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/todo-multi-in-progress

Browse Source
This commit is contained in:
Chinesezjc
2026-07-27 23:17:37 +08:00
parent 23ea010789 5b5f089bd0
commit a0675cb6c9
456 changed files with 8495 additions and 10470 deletions
Download Patch File Download Diff File Expand all files Collapse all files

86
docs/core-data-structures/session.md
View File

@@ -12,28 +12,17 @@ The append-only event types. Merge-extensible: a plugin declares extra event typ
```ts type-equiv
/**
* Shared payload for user, injected-context, and steering prompt messages. A
* Shared payload for user, injected-context, and steering messages. A
* direct human prompt, a synthetic `agent.inject()` context, and mid-turn
* steering all project into the model transcript as verbatim user-role content;
* they are told apart by `source` (a non-`user` kind marks injected context),
* not by event type. `meta` carries durable model-hidden producer state.
* not by event type.
*/
interface PromptMessageData {
/** Exact model-facing blocks, including any baked prompt-prefix contexts. */
interface UserMessageData {
/** Exact model-facing blocks. */
content: ContentBlock[]
/** Producer provenance for the direct prompt. */
/** Producer provenance. */
source: MessageSource
/** Present only when prompt-prefix contexts were baked into `content`. */
envelope?: PromptMessageEnvelope
/**
* Opaque durable JSON state retained on the event but hidden from the model
* projection. It is the intended channel for a future framing directive (a
* producer declares the frame, a dedicated renderer applies it — see the
* deferred note in
* ../../../../.agents/notes/implemented/simplification/2026-07-20-unwrap-injected-content-envelopes.md),
* so the surface keeps projecting `content` verbatim rather than wrapping it.
*/
meta?: JsonValue
}
```
@@ -46,10 +35,7 @@ interface PromptMessageData {
*/
interface SessionEventMap {
/**
* Opens turn `turn`. `trigger` records what started it — one claimed queued
* message or an idle-time injection. The turn is the durability/replay
* boundary: every event sits between a `turn/start` and its matching
* `turn/end` (the turn-enclosure invariant).
* Opens turn `turn`. `trigger` records what started the model loop.
*/
'turn/start': { turn: number; trigger: TurnTrigger }
/**
@@ -68,16 +54,10 @@ interface SessionEventMap {
* (the queued message claimed for this turn), a synthetic `agent.inject()`
* context (file-change notices, subdir AGENTS.md, skill content, cron
* notifications, …), or an admitted goal continuation round. All three
* project their `content` verbatim; `source` (with a non-`user` kind marking
* injected context) is the only channel that tells them apart. An idle
* injection wraps this event in a one-shot turn so the log stays turn-enclosed.
* project their `content` verbatim; `source` tells them apart. An idle
* injection may append this event between turns without running the model.
*/
'user/message': PromptMessageData
/**
* Durable record of a prompt veto and its reason. It is log-only: the blocked
* prompt never enters the model-visible surface, and its turn runs zero steps.
*/
'prompt/blocked': { content: ContentBlock[]; source: MessageSource; reason: string }
'user/message': UserMessageData
/** Raw stream chunk — token-level replay fidelity. */
'assistant/chunk': { turn: number; step: number; chunk: StreamChunk }
/**
@@ -114,7 +94,7 @@ interface SessionEventMap {
meta?: JsonValue
}
/** Steering content injected between steps of a running turn. */
'steering/message': PromptMessageData & { turn: number }
'steering/message': UserMessageData & { turn: number }
/** Whole-list snapshot; latest write wins on replay. Log-only UI state; never derived history. */
'todo/write': { todos: TodoItem[] }
/**
@@ -125,7 +105,7 @@ interface SessionEventMap {
}
```
`PromptMessageData.content` is always the exact model-facing content. When attached context declares `prompt-prefix` placement, AgentLoop concatenates its blocks, a `## My request:` delimiter, and the effective direct prompt into that array. The optional model-hidden `envelope` retains `displayContent` plus ordered prefix-context source/metadata descriptors, so transcript, title, and re-reference consumers can present the human prompt without changing reconstructable history. `displayPromptContent()` performs that selection and falls back to `content` for ordinary and older events.
`UserMessageData` is the durable `content` and `source` base shared by ordinary prompts, injected context, and steering. Live inbox events extend the same shape with an `AgentMessageId`; the loop adds only driver-owned routing state while an item remains pending.
