implement

Multi-Agent Implement v3.0.0

TDD Gate → Parallel Implementation → Real-time Review → Quality Guardian

Usage

/multi-implement [任務路徑]
/multi-implement .claude/memory/tasks/user-auth/tasks.yaml

Flags: --from-tasks ID | --skip-tdd-check | --parallel N

Role Configuration

ID	Name	Model	Focus
main_agent	Main Implementer	sonnet	Feature Implementation
tdd-enforcer	TDD Guardian	haiku	Test-First Check
security-auditor	Security Auditor	sonnet	Security Vulnerability Detection
maintainer	Maintainability Review	haiku	Code Quality

→ Model Routing Configuration:shared/config/model-routing.yaml

Execution Flow

Phase 0: 載入任務 → 解析 tasks.yaml、排序 waves
    ↓
For each task in waves:
    ↓
    Phase 1: TDD 閘門（PRE-TASK）
    ├── 測試檔案存在？
    ├── 測試可執行？
    └── ❌ 失敗 → BLOCKER，無法繼續
    ↓
    Phase 2: 並行實作 + 審查
    ┌──────────┬──────────┬──────────┐
    │ 主實作者 │安全審計員│可維護性  │
    └──────────┴──────────┴──────────┘
    ↓
    Phase 3: TDD 閘門（POST-TASK）
    ├── 測試通過？
    └── 覆蓋率達標？
    ↓
    Phase 4: 自我審查 + 提交
    ↓
End for
    ↓
Phase 5: 品質閘門 → 整體驗證

TDD Mandatory Gate

Pre-Task Gate（BLOCKER）：

• Test files must exist
• Tests must be executable (even if failing)

Post-Task Gate（BLOCKER）：

• Tests must pass
• Coverage ≥ 80%

Validation Script：shared/tools/tdd-validator.sh

→ Configuration:shared/quality/tdd-enforcement.yaml

Security Audit

Use OWASP Top 10 and CWE Top 25 frameworks:

• SQL Injection
• XSS
• Authentication/Authorization Issues
• Sensitive Data Handling

→ Framework:shared/perspectives/expertise-frameworks/security.yaml

CP4: Task Commit

After each task is completedCP4 Task Commit (incremental commit) must be executed.

Phase 4: 自我審查 + 提交
    ↓
CP4: Task Commit（每個 task）
    ├── git add {changed_files}
    ├── git add .claude/memory/implement/{id}/task-results/{task-id}.yaml
    └── git commit -m "feat(implement): complete {task-id} - {description}"

Note: The implement skill's CP4 is incrementally triggered, committing after each task is completed to ensure progress is not lost.

→ Protocol:shared/git/commit-protocol.md

Quality Gate

Pass Conditions (IMPLEMENT phase):

• ✅ All tasks completed
• ✅ Tests pass
• ✅ No BLOCKER issues
• ✅ Quality score ≥ 80

→ Gate Configuration:shared/quality/gates.yaml

Output Structure

.claude/memory/implement/[tasks-id]/
├── meta.yaml               # 元數據
├── perspectives/           # 角色分析報告（並行審查產出，保留）
│   ├── security-auditor.md     # 安全審計完整報告
│   └── maintainer.md           # 可維護性完整報告
├── summaries/              # 結構化摘要（供快速查閱）
│   ├── security-auditor.yaml
│   └── maintainer.yaml
├── task-results/           # 每個任務的結果
│   ├── T-F-001.yaml
│   └── T-F-002.yaml
├── security-report.md      # 安全審計報告（匯總）
├── coverage-report.md      # 覆蓋率報告
└── summary.md              # 實作摘要（主輸出）

⚠️ perspectives/ stores complete role analyses, summaries/ stores structured summaries.

Agent Capability Limits

Review Agents should not open Tasks：

Allowed Operations	Description
✅ Read	Read code
✅ Glob/Grep	Search files and content
✅ Bash	Run tests
✅ Write	Write reports
❌ Task	Spawn sub-agents

Action Log

After each tool call is completed, log to .claude/workflow/{workflow-id}/logs/actions.jsonl。

Logging Timing：

• Success: log tool、input、output_preview、duration_ms、status: success
• Failure: log tool、input、error、stderr(if any),status: failed

Key Actions (IMPLEMENT phase)：

Action	Logging Focus
Read (read existing code)	file_path、output_size
Edit (modify code)	file_path、old_string (truncated)、new_string (truncated)
Write (create new file)	file_path、content_size
Bash (run tests)	command、exit_code、stdout、stderr
Task (launch review Agent)	subagent_type、prompt (truncated)、agent_id

Troubleshooting：

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

# 查看測試失敗的詳情
jq 'select(.tool == "Bash" and .input.command | contains("test"))' actions.jsonl | jq '{command: .input.command, exit_code: .exit_code, stderr: .stderr}'

# 查看 Edit 操作失敗（找不到 old_string）
jq 'select(.tool == "Edit" and .status == "failed")' actions.jsonl

→ Logging Specification:shared/communication/execution-logs.md

Shared Modules

Module	Purpose
coordination/map-phase.md	Parallel Coordination
quality/tdd-enforcement.yaml	TDD Enforcement
quality/gates.yaml	Quality Gate
perspectives/expertise-frameworks/security.yaml	Security Framework
tools/tdd-validator.sh	TDD Validator

Workflow Position

RESEARCH → PLAN → TASKS → IMPLEMENT → REVIEW → VERIFY
                              ↑
                           你在這裡

• Input：tasks.yaml from tasks skill
• Output: implemented code, for review skill review