gpt 5.4 mini is not available in model picker when using API auth key

Not available on 0.116.0 codex CLI either, but updating to 0.117.0 (latest) fixes it

<img width="1000" height="539" alt="Image" src="https://github.com/user-attachments/assets/953ab482-0ac8-4d57-b0be-54f16277bc33" />

In refresh_available_models, API-key users hit an unconditional early return that prevents them from ever fetching the remote model catalog:

if self.auth_manager.auth_mode() != Some(AuthMode::Chatgpt)
    && !self.provider.has_command_auth()
{
    if matches!(
        refresh_strategy,
        RefreshStrategy::Offline | RefreshStrategy::OnlineIfUncached
    ) {
        self.try_load_cache().await;
    }
    return Ok(());  // ← always returns, never reaches fetch_and_update_models()
}

This creates a circular problem for API-key users (where auth_mode() == Some(AuthMode::ApiKey)):

1. fetch_and_update_models() is unreachable because of the early return at line 410
2. fetch_and_update_models() is the only place that calls persist_cache() (line 467-469), so the cache is never written by API-key users
3. try_load_cache() requires the cache to be fresh (5-min TTL) and version-matched, so any cache from a prior ChatGPT session quickly becomes unusable
4. The app falls back to the bundled models.json, which doesn't contain gpt-5.4-mini

In contrast, ChatGPT auth users and command-auth users proceed past line 411 to the normal OnlineIfUncached path, which falls through to fetch_and_update_models() on cache miss

The early return at line 410 should not be unconditional for API-key users who have valid credentials. The OnlineIfUncached and Online strategies should be allowed to reach fetch_and_update_models() when the user has a valid API key, just as they do for ChatGPT and command-auth users. The guard at lines 401-411 appears to have been intended for users with no auth at all (where a network fetch would fail anyway), but it also blocks authenticated API-key users.

I have encountered the same issue. I'm using the enterprise unified routing API endpoint, which may not support forwarding requests for the model list. I think it's worth considering updating the built-in models.json content, or providing a user-defined configuration file for models.json.