code-review

name	code-review
description	ACNH-wiki 프로젝트의 코드 변경사항을 다중 관점으로 리뷰하고, 버그, 성능, 보안, 코딩 컨벤션 위반사항을 JSON 형식으로 출력한다. 사용 시점: (1) PR 코드 리뷰 요청 시 ("이 PR 리뷰해줘", "코드 리뷰 부탁", "PR 검토해줘") (2) 커밋 리뷰 시 ("이 커밋 검토해줘", "커밋 리뷰해줘") (3) 현재 변경사항 리뷰 시 ("지금 변경사항 리뷰해줘", "내가 바꾼 코드 검토해줘", "코드 리뷰") (4) GitHub PR에 리뷰 코멘트 등록 요청 시

ACNH-wiki Code Review Skill

Multi-perspective code review using specialized reviewer agents in parallel.

Review Modes

FIRST STEP: Ask user to select review scope using AskUserQuestion tool:

{
  "questions": [
    {
      "question": "어떤 범위의 코드를 리뷰할까요?",
      "header": "리뷰 범위",
      "multiSelect": false,
      "options": [
        {
          "label": "현재 변경사항 (Recommended)",
          "description": "Staged + unstaged 모든 커밋되지 않은 변경사항을 리뷰합니다."
        },
        {
          "label": "특정 커밋",
          "description": "커밋 해시를 지정하여 해당 커밋의 변경사항만 리뷰합니다."
        },
        {
          "label": "Pull Request",
          "description": "GitHub PR 번호를 지정하여 PR의 모든 변경사항을 리뷰하고 GitHub에 코멘트를 등록합니다."
        }
      ]
    }
  ]
}

---

Core Workflow

Step 1: Gather Changes

# Current changes
.claude/skills/code-review/scripts/get-diff.sh

# Specific commit
.claude/skills/code-review/scripts/get-diff.sh <commit-hash>

# Pull Request (diff is sufficient for 99% of reviews)
.claude/skills/code-review/scripts/get-diff.sh <pr-number>

Step 2: Diff Size Guard

After gathering the diff, assess its size to ensure review quality:

# Check diff stats
.claude/skills/code-review/scripts/get-diff.sh --stat

Condition	Action
Swift files = 0 (only .md, .json, .xcconfig, etc.)	Run reviewer-conventions only, skip other 4 agents
Changed files > 30 OR diff > 5,000 lines	AskUserQuestion: "변경 범위가 큽니다. 전체 리뷰 vs 주요 파일만 선택할까요?"
Otherwise	Proceed normally with all 5 agents

Step 3: Load Review Guidelines

Read the following skill-local reference files (always available on current branch):

• references/coding-conventions.md - Team coding conventions
• references/architecture-guide.md - Architecture patterns (ReactorKit, Coordinator, etc.)

These contents will be included in the agent prompt (see Step 4).

PR reviews: If you need to access source files on the PR branch, see pr-review-workflow.md Stage 1.

Step 4: Phase 1 - Parallel Agent Execution

Run 5 reviewers in parallel using Task tool (project-local agents in .claude/agents/).

IMPORTANT: Call all 5 Task tools in a single message for parallel execution.

Agent Prompt Template

Each agent receives the same structured prompt:

You are reviewing code changes for the ACNH-wiki (너굴포털+) project.

## Review Metadata
- Mode: [Current Changes | Commit <hash> | PR #<number>]
- Changed Files: [file list extracted from diff headers]

## Coding Conventions Reference
<full contents of references/coding-conventions.md>

## Architecture Guide Reference
<full contents of references/architecture-guide.md>

## Diff to Review
Each diff line includes an explicit line number (e.g., `+  79: code`).
Use these numbers directly for `code_location.line_range` values. Do NOT count lines manually.
<full diff content from Step 1>

Invocation

Task(subagent_type="reviewer-swift-ios", prompt="<structured prompt>")
Task(subagent_type="reviewer-correctness", prompt="<structured prompt>")
Task(subagent_type="reviewer-security", prompt="<structured prompt>")
Task(subagent_type="reviewer-code-quality", prompt="<structured prompt>")
Task(subagent_type="reviewer-conventions", prompt="<structured prompt>")

Note: Agents are defined in .claude/agents/ directory. No plugin installation required.

Step 5: Phase 2 - False Positive Filter

After collecting all Phase 1 results, run the false-positive-filter agent for compound verification.

FP Filter Prompt Template

You are validating Phase 1 code review findings for ACNH-wiki.

## Phase 1 Findings (ALL priorities)
<JSON array of all findings from the 5 Phase 1 reviewers>

## Original Diff
<full diff content from Step 1>

Validate each finding using compound verification.
Output verdicts: TRUE_POSITIVE, FALSE_POSITIVE, DOWNGRADE, or UPGRADE.

