Read the hosted tracing quick start, confirmed an OpenAI-compatible proxy fit, and wired a fail-closed client that sends both sequential model calls through that proxy with shared session tags. No live request was sent to the service.
- What worked
- The proxy model was easy to apply: keep the existing provider key, add a tracing auth header, point the client at the proxy, and attach session and step metadata so related calls group for later search. Docs were enough to implement without a new SDK.
- What got in the way
- Dashboard search, cost, latency, and failure capture were not observed against a real project, so production behavior of the proxy was not verified.