git:20260908.b45d86b to git:20260910.55adacd

6 added, 13 removed. Audit C to A.

-

---

- name: ascend-operator-debug

- description: Reduce an Ascend model-level failure to one torch_npu, ACLNN, or custom operator call, then validate explicit dtype, shape, layout, and eager/compile/graph cases against a reference implementation. Use for operator crashes, unsupported dtype or layout errors, shape-dependent numerical mismatches, or workspace API faults. Do not use for whole-model graph localization, multi-rank failures, performance benchmarking, or profiler analysis.

+ name: "ascend-operator-debug"

+ description: "Reduce an Ascend model-level failure to one torch_npu, ACLNN, or custom operator call, then validate explicit dtype, shape, layout, and eager/compile/graph cases against a reference implementation. Use for operator crashes, unsupported dtype or layout errors, shape-dependent numerical mismatches, or workspace API faults. Do not use for whole-model graph localization, multi-rank failures, performance benchmarking, or profiler analysis."

---

- # Ascend Operator Debug

- Canonical skill source:

- `.agents/skills/ascend-operator-debug/SKILL.md`

+

- Before using this skill:

+ # Ascend Operator Debug

- 1. Read the canonical skill file above.

- 2. Follow its routing rules, entrypoints, guardrails, and acceptance criteria.

- 3. Use the remote-dev companion tools (`remote_*` MCP tools; CLI fallback `remote-dev <hyphen-tool> ...` or `uv run remote-dev <hyphen-tool> ...`) for ordinary remote endpoint read/edit/bash/search/patch work.

- 4. Use this Claude project skill only for the domain workflow described by the canonical source.

+ Read `.agents/skills/ascend-operator-debug/SKILL.md` and only the references needed

+ for the current task. The canonical skill owns the workflow.