Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use an explicitly selected AGY runtime for
.claude/skills/boshu2-agy-native/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 12% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -70% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -31% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -36% | 0% |
| case-14 | ✗→✓ | ▲ Improved | -9% | 0% |
Use AGY only when the caller explicitly selects that runtime. Discover its live command surface with agy --help (and agy models for the current model set) before acting, and scope every session to the supplied workspace and packet.
Discovering the live command surface before acting works because AGY's CLI changes faster than any skill text: a remembered flag is a guess, while a freshly listed one is evidence.
The posture a run gets is chosen by flags, so name it explicitly:
agy runs interactively and prompts for each tool permission.--dangerously-skip-permissions auto-approves every tool call — use it onlywhen the packet's declared effects and the caller's authorization cover that blast radius.
--sandbox restricts the session's terminal access.agy -p / --print) is the sanctioned headless path and carriesa built-in --print-timeout (default 5m); a run that exceeds it is killed and reported as timed out, not as a result.
A reader who sets none of these gets AGY's interactive default, not a scoped run. Match the posture to the declared effects and disclose which one was used.
If agy is not installed or its command surface cannot be discovered, report the absence as a disclosed fact and stop. Do not fall back to another runtime, do not guess a command surface, and never route through claude -p.
Named failure mode — wrapper drift: invoking AGY through remembered syntax that silently changed, producing runs that look scoped but are not.
Anti-pattern: reusing one AGY session for both author and validator roles because starting a second session is slower. Corrective: keep the identities distinct; a shared session forfeits the fresh-judgment guarantee that makes the validator's evidence usable.
the core verdict directly.
and never become AgentOps phase, queue, or completion state.
claude -p through an AGY wrapper.Return evidence to the caller and stop. Installation, plugin mutation, and recurring scheduling require separate explicit authorization.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 8,744 | 17,849 | +104% | 1 | 1 | 0% | 596 | 2,196 | +268% | 0 | 0 | — |
case-02 | fail→fail | 14,272 | 17,802 | +25% | 1 | 1 | 0% | 2,706 | 3,505 | +30% | 0 | 0 | — |
case-03 | fail→fail | 15,505 | 5,768 | -63% | 1 | 1 | 0% | 1,972 | 811 | -59% | 0 | 0 | — |
case-04 | fail→pass | 14,443 | 4,531 | -69% | 1 | 1 | 0% | 1,140 | 1,282 | +12% | 0 | 0 | — |
case-05 | fail→pass | 19,191 | 2,210 | -88% | 1 | 1 | 0% | 3,022 | 914 | -70% | 0 | 0 | — |
case-06 | pass→pass | 14,709 | 6,645 | -55% | 1 | 1 | 0% | 1,901 | 1,091 | -43% | 0 | 0 | — |
case-07 | pass→pass | 12,401 | 2,262 | -82% | 1 | 1 | 0% | 1,925 | 792 | -59% | 0 | 0 | — |
case-08 | pass→pass | 6,715 | 3,024 | -55% | 1 | 1 | 0% | 956 | 1,066 | +12% | 0 | 0 | — |
case-09 | pass→pass | 11,534 | 4,275 | -63% | 1 | 1 | 0% | 1,825 | 1,296 | -29% | 0 | 0 | — |
case-10 | pass→pass | 12,248 | 3,755 | -69% | 1 | 1 | 0% | 2,043 | 1,135 | -44% | 0 | 0 | — |
case-11 | fail→pass | 12,061 | 4,517 | -63% | 1 | 1 | 0% | 1,773 | 1,230 | -31% | 0 | 0 | — |
case-12 | fail→pass | 14,881 | 4,866 | -67% | 1 | 1 | 0% | 2,033 | 1,298 | -36% | 0 | 0 | — |
case-13 | pass→pass | 9,208 | 3,764 | -59% | 1 | 1 | 0% | 1,414 | 1,135 | -20% | 0 | 0 | — |
case-14 | fail→pass | 7,804 | 2,783 | -64% | 1 | 1 | 0% | 1,054 | 957 | -9% | 0 | 0 | — |
case-15 | pass→pass | 10,131 | 2,949 | -71% | 1 | 1 | 0% | 1,487 | 938 | -37% | 0 | 0 | — |
case-16 | fail→pass | 12,234 | 3,632 | -70% | 1 | 1 | 0% | 1,733 | 1,066 | -38% | 0 | 0 | — |
case-17 | fail→pass | 13,934 | 3,336 | -76% | 1 | 1 | 0% | 2,234 | 992 | -56% | 0 | 0 | — |
case-18 | fail→pass | 32,531 | 2,060 | -94% | 1 | 1 | 0% | 5,533 | 808 | -85% | 0 | 0 | — |
case-19 | pass→pass | 11,497 | 2,098 | -82% | 1 | 1 | 0% | 1,650 | 850 | -48% | 0 | 0 | — |
case-20 | pass→pass | 2,991 | 2,733 | -9% | 1 | 1 | 0% | 494 | 867 | +76% | 0 | 0 | — |
case-21 | pass→pass | 4,221 | 3,890 | -8% | 1 | 1 | 0% | 680 | 1,118 | +64% | 0 | 0 | — |
case-22 | pass→pass | 2,049 | 2,850 | +39% | 1 | 1 | 0% | 274 | 892 | +226% | 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 21 counted toward the lift figure. The other 1 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 +36 percentage points is the difference between those two pass rates over the 21 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.