libinyam/dsh-vision-provider
Config-only DeepSeek Harness bundle for OpenAI-compatible vision models.
Gives DeepSeek Harness selectable vision choices under one 'DeepSeek + Vision' provider: GLM-4.6V-Flash, Qwen VL Max, GPT-4.1 mini (Vision). v0.3.0 brings vision selection into the Web UI: it reads every model advertising image input from Settings > Models, each vision model becomes a separate selectable DeepSeek combination; text-only requests go directly to deepseek-v4-flash while image-bearing messages are analyzed by the selected vision model, the visual description replaces the raw image before the request reaches DeepSeek, and DeepSeek still performs reasoning, tool use and final response generation. Repeated tool steps reuse cached image analysis.
Install
dsh plugin --profile web add github:libinyam/dsh-vision-providerGitHub install per README: dsh plugin --profile web add github:libinyam/dsh-vision-provider (from a harness source checkout: pnpm dsh plugin --profile web add github:libinyam/dsh-vision-provider; always use the same DSH_HOME for plugin management and startup). No npm package. Requires DeepSeek Harness 0.1.0-rc.5+ and Node.js >=22.19.
Compatibility
DeepSeek Harness 0.1.0-rc.5+; Node.js >=22.19. Two-model bridge, not native pixel input — final answer quality depends on both the vision sidecar and DeepSeek.
Details
- Repo: libinyam/dsh-vision-provider
- Category: Coding & Development
- Stars: 4
- Version: GitHub source v0.3.0 (no npm package)
- Last push: 2026-08-15
- First seen: 2026-08-13
Recent updates
v0.3.0: Web UI vision selection; per-combination model naming; image-replacement bridge; cached image analysis; community project (not official DeepSeek/OpenAI).
FAQ
- Which vision models are available?
- GLM-4.6V-Flash, Qwen VL Max and GPT-4.1 mini (Vision) — selectable as separate DeepSeek + Vision combinations in the model picker.
- Does the vision model answer directly?
- No — it runs as a sidecar producing a visual description; DeepSeek still performs reasoning, tool use, and the final response.
- What are the requirements?
- DeepSeek Harness 0.1.0-rc.5+ and Node.js >=22.19; configure the DeepSeek key normally and the vision key via the existing vision-openai route.
Alternatives
Sorwcyra/ds-vision-plugin · GXX182/dsh-vision-bridge · Einskyle/dsh-llm-vision-bridge