Merge f2bba9c1ba into c2c2a623c0

litellm/llms/vertex_ai/context_caching/transformation.py

@@ -158,7 +158,7 @@ def cached_messages_end_on_supported_turn(cached_messages: Sequence[AllMessageVa
     The cachedContents API rejects contents ending on a model turn, which is how it
     classifies both assistant messages and tool results, with HTTP 400
     "Requests ending with a model turn are not supported". System messages are
-    extracted into system_instruction before contents are built, so the terminal
+    extracted into systemInstruction before contents are built, so the terminal
     turn is the last non-system message.
     """
     non_system_messages: Final = tuple(message for message in cached_messages if message.get("role") != "system")
@@ -206,6 +206,6 @@ def transform_openai_messages_to_gemini_context_caching(
         data["ttl"] = ttl
 
     if transformed_system_messages is not None:
-        data["system_instruction"] = transformed_system_messages
+        data["systemInstruction"] = transformed_system_messages
 
     return data

litellm/llms/vertex_ai/gemini/transformation.py

@@ -1218,12 +1218,12 @@ def _transform_request_body(
                     generation_config["mediaResolution"] = media_resolution_value["level"]
 
         data: Final = RequestBody(contents=content)
-        # Vertex rejects system_instruction/tools/toolConfig alongside cachedContent.
+        # Vertex rejects systemInstruction/tools/toolConfig alongside cachedContent.
         # Treat dropping these fields as a request mutation guarded by modify_params.
         can_send_cache_incompatible_fields: Final = cached_content is None or litellm.modify_params is False
         if can_send_cache_incompatible_fields:
             if system_instructions is not None:
-                data["system_instruction"] = system_instructions
+                data["systemInstruction"] = system_instructions
             if tools is not None:
                 data["tools"] = tools
             if tool_choice is not None:

litellm/types/llms/vertex_ai.py

@@ -338,7 +338,7 @@ class CachedContent(TypedDict, total=False):
 
 class RequestBody(TypedDict, total=False):
     contents: Required[list[ContentType]]
-    system_instruction: SystemInstructions
+    systemInstruction: SystemInstructions
     tools: Tools
     toolConfig: ToolConfig
     safetySettings: list[SafetSettingsConfig]
@@ -350,7 +350,7 @@ class RequestBody(TypedDict, total=False):
 
 class CachedContentRequestBody(TypedDict, total=False):
     contents: Required[list[ContentType]]
-    system_instruction: SystemInstructions
+    systemInstruction: SystemInstructions
     tools: Tools
     toolConfig: ToolConfig
     model: Required[str]  # Format: models/{model}

tests/local_testing/test_amazing_vertex_completion.py

@@ -2906,7 +2906,7 @@ def test_gemini_function_call_parameter_in_messages():
                         ],
                     },
                 ],
-                "system_instruction": {
+                "systemInstruction": {
                     "parts": [{"text": "Use search for most queries."}]
                 },
                 "tools": [

tests/test_litellm/llms/vertex_ai/context_caching/test_context_caching_ttl.py

@@ -334,7 +334,7 @@ class TestTransformationWithTTL:
 
         assert "ttl" in result
         assert result["ttl"] == "7200s"
-        assert "system_instruction" in result
+        assert "systemInstruction" in result
 
         if custom_llm_provider == "gemini":
             assert result["model"] == "models/gemini-2.5-pro"

tests/test_litellm/llms/vertex_ai/context_caching/test_vertex_ai_context_caching.py

@@ -1599,6 +1599,29 @@ def test_cached_messages_end_on_supported_turn():
     assert cached_messages_end_on_supported_turn([]) is False
 
 
+@pytest.mark.parametrize("custom_llm_provider", ["gemini", "vertex_ai", "vertex_ai_beta"])
+def test_cached_content_system_instruction_uses_canonical_rest_key(custom_llm_provider):
+    """cachedContents.create takes systemInstruction; this was the last snake_case sender in the provider."""
+    from litellm.llms.vertex_ai.context_caching.transformation import (
+        transform_openai_messages_to_gemini_context_caching,
+    )
+
+    result = transform_openai_messages_to_gemini_context_caching(
+        model="gemini-2.5-pro",
+        messages=[
+            {"role": "system", "content": "You are a concise assistant."},
+            {"role": "user", "content": "Reply with exactly: ok"},
+        ],
+        custom_llm_provider=custom_llm_provider,
+        cache_key="test-cache-key",
+        vertex_project="test_project",
+        vertex_location="test_location",
+    )
+
+    assert "system_instruction" not in result
+    assert result["systemInstruction"]["parts"][0]["text"] == "You are a concise assistant."
+
+
 class TestCheckCachePagination:
     """Test pagination logic in check_cache and async_check_cache methods."""
 

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

@@ -123,7 +123,7 @@ def test_cached_content_respects_modify_params_for_cache_incompatible_fields():
             cached_content=cache_name,
         )
         assert result.get("cachedContent") == cache_name
-        assert "system_instruction" in result
+        assert "systemInstruction" in result
         assert "tools" in result
         assert "toolConfig" in result
         assert "contents" in result
@@ -139,7 +139,7 @@ def test_cached_content_respects_modify_params_for_cache_incompatible_fields():
             cached_content=cache_name,
         )
         assert result_modify_true.get("cachedContent") == cache_name
-        assert "system_instruction" not in result_modify_true
+        assert "systemInstruction" not in result_modify_true
         assert "tools" not in result_modify_true
         assert "toolConfig" not in result_modify_true
         assert "contents" in result_modify_true
@@ -153,13 +153,32 @@ def test_cached_content_respects_modify_params_for_cache_incompatible_fields():
             litellm_params={},
             cached_content=None,
         )
-        assert "system_instruction" in result_no_cache
+        assert "systemInstruction" in result_no_cache
         assert "tools" in result_no_cache
         assert "toolConfig" in result_no_cache
     finally:
         litellm.modify_params = original_modify_params
 
 
+@pytest.mark.parametrize("custom_llm_provider", ["gemini", "vertex_ai"])
+def test_system_instruction_uses_canonical_rest_key(custom_llm_provider):
+    """Regression for #37028: strict generateContent endpoints reject the snake_case alias."""
+    result = _transform_request_body(
+        messages=[
+            {"role": "system", "content": "You are a concise assistant."},
+            {"role": "user", "content": "Reply with exactly: ok"},
+        ],
+        model="gemini-2.5-pro",
+        optional_params={},
+        custom_llm_provider=custom_llm_provider,
+        litellm_params={},
+        cached_content=None,
+    )
+
+    assert "system_instruction" not in result
+    assert result["systemInstruction"]["parts"][0]["text"] == "You are a concise assistant."
+
+
 # Tests for issue #14556: Labels field provider-aware filtering
 def test_google_genai_excludes_labels():
     """Test that Google GenAI/AI Studio endpoints exclude labels when custom_llm_provider='gemini'"""

tests/test_litellm/llms/vertex_ai/test_vertex.py

@@ -1579,5 +1579,5 @@ def test_system_prompt_only_adds_blank_user_message():
     #########################################################
     # system message was passed in
     #########################################################
-    assert len(data["system_instruction"]) == 1
-    assert data["system_instruction"]["parts"][0]["text"] == SYSTEM_INSTRUCTION
+    assert len(data["systemInstruction"]) == 1
+    assert data["systemInstruction"]["parts"][0]["text"] == SYSTEM_INSTRUCTION