server : add --reasoning-cache

[aldehir]

gpt-oss requires reasoning content from previous interactions to perform well.

This implements a reasoning cache, controlled by --reasoning-cache to store the reasoning content from tool call messages. On subsequent requests, the reasoning content is injected into the message. Since regular messages don't have an associated id, it will not add reasoning content to them. This aligns with the recommendations by OpenAI

To use, pass in --reasoning-cache 128.

Unfortunately, it won't work well when placed behind llama-swap with a ttl, as it will kill the process and the cache with it.

[victorb]

When I tried this last week, I think it helped a lot with preventing tool hallucinations. After all the work that went into ggml-org#15181 in the end, might something like this reasoning-cachestill be needed for better tool usage by GPT-OSS? There currently seems to be a bunch of conflicts I don't know how to resolve, so unfortunately cannot check myself if it improves it or not.

[aldehir]

@victorb I have rebased the changes here onto the latest upstream master branch. Feel free to use this in the mean time!

[aldehir]

@victorb https://github.com/aldehir/gpt-oss-adapter this might of interest to you. I think it'll take a while before clients start responding back with reasoning content. I don't like the idea of running yet another proxy, but it might be easier than continuously rebasing onto llama.cpp.