›_ Codex Issues
Codex Issues / #9766

Shift+Tab mode toggle resets model to gpt-5.2-codex; multi-agent cannot choose model or reasoning depth

Open 💬 4 comments Opened Jan 23, 2026 by d7185540

What version of Codex is running?

v0.89.0

What subscription do you have?

team

Which model were you using?

gpt-5.2

What platform is your computer?

Windows 11

What terminal emulator and version are you using (if applicable)?

PowerShell

What issue are you seeing?

  1. When I toggle modes with Shift+Tab, the currently selected model is forcibly reset to gpt-

5.2-codex.

  1. When multi-agent mode is started, agents are locked to gpt-5.2-codex and I cannot choose a

model or set reasoning depth.

What steps can reproduce the bug?

  • Model reset breaks workflow consistency and makes model comparison difficult.
  • Multi-agent work can’t assign the most suitable model for each task, reducing efficiency and

quality.

What is the expected behavior?

  • Add a configuration entry point for multi-agent model selection (CLI flag / config file /

interactive menu).

  • Keep model selection independent from mode toggle.

Additional information

_No response_

View original on GitHub ↗

4 Comments

sonobo · 7 months ago

I've proposed solution at https://github.com/openai/codex/issues/9783 (just in case)

malyutki · 7 months ago

Codex 0.91.0 the problem still occurs. If you set the Code to high or xhigh and change to Plan and back to Code it changes back to gpt-5.2-codex medium even if you choose a different model. Plan also don't remember changed model, but it seems that reasoning model in the Plan is remembered

3ld0ct0r · 7 months ago

For OpenAI Codex (v0.92.0) , it has changed wherein
When toggling modes with Shift+Tab, the currently selected model for Code mode remains unchanged but on toggling to Plan mode it drops to Medium (default) (current) for that model - #9980

oxysoft · 5 months ago

Cross-linking #14356 here because it proposes the cleaner control semantics that seem to sit underneath bugs like this.

Two pieces feel especially relevant:

  • model/reasoning controls should be explicit and directly adjustable, including reasoning depth itself
  • the active request should effectively snapshot its model/reasoning config when sent, and any later changes should stage for the next turn rather than clobbering the running one

That does not replace the bug fix here, but it does seem like the stronger mental model for avoiding this class of confusion.