From 39def1c20b316f59769c54cf535b3bc5cef59dde Mon Sep 17 00:00:00 2001
From: xuhd <hdxu@kland.com.cn>
Date: Sat, 16 May 2026 15:50:15 +0800
Subject: [PATCH] =?UTF-8?q?feat(skills):=20curator=20=E5=81=A5=E5=A3=AE?=
 =?UTF-8?q?=E6=80=A7=E5=8A=A0=E5=9B=BA=20+=20=E8=AF=84=E5=AE=A1=E6=A8=A1?=
 =?UTF-8?q?=E5=9E=8B=E5=8F=AF=E9=85=8D=E7=BD=AE=20+=20=E6=89=8B=E5=8A=A8?=
 =?UTF-8?q?=E8=A7=A6=E5=8F=91=E4=BD=93=E9=AA=8C?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

调度健壮性:
- 技能进化总开关 skill_evolution.enabled 真正生效;CuratorConfig
  与 CuratorSubConfig 默认 enabled 改为 false(显式 opt-in)
- 新增 per-user 运行锁(.curator_state 的 running/started_at),
  防定时器与手动触发并发重复运行;6h stale 检测防进程残留死锁
- maybe_run_curator 改用 ThreadPoolExecutor 限并发(上限 5),
  防首次部署所有用户同时到期的惊群
- 首次运行持续失败时按 6h 退避重试,不再每个 tick 烧 LLM 配额
- should_run_now 时间解析容忍 naive 时间串;损坏的 last_run_at
  触发重跑而非永久死锁
- SkillStore 新增 list_published_names 批量接口,curator 一轮
  只查一次 published,取代逐技能查库

LLM 评审超时:
- _run_llm_review 给 agent.invoke 套整体超时
  (curator.review_timeout_seconds,默认 600s),超时中止本次评审
  并在 curator 日志输出,curator 正常收尾、释放运行锁
- copy_context 把 user 上下文传入工作线程

模型可配置:
- 新增 curator.model_name;curator 评审与技能安全审核共用该模型,
  未配置则回退主模型
- 移除独立的 moderation_model_name 配置
- 前端进化面板新增模型下拉选择器

手动触发体验:
- /api/curator/run 派发后台线程前先预检,返回 {triggered, reason},
  不再无论开关状态都提示"已开始"
- 前端:总开关/curator 开关未开时禁用"立即执行"按钮,
  被预检拦下时按原因(开关未开 / 已有任务在跑)给出准确提示

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .../skills/skill-evolution-panel.tsx          |  70 ++++-
 frontend-web/src/core/curator/api.ts          |  10 +-
 frontend-web/src/core/curator/types.ts        |   4 +-
 frontend-web/src/core/i18n/locales/en-US.ts   |  13 +-
 frontend-web/src/core/i18n/locales/types.ts   |   9 +-
 frontend-web/src/core/i18n/locales/zh-CN.ts   |   9 +-
 offline-backend-20260512/backend/CLAUDE.md    |  49 +++
 .../backend/app/gateway/routers/skills.py     |  25 +-
 .../deerflow/config/skill_evolution_config.py |  13 +-
 .../deerflow/persistence/skills/base.py       |  10 +
 .../deerflow/persistence/skills/memory.py     |   3 +
 .../deerflow/persistence/skills/sql.py        |   6 +
 .../harness/deerflow/skills/curator.py        | 281 ++++++++++++++++--
 .../deerflow/skills/security_scanner.py       |   6 +-
 .../tests/test_skill_evolution_runtime.py     |   2 +-
 .../backend/tests/test_skills_multitenancy.py |   3 +-
 16 files changed, 448 insertions(+), 65 deletions(-)

diff --git a/frontend-web/src/components/workspace/skills/skill-evolution-panel.tsx b/frontend-web/src/components/workspace/skills/skill-evolution-panel.tsx
index 1cd87920..b9549ed6 100644
--- a/frontend-web/src/components/workspace/skills/skill-evolution-panel.tsx
+++ b/frontend-web/src/components/workspace/skills/skill-evolution-panel.tsx
@@ -21,6 +21,13 @@ import {
   CardTitle,
 } from "@/components/ui/card";
 import { Input } from "@/components/ui/input";
+import {
+  Select,
+  SelectContent,
+  SelectItem,
+  SelectTrigger,
+  SelectValue,
+} from "@/components/ui/select";
 import { Switch } from "@/components/ui/switch";
 import {
   useCuratorConfig,
@@ -31,6 +38,7 @@ import {
 } from "@/core/curator/hooks";
 import type { CuratorConfig } from "@/core/curator/types";
 import { useI18n } from "@/core/i18n/hooks";
+import { useModels } from "@/core/models/hooks";
 import { cn } from "@/lib/utils";
 
 export function SkillEvolutionPanel() {
@@ -68,6 +76,7 @@ function StatusCard() {
   const { t } = useI18n();
   const tt = t.settings.skillEvolution.status;
   const { status, isLoading } = useCuratorStatus();
+  const { config } = useCuratorConfig();
   const runCurator = useRunCurator();
 
   const lastRunDisplay = useMemo(() => {
@@ -79,10 +88,25 @@ function StatusCard() {
     }
   }, [status?.last_run_at, tt.lastRunNever]);
 
+  // 总开关或 curator 子开关任一关闭时,手动触发跑不起来 —— 直接禁用按钮。
+  const runDisabled = !config || !config.enabled || !config.curator.enabled;
+
   const handleRun = async () => {
     try {
-      await runCurator.mutateAsync();
-      toast.success(tt.runStarted);
+      const res = await runCurator.mutateAsync();
+      if (res.triggered) {
+        toast.success(tt.runStarted);
+      } else {
+        // 后端预检拦下了(开关未开 / 已有任务在跑)—— 给出对应原因,而不是误报"已开始"。
+        const reasonMsg: Record<string, string> = {
+          evolution_disabled: tt.runBlockedEvolutionDisabled,
+          curator_disabled: tt.runBlockedCuratorDisabled,
+          already_running: tt.runBlockedAlreadyRunning,
+        };
+        toast.warning(
+          reasonMsg[res.reason ?? ""] ?? tt.runBlockedEvolutionDisabled,
+        );
+      }
     } catch (err) {
       toast.error(err instanceof Error ? err.message : String(err));
     }
@@ -107,7 +131,8 @@ function StatusCard() {
         <Button
           size="sm"
           onClick={handleRun}
-          disabled={runCurator.isPending}
+          disabled={runCurator.isPending || runDisabled}
+          title={runDisabled ? tt.runDisabledHint : undefined}
         >
           {runCurator.isPending ? (
             <Loader2Icon className="size-4 animate-spin" />
@@ -245,11 +270,15 @@ function StatRow({
   );
 }
 
