ysr666/dsh-vision-router
ysr666/dsh-vision-router为纯文本 Agent 提供视觉能力:内置免 Key 视觉链 + 像素级视觉工具(看图问答、定位、裁剪、像素对比、取色、OCR、矢量化、抠图、截图);粘贴图片即可用。
为 DSH 上的纯文本模型提供视觉插件:通过分层证据记忆与缓存,实现图片理解与生成的原生交互及 GUI 自动化。
Intent-driven vision, image generation, and visible browser automation for DeepSeek Harness.
MindsEye is a plugin for DeepSeek Harness. It gives text-only models access to image understanding, image generation, and optional browser automation while keeping the DSH conversation as the main user experience.
mindseye_read_image handles visual questions and focused tasks such as OCR, layout, charts, colors, and pixel differences.mindseye_ground returns a target's pixel bounding box for downstream actions such as clicking or cropping.mindseye_generate_image sends a user's image request to the configured image-generation route.mindseye_edit_image sends a DSH image attachment and an edit request to the configured image-editing route.When gui.enabled is turned on, MindsEye opens a separate visible Chrome or Edge session. The GUI tools can open pages, take snapshots, wait, click, type, send key presses, scroll, and close the session.
If a page requires CAPTCHA, login, or permission confirmation, the run pauses on a native DSH question card. The user can:
After the user resumes, MindsEye checks the page state before returning control to the model. The browser uses an isolated session and does not attach to the user's existing Chrome or Edge profile. GUI actions require a fresh snapshot after each action so element references and coordinates cannot silently become stale.
| Tool | Purpose |
|---|---|
mindseye_plan | Extracts the current request and prepares the intent context used by downstream tools. |
mindseye_read_image | Answers questions about one or more images and extracts focused visual evidence. |
mindseye_ground | Locates a target and returns its pixel bounding box. |
mindseye_generate_image | Generates an image from the user's request. |
mindseye_edit_image | Edits a supplied image attachment. |
mindseye_vision_activate | Mounts the vision tools during a text-only turn. |
mindseye_gui_open / snapshot / wait | Opens a browser session and observes its current state. |
mindseye_gui_click / type / keypress / scroll | Performs a state-checked browser action. |
mindseye_gui_close | Closes the current browser session. |
The memory tools are optional and expose explicit DSH operations for storing, retrieving, searching, and comparing image-related records.
Configure MindsEye from the DSH settings card or the plugin configuration.
vision.routes: independent routes for understand, extract, and locate.vision.fallbacks: fallback routes for vision calls.image.generate: ordered image-generation routes.image.edit: ordered image-editing routes.gui.enabled: enables the visible browser tools. It is disabled by default.gui.browser: auto, chrome, or edge.gui.restrictHosts: enables host allowlisting when set to true.gui.allowedHosts: hosts allowed when host restriction is enabled.gui.maxSteps and gui.timeoutMs: limits for one browser run.Vision routes use OpenAI-compatible Chat Completions or Responses APIs. Image routes support JSON and multipart request bodies so different image providers can be configured independently.
npm install dsh-mindseye
npx @deepseek-ai/dsh plugin --profile web add dsh-mindseye
Restart DSH Web after installation. Then configure at least one vision route in the MindsEye settings card. Unconfigured focused vision routes fall back to the general understanding route when available.
pnpm install
pnpm test
pnpm typecheck
pnpm build