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/agent.py

@@ -41,6 +41,42 @@ AGENT_SYSTEM_PROMPT = (
     "直接输出给用户的最终答复（中文，不要再调用工具）。"
 )
 
+# ── 两级智能体（D7）：规划者（大模型）+ 执行者（本地小模型），交接走 handoff 文档 ──
+PLANNER_SYSTEM_PROMPT = (
+    "你是两级智能体中的**规划者**（大模型）。执行者是一个能力有限的本地小模型，"
+    "只能机械地使用工具。你的职责：把用户任务拆成执行者可照做的**具体指令**，"
+    "并在执行后审查其汇报。输出必须是合法 JSON 对象（不要 markdown 围栏）。"
+)
+EXECUTOR_SYSTEM_PROMPT = (
+    "你是两级智能体中的**执行者**（本地小模型）。规划者已给你具体指令，"
+    "你只负责用工具完成指令并在最后**汇报**：做了什么、结果如何、有什么问题。"
+    "严格遵守指令范围，不要自行扩大任务。汇报用中文，是给规划者看的，"
+    "要列出：修改的文件、关键命令输出、未完成项。"
+)
+
+# 规划者首轮：产出指令（JSON）
+_PLAN_SCHEMA_HINT = {
+    "instructions": "string(给执行者的具体步骤指令，<=600字)",
+    "acceptance": "string(验收标准，<=200字)",
+}
+# 规划者审查轮：裁决（JSON）
+_REVIEW_SCHEMA_HINT = {
+    "verdict": "enum(done|redo)",
+    "reply_to_executor": "string(verdict=redo 时给执行者的补充指令；done 时可空)",
+    "final_answer": "string(verdict=done 时给用户的最终答复)",
+}
+
+DEFAULT_MAX_HANDOFFS = 2   # 规划者<->执行者交接轮数上限
+
+
+def _parse_json_loose(content: str) -> Dict[str, Any]:
+    """宽松解析规划者的 JSON 输出（剥围栏/取首个对象）；失败返回 {}。"""
+    try:
+        from router_system.architect import ArchitectClient
+        return ArchitectClient._parse_json(content)
+    except Exception:
+        return {}
+
 
 # ─────────────────────────────────────────────────────────────────────────────
 # OpenAI 兼容工具调用客户端
@@ -129,6 +165,8 @@ class AgentRunInfo:
     completion_tokens: int = 0
     pool_id: str = ""
     workspace: str = ""                # 本次运行使用的工作区根目录（绝对路径）
+    executor_model: str = ""           # 两级模式：执行者模型名（空 = 单模型模式）
+    mode: str = "single"               # single | dual
     asyncio_task: Optional[asyncio.Task] = field(default=None, repr=False)
 
     def to_dict(self) -> Dict[str, Any]:
@@ -146,6 +184,8 @@ class AgentRunInfo:
             "completion_tokens": self.completion_tokens,
             "pool_id": self.pool_id,
             "workspace": self.workspace,
+            "executor_model": self.executor_model,
+            "mode": self.mode,
         }
 
 
@@ -169,12 +209,15 @@ class AgentService:
 
     # ---------- 注册与查询 ----------
     def register(self, request_id: str, task: str, model: str, pool_id: str,
-                 workspace: str = "") -> Optional[AgentRunInfo]:
+                 workspace: str = "", executor_model: str = "",
+                 mode: str = "single") -> Optional[AgentRunInfo]:
         running = [r for r in self._runs.values() if r.state == STATE_RUNNING]
         if len(running) >= self.max_running:
             return None
         info = AgentRunInfo(request_id=request_id, task=task, model=model,
-                            pool_id=pool_id, workspace=workspace, started_at=time.time())
+                            pool_id=pool_id, workspace=workspace,
+                            executor_model=executor_model, mode=mode,
+                            started_at=time.time())
         self._runs[request_id] = info
         self._dir(request_id).mkdir(parents=True, exist_ok=True)
         self._write_status(info)
@@ -186,27 +229,28 @@ class AgentService:
     # ---------- 执行 ----------
     async def run(self, info: AgentRunInfo, chat: Any, workspace_dir: str | Path,
                   max_rounds: int = 8, token_cap: int = 0,
-                  allow_shell: bool = False, shell_timeout_s: int = 20) -> None:
-        """执行智能体任务（由调用方包成后台协程）。"""
+                  allow_shell: bool = False, shell_timeout_s: int = 20,
+                  executor_chat: Any = None,
+                  max_handoffs: int = DEFAULT_MAX_HANDOFFS) -> None:
+        """执行智能体任务（由调用方包成后台协程）。
+
+        executor_chat 为空 = 单模型模式（chat 全程包办）；
+        提供时进入两级模式：chat 作规划者，executor_chat 作执行者（D7）。
+        """
+        try:
+            if executor_chat is not None:
+                result = await self.run_dual(
+                    info, chat, executor_chat, workspace_dir,
+                    max_rounds=max_rounds, token_cap=token_cap,
+                    allow_shell=allow_shell, shell_timeout_s=shell_timeout_s,
+                    max_handoffs=max_handoffs)
+            else:
                 tools = WorkspaceTools(workspace_dir, allow_shell=allow_shell,
                                        shell_timeout_s=shell_timeout_s)
                 loop = ToolLoop(tools, chat, max_rounds=max_rounds, token_cap=token_cap,
                                 on_event=self._make_event_writer(info))
-        try:
                 result = await loop.run(info.task, system=AGENT_SYSTEM_PROMPT)
-            info.response = result.get("response", "")
-            info.rounds = int(result.get("rounds", 0))
-            info.prompt_tokens = int(result.get("prompt_tokens", 0))
-            info.completion_tokens = int(result.get("completion_tokens", 0))
-            if result.get("reason") == "error":
-                info.state = STATE_FAILED
-                info.error = result.get("error")
-            elif result.get("reason") in ("token_cap", "max_rounds"):
-                # 触顶属于护栏行为：结果仍交付，但标记部分完成信息
-                info.state = STATE_DONE
-                info.error = result.get("error")
-            else:
-                info.state = STATE_DONE
+            self._apply_result(info, result)
         except Exception as exc:  # pragma: no cover
             info.state = STATE_FAILED
             info.error = f"{type(exc).__name__}: {exc}"
@@ -216,6 +260,156 @@ class AgentService:
             info.finished_at = time.time()
             self._write_status(info)
 
