dsh-ollama-vision-bridge

In DSH's Web interface, attaching an image to a pure text model triggers an immediate rejection: "model does not support images." dsh-ollama-vision-bridge plugs exactly that gap. It applies an idempotent patch to DSH's session controller so that when a model step detects an image and the model lacks vision capability, the image is automatically handed to a local Ollama vision model (default qwen3-vl:8b) to produce a description, which is then injected into the same model step as a contextual message. One send, one answer, no extra round-trips, no cloud service, no API key. The image reference stays intact in your conversation history and the user message is never modified.

VRAM discipline is a design detail that matters in practice. The bridge request carries an Ollama keep_alive parameter, so the VL model lingers in memory only briefly after inference and then unloads on its own, without competing with your primary model for GPU memory. Configuration is hot-reloaded from the vision-bridge section of settings.yaml; delete the section and the bridge gracefully falls back to DSH's native rejection path. The entire plugin has zero npm dependencies, and its runtime is a minimal Cordis line that simply logs patch and config status at startup.

This plugin is for you if you day-to-day use DSH with a high-capability text model for coding, writing, or reasoning, but occasionally need to attach a screenshot, a formula image, or a UI design without switching to a dedicated vision model or sending image data to the cloud. As long as Ollama is running locally and your GPU can hold an 8B-class VL model, this bridge lets text model and local vision understanding work together in a single, natural conversation turn.