Start

/ar:start — begin an autoresearch run

Stand up the loop: dedicated branch, ./.ar/ state, a real harness, and a measured noise floor. Starting does not iterate — /ar:resume does that.

First action, always

Check for existing state before anything else:

ls -la ./.ar/ar.jsonl 2>/dev/null && git branch --show-current

If ar.jsonl already exists, this is not a fresh run. Print the status block and say so — offer /ar:resume to continue or an explicit new-segment start over. Never silently overwrite a run in progress.

Hard rules

1. Confirm before creating. If this fired on inferred rather than stated intent, describe what opening a loop would do and confirm it is wanted before step 2. Ordinary optimisation — making something faster, tuning a few hyperparameters, running a benchmark — is handled directly. A loop is for many measured iterations under a locked harness, and it is not free: it branches and commits.
2. Never act on a branch this run did not create. Create ar/<slug>-<DD-MM-YYYY>, carrying the current working tree across — uncommitted and untracked files included — then commit it immediately as the revert floor. Full sequence in ${CLAUDE_PLUGIN_ROOT}/references/protocol.md §0.
3. Never fabricate a benchmark. The harness is filled in by the operator, and the loop does not start until it emits a real number. An invented benchmark makes every number downstream of it fiction.
4. Locked harness. benchmark.sh, checks.sh, evaluator.py and the code emitting the metric are never edited, for the whole run. A score is only meaningful while the thing measuring it holds still.
5. State lives only under ./.ar/. Nothing is written outside the target repo, and no global or shared store is touched.
6. Defer execution. Print exact commands for long runs; do not launch training inside the session.
7. ./.ar/ is gitignored as .ar/* plus !.ar/final_report.md — which is also what keeps git clean -fd from eating the loop's own state.

Steps

1. Confirm the objective first — goal, metric name, minimize or maximize, target if any, budget. State it back and get agreement before touching anything. Nothing in this step has side effects, which is the point: a misfired invocation costs one sentence rather than a branch and a commit in somebody's repository.
2. Branch and snapshot — references/protocol.md §0. First side effect of the run; everything after it is recoverable from the snapshot.
3. Scaffold ./.ar/ from ${CLAUDE_PLUGIN_ROOT}/templates/ (ar.config.json, benchmark.sh, ar-loop.sh, plus checks.sh / evaluator.py if that contract applies). Write the agreed objective into ar.config.json.
4. Hand over the harness. Stop here and ask for benchmark.sh and the config to be completed. This is the one point where starting waits on work it cannot do itself.
5. Measure the noise floor — baselineRepeats baseline runs, reduced to baseline and noiseFloor per references/statistics.md.
6. Write the config header to ./.ar/ar.jsonl; seed research.md, worklog.md, ideas.md — see references/state-schema.md.
7. Print the status block and stop. Suggest /ar:resume.

Read on demand

Read a reference when the step that needs it is reached — not up front.

Completion status

End with a terminal status token as the last line of your reply — DONE, DONE_WITH_CONCERNS, BLOCKED, or NEEDS_CONTEXT — per references/completion-status.md. For /ar:start, NEEDS_CONTEXT is the expected outcome at the harness hand-over (step 4, waiting on benchmark.sh and the config); DONE once the noise floor is measured and the config header is written; BLOCKED if no clean baseline can be established.