New install: codex stream and install assistant errors (with Codex CLI already configured for web account usage)

[darlingm]

EDIT: Cause found

TLDR: See the last comment and the linked issue. DNS was partially malfunctioning. Other programs and Code compiled by source weren't affected. I am confused still in settings why agents show code-gpt and -codex are not installed, even though I'm able to use these models now, and claude and gemini show enabled but I've never installed a custom agent for them.

What version of Code is running?

code 0.4.4

Which model were you using?

GPT-5-Codex (high)

What platform is your computer?

Linux 6.17.5-arch1-1 x86_64 unknown

What steps can reproduce the bug?

1. Install OpenAI Codex CLI v0.50.0
2. Run codex and configure it to use an OpenAI account (not the API.)
3. Observe that Codex CLI works fine, and that ~/codex/auth.json contains id_token, access_token, refresh_token, and account_id.
4. Install @just-every/code v0.4.4.
5. Run code.
6. Observe the banner says: Code • Model: GPT-5-Codex • Reasoning: high • Directory: /mnt/...
7. Type hello and press enter.

What is the expected behavior?

I expect to get a response back like Hey there! What can I help you with today?

What do you see instead?

1. Type hello and press enter. Observe Thinking...
2. After a number of seconds, observe stream error: error sending request for url (https://chatgpt.com/backend-api/codex/responses); retrying 1/5 in 217ms, then 2/5 in 363ms, 3/5 in 728ms, 4/5 in 1.697s, 5/5 in 3.157s, and error sending request for url (https://chatgpt.com/backend-api/codex/responses).
3. Run settings, and observe Model shows Model: gpt-5-codex Effort: High.
4. Enter Agents, and observe that code-gpt-5-codex and code-gpt-5 say not installed Enter to install.
5. Observe Terminal - Install code-gpt-5-codex\nRunning... Preparing install assistant... After a number of seconds, observe Install helper error: model stream failed. After several minutes of retries, I hit Esc and Ctrl+C.
6. Run curl 'https://chatgpt.com/backend-api/codex/responses', and observe {"detail":"Method Not Allowed"}. This of course isn't posting the required data so also gets an error, but shows connecting to the endpoint is fine.

Additional information

I have a straightforward networking setup. (No VPNs, etc.)

I also tried running OPENAI_WIRE_API=chat code but it had no effect.

I'm a little surprised that a fork of OpenAI Codex CLI needs something extra to be installed to use code-gpt-5-codex, but that's not a big deal.

[darlingm]

This looked a little like #371 but I don't think it's actually related.

[darlingm]

I see ~/.code/log/codex-tui.log includes WARN stream disconnected - retrying turn (1/5 in 217ms)... and so on, but doesn't have additional details.

[darlingm]

I'm on Arch Linux, running Node.js v25.0.0. I think this is some kind of incompatibility with newer system packages than the npm code versions were compiled against, but that npm codex versions run fine. Running code via npm fails, but compiling from source works.

I moved ~/.codex and ~/.code and just ran code, selecting to sign in with ChatGPT for my paid plan. It sent me to (I typed this, so I could have a typo, and omitted parts of course had the full parts) https://auth.openai.com/oauth/authorize?response_type=code&client_id=app_EM...&redirect_uri=http%3A%2F%2Flocalhost%3A1455%2auth%2Fcallback&scope=openid%20profile%20email%20offline_access&code_challenge=omitted&code_challenge_method=S256&id_token_aa_organizations=true&code_cli_simplified_flow=true&state=omitted&originator=code_cli_rs. I selected Continue with Google, then my account, then gave the authenticator PIN number. It redirected to http://localhost:1455/auth/callback?code=omitted&scope=openid+profile+email+offline_access&state=omitted which shows: Token exchange failed: error sending request for url (https://auth.openai.com/oauth/token). code still shows finish signing in via your browser.

I ran with OPENAI_API_KEY=<my API key> and I had the same problems.

I went back to code 0.1.3 through npm, and had the same problem connecting the OpenAI account, and the timeout with the API key.

I built 0.1.3 from source, intending to git bisect through the few commits between it and upstream's Codex CLI, and surprisingly it worked.

I built 0.4.4 from source, and surprisingly it worked.

I've double checked, and the npm versions fail while the from source versions work. The codex npm versions work.

[darlingm]

Figured it out. The default empty /etc/resolv.conf was somehow written a few days ago. Apparently I haven't ran anything else in the past few days that directly uses it, and maybe building from source somehow avoids needing it. Fixing this to be a symlink to the systemd-resolved stub-resolv.conf (a way systemd-resolved can be setup) makes npm-installed code work fine.

I have no idea how on #376 code was seeing that Codex's newest version was 0.50.0 when it couldn't resolve DNS. Maybe there's an IP address fallback.

I'll leave this open for you to decide if there's any changes you'd like to make to code in case other users have DNS not fully setup. Maybe there's not. I am confused still in settings why agents show code-gpt and -codex are not installed, even though I'm able to use these models now, and claude and gemini show enabled but I've never installed a custom agent for them.

I had some corruption and needed to reinstall all system files. Arch properly has the file marked to not clobber it when doing this, and I have no idea how this happened.