Control and review
Undo agent changes safely
Undo is a scoped recovery tool, not a substitute for version control. FireCursor can restore the session checkpoint so changes from the controlled run can be removed when they no longer match the task.
Quick answer
The shortest useful version.
Stop the run, inspect the changed files, confirm the checkpoint belongs to that session, protect unrelated user work, restore, and then verify repository status before starting again.
Step by step
Do this with confidence.
- 01
Stop further mutation
Cancel or pause the active run before evaluating recovery. Continuing to edit makes the boundary harder to reason about.
- 02
Inspect scope
List every changed file and identify pre-existing user work, concurrent edits, or generated artifacts.
- 03
Confirm the checkpoint
Make sure the restore target was created for the session and project you intend to undo.
- 04
Restore and re-check
Run the recovery, then inspect repository status and the affected files. Start a new task with tighter boundaries if needed.
Recovery has limits
A session checkpoint can address tracked workspace changes in its scope. It may not reverse external messages, cloud changes, billing actions, published releases, or other side effects. Those require service-specific recovery.
If unrelated work overlaps the same files, do not assume a one-tap restore is safe. Escalate to a desktop review and preserve the user’s changes.
Prevent the next drift
- Reduce the number of editable files
- Add an explicit do-not-touch list
- Require a plan before implementation
- Set verification and stop conditions
- Use a separate worktree for riskier exploration