fix: preserve Gemini thought_signature in tool calls and add ddgs dependency

[SoufianoDev]

Summary

Fixes #25 — two bugs that cause Gemini provider to fail when using tool calls.

Bug 1: Missing ddgs dependency

WebSearchTool imports ddgs/duckduckgo_search at runtime, but the package was not listed in pyproject.toml or agent/requirements.txt. This caused web_search to always return a JSON error, which then triggered Bug 2.

Bug 2: Gemini thought_signature not preserved across turns

Gemini's API returns extra_content.google.thought_signature on tool call responses. This signature must be echoed back in subsequent conversation turns. However, ChatOpenAI from langchain-openai strips extra_content during parsing. When the conversation history was sent back to Gemini without the signature, it returned a 400 INVALID_ARGUMENT error:

Function call is missing a thought_signature in functionCall parts.

Changes

• pyproject.toml / agent/requirements.txt — Added ddgs>=6.0.0 dependency
• agent/src/providers/llm.py — Added GeminiChatOpenAI subclass of ChatOpenAI that overrides _create_chat_result() to capture extra_content from raw API responses and store it in AIMessage.additional_kwargs["tool_call_extras"]. Updated build_llm() to use this class when LANGCHAIN_PROVIDER=gemini.
• agent/src/providers/chat.py — Added thought_signature field to ToolCallRequest. Updated _parse_response() to extract signatures from tool_call_extras. Updated stream_chat() to bypass streaming for Gemini (since stream chunks don't preserve extra_content) and use chat() directly.
• agent/src/agent/context.py — Updated format_assistant_tool_calls() to include extra_content.google.thought_signature in formatted assistant messages when present.
• agent/tests/test_gemini_thought_signature.py — Added 23 tests covering the full thought_signature preservation flow, including end-to-end tests with real OpenAI SDK Pydantic models.

Testing

Platform	Tests	Result
Linux (native)	23 new + 35 existing	🗹 All pass
Wine (Windows compat)	23 new + 35 existing	🗹 All pass
Wine (live)	WebSearchTool with ddgs	🗹 Returns results

[warren618]

Hi @SoufianoDev, thanks for the thorough PR and the detailed tests!

Unfortunately, this approach won't work due to a limitation in langchain-openai's message conversion pipeline. We verified this by running the code against a real API:

When the formatted assistant message (with extra_content) is passed back to ChatOpenAI.invoke(), LangChain converts the dict through convert_to_messages() → AIMessage → _convert_message_to_dict(). During this round-trip, all non-standard fields on tool calls are stripped:

# langchain-openai/chat_models/base.py
tool_call_supported_props = {"id", "type", "function"}
message_dict["tool_calls"] = [
    {k: v for k, v in tool_call.items() if k in tool_call_supported_props}
    for tool_call in message_dict["tool_calls"]
]

So even though the PR correctly captures and re-attaches thought_signature, it gets silently dropped before reaching the Gemini API — the 400 error would still occur.

We've documented this on the issue (#25) — this is an upstream LangChain limitation that can't be worked around from the application side. The current recommendation for Gemini users is to switch to a provider that doesn't require thought_signature (DeepSeek, Qwen, etc.).

The ddgs dependency fix has already been merged in b17aeec.

Appreciate the effort — happy to revisit if LangChain adds support for passthrough fields!