Fix Google ADK agents with output_schema

[devjoinedthechat]

What was changed

TemporalModel now replaces a type-valued response_schema on the LlmRequest with its JSON schema (TypeAdapter(schema).json_schema()) before scheduling invoke_model / invoke_model_streaming. Dicts and types.Schema pass through unchanged. The conversion runs on a copy after summary_fn, so the request ADK and callbacks see is not modified.

Why?

An LlmAgent with output_schema can't run under the plugin today. ADK's basic request processor calls llm_request.set_output_schema(agent.output_schema), which stores the Pydantic class itself in config.response_schema, and the activity input then fails to serialize:

pydantic_core._pydantic_core.PydanticSerializationError: Unable to serialize unknown type: <class 'pydantic._internal._model_construction.ModelMetaclass'>

That fails the workflow task, so it retries until the workflow times out. Reproduced with google-adk 2.4.0 (the locked version) and 2.9.2; ADK main still stores the type.

Sending the JSON schema doesn't change what the model receives. google-genai converts a model class through its JSON schema anyway: I compared the generated responseSchema for the Gemini API and Vertex with flat, nested, enum and recursive models and they're identical. For list[Model] the only difference is a title: "Placeholder" that google-genai adds for the generic alias. ADK's LiteLlm builds the same response_format from the dict.

If you'd rather handle this in _AdkPayloadConverter, I'm happy to move it there.

Checklist

1. Closes: no existing issue (small bug fix)

2. How was this tested:

   • New test_agent_with_output_schema runs an LlmAgent(output_schema=..., output_key=...) in a workflow with a local test model and checks the parsed result in session state. Without the change it fails with the error above until the workflow times out.
   • New unit tests for the conversion (model class, list[Model], dict passthrough) through the plugin's payload converter.
   • pytest tests/contrib/google_adk_agents: 55 passed, 5 skipped.
   • ruff, pyright, basedpyright, mypy and pydocstyle pass on the changed files.

3. Any docs updates needed? No. Added a changelog entry under Fixed.

[CLAassistant]

All committers have signed the CLA.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}