Codex ships a flag that puts a model, not a human, in the approval seat, and says so in its name
The flag routes approvals through automatic review and pins workspace-write as the default. It is named honestly, which matters: this is the same substitution several vendors are making, stated plainly rather than buried in a default. Note the ordering caveat from the merged PR: later subcommand permission overrides are preserved, so the flag does not unconditionally determine the effective policy.
What this changes for operators
- Decide deliberately whether a model reviewing a model is the control you want, and write it down. This is the choice of the fortnight, and taking it by default is different from taking it on purpose.
- Do not describe the flag as having the final say. Later subcommand overrides still apply, so the effective policy is a composition rather than a single setting.
Primary sources
Signal metadata
Source findings
- 2026-08-10-codex-approve-for-me-puts-a-model-not-a-human-in-the-approval 2026-08-10-codex-approve-for-me-puts-a-model-not-a-human-in-the-approval
Run: 2026-08-10-weekly-digest-2026-08-03_2026-08-10-frontier-v0
Schema: bitter.frontier_signals.v0 / ID: 2026-08-10-codex-approve-for-me-puts-a-model-in-the-seat
Research evidence and publication history are open in the repository.