@su-record/vibe 2.9.21 → 2.9.23

package/commands/vibe.review.md

@@ -1,624 +1,624 @@
----
-description: Multi-agent parallel code review with priority-based findings
-argument-hint: "PR number, branch name, or file path"
----
-
-# /vibe.review
-
-**Parallel Agent Code Review** - 13+ specialists review simultaneously
-
-## Usage
-
-```
-/vibe.review                         # Review current branch
-/vibe.review PR#123                  # Review specific PR
-/vibe.review feature/login           # Review specific branch
-/vibe.review src/api/                # Review specific path
-/vibe.review --race                  # Multi-LLM race mode (GPT + Gemini)
-/vibe.review --race security         # Race mode for specific review type
-```
-
-> **⏱️ Timer**: Call `getCurrentTime` tool at the START. Record the result as `{start_time}`.
-
-## Codex Plugin Integration
-
-> **Codex 플러그인 감지**: 워크플로우 시작 시 아래 명령으로 자동 감지.
->
-> ```bash
-> CODEX_AVAILABLE=$(node "{{VIBE_PATH}}/hooks/scripts/codex-detect.js" 2>/dev/null || echo "unavailable")
-> ```
->
-> `available`이면 `/codex:review`, `/codex:rescue` 자동 호출. `unavailable`이면 기존 GPT+Gemini Race 모드로 동작.
-
-## Race Mode (v2.6.9)
-
-**Multi-LLM competitive review** - Same review task runs on GPT + Gemini in parallel, results are cross-validated.
-
-### How It Works
-
-```
-/vibe.review --race
-
-security-review:
-├─ GPT Codex  → [SQL injection, XSS]
-└─ Gemini     → [SQL injection, CSRF]
-         ↓
-   Cross-validation:
-   - SQL injection (2/2) → 🔴 P1 (100% confidence)
-   - XSS (1/2) → 🟡 P2 (50% confidence)
-   - CSRF (1/2) → 🟡 P2 (50% confidence)
-```
-
-### Confidence-Based Priority
-
-| Confidence | Priority | Meaning |
-|------------|----------|---------|
-| 100% (2/2) | P1 | Both models agree - high confidence |
-| 50% (1/2) | P2 | One model found - needs verification |
-
-### Race Mode Options
-
-```
-/vibe.review --race                  # All review types
-/vibe.review --race security         # Security only
-/vibe.review --race performance      # Performance only
-/vibe.review --race architecture     # Architecture only
-```
-
-### Race Mode Output
-
-```
-## SECURITY Review (Race Mode)
-
-**Duration**: 3420ms
-**Models**: GPT Codex, Gemini
-
-### Model Results
-
-| Model | Issues Found | Duration | Status |
-|-------|--------------|----------|--------|
-| gpt | 3 | 1823ms | OK |
-| gemini | 2 | 2156ms | OK |
-
-### Cross-Validated Issues
-
-**Summary**: 3 issues (P1: 1, P2: 2)
-**Consensus Rate**: 67%
-
-#### 🔴 P1 - SQL Injection in user query
-
-- **Confidence**: 100% (gpt, gemini)
-- **Severity**: critical
-- **Location**: `src/api/users.ts:42`
-- **Suggestion**: Use parameterized queries
-
-#### 🟡 P2 - XSS vulnerability in render
-
-- **Confidence**: 50% (gpt)
-- **Severity**: high
-- **Location**: `src/components/Comment.tsx:15`
-```
-
-### Codex Review (Codex 플러그인 활성화 시)
-
-Race Mode에서 GPT+Gemini와 **동시에** Codex review 실행하여 3중 교차 검증:
-
-```
-/codex:review
-```
-
-교차 검증 테이블:
-
-```markdown
-| Issue | GPT | Gemini | Codex | Confidence |
-|-------|-----|--------|-------|------------|
-| {이슈} | ✅/❌ | ✅/❌ | ✅/❌ | {%} |
-```
-
-- 3개 모델 중 2개 이상 동의 → **High Confidence** (P1 자동 수정)
-- Codex만 발견 → **Medium Confidence** (P2 수동 검토)
-
-### When to Use Race Mode
-
-| Scenario | Recommended |
-|----------|-------------|
-| Critical security review | ✅ `--race security` |
-| Pre-production audit | ✅ `--race` |
-| Quick iteration | ❌ Standard review |
-| API cost concerns | ❌ Standard review |
-
-### Tool Invocation (Race Mode - GPT + Gemini in parallel via Bash)
-
-**🚨 Use --input file to avoid CLI argument length limits and Windows pipe issues.**
-
-1. Save code to review into `[SCRATCHPAD]/review-code.txt` (using Write tool)
-2. Write JSON input file `[SCRATCHPAD]/review-input.json` (using Write tool):
-   - `{"prompt": "Review this code for [REVIEW_TYPE]. Return JSON: {issues: [{id, title, description, severity, suggestion}]}. Code: [CODE_CONTENT]"}`
-   - Where `[CODE_CONTENT]` is the code text (properly JSON-escaped inside the prompt string)
-3. Script path: `[LLM_SCRIPT]` = `{{VIBE_PATH}}/hooks/scripts/llm-orchestrate.js`
-4. Run GPT + Gemini in PARALLEL (two Bash tool calls at once):
-
-```bash
-# GPT review (Bash tool call 1)
-node "[LLM_SCRIPT]" gpt orchestrate-json --input "[SCRATCHPAD]/review-input.json"
-```
-
-```bash
-# Gemini review (Bash tool call 2 - run in parallel)
-node "[LLM_SCRIPT]" gemini orchestrate-json --input "[SCRATCHPAD]/review-input.json"
-```
-
-## File Reading Policy (Mandatory)
-
-- **소스코드 파일**: 리뷰 대상 파일은 반드시 `Read` 도구로 전체 파일을 읽은 후 리뷰할 것 (Grep으로 훑어보기 금지)
-- **Grep 사용 제한**: 파일 위치 탐색(어떤 파일에 있는지 찾기)에만 사용. 파일 내용 파악 및 리뷰에는 반드시 Read 사용
-- **에이전트 spawn 시**: 프롬프트에 "대상 파일을 Read 도구로 전체 읽은 후 분석하라"를 반드시 포함할 것
-- **부분 읽기 금지**: Grep 결과의 주변 몇 줄만 보고 판단하지 말 것. 전체 맥락을 파악해야 정확한 리뷰 가능
-
-## Priority System
-
-| Priority | Criteria | Action |
-|----------|----------|--------|
-| P1 | Security vulnerabilities, data loss, crashes | Block merge, fix immediately |
-| P2 | Performance issues, architecture violations, missing tests | Fix before merge |
-| P3 | Style, refactoring suggestions, documentation | Add to backlog |
-
-## Convergence Rules (Over-Diagnosis Prevention)
-
-> **Principle**: Reviews must converge. A review that always finds more issues is broken, not thorough.
