local`: `applyLiteralEdit`, `listDirectory`, `probe`, `readForEdit`, `readTextForDiff`, `readWholeText`, `resolveLocalTarget`, `restoreLineEndings`, `streamWholeText`, and `writeFileAtomic`.
- `dsh-web-fetch-local`: `classifyContentType`, `decoderForCharset`, `isSameOrigin`, `parseCharset`, and `validateFetchUrl`; `dsh-web-search-exa`: `mapExaResponse` and `mapExaResult`; `dsh-web-search-deepseek`: `citationSnippets` and `mapAnthropicResponse`; `dsh-web-search-perplexity`: `mapPerplexityResponse` and `mapPerplexityResult`.
- `dsh-tool-fs`: `READ_LIMIT`, `STREAM_MIN_SIZE`, `READ_MAX_BYTES`, `READ_MAX_LINE_LENGTH`, `DIFF_CONTEXT`, `applyReadTool`, `parseReadArgs`, `applyWriteTool`, `formatWriteOutput`, `parseWriteArgs`, `applyEditTool`, `formatEditOutput`, `parseEditArgs`, `buildWindow`, `formatReadOutput`, `computeHunkDiffs`, and `diffsFromMeta`.
- `dsh-tool-web`: `WEB_SEARCH_MAX_RESULTS`, `applyWebSearchTool`, `formatSearchOutput`, `parseSearchArgs`, `presentSearchCall`, `applyWebFetchTool`, `formatFetchOutput`, `parseFetchArgs`, `presentFetchCall`, `renderBody`, and `htmlToMarkdown`; `dsh-timeout-policy`: `toolTimeoutResult`; `dsh-compact-basic`: `resolveConfig`; `dsh-tool-bash`: `renderResult`.
## Proposal
Delete the method and its test; delete the three export lines and their `packages/core/agent-loop/README.md` rows, pointing the inbox spec at the source module; drop the result field from the type, the registry's construction sites (the deny result, the dispatch result, `toolErrorResult`, and the post-execute snapshot's `callId` leg), the loop's ignore-comment, the proves-ignored regression test, and the mutation guard's `callId` assertions — the hazard they all pin disappears with the field, while the result's `additionalContext` ferry (a consumed post-execute channel) stays untouched. Update the `ToolExecutionResult` paste in [tools.md](../../../core-data-structures/tools.md) (and its `scripts/type-equiv.manifest.json` row) and the result-shape row in `packages/core/tools/README.md`; for the `invalidate()` removal, amend the [session-surface RFC](../../implemented/architecture/2026-06-18-session-surface.md)'s full-rebuild-after-wholesale-replacement sentence per [implemented/AGENTS.md](../../implemented/AGENTS.md).
Sequencing: the in-flight surface-cache work (tool-pairing balance caching) neither uses nor touches `invalidate`, so that removal lands after or alongside it mechanically. The execute pipeline is `tools/pre-execute` → dispatch → `tools/post-execute`, and post-execute listeners receive the execution object alongside the result — nothing needs the result's own id.
Remove or demote every row as one bounded coordinated public-surface cleanup. Update package READMEs, JSDoc, generated API/event catalogs, type-equivalence records, exports maps where needed, and tests so they exercise the owning public seam instead of preserving test-only entry points. Do not collapse any capability seam, LLM adapter, persistence backend, or lifecycle quiescence contract.
## Alternatives considered
### Why not keep them?
**Keep test conveniences and self-contained results public.** Public helpers can make white-box tests convenient, self-contained result fields can look ergonomic, and future embedders might want the concrete loop or enumeration methods. Those benefits are hypothetical; today they make every implementation and document explain states that no shipped caller can observe. A real consumer can introduce the smallest contract it needs, with its ownership and failure semantics known.
A future consumer that swaps a session's log in place would want a reset primitive — it re-adds `invalidate` with itself. A replacement-loop author might want to reuse the inbox or the driver — the architecture already answers that a replacement loop is a different bundle. An isolated result-logging listener might want self-contained correlation on the result — the execution object is in scope at every listener, and a field that exists only to be ignored is worse than absent: it invites exactly the orphaned-pairing bug the loop comment warns about.
**Keep every catalogued member for model-written mounts.** The self-referential toolset is a real generic consumer route, not generated-doc noise. Its value comes from an accurate, composable service surface, however, not from preserving duplicate fields or incoherent argument pairs indefinitely; each catalogued contraction above removes a fact available elsewhere on the same execution, agent, or result and updates the API reference in the same change.
## Acceptance criteria
- `invalidate()` and the result `callId` appear only in this RFC; `runLoop`/`Inbox`/`InboxMessage` remain package-internal only — no re-export from the package index and no outside-package importer; the agent-loop README lists only the consumed public surface; the inbox spec imports the source module.
- The pre-/post-execute pipeline contract tests pass with the shrunk result type; the mutation-guard and proves-ignored tests shed their `callId` legs with the hazard they pin.
- Exact-symbol searches show no removed surface outside this RFC and any implemented-RFC amendments.
- Every surface listed in this RFC is absent or demoted as specified; deliberately retained extension/test contracts outside the inventory are unchanged.
- Tool execution, compaction, both LLM adapters, both persistence backends, workflow isolation, and agent creation/resume retain their shipped behavior.
- Typecheck, coverage, snapshots, doc-sync, module-graph verification, build, and hygiene pass.
## Risks
All three are compile-visible removals with no runtime behavior change on any shipped path.
Most removals are compile-visible but runtime-neutral. The compaction argument cleanup deliberately forbids a session/context mismatch while retaining the manual-region seam. External pre-release embedders and existing model-written mounts may import fewer helpers, pass fewer arguments, or receive narrower result shapes; this is an intentional product-surface contraction, not merely generated-catalog cleanup. The repository is unreleased, so carrying unsupported surface is the larger foundation cost.

32
docs/rfc/proposed/simplification/2026-07-12-drop-unconsumed-skill-provider-events.md Normal file
View File

@@ -0,0 +1,32 @@
# RFC: Drop unconsumed skill provider events
Status: proposed
## Problem
Two skill-registry notifications are produced but have no production listener. The generated producer/consumer matrix and exact event-name searches find only declarations, emit sites, tests, generated catalogs, and prose for `skill/provider-added` and `skill/provider-removed`.
Skill discovery reads the current provider map on demand, provider registration synchronously clears completed catalogs, and the post-await revision check prevents stale discovery from entering the cache. No sibling plugin waits for a skill provider through these events, unlike the live `subagent/provider-added` consumer that tolerates concurrent sibling loading.
`tools/change` and `system-prompt/change` are explicitly outside this proposal. Existing simplification decisions retain them as intentional observation points for live tool and prompt UIs, and self-referential mounted plugins already use `tools/change`. This proposal also leaves `subagent/provider-added`/`removed` unchanged because `tool-subagent` has a production lifecycle consumer.
## Proposal
Delete the two skill-provider declarations and every emit path, rollback-order branch, test, and generated catalog/matrix row that exists only for them. Remove the corresponding skill-registry README/JSDoc contract. Where tests used an event to observe cleanup, assert provider lookup or collected output instead.
Amend the skill-system RFC and package documentation so provider registration is described as direct effect-owned state with cache invalidation, not as a lifecycle notification contract.
## Alternatives considered
**Keep skill-provider notifications for future plugins.** A third-party plugin could observe provider availability, but direct provider registration and on-demand lookup are the extension contract; no current consumer needs a push signal. If a future sibling-load race appears, it can introduce a notification with the identity and readiness semantics that consumer requires, as the subagent registry did.
## Acceptance criteria
- The generated event matrix contains no row for `skill/provider-added` or `skill/provider-removed`.
- Skill discovery, direct runtime registration, provider effect rollback/disposal, cache invalidation, and registry lookup cleanup behave unchanged; listener-triggered rollback disappears with the events.
- `tools/change`, `system-prompt/change`, and the real subagent provider lifecycle consumer remain documented and covered.
- Typecheck, coverage, snapshots, doc-sync, module-graph verification, build, and hygiene pass.
## Risks
This removes pre-release skill-provider observation points while retaining both ways third-party plugins contribute skills: direct runtime registration and provider registration. A future consumer that needs live provider availability must add a purpose-built notification rather than relying on these generic events.

