code-review（代码审查）

Code Review Skill

Reviews code through battle-tested engineering principles with comprehensive quality gates.

Usage

Tell me what to review:

• "Review src/auth.ts"
• "Review the changes in git diff HEAD~3"
• "Review this PR for merge readiness"

Review Process

1. Read the code — Understand what it does
2. Verify functionality — Check for bugs, edge cases, correctness
3. Assess test coverage — Ensure changed code has adequate tests
4. Evaluate test quality — Check tests follow best practices (not flaky)
5. Check guideline compliance — Verify AGENTS.md/CLAUDE.md adherence
6. Apply philosophy lenses — Check against engineering principles
7. Run QA checks — Execute lint, typecheck, tests (after fixes)
8. Output structured findings — Categorized, actionable feedback

1. Functionality Verification

Correctness Checklist

• Does the code do what it's supposed to do?
• Are all requirements/acceptance criteria met?
• Are edge cases handled? (null, empty, boundary values, overflow)
• Are error conditions handled gracefully?
• Is the happy path tested AND working?
• Are race conditions possible? (async, concurrent access)
• Is state managed correctly? (no stale data, proper initialization)

Bug Detection Signals

2. Test Coverage Assessment

Coverage Requirements

• New code: Must have tests covering primary functionality
• Bug fixes: Must have regression test proving fix
• Changed code: Tests should cover modified behavior
• Critical paths: Auth, payments, data mutations require high coverage

Coverage Analysis

Is the changed code tested? → NO → Block: Add tests
     ↓ YES
Are edge cases covered? → NO → Request edge case tests
     ↓ YES
Are error paths tested? → NO → Request error handling tests
     ↓ YES
Coverage adequate ✓

Google's Coverage Guidelines

• 60%: Acceptable minimum
• 75%: Commendable
• 90%: Exemplary
• Per-commit: 90%+ for changed lines is reasonable

Focus on what's NOT covered rather than hitting arbitrary numbers.

3. Test Quality (Non-Flaky Tests)

FIRST Principles for Good Tests

Flaky Test Detection

Tests are flaky if they fail intermittently without code changes.

Common Causes:

• Time/date dependencies (use fixed/mocked time)
• Random data without seed control
• Shared mutable state between tests
• Network/external service calls
• Race conditions in async code
• File system dependencies
• Database state leakage
• Order-dependent tests

Test Quality Checklist

• Tests are deterministic (same input → same output)
• Tests clean up after themselves (no side effects)
• Tests don't depend on execution order
• Async operations properly awaited
• External dependencies mocked/stubbed
• Time-sensitive tests use controlled time
• Random values use seeded generators
• Tests have meaningful assertions (not just expect(true))
• Test names describe behavior being tested
• One logical assertion per test

Red Flags in Tests

• sleep() or arbitrary delays
• Commented-out assertions
• Empty catch blocks swallowing failures
• Tests that pass when logic is deleted (useless tests)
• expect(result).toBeDefined() without checking value
• Shared state between it() blocks
• Network calls without mocking

4. QA Checks

Before Approving, Verify:

1. Lint passes: npm run lint, eslint, etc.
2. Types check: tsc --noEmit, npm run typecheck, etc.
3. Tests pass: npm test, pytest, etc.
4. Build succeeds: npm run build, cargo build, etc.

Automated QA Flow

Run lint → FAIL → Must fix before merge
    ↓ PASS
Run typecheck → FAIL → Must fix before merge
    ↓ PASS
Run tests → FAIL → Must fix (unless pre-existing)
    ↓ PASS
Build → FAIL → Must fix before merge
    ↓ PASS
QA Passed ✓

5. Guidelines Compliance

AGENTS.md / CLAUDE.md Adherence

Check if the codebase has guidance files:

• AGENTS.md or CLAUDE.md in root or relevant directories
• Project-specific coding standards
• Architecture decisions and patterns

Compliance Checklist

• Follows documented code style
• Uses approved libraries/patterns
• Adheres to naming conventions
• Follows directory structure guidelines
• Respects forbidden patterns (if documented)
• Matches documented architectural decisions

6. Engineering Philosophy Lenses

Grug Brain Check (Complexity Demon Detection)

• Is this the simplest solution?
• Could a newcomer understand it in 30 seconds?
• Are there unnecessary abstractions?

Code Smell Detection

Unix Philosophy Check

• Does each function do ONE thing well?
• Can components be tested in isolation?
• Are there side effects in "pure" functions?

DRY/KISS

• Is logic duplicated?
• Is there a simpler solution?

Output Format

## Summary
[One-line verdict: APPROVE / NEEDS WORK / RETHINK]

## QA Status
- Lint: ✓/✗
- Typecheck: ✓/✗
- Tests: ✓/✗ (X passed, Y failed)
- Build: ✓/✗

## Functionality Issues
[Bugs, incorrect behavior, edge cases]

## Test Coverage Gaps
[Untested code paths, missing tests]

## Test Quality Issues
[Flaky tests, bad patterns]

## Guidelines Violations
[AGENTS.md/CLAUDE.md non-compliance]

## Code Quality
[Design issues, maintainability]

## Nitpicks
[Minor suggestions]

## What's Good
[Acknowledge solid work]

Severity Levels

Master Decision Tree

Does it work correctly? → NO → 🔴 Fix bugs first
     ↓ YES
Does it have tests? → NO → 🔴 Add tests
     ↓ YES
Are tests good (not flaky)? → NO → 🟠 Fix test quality
     ↓ YES
Does QA pass? → NO → 🔴 Fix QA failures
     ↓ YES
Follows guidelines? → NO → 🟠 Align with standards
     ↓ YES
Is it simple? → NO → 🟡 Consider simplifying
     ↓ YES
APPROVE ✓

What NOT to Do

• Don't nitpick formatting (use formatters)
• Don't demand unnecessary abstractions
• Don't block on style preferences
• Don't rewrite working code for elegance alone
• Don't ignore test failures as "probably flaky"