-
-### Scope Limiting
-
-- **Review ONLY changed files** — based on `git diff --name-only`. Never scan the entire project
-- **If no git diff** (first review) — review only files in the target path
-
-### Severity Filtering by Round
-
-| Round | What to Report |
-|-------|---------------|
-| 1st review | P1 + P2 + P3 (all) |
-| 2nd review (same code) | P1 + P2 only (skip P3) |
-| 3rd+ review | P1 only (report only new P1s) |
-
-### Stop Conditions
-
-- **P1 = 0 means MERGE READY** — mergeable even with remaining P2/P3
-- **P1 = 0 after auto-fix means DONE** — record P2 auto-fix failures as TODO and stop
-- **Final P1 list unchanged after Review Debate → DONE** — no new findings = converged
-
-### Anti-Patterns (FORBIDDEN)
-
-- "All items must be verified" → Only P1 is mandatory, P2/P3 are best-effort
-- "Found one more issue" (repeated) → Only report P1s not mentioned in previous review
-- Forcing code changes for P3 issues → P3 goes to TODO files only, never force code changes
-- Infinite retries on auto-fix failure → max 1 retry then move to TODO
-
-## Process
-
-### Phase 1: Tech Stack Detection
-
-Detect project tech stack FIRST before launching reviewers:
-
-```
-Read package.json      -> TypeScript, React, Node.js
-Read pyproject.toml    -> Python, FastAPI, Django
-Read Gemfile           -> Ruby, Rails
-Read pubspec.yaml      -> Flutter, Dart
-Read go.mod            -> Go
-Read CLAUDE.md         -> Explicit tech stack declaration
-```
-
-### Phase 1.5: SPEC ↔ Code Alignment Check
-
-> When SPEC files exist, verify that code changes align with the SPEC
-
-```
-1. Search .claude/vibe/specs/ for related SPEC files (based on git diff filenames)
-2. Compare SPEC REQ-* list against functionality in changed code
-3. If functionality added that's not in SPEC → P2 finding: "Feature added without SPEC"
-4. If implementation differs from SPEC → P1 finding: "SPEC ↔ code mismatch"
-5. If no SPEC files exist → Skip (reviews work without SPEC too)
-```
-
-### Phase 2: Parallel Agent Review (STACK-AWARE) via Orchestrator
-
-**Execution via Orchestrator (12+ agents in parallel):**
-```bash
-node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/infra/orchestrator/index.js').then(o => o.review(['FILE_PATHS'], ['DETECTED_STACKS']).then(r => console.log(r.content[0].text)))"
-```
-
-**Example:**
-```bash
-# Review changed files with TypeScript + React stack
-node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/infra/orchestrator/index.js').then(o => o.review(['src/api/users.ts', 'src/components/Login.tsx'], ['TypeScript', 'React']).then(r => console.log(r.content[0].text)))"
-```
-
-**Core Reviewers (Always Run):**
-| Agent | Focus |
-|-------|-------|
-| security-reviewer | OWASP Top 10, vulnerabilities |
-| data-integrity-reviewer | Data validation, constraints |
-| performance-reviewer | N+1 queries, memory leaks |
-| architecture-reviewer | Layer violations, cycles |
-| complexity-reviewer | Cyclomatic complexity, length |
-| simplicity-reviewer | Over-abstraction, dead code |
-| git-history-reviewer | Churn files, risk patterns |
-| test-coverage-reviewer | Missing tests, edge cases |
-
-**Stack-Specific Reviewers (Conditional):**
-| Agent | Condition |
-|-------|-----------|
-| python-reviewer | .py files in diff |
-| typescript-reviewer | .ts/.tsx files OR tsconfig |
-| rails-reviewer | Gemfile has rails |
-| react-reviewer | package.json has react |
-
-### Phase 2.5: UI/UX Review Agents (Auto-triggered)
-
-> **활성화 조건**: 변경된 파일 중 UI 파일 존재 (`.tsx`, `.jsx`, `.vue`, `.svelte`, `.html`, `.css`, `.scss`)
-> **비활성화**: `.claude/vibe/config.json`에 `"uiUxAnalysis": false` 설정
-
-**기존 12+ 리뷰 에이전트와 병렬 실행:**
-
-| Agent | Role | Output |
-|-------|------|--------|
-| ⑥ ux-compliance-reviewer | UX 가이드라인 준수 검증 | P1/P2/P3 findings |
-| ⑦ ui-a11y-auditor | WCAG 2.1 AA 접근성 감사 | P1/P2/P3 findings |
-| ⑧ ui-antipattern-detector | UI 안티패턴 + 디자인 시스템 일관성 | P1/P2/P3 findings |
-
-**실행 방법 — 기존 Phase 2 에이전트와 병렬 실행:**
-
-```text
-# ⑥ UX 준수 검증 (Haiku)
-Task(subagent_type="ux-compliance-reviewer",
-  prompt="Review UI files for UX guideline compliance: {changed_ui_files}. Use core_ui_search against ux-guidelines and web-interface domains.")
-
-# ⑦ 접근성 감사 (Haiku)
-Task(subagent_type="ui-a11y-auditor",
-  prompt="Audit UI files for WCAG 2.1 AA compliance: {changed_ui_files}.")
-
-# ⑧ 안티패턴 검출 (Haiku)
-Task(subagent_type="ui-antipattern-detector",
-  prompt="Detect UI anti-patterns in: {changed_ui_files}. Check against MASTER.md if exists at .claude/vibe/design-system/{project}/MASTER.md.")
-```
-
-**findings 통합**: ⑥⑦⑧ findings를 기존 findings[]와 병합 → P1/P2/P3 통합 정렬
-
-**⑦ Critical finding 에스컬레이션**: ui-a11y-auditor의 P1 finding은 Review Debate Team(Phase 4.5)에 자동 포함
-
-### Phase 2.7: Boundary Mismatch Detection (Integration Coherence)
-
-> **활성화 조건**: 변경된 파일 중 API route + 프론트엔드 훅/컴포넌트가 함께 존재
-> 경계면 불일치는 개별 파일 리뷰로는 발견 불가 — **양쪽을 동시에 읽어야** 잡힘
-
-**검증 방법: "양쪽 동시 읽기"**
-
-반드시 **생산자와 소비자 코드를 동시에** Read하여 교차 비교한다.
