Windows Codex tasks repeatedly get stuck in inProgress, reconnect, and end as interrupted; feedback upload also fails

Summary

Codex desktop tasks have repeatedly become stuck over several days during local game development.

The affected task remains in active / inProgress, while the UI shows “reconnecting.” No assistant response is generated, no tool starts, no approval is requested, and no diagnostic error is shown. The task eventually becomes interrupted.

The in-app feedback uploader also failed 4–5 consecutive times. While attempting to report the problem through the GitHub form, the Codex in-app browser also crashed or disconnected more than once.

This is seriously disrupting long-running development work.

Environment

• Operating system: Windows 10 Pro
• Windows version: 10.0.19045
• Codex desktop app package version: 26.825.4187.0
• ChatGPT.exe file version: 151.0.7922.174
• Project type: Local Git repository, Cocos Creator, TypeScript
• Time zone: Asia/Shanghai
• Model: Not shown in the affected task UI

Affected session

• Task title: 检查任务是否卡住
• Session/thread ID: 01a04bb8-e405-7971-9f72-778a29bc8f80

Affected turns:

1. Turn ID: 01a04c45-4e17-7bf3-ab0b-365ce897c80d

   • Final state: interrupted
   • Duration: 416425 ms (approximately 6 minutes 56 seconds)

2. Turn ID: 01a04c4c-0c5d-79d1-aa9e-f97efcd5b3d6

   • Final state: interrupted
   • Duration: 638482 ms (approximately 10 minutes 38 seconds)
   • No assistant message
   • No tool execution marker
   • No visible error code

Steps to reproduce

1. Open an existing Codex desktop task connected to a local Git project.
2. Send a follow-up request that requires source inspection and tool execution.
3. The task enters active / inProgress.
4. The UI starts showing “reconnecting.”
5. Wait several minutes.
6. No assistant message or tool execution appears.
7. The task eventually becomes interrupted.
8. Send another follow-up message.
9. The same behavior may occur again.

Actual behavior

• The task remains in inProgress without performing observable work.
• No assistant response is generated.
• No shell command or tool starts.
• No approval request is displayed.
• No actionable error message or error code is provided.
• The UI repeatedly displays a reconnecting state.
• The task eventually records only turn_aborted / interrupted.

A bounded 30-second status check showed:

• Status remained active / inProgress
• No state revision
• No assistant message
• No tool marker
• No reported error

Evidence that this is not a project-code failure

Before the interruption, the same affected task successfully:

• Read local source files
• Ran PowerShell commands
• Ran ffmpeg video-frame extraction
• Modified TypeScript source files
• Completed multiple automated Node/TypeScript test suites
• Accessed the Git worktree normally

Another independent Codex task in the same desktop application continued replying and running tools.

This suggests a task-specific execution, connection, or message-flow failure rather than a Cocos project failure, terminal failure, or complete network outage.

In-app feedback upload failure

I attempted to report this problem using the Codex / feedback interface.

The upload failed 4–5 consecutive times with the message:

反馈上传未成功 / Feedback upload failed

The dialog stated that the feedback ID remained unchanged.

Feedback ID:

01a04722-83e4-7f92-a809-8184fea3c94e

Because the UI explicitly reported upload failure, I cannot confirm that the feedback or attached session data reached OpenAI.

Additional reporting failure

After the in-app feedback uploader failed, I opened the official GitHub Bug Report form through the Codex in-app browser.

During attempts to inspect or fill the form, the in-app browser crashed or disconnected more than once. The GitHub page remained open after recovery, but the browser-control session was lost.

No issue was submitted automatically.

Expected behavior

Codex should do one of the following:

1. Continue generating a response and running tools.
2. Clearly indicate that it is waiting for user approval.
3. Fail with a visible diagnostic error code.
4. Automatically recover the task without losing the active turn.
5. Provide a reliable retry or resume mechanism.

A task should not remain indefinitely in inProgress / reconnecting with no assistant or tool activity.

The in-app feedback uploader should also either upload successfully or provide a specific diagnostic error instead of repeatedly returning a generic upload failure.

Impact

This issue has happened repeatedly over several days.

Long-running game-development work must be moved into new tasks. Each migration requires another audit of:

• The authoritative worktree
• Existing Git changes
• Completed and incomplete source modifications
• Automated test results
• Pending manual Preview verification

This significantly disrupts development continuity and creates a risk of duplicated or incomplete work.

Questions for investigation

1. Why can a task remain inProgress with no assistant message or tool activity?
2. What subsystem does “reconnecting” refer to: the desktop client, backend execution process, or session message stream?
3. Why does the transcript contain only turn_aborted / interrupted without the underlying failure code?
4. Can Codex automatically detect and recover from this inactive state?
5. Why did the in-app feedback upload fail repeatedly?
6. Why did the in-app browser crash or lose its session while opening the GitHub report form?
7. Can clearer diagnostics and a reliable resume mechanism be added?

No source files, tests, or entry points are named. Start by tracing the desktop task lifecycle from active/inProgress through reconnecting and interrupted, then compare the feedback uploader and in-app browser failures. Done would require a reproducible diagnosis and a defined recovery or diagnostic behavior for the affected flows.