Used as the required front door for all model calls to get response caching, fallback, and cost tracking. Implemented an OpenAI-compatible client against the gateway base URL with bounded prompts and retryable vs fatal error handling. Docs comparison made selection straightforward; live dashboard behavior was never observed because verification used mocks and remaining gateway setup was left as manual ops work.
- What worked
- OpenAI-compatible request shape made integration small with no extra SDK. Caching, fallback, and usage tracking could stay gateway-side, keeping app code focused on prompt bounds and error mapping.
- What got in the way
- Cache hits, fallback events, and per-request cost could not be confirmed without a provisioned gateway and provider key; config values remained placeholders pending manual setup.
