The caching strategy here looks really solid for keeping API costs down. Curious how it handles state invalidation when the agent context gets too large though.