fix: Fixes a bug that causes intermittent pydantic validation errors when uploading files

The root cause is an unsafe in-memory mutation. The `SaveFilesAsArtifactsPlugin` was saving a direct reference to the message part and then modifying the message list in-place. This created a race condition where downstream code could alter the original part *after* it had been saved as an artifact, leading to a corrupted state.
This CL saves a `copy.copy()` of the artifact, which create a snapshot of the data.
Also Changes the plugin to return a new `types.Content` object instead of modifying the original message in-place

PiperOrigin-RevId: 814308070

Author: GWeale

src/google/adk/models/google_llm.py

@@ -16,6 +16,7 @@
 from __future__ import annotations
 
 import contextlib
+import copy
 from functools import cached_property
 import logging
 import os
@@ -300,8 +301,13 @@ async def _preprocess_request(self, llm_request: LlmRequest) -> None:
           if not content.parts:
             continue
           for part in content.parts:
-            _remove_display_name_if_present(part.inline_data)
-            _remove_display_name_if_present(part.file_data)
+            # Create copies to avoid mutating the original objects
+            if part.inline_data:
+              part.inline_data = copy.copy(part.inline_data)
+              _remove_display_name_if_present(part.inline_data)
+            if part.file_data:
+              part.file_data = copy.copy(part.file_data)
+              _remove_display_name_if_present(part.file_data)
 
     # Initialize config if needed
     if llm_request.config and llm_request.config.tools:

src/google/adk/plugins/save_files_as_artifacts_plugin.py

@@ -14,6 +14,7 @@
 
 from __future__ import annotations
 
+import copy
 import logging
 from typing import Optional
 
@@ -29,14 +30,15 @@ class SaveFilesAsArtifactsPlugin(BasePlugin):
   """A plugin that saves files embedded in user messages as artifacts.
 
   This is useful to allow users to upload files in the chat experience and have
-  those files available to the agent.
-
-  We use Blob.display_name to determine
-  the file name. Artifacts with the same name will be overwritten. A placeholder
-  with the artifact name will be put in place of the embedded file in the user
-  message so the model knows where to find the file. You may want to add
-  load_artifacts tool to the agent, or load the artifacts in your own tool to
-  use the files.
+  those files available to the agent within the current session.
+
+  We use Blob.display_name to determine the file name. By default, artifacts are
+  session-scoped. For cross-session persistence, prefix the filename with
+  "user:".
+  Artifacts with the same name will be overwritten. A placeholder with the
+  artifact name will be put in place of the embedded file in the user message
+  so the model knows where to find the file. You may want to add load_artifacts
+  tool to the agent, or load the artifacts in your own tool to use the files.
   """
 
   def __init__(self, name: str = 'save_files_as_artifacts_plugin'):
@@ -62,10 +64,14 @@ async def on_user_message_callback(
       return user_message
 
     if not user_message.parts:
-      return user_message
+      return None
+
+    new_parts = []
+    modified = False
 
     for i, part in enumerate(user_message.parts):
       if part.inline_data is None:
+        new_parts.append(part)
         continue
 
       try:
@@ -77,23 +83,32 @@ async def on_user_message_callback(
               f'No display_name found, using generated filename: {file_name}'
           )
 
+        # Store original filename for display to user/ placeholder
+        display_name = file_name
+
+        # Create a copy to stop mutation of the saved artifact if the original part is modified
         await invocation_context.artifact_service.save_artifact(
             app_name=invocation_context.app_name,
             user_id=invocation_context.user_id,
             session_id=invocation_context.session.id,
             filename=file_name,
-            artifact=part,
+            artifact=copy.copy(part),
         )
 