30
docs/rfc/proposed/simplification/2026-07-12-prune-unused-web-seam-fields.md Normal file
View File

@@ -0,0 +1,30 @@
# RFC: Prune unused web seam fields
Status: proposed
## Problem
The web capability carries request/result/status values that every shipped implementation populates but no production consumer reads. `WebSearchResult.providerId` and `query` and `WebFetchResult.providerId` are result echoes; `tool-web` formats only content/sources/truncation or final URL/status/body/truncation, and no other runtime reads them. Search providers return `WebProviderStatus.reason`, but resolution checks only `available` and intentionally emits a generic unavailable diagnostic.
`WebFetchRequest.timeoutMs` is likewise never set by a production caller. `tool-web` supplies only the URL, uses the tool definition's timeout plus `exec.signal` for the caller deadline, and relies on the local provider's configured default as a backstop. The unused per-request override forces `web-fetch-local` to expose `maxTimeoutMs`, clamp two timeout sources, and document/test precedence no product path can select. `WebExecContext` is another one-field wrapper: every caller allocates `{ signal }` and every provider immediately unwraps `exec?.signal`; no second execution-control field exists.
## Proposal
Remove the search/fetch `providerId` result echoes and search `query` echo; callers already own the request and provider selection. Shrink provider status to availability alone, preferably a boolean-returning method if that produces the clearest seam. Remove per-request fetch timeout, `maxTimeoutMs`, and their clamp/validation branches while retaining the provider's configurable default timeout and tool-level deadline. Replace `WebExecContext` with a direct optional `AbortSignal` parameter.
Update all web implementations, the model-facing tool, package READMEs/JSDoc, type-equivalence records, and tests. Keep the interface/implementation/consumer package split, provider selection, source citations, final-URL/status data, truncation reporting, and all safety limits.
## Alternatives considered
**Keep self-describing results, per-request deadlines, and an extensible execution-context object.** Result echoes can help generic telemetry, a request timeout can help trusted programmatic callers, and the wrapper leaves room for future controls. No such consumer/second field exists; carrying duplicate identity, a second deadline policy, and wrap/unwrap plumbing through every provider makes the current contract harder to implement and explain. If telemetry or per-call budget control arrives, it should define which deadline wins, where provider identity is observed, and whether multiple controls justify a context object.
## Acceptance criteria
- Every retained web request/result/status field has a production reader or is required to execute the provider request.
- Tool-visible search/fetch output, provider fallback, abort behavior, configured timeout backstop, truncation, and citations remain covered.
- No `maxTimeoutMs`, request-timeout precedence branch, or one-field execution-context wrapper remains.
- Typecheck, coverage, snapshots, doc-sync, module-graph verification, build, and hygiene pass.
## Risks
Pre-release programmatic callers lose result provenance echoes and per-request fetch deadlines. The provider still has a deployment-configurable timeout and respects cancellation, so the simplification removes configurability rather than a safety bound.

36
docs/rfc/proposed/simplification/2026-07-12-simplify-session-log-representation.md Normal file
View File

@@ -0,0 +1,36 @@
# RFC: Simplify session-log representation
Status: proposed
## Problem
The session log maintains two representations that cost more machinery than their consumers require: a pseudo-linked surface and custom request-header deltas.
`SurfaceManager` stores the same order in an array, a seq map, and mutable `prev`/`next` links. Production never reads `prev`; compact's sole `next` read is the successor of an array position. Replacement already uses `indexOf`, so the links do not make its dominant operation constant-time. A seq array with linear replacement lookup has the same asymptotic replacement cost and one representation to validate.
The request-header subsystem implements a custom system/tool delta codec and transmission-decision layer even though its contract says deltas are an encoding optimization, not a reconstructability requirement. Retaining the initial/resume full snapshot at each loop-instance boundary, then writing a canonical full `request/header` whenever that instance's assembled header changes, preserves replay while deleting `SystemDelta`, `ToolsDelta`, round-trip fallback, and the durable `request/header-delta` variant. Codec-only vocabulary disappears with the codec, not because its individual arms were invalid.
This proposal deliberately retains append and replacement `sourceEventSeqs`, crash-repair provenance, and all `SessionStartSource` variants: implemented RFCs give those fields an audit/interception role that zero current readers does not overturn.
## Proposal
Make `SurfaceManager.nodes` a `readonly number[]` of event sequences and remove the public `SurfaceNode` shape. Keep the internal replace-generation signal; update tool-pairing balance and compaction callers to use array values/indices for predecessor, successor, and replacement ranges, removing node links and the seq-to-node map. Replace post-anchor header deltas with canonical full changed-header snapshots and remove the delta codec/event/tests; initial and resume anchors remain full snapshots even when the folded header is unchanged.
Amend the session-surface and reconstructable-request RFCs where they describe the removed encoding. Update event types/invariants, request logging/replay, persistence fixtures, generated catalogs, package docs, and snapshots. Replace the codec-only `fallback` reason with an explicit `change` reason for post-anchor full snapshots, distinguishing them from the retained `initial` and `resume` anchors.
`SESSION_FORMAT_VERSION` is deliberately pinned at `0`, so an old v0 log containing `request/header-delta` would otherwise pass the version check and silently lose header changes after the delta fold is deleted. Seed/load validation must reject that legacy event fail-loud at the format boundary; no compatibility fold or migration is added.
## Alternatives considered
**Keep linked nodes and compact deltas for possible scale.** Links could help a future cursor API, and deltas can reduce logs when large tool schemas change by a small amount. No shipped cursor uses the links, while full snapshots trade disk size for substantially simpler correctness. If header volume proves material, compression or a measured canonical-delta scheme can be designed around real traces.
## Acceptance criteria
- `SurfaceManager.nodes` is one ordered seq array with no `SurfaceNode`, link fields, or seq-to-node map; incremental append processing and the internal replace-generation signal remain, while the separate public `invalidate()` deletion stays owned by the dead-surface RFC.
- Replaying full changed-header snapshots reconstructs exactly the same requests; no header-delta event/type/codec remains.
- A v0 seed or persisted log containing legacy `request/header-delta` is rejected before replay, with coverage for JSONL and SQLite load paths.
- New-shape v0 JSONL/SQLite replay, provenance, crash repair, compaction, snapshots, invariants, typecheck, coverage, doc-sync, build, and hygiene pass.
## Risks
Full headers increase log volume, and linear replacement lookup could be slower on very large surfaces. Replacements are already linear because the implementation calls `indexOf`; benchmarks should be added only if real traces show the simpler array is a bottleneck. Because the format version remains `0`, forgetting the explicit legacy-event rejection would be silent data corruption rather than a type error; the fail-loud load test is therefore part of the proposal, not optional cleanup.

