Codex CLI 0.146.1 auto-reviews cyber-capable models by default
The stable patch lands as labs wrestle with cyber-capable agents: Codex now explains permission changes inline and reviews advanced models before they act.
Copy markdown
Stricter review is now the default, not a flag
Codex 0.146.1 ships safer automatic-review defaults for cyber-capable models, so an advanced model's actions get an extra review pass before they land — with no config change on your part. It's the new stable `latest`, published around 16:00 UTC (9 PM Dubai).
Permission changes now show up in the terminal
Instead of quietly escalating, Codex now spells out permission changes in the terminal UI, so you see the moment a session's access shifts. It rhymes with this week's Claude Code 2.1.222 permission-bypass fixes — the whole agent stack is hardening at once.
Upgrading is a one-liner
Run `codex --upgrade` or `npm i -g @openai/codex@latest`. It's a patch over 0.146.0, so a CI bump is low-risk; the faster-moving 0.147 line is still alpha-only if you want the newer thread and agent-plugin work.
Elsewhere: ChatGPT stops big pastes eating your context
ChatGPT now auto-converts pastes over ~10,000 characters into attachments, keeping the composer clean and stopping one giant paste from swallowing your context window; a 'Show in text field' button restores it. Rolling out across plans.