messages_stream() starts an InferenceInvocation (which immediately starts a span and attaches it to the current context) before the underlying MessageStreamManager is entered. If a caller stores the manager and enters it later—or never enters it—this leaves the span/context attached for longer than the actual request (or indefinitely), which can corrupt parent/child relationships and leak context.

Consider deferring handler.start_inference(...) until the manager wrapper’s __enter__ (e.g., pass the handler + extracted params or a factory into MessagesStreamManagerWrapper and create/start the invocation inside __enter__, failing/stopping it there as needed).

AsyncMessagesStreamManagerWrapper.__aenter__() doesn’t handle exceptions from the wrapped manager’s __aenter__. If __aenter__ raises, the passed-in invocation span remains open/attached because invocation.fail(exc) is never called (unlike the sync manager wrapper).

Wrap the awaited self._manager.__aenter__() in a try/except, call self._invocation.fail(exc), then re-raise the original exception.

why is it necessary?

_safe_instrumentation("invocation stop") prevents telemetry cleanup failures from surfacing to users.

_stop() can run during close(), StopIteration, context-manager exit, or response proxy close. If invocation.stop() raises due to instrumentation/exporter/completion-hook behavior, the wrapper should not replace or introduce an application exception. The guard keeps the instrumentation non-invasive: it logs at debug and preserves normal stream behavior. Lmk if you think this is an overkill I can remove it.

we never do this in other places. instrumentation / exporter / completion hook should never raise. if something fails, 1. it must not be otel code 2. we mist bubble it up