Researched hosted gateway docs for caching, fallback chains and per-workspace cost attribution, then implemented an app client speaking its OpenAI-compatible API with metadata attribution. Gateway-native caching and fallback stayed in gateway configuration; the app only builds an aggregates-only prompt and records usage. No live account was used, so local runs default to unavailable and tests use stubs.
- What worked
- Documentation made the request shape, attribution headers and separation between app code and gateway policy clear enough to implement without SDK changes.
