Global context settings appear to remain active after switching from gpt-5.4 to gpt-5.4-mini

What version of Codex is running?

0.118.0

What subscription do you have?

ChatGPT / Codex App

Which model were you using?

Primarily gpt-5.4, then switched to gpt-5.4-mini

What platform is your computer?

macOS 15

What issue are you seeing?

model_context_window and model_auto_compact_token_limit appear to be treated as global values even after switching to a model with a smaller real context window.

In my config I use:

model = "gpt-5.4"
model_context_window = 1000000
model_auto_compact_token_limit = 500000

This is valid for gpt-5.4, but after switching in the app to gpt-5.4-mini, the UI still appears to reflect the larger context window instead of the model's actual limit.

As a result:

• the app visually behaves as if gpt-5.4-mini still has the large configured context window
• auto-compaction does not happen early enough
• the session eventually fails near the smaller real model limit with a context window exceeded error

What steps can reproduce the bug?

1. Set the config to:

``toml
model = "gpt-5.4"
model_context_window = 1000000
model_auto_compact_token_limit = 500000
``

1. Start Codex App.
2. Use gpt-5.4 successfully.
3. Switch the model in the app to gpt-5.4-mini.
4. Continue a long session until the token count approaches the smaller real limit for gpt-5.4-mini.
5. Observe that auto-compaction does not trigger early enough and the session eventually errors due to context overflow.

What is the expected behavior?

When switching from gpt-5.4 to gpt-5.4-mini, Codex should do one of the following automatically:

• clamp the effective context window to the selected model's real max
• clamp the effective auto-compact threshold below that model's max
• update the UI to show the effective per-model context window instead of the stale global configured value
• warn the user that the current global settings exceed the selected model's supported context

Additional information

This looks especially confusing because the config is valid for gpt-5.4, but becomes unsafe when the model is switched to gpt-5.4-mini in the UI.

The core problem is not just the error itself, but that the app appears to keep showing/using the larger configured context semantics after the model switch.