Add /reasoning slash command to change reasoning effort mid-conversation
Resolved 💬 4 comments Opened Aug 9, 2025 by zemaj Closed Sep 3, 2025
💡 Likely answer: A maintainer (dedrisian-oai, contributor)
responded on this thread — see the highlighted reply below.
Feature Request
Problem
Currently, users need to restart their Codex session or modify the config.toml file to change the reasoning effort level. This is inconvenient when working on tasks that vary in complexity during a single session.
Solution
Add a /reasoning slash command that allows users to change the reasoning effort level mid-conversation.
Features
- Parameter support:
/reasoning <level>to set a specific level directly /reasoning high- Sets reasoning to high/reasoning mediumor/reasoning med- Sets reasoning to medium/reasoning low- Sets reasoning to low/reasoning noneor/reasoning off- Disables reasoning
- Interactive UI:
/reasoningwithout parameters shows an interactive selection UI - Shows current reasoning level highlighted in yellow
- Use arrow keys (↑↓) to navigate options
- Press Enter to select a level
- Press Esc to cancel
- Each option shows a helpful description
Benefits
- Flexibility: Change reasoning effort without restarting the session
- Control: Adjust the model's reasoning depth based on task complexity
- Efficiency: Use lower effort for simple tasks, higher for complex ones
- Transparency: Clear feedback about the current reasoning level
Implementation Details
The implementation modifies several files in the TUI module to add the new command, handle parameter parsing, create an interactive selection UI, and update the session configuration dynamically.
4 Comments
I was about to file this - this is necessary! It would be less necessary if we had a good way to start new codex sessions resuming from the history/context of a previous session, though.
Huge improvement! Can this also get expanded to verbosity as well?
This is already mostly usable with the
/modelif you're using gpt-5. Closing for now.Seems unreasonable, it should be compatible across all models and providers.