botrun-flow-lang 5.12.263__py3-none-any.whl → 6.2.21__py3-none-any.whl

botrun_flow_lang/api/langgraph_api.py

@@ -1,811 +1,816 @@
-import logging
-import uuid
-import json
-import random
-import time
-import re
-
-from fastapi import APIRouter, HTTPException
-
-from pydantic import BaseModel
-
-from typing import Dict, Any, List, Optional
-
-from fastapi.responses import StreamingResponse
-
-from botrun_flow_lang.constants import ERR_GRAPH_RECURSION_ERROR, LANG_EN, LANG_ZH_TW
-
-from botrun_flow_lang.langgraph_agents.agents.agent_runner import (
-    agent_runner,
-    langgraph_runner,
-)
-
-from botrun_flow_lang.langgraph_agents.agents.gov_researcher.gov_researcher_graph import (
-    GovResearcherGraph,
-    get_content_for_gov_researcher,
-)
-from botrun_flow_lang.langgraph_agents.agents.gov_researcher.gov_researcher_2_graph import (
-    TAIWAN_SUBSIDY_SUPERVISOR_PROMPT,
-    create_taiwan_subsidy_agent_graph,
-    taiwan_subsidy_agent_graph,
-)
-from botrun_flow_lang.langgraph_agents.agents.gov_researcher.gemini_subsidy_graph import (
-    TAIWAN_SUBSIDY_SUPERVISOR_PROMPT as GEMINI_SUBSIDY_PROMPT,
-    create_gemini_subsidy_agent_graph,
-    gemini_subsidy_agent_graph,
-)
-from botrun_flow_lang.langgraph_agents.agents.langgraph_react_agent import (
-    create_react_agent_graph,
-    get_react_agent_model_name,
-)
-
-from botrun_flow_lang.langgraph_agents.cache.langgraph_botrun_cache import (
-    get_botrun_cache,
-)
-
-from botrun_flow_lang.models.token_usage import TokenUsage
-
-from botrun_flow_lang.utils.botrun_logger import (
-    get_session_botrun_logger,
-    default_logger,
-)
-
-# 放到要用的時候才 init，不然loading 會花時間
-# 因為要讓 langgraph 在本地端執行，所以這一段又搬回到外面了
-from langgraph.errors import GraphRecursionError
-import anthropic  # Keep relevant imports if needed for error handling here
-
-# ==========
-
-from botrun_flow_lang.langgraph_agents.agents.search_agent_graph import (
-    SearchAgentGraph,
-    # graph as search_agent_graph,
-)
-
-from botrun_flow_lang.utils.langchain_utils import (
-    extract_token_usage_from_state,
-    langgraph_event_to_json,
-    litellm_msgs_to_langchain_msgs,
-)
-
-
-router = APIRouter(prefix="/langgraph")
-
-
-class LangGraphRequest(BaseModel):
-    graph_name: str
-    # todo LangGraph 應該要傳 thread_id，但是因為現在是 cloud run 的架構，所以 thread_id 不一定會讀的到 (auto scale)
-    thread_id: Optional[str] = None
-    user_input: Optional[str] = None
-    messages: List[Dict[str, Any]] = []
-    config: Optional[Dict[str, Any]] = None
-    stream: bool = False
-    # LangGraph 是否需要從 checkpoint 恢復
-    need_resume: bool = False
-    session_id: Optional[str] = None
-
-
-class LangGraphResponse(BaseModel):
-    """
-    @param content: 這個是給評測用來評估結果用的
-    @param state: 這個是graph的 final state，如果需要額外資訊可以使用
-    @param token_usage: Token usage information for the entire graph execution
-    """
-
-    id: str
-    object: str = "chat.completion"
-    created: int
-    model: str
-    content: Optional[str] = None
-    state: Optional[Dict[str, Any]] = None
-    token_usage: Optional[TokenUsage] = None
-
-
-class SupportedGraphsResponse(BaseModel):
-    """Response model for listing supported graphs"""
-
-    graphs: List[str]
-
-
-class GraphSchemaRequest(BaseModel):
-    """Request model for getting graph schema"""
-
-    graph_name: str
-
-
-PERPLEXITY_SEARCH_AGENT = "perplexity_search_agent"
-CUSTOM_WEB_RESEARCH_AGENT = "custom_web_research_agent"
-LANGGRAPH_REACT_AGENT = "langgraph_react_agent"
-DEEP_RESEARCH_AGENT = "deep_research_agent"
-# GOV_RESEARCHER_AGENT = "gov_researcher_agent"
-GOV_SUBSIDY_AGENT = "gov_subsidy_agent"
-GEMINI_SUBSIDY_AGENT = "gemini_subsidy_agent"
-
-
-SUPPORTED_GRAPH_NAMES = [
-    # PERPLEXITY_SEARCH_AGENT,
-    LANGGRAPH_REACT_AGENT,
-    GOV_SUBSIDY_AGENT,
-    GEMINI_SUBSIDY_AGENT,
-    # GOV_RESEARCHER_AGENT,
-]
-SUPPORTED_GRAPH = {
-    GOV_SUBSIDY_AGENT: taiwan_subsidy_agent_graph,
-    GEMINI_SUBSIDY_AGENT: gemini_subsidy_agent_graph,
-}
-
-
-def contains_chinese_chars(text: str) -> bool:
-    """Check if the given text contains any Chinese characters."""
-    if not text:
-        return False
-    # This pattern matches Chinese characters (both simplified and traditional)
-    pattern = re.compile(
-        r"[\u4e00-\u9fff\u3400-\u4dbf\U00020000-\U0002a6df\U0002a700-\U0002ebef]"
-    )
-    return bool(pattern.search(text))
-
-
-async def get_cached_or_create_react_graph(
-    botrun_id: Optional[str],  # Key parameter - can be None/empty
-    config: Optional[Dict[str, Any]] = None,
-    messages: Optional[List[Dict]] = None,
-    user_input: Optional[str] = None,
-    logger: logging.Logger = default_logger,
-) -> Any:
-    """
-    Get cached graph for LANGGRAPH_REACT_AGENT or create new one.
-    Handles botrun_id-based caching with parameter validation.
-
-    Args:
-        botrun_id: The botrun ID for cache lookup. If None/empty, skips caching.
