dsh-herdr

In DeepSeek Harness, root agents and multiple sub-agents often run in parallel, each in its own phase—waiting for approval, actively processing, or fully idle. Herdr, as a multi-pane workflow manager, needs a single semantic status per pane to give you an at-a-glance picture: is this session blocked on approval, working through tasks, or sitting idle? dsh-herdr exists to bridge that gap. It runs inside the DSH TUI process and aggregates the state of every agent into one unified Herdr pane signal.

The mapping is straightforward: if any agent has a pending approval, the pane is marked blocked; if no approvals are pending and at least one agent is running, it shows working; when all agents are idle, the pane goes idle; and once the last agent is disposed, the pane is released. The plugin continuously listens for agent creation, status changes, disposal, and session events. On session restore it folds historical approval records so pending approvals are never lost after a restart.

On the communication side, the plugin prefers a persistent NDJSON connection over HERDR_SOCKET_PATH (Unix domain socket or Windows named pipe), avoiding a forked CLI process for every state change. If the socket times out it falls back to the Herdr CLI, and only logs a warning when both paths fail—never blocking agent execution. A cross-process epoch-based sequence number ensures that after a pane restart the counter does not reset to 1, and Herdr will not discard the report as stale. If you manage multiple AI workflow panes with Herdr and want real-time, zero-friction status feedback at the TUI level, this is a light yet practical fit.