miles990/multi-agent-workflow · Archived

tasks

多 Agent 任務分解框架 - 將計劃轉化為可執行的 DAG 任務?

First seen May 12, 2026

Installation

$ npx skills add miles990/multi-agent-workflow --skill tasks

Stronger alternatives

This repository is archived — consider an actively maintained alternative.

Similar popular skills

Related neighbors and high-traction skills in the same topics — useful to compare before installing.

Also in this package

Other skills from miles990/multi-agent-workflow · top by installs.

npx skills add miles990/multi-agent-workflow

Browse all from miles990/multi-agent-workflow

More details

Agent compatibility

Declared targets from SKILL.md / docs. Unmarked agents are not listed — the skill may still install via the CLI.

Claude Code Declared
Cursor Not declared
Codex Not declared
GitHub Copilot Not declared
Windsurf Not declared
Gemini CLI Not declared
Cline Not declared
OpenCode Not declared

Repository health

Stars 4
License LICENSE
Default branch main
Open issues 0
Status Archived

Skill metadata

Parsed from SKILL.md frontmatter.

Version3.2.0
Allowed toolsRead, Grep, Glob, Write, Bash, TaskCreate, TaskUpdate, TaskList, TaskGet
Declared agents claude-code

Package contents

Files included with this skill beyond the listing page.

  • skill md SKILL.md 7,604 B
  • docs SUMMARY.md 91 B

History

  1. First seen on skills.sh
  2. First recorded snapshot · 1 installs

SKILL.md

Multi-Agent Tasks v3.2.0

計劃分解 → DAG 驗證 → TDD 對應 → 可執行任務清單

Claude Code Tasks 整合

v3.2.0 支援原生 Tasks API,實現跨 session 任務協作

本 Skill 可使用 Claude Code 原生 Tasks 系統:

# 建立任務(自動生成 taskId)
TaskCreate({ subject: "實作登入功能", description: "..." })

# 設定依賴關係
TaskUpdate({ taskId: "2", addBlockedBy: ["1"] })

# 跨 session 共享
export CLAUDE_CODE_TASK_LIST_ID=my-workflow
API 用途
TaskCreate 建立任務
TaskUpdate 更新狀態、設定依賴(addBlockedBy)
TaskList 列出所有任務
TaskGet 取得任務詳情

優勢:

  • 跨 session 持久化
  • 多 Agent 協作同一任務清單
  • 自動依賴解除(完成時自動 unblock 下游任務)

使用方式

/multi-tasks [計劃路徑]
/multi-tasks .claude/memory/plans/user-auth/implementation-plan.md

Flags: --from-plan ID | --validate-only | --no-tdd-check

預設 4 視角

ID 名稱 模型 聚焦
dependency-analyst 依賴分析師 sonnet 任務依賴、執行順序
task-decomposer 任務分解師 haiku 粒度切分、並行識別
test-planner 測試規劃師 haiku TDD 對應、測試策略
risk-preventor 風險預防師 haiku 風險任務、預防措施

→ 模型路由配置:[shared/config/model-routing.yaml](_shared/config/model-routing.yaml)

執行流程

Phase 0: 載入計劃 → 解析 implementation-plan.md
    ↓
Phase 1: MAP(並行分解)
    ┌──────────┬──────────┬──────────┬──────────┐
    │依賴分析師│任務分解師│測試規劃師│風險預防師│
    └──────────┴──────────┴──────────┴──────────┘

    ⚠️ **並行執行關鍵**:
       在單一訊息中發送 4 個 Task 工具呼叫:
       - Task({description: "依賴分析", ...})
       - Task({description: "任務分解", ...})
       - Task({description: "測試規劃", ...})
       - Task({description: "風險預防", ...})
       這樣才能真正並行執行!
    ↓
Phase 2: REDUCE(合併 + 排序)
    ↓
Phase 3: TDD 驗證 → 確保每個 T-F-* 有對應 TEST-*
    ↓
Phase 4: DAG 驗證 → 無循環、無孤立任務
    ↓
Phase 5: 依賴偵測 → 隱含依賴分析
    ↓
Phase 6: 品質閘門 → 輸出 tasks.yaml

TDD 強制執行

必要條件(BLOCKER):

  • 所有 T-F- 任務必須有對應的 TEST- 任務
  • TEST- 任務必須 block 對應的 T-F- 任務

驗證腳本:shared/tools/tdd-validator.sh

→ 配置:[shared/quality/tdd-enforcement.yaml](_shared/quality/tdd-enforcement.yaml)

