[Windows][Codex App][Blocker] Repeated freezes and full application crashes consumed 100% of my weekly Pro usage within 24 hours
What version of the Codex App are you using (From “About Codex” dialog)?
26.721.41059
What subscription do you have?
ChatGPT Pro
What platform is your computer?
Microsoft Windows NT 10.0.22631.0 x64
What issue are you seeing?
The Codex App on Windows has become unusable because tasks repeatedly freeze, abort, or crash the entire application.
This is a separate and broader issue in addition to a previously reported tool-output/context failure. That earlier issue produced the message:
“Output exceeded the available model context and was truncated”
even for a one-line Git command.
The current issue affects the complete desktop application:
- Tasks stop progressing and freeze.
- Tasks abort before completing any meaningful work.
- Failed task windows suggest restarting, updating, or creating another task.
- Newly created tasks also fail.
- The complete Codex application closes unexpectedly.
- After reopening it, the application may report that it is not working.
- I am then required to open the Microsoft Store and use Update or Restore before the application can be used again.
- After restarting, updating, or restoring the application, the same instability can return.
The problem is not limited to a specific repository or an old task. It also occurs in newly created tasks.
The local repository remains accessible in a normal terminal. No repository corruption, Git permission failure, or filesystem permission error is reported.
The application currently cannot be used reliably for development work.
Within approximately 24 hours, these repeated crashes, frozen tasks, restarts, and replacement tasks consumed 100% of my weekly ChatGPT Pro usage allowance while producing almost no meaningful project progress.
This is an unacceptable state for a paid Pro subscription. Please investigate both the application crashes and whether the usage consumed by failed runs can be restored or credited.
What steps can reproduce the bug?
Feedback ID: no-active-thread-019fa37d-cbcf-7f60-8cbd-29f6db561fc3
What is the expected behavior?
Codex should remain stable while completing development tasks.
A failed tool call or task should not freeze or terminate the entire desktop application.
If an internal error occurs, Codex should:
- stop the failed operation cleanly
- preserve the current workspace and task state
- display a useful diagnostic message
- allow the task to be retried
- avoid requiring the complete task context to be recreated
- avoid consuming substantial paid usage for repeated failed operations
- remain launchable without requiring Microsoft Store repair or updates
Creating a new task should provide a reliable recovery path.
Usage consumed by application crashes and failed internal operations should not be charged against the user’s weekly allowance.
Additional information
The issue occurred approximately 40 times within the last 24 hours across multiple tasks, restarts, and newly created chats.
This is not an isolated crash. It is a persistent failure pattern that has made Codex effectively unusable and consumed 100% of my weekly Pro usage allowance within approximately one day.
A separate concern is that Codex has created approximately 60 GB of application data on the system drive, even though my development workspace is located on another partition. The application does not clearly explain what these files contain, which projects they belong to, whether they are safe to remove, or how the storage location can be changed.
3 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Update: the application is currently running stably again after resetting the Codex configuration and testing the plugins individually.
What I did:
The application is currently working again, but I cannot yet confirm that the underlying issue is permanently resolved.
The behavior suggests that the crashes may have been related to a corrupted or inconsistent local runtime/plugin state, a stale background process, or the interaction between multiple enabled plugins rather than Cloudflare alone.
I will continue monitoring the application. I would not close this issue as resolved yet because the failure previously returned after restarts and Microsoft Store updates.
No repository files or Git working state were changed during this troubleshooting.
Update: unfortunately, the problems described in this issue are not resolved. The temporary stability I reported after resetting the Codex configuration did not last, and I am again experiencing serious problems with the Codex desktop app and with Codex itself.
I want to add some context because at this point the problem is broader than individual application crashes.
I am a ChatGPT Pro subscriber paying approximately €100 per month and I use Codex extensively for software development. Over the last days and weeks, the combination of application instability, poor output quality, unreliable transcription, repeated retries, and very rapid consumption of the weekly Codex allowance has made the service extremely difficult to use productively.
The application is still crashing
The crashes described in my original report have returned.
In addition, I am now repeatedly experiencing another particularly problematic failure: Codex crashes when I try to submit a bug report about Codex.
I open the bug reporting flow, describe the problem, attempt to submit the report, and the complete desktop application terminates.
The crash screen says:
“ChatGPT wurde unerwartet beendet”
and offers the options to update ChatGPT, open Config.toml, or restart.
I have reproduced this repeatedly, approximately ten times while trying to report the current problems.
This creates a rather serious circular problem: Codex is malfunctioning, I try to report the malfunction through the provided reporting mechanism, and the reporting mechanism itself causes the desktop application to crash.
I have attached a screenshot of the crash screen.
Codex output quality has also degraded significantly
The application crashes are currently only one part of the problem.
During actual development work I am seeing a noticeable deterioration in Codex output quality.
Codex frequently misunderstands instructions that should be straightforward, loses important context, produces nonsensical or partially gibberish text, and sometimes makes changes that are unrelated to what was requested.
More importantly for development work, Codex is regularly introducing new bugs while attempting to fix existing ones.
