tdf1995/dsh-plugin-vision
Vision for text-only LLMs in DeepSeek Harness (DSH): describe images / OCR / VQA via free Gemini & GLM vision
Vision for text-only LLMs inside DSH through free Gemini/GLM APIs: see_image analyzes local images (png/jpg/jpeg/webp/gif, ≤ 20MB) with custom questions; dual-provider auto mode picks whichever key is available; sticky routing remembers the last successful provider; automatic failover on network errors/rate limits; 429 backoff retry (default 3); images over 4MB auto-compressed to 1920px (JPEG 85) before upload. vision_set_key stores an API key in-session (into the DSH credential store, immediate effect); vision_status shows per-provider key state without leaking keys. Built-in browser Ctrl+V / drag / 🖼️ button attachment cards survive restarts.
Install
npm i -D dsh-plugin-vision && dsh web --patch node_modules/dsh-plugin-vision/cordis.patch.ymlOfficial DSH plugin format (dsh.bundle.patch + cordis.patch.yml): npm i -D dsh-plugin-vision (npm 0.1.0 verified on registry), then dsh web --patch node_modules/dsh-plugin-vision/cordis.patch.yml, or merge the plugin's cordis.patch.yml into your profile patch. Requires Node.js >= 20 and at least one vision API key (Gemini or Zhipu GLM; GLM glm-4.6v-flash is free and China-direct). No API key ever enters code or the repo.
Compatibility
DeepSeek Harness with Node.js >= 20; dual provider (Gemini gemini-3.6-flash + Zhipu GLM glm-4.6v-flash); browser paste/drop built in.
Details
- Repo: tdf1995/dsh-plugin-vision
- Category: Files & Data
- Stars: 2
- Version: npm 0.1.0 (registry-verified 2026-08-29)
- Last push: 2026-08-15
- First seen: 2026-08-14
Recent updates
see_image with Gemini/GLM free vision; dual-provider auto + failover; vision_set_key/vision_status; built-in paste/drop attachments.
FAQ
- Which vision APIs are used?
- Gemini (gemini-3.6-flash) and Zhipu GLM (glm-4.6v-flash — free and China-direct); auto mode picks whichever key is available.
- Do API keys go into the repo?
- No — keys are stored in-session via vision_set_key into the DSH credential store; none are ever in code or the repository.
- What happens on rate limits?
- Automatic failover to the other provider plus 429 backoff retry (default 3 attempts).
Alternatives
54xkeee/dsh-vision · tiefeiyu/dsh-see-image · FuzzySoul/dsh-free-vision