dsh-subagent-concurrency-limit

In DeepSeek Harness, the workflow engine tracks maxConcurrentAgents and maxTotalAgents for agent() calls inside a run, yet a root-level direct call to the subagent tool is essentially unconstrained at the process level. Fan out a few delegations, let them spawn nested children, and CPU and memory can spike before you even notice. dsh-subagent-concurrency-limit exists to close exactly that gap.\n\nThe plugin maintains a single shared permit pool inside the process, capped at two by default. A root Agent requesting delegation enqueues in FIFO order and waits until a permit is free. Nested children that invoke the same delegation tool also draw from the pool; when it is exhausted a child fails immediately rather than blocking, so a descendant can never deadlock the ancestor it is serving. A permit is released the moment the child turn goes idle, the run ends, or the Agent is disposed. For remote-process providers such as Codex, Claude Code, and ACP, the plugin can only observe the configured root-delegation admission and the subagent/start and subagent/end boundaries; nested execution inside the remote process is invisible.\n\nPick this plugin up if your Harness workflows frequently fan out to multiple subagents, if you register non-standard delegation tool names, or if you simply want to keep a modest hardware footprint from getting overwhelmed. It does not alter any existing behavior—it just adds a gate at the permit level, and the default of two concurrent runs is enough to keep most scenarios well under control.