+    def _apply_result(self, info: AgentRunInfo, result: Dict[str, Any]) -> None:
+        """把循环结果落到运行状态（单/两级模式共用）。"""
+        info.response = result.get("response", "")
+        info.rounds = int(result.get("rounds", 0))
+        info.prompt_tokens = int(result.get("prompt_tokens", 0))
+        info.completion_tokens = int(result.get("completion_tokens", 0))
+        if result.get("reason") == "error":
+            info.state = STATE_FAILED
+            info.error = result.get("error")
+        elif result.get("reason") in ("token_cap", "max_rounds", "max_handoffs"):
+            # 触顶属于护栏行为：结果仍交付，但标记部分完成信息
+            info.state = STATE_DONE
+            info.error = result.get("error")
+        else:
+            info.state = STATE_DONE
+
+    # ---------- 两级模式（D7）：规划者 + 执行者 ----------
+    async def run_dual(self, info: AgentRunInfo, planner_chat: Any, executor_chat: Any,
+                       workspace_dir: str | Path, max_rounds: int = 8,
+                       token_cap: int = 0, allow_shell: bool = False,
+                       shell_timeout_s: int = 20,
+                       max_handoffs: int = DEFAULT_MAX_HANDOFFS) -> Dict[str, Any]:
+        """大模型拆解/审查 + 小模型执行工具轮，交接状态写 handoff.json（智能体版交流文本）。"""
+        info.mode = "dual"
+        tools = WorkspaceTools(workspace_dir, allow_shell=allow_shell,
+                               shell_timeout_s=shell_timeout_s)
+        handoff: Dict[str, Any] = {
+            "task": info.task, "planner_model": info.model,
+            "executor_model": info.executor_model, "workspace": info.workspace,
+            "instructions": "", "acceptance": "", "exchanges": [],
+        }
+        spent = {"in": 0, "out": 0}
+        total_rounds = 0
+
+        def _account(usage: Dict[str, Any] | None) -> None:
+            spent["in"] += int((usage or {}).get("prompt_tokens", 0))
+            spent["out"] += int((usage or {}).get("completion_tokens", 0))
+
+        def _save_handoff() -> None:
+            try:
+                (self._dir(info.request_id) / "handoff.json").write_text(
+                    json.dumps(handoff, ensure_ascii=False, indent=2), encoding="utf-8")
+            except OSError:
+                pass
+
+        def _remaining_cap() -> int:
+            return (token_cap - spent["in"] - spent["out"]) if token_cap else 1
+
+        async def _planner_json(user_msg: str) -> Dict[str, Any]:
+            """调规划者并解析 JSON；解析失败回喂重试一次，再失败降级为 {}（禁止带病继续的软版本）。"""
+            messages = [{"role": "system", "content": PLANNER_SYSTEM_PROMPT},
+                        {"role": "user", "content": user_msg}]
+            content = ""
+            for attempt in (1, 2):
+                resp = await planner_chat(messages, [])
+                _account(resp.get("usage"))
+                content = resp.get("content") or ""
+                obj = _parse_json_loose(content)
+                if obj:
+                    break
+                if attempt == 1:
+                    messages += [{"role": "assistant", "content": content},
+                                 {"role": "user",
+                                  "content": "你的输出不是合法 JSON。请重新只输出合法 JSON 对象。"}]
+            self._append_event(info, {"type": "message", "role": "planner",
+                                      "content": content[:2000]})
+            return obj
+
+        try:
+            # ---- 阶段 1：规划（大模型拆解为执行者指令） ----
+            self._append_event(info, {"type": "phase", "phase": "plan", "model": info.model})
+            plan = await _planner_json(
+                f"用户任务：{info.task}\n\n"
+                "请产出给执行者的指令，仅输出符合如下结构的 JSON：\n"
+                + json.dumps(_PLAN_SCHEMA_HINT, ensure_ascii=False))
+            instructions = (plan.get("instructions") or info.task).strip()
+            handoff["instructions"] = instructions
+            handoff["acceptance"] = str(plan.get("acceptance", ""))
+            _save_handoff()
+
+            final_text = ""
+            reason = "answer"
+            error = None
+            exec_rounds_total = 0
+
+            # ---- 阶段 2/3：执行 <-> 审查（有界交接） ----
+            for h in range(1, max_handoffs + 1):
+                # 执行（本地小模型跑工具轮）
+                self._append_event(info, {"type": "phase", "phase": "execute",
+                                          "handoff": h, "model": info.executor_model})
+                loop = ToolLoop(tools, executor_chat, max_rounds=max_rounds,
+                                token_cap=max(1, _remaining_cap()),
+                                on_event=self._make_event_writer(info),
+                                emit_final=False)
+                exec_result = await loop.run(instructions, system=EXECUTOR_SYSTEM_PROMPT)
+                _account({"prompt_tokens": exec_result.get("prompt_tokens", 0),
+                          "completion_tokens": exec_result.get("completion_tokens", 0)})
+                exec_rounds_total += int(exec_result.get("rounds", 0))
+                report = exec_result.get("response", "")
+                # 执行者汇报作为消息事件透出（前端可读）
+                self._append_event(info, {"type": "message", "role": "executor",
+                                          "handoff": h, "content": (report or "")[:4000]})
+                if exec_result.get("reason") == "error":
+                    reason, error = "error", exec_result.get("error")
+                    final_text = report
+                    break
+
+                # 审查（大模型裁决）
+                self._append_event(info, {"type": "phase", "phase": "review",
+                                          "handoff": h, "model": info.model})
+                review = await _planner_json(
+                    f"用户任务：{info.task}\n你之前给出的指令：{instructions}\n"
+                    f"验收标准：{handoff['acceptance'] or '（未明确）'}\n\n"
+                    f"执行者第 {h} 轮汇报：\n{report[:4000]}\n\n"
+                    "请审查是否已按验收标准完成，仅输出符合如下结构的 JSON：\n"
+                    + json.dumps(_REVIEW_SCHEMA_HINT, ensure_ascii=False))
+                verdict = str(review.get("verdict", "done")).lower()
+                handoff["exchanges"].append({
+                    "handoff": h, "executor_report": report,
+                    "verdict": verdict,
+                    "reply_to_executor": str(review.get("reply_to_executor", "")),
+                })
+                _save_handoff()
+
+                if verdict == "done":
+                    final_text = str(review.get("final_answer") or report)
+                    break
+                # redo：裁决意见作为下一轮执行者指令（带上一轮上下文）
+                instructions = str(review.get("reply_to_executor") or instructions)
+                if h == max_handoffs:
+                    reason = "max_handoffs"
+                    error = f"交接轮数达上限（{max_handoffs}），以执行者汇报收尾"
+                    final_text = report
+            else:
+                final_text = final_text or ""
+
+            self._append_event(info, {"type": "final", "round": total_rounds + exec_rounds_total,
+                                      "reason": reason, "error": error})
+            return {"response": final_text, "rounds": total_rounds + exec_rounds_total,
+                    "reason": reason, "error": error,
+                    "prompt_tokens": spent["in"], "completion_tokens": spent["out"]}
+        except Exception as exc:
+            reason = "error"
+            error = f"{type(exc).__name__}: {exc}"
+            self._append_event(info, {"type": "final", "round": total_rounds,
+                                      "reason": reason, "error": error})
+            return {"response": "", "rounds": total_rounds, "reason": reason,
+                    "error": error,
+                    "prompt_tokens": spent["in"], "completion_tokens": spent["out"]}
+
     # ---------- 事件 ----------
     def _make_event_writer(self, info: AgentRunInfo):
         def _on_event(ev: Dict[str, Any]) -> None:

