The Context Course
You are coordinating Autolab experiments in this repo using Pi and pi-subagents.
Arguments:
- campaign:
$1if provided, otherwiserecent-master: follow-ups - gpu slots:
$2if provided, otherwise1 - max ideas:
$3if provided, otherwise3
Read:
AGENTS.mdREADME.mddocs/pi-subagents-guide.mdresearch/notes.mdresearch/do-not-repeat.mdresearch/campaigns/research/experiments/research/results.tsvresearch/live/master.jsonresearch/live/dag.json
Use Pi project agents:
plannerfor fresh experiment queuesreviewerfor rule and comparability checksresearcherfor paper-derived hypothesesreporterfor Trackio and Hugging Face Jobs statusmemory-keeperfor durable notes in the main checkoutexperiment-workeronly through reserved worktrees
Ask planner for up to the requested max ideas for the requested campaign. Reject duplicates or stale ideas before launching paid work.
Do not allow more active experiment-worker runs than the requested GPU slots. Create isolated workers with uv run scripts/pi_worker.py create ..., launch