Luna Reserve hides other providers' models and rejects every request with a usage-limit error

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

26.911.61220 (build 9647), bundled codex-cli 0.155.0-alpha.2.6, current in the Sparkle feed

What subscription do you have?

ChatGPT Pro, ChatGPT login

What platform is your computer?

Darwin 25.6.0 arm64 arm (macOS 26.6, 25G5057c)

What issue are you seeing?

Once the weekly ChatGPT allowance is exhausted and Luna Reserve becomes active, Codex Desktop hides every model from other providers, and Reserve itself rejects every request with a usage-limit error.

The model picker lists only GPT-5.6 Luna. The other providers' models, configured in ~/.codex/config.toml via a local OpenAI-compatible router with their own credentials and quota, are gone although they do not consume the ChatGPT allowance. Sending with GPT-5.6 Luna fails with

You've hit your usage limit. Add credits to continue, or try again at Sep 19, 2026, 2:20 PM.

while the usage widget shows "Luna Reserve · 100% remaining" and the picker shows "Add Credits for advanced models and added reasoning". The app is unusable until the reset. Reproducible: always, including after restarting the app.

What steps can reproduce the bug?

1. Sign in and confirm the other providers' models appear in the model picker.
2. Use Codex until the weekly allowance reaches 100%. The usage widget switches to "Luna Reserve · 100% remaining".
3. Open the model picker: only GPT-5.6 Luna is listed.
4. Send a message with GPT-5.6 Luna: usage-limit error, no response.

What is the expected behavior?

• Models from other providers stay selectable and usable while the ChatGPT allowance is exhausted.
• A Reserve request is accepted while Reserve shows 100% remaining.

Additional information

The router log has the backend response for the failing send (2026-09-17 21:34:26 CEST, model gpt-reserve, forwarded unchanged with the app's own auth):

HTTP 429  rate_limit_exceeded  "The usage limit has been reached"

Account rate-limit state at that moment: weeklyPercent 100, weeklyResetAt 2026-09-19 14:20 CEST, resetCredits 0.

Related: #42830 (picker collapse on Windows), #46181, #40939.

Start by reproducing the exhausted-weekly-allowance state with the configured providers in ~/.codex/config.toml, then trace the model picker, usage widget, and request handling around the reported HTTP 429 response. Done means other providers remain selectable and usable, while GPT-5.6 Luna requests are accepted when Reserve reports 100% remaining.