review

Review — Multi-Agent Code Review

Three independent agents review code from different perspectives in parallel. Independence is intentional — agents don't share context, which reduces groupthink.

When to Apply

• After implementation, when reviewing code
• Checking code quality or security
• Verifying implementation matches the plan
• User invokes /review

Process

Step 1: Identify Review Scope

1. Determine diff range (use the first that applies):

Condition	Command	Rationale
User specified files or commit range	Use what the user said	Explicit intent
On a feature branch (not main/master)	git diff main...HEAD	Full branch changes
Uncommitted changes exist	git diff HEAD	Working tree changes
None of the above	Ask the user what to review	Avoid guessing

2. Gather context documents (best-effort — absence is normal):

Document	How to find	If missing
Implementation plan	docs/IMPL_PLAN_*.md matching changed modules	Use git log --oneline of the diff range to infer intent from commit messages
Technical spec	docs/SPEC.md	Use CLAUDE.md or project root config files (pyproject.toml, package.json, etc.) for conventions

Not every project uses the smart pipeline (/init → /spec → /plan → /impl). The review must work regardless — documents enrich the review but are not prerequisites.

3. Collect review inputs:

• CHANGED_FILES: list of modified/created files
• DIFF_CONTENT: output of the git diff command (this is what agents review — not the full files)
• IMPL_PLAN: implementation plan content, or commit messages summary if no plan exists
• SPEC_CONTEXT: technical spec sections, or project conventions from CLAUDE.md/config files
• PROJECT_TYPE: inferred project type (e.g., web API, CLI tool, batch pipeline, library) — agents use this to focus relevant checks

Step 2: Launch Review Agents

Environment Detection:

• If you have access to the Agent tool → use Option A (parallel)
• If the Agent tool is NOT available → use Option B (sequential)

Option A — Parallel (Agent tool available):

Run all three agents simultaneously using the Agent tool.

Agent A — Alignment Reviewer (계획-구현 정합성) Read agents/alignment-reviewer.md. Fill in {{DIFF_CONTENT}}, {{CHANGED_FILES}}, and {{IMPL_PLAN}}.

Agent B — Quality Reviewer (코드 품질 & 보안) Read agents/quality-reviewer.md. Fill in {{DIFF_CONTENT}}, {{CHANGED_FILES}}, {{SPEC_CONTEXT}}, and {{PROJECT_TYPE}}.

Agent C — Test Reviewer (테스트 & 성능) Read agents/test-reviewer.md. Fill in {{DIFF_CONTENT}}, {{CHANGED_FILES}}, and {{IMPL_PLAN}}.

Option B — Sequential (Agent tool NOT available):

Perform each review inline, one at a time:

1. Read agents/alignment-reviewer.md. Perform the Alignment Review inline. Write the result.
2. Read agents/quality-reviewer.md. Perform the Quality Review inline. Write the result.
3. Read agents/test-reviewer.md. Perform the Test Review inline. Write the result.

Step 3: Assemble Review Report

Combine all three agent outputs into a single report. Output to terminal only — do not save as a file (reviews are ephemeral, tied to specific code state).

Format:

## 코드 리뷰 결과

> 리뷰 대상: {변경 파일 요약}
> 참조 계획: {IMPL_PLAN 파일명 또는 "commit messages 기반"}
> 리뷰 일시: {YYYY-MM-DD}

---

### 요약

- **전체 평가**: {Pass / Pass with comments / Changes requested}
- **발견된 이슈**: 치명적 {N}개, 주의 {N}개, 제안 {N}개

---

### 계획-구현 정합성

{Agent A 결과}

---

### 코드 품질 & 보안

{Agent B 결과}

---

### 테스트 & 성능

{Agent C 결과}

---

### 문서 동기화

구현 과정에서 계획과 달라진 부분이 있다면:
- 어떤 부분이 달라졌는지
- 어떤 문서(PRD, SPEC, FEATURE_PLAN, IMPL_PLAN)를 업데이트해야 하는지
없으면 "문서 업데이트 불필요"
(계획 문서가 없는 프로젝트의 경우 이 섹션 생략)

---

### 추천 액션

1. **[치명적]** {즉시 수정 — file:line 포함}
2. **[주의]** {권장 수정 — file:line 포함}
3. **[제안]** {개선 — file:line 포함}

Overall rating criteria:

• Pass: no issues
• Pass with comments: no critical issues, only warnings/suggestions
• Changes requested: 1+ critical issues

Then ask: "리뷰 결과를 확인해 주세요. 수정이 필요한 항목이 있으면 말씀해 주세요."

Quality Rules

• Every issue must include specific location (file:line) and actionable suggestion
• Prioritize by severity — reviewers' time is limited
• Document sync section is mandatory when plan documents exist — catching plan drift early prevents documentation rot
• Avoid vague feedback like "improve this" — suggest specific changes like "change X to Y"
• Agents should review the diff to focus on what changed, reading full files only when surrounding context is needed to understand the change