autopilotlisted

# autopilot Take up to four independent ready tickets, work each one properly, hand back a single integration PR. You stay the merge gate. **Attribution**: see `.source`. --- ## What this skill does NOT do It owns no ticket cycle. Every worker runs the canonical `ticket-runner` skill, whole, once per ticket. If you find yourself writing "then the worker runs the tests, then reviews…" inside autopilot, stop: that behaviour has one owner, and duplicating it means the two copies drift and tickets get a different standard depending on how they were started. It also never merges. Not with a flag, not when the checks are green, not when the diff is small. `mergeGate: human` is the only value the active program accepts. --- ## The layers **L0 — this skill.** Reads the connectors, hydrates observations, calls the CLI, applies the action plans it gets back, re-observes, and persists nothing itself except through the CLI. It is the only layer allowed to talk to the tracker. **The CLI — `void-harness autopilot`.** Pure computation: selection, review budget, lease protocol, run state, next action. It contacts nothing — no tracker, no GitHub, no git — and spawns no agent. Every input and output carries `schemaVersion: 1`. **The adapters.** Claude executes the orchestration plan with its Workflow primitive, Codex with native subagents. Both read the *same* `OrchestrationPlan` and return the *same* `WorkerResult`. An adapter that is missing, or a permission that cannot be proven,