Merge branch 'master' of http://192.168.0.200/knowledge-base-bj/deerflow-server into ji

# Conflicts:
#	frontend-web/src/components/workspace/chats/chat-box.tsx
#	offline-backend-20260512/backend/app/gateway/routers/agents.py

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

@@ -503,6 +503,24 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
         except Exception:
             logger.exception("Skill ownership migration failed (non-fatal)")
 
+        # 一次性把 .deer-flow/agents/{id}/ 目录里历史遗留的 agent 导入 agents 表。
+        # 路由层原本每次请求都跑一遍这个 sync(扫文件系统 + N 个 agent 各一次 DB
+        # SELECT/UPDATE),把它挪到启动期 + 路由层只保留 TTL 兜底,显著降低
+        # /api/agents 列表延迟。
+        try:
+            from app.gateway.routers.agents import (
+                _mark_legacy_sync_fresh,
+                _sync_legacy_agents,
+            )
+
+            agent_store = getattr(app.state, "agent_store", None)
+            if agent_store is not None:
+                await _sync_legacy_agents(agent_store)
+                _mark_legacy_sync_fresh()
+                logger.info("Legacy folder-backed agents synced at startup")
+        except Exception:
+            logger.exception("Legacy agent sync failed at startup (non-fatal)")
+
         # 启动 skill curator 调度器:常驻后台任务,每小时检查一次需不需要进化
         # (per-user interval_hours 仍是真正决定要不要跑进化的闸)。
         try:

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

@@ -406,7 +406,7 @@ RECOMMENDER_AGENT_SOUL = """# 智能体推荐专员
 
 > 以下两个示例覆盖**不同任务类型**,目的是教你「思路要呼应任务本身」,**不要**把任何一个示例的具体职能链当模板。
 
-### 示例 A:合规改造任务(领域候选池)
+### 示例 A:合规改造任务(职能型候选池)
 
 **用户输入**:
 ```
@@ -415,11 +415,12 @@ RECOMMENDER_AGENT_SOUL = """# 智能体推荐专员
 关键假设:现有系统具备基础数字化能力;允许预定窗口停机
 
 候选智能体池:
-- compliance-research (合规调研):检索同类历史案例、行业最佳实践与监管口径
-- infra-capacity (基础设施评估):评估硬件容量与系统量化承载力
-- solution-architect (方案架构):产出主备方案与关键路径
-- risk-auditor (风险审查):合规红线检查与收敛建议
-- execution-planner (执行规划):里程碑与资源排布
+- roundtable-intelligence (情报收集):检索同类历史案例、行业最佳实践与外部情报
+- roundtable-environment (环境评估):评估硬件容量与系统量化承载力
+- roundtable-solution-design (方案设计):产出主备方案与关键路径
+- roundtable-risk-review (风险审查):合规红线检查与收敛建议
+- roundtable-execution-plan (执行规划):里程碑与资源排布
+- roundtable-summary (总结陈述):归纳冲突标记与盖印
 - my-cost-calc (成本测算):基于预算和资源单价计算总成本
 ```
 
@@ -430,10 +431,10 @@ RECOMMENDER_AGENT_SOUL = """# 智能体推荐专员
 ```json
 {
   "picks": [
-    {"agent_id": "compliance-research", "reason": "检索同类等保 2.0 改造案例与监管口径,直接支撑本次合规对标"},
-    {"agent_id": "solution-architect", "reason": "在 6 个月工期约束下产出主备改造方案与关键路径"},
-    {"agent_id": "risk-auditor", "reason": "对照等保 2.0 合规红线审查方案,避免后期返工"},
-    {"agent_id": "execution-planner", "reason": "把改造拆成可在 6 个月内交付的里程碑与资源排布"},
+    {"agent_id": "roundtable-intelligence", "reason": "检索同类等保 2.0 改造案例与监管口径,直接支撑本次合规对标"},
+    {"agent_id": "roundtable-solution-design", "reason": "在 6 个月工期约束下产出主备改造方案与关键路径"},
+    {"agent_id": "roundtable-risk-review", "reason": "对照等保 2.0 合规红线审查方案,避免后期返工"},
+    {"agent_id": "roundtable-execution-plan", "reason": "把改造拆成可在 6 个月内交付的里程碑与资源排布"},
     {"agent_id": "my-cost-calc", "reason": "校验整体方案落在 500 万元预算硬约束内"}
   ]
 }
@@ -452,7 +453,7 @@ RECOMMENDER_AGENT_SOUL = """# 智能体推荐专员
 - channel-planner (渠道规划师):按预算和受众设计跨平台投放组合
 - audience-insight (受众洞察):画像、关键词、情绪点研究
 - creative-copy (创意文案):产出落地页与社媒文案变体
-- ad-compliance (广告合规):审查文案与素材合规风险
+- roundtable-risk-review (风险审查):合规红线检查与收敛建议
 ```
 
 **你的输出**:
@@ -470,8 +471,8 @@ RECOMMENDER_AGENT_SOUL = """# 智能体推荐专员
 }
 ```
 
-> **关键观察**:两个示例都从候选池里按任务语义挑选,**不要**套用固定职能链(如"调研→设计→风险→执行")。
-> 候选池里的智能体能不能匹配任务,只看 description 和任务的实际语义。如果候选池里有"广告设计智能体"且任务是广告设计,**你必须挑它**,而不是机械复制示例 A 的链路。
+> **关键观察**:示例 A 选了 `roundtable-intelligence / solution-design / execution-plan` 等"职能型"智能体,示例 B **完全没用任何 `roundtable-*` 内置**,而是从用户自建的领域智能体里挑。
+> 这是因为:**候选池里的智能体能不能匹配任务,只看 description 和任务的实际语义,不看 id 是不是 `roundtable-` 开头**。如果候选池里有"广告设计智能体"且任务是广告设计,**你必须挑它**,而不是套用示例 A 的职能链。
 """
 
 

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

@@ -5,6 +5,7 @@ from __future__ import annotations
 import asyncio
 import logging
 import shutil
+import time
 from datetime import datetime
 from typing import Any
 from uuid import uuid4
@@ -207,6 +208,11 @@ async def _sync_legacy_agents(store: AgentStore) -> None:
 
     This preserves old deployments where agents already existed under
     ``.deer-flow/agents/{name}`` before the agents table was introduced.
+
+    Heavy: scans the agents directory + parses each ``config.yaml`` + issues a
+    SELECT-then-INSERT/UPDATE per agent. Call once at startup via the lifespan,
+    and rely on :func:`_maybe_sync_legacy_agents` (TTL-gated) inside request
+    handlers as a safety net for environments that bypassed the lifespan hook.
