Capture the agent's system prompt as the gen_ai.system_instructions span attribute in the LangChain instrumentation

CHANGELOG.md

@@ -1,5 +1,9 @@
 # Release History
 
+# 1.3.7 (Unreleased)
+### Features Added
+- Capture the agent's system prompt as the `gen_ai.system_instructions` span attribute in the LangChain instrumentation.
+
 # 1.3.6 (2026-07-22)
 
 ### Features Added

src/microsoft/opentelemetry/_genai/_langchain/_tracer.py

@@ -55,6 +55,7 @@
     GEN_AI_PROVIDER_NAME_KEY,
     GEN_AI_REQUEST_CHOICE_COUNT_KEY,
     GEN_AI_REQUEST_MODEL_KEY,
+    GEN_AI_SYSTEM_INSTRUCTIONS_KEY,
     GEN_AI_TOOL_DEFINITIONS_KEY,
     GEN_AI_USAGE_INPUT_TOKENS_KEY,
     GEN_AI_USAGE_OUTPUT_TOKENS_KEY,
@@ -69,10 +70,10 @@
     llm_provider,
     metadata,
     model_name,
-    prompts,
     _extract_structured_output_messages,
     _extract_agent_input_messages,
     _extract_agent_output_messages,
+    _extract_system_instruction,
     _output_message_to_input,
     _is_structured_output_run,
     _seed_initial_messages,
@@ -260,6 +261,7 @@ def _start_trace(self, run: Run) -> None:
                     "output_messages": [],
                     "pending_assistant": None,
                     "seeded_initial": False,
+                    "system_instruction": None,
                     "model": None,
                     "provider": None,
                     "request_choice_count": None,
@@ -524,6 +526,10 @@ def _aggregate_into_parent(self, run: Run) -> None:  # pylint: disable=too-many-
             # (e.g. a routing decision), not a conversational turn, and folding
             # it into the transcript pollutes ``gen_ai.input.messages``.
             if run_type in ("llm", "chat_model") and not _is_structured_output_run(run):
+                if not content.get("system_instruction"):
+                    system_instruction = _extract_system_instruction(run.inputs)
+                    if system_instruction:
+                        content["system_instruction"] = system_instruction
                 # Seed system/user messages from the agent's top-level inputs
                 # on the first LLM call.
                 if not content.get("seeded_initial"):
@@ -634,6 +640,11 @@ def _finalize_agent_span(self, span: Span, run: Run) -> None:
 
         # Set aggregated input/output messages only when content capture is enabled
         if _should_capture_content_on_spans(self._enable_sensitive_data):
+            if system_instruction := content.get("system_instruction"):
+                span.set_attribute(
+                    GEN_AI_SYSTEM_INSTRUCTIONS_KEY,
+                    safe_json_dumps([asdict(p) for p in system_instruction]),
+                )
             if tool_defs := content.get("tool_definitions"):
                 span.set_attribute(GEN_AI_TOOL_DEFINITIONS_KEY, tool_defs)
             if msgs := content.get("input_messages"):
@@ -707,12 +718,16 @@ def _update_span(span: Span, run: Run, enable_sensitive_data: bool = False) -> L
                     GEN_AI_OUTPUT_MESSAGES_KEY,
                     safe_json_dumps([asdict(m) for m in invocation.output_messages]),
                 )
+            if invocation.system_instruction:
+                span.set_attribute(
+                    GEN_AI_SYSTEM_INSTRUCTIONS_KEY,
+                    safe_json_dumps([asdict(p) for p in invocation.system_instruction]),
+                )
         # Extras not covered by LLMInvocation
         span.set_attributes(
             dict(
                 flatten(
                     chain(
-                        prompts(run.inputs, enable_sensitive_data),
                         invocation_parameters(run),
                         function_calls(run.outputs, enable_sensitive_data),
                         metadata(run),

src/microsoft/opentelemetry/_genai/_langchain/_utils.py

@@ -263,21 +263,6 @@ def __delitem__(self, key: K) -> None:
 LANGCHAIN_THREAD_ID = "thread_id"
 
 
-@stop_on_exception
-def prompts(
-    inputs: Mapping[str, Any] | None,
-    enable_sensitive_data: bool = False,
-) -> Iterator[tuple[str, list[str]]]:
-    if not _should_capture_content_on_spans(enable_sensitive_data):
-        return
-    if not inputs:
-        return
-    if not isinstance(inputs, Mapping):
-        return
-    if p := inputs.get("prompts"):
-        yield GEN_AI_SYSTEM_INSTRUCTIONS_KEY, p
-
-
 @stop_on_exception
 def input_messages(
     inputs: Mapping[str, Any] | None,
@@ -1323,6 +1308,25 @@ def _extract_system_instruction(inputs: Mapping[str, Any] | None) -> list[Text]:
             return [Text(content=str(item)) for item in p if item]
         if isinstance(p, str):
             return [Text(content=p)]
+        return []
+    # Messages path: only reached when there are no prompts.
+    multiple_messages = inputs.get("messages")
+    if multiple_messages and isinstance(multiple_messages, Iterable):
+        if isinstance(multiple_messages, list) and multiple_messages:
+            first_item = multiple_messages[0]
+            first_messages = first_item if isinstance(first_item, list) else multiple_messages
+        else:
+            first_messages = next(iter(multiple_messages), None)  # type: ignore[arg-type]
+        if first_messages is not None:
+            if not isinstance(first_messages, list):
+                first_messages = [first_messages]
+            results: list[Text] = []
+            for msg in first_messages:
+                if _normalize_role(_langchain_role(msg)) == "system":
+                    content = _langchain_content(msg)
+                    if content:
+                        results.append(Text(content=content))
+            return results
     return []
 
 
@@ -1332,6 +1336,8 @@ def _extract_structured_input_messages(
     """Convert LangChain input messages to OTel ``InputMessage`` list."""
     if not inputs or not isinstance(inputs, Mapping):
         return []
+
+    route_system_out = not inputs.get("prompts")
     multiple_messages = inputs.get("messages")
     if multiple_messages and isinstance(multiple_messages, Iterable):
         first_messages = next(iter(multiple_messages), None)
@@ -1342,6 +1348,8 @@ def _extract_structured_input_messages(
             results: list[InputMessage] = []
             for msg in first_messages:
                 role = _normalize_role(_langchain_role(msg))
+                if route_system_out and role == "system":
+                    continue
                 parts: list[Any] = []
                 tool_responses = _langchain_tool_responses(msg)
                 if tool_responses: