Codex CLI replaces the recommended stack's Claude Code. Then the differentiator: Codex Cloud takes a ticket from the queue, opens a PR autonomously, and pings you when it's ready for review. Pair it with CodeRabbit and the human reviewer mostly approves clean diffs.
Codex CLI handles the same role Claude Code does: terminal agent with file system access. Codex Cloud is the ChatGPT-bundled cloud lane that takes a task and opens a PR while you do something else. Use both: hands-on for the work you're thinking about, cloud for the cleanup tickets in the queue.
Especially valuable on Codex Cloud's autonomous PRs — those are the ones you didn't watch get written, so you want a second AI reading the diff before a human does.
If you're already thinking about the change, run the CLI. If it's a clear ticket from the backlog (refactor, dep bump, doc update), send it to Cloud and move on.
`codex` in your repo. Same shape as Claude Code: scope, plan, implement, run tests.
Plan first, then implement. Ask 1 to 2 clarifying questions before you start writing if anything is genuinely ambiguous.
Ticket / problem:
"""
{{paste ticket}}
"""
Constraints:
- Smallest viable diff.
- Match the existing code style. Read nearby files before introducing a new pattern.
- Run the project's test command. Don't report done if it fails.
- Treat AGENTS.md / CLAUDE.md / .cursorrules at the repo root as authoritative.
When done: print the file list, the test result, and a 2-line summary I can paste as the commit message.From ChatGPT, hand off the ticket with the repo selected. Cloud reads the repo, plans, opens a draft PR, and pings you. Read the PR description first; reject and retry if scope drifted.
CodeRabbit posts findings on both CLI-authored and Cloud-authored PRs. For Cloud PRs especially, treat CodeRabbit's summary as your first signal that the autonomous lane stayed on the rails.
For CLI work: paste findings into the same session. For Cloud PRs: kick the same task back via the Cloud `Address review feedback` action.
Half the PRs (the ones requiring product judgment) went through Codex CLI. The other half (dep updates, type-error cleanups, doc touch-ups) ran on Codex Cloud while the engineers worked on real product. The Cloud lane shipped 22 PRs in one week with 3 needing a retry for scope drift.
Autonomous lanes occasionally rewrite the world. Always read the PR description before the diff; if the description says 'also refactored…', kill the PR and retry with a tighter ticket.
Cloud tasks are metered separately from the chat subscription. A team that sends 30 cloud tasks/day will exceed the bundled budget within a week. Track it.
Codex CLI is supervised; Cloud isn't. The same prompt works very differently in the two lanes. Don't copy a CLI prompt into Cloud and expect the same result.