+/** Select 不允许空字符串 value,用此哨兵代表"用主模型"(对应 model_name=null)。 */
+const DEFAULT_MODEL_VALUE = "__default__";
+
 function ConfigCard() {
   const { t } = useI18n();
   const tt = t.settings.skillEvolution.config;
   const { config, isLoading } = useCuratorConfig();
   const updateConfig = useUpdateCuratorConfig();
+  const { models } = useModels();
 
   const [draft, setDraft] = useState<CuratorConfig | null>(null);
   useEffect(() => {
@@ -378,22 +407,35 @@ function ConfigCard() {
           />
         </FieldRow>
         <FieldRow
-          label={tt.moderationModelLabel}
-          helper={tt.moderationModelHelper}
+          label={tt.curatorModelLabel}
+          helper={tt.curatorModelHelper}
         >
-          <Input
-            type="text"
-            className="max-w-[20rem]"
-            value={draft.moderation_model_name ?? ""}
-            placeholder=""
-            onChange={(e) =>
+          <Select
+            value={draft.curator.model_name ?? DEFAULT_MODEL_VALUE}
+            onValueChange={(v) =>
               setDraft({
                 ...draft,
-                moderation_model_name:
-                  e.target.value.trim() === "" ? null : e.target.value,
+                curator: {
+                  ...draft.curator,
+                  model_name: v === DEFAULT_MODEL_VALUE ? null : v,
+                },
               })
             }
-          />
+          >
+            <SelectTrigger className="max-w-[20rem]">
+              <SelectValue />
+            </SelectTrigger>
+            <SelectContent>
+              <SelectItem value={DEFAULT_MODEL_VALUE}>
+                {tt.curatorModelDefault}
+              </SelectItem>
+              {models.map((m) => (
+                <SelectItem key={m.name} value={m.name}>
+                  {m.display_name || m.name}
+                </SelectItem>
+              ))}
+            </SelectContent>
+          </Select>
         </FieldRow>
 
         <div className="flex justify-end gap-2 pt-2">
diff --git a/frontend-web/src/core/curator/api.ts b/frontend-web/src/core/curator/api.ts
index 660f44cf..da6256bb 100644
--- a/frontend-web/src/core/curator/api.ts
+++ b/frontend-web/src/core/curator/api.ts
@@ -28,11 +28,17 @@ export async function updateCuratorConfig(
   return response.json() as Promise<CuratorConfig>;
 }
 
-export async function runCurator(): Promise<{ triggered: boolean; message: string }> {
+export interface RunCuratorResult {
+  triggered: boolean;
+  /** 未触发时的原因代码:evolution_disabled / curator_disabled / already_running。 */
+  reason: string | null;
+}
+
+export async function runCurator(): Promise<RunCuratorResult> {
   const response = await apiFetch(`${getBackendBaseURL()}/api/curator/run`, {
     method: "POST",
   });
-  return response.json();
+  return response.json() as Promise<RunCuratorResult>;
 }
 
 export async function getCuratorLogs(): Promise<CuratorLogs> {
diff --git a/frontend-web/src/core/curator/types.ts b/frontend-web/src/core/curator/types.ts
index 24ac064e..b8a2f9d8 100644
--- a/frontend-web/src/core/curator/types.ts
+++ b/frontend-web/src/core/curator/types.ts
@@ -3,11 +3,13 @@ export interface CuratorSubConfig {
   interval_hours: number;
   stale_after_days: number;
   archive_after_days: number;
+  review_timeout_seconds: number;
+  /** curator 评审使用的模型名;null = 用主模型。 */
+  model_name: string | null;
 }
 
 export interface CuratorConfig {
   enabled: boolean;
-  moderation_model_name: string | null;
   curator: CuratorSubConfig;
 }
 
