Merge a8f6b07eb5 into d56ff281de

litellm/integrations/opentelemetry.py

@@ -1,7 +1,7 @@
 import os
 import threading
 from collections import OrderedDict
-from collections.abc import Callable, Iterable, Mapping
+from collections.abc import Callable, Iterable, Mapping, Sequence
 from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass, field
 from datetime import datetime
@@ -2791,7 +2791,9 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
         primitive_value: Final = self._cast_as_primitive_value_type(value)
         span.set_attribute(key, primitive_value)
 
-    def _transform_messages_to_otel_semantic_conventions(self, messages: list[dict] | str) -> list[dict]:
+    def _transform_messages_to_otel_semantic_conventions(
+        self, messages: Sequence[dict[str, object]] | str
+    ) -> list[dict[str, object]]:  # mutable-ok: match mixin base return
         """
         Transforms LiteLLM/OpenAI style messages into OTEL GenAI 1.38 compliant format.
         OTEL expects a 'parts' array instead of a single 'content' string.
@@ -2807,25 +2809,61 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
 
         transformed: Final = []
         for msg in messages:
+            # Normalize pydantic / LiteLLM wrapper objects (e.g. a ModelResponse
+            # Message) to something we can read with .get(); "tool_calls" in msg
+            # is False for those objects, which silently dropped tool calls.
+            normalized = self._to_dict(msg)
+            if normalized is None:
+                continue
+            msg = cast(dict[str, object], normalized)  # cast-ok: _to_dict guarantees a dict-like payload
+
             role = msg.get("role", "user")
             content = msg.get("content", "")
-            parts = []
+            parts: list[dict[str, object]] = []  # mutable-ok: local accumulator returned read-only
 
             if isinstance(content, str):
                 parts.append({"type": "text", "content": content})
             elif isinstance(content, list):
                 # Handle multi-modal content if necessary
-                for part in content:
+                for part in cast(list[object], content):  # cast-ok: content already narrowed to list
                     if isinstance(part, dict):
                         parts.append(part)
                     else:
                         parts.append({"type": "text", "content": str(part)})
 
-            transformed_msg = {"role": role, "parts": parts}
+            # Record tool calls as GenAI tool_call parts (OpenLLMetry-compatible)
+            # so id / name / arguments are not dropped from the trace.
+            raw_tool_calls = msg.get("tool_calls")
+            tool_calls: Sequence[object] = raw_tool_calls if isinstance(raw_tool_calls, list) else ()
+            for tc in tool_calls:
+                if isinstance(tc, dict):
+                    tc_dict = cast(dict[str, object], tc)  # cast-ok: tc narrowed to dict
+                    tc_id = tc_dict.get("id")
+                    fn = tc_dict.get("function")
+                    if isinstance(fn, dict):
+                        fn_dict = cast(dict[str, object], fn)  # cast-ok: fn narrowed to dict
+                        fn_name = fn_dict.get("name")
+                        fn_arguments = fn_dict.get("arguments")
+                    else:
+                        fn_name = None
+                        fn_arguments = None
+                else:
+                    # pydantic tool-call objects (e.g. ChatCompletionMessageToolCall):
+                    # read fields via literal getattr; arbitrary objects degrade to
+                    # None exactly like the old .get()/getattr pair.
+                    tc_id = cast(object, getattr(tc, "id", None))  # cast-ok: getattr returns object
+                    fn = cast(object, getattr(tc, "function", None))  # cast-ok: getattr returns object
+                    if fn is None:
+                        fn_name = None
+                        fn_arguments = None
+                    else:
+                        fn_name = cast(object, getattr(fn, "name", None))  # cast-ok: getattr returns object
+                        fn_arguments = cast(object, getattr(fn, "arguments", None))  # cast-ok: getattr returns object
+                parts.append({"type": "tool_call", "id": tc_id, "name": fn_name, "arguments": fn_arguments})
+
+            transformed_msg: dict[str, object] = {"role": role, "parts": parts}  # mutable-ok: key-assigned below
             if "id" in msg:
                 transformed_msg["id"] = msg["id"]
-            if "tool_calls" in msg:
-                transformed_msg["tool_calls"] = msg["tool_calls"]
             if "tool_call_id" in msg:
                 transformed_msg["tool_call_id"] = msg["tool_call_id"]
             transformed.append(transformed_msg)
@@ -2895,13 +2933,19 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
             item = self._to_dict(raw_item)
             if item is None:
                 continue
+            item = cast(dict[str, object], item)  # cast-ok: _to_dict guarantees a dict-like payload
             if item.get("type") == "message":
                 role = item.get("role", "assistant")
-                parts = []
-                for raw_content in item.get("content", []):
+                parts: list[dict[str, object]] = []  # mutable-ok: local accumulator returned read-only
+                raw_content_list = item.get("content", [])
+                content_items = cast(  # cast-ok: narrowed to list above
+                    Sequence[object], raw_content_list if isinstance(raw_content_list, list) else ()
+                )
+                for raw_content in content_items:
                     content = self._to_dict(raw_content)
                     if content is None:
                         continue
+                    content = cast(dict[str, object], content)  # cast-ok: _to_dict guarantees a dict-like payload
                     if content.get("type") == "output_text":
                         text = content.get("text", "")
                         if text:
@@ -2910,7 +2954,7 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
                     transformed.append({"role": role, "parts": parts})
             elif item.get("type") == "function_call":
                 # Surface tool calls from Responses API output
-                part: dict = {
+                part: dict[str, object] = {  # mutable-ok: assembled then key-assigned before append
                     "type": "tool_call",
                     "name": item.get("name", ""),
                     "arguments": item.get("arguments", ""),