### `OutOfBandSessionEventMap` — narrow late-append opt-in
@@ -166,13 +146,13 @@ interface TodoItem {
### The request header event: `request/header`
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 Agent Note). A full `request/header` snapshot with reason `'initial'` or `'resume'` records each loop-instance boundary; a later changed request records another full snapshot with reason `'change'`. `foldRequestHeader(events)` reconstructs the header by selecting the latest snapshot. The event is not a `SurfaceEventType`: it produces no LLM message.
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 Agent Note). A full `request/header` snapshot with reason `'initial'` or `'resume'` records each loop-instance boundary; a later changed request records another full snapshot with reason `'change'`. `foldRequestHeader(events)` reconstructs the header by selecting the latest snapshot. The event is not a `SurfaceEventType`: it produces no LLM message.
```ts type-equiv
/**
* Logged request state outside derived history: call config, system prompt,
* tools, and prefix. The latest full `request/header` snapshot reconstructs it;
* canonical empty optional fields are absent.
* Logged request state outside derived history: call config, system prompt, and
* tools. The latest full `request/header` snapshot reconstructs it; canonical
* empty optional fields are absent.
*/
interface EpochHeader {
/** The conversation's call configuration (provider, model, reasoning effort, and sampling scalars). */
@@ -181,18 +161,10 @@ 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, 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`); it is composed once per loop instance and included in every full snapshot that instance records. Legacy v0 logs containing the removed `request/header-delta` event or its full-snapshot `fallback` reason are rejected at seed, append, and persistence-load boundaries rather than replayed incompletely.
Canonical form represents an empty system prompt or tool list as an absent field, matching how requests are built. Legacy v0 logs containing the removed `request/header-delta` event or its full-snapshot `fallback` reason are rejected at seed, append, and persistence-load boundaries rather than replayed incompletely.
## `SessionEvent<T>` — one log entry
@@ -484,7 +456,7 @@ declare class Session {
- `user/message` → a user message carrying exact `content`; an optional envelope remains log-only display metadata.
- `assistant/message` → an assistant message with the event's provider/model provenance and optional adapter-private replay state. Raw `assistant/chunk` events are replay/UI data and are **skipped** in derivation (the assembled message is authoritative). An **empty-content** `assistant/message` is also skipped — a max-tokens step cut off with no content still records an `assistant/message` to host its usage/provenance, but a content-less assistant turn must not enter the provider transcript.
- `tool/result` → a user message carrying a `tool-result` block.
- `user/message` (injected context, i.e. non-`user` source) → a user-role message carrying its `content` verbatim at its chronological position. Optional JSON `meta` remains in the event log and is never rendered.
- `user/message` (injected context, i.e. non-`user` source) → a user-role message carrying its `content` verbatim at its chronological position; provenance and domain data live in its typed source.
- `steering/message` → a user-role message carrying exact `content` at its chronological position; an optional envelope remains log-only display metadata.
Everything else (`turn/*`, `step/*`, plugin-owned `llm/retry`) is structural and does not project into a message. Token accounting reads per-step `assistant/chunk { type: 'usage' }` records and treats `assistant/message.usage` as the committed-step fallback when no usage chunk exists; failed model-request attempts have no assistant message, so their usage chunk is the durable accounting record. An operational error's step number is on `turn/end.reason` for `kind: 'error'`, with normalized `LlmFailure` facts for a final model-request failure and message/code for other live errors. Because this unreleased format intentionally has no compatibility promise, seed/load validation rejects request headers without provider+model and assistant messages without provider/model provenance instead of guessing a route for historical data.
@@ -506,14 +478,12 @@ An explicit `boundary` lets callers fork from a previous completed turn even if
*/
interface TurnTriggerMap {
message: { kind: 'message'; source: MessageSource }
/** Recovery turn reopened over the repaired current session log. */
retry: { kind: 'retry' }
/**
* An out-of-band context injection (`agent.inject()`) made while the agent
* was idle. The loop wraps the injected `user/message` (a non-`user` source,
* plugin by default) in a one-shot turn (`turn/start` → `user/message` →
* `turn/end`) so every event in the log stays turn-enclosed — the
* durability/replay boundary is the turn, and a bare event between turns would
* otherwise be indistinguishable from a crash tail on reload. The trigger's
* `source` mirrors that message's producer.
