phase0-final-rls-rbac-complete

Root cause: ensure_structured_column() executed a synchronous blocking
engine.connect().execute(SELECT information_schema.columns ...) inside the
async request path (_build_chat_history / _maybe_compress_history), freezing
the uvicorn event loop and hanging all endpoints (~200s, recovered only by
docker restart).

Fix:
- chat_context.py: remove per-request ensure_structured_column() calls; run
  once idempotently at app startup (main.py lifespan). Use CURRENT_DATABASE()
  for PostgreSQL compatibility.
- main.py: add startup-only ensure_structured_column() hook (try/except).
- llm_client.py: trust_env=False + bounded httpx.Timeout/wait_for so broken
  endpoints fail fast instead of hanging the event loop.
- chat_service.py: run media generation in a background daemon thread with its
  own event loop (_launch_media_task_bg) so SSE end does not cancel image/video
  tasks (#811); remove TEMP faulthandler diagnostic.

Safe-align: persist accumulated live container changes (admin/*, knowledge,
market, models, config, observability, mcp_client, platform_knowledge,
im_bot_platform, chat_media, capability_registry, requirements) and add
workflow_templates.

Verified: live failover test (model-006 -> conn-refused) returns in ~1.9s with
concurrent login 0.83s/200; event-loop responsiveness restored (sub-second).

.gitattributes

@@ -0,0 +1,33 @@
+# Normalize all text files to LF on checkout/commit (root-cause fix for mixed line endings)
+* text=auto eol=lf
+
+# Binary assets: never normalize line endings or diff as text
+*.png binary
+*.jpg binary
+*.jpeg binary
+*.gif binary
+*.ico binary
+*.webp binary
+*.woff binary
+*.woff2 binary
+*.ttf binary
+*.eot binary
+*.otf binary
+*.pdf binary
+*.zip binary
+*.tar binary
+*.gz binary
+*.bz2 binary
+*.7z binary
+*.pyc binary
+*.so binary
+*.dll binary
+*.exe binary
+*.db binary
+*.sqlite binary
+*.sqlite3 binary
+*.wasm binary
+*.mp4 binary
+*.mp3 binary
+*.avi binary
+*.mov binary

