The repository supports a local Codex review path that does not require an OPENAI_API_KEY. It uses the developer's existing Codex CLI login, including ChatGPT sign-in, and keeps the session ephemeral.
codex login status
scripts/codex-review.sh --base mainFor unstaged work only:
scripts/codex-review.sh --uncommittedThe wrapper forces a read-only sandbox and approval_policy="never", adds --ignore-user-config so user-configured MCP servers and hooks cannot create external side effects, uses Codex's built-in repository review target, and writes only the final report into ignored .codex-review/. Authentication is still loaded by Codex, but the wrapper does not read, copy, print, or commit authentication files. Codex CLI 0.145 does not allow a custom prompt together with review --base or review --uncommitted, so the project-specific checklist remains documented for maintainer review rather than being passed as a conflicting argument.
ChatGPT sign-in is appropriate for an interactive developer machine. GitHub-hosted automation is programmatic and requires a provider API key or another explicitly supported CI credential. Do not upload a local ChatGPT/Codex authentication file to Actions secrets and do not attempt to replay local login state in CI.
CI therefore validates all deterministic behavior without OpenAI access: syntax, generated configuration, secret invariants, offline bundle contents, the mock Responses-to-Chat-Completions bridge, and container startup. A cloud Codex review workflow can be added later if the project receives an API grant and the maintainer approves its permissions and cost controls.
Local Codex review is advisory. The maintainer remains responsible for checking the diff, test evidence, workflow permissions, release contents, and compatibility claims.