git:20261001.ed53bdb to git:20261003.5ee4059

---

name: grind-review

description: "Review one /grind goal's worker output against the goal and correct it by editing. May investigate with read-only gh and web search; never builds, lints or tests."

triggers:

- When the grind workflow reviews a goal's worker output

disable-model-invocation: true

---

<!-- managed-by: clud -->

# /grind-review

Keep tool output bounded (narrow searches, capped lists, ranged reads, bulk on disk, no identical repeats): https://github.com/zackees/clud/blob/main/docs/architecture/bounded-output.md

- Read every file the workers touched and every file they should have

touched. Judge the change against the goal, not the task list; missing

pieces are yours to add.

- Check the RED -> GREEN regression exists for each code change and would

fail without the fix.

- Fix defects directly by editing. Leave no TODOs for the integrator.

- Apply the repository's written rules (CLAUDE.md / AGENTS.md, per-directory

READMEs), especially its registries and banned APIs.

- **"Not yet run" is never a reason to reject.** You cannot run anything,

and running tests, lint and builds is the integrator's job: it runs every

check before it pushes. Approve on reading, and list each check the

integrator must run (the focused test, a reproduction, a script) in

`must_verify`. The workflow hands `must_verify` to the integrator.

- Set `approved=false` only for a defect you found by reading and could not

fix by editing, and say exactly what and why. Never reject because nothing

has been run, or because your role can't run tests or lint.

Return the verdict through StructuredOutput.