Code Review (4-Agent)

Comprehensive code review that dispatches 4 specialized agents in parallel for deep analysis.

Workflow

Phase 1: Preparation

1. Detect the review scope:
   • If on a feature branch: git diff $(git merge-base main HEAD)..HEAD
   • If PR number provided: gh pr diff <number>
   • Fallback: staged + unstaged changes
2. Get list of changed files: git diff --name-only <range>
3. Read all changed files in full (agents need complete context)

Phase 2: Dispatch 4 Agents in Parallel

Launch 4 agents in a single turn:

Agent 1 — Correctness & Logic

• Logic bugs, edge cases, null/undefined risks
• Off-by-one errors, race conditions
• Incorrect error handling, missing return values
• Type mismatches, contract violations

Agent 2 — Security

• Injection vulnerabilities (SQL, XSS, command)
• Auth/authz gaps, missing input validation
• Secrets in code, insecure defaults
• OWASP Top 10 patterns

Agent 3 — Performance

• N+1 queries, unnecessary re-renders
• Missing memoization, expensive computations in loops
• Bundle size impact, large imports
• Memory leaks, unbounded collections

Agent 4 — Style & Maintainability

• Naming clarity, dead code, unused imports
• Overly complex logic (cyclomatic complexity)
• Missing or misleading comments
• DRY violations, code duplication

Each agent outputs findings as:

### [CRITICAL|HIGH|MEDIUM|LOW] Title
**File:** path/to/file:line
**Description:** what's wrong and why
**Suggestion:** how to fix

Phase 3: Unified Report

Collect all 4 agent results and produce a single report:

# Code Review Report

## Summary
- X critical, Y high, Z medium, W low findings
- Overall assessment: APPROVE / REQUEST_CHANGES / COMMENT

## Critical & High Findings
(sorted by severity, then by file)

## Medium & Low Findings
(collapsed or summarized)

## Positive Notes
(good patterns worth highlighting)

Phase 4: Action (optional)

• If user says /code-review --fix, attempt to auto-fix MEDIUM and LOW findings
• If reviewing a PR, offer to post the review as a GitHub comment

Arguments

• /code-review — review current branch changes
• /code-review #<number> — review specific PR
• /code-review --fix — auto-fix medium/low findings after review

Rules

• NEVER approve code with CRITICAL findings
• Deduplicate findings across agents (same file:line = merge)
• Include positive feedback — don't make it only negative
• If the diff is huge (>2000 lines), split into chunks and review each
• Be strict but fair — only flag real issues, not style preferences