fix(ollama): end streamed tool calls with finish_reason tool_calls

ollama_chat streams tool calls in a chunk with done false and then sends a
separate done chunk whose message is empty and whose done_reason is stop. The
override that turns the finish reason into tool_calls only looked at the done
chunk's own tool_calls field, which is None by then, so the stream ended in
stop.

Clients that gate tool execution on finish_reason == tool_calls therefore
threw away the tool_calls delta they had just accumulated and treated the turn
as ordinary text. Non-streaming was fine; it reads message.tool_calls off the
full response.

Track whether any chunk in the stream carried tool calls and use that on the
done chunk. Streams that never produce a tool call keep whatever done_reason
Ollama sent, so stop and length are unchanged.

litellm/llms/ollama/chat/transformation.py

@@ -425,6 +425,7 @@ class OllamaChatConfig(BaseConfig):
 class OllamaChatCompletionResponseIterator(BaseModelResponseIterator):
     started_reasoning_content: bool = False
     finished_reasoning_content: bool = False
+    saw_tool_calls: bool = False
 
     def _is_function_call_complete(self, function_args: str | dict) -> bool:
         if isinstance(function_args, dict):
@@ -470,6 +471,7 @@ class OllamaChatCompletionResponseIterator(BaseModelResponseIterator):
             # process tool calls - if complete function arg - add id to tool call
             tool_calls = chunk["message"].get("tool_calls")
             if tool_calls is not None:
+                self.saw_tool_calls = True
                 for tool_call in tool_calls:
                     function_args = tool_call.get("function").get("arguments")
                     if function_args is not None and len(function_args) > 0:
@@ -510,9 +512,13 @@ class OllamaChatCompletionResponseIterator(BaseModelResponseIterator):
 
             if chunk["done"] is True:
                 finish_reason = chunk.get("done_reason") or "stop"
-                # Override finish_reason when tool_calls are present
+                # Override finish_reason when tool_calls are present anywhere in
+                # the stream. Ollama often sends them in a chunk with done false
+                # and then a separate done chunk with an empty message, so keying
+                # off this chunk's tool_calls alone leaves the stream ending in
+                # "stop".
                 # Fixes: https://github.com/BerriAI/litellm/issues/18922
-                if tool_calls is not None:
+                if self.saw_tool_calls:
                     finish_reason = "tool_calls"
                 choices = [
                     StreamingChoices(

tests/test_litellm/llms/ollama/test_ollama_chat_transformation.py

@@ -597,6 +597,52 @@ class TestOllamaFinishReasonLength:
             result.choices[0].finish_reason == "length"
         ), f"Expected 'length' when done_reason='length', got '{result.choices[0].finish_reason}'"
 
+    def test_finish_reason_tool_calls_when_split_across_chunks(self):
+        """Tool calls in an earlier chunk must still end the stream with 'tool_calls'.
+
+        Ollama commonly emits tool_calls in a chunk with done false and then a
+        separate done chunk whose message is empty and whose done_reason is
+        'stop'. Keying the override off the done chunk's own tool_calls left the
+        stream finishing as 'stop', so spec-strict OpenAI clients dropped the
+        accumulated tool_calls delta and never ran the tool.
+        """
+        iterator = OllamaChatCompletionResponseIterator(
+            streaming_response=iter([]),
+            sync_stream=True,
+        )
+
+        tool_call_chunk = {
+            "model": "glm-4.7-flash",
+            "message": {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {
+                        "function": {
+                            "name": "list_files",
+                            "arguments": {"path": "."},
+                        }
+                    }
+                ],
+            },
+            "done": False,
+        }
+        done_chunk = {
+            "model": "glm-4.7-flash",
+            "message": {"role": "assistant", "content": ""},
+            "done": True,
+            "done_reason": "stop",
+        }
+
+        tool_call_result = iterator.chunk_parser(tool_call_chunk)
+        assert tool_call_result.choices[0].delta.tool_calls is not None
+        assert tool_call_result.choices[0].finish_reason is None
+
+        done_result = iterator.chunk_parser(done_chunk)
+        assert (
+            done_result.choices[0].finish_reason == "tool_calls"
+        ), f"Expected 'tool_calls' after a tool call earlier in the stream, got '{done_result.choices[0].finish_reason}'"
+
     def test_finish_reason_stop_streaming(self):
         """Streaming: done_reason='stop' in final chunk must produce finish_reason='stop'."""
         iterator = OllamaChatCompletionResponseIterator(