dsh-view-image

Most text-only models (e.g. deepseek-v4-flash) have no vision capability at all. The moment you paste an image into a conversation, you get a hard "current model does not support images" error and the turn simply fails. dsh-view-image takes a different approach: it attaches a standalone OpenAI-compatible vision model outside the main dialogue, lets it do all the looking, and feeds back a plain-text description to the language model. Image bytes never enter the primary context window, so the main model only ever sees a natural-language summary, as if you had read the picture aloud to it. Your pure-text routing stays exactly as it is, but now "paste a screenshot and ask what the error means" just works.

On the UX side the plugin is polished: the Web input box accepts Ctrl+V paste, renders inline thumbnails (right-aligned, 240 px, multi-image supported), and offers click-to-enlarge, hover-to-copy, and drag-back-to-inputbox interactions. In terminal and headless workflows you simply pass a file path and the model calls view_image on its own. The vision endpoint is fully pluggable—Ollama, vLLM, LiteLLM, OpenRouter, Alibaba Cloud Bailian, opencode.ai, or any other OpenAI-compatible API can be pointed to via a single vision-model.json, so a small local VLM is all you need and there is no obligation to subscribe to a third-party multimodal API.

Attachments use content-addressed storage (SHA-256) with de-duplication and integrity checks, so thumbnails survive session restarts and replays without losing data. In short, the plugin is a good fit for three audiences: everyday users whose primary model is text-only but who occasionally need to describe images; developers scripting headless pipelines that should include an image-recognition step; and privacy-conscious users who prefer a locally deployed vision model over uploading screenshots to external multimodal services. Configure one endpoint, restart dsh once, and your text-only model gains a pair of eyes.