[naveenkarmegam]

Description

The LangGraph currency agent sample (samples/python/agents/langgraph) currently fails on every request with a JSON-RPC -32603 Internal error. This PR fixes the three underlying causes described in #722. There is one commit per cause, plus a small test-client fix.

Changes

1. fix(langgraph): use frankfurter.dev v1 endpoint
api.frankfurter.app now returns 301 to api.frankfurter.dev/v1. httpx doesn't follow redirects by default, so the tool always returned API request failed. This commit updates the tool URL and the README docs link.

2. fix(langgraph): upgrade to LangChain/LangGraph 1.x and gemini-3.8-flash

   • gemini-2.0-flash has been retired and returns 404 NOT_FOUND. Switched to gemini-3.8-flash.
   • Current Gemini models need thought signatures on function calls, which langchain-google-genai 2.x doesn't send (400 Function call is missing a thought_signature). Bumped langchain-google-genai>=3.0.0, langgraph>=1.0.0 and langchain-openai>=1.0.0, added langchain>=1.0.0, and regenerated uv.lock.
   • On LangGraph 1.x, create_react_agent(..., response_format=...) makes an extra model call that ends on a model turn, and Gemini rejects it (400 Requests ending with a model turn are not supported). Replaced it with langchain.agents.create_agent + ToolStrategy(ResponseFormat), which returns the structured response through a tool call.
   • Registered ResponseFormat with the checkpointer's JsonPlusSerializer. Without this, LangGraph 1.x warns that restoring it between turns "will be blocked in a future version", which would break multi-turn conversations.

   The model change and the dependency upgrade are in one commit because neither works without the other.

3. fix(langgraph): raise test client HTTP timeout
   A full request (model → tool → structured response) can take more than httpx's default 5 s, so test_client.py failed with A2AClientTimeoutError even though the server succeeded. Raised the timeout to 60 s.

Testing

Ran uv run app and app/test_client.py locally (Windows 11, Python 3.12, a2a-sdk 0.3.0):

• Single-turn message/send: "how much is 10 USD in INR?" → completed with the converted amount
• Multi-turn: "exchange rate for 1 USD?" → input-required → "CAD" → completed
• Streaming message/stream: submitted → working updates → artifact → completed
• No warnings or errors in the server log
• model_source=openai path still builds the agent with langchain-openai 1.x
• ruff check / ruff format: no new findings in changed lines (the file's pre-existing findings are unchanged)

Out of scope

About 22 other Python samples still reference gemini-2.0-flash and/or api.frankfurter.app and are likely affected too. I've kept this PR to the LangGraph sample and can follow up separately.

• Follow the CONTRIBUTING Guide.

Fixes #722 🦕