diff --git a/frontend-web/src/core/i18n/locales/en-US.ts b/frontend-web/src/core/i18n/locales/en-US.ts
index b5699e4d..db2001d4 100644
--- a/frontend-web/src/core/i18n/locales/en-US.ts
+++ b/frontend-web/src/core/i18n/locales/en-US.ts
@@ -557,6 +557,12 @@ export const enUS: Translations = {
         total: "Total",
         runNow: "Run now",
         runStarted: "Background run started",
+        runBlockedEvolutionDisabled:
+          "Skill evolution is off — enable it below to run the curator.",
+        runBlockedCuratorDisabled:
+          "The curator is off — enable it below to run it manually.",
+        runBlockedAlreadyRunning: "A curator run is already in progress.",
+        runDisabledHint: "Enable skill evolution and the curator below first.",
         lastActions: "Last run",
         llmSuggested: "LLM suggested (self-report)",
         actuallyDid: "Actually done (tool calls)",
@@ -581,9 +587,10 @@ export const enUS: Translations = {
         archiveAfterDaysLabel: "Archive threshold (days)",
         archiveAfterDaysHelper:
           "Stale skills inactive for this many additional days are archived.",
-        moderationModelLabel: "Moderation model (optional)",
-        moderationModelHelper:
-          "Leave blank to use the default chat model.",
+        curatorModelLabel: "Curator review model",
+        curatorModelHelper:
+          "Model for the curator's LLM review and skill security moderation. Defaults to the primary model.",
+        curatorModelDefault: "Primary model (default)",
         save: "Save changes",
         saveSuccess: "Configuration updated",
         reset: "Reset",
diff --git a/frontend-web/src/core/i18n/locales/types.ts b/frontend-web/src/core/i18n/locales/types.ts
index 6a5a0d9b..13c38425 100644
--- a/frontend-web/src/core/i18n/locales/types.ts
+++ b/frontend-web/src/core/i18n/locales/types.ts
@@ -482,6 +482,10 @@ export interface Translations {
         total: string;
         runNow: string;
         runStarted: string;
+        runBlockedEvolutionDisabled: string;
+        runBlockedCuratorDisabled: string;
+        runBlockedAlreadyRunning: string;
+        runDisabledHint: string;
         lastActions: string;
         llmSuggested: string;
         actuallyDid: string;
@@ -500,8 +504,9 @@ export interface Translations {
         staleAfterDaysHelper: string;
         archiveAfterDaysLabel: string;
         archiveAfterDaysHelper: string;
-        moderationModelLabel: string;
-        moderationModelHelper: string;
+        curatorModelLabel: string;
+        curatorModelHelper: string;
+        curatorModelDefault: string;
         save: string;
         saveSuccess: string;
         reset: string;
diff --git a/frontend-web/src/core/i18n/locales/zh-CN.ts b/frontend-web/src/core/i18n/locales/zh-CN.ts
index 1e8d9508..08ee4299 100644
--- a/frontend-web/src/core/i18n/locales/zh-CN.ts
+++ b/frontend-web/src/core/i18n/locales/zh-CN.ts
@@ -533,6 +533,10 @@ export const zhCN: Translations = {
         total: "合计",
         runNow: "立即执行",
         runStarted: "已触发后台执行",
+        runBlockedEvolutionDisabled: "自我进化未开启 —— 请先在下方打开开关。",
+        runBlockedCuratorDisabled: "curator 未开启 —— 请先在下方打开 curator 开关。",
+        runBlockedAlreadyRunning: "已有一次 curator 运行进行中。",
+        runDisabledHint: "请先在下方开启自我进化与 curator 开关。",
         lastActions: "最近一次运行",
         llmSuggested: "LLM 建议（自述）",
         actuallyDid: "实际执行（工具调用）",
@@ -552,8 +556,9 @@ export const zhCN: Translations = {
         staleAfterDaysHelper: "技能长时间不活跃后被标记为陈旧。",
         archiveAfterDaysLabel: "归档阈值（天）",
         archiveAfterDaysHelper: "标记陈旧后再经过这么多天会被自动归档。",
-        moderationModelLabel: "安全审核模型（可选）",
-        moderationModelHelper: "为空时使用默认主模型。",
+        curatorModelLabel: "技能优化模型",
+        curatorModelHelper: "用于 curator 技能评审和技能安全审核的模型，为空则使用主模型。",
+        curatorModelDefault: "主模型（默认）",
         save: "保存修改",
         saveSuccess: "配置已更新",
         reset: "重置",
diff --git a/offline-backend-20260512/backend/CLAUDE.md b/offline-backend-20260512/backend/CLAUDE.md
index 919c8c77..35b6be77 100644
--- a/offline-backend-20260512/backend/CLAUDE.md
+++ b/offline-backend-20260512/backend/CLAUDE.md
@@ -302,6 +302,55 @@ Custom agent metadata lives in the `agents` ORM table (`deerflow.persistence.age
 - **Injection**: Enabled skills listed in agent system prompt with container paths
 - **Installation**: `POST /api/skills/install` extracts .skill ZIP archive to custom/ directory
 
+#### Skill Curator (`packages/harness/deerflow/skills/curator.py`)
+
+Background skill lifecycle manager. The Gateway `lifespan` starts a resident
+scheduler task (`_curator_scheduler_loop` in `app/gateway/app.py`) that — after a
+5-minute startup delay — calls `maybe_run_curator()` every hour. A cross-process
+file lock (`.curator_scheduler.lock`) elects a single leader worker so the
+scheduler runs once per deployment. Per run:
+
+- **Master switch** — `skill_evolution.enabled` (default **`false`**) gates the
+  whole feature; `curator.enabled` (default **`false`**) gates the curator
+  specifically. Both must be `true` for the curator to act. `should_run_now()`
+  and `_curator_thread_target()` both enforce this — so manual `POST
+  /api/curator/run` also respects the master switch. Config is
+  `config.yaml.skill_evolution` deep-merged with per-user
+  `users/{user_id}/skill_evolution.json`.
+- **Per-user run lock** — `.curator_state` carries `running` / `started_at`;
+  `_try_acquire_run_lock()` prevents the same user's curator from running
+  concurrently (scheduler vs. manual trigger, or double manual clicks). A lock
+  whose `started_at` is older than `_CURATOR_STALE_LOCK_HOURS` (6h) is treated as
+  process-crash residue and can be taken over.
+- **Concurrency cap** — `maybe_run_curator()` collects all due users and hands
+  them to `_dispatch_curator_runs()`, which runs them through a
+  `ThreadPoolExecutor` capped at `_CURATOR_MAX_CONCURRENCY` (5) — avoids a
+  thundering herd of LLM calls when many users come due at once (e.g. first
+  deploy).
+- **Failure back-off** — `_release_run_lock()` always stamps `last_attempt_at`.
+  When a first run keeps crashing (`last_run_at` never set), `should_run_now()`
+  retries on `_CURATOR_RETRY_HOURS` (6h) instead of every hourly tick.
+- **`should_run_now()`** is the real gate: the hourly tick only checks "is the
+  per-user `interval_hours` (default 168h) elapsed". Timestamp parsing tolerates
+  naive (tz-less) strings — a corrupt `last_run_at` triggers a re-run rather than
+  a permanent deadlock.
+- **Published-skill filter** — `SkillStore.list_published_names()` is a batch
+  query; the curator fetches the published-name set once per run instead of one
+  DB round-trip per candidate skill.
+- **LLM review timeout** — `_run_llm_review()` runs `agent.invoke()` in a worker
+  thread bounded by `curator.review_timeout_seconds` (default 600s). The model's
+  per-request `request_timeout` can't bound a multi-turn React agent loop, so on
+  timeout the run is abandoned (a warning is written to `curator_log`, visible in
+  the Web UI) and the curator finishes cleanly + releases the run lock. The
+  worker context is copied via `contextvars.copy_context()` so tool calls still
+  resolve the correct `user_id`.
+- **Review model** — `_run_llm_review()` uses `curator.model_name` (a
+  `config.yaml` `models[]` entry name) if set, else the primary model; it is
+  editable from the Web UI evolution panel. The skill security scanner
+  (`scan_skill_content` in `security_scanner.py`) reuses the same
+  `skill_evolution.curator.model_name` setting — review and moderation share one
+  model. There is no separate moderation-model config field.
+
 ### Model Factory (`packages/harness/deerflow/models/factory.py`)
 