9
docs/rfc/rejected/architecture/2026-06-11-immutable-public-surfaces.md
View File

@@ -1,25 +1,24 @@
# RFC: Deep-readonly public surfaces
Status: rejected — the pervasive `DeepReadonly<T>` type flip was rejected in favor of an always-on `deriveMessages` clone plus dev-mode `Object.freeze` + invariants. The immutability *goal* shipped via that alternative; see [dev-mode invariants](../../implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md).
Status: rejected — the pervasive `DeepReadonly<T>` type flip is replaced by source-owned runtime immutability in `Session` plus relational development assertions. See [source-owned session immutability and dev-mode invariants](../../implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md).
## Problem
The session log is append-only by contract, but `session.events` returns `readonly SessionEvent[]` whose *elements* are mutable: a plugin can reach in and rewrite history (`events[0].data.content.push(...)`), silently breaking replay equivalence and the derived-history guarantee. The same applies to derived messages and prompt assemblies passed through waterfalls — mutation is sometimes the intended idiom (waterfall middleware mutates the request) and sometimes corruption (mutating a *logged* event), and the types don't distinguish.
The rejected proposal targeted an ownership hole that a `readonly SessionEvent[]` type alone cannot close: its elements remain mutable at runtime, so a cast or plain JavaScript can rewrite nested history. The implemented design closes that hole in `Session` by materializing and deep-freezing every accepted event and returning frozen array snapshots. In-flight prompt waterfalls remain intentionally transformable, so immutability is an ownership boundary rather than a blanket type rule.
## Proposal
> **Implemented differently — see the Status line and [dev-mode invariants](../../implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md).** The `DeepReadonly<T>` design below was rejected as written (compile-only, high type-noise, castable). What shipped: an always-on deep clone in `deriveMessages` (closing the request/adapter aliasing path) plus a dev-mode `Object.freeze` + invariants plugin. The proposal text is kept for the record.
> **Implemented differently — see the Status line and [source-owned session immutability and dev-mode invariants](../../implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md).** The `DeepReadonly<T>` design below is rejected as written: it is compile-only, noisy across consumers, and castable. `Session` instead snapshots and deep-freezes accepted events and public log snapshots in every composition; `deriveMessages()` returns detached frozen projections; the development plugin checks cross-record and cross-seam relationships.
Make immutability part of the type where mutation is corruption:
- `SessionEvent` data becomes `DeepReadonly` on the way OUT of a session (`events`, `session/event` listeners); `append()` keeps taking plain mutable input. A `DeepReadonly<T>` utility type lands in dsh-llm next to the brand/never helpers.
- `deriveMessages()` returns deep-readonly messages; the loop clones before handing a mutable request to the `agent/request` waterfall (mutation there is sanctioned — the clone makes the boundary explicit and cheap, once per step).
- `PromptAssembly` stays mutable through its waterfall (sanctioned) but the registry's internal section list is cloned per assembly (already true).
- Optionally, dev-mode `Object.freeze` of event data behind [the dev-mode invariants](../../implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md) flag, so sanctioned-mutation violations throw in tests rather than corrupting silently.
## Plan
Introduce `DeepReadonly`, flip the session read paths, fix resulting compile errors in consumers (expected: a handful in tests), add the freeze-in-dev option alongside [the dev-mode invariants](../../implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md) plugin.
Introduce `DeepReadonly`, flip the session read paths, and fix the resulting compile errors in consumers.
## Risks

4
docs/rfc/rejected/simplification/2026-06-20-single-session-acp-bridge.md
View File

@@ -4,7 +4,7 @@ Status: rejected — Zed is the current target ACP client and its ACP implementa
## Problem
The ACP bridge now supports multiple live sessions on one JSON-RPC connection. That capability brings multi-entry session maps, reverse session/agent lookups, per-session prompt state, loading ids, demux for every event, cross-session teardown, and isolation concerns for future permission prompts and background tasks. The older [multi-session ACP proposal](../../proposed/feature/2026-06-14-acp-multi-session.md) still tracks the unfinished permission-ownership piece; this RFC is the competing simplification path.
The ACP bridge now supports multiple live sessions on one JSON-RPC connection. That capability brings multi-entry session maps, reverse session/agent lookups, per-session prompt state, loading ids, demux for every event, cross-session teardown, and isolation concerns for future permission prompts and background tasks. The older [multi-session ACP proposal](../../implemented/feature/2026-06-14-acp-multi-session.md) still tracks the unfinished permission-ownership piece; this RFC is the competing simplification path.
The product target has proven it needs concurrent editor conversations over one harness process: Zed's ACP connection owns multiple sessions and load states. The snapshot replay tier still avoids concurrent model streams because its replay entries are positional; that is a test-fixture limitation, not a reason to remove bridge multiplexing.
@@ -20,7 +20,7 @@ Remove the multi-session maps and demux where a single `SessionRecord | undefine
- `session/new` and `session/load` reject while that record exists.
- Event handlers no longer demux across a `Map<sessionId, record>`.
- Multi-session tests are removed or moved under the proposal that continues to defend multiplexing.
- The existing [multi-session ACP proposal](../../proposed/feature/2026-06-14-acp-multi-session.md) is updated to link this RFC and remains the live direction.
- The existing [multi-session ACP proposal](../../implemented/feature/2026-06-14-acp-multi-session.md) is updated to link this RFC and remains the live direction.
## What we give up

37
docs/rfc/rejected/simplification/2026-07-12-collapse-workflow-to-foreground-core.md Normal file
View File

