dsh-vision-bridge

You are running your primary reasoning on a text-only model and a screenshot lands in the conversation the model simply cannot read. dsh-vision-bridge solves exactly that gap: as a dsh host plugin it hooks into the agent/pre-step waterfall, inspects each incoming message for image content blocks, and when the active model reports text-only input modalities while a vision endpoint is configured in the Web settings page, it transparently takes over that message so you never have to switch models by hand.

The core pipeline has three steps. First the current text model is asked to write a single precise vision prompt at temperature 0. Then that prompt, the scoped system prompt, and the image data are POSTed to the user-configured OpenAI-compatible /chat/completions endpoint. Finally the returned text replaces the original image blocks and is written into the session log as a plain-text user message. Throughout the process the text model never sees raw image bytes, the transcript stays fully reconstructable, and the KV-cache prefix is unchanged.

It is aimed at developers whose main inference loop runs on a text-only model but who occasionally need to handle image-bearing messages (product screenshots, error captures, whiteboard photos) without swapping the entire model stack for a single image, and at teams that consume OpenAI-compatible vision APIs and want vision to act as a transparent extension layer. The plugin exposes five settings fields (enabled, url, apiKey, model, timeoutMs) on the Web settings page, and all credentials are redacted on the wire.