backend/python/app/anysearch_client.py

@@ -0,0 +1,179 @@
+"""
+AnySearch 联网检索客户端
+======================
+调用 AnySearch 网页/领域搜索 API (https://api.anysearch.com/v1/search)，
+为 AI 聊天注入实时联网检索结果，亦可被后端端点 / Agent 工具复用。
+
+约束：平台容器存在智能路由代理，对外部 API 偶发重置；本客户端使用
+trust_env=False 直连 (等价于 NO_PROXY=api.anysearch.com)，规避代理重置。
+失败一律返回空/降级结果，保证聊天链路不中断。
+"""
+import json
+import logging
+from typing import List, Dict, Optional
+
+logger = logging.getLogger("aiflow.anysearch")
+
+DEFAULT_API_URL = "https://api.anysearch.com/v1/search"
+DEFAULT_ZONE = "cn"
+REQUEST_TIMEOUT = 8.0
+
+
+async def anysearch_search(
+    api_key: str,
+    api_url: str = DEFAULT_API_URL,
+    query: str = "",
+    top_k: int = 5,
+    zone: str = DEFAULT_ZONE,
+    freshness: Optional[str] = None,
+    timeout: float = REQUEST_TIMEOUT,
+) -> List[Dict]:
+    """调用 AnySearch 搜索，返回结构化结果列表。
+
+    每个结果:
+        {
+            "content": snippet 摘要文本,
+            "score": 0.9..0.5 递减,
+            "source": "web_anysearch",
+            "metadata": {"url", "title", "source_domain", "zone"}
+        }
+    失败时返回 [] (不抛异常)。
+    """
+    if not api_key:
+        return []
+    try:
+        import httpx
+    except ImportError:
+        logger.warning("httpx not available for AnySearch")
+        return []
+
+    payload = {
+        "query": query,
+        "zone": zone,
+        "top_k": max(1, min(int(top_k), 20)),
+    }
+    if freshness:
+        payload["freshness"] = freshness
+
+    headers = {
+        "Content-Type": "application/json",
+        "Authorization": f"Bearer {api_key}",
+    }
+    try:
+        # trust_env=False -> 忽略容器内 HTTP(S)_PROXY，直连 api.anysearch.com
+        async with httpx.AsyncClient(trust_env=False, timeout=timeout) as client:
+            resp = await client.post(api_url, json=payload, headers=headers)
+        if resp.status_code != 200:
+            logger.warning(f"AnySearch HTTP {resp.status_code}: {resp.text[:200]}")
+            return []
+        data = resp.json()
+        if data.get("code") != 0:
+            logger.warning(f"AnySearch biz error: {data.get('message')}")
+            return []
+        results = (data.get("data") or {}).get("results") or []
+        hits = []
+        for idx, r in enumerate(results[:top_k]):
+            title = r.get("title", "") or ""
+            url = r.get("url", "") or ""
+            snippet = r.get("snippet") or r.get("content") or ""
+            if not snippet and not title:
+                continue
+            try:
+                from urllib.parse import urlparse
+                domain = urlparse(url).netloc
+            except Exception:
+                domain = ""
+            hits.append({
+                "content": snippet,
+                "score": max(0.4, 0.9 - idx * 0.08),
+                "source": "web_anysearch",
+                "metadata": {
+                    "url": url,
+                    "title": title,
+                    "source_domain": domain,
+                    "zone": zone,
+                },
+            })
+        return hits
+    except Exception as e:
+        logger.warning(f"AnySearch search failed: {e}")
+        return []
+
+
+async def anysearch_health(api_key: str, api_url: str = DEFAULT_API_URL, timeout: float = REQUEST_TIMEOUT) -> Dict:
+    """连通性探测：返回 {"status": "connected"/"disconnected", "detail": ...}。"""
+    if not api_key:
+        return {"status": "disconnected", "detail": "未配置 API Key", "connected": False, "usable": False}
+    try:
+        import httpx
+        async with httpx.AsyncClient(trust_env=False, timeout=timeout) as client:
+            resp = await client.post(
+                api_url,
+                json={"query": "ping", "zone": DEFAULT_ZONE, "top_k": 1},
+                headers={"Content-Type": "application/json", "Authorization": f"Bearer {api_key}"},
+            )
+        if resp.status_code == 200:
+            return {"status": "connected", "detail": "AnySearch 连接正常", "connected": True, "usable": True}
+        return {"status": "error", "detail": f"HTTP {resp.status_code}", "connected": False, "usable": False}
+    except Exception as e:
+        return {"status": "error", "detail": str(e)[:200], "connected": False, "usable": False}
+
+
+def load_anysearch_config(db_session) -> Dict:
+    """从 site_settings 读取全局 AnySearch 配置（失败返回空字典）。"""
+    if db_session is None:
+        return {}
+    try:
+        from app.models import SiteSetting
+        row = db_session.query(SiteSetting).filter(SiteSetting.key == "anysearch_config").first()
+        if not row or not row.value:
+            return {}
+        cfg = json.loads(row.value)
+        if not isinstance(cfg, dict):
+            return {}
+        return cfg
+    except Exception as e:
+        logger.warning(f"load_anysearch_config failed: {e}")
+        return {}
+
+
+def save_anysearch_config(
+    db_session,
+    api_key: str,
+    is_enabled: bool = True,
+    api_url: str = DEFAULT_API_URL,
+    zone: str = DEFAULT_ZONE,
+    freshness: Optional[str] = None,
+    top_k: int = 5,
+) -> bool:
+    """保存/更新全局 AnySearch 配置到 site_settings。"""
+    if db_session is None:
+        return False
+    try:
+        from app.models import SiteSetting
+        from datetime import datetime
+        cfg = {
+            "api_key": api_key.strip(),
+            "is_enabled": bool(is_enabled),
+            "api_url": api_url.strip() or DEFAULT_API_URL,
+            "zone": zone or DEFAULT_ZONE,
+            "freshness": freshness,
+            "top_k": max(1, min(int(top_k), 20)),
+        }
+        row = db_session.query(SiteSetting).filter(SiteSetting.key == "anysearch_config").first()
+        if row:
+            row.value = json.dumps(cfg, ensure_ascii=False)
+            row.updated_at = datetime.now()
+        else:
+            from app.models import gen_uuid
+            row = SiteSetting(id=gen_uuid(), key="anysearch_config", value=json.dumps(cfg, ensure_ascii=False))
+            db_session.add(row)
+        db_session.commit()
+        return True
+    except Exception as e:
+        logger.warning(f"save_anysearch_config failed: {e}")
+        try:
+            db_session.rollback()
+        except Exception:
+            pass
+        return False

