Auto-compaction near high context usage can make Codex answer the previous message instead of the current one

What version of Codex are you using?

v0.120.0

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Windows

What issue are you seeing?

When context usage gets high and Codex triggers auto-compaction, Codex can sometimes answer the previous user message instead of the current one.

This is not just a normal misunderstanding. It looks like a conversation alignment bug after compaction. The assistant behaves as if it jumped back to the earlier turn and responds to the previous topic instead of the latest prompt.

This is severe enough that it happened during normal bug reporting itself: I was complaining about this bug, and Codex replied to my previous topic instead of the current complaint.

This is made worse by another issue in v0.120.0: I can no longer see the context capacity percentage, so I cannot tell when I am close to the auto-compaction threshold or avoid triggering it.

What steps can reproduce the bug?

1. Work in a long Codex session until context usage becomes very high.
2. Keep sending new messages across multiple topics in the same session.
3. Let Codex auto-compact near the high context threshold.
4. Send a new prompt about a different topic.
5. Sometimes Codex answers the previous message/topic instead of the current one.

What is the expected behavior?

After auto-compaction, Codex should still respond to the current user message, not a previous one.

Also, users should be able to see context capacity percentage (or an equivalent visible warning) so they know when auto-compaction is close.

Additional information

Observed symptoms include:

• I send a new message, but Codex replies to the previous topic.
• The reply does not match the current prompt and behaves as if it jumped back to an earlier turn.
• Sometimes it continues a topic that was already finished, instead of handling the newly submitted request.

Suggested areas to inspect:

• message/turn alignment after auto-compaction
• binding between the current user prompt and the compacted session state
• whether the assistant can continue using an outdated turn pointer after compaction

I previously reported the missing context percentage issue separately, but this auto-compaction / wrong-turn-reply behavior is the more serious reliability bug because it can lead to incorrect actions and decisions in coding workflows.