dsh-vision

Text-only models in DSH (DeepSeek Harness) simply cannot see images: the GUI blocks paste and the model accepts text alone. dsh-vision was built to close that gap — it intercepts image pastes and drags at the window level, uploads them locally, and injects a file path as plain text into the composer. The model never touches binary data, yet behind the scenes the plugin performs real image recognition and generation.

On the recognition side, describe_image returns structured JSON evidence (natural-language description, OCR lines with pixel boxes, layout regions, entities) or a plain-text summary when preferred. On the generation side, generate_image produces images at auto, HD, or 4K quality, displayed inline as a produced-file card with a lightbox, download, and Finder reveal. Both chains ride a Codex → Gemini → Zhipu failover order: if one engine is unavailable the next picks up automatically, and every temporary artifact follows a delete-after-use policy with uploads auto-cleaned after a configurable retention window.

If your DSH harness runs a text-only model, you already have a ChatGPT or Codex login, and you occasionally need to "look at a picture" or "generate one," dsh-vision is a zero-extra-account, fully settings-page-configurable, read-only-with-respect-to-existing-Codex-data visual layer for personal use.