support reasoning_content field in reasoning model api like deepseek-r1

[runningmaverick]

No description provided.

[juodumas]

One issue with this is that it would send the reasoning content back to the model which is undesirable.
I think one option is to ignore the reasoning content, but then the user does not see any activity until the model stops thinking. Another option is to add a <<< thinking section which is not sent back to the model on the next turn.

I am testing a <<< thinking section here juodumas-forks@b51b6a3

I didn't create a PR yet because I am not sure how many API providers use the reasoning_content field?

[runningmaverick]

One issue with this is that it would send the reasoning content back to the model which is undesirable. I think one option is to ignore the reasoning content, but then the user does not see any activity until the model stops thinking. Another option is to add a <<< thinking section which is not sent back to the model on the next turn.

I am testing a <<< thinking section here juodumas-forks@b51b6a3

I didn't create a PR yet because I am not sure how many API providers use the reasoning_content field?

👍 Your solution is better. As to the API providers' support, I think we can not wait for OAI to provide it.

[juodumas]

Agree that there is no need to wait for OAI, but so far I have not found any other provider besides deepseek using the reasoning_content field.

For example with Qwen/QwQ-32B, togetherai & groq simply return <think>...</think> in the content field.

So as the first step maybe you could slightly modify your PR to ignore reasoning_content but keep support for content: null because it is valid according to OAI and that's what is breaking deepseek-r1 with vim-ai. Of course I am not the maintainer, just thinking out loud :)

[juodumas]

I've found that sglang, vllm, llama.cpp all use the same reasoning_content field and opened the PR for vim-ai: #151

[runningmaverick]

I've found that sglang, vllm, llama.cpp all use the same reasoning_content field and opened the PR for vim-ai: #151

Great!