Request and model safety
- Pin exact production model identifiers.
- Validate the exact request shape you send in production.
- Re-test capability-specific behavior when switching providers or models.
Stateful behavior
- Decide how you will store conversation state.
- Verify tool loops, retries, and replay behavior if your app uses tools.
- Test streaming behavior if your UI or workers depend on it.
Error handling
- Handle synchronous API errors such as auth failures, validation errors, and insufficient credits.
- Handle mid-stream failures if you use streaming.
- Log request IDs, status codes, and failure modes in your own monitoring.
Access and cost control
- Separate development and production API keys.
- Confirm paid-model access and account balance before launch.
- Monitor costs and usage through account activity and usage data.