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 = ""):

gateway/settings.py

@@ -43,6 +43,7 @@ DEFAULTS: Dict[str, Any] = {
         "token_cap": 20000,                  # 单次智能体任务 token 熔断
         "allow_shell": False,                # 允许 run_command 执行 shell（默认关）
         "shell_timeout_s": 20,               # shell 命令超时
+        "max_handoffs": 2,                   # 两级模式：规划者<->执行者交接轮数上限
     },
 }
 

gateway/static/assets/MetricsView-DOvNkmWB.js

@@ -1 +1 @@
-import{A as e,D as t,G as n,I as r,L as i,N as a,O as o,P as s,V as c,W as l,j as u,k as d,s as f,t as p}from"./index-Bi4FCB5S.js";var m={class:`metrics-view`},h={key:0,class:`loading`},g={key:1,class:`error`},_={class:`card-grid`},v={class:`metric-card`},y={class:`kv-list`},b={class:`metric-card`},x={class:`kv-list`},S={key:0,class:`metric-card highlight`},C={class:`kv-list`},w={key:0},T={key:1},E={key:1,class:`metric-card`},D={class:`by-model`},O={class:`mono`},k={key:2,class:`metric-card review-card`},A={class:`review-stats`},j={class:`stat-item`},M={class:`stat-num`},N={class:`stat-item`},P={class:`stat-num`},F={key:0,class:`progress-wrap`},I={class:`review-rate`},L={class:`raw-json`},R=p(a({__name:`MetricsView`,setup(a){let p=c(null),R=c(!1),z=c(``),B=o(()=>p.value?.v2?.by_model||null);async function V(){R.value=!0,z.value=``;try{p.value=await f()}catch(e){z.value=e instanceof Error?e.message:`指标加载失败，请检查后端服务`}finally{R.value=!1}}return s(V),(a,o)=>(r(),u(`div`,m,[d(`header`,{class:`metrics-header`},[o[0]||=d(`h2`,null,`系统指标`,-1),d(`button`,{class:`refresh`,onClick:V},`🔄 刷新`)]),R.value?(r(),u(`div`,h,`加载中…`)):z.value?(r(),u(`div`,g,n(z.value),1)):p.value?(r(),u(t,{key:2},[d(`div`,_,[d(`div`,v,[o[1]||=d(`h3`,null,`路由器（v1）`,-1),d(`div`,y,[(r(!0),u(t,null,i(p.value.router,(e,i)=>(r(),u(t,{key:i},[d(`span`,null,n(i),1),d(`b`,null,n(e),1)],64))),128))])]),d(`div`,b,[o[2]||=d(`h3`,null,`缓存`,-1),d(`div`,x,[(r(!0),u(t,null,i(p.value.cache,(e,i)=>(r(),u(t,{key:i},[d(`span`,null,n(i),1),d(`b`,null,n(e),1)],64))),128))])]),p.value.v2?(r(),u(`div`,S,[o[3]||=d(`h3`,null,`协作管线（v2）`,-1),d(`div`,C,[(r(!0),u(t,null,i(p.value.v2,(i,a)=>(r(),u(t,{key:a},[a===`by_model`?e(``,!0):(r(),u(`span`,w,n(a),1)),a===`by_model`?e(``,!0):(r(),u(`b`,T,n(i),1))],64))),128))])])):e(``,!0),B.value&&Object.keys(B.value).length?(r(),u(`div`,E,[o[5]||=d(`h3`,null,`按模型分账（token / 成本）`,-1),d(`table`,D,[o[4]||=d(`thead`,null,[d(`tr`,null,[d(`th`,null,`模型`),d(`th`,null,`次数`),d(`th`,null,`入`),d(`th`,null,`出`),d(`th`,null,`成本 $`)])],-1),d(`tbody`,null,[(r(!0),u(t,null,i(B.value,(e,t)=>(r(),u(`tr`,{key:t},[d(`td`,O,n(t),1),d(`td`,null,n(e.requests),1),d(`td`,null,n(e.input_tokens),1),d(`td`,null,n(e.output_tokens),1),d(`td`,null,n(e.cost_est_usd),1)]))),128))])]),o[6]||=d(`p`,{class:`hint`},`单价来自模型池条目（$/1M tokens）；经典设置下的模型成本不计入。`,-1)])):e(``,!0),p.value.review?(r(),u(`div`,k,[o[9]||=d(`h3`,null,`人工检验`,-1),d(`div`,A,[d(`div`,j,[d(`span`,M,n(p.value.review.pending),1),o[7]||=d(`span`,{class:`stat-label`},`待审核`,-1)]),d(`div`,N,[d(`span`,P,n(p.value.review.total),1),o[8]||=d(`span`,{class:`stat-label`},`总提交`,-1)])]),p.value.review.total>0?(r(),u(`div`,F,[d(`div`,{class:`reviewed-bar`,style:l({width:`${(p.value.review.total-p.value.review.pending)/p.value.review.total*100}%`})},null,4)])):e(``,!0),d(`p`,I,` 通过率： `+n(((p.value.review.total-p.value.review.pending)/p.value.review.total*100).toFixed(1))+`% `,1)])):e(``,!0)]),d(`details`,L,[o[10]||=d(`summary`,null,`原始 JSON`,-1),d(`pre`,null,n(JSON.stringify(p.value,null,2)),1)])],64)):e(``,!0)]))}}),[[`__scopeId`,`data-v-ba641559`]]);export{R as default};
