Codex CLI Update: Voice, the /agents View, Resume and Worktrees

This research library uses AI-assisted source research and drafting. Linked sources support product claims; analysis and proposed exercises are our interpretation. Unless an article documents a test and its results, do not read it as a hands-on review or an independently verified benchmark.

The Codex CLI update announced at DevDay on 29 September 2026 lets you start and steer tasks with voice, track several tasks in a new /agents view, edit prompts, resume sessions, and work in Git worktrees, all in a cleaner terminal UI. OpenAI lists the refresh for all plans. Every command below comes from the Codex CLI docs, the CLI reference, or the Codex changelog, and I left out anything they do not document.

Install the Codex CLI update and check your version

On macOS and Linux, the standalone installer both installs and updates:

curl -fsSL https://chatgpt.com/codex/install.sh | sh

If you installed through npm, the changelog pins each release, for example npm install -g @openai/codex@0.159.0 for the DevDay build. codex update checks for and applies an update when your installed release supports self-update. Run codex in a project directory and choose Sign in with ChatGPT the first time.

The new workflows arrived over several releases, so your version decides what you see:

Release 0.156.0 also fixed two rough edges: resumed sessions restore Plan mode, and editing an earlier prompt keeps the thread identity and settings.

How do you use Codex CLI voice?

Start codex and press F8 to begin or end a voice conversation. Type /voice settings to choose the voice for future conversations. Release 0.156.0 bundles the audio runtimes for Linux and Windows.

Voice works best for dispatching and steering. Say the task in one go: which failing test, which directory, and what done looks like. Then read the plan and the diff on screen, because spoken summaries are a poor way to review code.

Check /permissions before you start talking. The desktop docs say ChatGPT Voice follows the same permissions as the tasks it directs, and your approval settings still decide which commands run without asking. Plan availability for voice in the CLI is not documented separately. The pricing page lists ChatGPT Voice for ChatGPT plans and not for API-key sign-in.

Track parallel work in the /agents view

OpenAI's recap calls it the /agents view, and the release notes call it the agent command center or agents overview. It lists your sessions so you can delegate work and follow several tasks from one screen. Recent releases added status filter tabs, token and usage estimates per task, model grouping, and hide, archive, and delete actions. You can also create a worktree session straight from it.

At the time of writing, the CLI slash command reference lists /agent and its alias /subagents for switching the active agent thread. Type / and read the popup to see what your build offers.

For work that Codex splits itself, ask for subagents directly:

Spawn one agent to map how the auth module refreshes tokens and another to list the failing tests in tests/auth. Wait for both, then summarize what they found before you change anything.

Use /agent to inspect a subagent thread while it runs. Approval requests can come from an inactive thread while you watch the main one, so read which thread is asking before you approve. Subagent workflows use more tokens than a single agent, because each one does its own model and tool work.

To cap how many spawned threads a session keeps open, add this to the project's .codex/config.toml:

[agents]
max_concurrent_threads_per_session = 4

Edit a prompt instead of piling on corrections

Press Escape twice with an empty composer to edit your previous message. Codex forks the chat from that point, so the wrong instruction and its output leave the context instead of sitting under a correction.

A few other composer habits pay off quickly:

• Up and Down restore draft history, and Ctrl+R searches earlier prompts.
• Enter while Codex is working injects new instructions into the current turn.
• Tab while Codex is working queues a follow-up for the next turn.
• /fork branches the current chat so you can try another approach and keep the transcript.

Resume sessions with codex resume

Give a session a name with /rename so you can find it again. Then pick it up from the shell:

# Choose from recent chats in this repository
codex resume

# Skip the picker and reopen the most recent chat here
codex resume --last

# Include sessions started in other directories
codex resume --all

# Reopen a specific session by name or ID
codex resume auth-token-refresh

# Branch your most recent session into a new chat
codex fork --last

Inside the TUI, /resume opens the same session list. If you resume from a different directory than the one the session used, Codex asks which to use. Set tui.resume_cwd to "current" or "session" in your config to answer that once.

Run parallel tasks in worktrees

A worktree gives each task its own checkout of the repository while sharing the .git metadata, so two sessions stop editing the same files. Worktree support is on by default since 0.156.0. The worktrees guide says Codex creates its managed worktrees in $CODEX_HOME/worktrees on a detached HEAD, which keeps your branch list clean.

Git's own rule still applies: a branch can be checked out in only one worktree at a time. Ignored files such as .env do not come along automatically. For desktop app managed worktrees, the guide documents a .worktreeinclude file that lists which ignored files to copy.

Try the new workflows on a scratch branch

1. Update the CLI, open a repository with a clean working tree, and create a Git checkpoint.
2. Start codex, press F8, and describe two independent jobs, such as a failing test and a README fix.
3. Open the agents view and start the second job as a worktree session.
4. When one session drifts, press Escape twice and rewrite the prompt.
5. Before you open a pull request, run /diff and then /review against the base branch in each session.
6. The next morning, run codex resume --last to continue where you stopped.

The review step is the one to keep. Our Delegate, Review, Own methodology explains why each parallel session should end with a diff a person can check in one sitting. For more terminal patterns, see the Codex CLI workflows topic.

Update the CLI now and try the Escape-twice edit on your next real prompt.

Further reading

• Codex CLI docs
• Codex CLI command reference
• Codex changelog
• Worktrees guide