deepseek-harness/docs/agent-lifecycle.zh.md

<!-- 英文源文件由 scripts/gen-doc-graphs.ts 生成；本中文文件是通过双语配对维护的经评审对侧。
     更新时先运行 `pnpm run gen-doc-graphs` 更新英文，再更新本文件并运行 `pnpm run verify-translation-pairing --write docs/agent-lifecycle.md` 重新记录配对。 -->

# Agent 轮次与步骤生命周期

[English](agent-lifecycle.md) | 中文

此时序图是 [architecture.md](architecture.zh.md#turn-flow) 的配套图示。持久的回放事实保存在 `session/event` 中，实时控制与状态则保存在 `agent/*` 中。

```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->>Prompt: <code>system-prompt/assemble</code> waterfall
  Driver->>Hooks: <code>agent/pre-step</code> waterfall
  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->>Hooks: <code>agent/request</code> waterfall
  Driver->>LLM: prepareCall(config, signal)
  Note over Driver,LLM: cancellation during either async phase 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
  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: prepare and reconcile the same rendered assembly 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
      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->>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/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
```

`assistant/message` 事件会记录每次成功的提供方调用，包括返回空内容或以 `max-tokens` 结束的调用，并嵌入精确的紧凑带时间 stream。空内容不会进入派生历史。失败、重试、取消或 stream error attempt 到达 settlement 时，如果没有 surface message，就会把 stream 记录为 `assistant/attempt`。实时 `agent/assistant-stream` chunk frame 是瞬态数据；回放读取任一种持久 settlement，如果进程在 settlement 前硬中断，则不会留下持久 attempt stream。

`dsh-compaction-basic` 在派生请求之前通过 `agent/pre-step` 处理压力，而 `agent/request-error` 仅用于规范的上下文溢出。任一触发条件满足后，系统都会先执行可选的工具结果剪枝，再选择摘要。恢复发生在仍打开的步骤内，只有剪枝或摘要生成推进 surface replacement generation 时才重试，否则仍以原始请求错误为准。每次重试都会准备调用，并在派生请求之前协调保留的已渲染组装结果，不重复组装、pre-step 或用户消息准入。

以返回的 `agent/pre-step` 决策为准；通过包装 `next()` 的监听器会保留下游消息与 `startsRequestSeries`，除非有意替换。steering（中途引导）和注入的上下文在后续的认领操作取得其下一步骤批次后，会经过同一 waterfall（瀑布式事件）。

需要可回放 transcript（文本记录）数据的 SDK 用户应当消费 `session/event`；`agent/*` 是用于队列与状态、提示词拦截、请求构造、steering、继续执行和错误处理的实时协调接口。

维护模式：英文源文件包含人工维护的 Mermaid 时序图，并由生成器写出；本中文文件作为经评审对侧通过双语配对维护。确切的事件签名位于生成的 Cordis 目录中。