autopilot

/autopilot — days-long autonomous operation

This is the "just keep going while I sleep" mode. It composes the pieces you already have (queue, ZCM, journal, calibration, budget, dead-ends, [LEARN]) into a single daemon that can drive an 8×H100 project for a week without human input.

Pre-flight checklist — DO NOT skip

Before starting autopilot, you MUST:

• Populate the queue. queue.py status shows ≥ target pending items.
• Set a budget. .codex/budget.toml exists with project + session ceilings.
• Have a launcher. autopilot.conf points at a LAUNCHER_CMD that takes a queue-item JSON on stdin, starts training in the background (nohup + &), and writes its PID + log paths.
• Configure ntfy. NTFY_TOPIC is set; you've subscribed in the app. Autopilot will wake you via ntfy on anomalies, ceiling-cross, and cooldown.
• A working dry-run gate. dry_run_gate.sh list shows sentinels for every training script the queue might reference.
• [DEAD-END] registry is current. Any director you've already killed is in the registry — autopilot queue-refill won't re-propose it.

The daemon

nohup bash .codex/scripts/autopilot_loop.sh > .codex/memory/autopilot.log 2>&1 &

Stop:

bash .codex/scripts/autopilot_loop.sh --stop

What the loop does

loop forever:
    item = queue.py next --json        # atomic claim, $0 LLM cost
    if item is None:
        if queue_under_target: ntfy(tier=2, "queue empty")
        sleep 300
        continue
    pid = LAUNCHER_CMD < item           # launcher backgrounds the job
    zcm.sh --pid $pid --log $log        # poll, $0 LLM cost, wake on anomaly
    match zcm_exit:
        0 (clean):       leave claimed — agent will close via post-run-review
        1 (crash):       kill process, increment failures
        2 (anomaly):     kill process, increment failures
        3 (timeout):     kill process, increment failures
    if consecutive_failures >= max:
        ntfy(tier=1, "cooldown, $seconds sleep")
        sleep $seconds                   # anti-burn
        seconds = min(seconds*2, max)