-
-| 검증 영역 | 생산자 (왼쪽) | 소비자 (오른쪽) | 검증 내용 |
-|----------|-------------|---------------|----------|
-| API ↔ 훅 타입 | route의 Response.json() shape | hooks의 fetch\<T\> 타입 | shape 일치, 래핑 unwrap, case 변환 |
-| 라우팅 정합성 | src/app/ page 파일 경로 | href, router.push 값 | 경로 매칭, route group 처리, 동적 세그먼트 |
-| 상태 전이 | STATE_TRANSITIONS 맵 | .update({ status }) 코드 | 죽은 전이, 무단 전이, 중간→최종 누락 |
-| 데이터 흐름 | DB 스키마 필드명 | API 응답 → 프론트 타입 | 필드명 일치, optional 처리 일관성 |
-
-**실행 방식 — 변경 파일 기반 자동 판별:**
-
-```text
-1. git diff에서 API route 파일과 대응 프론트 파일을 짝으로 매칭
-2. 짝이 있는 경우 → 아래 4개 검증 실행
-3. 짝이 없는 경우 (API만 또는 프론트만 변경) → "대응 파일 미변경" 경고 후 스킵
-```
-
-**검증 체크리스트:**
-
-- [ ] API 응답 shape과 대응 훅의 제네릭 타입이 일치
-- [ ] 래핑된 응답(`{ items: [...] }`)은 훅에서 unwrap하는지 확인
-- [ ] snake_case ↔ camelCase 변환이 일관되게 적용
-- [ ] 모든 API 엔드포인트에 대응하는 프론트 훅이 존재하고 실제 호출됨
-- [ ] 코드 내 모든 href/router.push 값이 실제 page 파일 경로와 매칭
-- [ ] 정의된 모든 상태 전이가 코드에서 실행됨 (죽은 전이 없음)
-- [ ] 프론트에서 상태 기반 분기의 값이 실제 도달 가능한 상태
-- [ ] DB 필드명 → API 응답 필드명 → 프론트 타입 정의 간 매핑 일관
-
-**Findings 분류:**
-- 경계면 불일치 → **P1** (런타임 에러의 주요 원인)
-- 대응 파일 미존재 (API 있으나 훅 없음) → **P2**
-- case 변환 비일관성 → **P2**
-
-### Phase 3: Deep Analysis
-
-After agent results:
-
-1. **System Context**: Component interactions, data flow, external dependencies
-2. **Stakeholder Perspectives**: Developers, Ops, Security, Business
-3. **Edge Cases**: Race conditions, resource exhaustion, network failures
-4. **Multiple Angles**: Technical excellence, business value, risk management
-
-### Phase 4: Findings Synthesis
-
-```
-REVIEW FINDINGS
-
-P1 CRITICAL (Blocks Merge) - N issues
-1. [SECURITY] SQL Injection in user query
-   Location: src/api/users.py:42
-   Fix: Use parameterized queries
-
-P2 IMPORTANT (Should Fix) - N issues
-2. [PERF] N+1 query in user list
-3. [ARCH] Circular dependency detected
-
-P3 NICE-TO-HAVE (Enhancement) - N issues
-4. [STYLE] Consider extracting helper function
-```
-
-### Phase 4.5: Agent Teams — Review Debate
-
-> **팀 정의**: `agents/teams/review-debate-team.md` 참조 (Code Review 컨텍스트)
-> 설정: `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` + `teammateMode: in-process` (`~/.claude/settings.json` 전역 — postinstall 자동 설정)
-
-**토론 결과 예시:**
-
-```
-🤝 REVIEW DEBATE RESULTS
-
-Team Consensus (4 reviewers):
-
-✅ Validated P1 (unanimous):
-  1. [SECURITY] SQL Injection — 4/4 agree critical
-
-⬆️ Upgraded P2→P1 (debate result):
-  2. [PERF] Unbounded query — architecture-reviewer pointed out
-     cascading failure risk under load → team agreed P1
-
-⬇️ Downgraded P1→P2 (debate result):
-  3. [SECURITY] CSRF on read-only endpoint — simplicity-reviewer
-     noted endpoint has no side effects → team agreed P2
-
-❌ Removed (false positive):
-  4. [ARCH] "Circular dependency" — architecture-reviewer confirmed
-     this is intentional bi-directional reference, not a cycle
-
-🆕 New findings (team discussion):
-  5. [DATA] Race condition in concurrent updates — emerged from
-     security + performance discussion
-```
-
-### Phase 5: Auto-Fix (P1/P2)
-
-**Auto-fixable issues are resolved immediately:**
-
-```
-🔧 AUTO-FIX Starting...
-
-P1 Critical:
-  1. [SECURITY] SQL Injection → Fixed with parameterized query ✅
-  2. [DATA] Missing transaction rollback → Added try-finally ✅
-
-P2 Important:
-  3. [PERF] N+1 query → Added select_related ✅
-  4. [ARCH] Circular dependency → Separated dependencies ✅
-  5. [TEST] Missing edge case → Added test ✅
-
-🔍 Re-validating...
-  ✅ Build successful
-  ✅ Tests passed
-
-✅ 5 issues auto-fixed!
-```
-
-**Cases that cannot be auto-fixed:**
-- Requires large-scale architecture changes
-- Requires business logic decisions
-- Requires user confirmation
-
-→ Manual handling instructions in Phase 6
-
-### Auto-Fix 실패 시 Codex Rescue (Codex 플러그인 활성화 시)
-
-P1/P2 auto-fix **3회 실패** 시, Codex에 위임:
-
-```
-/codex:rescue "Fix {priority} issue: {issue-description}. File: {file-path}"
-```
-
-Codex 수정 완료 후 해당 리뷰 에이전트가 재검증.
