fix(evaluation): skip invocations without user events to prevent ValidationError

EvaluationGenerator.convert_events_to_eval_invocations previously
initialized user_content as Content(parts=[]) for invocations without
user-authored events. While this is a valid Content object, creating
an Invocation with no meaningful user content is semantically incorrect
and can cause issues downstream in the evaluation pipeline.

This fix initializes user_content = None and skips the invocation if
no user event is found, emitting a debug log message for diagnostics.

Fixes #3760

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: 2 people

src/google/adk/evaluation/evaluation_generator.py

@@ -627,7 +627,7 @@ def convert_events_to_eval_invocations(
     for invocation_id, events in events_by_invocation_id.items():
       final_response = None
       final_event = None
-      user_content = Content(parts=[])
+      user_content = None
       invocation_timestamp = 0
       app_details = None
       if (
@@ -663,6 +663,18 @@ def convert_events_to_eval_invocations(
               events_to_add.append(event)
               break
 
+      if user_content is None:
+        # Skip invocations that have no user-authored event. Such invocations
+        # arise from internal/system-driven turns (e.g., background agent tasks)
+        # and are not meaningful for evaluation purposes. Including them would
+        # also cause a Pydantic ValidationError because Invocation.user_content
+        # requires a Content object.
+        logger.debug(
+            "Skipping invocation %s: no user-authored event found.",
+            invocation_id,
+        )
+        continue
+
       invocation_events = [
           InvocationEvent(author=e.author, content=e.content)
           for e in events_to_add