dsh-model-extended

Each model row in dsh natively carries only four fields, and reasoning effort plus input modality are pinned at the connection level, so every model under the same endpoint shares the same set of off / low / high / max tiers. In practice some models honour only two of those tiers, some have no reasoning at all, yet the UI still presents the full set - picking one either gets a server-side rejection or silently does nothing. Input modality cannot be scoped per model at all. dsh-model-extended moves both settings from the connection level back onto each model row itself.

In the DeepSeek catalogue or the custom-provider editor, two chip groups appear beside the existing capacity fields: Reasoning efforts and Accepted input. Checking them declares which tiers and modalities a model actually supports. Out-of-range efforts are rejected at runtime before the request leaves, reported as UNSUPPORTED_REASONING_EFFORT, and accepted modalities flow into the model picker and request-preparation pipeline. A No reasoning toggle can explicitly disable thinking for a model. Untouched models keep their adapter defaults, so nothing is forced.

It suits anyone running dsh with a mixed fleet of models whose capabilities differ. If you have ever got a 400 error after selecting a reasoning tier for a smaller model, or want to hide the thinking UI from a pure-text endpoint, this plugin exists for exactly that friction.