dsh-reasoning-effort

Tuning reasoning effort across different models in DeepSeek Harness used to mean opening a terminal, hunting down the llm-pi-ai config file, hand-editing JSON, and restarting the Web service. dsh-reasoning-effort moves that adjustment right into the DSH Web UI so you can dial reasoning depth as naturally as turning a volume knob.

Two surfaces work in concert. A Settings page under Reasoning Effort walks you through every explicitly configured custom pi-ai model and lets you toggle which of the seven levels (off, minimal, low, medium, high, xhigh, max) each one exposes. All other model fields and provider-specific wire values are preserved, and saves go through DSH's settings service so you never edit a config file by hand. The second surface is a reasoning-effort slider embedded directly in the composer s model picker: select a model, then drag the handle, use arrow keys, or scroll to choose the level for the current session. Both surfaces read and write the same llm-pi-ai configuration, so what you configure in Settings is exactly what the slider offers. Note that the full seven-level set applies to custom pi-ai providers only; the DeepSeek native adapter recognises just off, low, high, and max.

If you run several pi-ai custom providers in DSH and switch reasoning depths frequently, this plugin keeps you in the conversation loop instead of bouncing between a JSON editor and the browser.