Remove provider name from local agent span

instrumentation/opentelemetry-instrumentation-genai-langchain/.changelog/132.removed

@@ -0,0 +1 @@
+Stopped setting `gen_ai.provider.name` on internal agent spans.

instrumentation/opentelemetry-instrumentation-genai-langchain/src/opentelemetry/instrumentation/genai/langchain/callback_handler.py

@@ -92,9 +92,6 @@ def on_chain_start(
                 )
                 if suggested_agent_name_lower != agent_invocation_name_lower:
                     agent = self._telemetry_handler.invoke_local_agent(
-                        provider=metadata.get("ls_provider", "unknown")
-                        if metadata
-                        else "unknown",
                         agent_name=suggested_agent_name,
                     )
                     agent.input_messages = make_input_message(inputs)

instrumentation/opentelemetry-instrumentation-genai-langchain/tests/test_callback_handler.py

@@ -165,7 +165,6 @@ def test_new_agent_span_created(self):
         )
 
         telemetry.invoke_local_agent.assert_called_once_with(
-            provider="openai",
             agent_name="math_agent",
         )
         assert agent_inv.agent_name == "math_agent"

policies/genai_span_validation.rego

@@ -3,7 +3,7 @@
 # (name, type, presence) for spans matching its definitions; this file adds
 # cross-cutting span-level invariants the registry can't easily express.
 #
-# Two classes of rules, both keyed on `gen_ai.operation.name`:
+# Three classes of rules, all keyed on `gen_ai.operation.name`:
 #
 #   1. Span name format → `violation`
 #      (`{operation_name} {request_model}` for inference / embeddings,
@@ -17,6 +17,9 @@
 #      semantic-conventions/docs/gen-ai/gen-ai-spans.md and
 #      gen-ai-agent-spans.md (the MD flattens the YAML inheritance chain
 #      via `extends:`, so it's the right place to source from).
+#      `invoke_agent` is the one operation whose manifest also depends on
+#      span kind: semconv defines separate internal (same-process) and
+#      client (remote) spans, and only the client span carries server.*.
 #
 # The "set when known" Recommended subset (sampling parameters like
 # `frequency_penalty`, `max_tokens`; provider-side caches; conditionally-
@@ -99,21 +102,26 @@ deny contains _span_finding(
 
 # ─── Per-operation expected attributes (violation) ──────────────────────────
 
-_expected_for_op["chat"] := _inference_expected
+# `_expected_for_op(op, kind)` returns the expected-attribute manifest for a
+# span given its `gen_ai.operation.name` and span kind. Most operations ignore
+# kind (second arg is a wildcard); invoke_agent dispatches on it because
+# semconv splits it into separate internal and client spans. Undefined (→ no
+# violations) for an unmapped op or an unexpected agent span kind.
+_expected_for_op("chat", _) := _inference_expected
 
-_expected_for_op["generate_content"] := _inference_expected
+_expected_for_op("generate_content", _) := _inference_expected
 
-_expected_for_op["text_completion"] := _inference_expected
+_expected_for_op("text_completion", _) := _inference_expected
 
-_expected_for_op["embeddings"] := _embeddings_expected
+_expected_for_op("embeddings", _) := _embeddings_expected
 
-_expected_for_op["execute_tool"] := _execute_tool_expected
+_expected_for_op("execute_tool", _) := _execute_tool_expected
 
-_expected_for_op["invoke_agent"] := _invoke_agent_expected
+_expected_for_op("invoke_agent", kind) := _invoke_agent_expected[kind]
 
-_expected_for_op["create_agent"] := _create_agent_expected
+_expected_for_op("create_agent", _) := _create_agent_expected
 
-_expected_for_op["retrieval"] := _retrieval_expected
+_expected_for_op("retrieval", _) := _retrieval_expected
 
 # Inference (chat / generate_content / text_completion).
 # Required: gen_ai.operation.name, gen_ai.provider.name.
@@ -156,11 +164,17 @@ _execute_tool_expected := {
 	"gen_ai.tool.type",
 }
 
-# Invoke agent.
-# Required: gen_ai.operation.name, gen_ai.provider.name.
-_invoke_agent_expected := {
+# Invoke agent (internal)
+# Required: gen_ai.operation.name
+_invoke_agent_expected["internal"] := {
 	"gen_ai.operation.name",
-	"gen_ai.provider.name",
+}
+
+# Invoke agent (client)
+# Required: gen_ai.operation.name, server.address
+_invoke_agent_expected["client"] := {
+	"gen_ai.operation.name",
+	"server.address",
 }
 
