fix(vertex_ai): return a truncated choice when a Gemini candidate has no content

A content-less candidate (a thinking model that spends its whole token budget
on reasoning returns finishReason MAX_TOKENS with no content) was skipped,
leaving choices empty so callers hit IndexError on the first choice. Emit an
empty-content choice, mapping finishReason through the same path the mainline
uses; a missing finishReason is reported as length, never a false stop.

litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py

@@ -2219,6 +2219,29 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
 
         for idx, candidate in enumerate(_candidates):
             if "content" not in candidate:
+                # Content-less candidate (e.g. thinking model out of budget):
+                # emit an empty choice with a mapped truncation reason, not an
+                # empty choices list. Non-streaming only.
+                if isinstance(model_response, ModelResponse):
+                    finish_reason = candidate.get("finishReason")
+                    mapped_finish_reason = (
+                        VertexGeminiConfig._check_finish_reason(None, finish_reason)
+                        if finish_reason is not None
+                        else "length"
+                    )
+                    empty_message: ChatCompletionResponseMessage = {
+                        "role": "assistant",
+                        "content": "",
+                    }
+                    model_response.choices.append(
+                        litellm.Choices(
+                            finish_reason=mapped_finish_reason,
+                            index=candidate.get("index", idx),
+                            message=empty_message,
+                            logprobs=None,
+                            enhancements=None,
+                        )
+                    )
                 continue
 
             # Extract metadata using helper function

tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py

@@ -5553,3 +5553,56 @@ def test_accumulated_json_skips_non_dict_leading_value():
 
     assert len(out) == 1
     assert out[0].choices[0].delta.content == "a"
+
+
+def test_vertex_ai_content_less_candidate_rescued_as_length():
+    """A candidate with no "content" (e.g. a thinking model that spends its whole
+    token budget on reasoning -> finishReason=MAX_TOKENS, no content) must still
+    produce a choice, not be silently dropped. Dropping it leaves
+    model_response.choices empty and the caller hits IndexError on choices[0].
+    The rescued choice must carry empty content and finishReason mapped through
+    the SAME path the mainline uses (MAX_TOKENS -> "length"), never the raw
+    provider value.
+    """
+    v = VertexGeminiConfig()
+    model_response = ModelResponse()
+    model_response.choices = []
+
+    v._process_candidates(
+        _candidates=[
+            {
+                "finishReason": "MAX_TOKENS",
+                "index": 0,
+            }
+        ],
+        model_response=model_response,
+        standard_optional_params={},
+    )
+
+    assert len(model_response.choices) == 1
+    choice = model_response.choices[0]
+    # mapped, not the raw "MAX_TOKENS"
+    assert choice.finish_reason == "length"
+    assert choice.message.content == ""
+    assert choice.message.role == "assistant"
+
+
+def test_vertex_ai_content_less_candidate_without_finish_reason_not_stop():
+    """A content-less candidate with NO finishReason at all must not silently
+    become finish_reason="stop" (a false claim the model finished normally).
+    Budget-exhausted-with-no-marker is honestly reported as "length".
+    """
+    v = VertexGeminiConfig()
+    model_response = ModelResponse()
+    model_response.choices = []
+
+    v._process_candidates(
+        _candidates=[{"index": 0}],
+        model_response=model_response,
+        standard_optional_params={},
+    )
+
+    assert len(model_response.choices) == 1
+    assert model_response.choices[0].finish_reason == "length"
+    assert model_response.choices[0].finish_reason != "stop"
+    assert model_response.choices[0].message.content == ""