It seems like your approach depends on figuring out what the command is doing, classifying it into three tiers (deny/ask/approve) and then letting the agent run its command depending on that classification. Is that right?