Missing extra_content.google.thought_signature

Gemini 3 API calls broken due to non-spec attribute

Currently, all API calls to Gemini 3 models are broken.

Google introduced an additional attribute that is not part of the OpenAPI spec, which causes incompatibilities for clients that strictly follow the spec.

I noticed that related issues were already opened and addressed in specific OpenAI SDKs, and the fix has been merged there. However, this was applied only to the Python SDK. I’m working with a Golang application, so this workaround does not help in my case.

This raises the question:
Should this attribute be added to the OpenAPI spec to keep it consistent with the patches already applied in the Python SDK?

Related issues:

• https://github.com/openai/openai-agents-python/issues/2137
• https://github.com/inference-gateway/inference-gateway/issues/229

I’m also not fully convinced this is the right long-term approach, and I’m open to suggestions.

At the moment, at least 9 providers successfully follow the spec, and only Google diverged by introducing this additional attribute. I could implement this directly in the Inference Gateway as an optional, Google-specific field and gate it behind provider detection - but that feels like papering over the problem.

From my perspective, this should either:

• be addressed in the official spec, or
• Google should make this attribute optional and avoid breaking OpenAI-compatible clients.

Happy to hear thoughts on the preferred direction here.

Current spec:

    ChatCompletionStreamResponseDelta:
      type: object
      description: A chat completion delta generated by streamed model responses.
      properties:
        content:
          anyOf:
            - type: string
              description: The contents of the chunk message.
            - type: 'null'
        function_call:
          deprecated: true
          type: object
          description: >-
            Deprecated and replaced by `tool_calls`. The name and arguments of a function that should be
            called, as generated by the model.
          properties:
            arguments:
              type: string
              description: >-
                The arguments to call the function with, as generated by the model in JSON format. Note that
                the model does not always generate valid JSON, and may hallucinate parameters not defined by
                your function schema. Validate the arguments in your code before calling your function.
            name:
              type: string
              description: The name of the function to call.
        tool_calls:
          type: array
          items:
            $ref: '#/components/schemas/ChatCompletionMessageToolCallChunk'
        role:
          type: string
          enum:
            - developer
            - system
            - user
            - assistant
            - tool
          description: The role of the author of this message.
        refusal:
          anyOf:
            - type: string
              description: The refusal message generated by the model.
            - type: 'null'

Note that there is nowhere an attribute reflecting on the SDK change extra_content.google.thought_signature nor there is an indication about extensions for a specific provider.

Start with the ChatCompletionStreamResponseDelta schema shown in the issue and compare it with the related Python SDK and inference-gateway issues. First establish whether the provider-specific thought_signature belongs in the official spec or requires an extension mechanism; done means the direction is agreed and the schema or documented handling is updated consistently.