-
-### Phase 6: Todo File Creation (Items Requiring Manual Handling)
-
-Save **remaining** findings to `.claude/vibe/todos/`:
-
-```
-{priority}-{category}-{short-desc}.md
-
-Examples:
-- P2-arch-large-refactor.md  (Cannot be auto-fixed)
-- P3-style-extract-helper.md (Backlog)
-```
-
-## Output
-
-```
-CODE REVIEW SUMMARY
-PR #123: Add user authentication
-
-Reviewers: 13 agents
-⏱️ Started: {start_time}
-⏱️ Completed: {getCurrentTime 결과}
-
-Score: 92/100 (Good) ← Score after auto-fix
-
-Issues Found:
-- P1 Critical: 2 → 0 (✅ Auto-fixed)
-- P2 Important: 5 → 1 (✅ 4 auto-fixed)
-- P3 Nice-to-have: 3 (Backlog)
-
-Auto-Fixed: 6 issues
-- [SECURITY] SQL Injection ✅
-- [DATA] Transaction rollback ✅
-- [PERF] N+1 query ✅
-- [ARCH] Circular dependency ✅
-- [PERF] Unnecessary loop ✅
-- [TEST] Missing edge case ✅
-
-Remaining (Manual handling required):
-- P2-arch-large-refactor.md (Architecture decision required)
-- P3-style-extract-helper.md (Backlog)
-- P3-docs-add-readme.md (Backlog)
-
-✅ MERGE READY (P1/P2 resolved)
-```
-
-### Phase 7: Guide to Fix Workflow (Manual Handling Items)
-
-**Choose workflow when handling remaining issues:**
-
-```
-## Fix Workflow
-
-Choose a workflow to fix the discovered issues:
-
-| Task Scale | Recommended Approach |
-|------------|---------------------|
-| Simple fix (1-2 files) | Plan Mode |
-| Complex fix (3+ files, validation needed) | /vibe.spec |
-
-1. `/vibe.spec "fix: issue-name"` - VIBE workflow (SPEC validation + re-review)
-2. Plan Mode - Quick fix (for simple tasks)
-
-Which approach would you like to proceed with?
-```
-
-- Wait for user's choice before proceeding
-- If user chooses VIBE → wait for `/vibe.spec` command
-- If user chooses Plan Mode → proceed with EnterPlanMode
-
-## Core Tools (Code Analysis)
-
-### Tool Invocation
-
-All tools are called via:
-
-```bash
-node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/tools/index.js').then(t => t.TOOL_NAME({...args}).then(r => console.log(r.content[0].text)))"
-```
-
-### Recommended Tools for Review
-
-| Tool | Purpose | When to Use |
-|------|---------|-------------|
-| `validateCodeQuality` | Code quality check | Overall code quality scan |
-| `analyzeComplexity` | Complexity metrics | Check function complexity |
-| `findSymbol` | Find definitions | Locate implementations |
-| `findReferences` | Find all usages | Track symbol usage |
-| `saveMemory` | Save findings | Store important review findings |
-
-### Example Tool Usage in Review
-
-**1. Validate code quality before review:**
-
-```bash
-node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/tools/index.js').then(t => t.validateCodeQuality({targetPath: 'src/', projectPath: process.cwd()}).then(r => console.log(r.content[0].text)))"
-```
-
-**2. Analyze complexity of changed files:**
-
-```bash
-node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/tools/index.js').then(t => t.analyzeComplexity({targetPath: 'src/api/users.ts', projectPath: process.cwd()}).then(r => console.log(r.content[0].text)))"
-```
-
-**3. Save critical finding for reference:**
-
-```bash
-node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/tools/index.js').then(t => t.saveMemory({key: 'review-pr123-critical', value: 'SQL injection in users.py:42', category: 'review', projectPath: process.cwd()}).then(r => console.log(r.content[0].text)))"
-```
-
----
-
-## Quality Gate (Mandatory)
-
-### Review Quality Checklist
-
-Before completing review, check P1-critical items. P2/P3 items are best-effort:
-
-| Category | Check Item | Weight |
-|----------|------------|--------|
-| **Security** | OWASP Top 10 vulnerabilities scanned | 20% |
-| **Security** | Authentication/authorization verified | 10% |
-| **Security** | Sensitive data exposure checked | 10% |
-| **Performance** | N+1 queries detected and flagged | 10% |
-| **Performance** | Memory leaks checked | 5% |
-| **Architecture** | Layer violations detected | 10% |
-| **Architecture** | Circular dependencies checked | 5% |
-| **Code Quality** | Complexity limits enforced | 10% |
-| **Code Quality** | Forbidden patterns detected | 10% |
-| **Testing** | Test coverage gaps identified | 5% |
-| **Documentation** | Public API documentation checked | 5% |
-
-### Review Score Calculation
-
-```
-Score = 100 - (P1 × 20) - (P2 × 5) - (P3 × 1)
-
-Grades:
-- 95-100: ✅ EXCELLENT - Merge ready
-- 90-94:  ⚠️ GOOD - Minor fixes required before merge
-- 80-89:  ⚠️ FAIR - Must fix P2 issues
-- 0-79:   ❌ POOR - Block merge, fix P1/P2
-```
-
-### Merge Decision Matrix
-
-| P1 Count | P2 Count | Decision |
-|----------|----------|----------|
-| 0 | 0-2 | ✅ MERGE READY |
-| 0 | 3+ | ⚠️ FIX P2 FIRST |
-| 1+ | Any | ❌ BLOCKED |
-
-### Auto-Fix Capability Matrix
-
-| Issue Type | Auto-Fixable | Method |
-|------------|--------------|--------|
-| SQL Injection | ✅ Yes | Parameterized query |
-| Missing transaction | ✅ Yes | Add try-finally |
-| N+1 query | ✅ Yes | Add eager loading |
-| Circular dependency | ⚠️ Partial | Suggest restructure |
-| Missing tests | ✅ Yes | Generate test skeleton |
-| Hardcoded secrets | ❌ No | Flag for manual review |
-| Architecture violation | ❌ No | Suggest refactoring plan |
-
-### Forbidden Patterns (P1 Critical)
-
-| Pattern | Risk Level | Detection Method |
-|---------|------------|------------------|
-| Hardcoded credentials | Critical | Regex + entropy scan |
-| SQL string concatenation | Critical | AST analysis |
-| `eval()` or `exec()` | Critical | AST analysis |
-| Disabled CSRF protection | Critical | Config scan |
-| Debug mode in production | Critical | Config scan |
-| Unvalidated redirects | High | URL pattern scan |
-
-### Review Output Requirements
-
-Every review MUST produce:
-
-1. **Summary Statistics**
-   - Total issues by priority (P1/P2/P3)
-   - Auto-fixed count
-   - Remaining manual fixes
-
-2. **Detailed Findings**
-   - File path and line number
-   - Issue description
-   - Recommended fix
-   - Auto-fix status (applied/pending/manual)
-
-3. **Quality Score**
-   - Numerical score (0-100)
-   - Grade (EXCELLENT/GOOD/FAIR/POOR)
-   - Merge recommendation
-
----
-
-ARGUMENTS: $ARGUMENTS
+---
+description: Multi-agent parallel code review with priority-based findings
+argument-hint: "PR number, branch name, or file path"
+---
+
+# /vibe.review
+
+**Parallel Agent Code Review** - 13+ specialists review simultaneously
+
+## Usage
+
+```
+/vibe.review                         # Review current branch
+/vibe.review PR#123                  # Review specific PR
+/vibe.review feature/login           # Review specific branch
+/vibe.review src/api/                # Review specific path
+/vibe.review --race                  # Multi-LLM race mode (GPT + Gemini)
+/vibe.review --race security         # Race mode for specific review type
+```
+
+> **⏱️ Timer**: Call `getCurrentTime` tool at the START. Record the result as `{start_time}`.
