autopilot

name

autopilot

description

Continuous-operation mode for rockie — runs the experiment queue autonomously, using Zero-Cost Monitoring ($0 LLM cost during training), anti-burn exponential cooldown on failures, and ntfy to wake the human only when a decision is needed. Use when you want agent-driven research to proceed for days without human input. Not appropriate for unproven projects — only enable after you have a populated queue, budget ceilings, and a working launcher.

/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. .claude/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 .claude/scripts/autopilot_loop.sh > .claude/memory/autopilot.log 2>&1 &

Stop:

bash .claude/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)

Cost model (measured pattern from arXiv 2604.05854)

Conventional polling (LLM call per tick): ~$1.60/24h. rockie ZCM polling (LLM only on event): ~$0.08/24h.

What autopilot does NOT do

• Does NOT decide which experiment to run next semantically. It just pops priority order from the queue. If the queue is full of bad ideas, autopilot will dutifully run bad ideas. Quality of the queue is your responsibility (or the /queue-refill skill's).
• Does NOT interpret results. When ZCM returns, the journal node is still in running state. The agent (woken by ntfy on anomaly, or explicitly run after a clean completion) invokes /post-run-review to close it.
• Does NOT self-modify. No autonomous [LEARN] authorship, no CLAUDE.md rewriting, no rule changes. Those remain user/agent turns. (See docs/PORTS.md → GVU operator separation for the safe self-improvement roadmap item.)

Failure modes and their backstops

Failure	Backstop
Queue drains, no refill happens	ntfy tier-2 when queue goes empty + idle 5 min
Launcher crashes repeatedly	anti-burn cooldown (exponential), eventually ntfy tier-1
Training never dies (hang)	ZCM --max-seconds terminates with exit 3
Training dies with silent exit 0 but bad output	post-run-review catches (see journal failure_class)
Out of budget mid-run	budget-gate.sh blocks next tool call; autopilot notices launcher refuses next item
Ntfy server outage	notify.sh records NULL ntfy_id in notifications table; autopilot keeps running; humans lose realtime signal until back online
Clock drift / timestamp confusion	all hooks use datetime('now') SQLite-local; no dependence on wallclock math across machines

When to disable autopilot

• You're actively developing the launcher.
• You're exploring a new research direction (low queue signal).
• A fundamental harness bug is suspected (manual mode is faster to diagnose).
• You're on a budget near its ceiling and want to stop the line.