gpt-oss-120b and tools support

[jfayad]

I tried to use tool calls with gpt-oss-120b (hosted on Scaleway inference API) and it didn't work.

The same exact pipeline works well with llama-3.3-70b-instruct so I'm wondering if there's a few changes to perform on Langchain's side to support this model with tool calling ?

[jfayad]

After more carefull reading of the scaleway API docs, I found this: https://www.scaleway.com/en/docs/generative-apis/how-to/use-structured-outputs/#chat-completions-api-or-responses-api

The relevant part is the last paragraph right after the Note.

I'll do further tests using the Response API

[jfayad]

Basically Scaleway does not support function calling on gpt-oss-120b with the chat/completions endpoints, it only works with the new responses API.

The current implementation of responses in the library had a few flows though:

1. The reasoning messages were not relayed to the output (OpenAI docs explicitly state they should be)
2. The function_call message was sent back to the API with a status of completed, it needs to be nil for the LLM to pick up the function_call_output
3. The streaming logic was missing certain event types to :skip
4. LLM messages were being added with the role user instead of assistant
5. The top_logprobs was not supported at all