kmmuntasir/pi-nested-subagents

This is a plugin for a coding environment called pi that adds the ability to run multiple AI agents working in parallel or in a hierarchy, with one agent able to delegate tasks to other agents below it. The design is inspired by how Claude Code handles subagents, so the tool names, calling conventions, and interface patterns feel familiar to anyone who has used that system. The core idea is that you can have an orchestrating agent that breaks down a large task and assigns parts of it to specialized sub-agents. Those sub-agents can, in turn, spawn their own sub-agents up to a configurable depth limit (the default is five levels deep). This makes it possible to build workflows like orchestrator, then task planner, then leaf workers, all operating at the same time, with results flowing back up the tree in the correct order. The plugin adds a live status widget inside the editor showing animated spinners, token usage, and colored status icons for every running agent. A FleetView panel lists all active agents and lets you navigate into any one to see its conversation in real time, send it a message to redirect its work, or stop it early. Agent types can be customized by creating a Markdown file with a YAML header that defines the system prompt, model choice, and which tools the agent is allowed to use. Other features include git worktree isolation so each agent works in its own copy of the repo, skill preloading to inject reusable instruction sets into agents, scheduled agents using cron expressions or relative time offsets, and an event bus that lets other extensions react to agent lifecycle events. This is a fork of an earlier project, adding safe deep nesting and a built-in verification tool for auditing the agent tree. The full README is longer than what was shown.