This creates repeated correction loops. I ask Codex to fix a problem, the result introduces another issue, I then have to explain the problem again, Codex performs another operation, and additional usage is consumed simply trying to restore the state that existed before the failed attempt.
This is not limited to one repository or one project.
I work with multiple repositories and tasks, and the general pattern has become increasingly noticeable.
Typical problems I am currently experiencing include:
I am currently using the latest Codex model available in the application.
Most of the time I use medium reasoning effort. I have also repeatedly switched to high reasoning effort because I hoped the additional reasoning would improve reliability.
Unfortunately, this has not consistently improved the quality.
In several cases I simply consumed substantially more of my available usage while still receiving incorrect, incomplete, or unusable results.
Voice transcription inside Codex is unreliable
The voice input and transcription functionality inside Codex is also currently causing substantial problems.
I use voice input frequently when explaining complex development tasks.
The transcription has become unreliable enough that words, technical terms, names, and sometimes complete parts of sentences are interpreted incorrectly.
This is not just a cosmetic transcription issue.
Sometimes the transcription changes the meaning of the instruction enough that Codex begins working on the wrong task or operates under incorrect assumptions.
I then have to stop the task, correct the misunderstanding, explain the instruction again, and spend additional usage.
This significantly reduces the usefulness of voice interaction with Codex, particularly for longer technical instructions.
Weekly Codex usage is being consumed extremely quickly
The usage situation is now one of my biggest concerns.
My weekly Codex allowance currently lasts approximately two days of active development work.
In practical terms, with a subscription costing approximately €100 per month, I currently get roughly eight days per month in which I can use Codex at my normal working intensity before running into the weekly limits.
That would already be difficult to justify if all of that usage represented productive development work.
However, a significant amount of the allowance is currently being spent on correcting Codex failures, repeating instructions, recovering from aborted operations, repairing bugs introduced by Codex, restarting failed tasks, and trying higher reasoning settings in the hope of obtaining usable results.
This means the effective amount of productive Codex work available within the Pro allowance is substantially lower than the nominal allowance suggests.
I want to emphasize that I am not simply complaining that I use Codex heavily and therefore reach a usage limit.
The issue is that a significant percentage of the paid allowance is currently being consumed by unsuccessful Codex operations.
If Codex misunderstands a task, produces broken code, crashes, freezes, or aborts and I then have to spend another run correcting the previous run, I am effectively paying usage for the failure and then paying usage again for the attempted recovery.
This becomes particularly problematic when using high reasoning effort, because a failed or incorrect result can consume a substantial amount of the remaining weekly allowance.
The previous configuration reset was only a temporary improvement
I previously reported that the application became stable again after I reset the Codex configuration, disabled plugins and runtime components, tested them individually, and then restored the configuration.
At that point I explicitly did not consider the issue resolved because the crashes had previously returned after apparently successful recovery attempts.
Unfortunately, that concern was justified.
The temporary improvement did not resolve the broader issue.
The current problems therefore cannot simply be considered fixed by resetting Config.toml or disabling and re-enabling plugins.
Those steps may have temporarily changed the runtime state, but the application and Codex experience have again become unreliable.
Why this is particularly problematic for a paid Pro subscription
I understand that individual model responses can fail and that software can occasionally crash.
The current situation is different because several independent reliability problems are occurring at the same time:
application crashes,
task freezes and aborted runs,
poor or nonsensical Codex output,
new bugs introduced during attempted fixes,
unreliable voice transcription,
repeated corrective runs,
very rapid usage consumption,
and now crashes while trying to submit a bug report about those exact problems.
Taken together, this makes Codex unreliable as a professional development tool.
For a service at the Pro subscription price, I would expect failed internal operations and application crashes to be exceptional cases rather than something that consumes a significant part of the available weekly usage.
Request
Please investigate not only the desktop application crashes, but also whether there has been a recent regression affecting Codex response quality, task reliability, context handling, and voice transcription on Windows.
I would also appreciate clarification on how usage caused by failed operations is handled.
Specifically, I would like to know whether usage consumed by application crashes, aborted tasks, internal failures, or repeated attempts required to recover from incorrect Codex output can be restored or credited.
In my case, this is not a small amount of usage. The failures have repeatedly contributed to exhausting essentially the complete weekly Pro allowance while significantly limiting actual project progress.
The service is currently not providing a reasonable relationship between the cost of the Pro subscription, the available productive usage, and the reliability of the results.
Current environment
Codex App version: 26.721.41059
Subscription: ChatGPT Pro
Platform: Microsoft Windows NT 10.0.22631.0 x64
The problems occur across multiple tasks and repositories and are not isolated to a single repository.
Normal terminal and Git access to the repositories continue to work, so there is no indication that repository corruption or normal filesystem permissions are responsible for the failures.
I am happy to provide logs, additional screenshots, task examples, or diagnostic information if this helps identify the underlying problem.
At this point, however, I would strongly prefer that this issue not be considered resolved based on the temporary stability after the previous configuration reset. The underlying reliability problems have clearly returned.