dsh-vision-router
A tool that provides vision for text-only agents on DeepSeek Harness.
Verified
Core facts
Read by a script from the official API, not rewritten by a model.
- License
- MIT
- Languages
- JavaScript
- Maintainer
- ysr666
- Verified
What it is
dsh-vision-router provides vision, described as “eyes”, for text-only agents on DeepSeek Harness. Its vision tools are built on sharp, potrace, tesseract and system Chrome, without Python.
Main features
Vision answers are cached by attachment content hash.
The vision chain reads the image at original resolution and downscales it automatically only to protect latency and quota.
The pixel loop runs reference, vision_html_screenshot, vision_pixel_diff (ratio and worst 8×8-grid regions), then fix, and repeats until the mismatch converges.
vision_ground locates a target and gives an original-pixel box x1/y1/x2/y2.
vision_crop crops and zooms into a pixel box.
vision_pixel_diff compares pixels and reports the diff ratio and the worst 8×8-grid regions.
Getting started
Install into an existing profile that already serves the Web UI.
The dsh plugin add command only adds the plugin to the profile you name.
For Desktop, change the web profile option to the desktop profile option and reload the app.
Keep the usual chat model, click the Vision control beside the composer, and check that it shows a check mark.
Pasting or uploading an image does not enable Vision automatically.
The built-in anonymous OVH vision fallback is already configured, so normal image use needs no signup or API key.
npx @deepseek-ai/dsh plugin --profile web add dsh-vision-routernpx @deepseek-ai/dsh plugin --profile web add dsh-vision-routerWho it is for
The requirements include a DeepSeek Harness Web profile. The requirements include Node 22 or later on the host side. The default free chain needs no API key, and a credential reference, apiKeyEnv, is needed only for paid httpProviders.
Chrome, Chromium or Edge is needed only for vision_html_screenshot, and every other tool works without a browser. While Vision is off, the Agent cannot use vision_colors, vision_ocr or the other Vision Router tools, even with tool set to true in Settings. In the default mode with progressiveTools set to false, vision_activate is not registered.
Progressive mounting is opted into by explicitly setting progressiveTools to true in the profile or composition patch file. vision_screenshot is disabled by default and requires explicit opt-in. Vision Router does not operate fully offline in its default configuration.
To stop Vision Router itself from sending image data to remote vision endpoints, enable Local-only vision under Settings, Vision Router, General. OVH anonymous limits are 2 requests per minute per IP per model. The vision tools try backends in order and surface an error only after all of them fail.
Failures are classified as region, tos, quota, rate-limit, context or network, and the final error carries advice. Exact Test Vision sends one request to one exact model. Oversized uploads are downscaled before the call, with a default budget of 4 MP, to keep tool calls fast.
Related tools
Related tools listed on this site.
LongCat
Model and API platform. This review reads the official LongCat-2.0 technical article and its linked API product page, covering code understanding/generation, reasoning and interaction purposes; the original chat UI was not reviewed in operation.
豆包工作
Doubao Work is Doubao’s productivity-focused AI product. Its primary official description presents goal-driven task decomposition and tool use for research, content creation, data processing and application building.
InfCode
Enterprise AI coding agent from TokensInfinity. Its site describes using enterprise code and knowledge through IDE, extension and CLI forms for completion, cross-file changes and engineering tasks.
提示工程指南
Prompt Engineering Guide is a learning resource for prompt design and LLM applications; its Chinese site covers techniques, examples, agents and limitations.
图改改
TuGaiGai is an online editor focused on text within images. Its official page describes changing, adding and removing image text, font recognition, image translation and image/PDF conversion.
多维视界
Duowei Shijie is an audiovisual analysis tool whose official site describes transcription, translation, summaries, interactive learning and content rewriting.
Winston AI
AI-content detection and plagiarism-checking tool whose site describes text scans, OCR input, sentence prediction maps and reports.
Sources
This page is written from the official sources below. Open them to check the original text.
- GitHub READMEhttps://github.com/ysr666/dsh-vision-routerFetched
Verified ·
This page is written by AI from the official sources above and checked automatically. This site has not tested it. Found a mistake? Tell us.