Files
deepseek-harness/docs/agent-lifecycle.md
T
Yichen Jiang d65835559c refactor(agent): commit plan selections in pre-step and drop agent/step-boundary
agent/pre-step already runs before step/start, and tool resolution and prompt
assembly now follow step/start, so plan-mode can keep committing pending
selections after downstream pre-step listeners accept the step. Tool
conditions and plan guidance still read the committed state, and a rejected
proposal keeps the selection pending as on master. No agent/model listener
depends on plan mode, so the separate serial event had no consumer left.

The loop now claims input, runs agent/model, and then agent/pre-step. The
agent/model order test replaces the boundary tests; plan-mode tests cover
the rejected-proposal case again. Docs, the Agent Note, the upgrade guide,
and generated catalogs follow.
2026-10-15 14:16:00 +08:00

96 lines
6.0 KiB
Markdown

<!-- Generated by scripts/gen-doc-graphs.ts - do not edit by hand.
Run `pnpm run gen-doc-graphs` to regenerate. -->
# Agent Turn And Step Lifecycle
This sequence is the visual companion to [architecture.md](architecture.md#turn-flow). It keeps durable replay facts on `session/event` and live control/status on `agent/*`.
```mermaid
sequenceDiagram
participant User
participant Agent
participant Driver
participant Hooks as hook listeners
participant Prompt as ctx.systemPrompt
participant LLM as ctx.llm
participant Tools as ctx.tools
participant Session
participant SDK as UI or SDK listener
User->>Agent: followup(content)
Agent-->>SDK: <code>agent/inbox/spliced</code>
Agent-->>SDK: <code>agent/inbox/inserted</code> { message }
Agent->>Driver: queued work wakes driver
Driver-->>SDK: <code>agent/status</code> running
Driver->>Session: <code>turn/start</code>
Note over Agent,Driver: claim pending next-step input plus one queued prompt
Driver-->>SDK: <code>agent/inbox/spliced</code> pure deletion
Driver-->>SDK: <code>agent/inbox/claimed</code> { message, turn } per message
Driver->>Hooks: <code>agent/model</code> waterfall chooses provider, model, and effort, published as agent.step
Driver->>Hooks: <code>agent/pre-step</code> waterfall reads agent.step
Hooks-->>Driver: authoritative reject or enter(messages)
alt proposed step rejected, first batch empty, or pre-step failed
Driver-->>Driver: claimed batch stays removed, the open turn spends no step
else enter proposed step
Driver->>Session: <code>step/start</code>
Driver->>Tools: resolve({ agent, signal }) once per step
Driver->>Prompt: <code>system-prompt/assemble</code> contexts, variables, sections, and the step tool set
Driver->>Hooks: <code>agent/request</code> waterfall with the step model selection
Driver->>LLM: prepareCall(config, signal)
Note over Driver,LLM: cancellation during preparation commits neither system nor users
Note over Driver,Session: synchronous admission using the prepared call capability
Driver->>Session: <code>system/message</code> ordered per-node reconciliation
Driver->>Session: <code>user/message</code> per entered message, then a changed runtime-context snapshot
Driver->>Session: <code>request/header</code> and <code>request/context</code> as needed
Driver->>Driver: derive and freeze request from the log
Driver->>LLM: bound prepared call through <code>llm/stream</code> waterfall
LLM-->>Driver: StreamChunk*
Driver-->>SDK: <code>agent/assistant-stream</code> chunk*
alt final adapter or terminal in-band request failure
Driver->>Session: <code>assistant/attempt</code>
Driver-->>SDK: <code>agent/assistant-stream</code> committed end
Driver->>Hooks: <code>agent/request-error</code> waterfall
Hooks-->>Driver: return retry action or preserve the original error
Note over Driver,LLM: retry in the open step: request and prepare again with the same model selection, tool set, and prompt without repeating pre-step or users
else model request succeeded
Driver->>Session: <code>assistant/message</code>
Driver-->>SDK: <code>agent/assistant-stream</code> committed end
Driver->>Tools: classify pending call by executionMode
loop barriers and bounded rolling pool, reclassify before start
opt call starts
Driver->>Session: <code>tool/call</code>
Driver->>Tools: ordered pre, concurrent execute with the step tool set
Tools-->>Session: tool-owned events when applicable
end
opt next model-order result ready
Driver->>Tools: ordered post
Driver->>Session: <code>tool/result</code>
end
end
Driver->>Tools: dispose the step tool set
Driver->>Session: <code>step/end</code>
opt natural stop and next-step inbox empty
Driver->>Hooks: <code>agent/turn-stopping</code> serial terminal checkpoint
end
opt next-step input is pending
Driver-->>Driver: claim pending next-step input
Driver-->>SDK: <code>agent/inbox/claimed</code> { message, turn } per message
Driver->>Hooks: <code>agent/model</code> waterfall
Driver->>Hooks: <code>agent/pre-step</code> waterfall
Hooks-->>Driver: authoritative reject or enter(messages)
end
end
end
Driver->>Session: <code>turn/end</code>
Driver-->>SDK: <code>agent/status</code> idle
```
The `assistant/message` event records every successful provider call, including content-less and `max-tokens` finishes, and embeds the exact compact timed stream. Empty content stays out of derived history. A failed, retried, cancelled, or stream-error attempt that reaches settlement without a surface message records its stream as `assistant/attempt`. Live `agent/assistant-stream` chunk frames are transient; replay reads either durable settlement, and a hard process loss before settlement leaves no durable attempt stream.
`dsh-compaction-basic` uses `agent/pre-step` for pressure before request derivation and `agent/request-error` only for canonical context overflow. Once either trigger qualifies, optional tool-result pruning runs before summary selection. Recovery runs within the open step and retries only when pruning or summarization advances the surface replacement generation; otherwise the original request error remains authoritative. Each retry prepares its call with the step model selection and reconciles the retained rendered assembly before request derivation, without repeating model selection, tool resolution, assembly, pre-step, or user admission.
The returned `agent/pre-step` decision is authoritative; listeners wrapping `next()` preserve downstream messages and `startsRequestSeries` unless replacement is intentional. Steering and injected context pass through the same waterfall after a later claim operation takes their next-step batch.
SDK users that need replayable transcript data should consume `session/event`; `agent/*` is the live coordination API for queue/status, prompt interception, request construction, steering, continuation, and errors.
Maintenance mode: curated Mermaid sequence; exact event signatures live in the generated Cordis catalog.