QuintinShaw/pi-dynamic-workflows ? reverse-engineered prompt

Build me a Pi package that takes one plain English request and turns it into a runnable workflow that can fan the work out across multiple subagents, route each task to the right model, verify the results, and return one clean answer.

I want it to feel like Claude Code style agentic workflows, with a background /workflows experience, a simple interactive panel showing progress, and support for resuming a run if it gets interrupted. It should keep intermediate work out of chat context, support isolated worktrees when agents need to edit safely, and track tokens and cost for each agent. Please include a way to save and reuse workflows, plus an opt in /ultracode style mode for standing deep work. If anything depends on current Pi docs or APIs, look them up online and make it work with the latest guidance.