dsh-plugin-vision

Most pure-text LLMs are blind by design: hand them a screenshot and you get either a wall of base64 or a routing-layer rejection that the model does not accept image inputs. The see tool registered by dsh-plugin-vision bridges exactly that gap, decomposing any image into three plain-text channels so that non-multimodal models like deepseek-v4-flash can genuinely read what is on the picture.

Each channel has a distinct job. OCR taps the macOS Vision framework for offline text extraction with automatic language detection, returning normalized coordinates and word-level accuracy for mixed Chinese-English layouts. The ASCII layout channel renders the image as an 88-column grayscale character painting, letting the model perceive spatial structure without a single pixel. The semantic description channel calls a vision-language model (mimo-v2.5 by default) through an OpenAI-compatible gateway, injecting the OCR output as ground truth so that every number, label, and symbol in the report matches the image verbatim. Channels fail independently, so one hiccup never sinks the whole result.

If you are using a text-only model for document review, UI walkthroughs, chart reading, or screenshot Q&A and do not want to pay the premium of a multimodal endpoint, this plugin is the lowest-friction bridge. It works out of the box with zero configuration, and every output is clean text the model can quote, reason over, and compare directly.