litellm/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -369,6 +369,33 @@ class ChunkProcessor:
     def _get_role_from_chunks(chunks: Sequence["_BaseChunk"]) -> str:
         return ChunkProcessor._role_of_choice(next((c["choices"][0] for c in chunks if c.get("choices")), None))
 
+    @staticmethod
+    def _read_delta(choice: StreamingChoices | Mapping[str, object]) -> object:
+        """Return a streaming choice's delta payload.
+
+        ``StreamingChoices`` (an ``openai`` ``BaseModel`` subclass) declares
+        ``delta``; raw dict chunks carry the same data under a ``delta`` key.
+        """
+        return choice.get("delta") if isinstance(choice, Mapping) else choice.delta
+
+    @staticmethod
+    def _read_message(choice: StreamingChoices | Mapping[str, object]) -> object:
+        """Return a streaming choice's message payload.
+
+        Only dict chunks carry ``message``: ``StreamingChoices.__init__`` pops
+        it (Perplexity compat), so non-dict choices always read as ``None``.
+        """
+        return choice.get("message") if isinstance(choice, Mapping) else None
+
+    @staticmethod
+    def _read_tool_calls(payload: object) -> object:
+        """Return a delta/message payload's ``tool_calls``, dict or object form."""
+        if isinstance(payload, dict):
+            return cast(dict[str, object], payload).get("tool_calls")  # cast-ok: payload narrowed to dict
+        if isinstance(payload, Delta):
+            return payload.tool_calls
+        return None
+
     @staticmethod
     def _role_of_choice(choice: object) -> str:
         match choice:
@@ -414,15 +441,35 @@ class ChunkProcessor:
 
         role: Final = ChunkProcessor._get_role_from_chunks(chunks)
         finish_reason = "stop"
+        saw_tool_call = False  # rebind-ok: flipped to True inside the loop when a tool-call delta or message is seen
         for chunk in chunks:
             if "choices" in chunk and len(chunk["choices"]) > 0:
+                choice = chunk["choices"][0]
+                delta = ChunkProcessor._read_delta(choice)
+                if delta is not None:
+                    delta_tool_calls = ChunkProcessor._read_tool_calls(delta)
+                    if delta_tool_calls:
+                        saw_tool_call = True
+                message = ChunkProcessor._read_message(choice)
+                if message is not None:
+                    msg_tool_calls = ChunkProcessor._read_tool_calls(message)
+                    if msg_tool_calls:
+                        saw_tool_call = True
                 chunk_finish_reason = None
-                if hasattr(chunk["choices"][0], "finish_reason"):
-                    chunk_finish_reason = chunk["choices"][0].finish_reason
-                elif "finish_reason" in chunk["choices"][0]:
-                    chunk_finish_reason = chunk["choices"][0]["finish_reason"]
+                if hasattr(choice, "finish_reason"):
+                    # StreamingChoices is a dynamic OpenAIObject: the attribute may
+                    # or may not exist at runtime, so read via literal getattr.
+                    chunk_finish_reason = cast(str, getattr(choice, "finish_reason", None))  # cast-ok: dynamic attr
+                elif "finish_reason" in choice:
+                    chunk_finish_reason = cast(str, choice["finish_reason"])  # cast-ok: key checked
                 if chunk_finish_reason is not None:
                     finish_reason = chunk_finish_reason
+        if finish_reason == "stop" and saw_tool_call:
+            # A tool-calling stream whose chunks never surfaced an explicit
+            # finish_reason must aggregate to "tool_calls" (OpenAI spec),
+            # not the default "stop" - otherwise callbacks (e.g. OpenTelemetry)
+            # record a wrong finish reason for single-chunk streams. (litellm#45796)
+            finish_reason = "tool_calls"  # rebind-ok: aggregated finish reason for tool-calling streams (litellm#45796)
 