 # Create agent. After creation completes the provider returns an agent.id;
@@ -193,7 +207,7 @@ deny contains _span_finding(
 ) if {
 	input.sample.span
 	op := _attr_value(input.sample.span, "gen_ai.operation.name")
-	expected := _expected_for_op[op]
+	expected := _expected_for_op(op, input.sample.span.kind)
 	some attr_name in expected
 	not _has_attr(input.sample.span, attr_name)
 }

util/opentelemetry-util-genai/.changelog/132.removed

@@ -0,0 +1 @@
+Removed the `provider` parameter from the internal agent invocation APIs and stopped emitting `gen_ai.provider.name` on internal agent spans and metrics.

util/opentelemetry-util-genai/src/opentelemetry/util/genai/_agent_invocation.py

@@ -43,8 +43,8 @@ def __init__(
         metrics_recorder: InvocationMetricsRecorder,
         logger: Logger,
         completion_hook: CompletionHook,
-        provider: str,
         *,
+        provider: str | None = None,
         span_kind: SpanKind = SpanKind.INTERNAL,
         request_model: str | None = None,
         server_address: str | None = None,
@@ -64,10 +64,10 @@ def __init__(
             else _operation_name,
             span_kind=span_kind,
         )
-        self.provider = provider
         self.request_model = request_model
         self.server_address = server_address
         self.server_port = server_port
+        self.provider = provider
 
         self.agent_name: str | None = agent_name
         self.agent_id: str | None = None
@@ -104,19 +104,20 @@ def __init__(
     def _get_base_attributes(self) -> dict[str, Any]:
         """Return sampling-relevant attributes available at span creation time."""
         optional_attrs = (
+            (GenAI.GEN_AI_PROVIDER_NAME, self.provider),
             (GenAI.GEN_AI_REQUEST_MODEL, self.request_model),
             (GenAI.GEN_AI_AGENT_NAME, self.agent_name),
             (server_attributes.SERVER_ADDRESS, self.server_address),
             (server_attributes.SERVER_PORT, self.server_port),
         )
         return {
             GenAI.GEN_AI_OPERATION_NAME: self._operation_name,
-            GenAI.GEN_AI_PROVIDER_NAME: self.provider,
             **{k: v for k, v in optional_attrs if v is not None},
         }
 
     def _get_common_attributes(self) -> dict[str, Any]:
         optional_attrs = (
+            (GenAI.GEN_AI_PROVIDER_NAME, self.provider),
             (GenAI.GEN_AI_REQUEST_MODEL, self.request_model),
             (server_attributes.SERVER_ADDRESS, self.server_address),
             (server_attributes.SERVER_PORT, self.server_port),
@@ -127,7 +128,6 @@ def _get_common_attributes(self) -> dict[str, Any]:
         )
         return {
             GenAI.GEN_AI_OPERATION_NAME: self._operation_name,
-            GenAI.GEN_AI_PROVIDER_NAME: self.provider,
             **{k: v for k, v in optional_attrs if v is not None},
         }
 

util/opentelemetry-util-genai/src/opentelemetry/util/genai/handler.py

@@ -389,7 +389,6 @@ def tool(
 
     def start_invoke_local_agent(
         self,
-        provider: str,
         *,
         request_model: str | None = None,
         agent_name: str | None = None,
@@ -409,7 +408,6 @@ def start_invoke_local_agent(
             self._metrics_recorder,
             self._logger,
             self._completion_hook,
-            provider,
             span_kind=SpanKind.INTERNAL,
             request_model=request_model,
             agent_name=agent_name,
@@ -439,7 +437,7 @@ def start_invoke_remote_agent(
             self._metrics_recorder,
             self._logger,
             self._completion_hook,
-            provider,
+            provider=provider,
             span_kind=SpanKind.CLIENT,
             request_model=request_model,
             agent_name=agent_name,
@@ -449,7 +447,6 @@ def start_invoke_remote_agent(
 
     def invoke_local_agent(
         self,
-        provider: str,
         *,
         request_model: str | None = None,
         agent_name: str | None = None,
@@ -469,7 +466,6 @@ def invoke_local_agent(
             self._metrics_recorder,
             self._logger,
             self._completion_hook,
-            provider,
             span_kind=SpanKind.INTERNAL,
             request_model=request_model,
             agent_name=agent_name,
@@ -499,7 +495,7 @@ def invoke_remote_agent(
             self._metrics_recorder,
             self._logger,
             self._completion_hook,
-            provider,
+            provider=provider,
             span_kind=SpanKind.CLIENT,
             request_model=request_model,
             agent_name=agent_name,