dsh-mmroute

Text-only models such as DeepSeek and GLM come to a standstill the moment a task requires reading a screenshot, a Figma render, or any other visual input. dsh-mmroute removes that barrier: at every step of the agent stream it intercepts each image, asks a multimodal understander model of your choosing to transcribe it into structured, detailed text—verbatim OCR, chart data, layout, color, and explicit uncertainty declarations—then passes the conversation on to your text model for the final answer. There is no route switching, no extra CLI, no special prompt entries; your original model group stays exactly where it is.

What distinguishes this plugin is its end-to-end coverage and self-healing. Whether images arrive from user uploads, tool results, or MCP renders, they are all caught before dispatch. If a text model—or a gateway that silently rejects images—returns an image-related error, the router automatically reroutes through the understander and retries rather than failing outright. The answering model can also issue a vision_relook call on any previously transcribed image for a focused re-check, creating a commander-and-executor loop between the two models you selected. In long sessions, historical images are replayed as compact digests so context size stays under control.

Who is this for? Developers whose primary models are text-only yet who regularly need vision assistance; teams that want a fully offline transcription path via a local Ollama vision model; and users who prefer two hand-picked models collaborating over relying on a shared free endpoint or external login session. The only prerequisite is a multimodal understander model already present in your DSH configuration—cloud or local both work. The plugin ships no external services and reports no data.