Filter out empty thinking parts from Gemini 3 when sending data to the frontend on integrations like Vercel

[psg2]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

Slack thread: https://pydanticlogfire.slack.com/archives/C083V7PMHHA/p1763583745967179

When running Gemini 3 Pro even if we don't include thoughts the model will send back signatures with empty thinking parts, and if we stream those to the frontend it will break existing messages in two parts that might break content display.

Example Code

google_thinking_config = ThinkingConfigDict(
        include_thoughts=False,
        thinking_budget=-1,
    )
    settings = GoogleModelSettings(
        google_thinking_config=google_thinking_config,
    )
    model = GoogleModel("gemini-3-pro-preview", provider=provider, settings=settings)
    agent = Agent(model)

    response = agent.run_stream_events(user_prompt="Hello, how are you?")
    async for event in response:
        print(event)


---

PartStartEvent(index=0, part=ThinkingPart(content='', signature='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', provider_name='google-vertex'))
PartEndEvent(index=0, part=ThinkingPart(content='', signature='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', provider_name='google-vertex'), next_part_kind='text')
PartStartEvent(index=1, part=TextPart(content="I'm doing well, thank you for asking! How can I help you today?"), previous_part_kind='thinking')
FinalResultEvent(tool_name=None, tool_call_id=None)
PartEndEvent(index=1, part=TextPart(content="I'm doing well, thank you for asking! How can I help you today?"))
AgentRunResultEvent(result=AgentRunResult(output="I'm doing well, thank you for asking! How can I help you today?"))

Python, Pydantic AI & LLM client version

Python 3.12
Pydantic 1.20
Gemini 3 Pro

[DouweM]

@psg2 Thanks for submitting the issue! #3481 describes the same underlying issue, but gets closer at the root so I'll close this as a duplicate of that. Will work on a fix tomorrow.