From 0c28b47057102720353c04fbafd11f64760eb196 Mon Sep 17 00:00:00 2001
From: Chesars <cesarponce19544@gmail.com>
Date: Tue, 17 Mar 2026 17:47:01 -0300
Subject: [PATCH] fix(vertex): streaming finish_reason="stop" instead of
 "tool_calls" for gemini-3.1-flash-lite-preview

Models like gemini-3.1-flash-lite-preview send the final streaming chunk
with empty content (text:"") alongside finishReason:"STOP", instead of
omitting content entirely. The existing fix (PR #21577) only handled
chunks without content, so this case was missed.

Now, after processing candidates, if tool_calls were seen in earlier
chunks and a choice has finish_reason="stop", it is overridden to
"tool_calls" to match the OpenAI spec.

Fixes #22900
---
 .../vertex_and_google_ai_studio_gemini.py     | 10 +++
 ...emini_streaming_tool_call_finish_reason.py | 72 +++++++++++++++++++
 2 files changed, 82 insertions(+)

diff --git a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
index 3f1bccaccf..1054b311d0 100644
--- a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
+++ b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
@@ -3001,6 +3001,16 @@ class ModelResponseIterator:
                             )
                             model_response.choices.append(choice)
 
+                # Also handle the case where the final chunk has empty
+                # content (e.g. text:"") WITH finishReason. In this case
+                # _process_candidates DOES create a choice, but maps
+                # finishReason="STOP" to "stop" because the current chunk
+                # has no tool_calls. Override if we saw tool_calls earlier.
+                if self.has_seen_tool_calls:
+                    for choice in model_response.choices:
+                        if choice.finish_reason == "stop":
+                            choice.finish_reason = "tool_calls"
+
                 setattr(model_response, "vertex_ai_grounding_metadata", grounding_metadata)  # type: ignore
                 setattr(model_response, "vertex_ai_url_context_metadata", url_context_metadata)  # type: ignore
                 setattr(model_response, "vertex_ai_safety_ratings", safety_ratings)  # type: ignore
diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_gemini_streaming_tool_call_finish_reason.py b/tests/test_litellm/llms/vertex_ai/gemini/test_gemini_streaming_tool_call_finish_reason.py
index 3f8efd47fa..d4d76ab307 100644
--- a/tests/test_litellm/llms/vertex_ai/gemini/test_gemini_streaming_tool_call_finish_reason.py
+++ b/tests/test_litellm/llms/vertex_ai/gemini/test_gemini_streaming_tool_call_finish_reason.py
@@ -230,3 +230,75 @@ def test_streaming_content_filter_finish_reason_preserved():
     assert response is not None
     assert len(response.choices) == 1
     assert response.choices[0].finish_reason == "content_filter"
+
+
+def test_streaming_tool_call_finish_reason_with_empty_content_in_final_chunk():
+    """
+    When Gemini streams tool calls and the final chunk has BOTH empty content
+    (e.g. parts: [{text: ""}]) AND finishReason="STOP", the finish_reason
+    must still be "tool_calls".
+
+    This covers models like gemini-3.1-flash-lite-preview that send the
+    final chunk with content (empty text) instead of omitting it entirely.
+
+    Ref: https://github.com/BerriAI/litellm/issues/22900
+    """
+    logging_obj = _make_logging_obj()
+    iterator = ModelResponseIterator(
+        streaming_response=iter([]),
+        sync_stream=True,
+        logging_obj=logging_obj,
+    )
+
+    # Chunk 1: tool call with no finishReason
+    chunk_with_tool_calls = {
+        "candidates": [
+            {
+                "content": {
+                    "parts": [
+                        {
+                            "functionCall": {
+                                "name": "get_weather",
+                                "args": {"location": "San Francisco"},
+                            }
+                        }
+                    ],
+                    "role": "model",
+                },
+                "index": 0,
+            }
+        ],
+    }
+
+    # Chunk 2: finishReason="STOP" WITH empty content (text: "")
+    chunk_with_empty_content_and_finish = {
+        "candidates": [
+            {
+                "content": {
+                    "parts": [{"text": ""}],
+                    "role": "model",
+                },
+                "finishReason": "STOP",
+                "index": 0,
+            }
+        ],
+        "usageMetadata": {
+            "promptTokenCount": 50,
+            "candidatesTokenCount": 20,
+            "totalTokenCount": 70,
+        },
+    }
+
+    # Process chunk 1
+    response1 = iterator.chunk_parser(chunk_with_tool_calls)
+    assert response1 is not None
+    assert len(response1.choices) == 1
+    assert response1.choices[0].delta.tool_calls is not None
+    assert iterator.has_seen_tool_calls is True
+
+    # Process chunk 2 (final chunk with empty content)
+    response2 = iterator.chunk_parser(chunk_with_empty_content_and_finish)
+    assert response2 is not None
+    assert len(response2.choices) == 1
+    # Must be "tool_calls", NOT "stop"
+    assert response2.choices[0].finish_reason == "tool_calls"