Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.
Python agent-contextagent-memoryagent-pluginsagentic-ragcontext-databasedsh-pluginknowledge-basememorymemory-managementself-evolvingvector-database
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.pyindex 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"] = thinkingkwargs["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 kwargsdef _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.pynew file mode 100644index 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]