test: cover dict server tool usage chunks

litellm/tests/test_cost_calculator_attributeerror.py

@@ -4,8 +4,8 @@ import os
 # Add the project root to the path
 sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../..")))
 
-from litellm.cost_calculator import completion_cost
-from litellm.types.utils import Usage, ServerToolUse, ModelResponse
+from litellm.litellm_core_utils.streaming_chunk_builder_utils import ChunkProcessor
+from litellm.types.utils import Usage, ServerToolUse
 
 
 def test_usage_coercion_from_dict():
@@ -22,28 +22,45 @@ def test_usage_coercion_from_dict():
     assert usage.server_tool_use.web_search_requests == 5
 
 
-def test_completion_cost_with_dict_usage():
+def test_chunk_processor_coerces_dict_server_tool_use_from_stream_usage():
     """
-    Verify that completion_cost handles a response where server_tool_use is a dict.
-    This simulates the bug reported in #26153.
-    """
-    # Create a usage object and manually set server_tool_use to a dict to simulate the state after stream assembly
-    usage = Usage(prompt_tokens=10, completion_tokens=20)
-    usage.server_tool_use = {
-        "web_search_requests": 5
-    }  # Manually bypass the __init__ coercion for testing defensive checks
+    Verify streaming usage assembly handles server_tool_use from a dict.
 
-    response = ModelResponse(
-        id="test-id",
-        choices=[{"message": {"role": "assistant", "content": "hello"}}],
-        usage=usage,
+    Anthropic streaming chunks can arrive with usage as plain dicts. This
+    exercises the exact dict path that previously raised AttributeError when
+    server_tool_use was accessed with attribute syntax.
+    """
+    chunks = [
+        {
+            "usage": {
+                "prompt_tokens": 10,
+                "completion_tokens": 0,
+                "total_tokens": 10,
+                "server_tool_use": {"web_search_requests": 5},
+            }
+        },
+        {
+            "usage": {
+                "prompt_tokens": 0,
+                "completion_tokens": 20,
+                "total_tokens": 20,
+            }
+        },
+    ]
+
+    usage = ChunkProcessor(chunks=chunks).calculate_usage(
+        chunks=chunks,
+        model="claude-sonnet-4-5-20250929",
+        completion_output="hello",
     )
 
-    # This should not raise AttributeError
-    cost = completion_cost(completion_response=response, model="gpt-3.5-turbo")
-    assert cost is not None
+    assert usage.prompt_tokens == 10
+    assert usage.completion_tokens == 20
+    assert usage.total_tokens == 30
+    assert isinstance(usage.server_tool_use, ServerToolUse)
+    assert usage.server_tool_use.web_search_requests == 5
 
 
 if __name__ == "__main__":
     test_usage_coercion_from_dict()
-    test_completion_cost_with_dict_usage()
+    test_chunk_processor_coerces_dict_server_tool_use_from_stream_usage()