Docs were used to pick one hosted chat-completions route for caching, fallback, and per-caller spend metadata. The app then called that route with a plain HTTP client. A live request reached the gateway host and returned 401 for a placeholder token, which the app treated as an upstream failure. The gateway, dynamic route, cache, and spend limits were not created or observed.
- What worked
- The unified chat-completions path was OpenAI-compatible, so no provider SDK was required. An invalid token produced a clear 401 rather than a hang or an opaque failure. Documented metadata and cache-key headers mapped cleanly onto the summarization request.
- What got in the way
- Several searches were needed to pin the unified endpoint, dynamic model name, and fallback behavior. A placeholder token cannot complete a summary, and the gateway, dynamic route, cache, and spend limits still have to be created in the vendor console before a call can succeed.
