Skip to content

Agent Panel: Undo / Redo around checkpoints - aka checkpoint restore is irreversibly destructive #28676

Closed

Description

@lougreenwood

Summary

Checkpoints are nice... but since we can't manually edit the thread like a usual Zed Buffer like in the old AI panel, checkpoints are the only way to walk a thread back to a previous state and save the context size.

But... when restoring to a checkpoint, all work is lost and unless everything is committed to git before checking out - all of the work is lost.

This is to be expected, since the checkpoint was restored - however, there is no way to undo the checkpoint restore and revert back to the thread and file changes before the checkpoint restore and recover the work.

Description

Steps to trigger the problem:

  1. Build up a long context, make a lot of edits with the ai agent
  2. don't commit the work
  3. restore to a previous checkpoint
  4. see that all of the previous work was lost - realise you want the previous work back
  5. shit your pants
  6. go looking for an undo for the "checkpoint restore" to get the work back
  7. see there isn't any
  8. cry 😢

Actual Behavior: after restoring to a check point, there is no way to "undo" the checkpoint restore
Expected Behavior: if a user accidentally restores to a checkpoint, the user should be able to undo this and see that both the thread reverts to it's previous state AND the file edits revert back to their previous state before the checkpoint was restored (aka the original, long context at the end of point #1 above)

Zed Version and System Specs

Zed: v0.182.3 (Zed Preview)
OS: macOS 15.3.2
Memory: 32 GiB
Architecture: aarch64

Activity

  1. added
    area:aiRelated to Agent Panel, Edit Prediction, Copilot, or other AI features
    on Apr 14, 2025
  2. added theissue type on Apr 14, 2025
  3. JosephTLyons commented on Apr 22, 2025

    @JosephTLyons
    Member

    Hey @lougreenwood - thanks for flagging. We've seen this as well and we'll be working to improve the checkpoint system.

  4. lesderid commented on Jun 20, 2025

    @lesderid

    Perhaps an undo tree like in vim would be a good way to handle this?

    Related to this, it's rather inconvenient that editing a message to the agent is destructive: in the ChatGPT interface for example, this essentially just creates a branching point instead.

  5. rohan-patra commented on Jul 6, 2025

    @rohan-patra

    Maybe just create a new thread if an earlier message is edited?

  6. saatvik333 commented on Oct 12, 2025

    @saatvik333

    Just experienced this issue. T_T all my work is gone

    No way to recover what's lost too. AGGGHHHH...

  7. tifandotme commented on Oct 12, 2025

    @tifandotme

    agree with @rohan-patra

    would make more sense to create new thread if editing earlier prompt.

    I guess this would also require zed to remember thread's changes even if we switch threads from history?

  8. Kreijstal commented on Oct 13, 2025

    @Kreijstal

    I literally screamed FUCK when I noticed that checkpoint restore triggered, unintentionally by changing the past conversation because I am trying to be careful of the LLM context window by not being a token glutton and actually managing what I send the llm

  9. hugodemenez commented on Oct 22, 2025

    @hugodemenez

    @saatvik333 same here today lol. Just looked at reported issues in GitHub. Please we need a fix / improvements here... That's pretty much the most important part of agent coding, being able to revert / compare changes.

  10. alexugthub commented on Nov 8, 2025

    @alexugthub
  11. added
    state:needs triagetriage is not complete: no `area:*` label, or no `severity:*`/`reach:*` for Bugs/Crashes
    on Nov 27, 2025
  12. added
    state:needs reproNeeds reproduction steps and/or someone to reproduce
    and removed
    state:needs triagetriage is not complete: no `area:*` label, or no `severity:*`/`reach:*` for Bugs/Crashes
    on Dec 24, 2025
  13. 1 remaining item

  14. added
    state:reproducibleVerified steps to reproduce included or someone on the team managed to reproduce
    reach:many usersBugs that happen for at least a third of the users across all platforms and kinds of usage
    severity:S1Security holes w/o exploit, crash, install/update, sign-in, badly broken big features
    on Dec 24, 2025
  15. github-actions commented on Mar 27, 2026

    @github-actions

    Zed development moves fast and a significant number of bugs become outdated. If you can reproduce this bug on the latest stable Zed, please let us know by leaving a comment with the Zed version, it helps us focus on the right issues. If the bug doesn't appear for you anymore, feel free to close the issue yourself; otherwise, the bot will close it in a couple of weeks. But even after it's closed by the bot, you can leave a comment with the version where the bug is reproducible and we'll reopen the issue. Thanks!

  16. lesderid commented on Mar 27, 2026

    @lesderid

    It's wild to me that a confirmed reproducible, P1 priority bug can get auto-closed by a bot.

  17. Kreijstal commented on Mar 27, 2026

    @Kreijstal

    deleting all your work is not a priority [wontfix]

  18. github-actions commented on Jul 3, 2026

    @github-actions

    Zed development moves fast and a significant number of bugs become outdated. If you can reproduce this bug on the latest stable Zed, please let us know by leaving a comment with the Zed version, it helps us focus on the right issues. If the bug doesn't appear for you anymore, feel free to close the issue yourself; otherwise, the bot will close it in a couple of weeks. But even after it's closed by the bot, you can leave a comment with the version where the bug is reproducible and we'll reopen the issue. (This bot will only ask about this issue once) Thanks!

  19. sruggier commented on Jul 3, 2026

    @sruggier

    Still not stale.

  20. added
    never staleto keep the stalebot away
    and removed
    staleLabel used by `stale` action
    on Jul 10, 2026
  21. zelenenka commented on Nov 2, 2026

    @zelenenka
    Contributor

    we are removing the checkpoints in #65989.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:aiRelated to Agent Panel, Edit Prediction, Copilot, or other AI featuresnever staleto keep the stalebot awayreach:many usersBugs that happen for at least a third of the users across all platforms and kinds of usageseverity:S1Security holes w/o exploit, crash, install/update, sign-in, badly broken big featuresstate:reproducibleVerified steps to reproduce included or someone on the team managed to reproduce

    Type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions