Claude Code使用gemini模型报错

[teaim]

Claude Code for VS Code最新版 v2.1.74
CLIProxyAPI最新版 v6.8.51

通过修改 AI 提供商配置的接入标准的Gemini接口或Gemini的OpenAI兼容接口
在Claude Code中使用gemini-3-flash-preview可以正常对话, 但是遇到工具调用就会报错

API Error: 400 [{
"error": {
"code": 400,
"message": "Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call default_api:Bash , position 15. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.",
"status": "INVALID_ARGUMENT"
}
}
]

[luispater]

I reviewed the current repo and could not find an active code path here that would forward Gemini tool functionCall parts without a thoughtSignature/sentinel.

The current request translators already inject skip_thought_signature_validator for Gemini-native, OpenAI-compatible chat, OpenAI Responses, and Claude-to-Gemini tool-call flows, and the relevant translator tests pass on the current branch.

If this still reproduces, please reopen with a minimal repro that includes: the exact provider mode (Gemini native vs OpenAI-compatible), the endpoint used (/v1beta/..., /v1/chat/completions, or /v1/responses), a sanitized provider config, and request/response logs around the first tool call. Without a reproducible request path in this repo, there is not enough evidence here to keep this issue actionable.

[weijiafu14]

Claude Code + Gemini 报错通常是协议转换的问题。LiteLLM 在 Gemini 方向有几个已知 bug：

• thinking 模式下 thoughtSignature 混合 → 400 错误 (litellm#17949)
• tool_result 图片被静默丢弃 (litellm#23712)

我做了一个专门的 Gemini ↔ Claude 转换 proxy，直接调 Gemini API 不走 LiteLLM：

pip install "gemini-claude-bridge[proxy]"
GEMINI_API_KEY=your-key BIG_MODEL=gemini-3-flash-preview python examples/proxy_server.py
ANTHROPIC_BASE_URL=http://localhost:8082 claude

GitHub: gemini-claude-bridge

已用 Claude Code CLI 端到端测试通过（文本、工具调用、多轮、流式）。