-        config: Configuration dictionary
-        messages: List of message dictionaries
-        user_input: User input string
-        logger: Logger instance
-
-    Returns:
-        Graph instance (cached or newly created)
-    """
-    config = config or {}
-
-    # Extract parameters for hash calculation (moved from get_graph)
-    system_prompt = config.get("system_prompt", "")
-    if messages:
-        for message in messages:
-            if message.get("role") == "system":
-                system_prompt = message.get("content", "")
-
-    botrun_flow_lang_url = config.get("botrun_flow_lang_url", "")
-    user_id = config.get("user_id", "")
-    model_name = config.get("model_name", "")
-
-    # Determine language (moved from get_graph)
-    has_chinese = contains_chinese_chars(system_prompt)
-    if not has_chinese and user_input:
-        has_chinese = contains_chinese_chars(user_input)
-    lang = LANG_ZH_TW if has_chinese else LANG_EN
-
-    mcp_config = config.get("mcp_config")
-
-    # CRITICAL: Check if botrun_id is provided and not empty
-    if not botrun_id:
-        # If botrun_id is None or empty, skip caching entirely
-        logger.info("No botrun_id provided, creating new graph without caching")
-        graph = await create_react_agent_graph(
-            system_prompt=system_prompt,
-            botrun_flow_lang_url=botrun_flow_lang_url,
-            user_id=user_id,
-            model_name=model_name,
-            lang=lang,
-            mcp_config=mcp_config,
-        )
-        return graph
-
-    # If botrun_id is provided, use caching logic
-    cache = get_botrun_cache()
-    params_hash = cache.get_params_hash(
-        system_prompt, botrun_flow_lang_url, user_id, model_name, lang, mcp_config
-    )
-
-    # Try to get cached graph
-    cached_graph = cache.get_cached_graph(botrun_id, params_hash)
-    if cached_graph:
-        logger.info(f"Using cached graph for botrun_id: {botrun_id}")
-        return cached_graph
-
-    # Create new graph (same logic as in get_graph)
-    logger.info(f"Creating new graph for botrun_id: {botrun_id}")
-    graph = await create_react_agent_graph(
-        system_prompt=system_prompt,
-        botrun_flow_lang_url=botrun_flow_lang_url,
-        user_id=user_id,
-        model_name=model_name,
-        lang=lang,
-        mcp_config=mcp_config,
-    )
-
-    # Cache the new graph
-    cache.cache_graph(botrun_id, params_hash, graph)
-
-    return graph
-
-
-async def get_graph(
-    graph_name: str,
-    config: Optional[Dict[str, Any]] = None,
-    stream: bool = False,
-    messages: Optional[List[Dict]] = [],
-    user_input: Optional[str] = None,
-):
-    if (
-        graph_name not in SUPPORTED_GRAPH_NAMES
-        and graph_name not in SUPPORTED_GRAPH.keys()
-    ):
-        raise ValueError(f"Unsupported graph from get_graph: {graph_name}")
-    if graph_name == PERPLEXITY_SEARCH_AGENT:
-        graph = SearchAgentGraph().graph
-        graph_config = {
-            "search_prompt": config.get("search_prompt", ""),
-            "model_name": config.get("model_name", "sonar-reasoning-pro"),
-            "related_prompt": config.get("related_question_prompt", ""),
-            "search_vendor": config.get("search_vendor", "perplexity"),
-            "domain_filter": config.get("domain_filter", []),
-            "user_prompt_prefix": config.get("user_prompt_prefix", ""),
-            "stream": stream,
-        }
-    elif graph_name == GOV_SUBSIDY_AGENT:
-        graph = create_taiwan_subsidy_agent_graph(
-            config.get("prompt_template", TAIWAN_SUBSIDY_SUPERVISOR_PROMPT)
-        )
-        graph_config = {
-            "prompt_template": config.get("prompt_template", ""),
-            "legal_extraction_prompt": config.get("legal_extraction_prompt", ""),
-            "faq_extraction_prompt": config.get("faq_extraction_prompt", ""),
-            "calculation_analysis_prompt": config.get(
-                "calculation_analysis_prompt", ""
-            ),
-        }
-    elif graph_name == GEMINI_SUBSIDY_AGENT:
-        graph = create_gemini_subsidy_agent_graph(
-            config.get("prompt_template", GEMINI_SUBSIDY_PROMPT)
-        )
-        graph_config = {
-            "prompt_template": config.get("prompt_template", ""),
-        }
-    else:
-        raise ValueError(f"Unsupported graph type: {graph_name}")
-    return graph, graph_config
-
-
-def get_init_state(
-    graph_name: str,
-    user_input: str,
-    config: Optional[Dict[str, Any]] = None,
-    messages: Optional[List[Dict]] = [],
-    enable_prompt_caching: bool = False,
-):
-    if graph_name == PERPLEXITY_SEARCH_AGENT:
-        if len(messages) > 0:
-            return {"messages": litellm_msgs_to_langchain_msgs(messages)}
-        if config.get("user_prompt_prefix", ""):
-            return {
-                "messages": [
-                    {
-                        "role": "user",
-                        "content": config.get("user_prompt_prefix", "")
-                        + "\n\n"
-                        + user_input,
-                    }
-                ]
-            }
-
-        return {"messages": [user_input]}
-    elif graph_name == CUSTOM_WEB_RESEARCH_AGENT:
-        if len(messages) > 0:
-            return {
-                "messages": litellm_msgs_to_langchain_msgs(messages),
-                "model": config.get("model", "anthropic"),
-            }
-        return {
-            "messages": [user_input],
-            "model": config.get("model", "anthropic"),
-        }
-    elif graph_name == LANGGRAPH_REACT_AGENT:
-        if len(messages) > 0:
-            new_messages = []
-            for message in messages:
-                if message.get("role") != "system":
-                    new_messages.append(message)
-
-            return {
-                "messages": litellm_msgs_to_langchain_msgs(
-                    new_messages, enable_prompt_caching
-                )
-            }
-        else:
-            return {
-                "messages": [user_input],
-            }
-    elif graph_name == DEEP_RESEARCH_AGENT:
-        if len(messages) > 0:
-            return {
-                "messages": litellm_msgs_to_langchain_msgs(messages),
-                "topic": user_input,
-            }
-        return {
-            "messages": [user_input],
-            "topic": user_input,
-        }
-    else:
-        if len(messages) > 0:
-            return {"messages": litellm_msgs_to_langchain_msgs(messages)}
-        return {"messages": [user_input]}
-
-
-def get_content(graph_name: str, state: Dict[str, Any]):
-    if graph_name == PERPLEXITY_SEARCH_AGENT:
-        return state["messages"][-3].content
-    elif graph_name == CUSTOM_WEB_RESEARCH_AGENT:
-        content = state["answer"].get("markdown", "")
-        content = content.replace("\\n", "\n")
-        if state["answer"].get("references", []):
-            references = "\n\n參考資料：\n"
-            for reference in state["answer"]["references"]:
-                references += f"- [{reference['title']}]({reference['url']})\n"
-            content += references
-        return content
-    elif graph_name == DEEP_RESEARCH_AGENT:
-        sections = state["sections"]
-        sections_str = "\n\n".join(
-            f"章節: {section.name}\n"
-            f"描述: {section.description}\n"
-            f"需要研究: {'是' if section.research else '否'}\n"
-            for section in sections
-        )
-        sections_str = "預計報告結構：\n\n" + sections_str
-        return sections_str
-    # elif graph_name == GOV_RESEARCHER_AGENT:
-    #     return get_content_for_gov_researcher(state)
-    elif graph_name == GOV_SUBSIDY_AGENT or graph_name == GEMINI_SUBSIDY_AGENT:
-        messages = state["messages"]
-        # Find the last AI message
-        for msg in reversed(messages):
-            if msg.type == "ai":
-                if isinstance(msg.content, list):
-                    return msg.content[0].get("text", "")
-                else:
-                    return msg.content
-        return ""  # If no AI message found
-    else:
-        messages = state["messages"]
-        # Find the last human message
-        last_human_idx = -1
-        for i, msg in enumerate(messages):
-            if msg.type == "human":
-                last_human_idx = i