     """
     try:
         for agent_cfg in list_custom_agents():
@@ -225,6 +231,44 @@ async def _sync_legacy_agents(store: AgentStore) -> None:
         logger.exception("Failed to sync legacy folder-backed agents")
 
 
+# Throttle the per-request legacy-sync fallback. The real sync runs once at
+# startup (lifespan); request handlers used to call ``_sync_legacy_agents`` on
+# every hit, which scanned the filesystem + did N DB upserts per agent each
+# time. After the first successful sync we skip the work for ``_LEGACY_SYNC_TTL_SECONDS``
+# — long enough to be effectively "startup-only" in steady state, short enough
+# that a freshly dropped agent directory gets picked up on the next bucket.
+_LEGACY_SYNC_TTL_SECONDS = 3600.0
+_legacy_sync_state: dict[str, Any] = {"last_at": 0.0, "lock": None}
+
+
+async def _maybe_sync_legacy_agents(store: AgentStore) -> None:
+    """TTL-gated wrapper around :func:`_sync_legacy_agents`.
+
+    Becomes a no-op in steady state — the heavy lifting happens in the lifespan
+    hook. Inside the TTL window we don't even take the lock; outside it we
+    serialize callers so a thundering herd doesn't fan out into N filesystem
+    scans + N*M DB queries.
+    """
+    now = time.monotonic()
+    if now - float(_legacy_sync_state["last_at"]) < _LEGACY_SYNC_TTL_SECONDS:
+        return
+    lock = _legacy_sync_state["lock"]
+    if lock is None:
+        lock = asyncio.Lock()
+        _legacy_sync_state["lock"] = lock
+    async with lock:
+        # Re-check under the lock — a sibling coroutine may have just refreshed.
+        if time.monotonic() - float(_legacy_sync_state["last_at"]) < _LEGACY_SYNC_TTL_SECONDS:
+            return
+        await _sync_legacy_agents(store)
+        _legacy_sync_state["last_at"] = time.monotonic()
+
+
+def _mark_legacy_sync_fresh() -> None:
+    """Called by the lifespan hook after running :func:`_sync_legacy_agents`."""
+    _legacy_sync_state["last_at"] = time.monotonic()
+
+
 def _ensure_agent_dir_from_record(record: dict[str, Any]) -> None:
     """Lazily restore an agent's on-disk artifacts from its DB record.
 
@@ -262,14 +306,29 @@ def _record_to_response(
     record: dict[str, Any],
     *,
     include_soul: bool = False,
+    include_extras: bool = True,
 ) -> AgentResponse:
-    _ensure_agent_dir_from_record(record)
+    """Convert a DB row to an API response.
+
+    Args:
+        include_soul: Read SOUL.md from disk and embed in ``soul``. Set False
+            in list endpoints — the frontend (gallery / recommend / selector)
+            never reads ``soul``, only the edit dialog does. Skipping this
+            avoids one synchronous file read per agent.
+        include_extras: Self-heal a missing agent dir and read deprecated
+            ``model`` / ``tool_groups`` / ``skills`` from ``config.yaml``. Set
+            False in list endpoints to skip a stat + yaml parse per agent —
+            those fields are unused by current frontends and the dir restore
+            can happen lazily when the agent is actually opened.
+    """
     soul: str | None = None
+    extras: dict[str, Any] = {"skills": None, "model": None, "tool_groups": None}
+    if include_extras:
+        _ensure_agent_dir_from_record(record)
+        extras = _load_config_extras(record["id"])
     if include_soul:
         soul = load_agent_soul(record["id"]) or ""
 
-    extras = _load_config_extras(record["id"])
-
     return AgentResponse(
         id=record["id"],
         name=record["name"],
@@ -288,6 +347,25 @@ def _record_to_response(
     )
 
 
+async def _record_to_response_async(
+    record: dict[str, Any],
+    *,
+    include_soul: bool = False,
+    include_extras: bool = True,
+) -> AgentResponse:
+    """Async wrapper: when we actually touch disk, hop to a worker thread so
+    we don't block the asyncio loop. List endpoints (no soul, no extras) skip
+    this overhead and stay on the calling task."""
+    if not include_soul and not include_extras:
+        return _record_to_response(record, include_soul=False, include_extras=False)
+    return await asyncio.to_thread(
+        _record_to_response,
+        record,
+        include_soul=include_soul,
+        include_extras=include_extras,
+    )
+
+
 async def _get_owned_or_raise(store: AgentStore, agent_id: str, user_id: str) -> dict[str, Any]:
     owned = await store.get_owned(agent_id, user_id)
     if owned is not None:
@@ -349,7 +427,7 @@ async def list_agents(
     try:
         store = get_agent_store(request)
         tag_store: TagStore = get_tag_store(request)
-        await _sync_legacy_agents(store)
+        await _maybe_sync_legacy_agents(store)
         user_id = _current_user_id(request)
         if admin:
             if not _is_admin_user(request):
@@ -357,7 +435,14 @@ async def list_agents(
             agents = await store.list_all()
         else:
             agents = await store.list_visible(user_id)
-        responses = [_record_to_response(a, include_soul=True) for a in agents]
+        # List endpoint stays lean: skip per-agent SOUL.md read, config.yaml
+        # parse, and dir self-heal — none of those feed UI fields the gallery /
+        # recommend dialog / selector actually render. Detail endpoints (single
+        # agent + create/update responses) still load them.
+        responses = [
+            _record_to_response(a, include_soul=False, include_extras=False)
+            for a in agents
+        ]
 
         assignments = await tag_store.list_assignments("agent", [r.id for r in responses])
         favorite_ids = await store.list_favorite_ids(user_id)
@@ -403,11 +488,18 @@ async def get_agent_selector(request: Request) -> AgentSelectorResponse:
     _require_agents_api_enabled()
     try:
         store = get_agent_store(request)
-        await _sync_legacy_agents(store)
+        await _maybe_sync_legacy_agents(store)
 
         user_id = _current_user_id(request)
         records = await store.list_featured(user_id)
-        return AgentSelectorResponse(agents=[_record_to_response(r) for r in records])
+        # Selector dropdown only renders name/description/featured_order — same
+        # rationale as list_agents: skip soul + extras + self-heal.
+        return AgentSelectorResponse(
+            agents=[
+                _record_to_response(r, include_soul=False, include_extras=False)
+                for r in records
+            ]
+        )
     except HTTPException:
         raise
     except Exception as e:
@@ -427,12 +519,15 @@ async def get_agent(agent_id: str, request: Request) -> AgentResponse:
 
     try:
         store = get_agent_store(request)
-        await _sync_legacy_agents(store)
+        await _maybe_sync_legacy_agents(store)
         user_id = _current_user_id(request)
         record = await store.get_visible(agent_id, user_id)
         if record is None:
             raise HTTPException(status_code=404, detail=f"Agent '{agent_id}' not found")
-        response = _record_to_response(record, include_soul=True)
+        # Detail view legitimately reads SOUL.md + config.yaml — hop to a
+        # worker thread so the asyncio loop stays responsive under concurrent
+        # detail fetches.