 - `create_chat_model(name, thinking_enabled)` instantiates LLM from config via reflection
diff --git a/offline-backend-20260512/backend/app/gateway/routers/skills.py b/offline-backend-20260512/backend/app/gateway/routers/skills.py
index d80ac26f..5cc9a5a3 100644
--- a/offline-backend-20260512/backend/app/gateway/routers/skills.py
+++ b/offline-backend-20260512/backend/app/gateway/routers/skills.py
@@ -967,12 +967,21 @@ async def curator_status(request: Request) -> dict:
 @router.post("/curator/run", summary="Trigger Curator Run")
 async def run_curator(request: Request) -> dict:
     """Manually trigger a curator review pass for the authenticated user
-    (bypasses interval check, runs in a background thread)."""
+    (bypasses the interval check, runs in a background thread).
+
+    Pre-checks the master switch / curator switch / run lock so the response
+    accurately reflects whether a run actually started — instead of always
+    reporting success. Returns ``{"triggered": bool, "reason": str | None}``.
+    """
     try:
         import threading
-        from deerflow.skills.curator import _curator_thread_target
+
+        from deerflow.skills.curator import _curator_thread_target, manual_trigger_blocked_reason
 
         user_id = _current_user_id(request)
+        reason = manual_trigger_blocked_reason(user_id)
+        if reason is not None:
+            return {"triggered": False, "reason": reason}
         thread = threading.Thread(
             target=_curator_thread_target,
             args=(user_id,),
@@ -980,7 +989,7 @@ async def run_curator(request: Request) -> dict:
             name=f"skill-curator-manual-{user_id}",
         )
         thread.start()
-        return {"triggered": True, "message": "Curator run started in background"}
+        return {"triggered": True, "reason": None}
     except Exception as e:
         logger.error("Failed to trigger curator: %s", e, exc_info=True)
         raise HTTPException(status_code=500, detail=f"Failed to trigger curator: {str(e)}")
@@ -989,17 +998,19 @@ async def run_curator(request: Request) -> dict:
 class CuratorSubConfig(BaseModel):
     """Nested curator configuration mirroring CuratorConfig."""
 
-    enabled: bool = Field(default=True, description="Whether the curator runs automatically.")
+    # 与 CuratorConfig.enabled 保持一致:默认关闭,需显式 opt-in。
+    enabled: bool = Field(default=False, description="Whether the curator runs automatically.")
     interval_hours: int = Field(default=168, ge=1, description="Hours between curator runs.")
     stale_after_days: int = Field(default=30, ge=1, description="Days of inactivity before a skill is marked stale.")
     archive_after_days: int = Field(default=90, ge=1, description="Days of inactivity before a stale skill is archived.")
+    review_timeout_seconds: int = Field(default=600, ge=1, description="Overall timeout (seconds) for the LLM review pass.")
+    model_name: str | None = Field(default=None, description="Model name for the curator LLM review pass; null = primary model.")
 
 
 class CuratorConfigResponse(BaseModel):
     """Response model exposing the effective skill_evolution config."""
 
     enabled: bool
-    moderation_model_name: str | None
     curator: CuratorSubConfig
 
 
@@ -1011,7 +1022,6 @@ class CuratorConfigRequest(BaseModel):
     """
 
     enabled: bool = Field(..., description="Whether the agent can create and modify skills under skills/custom.")
-    moderation_model_name: str | None = Field(default=None, description="Optional model name for skill security moderation.")
     curator: CuratorSubConfig
 
 
@@ -1035,12 +1045,13 @@ def _build_curator_config_response(evo) -> CuratorConfigResponse:
     """
     return CuratorConfigResponse(
         enabled=evo.enabled,
-        moderation_model_name=evo.moderation_model_name,
         curator=CuratorSubConfig(
             enabled=evo.curator.enabled,
             interval_hours=evo.curator.interval_hours,
             stale_after_days=evo.curator.stale_after_days,
             archive_after_days=evo.curator.archive_after_days,
+            review_timeout_seconds=evo.curator.review_timeout_seconds,
+            model_name=evo.curator.model_name,
         ),
     )
 
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/config/skill_evolution_config.py b/offline-backend-20260512/backend/packages/harness/deerflow/config/skill_evolution_config.py
index ec9b9540..a6813313 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/config/skill_evolution_config.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/config/skill_evolution_config.py
@@ -4,10 +4,17 @@ from pydantic import BaseModel, Field
 class CuratorConfig(BaseModel):
     """Configuration for the background skill curator."""
 
-    enabled: bool = Field(default=True, description="Whether the curator runs automatically.")
+    # 默认关闭:curator 仅在用户显式开启时才运行(配合 SkillEvolutionConfig.enabled
+    # 总开关,二者均为安全默认 False —— 未显式 opt-in 则整套技能进化不动作)。
+    enabled: bool = Field(default=False, description="Whether the curator runs automatically.")
     interval_hours: int = Field(default=168, description="Hours between curator runs (default: 7 days).")
     stale_after_days: int = Field(default=30, description="Days of inactivity before a skill is marked stale.")
     archive_after_days: int = Field(default=90, description="Days of inactivity before a stale skill is archived.")
+    # LLM 评审整体超时:agent.invoke 内部可能多轮调 LLM,模型层的单次 request_timeout
+    # 兜不住整体;超过此秒数则中止本次评审,curator 正常收尾并释放运行锁。
+    review_timeout_seconds: int = Field(default=600, description="Overall timeout (seconds) for the LLM review pass.")
+    # curator LLM 评审使用的模型名(对应 config.yaml models[].name);为空则用主模型。
+    model_name: str | None = Field(default=None, description="Model name for the curator LLM review pass. Defaults to the primary chat model.")
 
