Add initial version of invoke_agent tracing instrumentation

Author: stanek-michal

instrumentation/opentelemetry-instrumentation-botocore/src/opentelemetry/instrumentation/botocore/extensions/__init__.py

@@ -33,6 +33,7 @@ def loader():
 
 _KNOWN_EXTENSIONS = {
     "bedrock-runtime": _lazy_load(".bedrock", "_BedrockRuntimeExtension"),
+    "bedrock-agent-runtime": _lazy_load(".bedrock", "_BedrockRuntimeExtension"),
     "dynamodb": _lazy_load(".dynamodb", "_DynamoDbExtension"),
     "lambda": _lazy_load(".lmbd", "_LambdaExtension"),
     "sns": _lazy_load(".sns", "_SnsExtension"),

instrumentation/opentelemetry-instrumentation-botocore/src/opentelemetry/instrumentation/botocore/extensions/bedrock.py

@@ -56,10 +56,16 @@
     GEN_AI_TOKEN_TYPE,
     GEN_AI_USAGE_INPUT_TOKENS,
     GEN_AI_USAGE_OUTPUT_TOKENS,
+    GEN_AI_AGENT_ID,
+    GEN_AI_AGENT_NAME,
+    GEN_AI_TOOL_CALL_ID,
+    GEN_AI_TOOL_NAME,
+    GEN_AI_TOOL_TYPE,
     GenAiOperationNameValues,
     GenAiSystemValues,
     GenAiTokenTypeValues,
 )
+
 from opentelemetry.semconv._incubating.metrics.gen_ai_metrics import (
     GEN_AI_CLIENT_OPERATION_DURATION,
     GEN_AI_CLIENT_TOKEN_USAGE,
@@ -118,6 +124,7 @@ class _BedrockRuntimeExtension(_AwsSdkExtension):
         "ConverseStream",
         "InvokeModel",
         "InvokeModelWithResponseStream",
+        "InvokeAgent",
     }
     _DONT_CLOSE_SPAN_ON_END_OPERATIONS = {
         "ConverseStream",
@@ -147,6 +154,9 @@ def setup_metrics(self, meter: Meter, metrics: dict[str, Instrument]):
     def _extract_metrics_attributes(self) -> _AttributeMapT:
         attributes = {GEN_AI_SYSTEM: GenAiSystemValues.AWS_BEDROCK.value}
 
+        if self._call_context.operation == "InvokeAgent":
+            attributes[GEN_AI_OPERATION_NAME] = "invoke_agent"
+
         model_id = self._call_context.params.get(_MODEL_ID_KEY)
         if not model_id:
             return attributes
@@ -170,6 +180,19 @@ def extract_attributes(self, attributes: _AttributeMapT):
 
         attributes[GEN_AI_SYSTEM] = GenAiSystemValues.AWS_BEDROCK.value
 
+        # Handle InvokeAgent
+        if self._call_context.operation == "InvokeAgent":
+            attributes[GEN_AI_OPERATION_NAME] = "invoke_agent"
+
+            # Set agent attributes
+            agent_id = self._call_context.params.get("agentId")
+            agent_alias_id = self._call_context.params.get("agentAliasId")
+
+            self._set_if_not_none(attributes, GEN_AI_AGENT_ID, agent_id)
+            self._set_if_not_none(attributes, GEN_AI_AGENT_NAME, agent_alias_id)
+            return
+
+        # Handle non-agent chat completions
         model_id = self._call_context.params.get(_MODEL_ID_KEY)
         if model_id:
             attributes[GEN_AI_REQUEST_MODEL] = model_id
@@ -329,10 +352,14 @@ def before_service_call(
 
         if span.is_recording():
             operation_name = span.attributes.get(GEN_AI_OPERATION_NAME, "")
-            request_model = span.attributes.get(GEN_AI_REQUEST_MODEL, "")
-            # avoid setting to an empty string if are not available
-            if operation_name and request_model:
-                span.update_name(f"{operation_name} {request_model}")
+            if self._call_context.operation == "InvokeAgent":
+                if operation_name:
+                    span.update_name(f"{operation_name}")
+            else:
+                request_model = span.attributes.get(GEN_AI_REQUEST_MODEL, "")
+                # avoid setting to an empty string if are not available
+                if operation_name and request_model:
+                    span.update_name(f"{operation_name} {request_model}")
 
         # this is used to calculate the operation duration metric, duration may be skewed by request_hook
         # pylint: disable=attribute-defined-outside-init
@@ -472,6 +499,65 @@ def _on_stream_error_callback(
                 attributes=metrics_attributes,
             )
 
