Autonomous run

You own the exit condition. Define done, then drive to it without stopping.

Steps

1. State the exit predicate. Before the first iteration, write done as a predicate you can check by running something: tests green, the repro fixed, all N PRs merged, pixel-diff zero. Done when the predicate names the command or artifact that proves it.
2. Pick the wake mechanism. Use your harness's loop or scheduling command if it has one. Otherwise:
   • When there is an event to watch (CI finishing, a merge, a ref advancing), start a subagent in the background that watches for the event and wakes you, with a long time-based heartbeat as the fallback.
   • When there is no event, use a fixed-interval heartbeat sized to when the result is worth checking again.
3. Iterate. Each iteration makes the smallest change the evidence justifies, checks it against the predicate, commits it if the predicate advanced, and discards it if it did not help. Revert a belt-and-suspenders change that "might help" instead of leaving it in. Sequence the work by the sequence-verifiable-units principle (call the Skill tool with "principles"). Verify each unit before starting the next, never in one batch at the end.
4. Own mid-run discoveries. Fix what you find along the way yourself, each through the skill that owns that kind of work:
   • A broken skill: Call the Skill tool with "writing-for-agents" and fix it.
   • A related bug, a flaky verifier, or a tooling failure: Call the Skill tool with "diagnosing-bugs".
   • Red CI, review noise and bot comments on a PR: call the Skill tool with "babysit" in drive mode. Its rules (no rebase or retarget, one fresh build for a flake, conflicts go to the branch owner) apply inside this run too, and it never merges.
   • Orphaned follow-ups and fixable drift: finish them with the skill that matches the work.