PR Code Review

Performs comprehensive, multi-dimensional code review for GitHub pull requests using 5 parallel Sonnet agents and confidence-based issue scoring.

When to use

• User asks to review a specific PR (e.g., "review PR #123")
• User mentions "code review", "check this PR", "review pull request"
• User provides a PR number or URL
• User says "review all PRs" or "check open PRs"

Quick start

User: "Review PR #945"
Assistant: [Launches this skill with PR number 945]

Review process

This skill implements an 8-step comprehensive review:

1. Eligibility check (Haiku agent)

   • Verify PR is open, not draft, not already reviewed
   • Skip if ineligible

2. Find CLAUDE.md files (Haiku agent)

   • Locate root and directory-specific CLAUDE.md files
   • Use for compliance checking

3. Get PR summary (Haiku agent)

   • Extract PR description, changes, affected files
   • Understand the context

4. Launch 5 parallel review agents (Sonnet agents)

   • Agent #1: CLAUDE.md compliance
   • Agent #2: Shallow bug scan
   • Agent #3: Git history analysis
   • Agent #4: Previous PR patterns
   • Agent #5: Code comment compliance

5. Score all issues (parallel Haiku agents)

   • Each issue gets 0-100 confidence score
   • Use standardized rubric
   • Verify CLAUDE.md citations

6. Filter issues

   • Keep only issues with score >= 80
   • Skip review if no high-confidence issues

7. Re-verify eligibility (Haiku agent)

   • Ensure PR is still open and not already reviewed

8. Post review comment (gh command)

   • Format with issue descriptions
   • Include file links with full SHA
   • Add feedback footer

Instructions

Step 1: Create todo list

Use TodoWrite to create tracking list:
- Check PR eligibility
- Find CLAUDE.md files
- Get PR summary
- Launch 5 review agents
- Score issues
- Filter and verify
- Post review comment

Step 2: Eligibility check

Launch Haiku agent to verify:

- Is PR closed? → SKIP
- Is PR draft? → SKIP
- Doesn't need review (automated/trivial)? → SKIP
- Already has Claude Code review? → SKIP

Use: gh pr view {number}
     gh api repos/{owner}/{repo}/pulls/{number}/comments

Step 3: Find CLAUDE.md files

Launch Haiku agent to locate:

- Root CLAUDE.md
- Directory-specific CLAUDE.md files for modified paths

Use: gh pr view {number} --json files
     Check for CLAUDE.md in each directory

Return list of CLAUDE.md file paths (not contents).

Step 4: Get PR summary

Launch Haiku agent to extract:

- What the PR adds/changes/fixes
- Key files modified
- Main functionality changes

Use: gh pr view {number}
     gh pr diff {number}

Step 5: Launch 5 parallel review agents

IMPORTANT: Launch all 5 agents in parallel (single message, multiple Task calls).

Agent #1: CLAUDE.md Compliance (Sonnet)

Review PR for CLAUDE.md compliance:
- Read CLAUDE.md files
- Check code style (type hints, docstrings, line length 88, PEP 8)
- Check testing requirements
- Check package management (uv only, no pip)
- Note: CLAUDE.md is guidance for writing code, not all applies to review

Return issues with:
- Issue description
- CLAUDE.md rule violated
- File and line reference

Agent #2: Shallow Bug Scan (Sonnet)

Review PR for obvious bugs:
- Use gh pr diff {number}
- Focus ONLY on changed lines
- Look for: logic errors, incorrect variable usage, missing critical error handling, runtime errors
- Avoid: nitpicks, style issues, false positives, pre-existing issues

Return issues with:
- Issue description
- Bug type
- File and line reference

Agent #3: Git History Analysis (Sonnet)

Review PR against git history:
- Get modified files
- Use git blame and git log
- Look for: violated patterns, repeated historical bugs, relevant commit context

Return issues with:
- Issue description
- Historical context
- File and line reference

Agent #4: Previous PR Patterns (Sonnet)

Review PR against previous PR patterns:
- Get modified files
- Search for recent PRs touching same files
- Check comments on those PRs for applicable feedback

Use: gh pr list --search "path:{file}" --state merged

Return issues with:
- Issue description
- Pattern from PR #X
- File and line reference

Agent #5: Code Comment Compliance (Sonnet)

Review PR for code comment compliance:
- Read modified files
- Check for TODO comments, warnings, documented patterns
- Verify changes follow comment guidance

Return issues with:
- Issue description
- Comment guidance violated
- File and line reference

Step 6: Score all issues

For EACH issue found by the 5 agents, launch a parallel Haiku agent to score it.