+    def _invoke_agent_on_success(
+        self,
+        span: Span,
+        result: dict,
+        instrumentor_context: _BotocoreInstrumentorContext,
+    ):
+        try:
+            if "completion" in result and isinstance(result["completion"], EventStream):
+                event_stream = result["completion"]
+
+                # Drain the stream so we can instrument AND keep events
+                all_events = list(event_stream)
+
+                # A replay generator so user code can still iterate
+                result["completion"] = _replay_events(all_events)
+
+                for event in all_events:
+                    if "returnControl" in event:
+                        self._handle_return_control(span, event)
+
+            # Record metrics
+            metrics = instrumentor_context.metrics
+            metrics_attributes = self._extract_metrics_attributes()
+            if operation_duration_histogram := metrics.get(GEN_AI_CLIENT_OPERATION_DURATION):
+                duration = max((default_timer() - self._operation_start), 0)
+                operation_duration_histogram.record(
+                    duration,
+                    attributes=metrics_attributes,
+                )
+
+        except json.JSONDecodeError:
+            _logger.debug("Error: Unable to parse the response body as JSON")
+        except Exception as exc:  # pylint: disable=broad-exception-caught
+            _logger.debug("Error processing response: %s", exc)
+
+    def _handle_return_control(self, span: Span, event: dict):
+        return_control = event["returnControl"]
+        invocation_id = return_control.get("invocationId")
+        invocation_inputs = return_control.get("invocationInputs", [])
+
+        if span.is_recording() and invocation_id:
+            span.set_attribute(GEN_AI_TOOL_CALL_ID, invocation_id)
+
+            for input_item in invocation_inputs:
+                # Handle function invocation
+                if "functionInvocationInput" in input_item:
+                    func_input = input_item["functionInvocationInput"]
+                    action_group = func_input.get("actionGroup")
+                    function = func_input.get("function")
+                    span.set_attribute(GEN_AI_TOOL_NAME, action_group)
+                    span.set_attribute(GEN_AI_TOOL_TYPE, "function")
+
+                # Handle API invocation
+                elif "apiInvocationInput" in input_item:
+                    api_input = input_item["apiInvocationInput"]
+                    action_group = api_input.get("actionGroup")
+                    span.set_attribute(GEN_AI_TOOL_NAME, action_group)
+                    span.set_attribute(GEN_AI_TOOL_TYPE, "extension")
+
     def on_success(
         self,
         span: Span,
@@ -481,6 +567,12 @@ def on_success(
         if self._call_context.operation not in self._HANDLED_OPERATIONS:
             return
 
+        # Handle InvokeAgent
+        if self._call_context.operation == "InvokeAgent":
+            self._invoke_agent_on_success(span, result, instrumentor_context)
+            return
+
+        # Handle non-agent chat completions
         capture_content = genai_capture_message_content()
 
         if self._call_context.operation == "ConverseStream":
@@ -754,3 +846,10 @@ def on_error(
                 duration,
                 attributes=metrics_attributes,
             )
+
+def _replay_events(events):
+    """
+    Helper so that user can still iterate EventStream
+    """
+    for e in events:
+        yield e

instrumentation/opentelemetry-instrumentation-botocore/test-requirements-0.txt

@@ -37,3 +37,5 @@ zipp==3.19.2
 -e opentelemetry-instrumentation
 -e propagator/opentelemetry-propagator-aws-xray
 -e instrumentation/opentelemetry-instrumentation-botocore
+git+https://github.com/open-telemetry/opentelemetry-python.git@main#subdirectory=opentelemetry-semantic-conventions
+

instrumentation/opentelemetry-instrumentation-botocore/test-requirements-1.txt

@@ -37,3 +37,5 @@ zipp==3.19.2
 -e opentelemetry-instrumentation
 -e propagator/opentelemetry-propagator-aws-xray
 -e instrumentation/opentelemetry-instrumentation-botocore
+git+https://github.com/open-telemetry/opentelemetry-python.git@main#subdirectory=opentelemetry-semantic-conventions
+

instrumentation/opentelemetry-instrumentation-botocore/tests/bedrock_utils.py

@@ -273,6 +273,29 @@ def assert_message_in_logs(log, event_name, expected_content, parent_span):
     assert_log_parent(log, parent_span)
 
 
+def assert_invoke_agent_attributes(span, agent_id, agent_alias_id, session_id, has_tool_call=False, is_result_call=False):
+    # Check system and operation name
+    assert span.attributes.get(GenAIAttributes.GEN_AI_SYSTEM) == GenAIAttributes.GenAiSystemValues.AWS_BEDROCK.value
+    assert span.attributes.get(GenAIAttributes.GEN_AI_OPERATION_NAME) == "invoke_agent"
+
+    # Check agent attributes
+    assert span.attributes.get(GenAIAttributes.GEN_AI_AGENT_ID) == agent_id
+    assert span.attributes.get(GenAIAttributes.GEN_AI_AGENT_NAME) == agent_alias_id
+
+    # If tool call exists, check tool attributes
+    if has_tool_call:
+        assert GenAIAttributes.GEN_AI_TOOL_CALL_ID in span.attributes
+        assert GenAIAttributes.GEN_AI_TOOL_NAME in span.attributes
+        assert GenAIAttributes.GEN_AI_TOOL_TYPE in span.attributes
+        allowed_tool_types = {"extension", "function", "datastore"}
+        assert span.attributes.get(GenAIAttributes.GEN_AI_TOOL_TYPE) in allowed_tool_types, \
+            f"Unexpected tool type in span: {span.attributes.get(GenAIAttributes.GEN_AI_TOOL_TYPE)}"
+    elif is_result_call:
+        assert GenAIAttributes.GEN_AI_TOOL_CALL_ID not in span.attributes
+        assert GenAIAttributes.GEN_AI_TOOL_NAME not in span.attributes
+        assert GenAIAttributes.GEN_AI_TOOL_TYPE not in span.attributes
+
+
 def assert_all_metric_attributes(
     data_point, operation_name: str, model: str, error_type: str | None = None
 ):
@@ -286,8 +309,9 @@ def assert_all_metric_attributes(
         data_point.attributes[GenAIAttributes.GEN_AI_SYSTEM]
         == GenAIAttributes.GenAiSystemValues.AWS_BEDROCK.value
     )
-    assert GenAIAttributes.GEN_AI_REQUEST_MODEL in data_point.attributes
-    assert data_point.attributes[GenAIAttributes.GEN_AI_REQUEST_MODEL] == model
+    if model is not None:
+        assert GenAIAttributes.GEN_AI_REQUEST_MODEL in data_point.attributes
+        assert data_point.attributes[GenAIAttributes.GEN_AI_REQUEST_MODEL] == model
 
     if error_type is not None:
         assert ERROR_TYPE in data_point.attributes