A solution is to have the tool run first in a container or sandbox where any changes can be rolled back ("test mode").
Once the command looks promising on the test container, run the command again in a fresh container to confirm, and maybe see a list of affected files. Finally run on the main filesystem outside of any test container.
Edit: similar ideas with family snapshots already suggested elsewhere in the thread
Once the command looks promising on the test container, run the command again in a fresh container to confirm, and maybe see a list of affected files. Finally run on the main filesystem outside of any test container.
Edit: similar ideas with family snapshots already suggested elsewhere in the thread