ascendc-port-orchestrator

Generate per-op `run_a3_reference.py` + `input_gen.py` + `manifest.json` for an arch22→arch35 migration workspace by parsing the upstream ops-nn op directory's `examples/test_aclnn_{op}.cpp` (aclnn signature) and `op_host/{op}_proto.cpp` (output shape inference). Eliminates the per-op manual-authoring step that blocks `phase_o25_a3_ref` from running on new ops. Use when Phase O2.4 needs to construct a missing A3 reference runner. Usage: /aog-a3-author {op_dir} {workspace} # parse op_dir, emit scripts to workspace

Phase O2.5 input_gen.py + edge dataset generator for the bundled orchestrator. Reads a source-architecture AscendC package or a differentiable forward spec, infers the case_gen SCHEMA (tensor_inputs, scalar_inputs, shape_derive, invariants), and emits a ready-to-run input_gen.py under `workspace/{op}/`. Running that script produces the edge_inputs.pt + manifest.json artifacts the workflow_critic enforces at `O2_5.B.*` — removing the "hand-write input_gen.py per op" burden that used to make orchestrators reach for `.workflow_exception_O2_5` waivers. Use when Phase O2.5 needs deterministic edge inputs for a new operator. Status: V1 (2026-04-24). Three templates (simple / fused / scalar-shaped) cover the patterns seen in ops #3, #5, #6, #11, #20, #24, #25. Complex cases (FFT op#23, variable-length lists) still need manual authoring but should be rare.

Review AscendC runtime findings and stage user-local c-tier entries; audit the release-owned bundled knowledge base without mutating it. Four modes: "update" (fast), "scan" (thorough), "validate" (single entry), "learn" (web scraping). Use when an operator run produces new knowledge or the AscendC KB needs maintenance.

Classify an op by reading its source (Python / PyTorch / AscendC) and emit `op_classification.json` with KB recommendations as the load-bearing output. Invoked by the Python orchestrator at Phase O1.7, or by a human for one-off inspection: `Skill(name="aog-op-classify", args="workspace/{op}")`. Use when Phase O1.7 needs operator classification and KB recommendations.

AscendC 算子双方法性能评估（profiler + wall clock）。适用于 op-gen 输出目录（output/{project}/src/kernels/{op}/）。 使用 torch_npu.profiler（V5: kernel_details.csv, AI_VECTOR_CORE/AI_CORE 含 TensorMove）采集 device 侧 kernel 时间， 同时使用 time.perf_counter() wall-clock 采集 host 侧端到端时间。生成双方法并排对比的 HTML 报告。 调用方式: /aog-perf-eval {output_dir}

Use when an arch22→arch35 migration can reuse an existing arch35 candidate. Scan, provenance-stage, build, measure, and learn from it without replacing the selected independent truth or customer-facing verification.

Generate or refresh a project-level REPORT.md for an output/{project}/ directory (cross-generation migration or backward-generation project). Wraps src/scripts/gen_report_tables.py (table injection via <!-- BEGIN-GEN:* --> markers) and the canonical 9-section structure defined in OUTPUT_PROJECT_LAYOUT.md §4. It is restricted to arch22→arch35 migration and backward-generation projects under `output/`. Use when an AscendC kernel project needs its report initialized, refreshed, or audited. Usage: /aog-report-gen {project_name} # refresh tables in existing REPORT.md /aog-report-gen {project_name} --init # create REPORT.md from template if absent /aog-report-gen {project_name} --audit # dry-run: check sections + freshness, no edits

Self-supervision skill — invoke to audit the current working session against recurring failure patterns user has had to correct across prior sessions. Goal: catch reward-hacking, priority drift, infrastructure bypass, and premature-conclusion smells BEFORE the user has to correct them again. Use when auditing an AscendC operator-generation session before a major decision. Status: MVP skeleton (2026-04-21). Full pattern catalog pending DEBT-031 cross-session retrospective analysis — until then, the checks below are seeded from the user's explicit feedback memories.

正向→反向 AscendC 算子生成入口。由一个可微 PyTorch 正向算子，自动生成其反向（梯度）AscendC 算子并在 NPU 上验证精度。可用自然语言指定目标芯片（a3/a5 或 arch22/arch35）。 触发：当用户需要为某正向算子生成对应反向算子时使用。

跨代际 AscendC 算子移植入口。把一个 AscendC 算子从来源架构移植到用户指定的目标架构/产品 （当前 arch22→arch35，如 910C/V220→950PR/V300）。用户用自然语言指定目标架构（arch35 / 950PR / A5 / SoC编号 / 代际皆可）；来源架构由代码分析自动识别。参数包括：1） 必选。待port的arch22算子实现目录 2) 推荐。KernelBench风格 (model.py 和test_case.json) 的算子golden与测试集合。

AscendC 算子轻量迁移 skill（ascendc-cross-gen-port 的无 golden 轻量入口）：把已有 DAV_2201（arch22）平台的 AscendC 算子工程按 Stage 0-5 阶段门禁改造迁移到 DAV_3510（arch35）平台，覆盖 L1 基础适配 / L2 RegBase MicroAPI 重写（含 AIC 低阶直跑评估）/ L3 SIMT 优化三层级判定与 Cube 类算子迁移（分形 ZZ→NZ、跨核同步协议），精度标杆由 agent 逆向源码自合成并与 A5 实测双向互检，不经编排引擎。当用户无 KernelBench golden 输入、或希望基于已有 910b/910_93 算子修改后快速迁移到 950/arch35 时使用；需引擎驱动的端到端自动移植（自动构建/精度/性能闭环与报告）请改用 ascendc-cross-gen-port。