fix(mcp+anthropic): vanilla mcp crashed due to argument manipulation (#2881)

Author: nirga

packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/__init__.py

@@ -15,6 +15,7 @@
     build_from_streaming_response,
 )
 from opentelemetry.instrumentation.anthropic.utils import (
+    JSONEncoder,
     acount_prompt_tokens_from_request,
     dont_throw,
     error_metrics_attributes,
@@ -28,7 +29,9 @@
 from opentelemetry.instrumentation.instrumentor import BaseInstrumentor
 from opentelemetry.instrumentation.utils import _SUPPRESS_INSTRUMENTATION_KEY, unwrap
 from opentelemetry.metrics import Counter, Histogram, Meter, get_meter
-from opentelemetry.semconv._incubating.attributes.gen_ai_attributes import GEN_AI_RESPONSE_ID
+from opentelemetry.semconv._incubating.attributes.gen_ai_attributes import (
+    GEN_AI_RESPONSE_ID,
+)
 from opentelemetry.semconv_ai import (
     SUPPRESS_LANGUAGE_MODEL_INSTRUMENTATION_KEY,
     LLMRequestTypeValues,
@@ -145,7 +148,7 @@ async def _dump_content(message_index, content, span):
             for j, item in enumerate(content)
         ]
 
-        return json.dumps(content)
+        return json.dumps(content, cls=JSONEncoder)
 
 
 @dont_throw
@@ -207,10 +210,14 @@ async def _aset_input_attributes(span, kwargs):
             for i, tool in enumerate(kwargs.get("tools")):
                 prefix = f"{SpanAttributes.LLM_REQUEST_FUNCTIONS}.{i}"
                 set_span_attribute(span, f"{prefix}.name", tool.get("name"))
-                set_span_attribute(span, f"{prefix}.description", tool.get("description"))
+                set_span_attribute(
+                    span, f"{prefix}.description", tool.get("description")
+                )
                 input_schema = tool.get("input_schema")
                 if input_schema is not None:
-                    set_span_attribute(span, f"{prefix}.input_schema", json.dumps(input_schema))
+                    set_span_attribute(
+                        span, f"{prefix}.parameters", json.dumps(input_schema)
+                    )
 
 
 def _set_span_completions(span, response):
@@ -315,7 +322,9 @@ async def _aset_token_usage(
         completion_tokens = 0
         if hasattr(anthropic, "count_tokens"):
             if response.get("completion"):
-                completion_tokens = await anthropic.count_tokens(response.get("completion"))
+                completion_tokens = await anthropic.count_tokens(
+                    response.get("completion")
+                )
             elif response.get("content"):
                 completion_tokens = await anthropic.count_tokens(
                     response.get("content")[0].text
@@ -357,7 +366,9 @@ async def _aset_token_usage(
         span, SpanAttributes.LLM_USAGE_CACHE_READ_INPUT_TOKENS, cache_read_tokens
     )
     set_span_attribute(
-        span, SpanAttributes.LLM_USAGE_CACHE_CREATION_INPUT_TOKENS, cache_creation_tokens
+        span,
+        SpanAttributes.LLM_USAGE_CACHE_CREATION_INPUT_TOKENS,
+        cache_creation_tokens,
     )
 
 
@@ -402,7 +413,9 @@ def _set_token_usage(
             if response.get("completion"):
                 completion_tokens = anthropic.count_tokens(response.get("completion"))
             elif response.get("content"):
-                completion_tokens = anthropic.count_tokens(response.get("content")[0].text)
+                completion_tokens = anthropic.count_tokens(
+                    response.get("content")[0].text
+                )
 
     if token_histogram and type(completion_tokens) is int and completion_tokens >= 0:
         token_histogram.record(
@@ -440,7 +453,9 @@ def _set_token_usage(
         span, SpanAttributes.LLM_USAGE_CACHE_READ_INPUT_TOKENS, cache_read_tokens
     )
     set_span_attribute(
-        span, SpanAttributes.LLM_USAGE_CACHE_CREATION_INPUT_TOKENS, cache_creation_tokens
+        span,
+        SpanAttributes.LLM_USAGE_CACHE_CREATION_INPUT_TOKENS,
+        cache_creation_tokens,
     )
 
 

packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/utils.py

@@ -1,4 +1,5 @@
 import asyncio
+import json
 import os
 import logging
 import threading
@@ -133,3 +134,19 @@ def run_async(method):
         thread.join()
     else:
         asyncio.run(method)
+
+
+class JSONEncoder(json.JSONEncoder):
+    def default(self, o):
+        if hasattr(o, "to_json"):
+            return o.to_json()
+
+        if hasattr(o, "model_dump_json"):
+            return o.model_dump_json()
+
+        try:
+            return str(o)
+        except Exception:
+            logger = logging.getLogger(__name__)
+            logger.debug("Failed to serialize object of type: %s", type(o).__name__)
+            return ""

