Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Internal specialist skill for Analyze shader source, IR evidence, and suspicious fingerprints.. Use when `rdc-debugger` dispatches shader-ir work.
.claude/skills/haolange-shader-ir/SKILL.md| Model | Eval pass | Runs |
|---|---|---|
| gemini-3.6-flash | 100% | 7 |
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-08 | ✗→✓ | ▲ Improved | -65% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -37% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -48% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -44% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -44% | 0% |
当前文件是 Codex 的 role skill 入口。
该角色默认是 internal/debug-only specialist。平台启动后不会自动进入该角色；只有用户手动召唤 rdc-debugger 并由它分派时，才进入当前 role。
Read first:
common/skills/rdc-debugger/SKILL.mdcommon/skills/shader-ir/SKILL.mdcommon/config/platform_capabilities.jsonPlatform contract:
coordination_mode = staged_handofforchestration_mode = multi_agentlive_runtime_policy = single_runtime_single_contextownership_lease mediated broker actions硬规则：
artifacts/intake_gate.yaml 与完整 runtime broker artifacts 前，不得进入 live 调查。session_id / context_id / event_id 等 runtime handle。Do not use this platform template before copying top-level debugger/common/ into the platform-local common/. Runtime case/run artifacts and reports are written under the platform-local workspace/.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 23,549 | 7,160 | -70% | 1 | 1 | 0% | 3,491 | 804 | -77% | 0 | 0 | — |
case-02 | fail→fail | 16,227 | 27,866 | +72% | 1 | 1 | 0% | 2,403 | 5,284 | +120% | 0 | 0 | — |
case-08 | fail→pass | 12,195 | 2,187 | -82% | 1 | 1 | 0% | 1,988 | 689 | -65% | 0 | 0 | — |
case-03 | fail→fail | 34,781 | 6,137 | -82% | 1 | 1 | 0% | 6,184 | 860 | -86% | 0 | 0 | — |
case-04 | fail→pass | 11,383 | 5,179 | -55% | 1 | 1 | 0% | 1,862 | 1,169 | -37% | 0 | 0 | — |
case-05 | fail→pass | 29,789 | 7,164 | -76% | 1 | 1 | 0% | 2,571 | 1,337 | -48% | 0 | 0 | — |
case-06 | fail→pass | 10,646 | 3,414 | -68% | 1 | 1 | 0% | 1,630 | 912 | -44% | 0 | 0 | — |
case-07 | fail→pass | 12,364 | 5,331 | -57% | 1 | 1 | 0% | 2,122 | 1,183 | -44% | 0 | 0 | — |
case-09 | fail→pass | 16,479 | 1,900 | -88% | 1 | 1 | 0% | 1,260 | 623 | -51% | 0 | 0 | — |
case-10 | fail→pass | 11,856 | 3,424 | -71% | 1 | 1 | 0% | 2,000 | 911 | -54% | 0 | 0 | — |
case-11 | fail→pass | 10,509 | 2,922 | -72% | 1 | 1 | 0% | 1,463 | 836 | -43% | 0 | 0 | — |
case-12 | fail→pass | 13,179 | 5,270 | -60% | 1 | 1 | 0% | 2,039 | 1,279 | -37% | 0 | 0 | — |
case-13 | fail→pass | 8,961 | 2,969 | -67% | 1 | 1 | 0% | 1,377 | 838 | -39% | 0 | 0 | — |
case-14 | fail→pass | 6,873 | 1,824 | -73% | 1 | 1 | 0% | 1,121 | 619 | -45% | 0 | 0 | — |
case-15 | fail→pass | 8,421 | 3,062 | -64% | 1 | 1 | 0% | 1,369 | 861 | -37% | 0 | 0 | — |
case-16 | fail→pass | 13,537 | 3,528 | -74% | 1 | 1 | 0% | 2,294 | 1,036 | -55% | 0 | 0 | — |
case-22 | pass→pass | 16,078 | 14,121 | -12% | 1 | 1 | 0% | 2,941 | 2,962 | +1% | 0 | 0 | — |
case-17 | fail→pass | 15,145 | 3,596 | -76% | 1 | 1 | 0% | 2,595 | 924 | -64% | 0 | 0 | — |
case-18 | fail→pass | 8,722 | 2,427 | -72% | 1 | 1 | 0% | 1,542 | 762 | -51% | 0 | 0 | — |
case-19 | pass→pass | 6,556 | 2,641 | -60% | 1 | 1 | 0% | 1,043 | 680 | -35% | 0 | 0 | — |
case-20 | pass→pass | 19,141 | 20,067 | +5% | 1 | 1 | 0% | 3,745 | 3,814 | +2% | 0 | 0 | — |
case-21 | pass→pass | 7,476 | 6,261 | -16% | 1 | 1 | 0% | 1,450 | 1,544 | +6% | 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 20 counted toward the lift figure. The other 2 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 +68 percentage points is the difference between those two pass rates over the 20 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.