 
 class SkillEvolutionConfig(BaseModel):
@@ -17,10 +24,6 @@ class SkillEvolutionConfig(BaseModel):
         default=False,
         description="Whether the agent can create and modify skills under skills/custom.",
     )
-    moderation_model_name: str | None = Field(
-        default=None,
-        description="Optional model name for skill security moderation. Defaults to the primary chat model.",
-    )
     curator: CuratorConfig = Field(
         default_factory=CuratorConfig,
         description="Configuration for the background skill curator.",
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/base.py b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/base.py
index 52d48a90..4f82cb4d 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/base.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/base.py
@@ -67,3 +67,13 @@ class SkillStore(ABC):
     async def count_owned(self, user_id: str) -> int:
         """Return the number of skills owned by ``user_id`` (owner_user_id IS NOT NULL)."""
         raise NotImplementedError
+
+    @abstractmethod
+    async def list_published_names(self) -> set[str]:
+        """Return the set of skill names with ``published = True`` (any owner).
+
+        Batch alternative to calling ``get_any`` per skill — the curator uses
+        this to filter out published skills in a single query instead of one
+        DB round-trip per candidate skill.
+        """
+        raise NotImplementedError
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/memory.py b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/memory.py
index 15bf5a53..48eedca5 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/memory.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/memory.py
@@ -127,3 +127,6 @@ class MemorySkillStore(SkillStore):
 
     async def count_owned(self, user_id: str) -> int:
         return sum(1 for s in self._skills.values() if s.get("owner_user_id") == user_id)
+
+    async def list_published_names(self) -> set[str]:
+        return {name for name, record in self._skills.items() if record.get("published")}
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/sql.py b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/sql.py
index 5997daca..104b951a 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/sql.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/persistence/skills/sql.py
@@ -142,3 +142,9 @@ class SkillRepository(SkillStore):
         stmt = select(func.count()).select_from(SkillRow).where(SkillRow.owner_user_id == user_id)
         async with self._sf() as session:
             return await session.scalar(stmt) or 0
+
+    async def list_published_names(self) -> set[str]:
+        stmt = select(SkillRow.name).where(SkillRow.published.is_(True))
+        async with self._sf() as session:
+            result = await session.execute(stmt)
+            return {row[0] for row in result.all()}
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/skills/curator.py b/offline-backend-20260512/backend/packages/harness/deerflow/skills/curator.py
index c493f2b6..ce70133f 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/skills/curator.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/skills/curator.py
@@ -10,6 +10,7 @@ from __future__ import annotations
 
 import asyncio
 import concurrent.futures
+import contextvars
 import json
 import logging
 import os
@@ -25,6 +26,10 @@ logger = logging.getLogger(__name__)
 
 _CURATOR_STATE_FILENAME = ".curator_state"
 
+# 周期调度一轮里同时执行的 curator 数量上限。到期用户多时(尤其首次部署所有用户一起
+# 到期)用线程池排队消费,避免一次性起 N 个线程齐调 LLM 打爆 provider 速率限制。
+_CURATOR_MAX_CONCURRENCY = 5
+
 _T = TypeVar("_T")
 
 
@@ -119,6 +124,81 @@ def _write_curator_state(state: dict, user_id: str | None) -> None:
         logger.debug("Failed to write curator state", exc_info=True)
 
 
+def _parse_state_time(value: str | None) -> datetime | None:
+    """解析 ``.curator_state`` 里的 ISO 时间字符串,返回 aware datetime(UTC)。
+
+    naive(无时区)字符串按 UTC 补全 —— 与 ``apply_automatic_transitions`` 的处理
+    一致,避免 naive 时间串与 ``datetime.now(UTC)`` 直接比较时抛异常。无法解析时
+    返回 None,由调用方决定降级行为。
+    """
+    if not value:
+        return None
+    try:
+        dt = datetime.fromisoformat(value)
+    except Exception:
+        return None
+    if dt.tzinfo is None:
+        dt = dt.replace(tzinfo=UTC)
+    return dt
+
+
+# Curator 运行锁:防止同一用户的 curator 并发执行(定时器 vs 手动触发、连点手动触发)。
+# 锁状态写进 .curator_state 的 running/started_at 字段,跨 worker 可见(手动请求可能
+# 落到非调度 worker)。started_at 超过下面的阈值视为进程残留,允许接管以免永久死锁。
+_CURATOR_STALE_LOCK_HOURS = 6
+# 上一轮"尝试"过但未成功(last_run_at 始终为空)时的重试间隔 —— 比正常 interval_hours
+# 短得多,但也不至于每个调度 tick(1h)都重试、每次都烧一份 LLM 配额。
+_CURATOR_RETRY_HOURS = 6
+# 进程内串行化运行锁的"读-改-写",保证同进程内的抢锁是原子的(跨进程的极小竞态窗口
+# 可接受 —— curator 触发频率极低)。
+_run_lock_guard = threading.Lock()
+
+
+def _run_lock_active(state: dict, now: datetime) -> bool:
+    """判断 ``.curator_state`` 里的运行锁当前是否有效(running 且未过期)。
+
+    ``started_at`` 缺失或不可解析、或距今已超过 ``_CURATOR_STALE_LOCK_HOURS`` 时,
+    视为进程残留留下的死锁,返回 False 以便后续触发能接管。
+    """
+    if not state.get("running"):
+        return False
+    started_dt = _parse_state_time(state.get("started_at"))
+    if started_dt is None:
+        return False
+    return now < started_dt + timedelta(hours=_CURATOR_STALE_LOCK_HOURS)
+
+
+def _try_acquire_run_lock(user_id: str, now: datetime) -> bool:
+    """尝试为 ``user_id`` 抢占 curator 运行锁。
+
+    抢到锁(已把 running=True / started_at 写回)返回 True;已有未过期的运行时返回
+    False。读-改-写在 ``_run_lock_guard`` 下进行,保证同进程内原子。
+    """
+    with _run_lock_guard:
+        state = _read_curator_state(user_id)
+        if _run_lock_active(state, now):
+            return False
+        state["running"] = True
+        state["started_at"] = now.isoformat()
+        _write_curator_state(state, user_id)
+        return True
+
+
+def _release_run_lock(user_id: str) -> None:
+    """释放 curator 运行锁并记录本次尝试时间。无论运行成功或异常都应调用。
+
+    除了把 running 置回 False,还写入 ``last_attempt_at`` —— 用于失败退避:首次运行
+    若一直崩溃,``last_run_at`` 永远为空,``should_run_now`` 改用
+    ``last_attempt_at + _CURATOR_RETRY_HOURS`` 判定,避免每个 tick 都重试。
+    成功跑完后 ``last_run_at`` 已被写入,``should_run_now`` 不再看 ``last_attempt_at``。
+    """
+    with _run_lock_guard:
+        state = _read_curator_state(user_id)
+        state["running"] = False
+        state["last_attempt_at"] = datetime.now(UTC).isoformat()
+        _write_curator_state(state, user_id)
+
+
 def _get_curator_config(user_id: str | None = None):
     """Return the curator config object for ``user_id``.
 
