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

docs: address model experience review

Browse Source
This commit is contained in:
Tianyi Cui
2026-07-12 02:55:26 +08:00
parent a08485fc80
commit b5cd511f35
64 changed files with 443 additions and 410 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
packages/subagent/subagent-acp/README.md
View File

@@ -4,13 +4,6 @@ The out-of-process **ACP subagent backend**: runs each child agent in a spawned
It is the direction-inverted twin of the server-side bridge in [`@deepseek-ai/dsh-acp`](../../ui/acp): that package is the ACP *agent* (it answers `initialize`/`newSession`/`prompt`); this one is the ACP *client* (it *calls* them and implements the `sessionUpdate`/`requestPermission` callbacks). Point the configured command at the `acp-agent` example to "talk to our own process".
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| Child-agent request | The remote child receives the standalone task through ACP plus its own process's configured system prompt, tools, and fresh session. It receives no parent conversation and cannot enforce the parent's scoped persona or tool filter. | The child pays for an independent full context and its own multi-step history. These tokens never enter the parent's context. |
| Parent tool result, indirectly | Through `dsh-tool-subagent`, the parent receives only the child's final streamed assistant text or a stop-reason error, not intermediate messages or tool traffic. | Parent input grows only by the final result, which is data-dependent and retained until compaction. This provider adds no parent schema itself. |
## What it does
`start(request)` spawns the configured command, wraps its stdio in an ACP `ClientSideConnection`, and drives one session: `initialize` → `newSession` → `prompt`. `run.started` resolves after `newSession` publishes the remote session and rejects when initialization fails or cancellation wins first; the service emits no start/end pair for a child that never became live. The child's streamed `agent_message_chunk` text becomes the `SubagentResult.output`; the prompt's terminal `StopReason` maps to the stop reason. `dispose()` kills the subprocess and awaits its exit.
@@ -77,6 +70,13 @@ The child env is built by [`buildChildEnv` from `@deepseek-ai/dsh-subagent-subpr
Named `name` / `inject` / `Config` / `apply`, with **no default export**: the cordis Loader's `unwrapExports` does `exports.default ?? exports`, so a stray default would collapse the module to the bare function and drop the `inject` namespace (see [docs/postmortem/0001](../../../docs/postmortem/0001-acp-default-export-drops-inject.md)).
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| Child-agent request | The remote child receives the standalone task through ACP plus its own process's configured system prompt, tools, and fresh session. It receives no parent conversation and cannot enforce the parent's scoped persona or tool filter. | The child pays for an independent full context and its own multi-step history. These tokens never enter the parent's context. |
| Parent tool result, indirectly | Through `dsh-tool-subagent`, the parent receives only the child's final streamed assistant text or a stop-reason error, not intermediate messages or tool traffic. | Parent input grows only by the final result, which is data-dependent and retained until compaction. This provider adds no parent schema itself. |
## Known Limitations and Deferred Work
- **A fresh process per run** — persistent-process pooling is a future optimization ([the seam RFC](../../../docs/rfc/implemented/feature/2026-06-21-subagent-capability-seam.md)).

14
packages/subagent/subagent-fork/README.md
View File

@@ -2,13 +2,6 @@
The in-process **fork** subagent backend: a [`SubagentProvider`](../subagent/README.md) that runs each child as a child [`Agent`](../../core/agent) **seeded with a prefix of the parent's session log** — so the child inherits the parent's conversation context instead of starting fresh. Shares the run driver (`startInProcessRun`) with [`dsh-subagent-spawn`](../subagent-spawn/README.md); the only difference is the seed. The shared `run.started` boundary resolves only after the seeded child is published, so `subagent/start` observers see a live registry entry.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| Child-agent history | The child receives the parent's balanced completed-turn surface prefix, then the new task, along with its own scoped persona, tool filter, and optional structured-output contract. The parent's current in-flight turn is excluded. | Forking duplicates the retained completed history into a separate child's requests; the child then accumulates its own tokens independently. A first-turn fork has no inherited history. |
| Parent tool result, indirectly | The parent receives only the child's own final output through `dsh-tool-subagent`, not the inherited prefix or intermediate work. | Parent input grows by one data-dependent final result retained until compaction. |
## The seed boundary (the crux)
At the moment a subagent tool's `execute` runs, the parent's CURRENT turn is open and unbalanced: the log holds the `assistant/message` carrying this spawn's tool-call and the dangling `tool/call` with no `tool/result` yet. Seeding that raw prefix would give the child an open turn that the session constructor and the dev-mode [invariants](../../support/invariants) replay **reject**.
@@ -29,6 +22,13 @@ The seam this rides on: `CreateAgentOptions.seed` (added on `dsh-agent`, threade
See [`dsh-subagent-spawn`](../subagent-spawn/README.md) for the run lifecycle, model inheritance, and depth tracking — all shared.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| Child-agent history | The child receives the parent's balanced completed-turn surface prefix, then the new task, along with its own scoped persona, tool filter, and optional structured-output contract. The parent's current in-flight turn is excluded. | Forking duplicates the retained completed history into a separate child's requests; the child then accumulates its own tokens independently. A first-turn fork has no inherited history. |
| Parent tool result, indirectly | The parent receives only the child's own final output through `dsh-tool-subagent`, not the inherited prefix or intermediate work. | Parent input grows by one data-dependent final result retained until compaction. |
## Known Limitations and Deferred Work
- **Runs expose no `sendMessage`/`resume`** — the optional runtime capabilities are absent on in-process runs; the consumer collects synchronously.

