0.157.0 is broken on macOS

[dbalatero]

The most recent commit 6778b3d (codex 0.157.0) is broken on my machine out of the box when I nix flake update to that SHA.

I get this error message:

Error: this CLI has no complete local package; install a packaged Codex CLI or use the standalone installer
To work without the background server, rerun the same command with --no-daemon (including resume or fork and its arguments).

When I revert to the version 2 days prior (codex 0.156.1, 515f02b) it works 👍.

Did OpenAI do a breaking change in 0.157.0 that needs to be worked around?

I'm on macOS Tahoe 26.5.

[fizzAI]

in the meantime you can just add --no-daemon to make it run still

openai/codex#47318 its apparently supposed to nicely error and offer to disable the daemon but codex is entirely vibecoded at this point so i wouldn't be surprised if that is just broken

[dbalatero]

Blegh love having to remember special flags, thanks Codex. Maybe I'll alias it or something.

[JonnieCache]

this occurs on linux as well

[sadjow]

Thanks for the report, @dbalatero, and thanks @fizzAI and @JonnieCache for the workaround and the Linux confirmation.

I reproduced this on the current release (Codex 0.159.0) on aarch64-darwin, starting from an empty CODEX_HOME. It affects every platform.

Since 0.157, Codex starts its background server by copying the package it runs from into $CODEX_HOME/packages/app-server-daemon/, and that needs the complete package layout OpenAI ships in codex-package-<platform>.tar.gz. This flake still installs the per-binary release assets, so any command that needs the server fails with this error. CI only checked codex --version, which is how these releases got through.

@srid opened #165, which builds from the package archive and adds a CI check that starts the server. I built it on aarch64-darwin: both interactive codex and codex agents start the server from a clean CODEX_HOME, and the server reports the same version as the CLI. It needs a rebase onto the current release before it can merge.

Until then, either of these works:

• Run Codex with --no-daemon. I confirmed this on 0.159.0.

• Pin the input to 0.156.1, the last release before the change:

  inputs.codex-cli-nix.url = "github:sadjow/codex-cli-nix/515f02b2fa319fc42075f456345a6c4e4508ebb2";

[JonnieCache]

Personally I don't want it to run any kind of server or resident process. Is it at least using systemd to manage it?

[sadjow]

@JonnieCache It doesn't use systemd. Codex starts the server itself as a detached background process, tracked by a pid file in $CODEX_HOME/app-server-daemon/, and next to it a separate updater loop. Neither survives a reboot, but Codex starts them again on the next launch that needs them. codex app-server daemon stop stops the server and leaves the updater running.

The updater matters for Nix users. Five minutes after it starts, and hourly after that, it downloads OpenAI's installer from https://chatgpt.com/codex/install.sh and switches the server to the latest upstream release. It runs the installer in daemon-only mode, so it only touches $CODEX_HOME/packages/app-server-daemon. It doesn't change ~/.local/bin or your shell profile. Still, the server stops being the Nix build: in my test, a 0.158.0 server started from this flake was running OpenAI's 0.159.0 within seven minutes. The DISABLE_AUTOUPDATER variable that our wrapper sets has no effect on it.

To keep the server on the Nix build, run this once Codex has started the server:

codex app-server daemon update --from-cli --yes

It copies the package of the codex you run it with and pins the server to it, and the updater stops. It restarts a running server, so work in progress there can be interrupted. A Nix upgrade doesn't move the pinned server either, so run it again after upgrading.

To turn it all off, put this in ~/.codex/config.toml:

[features]
daemon_auto_start = false

I tested this on 0.159.0 and with the build from #165. Interactive codex works normally, and neither the server nor the updater starts. It also gets around this issue without passing --no-daemon every time. The exception is codex agents, which needs the shared server.

If you want to keep the server but drop the updater, put {"updater": {"autoUpdateEnabled": false}} in $CODEX_HOME/app-server-daemon/settings.json. With that setting, codex app-server daemon start starts only the server.

#165 is merged now, so the flake's codex starts the server again once you update your lock.

[sadjow]

Fixed by #165, which is on main now. After updating your lock to the latest main (for example nix flake update codex-cli), codex starts normally again without --no-daemon. I tested it on aarch64-darwin, and the PR's CI runs the same check on Linux and macOS. The v0.159.0 tag predates the fix, so if you pin tags, the next release tag will include it. Closing, but reopen if it still fails for you.