Build permission prompts for agents from action scope, data class, user intent, risk level, and expiry