pr-review

name	pr-review
description	Pull request review workflow — orchestrates multi-agent review with confidence scoring, parallel analysis, and PR comment formatting.
user-invocable	false

PR Review Workflow

Overview

Multi-perspective PR review using parallel specialized agents with confidence-based filtering.

Review Pipeline

1. Gather context — PR diff, CLAUDE.md guidelines, related files
2. Launch parallel reviews:
   • Convention compliance (vs CLAUDE.md and coding standards)
   • Bug detection (focused on changes only, not pre-existing issues)
   • Historical context (git blame, related changes)
   • Test coverage assessment (are changes tested?)
3. Score findings — Each finding gets a confidence score 0-100
4. Filter — Only findings ≥80 confidence are reported (configurable)
5. Format — Output as terminal summary or PR comment

Finding Format

### {severity}: {title}
**Confidence:** {score}/100
**File:** {path}#L{start}-L{end}
**Category:** Bug | Convention | Security | Coverage | Performance

{description}

**Suggestion:**
{specific fix or recommendation}

Severity Levels

• BLOCKER (confidence ≥90): Breaks functionality, security vulnerability
• MAJOR (confidence ≥80): Significant quality issue, missing coverage
• MINOR (confidence ≥70): Style, naming, minor improvement
• NIT (confidence ≥60): Optional, author's discretion

Auto-Skip Conditions

Don't review if:

• PR is closed or draft
• PR is trivial/automated (dependabot, version bumps)
• PR already has a review from this session
• All changes are documentation-only (use doc-updater instead)

Comment Modes

• --terminal (default): Output to terminal for local review
• --comment: Post as PR comment via gh CLI
• --inline: Post inline comments on specific code lines