Detects whether you're reviewing a plan or code, assembles the right panel of specialized reviewers based on what it finds (architecture changes, auth logic, UI work), then runs a bounded review-fix loop until clean or capped. Gate defaults are opinionated: P2 for plans, P3 for code, on the theory that local nits are cheaper to fix now than file as issues later. The terminal P3 sweep is clever but feels like a patch for the real problem, which is that fixing nits generates more nits. Still, it's the first review automation I've seen that actually tries to close the loop instead of dumping findings into a file and wishing you luck.
npx skills add https://github.com/camacho/ai-skills --skill review