修复总控智能体共识后仍被错判为继续派活的死循环

_find_dispatch_message 原实现会一路向前翻历史,直到撞到任意一条带
tool_calls 的 AIMessage,导致本轮 LLM 即使共识(tool_calls=[]),也会
被上一轮的派活 AI 翻出来再用一次,前端因此把同样的 task 又发给子智能体,
形成"嘴上共识、手上派活"的死循环。

改成只看最新一条 AIMessage(跳过末尾 skill_stop / clarification ack
ToolMessage 即可),让 tool_calls 是否为空作为本轮的唯一权威信号。

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

.claude/settings.json

@@ -0,0 +1,10 @@
+{
+  "permissions": {
+    "allow": [
+      "PowerShell(uv --version)",
+      "PowerShell(netstat -ano)",
+      "PowerShell(Get-NetTCPConnection -LocalPort 8001 -ErrorAction SilentlyContinue)",
+      "PowerShell(Get-Process -Id 20560 -ErrorAction SilentlyContinue | Select-Object Id,ProcessName,Path,StartTime; try { \\(Invoke-WebRequest -Uri \"http://127.0.0.1:8001/health\" -UseBasicParsing -TimeoutSec 5\\).Content } catch { \"health check failed: $_\" })"
+    ]
+  }
+}

offline-backend-20260512/backend/app/gateway/routers/multi_agent.py

@@ -764,25 +764,30 @@ def _find_trailing_ai_text(messages: list[dict[str, Any]], *, exclude_text: str
 
 
 def _find_dispatch_message(messages: list[dict[str, Any]]) -> dict[str, Any]:
-    """Locate the latest AIMessage that carries dispatch tool calls.
+    """Locate the latest AIMessage from THIS turn.
 
-    The very last message can be a skill-stop ToolMessage injected by
-    ``SkillStopMiddleware`` when it intercepts the dispatch, or a clarification
-    ToolMessage. Either way, the actionable tool_calls live on the AIMessage
-    *before* it, so we walk backwards until we hit one with ``tool_calls``.
-    Returns the latest AIMessage when none of them carry tool_calls so the
-    caller can still extract its text content.
+    Trailing ToolMessages (skill_stop / clarification ack injected by
+    ``SkillStopMiddleware`` or ``ClarificationMiddleware``) are skipped because
+    the actionable tool_calls live on the AIMessage that emitted them. We stop
+    at the **first AIMessage encountered** — its ``tool_calls`` (or lack
+    thereof) is the authoritative signal for this turn.
+
+    历史踩坑:之前实现是"一直往前翻直到撞到带 tool_calls 的 AI",看似优雅,
+    实际上让 leader 永远收敛不了:本轮 LLM 不论怎么共识,只要历史上派过活,
+    上一轮的派活 AI 就会被翻出来再用一次,前端会把同样的 task 又发给子智能体,
+    陷入死循环。现在停在最新一条 AI 上,把判定交给调用方:`tool_calls=[]` →
+    共识分支,有 tool_calls → 派活分支。
     """
-    last_ai: dict[str, Any] = {}
     for msg in reversed(messages):
         if not isinstance(msg, dict):
             continue
-        if (msg.get("type") or "").lower() in ("ai", "aimessage", "aimessagechunk"):
-            if not last_ai:
-                last_ai = msg
-            if msg.get("tool_calls"):
+        msg_type = (msg.get("type") or "").lower()
+        if msg_type in ("tool", "toolmessage"):
+            continue  # 跳过末尾 skill_stop / clarification ack
+        if msg_type in ("ai", "aimessage", "aimessagechunk"):
             return msg
-    return last_ai
+        break  # 撞到 Human/其他类型,本轮没 AI(异常路径,留空让上游报错)
+    return {}
 
 
 async def _leader_run(