[Windows App] GPT-5.6-Sol Max is filtered out and config max falls back to Light

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

Codex App 26.707.9981.0 (release 26.707.72221; bundled codex-cli/app-server 0.144.2)

What subscription do you have?

API-key authentication. This appears to be local App picker/config filtering rather than plan-specific behavior.

What platform is your computer?

Microsoft Windows NT 10.0.19045.0 x64

What issue are you seeing?

On the current Windows Codex App, GPT-5.6-Sol's full Reasoning picker only contains:

Light
Medium
High
Extra High

Max is absent. Ultra is also not visible, but it has separate UI gating; the key bug here is that Max is a model-supported reasoning effort and is still filtered out.

When the global config contains a valid Max selection:

model = "gpt-5.6-sol"
model_reasoning_effort = "max"
plan_mode_reasoning_effort = "max"

restarting the App and creating a new thread does not preserve that selection in the Desktop UI. The new composer shows Light, and there is no Max option available to select.

The Settings search can find Model features, but opening the matched General settings area only renders the Configuration section; it does not expose Available reasoning efforts on this Windows build.

What steps can reproduce the bug?

1. On Windows, install/open Codex App 26.707.9981.0.
2. Put the following in ~/.codex/config.toml:

``toml
model = "gpt-5.6-sol"
model_reasoning_effort = "max"
plan_mode_reasoning_effort = "max"
``

1. Fully restart Codex App.
2. Create a new local thread.
3. Open the Reasoning picker above the composer.
4. Observe that the picker stops at Extra High and the active value is Light.
5. Search Settings for Model features / Available reasoning efforts and observe that there is no rendered control that allows Max to be enabled.

What is the expected behavior?

• Max should be present in the full Reasoning picker for GPT-5.6-Sol because the installed app-server's model metadata reports it as supported.
• A valid model_reasoning_effort = "max" config value should be honored, or at minimum produce a visible validation warning; it should not silently appear to fall back to Light.
• The Windows App should expose the Available reasoning efforts setting if the picker is intentionally controlled by that allowlist.

Additional information

The same installed runtime provides two useful controls:

1. codex debug models / app-server model metadata reports GPT-5.6-Sol reasoning efforts as:

``text
low, medium, high, xhigh, max, ultra
``

1. Starting the bundled CLI with model_reasoning_effort=max works. The generated session turn_context records both reasoning_effort: "max" and effort: "max".

Inspection of the installed Windows App bundle also explains the four visible options: the Desktop default enabled-reasoning-efforts list contains low, medium, high, xhigh, and ultra, but omits max; ultra is separately hidden unless enabled. This leaves only the four options shown above and causes a configured max value to fail the Desktop allowlist.

Related: #31968 tracks the Advanced shortcut/power slider skipping Max. This report is distinct: on Windows, Max is also absent from the full Reasoning picker, the UI control for enabling it is unavailable, and a global Max config is not preserved by the Desktop UI.