Last verified:
Codex not working? Find your error and fix it
Most “Codex not working” reports fall into six buckets: a bad extension update, an exhausted usage limit, the Windows sandbox (os error 32) bug, sign-in, network, or an old/mismatched version. Find the message you see below.
VS Code extension broke after an update (“Codex could not start”)
Reported on VS Code extension 26.1021.22027: the webview misses a 30-second ready timeout. Downgrade to 26.1007.21434 and reload.
code --install-extension [email protected] --force“You've hit your usage limit” / rate limit errors
Codex shares a five-hour and/or weekly allowance with your ChatGPT plan. Check Settings → Usage or run /status, then wait for the reset, switch to a lighter model, buy credits, or use an API key.
Windows: “setup refresh had errors” / os error 32
On the Windows Codex app (reported from 26.1002.x builds in October 2026), every command can fail before it starts with helper_unknown_error: setup refresh had errors. The sandbox log (%USERPROFILE%\.codex\.sandbox\) shows The process cannot access the file because it is being used by another process. (os error 32) while validating Codex's own runtime files.
Tracked in #51601. Updating to 26.1007.21434 is the reported fix; a temporary [windows] sandbox = "unelevated" fallback also helps some users.
Mac desktop app sends no requests
Some Mac users on desktop build 26.1002.51308 reported that the app sends no requests at all, and that checking for updates finds nothing. Fully quit the app, download the current build from OpenAI directly instead of relying on the in-app updater, and reinstall. Check the openai/codex issues for your build number.
Config problems: profiles and MCP servers
- Profiles ignored after an update: since Codex 0.134.0,
--profileno longer reads[profiles.name]blocks insideconfig.toml. Each profile is its own file, for example~/.codex/deep.config.toml, used withcodex --profile deep(config reference). - An MCP server breaks startup: a bad
[mcp_servers.*]entry (wrong command, missing binary, server that hangs) can stop sessions from starting. Comment out MCP servers one at a time inconfig.tomlto find the culprit. - Back up
config.tomlbefore editing. A TOML syntax error can break everything.
Sign-in and authentication errors
- Sign out and back in with the ChatGPT account that holds your plan. In the CLI:
codex logout
codex login- Business/Enterprise: confirm your workspace admin has Codex enabled for you.
- API key users: check the key is valid, belongs to the right organization, and has billing enabled.
Network, “reconnecting” and timeout errors
- Check status.openai.com for an outage first.
- VPN: users report a VPN can break all sessions at once. Disconnect it and retry before anything else.
- Behind a corporate proxy or VPN, make sure HTTPS to OpenAI domains is allowed; try another network to rule it out.
- For “reconnecting”, “process not available” or “error starting conversation” with CLI 0.163 or later, restart the background daemon:
codex app-server daemon restartCodex can't read or write files
File failures usually come from the sandbox or approval mode, not a bug. Changing access with /permissions applies only to the current session. To make it stick, set sandbox_mode (for example workspace-write) and approval_policy in config.toml. On Windows, see the sandbox guide.
Only one feature is broken
Users also report isolated failures while everything else works: browser access, suggestions, voice dictation, remote pairing, and the copy shortcut in the Linux Konsole terminal. Update to the latest version, toggle the feature off and on, and search the openai/codex issues for the feature name. These usually need a fix from OpenAI.
Is it broken, or did I hit a limit?
An exhausted allowance can look like Codex stopped working. Check Settings → Usage first. See Codex usage limit.
Pin a working extension version
When an update breaks Codex: Extensions → gear on the Codex extension → Install Specific Version → pick the last version that worked → reload. Then uncheck Auto Update in the same gear menu so only Codex stops updating. Re-enable it once the related GitHub issue is closed. On Windows, also check that a CODEX_CLI_PATH environment variable isn't pointing at an old codex.exe.
Sources
FAQ
Why is Codex not working today?
The most common causes are a broken extension update (Codex could not start on 26.1021.22027), an exhausted five-hour or weekly usage limit, a Windows sandbox regression (os error 32), an expired sign-in, or a network/proxy problem. Match your error message to the sections on this page.
How do I fix 'Codex could not start' in VS Code?
Downgrade the extension to 26.1007.21434 with Install Specific Version, reload the window, and turn off auto-update for that extension only. Full steps are on the could-not-start page.
What does os error 32 mean in Codex on Windows?
It is a Windows sharing violation: the sandbox setup tries to open files Codex itself is running and fails with 'setup refresh had errors'. It is tracked in GitHub issue #51601; some users report that switching windows.sandbox from elevated to unelevated and fully restarting the app restores commands.
Codex says I hit a limit. Is it broken?
No. Check Settings → Usage or run /status in the CLI to see which allowance is used up and when it resets.
How do I fix Codex login problems?
Sign out and sign in again with the ChatGPT account that has your plan, or use codex login in the CLI. If you use an API key, check that the key is valid and has billing set up.
Should I update or roll back the Codex extension?
If a new version breaks, pin the last working version and disable auto-update for Codex only. Re-enable updates once the GitHub issue for the bug is closed and a fixed release ships.
This is an independent, unofficial help site. It is not affiliated with or endorsed by OpenAI. Plans, limits and versions change; always check the official pages linked here.