dsh-see-image

English | 中文

A DSH (DeepSeek Harness) plugin that registers a see_image tool — route image files to any OpenAI-compatible vision model and get a text description back, so text-only models can "see" pictures without switching models or orchestrating subagents. Default backend: GitHub Copilot gpt-4.1 (works on the Free plan) with built-in token auto-refresh. Install = copy the folder into your DSH profile + one cordis.patch.yml entry; hot-reload, zero npm installs.

model → see_image(file_path, question?)
        │ 1. read the image through DSH's file service (sandbox / observation policy applies)
        │ 2. base64-encode it
        │ 3. POST {baseURL}/chat/completions
        ▼
    vision model → text description → model answers from the description

Live demo

Feed it assets/demo-chart.png (shipped in this repo) with the question "Read this bar chart: the value and color of each quarter, and which quarter is highest?":

see_image returns:

Title: Quarterly Revenue (M$) Each quarter's value and color: Q1: 85 (blue), Q2: 140 (red), Q3: 115 (green), Q4: 175 (orange) Highest quarter: Q4 (175, orange)

(Real output from GitHub Copilot Free plan gpt-4.1, ~5 seconds.)

Features

• Any OpenAI-compatible VLM: Zhipu / Bailian / Doubao / SiliconFlow / local Ollama / OpenRouter… just change three config keys
• First-class GitHub Copilot support: default backend gpt-4.1 (Free plan eligible) with built-in token auto-refresh — no key management at all
• No tool-count issues: requests are made by the plugin directly and don't carry DSH's tool list, so OpenAI-style gateway limits (128 tools) don't apply
• Zero npm dependencies: reuses the host's @deepseek-ai/dsh-tools / schemastery (the profile's flat node_modules resolves them automatically)
• Hot reload: editing cordis.patch.yml takes effect immediately (verified without restart; in conservative scenarios restart dsh web)

Install (3 steps)

1. (Optional) Clone the repository instead of unzipping:

   git clone https://github.com/tiefeiyu/dsh-see-image
   

2. Copy the plugin into your profile directory ($DSH_HOME is usually ~/.dsh; example shows the web profile):

   $DSH_HOME/profiles/web/plugins/dsh-see-image/
   ├── index.js        # plugin entry
   ├── scripts/        # Copilot token scripts (optional; only needed for the Copilot backend)
   ├── package.json
   ├── README.md
   └── LICENSE
   

3. Mount it — append to $DSH_HOME/profiles/web/cordis.patch.yml:

   - insert:
       - id: local-see-image
         name: './plugins/dsh-see-image/index.js'
         config:
           baseURL: 'https://api.individual.githubcopilot.com'   # default: GitHub Copilot
           model: gpt-4.1
           apiKeyEnv: VISION_API_KEY    # non-Copilot backends read this env var
           maxTokens: 1024
           timeoutMs: 90000
           maxBytes: 15728640
   

4. Verify: wait ~6 seconds (hot reload); see_image appears in the tool list of new sessions.

Configuration

Key	Default	Description
baseURL	https://api.individual.githubcopilot.com	OpenAI-compatible endpoint; the plugin appends /chat/completions
model	gpt-4.1	Vision model ID
apiKeyEnv	VISION_API_KEY	Env var name holding the API key for non-Copilot backends; empty sends no Authorization header (keyless local endpoints like Ollama)
maxTokens	1024	Max output tokens (Zhipu glm-4v-flash caps at 1024; raise it for larger models)
timeoutMs	90000	Request timeout
maxBytes	15728640	Max image size (15 MB)
prompt	(detailed Chinese description instruction)	Default question; a question argument passed at call time takes precedence

Backend switching examples

Just change three lines of config (baseURL / model / apiKeyEnv) and set the env var (setx VISION_API_KEY "your-key" or write it to ~/.dsh/.env):

Backend	baseURL	model	Notes
GitHub Copilot (default, free quota)	https://api.individual.githubcopilot.com	gpt-4.1	Uses the built-in token flow (below); no VISION_API_KEY needed
Zhipu (free tier)	https://open.bigmodel.cn/api/paas/v4	glm-4v-flash	Register at bigmodel.cn for a key (id.secret format)
Alibaba Bailian	https://dashscope.aliyuncs.com/compatible-mode/v1	qwen3-vl-flash	For screenshot/GUI scenarios try qwen3-vl-plus
SiliconFlow	https://api.siliconflow.cn/v1	Qwen/Qwen2.5-VL-32B-Instruct
Volcano Doubao	https://ark.cn-beijing.volces.com/api/v3	doubao-seed-2-1-turbo-260628	Ark model IDs carry a date suffix
Ollama local (offline)	http://localhost:11434/v1	qwen3-vl:4b	apiKeyEnv: '' — fully offline and free

GitHub Copilot token (default backend only)

The plugin refreshes automatically: Copilot access tokens expire after 30 minutes, so the plugin checks exp before every call, mints a fresh one from the OAuth refresh token when needed, and writes it back to ~/.dsh/.credentials.yaml.

One-time setup (~1 minute, needs GitHub authorization in a browser):

1. Run the bundled login script:

   node scripts/copilot-login.mjs
   

2. The script prints an authorization URL and a code (e.g. https://github.com/login/device + XXXX-XXXX). Open the URL in a browser, sign in to GitHub (an account with a Copilot subscription — the Free plan counts), enter the code and authorize.

3. The script saves two things:

   • ~/.dsh/copilot-oauth.json — the OAuth refresh token (long-lived, keep it private)
   • COPILOT_GITHUB_TOKEN in ~/.dsh/.credentials.yaml — the current access token (auto-refreshed by the plugin)

If the OAuth token ever goes stale (rare), rerun the login script. You can also force an immediate refresh with node scripts/refresh-copilot-token.mjs.

Usage

In conversation, just say "look at this image / read this screenshot" — the model locates the file and calls the tool itself. You can also state the question explicitly:

see_image(file_path="C:\\Users\\me\\Desktop\\error.png", question="What is the full text of this error?")

question may be omitted; the default output covers: scene → verbatim text transcription → color/shape/layout details → explanation of charts/UI/errors (in Chinese).

Compatibility & known notes

• Tested with @deepseek-ai/dsh 0.1.0-rc.6 (web profile, Windows; other platforms should behave the same)
• The plugin registers at the host-global layer, so it is visible to every DSH entry point (web / TUI / remote)
• Does not cover Anthropic / Gemini native protocols (Gemini can be reached through its OpenAI-compatible entry)
• If hot reload doesn't pick up plugin code changes (ESM module cache), rename the file to a new path (e.g. index.v2.js) and update cordis.patch.yml's name accordingly, or restart dsh web

License

MIT