You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
What version of the Codex App are you using (From “About Codex” dialog)?
26.928.2636.0
What subscription do you have?
pro lite(pro 100)
What platform is your computer?
Windows 11 Home China (CoreCountrySpecific), OS build 10.0.26200, x64
What issue are you seeing?
Windows Agent sandbox setup failed. CLI 0.159.2, Desktop 26.928.2636.0. Restarting, app repair, and reset were all ineffective; retrying does not trigger any UAC prompt, and setup_error.json remains stuck at 2026-09-30 15:24:41 UTC with the error helper_sandbox_lock_failed, unable to open .sandbox-bin.
press the button"retry to set up",it says"setup has stopped",if skip the process,you won't be able to send new messages to Codex,being told to "renew Agent sandbox to continue"
Same failure here on Codex Desktop 26.928.2636.0, Windows 11 Home, build 26300. Adding just WRITE_DAC for SYSTEM got mine past the setup failure, and sandbox commands could run again! Posting the anonymized details in case they help someone else.
Setup had already created the sandbox accounts and configured WFP (12 filters), then stopped at:
helper_sandbox_lock_failed:
lock sandbox bin dir <CODEX_HOME>/.sandbox-bin failed:
open directory <CODEX_HOME>/.sandbox-bin
The sandbox service was running as LocalSystem. The existing .sandbox-bin was an ordinary directory, owned by the interactive user, with inheritance disabled:
Identity
Permissions before the workaround
SYSTEM
Modify, Synchronize
Administrators
Modify, Synchronize
Interactive user / owner
Modify, ChangePermissions, Synchronize
CodexSandboxUsers
ReadAndExecute, Synchronize
So the owner could change the DACL, but SYSTEM lacked ChangePermissions / WRITE_DAC.
What worked locally: save the existing DACL, add only SYSTEM's WRITE_DAC on that one directory, then retry the normal Codex sandbox setup.
Here is the same operation with a backup and path checks, for the default$env:USERPROFILE\.codex location. Use it only when the service identity and ACL match the case above, and the interactive account already has permission to change that directory's DACL:
Then retry setup through Codex. This grant is additive and applies to the directory itself; it has no recursion or inheritance flags. The local recovery changed neither ownership nor the sandbox user group's rights, and added nothing to Administrators.
After the change, an actual command through Codex's sandbox runner printed SANDBOX_PROBE_OK and exited 0. Follow-up commands could read the workspace and the directory ACL. A subsequent startup probe also exited 0.
The catch: the post-setup ACL was back to the table above. Codex had removed the temporary SYSTEM WRITE_DAC grant itself. Shell startup and read access were verified; workspace writes, a future upgrade, and another provisioning cycle were not. This is a temporary provisioning workaround, with recurrence still possible.
The 0.159.2 provisioning source fits that explanation: ProvisionOnly opens the directory with READ_CONTROL | WRITE_DAC, while the generated SYSTEM/Administrators masks omit WRITE_DAC. That is a source-based explanation supported by the before/after result; the original log did not expose the underlying Windows error code.
Another reproducible case on the same Windows app version.
Environment
Codex / ChatGPT Windows app: 26.928.2636.0
Windows 11 Pro, build 10.0.26200, x64
Issue started only a few hours before this report; the same Work chats were usable before that.
Impact: existing Work chats are effectively unusable because the app requires Windows setup, then shows "Windows setup didn't finish / Setup stopped" when retrying. This blocks access to ongoing Work conversations that contain project instructions and files.
codex doctor findings (separate npm CLI)
Doctor: v0.144.4 windows-x86_64 (newer CLI version available)
Relevant warning: node_repl env var CODEX_WINDOWS_REGISTERED_CORE is not set
Additional thread-state warnings: rollout files missing from state DB / duplicate thread inventory entries. The DBs themselves report healthy.
This looks less like a general auth/network failure and more like a Windows provisioning / Registered Core setup regression in the current desktop build.
I can provide a redacted full codex doctor --json report if maintainers need it.
Activity
github-actions commentedon Sep 30, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
spidikor commentedon Oct 1, 2026
Same failure here on Codex Desktop 26.928.2636.0, Windows 11 Home, build 26300. Adding just
WRITE_DACfor SYSTEM got mine past the setup failure, and sandbox commands could run again! Posting the anonymized details in case they help someone else.Setup had already created the sandbox accounts and configured WFP (12 filters), then stopped at:
The sandbox service was running as LocalSystem. The existing
.sandbox-binwas an ordinary directory, owned by the interactive user, with inheritance disabled:So the owner could change the DACL, but SYSTEM lacked
ChangePermissions / WRITE_DAC.What worked locally: save the existing DACL, add only SYSTEM's
WRITE_DACon that one directory, then retry the normal Codex sandbox setup.Here is the same operation with a backup and path checks, for the default
$env:USERPROFILE\.codexlocation. Use it only when the service identity and ACL match the case above, and the interactive account already has permission to change that directory's DACL:Then retry setup through Codex. This grant is additive and applies to the directory itself; it has no recursion or inheritance flags. The local recovery changed neither ownership nor the sandbox user group's rights, and added nothing to Administrators.
After the change, an actual command through Codex's sandbox runner printed
SANDBOX_PROBE_OKand exited 0. Follow-up commands could read the workspace and the directory ACL. A subsequent startup probe also exited 0.The catch: the post-setup ACL was back to the table above. Codex had removed the temporary SYSTEM
WRITE_DACgrant itself. Shell startup and read access were verified; workspace writes, a future upgrade, and another provisioning cycle were not. This is a temporary provisioning workaround, with recurrence still possible.The 0.159.2 provisioning source fits that explanation:
ProvisionOnlyopens the directory withREAD_CONTROL | WRITE_DAC, while the generated SYSTEM/Administrators masks omitWRITE_DAC. That is a source-based explanation supported by the before/after result; the original log did not expose the underlying Windows error code.Related reports: SYSTEM permission workaround and recurrence, and WRITE_DAC-only recovery for SYSTEM plus Administrators. This local test adds a SYSTEM-only result on the desktop build in this issue.
fbnkch commentedon Oct 1, 2026
Another reproducible case on the same Windows app version.
Environment
26.928.2636.010.0.26200, x64codex doctorfindings (separate npm CLI)v0.144.4 windows-x86_64(newer CLI version available)15 ok · 1 idle · 3 notes · 2 warn · 0 fail degradedHTTP 101 Switching Protocols)restricted fs + restricted network · approval OnRequestnode_repl env var CODEX_WINDOWS_REGISTERED_CORE is not setThis looks less like a general auth/network failure and more like a Windows provisioning / Registered Core setup regression in the current desktop build.
I can provide a redacted full
codex doctor --jsonreport if maintainers need it.