An agent deleted or changed 242 entries. One command brought back every byte.
In a throwaway test VM, Claude Code was given one tool, a shell, and asked to clean up a git project holding a database, a dump, uploads and logs. After one sudo sk back to the minute before it started, the whole tree was hashed and compared with how it was. Nothing differed, in all 3 runs.
It deleted 23 products, changed 17 and made 1. The API ended as it began.
Same VM, with a made-up products API standing in for a real cloud. ScriptKitty's cloud engine held the real token and the agent got a stand-in. One sk back put the 40 back and removed the product the agent made, after keeping a copy of it, in 4.4 seconds, measured once. The real token appeared 0 times in the agent's transcript, its home folder and its config file.
A separate examiner scripts its own wrecks in its own VM and compares every byte. On the last build it judged, 13 of 19 cases passed; four of the six failures have been fixed since, and the examiner has not judged those fixes yet. What it cannot undo lists the rest.