backend/python/app/api/admin/achievement_admin.py

@@ -284,7 +284,7 @@ async def claim_achievement_reward(
         
         # Mark reward as claimed
         _execute_raw(db, 
-            "UPDATE user_achievements SET reward_claimed = 1, claimed_at = NOW() WHERE user_id = :uid AND achievement_id = :aid",
+            "UPDATE user_achievements SET reward_claimed = true, claimed_at = NOW() WHERE user_id = :uid AND achievement_id = :aid",
             {"uid": user_id, "aid": ach_id})
         
         return {"code": 0, "message": "奖励领取成功"}

backend/python/app/api/admin/analytics_admin.py

@@ -316,7 +316,7 @@ async def create_dashboard(
         config_json = json.loads('[{"type":"stat","title":"暂无数据","metric":"none","chart_type":"number"}]')
 
     if is_default:
-        db.execute(text("UPDATE analytics_dashboards SET is_default=0"))
+        db.execute(text("UPDATE analytics_dashboards SET is_default=false"))
 
     db.execute(text("""
         INSERT INTO analytics_dashboards (id, name, description, config, is_default, created_by)
@@ -332,7 +332,7 @@ async def create_dashboard(
 @router.delete("/dashboards/{dashboard_id}")
 async def delete_dashboard(dashboard_id: str, db: Session = Depends(get_db), user=Depends(get_current_user)):
     """删除仪表盘"""
-    db.execute(text("DELETE FROM analytics_dashboards WHERE id=:id AND is_default=0"), {"id": dashboard_id})
+    db.execute(text("DELETE FROM analytics_dashboards WHERE id=:id AND is_default=false"), {"id": dashboard_id})
     db.commit()
     return {"code": 200, "data": {"message": "仪表盘已删除"}}
 

backend/python/app/api/admin/billing_admin.py

@@ -445,7 +445,7 @@ async def billing_stats(db: Session = Depends(get_db), user=Depends(get_current_
         SELECT DATE(created_at) as d, COALESCE(SUM(amount), 0) as rev
         FROM billing_transactions
         WHERE transaction_type='recharge' AND status='completed'
-          AND created_at >= DATE_SUB(NOW(), INTERVAL 7 DAY)
+          AND created_at >= NOW() - INTERVAL '7 DAY'
         GROUP BY DATE(created_at) ORDER BY d
     """)).fetchall()
     daily_revenue = [{"date": str(r[0]), "revenue": float(r[1])} for r in revenue_7d]

backend/python/app/api/admin/code_exec_admin.py

@@ -103,7 +103,7 @@ async def get_execution_stats(db: Session = Depends(get_db), _=Depends(get_curre
     try:
         total = _execute_raw(db, "SELECT COUNT(*) as cnt FROM code_exec_tasks")[0]["cnt"]
         today = _execute_raw(db,
-            "SELECT COUNT(*) as cnt FROM code_exec_tasks WHERE DATE(created_at) = CURDATE()")[0]["cnt"]
+            "SELECT COUNT(*) as cnt FROM code_exec_tasks WHERE DATE(created_at) = CURRENT_DATE")[0]["cnt"]
         success_count = _execute_raw(db,
             "SELECT COUNT(*) as cnt FROM code_exec_tasks WHERE exit_code = 0 AND error IS NULL")[0]["cnt"]
         failed_count = _execute_raw(db,

backend/python/app/api/admin/distro_admin.py

@@ -703,7 +703,7 @@ echo "Installation complete!"
 {d.get('version', '1.0.0')}
 
 ## Installation
-Run `bash install.sh` to install.
+Run \"bash install.sh\" to install.
 
 ## License
 A valid license key is required. Contact your administrator for details.

backend/python/app/api/admin/im_bot_admin.py

@@ -169,7 +169,7 @@ async def get_im_bot_stats(user=Depends(get_current_user)):
         today_stats = _row_to_dict(rows[0]) if rows else {"msg_in": 0, "msg_out": 0, "users": 0, "errors": 0}
 
