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
{{ message }}
Repository navigation
[Bug]: Codex provider is marked Unavailable #14333
I searched existing issues and did not find a duplicate.
I included enough detail to reproduce or investigate the problem.
Area
apps/web
Steps to reproduce
On T3 Code v0.0.44, the Codex provider is marked Unavailable with “Codex app-server provider probe failed: Codex App Server process exited with code 1.” Settings show binary path codex, CODEX_HOME ~/.codex, and shadow home ~/.codex-t3/personal. This prevents me from using GPT models through the Codex provider.
This is not a confirmed T3 defect, and it is not #14325 (open, Windows shadow-home symlink EPERM).
The line in the screenshot is produced only after codex app-server has already started. checkCodexProviderStatus in apps/server/src/provider/Layers/CodexProvider.ts maps CodexAppServerProcessExitedError to Codex app-server provider probe failed: Codex App Server process exited with code 1. A binary that cannot be spawned takes the other branch and says Could not start Codex CLI (`codex`). On this machine the command was found, started, and then exited 1.
#14325 dies earlier, inside materializeCodexShadowHome, when Windows refuses the shadow-home symlink (EPERM) unless Developer Mode is on. That path never launches the app-server. The managed ChatGPT instance shows "Could not check Codex right now." A user-configured instance that fails there shows Driver 'codex' failed to create instance: Codex shadow home filesystem operation 'symlink' failed.... Neither is the text in this screenshot. The built-in CLI instance only builds that overlay when Shadow home path is non-empty.
The three paths in the report are the settings placeholders: binary codex, CODEX_HOME ~/.codex, shadow ~/.codex-t3/personal (packages/contracts/src/settings.ts). A managed ChatGPT shadow lives under the state directory (providers/codex/<id>/shadow). An empty shadow field renders as "Not used", not as that placeholder. If ~/.codex-t3/personal was actually saved, it is not a second login. The documented setup is a directory you have already signed into, for example ~/.codex_personal. A shadow directory with no auth.json still returns from account/read as signed out. It does not exit the process. Saving that path on Windows without Developer Mode would fail at symlink creation with the #14325 error, which is not what this screenshot shows.
The status probe discards the child's stderr (stderr: () => Sink.drain in packages/effect-codex-app-server/src/_internal/stdio.ts), so the UI cannot show why Codex quit. Nothing after v0.0.44 (451afcb) changes this message. There is no open PR.
Please reply with home paths and tokens removed:
Whether Shadow home path is a saved value or the gray placeholder. If you only use one Codex account, clear it and refresh the provider.
The first error lines from that process are the missing diagnosis. If codex exits 1 there, T3 is reporting the CLI (or that CODEX_HOME), not a separate provider failure.
On Windows with T3 Code 0.0.44, codex --version and codex app-server both exit 1 with “Missing optional dependency @openai/codex-win32-x64.” where.exe codex finds npm shims first and a separate OpenAI Codex executable later. That executable reports codex-cli 0.159.0, and its app server starts successfully on localhost. Home paths and tokens omitted.
Thanks for taking the time to report this and provide the details. We revisited it during the orchestrator V2 cleanup.
Your follow-up narrowed this down to the npm Codex shim missing its Windows dependency, while the separate Codex executable starts successfully. Pointing Settings → Providers → Codex → Binary path at that working executable, or repairing the npm install, is the next step. Thank you for checking both executables.
I’m closing this as answered, rather than attributing it to a new V2 fix.
Resolved diagnosis, not confirmed repaired installation. Reporter should point Binary path to the working executable or repair the npm installation.
If you still hit this on a current build, please reply with the app/server versions and the steps that reproduce it. We can reopen this if the original problem is still there.
Activity
[-][Bug]:[/-][+][Bug]: Codex provider is marked Unavailable [/+]juliusmarminge commentedon Sep 29, 2026
Triage
This is not a confirmed T3 defect, and it is not #14325 (open, Windows shadow-home symlink
EPERM).The line in the screenshot is produced only after
codex app-serverhas already started.checkCodexProviderStatusinapps/server/src/provider/Layers/CodexProvider.tsmapsCodexAppServerProcessExitedErrortoCodex app-server provider probe failed: Codex App Server process exited with code 1. A binary that cannot be spawned takes the other branch and saysCould not start Codex CLI (`codex`). On this machine the command was found, started, and then exited 1.#14325 dies earlier, inside
materializeCodexShadowHome, when Windows refuses the shadow-home symlink (EPERM) unless Developer Mode is on. That path never launches the app-server. The managed ChatGPT instance shows "Could not check Codex right now." A user-configured instance that fails there showsDriver 'codex' failed to create instance: Codex shadow home filesystem operation 'symlink' failed.... Neither is the text in this screenshot. The built-in CLI instance only builds that overlay when Shadow home path is non-empty.The three paths in the report are the settings placeholders: binary
codex, CODEX_HOME~/.codex, shadow~/.codex-t3/personal(packages/contracts/src/settings.ts). A managed ChatGPT shadow lives under the state directory (providers/codex/<id>/shadow). An empty shadow field renders as "Not used", not as that placeholder. If~/.codex-t3/personalwas actually saved, it is not a second login. The documented setup is a directory you have already signed into, for example~/.codex_personal. A shadow directory with noauth.jsonstill returns fromaccount/readas signed out. It does not exit the process. Saving that path on Windows without Developer Mode would fail at symlink creation with the #14325 error, which is not what this screenshot shows.The status probe discards the child's stderr (
stderr: () => Sink.draininpackages/effect-codex-app-server/src/_internal/stdio.ts), so the UI cannot show why Codex quit. Nothing afterv0.0.44(451afcb) changes this message. There is no open PR.Please reply with home paths and tokens removed:
codex --version where codex codex app-serverIf that shadow path is really saved, also:
The first error lines from that process are the missing diagnosis. If
codexexits 1 there, T3 is reporting the CLI (or thatCODEX_HOME), not a separate provider failure.TheDarkSkyXD commentedon Sep 30, 2026
On Windows with T3 Code 0.0.44, codex --version and codex app-server both exit 1 with “Missing optional dependency @openai/codex-win32-x64.” where.exe codex finds npm shims first and a separate OpenAI Codex executable later. That executable reports codex-cli 0.159.0, and its app server starts successfully on localhost. Home paths and tokens omitted.
juliusmarminge commentedon Oct 2, 2026
Thanks for taking the time to report this and provide the details. We revisited it during the orchestrator V2 cleanup.
Your follow-up narrowed this down to the npm Codex shim missing its Windows dependency, while the separate Codex executable starts successfully. Pointing Settings → Providers → Codex → Binary path at that working executable, or repairing the npm install, is the next step. Thank you for checking both executables.
I’m closing this as answered, rather than attributing it to a new V2 fix.
Source reviewed.
Resolved diagnosis, not confirmed repaired installation. Reporter should point Binary path to the working executable or repair the npm installation.
If you still hit this on a current build, please reply with the app/server versions and the steps that reproduce it. We can reopen this if the original problem is still there.