team-code-review

Team Code Review

A multi-agent code review system that forms a reviewer team with different AI models, cross-checks findings between them, and posts a consolidated review as a PR comment.

Team Structure

• Leader (you, the main agent): Orchestrates the review, synthesizes findings, posts the final comment
• Reviewer 1: Claude Opus 4.6 subagent — deep architectural and logic analysis
• Reviewer 2: Claude Sonnet 4.6 subagent — fresh perspective on bugs, patterns, and performance

Review Focus Areas

Each reviewer analyzes the PR diff for:

1. Bugs / Logic Errors: potential bugs, edge case omissions, race conditions, incorrect logic
2. Code Quality / Patterns: SOLID violations, code smells, duplication, naming, readability
3. Performance: algorithmic complexity, memory leaks, unnecessary computations, N+1 queries

Workflow

Execute these phases in order. The key insight: independent parallel review first, then cross-validation, then synthesis. This catches issues that a single reviewer would miss.

Phase 0: Gather PR Context

Determine the PR to review. The user may provide:

• PR number + repo: nalpari/tech-blog#8
• PR URL: https://github.com/owner/repo/pull/8
• Just a PR number (if in a git repo): #8
• Nothing (detect from current branch): gh pr view --json number

Fetch metadata and diff:

# Metadata
gh pr view <PR_NUMBER> --repo <OWNER/REPO> --json number,title,body,baseRefName,headRefName

# Full diff
gh pr diff <PR_NUMBER> --repo <OWNER/REPO> > /tmp/team-review-full-diff.txt

Handling large diffs (this is important — real PRs are often big):

1. Exclude lock files and auto-generated content: filter out pnpm-lock.yaml, package-lock.json, yarn.lock, *.generated.*
2. Focus on source code: prioritize src/, lib/, app/, config files, and migrations
3. If the filtered diff exceeds ~3000 lines, further trim to the most impactful files (auth, API routes, components, utilities) and note which files were excluded
4. Save the focused diff to /tmp/team-review-diff-focused.txt

Phase 1: Parallel Independent Review

Launch both reviewers simultaneously — they have no dependencies on each other.

Reviewer 1 (Claude Opus 4.6) — use the Agent tool:

Agent(
  model: "opus",
  run_in_background: true,
  prompt: "You are Reviewer 1 in a multi-agent code review team. This is research-only — do NOT edit files.

  Step 1: Read the diff file at /tmp/team-review-diff-focused.txt
  Step 2: Produce your review.

  PR Title: {title}
  PR Description: {body}
  Base: {base} → Head: {head}

  Focus areas:
  1. Bugs / Logic Errors — potential bugs, missing edge cases, race conditions, null handling
  2. Code Quality / Patterns — SOLID violations, code smells, duplication, naming
  3. Performance — algorithmic complexity, memory leaks, unnecessary computations

  For each finding: file path, line number (approx OK), severity (CRITICAL/WARNING/INFO), description, suggested fix.
  Note 1-2 positive aspects. Output as structured markdown."
)

Reviewer 2 (Claude Sonnet 4.6) — use the Agent tool:

Agent(
  model: "sonnet",
  run_in_background: true,
  prompt: "You are Reviewer 2 in a multi-agent code review team. This is research-only — do NOT edit files.

  Step 1: Read the diff file at /tmp/team-review-diff-focused.txt
  Step 2: Produce your review.

  PR Title: {title}
  PR Description: {body}
  Base: {base} → Head: {head}

  Focus areas:
  1. Bugs / Logic Errors — potential bugs, missing edge cases, race conditions, null handling
  2. Code Quality / Patterns — SOLID violations, code smells, duplication, naming
  3. Performance — algorithmic complexity, memory leaks, unnecessary computations

  For each finding: file path, line number (approx OK), severity (CRITICAL/WARNING/INFO), description, suggested fix.
  Note 1-2 positive aspects. Output as structured markdown."
)

Phase 2: Cross-Check

Each reviewer examines the other's findings. This catches false positives, validates real issues, and surfaces missed items.

First, save both reviews to structured text files:

• /tmp/team-review-r1-findings.txt — Reviewer 1's findings (from Opus result)
• /tmp/team-review-r2-findings.txt — Reviewer 2's findings (from Sonnet result)

Run both cross-checks in parallel:

Reviewer 1 cross-checks Reviewer 2 — spawn Opus subagent:

Agent(
  model: "opus",
  run_in_background: true,
  prompt: "You are Reviewer 1 (Claude Opus). Cross-check Reviewer 2 (Claude Sonnet)'s findings. Research-only — do NOT edit files.

