`reasoning_content` support for models other than DeepSeek (GLM, Kimi, etc.)

[hvppycoding]

Problem

When using tool-calling agents with Moonshot AI Kimi models, multi-turn conversations fail with a 400 error on the second turn.

These models include a reasoning_content field in their responses. When tool calls are also present, the reasoning_content must be included in the assistant message sent back to the API on the next turn — otherwise the API rejects the request.

Currently this behaviour is only implemented for DeepSeek models.

Affected models:

• Moonshot AI: kimi-k2, kimi-k2.5, and other Kimi reasoning variants — reasoning_content is required for multi-turn tool-calling; omitting it causes a 400 error.
• Zhipu AI: GLM-5, etc. — reasoning_content is not strictly required (no error without it), but including it is beneficial for model performance.

Question

Is there a recommended way to support reasoning_content for these additional providers?

[seratch]

Thanks for asking this! Currently, there is no way to customize the behavior, so I've added a new opt-in predicate function for the customization on the user side: #2670 Please note that merging this feature may take a while due to our other priorities.

[yang1002378395-cmyk]

Hi, I recently implemented GLM-4/GLM-5 support in another project (khoj-ai/khoj#1275) and can share some learnings about reasoning_content handling.

For Zhipu GLM models:

• GLM-4/GLM-5 returns reasoning_content in the response, but it's optional for subsequent turns (no 400 error if omitted)
• However, including it improves response quality for complex reasoning tasks
• The key is to check if the model response contains reasoning_content and preserve it in the conversation history

Implementation approach:

# When building assistant message for next turn
assistant_message = {
    "role": "assistant",
    "content": response.content,
}
if hasattr(response, "reasoning_content") and response.reasoning_content:
    assistant_message["reasoning_content"] = response.reasoning_content

For Kimi models:

• As noted in the issue, Kimi requires reasoning_content to be included in subsequent turns
• This is stricter than GLM's behavior

Happy to contribute a PR if this aligns with the project's roadmap. The milestone 0.14.x seems like a good target.