dsh-subagent-registry

Keeping a roster of specialist agents in dsh — a rigorous engineer, a nitpicking reviewer, a rubber-duck explainer — naturally leads to a folder of markdown files under ~/.dsh/agents/. But the main conversation has had no way to invoke them by name, and long-running tasks carry a painful risk: an API timeout or a crashed process means the next dispatch starts from zero, discarding every step already completed.

dsh-subagent-registry fills that gap. It scans the agents directory, reads each .md file (frontmatter metadata plus a markdown body used verbatim as the subagent persona), and registers them inside the use_agent tool so the parent conversation can dispatch any of them by name — no dsh core patching required. Before every dispatch, the plugin checks whether the same agent has a prior interrupted run; if one is detected, it resumes from the saved partial work instead of restarting. For long background tasks, setting background: true returns the child subagent id immediately while the parent keeps working, and ask_agent closes the loop by sending a follow-up message and waiting for the child reply, giving you full bidirectional parent-child conversation.

It is built for dsh users (0.1.2-rc.1 or later) who manage a team of markdown-defined agents, want long tasks to survive interruptions without losing progress, or need background subagents they can query interactively. Whether summoning a code reviewer before a commit, letting a rubber-duck explain a tricky module step by step, or dispatching a research agent that reports back when done, this plugin wires the whole flow together.