Feedback on Codex CLI User Experience (In-Progress Interaction)

What feature would you like to see?

I’ve been using Claude for over six months and recently started using Codex CLI as well.
From a model capability perspective, Codex is solid: it follows user instructions well, and the final outputs are generally good.

However, there is a serious UX problem during task execution, and it significantly hurts the overall experience.

1. Lack of Transparency During Long-Running Tasks

While a task is in progress, Codex CLI provides almost no feedback to the user.

In contrast, Claude Code continuously explains:
• what it is currently doing,
• why it is doing it,
• and how the user should interpret the current step.

This ongoing narration is extremely important, especially for long or complex tasks.
With Codex CLI, the system often appears to “hang” for long periods. Even if the model is actively reasoning or planning internally, the user has no visibility into that process.

From the user’s perspective, this feels like:
• uncertainty about whether the system is still working,
• no way to understand the current direction of the solution,
• and no opportunity to correct misunderstandings early.

At minimum, the CLI should periodically communicate:
• the current phase of work,
• the high-level direction of reasoning,
• or a short status update indicating progress.

Without this, long-running tasks feel opaque and unreliable.

⸻

1. Poor Handling of Mid-Task User Input

Another major issue is how Codex CLI handles user feedback during execution.

With Claude Code:
• the user can provide new input (e.g., change of direction, clarification, constraint),
• and the system naturally incorporates it at the next reasoning step,
• without requiring the user to interrupt the process.

Codex CLI behaves very differently:
• user input is only processed after the entire task completes,
• or the user must press ESC to interrupt execution,
• which completely aborts the ongoing work.

This design strongly discourages interaction.
Users hesitate to press ESC because it means throwing away progress, even when they notice a small but important correction early.

For an interactive CLI tool, this is a critical flaw.
Mid-task guidance should refine the solution—not force a full restart.

1. Overall Impression

In summary:
• The model quality and final outputs are good.
• But the experience feels like “a black box that produces results” rather than “a collaborative tool that works with the user.”

Codex CLI currently prioritizes output over communication.
What’s missing is the sense that the system is:
• explaining its thinking at a high level,
• checking alignment with the user,
• and inviting feedback while the solution is being formed.

Adding continuous, lightweight progress explanations and allowing non-destructive mid-task input would dramatically improve trust, usability, and perceived intelligence.

This is not a cosmetic improvement — it’s fundamental to making Codex CLI feel like a tool developers want to work with, not just wait for.

Additional information

_No response_