-
-        # Combine all AI messages after the last human message
-        ai_contents = ""
-        for msg in messages[last_human_idx + 1 :]:
-            if msg.type == "ai":
-                if isinstance(msg.content, list):
-                    ai_contents += msg.content[0].get("text", "")
-                else:
-                    ai_contents += msg.content
-
-        return ai_contents
-
-
-async def process_langgraph_request(
-    request: LangGraphRequest,
-    retry: bool = False,  # Keep retry logic for non-streaming path if needed
-    logger: logging.Logger = default_logger,
-) -> Any:  # Return type can be LangGraphResponse or StreamingResponse
-    """處理 LangGraph 請求的核心邏輯"""
-    # --- Streaming Case ---
-    if request.stream:
-        logger.info(f"Processing STREAM request for graph: {request.graph_name}")
-        # Use the new wrapper generator that handles resource management
-        return StreamingResponse(
-            managed_langgraph_stream_wrapper(request, logger),
-            media_type="text/event-stream",
-        )
-
-    # --- Non-Streaming Case ---
-    logger.info(f"Processing NON-STREAM request for graph: {request.graph_name}")
-    try:
-        config = request.config or {}
-        mcp_config = config.get("mcp_config")
-        user_id = config.get("user_id")
-
-        # Get botrun_id from config
-        botrun_id = config.get("botrun_id")  # Can be None/empty
-
-        # --- Graph and State Initialization (OUTSIDE of MCP client context) ---
-        # Cache logic for LANGGRAPH_REACT_AGENT only
-        if request.graph_name == LANGGRAPH_REACT_AGENT:
-            graph = await get_cached_or_create_react_graph(
-                botrun_id=botrun_id,  # Pass botrun_id (can be None)
-                config=request.config,
-                messages=request.messages,
-                user_input=request.user_input,
-                logger=logger,
-            )
-            graph_config = request.config
-        else:
-            # Existing logic for other graph types (calls modified get_graph)
-            graph, graph_config = await get_graph(
-                request.graph_name,
-                request.config,
-                False,  # stream=False
-                request.messages,
-                request.user_input,
-            )
-
-        # Determine model name for init_state caching logic if needed
-        # user_input_model_name = request.config.get("model_name", "")
-        # enable_caching = get_react_agent_model_name(user_input_model_name).startswith("claude-")
-
-        init_state = get_init_state(
-            request.graph_name,
-            request.user_input,
-            request.config,
-            request.messages,
-            False,  # enable_prompt_caching=enable_caching
-        )
-
-        thread_id = request.thread_id if request.thread_id else str(uuid.uuid4())
-        logger.info(f"Running non-stream with thread_id: {thread_id}")
-
-        # --- Run the agent (no MCP client needed during execution) ---
-        logger.info("Executing agent_runner for non-stream request...")
-        async for _ in agent_runner(
-            thread_id,
-            init_state,
-            graph,
-            request.need_resume,
-            extra_config=graph_config,
-        ):
-            pass  # Just consume the events
-
-        logger.info(
-            f"agent_runner completed for thread_id: {thread_id}. Fetching final state."
-        )
-
-        # --- Get Final State and Prepare Response (OUTSIDE of MCP client context) ---
-        config_for_state = {"configurable": {"thread_id": thread_id}}
-        state = await graph.aget_state(config_for_state)
-
-        try:
-            state_values_json = langgraph_event_to_json(state.values)
-            logger.info(
-                f"Final state fetched for {thread_id}: {state_values_json[:500]}..."
-            )  # Log truncated state
-        except Exception as e_log:
-            logger.error(f"Error serializing final state for logging: {e_log}")
-            logger.info(
-                f"Final state keys for {thread_id}: {list(state.values.keys())}"
-            )
-
-        content = get_content(request.graph_name, state.values)
-
-        model_name_config = (
-            request.config.get("model_name", "") if request.config else ""
-        )
-        final_model_name = model_name_config  # Default to config model name
-        if request.graph_name == LANGGRAPH_REACT_AGENT:
-            final_model_name = get_react_agent_model_name(model_name_config)
-            token_usage = extract_token_usage_from_state(state.values, final_model_name)
-        else:
-            token_usage = TokenUsage(
-                total_input_tokens=0,
-                total_output_tokens=0,
-                total_tokens=0,
-                nodes=[],
-            )
-
-        return LangGraphResponse(
-            id=thread_id,
-            created=int(time.time()),
-            model=request.graph_name,  # Or final_model_name? Check requirements
-            content=content,
-            state=state.values,  # Consider serializing state here if needed client-side
-            token_usage=token_usage,
-        )
-
-    except anthropic.RateLimitError as e:
-        if retry:
-            logger.error(
-                "Retry failed with Anthropic RateLimitError (non-stream)", exc_info=True
-            )
-            raise HTTPException(
-                status_code=429, detail=f"Rate limit exceeded after retry: {e}"
-            )  # 429 is more appropriate
-
-        logger.warning(
-            f"Anthropic RateLimitError occurred (non-stream): {e}. Retrying..."
-        )
-        retry_delay = random.randint(7, 20)
-        time.sleep(
-            retry_delay
-        )  # Note: time.sleep blocks async. Consider asyncio.sleep(retry_delay) if this becomes an issue.
-        logger.info(f"Retrying non-stream request after {retry_delay}s delay...")
-        return await process_langgraph_request(
-            request, retry=True, logger=logger
-        )  # Recursive call for retry
-
-    except GraphRecursionError as e:
-        logger.error(f"GraphRecursionError (non-stream): {e}", exc_info=True)
-        raise HTTPException(
-            status_code=500, detail=f"Graph execution exceeded maximum depth: {e}"
-        )
-
-    except Exception as e:
-        import traceback
-
-        tb_str = traceback.format_exc()
-        logger.error(
-            f"Unhandled exception in process_langgraph_request (non-stream): {e}",
-            exc_info=True,
-        )
-        raise HTTPException(
-            status_code=500, detail=f"Internal Server Error during graph execution: {e}"
-        )
-
-
-@router.post("/invoke")
-async def invoke(request: LangGraphRequest):
-    """
-    執行指定的 LangGraph，支援串流和非串流模式
-
-    Args:
-        request: 包含 graph_name 和輸入數據的請求
-
-    Returns:
-        串流模式: StreamingResponse
-        非串流模式: LangGraphResponse
-    """
-    session_id = request.session_id
-    user_id = request.config.get("user_id", "")
-
-    # *** Create a session-specific BotrunLogger for this specific request ***
-    # This ensures Cloud Logging and session/user context
-    logger = get_session_botrun_logger(session_id=session_id, user_id=user_id)
-
-    logger.info(
-        "invoke LangGraph API",
-        request=request.model_dump(),
-    )
-
-    # Pass the request-specific BotrunLogger down
-    return await process_langgraph_request(request, logger=logger)
-
-
-# NEW: Wrapper generator for managing resources during streaming
-async def managed_langgraph_stream_wrapper(
-    request: LangGraphRequest, logger: logging.Logger
-):
-    """
-    Manages AsyncExitStack and MCPClient lifecycle for streaming responses.
-    Initializes graph and then yields events from langgraph_stream_response_generator.
-    """
-    try:
-        config = request.config or {}
-        mcp_config = config.get("mcp_config")
-        user_id = config.get("user_id")
-        print(f"mcp_config: {mcp_config}, user_id: {user_id}")
-
-        # Get botrun_id from config
-        botrun_id = config.get("botrun_id")  # Can be None/empty
-
-        # --- Graph and State Initialization (OUTSIDE of MCP client context) ---
-        logger.info("Getting graph and initial state for stream...")