14
packages/subagent/subagent-inprocess/README.md
View File

@@ -2,13 +2,6 @@
The shared **in-process subagent run driver**. A library with no provider or import-time registration that the in-process backends — [spawn](../subagent-spawn/README.md) (a fresh child) and [fork](../subagent-fork/README.md) (a child seeded with a prefix of the parent's log) — both build on. Each accepted run installs one provider-owned cleanup effect. The backends are thin shells that differ ONLY in the session seed they pass; everything downstream lives here, so neither backend depends on the other.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| Child-agent request | The shared driver sends the task as the child's user message and composes per-child scoped persona and tool restrictions. Structured runs add a scoped instruction plus `structured_output` in the visible schema or Code Mode SDK, then stop after a committed capture. Spawn supplies no history; fork supplies its balanced seed. | Child input is isolated from the parent and grows through the child's own steps. Structured output adds fixed instruction and capability tokens only to that child for that run. |
| Parent result, indirectly | The driver extracts only the child's own last assistant output or captured structured value; seeded parent messages and intermediate child work do not become the result. | The parent receives one data-dependent result through the consumer; all other child tokens stay in the child session. |
## What it exports
### `startInProcessRun(ctx, request, options): SubagentRun`
@@ -45,6 +38,13 @@ Delegation depth rides on a merge-extensible `AgentOptions.subagentDepth` field
Thrown by `startInProcessRun` when a spawn would exceed the request's `maxDepth` cap; carries `attemptedDepth` and `maxDepth`.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| Child-agent request | The shared driver sends the task as the child's user message and composes per-child scoped persona and tool restrictions. Structured runs add a scoped instruction plus `structured_output` in the visible schema or Code Mode SDK, then stop after a committed capture. Spawn supplies no history; fork supplies its balanced seed. | Child input is isolated from the parent and grows through the child's own steps. Structured output adds fixed instruction and capability tokens only to that child for that run. |
| Parent result, indirectly | The driver extracts only the child's own last assistant output or captured structured value; seeded parent messages and intermediate child work do not become the result. | The parent receives one data-dependent result through the consumer; all other child tokens stay in the child session. |
## Known Limitations and Deferred Work
- **Runs expose no `sendMessage`/`resume`** — the optional runtime capabilities are absent on in-process runs; the consumer collects synchronously.

