code-review（代码审查）

Code Review Skill

Review code for correctness, security, performance, testing, documentation, and architecture. Produces two outputs:

1. Structured findings for Claude Code to act on
2. REVIEW.md human-readable summary

Review Workflow

0. MANDATORY: Use context7 Before Any Code Changes

CRITICAL: Before creating, editing, or suggesting code changes, ALWAYS use context7 to look up current documentation.

When to use context7:

• Before suggesting fixes or improvements
• When reviewing library/framework usage
• Before writing code examples or snippets
• When uncertain about API behavior or best practices

How to use:

1. Resolve library: context7 resolve <library-name>
2. Get docs: context7 get-library-docs --library <library> --topic <topic>

Example:

context7 resolve react
context7 get-library-docs --library react --topic "useEffect dependencies"

1. Determine Review Scope

Identify what's being reviewed:

• Single file: Review that file
• Directory: Review all relevant files
• Diff/PR: Focus on changed lines with surrounding context
• Entire codebase: Start with entry points, follow dependencies

2. Select Review Depth

Choose automatically based on context, or accept user override:

3. Detect Languages and Load References

Identify languages present, then load relevant reference files:

• R → references/r.md
• Python → references/python.md
• JavaScript → references/javascript.md
• SQL → references/sql.md
• C++ → references/cpp.md
• Rust → references/rust.md
• Go → references/go.md
• Ansible → references/ansible.md
• Kubernetes/Kustomize → references/kubernetes.md
• Dockerfile → references/dockerfile.md
• Docker Compose → references/docker-compose.md
• Bash → references/bash.md

4. Use context7 MCP for Documentation (MANDATORY)

ALWAYS query context7 when:

• Reviewing ANY library/framework usage (not just unfamiliar ones)
• Before suggesting code changes or fixes
• Checking if APIs are used correctly
• Verifying deprecated patterns
• Confirming best practices for specific versions
• Writing code examples or snippets in review feedback

Process:

1. Identify libraries/frameworks in the code being reviewed
2. Use context7 resolve <library> for each one
3. Use context7 get-library-docs to verify API usage and patterns
4. Only THEN proceed with review findings

Example queries:

• context7 resolve react then context7 get-library-docs --library react --topic "hooks"
• context7 resolve tensorflow then context7 get-library-docs --library tensorflow --topic "layers"
• context7 resolve tidyverse for R tidyverse patterns
• context7 resolve kubernetes for K8s manifest validation
• context7 resolve express for Node.js API patterns

Never skip this step - outdated or incorrect documentation can lead to poor review suggestions.

5. Spawn Subagents for Parallel Review

Use subagents to parallelize review work:

Language Subagent (one per language detected):

Task: Review [language] code in [files] for idioms, patterns, and language-specific issues.
Focus: Style, idioms, language-specific performance, common pitfalls.
Reference: Load references/[language].md
Output: Structured findings list

Security Subagent:

Task: Analyze [files] for security vulnerabilities.
Focus: Injection, auth issues, secrets exposure, unsafe operations, dependency risks.
Output: Security findings with severity and remediation

Architecture Subagent:

Task: Review overall structure and design of [files/project].
Focus: Coupling, cohesion, separation of concerns, design patterns, testability.
Output: Architecture findings and recommendations

6. Review Categories

Each category produces findings with severity ratings.

Correctness

• Logic errors and bugs
• Edge cases not handled
• Off-by-one errors
• Null/undefined handling
• Type mismatches
• Race conditions

Security

• Injection vulnerabilities (SQL, command, XSS)
• Authentication/authorization flaws
• Secrets in code
• Unsafe deserialization
• Path traversal
• Dependency vulnerabilities

Performance

• Algorithmic complexity issues
• Unnecessary allocations
• N+1 queries
• Missing caching opportunities
• Blocking operations
• Memory leaks

Testing

• Missing test coverage
• Untested edge cases
• Brittle tests
• Missing integration tests
• Inadequate mocking

Documentation

• Missing function/class docstrings
• Outdated comments
• Unclear variable names
• Missing README updates
• Undocumented public APIs

Architecture

• Tight coupling
• God objects/functions
• Circular dependencies
• Layer violations
• Missing abstractions
• Poor separation of concerns

7. Classify Findings

Rate each finding:

8. Generate Outputs

Output 1: Claude Code Action Format

Produce structured findings Claude Code can act on directly:

## File: [filepath]

### [Line X-Y]: [Brief title]
**Severity**: Critical|Major|Minor|Nitpick
**Category**: Correctness|Security|Performance|Testing|Documentation|Architecture

**Issue**: [Clear description of the problem]

**Current code**:
[relevant code snippet]

**Suggested fix**:
[corrected code snippet]

**Rationale**: [Why this change improves the code]

---

Group findings by file, ordered by severity (Critical first).

Output 2: REVIEW.md Human Summary

Write to REVIEW.md in the project root:

# Code Review Summary

**Reviewed**: [files/scope]
**Depth**: Quick|Standard|Deep
**Date**: [timestamp]

## Overview

[2-3 sentence summary of overall code quality and key concerns]

## Findings by Severity

### Critical ([count])
- [one-line summary with file:line reference]

### Major ([count])
- [one-line summary with file:line reference]

### Minor ([count])
- [one-line summary with file:line reference]

### Nitpicks ([count])
- [one-line summary with file:line reference]

## Category Breakdown

| Category | Critical | Major | Minor | Nitpick |
|----------|----------|-------|-------|---------|
| Correctness | X | X | X | X |
| Security | X | X | X | X |
| Performance | X | X | X | X |
| Testing | X | X | X | X |
| Documentation | X | X | X | X |
| Architecture | X | X | X | X |

## Recommendations

[Prioritized list of recommended actions]

## Positive Observations

[Things done well - important for balanced feedback]

9. Iterate on Critical/Major Issues

After generating outputs:

1. Use context7 to verify fix approaches before implementing
2. If user confirms, apply fixes for Critical and Major issues
3. Re-review changed code to verify fixes don't introduce new issues
4. Update REVIEW.md with resolution status

REMINDER: Use context7 to look up correct API usage before suggesting or implementing any code changes.