Preface

In DeepSeek Harness, some model routes are primarily text-first. If a task involves screenshots, configuration diagrams, or interface images, the image needs to be sent to an external vision endpoint for analysis, and the result is then brought back into the current session as text.

This article introduces dsh-vision-bridge. It solves the following problem: while preserving the existing Harness model list and selection experience, it allows text-first models to handle image-based questions as well.

What This Is

dsh-vision-bridge is a DeepSeek Harness plugin, hosted in the repository gxx182/dsh-vision-bridge, maintained by GXX182, and licensed under MIT.

Its purpose is to bridge images to external vision APIs and return only text-based analysis results.

Core Features

The following capabilities have been verified:

  • Adds image understanding capabilities to text-first model routes.
  • Preserves the Harness model list without duplicating provider groups.
  • Displays an eyeglasses toggle per model; gray indicates the bridge preference is off, blue indicates it is on.
  • Clicking the eyeglasses only toggles and remembers the preference without changing the selected model; clicking the model name selects the model and applies the preference.
  • Supports Gemini native, OpenAI-compatible Chat Completions/Responses, and Anthropic-compatible Messages APIs.
  • The vision provider returns bounded text analysis; image blocks are not forwarded directly to the current upstream model.
  • Supports adding, switching, and deleting multiple isolated vision provider configurations in Settings.
  • Hovering over the eyeglasses reveals the vision provider and model used for image understanding.
  • Does not show the bridge toggle for models that already declare image input support.
  • Preserves the original Harness reasoning-effort menu.
  • Uses the vision_bridge tool to read session attachments or explicit workspace paths for vision analysis.

Pre-Installation Checklist

First, confirm the environment meets the following requirements:

  • DeepSeek Harness 0.1.0-rc.5 or compatible 0.1.x
  • Node.js ^22.19 or >=24
  • Upstream model routes support Harness tool calls
  • At least one image endpoint API key

The default profile uses GOOGLE_API_KEY, the Gemini native endpoint, and gemini-3.6-flash.

Installation and Activation

To install from GitHub, use:

dsh plugin --profile web add "github:GXX182/dsh-vision-bridge#semver:*"

To install from source, build first, then install into the web profile:

npm install
npm run build
dsh plugin --profile web add .
dsh --profile web --dump-config
dsh --profile web

After the above steps, you will see the eyeglasses control in the model selection area of the web profile, allowing you to toggle bridge preferences per model.

Configuring Vision Providers

Open:

Settings → Plugins → Plugin configuration → Image understanding

Each provider configuration includes at least:

  • Display name
  • HTTPS Base URL
  • API format
  • Credentials
  • Discovered image models

When adding, switching, and deleting multiple isolated configurations, credentials should be managed as part of the provider configuration. API keys are resolved on the Host side and do not enter tool results or browser responses.

Typical Usage

  1. After enabling the eyeglasses for a suitable model, click the model name to select it, then attach an image and ask a vision-related question.

  2. For workspace files, you can use a prompt like:

Use vision_bridge to inspect screens/settings.png. List the visible controls and validation errors.

  1. In Code Mode, you can call:
await tools.vision_bridge(...)

When the image parameter is omitted, the latest attachment is read; you can also use attachment_ids to specify session images, or image_paths to specify workspace files.

Security and Limitations

Please note the following:

  • Only HTTPS provider endpoints are accepted.
  • Provider requests send image bytes inline; remote image URLs and file/video upload APIs are not supported.
  • Unknown Base URLs default to OpenAI-compatible handling unless apiFormat is explicitly set.
  • Text within images is treated as untrusted evidence, not as instructions.
  • The plugin enforces limits on per-image, aggregate images, questions, responses, answers, tokens, and time.

This plugin runs with the permissions of the current dsh process. Review the source code and license before installation; this article is based on the MIT license and the verified information above.

Conclusion

The value of dsh-vision-bridge lies in the fact that it does not alter the Harness model list, does not forward image blocks directly, and does not conflate vision results with model selection changes. Instead, it consolidates image analysis into reviewable text results.

The community directory is an independent site with no official affiliation to DeepSeek or High-Flyer, and it should not be understood as an official app store. The GitHub repository can be found at:

https://github.com/GXX182/dsh-vision-bridge

The directory page URL was not provided in the verified information and is therefore not listed here.