@@ -135,6 +215,21 @@ def _get_curator_config(user_id: str | None = None):
         return None
 
 
+def _skill_evolution_enabled(user_id: str | None = None) -> bool:
+    """返回技能进化总开关 ``skill_evolution.enabled``(已合并 per-user 覆盖)。
+
+    这是整套技能进化功能的总闸:关闭时 curator 完全不动作。任何读取异常都按
+    "关闭"处理(fail-safe)—— 宁可不跑,也不在配置不可读时偷偷运行。
+    """
+    try:
+        from deerflow.config.skill_evolution_runtime import load_skill_evolution_config_for_user
+
+        skill_evo = load_skill_evolution_config_for_user(user_id)
+        return bool(getattr(skill_evo, "enabled", False))
+    except Exception:
+        return False
+
+
 def _list_user_owned_skill_names(user_id: str) -> set[str] | None:
     """Return skill names owned by ``user_id`` (used to filter transitions).
 
@@ -231,12 +326,39 @@ def _is_skill_published(skill_name: str) -> bool:
         return False
 
 
+def _list_published_skill_names() -> set[str]:
+    """一次性返回所有 ``published=True`` 的技能名集合。
+
+    取代"每个技能调一次 ``_is_skill_published``"的逐条查库 —— curator 一轮里会遍历
+    所有候选技能,逐条查会产生 ≈2N 次独立 event loop + DB 往返。出错时返回空集
+    (fail-open:与 ``_is_skill_published`` 出错返回 False 语义一致,即 DB 不可用时
+    所有技能都视为未发布,回退到 curator 的 legacy 行为)。
+    """
+    try:
+        from deerflow.persistence.engine import get_session_factory
+        from deerflow.persistence.skills import make_skill_store
+
+        factory = get_session_factory()
+        if factory is None:
+            return set()
+        store = make_skill_store(factory)
+
+        async def _run() -> set[str]:
+            return await store.list_published_names()
+
+        return _run_async_blocking(_run())
+    except Exception:
+        logger.debug("_list_published_skill_names failed", exc_info=True)
+        return set()
+
+
 def _eligible_for_curation(
     name: str,
     entry: dict,
     *,
     owned_set: set[str] | None,
     referenced_set: set[str],
+    published_set: set[str],
 ) -> bool:
     """Three-condition gate for whether curator may touch a skill.
 
@@ -244,11 +366,15 @@ def _eligible_for_curation(
     1. source == "agent" (caller is expected to have already filtered, but
        we double-check here for safety)
     2. NOT referenced by any agent's skills config
-    3. NOT published (queried via SkillStore)
+    3. NOT published
 
     ``owned_set`` is the set of skill names owned by the running user. When
     None, the legacy "process every agent-created skill" fallback applies
     for ownership only (the other two conditions still hold).
+
+    ``published_set`` is the set of published skill names, pre-fetched once by
+    the caller via ``_list_published_skill_names()`` — avoids a per-skill DB
+    round-trip.
     """
     if entry.get("source") != "agent":
         return False
@@ -256,7 +382,7 @@ def _eligible_for_curation(
         return False
     if name in referenced_set:
         return False
-    if _is_skill_published(name):
+    if name in published_set:
         return False
     return True
 
@@ -264,21 +390,52 @@ def _eligible_for_curation(
 def should_run_now(user_id: str) -> bool:
     """Return True if the curator interval has elapsed for ``user_id`` and it
     is not paused."""
+    # 总开关:技能进化整体关闭时,curator 一律不跑。
+    if not _skill_evolution_enabled(user_id):
+        return False
     curator_cfg = _get_curator_config(user_id)
-    if curator_cfg is None or not getattr(curator_cfg, "enabled", True):
+    if curator_cfg is None or not getattr(curator_cfg, "enabled", False):
         return False
     state = _read_curator_state(user_id)
     if state.get("paused"):
         return False
+    # 运行锁:已有 curator 在跑(且锁未过期)时不再触发新的一轮。
+    if _run_lock_active(state, datetime.now(UTC)):
+        return False
     last_run = state.get("last_run_at")
     if not last_run:
-        return True
+        # 从未成功跑完过。若此前已尝试过(last_attempt_at 存在),按较短的重试间隔
+        # 退避 —— 避免首次运行持续失败时每个 tick(1h)都重试、每次烧一份 LLM 配额。
+        attempt_dt = _parse_state_time(state.get("last_attempt_at"))
+        if attempt_dt is None:
+            return True
+        return datetime.now(UTC) >= attempt_dt + timedelta(hours=_CURATOR_RETRY_HOURS)
     interval_hours = getattr(curator_cfg, "interval_hours", 168)
-    try:
-        last_dt = datetime.fromisoformat(last_run)
-        return datetime.now(UTC) >= last_dt + timedelta(hours=interval_hours)
-    except Exception:
-        return False
+    last_dt = _parse_state_time(last_run)
+    if last_dt is None:
+        # 时间串损坏 —— 视同"从未跑过",允许重跑并重写正确时间戳,而不是像旧逻辑那样
+        # except 返回 False 导致该用户永久不再触发。
+        return True
+    return datetime.now(UTC) >= last_dt + timedelta(hours=interval_hours)
+
+
+def manual_trigger_blocked_reason(user_id: str) -> str | None:
+    """手动触发 curator 的预检。返回 None 表示可以跑;否则返回阻塞原因代码:
+    ``evolution_disabled`` / ``curator_disabled`` / ``already_running``。
+
+    供 ``POST /api/curator/run`` 在派发后台线程**前**同步判定 —— 后台线程
+    (``_curator_thread_target``)本身也会再查一遍,这里只是为了让前端拿到准确反馈,
+    而不是无论开关状态如何都提示"已开始"。
+    """
+    if not _skill_evolution_enabled(user_id):
+        return "evolution_disabled"
+    curator_cfg = _get_curator_config(user_id)
+    if curator_cfg is None or not getattr(curator_cfg, "enabled", False):
+        return "curator_disabled"
+    state = _read_curator_state(user_id)
+    if _run_lock_active(state, datetime.now(UTC)):
+        return "already_running"
+    return None
 
