dsh-plugin-mm-vision

Text-only LLMs have one fundamental blind spot: they cannot see. A candlestick chart, a UI screenshot, a landscape photo is nothing more than unparseable bitstream to DeepSeek, GPT-4 base, or Claude. dsh-plugin-mm-vision is a DeepSeek Harness client plugin that gives any text-only LLM the ability to interpret images. It calls a vision model of your choosing to translate the image into compact, structured spatial text, which the text model then uses to reconstruct spatial cognition and reason about positional relationships.

The core capability is synesthesia encoding: every key element carries precise percentage coordinates, covering chart inflection points, annotations, buttons, and text blocks. When the prompt contains keywords like pixel or reconstruct, it can also output a 40x30 RGB color-block grid for original-image reconstruction. Auto mode detects whether the input is a chart or a natural photo and adjusts the encoding strategy accordingly; manual brief, full, coords, and pixel modes are also available. A built-in TTL cache ensures repeated analyses return instantly, and model, baseUrl, and API key are all configurable with zero hard-coded values.

It suits developers and teams whose workflows require a text model to understand image content: quantitative analysts who want an LLM to read market screenshots, front-end engineers who need a model to parse UI layouts, or researchers who want a model to describe experimental charts. As long as your pipeline intersects text models with visual content, this plugin slots in seamlessly.