bugfix：修复部分提示词导致的格式错误问题及langfuse客户端复用bug

app/observability/tracing.py

@@ -77,7 +77,8 @@ async def ainvoke_screen(
     from langfuse.langchain import CallbackHandler
     from langgraph.types import Command
 
-    handler = CallbackHandler()
+    # 明确回调所属项目，避免进程中存在多个客户端时跳过追踪
+    handler = CallbackHandler(public_key=os.getenv("LANGFUSE_PUBLIC_KEY"))
     # 在当前异步上下文执行回调，手动记录的 Jev generation 才会挂到所属节点。
     handler.run_inline = True
     resume = isinstance(payload, Command) and payload.resume is not None
@@ -156,16 +157,16 @@ def observe_decision(request: Any) -> Iterator[Any]:
     只有图回调已经挂上当前 span 时才记。没有父 span 就跳过，避免 Jev 单独变成一条 trace
     这里不吞异常：失败时标 ERROR 后原样抛出，调用方的 DecisionError 类型不变
     """
-    if not tracing_enabled() or _ensure_client() is None or not _has_active_span():
+    client = _ensure_client() if tracing_enabled() else None
+    if client is None or not _has_active_span():
         yield None
         return
 
-    from langfuse import get_client
-
     model = request.model or settings.typesafe_model
     # 意图判断用固定名字，方便以后按名字做评测。别的决策题不要都叫 classify-intent
     name = "classify-intent" if set(request.decisions) == {"intent"} else "run-decision"
-    with get_client().start_as_current_observation(
+    # 复用本应用已经选定的客户端
+    with client.start_as_current_observation(
         as_type="generation",
         name=name,
         model=model,