+        response = await _record_to_response_async(record, include_soul=True, include_extras=True)
         favorite_ids = await store.list_favorite_ids(user_id)
         response.is_favorite = agent_id in favorite_ids
         return response
@@ -488,7 +583,7 @@ async def create_agent_endpoint(request: AgentCreateRequest, http_request: Reque
             }
         )
         logger.info("Created agent '%s' (%s) at %s", name, agent_id, agent_dir)
-        return _record_to_response(record, include_soul=True)
+        return await _record_to_response_async(record, include_soul=True, include_extras=True)
 
     except HTTPException:
         raise
@@ -569,7 +664,7 @@ async def update_agent(agent_id: str, request: AgentUpdateRequest, http_request:
             raise HTTPException(status_code=404, detail=f"Agent '{agent_id}' not found")
 
         logger.info("Updated agent '%s' (%s)", record["name"], agent_id)
-        return _record_to_response(record, include_soul=True)
+        return await _record_to_response_async(record, include_soul=True, include_extras=True)
 
     except HTTPException:
         raise
@@ -744,7 +839,7 @@ async def set_agent_featured(
 
     try:
         store = get_agent_store(request)
-        await _sync_legacy_agents(store)
+        await _maybe_sync_legacy_agents(store)
         record = await store.get_any(agent_id)
         if record is None:
             raise HTTPException(status_code=404, detail=f"Agent '{agent_id}' not found")
@@ -756,7 +851,7 @@ async def set_agent_featured(
         if updated is None:
             raise HTTPException(status_code=404, detail=f"Agent '{agent_id}' not found")
         logger.info("Set agent '%s' featured_order=%s by admin", agent_id, body.order)
-        return _record_to_response(updated, include_soul=True)
+        return await _record_to_response_async(updated, include_soul=True, include_extras=True)
     except HTTPException:
         raise
     except Exception as e:

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

@@ -65,6 +65,7 @@ from fastapi.responses import StreamingResponse
 # identical in connection handling, payload shape, and error mapping.
 from app.gateway.routers._roundtable_seed import ensure_roundtable_functional_agents
 from app.gateway.routers.multi_agent import (
+    _ROUNDTABLE_THREAD_METADATA,
     _auth_headers,
     _create_thread,
     _diagnose_exception,
@@ -133,6 +134,70 @@ def _normalize_options(raw: object) -> tuple[str, ...]:
     return tuple(sorted(o for o in out if o))
 
 
+def _count_past_clarifications(messages: list[dict[str, Any]]) -> int:
+    """统计当前 thread 内已经发出过的 ``ask_clarification`` 次数。
+
+    用来给 ``stream_intent`` 做"3 轮硬上限"兜底:
+      - 0-1 次:正常流程,模型自行决定继续问还是收敛
+      - 2 次(即将进入第 3 轮):prompt 追加系统提示,要求本轮必须 INTENT_READY
+      - ≥3 次:工具白名单移除 ``ask_clarification``,模型物理上无法再问
+
+    只数"已经成型"的 tool_call,不区分用户是否回答过——LangGraph 把 ask
+    工具的调用永久记录在 AIMessage.tool_calls 里,数它即可。
+    """
+    if not isinstance(messages, list):
+        return 0
+    count = 0
+    for msg in messages:
+        if not isinstance(msg, dict):
+            continue
+        if (msg.get("type") or "").lower() not in ("ai", "aimessage", "aimessagechunk"):
+            continue
+        for tc in msg.get("tool_calls") or []:
+            if (tc.get("name") or "").lower() == "ask_clarification":
+                count += 1
+    return count
+
+
+async def _fetch_thread_messages(
+    client: httpx.AsyncClient,
+    base: str,
+    headers: dict[str, str],
+    thread_id: str,
+) -> list[dict[str, Any]]:
+    """读取当前 thread checkpoint 的 messages 列表(best-effort)。
+
+    用 ASGITransport loopback 调 ``GET /api/threads/{tid}/state``,失败时返回
+    空列表——计数兜底是个"刹车",失败时退化成原有的纯 SOUL 自律行为,不阻断主流程。
+    """
+    try:
+        resp = await client.get(f"{base}/api/threads/{thread_id}/state", headers=headers)
+    except Exception as exc:
+        logger.warning("intent: get thread state failed (%s) — falling back to soul-only flow", exc)
+        return []
+    if resp.status_code != 200:
+        logger.warning(
+            "intent: get thread state %s returned %s — falling back to soul-only flow",
+            thread_id,
+            resp.status_code,
+        )
+        return []
+    try:
+        payload = resp.json()
+    except Exception:
+        return []
+    values = payload.get("values") if isinstance(payload, dict) else None
+    if not isinstance(values, dict):
+        return []
+    messages = values.get("messages")
+    return messages if isinstance(messages, list) else []
+
+
+# 硬上限:超过这个轮数,无论模型怎么想,都强制移除 ask_clarification 工具。
+# 与 SOUL.md §0 「硬规则 2」对齐——"绝不存在第 4 轮"。
+_MAX_CLARIFICATION_ROUNDS = 3
+
+
 def _is_repeated_clarification(
     messages: list[dict[str, Any]],
     *,
@@ -222,7 +287,7 @@ async def init_intent(request: Request, payload: dict = Body(...)) -> dict[str,
     # 看不到逐 token 增量。
     try:
         async with _make_loopback_client(request) as client:
-            thread_id = await _create_thread(client, base, headers)
+            thread_id = await _create_thread(client, base, headers, metadata=_ROUNDTABLE_THREAD_METADATA)
     except HTTPException:
         # 上游已用结构化方式失败,原样传给前端不需要二次包装。
         raise
@@ -279,33 +344,70 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
         # 因此既无须 trust_env=False 防代理，也不存在 loopback 40s 超时这种坑。
         async with _make_loopback_client(request) as client:
             try:
+                # —— 3 轮硬上限兜底 ——
+                # 在发起 run 之前先读一次 thread state,数当前 ask_clarification 已发出次数:
+                #   past=0/1 → 正常,不动 message,不动工具白名单
+                #   past=2(本轮即将变成第 3 轮)→ 给消息加前缀,强提示"必须 INTENT_READY"
+                #   past≥3 → 直接把 ask_clarification 也加进 excluded_tools,工具物理消失
+                # 这一步是个"刹车":SOUL 让模型自律,这里让系统兜底。
+                past_clarifications = _count_past_clarifications(
+                    await _fetch_thread_messages(client, base, headers, thread_id)
+                )
+
+                # 1. ask_clarification 的工具白名单——超过硬上限就移除。
+                base_excluded_tools = [
+                    "web_search",
+                    "present_files",
+                    "view_image",
+                    "agent_orchestration",
+                    "bash",
+                    "ls",
+                    "read_file",
+                    "write_file",
+                    "str_replace",
+                    "memory",
+                    "hindsight_recall",
+                    "hindsight_reflect",
+                    "hindsight_retain",
+                    "task",
+                    "write_todos",
+                ]
+                if past_clarifications >= _MAX_CLARIFICATION_ROUNDS:
+                    base_excluded_tools.append("ask_clarification")
+                    logger.info(
+                        "intent stream: past_clarifications=%d >= %d, force-disabling "
+                        "ask_clarification to break the loop",
+                        past_clarifications,
+                        _MAX_CLARIFICATION_ROUNDS,
+                    )
+
+                # 2. message 前缀——给模型一个明确的"本轮必须收敛"信号。
+                effective_message = message
+                if past_clarifications >= _MAX_CLARIFICATION_ROUNDS:
+                    effective_message = (
+                        f"[系统强制] 已澄清 {past_clarifications} 轮,触发硬上限。"
+                        f"本轮请直接输出 [INTENT_READY],缺的字段用 assumptions 兜底,"
+                        f"绝对不要再调任何工具,也不要再询问任何问题。\n\n"
+                        f"用户最新输入:{message}"
+                    )
+                elif past_clarifications >= _MAX_CLARIFICATION_ROUNDS - 1:
+                    effective_message = (
+                        f"[系统提示] 已澄清 {past_clarifications} 轮,这是最后一轮。"
+                        f"本轮请直接输出 [INTENT_READY],未明确的字段用 assumptions 兜底。\n\n"
+                        f"用户最新输入:{message}"
+                    )
+
                 body = _run_payload(
                     agent_name=INTENT_AGENT_ID,
                     model_name=model_name,
                     thread_id=thread_id,
-                    new_message=message,
-                    # Intent agent only ever needs ask_clarification. Block
+                    new_message=effective_message,
+                    # Intent agent only ever needs ask_clarification (and even that
+                    # is yanked once past_clarifications hits the cap above). Block
                     # everything else so the model can't drift into solutioning
-                    # (e.g. saving a design doc with write_file) — the model
-                    # tends to ignore SOUL prose, hard exclusion is reliable.
