burtenshaw/multiautoresearch/autolab-hermes-delegation

delegate_task(
    goal="Propose up to 3 fresh Autolab experiments against the current local promoted master.",
    context="""Read AGENTS.md, README.md, research/notes.md, research/do-not-repeat.md,
research/campaigns/, research/experiments/, research/results.tsv, research/live/master.json,
and research/live/dag.json.
Return a ranked queue of 1-3 fresh experiments. Each must include:
- short title
- one-sentence hypothesis
- parent master hash
- exact single variable being changed
- expected upside
- reason it is not a duplicate
Do not run commands that mutate the repo. Do not propose multi-change ideas.""",
    toolsets=["file"],
    max_iterations=20,
)

delegate_task(
    goal="Review this Autolab plan or result for rule violations and comparability risk.",
    context="""Read AGENTS.md and the provided experiment details.
Prioritize:
- hard-rule violations
- stale-master risk
- duplicate experiments
- multi-change patches
- missing benchmark evidence
- incorrect submit or no-submit decisions
Return concise findings with exact file or evidence references.""",
    toolsets=["file"],
    max_iterations=20,
)

delegate_task(
    goal="Find up to 3 paper-derived single-change Autolab ideas that map cleanly to train.py.",
    context="""Read AGENTS.md, research/notes.md, research/do-not-repeat.md,
research/paper-ideas.md, research/results.tsv, research/live/master.json, and research/live/dag.json.
Use the repo's Hugging Face skills when useful. Reject ideas already present in code or already ruled out.
Return the smallest credible change to test for each idea and the main risk if it fails.""",
    toolsets=["web", "file", "skills", "terminal"],
    max_iterations=30,
)

delegate_task(
    goal="Summarize current Autolab fleet status and call out duplicate or stale active jobs.",
    context="""Use the repo reporter workflow:
- . ~/.autolab/credentials
- uv run scripts/trackio_reporter.py summary --max-jobs 25
- uv run scripts/trackio_reporter.py sync --project ${AUTOLAB_TRACKIO_PROJECT:-autolab} when needed
Treat Trackio plus HF Jobs metadata as the source of truth.
Do not edit repo markdown or code.""",
    toolsets=["terminal", "file", "skills"],
    max_iterations=25,
)

delegate_task(
    goal="Update the durable Autolab markdown after a worker completed.",
    context="""Read AGENTS.md plus research/notes.md, research/do-not-repeat.md,
research/campaigns/, research/experiments/, and the worker's final summary.
Update only the durable markdown in the main checkout.
Preserve:
- hypothesis tested
- parent master hash
- local val_bpb or failure state
- submit decision
- one short interpretation""",
    toolsets=["file"],
    max_iterations=25,
)