IndexAutomatic review
Automatic review
Automatic review in Codex is a second agent that approves or denies actions the sandbox will not run on its own. It replaces the person at that prompt. It does not turn the sandbox off.
Updated 2026-09-27
The docs are Auto-review. Set `approvals_reviewer = "auto_review"` while approvals are still interactive (`approval_policy = "on-request"`, or a granular policy that still asks). If the policy is `never`, nothing is left to review.
What the reviewer sees
- A command that needs to leave the writable roots
- A network request the sandbox blocked
- A side-effecting app or MCP call that still requires approval
- Not the ordinary edits that workspace-write already allows
A denial comes back with a reason, and Codex is told to find a safer path. OpenAI’s alignment note says sessions stop for a person far less often than manual approval, and that the reviewer is not a guarantee. Repeated attempts to talk the reviewer into a bad action can stop the run. In the desktop app these show up as review items: Reviewing, Approved, Denied, Aborted, or Timed out. Extra model calls mean extra Codex usage.
Questions
Does automatic review disable the sandbox?
No. Writable roots, network limits, and protected paths stay as they were. Only the approver changes.
Is this a code-review bot on pull requests?
Codex can also review PRs when you turn that on for GitHub. Automatic review in the sandbox docs is the approval-boundary reviewer, which is the feature this query maps to on OpenAI’s auto-review page.
Related notes
- Codex status and usage
Check Codex uptime on status.openai.com, plan limits at chatgpt.com/codex/settings/usage, and the CLI with /status. API spend is a different page.
- OpenAI Symphony
Symphony is OpenAI’s open-source orchestrator for Codex, tied to Linear. It runs tasks and can expose a local status dashboard.
- OpenAI Agent Builder
Agent Builder is the visual canvas in the OpenAI API docs for composing agents, tools, and handoffs. It is not the ChatGPT Agent button.