-        # Replace the inline data with a placeholder text
-        user_message.parts[i] = types.Part(
-            text=f'[Uploaded Artifact: "{file_name}"]'
+        # Replace the inline data with a placeholder text (using the clean name)
+        new_parts.append(
+            types.Part(text=f'[Uploaded Artifact: "{display_name}"]')
         )
+        modified = True
         logger.info(f'Successfully saved artifact: {file_name}')
 
       except Exception as e:
         logger.error(f'Failed to save artifact for part {i}: {e}')
         # Keep the original part if saving fails
+        new_parts.append(part)
         continue
 
-    return user_message
+    if modified:
+      return types.Content(role=user_message.role, parts=new_parts)
+    else:
+      return None

src/google/adk/tools/load_artifacts_tool.py

@@ -15,6 +15,7 @@
 from __future__ import annotations
 
 import json
+import logging
 from typing import Any
 from typing import TYPE_CHECKING
 
@@ -27,6 +28,8 @@
   from ..models.llm_request import LlmRequest
   from .tool_context import ToolContext
 
+logger = logging.getLogger('google_adk.' + __name__)
+
 
 class LoadArtifactsTool(BaseTool):
   """A tool that loads the artifacts and adds them to the session."""
@@ -96,7 +99,18 @@ async def _append_artifacts_to_llm_request(
       if function_response and function_response.name == 'load_artifacts':
         artifact_names = function_response.response['artifact_names']
         for artifact_name in artifact_names:
+          # Try session-scoped first (default behavior)
           artifact = await tool_context.load_artifact(artifact_name)
+
+          # If not found and name doesn't already have user: prefix,
+          # try cross-session artifacts with user: prefix
+          if artifact is None and not artifact_name.startswith('user:'):
+            prefixed_name = f'user:{artifact_name}'
+            artifact = await tool_context.load_artifact(prefixed_name)
+
+          if artifact is None:
+            logger.warning('Artifact "%s" not found, skipping', artifact_name)
+            continue
           llm_request.contents.append(
               types.Content(
                   role='user',

tests/unittests/artifacts/test_artifact_service.py

@@ -277,3 +277,48 @@ async def test_list_versions(service_type):
   )
 
   assert response_versions == list(range(4))
+
+
+@pytest.mark.asyncio
+async def test_list_keys_preserves_user_prefix():
+  """Tests that list_artifact_keys preserves 'user:' prefix in returned names."""
+  artifact_service = InMemoryArtifactService()
+  artifact = types.Part.from_bytes(data=b"test_data", mime_type="text/plain")
+  app_name = "app0"
+  user_id = "user0"
+  session_id = "123"
+
+  # Save artifacts with "user:" prefix (cross-session artifacts)
+  await artifact_service.save_artifact(
+      app_name=app_name,
+      user_id=user_id,
+      session_id=session_id,
+      filename="user:document.pdf",
+      artifact=artifact,
+  )
+
+  await artifact_service.save_artifact(
+      app_name=app_name,
+      user_id=user_id,
+      session_id=session_id,
+      filename="user:image.png",
+      artifact=artifact,
+  )
+
+  # Save session-scoped artifact without prefix
+  await artifact_service.save_artifact(
+      app_name=app_name,
+      user_id=user_id,
+      session_id=session_id,
+      filename="session_file.txt",
+      artifact=artifact,
+  )
+
+  # List artifacts should return names with "user:" prefix for user-scoped artifacts
+  artifact_keys = await artifact_service.list_artifact_keys(
+      app_name=app_name, user_id=user_id, session_id=session_id
+  )
+
+  # Should contain prefixed names and session file
+  expected_keys = ["user:document.pdf", "user:image.png", "session_file.txt"]
+  assert sorted(artifact_keys) == sorted(expected_keys)

tests/unittests/plugins/test_save_files_as_artifacts.py

@@ -57,7 +57,7 @@ async def test_save_files_with_display_name(self):
         invocation_context=self.mock_context, user_message=user_message
     )
 
-    # Verify artifact was saved with correct filename
+    # Verify artifact was saved with correct filename (session-scoped by default)
     self.mock_context.artifact_service.save_artifact.assert_called_once_with(
         app_name="test_app",
         user_id="test_user",
@@ -66,7 +66,7 @@ async def test_save_files_with_display_name(self):
         artifact=original_part,
     )
 