Scoring rubric (give to agent verbatim):

Score on 0-5 scale:
- 0: False positive. Not a real issue or pre-existing.
- 1: Low confidence. Might be real but likely false positive.
- 2: Uncertain. Could be real but might be nitpick or minor.
- 3: Moderate confidence. Real issue but not critical.
- 4: High confidence. Real issue that will impact functionality.
- 5: Critical. Definitely a real bug that will cause problems.

Agent instructions:

Score this issue on 0-5 scale using the rubric above.

Issue: [issue description]

For CLAUDE.md issues: Verify CLAUDE.md actually says this.
For bugs: Verify with gh pr diff that this is real.
Check for false positives: pre-existing issues, lines not modified, linter-catchable.

Return: score (0-5) and 1-2 sentence explanation.

Step 7: Filter and re-verify

1. Filter issues: Keep only score >= 4 (high confidence or critical)
2. If no issues >= 4: STOP, do not post review
3. Re-verify eligibility with Haiku agent (same as Step 2)
4. If not eligible: STOP

Step 8: Post review comment

Get PR head commit SHA:

gh pr view {number} --json headRefOid --jq '.headRefOid'

Format comment:

### Code review

Found {N} issues:

1. {brief description} ({CLAUDE.md says "..."} or {bug due to ...})

{link with full SHA}

2. ...

🤖 Generated with [Claude Code](https://claude.ai/code)

<sub>- If this code review was useful, please react with 👍. Otherwise, react with 👎.</sub>

Or if no issues:

### Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

🤖 Generated with [Claude Code](https://claude.ai/code)

Link format:

https://github.com/{owner}/{repo}/blob/{FULL_SHA}/{file_path}#L{start}-L{end}

Requirements:
- Use FULL commit SHA (not short)
- # before line numbers
- L prefix for lines
- Range format: L{start}-L{end}
- Include 1+ lines of context around issue

Post with:

gh pr comment {number} --body "{formatted_comment}"

False positives to avoid

Do NOT flag:

• Pre-existing issues (on lines not modified in PR)
• Linter/typechecker catchable issues (will run in CI)
• Pedantic nitpicks
• Style issues not in CLAUDE.md
• Issues silenced with lint ignore comments
• Intentional functionality changes
• General code quality (unless CLAUDE.md requires it)

Examples

Example 1: Review single PR

User: Review PR #945
Assistant: [Creates todo list, runs 8-step process, posts review]

Example 2: Review multiple PRs

User: Review all open PRs
Assistant: [Gets list with gh pr list, reviews each in sequence]

Example 3: PR is not ready

User: Review PR #959
Assistant: [Checks eligibility, finds incomplete description, skips with explanation]

Best practices

1. Always use parallel agents where possible (Steps 5 and 6)
2. Trust agent scores - they do deep analysis
3. Skip low-confidence issues - only report score >= 80
4. Verify eligibility twice - before and after scoring
5. Use full SHA in links - required for GitHub rendering
6. Keep comments brief - focus on high-impact issues
7. Update todos frequently - show progress to user

Common issues

Agents give contradictory scores: Trust the explanation, not just the number. If explanation says "not a problem" but score is 75, treat as false positive (score should be 0).

Too many issues found: This is good! The parallel agent approach catches more. The scoring step filters to only high-confidence issues (score >= 4).

PR changed during review: The re-verification step (Step 7) catches this.

No CLAUDE.md files: Review still works, just skips CLAUDE.md compliance checks.

Requirements

• GitHub CLI (gh) installed and authenticated
• Access to repository
• Ability to post comments on PRs

Tool usage

• Task tool: Launch all Haiku and Sonnet agents
• TodoWrite: Track progress through 8 steps
• Bash (gh): Interact with GitHub (no WebFetch)
• Read: Access CLAUDE.md files if needed

Output format

After completing review:

✓ Reviewed PR #{number}
- Found {N} high-confidence issues (score >= 4)
- Posted review comment: {URL}

Or:

✓ Reviewed PR #{number}
- No issues found
- Posted approval comment: {URL}

Or:

⊘ Skipped PR #{number}
- Reason: {eligibility issue}