@toyako/hive 1.0.0

package/docs/AGENT-HIVE-PHASE-5.1-PLANNER-ERROR-AUDIT.md

@@ -0,0 +1,323 @@
+# AGENT-HIVE-PHASE-5.1-PLANNER-ERROR-AUDIT.md
+
+**Audit Date**: 2026-06-24
+**Auditor**: 小白🤖 (AI Agent)
+**Audit Type**: Planner Integration & Error Propagation Audit
+**Package**: @toyako/agent-hive@1.3.0
+
+---
+
+## Part A — Planner Integration Audit
+
+### Step A1: Topology映射检查
+
+**源码**: `src/product/AutoTopologySelector.ts`
+
+```
+planning    → topology: "planExecuteReview"
+architecture → topology: "planExecuteReview"
+research    → topology: "planExecuteReview"
+```
+
+**PLANNER_TOPOLOGY_MAPPING: PASS**
+
+planning/architecture/research 三个意图均映射到 planExecuteReview topology。
+
+---
+
+### Step A2: Planner是否进入Runtime Graph
+
+**源码**: `src/commands/cli.ts` 第99-101行
+
+```typescript
+broker.addAgentProfile({ id: selection.executor, runtimeId: selection.executor, role: "executor", maxConcurrency: 1, status: "idle" });
+broker.addAgentProfile({ id: selection.reviewer, runtimeId: selection.reviewer, role: "reviewer", maxConcurrency: 1, status: "idle" });
+broker.addGraphEdge(selection.executor, selection.reviewer, "reviews", 10);
+```
+
+**证据**: 只添加 executor 和 reviewer，无 planner。
+
+**PLANNER_GRAPH_INTEGRATION: NO**
+
+---
+
+### Step A3: 调用链分析
+
+**全文搜索结果**:
+
+| 文件 | planner/Planner 出现次数 | 是否在执行链中 |
+|------|--------------------------|----------------|
+| cli.ts | 6次 | 仅定义角色/描述，不调用 |
+| Broker.ts | 0次 | ❌ |
+| TaskProcessor.ts | 0次 | ❌ |
+| GraphOperations.ts | 0次 | ❌ |
+
+**CALL_CHAIN**:
+```
+cli.ts (cmdRunWithTask)
+  ↓
+  AutoTopologySelector.select() → topology: "planExecuteReview"
+  ↓
+  broker.addAgentProfile(executor)     ← 只添加executor
+  broker.addAgentProfile(reviewer)     ← 只添加reviewer
+  ↓                                      无planner
+  Broker.submit() → TaskProcessor.process()
+  ↓
+  TaskProcessor.processWithGraph()
+    → executor.execute()
+    → reviewer.review()
+    ↓
+    无planner步骤
+```
+
+---
+
+### Step A4: 真实执行验证
+
+**任务**: "Design architecture for scalable SaaS platform"
+**意图分类**: planning (confidence: 0.48)
+**选择的Topology**: planExecuteReview
+
+**执行日志**:
+```
+[classify] Intent: planning (confidence: 0.48)
+[classify] Keywords: design, architect
+[select]   Intent="planning" → executor=claude(1) reviewer=hermes(1) topology=planExecuteReview
+[select]   Topology: planExecuteReview
+
+[→ claude] Executing...        ← Executor是Claude，不是Planner
+[✓ claude] Done: ...
+[→ hermes] Reviewing...        ← Reviewer是Hermes，不是Planner
+[✓ hermes] PASS (score: 85)
+```
+
+**观察**:
+- ❌ 无 `[→ planner]` 阶段
+- ❌ 无 `Hermes Planning` 阶段
+- ❌ 无 `OpenClaw Planning` 阶段
+- executor=claude, reviewer=hermes（hermes作为reviewer，不是planner）
+
+**PLANNER_RUNTIME_EXECUTION: NO**
+
+---
+
+### Step A5: Planner Output
+
+**PLANNER_OUTPUT = NULL**
+
+无任何Planner输出。
+
+---
+
+## Part B — Error Propagation Audit
+
+### Step B1: 错误控制流图
+
+**源码**: `src/broker/TaskProcessor.ts` 第66-84行
+
+```
+processSimple() / processWithGraph():
+  │
+  ├─ try {
+  │    execResult = await executor.execute(task)
+  │  } catch (err) {
+  │    → 打印 [✗ executor] ERROR
+  │    → 停止pipeline
+  │    → return
+  │  }
+  │
+  ├─ 打印 [✓ executor] Done    ← 不检查execResult.success
+  │
+  └─ reviewLoop()              ← 继续执行Reviewer
+```
+
+**ERROR_CONTROL_FLOW_MAP**:
+- throw Error() → catch块 → 停止pipeline → return
+- success=false → 无catch → 打印"Done" → 继续reviewLoop
+
+---
+
+### Step B2: Executor返回success=false的行为
+
+**源码**: `src/adapters/CodexAdapter.ts` 第68-76行
+
+```typescript
+async execute(task: Task, instruction?: string): Promise<AgentResult> {
+    try {
+      const output = await this.chat(prompt);
+      return { success: true, output: output || "[Codex] Completed" };
+    } catch (err: any) {
+      return { success: false, output: "", error: err.message };  // ← 捕获异常，返回success=false
+    }
+}
+```
+
+**关键发现**: CodexAdapter 捕获所有异常，返回 `{ success: false }`，**永远不抛出异常**。
+
+---
+
+### Step B3: 真实执行验证
+
+**使用无效API key的执行日志**:
+```
+[→ codex] Executing...