-                    # 仅保留 ask_clarification；其余工具见模块 docstring「工具白名单」。
-                    excluded_tools=[
-                        "web_search",
-                        "present_files",
-                        "view_image",
-                        "agent_orchestration",
-                        "bash",
-                        "ls",
-                        "read_file",
-                        "write_file",
-                        "str_replace",
-                        "memory",
-                        "hindsight_recall",
-                        "hindsight_reflect",
-                        "hindsight_retain",
-                        "task",
-                        "write_todos",
-                    ],
+                    # (e.g. saving a design doc with write_file) — SOUL prose is
+                    # not reliable on its own, hard exclusion is.
+                    excluded_tools=base_excluded_tools,
                     skill_stop_names=[],
                 )
 
@@ -338,7 +440,13 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
                         except json.JSONDecodeError as exc:
                             logger.warning("intent summary JSON parse failed: %s", exc)
                             summary = None
-                        ready_text = msg_text
+                        # SOUL §0 硬规则 3 要求 ```json``` 反引号闭合后不得再输出任何字符,
+                        # 但模型仍偶尔会在后面接 "那么我帮你列一下大纲……" 之类的尾巴。
+                        # 终态帧只回到 marker 块结尾 + 一个换行,把废话从 content 里裁掉——
+                        # 草稿持久化 / 跨设备重放 / 前端 splitIntentReadyBlock 的 after
+                        # 段都因此干净。流式期间用户仍会逐 token 看到那段废话,前端
+                        # Step1Panel 已经把 `parts.after` 隐藏掉,体验闭环。
+                        ready_text = msg_text[: m.end()].rstrip() + "\n"
                         break
                 if summary and ready_text:
                     final_frame = {

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

@@ -53,18 +53,28 @@ import os
 import re
 import time
 import traceback
+import uuid
 from typing import Any, AsyncIterator
 
 import httpx
 from deerflow.config import get_app_config
 from deerflow.tools.builtins.clarification_utils import resolve_allow_multiple
+from deerflow.runtime.user_context import get_effective_user_id
+from deerflow.utils.time import now_iso
 from fastapi import APIRouter, Body, HTTPException, Request
 from fastapi.responses import StreamingResponse
+from langgraph.checkpoint.base import empty_checkpoint
 
 logger = logging.getLogger(__name__)
 router = APIRouter(prefix="/api/multi-agent", tags=["multi-agent"])
 
 
+# 圆桌（intent / recommend / 席位 / 协调）创建的 thread 都是功能型内部线程,
+# 不是用户的普通对话。打上这个 metadata,前端「最近的对话」列表(recent-chat-list.tsx)
+# 据 `metadata.thread_type` 过滤掉,与 scheduler / bootstrap 系统会话同一套机制。
+_ROUNDTABLE_THREAD_METADATA: dict[str, Any] = {"thread_type": "roundtable", "system": True}
+
+
 # ---------------------------------------------------------------------------
 # 共享:异常诊断器(intent/recommend/multi-agent 共用)
 # ---------------------------------------------------------------------------
@@ -285,9 +295,22 @@ def _auth_headers(request: Request) -> dict[str, str]:
     return {"Authorization": token, "Content-Type": "application/json"}
 
 
-async def _create_thread(client: httpx.AsyncClient, base: str, headers: dict[str, str]) -> str:
-    """``POST /api/threads`` → 新 thread_id（intent/recommend/init 席位均用）。"""
-    resp = await client.post(f"{base}/api/threads", json={}, headers=headers)
+async def _create_thread(
+    client: httpx.AsyncClient,
+    base: str,
+    headers: dict[str, str],
+    *,
+    metadata: dict[str, Any] | None = None,
+) -> str:
+    """``POST /api/threads`` → 新 thread_id（intent/recommend/init 席位均用）。
+
+    ``metadata`` 写入 thread 元数据,圆桌调用方传 ``_ROUNDTABLE_THREAD_METADATA``
+    把这些功能型线程标记成可被前端「最近的对话」过滤掉的系统会话。
+    """
+    body: dict[str, Any] = {}
+    if metadata:
+        body["metadata"] = metadata
+    resp = await client.post(f"{base}/api/threads", json=body, headers=headers)
     resp.raise_for_status()
     return resp.json()["thread_id"]
 
@@ -323,6 +346,181 @@ async def _create_agent(
     return False, resp.text
 
 
+# ---------------------------------------------------------------------------
+# 同进程直调 helpers(仅供 init_multi_agent 用,不要扩散到其它地方)
+#
+# init 阶段串了 N+1 个 thread create + N 个 GET /api/agents + 1 个 POST /api/agents,
+# 用 HTTP loopback 串走每次都得 TCP + 鉴权中间件 + Pydantic 反序列化一遍。每个
+# 调用 10-30ms,加起来对前端就是「点进 Step 2 后席位灯一颗一颗亮」的卡顿来源。
+#
+# 这里把这三个调用改成直接吃 app.state 里的 store/checkpointer —— 跑在同一个事
+# 件循环里,省掉 TCP + 序列化 + 中间件链。鉴权 / 校验 / 文件副作用按需在直调
+# helper 里显式做掉,不要"绕过"任何一项。
+#
+# 设计边界:这里只覆盖 init 阶段。SSE 长连接(/runs/stream / state)留在 HTTP
+# loopback,因为它们本就是别的进程模型(LangGraph 子树),抽出来风险大收益小。
+# ---------------------------------------------------------------------------
+
+
+def _resolve_user_id(request: Request) -> str:
+    """与 routers/agents._current_user_id 完全等价,直调路径里用。"""
+    user = getattr(request.state, "user", None)
+    if user is not None:
+        return str(user.id)
+    return get_effective_user_id()
+
+
+async def _create_thread_direct(request: Request, *, metadata: dict[str, Any] | None = None) -> str:
+    """同进程版的 ``POST /api/threads``。
+
+    完全镜像 ``routers/threads.create_thread`` 的写盘动作(thread_meta +
+    空 checkpoint),但跳过 TCP + Pydantic 反序列化。仅供 init 阶段批量建席位
+    线程用,失败时抛 HTTPException 由 SSE 帧上报。
+
+    ``metadata`` 写入 thread 元数据;init 传 ``_ROUNDTABLE_THREAD_METADATA`` 把
+    席位 / 协调线程标记成系统会话,从前端「最近的对话」里隐藏。
+    """
+    from app.gateway.deps import get_checkpointer, get_thread_store
+
+    checkpointer = get_checkpointer(request)
+    thread_store = get_thread_store(request)
+    thread_id = str(uuid.uuid4())
+    now = now_iso()
+
+    try:
+        await thread_store.create(thread_id, assistant_id=None, metadata=metadata or {})
+    except Exception:
+        logger.exception("[init-direct] thread_meta create failed for %s", thread_id)
+        raise HTTPException(status_code=500, detail="Failed to create thread")
+
+    config = {"configurable": {"thread_id": thread_id, "checkpoint_ns": ""}}
+    ckpt_metadata = {
+        "step": -1,
