Feature Request: Add "/whisper" mode for lightweight, non-disruptive suggestions

Resolved 💬 2 comments Opened May 27, 2026 by samkit Closed May 27, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What variant of Codex are you using?

Desktop App

What feature would you like to see?

Problem

When using Codex, there are situations where it starts going down a suboptimal path. Currently, the only way to correct it is to interrupt the flow entirely by steering the conversation, which stops progress and forces a full context reset. This feels heavy-handed for minor course corrections.

Proposed Solution

Introduce a new /whisper command (or similar lightweight mechanism) that allows users to give gentle suggestions without interrupting Codex's current progress.

Example usage:

/whisper: Consider using a more efficient algorithm here
/whisper: Maybe add error handling for this edge case

Codex can then decide whether to incorporate the suggestion or continue with its current approach — similar to casually advising a fellow programmer.

Benefits

  • Enables subtle influence rather than full redirection
  • Reduces unnecessary conversation resets
  • Creates a more natural, collaborative feel (human-AI pair programming)
  • Keeps momentum while still guiding the model

Use Cases

  • Suggesting better variable names or structure mid-generation
  • Hinting at edge cases without breaking flow
  • Suggesting performance improvements while Codex is writing code
  • Guiding style or architecture preferences gently

This would make the interaction with Codex feel significantly more fluid and human-like.

Additional information

_No response_

View original on GitHub ↗

2 Comments

github-actions[bot] contributor · 3 months ago

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

  • #24557

Powered by Codex Action

etraut-openai contributor · 3 months ago

This is what steering is designed to handle. A steered message should not generally pull the agent completely off course but subtly redirect it. In most cases this works, but I sometimes see the behavior that you're describing. If you see that, please consider using /feedback to report it. That will help our team refine the training of future models to better handle steered messages.

You might also want to play with different phrasings of the steered prompt to help the model understand whether it's a subtle nudge or suggestion versus a forceful interjection or command.