---

name: operator-spec-maintenance description: "Guides ONNX operator-schema maintenance, function bodies, shape inference, reference ops, node tests, proto regeneration, and maintainer gates." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

ONNX Operator Spec Maintenance

Use this sub-skill when a task changes ONNX operator schemas, function bodies, shape inference, reference implementations, backend node tests, version-converter adapters, .in.proto sources, generated docs, or maintainer validation gates.

Workflow

1. Read the operator-change checklist before editing schema files.
2. Decide whether the requested behavior belongs in a new function, a new operator version, or a helper/refactor that preserves compatibility.
3. Edit source-of-truth files first: onnx/defs/<domain>/defs.cc, old.cc, *.h, .in.proto, reference ops, or node tests.
4. Use named shape-inference helpers rather than inline lambdas when practical.
5. Add or update reference implementations and node tests for user-visible behavior.
6. Regenerate docs/protos/coverage only after the source files are correct.
7. Run the focused Python/C++/lint gates that match the change.

Route to References and Helpers

• Read references/operator-maintenance.md for the edit checklist and file map.
• Read references/shape-inference-patterns.md for C++ inference helper patterns and test guidance.
• Read references/onnx-text-syntax.md for function-body and parser syntax conventions.
• Read references/troubleshooting.md before retrying a build, test, or generated-file failure.
• Run scripts/operator_change_checklist.py --help for a task-specific checklist summary.

Scope Boundaries

• Use model-authoring for ordinary graph/tensor/model creation.
• Use validation-and-conversion for checker/parser/printer/version-converter usage on existing models.
• Use reference-and-backend-tests when the task is only to compare outputs, not change the spec or its implementation.