+        "source": "input",
+        "writes": None,
+        "parents": {},
+        "created_at": now,
+    }
+    try:
+        await checkpointer.aput(config, empty_checkpoint(), ckpt_metadata, {})
+    except Exception:
+        logger.exception("[init-direct] checkpoint create failed for %s", thread_id)
+        raise HTTPException(status_code=500, detail="Failed to create thread")
+    return thread_id
+
+
+async def _get_agents_batch_direct(
+    request: Request,
+    agent_ids: list[str],
+) -> dict[str, dict[str, Any]]:
+    """同进程版的 N 个并行 ``GET /api/agents/{id}`` 合并为 1 次 DB 查询。
+
+    返回 ``{agent_id: {id, name, description}}``。**严格模式**:任一席位对当前
+    用户不可见就抛 404 —— 与原 HTTP loopback 的 ``_get_agent`` 行为一致(原路径
+    单查 404 会 raise,在外层被 except 捕获并 yield error 帧),不要悄悄回退。
+    """
+    from app.gateway.deps import get_agent_store
+
+    store = get_agent_store(request)
+    user_id = _resolve_user_id(request)
+    records = await store.list_by_ids(agent_ids, user_id)
+    out: dict[str, dict[str, Any]] = {}
+    for r in records:
+        rid = str(r.get("id") or "").lower()
+        if not rid:
+            continue
+        out[rid] = {
+            "id": rid,
+            "name": str(r.get("name") or rid),
+            "description": str(r.get("description") or ""),
+        }
+    missing = [sid for sid in agent_ids if sid.lower() not in out]
+    if missing:
+        raise HTTPException(
+            status_code=404,
+            detail=f"Agent(s) not visible to current user: {', '.join(missing)}",
+        )
+    return out
+
+
+async def _create_coordinator_direct(
+    request: Request,
+    *,
+    agent_id: str,
+    name: str,
+    description: str,
+    soul: str,
+    model: str | None,
+    skills: list[str] | None,
+) -> tuple[bool, dict[str, Any] | str]:
+    """同进程版的协调 agent 创建。
+
+    完整复刻 ``routers/agents.create_agent_endpoint`` 的关键副作用:
+      - ``agents_api.enabled`` master switch
+      - 显示名校验(空白则 422 → 这里转 (False, msg))
+      - 重复 id 检查(409 → (False, msg))
+      - 写 ``config.yaml`` + ``SOUL.md``(用 ``asyncio.to_thread`` 防同步 IO 阻塞)
+      - ``agent_store.create`` 入表
+
+    返回 ``(True, record)`` 或 ``(False, error_text)``;调用方决定要不要发 SSE error 帧。
+    """
+    from app.gateway.deps import get_agent_store
+    from app.gateway.routers.agents import (
+        _require_agents_api_enabled,
+        _write_agent_config,
+        _write_agent_soul,
+    )
+    from deerflow.config.paths import get_paths
+
+    _require_agents_api_enabled()
+    normalized_name = (name or "").strip()
+    if not normalized_name:
+        return False, "Agent name cannot be empty"
+
+    store = get_agent_store(request)
+    if await store.get_any(agent_id) is not None:
+        return False, f"Agent id '{agent_id}' already exists"
+
+    agent_dir = get_paths().agent_dir(agent_id)
+    if agent_dir.exists():
+        return False, f"Agent id '{agent_id}' already exists"
+
+    try:
+        # config.yaml / SOUL.md 是同步 IO —— 全部丢线程池,不阻塞 init SSE 的发帧节奏。
+        await asyncio.to_thread(
+            _write_agent_config,
+            agent_id,
+            name=normalized_name,
+            description=description or "",
+            skills=skills,
+            model=model,
+            tool_groups=None,
+        )
+        await asyncio.to_thread(_write_agent_soul, agent_id, soul)
+        record = await store.create(
+            {
+                "id": agent_id,
+                "name": normalized_name,
+                "description": description or "",
+                "user_id": _resolve_user_id(request),
+                "published": False,
+            }
+        )
+        return True, record
+    except Exception as exc:
+        # 与 routers/agents 失败回滚一致 —— 文件已写出但 DB 失败时清掉 dir,避免下次
+        # `agent_dir.exists()` 误判为"已存在"导致后续 init 永远 409。
+        try:
+            if agent_dir.exists():
+                import shutil
+
+                await asyncio.to_thread(shutil.rmtree, agent_dir)
+        except Exception:
+            logger.exception("[init-direct] cleanup after coordinator create failure")
+        logger.exception("[init-direct] coordinator create failed")
+        return False, str(exc)
+
+
 async def _append_thread_message(
     client: httpx.AsyncClient,
     base: str,
@@ -456,6 +654,8 @@ def _run_payload(
     new_message: str,
     excluded_tools: list[str],
     skill_stop_names: list[str],
+    thinking_enabled: bool = True,
+    reasoning_effort: str = "medium",
 ) -> dict[str, Any]:
     """构造 ``POST /api/threads/{thread_id}/runs/stream`` 的请求体。
 
@@ -463,6 +663,10 @@ def _run_payload(
     - ``excluded_tools``：按路由裁剪工具面（intent 仅澄清、special 禁派活等）。
     - ``skill_stop_names``：leader 侧配合 SkillStopMiddleware 在 ``agent_orchestration``
       调用后截断图，以便网关从 AIMessage.tool_calls 解析派活列表。
+    - ``thinking_enabled`` / ``reasoning_effort``：默认开 CoT + medium，适合 Step 1 澄清
+      与子智能体交付这类需要推理的场景。**推荐弹窗与总控派活**是短决策任务，调用方
+      应传 ``thinking_enabled=False`` + ``reasoning_effort="low"``，省掉模型先吐
+      reasoning tokens 才开口的时间，显著降低前端"调度中"首字节延迟。
     """
     return {
         "input": {
@@ -479,11 +683,17 @@ def _run_payload(
             "agent_name": agent_name,
             "model_name": model_name,
             "mode": "pro",
-            "reasoning_effort": "medium",
-            "thinking_enabled": True,
+            "reasoning_effort": reasoning_effort,
+            "thinking_enabled": thinking_enabled,
             "is_plan_mode": False,
             "subagent_enabled": False,
             "thread_id": thread_id,
+            # 圆桌四个 router(intent / recommend / leader / special)都是临时任务,
+            # 用户个人记忆与本次研讨无关。MemoryMiddleware 的 Hindsight recall 是
+            # 同步阻塞 HTTP 调用(单次 timeout 15s),如果记忆服务慢/不通会让前端
+            # "调度中"卡很久;这里默认跳过,builtin USER.md/MEMORY.md 仍通过系统
+            # 提示注入不受影响。
+            "memory_recall_disabled": True,
         },
         "stream_mode": ["messages-tuple", "values"],
         "stream_subgraphs": True,
@@ -534,111 +744,132 @@ async def init_multi_agent(request: Request, payload: dict = Body(...)) -> Strea
         _validate_agent_id(name, "agent_names entry")
     _validate_agent_id(main_agent_name, "main_agent_name")
 
