fix(anthropic): capture server-side tool usage metrics

py/src/braintrust/integrations/anthropic/_utils.py

@@ -27,11 +27,6 @@ def __getattr__(self, name: str) -> Any:
     ("ephemeral_1h_input_tokens", "prompt_cache_creation_ephemeral_1h_tokens"),
 )
 
-_ANTHROPIC_SERVER_TOOL_USE_METRIC_FIELDS = (
-    ("web_search_requests", "server_tool_use_web_search_requests"),
-    ("web_fetch_requests", "server_tool_use_web_fetch_requests"),
-)
-
 _ANTHROPIC_USAGE_METADATA_FIELDS = frozenset(
     {
         "service_tier",
@@ -95,8 +90,8 @@ def extract_anthropic_usage(usage: Any) -> tuple[dict[str, float], dict[str, Any
 
     server_tool_use = _try_to_dict(usage.get("server_tool_use"))
     if server_tool_use is not None:
-        for source_name, metric_name in _ANTHROPIC_SERVER_TOOL_USE_METRIC_FIELDS:
-            _set_numeric_metric(metrics, metric_name, server_tool_use.get(source_name))
+        for source_name, value in server_tool_use.items():
+            _set_numeric_metric(metrics, f"server_tool_use_{source_name}", value)
 
     if "prompt_cache_creation_tokens" not in metrics and cache_creation_breakdown:
         metrics["prompt_cache_creation_tokens"] = sum(cache_creation_breakdown)

py/src/braintrust/integrations/anthropic/test_anthropic.py

@@ -53,6 +53,10 @@ def test_log_message_to_span_includes_stop_reason_and_stop_sequence():
             "output_tokens": 7,
             "cache_read_input_tokens": 0,
             "cache_creation_input_tokens": 0,
+            "server_tool_use": {
+                "web_search_requests": 2,
+                "web_fetch_requests": 1,
+            },
         },
     )
 
@@ -71,13 +75,43 @@ def test_log_message_to_span_includes_stop_reason_and_stop_sequence():
             "completion_tokens": 7.0,
             "prompt_cached_tokens": 0.0,
             "prompt_cache_creation_tokens": 0.0,
+            "server_tool_use_web_search_requests": 2.0,
+            "server_tool_use_web_fetch_requests": 1.0,
             "tokens": 18.0,
             "time_to_first_token": 0.123,
         },
         metadata={},
     )
 
 
+def test_extract_anthropic_usage_includes_server_tool_use_metrics_from_objects():
+    usage = SimpleNamespace(
+        input_tokens=11,
+        output_tokens=7,
+        cache_read_input_tokens=3,
+        cache_creation_input_tokens=2,
+        server_tool_use=SimpleNamespace(
+            web_search_requests=2,
+            web_fetch_requests=1,
+            code_execution_requests=4,
+        ),
+    )
+
+    metrics, metadata = extract_anthropic_usage(usage)
+
+    assert metrics == {
+        "prompt_tokens": 16.0,
+        "completion_tokens": 7.0,
+        "prompt_cached_tokens": 3.0,
+        "prompt_cache_creation_tokens": 2.0,
+        "server_tool_use_web_search_requests": 2.0,
+        "server_tool_use_web_fetch_requests": 1.0,
+        "server_tool_use_code_execution_requests": 4.0,
+        "tokens": 23.0,
+    }
+    assert metadata == {}
+
+
 @pytest.mark.vcr
 def test_anthropic_messages_create_stream_true(memory_logger):
     assert not memory_logger.pop()