---

name: pypto-op-plan description: Stage 1 requirement planning and triage. Covers complex-kernel workflow triage, API availability checking, structurally-similar example search, and planning file (MEMORY.md) initial setup. Golden preparation rules live in pypto-golden-generate §13 (mathematician's skill).

PyPTO Complex Kernel — Stage 1 Planning

Contents

File	Purpose
This file (SKILL.md)	Stage 1 triage + planning file setup workflow

For PyPTO API lookups during planning, read docs/zh/api/operation/index.md for the op category index and docs/zh/api/operation/pypto-<op>.md for per-op signatures. For semantic exploration, load skill pypto-api-explore or grep docs/zh/api/operation/.

Triage and Planning

Before any code generation, decide whether the complex kernel workflow applies.

Use this workflow if any of the following are true:

• backward kernel,
• recurrent or stateful kernel,
• scan-like kernel,
• multiple dependent math stages,
• multiple layout transitions,
• multiple reductions,
• nested loop structure,
• previous failed attempts,
• temptation to "just implement the whole thing now".

Step 0.1: API availability check

Decompose the formula into atomic operations and verify each one exists in PyPTO:

# Browse all PyPTO ops by category
cat docs/zh/api/operation/index.md

# Look up specific API signatures and constraints
grep -l '<op1>\|<op2>' docs/zh/api/operation/*.md
cat docs/zh/api/operation/pypto-<op1>.md

For any operation not found by name, search by category or semantically:

grep -l "<relevant_category>" docs/zh/api/operation/*.md
grep -rn "<formula step keyword>" docs/zh/api/operation/

Mark each formula step as supported, needs substitute, or unsupported before proceeding. Do not begin implementation for unsupported operations without a confirmed substitute.

Subskill delegation: requirements and environment (optional)

If requirements are unstructured: Read skill pypto-intent-understand (SKILL.md auto-loads) and follow its workflow to produce a SPEC.md. Use the SPEC.md output to populate the memory file's task summary and API map.

If environment issues arise: Read skill pypto-environment-setup (SKILL.md auto-loads) and follow its workflow to diagnose and fix CANN, torch_npu, or build-chain problems.

Subskill delegation: enhanced API exploration

After completing Step 0.1, if deeper constraint verification is needed (3-layer validation, reference implementation search across models/ and examples/), read skill pypto-api-explore (SKILL.md auto-loads) and produce an API_REPORT.md. Merge the API_REPORT.md findings into the memory's API map section.

Step 0.2: Find structurally similar examples

Search for existing kernels with similar structure:

grep -rn "<kernel type>" examples/ custom/ models/

Record the most relevant example path in the memory file. Note: examples/ is an API-usage reference only, not the production implementation standard — its simplified forms (e.g. pypto.Tensor([])) may violate lint / gates. When an example conflicts with lint, lint takes priority; do NOT cite an example to declare lint a false positive.

Mandatory planning file

Create custom/<operator_name>/MEMORY.md first, then continue implementation immediately.

The memory must contain:

• task summary,
• reference locations (including example paths from Step 0.2),
• API availability map (from Step 0.1),
• normalized golden status,
• module list,
• module contracts,
• frozen items,
• attempt history,
• integration status,
• optimization status,
• blocker list,
• design format compliance: which layers from skill pypto-op-develop's references/pypto-kernel-design-format.md (A–L) apply, which are omitted and why,
• active_module and modules_pypto_verified (see skill pypto-orchestration-manual's references/rules.md → Module-at-a-time enforcement).