+
+## Codex Plugin Integration
+
+> **Codex 플러그인 감지**: 워크플로우 시작 시 아래 명령으로 자동 감지.
+>
+> ```bash
+> CODEX_AVAILABLE=$(node "{{VIBE_PATH}}/hooks/scripts/codex-detect.js" 2>/dev/null || echo "unavailable")
+> ```
+>
+> `available`이면 `/codex:review`, `/codex:rescue` 자동 호출. `unavailable`이면 기존 GPT+Gemini Race 모드로 동작.
+
+## Race Mode (v2.6.9)
+
+**Multi-LLM competitive review** - Same review task runs on GPT + Gemini in parallel, results are cross-validated.
+
+### How It Works
+
+```
+/vibe.review --race
+
+security-review:
+├─ GPT Codex  → [SQL injection, XSS]
+└─ Gemini     → [SQL injection, CSRF]
+         ↓
+   Cross-validation:
+   - SQL injection (2/2) → 🔴 P1 (100% confidence)
+   - XSS (1/2) → 🟡 P2 (50% confidence)
+   - CSRF (1/2) → 🟡 P2 (50% confidence)
+```
+
+### Confidence-Based Priority
+
+| Confidence | Priority | Meaning |
+|------------|----------|---------|
+| 100% (2/2) | P1 | Both models agree - high confidence |
+| 50% (1/2) | P2 | One model found - needs verification |
+
+### Race Mode Options
+
+```
+/vibe.review --race                  # All review types
+/vibe.review --race security         # Security only
+/vibe.review --race performance      # Performance only
+/vibe.review --race architecture     # Architecture only
+```
+
+### Race Mode Output
+
+```
+## SECURITY Review (Race Mode)
+
+**Duration**: 3420ms
+**Models**: GPT Codex, Gemini
+
+### Model Results
+
+| Model | Issues Found | Duration | Status |
+|-------|--------------|----------|--------|
+| gpt | 3 | 1823ms | OK |
+| gemini | 2 | 2156ms | OK |
+
+### Cross-Validated Issues
+
+**Summary**: 3 issues (P1: 1, P2: 2)
+**Consensus Rate**: 67%
+
+#### 🔴 P1 - SQL Injection in user query
+
+- **Confidence**: 100% (gpt, gemini)
+- **Severity**: critical
+- **Location**: `src/api/users.ts:42`
+- **Suggestion**: Use parameterized queries
+
+#### 🟡 P2 - XSS vulnerability in render
+
+- **Confidence**: 50% (gpt)
+- **Severity**: high
+- **Location**: `src/components/Comment.tsx:15`
+```
+
+### Codex Review (Codex 플러그인 활성화 시)
+
+Race Mode에서 GPT+Gemini와 **동시에** Codex review 실행하여 3중 교차 검증:
+
+```
+/codex:review
+```
+
+교차 검증 테이블:
+
+```markdown
+| Issue | GPT | Gemini | Codex | Confidence |
+|-------|-----|--------|-------|------------|
+| {이슈} | ✅/❌ | ✅/❌ | ✅/❌ | {%} |
+```
+
+- 3개 모델 중 2개 이상 동의 → **High Confidence** (P1 자동 수정)
+- Codex만 발견 → **Medium Confidence** (P2 수동 검토)
+
+### When to Use Race Mode
+
+| Scenario | Recommended |
+|----------|-------------|
+| Critical security review | ✅ `--race security` |
+| Pre-production audit | ✅ `--race` |
+| Quick iteration | ❌ Standard review |
+| API cost concerns | ❌ Standard review |
+
+### Tool Invocation (Race Mode - GPT + Gemini in parallel via Bash)
+
+**🚨 Use --input file to avoid CLI argument length limits and Windows pipe issues.**
+
+1. Save code to review into `[SCRATCHPAD]/review-code.txt` (using Write tool)
+2. Write JSON input file `[SCRATCHPAD]/review-input.json` (using Write tool):
+   - `{"prompt": "Review this code for [REVIEW_TYPE]. Return JSON: {issues: [{id, title, description, severity, suggestion}]}. Code: [CODE_CONTENT]"}`
+   - Where `[CODE_CONTENT]` is the code text (properly JSON-escaped inside the prompt string)
+3. Script path: `[LLM_SCRIPT]` = `{{VIBE_PATH}}/hooks/scripts/llm-orchestrate.js`
+4. Run GPT + Gemini in PARALLEL (two Bash tool calls at once):
+
+```bash
+# GPT review (Bash tool call 1)
+node "[LLM_SCRIPT]" gpt orchestrate-json --input "[SCRATCHPAD]/review-input.json"
+```
+
+```bash
+# Gemini review (Bash tool call 2 - run in parallel)
+node "[LLM_SCRIPT]" gemini orchestrate-json --input "[SCRATCHPAD]/review-input.json"
+```
+
+## File Reading Policy (Mandatory)
+
+- **소스코드 파일**: 리뷰 대상 파일은 반드시 `Read` 도구로 전체 파일을 읽은 후 리뷰할 것 (Grep으로 훑어보기 금지)
+- **Grep 사용 제한**: 파일 위치 탐색(어떤 파일에 있는지 찾기)에만 사용. 파일 내용 파악 및 리뷰에는 반드시 Read 사용
+- **에이전트 spawn 시**: 프롬프트에 "대상 파일을 Read 도구로 전체 읽은 후 분석하라"를 반드시 포함할 것
+- **부분 읽기 금지**: Grep 결과의 주변 몇 줄만 보고 판단하지 말 것. 전체 맥락을 파악해야 정확한 리뷰 가능
+
+## Priority System
+
+| Priority | Criteria | Action |
+|----------|----------|--------|
+| P1 | Security vulnerabilities, data loss, crashes | Block merge, fix immediately |
+| P2 | Performance issues, architecture violations, missing tests | Fix before merge |
+| P3 | Style, refactoring suggestions, documentation | Add to backlog |
+
+## Convergence Rules (Over-Diagnosis Prevention)
+
+> **Principle**: Reviews must converge. A review that always finds more issues is broken, not thorough.
