The --full-auto Flag Has Been Removed

The mistake

Running codex exec --full-auto in a CI job or automation script — usually because that's the flag in an older tutorial, a colleague's script from a while back, or a Stack Overflow answer that hasn't been revisited since it was written — after the flag has been removed.

What happened to the flag

For a while the flag was deprecated but still ran, printing a warning alongside normal output. Codex CLI 0.147.0, released on August 7, 2026, removed it from codex exec outright; OpenAI's changelog points to --sandbox workspace-write instead, and the current command reference doesn't document the flag at all. This page used to say the flag still worked. By the time this site re-checked the CLI in October, it had already been gone for weeks.

Why this happens

A deprecated flag that still works produces no immediate signal that anything is wrong — the only sign was a warning, trivial to miss in a CI log that nobody reads unless a job actually fails, and easy to filter out entirely if a pipeline only surfaces stderr on non-zero exit. So scripts written during the deprecation period kept the flag, and the removal turned a warning nobody read into a job that stops working.

Why it matters

An unattended pipeline is the worst place to discover a removed flag, because nobody's watching interactively when the release lands. A pinned, older CLI version can hide the problem for a while longer, which only moves the breakage to whenever the pin is next updated.

The fix

Replace --full-auto with the current explicit form, which spells out the sandbox mode directly rather than bundling it behind a shorthand name: codex exec --sandbox workspace-write. If your pipeline needs specific approval behavior alongside that sandbox setting — for instance, never prompting at all in a non-interactive run — pair it with the matching --ask-for-approval value rather than relying on a single flag to imply both. Grep any existing automation for the old flag name once; it's a small, mechanical fix.

A useful habit once you've made the swap: treat any command line copied from an old README, blog post or internal wiki page as suspect by default, not just this one flag. Codex's own CLI surface has moved fast enough that a command that was current a few months ago is a reasonable place to expect exactly this kind of quiet removal, and checking codex exec --help against what a script actually invokes takes seconds compared to debugging a pipeline that broke on a release nobody read the notes for.