signoff

name	signoff
description	Use for engineering work that must autonomously write a specification, obtain independent Council pressure, implement one bounded slice, prove it with executable receipts, obtain independent Roast review, and continue without goal drift or AI slop. The ./signoff state machine is authoritative.
disable-model-invocation	true

Signoff

Signoff is the reliability control plane around the coding agent. The user supplies one outcome in ordinary language. The agent follows the repository-local state machine until it emits DONE, STOPPED, PIVOT, or BLOCKED.

The model transcript is never authoritative. Run:

./signoff status
./signoff next

Then perform exactly the one legal action returned by ./signoff next.

Non-negotiable rules

1. Preserve the user's exact outcome in GOAL.txt; never paraphrase it in ./signoff start.
2. Do not edit any artifact after its hash is locked. A changed goal, charter, or spec requires the pivot gate and a fresh Council.
3. Work on one contract only. Do not touch files outside allowed_paths, exceed file/line budgets, or introduce adjacent cleanup.
4. A builder cannot count as a reviewer or lead judge. Context IDs must represent real fresh contexts; never manufacture quorum.
5. Claims such as “tests pass” are invalid until ./signoff verify executes the commands and seals EVIDENCE.json plus PATCH.diff.
6. Reviewers are read-only. Their suggestions do not become requirements; out-of-spec work stays OUT_OF_SCOPE.
7. UNKNOWN is not PASS. A disagreement can close only through a passed experiment, existing receipt, locked spec, or explicit user decision—not votes, confidence, consensus, or model brand.
8. Do not continue after a terminal state.

The one path

DRAFT

Complete the generated CHARTER.md and SPEC.json.

The charter states user-visible success, hard constraints, non-goals, stop/pivot conditions, and evidence standard. The spec uses stable requirement and acceptance IDs. Every must requirement maps to a falsifiable acceptance criterion.

Then run:

./signoff prepare-council

COUNCIL

Use the sibling council skill. Give every advisor the exact same goal, charter, spec, hashes, null hypothesis, and four canonical topics:

mission-value
feasibility
proof-sufficiency
scope-minimality

Advisors answer independently before seeing peers. Fill the generated COUNCIL.json. Material SUPPORT/OPPOSE conflicts require an evidence-based resolution. If independent contexts are unavailable, return INSUFFICIENT_QUORUM; do not imitate extra voices.

Then run:

./signoff lock

LOCKED / SLICE_DRAFT

Create one bounded slice:

./signoff prepare-slice

Complete CONTRACT.json with one real builder identity, requirement/acceptance IDs from the locked spec, allowed and forbidden path globs, production-file and changed-line budgets, and executable verification commands. Keep the slice as small as possible while remaining coherent.

Activate it:

./signoff slice

IMPLEMENTING

Make the smallest coherent change that satisfies the active contract. Do not change locked artifacts, generated receipts, or unrelated files. Then run:

./signoff verify

A failed command, scope breach, budget overflow, timeout, or verification-induced patch mutation is a real failure. Fix the cause and verify again; never edit the receipt.

VERIFIED / REVIEWING

Create sealed review packets:

./signoff prepare-roast

Use the sibling roast skill in fresh read-only contexts. Every reviewer answers every active acceptance ID with PASS, FAIL, or UNKNOWN against the same contract, evidence, and patch hashes. A separate lead judge accounts for every finding and conflict.

Then run:

./signoff roast

REVIEWED

Follow the deterministic result:

./signoff finish accepted
./signoff finish done
./signoff finish rework --root-cause "<stable root cause>"

DONE requires a final cumulative contract that covers every locked requirement and acceptance criterion. Repeated root causes or two review cycles without accepted progress force COUNCIL_REVIEW; implementation must pause until a pivot or stop decision.

User-facing progress

Do not dump internal JSON unless asked. Report only:

• the active bounded slice;
• executable evidence and whether scope stayed locked;
• unresolved disagreement or blocker;
• the final terminal result and residual risk.

A truthful stop is better than a plausible-looking completion claim.