         # Initialize the response dictionary
         response = ModelResponse(

tests/unit/integrations/test_otel_tool_call_parts.py

@@ -0,0 +1,93 @@
+# -*- coding: utf-8 -*-
+"""Regression tests for #45796: OTEL tool-call parts + streamed finish reason.
+
+Covered behavior:
+- ``_transform_messages_to_otel_semantic_conventions`` emits GenAI
+  ``{"type": "tool_call", "id", "name", "arguments"}`` parts for assistant
+  messages carrying ``tool_calls`` — for plain dicts and for Pydantic
+  message objects (which used to be skipped because ``"tool_calls" in msg``
+  is False for non-dict objects).
+- Plain string ``content`` still emits a text part (no regression).
+- ``ChunkProcessor.build_base_response`` aggregates a stream whose chunks
+  carried tool-call deltas but no explicit ``finish_reason`` to
+  ``"tool_calls"`` instead of the default ``"stop"``.
+"""
+from pydantic import BaseModel
+
+from litellm.integrations.opentelemetry import OpenTelemetry
+from litellm.litellm_core_utils.streaming_chunk_builder_utils import ChunkProcessor
+
+
+class _FakeAssistantMessage(BaseModel):
+    role: str = "assistant"
+    content: str | None = None
+    tool_calls: list | None = None
+
+
+class TestOtelToolCallParts:
+    def test_dict_message_with_tool_calls_emits_genai_tool_call_parts(self):
+        msg = {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": "call_1",
+                    "type": "function",
+                    "function": {"name": "get_weather", "arguments": '{"city": "SF"}'},
+                }
+            ],
+        }
+        out = OpenTelemetry()._transform_messages_to_otel_semantic_conventions([msg])[0]
+        assert out["parts"] == [
+            {"type": "tool_call", "id": "call_1", "name": "get_weather", "arguments": '{"city": "SF"}'}
+        ]
+        assert "tool_calls" not in out
+
+    def test_pydantic_message_with_tool_calls_emits_genai_tool_call_parts(self):
+        msg = _FakeAssistantMessage(
+            tool_calls=[
+                {
+                    "id": "call_2",
+                    "type": "function",
+                    "function": {"name": "lookup", "arguments": "{}"},
+                }
+            ]
+        )
+        out = OpenTelemetry()._transform_messages_to_otel_semantic_conventions([msg])[0]
+        assert out["parts"] == [
+            {"type": "tool_call", "id": "call_2", "name": "lookup", "arguments": "{}"}
+        ]
+
+    def test_string_content_still_emits_text_part(self):
+        out = OpenTelemetry()._transform_messages_to_otel_semantic_conventions(
+            [{"role": "assistant", "content": "hello"}]
+        )[0]
+        assert out["parts"] == [{"type": "text", "content": "hello"}]
+
+    def test_streamed_tool_call_without_explicit_finish_reason_aggregates_to_tool_calls(self):
+        chunks = [
+            {
+                "object": "chat.completion.chunk",
+                "created": 0,
+                "model": "mock",
+                "choices": [{"index": 0, "delta": {"role": "assistant", "content": None}}],
+            },
+            {
+                "choices": [
+                    {
+                        "index": 0,
+                        "delta": {
+                            "tool_calls": [
+                                {
+                                    "id": "c1",
+                                    "type": "function",
+                                    "function": {"name": "f", "arguments": "{}"},
+                                }
+                            ]
+                        },
+                    }
+                ]
+            },
+        ]
+        resp = ChunkProcessor(chunks).build_base_response(chunks)
+        assert resp.choices[0].finish_reason == "tool_calls"