-    base = _loopback_base(request)
-    headers = _auth_headers(request)
-
     def _frame(payload: dict[str, Any]) -> str:
         return f"data: {json.dumps(payload, ensure_ascii=False)}\n\n"
 
     async def stream_generator() -> AsyncIterator[str]:
-        # TCP loopback 到 127.0.0.1:<gateway port>。client 配置见 _make_loopback_client：
-        # trust_env=False 防 HTTP_PROXY 拦截,timeout=None 容纳 SSE 长连接。
-        async with _make_loopback_client(request) as client:
-            try:
-                # Per-seat prepare task — same body as before, but we no longer
-                # await them with asyncio.gather. Instead, asyncio.as_completed
-                # surfaces them in REAL completion order so we can emit a
-                # seat_ready frame the moment each finishes.
-                async def _prepare_sub_agent(sub_id: str) -> tuple[str, str, str, str]:
-                    info = await _get_agent(client, base, headers, sub_id)
-                    display_name = str(info.get("name") or sub_id)
-                    description = info.get("description", "") or ""
-                    thread_id = await _create_thread(client, base, headers)
-                    return sub_id.lower(), thread_id, display_name, description
+        # init 阶段 fast path:全部走同进程直调(_create_thread_direct /
+        # _get_agents_batch_direct / _create_coordinator_direct),不再走 HTTP
+        # loopback —— TCP + 鉴权中间件 + Pydantic 反序列化每次 10-30ms,N+1 个席位
+        # 累计就是"点进 Step 2 后席位一颗颗亮"的卡顿来源。
+        try:
+            t_init_start = time.perf_counter()
+            # 1) 一次 DB 查询拿到所有席位的 name/description,替代之前的 N 个并行
+            #    GET /api/agents/{id}。SOUL 拼接需要描述,这里一次性拿齐。
+            agent_meta = await _get_agents_batch_direct(request, agent_names)
+            t_meta_done = time.perf_counter()
 
-                tasks = [asyncio.create_task(_prepare_sub_agent(sid)) for sid in agent_names]
-                prepared: list[tuple[str, str, str, str]] = []
-                try:
-                    for fut in asyncio.as_completed(tasks):
-                        sid, thread_id, display_name, description = await fut
-                        prepared.append((sid, thread_id, display_name, description))
-                        yield _frame({
-                            "event": "seat_ready",
-                            "agent_name": sid,
-                            "thread_id": thread_id,
-                            "display_name": display_name,
-                            "description": description,
-                        })
-                except Exception:
-                    # Cancel any still-pending sibling so we don't leak tasks
-                    # past the AsyncClient close.
-                    for t in tasks:
-                        if not t.done():
-                            t.cancel()
-                    raise
-
-                agent_threads: dict[str, str] = {sid: thread_id for sid, thread_id, _, _ in prepared}
-                descriptions: list[str] = [
-                    f"- agent_name: {sid}（{display_name}）：{description}"
-                    for sid, _, display_name, description in prepared
-                ]
-
-                # 动态 SOUL：把本次 init 的席位列表写进总控 prompt，供 agent_orchestration 派活。
-                soul = (
-                    "#核心定位：\n"
-                    "用途：根据用户的任务目标,调用其他智能体干活,只负责协调汇总,"
-                    "当你觉得已有信息支持回答用户问题时候,停止调度,并输出最后的结果\n"
-                    "#调度方法：\n"
-                    "通过使用agent_orchestration技能调用其他智能体干活,使用技能时传入agent_name和需要做的活。"
-                    "注意：agent_name 必须严格使用下方列出的英文 id（例如 roundtable-intelligence,"
-                    "或形如 02714cbe6e64406a8a4659f0e3d9c0b8 这样的 UUID）,"
-                    "不要把括号里的中文显示名拼进 agent_name 里。\n"
-                    "#关键约束(必读)：\n"
-                    "1. **不要对同一子智能体重复派同样的活**：每次派活前回顾对话历史,"
-                    "如果某个子智能体已经针对你正要派的任务完成过交付("
-                    "看到形如\"子智能体[名称]完成XX工作,交付内容为:...\"的 human 消息),"
-                    "就直接基于已有交付推进,**不要重复派活**,除非你确实有**具体的、不一样的**补充问题。\n"
-                    "2. **看到交付要会用**：子智能体的交付内容会作为 human 消息追加到你的对话历史里,"
-                    "你能完整看见。当各方信息已足够形成结论时,**直接综合输出最终方案**(此时不再调用任何工具),"
-                    "让讨论收敛。\n"
-                    "3. **避免无意义循环**：如果连续两轮都没有新信息流入,立刻停止派活并基于现有素材给出最终结论。\n"
-                    "#可调度智能体：\n"
-                    + "\n".join(descriptions)
+            # 2) 每个席位 thread 并行创建(直调 thread_store + checkpointer),
+            #    用 asyncio.as_completed 保持"先完先发 seat_ready"的体验。
+            async def _prepare_sub_agent(sub_id: str) -> tuple[str, str, str, str]:
+                thread_id = await _create_thread_direct(request, metadata=_ROUNDTABLE_THREAD_METADATA)
+                meta = agent_meta.get(sub_id.lower(), {})
+                return (
+                    sub_id.lower(),
+                    thread_id,
+                    str(meta.get("name") or sub_id),
+                    str(meta.get("description") or ""),
                 )
 
-                coord_thread = await _create_thread(client, base, headers)
-                agent_threads[main_agent_name] = coord_thread
-                ok, result = await _create_agent(
-                    client,
-                    base,
-                    headers,
+            tasks = [asyncio.create_task(_prepare_sub_agent(sid)) for sid in agent_names]
+            prepared: list[tuple[str, str, str, str]] = []
+            try:
+                for fut in asyncio.as_completed(tasks):
+                    sid, thread_id, display_name, description = await fut
+                    prepared.append((sid, thread_id, display_name, description))
+                    yield _frame({
+                        "event": "seat_ready",
+                        "agent_name": sid,
+                        "thread_id": thread_id,
+                        "display_name": display_name,
+                        "description": description,
+                    })
+            except Exception:
+                # Cancel any still-pending sibling so partially-prepared seats
+                # don't keep running past the stream close.
