multi-pr-review_skill

Overview

This skill runs a multi-agent code review for a GitHub pull request by spawning three independent sub-agents that inspect the PR diff from different perspectives. It automatically validates and aggregates findings with reasoned analysis, deduplicates against existing comments, and posts a structured summary plus inline comments for important issues.

How this skill works

The skill fetches the PR diff and existing comments, then spawns three parallel sub-agents (Correctness, Code Health, UX) each receiving the same diff in a randomized file order to reduce ordering bias. Each agent outputs structured JSON issues; the parent process groups similar findings, validates them with reasoning (not simple voting), adjusts severities or drops false positives, determines a merge verdict (YES / NOT SURE / NO), and posts a summary plus inline HIGH/MEDIUM comments while collapsing low-priority notes.

When to use it

• Before merging a PR that touches critical logic, security, or UX surfaces
• When you want a broader review perspective than a single reviewer
• To catch maintainability regressions and dead infrastructure changes
• When you need automated, deduplicated feedback to avoid noise
• In CI or local workflows where a thorough pre-merge check is desired

Best practices

• Provide PR number or URL; otherwise the skill will infer from git context
• Keep PR descriptions clear so agents can assess intent and avoid false positives
• Ensure GITHUB_TOKEN is available for posting comments and fetching metadata
• Review the summarized dropped issues to learn why items were filtered
• Treat LOW items as suggestions; focus on resolving HIGH and MEDIUM findings

Example use cases

• Protecting a release branch by blocking merges with HIGH-severity security bugs
• Identifying dead migrations or unused infrastructure added in a PR
• Improving cross-cutting quality (consistency, accessibility, loading states)
• Automating reviews in a small team to surface maintainability hotspots
• Running nightly scans of open PRs to prioritize engineering attention