dsh-image-model-router

In a DeepSeek Harness agent workflow, one turn may need only text reasoning while the next carries a screenshot, a photo, or a diagram. Sticking to a text-only model discards image context; switching everything to a vision model pays extra compute on every plain-text turn.

dsh-image-model-router inspects each agent request: it walks nested tool-result content for images the same way the harness does, routes to a configurable multimodal model (default deepseek-v4-flash-vision-exp) when an image is found, and keeps the original model for text-only turns. The routing flag is keyed by the agent object in a WeakMap, so state can never leak across agents or turns.

Built for dsh users who mix text reasoning with image-heavy workflows such as debugging with screenshots, reading diagrams, or analyzing photos; the plugin handles model selection automatically without manual provider switches.