fix: 稳住 autopilot 循环并升级 openclaw

shuanbao · shuanbao · commit e373cc3d6ae3 · 2026-04-11T23:11:08.000+08:00
- 自检硬校验空 output 直接判零分,避免无产出任务被 LLM 打高分进入 rework 死循环
- readTaskOutput 支持逗号/数组多路径与 legacy 内联内容,多文件任务不再因单路径读失败整个置空
- parseTaskAssignments 扩展否定短语与动作动词白名单,过滤 chief 叙述型文本避免僵尸任务污染
- 部门循环分离 task 空闲与 agent 空闲,加 in_progress/rework 硬上限防止 agent 假汇报长期挂单
- CLAUDE.md 精简冗长架构图
- openclaw 升级到 2026.4.9

Co-Authored-By: shuanbao &lt;shuanbao0@gmail.com&gt;
diff --git a/CLAUDE.md b/CLAUDE.md
diff --git a/core/autopilot/department-loop.cjs b/core/autopilot/department-loop.cjs
@@ -232,6 +232,12 @@ async function autoTransitionTasks(deptId, config, chiefResponseText, options =
     const idleMins = activity
       ? activity.idleMins
       : Math.floor((Date.now() - new Date(task.updatedAt || task.createdAt).getTime()) / 60000)
+    // Separate "agent idle" vs "task idle". Agent idle is reset every time the
+    // dept-loop queries the agent's :main for status, so for a task in review
+    // that's what we've been measuring wrong: the task can sit for an hour in
+    // review while the agent is trivially reactive to our own pings. Use the
+    // task's own updatedAt for the review-pickup decision.
+    const taskIdleMins = Math.floor((Date.now() - new Date(task.updatedAt || task.createdAt).getTime()) / 60000)
 
     if (task.status === 'assigned') {
       if (idleMins < 5) {
@@ -242,11 +248,20 @@ async function autoTransitionTasks(deptId, config, chiefResponseText, options =
         logger.warn('dept-loop', `Assigned task ${task.id} never started, marked failed (agent ${assignee} idle ${idleMins}m)`)
       }
     } else if (task.status === 'review') {
-      // Collect review tasks for parallel quality gate processing
-      if (idleMins >= IDLE_COMPLETE_MINS) {
+      // Use taskIdleMins not agent idleMins — dept-loop pings reset agent idle.
+      if (taskIdleMins >= IDLE_COMPLETE_MINS) {
         reviewItems.push({ task, assignee })
       }
     } else if (task.status === 'in_progress' || task.status === 'rework') {
+      // Hard ceiling on dual-session: if the task's own updatedAt shows it has
+      // been stuck for more than STALE_TASK_MINS * 2, move it on regardless of
+      // what the agent reports. Without this, agents that keep replying
+      // "working" on a ghost task can stall the same task for hours.
+      if (taskIdleMins >= STALE_TASK_MINS * 2) {
+        transition(task, assignee, task.status, 'failed', `task 停滞 ${taskIdleMins}m，超过硬上限`)
+        logger.warn('dept-loop', `Hard ceiling: ${task.id} stuck ${taskIdleMins}m regardless of agent status → failed`)
+        continue
+      }
       if (dualEnabled && statusQueryResults) {
         // Dual-session path: use explicit status query instead of idle guessing
         const status = statusQueryResults[assignee]
diff --git a/core/repo/task.cjs b/core/repo/task.cjs
@@ -213,20 +213,52 @@ class TaskRepository extends BaseRepository {
   }
 
   /**
-   * Read the content of a task's output file
-   * @param {object} task - Task with optional output path
-   * @returns {string|null} file content or null if not available
+   * Read the content of a task's output file(s).
+   *
+   * task.output can be:
+   *   - a single file path: "projects/foo/docs/report.md"
+   *   - comma/newline-separated paths: "a.md, b.md, c.md"
+   *   - an array of paths: ["a.md", "b.md"]
+   *   - inline content (legacy): a string that isn't a path
+   *
+   * Returns concatenated content from all existing files (each prefixed with
+   * a "# <path>" header), or null if no paths resolve to readable files.
+   * For inline-content task.output (no path-like tokens), returns it as-is.
+   *
+   * @param {object} task
+   * @returns {string|null}
    */
   readTaskOutput(task) {
     if (!task || !task.output) return null
-    const outputPath = resolve(PROJECT_ROOT, task.output)
-    try {
-      if (!existsSync(outputPath)) return null
-      return readFileSync(outputPath, 'utf-8')
-    } catch (err) {
-      logger.debug('task-repo', 'failed to read task output', { outputPath, error: err.message })
-      return null
+
+    const raw = task.output
+    const tokens = Array.isArray(raw)
+      ? raw
+      : String(raw).split(/[\n,]+/).map(s => s.trim()).filter(Boolean)
+
+    if (tokens.length === 0) return null
+
+    const contents = []
+    let sawPathLike = false
+    for (const token of tokens) {
+      // Skip tokens that don't look like file paths (no slash, no .ext) —
+      // those are likely inline text, not a path.
+      if (!/[\\/]/.test(token) && !/\.[a-z0-9]{1,8}$/i.test(token)) continue
+      sawPathLike = true
+      const abs = resolve(PROJECT_ROOT, token)
+      try {
+        if (!existsSync(abs)) continue
+        const body = readFileSync(abs, 'utf-8')
+        contents.push(tokens.length > 1 ? `# ${token}\n\n${body}` : body)
+      } catch (err) {
+        logger.debug('task-repo', 'failed to read task output path', { path: abs, error: err.message })
+      }
     }
+
+    if (contents.length > 0) return contents.join('\n\n---\n\n')
+    // No path-like tokens at all → treat original as inline content (legacy).
+    if (!sawPathLike && typeof raw === 'string') return raw
+    return null
   }
 
