English | 中文

chat-deepseek-vision

A native DeepSeek Harness (DSH) Cordis plugin that analyzes images through the reverse-engineered chat.deepseek.com vision mode (model_type=vision) — free, no third-party vision API key required.

• No stdio/MCP bridge (replaces the old @deepseek-ai/dsh-mcp-client + node dist/index.js setup)
• Tools are registered directly into the current agent's tool registry, isomorphic with the official @deepseek-ai/dsh-tool-* plugins
• Pure ESM JavaScript with zero runtime assets (the official PoW wasm is embedded as base64), no build step

Tools

Screenshots rely on Windows PowerShell (CopyFromScreen / GDI BitBlt / PrintWindow three-method fallback). Video/audio capabilities are outside the "analyze images via chat.deepseek.com" scope and are not included in this plugin.

Send-box Attachment Images (Direct Support)

Users can attach images directly in the send box (drag & drop / paste / file picker) — no file path needed:

1. Admission pass-through: DSH's image-attachment admission (the prompt entry) requires the current model to declare image input, otherwise it rejects the request outright (MODEL_DOES_NOT_SUPPORT_IMAGES, shown as "the current model does not support images"). The official DeepSeek adapter hardcodes inputModalities: ["text"] in resolveModelInfo with no config override, so this plugin installs a capability-declaration patch (lib/model.js): for deepseek* routes only, it appends "image" to the returned resolveModelInfo modalities, letting attachment admission pass. The patch changes only the declaration, never any real request; the original method is restored when the plugin is disposed.

2. Auto-describe (agent/pre-step hook): image blocks in session messages are projected to text placeholders (the placeholder is rendered by the plugin as [图片附件 id=sha256:... name=... WxH] — Chinese by design, carrying the attachment id; the DeepSeek adapter never receives an image block), while the plugin automatically runs chat.deepseek.com vision and injects the description as a user message:

   <attachment-image id="sha256:..." name="photo.png" mediaType="image/png" width="1200" height="900">
   This image is ... (full description)
   </attachment-image>

   The model can answer directly from the description; the same attachment is only analyzed once per session (durable marker + in-process cache).

3. Deep-dive analysis: the model can call the analyze_attachment tool with an id for custom analysis (extract text/data, compare details, translate, etc.).

4. Native image models: when using a model that genuinely accepts images (e.g. a pi-ai vision model), set nativeImageModels: true — the plugin then skips projection/description when the model declares image input, letting the native image pipeline handle it.

5. Disable auto-describe: set autoDescribe: false (the analyze_attachment manual tool remains available).

Note: after projection, the original image thumbnail position in the user message shows the attachment placeholder text (consistent with DSH's policy for MCP image results with text-only models); the attachment bytes themselves remain stored as session attachments and can be re-analyzed anytime via analyze_attachment.

How It Works

Pure HTTP calls against chat.deepseek.com internal endpoints (no browser):

POST /api/v0/chat_session/create          → chat_session_id
POST /api/v0/chat/create_pow_challenge    → PoW challenge (DeepSeekHashV1)
solve PoW locally (bundled official sha3_wasm_bg.wasm, 26KB, zero imports)
POST /api/v0/file/upload_file (multipart) → file_id (x-ds-pow-response header)
GET  /api/v0/file/fetch_files             → poll until status == SUCCESS
POST /api/v0/chat/completion (SSE stream) → reply text (model_type="vision")
POST /api/v0/chat_session/delete          → delete session after use

⚠️ Risk notice: this is a reverse-engineered, non-public interface for personal research only. DeepSeek may change the protocol or flag accounts at any time; evaluate the risk yourself. When the algorithm changes, replace wasm/sha3_wasm_bg.wasm and run npm run embed:wasm.

Configuration

Plugin config comes from the loader entry's config; unset fields fall back to the same-named environment variables:

Getting the token (64 chars):

1. Log in to chat.deepseek.com in a browser
2. F12 → Console, run: JSON.parse(localStorage.getItem('userToken')).value
3. Fill the output into token (or DSW_TOKEN)

Limits: images only (screenshots, local images, URLs, data URIs), single image ≤ 15MB.

Install into a DSH Profile

1. Install the plugin package into the profile

Run in the profile directory (dsh plugin forwards its arguments to pnpm):

dsh plugin --profile web add file:D:/mcp/chat-deepseek-vision

Equivalent to pnpm add file:D:/mcp/chat-deepseek-vision in %USERPROFILE%\.dsh\profiles\web.

2. Register the plugin in cordis.patch.yml

Edit %USERPROFILE%\.dsh\profiles\web\cordis.patch.yml and append a loader entry (token from the browser localStorage):

# chat-deepseek-vision: native DSH vision plugin (replaces the MCP deepseek-vision)
- insert:
    - id: chat-deepseek-vision
      name: chat-deepseek-vision
      config:
        token: 'your-64-char-token'
        # cookie: 'ds_session_id=xxx; ...'   # optional
        # baseUrl: 'https://chat.deepseek.com'
        # timeoutMs: 180000

You may also omit config and rely on the environment variables DSW_TOKEN/DSW_COOKIE (handier for headless/CI).

3. Restart DSH

After restart, verify: ask the agent to call dsv_status — "connected OK" means the install succeeded; analyze_image is ready to use.

Local Development

npm run smoke        # smoke test: wasm/PoW, config resolution, resolveImage, plugin shape, screenshots, attachment projection/auto-describe
npm run embed:wasm   # after DeepSeek updates the PoW algorithm, re-embed from wasm/sha3_wasm_bg.wasm
node scripts/smoke.mjs

Directory layout:

chat-deepseek-vision/
├── package.json          # name: chat-deepseek-vision (peer: @deepseek-ai/cordis, @deepseek-ai/dsh-llm, @deepseek-ai/dsh-tools)
├── lib/
│   ├── index.js          # Cordis plugin: name/inject/apply + 6 tool registrations + agent/pre-step attachment handling
│   ├── attachment.js     # send-box attachments: image-block projection / auto-describe / analyze_attachment ref lookup
│   ├── model.js          # capability-declaration patch: deepseek routes declare image (attachment admission pass-through)
│   ├── vision.js         # facade: config resolution + resolveImage/analyzeImage(Buffer) + self-check
│   ├── deepseek-web.js   # DeepSeekWebClient (session/PoW/upload/SSE, ported from client.ts)
│   ├── screen.js         # screenshot tool logic (ported from tools.ts: describe_screen etc.)
│   ├── platform.js       # screenshot & window enumeration (ported from platform.ts, PowerShell, execSync pipe capture)
│   ├── pow.js            # PoW solver (wasm-bindgen calling convention, ported from pow.ts)
│   └── wasm-base64.js    # generated: official sha3_wasm_bg.wasm embedded as base64
├── wasm/sha3_wasm_bg.wasm  # official PoW wasm (update source)
└── scripts/
    ├── embed-wasm.mjs    # regenerate wasm-base64.js
    └── smoke.mjs         # local smoke test (SMOKE_SCREENSHOT=1 also runs a real screen capture)

License

MIT