Codex stops execution mid turn and answer an old unrelated question
Open 💬 15 comments Opened Mar 12, 2026 by nachoal
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
What version of Codex CLI is running?
0.114.0
What subscription do you have?
200usd pro
Which model were you using?
gpt-5.4
What platform is your computer?
Darwin 25.2.0 arm64 arm
What terminal emulator and version are you using (if applicable)?
ghostty 1.3.0
What issue are you seeing?
Codex randomly stops what it is doing in relation to the current turn input prompt i sent and sidelines into a totally unrelated previous message/turn input prompt. For example:
- codex: i found the issue, it is this: blabla.
- me: please fix the issue
- codex: ok im going to fix the issue "tool calls here"
- ...
- codex: you are right i need to find this issue <-- like it just recently read another message in the conversation
What steps can reproduce the bug?
Uploaded thread: 019ce301-4b99-7df0-89b8-e26ec90271cb
What is the expected behavior?
_No response_
Additional information
_No response_
15 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
One thing to note: this usually resets if you rewind the conversation, but I'm hitting a properly stuck case now. Rewinding and re-running the prompt doesn't unstick it; cancelling and then duplicating the prompt doesn't either.
Had to force compact manually to free it
019ce550-98a1-7b42-9a2c-08748632d6fa
The assistant failed to follow my latest instruction and instead repeated an explanation that had already been resolved and was no longer relevant. Specifically, I had already clarified that I wanted it to proceed with the previously defined action, but it regressed to reiterating background context instead of executing the requested next step. This created unnecessary friction, wasted time, and broke the expected continuity of the interaction. Please investigate why the assistant ignored the most recent operative instruction and reverted to an earlier explanatory state, even though that discussion was already settled.
I've hit a case with this where even manual compact doesn't get this to stop and the model genuinely can't proceed -- I'm forced to go into a cold context, it seems.
019cce30-1373-7020-9da0-d10d166907a2
I have the same problem. Codex responds to an old message, and the entire message chain following that old message is broken. It seems the problem occurs more often in large context windows.
I also have the same problem. Can't provide any more useful context, but just to add a +1 to this
same here, keeps happening to me when above the 256k context more often.
same problem, getting more freuqently lately
Same problem here. Has been the case over the past 2-3 months, and is becoming more frequent w/ 1M context window.
Same issue here been getting it with gpt 5.2-codex, 5.3-codex, and now 5.4 (even on 1M context). Usually happens right after 256k context.
Same problem on gpt 5.5 it happened on 3 rounds of a conversation before compaction even so...
Any updates on this issue?
🫠
I've been using 5.5 very actively since its release and can say that this issue is no longer present.
Now that you mention it, I believe I can say the same.