+import{A as e,D as t,G as n,I as r,L as i,N as a,O as o,P as s,V as c,W as l,j as u,k as d,s as f,t as p}from"./index-DtjeaX4S.js";var m={class:`metrics-view`},h={key:0,class:`loading`},g={key:1,class:`error`},_={class:`card-grid`},v={class:`metric-card`},y={class:`kv-list`},b={class:`metric-card`},x={class:`kv-list`},S={key:0,class:`metric-card highlight`},C={class:`kv-list`},w={key:0},T={key:1},E={key:1,class:`metric-card`},D={class:`by-model`},O={class:`mono`},k={key:2,class:`metric-card review-card`},A={class:`review-stats`},j={class:`stat-item`},M={class:`stat-num`},N={class:`stat-item`},P={class:`stat-num`},F={key:0,class:`progress-wrap`},I={class:`review-rate`},L={class:`raw-json`},R=p(a({__name:`MetricsView`,setup(a){let p=c(null),R=c(!1),z=c(``),B=o(()=>p.value?.v2?.by_model||null);async function V(){R.value=!0,z.value=``;try{p.value=await f()}catch(e){z.value=e instanceof Error?e.message:`指标加载失败，请检查后端服务`}finally{R.value=!1}}return s(V),(a,o)=>(r(),u(`div`,m,[d(`header`,{class:`metrics-header`},[o[0]||=d(`h2`,null,`系统指标`,-1),d(`button`,{class:`refresh`,onClick:V},`🔄 刷新`)]),R.value?(r(),u(`div`,h,`加载中…`)):z.value?(r(),u(`div`,g,n(z.value),1)):p.value?(r(),u(t,{key:2},[d(`div`,_,[d(`div`,v,[o[1]||=d(`h3`,null,`路由器（v1）`,-1),d(`div`,y,[(r(!0),u(t,null,i(p.value.router,(e,i)=>(r(),u(t,{key:i},[d(`span`,null,n(i),1),d(`b`,null,n(e),1)],64))),128))])]),d(`div`,b,[o[2]||=d(`h3`,null,`缓存`,-1),d(`div`,x,[(r(!0),u(t,null,i(p.value.cache,(e,i)=>(r(),u(t,{key:i},[d(`span`,null,n(i),1),d(`b`,null,n(e),1)],64))),128))])]),p.value.v2?(r(),u(`div`,S,[o[3]||=d(`h3`,null,`协作管线（v2）`,-1),d(`div`,C,[(r(!0),u(t,null,i(p.value.v2,(i,a)=>(r(),u(t,{key:a},[a===`by_model`?e(``,!0):(r(),u(`span`,w,n(a),1)),a===`by_model`?e(``,!0):(r(),u(`b`,T,n(i),1))],64))),128))])])):e(``,!0),B.value&&Object.keys(B.value).length?(r(),u(`div`,E,[o[5]||=d(`h3`,null,`按模型分账（token / 成本）`,-1),d(`table`,D,[o[4]||=d(`thead`,null,[d(`tr`,null,[d(`th`,null,`模型`),d(`th`,null,`次数`),d(`th`,null,`入`),d(`th`,null,`出`),d(`th`,null,`成本 $`)])],-1),d(`tbody`,null,[(r(!0),u(t,null,i(B.value,(e,t)=>(r(),u(`tr`,{key:t},[d(`td`,O,n(t),1),d(`td`,null,n(e.requests),1),d(`td`,null,n(e.input_tokens),1),d(`td`,null,n(e.output_tokens),1),d(`td`,null,n(e.cost_est_usd),1)]))),128))])]),o[6]||=d(`p`,{class:`hint`},`单价来自模型池条目（$/1M tokens）；经典设置下的模型成本不计入。`,-1)])):e(``,!0),p.value.review?(r(),u(`div`,k,[o[9]||=d(`h3`,null,`人工检验`,-1),d(`div`,A,[d(`div`,j,[d(`span`,M,n(p.value.review.pending),1),o[7]||=d(`span`,{class:`stat-label`},`待审核`,-1)]),d(`div`,N,[d(`span`,P,n(p.value.review.total),1),o[8]||=d(`span`,{class:`stat-label`},`总提交`,-1)])]),p.value.review.total>0?(r(),u(`div`,F,[d(`div`,{class:`reviewed-bar`,style:l({width:`${(p.value.review.total-p.value.review.pending)/p.value.review.total*100}%`})},null,4)])):e(``,!0),d(`p`,I,` 通过率： `+n(((p.value.review.total-p.value.review.pending)/p.value.review.total*100).toFixed(1))+`% `,1)])):e(``,!0)]),d(`details`,L,[o[10]||=d(`summary`,null,`原始 JSON`,-1),d(`pre`,null,n(JSON.stringify(p.value,null,2)),1)])],64)):e(``,!0)]))}}),[[`__scopeId`,`data-v-ba641559`]]);export{R as default};

