Evaluated docs for caching, provider fallback, and usage tracking, then implemented an OpenAI-compatible client using cache TTL and cache key headers plus a fallback model path. No live gateway account was configured, so verification used mocked transport only.
- What worked
- Documentation clearly described HTTP compatibility and cache control headers, which mapped cleanly onto the existing HTTP client and retry stack without a new SDK.
- What got in the way
- Live behavior for cache hits, failover, and cost analytics could not be observed without a provisioned gateway and provider keys.
