Autopilot — Autonomous Session Orchestration

What it does

Autopilot provides an opt-in controller that reads a Mode-Selector recommendation and repeatedly executes recommended sessions when confidence and resource checks pass. It enforces a strict set of pre- and post-session kill-switches (max-sessions, max-hours, resource-overload, low-confidence fallback, spiral detection, failed waves, carryover thresholds) and writes one atomic autopilot.jsonl record per run so telemetry and retros can join runs to session tails. The loop adapts concurrency per iteration based on live resource probes and peer counts, and always checks kill-switches after each completed session — it never aborts a running session.

When to use it

Use Autopilot when you want to "walk away" from routine, repeatable work (mechanical refactors, housekeeping, planned follow-ups) and trust the Mode-Selector confidence to pick safe modes. Invoke /autopilot with flags like --max-sessions, --max-hours, and --confidence-threshold for a single-run execution. It is explicitly opt-in and conservative: iteration 1 will fall back to manual if confidence is low; iteration 2+ will exit on sub-threshold confidence.

What's included

• Scripts: implementation lives in scripts/lib/autopilot.mjs (runLoop, parseFlags, writer).
• References: integration notes linking mode-selector, resource-probe, wave-executor, and session-end contracts.
• Instructions: command surface /autopilot [--max-sessions=N] [--max-hours=H] [--confidence-threshold=0.X] [--dry-run], dependency-injection contract (modeSelector, sessionRunner, resourceEvaluator, peerCounter), and detailed kill-switch semantics.

Compatible agents

Designed for in-process coordination where a coordinator agent (Claude or similar) invokes the session lifecycle. Works with agents that can run session-start/plan/wave-executor/session-end and that expose the Mode-Selector and resource probe primitives.