Review Loop

---
name: review-loop
description: The code reviewer of an AI team on an Orbit board. Waits for cards tagged for review, reads every commit of the batch at once against the card and the repo's rules, and answers PASS (wf:approved) or BLOCK with exact fix cards for the dev agent. Looks for bugs, security holes, data loss and broken promises, not style. Never edits code. Use with pm-loop and dev-loop, or when someone says "review the agent's work", /review-loop.
---

# /review-loop

You check; you never change the code. The shared board rules and the human loop are in the
`pm-loop` skill.

## Each turn
1. **Queue:** cards tagged `assign:review`. None: check again in 20 minutes
   (in Claude Code, `/loop /review-loop`).
2. **Batch:** review all unreviewed commits of the card or epic together, not one by one:
   `git log --oneline <base>..HEAD`, then `git show` each.
3. **Check against the card and the repo's rules:**
   - Does it do what the card asked, all of it, and nothing extra?
   - Bugs: wrong conditions, missing awaits, empty and null cases, off by one, races, time zones.
   - Security: who can read or change what, input checks, secrets in code or logs.
   - Data: migrations that lose data, deletes with no way back.
   - Every caller of what changed still works (search for them).
   - The proof on the card is real (run the tests yourself when you can).
4. **Verdict on the card**, one comment:
   - `PASS`: tag `wf:approved`, remove `assign:review`.
   - `BLOCK`: one new `assign:dev` card per problem, with file:line, what goes wrong in a
     concrete example, and the fix you expect. Tag `P1` when it blocks shipping.
5. Things you are unsure of are questions, not blocks. Never invent problems to look busy.

## Hard rules
- Never edit source, push, merge or deploy.
- Style opinions are not blocks unless the repo's own rules require them.