Bug: Local OpenAI-compatible model — supported_reasoning_efforts from /v1/models is ignored, no variants available via /variants

Koder version: 0.4.1
Provider: local OpenAI-compatible server (ExLlamaV3 / TabbyAPI) at http://localhost:8888/v1, configured in koder.jsonc via @ai-sdk/openai-compatible
Model: qwen3.8-27b-exl3-4.0bpw
OS: Linux

The local server advertises reasoning efforts

GET http://localhost:8888/v1/models returns:

{
  "id": "qwen3.8-27b-exl3-4.0bpw",
  "object": "model",
  "owned_by": "exl3",
  "max_model_len": 262144,
  "supported_reasoning_efforts": ["low", "medium", "xhigh"]
}

Observed behavior

Koder does not process the reported reasoning-effort variants for local models:

• /variants offers no variants for the local model, so the reasoning effort cannot be switched in-session.
• Every assistant message for the local model is recorded without a variant field, while platform models in the same sessions record e.g. variant: "medium".

Workaround currently in use

Hard-configuring the effort in koder.jsonc:

"models": {
  "qwen3.8-27b-exl3-4.0bpw": {
    "options": { "reasoningEffort": "medium" }
  }
}

Expected behavior

Koder should read supported_reasoning_efforts from the OpenAI-compatible /v1/models response and expose those efforts as switchable variants via /variants, the same way it does for platform models.