DSH Cluster Plugin

Cluster mode for DeepSeek Harness — a canvas node-graph for orchestrating multi-agent collaboration and information flow. Every node is an agent (with its own name, persona, and mode); the directed edges between nodes decide who can message whom.

中文文档

The canvas

The three tools

• cluster_view — shows all running agent instances (sessions) and agent categories, so you can see who is available to receive messages.
• cluster_spawn — creates a brand-new instance of an agent category and delivers a first message to it (each call makes a fresh, memory-less instance).
• cluster_send — delivers a message into a specific running instance's inbox, which it processes like a new user message.

Features

• Canvas node-graph — add nodes, connect edges (one-way / two-way / none), bind agents, rename.
• Agent management — each agent has a name, prompt (persona), mode, and optional space; auto-saved to $DSH_HOME/cluster-agents/<id>/agent.json.
• Information-flow constraints — edges are directed and deny-by-default: only declared flows may cluster_send/cluster_spawn; an empty graph is fully open.
• Three agent modes:
  • single — address it with cluster_send (its most recent running session).
  • multi — address it with cluster_spawn (a fresh instance each time).
  • any — either.
• Export / load — export the graph and agent properties together (.txt JSON, v2), then load them elsewhere to restore agents and write them back to disk.
• Identity space — optionally attach a disk directory to an agent; its persona is told to browse that space first. Since the Agent panel on the canvas is intentionally simple, richer materials — Skills, scripts, long-term memory, personality files, agent.md, etc. — can live in each agent's own space.

Install

Prerequisite: the profile already has the official @deepseek-ai/dsh-base + @deepseek-ai/dsh-web-app bundles.

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

Start it:

dsh --profile web web   # or `dsh web`

Quick start

1. Open dsh web and click the cluster icon in the bottom-left corner to open the canvas.
2. In the Agents bar, add an agent (e.g. id math_teacher), give it a name and persona, pick a mode.
3. Double-click the canvas to add a node, bind it to an agent, then connect nodes (drag A → B means A can message B).
4. Enter a node: the agent receives your persona plus a workflow hint, and uses cluster_view / cluster_send / cluster_spawn to collaborate.
5. The last link (whose cluster_view shows only itself) is told to produce the final result directly, so the pipeline terminates.

How information flows

• An edge = an allowed flow. A → B means A can message B; B cannot message A back (unless the edge is two-way).
• cluster_view only shows yourself + your downstream, so a downstream agent cannot see upstream — it can only pass messages further down.
• No edges = fully open (handy to try things out before adding constraints).

Packages

Package	Role
@lanxi266/dsh-cluster-agent-fs	Host filesystem service: reads/writes agent.json, persists the graph, exposes the Remote endpoints
@lanxi266/dsh-tool-cluster	Host tools: cluster_view / cluster_send / cluster_spawn + persona injection
@lanxi266/dsh-client-ui-cluster	Browser canvas: cluster icon, node-graph, agent editor
@lanxi266/dsh-cluster-plugin	bundle: just a cordis.patch.yml that inserts the three rows above into a profile

Requirements

• Node ≥ 22.19 (official DSH requirement)
• Official DSH runtime: @deepseek-ai/dsh-base, @deepseek-ai/dsh-web-app (the @deepseek-ai/dsh-* series is provided by DSH itself)

License

MIT

Note

The author is a Java programmer, so this project was completed by DeepSeek-V4-Pro-High, costing 36.62 RMB / 5.39 USD (before the price increase). There may be many bugs — please be understanding, and feel free to fix bugs locally with Vibe-Coding. This project is open-sourced under the MIT license.