alloy-autopilot

name	alloy-autopilot
description	Use only when the user explicitly allows bounded unattended execution of an approved Alloy task.

Alloy Autopilot

Autopilot chains /execute and /verify under strict stop conditions. It does not add a scheduler, daemon, or extra persistent agent.

Entry Conditions

• User explicitly requests unattended or autopilot execution.
• .alloy/tasks/<task-id>/plan.md exists and is approved, or the user approves it in the current message.
• Verification commands are known.
• Risk level and iteration budget are understood.

Arguments

/autopilot <task-id> [--risk low|med|high] [--max-iters N] [--resume]

Workflow

1. Read plan.md and progress.md.
2. Confirm gates and stop conditions.
3. Run one /execute task at a time.
4. Use alloy-tdd and alloy-debug as required.
5. Append every iteration to progress.md.
6. Run /verify after implementation tasks are complete.
7. Stop and report status.

Risk Controls

Low risk: small scope, known tests, no secrets, no production services.

Medium risk: multi-file changes or generated artifacts. Require review before final DONE.

High risk: migrations, auth, billing, data loss, production infra, or secrets. Do not run unattended unless the user explicitly overrides and provides verification boundaries.

Stop Conditions

• Any required gate cannot be checked.
• A command fails twice for the same root cause.
• Scope expands beyond the approved plan.
• A required service, credential, or permission is missing.
• Max iterations reached.

Final Output

Return status, iterations used, files changed, commands run, gate state, remaining risk, and next safe command.