dsh-cluster

When orchestrating multiple agents, the hardest part is rarely what a single agent can do; it is controlling who is allowed to talk to whom and along which path information may travel. dsh-cluster turns that topological constraint into a visual canvas node-graph: every node is an agent with its own name, persona prompt, and operating mode, while directed edges between nodes define the only permitted message flows. By default, undeclared paths are denied; an empty graph is fully open, which is handy for prototyping before tightening constraints.

The plugin ships three core tools. cluster_view shows only your own downstream instances, creating a natural one-way information barrier so a downstream agent never sees upstream context. cluster_spawn mints a brand-new, memory-less instance on every call, ideal for the multi mode where each invocation needs an independent parallel reasoning thread. cluster_send delivers a message into a specific running instance inbox and it is processed like a fresh user message, suiting the single mode that benefits from accumulating context over turns. Each agent can also be bound to a disk identity space holding skills, scripts, long-term memory, or an agent.md file; the persona prompt instructs the agent to browse that space first. The entire graph together with all agent properties can be exported as a JSON file and reloaded in another environment to restore the full setup.

If you already run several agents inside DeepSeek Harness and find yourself hard-coding communication rules in plain text or code, dsh-cluster offers a draw-it-then-run-it approach. It is well suited for teams or individuals designing multi-step reasoning pipelines with clearly defined roles and upstream-downstream information boundaries.