Answer first
Xiaojian groups model identifiers containing Codex under the GPT family, but a specific Codex model is selectable only when the server publishes it. This guide explains the workflow; it does not claim that every deployment exposes a particular Codex release.Verified facts and limits
- The model picker follows the server-provided catalog and capability declarations.
- Changing a model applies to the next message while existing tasks remain independent.
- The repository does not provide enough production evidence to claim universal Codex availability.
Write the request as a bounded task
Start with the outcome, writable directories, forbidden areas, and proof of completion. “Only change public routes, build paired pages, and prove the signed-in workspace is untouched” is more actionable than “refactor the website.”
Ask for evidence before edits
Have the model read repository instructions, nearby implementation, and test configuration, then report the current state. In a dirty worktree, explicitly preserve existing changes and require path-level diff evidence at handoff.
- Authoritative source and design docs
- Writable and forbidden paths
- Tests, build checks, and screenshot expectations
End with a stop condition
Require fresh test output, known failures, unverified risks, and the exact commit scope. If Codex is absent from the catalog, select another available coding model or wait for an administrator to publish it; a family label is not an availability guarantee.
Sources and evidence
- Repository implementation evidence: apps/api/src/provider.ts and apps/web/app/chat/page.tsx (model catalog and family mapping, 2026-08-02)
- Specific Codex models and production availability must be confirmed per deployment
Capability status reflects repository evidence on 2026-08-02. It is not a production availability or commercial-readiness guarantee.