-        # Cache logic for LANGGRAPH_REACT_AGENT only
-        if request.graph_name == LANGGRAPH_REACT_AGENT:
-            graph = await get_cached_or_create_react_graph(
-                botrun_id=botrun_id,  # Pass botrun_id (can be None)
-                config=request.config,
-                messages=request.messages,
-                user_input=request.user_input,
-                logger=logger,
-            )
-            graph_config = request.config
-        else:
-            # Existing logic for other graph types (calls modified get_graph)
-            graph, graph_config = await get_graph(
-                request.graph_name,
-                request.config,
-                request.stream,  # Pass stream=True
-                request.messages,
-                request.user_input,
-            )
-
-        # Determine model name for init_state caching logic if needed
-        # user_input_model_name = request.config.get("model_name", "")
-        # enable_caching = get_react_agent_model_name(user_input_model_name).startswith("claude-") # Example
-
-        init_state = get_init_state(
-            request.graph_name,
-            request.user_input,
-            request.config,
-            request.messages,
-            False,  # enable_prompt_caching=enable_caching # Pass caching flag if used
-        )
-
-        thread_id = request.thread_id if request.thread_id else str(uuid.uuid4())
-        logger.info(f"Streaming with thread_id: {thread_id}")
-
-        # --- Yield from the actual stream response generator ---
-        async for event in langgraph_stream_response_generator(
-            thread_id,
-            init_state,
-            graph,
-            request.need_resume,
-            logger,
-            graph_config,
-        ):
-            yield event  # Yield the formatted event string
-
-    except anthropic.RateLimitError as e:
-        # Handle rate limit errors specifically for streaming if needed
-        # Note: Retry logic might be complex to implement correctly within a generator.
-        #       Consider if retry should happen at a higher level or if yielding an error is sufficient.
-        logger.error(
-            f"Anthropic RateLimitError during stream setup/execution: {e}",
-            exc_info=True,
-        )
-        error_payload = json.dumps(
-            {"error": f"Rate Limit Error: {e}", "retry_possible": False}
-        )  # Indicate no auto-retry here
-        yield f"data: {error_payload}\n\n"
-        yield "data: [DONE]\n\n"  # Ensure stream terminates correctly
-
-    except GraphRecursionError as e:
-        # Handle recursion errors specifically (can happen during graph execution)
-        logger.error(
-            f"GraphRecursionError during stream: {e} for thread_id: {thread_id}",
-            error=str(e),
-            exc_info=True,
-        )
-        try:
-            error_msg = json.dumps(
-                {"error": ERR_GRAPH_RECURSION_ERROR, "detail": str(e)}
-            )
-            yield f"data: {error_msg}\n\n"
-        except Exception as inner_e:
-            logger.error(
-                f"Error serializing GraphRecursionError for stream: {inner_e}",
-                exc_info=True,
-            )
-            yield f"data: {json.dumps({'error': ERR_GRAPH_RECURSION_ERROR})}\n\n"
-        yield "data: [DONE]\n\n"  # Ensure stream terminates correctly
-
-    except Exception as e:
-        # Catch-all for other errors during setup or streaming
-        import traceback
-
-        tb_str = traceback.format_exc()
-        logger.error(
-            f"Unhandled exception in managed_langgraph_stream_wrapper: {e}",
-            exc_info=True,
-            traceback=tb_str,
-        )
-        error_payload = json.dumps({"error": f"Streaming Error: {e}", "detail": tb_str})
-        yield f"data: {error_payload}\n\n"
-        yield "data: [DONE]\n\n"  # Ensure stream terminates correctly
-
-
-# RENAMED: Original langgraph_stream_response, now focused on generation
-async def langgraph_stream_response_generator(
-    thread_id: str,
-    init_state: Dict,
-    graph: Any,  # Receives the already configured graph
-    need_resume: bool = False,
-    logger: logging.Logger = default_logger,
-    extra_config: Optional[Dict] = None,
-):
-    """
-    Generates LangGraph stream events using langgraph_runner.
-    Handles formatting ('data: ...') and '[DONE]' signal.
-    Exception handling specific to langgraph_runner execution.
-    """
-    try:
-        logger.info(
-            "Starting langgraph_runner iteration",
-            thread_id=thread_id,
-            need_resume=need_resume,
-        )
-
-        final_event = None
-        first_event = True  # To potentially log first event differently if needed
-        async for event in langgraph_runner(
-            thread_id, init_state, graph, need_resume, extra_config
-        ):
-            final_event = event  # Keep track of the last event
-            event_json_str = langgraph_event_to_json(event)  # Serialize event safely
-            if first_event:
-                # Optional: Different logging for the very first event chunk
-                logger.info(
-                    f"First stream event for {thread_id}: {event_json_str[:200]}..."
-                )  # Log truncated first event
-                first_event = False
-
-            # print statement for local debugging if needed
-            # from datetime import datetime
-            # print(datetime.now().strftime("%Y-%m-%d %H:%M:%S"), event_json_str)
-
-            yield f"data: {event_json_str}\n\n"
-
-        # Log details about the final event if needed
-        if final_event:
-            logger.info(
-                "Finished langgraph_runner iteration",
-                thread_id=thread_id,
-                final_event_type=final_event.get("event"),
-                # final_event_data=langgraph_event_to_json(final_event) # Log full final event if useful
-            )
-        else:
-            logger.warning(
-                "langgraph_runner finished without yielding any events",
-                thread_id=thread_id,
-            )
-
-        yield "data: [DONE]\n\n"  # Signal end of stream
-
-    # Error handling remains here as these errors occur during langgraph_runner
-    except GraphRecursionError as e:
-        logger.error(
-            f"GraphRecursionError in stream generator: {e} for thread_id: {thread_id}",
-            error=str(e),
-            exc_info=True,
-        )
-        try:
-            error_msg = json.dumps(
