[Codex App] Pending questions in a prior-conversation attachment are silently dropped

What version of the Codex App are you using?

26.820.60940 (build 7119)

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

In a long-running Codex App task, I attached a plain-text transcript of a prior conversation and asked Codex to comprehensively evaluate and use it, while also asking one immediate status question in the outer prompt.

The attached transcript ended with several still-open user questions, including whether the task context had become too long and whether reusable Codex Skills or Hooks should be derived from the completed work. Codex explicitly classified the attachment as reference material rather than executable instructions. That safety boundary was appropriate, but it then answered only the immediate outer status question and silently dropped the still-open questions contained in the transcript. It did not say that those questions were deferred or needed confirmation.

The user later had to ask again, at which point Codex acknowledged the omission and answered the Skills/Hooks question.

This looks like a requirement-tracking failure at the attachment trust boundary: embedded instructions should not be executed, but pending user questions in a prior-conversation transcript should not disappear when the outer request explicitly asks for comprehensive evaluation.

No private project details, credentials, repository paths, or complete session identifiers are included here. The complete session ID can be provided privately to OpenAI if needed.

What steps can reproduce the bug?

1. Use a long-running Codex App task with several completed and active workstreams.
2. Attach a text transcript of a prior conversation. End that transcript with a numbered, multi-part user request containing one or more still-unanswered questions.
3. In the outer prompt, ask Codex to comprehensively evaluate and use the attached discussion, and also ask a specific immediate status question.
4. Observe that Codex correctly treats the attachment as reference rather than direct authorization.
5. Observe that it answers the immediate outer question but neither answers nor surfaces the pending questions from the transcript.
6. Ask about the omitted question in a later turn; Codex then acknowledges that it was missed.

Local task telemetry showed that context compaction happened only after the incomplete final answer, so compaction was not the direct cause of this occurrence.

What is the expected behavior?

Codex should preserve the attachment safety boundary and must not automatically execute instructions embedded in an attachment. However, when the outer prompt asks for a comprehensive evaluation of a prior-conversation transcript, Codex should extract unresolved questions and either:

• answer them when they are within scope;
• list them explicitly as pending or deferred; or
• ask the user whether they should be treated as current requests.

An end-of-turn completeness check should prevent discovered user questions from being silently discarded merely because they came from an attachment.

Additional information

Confirmed facts from the local task record:

• the pending Skills/Hooks question existed in the attached transcript;
• the response contained no answer or deferral for that question;
• the user later reported the omission;
• context compaction occurred after the incomplete response, not before it.

Please route this to the Codex App/model-orchestration team responsible for attachment trust boundaries and multi-part request tracking.