Support empty thinking tags for Nemotron outputs

[MaxFrax]

Pydantic-AI thinking separator issue

Problem

nvidia/Nemotron-120B-A12B emits thinking content at the start of the message and only closes it with an end tag like </think>.
There is no opening tag in the content.

Pydantic-AI’s thinking splitter expects both a start and end tag. If a profile is configured with an empty start tag (""), the splitter loops indefinitely and the output is not parsed correctly.

Impact

• The agent’s final response includes the model’s internal thinking.
• There is no clean separation between ThinkingPart and TextPart for these models.

Example

Raw content starts with thinking and ends with </think>, followed by the user-visible response.

Desired behavior

Treat content from the beginning until </think> as thinking, and return the remainder as the user-visible response.

Profile configuration

profile = OpenAIModelProfile(thinking_tags=('', '</think>'))

model = OpenAIChatModel(
    MODEL_NAME,
    provider=OpenAIProvider(api_key=settings.openai_api_key, base_url=BASE_URL),
    profile=profile,
)