Architecture & vision

rook is an experiment in the agent-native IDE: persistent workspaces, concurrent agent sessions, review, and attention management — built for a world where several agents work across repositories at once.

A few load-bearing decisions:

• A single command registry. Every action is a named command. Clicks, keybindings, and (later) the agent’s tool surface all dispatch the same commands, so the agent can do everything you can — by construction.
• A separate PTY host process. The UI can crash, rebuild, and reattach without killing shells — the tmux-server model.
• Shell integration (OSC 133) from the start. Semantic prompt marks give command boundaries, cwd, and exit codes as structure the UI and agent read.

For the full decision log, see the rook README.