[Windows app] “More details” Quick Chat ignores the saved model and submits with Auto

Summary

In the Microsoft Store Codex/ChatGPT desktop app, selecting text in a Codex conversation and clicking More details opens a ChatGPT Quick Chat and immediately submits the generated prompt using Auto.

The selected-text Quick Chat does not inherit either:

• the user's saved ChatGPT model and thinking-effort selection; or
• Codex's configured model / model_reasoning_effort.

As a result, there is currently no way to make this shortcut use a reasoning model by default before its first message is submitted.

Environment

• OS: Windows
• Distribution: Microsoft Store
• Package: OpenAI.Codex
• App version: 26.715.4045.0
• Executable display name: ChatGPT

Steps to reproduce

1. Open a Codex conversation in the Windows desktop app.
2. Select some text in a message.
3. Click More details.
4. Observe that a ChatGPT Quick Chat opens and automatically submits “Tell me more about this”.
5. Check the model used for the generated response.

Actual behavior

The first Quick Chat request uses Auto with no explicit thinking effort.

Changing Codex configuration does not affect it, for example:

model = "gpt-5.6-sol"
model_reasoning_effort = "high"

Selecting another model and thinking effort in the regular ChatGPT UI also does not reliably affect a newly created selected-text Quick Chat.

Expected behavior

The selected-text Quick Chat should do one of the following:

1. inherit the user's last saved ChatGPT model and thinking-effort selection;
2. provide a dedicated default-model setting for Quick Chat; or
3. allow model selection before automatically submitting the first message.

The smallest and least surprising behavior would be to inherit the existing saved ChatGPT model selection, including its thinking effort, before falling back to Auto.

Additional technical findings

I inspected the packaged desktop frontend to distinguish this behavior from the Codex app-server path.

• The action is identified as selectedTextOverlay.moreDetails.
• It creates a new ChatGPT consumer Quick Chat and submits model and optional thinking_effort fields.
• For a new selected-text Quick Chat, model resolution falls back to the server model-list default and ultimately auto.
• The app already has a persisted global ChatGPT model selection (chatgpt-last-selected-model-v1), but this selected-text initialization path does not consult it.
• This path is separate from Codex threads and does not read ~/.codex/config.toml.

This appears to be a model-state initialization bug specific to the selected-text Quick Chat path, rather than an account/model availability issue.

Suggested fix

When initializing the selected-text Quick Chat, prefer the persisted ChatGPT model selection, including thinkingEffort, and only fall back to the conversation/model-list default when no saved selection exists.

This would preserve entitlement and rollout checks from the existing model picker while preventing the shortcut from silently reverting to Auto.