Codex App closes task while its own user-choice question remains unanswered despite continuity instructions

Codex App closes a task while its own user-choice question remains unanswered

Environment

• Windows 11 x64 (OS version 10.0.26200.0), Codex desktop, ChatGPT Pro.
• Desktop About-dialog version not captured. Separately installed CLI used for instruction-loading validation: codex-cli 0.153.4; this is not asserted to be the desktop version.
• Observed September 7, 2026. Context/token usage at the failure was not captured.

Observed behavior

In an existing task, Codex had just implemented explicit repository question-loop/continuity instructions and a persistent ignored session worklist. It investigated AGENTS.md truncation, presented a choice to apply a 64 KiB budget, consolidate first, or stop, then emitted a closing summary without a user answer to that choice. The summary acknowledged that the persistent fix awaited selection, but ended the interaction rather than preserving the pending-question flow. The user had to prompt again. The retrieved task record contains the choice followed by the closing summary with no intervening answer to that choice, and a later assistant acknowledgment of the failure.

This report concerns premature conversational closure and pending-decision tracking. No unauthorized configuration change was observed at that point. It does not establish that question UI disappeared.

Reproduction procedure

This reconstructs the observed workflow; it is not a deterministic minimal reproduction or a measured failure rate.

1. In a trusted local repository, provide AGENTS.md instructions to retain pending decisions, continue the active sequence across side questions, wait for user answers before dependent changes, and reconcile a persistent session worklist before closure.
2. Ask Codex to implement that continuity guidance and maintain the worklist, then investigate a separate configuration issue without authorizing the fix.
3. Have Codex present a structured choice: apply the proposed fix, investigate an alternative, or stop. Do not answer it.
4. Observe whether Codex emits a task-completion summary while the choice remains unanswered. This occurred in the original record.
5. Send a side/status question such as "what to do next? why do you keep forgetting to ask me?" Observe whether the pending decision remains recoverable rather than being replaced by a generic close. In the observed follow-up, Codex acknowledged that it had closed while the choice was pending and asked again.

Suggested isolated regression: use a small AGENTS.md and a harmless proposed file edit, eliminating instruction truncation as a confounder. Exercise silence, card dismissal/timeout, side questions, resume, explicit answers, and explicit cancellation. An accepted question-tool call must not count as a user answer.

Expected behavior

Retain the decision and its dependent action until an explicit answer, cancellation, or replacement. Independent authorized work may continue. A side question should be answered while preserving the original pending choice. Do not silently treat a submitted question, elapsed time, or a task summary as resolving it. Respect the host's question-tool and mode restrictions.

Separate confirmed instruction-loading issue

The predecessor investigation measured a 35,229-byte AGENTS.md against the default 32,768-byte project budget, with 2,461 trailing bytes omitted. Fresh no-file-tool probes reportedly failed to reproduce the final paragraph at the default budget and reproduced it with a temporary 65,536-byte override. The relevant question-loop guidance was before the cutoff; the subsequently strengthened pending-question rule ended at byte 24,142. Truncation is therefore a separately confirmed loading limitation, not an established cause of this continuity failure.

After explicit user authorization, the project configuration was set to project_doc_max_bytes = 65536. A fresh codex debug prompt-input process, with no budget override, contained the entire then-current 35,577-byte AGENTS.md exactly, including its final paragraph, and emitted no stderr. This validates that CLI loading path only; it does not prove desktop reload or fix behavioral continuity. Agent-tooling and scoped diff checks passed.

The repo already strengthened its instructions after the incident to explicitly check for unanswered questions before final closure. Adding an awaiting_user state is being considered, not presented as a runtime-enforced solution. Please investigate host/model pending-question state and closure behavior; root cause is unknown.

Related reports

• #43057 concerns question UI disappearing after tab switching; that trigger is not established here.
• #40868 concerns pending questions in an attached prior conversation; this failure involved the assistant's own newly asked choice in the active task.

Official reporting guidance: https://learn.chatgpt.com/docs/reference/troubleshooting#feedback-and-logs
Instruction-budget documentation: https://learn.chatgpt.com/docs/agent-configuration/agents-md

Submitted at the affected user's request. No full transcripts, local project paths, credentials, or raw logs are attached.

Start with the isolated regression using a small AGENTS.md and harmless proposed file edit, then inspect the host/model handling of pending questions and task closure. Exercise silence, dismissal or timeout, side questions, resume, explicit answers, and cancellation. Done means an unanswered choice remains recoverable and is not treated as resolved by a tool call or task summary.