A quick question that would help us narrow this down: are you working over Remote SSH, or on files that live on a network mount (NFS or similar)?
Connection drops in those setups are the most common cause we’ve seen for a checkpoint’s saved state getting corrupted, which then makes Restore silently do nothing while the git diff stays put.
We have a fix for exactly that corruption case targeted for 3.18.
Thanks! Next time a restore doesn’t take effect, could you:
Open the developer console first (Help → Toggle Developer Tools → Console)
Try to restore the checkpoint
Copy everything in the console containing “[composer]” and paste it here
A restore that worked prints “Processed file … for revert operation” lines. If you instead see “No original file state found” or an error, that tells us exactly which case you’re hitting.
Two more things that would help:
Do the affected files include any .ipynb notebooks? (Those aren’t supported for checkpoint restore.)
Does it tend to happen in long conversations, after reloading the window or returning to an older chat, or in sessions where subagents made some of the edits?