[Feature] Local Endpoint: support custom request parameters (e.g. thinking / reasoning_effort) to disable reasoning models' thinking mode

[yinsinvy]

Summary

Request: allow the Local Endpoint (OpenAI-compatible) engine to send custom
request-body parameters — ideally a free-form "extra parameters" JSON field in
Settings, or first-class support for thinking / reasoning_effort.

Background

Reasoning models with "thinking on by default" are unusable for inline ghost-text
completion, because every suggestion waits for a full chain-of-thought pass first.

The most relevant case today is DeepSeek:

• The current DeepSeek lineup is deepseek-v4-flash / deepseek-v4-pro
  (deepseek-chat / deepseek-reasoner have been removed from the API docs,
  so they are no longer a long-term option).

• V4 models default to thinking mode ON with effort=high.

• The API supports both modes, but disabling thinking requires a request-body
  parameter that Cotabby currently has no way to send:

  {
    "model": "deepseek-v4-flash",
    "thinking": { "type": "disabled" }
  }
  
  

(or reasoning_effort: "none")

Cotabby's Local Endpoint settings only expose Server URL / API Key / Model /
API Format, so there is no way to pass this through. Verified with a direct API
call: with "thinking": {"type": "disabled"} the same prompt returns in ~1
completion token with no reasoning_content — perfect for inline completion.

Proposed solution
Any of these would solve it (in order of flexibility):

Free-form "Extra request parameters" JSON field in Local Endpoint
settings, merged into the request body. Most future-proof — covers
thinking toggles, reasoning_effort, Ollama options, etc.
A simpler "Disable reasoning/thinking" checkbox that sends
{"thinking": {"type": "disabled"}} (DeepSeek) or
"reasoning_effort": "none" (OpenAI-style) for the endpoint.
Per-model default in the Model dropdown for known endpoints.
Why it matters
With this, Cotabby works out-of-the-box with DeepSeek V4 (one of the most
popular OpenAI-compatible providers) as a near-zero-memory alternative to the
bundled local models — a great fit for 16GB machines. Without it, every
suggestion on a default-configured V4 endpoint stalls for seconds.

Environment: Cotabby 0.6.2-beta, macOS 26.

Thanks — really enjoying the app otherwise.