Multi Model Peer Review

peer-review — shared Claude Code skill

An independent **peer review** for specs, plans, and design docs, driven through two CLIs — OpenAI's Codex and a Google Gemini model — plus, optionally, an open-weight roster (DeepSeek, Moonshot's Kimi) on the Fireworks AI API. You finish a doc; the skill runs every available engine against it from the shell in parallel, merges their findings into one severity-ordered list tagged by which engine raised each, and applies the worthwhile ones — no copy-pasting between tools. If an engine is missing or down, the review still runs on the others **and says so**. The catches only one engine makes are the point: a flaw caught in a spec now is far cheaper than the same flaw in the code built from it later.

**It reports by default and edits only when you ask.** "Peer review this" gets you findings. "Review and fix it" gets you a revised file. Proactive reviews (below) always report only.

Editing is a normal targeted diff, not a guarded operation — it does **not** checkpoint for you, and it won't refuse on an uncommitted file, since the doc you just wrote is uncommitted by definition. There is one real check behind that: it hashes every file it might edit when the review starts and again before editing, so if a file moved underneath it — you edited in another window, a formatter ran — it stops rather than patching a version the engines never saw. That narrows the window rather than sealing it: the check and the write are separate steps, so it can't catch a change that lands between them. And it says nothing about work that was already sitting in the file when the run began; for that, and for any fix large enough to rewrite whole sections, it looks at `git status` and tells you what it would overwrite. **Commit or stash first if the file holds work you can't afford to lose.**

It's a *personal* Claude Code skill: `SKILL.md` plus one helper script — `roster-agent.js`, which drives the repo-aware roster route. Both live in the `peer-rev