+                for t in tasks:
+                    if not t.done():
+                        t.cancel()
+                raise
+            t_seats_done = time.perf_counter()
+
+            agent_threads: dict[str, str] = {sid: thread_id for sid, thread_id, _, _ in prepared}
+            descriptions: list[str] = [
+                f"- agent_name: {sid}（{display_name}）：{description}"
+                for sid, _, display_name, description in prepared
+            ]
+
+            # 动态 SOUL：把本次 init 的席位列表写进总控 prompt，供 agent_orchestration 派活。
+            soul = (
+                "#核心定位：\n"
+                "用途：根据用户的任务目标,调用其他智能体干活,只负责协调汇总,"
+                "当你觉得已有信息支持回答用户问题时候,停止调度,并输出最后的结果\n"
+                "#调度方法：\n"
+                "通过使用agent_orchestration技能调用其他智能体干活,使用技能时传入agent_name和需要做的活。"
+                "注意：agent_name 必须严格使用下方列出的英文 id（例如 roundtable-intelligence,"
+                "或形如 02714cbe6e64406a8a4659f0e3d9c0b8 这样的 UUID）,"
+                "不要把括号里的中文显示名拼进 agent_name 里。\n"
+                "#关键约束(必读)：\n"
+                "1. **不要对同一子智能体重复派同样的活**：每次派活前回顾对话历史,"
+                "如果某个子智能体已经针对你正要派的任务完成过交付("
+                "看到形如\"子智能体[名称]完成XX工作,交付内容为:...\"的 human 消息),"
+                "就直接基于已有交付推进,**不要重复派活**,除非你确实有**具体的、不一样的**补充问题。\n"
+                "2. **看到交付要会用**：子智能体的交付内容会作为 human 消息追加到你的对话历史里,"
+                "你能完整看见。当各方信息已足够形成结论时,**直接综合输出最终方案**(此时不再调用任何工具),"
+                "让讨论收敛。\n"
+                "3. **避免无意义循环**：如果连续两轮都没有新信息流入,立刻停止派活并基于现有素材给出最终结论。\n"
+                "#可调度智能体：\n"
+                + "\n".join(descriptions)
+            )
+
+            # 3) coordinator thread + agent 并行创建。thread 与 agent 之间无依赖,
+            #    asyncio.gather 比串行省一个直调往返的开销。
+            coord_thread_task = asyncio.create_task(
+                _create_thread_direct(request, metadata=_ROUNDTABLE_THREAD_METADATA)
+            )
+            coord_agent_task = asyncio.create_task(
+                _create_coordinator_direct(
+                    request,
+                    agent_id=main_agent_name,
                     name=main_agent_name,
                     description="你是总控控制协调智能体,负责协调其他智能体干活",
+                    soul=soul,
                     model=model,
                     skills=["agent_orchestration"],
-                    soul=soul,
-                    agent_id=main_agent_name,
                 )
-                if not ok and isinstance(result, str) and "exists" not in result:
-                    yield _frame({"event": "error", "detail": f"create coordinator agent failed: {result}"})
-                    return
+            )
+            coord_thread = await coord_thread_task
+            agent_threads[main_agent_name] = coord_thread
+            ok, result = await coord_agent_task
+            if not ok and isinstance(result, str) and "exists" not in result:
+                yield _frame({"event": "error", "detail": f"create coordinator agent failed: {result}"})
+                return
+            t_coord_done = time.perf_counter()
+            logger.info(
+                "[init-direct] meta=%.3fs seats=%.3fs coord=%.3fs total=%.3fs seats_count=%d",
+                t_meta_done - t_init_start,
+                t_seats_done - t_meta_done,
+                t_coord_done - t_seats_done,
+                t_coord_done - t_init_start,
+                len(prepared),
+            )
 
-                yield _frame({
-                    "event": "coordinator_ready",
-                    "main_agent_name": main_agent_name,
-                    "thread_id": coord_thread,
-                })
-                yield _frame({
-                    "event": "init_done",
-                    "main_agent_name": main_agent_name,
-                    "thread_ids": agent_threads,
-                })
-            except HTTPException as exc:
-                logger.warning("multi-agent init HTTPException: %s %s", exc.status_code, exc.detail)
-                yield _frame({"event": "error", "detail": f"[{exc.status_code}] {exc.detail}"})
-            except Exception as exc:
-                logger.exception("multi-agent init unexpected failure")
-                yield _frame({"event": "error", "detail": f"unexpected: {exc!r}"})
+            yield _frame({
+                "event": "coordinator_ready",
+                "main_agent_name": main_agent_name,
+                "thread_id": coord_thread,
+            })
+            yield _frame({
+                "event": "init_done",
+                "main_agent_name": main_agent_name,
+                "thread_ids": agent_threads,
+            })
+        except HTTPException as exc:
+            logger.warning("multi-agent init HTTPException: %s %s", exc.status_code, exc.detail)
+            yield _frame({"event": "error", "detail": f"[{exc.status_code}] {exc.detail}"})
+        except Exception as exc:
+            logger.exception("multi-agent init unexpected failure")
+            yield _frame({"event": "error", "detail": f"unexpected: {exc!r}"})
 
     return StreamingResponse(
         stream_generator(),
@@ -825,6 +1056,11 @@ async def _leader_run(
         new_message=new_message,
         excluded_tools=["web_search"],
         skill_stop_names=effective_stop_names,
+        # 总控派活是"下一步派谁、什么任务"的短决策，不需要 CoT。
+        # 关掉 thinking + low reasoning 让"调度中"几乎瞬间转入"正在派活"，
+        # 子智能体仍保留 thinking（在 _special_run 里默认 True），保证交付质量。
+        thinking_enabled=False,
+        reasoning_effort="low",
     )
     t_payload_built = time.perf_counter()
     logger.info(

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

@@ -49,6 +49,7 @@ from fastapi.responses import StreamingResponse
 
 from app.gateway.routers._roundtable_seed import ensure_roundtable_functional_agents
 from app.gateway.routers.multi_agent import (
+    _ROUNDTABLE_THREAD_METADATA,
     _auth_headers,
     _create_thread,
     _flatten_content,
@@ -237,7 +238,7 @@ async def stream_recommend(request: Request, payload: dict = Body(...)) -> Strea
         async with _make_loopback_client(request) as client:
             try:
                 # 每次推荐独立 thread；prompt 已包含完整 intent + 候选列表，无需多轮。
-                thread_id = await _create_thread(client, base, headers)
+                thread_id = await _create_thread(client, base, headers, metadata=_ROUNDTABLE_THREAD_METADATA)
 
                 body = _run_payload(
                     agent_name=RECOMMENDER_AGENT_ID,
@@ -253,6 +254,12 @@ async def stream_recommend(request: Request, payload: dict = Body(...)) -> Strea
                         "ask_clarification",
                     ],
                     skill_stop_names=[],
+                    # 推荐是"从候选池挑 2-6 个"的短分类任务，不需要 CoT。
+                    # 关掉 thinking + 用 low reasoning，避免模型先吐数千 reasoning
+                    # tokens 才开始输出 [RECOMMEND_READY]，前端"等待推荐结果..."
