model_reasoning_effort from config.toml seems overridden by last-used reasoning on fresh CLI launch

Resolved 💬 3 comments Opened Apr 11, 2026 by Michaelihc Closed Apr 11, 2026

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of Codex CLI is running?

0.120.0

What subscription do you have?

ChatGPT Pro 20x

Which model were you using?

gpt-5.4

What platform is your computer?

Microsoft Windows NT 10.0.26220.0 x64

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

PowerShell

What issue are you seeing?

A fresh codex session can start with the previously used reasoning effort instead of the value in ~/.codex/config.toml.

My config:

model = "gpt-5.4"
model_reasoning_effort = "high"

But after using xhigh in a prior session, a brand-new codex launch started with xhigh.

What steps can reproduce the bug?

Set in ~/.codex/config.toml:

model = "gpt-5.4"
model_reasoning_effort = "high"

Start codex
Change reasoning effort to xhigh
Exit
Start a fresh new codex session

What is the expected behavior?

A fresh CLI launch should respect model_reasoning_effort from ~/.codex/config.toml, or the docs should explicitly state that persisted last-used state overrides config on startup.

Additional information

This may be related to the CLI 0.116.0 changelog entry saying model and reasoning effort are persisted in sqlite.

View original on GitHub ↗

3 Comments

github-actions[bot] contributor · 4 months ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

#16984
#15853

Powered by Codex Action

etraut-openai contributor · 4 months ago

In your repro steps, step 3 says "Change the reasoning effort to xhigh". I presume you're doing this with the /model slash command in the TUI? This command allows you to select the model and reasoning effort, and then it writes the selection back to your global (user-level) config file. When you restart the CLI the next time — or create a new thread using /new or /clear, this will be the new default model and reasoning effort.

It sounds like this is working as intended, so I'm going to close the bug report. Let me know if I misunderstood anything or you have any follow-up questions.

DerekWW · 2 months ago

@etraut-openai

This is confusing behavior and is not what I would expect. In fact we have run into this confusion with several developers in my org. Writing a new value for those fields in the user level config.toml on model/reasoning change in a client is behavior that is listed anywhere in the docs, in fact the docs explicitly call it temporary.

https://developers.openai.com/codex/models#choosing-a-different-local-model-temporarily

The docs state Choose the model Codex uses by default in the CLI and IDE. and is under the header Default Model.

https://developers.openai.com/codex/config-basic#default-model

To me that sounds like I should be able to set a model via config.toml that is always used for new sessions, and does not get updated just because I needed to use a different model or reasoning level for something as a one off.

It seems that managed_config.toml does behave the way I would expect in that you can set a default model(same default terminology as the basic config file, but different behavior) and it will be used as the starting model for all new sessions.

https://developers.openai.com/codex/enterprise/managed-configuration#managed-defaults-managed_configtoml