Windows Codex loses foreground command execution handle after ~30 seconds —

[joeyb1102]

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

26.805.11740

What subscription do you have?

Plus

What platform is your computer?

Windows 11

What issue are you seeing?

Codex version: 26.805.11740
OS: Windows 11
Sandbox: Workspace write
Codex Workspace Diagnose: Dependencies look healthy

Problem

Codex repeatedly loses the execution/session handle for foreground commands after approximately 30 seconds.

The failure occurs with:

exec cell 11 not found

This initially appeared during Playwright tests, but I reproduced it with a minimal PowerShell command that does not use Node.js, Playwright, Git, or my project.

Minimal reproduction

Run a foreground PowerShell command that emits a heartbeat every 10 seconds for 90 seconds:

1..9 | ForEach-Object {
Write-Output ("CODEX-RUNNER-HEARTBEAT " + $_ + "/9 " + (Get-Date -Format "HH:mm:ss"))
Start-Sleep -Seconds 10
}
Write-Output "CODEX-RUNNER-DIAGNOSTIC-COMPLETE"
Result
RESULT: FAIL
HEARTBEATS OBSERVED: 3/9
FINAL MARKER OBSERVED: NO
EXIT CODE: unavailable
EXECUTION-HANDLE ERROR: exec cell 11 not found
TOTAL DURATION: approximately 30 seconds

The command is started successfully, but Codex loses the execution handle before the process completes.

Restarting Windows did not fix the problem.
Workspace dependencies diagnose as healthy.

The same long-running Playwright command completes normally when executed directly in Windows PowerShell outside Codex.

This appears to be a Windows Codex command-runner/session-management issue rather than a Node, Playwright, Git, or project-specific problem.

Reproduced: August 11, 2026, around 4:00 PM EDT.

What steps can reproduce the bug?

Codex version: 26.805.11740
OS: Windows 11
Sandbox: Workspace write
Codex Workspace Diagnose: Dependencies look healthy

Problem

Codex repeatedly loses the execution/session handle for foreground commands after approximately 30 seconds.

The failure occurs with:

exec cell 11 not found

This initially appeared during Playwright tests, but I reproduced it with a minimal PowerShell command that does not use Node.js, Playwright, Git, or my project.

Minimal reproduction

Run a foreground PowerShell command that emits a heartbeat every 10 seconds for 90 seconds:

1..9 | ForEach-Object {
Write-Output ("CODEX-RUNNER-HEARTBEAT " + $_ + "/9 " + (Get-Date -Format "HH:mm:ss"))
Start-Sleep -Seconds 10
}
Write-Output "CODEX-RUNNER-DIAGNOSTIC-COMPLETE"
Result
RESULT: FAIL
HEARTBEATS OBSERVED: 3/9
FINAL MARKER OBSERVED: NO
EXIT CODE: unavailable
EXECUTION-HANDLE ERROR: exec cell 11 not found
TOTAL DURATION: approximately 30 seconds

The command is started successfully, but Codex loses the execution handle before the process completes.

Restarting Windows did not fix the problem.
Workspace dependencies diagnose as healthy.

The same long-running Playwright command completes normally when executed directly in Windows PowerShell outside Codex.

This appears to be a Windows Codex command-runner/session-management issue rather than a Node, Playwright, Git, or project-specific problem.

Reproduced: August 11, 2026, around 4:00 PM EDT.

What is the expected behavior?

No response

Additional information

No response

[github-actions]

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• Codex can’t run long processes #36747

Powered by Codex Action

[kingsleyonoh]

Corroborating reproduction from T3 Code (Alpha) 0.0.33 using codex-cli 0.147.0 on Windows, observed 2026-08-15.

Workload: a foreground PowerShell shell_command running a quiet full pytest suite. The initial tool call yielded after about 11 seconds with exec cell ID 16. Two subsequent 30-second waits both returned "Script running with cell ID 16" and no output (pytest was invoked with -q). A later wait on the same ID failed immediately with:

exec cell 16 not found

A Win32_Process query immediately afterward found no remaining pytest process, so the command had ended but its final exit code and final pytest summary were unavailable to the agent. This was not an output-cap/high-volume case: the process emitted no visible output during the run.

This occurred through the T3 Code Codex harness, but the failure signature and Windows foreground-command behavior match this upstream report. A useful recovery would be to retain completed cell results long enough for a pending wait to retrieve them, or return the terminal result instead of "not found" when the cell completed between polls.

[kingsleyonoh]

Cross-reported the T3 Code integration/lifecycle symptom with a Windows repro and environment details at pingdotgg/t3code#6656. The failure includes both the lost resumable exec-cell handle and hidden in-flight tool state. A validated temporary workaround is to move long jobs under Windows Task Scheduler and persist a durable state/exit-code/JUnit manifest, rather than relying on a yielded exec cell.