feat: enable api/llm model to converse in between each step
Resolved 💬 4 comments Opened Apr 18, 2025 by seokw-Ryan Closed Aug 7, 2025
💡 Likely answer: A maintainer (tibo-openai, collaborator)
responded on this thread — see the highlighted reply below.
It would be beneficial to enhance the Codex CLI by allowing the LLM to engage in conversational interactions between execution steps. This feature would enable users to:
- Review and discuss intermediate outputs before proceeding.
- Provide additional context or corrections during multi-step tasks.
- Iteratively refine actions based on real-time feedback.
Implementing this capability would improve user control and adaptability, especially in complex workflows.
Proposed Implementation:
- Introduce an optional conversational mode that pauses after each step, awaiting user input.
- Allow users to confirm, modify, or skip subsequent actions based on the assistant's suggestions.
- Maintain a history of interactions to provide context for ongoing tasks.
This feature would align with the goal of creating a more interactive and user-friendly coding assistant within the terminal environment
4 Comments
Could you expand with a concrete example of what this would look like? Thanks!
When Codex runs it runs through multiple shell commands without explaining the rationale behind them. This goes similarly with the thinking process when using the reasoning models. This may be difficult to make but it would be nice to interrupt between each "step" the model goes through, enabling the user to ask the reason why the model decided to write a shell command or code. This could be done through a separate converse mode or letting the model to ask the user more throughout its process (maybe letting the user choose from 1-10 from how much detail you want the model to provide with the user). I have tried to achieve the above through instructions.md but without success.
@tibo-openai I think adding an Interactive Reasoning Mode to Codex could be a really valuable feature for the community. The idea is to make Codex feel more like a collaborative programming partner—by showing its reasoning and letting users jump in to guide the process when needed.
This helps in a bunch of ways: it builds trust by making Codex’s thought process transparent, gives users more control when things could go off track, and even helps folks learn by seeing the logic behind each step.
For Example:
<img width="365" alt="Image" src="https://github.com/user-attachments/assets/6d26255c-b3eb-4a92-b74a-b428ec706246" />
If the idea resonates, I’d love to start building it out and get your feedback along the way!
Thanks for the feedback! We're primarily focused on the native/Rust implementation of Codex CLI and would love to see if this is still a desired feature on the latest version. If it is, please let us know, and we can explore how to incorporate such functionality.