[App 26.707.62119] Running picker exposes unsupported gpt-5.2 and hides Spark while fresh model/list shows the opposite

What version of the Codex App are you using (From “About Codex” dialog)?

26.707.62119 (build 5211), bundled codex-cli 0.144.2

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

The running Codex Desktop model picker became inconsistent with a fresh model/list response from the exact same bundled app-server binary and ChatGPT account.

The Desktop picker currently shows:

GPT-5.6-Sol
GPT-5.6-Terra
GPT-5.6-Luna
GPT-5.5
GPT-5.4
GPT-5.4-Mini
GPT-5.2

GPT-5.3-Codex-Spark, which had been available earlier the same day, is missing. Selecting GPT-5.2 creates a real task with model = gpt-5.2, but the task produces no response and the app reports that the current ChatGPT account does not support the model.

This is not only a display-name alias. The local task index records the failed task as:

model: gpt-5.2
model_provider: openai
cli_version: 0.144.2
tokens_used: 0

However, starting a fresh instance of the bundled app-server and calling model/list on the same machine and account returns gpt-5.3-codex-spark and does not return gpt-5.2:

gpt-5.6-sol
gpt-5.6-terra
gpt-5.6-luna
gpt-5.5
gpt-5.4
gpt-5.4-mini
gpt-5.3-codex-spark

The persisted models_cache.json also contains gpt-5.3-codex-spark with supported_in_api: false and contains no gpt-5.2 entry. This suggests that the already-running Desktop process is rendering a stale or incorrectly merged model catalog that disagrees with both a fresh app-server response and the persisted cache.

What steps can reproduce the bug?

1. Sign in to Codex Desktop with a ChatGPT Pro account.
2. Keep the app running while the available model catalog changes or refreshes. In this case, Spark was usable earlier the same day.
3. Open the Desktop model picker.
4. Observe that GPT-5.3-Codex-Spark is missing and GPT-5.2 is visible.
5. Select GPT-5.2 and start a task.
6. Observe that the task is stored with model = gpt-5.2, produces no assistant output, and is rejected as unsupported for the ChatGPT account.
7. Without restarting the Desktop app, start the bundled app-server separately:

``bash
/Applications/ChatGPT.app/Contents/Resources/codex app-server --stdio
``

1. Send JSON-RPC initialize, followed by:

``json
{"id":2,"method":"model/list","params":{}}
``

1. Observe that the fresh app-server returns gpt-5.3-codex-spark and does not return gpt-5.2, which is the opposite of the running Desktop picker.

No model cache, authentication file, or Codex configuration was manually edited before the failure. I have not restarted the app so the inconsistent running state is still preserved.

What is the expected behavior?

The Desktop picker should use the same effective model catalog as its bundled app-server for the authenticated account.

In particular:

• It should not show gpt-5.2 when that model will be rejected for a ChatGPT-authenticated task.
• It should not hide gpt-5.3-codex-spark when a fresh model/list from the same bundled binary and account returns it.
• A catalog refresh should update or invalidate the running Desktop state instead of merging stale and built-in entries.

Additional information

• The Codex app process and its primary app-server had remained running since earlier in the evening when Spark was available.
• Related recent model-picker source-of-truth issues include #31873 and #32049, but neither reports this exact combination: the running Desktop picker exposing unsupported gpt-5.2 while a fresh 0.144.2 app-server returns Spark instead.
• A screenshot of the picker is available, but the complete model list is transcribed above as text.
• A task/session identifier can be provided privately if maintainers need it for server-side lookup.