@@ -0,0 +1,37 @@
# RFC: Collapse workflows to the exercised foreground core
Status: rejected — Workflow progress is an intentional observation surface; make it useful through a consumer instead of deleting it.
## Problem
The workflow capability executes foreground JavaScript that composes subagents, but it also carries an unconsumed progress-observation system. No production listener subscribes to any of the six `workflow/*` events; listeners exist only in workflow tests. Nevertheless the seam defines run/phase/agent outcome payloads, the worker sends phase/log/agent lifecycle protocol messages, the host forwards them through a `liveAgents` pairing ledger, and the engine maintains run ids solely to correlate those notifications.
The progress vocabulary is not merely unused; it cannot serve its only named future owner without redesign. `WorkflowRunInfo` contains `{id, meta}` but no parent agent, session, or tool-call identity, while the model-facing tool never exposes the run id. A global ACP listener could not route an event to the correct client session. `meta.phases` is never consulted, `phase(title)` does not validate against it, phase `detail`/`model` and agent `label`/`phase` feed only events, and `whenToUse` is validated and copied but never rendered or selected. `phase()` and `log()` still cross the worker boundary despite having no receiver.
The live handle repeats event-era data after those observers disappear. `WorkflowRun.id` has no non-event consumer, while the tool reads `run.meta.name` only to render a value it already owns as `args.meta.name`; neither belongs on the execution/cancellation handle.
Cancellation also has two public channels for one synchronous start. `WorkflowStartRequest.signal` is passed to the worker host, while the sole production caller separately bridges the same signal to `WorkflowRun.cancel()`. Because `start()` returns the run before control can yield, there is no readiness window that requires request-time cancellation; the duplicate signal adds host listener/disarm state without closing a race.
`WorkflowError.fatal` is the same speculative branch in miniature: every production construction is fatal, `fatal: false` exists only in tests, and combinators already distinguish workflow failures with `instanceof`.
## Proposal
Keep the exercised core: `agent(prompt, { schema, model })`, `parallel`, `pipeline`, `args`, concurrency/agent caps, cancellation, bounded disposal, structured results, worker isolation, and foreground tool collection. Remove all `workflow/*` events and their event-only info/outcome types; remove `phase()`, `log()`, agent `label`/`phase`, phase declarations, `whenToUse`, and their worker messages/host observers; collapse workflow metadata to the name the tool actually uses; remove event-only run ids/meta snapshots and the synthesized agent-end ledger. Shrink `WorkflowRun` to `result`, `cancel()`, and `dispose()`; the tool renders the request-owned name. Remove `WorkflowStartRequest.signal` and the worker host's input-signal listener/disarm state, retaining the caller-owned bridge from its abort signal to `run.cancel()`. Make `WorkflowError` one fatal error class without a boolean mode or `isFatalWorkflowError()` helper.
Amend the implemented dynamic-workflow RFC and update the seam/tool/worker READMEs, tool schema, generated catalogs and package graph, worker type-equivalence records, unit tests, and workflow snapshot/header fixtures. Progress UI work, if commissioned, starts from a correlation contract that names the parent agent/session/tool call instead of reviving this protocol unchanged.
## Alternatives considered
**Keep the prebuilt observation vocabulary for a future UI.** The current shape resembles Claude Code dynamic-workflow metadata, and the host deliberately pairs each forwarded agent start with either the worker's end or a synthesized terminal end. Removing it gives up compatibility-by-shape and makes progress UI a new design task, but the existing payloads still lack routable ownership, so balanced lifecycles alone cannot make the named ACP owner viable without redesign.
## Acceptance criteria
- The workflow public seam contains only execution, cancellation, result, and disposal contracts with a production consumer.
- No workflow event, phase/log protocol message, run-id generator, progress-only metadata, host pairing ledger, or fatal-mode branch remains.
- The run handle has no id/meta echoes, and cancellation has one holder-owned channel after synchronous `start()` returns.
- Parallel/pipeline behavior, caps, cancellation quiescence, worker containment, structured output, and the model-facing workflow scenarios retain coverage.
- Typecheck, coverage, snapshots, doc-sync, module-graph verification, build, and hygiene pass.
## Risks
This is a compile-visible contraction of the workflow DSL, event taxonomy, handle, and start request. Existing workflow calls that supply descriptive metadata, and scripts that use `phase`, `log`, or labels, must shrink; programmatic callers bridge their own abort source to the returned handle; and a future observer must add a better-correlated seam. The execution semantics that make workflows useful do not change.

27
docs/rfc/rejected/simplification/2026-07-12-prune-unused-skill-registry-surface.md Normal file
View File

@@ -0,0 +1,27 @@
# RFC: Prune unused skill registry surface
Status: rejected — Direct runtime skill registration is an intentional extension path for third-party plugins.
## Problem
The skill service's embedded-runtime subsystem has zero production caller of `ctx.skills.register()`. It adds a reserved `runtime` provider name, a runtime map/rank/source, duplicate policy, a second revision in cache keys, normalization, disposers, and tests alongside the provider seam every shipped skill already uses. `SkillSummary.whenToUse` and candidate/definition `path` are parsed and copied but never read by a production consumer: the model catalog renders name/description, resource loading uses `resourceBase`, and providers own their locator. The deliberately open `metadata` extension point stays.
## Proposal
Remove `SkillService.register()`, `SkillRegistration`, the runtime pseudo-provider and reserved-name rules, runtime revisions/cache branches, and runtime-only source/rank normalization. Tests that need an embedded skill register a small real provider. Retain `providerRevision` as the in-flight discovery epoch, but key completed catalogs by cwd alone: every provider mutation synchronously clears the cache, and the post-await revision comparison already prevents inserting stale work. Remove `whenToUse`, `SkillCandidate.path`, and `SkillDefinition.path` from the skill contract and local-provider copies while retaining provider locator/root paths; retain `metadata`, `disableModelInvocation`, `source`, `provider`, `locator`, and `resourceBase` as either deliberate extension vocabulary or production-consumed fields.
Amend the skill-system RFC, README, JSDoc, catalogs, and tests. Agent-scoped system-prompt sections, tool providers, and variables are explicitly outside this proposal: the [agent-scope contributor contract](../../implemented/architecture/2026-07-08-agent-scope-contexts.md) intentionally allows all three to be registered during `setup(agentCtx)` through the agent-owned context, so absence of a fixed in-repo scoped registration is not evidence of non-consumption.
## Alternatives considered
**Keep runtime skill registration for embedders.** It is a deliberate synchronous direct-definition convenience in the implemented skill RFC. A small provider wrapper can expose the same embedded data under effect-owned lifetime, but it must implement async `list()`/`get()`, carry provider identity, and accept provider duplicate semantics. The proposal chooses that one regular path over preserving a second ranking, validation, cache-invalidation, and lookup path.
## Acceptance criteria
- Skill collection has one provider-backed path, a cwd-only completed-cache key, and a revision epoch only for in-flight invalidation; retained skill fields have a production reader or a recorded deliberate extension contract.
- Agent-scoped prompt sections, variables, tool providers, tool guards, and structured-output commit behavior in native and Code Mode remain unchanged.
- Typecheck, coverage, snapshots, doc-sync, module-graph verification, build, and hygiene pass.
## Risks
This is a compile-visible contraction of the pre-release skill registry. External programmatic `list()`/`get()` consumers lose `whenToUse` routing hints and candidate/definition `path`; the shipped model catalog never renders them, and resource resolution keeps its explicit `resourceBase` plus provider-owned opaque locator, but those fields are not observationally identical. Skill-local frontmatter parsing must continue to preserve and validate the supported metadata schema, and external providers remain able to supply embedded, filesystem, remote, or other skill sources.

