Changelog

​

Added

• bentolabs_sdk.analytics module with init, track_ai, begin, flush, and shutdown. You no longer need to set up an OTel tracer yourself.
• bento.begin(...) for trajectories (multi-step interaction spans). Works as a context manager or imperatively with t.finish(). Out-of-order finish() calls raise RuntimeError.
• @bento.interaction and @bento.tool decorators. Both work on sync and async functions.
• @bento.tool records bound arguments as input.value and the return value as output.value automatically.
• bento.tool_span(...) for standalone tool spans, and interaction.tool_span(...) for tool spans inside an open trajectory. Both set openinference.span.kind="tool".

​

Improved

• Property types are preserved on the wire. int, float, bool, str, and homogeneous lists keep their type. Dicts and mixed lists fall back to JSON strings.

​

Changed

• track_ai and begin now detach from the caller’s active OTel context. BentoLabs spans will not get attached to FastAPI, Django, or other framework traces.

​

Added

• BentoLabsSpanProcessor: OTel SpanProcessor that ships spans to BentoLabs.
• BentoLabsTraceExporter: OTel SpanExporter for batch or custom processor setups.
• resolve_options: resolves API key and base URL from constructor args, env vars (BENTOLABS_API_KEY, BENTOLABS_BASE_URL), or defaults.
• BentoAuthError with a code field. Codes: missing_api_key, invalid_api_key_format, already_initialized.