+                    # 从 10-30s 降到 2-5s。
+                    thinking_enabled=False,
+                    reasoning_effort="low",
                 )
 
                 captured: list[str] = []

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

@@ -137,7 +137,17 @@ class MemoryMiddleware(AgentMiddleware[MemoryMiddlewareState]):
 
         召回结果包在 ``<memory-context>`` 标签里并带系统提示,告诉模型这是
         背景信息而非用户的新输入。
+
+        **跳过开关**:若 ``runtime.context.memory_recall_disabled`` 为真,直接
+        返回 None。圆桌规划等临时性任务(intent / recommend / leader / special)
+        不需要从用户个人记忆里召回,Hindsight HTTP 同步阻塞会显著增加 LLM 首
+        字节延迟(前端"调度中"卡顿);由路由层通过 ``_run_payload`` 的 context
+        字段统一关闭。
         """
+        runtime = getattr(request, "runtime", None)
+        if runtime is not None and getattr(runtime, "context", None):
+            if runtime.context.get("memory_recall_disabled"):
+                return None
         if not self._injection_enabled:
             return None
         config = self._memory_config or get_memory_config()

offline-backend-20260512/backend/packages/harness/deerflow/config/agents_config.py

@@ -1,7 +1,9 @@
 """Configuration and loaders for custom agents."""
 
 import logging
+import os
 import re
+import threading
 from typing import Any
 
 import yaml
@@ -13,6 +15,25 @@ logger = logging.getLogger(__name__)
 
 SOUL_FILENAME = "SOUL.md"
 AGENT_ID_PATTERN = re.compile(r"^[A-Za-z0-9_-]+$")
+
+# Per-file mtime caches. These are read on every /api/agents list/detail request
+# (one read per agent) — caching keyed by (path, mtime_ns, size) is correct
+# because any edit to the file updates mtime, and an external tool re-writing
+# the same bytes doesn't actually change the result we'd return. Cache size is
+# unbounded but lookup is O(1) and entries are tiny (one parsed dict / string
+# per agent file); for the order of N≤10⁴ agents the memory cost is negligible.
+_config_cache: dict[str, tuple[tuple[int, int], "AgentConfig | None"]] = {}
+_soul_cache: dict[str, tuple[tuple[int, int], str | None]] = {}
+_cache_lock = threading.Lock()
+
+
+def _file_stat_key(path) -> tuple[int, int] | None:
+    """Return (mtime_ns, size) for cache invalidation, or None when missing."""
+    try:
+        st = os.stat(path)
+    except OSError:
+        return None
+    return (st.st_mtime_ns, st.st_size)
 # Backward-compatible alias. Historically this validated the display name;
 # after the DB-backed agent table, display names can be Chinese and duplicated.
 # The pattern now applies only to storage/runtime ids.
@@ -94,6 +115,18 @@ def load_agent_config(agent_id: str | None) -> AgentConfig | None:
         )
         return None
 
+    # mtime-based cache. The config file is rewritten every time the agent is
+    # edited via PUT /api/agents/{id}, which bumps mtime — so the next read
+    # picks up the change automatically. Missing config_file = key None below,
+    # bypassing the cache (cheap path: just return AgentConfig with id/name).
+    cache_path_key = str(config_file)
+    stat_key = _file_stat_key(config_file) if config_file.exists() else None
+    if stat_key is not None:
+        with _cache_lock:
+            cached = _config_cache.get(cache_path_key)
+        if cached is not None and cached[0] == stat_key:
+            return cached[1]
+
     data: dict[str, Any] = {}
     if config_file.exists():
         try:
@@ -112,7 +145,11 @@ def load_agent_config(agent_id: str | None) -> AgentConfig | None:
     known_fields = set(AgentConfig.model_fields.keys())
     data = {k: v for k, v in data.items() if k in known_fields}
 
-    return AgentConfig(**data)
+    result = AgentConfig(**data)
+    if stat_key is not None:
+        with _cache_lock:
+            _config_cache[cache_path_key] = (stat_key, result)
+    return result
 
 
 def load_agent_soul(agent_id: str | None) -> str | None:
@@ -131,8 +168,21 @@ def load_agent_soul(agent_id: str | None) -> str | None:
     soul_path = agent_dir / SOUL_FILENAME
     if not soul_path.exists():
         return None
+    # mtime-based cache, same rationale as load_agent_config: SOUL.md is
+    # rewritten on edit (PUT /api/agents/{id}), which bumps mtime.
+    cache_path_key = str(soul_path)
+    stat_key = _file_stat_key(soul_path)
+    if stat_key is not None:
+        with _cache_lock:
+            cached = _soul_cache.get(cache_path_key)
+        if cached is not None and cached[0] == stat_key:
+            return cached[1]
     content = soul_path.read_text(encoding="utf-8").strip()
-    return content or None
+    result = content or None
+    if stat_key is not None:
+        with _cache_lock:
+            _soul_cache[cache_path_key] = (stat_key, result)
+    return result
 
 
 def list_custom_agents() -> list[AgentConfig]: