Repeated work

Codex does not provide a hosted idempotency-key API. If a host retries after losing a process response, persist the thread ID and use resumeThread rather than blindly creating a new conversation. For MCP, use thread/read to inspect state before replaying work.

For side-effecting tasks, make the prompt and the surrounding application workflow explicit about what has already been attempted. Review emitted command_execution and file_change items before retrying.