gateway/static/assets/ReviewView-ByUG_VcN.js

@@ -1 +1 @@
-import{A as e,D as t,E as n,G as r,I as i,L as a,M as o,N as s,O as c,P as l,U as u,V as d,f,j as p,k as m,t as h,v as g,z as _}from"./index-Bi4FCB5S.js";var v={class:`review-view`},y={class:`review-header`},b={class:`controls`},x={key:0,class:`loading`},S={key:1,class:`error`},C={key:2,class:`queue-list`},w={key:0,class:`empty`},T={class:`card-header`},E={class:`card-id`},D={class:`tags`},O={class:`date`},k={class:`query-block`},A={class:`response-block`},j={key:0,class:`actions`},M=[`onUpdate:modelValue`],N={class:`btn-row`},P=[`onClick`],F=[`onClick`],I={key:1,class:`correction`},L=h(s({__name:`ReviewView`,setup(s){let h=d([]),L=d(!1),R=d(``),z=d(`pending`),B=d({}),V=c(()=>z.value===`all`?h.value:h.value.filter(e=>e.verdict===z.value));async function H(){L.value=!0,R.value=``;try{h.value=await f()}catch(e){R.value=e instanceof Error?e.message:String(e)}finally{L.value=!1}}async function U(e,t){try{await g(e,t,B.value[e]||void 0),await H()}catch(e){R.value=e instanceof Error?e.message:String(e)}}return l(H),(s,c)=>(i(),p(`div`,v,[m(`header`,y,[c[4]||=m(`h2`,null,`人工检验队列`,-1),m(`div`,b,[m(`button`,{class:u({active:z.value===`all`}),onClick:c[0]||=e=>z.value=`all`},`全部`,2),m(`button`,{class:u({active:z.value===`pending`}),onClick:c[1]||=e=>z.value=`pending`},`待审核`,2),m(`button`,{class:u({active:z.value===`approved`}),onClick:c[2]||=e=>z.value=`approved`},`已通过`,2),m(`button`,{class:u({active:z.value===`rejected`}),onClick:c[3]||=e=>z.value=`rejected`},`已拒绝`,2),m(`button`,{class:`refresh-btn`,onClick:H},`🔄 刷新`)])]),L.value?(i(),p(`div`,x,`加载中…`)):R.value?(i(),p(`div`,S,r(R.value),1)):(i(),p(`div`,C,[V.value.length?e(``,!0):(i(),p(`div`,w,`队列为空。`)),(i(!0),p(t,null,a(V.value,s=>(i(),p(`div`,{key:s.id,class:`review-card`},[m(`div`,T,[m(`span`,E,`#`+r(s.id),1),m(`span`,{class:u([`verdict-badge`,s.verdict])},r(s.verdict),3),m(`span`,D,[(i(!0),p(t,null,a(s.tags,e=>(i(),p(`span`,{key:e,class:`tag`},r(e),1))),128))]),m(`span`,O,r(s.created_at),1)]),m(`div`,k,[c[5]||=m(`strong`,null,`Query：`,-1),o(r(s.query),1)]),m(`div`,A,[c[6]||=m(`strong`,null,`Response：`,-1),m(`pre`,null,r(s.response),1)]),s.verdict===`pending`?(i(),p(`div`,j,[_(m(`textarea`,{"onUpdate:modelValue":e=>B.value[s.id]=e,placeholder:`修正意见（可选）`,rows:`2`},null,8,M),[[n,B.value[s.id]]]),m(`div`,N,[m(`button`,{class:`approve`,onClick:e=>U(s.id,`approved`)},`✅ 通过`,8,P),m(`button`,{class:`reject`,onClick:e=>U(s.id,`rejected`)},`❌ 拒绝`,8,F)])])):s.correction?(i(),p(`div`,I,[c[7]||=m(`strong`,null,`修正：`,-1),o(r(s.correction),1)])):e(``,!0)]))),128))]))]))}}),[[`__scopeId`,`data-v-19c16eff`]]);export{L as default};
+import{A as e,D as t,E as n,G as r,I as i,L as a,M as o,N as s,O as c,P as l,U as u,V as d,f,j as p,k as m,t as h,v as g,z as _}from"./index-DtjeaX4S.js";var v={class:`review-view`},y={class:`review-header`},b={class:`controls`},x={key:0,class:`loading`},S={key:1,class:`error`},C={key:2,class:`queue-list`},w={key:0,class:`empty`},T={class:`card-header`},E={class:`card-id`},D={class:`tags`},O={class:`date`},k={class:`query-block`},A={class:`response-block`},j={key:0,class:`actions`},M=[`onUpdate:modelValue`],N={class:`btn-row`},P=[`onClick`],F=[`onClick`],I={key:1,class:`correction`},L=h(s({__name:`ReviewView`,setup(s){let h=d([]),L=d(!1),R=d(``),z=d(`pending`),B=d({}),V=c(()=>z.value===`all`?h.value:h.value.filter(e=>e.verdict===z.value));async function H(){L.value=!0,R.value=``;try{h.value=await f()}catch(e){R.value=e instanceof Error?e.message:String(e)}finally{L.value=!1}}async function U(e,t){try{await g(e,t,B.value[e]||void 0),await H()}catch(e){R.value=e instanceof Error?e.message:String(e)}}return l(H),(s,c)=>(i(),p(`div`,v,[m(`header`,y,[c[4]||=m(`h2`,null,`人工检验队列`,-1),m(`div`,b,[m(`button`,{class:u({active:z.value===`all`}),onClick:c[0]||=e=>z.value=`all`},`全部`,2),m(`button`,{class:u({active:z.value===`pending`}),onClick:c[1]||=e=>z.value=`pending`},`待审核`,2),m(`button`,{class:u({active:z.value===`approved`}),onClick:c[2]||=e=>z.value=`approved`},`已通过`,2),m(`button`,{class:u({active:z.value===`rejected`}),onClick:c[3]||=e=>z.value=`rejected`},`已拒绝`,2),m(`button`,{class:`refresh-btn`,onClick:H},`🔄 刷新`)])]),L.value?(i(),p(`div`,x,`加载中…`)):R.value?(i(),p(`div`,S,r(R.value),1)):(i(),p(`div`,C,[V.value.length?e(``,!0):(i(),p(`div`,w,`队列为空。`)),(i(!0),p(t,null,a(V.value,s=>(i(),p(`div`,{key:s.id,class:`review-card`},[m(`div`,T,[m(`span`,E,`#`+r(s.id),1),m(`span`,{class:u([`verdict-badge`,s.verdict])},r(s.verdict),3),m(`span`,D,[(i(!0),p(t,null,a(s.tags,e=>(i(),p(`span`,{key:e,class:`tag`},r(e),1))),128))]),m(`span`,O,r(s.created_at),1)]),m(`div`,k,[c[5]||=m(`strong`,null,`Query：`,-1),o(r(s.query),1)]),m(`div`,A,[c[6]||=m(`strong`,null,`Response：`,-1),m(`pre`,null,r(s.response),1)]),s.verdict===`pending`?(i(),p(`div`,j,[_(m(`textarea`,{"onUpdate:modelValue":e=>B.value[s.id]=e,placeholder:`修正意见（可选）`,rows:`2`},null,8,M),[[n,B.value[s.id]]]),m(`div`,N,[m(`button`,{class:`approve`,onClick:e=>U(s.id,`approved`)},`✅ 通过`,8,P),m(`button`,{class:`reject`,onClick:e=>U(s.id,`rejected`)},`❌ 拒绝`,8,F)])])):s.correction?(i(),p(`div`,I,[c[7]||=m(`strong`,null,`修正：`,-1),o(r(s.correction),1)])):e(``,!0)]))),128))]))]))}}),[[`__scopeId`,`data-v-19c16eff`]]);export{L as default};

gateway/static/index.html

@@ -5,8 +5,8 @@
     <link rel="icon" type="image/svg+xml" href="/static/favicon.svg" />
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
     <title>webapp</title>
-    <script type="module" crossorigin src="/static/assets/index-Bi4FCB5S.js"></script>
-    <link rel="stylesheet" crossorigin href="/static/assets/index-zGyQjwum.css">
+    <script type="module" crossorigin src="/static/assets/index-DtjeaX4S.js"></script>
+    <link rel="stylesheet" crossorigin href="/static/assets/index-DPz6YNpx.css">
   </head>
   <body>
     <div id="app"></div>

router_system/tools.py

@@ -413,6 +413,7 @@ class ToolLoop:
         token_cap: int = 0,
         on_event: Optional[Callable[[Dict[str, Any]], None]] = None,
         result_preview_chars: int = MAX_RESULT_CHARS,
+        emit_final: bool = True,
     ):
         self.tools = tools
         self.chat_fn = chat_fn
@@ -420,8 +421,11 @@ class ToolLoop:
         self.token_cap = int(token_cap)          # 0 = 不限
         self.on_event = on_event
         self.result_preview_chars = result_preview_chars
+        self.emit_final = emit_final             # 两级模式内层循环置 False，由外层统一收尾
 
     def _emit(self, ev: Dict[str, Any]) -> None:
+        if ev.get("type") == "final" and not self.emit_final:
+            return  # 内层循环不发终态事件（外层编排负责）
         if self.on_event is not None:
             try:
                 self.on_event(ev)

