dsh-vision-bridge

Running a text-only model such as DeepSeek inside DSH means that any image attached to a message is rejected outright at the host layer. You get a single line of refusal and lose the ability to leverage the vision tools already available in the session. dsh-vision-bridge exists to close exactly that gap.

At its core is a three-layer pipeline. The admission layer patches model-info so the bridged model reports an image input modality. The pre-step layer materializes the image into the session workspace and injects a [vision-bridge] instruction carrying the file paths. The streaming-dispatch layer strips every image block, including those nested inside tool results, and replaces them with a placeholder before the text-only adapter serializes the conversation. A system-prompt rule also requires the model to call the configured vision tool (default vision_glance, fuzzy-matched against the actual toolset of the agent) and answer from its analysis, so the tool-call card and its result appear inline in the conversation.

Ideal for users who run DeepSeek or other image-incapable models in DSH but still want to perform image understanding and analysis through native vision tools such as glance, ground, or detect. The plugin works on WSL, Ubuntu, and native Windows, and all settings are persisted through the DSH settings service across restarts.