New conversation ignores selected model and starts as GPT-5.4

Summary

In Codex Desktop, changing the model to gpt-5.4-mini before sending the first prompt in a new conversation can still create the thread on gpt-5.4. This doesn't happen in a newly opened Codex app, but happens after 15-30 minutes of use. Only fully quitting the app will fix it.

Repro

1. Create a new conversation.
2. Change the model to gpt-5.4-mini.
3. Send a simple prompt like test.

Expected

The new conversation should be created on gpt-5.4-mini.

Actual

The new conversation is created on gpt-5.4.
The UI shows GPT-5.4, and the persisted thread state also shows gpt-5.4.

Concrete evidence

• Conversation id: 019d911e-7997-7a82-b855-46bef5c8b853
• At 2026-04-15 14:29:46 CEST, the app updated the default config to model = "gpt-5.4-mini".
• The app logged: Setting default model and reasoning effort newEffort=high newModel=gpt-5.4-mini.
• Roughly 0.7s later, thread/start ran for the new conversation.
• The created thread persisted as model = gpt-5.4.
• First prompt in that thread was test.

Likely cause

thread/start appears to read a stale in-memory model selection instead of the just-updated default/config value.

Environment

• Codex Desktop build seen locally: 26.409.20454
• Telemetry included app.version=0.119.0-alpha.28