autonomous-looplisted

# Drive a work-item to proven done Take one work-item that already has a contract and carry it to done that is **proven, not asserted** - a real check went green, not a claim that it would. A feature and a fix run the *same* loop; only the front-end that produced the contract differs. This skill is the loop itself. It leans on other better-dev practices for the pieces around it and never re-does their jobs. This loop is a bounded goal-runner: a human initiates one work-item and the loop drives it to proven done. Recurring or unattended cadence work is a separate, opt-in layer better-dev doesn't yet build - not something this loop starts on its own. ## What it leans on - **The worktree** - `/worktree-branching` puts this work-item in its own git worktree off the integration branch, isolated from every other loop. - **The contract** - `/plan-grill` (feature) or `/diagnose` (fix) produces the done-criteria the loop grades against. - **Fresh workers and an independent reviewer** - `/orchestrating-agents` dispatches an isolated worker per task and the reviewer that grades it. This skill decides *what* to dispatch and *when*; it does not re-specify how dispatch works. - **The verdict** - `/review` reads the diff and distrusts the report. - **Overrides first** - read `.better-dev/overrides.md` before applying any default here; a project override wins. ## Two gates before the loop The loop is only as legitimate as what it starts from. Clear both, or it has nothing rea