From 7b6149ea3fbd09abf660a0275294590d49162239 Mon Sep 17 00:00:00 2001
From: BillionClaw <267901332+BillionClaw@users.noreply.github.com>
Date: Tue, 24 Mar 2026 10:31:09 +0800
Subject: [PATCH] fix(gemini): handle empty response when finish_reason is STOP

gemini-2.5-flash-lite can return finish_reason=STOP with empty
content in long-running agentic tasks. Fixed response parsing to
return empty string instead of None. Fixes #24442
---
 .../vertex_and_google_ai_studio_gemini.py     |  6 +++
 ...test_vertex_and_google_ai_studio_gemini.py | 44 +++++++++++++++++++
 2 files changed, 50 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 36f51c5b2f5..c002f6eab5d 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
@@ -2198,6 +2198,12 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
                     cumulative_tool_call_idx=cumulative_tool_call_index,
                     is_function_call=is_function_call(standard_optional_params),
                 )
+            else:
+                # Handle case where finish_reason=STOP but no parts/content
+                # e.g. gemini-2.5-flash-lite in long-running agentic tasks
+                finish_reason = candidate.get("finishReason")
+                if finish_reason == "STOP":
+                    chat_completion_message["content"] = ""
 
             if "logprobsResult" in candidate:
                 chat_completion_logprobs = VertexGeminiConfig._transform_logprobs(
diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py
index 3102a695961..d15618d7edc 100644
--- a/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py
+++ b/tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py
@@ -3825,3 +3825,47 @@ def test_sync_streaming_uses_custom_client():
     # Verify that gemini_client is in the partial's keywords
     assert "gemini_client" in partial_make_sync_call.keywords
     assert partial_make_sync_call.keywords["gemini_client"] is mock_client
+
+
+def test_vertex_ai_empty_response_with_stop_finish_reason():
+    """
+    Test that empty responses with finish_reason=STOP return empty string content.
+    
+    Regression test for https://github.com/BerriAI/litellm/issues/24442
+    gemini-2.5-flash-lite can return finish_reason=STOP with no parts/text content
+    in long-running agentic tasks. Response looks like:
+    {"candidates":[{"content":{"role":"model"},"finishReason":"STOP","index":0}]}
+    """
+    completion_response = {
+        "candidates": [
+            {
+                "content": {"role": "model"},
+                "finishReason": "STOP",
+                "index": 0,
+            }
+        ],
+        "usageMetadata": {
+            "promptTokenCount": 50,
+            "candidatesTokenCount": 0,
+            "totalTokenCount": 50,
+        },
+    }
+
+    raw_response = MagicMock()
+    raw_response.json.return_value = completion_response
+
+    result = VertexGeminiConfig().transform_response(
+        model="gemini-2.5-flash-lite",
+        raw_response=raw_response,
+        model_response=ModelResponse(),
+        logging_obj=MagicMock(),
+        request_data={},
+        messages=[],
+        optional_params={},
+        litellm_params={},
+        encoding=None,
+    )
+
+    # Should return empty string content, not None
+    assert result.choices[0].message.content == ""
+    assert result.choices[0].finish_reason == "stop"