10
docs/testing.md
View File

@@ -1,17 +1,17 @@
# Testing policy
How this repo tests, tier by tier, and the rules that keep a green suite meaning something. Commands live in the root [AGENTS.md](../AGENTS.md) § Commands; the RFCs linked per tier carry the design rationale.
How this repo tests, tier by tier, and the rules that keep a green suite meaningful. Commands live in root [AGENTS.md](../AGENTS.md); linked RFCs carry the rationale.
## Tiers
- **Unit** (`pnpm run test`): vitest over `packages|examples/*/tests/**/*.spec.ts`, colocated with what they test. Every registry gets an HMR-safety test (dispose the contributing fiber, assert cleanup). Excessive tests are welcome — err toward covering edge cases, error paths, event ordering, and concurrency races; review findings get regression tests (see `packages/core/agent-loop/tests/review-fixes.spec.ts`).
- **Unit** (`pnpm run test`): vitest over `packages|examples/*/tests/**/*.spec.ts`, colocated with what they test. Every registry gets an HMR-safety test (dispose the contributing fiber, assert cleanup). Prefer edge cases, error paths, event ordering, and concurrency races; review findings get regression tests (see `packages/core/agent-loop/tests/review-fixes.spec.ts`).
- **Coverage gate** (`pnpm run test:coverage`): the gating run, per-file 100% on `packages/*/*/src`. An uncovered line is often dead code the gate is correctly flagging for deletion, not a missing test to bolt on. Line coverage is necessary, never sufficient — it proves lines ran, not that the feature works as shipped.
- **Real-API e2e** (`pnpm run test:e2e`): with-key tests against live provider APIs — the DeepSeek model plus provider-specific smokes that gate on their own keys (`EXA_API_KEY`, `PERPLEXITY_API_KEY`, …); each suite self-skips without its key so keyless CI stays green ([real-API e2e RFC](rfc/implemented/testing/2026-06-19-real-api-e2e-ci.md)).
- **Snapshot** (`pnpm run test:snapshot`): boots the real example subprocess, replays a recorded session keyless, diffs normalized stdout + the re-persisted log against committed goldens ([snapshot RFC](rfc/implemented/testing/2026-06-19-acp-snapshot-tests.md)). Re-record with `pnpm run test:snapshot:record`; reviewing the golden diff is part of the review. System-prompt/tool-schema content is pinned by ONE scenario (`text-turn`) and tokenized in every other fixture, so a prompt or schema edit churns one committed line ([pinned-header RFC](rfc/implemented/testing/2026-07-06-pin-request-header-content-in-one-scenario.md)).
- **Snapshot** (`pnpm run test:snapshot`): boots the real example subprocess, replays a recorded session keyless, diffs normalized stdout + the re-persisted log against committed goldens ([snapshot RFC](rfc/implemented/testing/2026-06-19-acp-snapshot-tests.md)). Use `pnpm run test:snapshot:record` when the model transcript should change; use `pnpm run test:snapshot:refresh` when the committed transcript is still the right mock LLM input and replay goldens need keyless rewrite. Review the golden diff. System-prompt/tool-schema content is pinned by ONE scenario (`text-turn`) and tokenized in every other fixture, so a prompt or schema edit churns one committed line ([pinned-header RFC](rfc/implemented/testing/2026-07-06-pin-request-header-content-in-one-scenario.md)).
## The with-key policy: inference is cheap here
We are DeepSeek — do not ration real-API tests. A no-key test proves the plumbing; only a with-key run proves the agent works against a real model. Write many: real prompts that write files, multi-turn conversations, tool use, cancellation mid-stream. Cheapest and highest-value are **smoke tests** that boot the real example, send one real prompt, and check the world — they catch the "green unit tests, broken product" class that mocks structurally cannot ([postmortem 0001](postmortem/0001-acp-default-export-drops-inject.md)). The self-skip exists only so secretless CI and keyless contributors aren't blocked; it is not a cost signal. Every example ships a keyless smoke and — unless keyless-by-nature — a with-key smoke ([examples/AGENTS.md](../examples/AGENTS.md)).
We are DeepSeek — do not ration real-API tests. A no-key test proves plumbing; only a with-key run proves the agent works against a real model. Write many: file-writing prompts, multi-turn conversations, tool use, cancellation mid-stream. Highest-value are **smoke tests** that boot the real example, send one real prompt, and check the world — they catch the "green unit tests, broken product" class that mocks structurally cannot ([postmortem 0001](postmortem/0001-acp-default-export-drops-inject.md)). The self-skip exists only so secretless CI and keyless contributors aren't blocked; it is not a cost signal. Every example ships a keyless smoke and — unless keyless-by-nature — a with-key smoke ([examples/AGENTS.md](../examples/AGENTS.md)).
## Prefer the real implementation over a mock
@@ -25,7 +25,7 @@ An e2e assertion re-runs the command or re-reads the file externally; a keyword
- A plugin shipped via `cordis.yml` needs at least one test through the REAL Loader path: hand-built `ctx.plugin({...})` mounts bypass `unwrapExports` and cannot catch a broken export shape ([postmortem 0001](postmortem/0001-acp-default-export-drops-inject.md); export-shape rules in [packages/AGENTS.md](../packages/AGENTS.md)).
- A guard only guards if the regression actually fails it. For a plugin without `inject` (bundle/composition plugins), a Loader smoke stays green under a broken export shape — add an explicit `expect('default' in mod).toBe(false)` plus an `unwrapExports` round-trip assertion, and prove it: introduce the regression, watch red, revert.
- "Real entry path" means the published artifact: the package `bin` points at built `lib/bin.js` under plain `node`, which tsx masks (settle races, module resolution, a swallowed load failure exiting 0). Keep the built-bin smokes green (`packages/ui/*/tests/built-bin.e2e.ts`), and assert a genuinely-missing config exits non-zero.
- "Real entry path" means the published artifact: the package `bin` points at built `lib/bin.js` under plain `node`, which tsx masks (settle races, module resolution, a swallowed load failure exiting 0). The same applies to any non-index runtime entry the built package resolves at run time (the worker-thread runtime's sibling `lib/worker.js`). Keep the built-artifact smokes green (`packages/ui/*/tests/built-bin.e2e.ts`, `packages/code-runtime/code-runtime-worker/tests/built-lib.e2e.ts`), and assert a genuinely-missing config exits non-zero.
- An e2e that spawns an example from a temp cwd sets `TSX_TSCONFIG_PATH` to the repo-root tsconfig, or it silently falls back to stale built `lib/` ([examples/AGENTS.md](../examples/AGENTS.md)).
## When a snapshot test is required

