Foreword

DSH (DeepSeek Harness) follows the “everything is a plugin” route; if the model lacks a capability, the plugin fills the gap. However, pure text DeepSeek models are helpless when encountering images: during math modeling tasks, they cannot read information from geometric diagrams, flowcharts, or data charts within images, relying solely on manual transcription.

Existing vision plugins mostly rely on Chrome + Gemini, requiring API Key application and configuration, which involves too many steps for those who just want to get it running quickly. dsh-vision-edge-doubao takes a different path: it drives the local Edge browser and reuses Doubao’s web login state for vision recognition, zero cost, no API key needed. Below is a specific introduction to how this plugin works.

What is this

dsh-vision-edge-doubao is a DSH vision plugin, author iceBear662, MIT license, current version 0.1.1. It starts an Edge debugging instance locally to open the Doubao web version, forwarding images and recognition requests from DSH sessions to Doubao, and then extracting the answers back to the model, thereby granting the model vision capabilities.

The project borrows the architecture and bridging scheme from dsh-vision (author 54xkeee, MIT License), with the main changes being switching the bridging browser from Chrome to Edge, and fixing compatibility issues such as input box positioning, reply extraction, and long text sending after the Doubao web redesign.

Core Features

  • General Vision: Photos, screenshots, OCR text recognition, region inspection (region), multi-image comparison (compare).
  • Math Modeling Special (mode=math): Deep structural recognition of five common types in math modeling—geometric shapes, flowcharts, data charts, tables, and formulas, where formulas output LaTeX transcription.
  • Clarification Loop: When recognition encounters an uncertain item, the plugin returns a clarify question (e.g., “Is point D the midpoint of BC?”) in the result. After you answer, the plugin re-recognizes with supplementary information to obtain a more precise result.
  • Maintenance-free Login: Doubao login state is saved in an independent browser profile, allowing for one-time login for long-term use without needing to log in again even after restarting the computer.
  • Zero Cost: Reuses browser login state, no need to apply for any API Key.
  • Expert Mode Auto-enforcement: The plugin automatically ensures Doubao uses Expert mode for vision (a persistent setting that remains after one switch), performing an idempotent check before each vision. To switch back to Quick mode, set the environment variable DOUBAO_MODE=quick before starting the bridge; if an exception is encountered, it silently degrades and continues using the current mode for recognition.
  • Cache: Returns cached results for the same image to save traffic; try a different phrasing to re-recognize.
  • Doubao Bridge Compatibility: Handles ProseMirror input box, long text insertion in one go, virtual scrolling reply extraction, and multi-segment reply merging; prompts use answer priority + disable tables (to avoid Doubao table rendering loss) + anti-injection headers.

Environment Requirements

  • Windows, requires Edge browser
  • DeepSeek Harness (DSH) Web GUI installed
  • Node.js ≥ 18 (bridging component needed)

Installation and Enablement

Installation uses the recommended method marked in the README, adding from GitHub:

dsh plugin --profile web add github:iceBear662/dsh-vision-edge-doubao

After installation, restart dsh web to take effect; the plugin loads at startup.

To uninstall, execute:

dsh plugin --profile web remove dsh-vision-edge-doubao

First-time Configuration (Only once)

  1. Run the one-click startup script:
powershell -ExecutionPolicy Bypass -File <Plugin Directory>\start-vision-edge.ps1

The script automatically completes three things: starting the Edge debugging instance (independent configuration, does not affect the daily browser), starting the Doubao bridge, and verifying the environment. After installation, the script is located at node_modules\dsh-vision-edge-doubao\start-vision-edge.ps1.

  1. Login to Doubao: Open https://www.doubao.com in the popped-up Edge window, scan the code or login with a phone number once. The login state is saved in an independent configuration; you won’t need to log in again later.

  2. Start using: In a DSH session, simply send an image to the model, or say “look at this picture”, and the model will automatically invoke the vision tool.

Vision Parameters

mode controls the recognition mode: glance (general) / ocr (text transcription) / region (region inspection) / compare (multi-image comparison) / math (math modeling special).

detail controls the precision: auto / fast / standard / deep, where auto automatically upgrades for complex images.

Applicable Scenarios and Notes

Suitable for: Windows + Edge environment, DSH users who don’t want to apply for an API Key for vision. The benefits are most obvious in scenarios such as math modeling, where recognizing geometric shapes / flowcharts / charts / formulas is needed.

There are a few points that must be known before use:

  • The plugin drives the browser through automation, reusing your already logged-in Doubao session to recognize images. This may violate Doubao/ByteDance’s Terms of Service and carries a risk of account restriction or banning. It is recommended to use a non-primary account, do not use for commercial purposes or bulk scraping, and reasonably control the call frequency.
  • This project has no affiliation with Doubao or ByteDance; the Doubao web redesign may cause the plugin to fail, and long-term availability is not guaranteed.
  • Recognition timeout is mostly due to Edge being closed; simply run start-vision-edge.ps1 again to recover; Doubao occasionally triggers a CAPTCHA, just wait and retry.
  • The plugin runs with the permissions of the current dsh process; it is recommended to read the source code before installation to confirm that the MIT license and code content are acceptable.

Conclusion

One-sentence summary: dsh-vision-edge-doubao uses local Edge + Doubao web version to plug vision capabilities into DSH at zero cost; its distinguishing features from general vision plugins are the math modeling special mode and the clarification loop.

  • GitHub: https://github.com/iceBear662/dsh-vision-edge-doubao
  • Community Directory: https://www.skillhub.cn/plugins/iceBear662/dsh-vision-edge-doubao (Independent community site, no official affiliation with DeepSeek / Moonshot AI)