Codex Lets You Answer Its Questions Inline While It Keeps Working

CodexView original changelog

Codex CLI 0.154.0 added inline question answering, letting developers respond to a question Codex raises mid-task using either suggested choices or custom text, without losing the main draft they were already composing. Previously an agent question competed with whatever a developer had typed into the prompt, forcing a choice between answering and preserving in-progress work. Codex keeps working while the answer is given.

Key Takeaways

  • Codex CLI 0.154.0 lets developers answer an agent question inline while Codex keeps working on the task.
  • Answers can be given via suggested choices for known options or custom text for anything else.
  • The key detail is that the main draft is preserved, so answering no longer costs a half-written instruction.
  • The change separates two input surfaces that previously competed for the same prompt.
  • It targets long interactive sessions, where mid-task questions arrive most often.
  • It shipped alongside experimental worktree support and GPT-6 Astra's addition to the model picker in the same release.

Sources & Mentions

1 external resource covering this update


Answering Without Losing Your Draft

When an agent needs a decision from the developer mid-task, the interruption usually arrives at the worst moment: halfway through composing the next instruction. Codex CLI 0.154.0, released September 9, 2026, addressed that directly by letting developers answer a question inline while Codex continues working, without disturbing the main draft already in the prompt.

The answer can take either of two forms. Codex presents suggested choices for questions where the options are known ahead of time, so a developer can pick one without typing anything. For anything the suggestions do not cover, custom text is accepted instead. Either way, the response goes to the question rather than into the draft, and the draft survives.

Why the Draft Preservation Is the Point

The headline capability sounds small, but the detail that makes it worth noting is the separation of two input surfaces that used to be one. Previously a developer facing an agent question had to either abandon what they were typing or juggle it somewhere else before responding. Treating the question as its own input target removes a genuine and repeated annoyance in long, interactive sessions.

Part of a Broader 0.154.0 Release

Inline answering shipped alongside experimental worktree support, GPT-6 Astra's arrival in the model picker and Amazon Bedrock catalogs, a shared background Codex server for Windows sessions, Vim R replace mode, and formatting-preserving copy. It is a substantial release for the CLI's interactive surface specifically, with several of its headline items aimed at developers who keep a session open and conversational rather than firing one-shot tasks.


Mentioned onOpenAI