Gemini-3-pro return Corrupted thought signature on multi step tool call

[Smilexuhc]

I am using the Vercel AI SDK with the OpenRouter provider to access Google Gemini-3-pro-preview model.

During a conversation, specifically when context history is involved (multi-step tool call), the API returns error with the message "Corrupted thought signature."

version: @openrouter/ai-sdk-provider 1.3.0
provider: google ai studio

[mattapperson]

Could you provide an example of this happening to help us reproduce and resolve the issue?

[SteffenDE]

I think the issue is that the reasoning_details in the providerMetadata gets replaced instead of merged, which can be seen in this demo repo:

https://github.com/SteffenDE/openrouter-ai-sdk-issue

If you run this and send a message, you can see that the reasoning chunk gets updated, but the details array is always overwritten. In the screenshot you can see that the reasoning text starts with "**Evaluating..." (the first chunk, later ones are appended to the text), but the details only contain the last chunk (starting with "**Composing...").

[robert-j-y]

This issue has been resolved. The original "Corrupted thought signature" error was fixed in v1.5.2 (Dec 9, 2025).

Regarding the streaming behavior @SteffenDE observed: During streaming, each chunk's providerMetadata contains only that chunk's reasoning_details (the delta). This is by design for real-time display. The accumulated reasoning_details are attached to:

1. The first tool call's providerMetadata.openrouter.reasoning_details
2. The finish event's providerMetadata.openrouter.reasoning_details

For multi-turn conversations, use response.messages directly - this contains the accumulated reasoning_details needed for Gemini to continue:

const firstResult = await streamText({ model, prompt, tools });
const firstResponse = await firstResult.response;

const secondResult = await streamText({
  model,
  messages: firstResponse.messages, // Contains accumulated reasoning_details
  tools,
});

An e2e test verifying this behavior exists at e2e/gemini/reasoning-multiturn.test.ts. If you're still experiencing issues on v2.0.4+, please open a new issue with a minimal reproduction.