dsh-vision-skill

A lot of DSH users run purely text-based models, and the moment they paste an image into the conversation the host throws MODEL_DOES_NOT_SUPPORT_IMAGES. dsh-vision-skill is the bridge for exactly that gap. It wraps Qwen's official dynamic-resolution pipeline into a native DSH plugin, touches only official extension seams (no framework patches), and lets any OpenAI-compatible multimodal backend (MiniMax-M3 by default) serve as the "eyes" of a text-only model.

The plugin ships eight tools plus one runtime skill covering general image analysis (including a structured evidence mode), standalone OCR with layout preservation, object grounding that returns both pixel and normalised coordinates, element enumeration, dominant-colour detection, chunked OCR for very long screenshots, and a clipboard-image fallback. Since v0.4 the paste-to-path channel uploads pasted images into the workspace automatically, so no image block ever appears in the message and the host stops rejecting the turn. A progressive-exposure design keeps only a single lightweight activation tool in the global namespace; the full tool set is bound to the agent only after the skill loads, saving a meaningful chunk of context.

Production-oriented engineering is baked in as well: ordered multi-provider failover with 429 back-off, a SHA-256-keyed image cache parameterised by mode/budget/crop/prompt, realpath-based path-fence validation to block traversal, and credential-based key management instead of raw strings in config. The evidence mode of vision_analyze returns structured JSON with summary, full OCR text, reading order, entity–relation semantics, and uncertainty flags, making it easy for downstream pipelines to consume directly.

If you are running a text-only model in DSH, need OCR or visual grounding on the fly, and would rather not maintain framework patches or a pile of ad-hoc scripts, this plugin converges the whole workflow into a single install. Point it at any compatible vision API, set the credential, and go.