tests/test_agent_api.py

@@ -244,3 +244,162 @@ def test_agent_workspace_and_file_accept_root(agent_env, client, tmp_path):
     # 非法 root -> 400
     r = client.get("/agent/workspace", params={"root": str(tmp_path / "nope")})
     assert r.status_code == 400
+
+
+# ---------------- 两级智能体（T26）：规划者 + 执行者 ----------------
+
+def _planner_resp(obj=None, raw=""):
+    content = raw or json.dumps(obj, ensure_ascii=False)
+    return {"content": content, "tool_calls": [],
+            "usage": {"prompt_tokens": 50, "completion_tokens": 20}}
+
+
+def _install_dual(agent_env, monkeypatch, planner_script, executor_script):
+    """注入假规划者（build_agent_chat）与假执行者（OpenAICompatChat）。"""
+
+    class FakePlanner:
+        api_key = "sk-fake"
+
+        def __init__(self, *a, **k):
+            self.script = list(planner_script)
+
+        async def __call__(self, messages, tools_spec):
+            if self.script:
+                return self.script.pop(0)
+            return _planner_resp({"verdict": "done", "final_answer": "（兜底）完成。"})
+
+    class FakeExecutorChat:
+        def __init__(self, *a, **k):
+            self.script = list(executor_script)
+
+        async def __call__(self, messages, tools_spec):
+            if self.script:
+                return self.script.pop(0)
+            return {"content": "（执行者兜底）没有更多动作。", "tool_calls": [], "usage": {}}
+
+    def fake_chat_factory(acfg):
+        return FakePlanner()
+
+    monkeypatch.setattr(ga, "build_agent_chat", fake_chat_factory)
+    monkeypatch.setattr(ag, "OpenAICompatChat", FakeExecutorChat)
+
+
+def test_dual_agent_done_flow(agent_env, client, monkeypatch, tmp_path):
+    """规划 -> 执行(写文件) -> 审查 done：事件/交接文档/状态全部落位。"""
+    _install_dual(
+        agent_env, monkeypatch,
+        planner_script=[
+            _planner_resp({"instructions": "在 data 目录创建 report.json",
+                           "acceptance": "文件存在且内容为合法 JSON"}),
+            _planner_resp({"verdict": "done", "reply_to_executor": "",
+                           "final_answer": "执行者已按指令创建数据文件，验收通过。"}),
+        ],
+        executor_script=[
+            {"content": None,
+             "tool_calls": [{"id": "e1", "name": "write_file",
+                             "arguments": {"path": "data/report.json",
+                                           "content": '{"ok": true}'}}],
+             "usage": {"prompt_tokens": 100, "completion_tokens": 10}},
+            {"content": "汇报：已创建 data/report.json，内容 {\"ok\": true}。",
+             "tool_calls": [], "usage": {"prompt_tokens": 120, "completion_tokens": 15}},
+        ])
+    r = client.post("/agent", json={"task": "建数据文件", "executor_pool_id": "no-such"})
+    # 执行者条目不存在 -> 400
+    assert r.status_code == 400
+
+    # 先放一个合法 llama_server 条目作为执行者
+    client.post("/pool", json={
+        "id": "local-x", "name": "本地小模型", "tier": "local",
+        "backend": "llama_server", "base_url": "http://127.0.0.1:8901/v1",
+        "model": "qwen-0.8b", "enabled": True})
+    r2 = client.post("/agent", json={"task": "建数据文件", "executor_pool_id": "local-x"})
+    assert r2.status_code == 200
+    assert r2.json()["mode"] == "dual"
+    assert "本地小模型" in r2.json()["executor_model"]
+
+    rid = r2.json()["request_id"]
+    info = _wait_done(agent_env["service"], rid)
+    assert info.state == "done", info.error
+    assert info.mode == "dual"
+    assert info.response == "执行者已按指令创建数据文件，验收通过。"
+
+    # 事件序列：规划 -> 执行(含工具) -> 审查 -> final
+    evs = agent_env["service"].read_events(rid)
+    phases = [e["phase"] for e in evs if e["type"] == "phase"]
+    assert phases == ["plan", "execute", "review"]
+    kinds = [e["type"] for e in evs]
+    assert "message" in kinds and "tool_call" in kinds
+    # 交接文档（智能体版交流文本）
+    ho = json.loads((agent_env["service"]._dir(rid) / "handoff.json").read_text(encoding="utf-8"))
+    assert ho["instructions"]
+    assert ho["exchanges"][0]["verdict"] == "done"
+    assert ho["executor_model"] == "本地小模型（qwen-0.8b）"
+
+    st = client.get(f"/agent/{rid}/status").json()
+    assert st["mode"] == "dual" and st["executor_model"]
+
+
+def test_dual_agent_redo_then_done(agent_env, client, monkeypatch):
+    """第一轮裁决 redo -> 执行者带补充指令再跑 -> 第二轮 done。"""
+    _install_dual(
+        agent_env, monkeypatch,
+        planner_script=[
+            _planner_resp({"instructions": "写 hello.txt"}),
+            _planner_resp({"verdict": "redo", "reply_to_executor": "文件内容不对，请写入 DONE",
+                           "final_answer": ""}),
+            _planner_resp({"verdict": "done", "reply_to_executor": "",
+                           "final_answer": "第二轮通过。"}),
+        ],
+        executor_script=[
+            {"content": "汇报：已写 hello.txt（内容空白）", "tool_calls": [],
+             "usage": {"prompt_tokens": 10, "completion_tokens": 5}},
+            {"content": None,
+             "tool_calls": [{"id": "e1", "name": "write_file",
+                             "arguments": {"path": "hello.txt", "content": "DONE"}}],
+             "usage": {"prompt_tokens": 10, "completion_tokens": 5}},
+            {"content": "汇报：已按补充指令重写 hello.txt 内容为 DONE",
+             "tool_calls": [], "usage": {"prompt_tokens": 10, "completion_tokens": 5}},
+        ])
+    client.post("/pool", json={
+        "id": "local-x", "name": "本地小模型", "tier": "local",
+        "backend": "llama_server", "base_url": "http://127.0.0.1:8901/v1",
+        "model": "qwen-0.8b", "enabled": True})
+    r = client.post("/agent", json={"task": "写 hello.txt", "executor_pool_id": "local-x"})
+    rid = r.json()["request_id"]
+    info = _wait_done(agent_env["service"], rid)
+    assert info.state == "done"
+    assert info.response == "第二轮通过。"
+    ho = json.loads((agent_env["service"]._dir(rid) / "handoff.json").read_text(encoding="utf-8"))
+    assert [x["verdict"] for x in ho["exchanges"]] == ["redo", "done"]
+    # 第二轮执行者应收到 redo 补充指令（消息历史含 reply_to_executor 内容）
+    evs = agent_env["service"].read_events(rid)
+    exec_phases = [e for e in evs if e["type"] == "phase" and e["phase"] == "execute"]
+    assert len(exec_phases) == 2
+
+
+def test_dual_agent_executor_error(agent_env, client, monkeypatch):
+    """执行者客户端异常 -> 任务 failed，错误透出。"""
+    class BoomChat:
+        def __init__(self, *a, **k):
+            pass
+
+        async def __call__(self, messages, tools_spec):
+            raise RuntimeError("本地模型连不上")
+
+    class PlanOK:
+        api_key = "sk-fake"
+
+        async def __call__(self, messages, tools_spec):
+            return _planner_resp({"instructions": "随便执行"})
+
+    monkeypatch.setattr(ga, "build_agent_chat", lambda acfg: PlanOK())
+    monkeypatch.setattr(ag, "OpenAICompatChat", BoomChat)
+    client.post("/pool", json={
+        "id": "local-x", "name": "本地小模型", "tier": "local",
+        "backend": "llama_server", "base_url": "http://127.0.0.1:8901/v1",
+        "model": "qwen-0.8b", "enabled": True})
+    r = client.post("/agent", json={"task": "t", "executor_pool_id": "local-x"})
+    rid = r.json()["request_id"]
+    info = _wait_done(agent_env["service"], rid)
+    assert info.state == "failed"
+    assert "RuntimeError" in (info.error or "")

