[Windows][Voice Chat] Starting voice silently resets GPT-5.6 Sol High to Low

What version of Codex are you using?

Codex Desktop on Windows. The bundled runtime recorded in session metadata is codex-cli 0.151.0-alpha.7.2.

What issue are you seeing?

The user-level default is explicitly configured as:

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

Starting Realtime Voice silently creates or updates the voice task with gpt-5.6-sol at low reasoning effort. The user then has to manually switch it back to High every time.

This is not a display-only issue: local session events record thread_settings_applied with reasoning_effort: "low" immediately after voice starts. The configuration file remains unchanged at High.

Steps to reproduce

1. In Codex Desktop on Windows, select GPT-5.6 Sol with High reasoning.
2. Confirm the user-level configuration also uses GPT-5.6 Sol and High reasoning.
3. Start a new Realtime Voice chat.
4. Inspect the model/reasoning selector or the resulting thread_settings_applied session event.
5. Observe that the voice task starts as GPT-5.6 Sol with Low reasoning.
6. Manually change it back to High.
7. Start another voice chat and observe the same reset.

Expected behavior

Realtime Voice should preserve the user's selected/default model and reasoning effort. If the product needs a lower-latency internal voice model, that internal choice should not overwrite the Codex task's selected model or reasoning level.

Actual behavior

Voice startup silently overrides High with Low.

Reproduction evidence

This reproduced in three separate voice sessions on the same day:

In one longer-running session, the setting later changed from High back to Low again and then had to be restored manually.

No session IDs, local paths, account details, or private logs are included here. The exact event timestamps and redacted session evidence can be provided privately if maintainers need them.

Related

This appears related to voice-task configuration inheritance problems reported in #39269, but this reproduction is narrower: even a valid persistent default of GPT-5.6 Sol / High is overridden to Low whenever voice starts.

Start with Realtime Voice startup and the thread_settings_applied event, then trace how the selected model and reasoning effort are inherited when a voice task is created. Compare the persistent High setting with the Low value applied after voice starts; done means voice preserves the user's selected model and reasoning effort without changing the configuration.