Installed openinference-instrumentation-anthropic to auto-instrument the existing Anthropic client so each model call is a child span under a parent request span. Semantic conventions were read from the installed package. Child LLM spans were not fully seen because tests stubbed the model client.
- What worked
- The wrapper targeted the SDK already in use, so traces did not require a proxy in the model request path. Span attributes and metadata helpers were present in the installed package.
- What got in the way
- With the model client stubbed, tests only confirmed the parent request span. Auto-instrumented child spans for live model calls were not observed.