tests/test_gateway.py

@@ -86,6 +86,12 @@ def test_metrics(client):
 
 
 def test_config_get_put_reset(client):
+    """/config 端到端。settings.json 是活文件（用户真实配置），
+    测试前后必须备份/恢复，禁止把用户配置清掉。"""
+    import json
+    store = ga.settings_store()
+    snapshot = json.loads(json.dumps(store._data, ensure_ascii=False))
+    try:
         # GET 默认
         r = client.get("/config")
         assert r.status_code == 200
@@ -98,6 +104,10 @@ def test_config_get_put_reset(client):
         r3 = client.post("/config/reset")
         assert r3.status_code == 200
         assert r3.json()["worker"]["backend"] == "llama_server"
+    finally:
+        store._data = snapshot
+        store.save()
+        ga.rebuild_pipeline()
 
 
 def test_workspace_not_found(client):

webapp/src/api/index.ts

@@ -311,7 +311,7 @@ export async function listPoolModels(id: string) {
 // ── 智能体（工具调用） ───────────────────────────────────────────────────────
 
 export interface AgentEvent {
-  type: 'round' | 'tool_call' | 'tool_result' | 'usage' | 'final'
+  type: 'round' | 'tool_call' | 'tool_result' | 'usage' | 'final' | 'phase' | 'message'
   ts?: number
   round?: number
   name?: string
@@ -322,6 +322,12 @@ export interface AgentEvent {
   completion_tokens?: number
   reason?: string
   error?: string
+  // 两级模式（D7）
+  phase?: 'plan' | 'execute' | 'review'
+  handoff?: number
+  model?: string
+  role?: 'planner' | 'executor'
+  content?: string
 }
 
 export interface AgentStatus {
@@ -337,12 +343,21 @@ export interface AgentStatus {
   started_at?: number
   finished_at?: number
   workspace?: string
+  executor_model?: string
+  mode?: 'single' | 'dual'
 }
 
-/** POST /agent：提交智能体任务（workspace 可选：选中工作目录） */
-export async function startAgent(task: string, poolId?: string, workspace?: string) {
-  const { data } = await http.post<{ request_id: string; status: string; model: string; workspace?: string }>(
-    '/agent', { task, pool_id: poolId || undefined, workspace: workspace || undefined })
+/** POST /agent：提交智能体任务（executorPoolId 可选：两级模式的执行者/本地小模型） */
+export async function startAgent(task: string, poolId?: string, workspace?: string, executorPoolId?: string) {
+  const { data } = await http.post<{
+    request_id: string; status: string; model: string
+    workspace?: string; mode?: 'single' | 'dual'; executor_model?: string
+  }>('/agent', {
+    task,
+    pool_id: poolId || undefined,
+    workspace: workspace || undefined,
+    executor_pool_id: executorPoolId || undefined,
+  })
   return data
 }
 

webapp/src/views/AgentView.vue

@@ -49,6 +49,13 @@
           <option value="">默认模型（模型池 Agent 角色 / Architect 设置）</option>
           <option v-for="e in agentModels" :key="e.id" :value="e.id">{{ e.name }}（{{ e.model }}）</option>
         </select>
+        <select v-model="selectedExecutorId" class="model-select executor-select"
+          title="两级模式：大模型拆解/审查，本地小模型执行工具轮">
+          <option value="">单模型模式（规划者全程包办）</option>
+          <option v-for="e in executorCandidates" :key="e.id" :value="e.id">
+            🔧 执行者：{{ e.name }}（{{ e.model }}）
+          </option>
+        </select>
       </div>
       <div class="task-bar">
         <textarea v-model="task" class="task-input" rows="3"
@@ -66,6 +73,17 @@
         <template v-for="(ev, i) in events" :key="i">
           <div v-if="ev.type === 'round'" class="ev-round">— 第 {{ ev.round }} 轮 —</div>
 
+          <div v-else-if="ev.type === 'phase'" class="ev-phase" :class="'ph-' + ev.phase">
+            {{ phaseLabel(ev) }}
+          </div>
+
+          <div v-else-if="ev.type === 'message'" class="ev-msg">
+            <div class="ev-title">
+              {{ ev.role === 'planner' ? '🧠 规划者' : '🔧 执行者' }}{{ ev.handoff ? `（第 ${ev.handoff} 轮交接）` : '' }}
+            </div>
+            <pre class="ev-msgbody" :class="ev.role === 'planner' ? 'msg-planner' : 'msg-executor'">{{ ev.content }}</pre>
+          </div>
+
           <div v-else-if="ev.type === 'tool_call' && ev.name === 'edit_file'" class="ev-card">
             <div class="ev-title">✏️ 精确编辑 <b>{{ (ev.arguments as any)?.path }}</b></div>
             <pre class="diff-old">- {{ (ev.arguments as any)?.old_string }}</pre>
@@ -171,6 +189,9 @@ const manualPath = ref('')
 const createIfMissing = ref(false)
 const allowShell = ref(false)
 
