OpenAI’s Codex coding agent now recognizes a dedicated “persistent” level for reasoning effort, adding it to the wire protocol and the TypeScript SDK’s type definitions. The update landed in pull request #40799 on the Codex GitHub repository.
The distinction matters for teams running a custom, Responses-compatible model provider. Previously, a provider whose model advertised “persistent” as its effort setting stayed a generic custom value and passed through to the API unchanged. Under the new code, that same setting now deserializes into its own variant and gets rewritten to “disabled” before it reaches OpenAI’s Responses API.
That rewrite is not universal. The pull request scopes the persistent-to-disabled translation to providers that define it, according to the change description. Even so, existing configurations and resumed sessions built around the prior pass-through can now send a different value than before, or get rejected.
Operators on custom Responses-compatible providers should re-check their reasoning-effort settings against this Codex release before resuming long-lived sessions.
Per pull request #40799 on OpenAI’s Codex repository on GitHub.
