dsh-model-switch

A single DeepSeek Harness conversation can touch the main model, subagents, web search, image generation, and Plan execution at the same time, and each of those paths may call for a different model. Model Switch gives you an explicit provider-and-model choice for all six routes, so every call lands on the model you actually intended rather than silently defaulting somewhere else.

Its most distinctive principle is loud failure: an invalid route, an unpublished variant, or a missing provider errors out explicitly, and Model Switch never quietly substitutes another provider or model. Variant axes such as Fast, Context 128K / 1M, and Thinking are rendered entirely from entries the provider has published to the official DSH model catalog, and the Picker never synthesizes a model id the provider has not published. Plan Review keeps an execution-model draft independent of the Main default, committing it to the active session only after you confirm.

The plugin is a good fit for DeepSeek Harness users who switch between multiple providers, pin subagents to a fixed model, want to verify the execution model before approving a Plan, or need web search and image generation to follow a single consistent route. It works through provider-owned adapters, does not patch DSH Core, and does not manage provider credentials.