Kimi CLI has a mechanism with checkpoints and the ability for the agent to revert to a checkpoint + a message of how to continue based on what went right/wrong. I don't know if that's the cause of what you've seen, but it's plausible.