Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Public main skill for the RenderDoc/RDC GPU debugger framework. Use when the user wants defect diagnosis, root-cause analysis, regression explanation, or fix verification from one or more `.rdc` captures. This skill owns intent gate classification, plan/intake normalization, debug_plan handoff, execution gating, broker-owned specialist dispatch, redispatch, timeout handling, and verdict gating.
.claude/skills/haolange-rdc-debugger/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | -12% | 0% |
| case-01 | ✗→✓ | ▲ Improved | -50% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 8% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 25% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 55% | 0% |
RDC Debugger 主技能包装说明当前文件是 Codex 的 public main skill 入口。
平台启动后默认保持普通对话态。只有用户手动召唤 rdc-debugger，才进入 RenderDoc/RDC GPU Debug 调试框架。
rdc-debugger 是单一公开入口，但内部固定拆成两段：
Plan / Intake PhaseAudited Execution Phase当前阶段兼容宿主 Plan Mode，负责：
intent_gatereference_contract 生成与 readiness 判定debug_plan 编译当前阶段对外只允许要求用户补充自然语言事实、目标、现象、环境和可访问材料。
不得把以下内部对象直接当成用户输入要求：
reference_contractstrict_readymissing_inputsdebug_plan如果 execution 仍被 BLOCKED_MISSING_FIX_REFERENCE 阻断，必须明确说明缺的是哪类参考事实或基线材料，而不是让用户按内部字段名补 YAML。
当前阶段默认通过轻量 sub-agent 收敛输入：
clarification_agentreference_contract_agentplan_compiler_agent当前阶段不允许：
rd.* 分析只有当 debug_plan.execution_readiness = ready 时，当前 skill 才允许进入 execution。
固定顺序：
entry_gateaccept_intake，内部完成 capture import、case/run bootstrap、intake_gate 与 broker startuptriagedispatch/specialist loopskepticcuratorfinal_audit在 artifacts/intake_gate.yaml 通过，且 artifacts/runtime_session.yaml、artifacts/runtime_snapshot.yaml、artifacts/ownership_lease.yaml、artifacts/runtime_failure.yaml 产出前，不得进入 specialist dispatch 或 live rd.* 分析。
本 skill 只引用当前平台根目录的 common/：
common/skills/rdc-debugger/SKILL.mdcommon/config/platform_adapter.jsoncommon/config/platform_capabilities.json 的当前平台定义为准当前平台固定 contract：
coordination_mode = staged_handofforchestration_mode = multi_agentlive_runtime_policy = single_runtime_single_contexthook_ssot = shared_harness执行约束：
ownership_lease 下通过 broker action 请求访问 live runtime。triage + specialist + skeptic + curator 明确必须通过 sub-agent 执行。entry_gate、accept_intake、intake_gate、dispatch_readiness、final_audit 属于控制面，不 agent 化。waiting_for_specialist_brief 并持续汇总阶段回报；短时 silence 不得触发 orchestrator 抢活。BLOCKED_SPECIALIST_FEEDBACK_TIMEOUT 或等价阻断状态，而不是让 orchestrator 抢做 specialist live investigation。未先将顶层 debugger/common/ 拷入当前平台根目录的 common/ 之前，不允许在宿主中使用当前平台模板。
运行时 case/run 现场与第二层报告统一写入平台根目录下的 workspace/。
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-03 | fail→pass | 14,523 | 7,636 | -47% | 1 | 1 | 0% | 2,498 | 2,199 | -12% | 0 | 0 | — |
case-02 | fail→fail | 22,968 | 25,763 | +12% | 1 | 1 | 0% | 3,536 | 5,551 | +57% | 0 | 0 | — |
case-01 | fail→pass | 32,455 | 10,924 | -66% | 1 | 1 | 0% | 6,007 | 3,004 | -50% | 0 | 0 | — |
case-04 | fail→fail | 18,553 | 16,268 | -12% | 1 | 1 | 0% | 3,129 | 3,369 | +8% | 0 | 0 | — |
case-05 | fail→pass | 9,672 | 5,809 | -40% | 1 | 1 | 0% | 1,741 | 1,887 | +8% | 0 | 0 | — |
case-06 | fail→pass | 8,009 | 5,262 | -34% | 1 | 1 | 0% | 1,460 | 1,820 | +25% | 0 | 0 | — |
case-07 | fail→pass | 9,082 | 7,252 | -20% | 1 | 1 | 0% | 1,470 | 2,276 | +55% | 0 | 0 | — |
case-08 | fail→pass | 3,105 | 6,094 | +96% | 1 | 1 | 0% | 402 | 1,984 | +394% | 0 | 0 | — |
case-09 | fail→pass | 16,315 | 10,423 | -36% | 1 | 1 | 0% | 2,533 | 2,790 | +10% | 0 | 0 | — |
case-10 | fail→pass | 15,567 | 5,942 | -62% | 1 | 1 | 0% | 2,824 | 1,966 | -30% | 0 | 0 | — |
case-11 | pass→pass | 15,358 | 5,492 | -64% | 1 | 1 | 0% | 2,394 | 1,935 | -19% | 0 | 0 | — |
case-12 | fail→pass | 11,143 | 4,492 | -60% | 1 | 1 | 0% | 1,741 | 1,722 | -1% | 0 | 0 | — |
case-13 | fail→pass | 13,570 | 5,869 | -57% | 1 | 1 | 0% | 2,830 | 1,831 | -35% | 0 | 0 | — |
case-14 | pass→pass | 10,032 | 5,134 | -49% | 1 | 1 | 0% | 1,567 | 2,002 | +28% | 0 | 0 | — |
case-15 | fail→pass | 14,730 | 3,337 | -77% | 1 | 1 | 0% | 2,413 | 1,560 | -35% | 0 | 0 | — |
case-16 | fail→pass | 8,388 | 3,072 | -63% | 1 | 1 | 0% | 1,421 | 1,471 | +4% | 0 | 0 | — |
case-17 | fail→pass | 14,584 | 5,246 | -64% | 1 | 1 | 0% | 2,180 | 1,970 | -10% | 0 | 0 | — |
case-18 | pass→pass | 6,619 | 3,983 | -40% | 1 | 1 | 0% | 1,115 | 1,746 | +57% | 0 | 0 | — |
case-19 | fail→pass | 12,471 | 4,168 | -67% | 1 | 1 | 0% | 1,938 | 1,640 | -15% | 0 | 0 | — |
case-20 | pass→pass | 20,619 | 21,286 | +3% | 1 | 1 | 0% | 4,214 | 4,675 | +11% | 0 | 0 | — |
case-21 | pass→pass | 17,722 | 13,456 | -24% | 1 | 1 | 0% | 3,863 | 4,518 | +17% | 0 | 0 | — |
case-22 | pass→pass | 6,470 | 7,579 | +17% | 1 | 1 | 0% | 1,279 | 2,341 | +83% | 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. The headline lift of +64 percentage points is the difference between those two pass rates over the 22 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.