-                {"error": ERR_GRAPH_RECURSION_ERROR, "detail": str(e)}
-            )
-            yield f"data: {error_msg}\n\n"
-        except Exception as inner_e:
-            logger.error(
-                f"Error serializing GraphRecursionError msg: {inner_e}", exc_info=True
-            )
-            yield f"data: {json.dumps({'error': ERR_GRAPH_RECURSION_ERROR})}\n\n"
-        # Ensure [DONE] is sent even after handled error to terminate client side
-        yield "data: [DONE]\n\n"
-
-    except Exception as e:
-        # Catch errors specifically from langgraph_runner or event processing
-        import traceback
-
-        tb_str = traceback.format_exc()
-        logger.error(
-            f"Exception in stream generator: {e} for thread_id: {thread_id}",
-            error=str(e),
-            exc_info=True,
-            traceback=tb_str,
-        )
-        error_response = {"error": f"Stream Generation Error: {e}", "detail": tb_str}
-        yield f"data: {json.dumps(error_response)}\n\n"
-        # Ensure [DONE] is sent even after handled error
-        yield "data: [DONE]\n\n"
-
-
-@router.get("/list", response_model=SupportedGraphsResponse)
-async def list_supported_graphs():
-    """
-    列出所有支援的 LangGraph names
-
-    Returns:
-        包含所有支援的 graph names 的列表
-    """
-    return SupportedGraphsResponse(graphs=list(SUPPORTED_GRAPH.keys()))
-
-
-@router.post("/schema", response_model=dict)
-async def get_graph_schema(request: GraphSchemaRequest):
-    """
-    取得指定 graph 的 schema
-    """
-    if request.graph_name not in SUPPORTED_GRAPH:
-        raise HTTPException(status_code=404, detail="Graph not found")
-    return SUPPORTED_GRAPH[request.graph_name].get_context_jsonschema()
+import logging
+import uuid
+import json
+import random
+import time
+import re
+
+from fastapi import APIRouter, HTTPException
+
+from pydantic import BaseModel
+
+from typing import Dict, Any, List, Optional
+
+from fastapi.responses import StreamingResponse
+
+from botrun_flow_lang.constants import ERR_GRAPH_RECURSION_ERROR, LANG_EN, LANG_ZH_TW
+
+from botrun_flow_lang.langgraph_agents.agents.agent_runner import (
+    agent_runner,
+    langgraph_runner,
+)
+
+from botrun_flow_lang.langgraph_agents.agents.gov_researcher.gov_researcher_graph import (
+    GovResearcherGraph,
+    get_content_for_gov_researcher,
+)
+from botrun_flow_lang.langgraph_agents.agents.gov_researcher.gov_researcher_2_graph import (
+    TAIWAN_SUBSIDY_SUPERVISOR_PROMPT,
+    create_taiwan_subsidy_agent_graph,
+    taiwan_subsidy_agent_graph,
+)
+from botrun_flow_lang.langgraph_agents.agents.gov_researcher.gemini_subsidy_graph import (
+    TAIWAN_SUBSIDY_SUPERVISOR_PROMPT as GEMINI_SUBSIDY_PROMPT,
+    create_gemini_subsidy_agent_graph,
+    gemini_subsidy_agent_graph,
+)
+from botrun_flow_lang.langgraph_agents.agents.langgraph_react_agent import (
+    create_react_agent_graph,
+    get_react_agent_model_name,
+)
+
+from botrun_flow_lang.langgraph_agents.cache.langgraph_botrun_cache import (
+    get_botrun_cache,
+)
+
+from botrun_flow_lang.models.token_usage import TokenUsage
+
+from botrun_flow_lang.utils.botrun_logger import (
+    get_session_botrun_logger,
+    default_logger,
+)
+
+# 放到要用的時候才 init，不然loading 會花時間
+# 因為要讓 langgraph 在本地端執行，所以這一段又搬回到外面了
+from langgraph.errors import GraphRecursionError
+import anthropic  # Keep relevant imports if needed for error handling here
+
+# ==========
+
+from botrun_flow_lang.langgraph_agents.agents.search_agent_graph import (
+    SearchAgentGraph,
+    # graph as search_agent_graph,
+)
+
+from botrun_flow_lang.utils.langchain_utils import (
+    extract_token_usage_from_state,
+    langgraph_event_to_json,
+    litellm_msgs_to_langchain_msgs,
+)
+
+
+router = APIRouter(prefix="/langgraph")
+
+
+class LangGraphRequest(BaseModel):
+    graph_name: str
+    # todo LangGraph 應該要傳 thread_id，但是因為現在是 cloud run 的架構，所以 thread_id 不一定會讀的到 (auto scale)
+    thread_id: Optional[str] = None
+    user_input: Optional[str] = None
+    messages: List[Dict[str, Any]] = []
+    config: Optional[Dict[str, Any]] = None
+    stream: bool = False
+    # LangGraph 是否需要從 checkpoint 恢復
+    need_resume: bool = False
+    session_id: Optional[str] = None
+
+
+class LangGraphResponse(BaseModel):
+    """
+    @param content: 這個是給評測用來評估結果用的
+    @param state: 這個是graph的 final state，如果需要額外資訊可以使用
+    @param token_usage: Token usage information for the entire graph execution
+    """
+
+    id: str
+    object: str = "chat.completion"
+    created: int
+    model: str
+    content: Optional[str] = None
+    state: Optional[Dict[str, Any]] = None
+    token_usage: Optional[TokenUsage] = None
+
+
+class SupportedGraphsResponse(BaseModel):
+    """Response model for listing supported graphs"""
+
+    graphs: List[str]
+
+
+class GraphSchemaRequest(BaseModel):
+    """Request model for getting graph schema"""
+
+    graph_name: str
+
+
+# 從常數檔案匯入，避免外部模組為了取得常數而觸發重型 import
+from botrun_flow_lang.api.langgraph_constants import (
+    LANGGRAPH_REACT_AGENT,
+    GOV_SUBSIDY_AGENT,
+    PERPLEXITY_SEARCH_AGENT,
+)
+
+# 僅在此檔案內部使用的常數
+CUSTOM_WEB_RESEARCH_AGENT = "custom_web_research_agent"
+DEEP_RESEARCH_AGENT = "deep_research_agent"
+# GOV_RESEARCHER_AGENT = "gov_researcher_agent"
+GEMINI_SUBSIDY_AGENT = "gemini_subsidy_agent"
+
+
+SUPPORTED_GRAPH_NAMES = [
+    # PERPLEXITY_SEARCH_AGENT,
+    LANGGRAPH_REACT_AGENT,
+    GOV_SUBSIDY_AGENT,
+    GEMINI_SUBSIDY_AGENT,
+    # GOV_RESEARCHER_AGENT,
+]
+SUPPORTED_GRAPH = {
+    GOV_SUBSIDY_AGENT: taiwan_subsidy_agent_graph,
+    GEMINI_SUBSIDY_AGENT: gemini_subsidy_agent_graph,
+}
+
+
+def contains_chinese_chars(text: str) -> bool:
+    """Check if the given text contains any Chinese characters."""
+    if not text:
+        return False
+    # This pattern matches Chinese characters (both simplified and traditional)
+    pattern = re.compile(
+        r"[\u4e00-\u9fff\u3400-\u4dbf\U00020000-\U0002a6df\U0002a700-\U0002ebef]"
+    )
+    return bool(pattern.search(text))
+
+
+async def get_cached_or_create_react_graph(
+    botrun_id: Optional[str],  # Key parameter - can be None/empty
+    config: Optional[Dict[str, Any]] = None,
+    messages: Optional[List[Dict]] = None,
+    user_input: Optional[str] = None,
+    logger: logging.Logger = default_logger,
+) -> Any:
+    """
+    Get cached graph for LANGGRAPH_REACT_AGENT or create new one.
+    Handles botrun_id-based caching with parameter validation.
+
+    Args:
+        botrun_id: The botrun ID for cache lookup. If None/empty, skips caching.
