This is an ops protocol for running multiple AI coding CLIs in parallel in tmux.
Core idea: YAML inbox files as the source of truth (auditable via git diffs) + inotify-driven wakeups (no polling). tmux only gets a tiny “inboxN” nudge.
I’d love feedback on:
1) mailbox protocol design (what would you change?)
2) failure recovery / redo protocol
3) safe wakeup patterns for TUIs (avoid clobbering human input)