 
 def apply_automatic_transitions(user_id: str, now: datetime | None = None) -> dict:
@@ -303,6 +460,7 @@ def apply_automatic_transitions(user_id: str, now: datetime | None = None) -> di
 
     owned = _list_user_owned_skill_names(user_id)
     referenced = _collect_referenced_skill_names()
+    published = _list_published_skill_names()  # 一次性查出,避免逐技能查库
     raw_entries = agent_created_entries()
 
     skipped_referenced = 0
@@ -314,7 +472,7 @@ def apply_automatic_transitions(user_id: str, now: datetime | None = None) -> di
         if name in referenced:
             skipped_referenced += 1
             continue
-        if _is_skill_published(name):
+        if name in published:
             skipped_published += 1
             continue
         eligible[name] = entry
@@ -377,9 +535,10 @@ def _render_candidate_list(user_id: str) -> str:
 
     owned = _list_user_owned_skill_names(user_id)
     referenced = _collect_referenced_skill_names()
+    published = _list_published_skill_names()  # 一次性查出,避免逐技能查库
     entries: dict[str, dict] = {}
     for k, v in all_entries().items():
-        if not _eligible_for_curation(k, v, owned_set=owned, referenced_set=referenced):
+        if not _eligible_for_curation(k, v, owned_set=owned, referenced_set=referenced, published_set=published):
             continue
         if v.get("state") == STATE_ARCHIVED:
             continue
@@ -666,7 +825,11 @@ def _run_llm_review(candidate_list: str, *, user_id: str) -> tuple[str, list[dic
         from deerflow.models import create_chat_model
         from deerflow.tools.builtins.skill_tools import skill_list_tool, skill_view_tool
 
-        model = create_chat_model()
+        curator_cfg = _get_curator_config(user_id)
+        # curator 评审用的模型:per-user 可配 curator.model_name(对应 config.yaml
+        # models[].name);为空则用主模型。
+        model_name = getattr(curator_cfg, "model_name", None) if curator_cfg else None
+        model = create_chat_model(name=model_name)
         # 用 curator 受限版 skill_manage(屏蔽 delete),不用通用版 —— 见 _make_curator_skill_manage_tool。
         tools = [skill_list_tool, skill_view_tool, _make_curator_skill_manage_tool(), _make_skill_archive_tool(user_id)]
         # Reuse the lead-agent error middleware: when the LLM mis-calls a tool
@@ -677,9 +840,36 @@ def _run_llm_review(candidate_list: str, *, user_id: str) -> tuple[str, list[dic
         # middleware kwarg — use LangChain 1.x's ``create_agent`` which does.
         agent = create_agent(model=model, tools=tools, middleware=[ToolErrorHandlingMiddleware()])
 
-        curator_log.record("info", "LLM 评审:开始调用 React Agent", user_id=user_id)
+        timeout_seconds = getattr(curator_cfg, "review_timeout_seconds", 600) if curator_cfg else 600
+        curator_log.record(
+            "info",
+            f"LLM 评审:开始调用 React Agent(模型 {model_name or '主模型'},整体超时 {timeout_seconds}s)",
+            user_id=user_id,
+        )
         prompt = _CURATOR_REVIEW_PROMPT + f"\n\nSkill candidates for this pass:\n{candidate_list}"
-        result = agent.invoke({"messages": [HumanMessage(content=prompt)]})
+
+        # 给整个 React Agent 评审套一个总超时:agent.invoke 内部可能多轮调 LLM,模型层的
+        # 单次 request_timeout 兜不住整体。超时后本 curator 线程能正常收尾、释放运行锁,
+        # 而不是无限期挂起。注意:Python 无法强杀线程 —— 超时后那个 agent.invoke 线程会
+        # 成为孤儿(最终被模型层 request_timeout 或自然返回所终结),但不再拖住 curator。
+        # copy_context():把 current_user 等 ContextVar 传进工作线程,否则工具调用会把
+        # user_id 误解析为 DEFAULT_USER_ID。
+        ctx = contextvars.copy_context()
+        pool = concurrent.futures.ThreadPoolExecutor(max_workers=1, thread_name_prefix="curator-llm")
+        future = pool.submit(ctx.run, agent.invoke, {"messages": [HumanMessage(content=prompt)]})
+        try:
+            result = future.result(timeout=timeout_seconds)
+        except concurrent.futures.TimeoutError:
+            curator_log.record(
+                "warning",
+                f"LLM 评审超时(超过 {timeout_seconds}s),已中止本次评审",
+                user_id=user_id,
+            )
+            logger.warning("Curator[%s] LLM review timed out after %ss", user_id, timeout_seconds)
+            return "", []
+        finally:
+            # wait=False:超时时不阻塞等待挂死的 invoke 线程。
+            pool.shutdown(wait=False)
 
         messages = result.get("messages", [])
         tool_actions = _extract_tool_actions(messages)
@@ -876,6 +1066,23 @@ def _curator_thread_target(user_id: str) -> None:
     from deerflow.runtime.user_context import reset_current_user, set_current_user
     from deerflow.skills import curator_log
 