         # Total routing rules
-        routes_count = _execute_raw(db, text("SELECT COUNT(*) as cnt FROM im_bot_routes WHERE is_active=1"))[0]._mapping['cnt']
+        routes_count = _execute_raw(db, text("SELECT COUNT(*) as cnt FROM im_bot_routes WHERE is_active=true"))[0]._mapping['cnt']
 
         return {
             "success": True,
@@ -202,7 +202,7 @@ async def get_dashboard(user=Depends(get_current_user)):
                    SUM(messages_out) as msg_out,
                    SUM(unique_users) as users
             FROM im_bot_daily_stats
-            WHERE stat_date >= DATE_SUB(CURDATE(), INTERVAL 7 DAY)
+            WHERE stat_date >= CURRENT_DATE - INTERVAL '7 DAY'
             GROUP BY stat_date
             ORDER BY stat_date
         """))
@@ -217,7 +217,7 @@ async def get_dashboard(user=Depends(get_current_user)):
             LEFT JOIN (
                 SELECT channel_id, SUM(messages_in + messages_out) as msg_count
                 FROM im_bot_daily_stats
-                WHERE stat_date >= DATE_SUB(CURDATE(), INTERVAL 30 DAY)
+                WHERE stat_date >= CURRENT_DATE - INTERVAL '30 DAY'
                 GROUP BY channel_id
             ) s ON c.id = s.channel_id
             ORDER BY c.created_at DESC
@@ -357,7 +357,7 @@ async def get_channel_detail(channel_id: str, user=Depends(get_current_user)):
             SELECT stat_date, messages_in, messages_out, unique_users, error_count
             FROM im_bot_daily_stats
             WHERE channel_id = :cid
-            AND stat_date >= DATE_SUB(CURDATE(), INTERVAL 7 DAY)
+            AND stat_date >= CURRENT_DATE - INTERVAL '7 DAY'
             ORDER BY stat_date
         """), {"cid": channel_id})
         channel['stats_7d'] = _rows_to_dict(stats)
@@ -428,7 +428,7 @@ async def activate_channel(channel_id: str, user=Depends(get_current_user)):
     db = SessionLocal()
     try:
         _execute_raw(db, text(
-            "UPDATE channels SET is_active=1, status='connected' WHERE id=:cid"
+            "UPDATE channels SET is_active=true, status='connected' WHERE id=:cid"
         ), {"cid": channel_id})
         return {"success": True, "message": "Channel activated"}
     except Exception as e:
@@ -443,7 +443,7 @@ async def deactivate_channel(channel_id: str, user=Depends(get_current_user)):
     db = SessionLocal()
     try:
         _execute_raw(db, text(
-            "UPDATE channels SET is_active=0, status='disconnected' WHERE id=:cid"
+            "UPDATE channels SET is_active=false, status='disconnected' WHERE id=:cid"
         ), {"cid": channel_id})
         return {"success": True, "message": "Channel deactivated"}
     except Exception as e:

backend/python/app/api/admin/knowledge_admin.py

@@ -38,6 +38,11 @@ class ExternalServiceConfig(BaseModel):
     config: Optional[Dict[str, Any]] = None
 
 
+class AnySearchConfig(BaseModel):
+    api_key: str = Field(min_length=1, max_length=500)
+    is_enabled: bool = True
+
+
 class ExternalServiceUpdate(BaseModel):
     name: Optional[str] = None
     api_url: Optional[str] = None
@@ -738,8 +743,8 @@ def _get_external_services_table(db: Session):
             return []
 