Invocation

Task(subagent_type="false-positive-filter", prompt="<structured prompt>")

FP Filter Verdicts

Verdict	Action
TRUE_POSITIVE	Keep finding as-is
FALSE_POSITIVE	Remove finding from final output
DOWNGRADE	Lower priority (e.g., P1 -> P2) with reason
UPGRADE	Raise priority (e.g., P2 -> P0) with evidence

Step 6: Aggregate & Deduplicate

Collect validated findings from Phase 2 and deduplicate:

Deduplication Rules

1. Same file + overlapping line range (+-3 lines) -> Merge as single issue

   • priority: Use the higher (more severe) value
   • body: Use the more detailed description
   • reviewer: List all contributing reviewers (e.g., "reviewer-swift-ios, reviewer-correctness")

2. Same file, non-overlapping lines -> Keep as separate issues

3. Same code location, different categories -> Keep both if perspectives genuinely differ

   • e.g., Safety (swift-ios) + Correctness -> two separate findings

Validation

• P0/P1 findings MUST include all 3 evidence items (Pattern, Proof, Impact)
• If ANY evidence missing -> Downgrade to P2/P3
• Remove all FALSE_POSITIVE findings from final output
• Apply DOWNGRADE/UPGRADE priority adjustments from FP Filter

Step 7: Generate Final Output

{
  "findings": [
    {
      "reviewer": "reviewer-swift-ios",
      "category": "Memory Safety",
      "title": "[P1] Issue title (<=80 chars)",
      "body": "Description with evidence, reference file/line",
      "confidence_score": 0.95,
      "priority": 1,
      "code_location": {
        "absolute_file_path": "/path/to/file.swift",
        "line_range": {"start": 45, "end": 47}
      }
    }
  ],
  "overall_correctness": "patch is correct | patch is incorrect",
  "overall_explanation": "1-3 sentences summary",
  "overall_confidence_score": 0.9
}

Review State Decision:

Condition	Result
P0 > 0 OR P1 > 0	"patch is incorrect"
P0 = 0 AND P1 = 0	"patch is correct"

Step 8: Post to GitHub (PR Reviews Only)

If findings count is 0:

• Inform user: "코드 리뷰 완료! 발견된 이슈가 없습니다."
• END workflow (no GitHub comment needed)

If findings count > 0:

Execute script:

echo '<json>' | .claude/skills/code-review/scripts/post-review-comments.sh --pr <PR_NUMBER>

리뷰 코멘트가 PR에 즉시 등록됩니다.

---

Reviewer Agents

Project-local agents defined in .claude/agents/:

Agent	Model	Focus	File
reviewer-swift-ios	sonnet	Memory safety, RxSwift safety, Concurrency, Performance safety	reviewer-swift-ios.md
reviewer-correctness	sonnet	Logic errors, RxSwift stream correctness, ReactorKit flow correctness	reviewer-correctness.md
reviewer-security	opus	OWASP M1-M10, Privacy, Data protection, CoreData/CloudKit security	reviewer-security.md
reviewer-code-quality	sonnet	ReactorKit architecture, Coordinator pattern, Networking, Maintainability	reviewer-code-quality.md
reviewer-conventions	sonnet	Coding conventions (naming, formatting, RxSwift style, localization)	reviewer-conventions.md
false-positive-filter	opus	Agent-specific validation, compound verification, UPGRADE/DOWNGRADE	false-positive-filter.md

---

Priority Classification

CRITICAL: Same issue = same priority regardless of context.

P0 (Blocker)

• Force unwrap on nil / Force try in production
• Retain cycle / Strong delegate
• Data race / Deadlock
• Data loss / Security bypass
• Hardcoded credentials

P1 (Must Fix)

• Missing [weak self] in async closure
• guard let self = self (should be owner)
• Empty catch block
• Wrong output for common input
• Missing auth check

P2 (Should Fix)

• disposed(by:) on same line as }
• Missing guard blank line
• Edge case failure
• Localization key missing in one lproj

P3 (Nice to Have)

• Naming suggestions
• Style improvements
• Theoretical issues

---

Evidence Requirements (P0/P1)

ALL P0/P1 findings MUST include:

1. Pattern/Violation: What rule/pattern is violated
2. Proof: Code path, crash scenario, or test case
3. Impact: User-facing? Data affected? Scope?

If ANY missing -> Downgrade to P2/P3

---

Quick Reference: Common Violations

Violation	Wrong	Correct	Priority
weak self	guard let self = self	guard let owner = self	P1
disposed	}.disposed(by:)	}\n.disposed(by:)	P2
guard	guard...\ncode()	guard...\n\ncode()	P2
localization	NSLocalizedString(...)	"key".localized	P2
Coordinator bypass	present(vc, animated:)	coordinator?.transition(for:)	P1

---

Scripts

get-diff.sh

.claude/skills/code-review/scripts/get-diff.sh                    # Current changes
.claude/skills/code-review/scripts/get-diff.sh <commit-hash>      # Specific commit
.claude/skills/code-review/scripts/get-diff.sh <pr-number>        # Pull Request

post-review-comments.sh

echo '<json>' | .claude/skills/code-review/scripts/post-review-comments.sh --pr <pr-number>

Flags: --pr <number>, --input <file>, --force

---

References

File	Contents
coding-conventions.md	ACNH-wiki coding conventions
architecture-guide.md	ReactorKit, Coordinator, Data flow patterns
finding-guidelines.md	8 criteria, JSON schema
pr-review-workflow.md	6-stage PR workflow
pending-review-guide.md	GitHub comment format