Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Manage `@metaharness/darwin` bench suites — `bench create <repo>` scaffolds a JSON suite from a repo's test corpus; `bench verify <suite.json>` checks suite well-formedness. Bench suites are the fixed evaluation corpora that `harness-evolve --bench <suite.json>` scores variants against, decoupling evolution from the repo's natural tests. Degrades gracefully when @metaharness/darwin is absent.
.claude/skills/ruvnet-harness-bench/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-09 | ✗→✓ | ▲ Improved | 1% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -43% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -62% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -50% | 0% |
| case-16 | ✗→✓ | ▲ Improved | -30% | 0% |
Surfaces metaharness-darwin bench <create|verify> — the supporting verb for harness-evolve --bench. Use when you want evolution scored against a fixed corpus (independent of npm test) so champion fitness is comparable across commits or across forks of the same harness.
npm test isflaky, slow, or undersized — scaffold a deterministic bench suite once, then evolve against it repeatedly.
bench verify the checked-in suite on every PR that touches it(cheap; ~5s).
the evaluation without losing comparability to the parent.
Implementation: scripts/bench.mjs.
--op create--repo path; reject if missing.metaharness-darwin bench create <repo> [--out <suite.json>].<repo>/.metaharness/bench/suite.json (chosen by upstream).{ input, expectedOutput, weight } tasksderived from existing test cases.
--op verify--suite path; reject if missing.metaharness-darwin bench verify <suite.json>.json{ "success": true, "data": { "op": "verify", "taskCount": 42, "wellFormed": true, "durationMs": 870 } }
| Code | Meaning | |---|---| | 0 | OK (or degraded — Darwin absent) | | 1 | --op verify and suite malformed | | 2 | Config error or upstream invocation failure |
When @metaharness/darwin is absent, emits the standard {degraded: true, reason: 'metaharness-darwin-not-available'} payload and exits 0.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 3,913 | 2,122 | -46% | 1 | 1 | 0% | 233 | 875 | +276% | 0 | 0 | — |
case-02 | fail→fail | 5,462 | 5,306 | -3% | 1 | 1 | 0% | 327 | 823 | +152% | 0 | 0 | — |
case-03 | fail→fail | 2,147 | 3,554 | +66% | 1 | 1 | 0% | 258 | 1,204 | +367% | 0 | 0 | — |
case-04 | pass→pass | 11,220 | 7,390 | -34% | 1 | 1 | 0% | 2,070 | 1,860 | -10% | 0 | 0 | — |
case-09 | fail→pass | 4,921 | 1,811 | -63% | 1 | 1 | 0% | 813 | 818 | +1% | 0 | 0 | — |
case-05 | pass→pass | 6,523 | 6,524 | +0% | 1 | 1 | 0% | 1,371 | 1,729 | +26% | 0 | 0 | — |
case-06 | pass→pass | 10,534 | 5,772 | -45% | 1 | 1 | 0% | 1,819 | 1,608 | -12% | 0 | 0 | — |
case-07 | pass→pass | 6,614 | 1,252 | -81% | 1 | 1 | 0% | 1,227 | 681 | -44% | 0 | 0 | — |
case-08 | pass→pass | 12,120 | 1,975 | -84% | 1 | 1 | 0% | 2,335 | 688 | -71% | 0 | 0 | — |
case-10 | fail→pass | 7,933 | 1,806 | -77% | 1 | 1 | 0% | 1,426 | 814 | -43% | 0 | 0 | — |
case-11 | fail→pass | 11,657 | 1,576 | -86% | 1 | 1 | 0% | 2,018 | 775 | -62% | 0 | 0 | — |
case-12 | pass→pass | 12,383 | 2,092 | -83% | 1 | 1 | 0% | 2,427 | 869 | -64% | 0 | 0 | — |
case-13 | pass→pass | 8,295 | 2,208 | -73% | 1 | 1 | 0% | 1,509 | 891 | -41% | 0 | 0 | — |
case-14 | pass→pass | 10,080 | 2,108 | -79% | 1 | 1 | 0% | 1,741 | 835 | -52% | 0 | 0 | — |
case-15 | fail→pass | 10,044 | 2,092 | -79% | 1 | 1 | 0% | 1,700 | 847 | -50% | 0 | 0 | — |
case-16 | fail→pass | 14,910 | 6,541 | -56% | 1 | 1 | 0% | 2,555 | 1,793 | -30% | 0 | 0 | — |
case-17 | fail→pass | 11,564 | 2,758 | -76% | 1 | 1 | 0% | 1,876 | 1,022 | -46% | 0 | 0 | — |
case-18 | fail→pass | 54,843 | 2,194 | -96% | 1 | 1 | 0% | 4,210 | 911 | -78% | 0 | 0 | — |
case-19 | fail→pass | 10,804 | 1,618 | -85% | 1 | 1 | 0% | 2,006 | 693 | -65% | 0 | 0 | — |
case-20 | fail→pass | 8,950 | 1,553 | -83% | 1 | 1 | 0% | 1,533 | 765 | -50% | 0 | 0 | — |
case-21 | fail→pass | 14,330 | 3,769 | -74% | 1 | 1 | 0% | 2,373 | 1,286 | -46% | 0 | 0 | — |
case-22 | pass→pass | 4,208 | 1,530 | -64% | 1 | 1 | 0% | 658 | 660 | +0% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +45 percentage points is the difference between those two pass rates over the 19 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.