| name | harness-optimization-audit |
| description | Report-first audit for AILI/ROSE harness routing, context cost, review fan-out, subagent parallelism, trigger noise, false PASS risk, and evidence-loss risk. Use when optimizing harness quality or cost; do not edit core harness controls and route approved changes to harness-evolution. |
Harness Optimization Audit
Purpose
Use this skill to audit whether the AILI/ROSE harness is spending attention, tokens, subagents, and review lanes on the right work.
The default output is a report with evidence and recommendations. It is not an implementation workflow and must not edit core harness controls.
Trigger
- User asks to optimize or audit agent/skill routing, trigger noise, subagent usage, review fan-out, or context/token cost.
- A workflow repeatedly over-triggers skills, misses required skills, dispatches too many or too few subagents, or loses evidence between lanes.
- Review-pipeline behavior appears too expensive, too shallow, or poorly matched to changed surfaces.
- ROSE or a subagent claims PASS despite skipped gates, stale evidence, partial lane output, or missing
Unverified fields.
- A harness change proposal needs cost/quality tradeoff analysis before
harness-evolution edits.
Near Misses
- User reports one concrete harness bug and wants localization: use
harness-issue-triage first.
- User has approved exact harness edits: use
harness-evolution after this report, not this skill.
- Ordinary product-code performance tuning: use
performance-optimization.
- Ordinary code review or coverage review: use
code-review-and-quality, review-pipeline, or coverage-review.
- False-success review for one implemented change: use
silent-failure-hunting unless the question is about systemic harness routing or cost.
Required Routing
- Owner lane:
subagent:review when delegated; ROSE may run it directly for a small report-only audit.
- Mode: read-only report first.
- Inputs: active user goal, relevant workflow files or docs, changed harness surfaces, observed failure or cost signal, current verification/review evidence, and any accepted OpenSpec or proposal constraints.
- Handoff: approved edits to core harness controls route to
harness-evolution; unlocalized defects route to harness-issue-triage.
Audit Dimensions
1. Trigger Precision
- Which skills or agents fire too often, too late, or not at all?
- Are descriptions narrow enough to avoid catch-all activation?
- Do near misses point to the correct adjacent skill?
- Does routing preserve
/ideate, /define, /build, and /ship as top-level lifecycle entrypoints?
2. Context and Token Cost
- Does the workflow load large skills, docs, logs, or references before they are needed?
- Can a shorter trigger, progressive disclosure, or evidence-only scout reduce context without losing safety?
- Are repeated reads, duplicate searches, or raw logs being carried across phases unnecessarily?
- Are completion reports concise while still preserving evidence,
Unverified, and skipped-check fields?
3. Subagent Parallelism
- Are independent research, review, test, security, and implementation lanes dispatched in parallel when safe?
- Are overlapping edit scopes, hidden dependencies, or shared mutable state forcing serial execution?
- Are subagent packets complete enough to prevent rework and missing evidence?
- Are lane outputs reconciled by ROSE instead of copied as final truth?
4. Review-Pipeline Fan-Out
- Are code, test, coverage, security, silent-failure, browser, E2E, and AI-regression lanes selected based on changed surface and risk?
- Are irrelevant lanes skipped with a reason instead of being always-on?
- Are relevant high-risk lanes missing because the change looked small?
- Does the fix loop rerun only the lanes affected by the fix?
5. False PASS and Evidence Loss
- Could stale logs, partial command output, missing exit codes, skipped tests, or stale screenshots be mistaken for fresh proof?
- Are
PASS, NEEDS_REVIEW, blocked statuses, and Unverified used consistently?
- Is evidence lost when moving between ROSE, subagents, review-pipeline, test reports, and final completion reports?
- Are clean reviews required to name residual risk rather than implying zero risk?
Report Workflow
- State the optimization question in one sentence.
- List the harness surfaces inspected and the evidence anchors used.
- Classify each issue as
cost, quality, routing, parallelism, fan-out, false-pass, or evidence-loss.
- Rank each issue by impact: Critical, Important, Suggestion, or Observation.
- Recommend the smallest safe change, or say no change when the current cost buys necessary safety.
- For every recommendation, name the owner path: no edit, docs-only, skill text, command prompt, agent prompt, installer/config, fixture/test, or
harness-evolution approval needed.
- Mark anything not proven by inspected evidence as
Unverified.
Output Contract
HARNESS OPTIMIZATION AUDIT: PASS | NEEDS_CHANGES | BLOCKED | PARTIAL
QUESTION:
- <routing/cost/quality question>
EVIDENCE:
- Surfaces inspected:
- Commands/logs inspected:
- Not inspected / Unverified:
FINDINGS:
- [Critical|Important|Suggestion|Observation][cost|quality|routing|parallelism|fan-out|false-pass|evidence-loss] evidence - issue - recommendation
TRADEOFFS:
- Safety kept:
- Cost reduced or not reduced:
- Risk introduced:
APPROVED EDIT PATH:
- No edit needed | route to harness-evolution | ask user for approval | run harness-issue-triage first
VERIFICATION PLAN:
- <fixture, smoke check, script, review lane, or manual/static check>
Boundaries
- Do not edit core harness controls from this skill.
- Core harness controls include ROSE/runtime rules, command prompts, agent prompts, skill routing, subagent contracts, memory policy, installer/config behavior, review gates, and harness docs.
- Do not weaken verification, security, review, or evidence gates only to reduce token cost.
- Do not add dependencies, mutate memory databases, change manifests, commit, push, or rewrite workflow history.
- Do not treat a cheaper route as better unless it preserves required safety, evidence, and lifecycle ownership.
Verification
Before returning an audit:
- Confirm every recommendation names an evidence anchor and a safe edit path.
- Confirm cost-saving suggestions do not remove required review, security, verification, or
Unverified reporting.
- Confirm approved core-harness edits are handed to
harness-evolution, not applied here.
- Confirm the report states what was not inspected and what remains
Unverified.