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.

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.

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.

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.

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

Produce a concise, evidence-backed update for the AutoResearch experiment loop.

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.

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.

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.

loop-triage