+        config: Configuration dictionary
+        messages: List of message dictionaries
+        user_input: User input string
+        logger: Logger instance
+
+    Returns:
+        Graph instance (cached or newly created)
+    """
+    config = config or {}
+
+    # Extract parameters for hash calculation (moved from get_graph)
+    system_prompt = config.get("system_prompt", "")
+    if messages:
+        for message in messages:
+            if message.get("role") == "system":
+                system_prompt = message.get("content", "")
+
+    botrun_flow_lang_url = config.get("botrun_flow_lang_url", "")
+    user_id = config.get("user_id", "")
+    model_name = config.get("model_name", "")
+
+    # Determine language (moved from get_graph)
+    has_chinese = contains_chinese_chars(system_prompt)
+    if not has_chinese and user_input:
+        has_chinese = contains_chinese_chars(user_input)
+    lang = LANG_ZH_TW if has_chinese else LANG_EN
+
+    mcp_config = config.get("mcp_config")
+
+    # CRITICAL: Check if botrun_id is provided and not empty
+    if not botrun_id:
+        # If botrun_id is None or empty, skip caching entirely
+        logger.info("No botrun_id provided, creating new graph without caching")
+        graph = await create_react_agent_graph(
+            system_prompt=system_prompt,
+            botrun_flow_lang_url=botrun_flow_lang_url,
+            user_id=user_id,
+            model_name=model_name,
+            lang=lang,
+            mcp_config=mcp_config,
+        )
+        return graph
+
+    # If botrun_id is provided, use caching logic
+    cache = get_botrun_cache()
+    params_hash = cache.get_params_hash(
+        system_prompt, botrun_flow_lang_url, user_id, model_name, lang, mcp_config
+    )
+
+    # Try to get cached graph
+    cached_graph = cache.get_cached_graph(botrun_id, params_hash)
+    if cached_graph:
+        logger.info(f"Using cached graph for botrun_id: {botrun_id}")
+        return cached_graph
+
+    # Create new graph (same logic as in get_graph)
+    logger.info(f"Creating new graph for botrun_id: {botrun_id}")
+    graph = await create_react_agent_graph(
+        system_prompt=system_prompt,
+        botrun_flow_lang_url=botrun_flow_lang_url,
+        user_id=user_id,
+        model_name=model_name,
+        lang=lang,
+        mcp_config=mcp_config,
+    )
+
+    # Cache the new graph
+    cache.cache_graph(botrun_id, params_hash, graph)
+
+    return graph
+
+
+async def get_graph(
+    graph_name: str,
+    config: Optional[Dict[str, Any]] = None,
+    stream: bool = False,
+    messages: Optional[List[Dict]] = [],
+    user_input: Optional[str] = None,
+):
+    if (
+        graph_name not in SUPPORTED_GRAPH_NAMES
+        and graph_name not in SUPPORTED_GRAPH.keys()
+    ):
+        raise ValueError(f"Unsupported graph from get_graph: {graph_name}")
+    if graph_name == PERPLEXITY_SEARCH_AGENT:
+        graph = SearchAgentGraph().graph
+        graph_config = {
+            "search_prompt": config.get("search_prompt", ""),
+            "model_name": config.get("model_name", "sonar-reasoning-pro"),
+            "related_prompt": config.get("related_question_prompt", ""),
+            "search_vendor": config.get("search_vendor", "perplexity"),
+            "domain_filter": config.get("domain_filter", []),
+            "user_prompt_prefix": config.get("user_prompt_prefix", ""),
+            "stream": stream,
+        }
+    elif graph_name == GOV_SUBSIDY_AGENT:
+        graph = create_taiwan_subsidy_agent_graph(
+            config.get("prompt_template", TAIWAN_SUBSIDY_SUPERVISOR_PROMPT)
+        )
+        graph_config = {
+            "prompt_template": config.get("prompt_template", ""),
+            "legal_extraction_prompt": config.get("legal_extraction_prompt", ""),
+            "faq_extraction_prompt": config.get("faq_extraction_prompt", ""),
+            "calculation_analysis_prompt": config.get(
+                "calculation_analysis_prompt", ""
+            ),
+        }
+    elif graph_name == GEMINI_SUBSIDY_AGENT:
+        graph = create_gemini_subsidy_agent_graph(
+            config.get("prompt_template", GEMINI_SUBSIDY_PROMPT)
+        )
+        graph_config = {
+            "prompt_template": config.get("prompt_template", ""),
+        }
+    else:
+        raise ValueError(f"Unsupported graph type: {graph_name}")
+    return graph, graph_config
+
+
+def get_init_state(
+    graph_name: str,
+    user_input: str,
+    config: Optional[Dict[str, Any]] = None,
+    messages: Optional[List[Dict]] = [],
+    enable_prompt_caching: bool = False,
+):
+    if graph_name == PERPLEXITY_SEARCH_AGENT:
+        if len(messages) > 0:
+            return {"messages": litellm_msgs_to_langchain_msgs(messages)}
+        if config.get("user_prompt_prefix", ""):
+            return {
+                "messages": [
+                    {
+                        "role": "user",
+                        "content": config.get("user_prompt_prefix", "")
+                        + "\n\n"
+                        + user_input,
+                    }
+                ]
+            }
+
+        return {"messages": [user_input]}
+    elif graph_name == CUSTOM_WEB_RESEARCH_AGENT:
+        if len(messages) > 0:
+            return {
+                "messages": litellm_msgs_to_langchain_msgs(messages),
+                "model": config.get("model", "anthropic"),
+            }
+        return {
+            "messages": [user_input],
+            "model": config.get("model", "anthropic"),
+        }
+    elif graph_name == LANGGRAPH_REACT_AGENT:
+        if len(messages) > 0:
+            new_messages = []
+            for message in messages:
+                if message.get("role") != "system":
+                    new_messages.append(message)
+
+            return {
+                "messages": litellm_msgs_to_langchain_msgs(
+                    new_messages, enable_prompt_caching
+                )
+            }
+        else:
+            return {
+                "messages": [user_input],
+            }
+    elif graph_name == DEEP_RESEARCH_AGENT:
+        if len(messages) > 0:
+            return {
+                "messages": litellm_msgs_to_langchain_msgs(messages),
+                "topic": user_input,
+            }
+        return {
+            "messages": [user_input],
+            "topic": user_input,
+        }
+    else:
+        if len(messages) > 0:
+            return {"messages": litellm_msgs_to_langchain_msgs(messages)}
+        return {"messages": [user_input]}
+
+
+def get_content(graph_name: str, state: Dict[str, Any]):
+    if graph_name == PERPLEXITY_SEARCH_AGENT:
+        return state["messages"][-3].content
+    elif graph_name == CUSTOM_WEB_RESEARCH_AGENT:
+        content = state["answer"].get("markdown", "")
+        content = content.replace("\\n", "\n")
+        if state["answer"].get("references", []):
+            references = "\n\n參考資料：\n"
+            for reference in state["answer"]["references"]:
+                references += f"- [{reference['title']}]({reference['url']})\n"
+            content += references
+        return content
+    elif graph_name == DEEP_RESEARCH_AGENT:
+        sections = state["sections"]
+        sections_str = "\n\n".join(
+            f"章節: {section.name}\n"
+            f"描述: {section.description}\n"
+            f"需要研究: {'是' if section.research else '否'}\n"
+            for section in sections
+        )
+        sections_str = "預計報告結構：\n\n" + sections_str
+        return sections_str
+    # elif graph_name == GOV_RESEARCHER_AGENT:
+    #     return get_content_for_gov_researcher(state)
+    elif graph_name == GOV_SUBSIDY_AGENT or graph_name == GEMINI_SUBSIDY_AGENT:
+        messages = state["messages"]
+        # Find the last AI message
+        for msg in reversed(messages):
+            if msg.type == "ai":
+                if isinstance(msg.content, list):
+                    return msg.content[0].get("text", "")
+                else:
+                    return msg.content
+        return ""  # If no AI message found
+    else:
+        messages = state["messages"]
+        # Find the last human message
+        last_human_idx = -1
+        for i, msg in enumerate(messages):
+            if msg.type == "human":
+                last_human_idx = i
