dsh-vision-bridge

In DeepSeek Harness, text-only models (deepseek-v3, Qwen text variants, etc.) hit a hard wall: the session controller modality check rejects any prompt carrying an image, so users simply cannot attach screenshots, charts, or design files to a conversation. dsh-vision-bridge closes that gap by transparently extending the model input declaration inside the Cordis runtime, letting text models accept images and then auto-translating them into structured text descriptions before handing the turn back to the chat model.

Three layers of capability do the heavy lifting. First, a service-mode bridge: natively visual models get direct passthrough; everything else is routed to a configurable backend chain (dsh-catalog, OpenAI-compatible endpoints, Ollama, custom webhooks) with automatic failover, parallel racing, and circuit-breaker protection. Second, a high-performance LRU cache keyed on content hash eliminates redundant vision-API calls and saves token quota on repeated questions about the same image. Third, a suite of 26 ready-to-use visual tools spans offline Tesseract OCR, visual question answering, bounding-box grounding, UI layout breakdown, pixel differencing, SVG tracing, and more, covering the full pipeline from reading a single screenshot to comparing multiple images.

If you routinely work with screenshots, UI mockups, charts, or any visual material inside a text-only DSH conversation, this plugin is close to a must-have. Three processing modes (hybrid, llm, tools) let you dial in the level of automation without changing your workflow, giving a plain-text chat full multimodal understanding out of the box.