motongv/dsh-vision-adapter
? DeepSeek Harness ?????:???? Kimi????? DeepSeek ? DSH ??
dsh-vision-adapter adds vision capability to DeepSeek Harness: after installing you can drag and drop an image into the chat and have it recognized, with the chat entry point unchanged -- the image goes to a vision model, the text goes to DeepSeek for reasoning. A visual settings page (Settings -> Vision) lets you pick one of eight pre-configured providers (Kimi/Moonshot, Zhipu GLM, Qwen/Tongyi, OpenAI, Google Gemini, Doubao/Volcengine, MiniMax, StepFun) and paste an API key, after which two models appear in the model picker: DeepSeek V4 Pro Vision (X) and DeepSeek V4 Flash Vision (X). It runs an in-process HTTP bridge inside DSH and registers the selected vision model as an llm-pi-ai provider, merging incrementally so it never overwrites existing vision bridges. It caches image descriptions (the same image calls the vision model once), retries with backoff on 429/concurrency limits, and applies compatibility fixes.
Install
dsh plugin --profile web add dsh-vision-adapternpm package dsh-vision-adapter 0.6.4 (registry re-verified 2026-09-30 -- note the registry entry carries NO repository field, so the package name is taken from the README rather than confirmed by a registry back-link, MIT). After installing, restart DSH (the README notes dsh plugin add only changes config and plugins mount at startup), then open Settings -> Vision, pick a provider, paste the API key, and click Save & Apply. The pasted key is moved into DSH's credential store (~/.dsh/.credentials.yaml) with only a reference name kept in settings.
Compatibility
DeepSeek Harness Web UI (dsh web). Requires an API key for the chosen vision provider. Scope is DSH specifically, not a generic OpenAI client tool.
Details
- Repo: motongv/dsh-vision-adapter
- Category: Other
- Stars: 1
- Version: npm dsh-vision-adapter 0.6.4 (registry re-verified 2026-09-30; no repository field on the registry entry); MIT
- Last push: 2026-08-18
- First seen: 2026-08-14
Recent updates
The README documents the eight built-in providers with their preset vision models, endpoints and credential reference names (VISION_API_KEY, ZHIPU_API_KEY, DASHSCOPE_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY, ARK_API_KEY, MINIMAX_API_KEY, STEPFUN_API_KEY), the secure-key handling (plaintext moved into DSH's credential file), the image-description cache and 429 backoff, and the advanced section's DeepSeek text-backend settings (default deepseek-v4-pro + DEEPSEEK_API_KEY).
FAQ
- How do I install dsh-vision-adapter?
- Run dsh plugin --profile web add dsh-vision-adapter, then restart DSH. Open Settings -> Vision, choose a provider (endpoint and model are pre-filled), paste the API key, and click Save & Apply.
- Which providers are supported?
- Per the README: eight built-in providers (Kimi/Moonshot, Zhipu GLM, Qwen/Tongyi, OpenAI, Google Gemini, Doubao/Volcengine Ark, MiniMax, StepFun), each exposing two DSH models that route the text to DeepSeek V4 Pro or V4 Flash.
- Where is my API key stored?
- The README states a pasted plaintext key is written into DSH's credential store (~/.dsh/.credentials.yaml) and only a reference name is kept in the settings file, with a Configured/Missing status badge shown on the page.
Alternatives
oil-oil/dsh-vision · ysr666/dsh-vision-router · Einskyle/dsh-llm-vision-bridge