Introduction

DSH can extend its capabilities through plugins. For multi-agent scenarios, the difficulty lies not only in writing a persona for each agent, but also in deciding who can send messages to whom, who can only see downstream information, and who ultimately outputs the result.

The DSH Cluster Plugin provides a set of cluster modes: open the canvas in dsh web, use nodes to represent agents, use directed edges to declare allowed message flows, and allow agents to collaborate via cluster_view, cluster_send, and cluster_spawn.

What is this

The package name for the DSH Cluster Plugin is @lanxi266/dsh-cluster-plugin, the repository is https://github.com/Lanxi26/dsh-cluster, the maintainer is Lanxi26, and the license is MIT.

It is tailored for the cluster mode of DeepSeek Harness and provides three things:

  • Canvas node-graph
  • Agent management
  • Flow-constrained message routing

One-sentence positioning: It is a DSH plugin that visualizes multi-agent collaboration as a graph and limits message reachability.

Core Features

Canvas Node-Graph

In the cluster canvas of dsh web, you can perform the following operations:

  • Add nodes
  • Connect edges; edges can be one-way, two-way, or none
  • Bind nodes to agents
  • Rename nodes or agents

Here, “nodes” are not ordinary workflow blocks, but rather the runtime locations that host an agent.

Agent Management

Each agent has:

  • name
  • prompt, which is the persona
  • mode
  • An optional space

Agent data will be automatically saved to:

$DSH_HOME/cluster-agents/<id>/agent.json

The three modes are respectively:

mode Behavior
single Addressed via cluster_send, targeting the most recent run session of that agent
multi Addressed via cluster_spawn, creating a new instance each time
any Both cluster_send and cluster_spawn are available

Flow Constraint

Edges are directed and follow a “deny-by-default” policy: only declared flows can use cluster_send / cluster_spawn.

If there are no edges, the graph is completely open, making it suitable for initial workflow testing.

cluster_view only displays itself and its downstream agents. Therefore, downstream agents cannot see upstream information and can only propagate messages further down.

Runtime tools include:

  • cluster_view: View currently available agent instances and categories that can receive messages
  • cluster_spawn: Create a new agent instance and send the first message
  • cluster_send: Deliver a message to the inbox of a specific running instance

Export / Load

Graphs and agent properties can be exported together as .txt JSON (version v2). Once loaded at another location, agents can be restored and written back to disk.

Identity space

You can optionally mount a disk directory for a specific agent. The agent’s persona will be prompted to browse this space first. This allows materials such as Skills, scripts, long-term memory, personality files, and agent.md to be placed in the respective space of each agent.

Installation and Enablement

Prerequisites:

  • Node ≥ 22.19 (official DSH requirement)
  • The current profile already includes the official @deepseek-ai/dsh-base and @deepseek-ai/dsh-web-app bundles

First, add the plugin:

dsh plugin --profile web add @lanxi266/dsh-cluster-plugin

Then start the web profile:

dsh --profile web web

Alternatively, you can write:

dsh web

The plugin package itself is a bundle and introduces:

  • @lanxi266/dsh-cluster-agent-fs
  • @lanxi266/dsh-tool-cluster
  • @lanxi266/dsh-client-ui-cluster

Typical Usage

  1. Open dsh web, click the cluster icon in the bottom left to open the canvas.
  2. Add an agent in the Agents bar, for example with id math_teacher, fill in the name and persona, and select a mode.
  3. Double-click on the canvas to add nodes, bind the nodes to the agent, and then connect the nodes; dragging A → B indicates that A can send messages to B.
  4. Enter a node: the entry agent will receive the persona and workflow hint, and use cluster_view, cluster_send, and cluster_spawn to collaborate.
  5. The final linked cluster_view only shows itself; at this point, the agent is prompted to directly produce the final result, terminating the workflow.

Applicable Scenarios and Notes

Suitable for:

  • Need to visualize multi-agent collaboration within DSH
  • Need to restrict who can send messages to whom between agents
  • Need to export and load the graph and agent configuration
  • Need to mount an independent directory for an agent so it reads local materials first

Notes:

  • The plugin runs with the permissions of the current dsh process; you should check the source code, dependencies, and MIT license before installation.
  • An empty graph is completely open; before formal use, you should explicitly draw the required directed edges.
  • Agent data will be written to $DSH_HOME/cluster-agents/<id>/agent.json; include this path when backing up, migrating, and cleaning up.
  • The visibility of cluster_view extends only to downstream; do not rely on upstream information appearing downstream.
  • The author states that the project may still contain many bugs; it is recommended to verify key paths locally first.

Links

GitHub: https://github.com/Lanxi26/dsh-cluster

Package name: @lanxi266/dsh-cluster-plugin

The Community Directory page URL was not confirmed in this material; you can search using the package name above.