+[✓ codex] Done:                              ← success=false，空输出，仍显示"Done"
+[→ claude] Reviewing...                      ← 继续执行Reviewer
+[✗ claude] FAIL (score: 0)
+  issues: Review execution failed: 401 Invalid API Key
+[↻ codex] Revision #1...
+[✓ codex] Revision done:                     ← 仍然显示"Done"
+[→ claude] Reviewing...                      ← 继续执行Reviewer
+[✗] Max revisions reached. FAILED.
+```
+
+**EXECUTOR_FAILURE_BEHAVIOR**:
+- STOPPED_PIPELINE: **NO**
+- REVIEWER_EXECUTED: **YES**
+- FALSE_SUCCESS_MESSAGE: **YES**
+
+---
+
+### Step B4: 失败模式矩阵
+
+| Failure Mode | Pipeline Stop | Reviewer Runs | Status |
+|--------------|---------------|---------------|--------|
+| throw Error() | ✅ YES | ❌ NO | CORRECT |
+| success=false | ❌ NO | ✅ YES | **BROKEN** |
+
+**源码证据**:
+
+**throw Error()**:
+```typescript
+// TaskProcessor.ts 第66-75行
+try {
+  execResult = await this.execWithTimeout(() => executor.execute(task), task.timeout);
+} catch (err: any) {
+  // 捕获异常 → 停止pipeline
+  console.log(`  [✗ ${task.executor}] ERROR: ${err.message}`);
+  this.saveFinal(task, actor);
+  actor.stop();
+  return;  // ← 停止
+}
+```
+
+**success=false**:
+```typescript
+// TaskProcessor.ts 第77-84行
+// 不检查execResult.success
+console.log(`  [✓ ${task.executor}] Done: ${execResult.output.slice(0, 80)}`);
+
+// 继续执行Reviewer
+await this.reviewLoop(task, actor, executor, reviewerName, log, execResult);
+```
+
+---
+
+## Part C — Trust Risk Assessment
+
+### Planner Risk
+
+| Condition | Status |
+|-----------|--------|
+| Planner存在且执行 | - |
+| Planner存在但不执行 | **SPEC_ONLY** |
+| Planner不存在 | - |
+
+**证据**: planExecuteReview topology存在于代码中，但TaskProcessor不调用Planner。
+
+---
+
+### Error Propagation Risk
+
+| Condition | Status |
+|-----------|--------|
+| success=false正确中断 | - |
+| success=false继续执行 | **BROKEN** |
+
+**证据**: CodexAdapter返回success=false时，TaskProcessor打印"Done"并继续执行Reviewer。
+
+---
+
+## Part D — Marketing Claim Validation
+
+### Claim A: "Hermes规划 → Claude实现 → Codex审查"
+
+**状态: UNSUPPORTED**
+
+**证据**:
+- Hermes从未作为Planner被调用
+- cli.ts只添加executor和reviewer到graph
+- TaskProcessor中无planner代码
+- 真实执行日志无planner阶段
+
+---
+
+### Claim B: "OpenClaw规划 → Codex实现 → Claude审查"
+
+**状态: UNSUPPORTED**
+
+**证据**: 同Claim A，Planner未集成到执行流程。
+
+---
+
+### Claim C: "Multi-Agent Workflow"
+
+**状态: PARTIALLY_SUPPORTED**
+
+**证据**:
+- ✅ simpleChain (executor→reviewer) 真实工作
+- ❌ planExecuteReview (planner→executor→reviewer) 未实现planner阶段
+- ❌ Error propagation有缺陷
+
+---
+
+## Final Classification
+
+### D. BOTH_PLANNER_AND_ERROR_CONTROL_BROKEN
+
+**理由**:
+
+1. **Planner未集成**: planExecuteReview topology存在但TaskProcessor不调用Planner。cli.ts只添加executor和reviewer到graph。
+
+2. **Error Propagation缺陷**: CodexAdapter返回success=false时，TaskProcessor打印"[✓] Done"并继续执行Reviewer，不中断pipeline。
+
+---
+
+## 核心结论
+
+### 1. Planner是否真实存在于执行链？
+
+**NO**
+
+- TopologyTemplates.ts定义了planExecuteReview topology
+- AutoTopologySelector为planning/architecture/research选择此topology
+- 但cli.ts只添加executor和reviewer到graph
+- TaskProcessor中无planner代码（0次引用）
+- 真实执行无planner阶段
+
+### 2. Executor success=false是否会停止系统？
+
+**NO**
+
+- CodexAdapter捕获异常，返回success=false
+- TaskProcessor只catch异常，不检查success字段
+- 打印"[✓ executor] Done"即使success=false
+- 继续执行reviewLoop
+
+---
+
+## 证据清单
+
+| 证据 | 来源 | 结论 |
+|------|------|------|
+| Topology映射 | AutoTopologySelector.ts | planning→planExecuteReview |
+| Graph添加 | cli.ts:99-100 | 只添加executor+reviewer |
+| TaskProcessor引用 | TaskProcessor.ts | 0次引用planner |
+| 真实执行日志 | hive run输出 | 无planner阶段 |
+| CodexAdapter错误处理 | CodexAdapter.ts:73-74 | 返回success=false，不抛异常 |
+| TaskProcessor错误处理 | TaskProcessor.ts:66-84 | 只catch异常，不检查success |
+| 真实执行日志 | hive run输出(无效key) | 打印"Done"并继续Reviewer |
+
+---
+
+**审计完成**