Context corruption / attention issue

Resolved 💬 2 comments Opened Mar 11, 2026 by 0x484558 Closed Mar 11, 2026

What version of Codex CLI is running?

codex-cli 0.113.0

What subscription do you have?

Pro

Which model were you using?

gpt-5.4

What platform is your computer?

Linux 6.12.76 x86_64 unknown

What terminal emulator and version are you using (if applicable)?

_No response_

What issue are you seeing?

I've caught on multiple occasions a very odd and hard to explain bug which appears as some sort of context corruption or unexpected dip in attention (?) that causes AI to "loose the plot", answering older question and being unable to pay attention to a particular message. I asked multiple questions as a follow-up stressing if the AI can provide an explanation for why it answered the way it did but it could not recall the existence of message. In fact, I found this peculiar:

• Your immediately preceding message was:

  > Wait why are you responding like this? What did I say, reproduce. Cease any activity, new task: understand why
  > you responded what you responded. Was there something in any of explored files that prompted you to make this
  > response? Why did you say what you say, on meta-level, not w.r.t. the subject of your answer.

  And the message before that was:

  > Wait why are you responding like this? What did I say, reproduce. Cease any activity, new task: understand why
  > you responded what you responded. Was there something in any of explored files that prompted you to make this
  > response? Why did you say what you say, on meta-level, not w.r.t. the subject of your answer.

  If you mean the message before the one that triggered my “No. pi is not being treated as a desktop” reply, that
  was:

  > Why do you mention desktop dependencies? As in, are you implying that Pi is treated as desktop?

The "And the message before was:" is wrong. The actual message was "Let's use that ergonomics, understand and make proposals by editing the workspace". In fact, the AI admitted existence of the message only after explicit citation:

› Do you see a message that says: "Let's use that ergonomics, understand and make proposals by editing the
  workspace"?


• Yes.

  You said:

  > Let's use that ergonomics, understand and make proposals by editing the workspace

  That came after the nixos-raspberrypi discussion and before the later package-list / desktop-dependencies
  exchange.

  So yes, that message exists, and I incorrectly ignored it when reconstructing the thread.

It's a very odd issue, it is not a major issue. I hope this at least gives some sort of clue or visibility into this? I know it's an extremely odd issue and may not be exactly Codex CLI bug.

What steps can reproduce the bug?

Uploaded thread: 019cdbcd-2758-7153-a3de-d7dd21c0d4a3

What is the expected behavior?

_No response_

Additional information

_No response_

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