+
+### Scope Limiting
+
+- **Review ONLY changed files** — based on `git diff --name-only`. Never scan the entire project
+- **If no git diff** (first review) — review only files in the target path
+
+### Severity Filtering by Round
+
+| Round | What to Report |
+|-------|---------------|
+| 1st review | P1 + P2 + P3 (all) |
+| 2nd review (same code) | P1 + P2 only (skip P3) |
+| 3rd+ review | P1 only (report only new P1s) |
+
+### Stop Conditions
+
+- **P1 = 0 means MERGE READY** — mergeable even with remaining P2/P3
+- **P1 = 0 after auto-fix means DONE** — record P2 auto-fix failures as TODO and stop
+- **Final P1 list unchanged after Review Debate → DONE** — no new findings = converged
+
+### Anti-Patterns (FORBIDDEN)
+
+- "All items must be verified" → Only P1 is mandatory, P2/P3 are best-effort
+- "Found one more issue" (repeated) → Only report P1s not mentioned in previous review
+- Forcing code changes for P3 issues → P3 goes to TODO files only, never force code changes
+- Infinite retries on auto-fix failure → max 1 retry then move to TODO
+
+## Process
+
+### Phase 1: Tech Stack Detection
+
+Detect project tech stack FIRST before launching reviewers:
+
+```
+Read package.json      -> TypeScript, React, Node.js
+Read pyproject.toml    -> Python, FastAPI, Django
+Read Gemfile           -> Ruby, Rails
+Read pubspec.yaml      -> Flutter, Dart
+Read go.mod            -> Go
+Read CLAUDE.md         -> Explicit tech stack declaration
+```
+
+### Phase 1.5: SPEC ↔ Code Alignment Check
+
+> When SPEC files exist, verify that code changes align with the SPEC
+
+```
+1. Search .claude/vibe/specs/ for related SPEC files (based on git diff filenames)
+2. Compare SPEC REQ-* list against functionality in changed code
+3. If functionality added that's not in SPEC → P2 finding: "Feature added without SPEC"
+4. If implementation differs from SPEC → P1 finding: "SPEC ↔ code mismatch"
+5. If no SPEC files exist → Skip (reviews work without SPEC too)
+```
+
+### Phase 2: Parallel Agent Review (STACK-AWARE) via Orchestrator
+
+**Execution via Orchestrator (12+ agents in parallel):**
+```bash
+node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/infra/orchestrator/index.js').then(o => o.review(['FILE_PATHS'], ['DETECTED_STACKS']).then(r => console.log(r.content[0].text)))"
+```
+
+**Example:**
+```bash
+# Review changed files with TypeScript + React stack
+node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/infra/orchestrator/index.js').then(o => o.review(['src/api/users.ts', 'src/components/Login.tsx'], ['TypeScript', 'React']).then(r => console.log(r.content[0].text)))"
+```
+
+**Core Reviewers (Always Run):**
+| Agent | Focus |
+|-------|-------|
+| security-reviewer | OWASP Top 10, vulnerabilities |
+| data-integrity-reviewer | Data validation, constraints |
+| performance-reviewer | N+1 queries, memory leaks |
+| architecture-reviewer | Layer violations, cycles |
+| complexity-reviewer | Cyclomatic complexity, length |
+| simplicity-reviewer | Over-abstraction, dead code |
+| git-history-reviewer | Churn files, risk patterns |
+| test-coverage-reviewer | Missing tests, edge cases |
+
+**Stack-Specific Reviewers (Conditional):**
+| Agent | Condition |
+|-------|-----------|
+| python-reviewer | .py files in diff |
+| typescript-reviewer | .ts/.tsx files OR tsconfig |
+| rails-reviewer | Gemfile has rails |
+| react-reviewer | package.json has react |
+
+### Phase 2.5: UI/UX Review Agents (Auto-triggered)
+
+> **활성화 조건**: 변경된 파일 중 UI 파일 존재 (`.tsx`, `.jsx`, `.vue`, `.svelte`, `.html`, `.css`, `.scss`)
+> **비활성화**: `.claude/vibe/config.json`에 `"uiUxAnalysis": false` 설정
+
+**기존 12+ 리뷰 에이전트와 병렬 실행:**
+
+| Agent | Role | Output |
+|-------|------|--------|
+| ⑥ ux-compliance-reviewer | UX 가이드라인 준수 검증 | P1/P2/P3 findings |
+| ⑦ ui-a11y-auditor | WCAG 2.1 AA 접근성 감사 | P1/P2/P3 findings |
+| ⑧ ui-antipattern-detector | UI 안티패턴 + 디자인 시스템 일관성 | P1/P2/P3 findings |
+
+**실행 방법 — 기존 Phase 2 에이전트와 병렬 실행:**
+
+```text
+# ⑥ UX 준수 검증 (Haiku)
+Task(subagent_type="ux-compliance-reviewer",
+  prompt="Review UI files for UX guideline compliance: {changed_ui_files}. Use core_ui_search against ux-guidelines and web-interface domains.")
+
+# ⑦ 접근성 감사 (Haiku)
+Task(subagent_type="ui-a11y-auditor",
+  prompt="Audit UI files for WCAG 2.1 AA compliance: {changed_ui_files}.")
+
+# ⑧ 안티패턴 검출 (Haiku)
+Task(subagent_type="ui-antipattern-detector",
+  prompt="Detect UI anti-patterns in: {changed_ui_files}. Check against MASTER.md if exists at .claude/vibe/design-system/{project}/MASTER.md.")
+```
+
+**findings 통합**: ⑥⑦⑧ findings를 기존 findings[]와 병합 → P1/P2/P3 통합 정렬
+
+**⑦ Critical finding 에스컬레이션**: ui-a11y-auditor의 P1 finding은 Review Debate Team(Phase 4.5)에 자동 포함
+
+### Phase 2.7: Boundary Mismatch Detection (Integration Coherence)
+
+> **활성화 조건**: 변경된 파일 중 API route + 프론트엔드 훅/컴포넌트가 함께 존재
+> 경계면 불일치는 개별 파일 리뷰로는 발견 불가 — **양쪽을 동시에 읽어야** 잡힘
+
+**검증 방법: "양쪽 동시 읽기"**
+
+반드시 **생산자와 소비자 코드를 동시에** Read하여 교차 비교한다.
