AI: Re-sending message to agent reverts files on disk! (checkpoints)

[niemeyer]

Summary

After mysteriously losing changes in a few different cases, I finally found out that Zed has been reverting files on disk to a prior state, when re-running a thread from an earlier point.

The problem is easier to convey (and believe) with a video:

Screencast.From.2025-09-05.10-54-52.webm

Expected Behavior:

Never ever silently revert content in files, even more after they've been written down to disk.

Zed Version and System Specs

Zed: v0.203.1 (Zed Preview)
OS: Linux Wayland ubuntu 25.04
Memory: 30.3 GiB
Architecture: x86_64

[notpeter]

I have experienced this as well, but had not been able to come up with steps to reliably reproduce.

But I think I've got something:

1. Run

mkdir -p /tmp/badcheckpoint
cd $_
touch README.md
git init
git add README.md
git commit -m "README"
echo "original" > example.txt
zed . example.txt

2. Open the agent panel and ask Please append 'robot' to example.txt
3. In the example.txt buffer replace original with changed.
4. Save the example.txt buffer
5. Click on the previous message in the agent and change it to "wait here".

Actual behavior: Zed agent panel reverts to the checkpoint, which discards the changes made to the buffer. Buffer is reset to original. Additionally, it now incorrectly is displaying the suggested changes as adding original to replace changed.

@niemeyer If you have explicit steps which better demonstrate this, please include them. Specifically you example doesn't show the incorrect keep/reject lines which matches what I saw previously. I imagine it depends upon the previous agent steps from prior to the video recording.

[niemeyer]

It's unfortunately much easier to reproduce here, and it doesn't require that the agent interact with the affected file in any way, which is somewhat concerning.

I've just re-recorded the video with a fresh environment, new git repository with single revision with README.md, and simply edit > chat > edit > chat, on a new file is enough to reproduce.

Here is the full interaction:

Screencast.From.2025-09-05.17-40-26.webm

[ConradIrwin]

Ugh, this is not great.

Currently we take a snapshot before letting the AI do anything (just in case), and whenever you edit a message we restore that Snapshot.

This is often what you want in terms of the Agent's behavior: by going back in time and giving it new instructions you're implicitly saying "that didn't go well, please try again" (as opposed to just adding more messages to the thread).

It is almost never what you want in terms of user edits though.

Unfortunately today we can't easily distinguish between the two (partly because we're not tracking everything well enough in-memory, and also because we can't really be sure if a file changes on disk if it was because of the AI running tools or you changing things).

The way cursor handles this is to prompt if you want to revert or not. I'm planning to copy that behavior so that:

• If there are changes since the snapshot, and we can't be sure that the agent made all of them, we prompt you with "do you want to restore the codebase back to this point in time?" or similar when you hit enter.

Does that seem like it would fix your problem?

It's not great, because I think we'll be prompting people more often than we really need to; but it seems better than losing your work.

[ConradIrwin]

This should be fixed by #37801.

Now if you want to restore a checkpoint you'll have to click "restore checkpoint", and editing a message will merely reject any pending agent edits.