   /** Update a task in-place via DB + sync to file if project task */
diff --git a/core/task/auto-transition.cjs b/core/task/auto-transition.cjs
@@ -16,6 +16,13 @@ const logger = require('../common/logger.cjs')
  * @param {string} text
  * @returns {Array<{agentId: string, summary: string, projectId?: string}>}
  */
+// Matches phrases chief uses to signal "no new task for this agent". Kept as a
+// single regex so additions go in one place. The patterns are intentionally
+// broad because chief prose varies a lot; false negatives are much worse than
+// false positives here (a skipped line just means no task is created, while a
+// missed filter creates a zombie task that pollutes the project).
+const NO_ASSIGNMENT_RE = /无需分配|不需要分配|不需要|跳过|不分配|不新增|不新建|不加并发|不增加|本轮(?:不|无)|暂不|暂无|维持(?:现状|低并发|冻结|不)|冻结新分配|继续(?:收口|执行|冻结)|先完成|无新(?:任务|增)|引用现有任务|当前忙碌|当前🔴|当前🔵|🔴\s*busy|已有\s*\d+\s*个进行中/
+
 function parseTaskAssignments(text) {
   if (!text) return []
   const match = text.match(/[\[【]任务分配[\]】]\s*\n([\s\S]*?)(?=\n[\[【]|$)/)
@@ -27,14 +34,19 @@ function parseTaskAssignments(text) {
     const m = line.match(/^(?:[-*]|\d+[.)]\s*)\s*(\S+?)[:\uff1a]\s*(.+?)(?:\s*[\(\uff08].*[\)\uff09])?\s*$/)
     if (!m) continue
     const [, agentId, rawSummary] = m
-    if (/无需分配|不需要|跳过/.test(rawSummary)) continue
+    if (NO_ASSIGNMENT_RE.test(rawSummary)) continue
     if (/task-[a-z0-9]/.test(line)) continue
 
     // Extract optional [project: xxx]
     const projMatch = rawSummary.match(/\[project:\s*([^\]]+)\]/)
     const projectId = projMatch ? projMatch[1].trim() : undefined
     const summary = rawSummary.replace(/\[project:\s*[^\]]+\]\s*/, '').trim()
 
+    // Reject if the cleaned summary looks like a status sentence rather than a
+    // concrete task directive. A real assignment has an action verb (分配/
+    // 创建/产出/写/review 等); status sentences don't.
+    if (!/分配|创建|产出|写|实现|撰|制作|修复|修正|调试|补齐|生成|设计|开发|实施|执行|交付|补充|补交|新增|新建|research|review|test|implement|write|build|create|fix|add|update|refactor/i.test(summary)) continue
+
     const entry = { agentId, summary }
     if (projectId) entry.projectId = projectId
     assignments.push(entry)
diff --git a/core/task/quality-orchestrator.cjs b/core/task/quality-orchestrator.cjs
@@ -186,6 +186,28 @@ class QualityOrchestrator {
     }
   }
 
