dsh-hive

DeepSeek Harness ships with a subagent system that works well for ephemeral fan-out: a parent spawns a child, they share context, the child dies when done. But many real workflows need the opposite: several independent, long-lived top-level conversations that hand work to each other like colleagues, each thinking in its own context without polluting the other. dsh-hive exists for exactly that.

It exposes two model-callable tools. list_sessions returns every live conversation in the current process with its id, status, and working directory, so you can discover a target. send_to_session delivers a text message directly into that conversation and wakes it for a new turn immediately. The delivered message automatically carries a send-the-result-back-when-done instruction pre-filled with the session id of the sender, so the receiver works independently and then replies with a short summary; the sender is woken by that reply in its next turn. No human relay, no one sitting and waiting.

The plugin also handles the small details that make multi-conversation coordination feel natural: batch sends tag each reply with a task label so the coordinator can tell them apart; after a successful delivery the sender concludes its turn (send-and-yield) rather than blocking on a reply it cannot see; and self-delivery is explicitly rejected with a clear error to prevent infinite loops.

It suits workflow-orchestration setups where a coordinator conversation decomposes a complex task, dispatches subtasks to several worker conversations that each process them in their own context, and then aggregates the results. If you already run multiple parallel conversations in DeepSeek Harness, dsh-hive turns them from isolated sessions into a collaborative team.