dsh-vision-tool

In DeepSeek Harness, text-only models like deepseek-v4-flash have no native ability to interpret images—pasted screenshots are opaque binary blobs the model simply cannot parse. dsh-vision-tool bridges that gap with two coordinated plugins: vision-prompt intercepts the session.prompt API call, persists pasted images as content-addressed attachments, and rewrites them in place into text prompts that carry the full attachment-reference JSON; vision-tool then registers a global analyze_image tool that lets the model route any image to a vision model and receive a plain-text description back. The entire pipeline is verified end-to-end: paste an image, it is stored as an attachment, the model calls analyze_image, and the vision model returns a description.

Out of the box the tool targets the kimi-for-coding endpoint, but any OpenAI-compatible vision API works—just swap the base URL and model name in your profile config. Accepted inputs include the attachment-reference JSON produced by the paste-rewrite mechanism and local file paths (png, jpg, jpeg, webp, gif; local files capped at 20 MB). On the security side, vision-prompt re-implements the official /api trust fence (loopback and trustedHosts checks, sec-fetch-site and Origin validation, 160 MB body limit), and any failure degrades to a clean passthrough so the original request always reaches the model intact.

This plugin is a good fit for developers who work in DSH with text-only models but still need to read screenshots, charts, or diagrams on the fly. Once installed the workflow is invisible: paste an image and keep typing—the model decides for itself whether to call the vision tool, and sessions that already handle images natively are forwarded without a single byte changed.