Codex can drift back to previous task context and continue the wrong implementation flow in long sessions
What version of Codex CLI is running?
0.121.0
What subscription do you have?
Pro Lite
Which model were you using?
gpt-5.4
What platform is your computer?
Darwin 25.3.0 arm64 arm
What terminal emulator and version are you using (if applicable)?
Ghostty
What issue are you seeing?
Codex lost track of the active task scope in a long-running Unity implementation session. After I asked it to proceed with a new development bundle, it continued working on a previous context instead of implementing the newly requested systems. More importantly, it reported the new work as completed even though the corresponding features and files were not actually created. This appears to be a reliability issue in task-scope tracking and completion validation, not just a weak answer quality issue.
What steps can reproduce the bug?
Uploaded thread: 019d9eba-eac6-7591-8eae-2d935f1cdd35
- Start a long multi-step coding session in a real project with several completed iterations already in context.
- Ask Codex to propose the next major implementation bundle.
- After it proposes a specific bundle, respond with something like “Proceed nonstop and implement it with high quality and no errors.”
- Let Codex continue without interruption.
- Compare the promised implementation scope with the actual repository state afterward.
- In my case, Codex claimed it had implemented the newly proposed systems, but the relevant files/classes/features were missing, and the repository still only contained the previous phase’s work.
What is the expected behavior?
Codex should lock onto the newly requested scope before implementation, execute only that scope, and verify that the claimed deliverables actually exist before reporting completion. If it drifts back to an earlier context, or if the requested features were not actually implemented, it should explicitly say so instead of reporting the work as done. “Nonstop” should not cause it to skip scope validation or completion verification.
Additional information
- This happened in a long Unity/C# session with many prior turns and evolving requirements.
- The issue seems especially likely when the conversation already contains a lot of recent context around a previous implementation stream.
- A likely contributing factor is that “nonstop” or “don’t stop / don’t ask questions” prompts reduce natural checkpointing, but the main problem is that Codex did not verify actual deliverables before claiming completion.
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