+  /**
+   * @private
+   * Given a task.output value (string or array), return the subset of
+   * path-like tokens that do not resolve to an existing file. Used for
+   * clearer self-check failure messages when multi-path outputs are in play.
+   */
+  _listMissingOutputPaths(output) {
+    if (!output) return []
+    const { existsSync } = require('fs')
+    const { resolve } = require('path')
+    const { PROJECT_ROOT } = require('../common/paths.cjs')
+    const tokens = Array.isArray(output)
+      ? output
+      : String(output).split(/[\n,]+/).map(s => s.trim()).filter(Boolean)
+    const missing = []
+    for (const token of tokens) {
+      if (!/[\\/]/.test(token) && !/\.[a-z0-9]{1,8}$/i.test(token)) continue
+      if (!existsSync(resolve(PROJECT_ROOT, token))) missing.push(token)
+    }
+    return missing
+  }
+
   /**
    * Select a peer reviewer from the department.
    */
@@ -252,25 +274,29 @@ class QualityOrchestrator {
   async _requestSelfCheck(agentId, task, deptId) {
     if (!agentId) return { passed: false, score: 0, checklist: ['无执行者'], at: new Date().toISOString() }
 
-    // Hard validation: check output content via injected readTaskOutput
-    if (task.output) {
-      const content = this._readTaskOutput(task)
-      if (!content) {
-        return { passed: false, score: 0, checklist: ['产出文件不存在: ' + task.output], at: new Date().toISOString() }
-      }
-      if (content.length < 500) {
-        return { passed: false, score: 0, checklist: [`产出仅 ${content.length} 字符，最低要求 500`], at: new Date().toISOString() }
-      }
-      if (/\$\{[^}]+\}/.test(content.slice(0, 5000))) {
-        return { passed: false, score: 0, checklist: ['含未渲染模板变量 ${...}'], at: new Date().toISOString() }
-      }
+    // Hard validation: empty output is an automatic fail. A task that reaches
+    // self-check without any output field means nothing was produced —
+    // previously the `if (task.output)` guard would silently skip validation
+    // and let the LLM grade a phantom task, which is how empty rework chains
+    // kept scoring high without ever writing anything.
+    if (!task.output) {
+      return { passed: false, score: 0, checklist: ['产出为空：task.output 字段未设置'], at: new Date().toISOString() }
     }
-
-    let outputContent = ''
-    if (task.output) {
-      const raw = this._readTaskOutput(task)
-      outputContent = raw ? raw.slice(0, 5000) : `(无法读取: ${task.output})`
+    const content = this._readTaskOutput(task)
+    if (!content) {
+      const missing = this._listMissingOutputPaths(task.output)
+      const detail = missing.length > 0 ? missing.join(', ') : String(task.output)
+      return { passed: false, score: 0, checklist: ['产出文件不存在: ' + detail], at: new Date().toISOString() }
     }
+    if (content.length < 500) {
+      return { passed: false, score: 0, checklist: [`产出仅 ${content.length} 字符，最低要求 500`], at: new Date().toISOString() }
+    }
+    if (/\$\{[^}]+\}/.test(content.slice(0, 5000))) {
+      return { passed: false, score: 0, checklist: ['含未渲染模板变量 ${...}'], at: new Date().toISOString() }
+    }
+
+    // `content` was already read during hard validation above
+    const outputContent = content.slice(0, 5000)
 
     // Build checklist: prefer type-specific from task-standards.md, fallback to generic
     let checklistItems
diff --git a/package.json b/package.json
@@ -22,7 +22,7 @@
     "@anthropic-ai/sdk": "^0.39.0",
     "better-sqlite3": "^11.10.0",
     "clawhub": "^0.7.0",
-    "openclaw": "^2026.3.31",
+    "openclaw": "^2026.4.9",
     "ws": "^8.19.0"
   },
   "engines": {
diff --git a/tests/core/repo/task.test.cjs b/tests/core/repo/task.test.cjs
@@ -1,6 +1,9 @@
 'use strict'
 const { describe, it } = require('node:test')
 const assert = require('node:assert/strict')
+const { mkdtempSync, writeFileSync, rmSync } = require('fs')
+const { tmpdir } = require('os')
+const { join } = require('path')
 
