What happened

OpenAI shipped Codex version 0.147.0 on August 7. Codex is OpenAI’s command-line coding agent, the closest tool it has to Claude Code: you run it in a terminal and it reads, writes, and executes code on your machine.

The headline change is a new flag: --approve-for-me. Codex normally pauses before running a command that could change your project or your system, like installing a package or deleting a file, and asks you to approve it. The new flag changes that. According to the public code change on GitHub that added it, --approve-for-me routes approval requests “through automatic review” instead of asking you directly.

That’s a distinct setting from full autonomy. Codex already has flags like --full-auto that skip approval checks entirely. This one keeps a review step, but hands that review to an automated process instead of you. OpenAI’s public release notes and the pull request don’t spell out what that automated reviewer checks for or what makes it approve or reject a given command. That detail simply isn’t published yet.

The same release tightens control in a different spot. Codex now requires you to explicitly confirm you trust a project folder before it will act on it, if it doesn’t recognize that folder already. It also improves redaction of secrets and bearer tokens, credentials that let an app act on your behalf without you typing a password each time, so they stop showing up in your terminal history or command logs.

Why it matters

--approve-for-me exists to remove a real annoyance: babysitting a stream of approval prompts during a long coding session. That’s a legitimate problem, and an automated review step is a reasonable middle ground between approving every command by hand and turning approvals off entirely.

The catch is that you can’t currently verify what the automated reviewer would catch. Claude Code shipped two rounds of permission-check patches this week after finding that its own approval prompts could be evaded or made to show less than what would actually run. The lesson from that isn’t that Claude Code is uniquely risky. It’s that an approval step is only as good as your ability to trust what it’s actually checking. Codex’s new flag asks you to trust a review process OpenAI hasn’t documented.

Who should care

Anyone running Codex who finds themselves approving the same categories of command over and over, especially in unattended workflows that run without anyone watching each step. If you’ve been tempted to reach for --full-auto just to stop the interruptions, --approve-for-me is the option OpenAI is offering as a safer middle step.

What builders should do next

Don’t turn on --approve-for-me in a project with real credentials, production access, or anything you can’t afford to lose, until OpenAI documents what the automatic reviewer actually checks. Test it somewhere disposable first.

A useful test: pick a bounded task, like adding a new feature and its automated tests, and run it twice on a throwaway copy of your project, once with --approve-for-me and once with Codex’s default manual approval. Compare three things: how many commands the automatic review let through that you would have blocked by hand, how much time the automated version saved, and whether both runs produced code that passed the same tests. If the automatic review waves through something you’d have stopped, that tells you more about its behavior than anything in the release notes.

To update, run npm install -g @openai/codex@latest if you installed Codex through npm, brew upgrade --cask codex if you installed it through Homebrew, or codex update to use Codex’s own built-in updater. Then confirm with codex --version that you’re on 0.147.0 or later.

End of article