feat(v3): T26 两级智能体（大模型规划/审查 + 本地小模型执行，D7）

- run_dual 编排：规划者两阶段 JSON（plan/review，失败回喂重试一次再降级），
  redo 时裁决意见回喂执行者，交接上限 agent.max_handoffs（默认 2）
- 交接文档 agent_runs/{id}/handoff.json（智能体版交流文本：instructions/acceptance/exchanges）
- /agent 新增 executor_pool_id；规划者==执行者条目拒绝；整体 token_cap 覆盖两级调用
- ToolLoop 增 emit_final 开关（内层循环不发终态，防前端 SSE 提前收口）
- 前端：执行者选择器 + phase/message 事件渲染（阶段徽标 + 双色消息卡）
- 测试 +3（done/redo/执行者故障），全量 277 passed
- fix(tests): test_config_get_put_reset 增加设置备份/恢复隔离，防止清掉用户真实配置

gateway/api.py

@@ -530,13 +530,44 @@ try:
             workspace_dir = agent_cfg.get("workspace_dir", "agent_workspace")
 
         chat, model, used_pool_id = _resolve_agent_chat(pool_id)
+
+        # 两级模式（D7）：显式指定执行者（本地小模型）时，规划=chat、执行=executor_chat
+        executor_pool_id = str((req or {}).get("executor_pool_id") or "").strip()
+        executor_chat = None
+        executor_model = ""
+        if executor_pool_id:
+            entry = get_pool().get(executor_pool_id)
+            if entry is None:
+                raise HTTPException(status_code=400,
+                                    detail=f"执行者条目不存在: {executor_pool_id}")
+            if entry["backend"] == "mock":
+                raise HTTPException(status_code=400,
+                                    detail="mock 模型不能担任执行者，请选择 llama_server 或 openai 条目")
+            from gateway.agent import OpenAICompatChat
+            executor_chat = OpenAICompatChat(
+                base_url=entry["base_url"] or "http://127.0.0.1:8901/v1",
+                api_key=entry.get("api_key") or None,
+                model=entry["model"],
+                temperature=float(entry.get("temperature", 0.3)),
+                max_tokens=int(entry.get("max_tokens", 4096)))
+            executor_model = f"{entry['name']}（{entry['model']}）"
+
+        if executor_chat is not None and executor_pool_id == used_pool_id:
+            raise HTTPException(status_code=400,
+                                detail="规划者与执行者是同一个模型，两级模式无意义；请更换执行者条目")
+
         service = get_agent_service()
         request_id = new_request_id()
+        mode = "dual" if executor_chat is not None else "single"
         info = service.register(request_id, task, model, used_pool_id,
-                                workspace=workspace_dir)
+                                workspace=workspace_dir,
+                                executor_model=executor_model, mode=mode)
         if info is None:
             raise HTTPException(status_code=503, detail="智能体同时运行任务已达上限")
 
+        s = settings_store().to_dict()
+        agent_cfg = s.get("agent", {})
+
         async def _run():
             try:
                 await service.run(
@@ -546,6 +577,8 @@ try:
                     token_cap=int(agent_cfg.get("token_cap", 20000)),
                     allow_shell=bool(agent_cfg.get("allow_shell", False)),
                     shell_timeout_s=int(agent_cfg.get("shell_timeout_s", 20)),
+                    executor_chat=executor_chat,
+                    max_handoffs=int(agent_cfg.get("max_handoffs", 2)),
                 )
             except Exception as exc:
                 import traceback
@@ -557,7 +590,8 @@ try:
 
         info.asyncio_task = asyncio.create_task(_run())
         return {"request_id": request_id, "status": "running", "model": model,
-                "workspace": workspace_dir}
+                "workspace": workspace_dir, "mode": mode,
+                "executor_model": executor_model}
 
     @app.get("/agent/fs", tags=["agent"])
     async def agent_fs_browse(path: str = ""):