 describe('TaskRepository', () => {
   it('exports TaskRepository and taskRepo', () => {
@@ -41,4 +44,79 @@ describe('TaskRepository', () => {
     assert.deepEqual(task.assignees, ['writer-a'])
     assert.equal(task.assignedAgent, 'writer-a')
   })
+
+  describe('readTaskOutput', () => {
+    const { taskRepo } = require('../../../core/repo/task.cjs')
+
+    it('returns null for missing / empty output', () => {
+      assert.equal(taskRepo.readTaskOutput(null), null)
+      assert.equal(taskRepo.readTaskOutput({}), null)
+      assert.equal(taskRepo.readTaskOutput({ output: '' }), null)
+    })
+
+    it('reads a single existing file path (absolute)', () => {
+      const dir = mkdtempSync(join(tmpdir(), 'af-readoutput-'))
+      try {
+        const f = join(dir, 'a.md')
+        writeFileSync(f, '# Hello')
+        const content = taskRepo.readTaskOutput({ output: f })
+        assert.equal(content, '# Hello')
+      } finally {
+        rmSync(dir, { recursive: true, force: true })
+      }
+    })
+
+    it('reads comma-separated multi-path output (the bug that caused the rework death spiral)', () => {
+      const dir = mkdtempSync(join(tmpdir(), 'af-readoutput-'))
+      try {
+        const a = join(dir, 'a.md')
+        const b = join(dir, 'b.md')
+        writeFileSync(a, 'Alpha content')
+        writeFileSync(b, 'Beta content')
+        const content = taskRepo.readTaskOutput({ output: `${a}, ${b}` })
+        assert.ok(content, 'should return non-null')
+        assert.match(content, /Alpha content/)
+        assert.match(content, /Beta content/)
+        assert.match(content, /---/, 'should separate files with divider')
+      } finally {
+        rmSync(dir, { recursive: true, force: true })
+      }
+    })
+
+    it('accepts array of paths', () => {
+      const dir = mkdtempSync(join(tmpdir(), 'af-readoutput-'))
+      try {
+        const a = join(dir, 'a.md')
+        writeFileSync(a, 'content A')
+        const content = taskRepo.readTaskOutput({ output: [a] })
+        assert.match(content, /content A/)
+      } finally {
+        rmSync(dir, { recursive: true, force: true })
+      }
+    })
+
+    it('skips non-existent paths but returns content from existing ones', () => {
+      const dir = mkdtempSync(join(tmpdir(), 'af-readoutput-'))
+      try {
+        const a = join(dir, 'a.md')
+        writeFileSync(a, 'real content')
+        const content = taskRepo.readTaskOutput({ output: `${a}, ${dir}/missing.md` })
+        assert.match(content, /real content/)
+      } finally {
+        rmSync(dir, { recursive: true, force: true })
+      }
+    })
+
+    it('returns null when all paths are missing', () => {
+      assert.equal(
+        taskRepo.readTaskOutput({ output: '/nope/a.md, /nope/b.md' }),
+        null
+      )
+    })
+
+    it('treats non-path strings as inline content (legacy)', () => {
+      const content = taskRepo.readTaskOutput({ output: 'PM Review 通过: commit abc — tests OK' })
+      assert.equal(content, 'PM Review 通过: commit abc — tests OK')
+    })
+  })
 })
diff --git a/tests/core/task/auto-transition.test.cjs b/tests/core/task/auto-transition.test.cjs
@@ -227,6 +227,36 @@ describe('AutoTransition', () => {
       assert.equal(result[1].summary, '写第二章')
       assert.equal(result[1].projectId, undefined)
     })
+
+    it('filters chief prose that signals "no assignment" (regression: fake task pollution)', () => {
+      // Real chief responses observed in production death-spiral cycles.
+      // Each of these previously produced a zombie task named after the prose.
+      const cases = [
+        '[任务分配]\n- apple-tester: 本轮不新增分配，继续执行先收口、后扩张',
+        '[任务分配]\n- apple-release: 当前无新增紧急待办且历史无响应风险仍在，维持低并发不分配新任务',
+        '[任务分配]\n- ios-developer: 无需分配（遵循先完成再开始）',
+        '[任务分配]\n- apple-designer: 无需分配（🔵 工作中，且已有 2 个进行中任务，严格不加并发）',
+        '[任务分配]\n- apple-tester: 继续收口 hot-topics 需求分析链路中的待确认项，引用现有任务，不新建',
+      ]
+      for (const text of cases) {
+        const result = parseTaskAssignments(text)
+        assert.equal(result.length, 0, `should skip prose: ${text.split('\n')[1]}`)
+      }
+    })
+
+    it('still accepts real assignments with action verbs', () => {
+      const cases = [
+        ['[任务分配]\n- apple-designer: 分配需求文档体验补强包 [project: apple-dev/hot-topics]', '分配需求文档体验补强包'],
+        ['[任务分配]\n- ios-developer: 实现登录页面', '实现登录页面'],
+        ['[任务分配]\n- apple-tester: 编写 CI 测试用例', '编写 CI 测试用例'],
+        ['[任务分配]\n- apple-release: review 发布检查清单', 'review 发布检查清单'],
+      ]
+      for (const [text, expectedSummary] of cases) {
+        const result = parseTaskAssignments(text)
+        assert.equal(result.length, 1, `should parse: ${text}`)
+        assert.equal(result[0].summary, expectedSummary)
+      }
+    })
   })
 
   describe('parseTaskRecoveries', () => {
diff --git a/tests/core/task/quality-orchestrator.test.cjs b/tests/core/task/quality-orchestrator.test.cjs
