Repository navigation
Interrupting requests makes the model forget everything it has done #2554
Description
When interrupting a request because a model goes off rails, then adding an instruction, the model forgets everything it has read or edited thus far. It will always start again reading all the files, making the same edits again, wondering why some changes are already there etc.
Expected behavior:
Interrupting a request and adding a follow up request, the knowledge about what has been read and edited should be present in the context, allowing it to seemlessly pick up where it stopped.
Metadata
Metadata
Assignees
Labels
No labels
Activity
github-actions commentedon Sep 11, 2025
This issue might be a duplicate of existing issues. Please check:
Feel free to ignore if none of these address your specific case.
ribelo commentedon Sep 11, 2025
Yeah, I noticed it today and I can't tell if I saw it before. It basically makes work impossible or at least expansive
thdxr commentedon Sep 12, 2025
couldn't easily recreate but can you recreate and drop a share link so i can inspect? ty
rekram1-node commentedon Sep 12, 2025
related: #2494
ribelo commentedon Sep 12, 2025
@thdxr #2563
I made it before you asked. If I will have a session I can share, I’ll do it. For now, you can check if the code I fixed was actually broken, in my opinion it was.
henrikklee commentedon Sep 14, 2025
I cannot share the session of the project where it occurs the most. So far I haven't been able to create an optimal reproduction. It seems to be at least somehow model specific (esp. Grok Code and GPT 4.1).
rekram1-node commentedon Sep 14, 2025
this should actually be fixed on latest
Dax made some message structure changes
rekram1-node commentedon Sep 14, 2025
Going to close, if anyone has this issue on latest lmk and we can reopen
fix: preserve context after interrupts and propagate subagent interru…