›_ Codex Issues
Codex Issues / #3365

Loss of Intermediate Steps on Interrupt in VS Extension

Resolved 💬 3 comments Opened Sep 9, 2025 by dpolivaev Closed Sep 10, 2025
💡 Likely answer: A maintainer (gpeal, contributor) responded on this thread — see the highlighted reply below.

What version of the VS Code extension are you using?

_No response_

Which IDE are you using?

Cursor

What platform is your computer?

Darwin 23.6.0 arm64 arm

What steps can reproduce the bug?

  1. Start a long-running agent task in the Codex VS extension.
  2. Interrupt/cancel the task before it finishes.
  3. Try to resume by sending “continue”.

What is the expected behavior?

  • Intermediate reasoning steps (actions and thoughts after the last user message) should remain visible after interruption.
  • Resuming with “continue” should take those steps into account instead of starting from scratch.

What do you see instead?

  • All intermediate steps are shown in the "cancelled" group.
  • Resuming causes the agent to ignore its prior thoughts and actions, starting a new reasoning chain.

<img width="414" height="630" alt="Image" src="https://github.com/user-attachments/assets/34ec8edc-03a7-4849-a904-d02f6e7fafe9" />

Additional information

In Codex-CLI, intermediate steps remain visible, and I can interrupt, adjust, and resume without losing context.
The lack of this functionality in the VS extension makes it significantly less useful for iterative work.

View original on GitHub ↗

3 Comments

gpeal contributor · 11 months ago

Can you elaborate what you mean? The behavior should be the same, in theory, at least.

dpolivaev · 11 months ago

Thanks for clarifying. I checked again and confirmed the model doesn’t lose its internal context after interruption. It still knows which files it has read and which tools it has used.

The issue turned out to be about the way this state is presented to the user:

  1. Labeling – All intermediate steps are currently shown as “cancelled”. That wording suggests they were discarded, even though in fact they remain in context. It would be clearer if they were marked as “stopped” (or similar), to indicate that the chain was interrupted but not lost.
  2. Summary after stop – At the end of an interrupted run, there’s no consolidated summary of what the agent has already accomplished. Without that, the only way to reconstruct its current state is to open the “cancelled” group and read through the steps manually. I also would not have started comparing the “cancelled” and “finished” groups at all if such a summary had already been given after the first stopped group.
  3. Duplication in UI – As shown in the screenshot I attached, the same reasoning block appears both under “Cancelled” and again under “Finished working.” This duplication made it look like the agent had restarted from scratch, reinforcing my (wrong) assumption that prior steps were lost.

Because of this presentation (everything under “cancelled,” no summary, and duplicated blocks), I initially assumed the agent had lost all its recent work, while in fact it hadn’t.

Could you adjust the UI so that interrupted steps are shown as stopped rather than cancelled and provide a summary of completed work after stop? That would remove the ambiguity, increase confidence, and make the progress easier to follow.

gpeal contributor · 11 months ago

@dpolivaev We can look into the copy here. As far as all of the sections. That is the model's chain of thought so don't worry too much about it duplicating work and expect things like that to improve with the models over time.