Codex
OpenAI Codex agents as an alternative operator backend for hillclimb searches.
Codex is an alternative operator backend for
hillclimb searches: the same search loop
(draft, debug, improve, ensemble), with candidates authored by
headless OpenAI Codex agent calls (codex exec) instead of Claude Code.
Status: local searches. rebase hillclimb start <target> --local --backend codex
runs every operator as a headless codex exec call, using your codex login
(or OPENAI_API_KEY with backend_auth: api-key in hillclimb/config.yaml).
Hosted Codex searches are not released yet.
What It Provides
| Capability | Behavior |
|---|---|
| Search operators | Candidates authored by headless codex exec calls in sandboxed workspaces. |
| ChatGPT-plan billing | Authenticate with a ChatGPT subscription via codex login. |
| API-key billing | Alternatively, an OPENAI_API_KEY bills the OpenAI API directly. |
| Credential hygiene | Same boundary as Claude Code: agent-authored code runs with credentials scrubbed; only the agent process sees the OpenAI credential. |
| Backend selection | rebase hillclimb start --local --backend codex. |
Why Multiple Agent Backends
Different coding agents draft meaningfully different solutions to the same problem — exactly the diversity the search's draft stage wants. A longer-term extension is per-operator backend mixing (for example, Codex drafts with Claude Code debugging), so one search can combine agents' strengths.
Using Codex Today
Outside hillclimb, nothing stops a Rebase function or workflow from invoking the Codex CLI as part of its own logic — the integration described here is specifically about Codex as a first-class search operator backend.