14
packages/subagent/subagent-spawn/README.md
View File

@@ -4,13 +4,6 @@ The in-process **spawn** subagent backend: a [`SubagentProvider`](../subagent/RE
The run mechanics live in the shared [`@deepseek-ai/dsh-subagent-inprocess`](../subagent-inprocess/README.md) driver (`startInProcessRun`); this backend just passes **no seed** (a fresh child). The [fork](../subagent-fork/README.md) backend is an independent peer over the same driver — neither knows about the other.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| Child-agent request | The fresh child receives the standalone task, inherits the parent model and workspace by default, and sees the globally composed prompt and tools after its scoped persona and tool filter. It receives zero parent conversation messages. | The child pays for a new independent context and history; no parent-history tokens are duplicated. |
| Parent tool result, indirectly | Through `dsh-tool-subagent`, the parent receives only the child's final output or stop-reason error. | Parent input grows by one data-dependent result retained until compaction. |
## What it does
`start(request)` delegates to `startInProcessRun(ctx, request, {})` with no seed: a fresh child agent with the parent's `cwd`/`parentSession` lineage and (by default) the parent's model. The driver creates one run-owner fiber under `parent.ctx`; parent teardown, this provider's teardown, and manual disposal all converge there before child publication. Its `run.started` boundary resolves only after the fresh child is published, so `subagent/start` observers see a live registry entry. See the [driver README](../subagent-inprocess/README.md) for the full lifecycle (depth check, one-shot drive, result read, dispose).
@@ -25,6 +18,13 @@ The run mechanics live in the shared [`@deepseek-ai/dsh-subagent-inprocess`](../
|---|---|
| `providerName` | Registry name on `ctx.subagents` (default `spawn`). |
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| Child-agent request | The fresh child receives the standalone task, inherits the parent model and workspace by default, and sees the globally composed prompt and tools after its scoped persona and tool filter. It receives zero parent conversation messages. | The child pays for a new independent context and history; no parent-history tokens are duplicated. |
| Parent tool result, indirectly | Through `dsh-tool-subagent`, the parent receives only the child's final output or stop-reason error. | Parent input grows by one data-dependent result retained until compaction. |
## Known Limitations and Deferred Work
- **Runs expose no `sendMessage`/`resume`** — the optional runtime capabilities are absent on in-process runs; the consumer collects synchronously.

12
packages/subagent/subagent-subprocess/README.md
View File

@@ -4,12 +4,6 @@ Shared machinery for **out-of-process subagent backends** — providers that spa
Every tunable is a **parameter**: the dispose ladder takes its grace periods per call, the config-dir helper takes an optional pinned path. Defaults live in each consuming plugin's Config (defaulted, validated fields changeable from `cordis.yml`), never in this library.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| None directly | This process utility registers no provider, prompt, tool, or message. A consuming backend's child application decides the child's model context; environment scrubbing and isolated config directories prevent ambient credentials and user state from silently changing that composition. | Zero direct tokens. It can indirectly stabilize child context, but it adds no text to parent or child requests. |
## What it exports
### `SENSITIVE_ENV_PATTERN` / `buildChildEnv(extra)`
@@ -45,6 +39,12 @@ A per-run isolated config directory for an external CLI child (the target of `CL
`tests/subagent-subprocess.spec.ts`: the env scrub and config-dir helpers run against the real process env and real filesystem (the rm-failure path injects its rejection at the fs boundary — a real recursive-rm failure is not portably provokable, and root ignores permission bits); the exit waits and the dispose ladder run against a scriptable fake child, driving each escalation tier deterministically. The [ACP backend suite](../subagent-acp/README.md) exercises the same ladder against real subprocesses (EOF-cooperative, EOF-ignoring, and SIGTERM-trapping children) end to end.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| None directly | This process utility registers no provider, prompt, tool, or message. A consuming backend's child application decides the child's model context; environment scrubbing and isolated config directories prevent ambient credentials and user state from silently changing that composition. | Zero direct tokens. It can indirectly stabilize child context, but it adds no text to parent or child requests. |
## Known Limitations and Deferred Work
- **The credential scrub is name-based** — only variables matching `KEY` / `SECRET` / `TOKEN` are removed; differently named secrets such as `PASSWORD` pass through unless the backend supplies a stricter environment.

