Multi-agent PR review orchestrator. Sizes the PR first, then runs GitHub Copilot and parallel Claude subagents (adversarial, operational, and reference-comparison lenses) at the same time, triages the combined findings into fix-now / stale re-flag / file-as-issue buckets, and caps Copilot at two rounds because round three is mostly stale re-flags. Run with /code-review [pr-number-or-branch].
Built from a real deployment: across five review rounds on a Vercel-hosted Next.js PR, Copilot caught the line-level issues while the subagents caught the P0 architectural miss (maxDuration) Copilot missed in every round, plus a persistent-replay vector and four operational issues. If you have access to a paid multi-agent review service, this skill is the in-flight free version of the same idea.
See SKILL.md for the full workflow and patterns/subagent-prompts.md for the paste-ready subagent briefs.