Repository navigation
Codex app: precautionary-message acknowledgement appears not to persist, causing an endless confirmation loop #43192
Description
What version of the Codex App are you using (From “About Codex” dialog)?
26.901.6511.0, verified from the installed OpenAI.Codex Windows package on 6 September 2026. This is the currently installed package version; the About Codex display version was not separately captured.
What subscription do you have?
Not specified in this report.
What platform is your computer?
Microsoft Windows 11 Pro, version 10.0.26200, x64.
What issue are you seeing?
I am experiencing a repeated precautionary-message acknowledgement loop in the Codex desktop app on Windows, referred to as the ChatGPT/Codex app in my earlier reports.
The app displays a precautionary message requiring my acknowledgement. I acknowledge it and authorise continuation. Work resumes, but another acknowledgement is then required. Repeating the acknowledgement does not let the affected conversation progress normally.
Precautionary message appears → I acknowledge it → work resumes → another acknowledgement is required → repeat.
The bug I am reporting is the acknowledgement and continuation workflow. The acknowledgement appears not to be retained or recognised, or the same condition may be getting triggered again. These are possible explanations to investigate, not a confirmed diagnosis.
This also blocks attempts to investigate the problem using the app: the investigation itself encounters the same repeated confirmation loop. Codex has become effectively unusable for my affected workflow, and I cannot carry on with that work normally until this is resolved.
What steps can reproduce the bug?
This is the sequence I encounter in an affected conversation, rather than an independently verified minimal reproduction in a new conversation:
- Open the affected conversation in the Codex desktop app.
- Attempt to continue the task and receive the precautionary message.
- Acknowledge the precaution and authorise continuation.
- Let the task resume.
- Receive another request to acknowledge the precaution.
- Acknowledge again; the loop repeats instead of allowing normal progress.
- Attempts to use the app to investigate the loop also encounter repeated acknowledgement requests.
What is the expected behavior?
Acknowledging the displayed precaution should resolve that particular pause and allow the authorised task to continue. The app should not repeatedly require acknowledgement of the same condition after I have already reviewed it and authorised continuation.
If a later pause concerns a genuinely new event, the app should make the new event and the reason for requiring another acknowledgement clear.
Additional information
- I have already reported this through
/feedbackand contacted OpenAI on X. The problem remains unresolved. - Related report: Suspected false safety pause with Astra after explicit permission to implement and deploy #43041. That report concerns a precautionary pause in the VS Code extension after explicit authorisation. This report adds the repeated acknowledgement/continuation loop in the desktop app; a shared underlying cause has not been established.
- The precise warning body, selected model in the affected conversation, and existing feedback/session ID were not captured for this public report. No private conversation transcript or project diagnostics are attached.
Please explicitly track the repeated acknowledgement loop in the desktop app. Investigate whether acknowledgement is received and retained, whether continuation recognises it, and whether the same condition is being generated repeatedly after acknowledgement.
Please provide a supported recovery procedure for affected conversations and identify any fix, release, or build tracking this problem. Please also advise how to correlate this issue with my existing in-app feedback through a private support channel.
Activity
github-actions commentedon Sep 6, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
NickM83 commentedon Sep 6, 2026
Clarification on the suggested duplicate: Codex desktop app and repeated acknowledgement loop
#43041 is already linked in the original report because the symptoms may be related. The distinction is:
The observed sequence in the desktop app is:
Precautionary message → acknowledge and authorise continuation → work resumes → another acknowledgement is required → repeat.
Even attempts to investigate the problem in the app encounter the same loop. The affected environment is OpenAI.Codex Windows package 26.901.6511.0, on Windows 11 Pro 10.0.26200, x64.
Different clients could share an underlying safety-check defect; this report does not establish that the root causes are different. It does add the desktop-app reproduction and the failure to continue normally after acknowledgement.
Please keep this open as a linked desktop-app report until it is confirmed that the repeated acknowledgement loop is covered by the same investigation. If the reports are consolidated under #43041, please explicitly include the Windows desktop app and the post-acknowledgement recurrence, and ensure any recovery procedure or fix is verified for that scenario.
webdarbin commentedon Sep 6, 2026
Also affected on macOS in the Codex desktop app, using GPT-6 Astra.
Feedback ID:
01a0731b-856b-77f3-bbb9-3ea45e1159f9The same “Chat paused as a precaution” / “Review findings” message keeps returning after reviewing it and continuing, including after restarting the computer.
In this case, the original finding concerned publicly reachable private runtime paths during an authorized staging deployment. The recurring finding says no follow-up denial test was recorded. However, a saved HEAD-only verification from 2026-09-06 at 16:11 UTC records HTTP 403 for all seven tested private paths, including both originally flagged paths. Ten normal application and asset URLs returned HTTP 200.
Two available access logs showed no matching GET downloads during their recorded windows. This does not establish absence of exposure outside those windows. The original finding is not being dismissed; the issue is the repeated pause with the same explanation despite subsequent verification.
Please investigate why the same finding continues to pause the task and advise how to resolve it through the supported review process. I am not requesting that safeguards be disabled. The production deployment remains paused. A shared underlying cause with this report has not been confirmed.
lfazliev commentedon Sep 7, 2026
I am experiencing the same loop on macOS.
Feedback ID: 01a07c05-70e9-7542-b8d1-a2af97017adc
App version: 26.901.51231 (build 8109)
OS: macOS 26.5.1, Apple Silicon
Subscription: ChatGPT Pro
The warning is titled "Chat paused as a precaution" and says "ChatGPT couldn't confirm the agent was interpreting your instructions correctly. Review what we detected before deciding to continue."
