fix(openai): serialize output schema activity input

[zawadzkidiana]

Summary

• Convert OpenAI AgentOutputSchemaBase instances into a concrete, data-conversion-friendly CadenceOutputSchema before scheduling the model Activity.
• Keep PydanticDataConverter unchanged instead of teaching the generic converter about one framework's abstract class.
• Preserve the existing 11-argument Activity payload shape while supporting both output_schema=None and non-null structured output.
• Keep output validation in the Workflow with the original OpenAI Agents schema; the Activity receives only the name, JSON schema, plain-text flag, and strictness needed for the provider request.
• Avoid repeated failed TypeAdapter construction by removing the converter-level fallback entirely.

Fixes #197.

Validation

• uv run pytest -q: passed; integration tests skipped unless explicitly enabled.
• uv run ruff check cadence tests: passed.
• uv run ruff format --check cadence tests: passed.
• uv run --all-extras mypy cadence/ tests/: passed.
• Full 11-argument round trips pass for null and structured output schemas.
• The actual CadenceModel handoff is tested to send CadenceOutputSchema to the Activity.
• A credential-free worker-side OpenAIActivities.invoke_model test reconstructs a function tool, carries structured-output metadata, and calls a mocked model.
• Unsupported non-null arbitrary types still raise PydanticSchemaGenerationError; generic converter strictness is unchanged.
• Responses API and Chat Completions request formats are identical before and after transport in strict and non-strict modes.
• 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: 12/12 passed.
• The previously failing Recurring AI Watch OpenAI release test passed.
• The upgraded cadence-samples OpenAI environment compiled, passed Ruff, and passed null and structured-output boundary checks.

Limitations

• No paid-provider or Windows live matrix was run.
• This PR does not change downstream pins. A published client release is required before cadence-ai-samples can replace its SHA requirements.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
cadence/contrib/openai/cadence_model.py	93.75% <100.00%> (+93.75%)	⬆️
cadence/contrib/openai/cadence_output_schema.py	100.00% <100.00%> (ø)
cadence/contrib/openai/openai_activities.py	100.00% <100.00%> (+100.00%)	⬆️

... and 4 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[gitar-bot]

Gitar is not allowed to push to this forked PR. Please enable "Allow edits from maintainers" to allow Gitar to push.

[zawadzkidiana]

The remaining 3/4 objective is intentionally not implemented as written because it describes the earlier converter-level approach:

Decode a JSON null value for a nullable Activity parameter without constructing a Pydantic schema for the non-null member.

The revised implementation removes AgentOutputSchemaBase | None from the Activity boundary. CadenceModel converts it before scheduling, and OpenAIActivities.invoke_model now receives CadenceOutputSchema | None, a concrete serializable type that Pydantic supports.

As a result:

• Pydantic successfully constructs and caches the adapter for CadenceOutputSchema | None.
• A null value is validated normally by that adapter.
• No PydanticSchemaGenerationError is raised or caught.
• There is no repeated failed-adapter path to cache.
• PydanticDataConverter remains generic and unchanged.

So the user-visible outcome is covered, but the literal requirement to avoid constructing a schema is superseded: constructing a valid schema for the concrete transport type is now the intended behavior. The full 11-argument null round trip, non-null structured-output round trip, provider-format equivalence, strict non-null rejection, and credential-free worker-side Activity path are all tested.

[gitar-bot]

Code Review ✅ Approved 1 closed / 1 findings

🟡 Medium risk · Converting OpenAI output schemas into activity-transport metadata could disrupt structured-output model requests if serialization or reconstruction mishandles schema details.

Converts OpenAI AgentOutputSchemaBase into a serializable CadenceOutputSchema before the model Activity is scheduled, keeping PydanticDataConverter unchanged. Failed TypeAdapter builds are no longer cached, so they rerun on each null-schema call. No open issues remain.

✅ 1 closed

✅ Performance: Failed TypeAdapter builds are never cached, so they rerun on every null

📄 cadence/contrib/pydantic/_data_converter.py:41-43 📄 cadence/contrib/pydantic/_data_converter.py:62-68
self._type_adapter is lru_cache(TypeAdapter), and functools.lru_cache does not store results for calls that raise. So when the hint is something like AgentOutputSchemaBase | None, every decode of null (for example output_schema=None on each invoke_model activity call) runs Pydantic schema generation again, raises PydanticSchemaGenerationError, and then catches it. Before this PR that path failed outright. Now it succeeds but repeats the schema-generation cost on every call. A fix is to remember hints that failed (a small set or dict keyed by hint) and skip the adapter build for them when the value is null.

Verification Objectives 3/4 Repos 1 inspected · PRs 1 inspected

Functional validation 3 covered / 4 objectives

#197 3 covered / 4 objectives

Other objectives on this issue, possibly covered elsewhere:

◻️ Decode a JSON null value for a nullable Activity parameter without constructing a Pydantic schema for the non-null member

✅ Add a credential-free mocked model and tool integration check exercising the worker-side Activity boundary
✅ Keep strict validation unchanged for non-null values, so unsupported arbitrary objects are not silently accepted
✅ Round-trip all 11 arguments of OpenAIActivities.invoke_model through PydanticDataConverter with output_schema set to None

Rules No rules evaluated

Broader impact Repos 1 inspected PRs 1 inspected

Cross-repo 1 repository inspected 0 open impacts

cadence-workflow/cadence-samples Inspected

Cross-PR 1 PR inspected 0 open impacts

PR #196: feat!: Standardize Error Handling and Support Custom Error Reasons Inspected

_{For more verification details, comment gitar display:verbose.}

Options

Auto-apply is off → Gitar will not commit updates to this branch.
Display: compact → Showing compact verification results.

Comment with these commands to change the behavior for this request:

Auto-apply	Compact
gitar auto-apply:on	gitar display:verbose

_{Was this helpful? React with 👍 / 👎 | Gitar}

[shijiesheng]

Good catch!