name

autoresearch-new-round

description

Arm a new autoresearch performance-research round for this repo's SASA kernel: pick/confirm a scope, scaffold autoresearch/runs/<date>_<slug>/ from the template (seeded with the already-closed leads from META.md so they're not re-tried), point the dispatcher KICKOFF at it, start the LOG, commit, and ready the /loop. Use when the user wants to start, arm, or set up a new autoresearch round / a new optimization investigation round on the molecular-surface kernel. NOT for executing an already-armed round (that is `/loop execute the autoresearch kickoff`).

Arm a new autoresearch round

Sets up a fresh, self-contained research round under autoresearch/runs/<id>/ and points the dispatcher at it, so /loop execute the autoresearch kickoff runs the new mission. See autoresearch/README.md for the directory conventions and autoresearch/META.md for the accumulated lessons.

$ARGUMENTS (optional) is the scope/mission for the round (free text, e.g. "whole-dataset batch throughput mode" or "tess-4 single-protein latency"). If empty or ambiguous, ask the user for it before scaffolding.

Procedure

Read the accumulated state FIRST (non-negotiable — this is the whole point of the harness):
- autoresearch/META.md — the permanently-closed leads (do NOT seed a round around them), the meta-lessons, and the machine/tooling gotchas.
- The most recent autoresearch/archive/*/REPORT.md — what the last round concluded.
- Confirm there is no active round already: check autoresearch/KICKOFF.md (the dispatcher). If it already points to an active runs/<id>/, stop and tell the user a round is in progress (offer to resume it via /loop or to conclude+archive it first).
Determine the scope. If $ARGUMENTS gives a clear scope, use it. Otherwise ask the user with AskUserQuestion — the productive out-of-prior-scope directions are listed in the dispatcher (batch/GPU throughput C4; single-protein latency C5; an area-only fast path; a new operating point). Reject a scope that only re-opens a closed lead (per META.md §1) unless the user explicitly wants to re-validate one — name the closure and confirm before proceeding.
Pick the round id: <YYYY-MM-DD>_<short-kebab-slug> (date = today from the environment's current date; slug = 2–4 words of the scope, e.g. 2026-07-01_batch-throughput).
Scaffold the round (these are mostly mechanical; do them, then verify):
- mkdir -p autoresearch/runs/<id>
- cp autoresearch/templates/KICKOFF.template.md autoresearch/runs/<id>/KICKOFF.md
- Fill in the template's <…> fields in runs/<id>/KICKOFF.md: Mission & scope (from step 2), Output contract (bit-exact vs the oracle by default; the tolerance envelope for opt-in variants), in/out of scope, Current state (seed the baselines + the relevant closed leads from META.md so the round starts informed), and a first ranked Leads list (3–5, each with a cheap load-immune kill-experiment). Keep the rules / per-phase workflow / stop conditions sections as-is.
- Create autoresearch/runs/<id>/LOG.md with a header and a Phase 0 — setup entry (mission, inherited state, the top lead, and "next: de-risk cheaply first").
- Create autoresearch/runs/<id>/results/ (empty; benchmark logs land here).
Point the dispatcher at the active round. Edit autoresearch/KICKOFF.md so its top clearly states the active round and instructs the loop to execute runs/<id>/KICKOFF.md (replace the "no active round / idle" status with an ACTIVE ROUND: runs/<id>/KICKOFF.md pointer + one-line scope). Keep the operating-rules footer.
Commit (type: subject, no Co-Authored-By; stay on the autoresearch branch; do not push): git add -A && git commit -m "autoresearch: arm round <id> — <scope>".
Hand off to the loop. Tell the user the round is armed and that running /loop execute the autoresearch kickoff will start it (self-paced: load-immune work anytime, timing gated on an idle box). Do not start the multi-hour loop yourself unless the user asks — arming and running are separate steps.

Guardrails

Read autoresearch/META.md closed-leads before seeding leads; never scaffold a round whose top leads are already closed.
Mechanical file ops (mkdir/cp/edit/commit) only — never edit a frozen surface/strategy class (root CLAUDE.md); the round itself instruments via test-scope clones.
One active round at a time. Conclude + git mv runs/<id> archive/<id> + reset the dispatcher to idle before arming the next (the round's own stop-condition workflow covers this).

Arm a new autoresearch round

Procedure

Read the accumulated state FIRST (non-negotiable — this is the whole point of the harness):

autoresearch/META.md — the permanently-closed leads (do NOT seed a round around them), the meta-lessons, and the machine/tooling gotchas.
The most recent autoresearch/archive/*/REPORT.md — what the last round concluded.
Confirm there is no active round already: check autoresearch/KICKOFF.md (the dispatcher). If it already points to an active runs/<id>/, stop and tell the user a round is in progress (offer to resume it via /loop or to conclude+archive it first).

Determine the scope. If $ARGUMENTS gives a clear scope, use it. Otherwise ask the user with AskUserQuestion — the productive out-of-prior-scope directions are listed in the dispatcher (batch/GPU throughput C4; single-protein latency C5; an area-only fast path; a new operating point). Reject a scope that only re-opens a closed lead (per META.md §1) unless the user explicitly wants to re-validate one — name the closure and confirm before proceeding.

Pick the round id: <YYYY-MM-DD>_<short-kebab-slug> (date = today from the environment's current date; slug = 2–4 words of the scope, e.g. 2026-07-01_batch-throughput).

Scaffold the round (these are mostly mechanical; do them, then verify):

mkdir -p autoresearch/runs/<id>
cp autoresearch/templates/KICKOFF.template.md autoresearch/runs/<id>/KICKOFF.md
Fill in the template's <…> fields in runs/<id>/KICKOFF.md: Mission & scope (from step 2), Output contract (bit-exact vs the oracle by default; the tolerance envelope for opt-in variants), in/out of scope, Current state (seed the baselines + the relevant closed leads from META.md so the round starts informed), and a first ranked Leads list (3–5, each with a cheap load-immune kill-experiment). Keep the rules / per-phase workflow / stop conditions sections as-is.
Create autoresearch/runs/<id>/LOG.md with a header and a Phase 0 — setup entry (mission, inherited state, the top lead, and "next: de-risk cheaply first").
Create autoresearch/runs/<id>/results/ (empty; benchmark logs land here).

Point the dispatcher at the active round. Edit autoresearch/KICKOFF.md so its top clearly states the active round and instructs the loop to execute runs/<id>/KICKOFF.md (replace the "no active round / idle" status with an ACTIVE ROUND: runs/<id>/KICKOFF.md pointer + one-line scope). Keep the operating-rules footer.

Commit (type: subject, no Co-Authored-By; stay on the autoresearch branch; do not push): git add -A && git commit -m "autoresearch: arm round <id> — <scope>".

Hand off to the loop. Tell the user the round is armed and that running /loop execute the autoresearch kickoff will start it (self-paced: load-immune work anytime, timing gated on an idle box). Do not start the multi-hour loop yourself unless the user asks — arming and running are separate steps.

Guardrails

Read autoresearch/META.md closed-leads before seeding leads; never scaffold a round whose top leads are already closed.

Mechanical file ops (mkdir/cp/edit/commit) only — never edit a frozen surface/strategy class (root CLAUDE.md); the round itself instruments via test-scope clones.

One active round at a time. Conclude + git mv runs/<id> archive/<id> + reset the dispatcher to idle before arming the next (the round's own stop-condition workflow covers this).