fix: set include_thoughts=False when thinking_budget is 0

[yaowubarbara]

Related Issues

Fixes #2845

Proposed Changes

When thinking_budget is set to 0 (to disable thinking), the include_thoughts parameter in ThinkingConfig was hardcoded to True. This caused Vertex AI to reject the request with a 400 INVALID_ARGUMENT error because you can't request thinking output when thinking is disabled.

The fix sets include_thoughts based on whether thinking is actually enabled (thinking_budget != 0).

How did you test it?

Updated the existing test_process_thinking_budget test to verify include_thoughts is correct for each scenario:

• thinking_budget=1024 -> include_thoughts=True
• thinking_budget=-1 (dynamic) -> include_thoughts=True
• thinking_budget=0 (disabled) -> include_thoughts=False
• thinking_budget=24576 -> include_thoughts=True

All existing tests still pass.

Checklist

• I have read the contributors guidelines and the code of conduct
• I have updated the related issue with new insights and changes
• I added unit tests and updated the docstrings
• I used one of the conventional commit types for my PR title

[julian-risch]

Hi @yaowubarbara Thank you for opening this pull request and also for creating the corresponding issue. It seems that your commit 103ca1a addresses the issue in the Google Gen AI integration but there are also other changes in the PR that change the Anthropic integration. Could you please create a separate PR for any changes to this other integration?

[julian-risch]

Ah, I just saw your other PR that already got merged: #2849
So could you please remove those changes from this open PR here? Thank you!

[fkdosilovic]

My two cents to the PR:

1. I believe that the most sensible option would be to allow for setting the include_thoughts in generation_kwargs explicitly by the user. Only boolean values should be allowed! If no value for include_thoughts is set, then the value should be set accordingly, taking into account the thinking_budget or thinking_level arguments.
2. The include_thoughts should be changed both for thinking_budget and thinking_level branches. The need for these two branches is due to the fact that Gemini 2.5 and Gemini 3 have different parameters for controlling thinking/reasoning effort of the models.

General logic for setting include_thoughts if thinking_budget or thinking_level is not set:

# for thinking_budget branch. this is added after the basic type validation of thinking_budget
include_thoughts = generation_kwargs.get('include_thoughts', None)
if include_thoughts is None:
    if thinking_budget == 0:
        include_thoughts = False:
    else:
        include_thoughts = True
        
# for thinking_level branch.  this is added after the basic validation of thinking_level
include_thoughts = generation_kwargs.get('include_thoughts', None)
if include_thoughts is None:
    if thinking_level == types.ThinkingLevel.THINKING_LEVEL_MINIMAL:
        include_thoughts = False:
    else:
        include_thoughts = True

Of course, @yaowubarbara, don't make any changes until core maintainers comment on my aforementioned suggestions.

[yaowubarbara]

Thanks for pointing that out! I've removed the Anthropic changes from this PR, now it only contains the Google GenAI fix commit. Those changes was already merged in #2849, sorry about the mix-up.

[yaowubarbara]

Thanks for the detailed suggestions @fkdosilovic! That makes alot of sense to allow users to explicitly set include_thoughts via generation_kwargs, and handling both thinking_budget and thinking_level branches separately. I'll hold off on making changes until the core maintainers weigh in on this approach.

[anakin87]

Related PR by @tstadel: #2737

[fkdosilovic]

Related PR by @tstadel: #2737

It seems to be a PR that added the thinking_level support, but it does not mention the subtle bug and edge case related to this PR.

[julian-risch]

Thanks for opening this PR!
I agree with @fkdosilovic 's suggestion shared in the comment above.
@yaowubarbara Could you please adjust the PR so that it works with series 3 models that use thinking_level and with series 2.5 models that use thinking_budget? In particular, if no value for include_thoughts is set by the user, then the value should be set automatically based on thinking_budget or thinking_level.

[yaowubarbara]

Hi @julian-risch @fkdosilovic, pushed the updates:

• Auto-derive include_thoughts for thinking_level branch (MINIMAL → False, others → True)
• Support explicit include_thoughts override via generation_kwargs in both branches
• Added tests for all cases

All passing locally. Ready for another look!

[anakin87]

Looks good!