Support for thought_signature in openai-compatible provider

[agamble-quora]

Description

Support was added to @ai-sdk/openai-compatible for retaining and passing thought signatures from extra_content in tool calls:

• Support Gemini Though Signatures for OpenAI compatibility vercel/ai#11590
• feat(openai-compat): add thoughtSignature handling for google models vercel/ai#11745

The version of this package that opencode is using is quite old. I tried upgrading the package and there are other errors.

Can this package be upgraded so that support for tools with thinking models will be supported?

Plugins

No response

OpenCode version

v1.1.19 (git / dev)

Steps to reproduce

Upgrade @ai-sdk/openai-compatible to 2.0.7 or later

Try tool calling with Gemini 3 Pro model using Poe as the provider.

• openai-compatible provider
• baseUrl: https://api.poe.com/v1

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• [BUG]: Gemini 3 Pro function calling fails - missing thoughtSignature support #4832: [BUG]: Gemini 3 Pro function calling fails - missing thoughtSignature support - This is directly related to the thought_signature preservation requirement for Gemini 3 Pro
• Gemini 3 issue : Function call is missing a thought_signature in functionCall parts #4481: Gemini 3 issue : Function call is missing a thought_signature in functionCall parts (closed) - Similar issue already reported
• Response terminates prematurely when using Gemini 3 via LiteLLM #6244: Response terminates prematurely when using Gemini 3 via LiteLLM - Related Gemini 3 tool calling issue
• @ai-sdk/openai-compatible max_tokens error for GPT 5.x #5421: @ai-sdk/openai-compatible max_tokens error for GPT 5.x - Related issue about @ai-sdk/openai-compatible package needing updates for newer models

The core issue appears to be that upgrading @ai-sdk/openai-compatible is needed to properly support Gemini 3 models with thought_signature preservation and other recent API requirements.

Feel free to ignore if none of these address your specific case.

[TaterDoge]

The same problem: thought_signature

[edenreich]

lol, why did they have to add thought_signature?

It's not an OpenAI compatible spec, I've looked on the official OpenAI specification it's not even there!

The issue is Googles "compatible" API

[renich]

I get this too often while using antigravity's gemini 3 pro models.

[requesty-JohnCosta27]

I'm also facing this issue. The latest openai-compatible package supports this. I'm wondering if Opencode could update the current package version.

[edenreich]

@requesty-JohnCosta27

I'm also facing this issue. The latest openai-compatible package supports this. I'm wondering if Opencode could update the current package version.

Hey, quick question, I'm curious if you can reference a link to where did you find it compatible, I was looking here:

https://github.com/openai/openai-openapi

Or more specifically here:

https://github.com/openai/openai-openapi/blob/manual_spec/openapi.yaml

I couldn't find any indication of thought_signature attribute in the spec and so I'm curious where do you look it up? thanks!
I'm having the same issue with Gemini Models in the Inference Gateway CLI, the API is basically broken since this attribute was added, I think they could at least make it optional until it's in the spec.

[requesty-JohnCosta27]

@requesty-JohnCosta27

I'm also facing this issue. The latest openai-compatible package supports this. I'm wondering if Opencode could update the current package version.

Hey, quick question, I'm curious if you can reference a link to where did you find it compatible, I was looking here:

https://github.com/openai/openai-openapi

Or more specifically here:

https://github.com/openai/openai-openapi/blob/manual_spec/openapi.yaml

I couldn't find any indication of thought_signature attribute in the spec and so I'm curious where do you look it up? thanks! I'm having the same issue with Gemini Models in the Inference Gateway CLI, the API is basically broken since this attribute was added, I think they could at least make it optional until it's in the spec.

I don't think it is on the OpenAPI, it's very ad hoc.

But it is on the openai-compatible AI SDK package:

Receiving the thought_signature

How openai-compatible adapts and forwards the thought_signature

[edenreich]

@requesty-JohnCosta27

I'm also facing this issue. The latest openai-compatible package supports this. I'm wondering if Opencode could update the current package version.

Hey, quick question, I'm curious if you can reference a link to where did you find it compatible, I was looking here:

https://github.com/openai/openai-openapi

Or more specifically here:

https://github.com/openai/openai-openapi/blob/manual_spec/openapi.yaml

I couldn't find any indication of thought_signature attribute in the spec and so I'm curious where do you look it up? thanks! I'm having the same issue with Gemini Models in the Inference Gateway CLI, the API is basically broken since this attribute was added, I think they could at least make it optional until it's in the spec.

I don't think it is on the OpenAPI, it's very ad hoc.

But it is on the openai-compatible AI SDK package:

Receiving the thought_signature

How openai-compatible adapts and forwards the thought_signature

Interesting, got it, I might need to implement it the same way.

[github-actions]

To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.