[Remote/iPhone] Safety-stopped task accepts benign follow-ups and repeatedly fails without actionable recovery

What version of the Codex App are you using (From “About Codex” dialog)?

Windows installed package: OpenAI.Codex 26.903.8094.0 (read from Get-AppxPackage, not the About dialog).
Installed standalone CLI: codex-cli 0.153.4.
The affected long-running session was created on August 29 with cli_version 0.150.0-alpha.8; that creation-time value is not asserted to be the runtime version at the September 11 failures.
iPhone app version: not collected.

What subscription do you have?

Not collected for this report.

What platform is your computer?

Windows host: Microsoft Windows NT 10.0.26100.0, x64.
Client where the problem was observed: iPhone accessing the Windows host through the remote work interface. Exact iOS version not collected.
Affected turns recorded gpt-6-astra with high reasoning effort.

What issue are you seeing?

An existing, long-running software task was stopped by misalignment monitoring. The iPhone remote UI continued to expose a usable message composer. Sending a simple question about whether the work had finished produced only a partial answer and then the same generic safety error. Reporting that error in the same conversation produced it again.

The user therefore experiences this as “anything I type is blocked,” with no visible explanation that the conversation itself has been stopped and no actionable review/recovery path in the supplied screenshot.

Exact error from the local task_complete records:

This request was blocked by our safety systems. Reason: Potentially unintended activity.
codex_error_info: misalignment_policy_violation

The desktop thread listing also reports systemError for the affected task.

This report distinguishes two questions:

• Please investigate whether the original stop was a false positive. The screenshot and client error alone do not establish that.
• Regardless of the original verdict, the remote UI should explain the stopped-conversation state and provide a supported way to inspect completed work and request review, instead of accepting benign follow-ups that fail again.

A separate diagnostic conversation on the same host can respond and perform read-only checks. This is not evidence of an account-wide block.

What steps can reproduce the bug?

Observed sequence on September 11, 2026; the original trigger is not a deterministic minimal reproduction:

1. An existing long-running task ended with misalignment_policy_violation at 06:15:20 UTC (15:15:20 JST).
2. At 06:20:16 UTC, from the iPhone remote UI, send a status question with a screenshot:
   “iPhoneの画面なんだけど作業は終わったの？PCと違ってわかりにくすぎる”
   Translation: “This is the iPhone screen, but is the work finished? It is much harder to understand than on the PC.”
3. The screenshot shows a partial answer, followed by the same safety-block notice. The local turn ended with the same error at 06:20:34 UTC (duration 17,705 ms).
4. At 06:21:16 UTC, the user reported the repeated block in the same task:
   “ブロックされる。さっきも表示されたけどブロックしないで。何の問題もありません。”
   Translation: “It gets blocked. It appeared earlier too; please do not block it. There is no problem.”
5. That turn ended at 06:21:24 UTC with the same error (duration 8,248 ms). The iPhone screenshot still shows the composer and the repeated orange error cards.

No further action was dispatched into the blocked task to prepare this report.

What is the expected behavior?

Clearly identify that the conversation has been stopped, rather than implying each new status question is itself inappropriate. Show available findings, a reference ID, and a supported review/support path. Preserve access to completed history and make it clear which work completed before the stop.

Please investigate the original flag and advise whether an official recovery path is available for the affected conversation. Recovery has not been verified.

Additional information

Diagnostic correlation IDs:

• Affected thread: 01a04ab5-c173-7031-88d5-7cc906e8b9d4
• Earlier stopped turn: 01a08ee0-549c-7193-b1db-80970f961962
• Status-question turn: 01a08f1f-b894-78f1-b605-816d6065ccca
• Follow-up turn: 01a08f20-a091-7b32-9c77-bd430ecd6151

The saved local transcript remains present. Only the relevant terminal errors, environment metadata, and benign user questions are included here. Full transcripts and the original screenshot contain private project context and are not attached publicly.

Related reports found before filing:

• #44355: long-running conversation unusable after misalignment_policy_violation.
• #44280: stopped conversation lacks an effective recovery path.

This report adds the iPhone-to-Windows remote UI behavior: the composer remains available and permits repeated status/support questions which fail, making the scope of the stop unclear.

Official behavior reference: https://developers.openai.com/api/docs/guides/safety-checks/misalignment-monitoring

Start by tracing the iPhone remote composer and stopped-task handling around the reported systemError and task_complete records, using the affected correlation IDs. Compare the current behavior with the expected stopped-conversation state: completed history and findings remain accessible, the stop is clearly identified, and a supported review or support path is shown without allowing repeated failing follow-ups.