Introduction

A common direction in the DSH ecosystem is splitting capabilities into plugins, allowing the main agent to call external tools around specific tasks. For agent developers, the actual pain point is usually not “whether a model can be called,” but rather: how to split tasks among different experts, how to compare model outputs side-by-side, and how the text-based main agent can obtain image content.

Addressing these issues, dsh-collaboration provides a suite of collaboration tools for DeepSeek Harness: a user-configurable specialist roster, on-demand scheduling, model comparison, and a vision bridge. Models still come from the official provider process, and the plugin itself does not bundle an LLM adapter.

What is this

dsh-collaboration is an independent third-party open-source plugin suite maintained by Socialist-Sister, licensed under MIT.

Its positioning can be summarized as:

Multi-agent collaboration suite for DeepSeek Harness

It mainly provides the following capabilities:

  • Specialist roster: User-configured list of experts.
  • Team console: On-demand scheduling of specialist instances.
  • Model comparison: Send the same prompt to multiple models and view results side-by-side.
  • Vision bridge: Hand over images to vision models and return text analysis.
  • Image inbox: Paste images in a collaboration session, save them as workspace files, and bring the path into the draft.
  • One-line preset: Provide enableable collaboration presets.

Model integration does not rely on built-in adapters in the plugin, but uses DeepSeek Harness’s official process:

Settings → Models → Add provider

Core Capabilities

Specialist Roster

The Specialist roster is the foundation of this plugin. It allows the main agent to organize tasks around a set of user-configured roles, rather than calling a single model each time.

Typical scheduling entry points include:

team_call
roundtable

Where team_call is used to hire persistent specialist instances, and clones can be attached.

Team Console

The Team Console provides a set of collaboration tools for managing subsequent interactions of specialist instances.

Common tools are as follows:

team_call
team_message
team_status
team_close
roundtable

Their responsibilities can be understood step-by-step:

  1. team_call hires persistent specialist instances, optionally with clones.
  2. team_message is used for follow-up questions or relaying messages.
  3. team_status views the real-time team dashboard.
  4. team_close closes an instance.
  5. roundtable is used for one-time multi-expert discussions.

Model Comparison

The Model comparison tool is used to view outputs of different models for the same input horizontally (side-by-side).

The tool name is:

model_compare

The basic usage is: send the same prompt to multiple models and display results side-by-side. It is suitable for determining differences between different models on a specific task, rather than just getting a single answer.

Vision Bridge

The Vision bridge targets scenarios where the main agent is text-centric but still needs image analysis results.

The tool name is:

vision

Its workflow is:

  1. The main agent sends the image to a vision-capable model.
  2. The vision model returns text analysis.
  3. The main agent continues working based on this text analysis.

This brings multimodal capabilities back into the main agent’s text workflow.

Image Inbox

The Image inbox solves the problem of text-based main agents processing image pastes.

In a collaboration session:

Paste an image

Subsequently:

  1. The image is saved as a workspace file.
  2. The path appears in the draft.
  3. The main agent can continue processing the image based on that path.

It should be noted that text-only DeepSeek routes declare that they do not have an image modality, so pasting images may be rejected during the admission phase. The value of the Image inbox lies in converting images into workspace file paths that the main agent can continue to process.

One-line Preset

The One-line preset provides entry points for collaboration-related presets, allowing users to enable a set of collaboration tools instead of manually configuring them item by item.

For specific enabling methods, please refer to the instructions in the repository README.

Installation and Usage

The verified materials do not provide directly usable installation commands, so this article does not fabricate installation commands. Before enabling, please read the Quick start and roster-related instructions in the repository README.

The runtime environment requirements given by current materials are:

node >=20

Model integration still uses the official process:

Settings → Models → Add provider

After enabling, core tools can be called in the main agent using the names listed above: team_call, team_message, team_status, team_close, model_compare, vision, etc.

Typical Usage

Below are a set of examples that can be used in the order of tasks.

Assembling an Expert Team

First, use team_call to hire persistent specialist instances. Clones can be used to facilitate the same role handling multiple subtasks.

team_call

Continuing Communication

When an instance needs additional information or further processing, use team_message for follow-up questions or relaying.

team_message

Checking Status

Use team_status to view the running status of the current team instances.

team_status

Closing an Instance

Use team_close to close instances that are no longer needed.

team_close

Comparing Model Outputs

Use model_compare on the same prompt to view the answers of different models side-by-side.

model_compare

Getting Image Analysis

Use vision to hand over images to vision models and retrieve the text analysis results.

vision

Pasting Images into a Collaboration Session

Paste images directly in a collaboration session. The images will be saved as workspace files, the paths will go into the draft, and the main agent can continue processing those paths.

Paste an image in a collaboration session

Use Cases and Notes

Suitable for the following users:

  • Have registered model routes in DeepSeek Harness and wish to organize them into a collaboration team.
  • Need the main agent to be responsible for task division, follow-up questions, status checking, and result aggregation.
  • Need to compare outputs of multiple models for the same prompt.
  • Need the text-based main agent to read image content with the help of a vision model.

Notes before use:

  • The plugin runs with the permissions of the current DSH process. Check the source code and MIT license before installing.
  • The plugin does not bundle LLM adapters. Models must be integrated via the official provider process.
  • Text-only DeepSeek routes do not declare an image modality, so pasting images may be rejected during the admission phase.
  • Current materials do not provide complete installation commands and directory page URLs; refer to the repository README before enabling.

Links

GitHub:

https://github.com/Socialist-Sister/dsh-collaboration

If listed in community directories, refer to the actual information on the directory page; the directory page URL was not provided in the verified materials.