dsh-client-paste-image

When using text-only models like DeepSeek or GLM, pasting a screenshot into the composer often returns a dead-end message - the model simply does not support vision. dsh-client-paste-image bridges that gap: it intercepts clipboard images at the capture stage, routes the bytes through the dsh-drop-caret upload endpoint to persist them as content-addressed files, and inserts the resulting file path back into the input. The conversation stays plain text; when an image is actually needed, the path can be handed to a vision-capable model.

The plugin only activates for routes listed in textOnlyRoutes within cordis.patch.yml. Vision models and every other unlisted route are completely unaffected - native thumbnails, draft attachments, and image-admission pipelines all run as usual. An empty list means no interception anywhere, a fail-safe that behaves as if the plugin were not installed. Alt + Paste bypasses the interception at any time, and plain-text pastes are never touched.

Ideal for developers who work daily with non-vision models like DeepSeek yet frequently need to reference screenshots or icons, and for teams building workflows where a text agent and a vision agent divide their responsibilities. No model switching required - paste and get a path, keeping the flow intact.