I think the idea here is to move the responsibility layer away from the agent, rather than trust the CLI will behave and have to learn specific configs for each (given OP's tool works for any agent, not just Claude), this standardizes and centralizes it.