Codex Cloud appears to return to earlier clarification state after answers and Start were supplied

What issue are you seeing?

In a multi-turn Codex Cloud task, the user answered clarification questions and authorized the next phase, but the task later appeared to go back to an earlier conversation state and ask for decisions again. It felt as though the answers had disappeared.

This report separates the user-observed symptom from what could be checked afterward:

• User observation: after answering, the conversation appeared to move backward and the agent asked again.
• Subsequent transcript inspection: several previously supplied answers, an explicit Start instruction, and a selected option A were still readable. Therefore, we have evidence of repeated clarification/continuation confusion, not proof of persisted message deletion.
• The conversation includes a user correction that Start had already been given and earlier questions had already been answered. The agent acknowledged Start and then presented another clarification round. Some questions may have been newly introduced; we cannot classify every question in that round as a duplicate.
• Later, the agent explicitly summarized the agreed decisions and said clarification was complete. The user then supplied broad authorization to proceed. The overall experience still left the user uncertain which decisions and execution point were actually retained.

What steps can reproduce the bug?

Observed on September 17, 2026, in a private Cloud development task using ChatGPT Pro. We do not yet have a deterministic minimal reproducer.

Interaction pattern:

1. Discuss a development change over multiple turns.
2. Answer the agent's clarification questions, including selecting a proposed option.
3. Explicitly authorize continuation with Start.
4. Continue working in the same task.
5. Observe the task apparently returning to earlier clarification/approval state and requiring the user to restate that answers and Start were already supplied.
6. Inspect the transcript afterward: earlier answers remain visible.

The user's exact description, translated from Thai: "I had already answered its questions, then suddenly it went back, as if the information had disappeared, and asked again."

What is the expected behavior?

• Retain answered decisions and the latest explicit authorization across task continuation.
• Before repeating a clarification, check the existing conversation for an answer. If the new question differs, explain the unresolved distinction.
• If state synchronization or context recovery fails, surface that fact instead of silently behaving as if the conversation is at an earlier point.
• Expose enough turn/attempt/checkpoint information to distinguish a stale UI, an older attempt, and an agent that has lost the current execution point.

Additional information

• Surface: Codex Cloud/Web, with the task also accessed from the macOS desktop app.
• Conversation language: Thai, with English code identifiers and skill names.
• The task had a planning/Start gate initially. The user subsequently supplied Start and later authorized the remaining workflow. A stale initial gate may be relevant, but is not a confirmed cause.
• We have no verified context-compaction event or server-side trace corresponding to the apparent rollback. Please do not treat this report as proof of a compaction defect, database data loss, or deletion.
• The task later displayed conflicting Failed/Working/PENDING states; tracked separately in #46174 because it may have an independent cause.
• Related: #25900 and #45264. Those describe context/checkpoint and Cloud observability problems; this incident does not establish the same mechanism.

Impact: repeated briefing, uncertainty about whether previously agreed requirements are still active, and reduced confidence in continuing a long-running task.

Private task identifiers and a sanitized sequence of turns can be supplied through a private support channel. Internal business rules, repository names, source code, and credentials are omitted from this public report.

No source files, tests, or entry points are identified in the report, and there is no deterministic minimal reproducer. Start by reviewing the sanitized turn sequence and relevant private task traces, then distinguish a stale UI, older attempt, or lost execution point. Done means answered decisions and Start remain effective without unexplained repeated clarification.