+
+| 검증 영역 | 생산자 (왼쪽) | 소비자 (오른쪽) | 검증 내용 |
+|----------|-------------|---------------|----------|
+| API ↔ 훅 타입 | route의 Response.json() shape | hooks의 fetch\<T\> 타입 | shape 일치, 래핑 unwrap, case 변환 |
+| 라우팅 정합성 | src/app/ page 파일 경로 | href, router.push 값 | 경로 매칭, route group 처리, 동적 세그먼트 |
+| 상태 전이 | STATE_TRANSITIONS 맵 | .update({ status }) 코드 | 죽은 전이, 무단 전이, 중간→최종 누락 |
+| 데이터 흐름 | DB 스키마 필드명 | API 응답 → 프론트 타입 | 필드명 일치, optional 처리 일관성 |
+
+**실행 방식 — 변경 파일 기반 자동 판별:**
+
+```text
+1. git diff에서 API route 파일과 대응 프론트 파일을 짝으로 매칭
+2. 짝이 있는 경우 → 아래 4개 검증 실행
+3. 짝이 없는 경우 (API만 또는 프론트만 변경) → "대응 파일 미변경" 경고 후 스킵
+```
+
+**검증 체크리스트:**
+
+- [ ] API 응답 shape과 대응 훅의 제네릭 타입이 일치
+- [ ] 래핑된 응답(`{ items: [...] }`)은 훅에서 unwrap하는지 확인
+- [ ] snake_case ↔ camelCase 변환이 일관되게 적용
+- [ ] 모든 API 엔드포인트에 대응하는 프론트 훅이 존재하고 실제 호출됨
+- [ ] 코드 내 모든 href/router.push 값이 실제 page 파일 경로와 매칭
+- [ ] 정의된 모든 상태 전이가 코드에서 실행됨 (죽은 전이 없음)
+- [ ] 프론트에서 상태 기반 분기의 값이 실제 도달 가능한 상태
+- [ ] DB 필드명 → API 응답 필드명 → 프론트 타입 정의 간 매핑 일관
+
+**Findings 분류:**
+- 경계면 불일치 → **P1** (런타임 에러의 주요 원인)
+- 대응 파일 미존재 (API 있으나 훅 없음) → **P2**
+- case 변환 비일관성 → **P2**
+
+### Phase 3: Deep Analysis
+
+After agent results:
+
+1. **System Context**: Component interactions, data flow, external dependencies
+2. **Stakeholder Perspectives**: Developers, Ops, Security, Business
+3. **Edge Cases**: Race conditions, resource exhaustion, network failures
+4. **Multiple Angles**: Technical excellence, business value, risk management
+
+### Phase 4: Findings Synthesis
+
+```
+REVIEW FINDINGS
+
+P1 CRITICAL (Blocks Merge) - N issues
+1. [SECURITY] SQL Injection in user query
+   Location: src/api/users.py:42
+   Fix: Use parameterized queries
+
+P2 IMPORTANT (Should Fix) - N issues
+2. [PERF] N+1 query in user list
+3. [ARCH] Circular dependency detected
+
+P3 NICE-TO-HAVE (Enhancement) - N issues
+4. [STYLE] Consider extracting helper function
+```
+
+### Phase 4.5: Agent Teams — Review Debate
+
+> **팀 정의**: `agents/teams/review-debate-team.md` 참조 (Code Review 컨텍스트)
+> 설정: `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` + `teammateMode: in-process` (`~/.claude/settings.json` 전역 — postinstall 자동 설정)
+
+**토론 결과 예시:**
+
+```
+🤝 REVIEW DEBATE RESULTS
+
+Team Consensus (4 reviewers):
+
+✅ Validated P1 (unanimous):
+  1. [SECURITY] SQL Injection — 4/4 agree critical
+
+⬆️ Upgraded P2→P1 (debate result):
+  2. [PERF] Unbounded query — architecture-reviewer pointed out
+     cascading failure risk under load → team agreed P1
+
+⬇️ Downgraded P1→P2 (debate result):
+  3. [SECURITY] CSRF on read-only endpoint — simplicity-reviewer
+     noted endpoint has no side effects → team agreed P2
+
+❌ Removed (false positive):
+  4. [ARCH] "Circular dependency" — architecture-reviewer confirmed
+     this is intentional bi-directional reference, not a cycle
+
+🆕 New findings (team discussion):
+  5. [DATA] Race condition in concurrent updates — emerged from
+     security + performance discussion
+```
+
+### Phase 5: Auto-Fix (P1/P2)
+
+**Auto-fixable issues are resolved immediately:**
+
+```
+🔧 AUTO-FIX Starting...
+
+P1 Critical:
+  1. [SECURITY] SQL Injection → Fixed with parameterized query ✅
+  2. [DATA] Missing transaction rollback → Added try-finally ✅
+
+P2 Important:
+  3. [PERF] N+1 query → Added select_related ✅
+  4. [ARCH] Circular dependency → Separated dependencies ✅
+  5. [TEST] Missing edge case → Added test ✅
+
+🔍 Re-validating...
+  ✅ Build successful
+  ✅ Tests passed
+
+✅ 5 issues auto-fixed!
+```
+
+**Cases that cannot be auto-fixed:**
+- Requires large-scale architecture changes
+- Requires business logic decisions
+- Requires user confirmation
+
+→ Manual handling instructions in Phase 6
+
+### Auto-Fix 실패 시 Codex Rescue (Codex 플러그인 활성화 시)
+
+P1/P2 auto-fix **3회 실패** 시, Codex에 위임:
+
+```
+/codex:rescue "Fix {priority} issue: {issue-description}. File: {file-path}"
+```
+
+Codex 수정 완료 후 해당 리뷰 에이전트가 재검증.
