how to show thinking using local models?

[svalaskevicius]

I'm using local qwen3.5, and opencode shows reasoning / thinking output for the same queries, while the codecompanion chat window - does not.

local local_openai_fn = function()
  local base = require("codecompanion.adapters").extend("openai_compatible", {
    env = {
      url = "http://127.0.0.1:8000",
      api_key = "TERM",
      chat_url = "/v1/chat/completions",
    },
    interactions = {
        chat = {
          adapter = "llama.cpp",
        },
        inline = {
          adapter = "llama.cpp",
        },
    },
    schema = {
      model = {
        default = "cog",
      },
      temperature = {
        default = 0.4
      },
      top_k = {
        default = 40
      },
      top_p = {
        default = 0.85
      },
      stop = {
      },
      num_ctx = {
        default = 262144,
      },
      repeat_penalty = { default = 1.05 },
      use_beam_search = { default = true },
    }
  })

I also tried setting:

      display = {
        chat = {
          fold_reasoning = true,
          show_reasoning = true,

no effect

[olimorris]

I don't think the openai_compatible has any reasoning builtin, so you'd need to add that in yourself or try extending from openai

[svalaskevicius]

right, is there no easy way to use llama.cpp with all features on? :) I'd imagine this is a standard setup so many people would be using it?

[olimorris]

Defo search the discussions because I know llama has come up a lot and the community did share a complete adapter.

[svalaskevicius]

found it! see https://codecompanion.olimorris.dev/configuration/adapters-http#llama-cpp-with-reasoning-format-deepseek - it applies for qwen models too, not just deepseek

[svalaskevicius]

i.e. it is needed to add the handler to copy the reasoning content