From aad4d335defd8713b3aa1f85275149c40d33af2c Mon Sep 17 00:00:00 2001
From: hcl <chenglunhu@gmail.com>
Date: Tue, 23 Jun 2026 21:36:11 +0800
Subject: [PATCH] fix(otel): don't crash set_attributes on non-dict (MCP)
 response_obj (#30660)

* fix(otel): don't crash set_attributes on non-dict (MCP) response_obj

MCP tool calls pass a Pydantic CallToolResult, but set_attributes accesses
response_obj via .get() throughout. The AttributeError was caught but skipped
writing the span output. Normalize a non-dict response_obj to a dict (model_dump)
so the span is fully emitted. Fixes #30651.

Co-Authored-By: Chenglun Hu <chenglunhu@gmail.com>

* test(otel): cover model_dump-raises and non-serializable fallbacks for #30651

* style: black-format the non-dict response test

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

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
---
 litellm/integrations/opentelemetry.py         | 13 +++++
 .../integrations/test_opentelemetry.py        | 50 +++++++++++++++++++
 2 files changed, 63 insertions(+)

diff --git a/litellm/integrations/opentelemetry.py b/litellm/integrations/opentelemetry.py
index de543fa042b..1aaf9d8286e 100644
--- a/litellm/integrations/opentelemetry.py
+++ b/litellm/integrations/opentelemetry.py
@@ -2117,6 +2117,19 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
             if standard_logging_payload is None:
                 raise ValueError("standard_logging_object not found in kwargs")
 
+            # MCP tool calls pass a Pydantic response (e.g. mcp.types.CallToolResult)
+            # rather than a dict. set_attributes accesses response_obj via .get()
+            # throughout, so a non-dict payload raises AttributeError and the span
+            # output is silently dropped. Normalize it to a dict here. See #30651.
+            if response_obj is not None and not hasattr(response_obj, "get"):
+                if hasattr(response_obj, "model_dump"):
+                    try:
+                        response_obj = response_obj.model_dump()
+                    except Exception:
+                        response_obj = None
+                else:
+                    response_obj = None
+
             # https://github.com/open-telemetry/semantic-conventions/blob/main/model/registry/gen-ai.yaml
             # Following Conventions here: https://github.com/open-telemetry/semantic-conventions/blob/main/docs/gen-ai/llm-spans.md
             #############################################
diff --git a/tests/test_litellm/integrations/test_opentelemetry.py b/tests/test_litellm/integrations/test_opentelemetry.py
index 7ffd09b931f..d6af11b6270 100644
--- a/tests/test_litellm/integrations/test_opentelemetry.py
+++ b/tests/test_litellm/integrations/test_opentelemetry.py
@@ -5852,3 +5852,53 @@ class TestOpenTelemetryMetricAttributeFiltering(unittest.TestCase):
                         exporter="console", attributes=attributes
                     )
                 )
+
+
+class PydanticLikeToolResult:
+    """Mimics mcp.types.CallToolResult: a Pydantic-style object with model_dump
+    and no dict .get (the shape that crashed set_attributes — see #30651)."""
+
+    def model_dump(self):
+        return {"content": [{"type": "text", "text": "ok"}], "isError": False}
+
+
+def test_set_attributes_does_not_crash_on_non_dict_response_obj():
+    from litellm.integrations.opentelemetry import OpenTelemetry
+
+    otel = OpenTelemetry()
+    otel.tracer = MagicMock()
+    span = MagicMock()
+    from collections import defaultdict
+
+    # defaultdict(dict) keeps the many standard_logging_payload[...] subscripts in
+    # set_attributes safe so execution reaches the response_obj.get(...) path,
+    # which is where the non-dict crash (#30651) happened.
+    kwargs = {
+        "standard_logging_object": defaultdict(dict, {"id": "log-1"}),
+        "litellm_params": {},
+        "optional_params": {},
+    }
+
+    def _set_attr_error_count(response_obj):
+        with patch("litellm.integrations.opentelemetry.verbose_logger") as vl:
+            otel.set_attributes(span, kwargs, response_obj)
+        return sum(1 for c in vl.exception.call_args_list if "set_attributes" in str(c))
+
+    # A dict response is the known-good baseline; a Pydantic (non-dict) response
+    # used to raise "'CallToolResult' object has no attribute 'get'" on top of it.
+    # After the fix the non-dict path must not add any set_attributes error.
+    baseline = _set_attr_error_count({"id": "resp-1"})
+    pydantic = _set_attr_error_count(PydanticLikeToolResult())
+    assert pydantic <= baseline, (pydantic, baseline)
+
+    # also cover the fallbacks: model_dump raising, and an object with neither
+    # .get nor model_dump — both normalize to None without adding errors.
+    class _RaisingDump:
+        def model_dump(self):
+            raise RuntimeError("boom")
+
+    class _Opaque:
+        pass
+
+    assert _set_attr_error_count(_RaisingDump()) <= baseline
+    assert _set_attr_error_count(_Opaque()) <= baseline