--- description: "AILI harness-audit command generated from the backend-neutral canonical body." agent: rose subtask: false --- # /harness-audit User input: `$ARGUMENTS` Purpose: Run one bounded, report-first audit of a named harness routing, context-cost, review fan-out, parallelism, trigger-noise, false-success, or evidence-loss concern. Required behavior: - Resolve the target surface, question, available evidence, and report destination before assessing the harness. - Inspect only evidence relevant to the named concern and distinguish observed behavior, missing evidence, and proposed improvements. - Return evidence-grounded findings, risks, and candidates for a separately approved harness change; preserve the current canonical owner and operation gates. Hard stops: - Do not edit harness controls, invoke another process loop, dispatch a review swarm, or turn the audit into a lifecycle transition. - Do not treat an audit finding as implementation authorization, acceptance, verification, or a final verdict. - Do not write a durable report without an existing repository-local path or an explicit placement decision. Output contract: - Target, evidence scope, inspected sources, and explicit evidence limits. - Findings grouped by routing, cost, evidence, or safety impact, with candidate follow-up owners. - Report path or chat-only result, plus `Unverified` items and the next decision needed.