+    # 总开关 + curator 子开关:任一关闭则直接跳过。手动触发 /api/curator/run 也经由
+    # 本函数,因此手动触发同样受总开关与 curator.enabled 约束。
+    if not _skill_evolution_enabled(user_id):
+        logger.info("Curator[%s]: 技能进化总开关关闭,跳过", user_id)
+        return
+    curator_cfg = _get_curator_config(user_id)
+    if curator_cfg is None or not getattr(curator_cfg, "enabled", False):
+        logger.info("Curator[%s]: curator 未启用,跳过", user_id)
+        return
+
+    # 运行锁:防止同一用户的 curator 并发(定时器 vs 手动触发、连点手动触发)。抢不到
+    # 锁说明已有一轮在跑,直接跳过本次触发。
+    if not _try_acquire_run_lock(user_id, datetime.now(UTC)):
+        logger.info("Curator[%s]: 已有 curator 正在运行,跳过本次触发", user_id)
+        curator_log.record("info", "已有 curator 正在运行,跳过本次触发", user_id=user_id)
+        return
+
     # Bind the per-thread user context for the entire curator run so every
     # downstream tool / repository call resolves the correct user_id.
     token = set_current_user(_CuratorUser(user_id))
@@ -886,6 +1093,8 @@ def _curator_thread_target(user_id: str) -> None:
         curator_log.record("error", f"策展后台任务失败:{exc}", user_id=user_id)
     finally:
         reset_current_user(token)
+        # 无论成功或异常,必定释放运行锁。
+        _release_run_lock(user_id)
 
 
 async def maybe_run_curator() -> None:
@@ -911,20 +1120,42 @@ async def maybe_run_curator() -> None:
         logger.exception("Curator: failed to enumerate users at startup")
         return
 
+    due_users: list[str] = []
     for user_id in user_ids:
         try:
-            if not should_run_now(user_id):
-                continue
-            logger.info("Curator[%s]: interval elapsed, spawning background thread", user_id)
-            thread = threading.Thread(
-                target=_curator_thread_target,
-                args=(user_id,),
-                daemon=True,
-                name=f"skill-curator-{user_id}",
-            )
-            thread.start()
+            if should_run_now(user_id):
+                due_users.append(user_id)
         except Exception:
-            logger.exception("Curator: failed to spawn for user %s", user_id)
+            logger.exception("Curator: should_run_now check failed for user %s", user_id)
+    _dispatch_curator_runs(due_users)
+
+
+def _dispatch_curator_runs(due_users: list[str]) -> None:
+    """用线程池(并发上限 ``_CURATOR_MAX_CONCURRENCY``)派发到期用户的 curator 执行。
+
+    到期用户排队消费,而非一次性全部起线程齐调 LLM —— 防首次部署所有用户同时到期时
+    的惊群。``shutdown(wait=False)``:提交后立即返回,线程池在所有任务跑完后自行回收。
+    抽成独立函数便于对并发上限做单元测试。
+    """
+    if not due_users:
+        return
+
+    logger.info(
+        "Curator: %d 个用户到期,以并发上限 %d 派发执行",
+        len(due_users),
+        _CURATOR_MAX_CONCURRENCY,
+    )
+    executor = concurrent.futures.ThreadPoolExecutor(
+        max_workers=_CURATOR_MAX_CONCURRENCY,
+        thread_name_prefix="skill-curator",
+    )
+    try:
+        for user_id in due_users:
+            executor.submit(_curator_thread_target, user_id)
+    except Exception:
+        logger.exception("Curator: failed to submit curator tasks")
+    finally:
+        executor.shutdown(wait=False)
 
 
 def get_curator_status(user_id: str) -> dict:
diff --git a/offline-backend-20260512/backend/packages/harness/deerflow/skills/security_scanner.py b/offline-backend-20260512/backend/packages/harness/deerflow/skills/security_scanner.py
index 5336f8ee..a1ee6af8 100644
--- a/offline-backend-20260512/backend/packages/harness/deerflow/skills/security_scanner.py
+++ b/offline-backend-20260512/backend/packages/harness/deerflow/skills/security_scanner.py
@@ -50,8 +50,10 @@ async def scan_skill_content(content: str, *, executable: bool = False, location
 
     try:
         config = app_config or get_app_config()
-        model_name = config.skill_evolution.moderation_model_name
-        model = create_chat_model(name=model_name, thinking_enabled=False, app_config=config) if model_name else create_chat_model(thinking_enabled=False, app_config=config)
+        # 安全审核复用 curator 的评审模型配置(skill_evolution.curator.model_name):
+        # 二者用同一个模型。未配置时 model_name 为 None,create_chat_model 回退主模型。
+        model_name = config.skill_evolution.curator.model_name
+        model = create_chat_model(name=model_name, thinking_enabled=False, app_config=config)
         response = await model.ainvoke(
             [
                 {"role": "system", "content": rubric},
diff --git a/offline-backend-20260512/backend/tests/test_skill_evolution_runtime.py b/offline-backend-20260512/backend/tests/test_skill_evolution_runtime.py
index cdc190a5..760a5065 100644
--- a/offline-backend-20260512/backend/tests/test_skill_evolution_runtime.py
+++ b/offline-backend-20260512/backend/tests/test_skill_evolution_runtime.py
@@ -25,12 +25,12 @@ def test_load_overrides_returns_none_when_missing(overrides_path):
 def test_save_and_load_overrides_roundtrip(overrides_path):
     payload = {
         "enabled": True,
-        "moderation_model_name": "moderation-mini",
         "curator": {
             "enabled": True,
             "interval_hours": 12,
             "stale_after_days": 7,
             "archive_after_days": 30,
+            "model_name": "curator-mini",
         },
     }
     written = skill_evolution_runtime.save_overrides(payload)
diff --git a/offline-backend-20260512/backend/tests/test_skills_multitenancy.py b/offline-backend-20260512/backend/tests/test_skills_multitenancy.py
index 4ade945f..33ebdad9 100644
--- a/offline-backend-20260512/backend/tests/test_skills_multitenancy.py
+++ b/offline-backend-20260512/backend/tests/test_skills_multitenancy.py
@@ -501,7 +501,8 @@ def test_curator_skips_published_skill(isolated_paths, monkeypatch) -> None:
 
     monkeypatch.setattr(curator, "_list_user_owned_skill_names", lambda uid: {"public-tool"})
     monkeypatch.setattr(curator, "_collect_referenced_skill_names", lambda: set())
-    monkeypatch.setattr(curator, "_is_skill_published", lambda name: name == "public-tool")
+    # curator 现在用批量接口一次性查 published 集合,而非逐技能 _is_skill_published。
+    monkeypatch.setattr(curator, "_list_published_skill_names", lambda: {"public-tool"})
     _isolated_skills_root(isolated_paths, monkeypatch)
     _seed_old_agent_skill("public-tool")