I didn't want to go back to allowlists and manually approving commands, but I also didn't want to risk losing data again, so I built Yolt (You Only Live Twice) as a failsafe.
Yolt is a lightweight CLI that watches your project directory and keeps restorable copies of files, so you can recover when an LLM deletes or overwrites them.
It focuses specifically on the cases that are hardest to recover from:
- File deletes
- File overwrites (where previous contents are lost)
Usage is simple:
cd ~/project-dir
yolt watch
Then if something gets deleted or overwritten: yolt restore path/to/file.ext
When you start watching a directory, it takes an initial snapshot. After that, it continuously captures changes and keeps a rolling history (default 30 minutes, configurable time window and disk space caps), so you can restore earlier versions.It still has limitations (it can't recover data if it wasn't already watching the project directory and it doesn’t handle external side effects), but in practice it's been enough to make me feel better about running agents with fewer guardrails.
Would love feedback, especially from those using agent workflows heavily. It's macOS-only right now, but should be pretty straightforward to port to Linux, as well.
Memory-safe, low overhead for a long-running foreground process, easy cross-platform path when Linux support is added.