Codex CLI v0.159.0 lets you steer Codex mid-answer

Codex CLI v0.159.0 lets you steer Codex while it's still answering. Send a new message in the middle of a response and it reaches the model right away, instead of waiting in the queue.

Slide text

01 02 03 04 Steer Codex mid-answer

openai/codex · GitHub release rust-v0.159.0 · openai/codex · pull requests #48135, #48141

Released September 29, 2026 Opt-in

It works on long running code mode calls too. Codex picks up your message as soon as you send it, and the code it started keeps running in the background. The setting is opt-in, and it's off by default.

Next, the terminal screen.

New sessions now open on a compact welcome screen with a short greeting. Resumed, forked and cleared sessions get the same borderless header, with the title, the version and your folder.

Codex now shows an occasional tip. One appears under the working status after thirty seconds, and a few show up after finished turns, at most two per session. The existing tooltips setting turns them off.

When you close the warnings viewer, it now dismisses the warnings you looked at, so the badge stops counting them. To keep one for later, press K and move on to the next.

When Codex asks whether to implement a plan, that prompt used to block scrolling. Now your mouse wheel scrolls the transcript behind it, so you can check a long plan's earlier steps before you decide.

Codex draws Mermaid diagrams right in the terminal, and it now understands more flowchart syntax. Dashed and two way arrows, labels on edges, punctuation inside labels, and node groups that fan out into every edge.

Copying from the transcript now keeps Markdown tables as tables, along with formatting and meaningful whitespace, like the spaces in code. And more terminals now copy automatically when you select text.

Then the fixes.

On Windows, starting MCP servers, the code mode host or piped commands no longer flashes stray console windows. And under launchers that stop background processes, Codex falls back to embedded mode instead of failing to open.

Your AWS folder is now protected by default inside folders Codex may write to, because AWS profiles can run credential helpers. And commands you approve still respect the files you've explicitly blocked.

Signing in with ChatGPT on your own machine now opens the browser reliably, and onboarding can copy the login link for you. Blank sessions keep their drafts when you switch tasks. And on macOS, secure connections work in sandboxes with network access.

And two removals.

Codex no longer suggests your next message after a turn, and the setting that controlled those suggestions is gone. The plugin creator skill no longer ships with Codex.

The release notes list every change in this version. You'll find them on the Codex releases page on GitHub, under v0.159.0.