~/workflow

High-level goals enter the system through local audio or command input, then normalize into structured tasks for the orchestrator.

Parallel git worktrees and isolated environments prevent cross-task contamination and ensure clean agent execution.

Tasks are assigned to a fleet of agents across multiple runtimes, allowing parallel problem solving and model redundancy.

Global and project memories are kept separate, with progressive skill activation to reduce token overhead and preserve deterministic behavior.

Independent verification runs adversarial review, evidence capture, and regression checks before output is considered production-ready.