* An out-of-band producer explicitly enclosed injected context in a one-shot
* turn. `Agent.inject()` appends idle context directly and does not use this
* trigger; the source mirrors the producer of the enclosed `user/message`.
*/
injection: { kind: 'injection'; source: MessageSource }
}
@@ -536,7 +506,8 @@ interface TurnEndReasonMap {
* step number the failure occurred on (the operational error's location — the
* single durable record of an in-turn failure; live diagnostics also fire via
* `agent/error`). Final model-request failures retain their normalized facts
* as one `failure`; other turn failures retain their live Error message/code.
* as one `failure`; other thrown values retain their rendered message and a
* real `HarnessError` code when present.
*/
error: { kind: 'error'; step: number } & (
| { failure: LlmFailure; message?: never; code?: never }
@@ -545,11 +516,6 @@ interface TurnEndReasonMap {
disposed: { kind: 'disposed' }
/** At least one step reached its output-token ceiling, even if a plugin continued the turn. */
'max-tokens': { kind: 'max-tokens' }
/**
* Policy blocked the turn's claimed prompt before the first step. The
* zero-step turn still records a balanced durable boundary and veto reason.
*/
rejected: { kind: 'rejected'; reason: string }
/**
* A persistence backend closed a crash-orphaned turn on reload. The loop never
* emits this marker, and the events recorded before the crash remain intact.
@@ -558,7 +524,7 @@ interface TurnEndReasonMap {
}
```
`max-tokens` mirrors the model-call `FinishReason` of the same name: any `max-tokens` step in a turn makes the whole turn end `max-tokens` rather than `completed` (the cut-short fact wins over a later continuation), so a consumer can tell a clean stop from a truncated one — but only over `completed`: the `disposed`/`aborted`/`error` outcomes take precedence. `rejected` is a zero-step turn whose claimed prompt an `agent/prompt-submit` hook blocked (the ACP bridge maps it to `cancelled`). `interrupted` is the one reason no loop emits — it is synthesized by crash recovery (see [persistence.md](persistence.md)). Both maps are merge-extensible.
`max-tokens` mirrors the model-call `FinishReason` of the same name: any `max-tokens` step in a turn makes the whole turn end `max-tokens` rather than `completed` (the cut-short fact wins over a later continuation), so a consumer can tell a clean stop from a truncated one — but only over `completed`: the `disposed`/`aborted`/`error` outcomes take precedence. `interrupted` is the one reason no loop emits — it is synthesized by crash recovery (see [persistence.md](persistence.md)). Both maps are merge-extensible.
## The turn-enclosure invariant
@@ -568,7 +534,7 @@ Every session event lives **inside** a turn (between a `turn/start` and its `tur
A plugin may declaration-merge extra `SessionEventMap` types. These are **log-only**: NOT `SurfaceEventType`s (they carry no `surfaceOp` and contribute nothing to derived history), but, like every event, they must sit inside an open turn. The full per-event enumeration — core and plugin-contributed alike, with payloads and provenance — is the generated [persistence log event catalog](../persistence-catalog.md); the compaction seam's `compact/*` semantics are discussed on [compaction.md](compaction.md).
The hook bridges' `hook/invoked` / `hook/result` provenance pairs (from `@deepseek-ai/dsh-hook-protocol`) correlate by `handlerId`. The mid-turn hook points (`PreToolUse`/`PostToolUse`/`UserPromptSubmit`/`Stop`) fire inside the loop's open turn, so their `hook/*` records are turn-enclosed by construction. `SessionStart` gets no `hook/*` record — its injected `user/message` is the durable evidence — because it has no open turn to enclose one (see [the hook-bridges Agent Note](../../.agents/notes/implemented/feature/2026-06-30-hook-bridges.md)).
The hook bridges' `hook/invoked` / `hook/result` provenance pairs (from `@deepseek-ai/dsh-hook-protocol`) correlate by `handlerId`. The mid-turn hook points (`PreToolUse`/`PostToolUse`/`Stop`) fire inside the loop's open turn, so their `hook/*` records are turn-enclosed by construction. `SessionStart` and the pre-turn `UserPromptSubmit` admission seam get no `hook/*` record because neither has an open turn to enclose one; allowed context is instead evidenced by its sourced `user/message` (see [the hook-bridges Agent Note](../../.agents/notes/implemented/feature/2026-06-30-hook-bridges.md)).
## Durability contract