dsh-image-bridge

Pasting a screenshot into DeepSeek Harness and asking a text-only model to read it normally hits a hard wall: the api-proxy refuses any image-bearing message the moment it sees the model modality list is text-only. dsh-image-bridge removes that wall, so pure-text models inside DeepSeek Harness can still receive pasted images and delegate recognition to whatever vision or MCP tools you already have wired up.

The plugin makes three targeted adjustments. First, it wraps the model-info pre-check so image messages are no longer rejected at the proxy layer. Second, it reads the image bytes through the attachments API and persists them into the session workspace .attachments/ folder. Third, it exploits the DSH surface-replace mechanism to swap the image block in the model-visible transcript for a lightweight placeholder such as [Image 1]:"", while the human-facing chat still renders the original thumbnail (click-to-zoom) rather than a bare file path. The model then passes that placeholder to an existing vision or MCP tool—vision_glance, mcp__mcp-vision__analyze_image, mcp__mcp-vision__ocr_extract, and the like—for actual pixel-level understanding. Models that natively support image input are left completely untouched on their native path.

If your daily workflow centers on a DeepSeek-class text model for code review, documentation drafting, or research, and you occasionally need it to look at a screenshot, an architecture diagram, or a scanned PDF page, this plugin is the missing bridge between what you paste and what the model can consume. Two practical caveats: avoid mixing slash-commands and images in the same submit, and consider clearing or compressing chat history before switching back to a text-only model so residual image blocks from a previous vision-model session do not trip the adapter.