motongv/dsh-opencodex-vision-toolkit
Adds native visual tools for text-only DeepSeek agents in DeepSeek Harness, delegating remote image understanding to a local OpenCodex instance while keeping cropping and pixel comparison local and model-free: vision_describe (description or focused Q&A), vision_ocr (visible text in reading order), vision_locate (estimated target boxes in original-image pixels), vision_crop (deterministic local crop to PNG) and vision_pixel_diff (local difference percentage plus a red heatmap). Only opencodex_vision_activate is visible initially; once the opencodex-vision skill is loaded and activation succeeds, the five execution tools are registered for that agent only. The README documents traversal/symlink checks, byte and decoded-pixel limits before remote calls, and that vision output is treated as untrusted evidence rather than instructions.
Install
git clone https://github.com/motongv/dsh-opencodex-vision-toolkit.git
cd dsh-opencodex-vision-toolkit
npm install --legacy-peer-deps
npm run build
dsh plugin --profile web add "$PWD"README Install block quoted verbatim; the README adds that you restart DSH Web, invoke /opencodex-vision, then call opencodex_vision_activate once, and that a preflight doctor can be run first with node ./bin/cli.mjs doctor. Registry check 2026-09-16: no dsh-opencodex-vision-toolkit npm package resolved, so the source path is the only published route.
Compatibility
README requirements: DeepSeek Harness 0.1.0-rc.6 or a compatible release, Node.js 22.19+ or 24+, OpenCodex on http://127.0.0.1:10100/v1, and an image-capable OpenCodex model route whose default is gpt-5.6-luna. The README states the plugin never reads OpenCodex configuration, API keys, OAuth tokens, cookies or DSH Credentials — it accepts only an HTTP loopback base URL. Repo verified 2026-09-16 (0★, MIT, last push 2026-08-13).
Details
- Repo: motongv/dsh-opencodex-vision-toolkit
- Category: Coding & Development
- Stars: 0
- Version: Source checkout (README: requires DeepSeek Harness 0.1.0-rc.6 or a compatible release)
- Last push: 2026-08-13
- First seen: 2026-08-13
Recent updates
The README documents the tool list, the activation flow and the security limits rather than a release-by-release table.
FAQ
- How do I install it?
- The README's source path is git clone → npm install --legacy-peer-deps → npm run build → dsh plugin --profile web add "$PWD"; then restart DSH Web, invoke /opencodex-vision and call opencodex_vision_activate once.
- Why do I not see the vision tools at first?
- Per the README only opencodex_vision_activate is initially visible; the five execution tools are registered for that agent after the opencodex-vision skill is loaded and activation succeeds.
- Does image understanding need a model call?
- Only the remote description/OCR/locate steps do — the README says cropping and pixel comparison run locally with no model call.