What problem does it solve? AscendC operator development generates hard-won lessons (build errors, precision pitfalls, performance patterns) that get lost between runs. This Skill reviews runtime findings, deduplicates them against the bundled knowledge base, and stages validated c-tier entries without ever mutating release-owned content. ## Core Features & Use Cases - Four maintenance modes: "update" (fast per-op review), "scan" (full KB dedup and validation), "validate" (test a single known issue against the current environment), and "learn" (scrape AscendC best practices from hiascend.com). - Semantic quality gates: generalization checks, applicability scope tagging (SoC/CANN/op-class), evidence cross-checks against actual kernel code, and regression-risk detection before any entry is admitted. - Tiered safety model: bundled b-tier cards are strictly read-only; all new knowledge is staged as user-local c-tier intake JSON for the orchestrator to admit. - Use Case: After an AscendC operator run finishes, point the Skill at workspace/{op}/knowledge_update.md to review the findings, generalize the lessons, and emit a validated intake file. ## Quick Start Run the knowledge maintenance skill in update mode on workspace/myop/knowledge_update.md to review and stage new AscendC findings.
