Repository navigation
Codex startup failure (model_not_found) leaves the run in running indefinitely, and switching the model in the UI has no effect #31467
Description
Version: T3 Code 0.1.6 · provider: Codex (ChatGPT account)
Steps
- Start a Codex thread with a model name that does not exist (e.g.
gpt-6). - Codex answers within seconds with
model_not_found: "The 'gpt-6' model is not supported when using Codex with a ChatGPT account." - Look at the thread in the UI.
- Pick a valid model in the model picker and send a message.
Expected
- The run ends in an error state when the provider fails at startup, and the thread shows the error.
- The model change in the picker applies to the next run.
Actual
- The run (requested 09:21:42) stays
runningfor 2 h 30. The provider error was received at 09:21:48. The run only becameinterruptedat 11:51:44, after a manual interrupt. - After switching the model, the new run still carries
gpt-6in its model selection (thread_settings_appliedstill showsgpt-6).
sequenceDiagram
participant U as User
participant T as T3 (UI + runs)
participant C as Codex
U->>T: start thread, model "gpt-6"
Note over T: 09:21:42 run 1 = running
T->>C: start turn
C-->>T: 09:21:48 error model_not_found
Note over T: the error does not close the run
Note over T: 09:21:48 to 11:51:44: still "running" (2 h 30)
U->>T: switch model in the UI
T->>C: run 2 started with "gpt-6" again
U->>T: 11:51:44 manual interrupt, run = interrupted
| Time | Codex | T3 |
|---|---|---|
| 09:21:42 | starts | run 1: running |
| 09:21:48 | error model_not_found |
still running |
| after model switch in UI | n/a | settings applied: still gpt-6 |
| 11:51:44 | n/a | manual interrupt: interrupted |
Impact
A typo in a model name looks like a very long "thinking" turn. When agents launch threads automatically, nobody notices the failure. Validating the model name against the provider catalogue would also help, but the main bug is that the run never leaves running.
Evidence
orchestration_v2_projection_runs row (status, requested_at, completed_at) and the provider error payload. I can attach the redacted rows.
Activity
juliusmarminge commentedon Nov 5, 2026
Thanks for the detailed report. This is fixed on main by 997e3f37f7 (#28040), which shipped in v0.1.7.
Cause: with a ChatGPT account, Codex reports this rejection with the error code
modelNotFound, which T3 Code 0.1.6 did not recognize. It dropped Codex's failedturn/completednotification, so the run stayedrunning. The second symptom follows from the first: while that run is still running, a message sent after switching models is delivered as a steer into the same run, so it keepsgpt-6.What I ran, using a local backend that rejects
gpt-6with the samedetailmessage, with Codex reportingmodelNotFound:turn/completedabout 8 s after the send, and the run stayedrunning. After switching to GPT-6.1-Sol, the next message was steered into the samegpt-6run and never reached the backend.failedabout 8 s after the send and the thread shows "The 'gpt-6' model is not supported when using Codex with a ChatGPT account." After switching the model, the next message starts a new run ongpt-6.1-sol.Updating to 0.1.7 or later should resolve both. Closing this; please reopen if you still see it on 0.1.7 or later.