review

Multi-Agent Review v3.0.0

Multi-perspective review → Issue classification → Prioritization → Fix suggestions

Usage

/multi-review [實作路徑]
/multi-review .claude/memory/implement/user-auth/

Flags: --from-implement ID | --focus security|quality|docs

Default 4 perspectives

ID	Name	Model	Focus
code-quality	Code quality	haiku	Naming, structure, readability
test-coverage	Test coverage	haiku	Coverage, test quality
documentation	Documentation check	haiku	Comments, README, API documentation
integration	Integration analysis	sonnet	Integration issues, dependency conflicts

→ Model routing configuration:shared/config/model-routing.yaml

Execution flow

Phase 0: 載入實作 → 收集變更檔案、測試結果
    ↓
Phase 1: MAP（並行審查）
    ┌──────────┬──────────┬──────────┬──────────┐
    │程式碼品質│測試覆蓋  │文檔檢查  │整合分析  │
    └──────────┴──────────┴──────────┴──────────┘

    ⚠️ **並行執行關鍵**：
       在單一訊息中發送 4 個 Task 工具呼叫：
       - Task({description: "程式碼品質審查", ...})
       - Task({description: "測試覆蓋審查", ...})
       - Task({description: "文檔檢查", ...})
       - Task({description: "整合分析", ...})
       這樣才能真正並行執行！
    ↓
Phase 2: REDUCE（問題匯總）
    ├── 問題分類（BLOCKER/HIGH/MEDIUM/LOW）
    ├── 優先排序
    └── 修復建議
    ↓
Phase 3: 品質評分 → 各維度評分
    ↓
Phase 4: 品質閘門 → 通過/回退決策

Issue classification

Level	Definition	Handling
BLOCKER	Must be fixed to continue	Roll back to IMPLEMENT
HIGH	Important issue, should be fixed	Fix within time limit
MEDIUM	Suggested improvement	Record and track
LOW	Style suggestion	Optional

Quality scoring

Three-dimension scoring:

• Completeness (30%): Complete functionality, no omissions
• Consistency (30%): Consistent style, no contradictions
• Actionability (40%): Specific and executable suggestions

→ Scoring configuration:shared/quality/scoring.yaml

CP4: Task Commit

After the quality gate check is complete,CP4 Task Commit must be executed。

Phase 4: 品質閘門 → 通過/回退決策
    ↓
CP4: Task Commit
    ├── git add .claude/memory/review/{implement-id}/
    └── git commit -m "docs(review): complete {feature} code review"

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

Quality gate

Pass conditions (REVIEW phase):

• ✅ No BLOCKER issues
• ✅ HIGH issues ≤ 2
• ✅ Quality score ≥ 75

Failure handling:

• BLOCKER exists → Roll back to IMPLEMENT
• HIGH > 2 → Require fix and re-review

→ Gate configuration:shared/quality/gates.yaml

Output structure

.claude/memory/review/[implement-id]/
├── meta.yaml           # 元數據
├── perspectives/       # 完整視角報告（MAP 產出，保留）
│   ├── code-quality.md
│   ├── test-coverage.md
│   ├── documentation.md
│   └── integration.md
├── summaries/          # 結構化摘要（REDUCE 產出，供快速查閱）
│   ├── code-quality.yaml
│   ├── test-coverage.yaml
│   ├── documentation.yaml
│   └── integration.yaml
├── issues.yaml         # 問題清單
├── scoring.yaml        # 評分詳情
└── review-summary.md   # 審查摘要（主輸出）

⚠️ perspectives/ stores the complete report, summaries/ stores the structured summary, both must be retained.

Agent capability limitations

Review Agent should not open Task：

Allowed operations	Description
✅ Read	Read code
✅ Glob/Grep	Search files and content
✅ Bash	Execute commands
✅ Write	Write reports
❌ Task	Open sub-Agent

Action log

After each tool call completes, record to .claude/workflow/{workflow-id}/logs/actions.jsonl。

Recording timing：

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

Key actions (REVIEW phase)：

Action	Recording focus
Read (read code)	file_path、output_size
Task (start review Agent)	subagent_type、prompt (truncated)、agent_id
Grep (search issue patterns)	pattern、match_count
Write (write review report)	file_path、content_size

Troubleshoot issues：

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

# 查看特定審查視角的行動
jq 'select(.agent_id == "agent_code-quality")' actions.jsonl

→ Log specification:shared/communication/execution-logs.md

Shared modules

Module	Purpose
coordination/map-phase.md	Parallel coordination
coordination/reduce-phase.md	Aggregation and integration, large file handling
quality/scoring.yaml	Quality scoring
quality/gates.yaml	Quality gate
quality/rollback-strategy.yaml	Rollback strategy
perspectives/expertise-frameworks/	Professional Framework

Workflow Position

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

• Input: Implemented code, from implement skill
• Output: Review report, for verify skill verification or rollback implement