🚦 The gates

Every run passes the same sequence of gates. Each is a real action with a real, inspectable result — no gate can be talked past, and none is simulated.

1. Red — a real failing test must exist. The agent writes slice.test.js and a real node --test run must prove it red.
2. Green — a real implementation makes it pass. At least one test must actually execute and pass — an exit code alone never counts, because the test runner exits clean even when it discovered zero tests. The runner parses the test-count summary.
3. Refactor — cleanup with the tests still green (re-run, re-proven).
4. Integrate — a real git commit in the sandbox repo. Only committed work moves forward.
5. Security & performance review — real model verdicts rendered on the real diff of the change, each naming its top risk.
6. Judged acceptance — a model judge compares the ask and acceptance criteria against the final diff and answers met / not_met with cited evidence. A not_met fails closed: a real repair agent must answer the judge before the run proceeds.
7. Staging — the workspace is cloned fresh (only committed files exist in the clone) and the tests re-run there. Depending on uncommitted state fails here, for real.
8. Canary — a stability bake: the suite must stay green across 3 consecutive runs in the clean copy. A real flake is a real wobble — and a wipe.
9. Rollout — a real smoke load: the delivered module is imported the way a consumer would import it, and must load cleanly. Nothing deploys from a local run, so no deployment is claimed.
10. Shipped— the result is handed to you: the delivered code is browsable in the App View on the world’s Live screen.