Structured protocol for AI agents working across many context windows: a .lra/ directory with an atomic feature list (id/priority/acceptance criteria/status) and a session progress log, plus a session protocol (read context → pick ONE feature → implement → test → checkpoint) and recovery from broken states. Based on Anthropic research on long-running agents. Triggers: 'lra', 'checkpoint', 'feature list', 'long running', 'продолжи работу', 'долгая сессия', 'план фич', 'статус проекта'.
