Claude Code 源码研究（八）推理逻辑：thinking 块、Plan Mode、4 种正交拆分

推理（Reasoning）= 模型行为
拆分（Decomposition）= 工程提供 4 种正交工具供模型调用
调度（Orchestration）= QueryEngine + query 循环

// src/utils/messages.ts:3011-3015
switch (message.event.content_block.type) {
  case 'thinking':
  case 'redacted_thinking':
    onSetStreamMode('thinking')   // UI 状态切到"思考中"
    return
  ...
}

// src/utils/messages.ts:2964-2974
// 完整 thinking 块捕获用于 transcript 模式
const thinkingBlock = message.message.content.find(b => b.type === 'thinking')
if (thinkingBlock) {
  onStreamingThinking?.(() => ({
    thinking: thinkingBlock.thinking,
    isStreaming: false,
    streamingEndedAt: Date.now(),
  }))
}

// src/query.ts:158 注释
// thinking blocks must be preserved for the duration of an assistant
// trajectory (a single turn, or if that turn includes a tool_use block
// then also its subsequent tool_result and the following assistant message)

export type ThinkingConfig =
  | { type: 'adaptive' }                       // 模型自适应（4.6+ 强制）
  | { type: 'enabled'; budgetTokens: number }  // 固定预算
  | { type: 'disabled' }

const RAINBOW_COLORS = [
  'rainbow_red', 'rainbow_orange', 'rainbow_yellow',
  'rainbow_green', 'rainbow_blue', 'rainbow_indigo', 'rainbow_violet',
]

// 进入计划模式后:
// 1. permissionContext.mode = 'plan'
// 2. 禁止写文件（FileWrite/FileEdit/Bash 写命令）
// 3. 只能 Read/Grep/Glob 探索
// 4. 必须用 ExitPlanModeTool 提交方案供用户审批

In plan mode, you should:
1. Thoroughly explore the codebase to understand existing patterns
2. Identify similar features and architectural approaches
3. Consider multiple approaches and their trade-offs
4. Use AskUserQuestion if you need to clarify the approach
5. Design a concrete implementation strategy
6. When ready, use ExitPlanMode to present your plan for approval

Remember: DO NOT write or edit any files yet.

主代理 query 循环
  └─ tool_use: AgentTool({ subagent_type: 'Explore', prompt: '...' })
       └─ AgentTool.call()
            └─ runAgent({ agentType, prompt, parentMessage, onProgress, thinkingConfig })
                  └─ LocalAgentTask (同进程)
                       ├─ registerTask(appState)
                       ├─ 独立 query 循环
                       │    ├─ updateProgressFromMessage()   追踪进度
                       │    └─ 完成 → SyntheticOutputTool 汇总
                       └─ AgentTool 返回 summary 给主代理

export function getPlanSlug(sessionId?: SessionId): string
export function setPlanSlug(sessionId, slug): void
export function clearPlanSlug(sessionId?): void
export function clearAllPlanSlugs(): void
export const getPlansDirectory: () => string  // memoized
export function getPlanFilePath(agentId?): string
export function getPlan(agentId?): string | null
export async function copyPlanForResume(...)
export async function copyPlanForFork(...)
export async function persistFileSnapshotIfRemote(): Promise<void>

isPlanModeInterviewPhaseEnabled()

用户输入
  ↓
processUserInput.ts        命令解析 / 文件附件 / mention
  ↓
QueryEngine.ask()
  ↓
query() → queryLoop()
  ↓
preProcess (budget / snip / microcompact / autocompact)
  ↓
fetchSystemPromptParts     动态拼 system prompt（含 thinking 提示）
  ↓
API stream                  (模型推理发生在这里)
  ├─ thinking_delta    →   onSetStreamMode('thinking')   ← 推理块
  ├─ text_delta        →   onSetStreamMode('responding') ← 回复
  └─ tool_use_delta    →   拼 tool_use 块                ← 工具
  ↓
stop_reason == 'tool_use'
  └─ toolExecutor 并发 call()
       ├─ Tool.checkPermissions() → 必要时弹用户对话框
       ├─ Tool.call() → 可能 spawn Task（如 AgentTool → LocalAgentTask）
       │              ├─ 子代理内部又跑 query 循环（递归推理）
       │              └─ 返回 summary
       └─ onProgress(yield) 流出进度
  ↓
拼回 tool_result → 下一轮 while（再次推理）
  ↓
stop_reason == 'end_turn' → return Terminal