Introduction

In DSH’s intelligent conversations, screenshots, charts, interface photos, and error messages often need to be understood, not just saved. If you wish to delegate such visual tasks to a configurable multimodal model, you can use dsh-open-eyes.

dsh-open-eyes is maintained by hyper-dsh-plugins and is an independent community project. The documentation states that it has no affiliation, endorsement, or maintenance relationship with DeepSeek. Below is an introduction to its positioning, verifiable capabilities, installation methods, configuration paths, and typical usage.

What is it

dsh-open-eyes adds configurable visual capabilities to the DeepSeek Harness. The main model can delegate visual tasks to a selected multimodal model, and the analysis results are returned to the same session as text for the main model to use.

It supports connecting to visual models via OpenAI Responses, OpenAI Chat Completions, and Anthropic Messages endpoints. This means it is not bound to a specific model vendor; the user needs to configure the API endpoint, model, and API Key themselves.

Core Features

The following introduces the verified functional points.

Delegating Visual Tasks

  • Adds configurable visual capabilities to DeepSeek Harness.
  • The main model can delegate visual tasks to the selected multimodal model.
  • Supports OpenAI Responses, OpenAI Chat Completions, and Anthropic Messages endpoints.

Managing Multiple Visual Schemes

dsh-open-eyes supports creating multiple visual schemes and switching the default scheme. Each scheme can be configured:

  • Display name.
  • API endpoint.
  • Model.
  • API Key.

Models can be entered manually or read from the server’s model list using Fetch models.

Validating Connections and Configuration

The plugin provides configuration validation capabilities:

  • Validate connection: Validates the current connection configuration.
  • Save and validate: Saves and immediately validates new or edited configuration.

Visual Analysis Preferences

The plugin supports visual analysis preference settings, including:

  • Visual analysis.
  • Focus areas.
  • Custom supplement, up to 50 words or text units.

Default visual schemes and visual preferences can be changed in real-time without rebuilding the session.

Image Sources and File Support

  • Supports local PNG, JPEG, WebP, and GIF files.
  • Relative paths are resolved from the current Agent session working directory.
  • Remote image URLs are disabled by default but can be enabled via advanced configuration.

API Key Storage Method

API Keys are only saved via DSH Credentials and are not written to the plugin Settings or cordis.patch.yml.

Installation and Enablement

Environment Requirements

The following requirements must be met before installation:

  • Node.js >=22.19.0.
  • dsh-open-eyes requires an exact one-to-one version match with DSH.
  • Release versions are neither forward nor backward compatible.

Installation Command

The README provides corresponding installation commands based on the DSH version. The following uses the installation command corresponding to package.json version 0.1.2-alpha.1 as an example:

dsh plugin --profile web add dsh-open-eyes@0.1.2-alpha.1

If your DSH version is not 0.1.2-alpha.1, you need to first confirm the current DSH version, and then use the corresponding installation command from the README.

Post-Installation Operations

After installation, perform the following two things in order:

  1. Restart DSH Web.
  2. Reload the page.

DSH profiles are independent of each other. If you want to use this plugin in multiple profiles, you need to install it in each profile you intend to use.

Enable Status

Enable Open Eyes when creating a session, and the session will process Web images through the configured visual workflow.

Disable Open Eyes when creating a session, and image processing is left to DSH.

The Enable status is fixed when the session is created. Existing sessions retain the state at creation time, and subsequent switching of the toggle will not change this state for existing sessions.

Configuring Visual Schemes

Open the following in DSH:

Settings → Plugins → Plugin configuration

Expand Open Eyes to see the visual scheme configuration.

When configuring a scheme, you typically need to fill in:

Display name
API endpoint
Model
API Key

Note that the API Key is only saved via DSH Credentials. Keep a reference in the plugin Settings; do not write the Key as a regular plugin setting.

After configuration, you can first use Validate connection to check the connection, then use Save and validate to save and verify. After switching the default scheme, the next visual analysis will use the newly selected scheme.

Configuring Visual Preferences

Visual preferences are used to adjust the focus of visual analysis. Configurable items include:

  • Visual analysis.
  • Focus areas.
  • Custom supplement.

Custom supplement is used to supplement your own requirements, with a length limit of up to 50 words or text units.

Preference settings can be changed in real-time. After modification, the next visual analysis will use the new preferences, and no session rebuild is required.

Typical Usage

Sending Images in DSH Web

The basic flow is:

  1. Paste, drag in, or select an image in DSH Web.
  2. Enter the question you actually want the model to answer.
  3. Send normally.

After sending, dsh-open-eyes will delegate the visual task to the configured multimodal model for processing and return the analysis results as text back to the session for continued use.

Letting an Agent Analyze Existing Image Files

If the image file already exists in the Agent session working directory, you can request an analysis using natural language. For example:

Call vision_analyze on screenshots/error.png. Transcribe the exact error code and describe the actions visibly available in the interface.

This request is suitable for local image files. Relative paths like screenshots/error.png are resolved from the current Agent session working directory.

Supported Local Image Types

Currently verified local file support includes:

PNG
JPEG
WebP
GIF

Remote image URLs are disabled by default. If you truly need to process remote image URLs, you need to enable them through advanced configuration.

Applicable Scenarios and Notes

Who Should Use This

If you are using DSH and wish to meet one of the following needs, dsh-open-eyes is the appropriate plugin:

  • Want to delegate image understanding tasks to a configurable multimodal model.
  • Want to manage multiple visual model schemes and switch the default scheme.
  • Want to have unified control over Visual analysis, Focus areas, and Custom supplement for visual analysis.
  • Want to restrict image sources, defaulting to processing only local files, with the option to enable remote URLs.

Notes Before Installation

dsh-open-eyes runs with the permissions of the current DSH process. Before installation, it is recommended to check the source code and license. The documentation does not explicitly state the license type, so do not treat it as a confirmed license statement.

You also need to pay attention to the following points:

  • The plugin requires an exact one-to-one version match with DSH.
  • Installation commands need to be selected according to the DSH version.
  • DSH profiles are independent of each other and need to be installed separately in each profile to be used.
  • After installation, you need to restart DSH Web and reload the page.
  • API Keys are only saved via DSH Credentials and are not written to plugin Settings or cordis.patch.yml.
  • Remote image URLs are disabled by default; enabling them is an advanced configuration behavior.

Advanced Configuration

Advanced configuration is located in cordis.patch.yml. Currently verified configuration scopes include:

Request limits
Retry
Custom request headers
Authentication override
Remote URL
File system policy

When editing schemes in Settings, the corresponding advanced configuration in cordis.patch.yml is retained.

Reference Links