+
+        # Combine all AI messages after the last human message
+        ai_contents = ""
+        for msg in messages[last_human_idx + 1 :]:
+            if msg.type == "ai":
+                if isinstance(msg.content, list):
+                    ai_contents += msg.content[0].get("text", "")
+                else:
+                    ai_contents += msg.content
+
+        return ai_contents
+
+
+async def process_langgraph_request(
+    request: LangGraphRequest,
+    retry: bool = False,  # Keep retry logic for non-streaming path if needed
+    logger: logging.Logger = default_logger,
+) -> Any:  # Return type can be LangGraphResponse or StreamingResponse
+    """處理 LangGraph 請求的核心邏輯"""
+    # --- Streaming Case ---
+    if request.stream:
+        logger.info(f"Processing STREAM request for graph: {request.graph_name}")
+        # Use the new wrapper generator that handles resource management
+        return StreamingResponse(
+            managed_langgraph_stream_wrapper(request, logger),
+            media_type="text/event-stream",
+        )
+
+    # --- Non-Streaming Case ---
+    logger.info(f"Processing NON-STREAM request for graph: {request.graph_name}")
+    try:
+        config = request.config or {}
+        mcp_config = config.get("mcp_config")
+        user_id = config.get("user_id")
+
+        # Get botrun_id from config
+        botrun_id = config.get("botrun_id")  # Can be None/empty
+
+        # --- Graph and State Initialization (OUTSIDE of MCP client context) ---
+        # Cache logic for LANGGRAPH_REACT_AGENT only
+        if request.graph_name == LANGGRAPH_REACT_AGENT:
+            graph = await get_cached_or_create_react_graph(
+                botrun_id=botrun_id,  # Pass botrun_id (can be None)
+                config=request.config,
+                messages=request.messages,
+                user_input=request.user_input,
+                logger=logger,
+            )
+            graph_config = request.config
+        else:
+            # Existing logic for other graph types (calls modified get_graph)
+            graph, graph_config = await get_graph(
+                request.graph_name,
+                request.config,
+                False,  # stream=False
+                request.messages,
+                request.user_input,
+            )
+
+        # Determine model name for init_state caching logic if needed
+        # user_input_model_name = request.config.get("model_name", "")
+        # enable_caching = get_react_agent_model_name(user_input_model_name).startswith("claude-")
+
+        init_state = get_init_state(
+            request.graph_name,
+            request.user_input,
+            request.config,
+            request.messages,
+            False,  # enable_prompt_caching=enable_caching
+        )
+
+        thread_id = request.thread_id if request.thread_id else str(uuid.uuid4())
+        logger.info(f"Running non-stream with thread_id: {thread_id}")
+
+        # --- Run the agent (no MCP client needed during execution) ---
+        logger.info("Executing agent_runner for non-stream request...")
+        async for _ in agent_runner(
+            thread_id,
+            init_state,
+            graph,
+            request.need_resume,
+            extra_config=graph_config,
+        ):
+            pass  # Just consume the events
+
+        logger.info(
+            f"agent_runner completed for thread_id: {thread_id}. Fetching final state."
+        )
+
+        # --- Get Final State and Prepare Response (OUTSIDE of MCP client context) ---
+        config_for_state = {"configurable": {"thread_id": thread_id}}
+        state = await graph.aget_state(config_for_state)
+
+        try:
+            state_values_json = langgraph_event_to_json(state.values)
+            logger.info(
+                f"Final state fetched for {thread_id}: {state_values_json[:500]}..."
+            )  # Log truncated state
+        except Exception as e_log:
+            logger.error(f"Error serializing final state for logging: {e_log}")
+            logger.info(
+                f"Final state keys for {thread_id}: {list(state.values.keys())}"
+            )
+
+        content = get_content(request.graph_name, state.values)
+
+        model_name_config = (
+            request.config.get("model_name", "") if request.config else ""
+        )
+        final_model_name = model_name_config  # Default to config model name
+        if request.graph_name == LANGGRAPH_REACT_AGENT:
+            final_model_name = get_react_agent_model_name(model_name_config)
+            token_usage = extract_token_usage_from_state(state.values, final_model_name)
+        else:
+            token_usage = TokenUsage(
+                total_input_tokens=0,
+                total_output_tokens=0,
+                total_tokens=0,
+                nodes=[],
+            )
+
+        return LangGraphResponse(
+            id=thread_id,
+            created=int(time.time()),
+            model=request.graph_name,  # Or final_model_name? Check requirements
+            content=content,
+            state=state.values,  # Consider serializing state here if needed client-side
+            token_usage=token_usage,
+        )
+
+    except anthropic.RateLimitError as e:
+        if retry:
+            logger.error(
+                "Retry failed with Anthropic RateLimitError (non-stream)", exc_info=True
+            )
+            raise HTTPException(
+                status_code=429, detail=f"Rate limit exceeded after retry: {e}"
+            )  # 429 is more appropriate
+
+        logger.warning(
+            f"Anthropic RateLimitError occurred (non-stream): {e}. Retrying..."
+        )
+        retry_delay = random.randint(7, 20)
+        time.sleep(
+            retry_delay
+        )  # Note: time.sleep blocks async. Consider asyncio.sleep(retry_delay) if this becomes an issue.
+        logger.info(f"Retrying non-stream request after {retry_delay}s delay...")
+        return await process_langgraph_request(
+            request, retry=True, logger=logger
+        )  # Recursive call for retry
+
+    except GraphRecursionError as e:
+        logger.error(f"GraphRecursionError (non-stream): {e}", exc_info=True)
+        raise HTTPException(
+            status_code=500, detail=f"Graph execution exceeded maximum depth: {e}"
+        )
+
+    except Exception as e:
+        import traceback
+
+        tb_str = traceback.format_exc()
+        logger.error(
+            f"Unhandled exception in process_langgraph_request (non-stream): {e}",
+            exc_info=True,
+        )
+        raise HTTPException(
+            status_code=500, detail=f"Internal Server Error during graph execution: {e}"
+        )
+
+
+@router.post("/invoke")
+async def invoke(request: LangGraphRequest):
+    """
+    執行指定的 LangGraph，支援串流和非串流模式
+
+    Args:
+        request: 包含 graph_name 和輸入數據的請求
+
+    Returns:
+        串流模式: StreamingResponse
+        非串流模式: LangGraphResponse
+    """
+    session_id = request.session_id
+    user_id = request.config.get("user_id", "")
+
+    # *** Create a session-specific BotrunLogger for this specific request ***
+    # This ensures Cloud Logging and session/user context
+    logger = get_session_botrun_logger(session_id=session_id, user_id=user_id)
+
+    logger.info(
+        "invoke LangGraph API",
+        request=request.model_dump(),
+    )
+
+    # Pass the request-specific BotrunLogger down
+    return await process_langgraph_request(request, logger=logger)
+
+
+# NEW: Wrapper generator for managing resources during streaming
+async def managed_langgraph_stream_wrapper(
+    request: LangGraphRequest, logger: logging.Logger
+):
+    """
+    Manages AsyncExitStack and MCPClient lifecycle for streaming responses.
+    Initializes graph and then yields events from langgraph_stream_response_generator.
+    """
+    try:
+        config = request.config or {}
+        mcp_config = config.get("mcp_config")
+        user_id = config.get("user_id")
+        print(f"mcp_config: {mcp_config}, user_id: {user_id}")
+
+        # Get botrun_id from config
+        botrun_id = config.get("botrun_id")  # Can be None/empty
+
+        # --- Graph and State Initialization (OUTSIDE of MCP client context) ---
+        logger.info("Getting graph and initial state for stream...")
