[ENHANCEMENT] Add a LiteLLM reasoning override for local models #2199

Type

Enhancement

Problem / Value

I use LiteLLM to access local reasoning models, such as Qwen 3.8 27B. Zoo Code uses model-name detection to decide whether to send earlier assistant reasoning back to the model. Local model names and custom proxy aliases may not match that detection. Users need a manual override so they can configure these models without renaming proxy models or changing Zoo Code.

Context

Add a per-profile setting in the LiteLLM provider settings named "Send reasoning back to the model", with three choices:

• Auto: Use the existing model detection. Keep this as the default.
• On: Send available plain-text reasoning from earlier assistant messages, regardless of model detection.
• Off: Exclude earlier plain-text reasoning from requests, regardless of model detection.

This gives local-model users explicit control when automatic detection does not match their setup. It does not assume that every local model needs reasoning replay.

Constraints/Preferences

• Preserve the current behavior for existing profiles by defaulting to Auto.
• Keep the choice separate for each provider profile.
• Apply the choice to outgoing conversation history only. Keep stored reasoning visible in the chat.
• Do not use this setting to enable reasoning generation.

An implementation with tests and a settings control is already prepared. I would like to contribute it after issue approval.