packages/opentelemetry-instrumentation-anthropic/tests/test_messages.py

@@ -53,7 +53,10 @@ def test_anthropic_message_create(exporter, reader):
         + anthropic_span.attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS]
         == anthropic_span.attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS]
     )
-    assert anthropic_span.attributes.get("gen_ai.response.id") == "msg_01TPXhkPo8jy6yQMrMhjpiAE"
+    assert (
+        anthropic_span.attributes.get("gen_ai.response.id")
+        == "msg_01TPXhkPo8jy6yQMrMhjpiAE"
+    )
 
     metrics_data = reader.get_metrics_data()
     resource_metrics = metrics_data.resource_metrics
@@ -120,7 +123,10 @@ def test_anthropic_multi_modal(exporter):
         + anthropic_span.attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS]
         == anthropic_span.attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS]
     )
-    assert anthropic_span.attributes.get("gen_ai.response.id") == "msg_01B37ySLPzYj8KY6uZmiPoxd"
+    assert (
+        anthropic_span.attributes.get("gen_ai.response.id")
+        == "msg_01B37ySLPzYj8KY6uZmiPoxd"
+    )
 
 
 @pytest.mark.vcr
@@ -184,7 +190,10 @@ async def test_anthropic_async_multi_modal(exporter):
         + anthropic_span.attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS]
         == anthropic_span.attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS]
     )
-    assert anthropic_span.attributes.get("gen_ai.response.id") == "msg_01DWnmUo9hWk4Fk7V7Ddfa2w"
+    assert (
+        anthropic_span.attributes.get("gen_ai.response.id")
+        == "msg_01DWnmUo9hWk4Fk7V7Ddfa2w"
+    )
 
 
 @pytest.mark.vcr
@@ -237,7 +246,10 @@ def test_anthropic_message_streaming(exporter, reader):
         + anthropic_span.attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS]
         == anthropic_span.attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS]
     )
-    assert anthropic_span.attributes.get("gen_ai.response.id") == "msg_01MXWxhWoPSgrYhjTuMDM6F1"
+    assert (
+        anthropic_span.attributes.get("gen_ai.response.id")
+        == "msg_01MXWxhWoPSgrYhjTuMDM6F1"
+    )
 
     metrics_data = reader.get_metrics_data()
     resource_metrics = metrics_data.resource_metrics
@@ -290,7 +302,10 @@ async def test_async_anthropic_message_create(exporter, reader):
         + anthropic_span.attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS]
         == anthropic_span.attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS]
     )
-    assert anthropic_span.attributes.get("gen_ai.response.id") == "msg_01UFDDjsFn5BPQnfNwmsMnAY"
+    assert (
+        anthropic_span.attributes.get("gen_ai.response.id")
+        == "msg_01UFDDjsFn5BPQnfNwmsMnAY"
+    )
 
     metrics_data = reader.get_metrics_data()
     resource_metrics = metrics_data.resource_metrics
@@ -349,7 +364,10 @@ async def test_async_anthropic_message_streaming(exporter, reader):
         + anthropic_span.attributes[SpanAttributes.LLM_USAGE_PROMPT_TOKENS]
         == anthropic_span.attributes[SpanAttributes.LLM_USAGE_TOTAL_TOKENS]
     )
-    assert anthropic_span.attributes.get("gen_ai.response.id") == "msg_016o6A7zDmgjucf5mWv1rrPD"
+    assert (
+        anthropic_span.attributes.get("gen_ai.response.id")
+        == "msg_016o6A7zDmgjucf5mWv1rrPD"
+    )
 
     metrics_data = reader.get_metrics_data()
     resource_metrics = metrics_data.resource_metrics
@@ -371,16 +389,16 @@ def test_anthropic_tools(exporter, reader):
                     "properties": {
                         "location": {
                             "type": "string",
-                            "description": "The city and state, e.g. San Francisco, CA"
+                            "description": "The city and state, e.g. San Francisco, CA",
                         },
                         "unit": {
                             "type": "string",
                             "enum": ["celsius", "fahrenheit"],
-                            "description": "The unit of temperature, either 'celsius' or 'fahrenheit'"
-                        }
+                            "description": "The unit of temperature, either 'celsius' or 'fahrenheit'",
+                        },
                     },
-                    "required": ["location"]
-                }
+                    "required": ["location"],
+                },
             },
             {
                 "name": "get_time",
@@ -390,19 +408,19 @@ def test_anthropic_tools(exporter, reader):
                     "properties": {
                         "timezone": {
                             "type": "string",
-                            "description": "The IANA time zone name, e.g. America/Los_Angeles"
+                            "description": "The IANA time zone name, e.g. America/Los_Angeles",
                         }
                     },
-                    "required": ["timezone"]
-                }
-            }
+                    "required": ["timezone"],
+                },
+            },
         ],
         messages=[
             {
                 "role": "user",
-                "content": "What is the weather like right now in New York? Also what time is it there now?"
+                "content": "What is the weather like right now in New York? Also what time is it there now?",
             }
-        ]
+        ],
     )
     try:
         client.messages.create(
@@ -428,69 +446,92 @@ def test_anthropic_tools(exporter, reader):
 
