Codex CLI v0.167.0 runs the agent command center without a daemon and marks unread results

Codex CLI v0.167.0 opens the agent command center without a shared background server. Run codex agents --no-daemon and it starts an embedded server, so you can switch between chats from a single session.

Slide text

01 02 03 Command center

openai/codex · release notes rust-v0.167.0 · #58238, #58239

Codex CLI v0.167.0 runs the agent command center without a daemon

codex agents --no-daemon

The command center now shows which results you haven't seen. A blue dot marks an unread result, and a gray hollow dot marks an idle task you've already read. The session picker shows the blue dot too.

The session picker lists conversations from every connected backend, and now you can open one that lives on another backend. Your draft is kept. You can't switch during an active turn.

Images and sign-in.

Attach a local image and a thumbnail shows while you type. Thumbnails follow attachment order, with a count for any that don't fit. It needs a supported terminal and a local workspace.

When Codex views an image, its file name in the transcript is now a link that opens the image. That works in local workspaces only, and the links are off in SSH sessions.

The sign-in screen keeps ChatGPT and device code on the main menu. An OpenAI API key, Amazon Bedrock and Google Cloud Vertex AI now sit under a submenu called Sign in another way.

Google Cloud users get a sign-in wizard for Vertex AI. It asks for your project, then uses Application Default Credentials or a credential file. It's behind a feature flag that's off by default.

If your network needs the proxy your operating system already knows about, Codex can now use it on macOS and Windows. It's off by default. You turn it on under /experimental and restart Codex.

When Codex asks you to approve a patch, you can expand its details right in the transcript first. After you answer, the transcript goes back to following new output.

Fixes.

Saved chats now stay in the list after you switch model providers. If one can't resume, you can still read it, and Codex drafts an editable prompt to carry it into a new chat.

Clearing the screen with Ctrl+L used to discard the earlier conversation too. Now the display starts fresh, and the earlier messages are still there in the detailed transcript under Ctrl+T.

A deletion that leaves an exclamation mark first keeps your draft as plain text. Network approval prompts go away when their request ends, one per endpoint. And compaction counts output from failed responses.

If you run several agents at once, I'd update for the unread dots alone. You see which one finished without opening every chat. Every change is in the v0.167.0 release notes on GitHub.