PR Review Skill

This skill provides comprehensive pull request analysis with deep semantic understanding of code changes using parallel specialized auditors.

Capabilities

1. Parallel Sub-Agent Architecture - 5 specialized auditors run concurrently for faster, deeper analysis:
   • Security Auditor - Injection vulnerabilities, auth gaps, data exposure, race conditions
   • Architecture Auditor - Import boundaries, monorepo structure, file organization, naming
   • Constitution Auditor - All 14 project principles compliance verification
   • Code Quality Auditor - Complexity, code smells, performance, testing gaps
   • API Stability Auditor - oRPC compliance (Principle IX), breaking changes (Principle XIII)
2. Smart Aggregation - Findings merged, deduplicated, and sorted by severity
3. PR Metadata Updates - Automatically updates PR title and description (never just recommends)
4. Logic Flow Visualization - Mermaid diagrams for complex workflows
5. Local Mode - Create task list for local fixes instead of posting to GitHub

Usage Modes

Standard Mode (default)

Posts review directly to GitHub PR:

• /pr-review or /pr-review 123

Local Mode (--local)

Creates a task list of issues to fix locally without posting to GitHub:

• /pr-review --local or /pr-review 123 --local

Use local mode when you want to:

• Fix issues before requesting a formal review
• Self-review changes before pushing
• Iterate on fixes without cluttering PR comments

Loop Mode (--loop)

Iteratively reviews, auto-fixes all findings, and re-reviews until 0 findings or max 5 iterations:

• /pr-review --loop or /pr-review 123 --loop

Use loop mode when you want to:

• Auto-fix all review findings without manual intervention
• Iterate until the PR is clean before posting to GitHub
• Get a fully automated review-fix-review cycle

On approval (0 findings), fixes are committed and the review is automatically posted to GitHub. On max iterations or stall, fixes are committed and a local task list of remaining findings is created. If no fixes were applied (all findings skipped), the commit is skipped.

Note: --loop and --local are mutually exclusive. If both are passed, the skill errors out.

Quick Start

When a pull request needs review, ask Claude to:

• "Review this pull request"
• "Analyze the changes in this PR"
• "Check if these changes follow our constitution"
• "Review this PR locally" (uses --local mode)

Claude will automatically use this skill to perform comprehensive analysis.

Autonomous Execution

This skill runs fully autonomously without user interaction. When invoked:

• Do NOT ask for user confirmation at any step
• Do NOT pause between tasks or wait for approval
• Proceed directly through all workflow tasks
• Only stop if a critical error occurs (gh CLI not authenticated, invalid PR number, or API failure)

Execute the entire workflow from start to finish:

• Standard mode: Posts review to GitHub automatically
• Local mode: Creates task list for local fixes (no GitHub posting)

Operating Constraints

• GitHub CLI Required: Must have gh authenticated and be in a git repository
• Parallel Execution: Uses Task tool to launch 5 specialized auditors concurrently
• AI-Driven Analysis: All findings generated through semantic analysis
• Constitution Knowledge: Understands all 14 project principles (v1.0.0)
• Completion Criteria: See WORKFLOW.md Tasks 7 and 8 for success criteria

Specialized Auditors

Each auditor runs independently via the Task tool and returns findings in a structured format for aggregation.

Workflow

For the detailed workflow with parallel auditors, see WORKFLOW.md.

Analysis Guide

For analysis criteria, finding formats, and templates, see ANALYSIS_GUIDE.md.

Output

The skill directly updates the PR (not just recommends):

1. PR Title - Updated via gh pr edit --title with conventional commit format
2. PR Description - Updated via gh pr edit --body with structured markdown
3. Review Comment - Posted via gh pr comment with categorized findings