| name | loop-triage |
| description | Read AutoResearch run evidence and loop state, then propose one bounded next experiment or escalate. L1 is report-only. |
| user_invocable | true |
AutoResearch Loop Triage
Produce a concise, evidence-backed update for the AutoResearch experiment loop.
Read first
LOOP.md, loop-constraints.md, loop-budget.md, and STATE.md.
.planning/STATE.md for project blockers.
- The latest available local run manifest, logs, metrics, report, and Git provenance.
Output
Return exactly these sections:
High Priority — at most one problem/opportunity and why it matters.
Evidence — run id, paths, metrics, failure signature, and confidence.
Candidate Hypothesis — one falsifiable hypothesis and smallest validation.
Risk / Budget / Authority — what is safe now and what needs approval.
Verdict — KEEP, REJECT, or ESCALATE_HUMAN.
State Updates — exact proposed edits for STATE.md and an append-only log entry.
Rules
- L1 is report-only: never modify source, run remote training, or write to external systems.
- Separate model/data/system/hardware/environment failure layers before proposing code work.
- Do not infer success from a single log line; prefer the immutable run bundle.
- Stop after two identical failure signatures or when evidence is not comparable.
- Never propose an architectural overhaul during triage.
- When escalating, include evidence, attempted actions, options, recommendation, and estimated cost.