-    # Verify message was modified with placeholder
+    # Verify message was modified with placeholder (clean name)
     assert result.parts[0].text == '[Uploaded Artifact: "test_document.pdf"]'
 
   @pytest.mark.asyncio
@@ -85,7 +85,7 @@ async def test_save_files_without_display_name(self):
         invocation_context=self.mock_context, user_message=user_message
     )
 
-    # Verify artifact was saved with generated filename
+    # Verify artifact was saved with generated filename (session-scoped by default)
     expected_filename = "artifact_test_invocation_123_0"
     self.mock_context.artifact_service.save_artifact.assert_called_once_with(
         app_name="test_app",
@@ -95,8 +95,12 @@ async def test_save_files_without_display_name(self):
         artifact=original_part,
     )
 
-    # Verify message was modified with generated filename
-    assert result.parts[0].text == f'[Uploaded Artifact: "{expected_filename}"]'
+    # Verify message was modified with generated filename (clean name)
+    generated_display_name = "artifact_test_invocation_123_0"
+    assert (
+        result.parts[0].text
+        == f'[Uploaded Artifact: "{generated_display_name}"]'
+    )
 
   @pytest.mark.asyncio
   async def test_multiple_files_in_message(self):
@@ -138,7 +142,7 @@ async def test_multiple_files_in_message(self):
     )
     assert second_call[1]["filename"] == "file2.jpg"
 
-    # Verify message parts were modified correctly
+    # Verify message parts were modified correctly (clean names)
     assert result.parts[0].text == '[Uploaded Artifact: "file1.txt"]'
     assert result.parts[1].text == "Some text between files"  # Unchanged
     assert result.parts[2].text == '[Uploaded Artifact: "file2.jpg"]'
@@ -174,9 +178,8 @@ async def test_no_parts_in_message(self):
         invocation_context=self.mock_context, user_message=user_message
     )
 
-    # Should return original message unchanged
-    assert result == user_message
-    assert result.parts == []
+    # Should return None to proceed with original message
+    assert result is None
 
     # Should not try to save any artifacts
     self.mock_context.artifact_service.save_artifact.assert_not_called()
@@ -193,10 +196,8 @@ async def test_parts_without_inline_data(self):
         invocation_context=self.mock_context, user_message=user_message
     )
 
-    # Should return original message unchanged
-    assert result == user_message
-    assert result.parts[0].text == "Hello world"
-    assert result.parts[1].text == "No files here"
+    # Should return None to proceed with original message
+    assert result is None
 
     # Should not try to save any artifacts
     self.mock_context.artifact_service.save_artifact.assert_not_called()
@@ -221,9 +222,8 @@ async def test_save_artifact_failure(self):
         invocation_context=self.mock_context, user_message=user_message
     )
 
-    # Should preserve original part when saving fails
-    assert result.parts[0] == original_part
-    assert result.parts[0].inline_data == inline_data
+    # Should return None when saving fails (no modifications made)
+    assert result is None
 
   @pytest.mark.asyncio
   async def test_mixed_success_and_failure(self):
@@ -264,7 +264,7 @@ def mock_save_artifact(*_args, **_kwargs):
         invocation_context=self.mock_context, user_message=user_message
     )
 
-    # First file should be replaced with placeholder
+    # First file should be replaced with placeholder (clean name)
     assert result.parts[0].text == '[Uploaded Artifact: "success.pdf"]'
 
     # Second file should remain unchanged due to failure
@@ -287,7 +287,7 @@ async def test_placeholder_text_format(self):
         invocation_context=self.mock_context, user_message=user_message
     )
 
-    # Verify exact format of placeholder text
+    # Verify exact format of placeholder text (clean name)
     expected_text = '[Uploaded Artifact: "test file with spaces.docx"]'
     assert result.parts[0].text == expected_text