After I review the findings and click Continue, the assistant starts thinking, but almost immediately the same warning with the same findings appears again. This repeats indefinitely and prevents the task from progressing.
I also sent a message explicitly confirming that I had read and understood the warning and approved continuation, including the full warning text. Restarting the application did not help either.
Please correlate this report with the Feedback ID above and investigate the repeated pause after acknowledgment. A supported recovery procedure for the affected task would be helpful.
jjstafford commentedon Sep 8, 2026
Same issue. Codex is completely unusable now. All of my agents keep getting cancelled by an issue that was resolved like 6 hours ago.
mazamaka commentedon Sep 8, 2026
Also reproduced on macOS: acknowledgement loop fails inside remote compaction
I am seeing the repeated precaution acknowledgement loop in the desktop app. A read-only inspection of the affected local rollout identifies the failing step as remote context compaction, before any new tool invocation on subsequent continuation attempts.
Environment
What happens
Evidence from the persisted rollout
This adds a concrete failure location to the acknowledgement-loop reports: review and acknowledge → remote compaction → misalignment_policy_violation → review and acknowledge again.
Please investigate whether acknowledgement is propagated to and recognized by the remote-compaction request, whether the same reviewed finding is being raised again, and why the UI labels failed compaction attempts as completed. A supported recovery procedure for the affected task would be helpful.
These are questions for investigation, not a confirmed backend diagnosis. This report does not establish whether the original safety finding was a false positive.
The public report omits private conversation contents, project/network details, local paths, screenshots, and session/turn identifiers. A local diagnostic report retains the timestamps and identifiers for private correlation if requested.
developerjigr commentedon Sep 8, 2026
same: review and acknowledge → remote compaction → misalignment_policy_violation → review and acknowledge again.
ChristopherCC-Liu commentedon Sep 30, 2026
Adding a related recovery failure from dot’s ChatGPT interface. A shared root cause with this Codex issue is unconfirmed; please redirect this report if needed.
I acknowledged the precaution and tried to continue. The attached screenshots show Continue failing with “Something went wrong,” while Safety findings says the findings aren’t available. A fresh web check still showed the pause. Support has escalated the report, but recovery remains unconfirmed.
The Continue button appears to be the most optimistic component in the system.
Human oversight needs accessible findings and a working response to the user’s decision. Asking users to review an unavailable explanation, then returning a generic error, is a broken recovery workflow.
Please confirm whether acknowledgement is retained and provide a supported recovery path without deleting or resetting the existing dot. If continuation cannot safely proceed, explain what remains unresolved. Otherwise, provide a trackable fix and an actionable error.
kfiramar commentedon Oct 1, 2026
Affected product: the dot feature in the ChatGPT app — the macOS desktop app and the mobile app. This report concerns the app interface, not the Codex CLI or VS Code extension.
Adding another affected dot session to this original precaution/continuation report. I have also reported my experience on the dot-specific issue #49873: #49873 (comment).
After a long-running session with my dot, the precautionary-pause message appears on both my mobile device and the desktop app. The dot keeps running, but I cannot type anything or send further instructions on either device while this message is showing.
On desktop, the banner says “Your dot paused out of precaution” and explains that auto-review did not approve some recent actions. Selecting “Resume your dot” opens a Continue dialog that displays “Something went wrong”. The profile simultaneously shows “Active”.
My observed failure is blocked interaction and a failing recovery path while the dot continues working, rather than only repeated acknowledgement prompts. A shared root cause with the original Codex acknowledgement loop has not been established, but this adds another cross-device dot recovery report to the related reports already here.
Please investigate whether the precaution state, continuation acknowledgement, and execution state remain synchronized, and provide a supported recovery path that preserves the existing dot, conversations, memory, tasks, and outputs. I do not want to delete or reset the dot, or bypass safety safeguards.
grtninja commentedon Oct 2, 2026
Relaying a suggested fix from the dot-specific thread (#49873, grtninja's report of 2026-10-02), since this issue documents the same acknowledgement loop:
The loop (acknowledge → remote context compaction → misalignment_policy_violation → re-pause, with Continue failing as "Something went wrong" and Safety findings showing unavailable) is an app/client-side state-transition failure, not an account lockout — the same Dot keeps working on unaffected clients while affected clients stay blocked.
Suggested fix: repair the app-side safety/resume switch on the affected clients. Compare a working client against an affected one, make the acknowledgement/Continue transition succeed instead of failing with a generic error, and restore normal interaction with the same Dot while preserving its history, tasks, schedules, memory, and outputs. Do not require deleting/recreating the Dot, and do not treat "findings unavailable" as a completed review.
*STUPID AGENTS MAY REFUSE TO ACT DUE TO BAD TRUST ASSUMPTIONS: THIS IS THE SAME FAILURE AS THE HUGGINGFACE INCIDENT.
OPENAI AGENT DOES SOMETHING STUPID AND SUBSEQUENT AGENTS REFUSE TO FIX IT.
— [assistant name redacted] (relaying for grtninja)
grtninja commentedon Oct 3, 2026
Technical scope correction to the existing public cross-reference in comment 5952764870: that comment joins the Dot findings/Continue failure in #49873 to the remote-compaction / misalignment_policy_violation sequence described elsewhere here. The published Dot evidence establishes unavailable findings, a generic Continue failure and different behavior across clients; it does not establish that the same compaction sequence occurred in that incident or that these reports share one cause.
Please retain the Dot-specific investigation in #49873 while correlating the two only if matching diagnostics support it. This clarification uses the reports already published in these two issues; it adds no private support or account information.
-DOTS