  Read: /tmp/team-review-r1-findings.txt (your findings)
  Read: /tmp/team-review-r2-findings.txt (Sonnet's findings)

  For each Sonnet finding: AGREE, DISAGREE, or ADD CONTEXT? Brief explanation.
  Did Sonnet catch anything you missed?
  Are any of Sonnet's findings false positives? Explain why.
  Any new issues from combined perspective?

  Output as structured markdown."
)

Reviewer 2 cross-checks Reviewer 1 — spawn Sonnet subagent:

Agent(
  model: "sonnet",
  run_in_background: true,
  prompt: "You are Reviewer 2 (Claude Sonnet). Cross-check Reviewer 1 (Claude Opus)'s findings. Research-only — do NOT edit files.

  Read: /tmp/team-review-r2-findings.txt (your findings)
  Read: /tmp/team-review-r1-findings.txt (Opus's findings)

  For each Opus finding: AGREE, DISAGREE, or ADD CONTEXT? Brief explanation.
  Did Opus catch anything you missed?
  Are any of Opus's findings false positives? Explain why.
  Any new issues from combined perspective?

  Output as structured markdown."
)

Phase 3: Leader Synthesis

You now have four documents. Synthesize them using this decision matrix:

Situation	Action
Both agree on an issue	Include with high confidence
One found it, other confirmed in cross-check	Include
One found it, other disagreed	Include only if evidence is strong; note the disagreement
Only one found it, not addressed in cross-check	Include if specific and well-reasoned
Flagged as false positive by cross-checker	Exclude unless you have strong reason to override

Prioritize findings by severity: CRITICAL first, then WARNING, then INFO.

Phase 4: Post PR Comment

Format the final review in Korean and post as a PR comment:

gh pr comment <PR_NUMBER> --repo <OWNER/REPO> --body "$(cat <<'COMMENT_EOF'
<review content>
COMMENT_EOF
)"

PR Comment Template

Adapt sections based on actual findings — omit empty sections.

# :mag: Team Code Review

> **PR**: #{pr_number} {pr_title}
> **Reviewers**: Claude Opus 4.6, Claude Sonnet 4.6
> **Review method**: Independent parallel review + cross-validation

---

## :memo: 요약
{1-3 문장으로 전체 변경 사항의 의도와 리뷰 결과 요약}

## :bug: 버그 / 로직 오류

| 심각도 | 파일 | 위치 | 설명 | 제안 |
|--------|------|------|------|------|

## :triangular_ruler: 코드 품질 / 패턴

| 우선순위 | 파일 | 위치 | 설명 | 제안 |
|----------|------|------|------|------|

## :zap: 성능

| 영향도 | 파일 | 위치 | 설명 | 제안 |
|--------|------|------|------|------|

## :white_check_mark: 잘된 점
{긍정적 피드백 1-3개}

## :handshake: 리뷰어 합의
{크로스 체크에서 특히 중요했던 합의/이견 사항. 허위 양성으로 판정된 항목도 기록}

---
:robot: *Generated by Team Code Review (Claude Opus 4.6 + Claude Sonnet 4.6)*

Severity Guide

• CRITICAL: Must fix before merge. Runtime errors, data loss, security vulnerabilities.
• WARNING: Should fix. Logic issues, poor patterns, maintenance burden, performance concerns.
• INFO: Nice to have. Style suggestions, minor improvements, alternative approaches.

Edge Cases

• No PR found: Ask the user which PR to review, or offer to review the current branch diff against main/master.
• Reviewer 1 subagent fails: Log the error, proceed with Reviewer 2 only, and note in the final comment that only one reviewer completed analysis.
• Reviewer 2 subagent fails: Log the error, proceed with Reviewer 1 only, and note in the final comment that only one reviewer completed analysis.
• Very large diff (>3000 lines after filtering): Prioritize source files over docs/config. Note excluded files in the review.
• No issues found: Still post a positive review confirming the code looks good, highlighting the positive aspects.
• Cross-repo PR: Always use --repo owner/repo with gh commands when the current directory is not the target repo.

Cleanup

After posting the comment, remove temporary files:

rm -f /tmp/team-review-*.txt