Codex seemed to re-answer an earlier prompt when I gave it a new prompt

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #14253
• #14408

Powered by Codex Action

Can confirm this. It's happening for roughly two days already. I had 1M context size before in config, removed it then as a supposed fix and it didn't help. Sent several feedbacks already via /feedback command.

Does this appear to be specific to gpt-5.4? Or are you able to repro this with gpt-5.3-codex as well? Just trying to understand whether this is model-specific behavior or a harness bug.

Haven't ever experienced it on anything other than 5.4 (xhigh both times). And only encountered it one time other than this like a week ago. In both cases I basically sent back "err I think you got confused and replied to an earlier question; <restate my last prompt as a summary>" and off it went.

The only thing that stands out is that I think in both cases, there was a long pause between interaction. So the last message was last night at say 8pm, and this morning was at ~10am. Same session, no resume.

I've experienced this too, only since 5.4/0.114 and have only been using 5.4 since the 0.114 release, but it was introduced by one of these. On earlier releases this has happened quite often on the first reply after compaction, but it's only wirh 5.4/0.114 that it has happened in the middle of a session, not right after compaction.

Hi, I run Agents Need Context and I am researching failures where a coding agent still has the model capability, but continuity drifts and it answers from stale session state.

Your report is useful because the wrong context seems to get reused after a pause even without an explicit resume, and other builders in the thread are seeing the same pattern.

If you are open to it, I would value one short postmortem here: https://www.agentsneedcontext.com/agent-failure-postmortem

I am looking for five things: what the agent was trying to do, what context or session state it answered from incorrectly, where the right context should have come from, what workaround you used, and what it cost. No pitch, just research.

I am also experiencing this, GPT 5.4. Codex app for macOS.

GPT-5.3-Codex is all good, it's just 5.4 for me that is showing this issue. I will revert to 5.3-Codex for now.

Same issue for me, having it for about a week.

This is becoming a major annoyance, I have to roll back the chat frequently. Makes me worried about context integrity in general.

5.4 xhigh on 0.114

@etraut-openai just had it happen again, and one commonality is this was a session I last interacted with about 4 days ago. I gave a prompt, it's first response clearly indicated it was working on what I'd just asked, but then the next thing I saw from it was clearly related to the prior prompt. That's another constant factor it seems: when the issue happens, it always appears to be the N-1 prompt that it starts responding to. If I do /feedback and get the GUID, can you view the thread from that? I don't mind providing log/transcript, although would prefer to do that offline if it's literally the log/transcript (email or otherwise).

this happens to me all the time now with 5.4, the product has become unusable as a result and this problem is getting worse and worse by the day

I am experiencing this same issue, but it is not just through the Codex Mac app; it is also via the CLI. It only seems to happen once I pass the 300k in context.