fix: 修复 setup_agent 创建专家流程的多个问题

后端 (setup_agent_approval_middleware.py)：
- 拦截策略从 wrap_tool_call + Command(goto=END) 改为
  after_model + jump_to="end"。在 tool_node 执行前直接终止图，
  避免模型并行调多个 tool_call 时 Command 输出被稀释、graph
  路由不稳定的问题（症状：审批 ToolMessage 发出后流不结束）。
- 用 @hook_config(can_jump_to=["end"]) 装饰器声明跳转能力，
  否则 LangChain 不会为该 hook 构造 conditional edge。

前端 (NewAgentPage.tsx)：
- 聊天阶段不再调 ensureAgent / POST /api/agents，placeholder
  agent 不再落库；agent 仅在用户点右上角“保存”时创建。
- 自动生成的占位名 "未命名专家-xxx" 不再回写到输入框，保持空白
  直到用户输入或 setup_agent 审批回填。
- handleApproveSetupAgent 仅把模型 propose 的 payload 填入表单，
  不再 PUT /api/agents/{id}；落库交给手动保存按钮。
- onToolEnd 不再 getAgentWithRetry —— 工具被 middleware 拦截、
  根本不会执行，没有 agent 状态可拉。清理对应死代码（getAgent
  import、wait()、AGENT_READ_RETRY_DELAYS_MS、getAgentWithRetry）。
- 路由跳转补齐 /page 前缀：返回按钮、“开始对话”按钮。
- 右侧表单：SOUL.md textarea 加 max-h-[400px]，技能列表外框加
  max-h-[240px]，各自 overflow-y-auto，避免长内容撑高整列。

前端 (agent-card.tsx)：
- 列表卡片“聊天”按钮的跳转路径补齐 /page 前缀，避免点击后 404。

无新依赖。

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/setup_agent_approval_middleware.py

@@ -3,16 +3,13 @@
 from __future__ import annotations
 
 import logging
-from collections.abc import Callable
 from hashlib import sha256
 from typing import Any, override
 
 from langchain.agents import AgentState
-from langchain.agents.middleware import AgentMiddleware
-from langchain_core.messages import ToolMessage
-from langgraph.graph import END
-from langgraph.prebuilt.tool_node import ToolCallRequest
-from langgraph.types import Command
+from langchain.agents.middleware import AgentMiddleware, hook_config
+from langchain_core.messages import AIMessage, ToolMessage
+from langgraph.runtime import Runtime
 
 logger = logging.getLogger(__name__)
 
@@ -26,15 +23,22 @@ class SetupAgentApprovalMiddlewareState(AgentState):
 class SetupAgentApprovalMiddleware(AgentMiddleware[SetupAgentApprovalMiddlewareState]):
     """Stop before ``setup_agent`` writes files and ask the user to approve.
 
-    The first ``setup_agent`` call is converted into a structured tool message
-    and the run ends. A subsequent request can set ``setup_agent_approved`` in
-    runtime context to allow the tool to execute.
+    The interception runs in ``after_model``: as soon as the model emits an
+    AIMessage that contains a ``setup_agent`` tool call (and the run has not
+    been pre-approved via ``setup_agent_approved=True`` in runtime context),
+    we append a structured approval ``ToolMessage`` and return
+    ``jump_to="end"`` to terminate the LangGraph run *before* the tool node
+    executes. This is the LangChain-recommended way to short-circuit after
+    the model; using ``wrap_tool_call`` + ``Command(goto=END)`` proved
+    unreliable when the AIMessage contained multiple parallel tool calls
+    (the ``Command`` got mixed with normal tool outputs and the graph
+    looped back to the model instead of terminating).
     """
 
     state_schema = SetupAgentApprovalMiddlewareState
 
-    def _is_approved(self, request: ToolCallRequest) -> bool:
-        context = getattr(request.runtime, "context", None) or {}
+    def _is_approved(self, runtime: Runtime) -> bool:
+        context = getattr(runtime, "context", None) or {}
         if not isinstance(context, dict):
             return False
         return context.get("setup_agent_approved") is True
@@ -54,37 +58,57 @@ class SetupAgentApprovalMiddleware(AgentMiddleware[SetupAgentApprovalMiddlewareS
             "skills": skills if isinstance(skills, list) else [],
         }
 
-    def _handle_setup_agent(self, request: ToolCallRequest) -> Command:
-        args = request.tool_call.get("args", {}) or {}
-        tool_call_id = request.tool_call.get("id", "")
+    def _intercept(self, last: AIMessage) -> dict[str, Any] | None:
+        """Return a state update if setup_agent must be gated; None to passthrough."""
+        tool_calls = list(last.tool_calls or [])
+        setup_call = next((tc for tc in tool_calls if tc.get("name") == "setup_agent"), None)
+        if setup_call is None:
+            return None
+
+        args = setup_call.get("args") or {}
+        tool_call_id = setup_call.get("id") or ""
         payload = self._build_payload(args)
         logger.info("Setup agent approval requested before writing agent files")
 
-        message = ToolMessage(
+        approval_msg = ToolMessage(
             id=self._stable_message_id(tool_call_id, args),
             content="请确认是否创建/更新这个专家配置。",
             tool_call_id=tool_call_id,
             name="setup_agent",
             additional_kwargs={"setup_agent_approval": payload},
         )
-        return Command(update={"messages": [message]}, goto=END)
 
-    @override
-    def wrap_tool_call(
-        self,
-        request: ToolCallRequest,
-        handler: Callable[[ToolCallRequest], ToolMessage | Command],
-    ) -> ToolMessage | Command:
-        if request.tool_call.get("name") != "setup_agent" or self._is_approved(request):
-            return handler(request)
-        return self._handle_setup_agent(request)
+        # Any other tool_calls on the same AIMessage are left dangling; the
+        # existing DanglingToolCallMiddleware will inject placeholder
+        # ToolMessages on the next model invocation, which matches the
+        # desired semantics: every action pauses until the user approves.
+        return {
+            "messages": [approval_msg],
+            "jump_to": "end",
+        }
 
+    @hook_config(can_jump_to=["end"])
     @override
-    async def awrap_tool_call(
+    def after_model(
         self,
-        request: ToolCallRequest,
-        handler: Callable[[ToolCallRequest], ToolMessage | Command],
-    ) -> ToolMessage | Command:
-        if request.tool_call.get("name") != "setup_agent" or self._is_approved(request):
-            return await handler(request)
-        return self._handle_setup_agent(request)
+        state: SetupAgentApprovalMiddlewareState,
+        runtime: Runtime,
+    ) -> dict[str, Any] | None:
+        if self._is_approved(runtime):
+            return None
+        messages = state.get("messages") or []
+        if not messages:
+            return None
+        last = messages[-1]
+        if not isinstance(last, AIMessage):
+            return None
+        return self._intercept(last)
+
+    @hook_config(can_jump_to=["end"])
+    @override
+    async def aafter_model(
+        self,
+        state: SetupAgentApprovalMiddlewareState,
+        runtime: Runtime,
+    ) -> dict[str, Any] | None:
+        return self.after_model(state, runtime)