Commands

Command	What it does
/gru <request>	The main one. First time: the interview. Then grill, cut into cards, preview, run, hand back.
/gru qa <target>	Run QA only on a PR, branch or card.
/gru review <target>	Run the reviewer only.
/gru check <target>	QA, then the reviewer rules on its findings.
/gru run <stage> <target>	Any single stage: plan, tests, build, gate, qa, review.
/gru feedback [note]	Report something Gru itself should do better.
/gru status [feature]	A table of cards with stage, models, replans, tokens and observations.
/gru steer <card> <...>	Change one card's course (below).
/gru resume	Pick up an unfinished feature after a closed terminal.
/gru setup	Redo the interview.
/gru selftest	Prove each configured slot and harness works here.
/gru done	Wrap up, clean leftovers, archive.

/gru <request>

1. Preflight. No local.yml (or an older version)? The interview comes first, then your request.
2. Grill. Gru reads the repo docs, then asks at most 7 sharp questions per round, each with a recommendation. It may say "don't build this".
3. Cards. Each with objective, definition of done, size and files owned.
4. Preview. Cards, models, schedule, cost, budget. You edit, then accept.
5. Run. The pipeline per card, then the hand-back.

A small request (a question, a one-file tweak) is one XS card with no grill.

/gru status

| Card | Title            | Size | Stage      | swe / plan / qa / rev   | Replans | Tokens | Observations                     |
| C1   | Rate limit core  | S    | qa+review  | weak / mid / mid / mid  | 0       | 38k    | tests pinned, 5 DoD lines        |
| C2   | Per-key config   | M    | replan     | mid / strong x3         | 1/2     | 96k    | 2 blocking from reviewer         |
| C3   | Docs             | XS   | planned    | weak / mid              | 0       | -      | waits for C1                     |

Below the table: budget used, what runs now, decisions needed from you, calibration notes. Read-only.

/gru steer

/gru steer C2 use the existing Redis client, do not add a dependency
/gru steer C2 bump            # swe up one slot (or: bump strong)
/gru steer C2 resize L
/gru steer C2 rerun build
/gru steer C2 stop            # park it, keep the work

Free text becomes a binding note delivered to the card's next stage (or to a live tether session). Changing the objective or definition of done goes back through the planner and a re-pin of the checks. Steering never exceeds a cap without asking.

/gru resume, selftest, done

resume lists active features, reads the chosen log and re-attaches or restarts in-flight minions at their recorded slot (logs are local and git-ignored, so a fresh clone or container has none). selftest runs one trivial card per slot and harness pair and records the result under verified, including the sandbox checks (hidden $HOME, worktree-only writes, network, Docker image and services). done writes the wrap-up and cleans only what the feature created: it kills leftover minions, removes worktrees of merged or pushed branches, deletes merged local task branches (git branch -d, never -D), and keeps and lists anything unpushed.