review-pr

• Use before creating PRs or when reviewing existing PRs
• Skip trivial changes, drafts, and bot PRs

1. Validate: Skip if closed, already reviewed, or trivial
2. Context: Get PR intent, changed files, and relevant CLAUDE.md files
3. Review (4 parallel agents):

- 2 agents: CLAUDE.md compliance (in-scope rules only, unambiguous violations)

- 2 agents: Bugs (compile errors, logic errors, security issues)

1. Validate issues: Secondary agent verifies each flagged issue
2. Post: Inline comments with reasoning, or "No issues found" summary

Flag only when certain:

• Code will fail to compile (syntax, types, imports)
• Code will produce wrong results (clear logic errors)
• Unambiguous CLAUDE.md violation (quote rule, verify scope)

Never flag:

• Style, quality, or subjective preferences
• Pre-existing issues or linter-catchable problems
• Potential issues dependent on inputs
• Explicitly silenced violations

Inline comments (use mcp__github_inline_comment__create_inline_comment):

• Issue description with reasoning
• Link to rule (CLAUDE.md) or code with full SHA: https://github.com/owner/repo/blob/[40-char-sha]/file.ts#L4-L7
• Committable suggestion for small fixes (< 6 lines); description for larger changes
• One comment per unique issue

Summary (if no issues):

```

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

```

• ❌ "This might cause issues" → ✅ "Variable x undefined at line 45, will throw ReferenceError"
• ❌ "Consider refactoring" → ✅ "Violates CLAUDE.md: 'Max 50 lines' (currently 67)"
• ❌ Multiple comments for same issue → ✅ One comment linking all locations

Every flagged issue should be something a senior engineer would catch.