dsh-model-cascade

In DeepSeek Harness, the model picker in the conversation input has always been a flat two-level list grouped by provider. Once you have three or more API route groups and a dozen-plus model IDs, scrolling through that list to find the right model gets tedious. Worse, choosing a model and adjusting the reasoning effort are chained together in the same cascade step, breaking the flow of a quick switch.

dsh-model-cascade replaces that with a clean three-level cascade: first the API route group actually configured in your harness (DeepSeek, SiliconFlow, Volcano Engine, and so on), then the model series (extracted automatically from the model ID), and finally the full model name. Tap the full name and the menu closes immediately, model selected, done. Reasoning effort, when the chosen model exposes its reasoning.efforts in the directory, lives on a standalone row at the top of the menu. You can adjust the effort level in isolation without re-walking the cascade, and a fixed footer always shows which route group and model you are currently running.

The plugin registers on the conversation.input.model slot at priority −1, overriding the built-in list while reusing the ui-model-selection modelDirectories service for data and submission. Selection state, connection-reset handling, composer blocking behavior, and theming tokens all stay consistent with the native experience in both light and dark mode. It is made for DSH users who run multiple route groups, deal with a large model catalogue, and fine-tune inference effort often—install, restart dsh web, and the cascade is yours.