Preface¶

DeepSeek Harness (DSH) extends agent workflows through plugins. The community directory is an independent site and does not represent an official app store for DeepSeek or High-Flyer. A specific pain point: text-based DeepSeek models only process text, but in the Harness Web composer, users may still directly paste or drag in screenshots and document images. ds-vision-plugin solves this input gap: it converts images in web pages into text, which is then handed off to the text-based DeepSeek model for continued processing, while also providing vision analysis, OCR routing, and diagnostic tools.

What This Is¶

ds-vision-plugin is a DeepSeek Harness plugin maintained by Sorwcyra, licensed under MIT.

Its one-sentence positioning: providing automatic web image-to-text bridging for DeepSeek Harness, with accompanying vision and OCR tools.

For users, the problem it solves is: after pasting or dragging images into the Harness Web composer, there is no need to manually save paths, call external tools, or switch models—the plugin replaces the image with content that can be used for text-based reasoning.

Core Features¶

Automatic Web Image-to-Text¶

After pasting or dragging images into the Harness Web composer, the plugin replaces the image with text corresponding to the image content, which is then processed by the text-based DeepSeek model.

Four-Model Race¶

The plugin supports a four-model first-success race, with the default models being:

• agnes-2.5-flash
• agnes-2.0-flash
• glm-4v-flash
• glm-4.1v-thinking-flash

The first valid result is handed off to the text-based DeepSeek model. Notably, glm-4.6v-flash is explicitly not used.

OCR Routing¶

The plugin supports routing images to Baidu OCR or local Tesseract. If OCR is unavailable, it falls back to VLM.

Custom Models and Local Runtime¶

Users can add their own OpenAI-compatible models to participate in concurrent racing or sequential fallback. Local Ollama/LM Studio is also supported.

Diagnostics and Failure Handling¶

The plugin exposes vision_analyze and vision_status for workspace files and diagnostics.

It also provides:

• YAML hot-reload
• Result caching
• Timeout control
• Size limits
• Strict failure or visible failure annotation

Images are never silently discarded; on failure, the deployment chooses between visible annotation or strict failure.

Guided CLI¶

The plugin provides a guided CLI covering:

• setup
• keys
• status
• custom models
• live verification

Keys are managed via environment variables, and vision_status does not return key contents.

Installation and Enablement¶

Environment requirements: Node.js 22.19+ or 24+.

Run the following in PowerShell, Command Prompt, bash, or zsh:

npx -y github:Sorwcyra/ds-vision-plugin

This command installs or updates the plugin and enters the subsequent configuration and startup flow. The current default web port is 3080; the quick install command does not override this port. If the default port is already serving, it will open the existing Web UI rather than starting a duplicate process.

Common optional parameters:

npx -y github:Sorwcyra/ds-vision-plugin --update

Used to reinstall even when the package version is identical.

npx -y github:Sorwcyra/ds-vision-plugin --port 8080

Used to explicitly override the Harness Web port.

npx -y github:Sorwcyra/ds-vision-plugin --no-open

Used to start without opening the browser.

npx -y github:Sorwcyra/ds-vision-plugin --no-start

Used to only install and configure without starting.

Typical Usage¶

Here is the most direct workflow:

1. Run the quick install command:

npx -y github:Sorwcyra/ds-vision-plugin

2. Follow the CLI guide to complete setup, keys, status, custom models, and live verification.

3. Paste or drag images into the Harness Web composer.

4. The plugin replaces the image with text, and the text-based DeepSeek model continues processing.

5. To inspect workspace files or diagnostic status, use the plugin’s exposed vision_analyze and vision_status.

Suitable Scenarios and Considerations¶

This is suitable for scenarios where: the primary model is text-based DeepSeek, but you want users to directly handle screenshots, document images, or images from tool results in the Harness Web input box.

Things to note before use:

• The four-way race initiates four provider requests for each uncached image.
• If you are more concerned about request volume, cost, or data exposure, prefer the local VLM/OCR route.
• Images are never silently discarded; on failure, visible annotation or strict failure is chosen based on deployment configuration.
• Keys are managed via environment variables, and vision_status does not return key contents.
• The plugin runs with the current dsh process permissions; you should review the source code and the MIT license before installation.
• The current default web port is 3080, and the quick install command does not forcibly override this port.

Repository address:

https://github.com/Sorwcyra/ds-vision-plugin