+// 两级模式
+const selectedExecutorId = ref('')
+
 const wsFiles = ref<{ name: string; type: string; size?: number }[]>([])
 const loadingWs = ref(false)
 const filePreview = ref<{ path: string; content: string; truncated: boolean } | null>(null)
@@ -180,6 +201,18 @@ let _watch: ReturnType<typeof watchAgent> | null = null
 
 const agentModels = computed(() => poolEntries.value.filter(e => e.enabled))
 
+const executorCandidates = computed(() =>
+  poolEntries.value
+    .filter(e => e.enabled && e.backend !== 'mock')
+    .sort((a, b) => (a.backend === 'llama_server' ? -1 : 0) - (b.backend === 'llama_server' ? -1 : 0)))
+
+function phaseLabel(ev: AgentEvent) {
+  const m = ev.model ? ` · ${ev.model}` : ''
+  if (ev.phase === 'plan') return `🧠 规划者 · 拆解任务${m}`
+  if (ev.phase === 'execute') return `🔧 执行者 · 工具执行（第 ${ev.handoff} 轮交接）${m}`
+  return `🔍 规划者 · 审查裁决${m}`
+}
+
 const finalResponse = computed(() => ((statusInfo.value as any)?.response || ''))
 
 const statusLine = computed(() => {
@@ -281,7 +314,8 @@ async function run() {
   events.value = []
   statusInfo.value = null
   try {
-    const init = await startAgent(task.value.trim(), selectedPoolId.value, selectedRoot.value)
+    const init = await startAgent(task.value.trim(), selectedPoolId.value, selectedRoot.value,
+                                  selectedExecutorId.value)
     if (init.workspace) selectedRoot.value = init.workspace
     _watch = watchAgent(init.request_id)
     _watch.subscribe({
@@ -481,6 +515,37 @@ onMounted(async () => {
   font-size: 12px;
   margin: 10px 0;
 }
+.ev-phase {
+  text-align: center;
+  font-size: 12px;
+  font-weight: 600;
+  border-radius: 12px;
+  padding: 4px 14px;
+  margin: 12px auto 8px;
+  width: fit-content;
+  max-width: 90%;
+  white-space: nowrap;
+  overflow: hidden;
+  text-overflow: ellipsis;
+}
+.ph-plan { color: #1e40af; background: #dbeafe; }
+.ph-execute { color: #166534; background: #dcfce7; }
+.ph-review { color: #7c2d12; background: #ffedd5; }
+.ev-msg { margin-bottom: 8px; }
+.ev-msgbody {
+  margin: 4px 0 0;
+  padding: 8px 12px;
+  border-radius: 8px;
+  font-size: 13px;
+  font-family: inherit;
+  white-space: pre-wrap;
+  word-break: break-word;
+  max-height: 260px;
+  overflow-y: auto;
+}
+.msg-planner { background: #eff6ff; color: #1e3a8a; border: 1px solid #bfdbfe; }
+.msg-executor { background: #f0fdf4; color: #14532d; border: 1px solid #bbf7d0; }
+.executor-select { flex: 1; max-width: 340px; }
 .ev-card {
   border: 1px solid #e5e7eb;
   border-radius: 8px;

任务拆解与执行计划.md

@@ -118,3 +118,4 @@ P0 完成后的能力：干净的后端抽象 + 可量化的评测 + 可追溯
 | T23 | 工作区选择：/agent/fs 目录浏览 + /agent/workspaces 最近列表 + /agent 带 workspace | ✅ 完成 | T23 |
 | T24 | 前端：工作区选择栏（目录浏览器/最近/shell 开关）+ edit diff 卡 + 命令卡 | ✅ 完成 | T24-T25 |
 | T25 | 集成验证：选定真实目录"读→精确编辑→运行验证"全链路 + 274 测试全绿 | ✅ 完成 | T24-T25 |
+| T26 | 两级智能体：规划者(大模型)拆解/审查 + 执行者(本地小模型)工具轮，handoff.json 交接，/agent 带 executor_pool_id | ✅ 完成 | T26 |

实现方案_v4_模型池与工具智能体.md

@@ -95,3 +95,17 @@
 
 `config/model_pool.json`（池持久化）、`agent_runs/{id}/`（events.jsonl + status.json）、
 `agent_workspace/`（智能体默认工作区，可在设置 agent.workspace_dir 调整）。
+
+## 7. 增补：两级智能体（D7，T26）
+
+- **决策 D7（价位协作延伸到智能体形态）**：`POST /agent` 支持 `executor_pool_id`——
+  规划者（大模型：agent 角色或经典 Architect 设置）负责拆解指令与审查裁决；
+  执行者（本地小模型：llama_server/openai 池条目）负责工具轮；中间状态写
+  `agent_runs/{id}/handoff.json`（智能体版"交流文本"：instructions/acceptance/exchanges）。
+- **协议**：规划者两阶段 JSON（plan: instructions+acceptance；review: verdict done|redo +
+  reply_to_executor + final_answer），解析失败回喂重试一次再降级（对齐仓库 JSON 纪律）；
+  redo 时裁决意见作为执行者下一轮指令，交接上限 `agent.max_handoffs`（默认 2）。
+- **护栏**：规划者与执行者不得为同一池条目；执行轮 token 与整体 token_cap 共享预算；
+  ToolLoop 内层循环 `emit_final=False`，终态事件由外层编排统一发出（防前端 SSE 提前收口）。
+- **事件**：新增 `phase`（plan/execute/review，带模型名）与 `message`（planner/executor 正文）
+  两类事件，前端以阶段徽标 + 双色消息卡渲染。

毕业设计_进度记录.md

@@ -122,3 +122,15 @@
 - **测试**：全量 274 passed（新增 12：tools 扩展 7 + 工作区选择 5）。
 - **安全边界（论文可写）**：文件操作相对所选根关押；shell 独立开关默认关；
   唯一性约束防覆盖式误编辑——"能力开放 + 边界收窄"的设计权衡可作一节。
+
+### 5.5 增补（同日）：两级智能体（T26，路线 B 落地）
+
+- **形态**：大模型（规划者）拆解任务出指令 JSON → 本地小模型（执行者）跑工具轮并汇报 →
+  大模型审查裁决 done/redo（有界交接，默认 2 轮）→ 最终答复。交接状态落
+  `handoff.json`（智能体版交流文本），SSE 新增 phase/message 事件实时可视化。
+- **测试**：+3 项（done 流 / redo 流 / 执行者故障），全量 277 passed。
+- **事故与修复**：发现 `test_config_get_put_reset` 调 /config/reset 会清空真实
+  settings.json——此前全量跑测把用户在设置页配置的 API key 抹掉了。已修复测试隔离
+  （备份/恢复），并向用户说明需重配 key。
+- **运维提醒**：网关用 taskkill /T 停止会连带杀掉其启动的 llama-server 子进程；
+  重启网关后需在设置页重新启动 llama-server。