resonance-ops-goal

Invoke /goal with the outcome you want reached. The skill frames a goal contract, gets one approval, decomposes the work into slices, then loops build-verify-decide until the Definition of Done is proven by executed checks.

Use a bounded loop where caps live in code, not prose. The loop_state.py script enforces max slice attempts, a total iteration ceiling, and a stuck detector, and its STOP directives cannot be overridden by the agent.

Only an executed check counts: a real test run that is green, clean validators, an audit returning APPROVE, or a verified browser render. The model's own judgment that code looks correct is explicitly rejected as a done signal.

No. Shipping stays a separate human-approved step, and the loop always stops before one-way doors like deploys, destructive changes, or schema migrations. It presents verified evidence and waits for explicit approval.

After the slice hits its attempt cap, loop_state.py returns STOP_SLICE. The skill re-plans that one slice once with lessons from the failure; if it fails again, the loop stops and escalates with the real output instead of retrying forever.

Skip it when the request is a single step or the path is fully known, since a direct action or fixed prompt chain is cheaper. The loop is the most expensive option and is reserved for work needing feedback-driven decisions.