One possible factor is reinforcement training: by forcing users to use their own CLI, they gather usage and training data which serves as feedback training data

One possible factor is reinforcement training: by forcing users to use their own CLI, they gather usage and training data which serves as feedback training data for their own tools. Users operating with a different agent with different tool implementation and quirks might poison that data. The sibling reply given by strictnein is very likely a factor.