287
docs/tool-catalog.md
View File

@@ -15,17 +15,117 @@ This table connects model-visible tool names to the plugin package and service s
| Tool package | Model-visible names | Requires | Writes / affects | Shipped aliases | Deployment note |
| --- | --- | --- | --- | --- | --- |
| `@deepseek-ai/dsh-tool-ask-user` | `ask_user_question` | `ctx.tools`, `ctx.userInteraction` | `tool/call`, `tool/result after a UI/provider answers the question` | - | ask_user_question pauses the tool call until the active UI provider returns a human answer. |
| `@deepseek-ai/dsh-tools` | `run_code` | `ctx.tools`, `ctx.codeRuntime (execution time)`, `ctx.systemPrompt` | `tool/call`, `one tool/code-dispatch per bridged sub-call`, `tool/result` | - | Owned by the tool registry as a reserved transport outside filterable capability layers under `mode: code` / `mode: both` (see the Code Mode RFC). Under `code` it is the registry's only wire contribution; the other visible capabilities are declared in a generated TypeScript SDK section, and a program calls them through serialized bindings that re-enter the complete guarded tool pipeline and link each nested execution to this outer result. |
| `@deepseek-ai/dsh-tool-bash` | `bash`, `bash_kill`, `bash_output` | `ctx.tools`, `ctx.bash` | `tool/call`, `tool/result`, `context/message via agent.inject() for background completion notices` | - | The bash/bash_output/bash_kill tools are model-facing consumers of the bash executor seam. |
| `@deepseek-ai/dsh-tool-cordis` | `cordis_inspect`, `cordis_mount`, `cordis_unmount` | `ctx.tools` | `tool/call`, `tool/result`, `live plugin-tree mutations (mount/unmount)` | - | Ships in examples/cordis-agent only (a deliberate opt-in — mounted code gets the real ctx, see docs/rfc/implemented/feature/2026-07-08-self-referential-cordis-toolset.md). Plugins the model mounts may register ADDITIONAL model-visible tools at runtime; the request-header ToolsDelta logs those tool-set changes. |
| `@deepseek-ai/dsh-tool-fs` | `edit`, `read`, `write` | `ctx.tools`, `ctx.fs`, `ctx.systemPrompt` | `tool/call`, `fs/write-intent or fs/edit-intent for mutations`, `fs/observed after successful file operations`, `tool/result` | - | The read-before-write/edit policy is added by `@deepseek-ai/dsh-fs-policy` (an `fs/*` event-gate plugin, no schema change); a deployment that loads these tools is expected to also load it. The tool schemas above are identical with or without the policy plugin. |
| `@deepseek-ai/dsh-tool-skill` | `skill` | `ctx.tools`, `ctx.skills` | `tool/call`, `tool/result` | - | - |
| `@deepseek-ai/dsh-tool-subagent` | `subagent` | `ctx.tools`, `ctx.subagents` | `tool/call`, `tool/result`, `child session events through the chosen provider` | `subagent`, `subagent_fork` | The registered tool name is the load-time `toolName` config (default `subagent`); the schema above is that default. The shipped example agents load this package once per subagent backend, so the model additionally sees `subagent_fork` (bound to the fork backend) with an identical schema — see `examples/coding-agent/cordis.yml` and `examples/acp-agent/cordis.yml`. |
| `@deepseek-ai/dsh-tool-todo` | `todo_write` | `ctx.tools`, `owning Agent session` | `tool/call`, `todo/write`, `tool/result` | - | todo_write is session-owned state; UIs render the latest todo/write event as a checklist or ACP plan. |
| `@deepseek-ai/dsh-tool-workflow` | `workflow` | `ctx.tools`, `ctx.workflows`, `ctx.systemPrompt`, `a calling Agent (exec.agent parents the script children)` | `tool/call`, `tool/result` | - | - |
| `@deepseek-ai/dsh-tool-web` | `web_fetch`, `web_search` | `ctx.tools`, `ctx.web`, `ctx.systemPrompt` | `tool/call`, `tool/result` | - | web_search and web_fetch keep provider selection behind ctx.web so model-visible schemas stay stable across backend swaps. |
## `@deepseek-ai/dsh-tool-ask-user`
### `ask_user_question`
Ask the user a concise question when you need confirmation, a choice, or missing information before proceeding. Send one or more questions, each with a stable id that will be echoed in the answer.
```json
{
"type": "object",
"properties": {
"questions": {
"type": "array",
"description": "Questions to ask the user before continuing.",
"items": {
"type": "object",
"properties": {
"id": {
"type": "string",
"description": "Stable id for this question; echoed in the answer."
},
"question": {
"type": "string",
"description": "The specific question to ask the user."
},
"header": {
"type": "string",
"description": "Optional short heading for the question, such as \"Confirm\" or \"Choose Mode\"."
},
"options": {
"type": "array",
"description": "Optional choices to show the user. If you recommend one, put it first and append \"(Recommended)\" to that label.",
"items": {
"type": "object",
"properties": {
"label": {
"type": "string",
"description": "Short user-facing option label."
},
"description": {
"type": "string",
"description": "One sentence explaining the tradeoff or impact."
}
},
"required": [
"label"
]
}
},
"multi_select": {
"type": "boolean",
"description": "Whether the user may select more than one option. Defaults to false."
}
},
"required": [
"id",
"question"
]
}
}
},
"required": [
"questions"
]
}
```
Source: [`packages/ui/tool-ask-user/src/index.ts`](../packages/ui/tool-ask-user/src/index.ts)
ask_user_question pauses the tool call until the active UI provider returns a human answer.
## `@deepseek-ai/dsh-tools`
### `run_code`
Execute a TypeScript program against the available tools. Write the BODY of an async function (erasable syntax only; top-level `await` and `return` work) and call tools as `await tools.name(args)` per the declarations in the system prompt. Only what you print or return comes back — curate it.
```json
{
"type": "object",
"properties": {
"code": {
"type": "string",
"description": "The program: the body of an async TypeScript function."
}
},
"required": [
"code"
]
}
```
Source: [`packages/core/tools/src/code-mode.ts`](../packages/core/tools/src/code-mode.ts)
Owned by the tool registry as a reserved transport outside filterable capability layers under `mode: code` / `mode: both` (see the Code Mode RFC). Under `code` it is the registry's only wire contribution; the other visible capabilities are declared in a generated TypeScript SDK section, and a program calls them through serialized bindings that re-enter the complete guarded tool pipeline and link each nested execution to this outer result.
## `@deepseek-ai/dsh-tool-bash`
### `bash`
Execute a bash command (`bash -c`) and return its stdout/stderr. Each call runs in a fresh shell: no state (cwd, variables, functions) persists between calls — pass `workdir` instead of using `cd`. Non-zero exits are reported as `[exit code: N]`. Long output is truncated to its tail; the full output is saved to a file whose path is reported when available. Set `run_in_background: true` for long-running commands: the call returns a task id immediately; poll it with `bash_output` and stop it with `bash_kill`.
Execute a bash command (`bash -c`) and return its stdout/stderr. Each call runs in a fresh shell: no state (cwd, variables, functions) persists between calls — pass `workdir` instead of using `cd`. Non-zero exits are reported as `[exit code: N]`. Commands may run under a file sandbox; a blocked file operation is reported as `[sandbox: file access denied under <mode> mode]` — a policy denial, not a bug in the command; do not retry another way (a background task reports the same marker via bash_output once it has finished). Long output is truncated to its tail; the full output is saved to a file whose path is reported when available. Set `run_in_background: true` for long-running commands: the call returns a task id immediately; poll it with `bash_output` and stop it with `bash_kill`.
```json
{
@@ -105,6 +205,78 @@ Source: [`packages/bash/tool-bash/src/index.ts`](../packages/bash/tool-bash/src/
