LoopBreaker is an early-stage developer tool, not a hardened production system. Read this before enabling file rollback.
loop-breaker install writes .loopbreaker/config.json with "mode": "warn".
In warn mode LoopBreaker:
- Detects doom loops
- Injects steering text to the agent
- Does not modify or delete any project files
This is the recommended mode for real repositories.
| Mode | Behavior |
|---|---|
warn |
Detect + steer only (default) |
restore |
Overwrite changed files back to last healthy checkpoint |
full |
restore + delete files that did not exist at checkpoint |
Enable explicitly:
loop-breaker install --mode restore # or full- Checkpoints are in-memory snapshots, not Git. A wrong “healthy” checkpoint can restore stale code.
fullmode deletes files created after the checkpoint. Use only on disposable copies.- Prefer Git stash / checkout for production recovery when possible.
Session files may contain copies of tracked source files from checkpoints (.py, .json, etc.). Treat .loopbreaker/ as sensitive and keep it gitignored.
The scanner now ignores .loopbreaker/ so state files are not re-snapshotted.
Workspace-relative paths are validated before any read/write. Paths containing .. are rejected.
- Detection relies mainly on shell output parsing; not all agent loops run tests.
- Test counts are parsed naively from
N passed/N failedstrings. - Diff churn metrics are not populated from Cursor hooks today.
CURSOR_AUDIT_REPORT.md(removed) was not an independent audit — only synthetic self-tests.
See bug report template.