Arize-ai/openinference

Describe the bug For converse and converse_stream, openinference-instrumentation-bedrock builds llm.invocation_parameters only from inferenceConfig (_converse_attributes.py). Two problems: Model-specific parameters are dropped. Converse passes parameters that are not in inferenceConfig through additionalModelRequestFields (e.g. top_k for Anthropic models; inferenceConfig has no top_k). These are sent to the model but never recorded. A response field is mixed into the request parameters. When the response arrives, the code copies inferenceConfig and then adds the response's stop reason to it: llm_attributes["invocation_parameters"] = dict(request_data["inferenceConfig"]) if "inferenceConfig" in request_data else {} llm_attributes["invocation_parameters"]["stop_reason"] = stop_reason So llm.invocation_parameters contains "stop_reason": "end_turn", which was not a request parameter (it is already recorded as llm.finish_reason). To Reproduce Self-contained: a botocore before-send h

Describe the bug With openinference-instrumentation-bedrock, the LLM span of converse_stream and invoke_model_with_response_stream is only ended when the event stream is read to the end. If the application stops early (break out of the loop, close() the stream, or simply drops it), the span is never ended, even after garbage collection, so it is never exported. Both streams are wrapped in _EventStream (utils/__init__.py); the span is ended by the callback only when the iterator raises StopIteration or an exception. Neither _EventStream nor _Iterator overrides close() or has a finalizer, so an early exit, an explicit close() (forwarded straight to botocore's EventStream), or garbage collection never ends the span. Same class of bug as #3790 (openai), #3784 (cohere), #3930 (anthropic) and #3931 (google-genai). To Reproduce Self-contained: a botocore before-send hook answers the requests with real AWS event-stream framing (no AWS account or network needed). import base64, gc, json, loggin

Describe the bug boto3 documents the body of invoke_model / invoke_model_with_response_stream as "bytes or seekable file-like object". openinference-instrumentation-bedrock parses it with json.loads(kwargs["body"]), which only accepts str / bytes: invoke_model_with_response_stream: _InvokeModelWithResponseStream.handle_response (_wrappers.py) calls json.loads(kwargs["body"]) after the wrapped call has returned. With a file-like body the request has already been sent (and billed), then TypeError: the JSON object must be str, bytes or bytearray, not BytesIO is raised into the application, which never gets the response stream. The span is started with end_on_exit=False, so it is never ended either. invoke_model: the same json.loads in _model_invocation_wrapper is inside a try/except, so the call succeeds, but no input is recorded, is_claude_message_api stays False, and the Anthropic response is then not recognised, so the output is not recorded either. The module docstring and comments

Describe the bug With @arizeai/openinference-instrumentation-openai-agents (0.3.2, also main), the span the processor creates for a function tool is never made active, so spans created while the tool runs (HTTP calls, database or RPC clients, any other instrumentation) are not its children. Inside the tool's execute, trace.getActiveSpan() is undefined, and a span started there begins a new trace. OpenInferenceTracingProcessor.onSpanStart builds a parent context and calls this.oiTracer.startSpan(...) (js/packages/openinference-instrumentation-openai-agents/src/processor.ts, around line 1225), keeping the span in a private otelSpans map; nothing calls context.with or attaches the context. The Python processor does attach it (_processor.py, line 180: self._tokens[span.span_id] = attach(set_span_in_context(otel_span)), detached in on_span_end), so in Python such child spans nest under the tool span. To Reproduce A scripted model (no API key, no network) calls one tool, which starts a span:

Summary get_llm_input_message_attributes yields a list-valued content straight into the string-typed message.content span attribute. The repo's own GenAI translation then renders it with str(), so the span reports the user's message as a Python repr: input: [{"role": "user", "content": [{"type": "text", "text": "hello"}]}] span: gen_ai.input.messages -> ... "content": "[{'type': 'text', 'text': 'hello'}]" With string content the same call produces "content": "hello". Where python/openinference-instrumentation/src/openinference/instrumentation/_attributes.py:680-681: if (content := message.get("content")) is not None: yield f"{base_key}.{message_index}.{MESSAGE_CONTENT}", content No type check. The contract it violates is declared in the same package, _types.py:87-92: class Message(TypedDict) declares content: str, contents: "Sequence[MessageContent]", tool_call_id: str, tool_calls: "Sequence[ToolCall]" — so contents is already the declared channel for content blocks. _genai_c