OpenAI model Activity cannot transport output_schema through PydanticDataConverter #197

Summary

OpenAIActivities.invoke_model accepts output_schema: AgentOutputSchemaBase | None. AgentOutputSchemaBase is an abstract OpenAI Agents class and is not a suitable Activity transport type for PydanticDataConverter.

With cadence-python-client==0.4.1 and OpenAI Agents 0.24.0, the Worker can fail while decoding the first model Activity input. A JSON null still causes Pydantic to construct a schema for the complete union, which raises PydanticSchemaGenerationError before the Activity or provider request runs. Non-null OpenAI schema objects also contain behavior that should remain in the Workflow rather than cross the Activity boundary.

Expected behavior

The OpenAI integration should transport null, plain-text, and structured-output metadata across the Activity boundary without weakening validation in the generic Pydantic converter. Provider request formatting must remain unchanged, and output validation must continue to run in the Workflow with the original OpenAI schema.

Resolution

Convert AgentOutputSchemaBase into a concrete, serializable CadenceOutputSchema in CadenceModel before scheduling the Activity. Transport only the metadata needed by the provider request:

• output type name;
• plain-text flag;
• JSON schema, when applicable; and
• strict JSON-schema flag.

PydanticDataConverter remains unchanged. The Activity uses the transported metadata for the provider request, while the Workflow retains the original schema for output validation.

Acceptance criteria

• Round-trip all 11 invoke_model arguments with output_schema=None.
• Round-trip a non-null structured output schema.
• Exercise the actual CadenceModel to Activity conversion path.
• Exercise the worker-side Activity boundary with a credential-free mocked model and tool.
• Preserve Responses API and Chat Completions request formatting in strict and non-strict modes.
• Keep strict validation unchanged for unsupported non-null arbitrary objects.
• Run OpenAI tests in the coverage job so the transport implementation is covered.

Validation

Implemented in PR #198.

• Unit tests, integration tests, Ruff, formatting, mypy, DCO, Gitar, and Codecov pass.
• OpenAI Agents 0.24.0 passes null, structured-output, and provider-format equivalence checks.
• A locally built wheel passed all 12 pinned cadence-ai-samples recipe/profile locations.
• The previously failing Recurring AI Watch OpenAI release test passed.
• The upgraded cadence-samples OpenAI environment passed compilation, Ruff, and boundary checks.

Downstream impact

OpenAI recipes should remain on their known-working client SHA until this fix is published in a client release. After that release, the SHA requirements can be replaced with the released version.

No paid-provider or Windows live matrix was run.