Ran into several situations where I'd be produce a useful test in claude while taking the implementation in a direction I rejected. /rewind skill exists but it discards everything after a checkpoint. This means I have to either start over or continue with the unwanted migration and then of course start over.

/recover skill creates an approval-based contract specifying what survives and what does not. It prepares a clean worktree and continuation prompt.

It cannot invoke or clear or invisibly start the replacement session

Currently in beta, tear it apart, leave me some feedback.

Curious to to hear from people who run long sessions how often this happens in their workflow and whether such a skill like /recover would be useful.