Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Detect drift between observed repository state and durable GRACE 4 .grace current state, then help create reconciliation changes.
.claude/skills/osovv-grace-refresh/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -30% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -78% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -59% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -55% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -58% | 0% |
<skill> <purpose> Compare code, tests, and file-local markup against .grace/context, .grace/graph, .grace/verification, and active change scopes. This skill reports drift and proposes a normal GraceChangeSpec/GraceChangePlan; it does not silently mutate current state. </purpose>
<workflow>
grace lint --path <project-root> and grace status --path <project-root>..grace/graph/index.xml and routed graph documents for stale, missing, or orphaned M-* and DF-* anchors..grace/verification/index.xml and routed verification documents for missing deterministic V-M-* coverage, stale commands, or missing evidence markers.LINKS:, module contracts, tests, and log markers against durable anchors.grace-spec and grace-plan instead of direct mutation.</workflow> </skill>
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-23 | pass→pass | 15,135 | 12,200 | -19% | 1 | 1 | 0% | 2,438 | 2,365 | -3% | 0 | 0 | — |
case-22 | pass→pass | 11,687 | 6,746 | -42% | 1 | 1 | 0% | 2,022 | 1,418 | -30% | 0 | 0 | — |
case-20 | fail→fail | 13,955 | 2,174 | -84% | 1 | 1 | 0% | 2,260 | 595 | -74% | 0 | 0 | — |
case-21 | pass→pass | 14,399 | 11,189 | -22% | 1 | 1 | 0% | 2,547 | 2,527 | -1% | 0 | 0 | — |
case-03 | fail→fail | 7,255 | 2,695 | -63% | 1 | 1 | 0% | 1,275 | 701 | -45% | 0 | 0 | — |
case-01 | fail→fail | 11,565 | 23,227 | +101% | 1 | 1 | 0% | 2,037 | 728 | -64% | 0 | 0 | — |
case-02 | fail→pass | 12,104 | 5,481 | -55% | 1 | 1 | 0% | 1,799 | 1,254 | -30% | 0 | 0 | — |
case-04 | fail→pass | 14,958 | 2,955 | -80% | 1 | 1 | 0% | 2,704 | 591 | -78% | 0 | 0 | — |
case-05 | fail→pass | 9,489 | 2,591 | -73% | 1 | 1 | 0% | 1,694 | 691 | -59% | 0 | 0 | — |
case-06 | fail→pass | 9,836 | 2,858 | -71% | 1 | 1 | 0% | 1,642 | 747 | -55% | 0 | 0 | — |
case-07 | fail→pass | 9,335 | 2,798 | -70% | 1 | 1 | 0% | 1,543 | 653 | -58% | 0 | 0 | — |
case-08 | fail→pass | 7,297 | 5,185 | -29% | 1 | 1 | 0% | 1,243 | 1,183 | -5% | 0 | 0 | — |
case-09 | fail→pass | 8,221 | 5,813 | -29% | 1 | 1 | 0% | 1,547 | 1,269 | -18% | 0 | 0 | — |
case-10 | fail→pass | 7,599 | 3,056 | -60% | 1 | 1 | 0% | 1,190 | 795 | -33% | 0 | 0 | — |
case-11 | fail→pass | 10,889 | 2,218 | -80% | 1 | 1 | 0% | 1,918 | 654 | -66% | 0 | 0 | — |
case-12 | fail→pass | 13,380 | 7,434 | -44% | 1 | 1 | 0% | 2,255 | 1,524 | -32% | 0 | 0 | — |
case-13 | fail→pass | 13,309 | 7,146 | -46% | 1 | 1 | 0% | 2,346 | 1,616 | -31% | 0 | 0 | — |
case-14 | fail→pass | 7,870 | 2,702 | -66% | 1 | 1 | 0% | 1,526 | 682 | -55% | 0 | 0 | — |
case-15 | fail→fail | 19,114 | 1,789 | -91% | 1 | 1 | 0% | 3,268 | 510 | -84% | 0 | 0 | — |
case-16 | fail→fail | 13,703 | 6,183 | -55% | 1 | 1 | 0% | 2,100 | 1,249 | -41% | 0 | 0 | — |
case-17 | fail→pass | 12,059 | 8,260 | -32% | 1 | 1 | 0% | 2,020 | 1,595 | -21% | 0 | 0 | — |
case-18 | fail→fail | 10,931 | 5,165 | -53% | 1 | 1 | 0% | 1,678 | 1,123 | -33% | 0 | 0 | — |
case-19 | fail→pass | 10,871 | 4,836 | -56% | 1 | 1 | 0% | 1,637 | 954 | -42% | 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. 23 cases were attempted. The headline lift of +61 percentage points is the difference between those two pass rates over the 23 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.