依賴偵測

四層依賴分析:

  1. L1_explicit: 顯式聲明的依賴
  2. L2fileoverlap: 修改相同檔案的任務
  3. L3_semantic: 語意資料流依賴
  4. L4_environment: 運行時環境依賴

驗證腳本:shared/tools/dag-validator.py

→ 配置:[shared/tasks/dependency-detection.yaml](_shared/tasks/dependency-detection.yaml)

CP4: Task Commit

品質閘門通過後,必須執行 CP4 Task Commit。

Phase 6: 品質閘門 → 輸出 tasks.yaml
    ↓
CP4: Task Commit
    ├── git add .claude/memory/tasks/{plan-id}/
    └── git commit -m "feat(tasks): decompose {plan} into executable tasks"

→ 協議:[shared/git/commit-protocol.md](_shared/git/commit-protocol.md)

品質閘門

通過條件(TASKS 階段):

  • ✅ DAG 驗證通過(無循環)
  • ✅ TDD 對應完整
  • ✅ 所有任務有估算
  • ✅ 品質分數 ≥ 80

→ 閘門配置:[shared/quality/gates.yaml](_shared/quality/gates.yaml)

輸出結構

.claude/memory/tasks/[plan-id]/
├── meta.yaml               # 元數據
├── perspectives/           # 完整視角分析(MAP 產出,保留)
│   ├── dependency-analyst.md
│   ├── task-decomposer.md
│   ├── test-planner.md
│   └── risk-preventor.md
├── summaries/              # 結構化摘要(REDUCE 產出,供快速查閱)
│   ├── dependency-analyst.yaml
│   ├── task-decomposer.yaml
│   ├── test-planner.yaml
│   └── risk-preventor.yaml
└── tasks.yaml              # 任務清單(主輸出)

⚠️ perspectives/ 保存完整分析,summaries/ 保存結構化摘要,tasks.yaml 是最終可執行任務清單。

Agent 能力限制

視角 Agent 不應該開啟 Task:

允許的操作 說明
✅ Read 讀取檔案
✅ Glob/Grep 搜尋檔案和內容
✅ Explore agent 輕量級探索
✅ Bash 執行命令
✅ Write 寫入報告
❌ Task 開子 Agent

tasks.yaml 格式

version: "1.0"
plan_ref: "plans/user-auth"
waves:
  - id: wave-1
    tasks:
      - id: TEST-001
        type: test
        description: "登入功能測試"
        estimate: 15
      - id: T-F-001
        type: feature
        description: "實作登入功能"
        depends_on: [TEST-001]
        estimate: 30
        files: [src/auth/login.ts]

行動日誌

每個工具調用完成後,記錄到 .claude/workflow/{workflow-id}/logs/actions.jsonl。

記錄時機:

  • 成功:記錄 tool、input、outputpreview、durationms、status: success
  • 失敗:記錄 tool、input、error、stderr(如有)、status: failed

關鍵行動(TASKS 階段):

行動 記錄重點
Read(讀取計劃) filepath、outputsize
Task(啟動 Agent) subagenttype、prompt (truncated)、agentid
Bash(執行 DAG 驗證) command、exit_code、stderr
Write(寫入 tasks.yaml) filepath、contentsize

排查問題:

# 查看 TASKS 階段所有失敗行動
jq 'select(.stage == "TASKS" and .status == "failed")' actions.jsonl

# 查看 DAG 驗證失敗的詳情
jq 'select(.tool == "Bash" and .status == "failed")' actions.jsonl | jq '{command: .input.command, error: .error, stderr: .stderr}'

→ 日誌規範:[shared/communication/execution-logs.md](_shared/communication/execution-logs.md)

共用模組

模組 用途
[coordination/map-phase.md](_shared/coordination/map-phase.md) 並行協調
[coordination/reduce-phase.md](_shared/coordination/reduce-phase.md) 匯總整合、大檔案處理
[quality/tdd-enforcement.yaml](_shared/quality/tdd-enforcement.yaml) TDD 強制
[tasks/dependency-detection.yaml](_shared/tasks/dependency-detection.yaml) 依賴偵測
[quality/gates.yaml](_shared/quality/gates.yaml) 品質閘門
[tools/dag-validator.py](_shared/tools/dag-validator.py) DAG 驗證器

工作流位置

RESEARCH → PLAN → TASKS → IMPLEMENT → REVIEW → VERIFY
                    ↑
                 你在這裡
  • 輸入:implementation-plan.md 來自 plan skill
  • 輸出:tasks.yaml 供 implement skill 使用