12
packages/subagent/subagent/README.md
View File

@@ -14,12 +14,6 @@ This package is the interface third of the capability seam, split so each concer
Unlike the bash seam (one executor per context, second load throws), **multiple providers coexist** here. Each registers under a unique name and a caller picks one by name — the shape mirrors the LLM adapter registry (`LlmService.registerAdapter`), not the single-service bash executor. This is the requirement that rules out the bash shape: an agent may want an in-process child for a cheap subtask and an out-of-process ACP child for an isolated one, in the same runtime.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| None directly | The provider registry registers no prompt or tool. Provider lifecycle makes a bound `dsh-tool-subagent` schema appear or disappear, and `inheritsParentContext` selects truthful fresh-versus-fork wording. Run events are observe-only. | Zero direct tokens. Child prompts and final results enter model contexts only through a provider and consumer. |
## Service API (`ctx.subagents`)
| Member | Semantics |
@@ -42,6 +36,12 @@ Beside `capabilities` sits one DESCRIPTIVE fact, not validated by the service: `
The service also announces provider lifecycle: `subagent/provider-added` (the frozen registry snapshot) fires after a registration and `subagent/provider-removed` (the accepted name) after an unregistration, so a consumer deriving state from a named provider (the model-facing tool wording) mirrors registry membership instead of assuming load order — the cordis Loader starts sibling plugins concurrently, so "listed earlier" does not mean "registered earlier". Run lifecycle is gated by provider readiness: `subagent/start` (payload `SubagentRunInfo`) fires only after `run.started` fulfills, and `subagent/end` (payload `SubagentRunEndInfo`) fires only for that announced run; readiness rejection emits neither. For spawn/fork, the start listener can resolve the published child via `ctx.agents.get(info.id)`; a remote provider need not have a local registry entry. Both events are **observe-only** plain emits. The service observes `result` immediately even while readiness is pending, clones its output before the caller can mutate it, and buffers that end payload until start has fired; a rejecting result cannot become an unhandled detached promise, start always precedes end, and a listener cannot corrupt the caller's result. `subagent/end` carries the cloned output as `lastAssistantMessage` on the settle path and omits it on infrastructure rejection. Any run-affecting decision is out of scope for this observe-only surface.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| None directly | The provider registry registers no prompt or tool. Provider lifecycle makes a bound `dsh-tool-subagent` schema appear or disappear, and `inheritsParentContext` selects truthful fresh-versus-fork wording. Run events are observe-only. | Zero direct tokens. Child prompts and final results enter model contexts only through a provider and consumer. |
## Known Limitations and Deferred Work
- **The consumer collects synchronously** — it starts a run and awaits `result`; steering (`sendMessage`) is part of the contract but intentionally unused, and background / poll / spill semantics are deferred to a future redesign unifying long-running-tool handling across subagents and bash ([the seam RFC](../../../docs/rfc/implemented/feature/2026-06-21-subagent-capability-seam.md)).

14
packages/subagent/tool-subagent/README.md
View File

@@ -2,13 +2,6 @@
The model-facing `subagent` tool: delegate a self-contained task to a child agent and return its final output. Pure schema + lifecycle shaping over the [`ctx.subagents`](../subagent/README.md) provider registry — an in-process, ACP, or future A2A backend swaps in without changing what the model sees.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| Tool schema | While the configured provider exists, the parent model sees one `{ description, prompt }` tool under `toolName`. Its description explicitly says whether the child inherits completed turns or needs a standalone prompt; persona, model, filter, depth, and provider choice remain deployment config. | Fixed schema cost per parent request while mounted. Removing the provider removes the whole schema; exposing multiple providers adds one independently named schema per load. |
| Tool-call history and result | The task description and full prompt remain in the parent assistant tool call. The result contains only the child's final text or a stop-reason error, never intermediate child steps. | Prompt and final output are data-dependent retained tokens. All child working context is paid in the child and omitted from the parent. |
## Provider selection is config, not model-facing
This plugin binds to **exactly one** provider (`Config.provider`). The model sees only `{ description, prompt }` — there is no provider/type parameter in the schema. To expose more than one transport, load the plugin more than once, each bound to a different provider **and a distinct `toolName`** (the tool registry rejects a duplicate name, so a second load that kept the default `subagent` name would throw). Keeping selection in config (not the schema) is the deliberate split: the *service* holds a multi-provider registry; the *tool* picks one.
@@ -32,6 +25,13 @@ The tool description and the `prompt` parameter description are DERIVED from the
Background / poll collection is deferred (see the [RFC](../../../docs/rfc/implemented/feature/2026-06-21-subagent-capability-seam.md)); this cut blocks the parent turn until the child finishes.
## Model Experience
| Context surface | What the model sees | Token effect |
|---|---|---|
| Tool schema | While the configured provider exists, the parent model sees one `{ description, prompt }` tool under `toolName`. Its description explicitly says whether the child inherits completed turns or needs a standalone prompt; persona, model, filter, depth, and provider choice remain deployment config. | Fixed schema cost per parent request while mounted. Removing the provider removes the whole schema; exposing multiple providers adds one independently named schema per load. |
| Tool-call history and result | The task description and full prompt remain in the parent assistant tool call. The result contains only the child's final text or a stop-reason error, never intermediate child steps. | Prompt and final output are data-dependent retained tokens. All child working context is paid in the child and omitted from the parent. |
## Known Limitations and Deferred Work
- **Delegation blocks the parent turn** — synchronous collect only; background start + poll collection is deferred to the long-running-runtime redesign.