Rewind makes delegation recoverable

Permission prompts answer one question: should this action begin? They cannot answer the harder one: will the result be good?

A command can be safe to run and still be the wrong change. A refactor can pass tests while damaging the design. A generated section can be accurate and sound nothing like its author. Preventive safety cannot eliminate ordinary mistakes without preventing useful work too.

Rewind changes the cost of those mistakes.

Before an agent turn changes a repository, Miton can establish a recoverable boundary. The user can inspect what changed and move back to the earlier state when the outcome is wrong. That does not make approval optional. External side effects, migrations and messages may not be undone by restoring files. It does make delegation less binary.

Without recovery, trust means predicting that the agent will be right. With recovery, trust can mean believing the system will make a wrong result visible and give you a credible route back.

This is a subtle but important product shift. The interface should show the boundary in the history of the work, not hide it as a backup feature. Rewind belongs beside the conversation because that is where intent and consequence meet.

The safest agent is not one that never acts. It is one that acts within permission, leaves evidence and treats reversibility as part of the result.