The bash/bash_output/bash_kill tools are model-facing consumers of the bash executor seam.
## `@deepseek-ai/dsh-tool-cordis`
### `cordis_inspect`
Inspect the live cordis runtime that is running THIS agent. Read-only. Sections: `services` (every provided ctx service and the plugin fiber that owns it), `plugins` (a flat list of the loaded plugins with their lifecycle states), `tools` (the model-facing tools currently registered, i.e. what you can call), `dynamic` (plugins you mounted via cordis_mount: id, name, state, provided services, awaited services), `api` (method signatures AND argument/return type shapes for every LIVE service — read this before writing plugin code that calls a service), `events` (every harness event with its dispatch mode and exact signature — pick listener targets here). Omit `what` to get all six sections.
```json
{
"type": "object",
"properties": {
"what": {
"type": "string",
"description": "Limit the report to one section. Omit for all sections.",
"enum": [
"services",
"plugins",
"tools",
"dynamic",
"api",
"events"
]
}
}
}
```
Source: [`packages/cordis/tool-cordis/src/index.ts`](../packages/cordis/tool-cordis/src/index.ts)
### `cordis_mount`
Mount a NEW cordis plugin into the live runtime that is running THIS agent (self-modification). `code` runs as the body of an async JavaScript function in an isolated sandbox and MUST `return` a plugin. Two forms: FUNCTION form `return (ctx) => { … }` — declares no inject, so it can register tools, listen to events, and provide services, but reaching ANY service (e.g. ctx.bash) throws; use it only when you need no services. OBJECT form `return { name?, inject: ['bash', 'llm', …], apply(ctx) { … } }` — declares dependencies, and cordis activates the plugin only after the services exist; PREFER this form. You may reach ONLY the services you list in inject: an undeclared service throws even if it exists, because an undeclared dependency would not be cleaned up if its provider is unmounted. BEFORE calling a service from your code, read cordis_inspect what:"api" — it lists method signatures AND the type shapes of their arguments/returns (do not guess a field's type; e.g. a bash run's stdout is an object, not a string). Inside `apply`, use the standard cordis API: `ctx.on(event, listener)` to observe events (see cordis_inspect what:"events"), or call `harness.registerTool(ctx, harness.defineTool({ name, description, parameters: { text: { type: 'string', required: true } }, async execute(args) { … } }))` to give yourself a new tool — it becomes callable on your NEXT step. Tool parameters: each key IS a property — { type: 'string'|'number'|'boolean'|'object'|'array', required?: true, description?, enum?, items?, properties? }; a JSON-Schema-style { type: 'object', properties, required: […] } wrapper and type 'integer' are also accepted and normalized. A tool's `execute` MUST return an ARRAY of content blocks, e.g. `return [{ type: 'text', text: someString }]` — never a bare string. Mounts can COMPOSE: one plugin may `ctx.provide('name', value)` a service and another may declare `inject: ['name']` to consume it — the consumer stays pending until the provider exists and returns to pending when the provider is unmounted. Everything registered inside `apply` is cleaned up automatically on unmount. Sandbox globals: `console` (tagged `[cordis:<id>]`, writes through to the harness terminal), `harness.defineTool`, `harness.registerTool`, `btoa`, `atob`, `TextEncoder`, `TextDecoder`. Node APIs are DISABLED — do filesystem/network/timer work through the cordis services, never Node built-ins: `require`, `setTimeout`/`setInterval`, and `fetch` throw redirect errors; `process` and `Buffer` are undefined. Instead use inject: ['fs'] + ctx.fs for files, inject: ['web'] + ctx.web for HTTP, inject: ['bash'] + ctx.bash for processes, and inject: ['timer'] + ctx.setTimeout/ctx.setInterval for timing (fiber effects, auto-cleaned on unmount) — cordis_inspect what:"api" shows what THIS runtime provides. Write PLAIN JavaScript, not TypeScript (no `as`, no type annotations). Cautions: (1) waterfall events (e.g. tools/pre-execute) hand the listener a trailing `next` callback which MUST be called — returning without `next()` VETOES the call; prefer plain notification events unless you intend to intercept. (2) Never await something that only resolves after the current turn (your code runs INSIDE a tool call of that turn — it would deadlock). (3) Your `ctx` is a restricted façade: you can register tools, observe events, provide/consume services, and use timers, but framework internals (ctx.root, ctx.fiber, ctx.extend, ctx.plugin, …) are withheld. It is not a security boundary though — the services you inject (e.g. ctx.bash) reach the real runtime.
```json
{
"type": "object",
"properties": {
"code": {
"type": "string",
"description": "Body of an async JS function; must `return` the plugin to mount."
}
},
"required": [
"code"
]
}
```
Source: [`packages/cordis/tool-cordis/src/index.ts`](../packages/cordis/tool-cordis/src/index.ts)
### `cordis_unmount`
Dispose a plugin previously mounted with cordis_mount, by id. All its registrations (event listeners, tools, services) are cleaned up through the cordis effect lifecycle. Returns only after disposal has fully completed (quiescence, not just a request to stop).
```json
{
"type": "object",
"properties": {
"id": {
"type": "string",
"description": "The dynamic mount id returned by cordis_mount (e.g. \"dyn-1\")."
}
},
"required": [
"id"
]
}
```
Source: [`packages/cordis/tool-cordis/src/index.ts`](../packages/cordis/tool-cordis/src/index.ts)
Ships in examples/cordis-agent only (a deliberate opt-in — mounted code gets the real ctx, see docs/rfc/implemented/feature/2026-07-08-self-referential-cordis-toolset.md). Plugins the model mounts may register ADDITIONAL model-visible tools at runtime; the request-header ToolsDelta logs those tool-set changes.
## `@deepseek-ai/dsh-tool-fs`
### `edit`
@@ -199,6 +371,29 @@ Source: [`packages/fs/tool-fs/src/index.ts`](../packages/fs/tool-fs/src/index.ts
The read-before-write/edit policy is added by `@deepseek-ai/dsh-fs-policy` (an `fs/*` event-gate plugin, no schema change); a deployment that loads these tools is expected to also load it. The tool schemas above are identical with or without the policy plugin.
## `@deepseek-ai/dsh-tool-skill`
### `skill`
Load the full instructions for an available skill. Call this with the exact skill name from the session skill catalog before acting on a task that names or clearly matches that skill.
```json
{
"type": "object",
"properties": {
"name": {
"type": "string",
"description": "The exact skill name from the available skills list."
}
},
"required": [
"name"
]
}
```
Source: [`packages/skill/tool-skill/src/index.ts`](../packages/skill/tool-skill/src/index.ts)
## `@deepseek-ai/dsh-tool-subagent`
### `subagent`
@@ -276,6 +471,92 @@ Source: [`packages/todo/tool-todo/src/index.ts`](../packages/todo/tool-todo/src/
todo_write is session-owned state; UIs render the latest todo/write event as a checklist or ACP plan.
## `@deepseek-ai/dsh-tool-workflow`
### `workflow`
Run a JavaScript workflow script that orchestrates subagents at scale. Use this for work that fans out across many independent pieces — an audit over many files, a migration, multi-angle research, adversarial verification of findings — where you write the orchestration as a script instead of delegating turn by turn.
The workflow's identity rides the `meta` parameter as JSON: required `name` (short kebab-case) and `description` strings, optional `whenToUse` string and `phases` array (`{title, detail?, model?}`). The `script` parameter is the plain JavaScript body ONLY (NOT TypeScript, and NO `export const meta` statement — meta is a parameter, not code), running with top-level await; end with `return <value>` — the value must be JSON-serializable and is this tool's result.
Script-body hooks:
