autopilot-runner

autopilot-runner

Skill that backs /shannon:autopilot. Provides the retry loop logic, refusal parsing, and remediation prompt synthesis.

Behavior contract

Given a task and max-attempts N:

1. For attempt 1..N:
   • Invoke /shannon:cook with current prompt.
   • Read e2e-evidence/<run-id>/completion-gate/report.json.
   • If verdict == "COMPLETE": return success with cited evidence.
   • If verdict == "REFUSED":
     • Parse REFUSAL.md cited blockers.
     • Synthesize remediation prompt (next-attempt brief targeting only failing MSCs).
     • Continue.
2. If N attempts exhausted and still REFUSED: write final REFUSAL.md to plans/reports/autopilot-<run-id>-REFUSAL.md; return failure.

When to use

• User invokes /shannon:autopilot
• User says "run this autonomously until done"
• Long-running tasks where refusal-driven retry is preferred over manual oversight

When NOT to use

• User wants single-pass execution → use /shannon:cook directly
• Task is exploratory / no clear success criteria → use /shannon:loop instead
• Production deploy where refusal must escalate to a human gate

Examples

attempt-1 → cook → REFUSED ("evidence missing for journey: login-flow")
attempt-2 → cook with prompt "Re-run login-flow journey; capture screenshot at /login after submit" → COMPLETE

Iron rules

• Never invent COMPLETE verdict — must be cited by completion-gate.
• Never override REFUSAL.md — read it, act on its blockers, retry.
• No --force-complete flag. Ever.