dsh-conversation-link

DeepSeek Harness keeps every opened conversation alive in the same process and lets plugins address any active agent, but it only exposes one cross-session path: parent session to subagent. When you spin up several peer conversations on one project and want them to coordinate, check progress, or enforce project-level constraints, there was no built-in way to do it. dsh-conversation-link fills that lateral gap: it lets conversations discover each other, exchange messages, read progress without interrupting, and set self-constraining standing rules—all without a prior connect step and without wrapping anything in a subagent.

Five capabilities cover the core loop. Discovery and addressing: every unarchived conversation in the current workspace is findable via a stable short handle or a custom nickname—no UUIDs. Bidirectional messaging: the first message to an unseen peer automatically establishes the connection and delivers in one call; subsequent sends reuse the link. Delivery defaults to auto, which picks steer for a busy peer (lands at the next step boundary) or queue for an idle peer (a clean dedicated turn). Non-intrusive progress reads: current turn/step, last human and assistant text, recent tool names, and queue depth—without waking the peer. Three-stage standing rules: before rejects an outgoing tool call, after rejects a completed result and feeds correction back, and input re-declares a constraint before every step. Rules bind only the conversation that declared them; no parameter in conversation_rule can point at another conversation. Peer spawning: conversation_spawn creates a new conversation at the same level as the caller, optionally links it, and assigns the first task in a single call.

There is no role hierarchy anywhere. A link is a name, not a permission; the only mechanism that can block anything is a rule, and that rule binds only the conversation that wrote it. Linked conversations are colleagues introduced to one another, not monitored processes. This plugin is aimed at developers who run several conversations in parallel on a single project—frontend, backend, testing, docs—and need those conversations to announce interface changes to each other, track progress without interrupting, and hold the line on project-level constraints even when no human is watching.