Preface

DeepSeek Harness (DSH) adopts a plugin architecture to provide flexible extension capabilities. When a single agent cannot easily handle complex tasks, multi-agent collaboration becomes the natural choice. dsh-agent-hub is a plugin designed for this purpose. It transforms the multi-agent collaboration process into a visual “group chat room”: requirements enter the room, deliverables leave the room, and the intermediate steps—task decomposition, dispatch, execution, acceptance, and knowledge consolidation—all occur within the message stream.

What It Is

dsh-agent-hub is a group chat–style multi-agent collaboration console maintained by user jax629321. It organizes multi-agent collaboration into a group chat room: requirements enter the room, deliverables leave the room, and all intermediate steps occur in the group message stream. The entire process is visible, traceable, and can be intervened in at any time.

Core Capabilities

Coordination Brain and Four-Question Decision-Making

Choose any member as the “brain.” The brain autonomously drives the process according to the state machine “decision → dispatch → result collection → acceptance → consolidation → next round.” In each round, it explicitly answers “Goal? Current status? Available members? Definition of Done (DoD)?” and outputs structured JSON to avoid vagueness.

Atomic Task Splitting and Directed Dispatch

Supports @member directed dispatch; the member list automatically appears while typing. Each member is assigned one atomic task, one deliverable, and one acceptance criterion. Dependency ordering and parallel execution without dependencies are supported.

Verification and Rework Loop

Reproducibility is the final acceptance criterion; acceptance criteria are checked item by item. If a result is not qualified, it is rejected and returned with a one-time, precisely targeted rework list. After it passes, the result is written to group memory and completion is announced.

Autonomous Optimization and Circuit-Breaking Mechanism

Group memory is persisted and reused across rounds and tasks, supporting multi-round autonomous iteration. A triple circuit-breaking mechanism (maximum rounds, single-round timeout, manual stop) is built in as a fallback.

Zero-Dependency Deployment

The built-in backend ships with the plugin, with zero external service dependencies and zero npm runtime dependencies.

Installation and Activation

It is recommended to enable it automatically with a single command:

dsh plugin --profile <你的profile名> add dsh-agent-hub

After installation, the plugin is automatically added to the plugin layer of the profile. After restarting DeepSeek Harness, an “Agent Collaboration Console” button appears at the bottom of the left sidebar.

Typical Usage

  1. Create a group and invite members: Create a group in the console, click “Invite Members” in the right-side member panel, and fill in base_url / api_key / model to add an OpenAI-compatible API member.
  2. Set the brain: Click “Set as Brain” for a member.
  3. Dispatch tasks:
    - Send instructions to a specific member using @member_name for directed dispatch.
    - Send instructions to the brain; the brain will autonomously decompose, assign, supervise, and accept tasks.
  4. Acceptance and consolidation: After checking “Automatically approve ongoing tasks,” the brain does not require manual confirmation in each round. During acceptance, deliverables are checked; if not qualified, they are rejected, and after passing they are written to group memory.

Notes

  • Environment requirements: DeepSeek Harness with a Web profile is required; Python 3.10+ is recommended.
  • Source review: The plugin runs third-party agent code. Review the source code before use.
  • Installation time: The first installation takes 1–2 minutes (create a venv and install dependencies).
  • Runtime behavior: It runs without a window throughout (using pythonw.exe).
  • Data retention: Uninstalling or reinstalling the plugin does not delete group data.