dsh-vision-adapter

Adds vision capability to DeepSeek Harness (DSH): after installing, you can drag & drop an image into the chat to have it recognized — the image goes to a vision model, the text goes to DeepSeek for reasoning, and the chat entry point stays exactly the same.

This is a plugin specifically for DeepSeek Harness (DSH), not a generic OpenAI client tool. It requires the DSH web UI (dsh web).

What you get

After installation, a visual settings page appears under Settings → Vision:

1. Pick a vision provider from the dropdown (endpoint and model are pre-filled automatically);
2. Paste your API Key into the field (masked by default, click the eye icon to reveal);
3. Click Save & Apply.

After saving, two models appear in the model picker at the bottom-right of the chat input — DeepSeek V4 Pro Vision (X) and DeepSeek V4 Flash Vision (X) (X = the provider you chose). Select one and drag an image in to recognize it.

Install (one command)

dsh plugin --profile web add dsh-vision-adapter

Then restart DSH.

Built-in providers (8)

Just pick a provider — endpoints and models are pre-configured. You only paste your API Key:

How the API Key works: paste the plaintext key for the provider you chose. On save, the plugin writes the key securely into DSH's credential store (~/.dsh/.credentials.yaml) and only keeps a reference name in the settings file — the plaintext never lands in settings. The page shows a Configured / Missing status badge. Once configured, you can leave the field empty and save again.

Configuration

1. After restart, open Settings → Vision;
2. Select a provider in the dropdown — endpoint and model are filled in automatically;
3. Paste your API Key (masked by default; click the eye icon to reveal);
4. Choose Output language (Chinese or English);
5. Click Save & Apply.

The advanced section lets you change the DeepSeek text backend (model name / endpoint / key reference). Defaults: deepseek-v4-pro + DEEPSEEK_API_KEY.

Models

Each provider exposes two models, backed by two DeepSeek reasoning backends:

• DeepSeek V4 Pro Vision (X) → text routed to DeepSeek V4 Pro
• DeepSeek V4 Flash Vision (X) → text routed to DeepSeek V4 Flash

When you switch providers in Settings, the model names update in sync (pick Zhipu → shows Zhipu, pick MiniMax → shows MiniMax).

How it works

image ──► vision model (Kimi/Zhipu/Qwen/OpenAI/Gemini/Doubao/MiniMax/StepFun) ──► text description
text + description ──► DeepSeek ──► reply

The plugin runs an in-process HTTP bridge inside DSH and registers the selected vision model as an llm-pi-ai provider (merged incrementally — it never overwrites other plugins' or your existing vision bridges). The image is described into text by the vision model, then the text + description are sent to DeepSeek for reasoning.

Details

• Image description cache: the same image only calls the vision model once;
• Rate-limit backoff: automatic retry with backoff on 429 / concurrency limits;
• Compatibility fixes: developer → system role normalization, max_completion_tokens → max_tokens, etc.;
• Secure key storage: plaintext keys pasted in Settings are moved into DSH's credential file, not kept in the settings file.

FAQ

Q: I don't see Settings → Vision or the models after installing? Restart DSH (dsh plugin add only changes config; plugins mount at startup).

Q: 401 when recognizing? The API Key is missing, wrong, or expired — check the Configured / Missing badge in Settings, and get a new key from the provider's console if needed.

Q: 429 when recognizing? The provider's free quota is rate-limited. The plugin already retries with backoff and caches images; if it keeps happening, top up the quota or switch to a provider with more headroom.

Q: Can I add another vision provider? Edit PROVIDERS in src/config.js (add one entry: endpoint, model, and credential reference name).

License

MIT