dsh-model-auto-hot-switch

Not every model in the DeepSeek Harness accepts image input. When you ask the agent to read a screenshot, run an OCR job, or handle any image file, that step should go to a vision-capable model — while plain chat and coding should stay on the fast default you chose. Manually flipping between models mid-session is tedious and easy to forget, especially when image and non-image steps alternate inside a single turn.

dsh-model-auto-hot-switch makes the switch automatic and free. For each step (session, turn, step) it checks locally whether the messages carry an image block: if yes, the step is routed to the first discovered vision model; if no, your default model is left untouched. The classification is pure local code — no classifier call, no prompt rewrite, no duplicated request — so it costs zero extra tokens. It replaces only the provider and model fields through the harness's own agent/request extension point; messages, system prompt, tools, and session state are never modified. One click on the lightning icon at the bottom-right of dsh Web enables it, and the toggle persists in config.json across restarts.

Built for dsh Web users who mix image tasks (screenshots, OCR, image files) with regular chat and coding in the same session. As long as a vision-capable model exists in your provider catalog, the plugin auto-discovers it with zero configuration. In headless mode the route endpoints are skipped, but the hot-switch logic still runs from the persisted toggle.