feat：完成重构和测试

app/processor/history.py

@@ -0,0 +1,39 @@
+"""会话历史：数据库保存完整对话，每次请求按 token 预算只带最近的完整轮次
+
+裁剪在服务层组装请求时做一次，意图判断、问答和 screen agent 看到的是同一段历史
+"""
+
+from langchain_core.messages import (
+    AIMessage,
+    BaseMessage,
+    HumanMessage,
+    trim_messages,
+)
+from langchain_core.messages.utils import count_tokens_approximately
+
+from app.schemas.base import JsonObject
+
+
+def history_messages(history: list[JsonObject]) -> list[BaseMessage]:
+    """每轮转成一条用户消息和一条助手消息，按时间从早到晚排列"""
+    messages: list[BaseMessage] = []
+    for turn in history:
+        messages.append(HumanMessage(content=turn["user"]))
+        messages.append(AIMessage(content=turn["assistant"]))
+    return messages
+
+
+def trim_history(history: list[JsonObject], max_tokens: int) -> list[JsonObject]:
+    """从最近一轮往前保留，总量不超过 max_tokens；只保留完整的轮次，不从一段回复中间截断
+
+    单独一轮就超过预算时这一轮也不保留
+    """
+    kept = trim_messages(
+        history_messages(history),
+        max_tokens=max_tokens,
+        token_counter=count_tokens_approximately,
+        strategy="last",
+        # 从用户消息开始，助手回复放不下时连同它的用户消息一起去掉
+        start_on="human",
+    )
+    return history[len(history) - len(kept) // 2 :]

app/processor/nodes/intent_node.py

@@ -14,8 +14,10 @@ class IntentNode(BaseNode):
 
     async def process(self, state: QueryState) -> QueryState:
         updated_state: QueryState = state.copy()
+        context = state["input_context"]
         request = DecisionRequest(
-            state=request_summary(state["input_context"]),
+            # Jev 只接收 JSON，会话历史作为 history 字段和本次请求放在一起
+            state={"history": context.history, **request_summary(context)},
             decisions={"intent": build_intent_question()},
         )
         response = await get_decision_engine().run(request)

app/processor/nodes/qa_node.py

@@ -6,6 +6,7 @@ from langgraph.config import get_stream_writer
 
 from app.clients.llm import get_llm
 from app.processor.base import BaseNode
+from app.processor.history import history_messages
 from app.processor.state import QueryState
 from app.prompt.qa import QA_SYSTEM_PROMPT
 
@@ -24,11 +25,16 @@ class QANode(BaseNode):
 
         parser = StrOutputParser()
         parts: list[str] = []
+        context = state["input_context"]
         async for chunk in get_llm().astream(
             [
                 SystemMessage(content=QA_SYSTEM_PROMPT),
+                # 会话历史作为此前的对话消息，本次请求是最后一条用户消息
+                *history_messages(context.history),
                 HumanMessage(
-                    content=state["input_context"].model_dump_json(exclude_unset=True)
+                    content=context.model_dump_json(
+                        exclude_unset=True, exclude={"history"}
+                    )
                 ),
             ]
         ):

app/processor/nodes/screen_agent_node.py

@@ -10,6 +10,7 @@ from langchain_core.messages import HumanMessage
 from app.agent.screen_agent import AgentReply, get_screen_agent
 from app.agent.tools import ScreenAgentContext
 from app.processor.base import BaseNode
+from app.processor.history import history_messages
 from app.processor.state import QueryState
 from app.processor.summary import request_summary
 from app.resources.component import get_catalog_cache
@@ -55,10 +56,12 @@ class ScreenAgentNode(BaseNode):
         agent = await get_screen_agent()
         result = await agent.ainvoke(
             {
+                # 会话历史作为此前的对话消息，本次请求是最后一条用户消息
                 "messages": [
+                    *history_messages(context.history),
                     HumanMessage(
                         content=json.dumps(request_summary(context), ensure_ascii=False)
-                    )
+                    ),
                 ]
             },
             context=ScreenAgentContext(screen=context.screen, catalog=catalog),

app/processor/summary.py

@@ -1,11 +1,14 @@
-"""请求摘要：意图判断和 screen agent 共用，只给组件的识别信息和布局，完整配置按需读取"""
+"""请求摘要：意图判断和 screen agent 共用，只给组件的识别信息和布局，完整配置按需读取
+
+只描述本次请求；会话历史由各使用方另行传入，见 history.py
+"""
 
 from app.schemas.base import JsonObject
 from app.schemas.input_context import InputContext
 
 
 def request_summary(context: InputContext) -> JsonObject:
-    """用户输入、会话历史、澄清记录、编辑上下文、附件和大屏摘要"""
+    """用户输入、澄清记录、编辑上下文、附件和大屏摘要"""
     nodes: list[JsonObject] = []
     for node in context.screen.nodes:
         # 图表标题帮助判断用户指的是哪个组件；没有 option 或标题格式不符时按无标题处理
@@ -28,7 +31,6 @@ def request_summary(context: InputContext) -> JsonObject:
 
     return {
         "prompt": context.prompt,
-        "history": context.history,
         "clarification_history": context.clarification_history,
         "context": context.context.model_dump(mode="json", exclude_unset=True),
         "screen_summary": {