+
+### Phase 6: Todo File Creation (Items Requiring Manual Handling)
+
+Save **remaining** findings to `.claude/vibe/todos/`:
+
+```
+{priority}-{category}-{short-desc}.md
+
+Examples:
+- P2-arch-large-refactor.md  (Cannot be auto-fixed)
+- P3-style-extract-helper.md (Backlog)
+```
+
+## Output
+
+```
+CODE REVIEW SUMMARY
+PR #123: Add user authentication
+
+Reviewers: 13 agents
+⏱️ Started: {start_time}
+⏱️ Completed: {getCurrentTime 결과}
+
+Score: 92/100 (Good) ← Score after auto-fix
+
+Issues Found:
+- P1 Critical: 2 → 0 (✅ Auto-fixed)
+- P2 Important: 5 → 1 (✅ 4 auto-fixed)
+- P3 Nice-to-have: 3 (Backlog)
+
+Auto-Fixed: 6 issues
+- [SECURITY] SQL Injection ✅
+- [DATA] Transaction rollback ✅
+- [PERF] N+1 query ✅
+- [ARCH] Circular dependency ✅
+- [PERF] Unnecessary loop ✅
+- [TEST] Missing edge case ✅
+
+Remaining (Manual handling required):
+- P2-arch-large-refactor.md (Architecture decision required)
+- P3-style-extract-helper.md (Backlog)
+- P3-docs-add-readme.md (Backlog)
+
+✅ MERGE READY (P1/P2 resolved)
+```
+
+### Phase 7: Guide to Fix Workflow (Manual Handling Items)
+
+**Choose workflow when handling remaining issues:**
+
+```
+## Fix Workflow
+
+Choose a workflow to fix the discovered issues:
+
+| Task Scale | Recommended Approach |
+|------------|---------------------|
+| Simple fix (1-2 files) | Plan Mode |
+| Complex fix (3+ files, validation needed) | /vibe.spec |
+
+1. `/vibe.spec "fix: issue-name"` - VIBE workflow (SPEC validation + re-review)
+2. Plan Mode - Quick fix (for simple tasks)
+
+Which approach would you like to proceed with?
+```
+
+- Wait for user's choice before proceeding
+- If user chooses VIBE → wait for `/vibe.spec` command
+- If user chooses Plan Mode → proceed with EnterPlanMode
+
+## Core Tools (Code Analysis)
+
+### Tool Invocation
+
+All tools are called via:
+
+```bash
+node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/tools/index.js').then(t => t.TOOL_NAME({...args}).then(r => console.log(r.content[0].text)))"
+```
+
+### Recommended Tools for Review
+
+| Tool | Purpose | When to Use |
+|------|---------|-------------|
+| `validateCodeQuality` | Code quality check | Overall code quality scan |
+| `analyzeComplexity` | Complexity metrics | Check function complexity |
+| `findSymbol` | Find definitions | Locate implementations |
+| `findReferences` | Find all usages | Track symbol usage |
+| `saveMemory` | Save findings | Store important review findings |
+
+### Example Tool Usage in Review
+
+**1. Validate code quality before review:**
+
+```bash
+node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/tools/index.js').then(t => t.validateCodeQuality({targetPath: 'src/', projectPath: process.cwd()}).then(r => console.log(r.content[0].text)))"
+```
+
+**2. Analyze complexity of changed files:**
+
+```bash
+node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/tools/index.js').then(t => t.analyzeComplexity({targetPath: 'src/api/users.ts', projectPath: process.cwd()}).then(r => console.log(r.content[0].text)))"
+```
+
+**3. Save critical finding for reference:**
+
+```bash
+node -e "import('{{VIBE_PATH_URL}}/node_modules/@su-record/vibe/dist/tools/index.js').then(t => t.saveMemory({key: 'review-pr123-critical', value: 'SQL injection in users.py:42', category: 'review', projectPath: process.cwd()}).then(r => console.log(r.content[0].text)))"
+```
+
+---
+
+## Quality Gate (Mandatory)
+
+### Review Quality Checklist
+
+Before completing review, check P1-critical items. P2/P3 items are best-effort:
+
+| Category | Check Item | Weight |
+|----------|------------|--------|
+| **Security** | OWASP Top 10 vulnerabilities scanned | 20% |
+| **Security** | Authentication/authorization verified | 10% |
+| **Security** | Sensitive data exposure checked | 10% |
+| **Performance** | N+1 queries detected and flagged | 10% |
+| **Performance** | Memory leaks checked | 5% |
+| **Architecture** | Layer violations detected | 10% |
+| **Architecture** | Circular dependencies checked | 5% |
+| **Code Quality** | Complexity limits enforced | 10% |
+| **Code Quality** | Forbidden patterns detected | 10% |
+| **Testing** | Test coverage gaps identified | 5% |
+| **Documentation** | Public API documentation checked | 5% |
+
+### Review Score Calculation
+
+```
+Score = 100 - (P1 × 20) - (P2 × 5) - (P3 × 1)
+
+Grades:
+- 95-100: ✅ EXCELLENT - Merge ready
+- 90-94:  ⚠️ GOOD - Minor fixes required before merge
+- 80-89:  ⚠️ FAIR - Must fix P2 issues
+- 0-79:   ❌ POOR - Block merge, fix P1/P2
+```
+
+### Merge Decision Matrix
+
+| P1 Count | P2 Count | Decision |
+|----------|----------|----------|
+| 0 | 0-2 | ✅ MERGE READY |
+| 0 | 3+ | ⚠️ FIX P2 FIRST |
+| 1+ | Any | ❌ BLOCKED |
+
+### Auto-Fix Capability Matrix
+
+| Issue Type | Auto-Fixable | Method |
+|------------|--------------|--------|
+| SQL Injection | ✅ Yes | Parameterized query |
+| Missing transaction | ✅ Yes | Add try-finally |
+| N+1 query | ✅ Yes | Add eager loading |
+| Circular dependency | ⚠️ Partial | Suggest restructure |
+| Missing tests | ✅ Yes | Generate test skeleton |
+| Hardcoded secrets | ❌ No | Flag for manual review |
+| Architecture violation | ❌ No | Suggest refactoring plan |
+
+### Forbidden Patterns (P1 Critical)
+
+| Pattern | Risk Level | Detection Method |
+|---------|------------|------------------|
+| Hardcoded credentials | Critical | Regex + entropy scan |
+| SQL string concatenation | Critical | AST analysis |
+| `eval()` or `exec()` | Critical | AST analysis |
+| Disabled CSRF protection | Critical | Config scan |
+| Debug mode in production | Critical | Config scan |
+| Unvalidated redirects | High | URL pattern scan |
+
+### Review Output Requirements
+
+Every review MUST produce:
+
+1. **Summary Statistics**
+   - Total issues by priority (P1/P2/P3)
+   - Auto-fixed count
+   - Remaining manual fixes
+
+2. **Detailed Findings**
+   - File path and line number
+   - Issue description
+   - Recommended fix
+   - Auto-fix status (applied/pending/manual)
+
+3. **Quality Score**
+   - Numerical score (0-100)
+   - Grade (EXCELLENT/GOOD/FAIR/POOR)
+   - Merge recommendation
+
+---
+
+ARGUMENTS: $ARGUMENTS