- `agent(prompt, opts?): Promise<any>` — run one subagent to completion. Without `opts.schema` it resolves to the child's final text; with `opts.schema` (an object-rooted JSON Schema using ONLY type/properties/required/additionalProperties/items/enum/const — no oneOf/pattern/format/numeric bounds) it resolves to the validated object. Resolves `null` when the child fails (filter with `.filter(Boolean)`). Other opts: `label` (display), `phase` (progress group), `model` (override). Anything else (`effort`/`isolation`/`agentType`) is rejected loudly.
- `pipeline(items, ...stages): Promise<any[]>` — run each item through the stages independently with NO barrier between stages (prefer this for multi-stage work). Each stage receives `(prev, item, index)`. An ordinary stage throw drops that ITEM to `null` and skips its remaining stages.
- `parallel(thunks): Promise<any[]>` — run zero-argument functions concurrently and await ALL of them (a barrier; use only when a stage genuinely needs every prior result together). A throwing thunk resolves to `null`.
- `phase(title)` — start a progress phase; `log(message)` — narrate progress; `args` — the tool call's `args` input, verbatim.
Misused hooks (bad arguments, unknown options, unsupported schemas, tripped caps) throw errors that ALWAYS kill the script — they never dissolve into a per-item `null`.
Constraints: concurrency and total-agent caps apply; no filesystem, network, timers, or Node.js APIs are provided — the agents do the work, the script only coordinates them. The run executes in the foreground: this call returns when the whole script finishes.
```json
{
"type": "object",
"properties": {
"script": {
"type": "string",
"description": "The plain-JS workflow script body (top-level await allowed; NO `export const meta` statement; end with `return <json-value>`)."
},
"meta": {
"type": "object",
"description": "The workflow identity block (plain JSON — never code).",
"properties": {
"name": {
"type": "string",
"description": "Short kebab-case workflow name."
},
"description": {
"type": "string",
"description": "One-line description of what the workflow does."
},
"whenToUse": {
"type": "string",
"description": "Optional guidance on when this workflow applies."
},
"phases": {
"type": "array",
"description": "Optional phase declarations matched by phase() calls.",
"items": {
"type": "object",
"properties": {
"title": {
"type": "string",
"description": "The phase title phase() calls match by exact string."
},
"detail": {
"type": "string",
"description": "Optional one-line description of the phase."
},
"model": {
"type": "string",
"description": "Optional model override this phase is expected to use."
}
},
"required": [
"title"
]
}
}
},
"required": [
"name",
"description"
]
},
"args": {
"type": "object",
"description": "Optional JSON input exposed to the script as the `args` global (wrap a bare list as a field, e.g. {\"files\": [...]})."
}
},
"required": [
"script",
"meta"
]
}
```
Source: [`packages/workflow/tool-workflow/src/index.ts`](../packages/workflow/tool-workflow/src/index.ts)
## `@deepseek-ai/dsh-tool-web`
### `web_fetch`
@@ -289,10 +570,6 @@ Fetch the content of a specific HTTP(S) URL and return it decoded to text.
"url": {
"type": "string",
"description": "The HTTP(S) URL to fetch."
},
"timeout_ms": {
"type": "number",
"description": "Optional fetch timeout in milliseconds (capped by the provider)."
}
},
"required": [

32
docs/tool-execution-pipeline.md
View File

@@ -3,7 +3,7 @@
# Tool Execution Pipeline
This graph shows where policy, hooks, sandboxing, filesystem guards, result rewriting, and UI rendering fit without changing the loop. The key extension points are the `tools/pre-execute` and `tools/post-execute` waterfalls.
This graph shows where policy, hooks, sandboxing, filesystem guards, result rewriting, final-outcome observation, and UI rendering fit without changing the loop. The transformable extension points are the `tools/pre-execute`, `tools/execute`, and `tools/post-execute` waterfalls; monotonic guards and `tools/result` are the owner-enforced boundaries around them.
```mermaid
flowchart TD
@@ -11,29 +11,43 @@ flowchart TD
toolCall["Session event: <code>tool/call</code><br/>logged before execution"]
presentCall["UI pending card<br/>presentCall(args)"]
pre["<code>tools/pre-execute</code> waterfall<br/>hooks, permission, sandbox"]
denied["deny or ask<br/>tool body skipped"]
guards["Registered monotonic guards<br/>deny or abstain; identity protected"]
denied["denied or approval refused<br/>tool body skipped"]
approval["<code>ctx.approval</code> one-shot prompt<br/>absent or unanswerable: deny"]
around["<code>tools/execute</code> waterfall<br/>timeout, retry, metrics (around dispatch)"]
toolBody["Registered tool execute() body"]
fsGate["<code>fs/write-intent</code> or <code>fs/edit-intent</code><br/>tool-fs mutations only"]
owned["Tool-owned session events<br/><code>todo/write</code>, <code>fs/observed</code>, <code>hook/invoked</code>, <code>hook/result</code>"]
owned["Tool-owned session events<br/><code>todo/write</code>, <code>fs/observed</code>, <code>hook/invoked</code>, <code>hook/result</code>, <code>tool/code-dispatch</code>"]
post["<code>tools/post-execute</code> waterfall<br/>accept, block, replace, add context"]
final["<code>tools/result</code> synchronous notification<br/>frozen authoritative outcome"]
context["Buffered additionalContext<br/>context/message after all tool results"]
toolResult["Session event: <code>tool/result</code><br/>single model-facing outcome"]
allResults["All calls in the step settled<br/>and tool/result events recorded"]
presentResult["UI completed card<br/>presentResult(args, result)"]
model --> toolCall
toolCall --> presentCall
toolCall --> pre
pre -->|allow| toolBody
pre -->|deny or ask| denied
pre -->|allow| guards
guards -->|allow| around
guards -->|deny| denied
around --> toolBody
pre -->|deny| denied
pre -->|ask| approval
approval -->|allowed-once| guards
approval -->|rejected, cancelled, unavailable| denied
denied --> post
toolBody --> fsGate
fsGate --> toolBody
toolBody --> owned
toolBody --> post
post --> context
post --> toolResult
toolBody --> around
around --> post
post --> final
final --> toolResult
toolResult --> presentResult
toolResult --> allResults
allResults --> context
```
Filesystem read-before-edit checks live below `tool-fs` on the `fs/*` event gate, while hook bridges and future permission prompts live on the generic tool waterfalls. That split lets the same hooks observe bash, fs, web, todo, and subagent calls without coupling those tools to one policy service.
Filesystem read-before-edit checks live below `tool-fs` on the `fs/*` event gate; hook bridges and approval-triggering permission policy enter through the generic pre/post tool waterfalls, while `ctx.approval` resolves an `ask` before the monotonic guards; owner policy that must not be reordered uses registered guards; and around-dispatch concerns like the tool-call timeout policy (`@deepseek-ai/dsh-timeout-policy`) wrap core dispatch on `tools/execute`. The synchronous `tools/result` notification observes the immutable final outcome after every transform, lossless-JSON validation, and outer error normalization. That split lets the same hooks observe bash, fs, web, todo, skill, and subagent calls without coupling those tools to one policy service. Code Mode rides the whole pipeline twice over: `run_code` is the reserved registry-owned transport whose body enters the pipeline, and each tool call its program makes re-enters `ctx.tools.execute()` — serialized one at a time, carrying the outer execution's opaque token for correlation, and logged as a `tool/code-dispatch` session event, with a deny surfacing to the program as a binding rejection (a sub-call's `additionalContext` is deliberately dropped — no safe outlet mid-run preserves call/result adjacency).
Maintenance mode: curated Mermaid flow; exact tool schemas and event signatures live in generated catalogs.