task: distill Ascend ModelAgent (model-agent) into source-grounded candidate skills #352

Current state: @wuhang2014 has proposed source-grounded candidate guidance from GitCode Ascend/model-agent. The Issue remains Proposed, with no assignee or accepted implementation delegation; accountable, implementing and acceptance owners are TBD. The initial review suggests a read-only revision/license/workflow inventory before skill text, then routing by existing owners; it asks about the canonical moving branch, document-generation scope and acceptance owner. Those are review suggestions/questions, not delegated work or a scope decision. Upstream revision, readset and license are not yet verified here. No runtime, model, performance or agent-efficacy result is established.
Decision needed: Maintainers/requester need to record whether the proposed intake scope is accepted and name its accountable, implementing and goal-acceptance owners, including routing to existing canonical owners.
Next action: Record the scope and ownership disposition before treating this proposal as delegated implementation. Any subsequent intake must retain the stated exact source-pin/readset and license-inspection requirements, candidate-only status, and not_run execution boundaries.

Goal and acceptance

• Problem / desired outcome: InferRSI's Ascend knowledge covers CannBot (fixed-stage report acceptance, via ascend-performance-analysis) and MindSpeed/MindIE provenance, but has no knowledge of Ascend ModelAgent (Ascend/model-agent, Python, master branch), Huawei's full-pipeline LLM deployment agent for Ascend NPU — adaptation checking, tuning, deployment, go-live stabilization and document generation. Distill its workflow/operating contracts into source-grounded candidate skills following the same intake pattern as the CannBot extension (issue [Tracking] SearchSpace distillation coverage and knowledge consolidation #91).
• Completion criteria:
  • Pin the inspected upstream revision from the canonical GitCode Git endpoint; preserve readsets, source license/notices (license currently shows "Others" — treat as UNKNOWN until inspected) and mechanism limits.
  • Cover the agent's stage/workflow structure, role instructions, registered callbacks/hooks and report-acceptance predicates, analogous to cannbot-stage-evidence.md: distinguish administrative completion from measured quality evidence.
  • Route mined knowledge to existing owners where applicable (e.g. ascend-performance-analysis, ascend-model-adaptation, draft-acceptance-issue) instead of duplicating procedures; new skills only where ownership is genuinely missing.
  • Integrate navigation, catalog discovery and both knowledge matrices; regenerate the skill index with tools/build_skill_index.py --check.
  • Preserve unverified boundaries: candidate state only, no performance claims, agent efficacy not evaluated, host/model/backend not_run.

Scope and constraints

• In scope / out of scope: Source-derived candidate guidance only. Out of scope: executing ModelAgent or its agent host, model downloads, numeric performance/quality claims, runtime engine integration, new CI lanes, default acceptance gates, changes to existing experiment evidence.
• Relevant dependencies, shared interfaces or limits: Source is hosted on GitCode (not GitHub); use the anonymous public Git endpoint and record exact commit/tree hashes as done for the canonical cann/cannbot pin. License field must be inspected before importing any content beyond reading; UNKNOWN license boundaries preserved as in the historical awesome-ascend-skills import.

Ownership and authorization

• Proposed by / actual requester: @wuhang2014 (fork maintainer) via agent session.
• Status: Proposed
• Accountable parent (Issue-level owner, if assigned): TBD
• Slice implementer(s) / acting contributor (if assigned): TBD
• Goal-acceptance owner: TBD
• Accepted decision or scoped delegation: Not yet granted