review（审查）

Code Review

Review the current branch's changes against main.

Usage

• /review or /review 2 - Standard review
• /review 1 - Quick sanity check
• /review 3 - Deep analysis (core checks)
• /review 4 - Experimental (Level 3 + advanced checks for evaluation)

Setup

# Get the diff
git diff main...HEAD

# Get list of changed files
git diff --name-only main...HEAD

Level 1: Quick

Fast sanity check. Only blockers.

• Types compile (tsc --noEmit)
• No obvious runtime errors
• No secrets or credentials in diff
• No console.log/debugger statements left in
• Imports resolve

Output: "No blockers found" or list of blockers. No grade.

Level 2: Standard

Full review with grade.

Checklist

Logic & Correctness

• Code does what the PR description says
• Edge cases handled (null, empty, zero, negative)
• Error paths handled appropriately

Types & Safety

• No any types introduced
• Null/undefined properly checked
• Type assertions (as) justified

React Patterns

• No useEffect for derived state (use useMemo or compute directly)
• No useEffect for data fetching (use React Query/tRPC/server components)
• Dependencies arrays correct
• No missing keys in lists

API & Data

• API returns only needed fields
• No N+1 queries visible in diff
• Mutations invalidate relevant caches

Security Basics

• User input validated/sanitized
• No SQL/command injection vectors
• Auth checks present where needed

Style

• Follows existing patterns in codebase
• No commented-out code
• Clear naming

Output Format

## Review: [PR/Branch name]

### Summary
[2-3 sentences on what this change does and overall impression]

### Issues
[List any problems found, grouped by severity]

### Suggestions
[Optional improvements, not blockers]

### Grade: [A-F] ([score]/100)

Level 3: Deep

Everything in Level 2, plus read and apply each check file in checks/ directory.

For each check file:

1. Read the file from ~/.claude/skills/review/checks/
2. Apply its rules to the diff
3. Report findings under a heading matching the check name

Check files to load

Core (always applied)

• security.md - Security audit
• async.md - Async/await patterns
• errors.md - Error handling coverage
• concurrency.md - Race conditions
• idempotency.md - Idempotent operations
• timezones.md - Date/timezone handling
• memory.md - Memory leaks

Shell/Infra (for.sh,.zsh,.yml files)

• shell.md - Shell script quality
• ansible.md - Ansible task quality
• dotfiles.md - Dotfiles/config management

React/JS (for.ts,.tsx,.js,.jsx files)

• nplus1.md - Database and API query patterns
• accessibility.md - a11y compliance
• hooks.md - React Rules of Hooks
• antipatterns.md - React antipatterns
• state.md - State mutation patterns
• performance.md - Performance issues
• boundaries.md - Error boundaries and fault tolerance

Output Format

Same as Level 2, but with additional sections for each check that found issues.

Level 4: Experimental

Everything in Level 3, plus experimental checks being evaluated for promotion to Level 3.

Run periodically to see if any of these should become standard checks.

Additional check files to load

• ast.md - Dead code, complexity, duplicates, dependency graphs
• advanced-security.md - Prototype pollution, ReDoS, timing attacks
• framework.md - Next.js, tRPC, Prisma, RSC patterns
• testing.md - Test coverage gaps and quality
• documentation.md - JSDoc, comments, README sync
• dependencies.md - Vulnerabilities, unused deps, licenses
• git.md - Commit messages, large files, conflict markers

Output Format

Same as Level 3, but with additional sections for experimental checks. Note which experimental checks found useful issues - candidates for promotion to Level 3.

Promoting Checks

If an experimental check consistently finds real issues, promote it to Level 3:

1. Move the check file entry from Level 4 list to Level 3 list
2. Update this file