dsh-vision-pro-bridge

DeepSeek-V4-Pro is a text-only model whose API rejects image bytes outright. Pasting a screenshot into a Pro session in DeepSeek Harness yields a hard "does not accept image input" error, and the official escape hatch is to downgrade the entire conversation to the weaker Flash-tier vision model. Yet many workflows demand Pro's coding strength while also needing it to "see" a UI mockup, an error trace, or a chart—having both without compromise was previously out of reach.

dsh-vision-pro-bridge solves this in roughly 250 lines. It registers a twin provider route (deepseek-vision-pro) that declares image input support. When an image is attached, the plugin first calls deepseek-v4-flash-vision-exp to transcribe it into structured text, then feeds that text—along with a stable local file path—to deepseek-v4-pro for the actual reasoning. Everything runs on a single DEEPSEEK_API_KEY; there is no Ollama, no openai or sharp dependency, and images are content-addressed and cached locally so repeated attachments skip re-transcription. With no image present, requests forward straight to Pro with zero overhead.

If you rely on V4-Pro daily for coding and log-reading, occasionally want it to inspect an error screenshot or a UI layout, and would rather not swap models, spin up a local VLM, or juggle a second API key, this plugin is the minimal, zero-dependency bridge. It is DeepSeek-only—one vendor, one key, one pipeline—with transparent, easy-audit code that is straightforward to fork and extend.