+        # Cache logic for LANGGRAPH_REACT_AGENT only
+        if request.graph_name == LANGGRAPH_REACT_AGENT:
+            graph = await get_cached_or_create_react_graph(
+                botrun_id=botrun_id,  # Pass botrun_id (can be None)
+                config=request.config,
+                messages=request.messages,
+                user_input=request.user_input,
+                logger=logger,
+            )
+            graph_config = request.config
+        else:
+            # Existing logic for other graph types (calls modified get_graph)
+            graph, graph_config = await get_graph(
+                request.graph_name,
+                request.config,
+                request.stream,  # Pass stream=True
+                request.messages,
+                request.user_input,
+            )
+
+        # Determine model name for init_state caching logic if needed
+        # user_input_model_name = request.config.get("model_name", "")
+        # enable_caching = get_react_agent_model_name(user_input_model_name).startswith("claude-") # Example
+
+        init_state = get_init_state(
+            request.graph_name,
+            request.user_input,
+            request.config,
+            request.messages,
+            False,  # enable_prompt_caching=enable_caching # Pass caching flag if used
+        )
+
+        thread_id = request.thread_id if request.thread_id else str(uuid.uuid4())
+        logger.info(f"Streaming with thread_id: {thread_id}")
+
+        # --- Yield from the actual stream response generator ---
+        async for event in langgraph_stream_response_generator(
+            thread_id,
+            init_state,
+            graph,
+            request.need_resume,
+            logger,
+            graph_config,
+        ):
+            yield event  # Yield the formatted event string
+
+    except anthropic.RateLimitError as e:
+        # Handle rate limit errors specifically for streaming if needed
+        # Note: Retry logic might be complex to implement correctly within a generator.
+        #       Consider if retry should happen at a higher level or if yielding an error is sufficient.
+        logger.error(
+            f"Anthropic RateLimitError during stream setup/execution: {e}",
+            exc_info=True,
+        )
+        error_payload = json.dumps(
+            {"error": f"Rate Limit Error: {e}", "retry_possible": False}
+        )  # Indicate no auto-retry here
+        yield f"data: {error_payload}\n\n"
+        yield "data: [DONE]\n\n"  # Ensure stream terminates correctly
+
+    except GraphRecursionError as e:
+        # Handle recursion errors specifically (can happen during graph execution)
+        logger.error(
+            f"GraphRecursionError during stream: {e} for thread_id: {thread_id}",
+            error=str(e),
+            exc_info=True,
+        )
+        try:
+            error_msg = json.dumps(
+                {"error": ERR_GRAPH_RECURSION_ERROR, "detail": str(e)}
+            )
+            yield f"data: {error_msg}\n\n"
+        except Exception as inner_e:
+            logger.error(
+                f"Error serializing GraphRecursionError for stream: {inner_e}",
+                exc_info=True,
+            )
+            yield f"data: {json.dumps({'error': ERR_GRAPH_RECURSION_ERROR})}\n\n"
+        yield "data: [DONE]\n\n"  # Ensure stream terminates correctly
+
+    except Exception as e:
+        # Catch-all for other errors during setup or streaming
+        import traceback
+
+        tb_str = traceback.format_exc()
+        logger.error(
+            f"Unhandled exception in managed_langgraph_stream_wrapper: {e}",
+            exc_info=True,
+            traceback=tb_str,
+        )
+        error_payload = json.dumps({"error": f"Streaming Error: {e}", "detail": tb_str})
+        yield f"data: {error_payload}\n\n"
+        yield "data: [DONE]\n\n"  # Ensure stream terminates correctly
+
+
+# RENAMED: Original langgraph_stream_response, now focused on generation
+async def langgraph_stream_response_generator(
+    thread_id: str,
+    init_state: Dict,
+    graph: Any,  # Receives the already configured graph
+    need_resume: bool = False,
+    logger: logging.Logger = default_logger,
+    extra_config: Optional[Dict] = None,
+):
+    """
+    Generates LangGraph stream events using langgraph_runner.
+    Handles formatting ('data: ...') and '[DONE]' signal.
+    Exception handling specific to langgraph_runner execution.
+    """
+    try:
+        logger.info(
+            "Starting langgraph_runner iteration",
+            thread_id=thread_id,
+            need_resume=need_resume,
+        )
+
+        final_event = None
+        first_event = True  # To potentially log first event differently if needed
+        async for event in langgraph_runner(
+            thread_id, init_state, graph, need_resume, extra_config
+        ):
+            final_event = event  # Keep track of the last event
+            event_json_str = langgraph_event_to_json(event)  # Serialize event safely
+            if first_event:
+                # Optional: Different logging for the very first event chunk
+                logger.info(
+                    f"First stream event for {thread_id}: {event_json_str[:200]}..."
+                )  # Log truncated first event
+                first_event = False
+
+            # print statement for local debugging if needed
+            # from datetime import datetime
+            # print(datetime.now().strftime("%Y-%m-%d %H:%M:%S"), event_json_str)
+
+            yield f"data: {event_json_str}\n\n"
+
+        # Log details about the final event if needed
+        if final_event:
+            logger.info(
+                "Finished langgraph_runner iteration",
+                thread_id=thread_id,
+                final_event_type=final_event.get("event"),
+                # final_event_data=langgraph_event_to_json(final_event) # Log full final event if useful
+            )
+        else:
+            logger.warning(
+                "langgraph_runner finished without yielding any events",
+                thread_id=thread_id,
+            )
+
+        yield "data: [DONE]\n\n"  # Signal end of stream
+
+    # Error handling remains here as these errors occur during langgraph_runner
+    except GraphRecursionError as e:
+        logger.error(
+            f"GraphRecursionError in stream generator: {e} for thread_id: {thread_id}",
+            error=str(e),
+            exc_info=True,
+        )
+        try:
+            error_msg = json.dumps(
+                {"error": ERR_GRAPH_RECURSION_ERROR, "detail": str(e)}
+            )
+            yield f"data: {error_msg}\n\n"
+        except Exception as inner_e:
+            logger.error(
+                f"Error serializing GraphRecursionError msg: {inner_e}", exc_info=True
+            )
+            yield f"data: {json.dumps({'error': ERR_GRAPH_RECURSION_ERROR})}\n\n"
+        # Ensure [DONE] is sent even after handled error to terminate client side
+        yield "data: [DONE]\n\n"
+
+    except Exception as e:
+        # Catch errors specifically from langgraph_runner or event processing
+        import traceback
+
+        tb_str = traceback.format_exc()
+        logger.error(
+            f"Exception in stream generator: {e} for thread_id: {thread_id}",
+            error=str(e),
+            exc_info=True,
+            traceback=tb_str,
+        )
+        error_response = {"error": f"Stream Generation Error: {e}", "detail": tb_str}
+        yield f"data: {json.dumps(error_response)}\n\n"
+        # Ensure [DONE] is sent even after handled error
+        yield "data: [DONE]\n\n"
+
+
+@router.get("/list", response_model=SupportedGraphsResponse)
+async def list_supported_graphs():
+    """
+    列出所有支援的 LangGraph names
+
+    Returns:
+        包含所有支援的 graph names 的列表
+    """
+    return SupportedGraphsResponse(graphs=list(SUPPORTED_GRAPH.keys()))
+
+
+@router.post("/schema", response_model=dict)
+async def get_graph_schema(request: GraphSchemaRequest):
+    """
+    取得指定 graph 的 schema
+    """
+    if request.graph_name not in SUPPORTED_GRAPH:
+        raise HTTPException(status_code=404, detail="Graph not found")
+    return SUPPORTED_GRAPH[request.graph_name].get_context_jsonschema()