     # verify request and inputs
     assert (
-        anthropic_span.attributes["gen_ai.prompt.0.content"] ==
-        "What is the weather like right now in New York? Also what time is it there now?"
+        anthropic_span.attributes["gen_ai.prompt.0.content"]
+        == "What is the weather like right now in New York? Also what time is it there now?"
     )
     assert anthropic_span.attributes["gen_ai.prompt.0.role"] == "user"
-    assert (
-        anthropic_span.attributes["llm.request.functions.0.name"] == "get_weather"
-    )
+    assert anthropic_span.attributes["llm.request.functions.0.name"] == "get_weather"
     assert (
         anthropic_span.attributes["llm.request.functions.0.description"]
         == "Get the current weather in a given location"
     )
-    assert (
-        anthropic_span.attributes["llm.request.functions.0.input_schema"]
-        == json.dumps({
+    assert anthropic_span.attributes[
+        "llm.request.functions.0.parameters"
+    ] == json.dumps(
+        {
             "type": "object",
             "properties": {
                 "location": {
                     "type": "string",
-                    "description": "The city and state, e.g. San Francisco, CA"
+                    "description": "The city and state, e.g. San Francisco, CA",
                 },
                 "unit": {
                     "type": "string",
                     "enum": ["celsius", "fahrenheit"],
-                    "description": "The unit of temperature, either 'celsius' or 'fahrenheit'"
-                }
+                    "description": "The unit of temperature, either 'celsius' or 'fahrenheit'",
+                },
             },
-            "required": ["location"]
-        })
+            "required": ["location"],
+        }
     )
     assert anthropic_span.attributes["llm.request.functions.1.name"] == "get_time"
     assert (
         anthropic_span.attributes["llm.request.functions.1.description"]
         == "Get the current time in a given time zone"
     )
-    assert (
-        anthropic_span.attributes["llm.request.functions.1.input_schema"]
-        == json.dumps({
+    assert anthropic_span.attributes[
+        "llm.request.functions.1.parameters"
+    ] == json.dumps(
+        {
             "type": "object",
             "properties": {
                 "timezone": {
                     "type": "string",
-                    "description": "The IANA time zone name, e.g. America/Los_Angeles"
+                    "description": "The IANA time zone name, e.g. America/Los_Angeles",
                 }
             },
-            "required": ["timezone"]
-        })
+            "required": ["timezone"],
+        }
     )
 
     # verify response and output
-    assert anthropic_span.attributes["gen_ai.completion.0.finish_reason"] == response.stop_reason
-    assert anthropic_span.attributes["gen_ai.completion.0.content"] == response.content[0].text
+    assert (
+        anthropic_span.attributes["gen_ai.completion.0.finish_reason"]
+        == response.stop_reason
+    )
+    assert (
+        anthropic_span.attributes["gen_ai.completion.0.content"]
+        == response.content[0].text
+    )
     assert anthropic_span.attributes["gen_ai.completion.0.role"] == "assistant"
 
-    assert (anthropic_span.attributes["gen_ai.completion.0.tool_calls.0.id"]) == response.content[1].id
-    assert (anthropic_span.attributes["gen_ai.completion.0.tool_calls.0.name"]) == response.content[1].name
+    assert (
+        anthropic_span.attributes["gen_ai.completion.0.tool_calls.0.id"]
+    ) == response.content[1].id
+    assert (
+        anthropic_span.attributes["gen_ai.completion.0.tool_calls.0.name"]
+    ) == response.content[1].name
     response_input = json.dumps(response.content[1].input)
-    assert (anthropic_span.attributes["gen_ai.completion.0.tool_calls.0.arguments"] == response_input)
+    assert (
+        anthropic_span.attributes["gen_ai.completion.0.tool_calls.0.arguments"]
+        == response_input
+    )
 
-    assert (anthropic_span.attributes["gen_ai.completion.0.tool_calls.1.id"]) == response.content[2].id
-    assert (anthropic_span.attributes["gen_ai.completion.0.tool_calls.1.name"]) == response.content[2].name
+    assert (
+        anthropic_span.attributes["gen_ai.completion.0.tool_calls.1.id"]
+    ) == response.content[2].id
+    assert (
+        anthropic_span.attributes["gen_ai.completion.0.tool_calls.1.name"]
+    ) == response.content[2].name
     response_input = json.dumps(response.content[2].input)
-    assert (anthropic_span.attributes["gen_ai.completion.0.tool_calls.1.arguments"] == response_input)
-    assert anthropic_span.attributes.get("gen_ai.response.id") == "msg_01RBkXFe9TmDNNWThMz2HmGt"
+    assert (
+        anthropic_span.attributes["gen_ai.completion.0.tool_calls.1.arguments"]
+        == response_input
+    )
+    assert (
+        anthropic_span.attributes.get("gen_ai.response.id")
+        == "msg_01RBkXFe9TmDNNWThMz2HmGt"
+    )
 
     # verify metrics
     metrics_data = reader.get_metrics_data()