review
并行子代理进行代码审查。
📖 中文导读:并行子代理进行代码审查。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。
name: review description: Code review using parallel subagents. allowed-tools: Read, Grep, Glob, Bash, Task
Code Review
Code review using 3 parallel subagents. Scope detected automatically.
Workflow
- Detect scope (see below)
- Launch 3 parallel reviewer subagents
- Aggregate findings by severity, deduplicate
- Report summary with actionable items
Scope Detection
# Priority order:
# 1. User provided specific files/paths? -> Review those
# 2. Feature branch? -> diff vs main
# 3. Upstream remote? -> diff vs upstream/main
# 4. Version tags? -> diff vs last tag
# 5. Unpushed commits? -> diff vs origin/main
# 6. Staged changes? -> git diff --cached
# 7. Unstaged changes? -> git diff
# 8. Nothing -> inform user
Parallel Subagents
Launch all 3 in parallel using Task tool with subagent_type: reviewer:
Agent 1: Correctness
Logic and functional correctness.
- Logic errors, off-by-one, boundary conditions
- Null/undefined handling
- State consistency (partial failures)
- Race conditions (if concurrent)
- Edge cases: empty, max, negative, zero
- Control flow correctness
- Algorithm correctness
Agent 2: Safety
Security and error handling (defensive thinking).
- Input validation at boundaries
- Hardcoded secrets, credentials
- Injection: SQL, XSS, command, path traversal
- Auth/authz checks
- Sensitive data in logs
- Silent failures, empty catch blocks
- Error propagation (swallowed vs bubbled)
- Fallback behavior hiding problems
- Resource cleanup on error
Agent 3: Quality
Design, performance, and idioms.
- Fits existing patterns?
- Single responsibility
- Over-engineering (future problems?)
- Code smells: long functions (>40 lines), large files (>400 lines)
- Dead code, unused variables
- Naming: intention-revealing, proportional to scope
- No
_v2,_newsuffixes - Unnecessary allocations (
Stringvs&str,Vecvs&[T]) - O(n²) where O(n) possible
- Blocking I/O in async, N+1 queries
- Language idioms (Rust/Python/Go/TS)
Subagent Prompt
Review code changes for [FOCUS: Correctness | Safety | Quality].
Scope: [diff or files]
Report issues with confidence >= 80%. Format:
[SEVERITY] file:line - Issue
-> Fix
Severities:
- ERROR: Must fix (bugs, security, silent failures)
- WARN: Should fix (smells, performance)
- NIT: Optional (style, minor)
Be thorough. No false positives. Only flag what you're confident about.
Aggregation
- Deduplicate (same issue from multiple agents = 1 issue)
- Sort: ERROR → WARN → NIT
- Group by file
Output Format
## Review Summary
Scope: [X files, Y lines]
Issues: ERROR: X, WARN: X, NIT: X
## Critical (must fix)
[ERROR] file:line - Issue
-> Fix
## Important (should fix)
[WARN] file:line - Issue
-> Fix
## Minor
[NIT] file:line - Issue
-> Fix
## Strengths
- What's done well
## Verdict
LGTM / LGTM with nits / Needs work
Quick Review (small changes)
For < 50 lines, review directly without subagents:
- Logic correct, edge cases handled
- Errors handled, not swallowed
- No security issues
- No debug statements
- Naming clear
- Fits existing patterns