Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

fix(vlm): strip cache_control for Gemini + tools to avoid LiteLLM CachedContent 400 (#1569)

Brian Le4c882afca61d
diff --git a/openviking/models/vlm/backends/litellm_vlm.py b/openviking/models/vlm/backends/litellm_vlm.py
index 913acd484..cc30b8a83 100644
--- a/openviking/models/vlm/backends/litellm_vlm.py
+++ b/openviking/models/vlm/backends/litellm_vlm.py
@@ -232,6 +232,20 @@ class LiteLLMVLMProvider(VLMBase):
extra["enable_thinking"] = thinking
kwargs["extra_body"] = extra
+ # Workaround for LiteLLM bug where Gemini context-caching path emits
+ # both `cachedContent` and `toolConfig`, which Gemini rejects with a
+ # 400 "CachedContent can not be used with GenerateContent request
+ # setting system_instruction, tools or tool_config".
+ # See BerriAI/litellm#17304 and PR #25659. Remove when LiteLLM ships
+ # the fix.
+ if provider == "gemini" and tools:
+ kwargs["messages"] = [
+ {k: v for k, v in msg.items() if k != "cache_control"}
+ if isinstance(msg, dict)
+ else msg
+ for msg in messages
+ ]
+
return kwargs
def _parse_tool_calls(self, message) -> List[ToolCall]:
diff --git a/tests/unit/test_litellm_vlm_gemini_cache.py b/tests/unit/test_litellm_vlm_gemini_cache.py
new file mode 100644
index 000000000..b07b9e870
--- /dev/null
+++ b/tests/unit/test_litellm_vlm_gemini_cache.py
@@ -0,0 +1,70 @@
+# Copyright (c) 2026 Beijing Volcano Engine Technology Co., Ltd.
+# SPDX-License-Identifier: AGPL-3.0
+"""Tests for Gemini cache_control stripping workaround.
+
+LiteLLM has an open bug (BerriAI/litellm#17304, PR #25659) where the Gemini
+context-caching path leaves `tool_choice` in optional_params while also
+setting `cachedContent`. Gemini then rejects the combination with:
+
+ "CachedContent can not be used with GenerateContent request setting
+ system_instruction, tools or tool_config."
+
+OpenViking's memory-extraction ReAct loop adds `cache_control: ephemeral`
+markers and passes `tool_choice="auto"`, so every Gemini call with tools
+hits this 400. The LiteLLM VLM backend works around it by stripping
+`cache_control` markers from messages when the resolved provider is
+`gemini`.
+"""
+
+from openviking.models.vlm.backends.litellm_vlm import LiteLLMVLMProvider
+
+
+def _vlm(model: str) -> LiteLLMVLMProvider:
+ return LiteLLMVLMProvider({"api_key": "fake", "model": model, "max_tokens": 256})
+
+
+class TestGeminiCacheControlStripping:
+ def test_cache_control_stripped_for_gemini_with_tools(self):
+ vlm = _vlm("gemini-3.1-flash-lite-preview")
+ messages = [{"role": "user", "content": "hi", "cache_control": {"type": "ephemeral"}}]
+ tools = [{"type": "function", "function": {"name": "read", "parameters": {}}}]
+ kwargs = vlm._build_kwargs(
+ model="gemini/gemini-3.1-flash-lite-preview",
+ messages=messages,
+ tools=tools,
+ )
+ assert all("cache_control" not in m for m in kwargs["messages"])
+ assert kwargs["messages"][0]["content"] == "hi"
+ assert kwargs["tools"] == tools
+
+ def test_cache_control_preserved_for_gemini_without_tools(self):
+ vlm = _vlm("gemini-3.1-flash-lite-preview")
+ messages = [{"role": "user", "content": "hi", "cache_control": {"type": "ephemeral"}}]
+ kwargs = vlm._build_kwargs(
+ model="gemini/gemini-3.1-flash-lite-preview",
+ messages=messages,
+ )
+ assert kwargs["messages"][0].get("cache_control") == {"type": "ephemeral"}
+
+ def test_cache_control_preserved_for_anthropic(self):
+ vlm = _vlm("claude-haiku-4-5")
+ messages = [{"role": "user", "content": "hi", "cache_control": {"type": "ephemeral"}}]
+ tools = [{"type": "function", "function": {"name": "read", "parameters": {}}}]
+ kwargs = vlm._build_kwargs(
+ model="anthropic/claude-haiku-4-5",
+ messages=messages,
+ tools=tools,
+ )
+ assert kwargs["messages"][0].get("cache_control") == {"type": "ephemeral"}
+
+ def test_non_dict_messages_left_intact(self):
+ vlm = _vlm("gemini-3.1-flash-lite-preview")
+ sentinel = object()
+ messages = [sentinel]
+ tools = [{"type": "function", "function": {"name": "read", "parameters": {}}}]
+ kwargs = vlm._build_kwargs(
+ model="gemini/gemini-3.1-flash-lite-preview",
+ messages=messages,
+ tools=tools,
+ )
+ assert kwargs["messages"] == [sentinel]