LlmAgent rejects temperature= with extra_forbidden — no discoverable path to generate_content_config

[CarSanoja]

Summary

Coming from google-genai (where GenerateContentConfig(temperature=...) is routine), the natural first attempt with ADK is:

from google.adk.agents import LlmAgent

agent = LlmAgent(
    name="grader",
    model="gemini-3.5-flash-lite",
    instruction="...",
    output_schema=MySchema,
    temperature=0.1,
)

On ADK 2.7.0 this fails at construction time:

pydantic_core._pydantic_core.ValidationError: 1 validation error for LlmAgent temperature Extra inputs are not permitted [type=extra_forbidden, input_value=0.1, input_type=float]

The working form is generate_content_config=genai_types.GenerateContentConfig(temperature=0.1), but nothing in the error, the docstring, or the validation message points there. In our project this crashed every structured agent in a production pipeline and cost a debugging session to trace (model_fields introspection).

Suggested fixes (any one would do)

1. Accept temperature as a convenience kwarg and fold it into generate_content_config.
2. Or add a custom pydantic error message for common generation kwargs (temperature, top_p, max_output_tokens) pointing to generate_content_config.
3. Or document the mapping prominently in the LlmAgent docstring.

Environment

google-adk 2.7.0, google-genai 2.18.1, Python 3.13, macOS. Found while building a K-12 exam-grading fleet for the All Things Agentic hackathon.

[a2105z]

Taking this right now, don't duplicate.

[sanketpatil06]

Hi @CarSanoja, thanks for reporting this. We were able to observe the same behavior you described.
We see that PR #6837 is already open to address this. Could you please confirm if it resolves the issue on your end before we proceed with reviewing and merging it?

[CarSanoja]

Confirmed — #6837 resolves it on my end.

I installed the PR branch into a clean venv (pip install "git+https://github.com/a2105z/adk-python@fix/llm-agent-generation-kwargs", Python 3.13, google-genai 2.18.1) and ran the original repro plus the cases I would want covered before calling it fixed:

1. The exact repro from this issue — LlmAgent(name=..., model=..., instruction=..., output_schema=..., temperature=0.1) constructs, and agent.generate_content_config.temperature == 0.1. ✅
2. Other generation kwargs — top_p=0.9, max_output_tokens=256 land correctly on the config. ✅
3. Merge instead of clobber — passing generate_content_config=GenerateContentConfig(top_p=0.5) and temperature=0.2 keeps both (top_p 0.5 preserved, temperature 0.2 applied). This was my main worry with a kwargs-folding approach; the PR handles it. ✅
4. Fields that already have a dedicated argument get a directed error rather than silently folding: system_instruction=... raises "system_instruction must be set via LlmAgent.instruction, not via LlmAgent(system_instruction=...)". That message is better than what I asked for in the issue. ✅
5. No regression on genuine typos — temperatur=0.1 still raises extra_forbidden, so the strictness that makes LlmAgent safe is intact. ✅

Context on why this mattered in practice: I hit it while building a multi-agent grading pipeline where every agent is constructed with an output_schema, so the failure took out all of them at construction time, and the extra_forbidden message gave no hint that generate_content_config was the intended path. With this PR the natural first attempt just works, and the two cases where it shouldn't (owned fields, typos) fail loudly with useful messages.

Thanks @a2105z for the quick turnaround and @sanketpatil06 for confirming. Happy to re-test if the branch changes before merge.

[DeanChensj]

Thanks @a2105z for the quick fix, I just left a comment in your PR and I think a cleaner fix would be the option2 to have better error message.