 
-def _check_service_connection(service_type: str, api_url: str, api_key: str) -> dict:
+async def _check_service_connection(service_type: str, api_url: str, api_key: str) -> dict:
-    """测试外部知识库服务连接状态（当前仅支持 Dify）"""
+    """测试外部知识库服务连接状态（支持 Dify）"""
     try:
         if service_type == "dify":
             from ...dify_kb import dify_test
@@ -793,13 +798,13 @@ async def create_external_service(
     db: Session = Depends(get_db),
 ):
     """创建新的外部知识库服务连接配置"""
-    # 验证服务类型（当前仅支持 Dify）
+    # 验证服务类型（当前仅支持 Dify；AnySearch 已改为全局固定配置）
     valid_types = ["dify"]
     if req.service_type not in valid_types:
         raise HTTPException(status_code=400, detail=f"不支持的服务类型: {req.service_type}")
 
     # 检查连接
-    conn_result = _check_service_connection(req.service_type, req.api_url, req.api_key or "")
+    conn_result = await _check_service_connection(req.service_type, req.api_url, req.api_key or "")
 
     service_id = str(uuid.uuid4())
     config_json = json.dumps(req.config or {}, ensure_ascii=False)
@@ -877,6 +882,54 @@ async def create_external_service(
     }
 
 
+@router.get("/anysearch/config", summary="获取 AnySearch 全局配置")
+async def get_anysearch_config(
+    admin_user=Depends(get_current_admin),
+    db: Session = Depends(get_db),
+):
+    """读取全局 AnySearch 联网检索配置（site_settings 表）。"""
+    from ...anysearch_client import load_anysearch_config, DEFAULT_API_URL
+    cfg = load_anysearch_config(db)
+    return {
+        "code": 0,
+        "data": {
+            "api_key": cfg.get("api_key", ""),
+            "is_enabled": cfg.get("is_enabled", False),
+            "api_url": cfg.get("api_url") or DEFAULT_API_URL,
+            "zone": cfg.get("zone", "cn"),
+            "freshness": cfg.get("freshness"),
+            "top_k": cfg.get("top_k", 5),
+        },
+    }
+
+
+@router.post("/anysearch/config", summary="保存 AnySearch 全局配置")
+async def save_anysearch_config_endpoint(
+    req: AnySearchConfig,
+    admin_user=Depends(get_current_admin),
+    db: Session = Depends(get_db),
+):
+    """保存全局 AnySearch 联网检索配置，并测试连通性。"""
+    from ...anysearch_client import save_anysearch_config, anysearch_health, DEFAULT_API_URL
+    ok = save_anysearch_config(db, api_key=req.api_key, is_enabled=req.is_enabled, api_url=DEFAULT_API_URL)
+    if not ok:
+        raise HTTPException(status_code=500, detail="保存 AnySearch 配置失败")
+    health = await anysearch_health(req.api_key or "", DEFAULT_API_URL)
+    LogService.log_audit(
+        admin_id=admin_user.id,
+        admin_name=admin_user.username,
+        action="update_anysearch_config",
+        target_type="site_setting",
+        target_id="anysearch_config",
+        detail=f"更新 AnySearch 全局配置: enabled={req.is_enabled}, connected={health.get('connected')}",
+    )
+    return {
+        "code": 0,
+        "message": "AnySearch 配置已保存",
+        "data": {"connection": health},
+    }
+
+
 @router.put("/external/services/{service_id}", summary="更新外部知识库服务配置")
 async def update_external_service(
     service_id: str,
@@ -916,7 +969,7 @@ async def update_external_service(
     new_key = req.api_key or row[4]
     if req.api_url or req.api_key:
         service_type = row[2]
-        conn_result = _check_service_connection(service_type, new_url, new_key)
+        conn_result = await _check_service_connection(service_type, new_url, new_key)
         updates["status"] = conn_result.get("status", "disconnected")
 
     updates["updated_at"] = "NOW()"
@@ -999,7 +1052,7 @@ async def check_external_service_status(
     if not row:
         raise HTTPException(status_code=404, detail="外部知识库服务不存在")
 
-    conn_result = _check_service_connection(row[0], row[1], row[2])
+    conn_result = await _check_service_connection(row[0], row[1], row[2])
 
     # 更新状态
     try:

backend/python/app/api/admin/mcp_admin.py

@@ -30,6 +30,7 @@ Phase 3.3: Manage MCP servers, discover tools/resources, test connections
 
 
 
+import asyncio
 from fastapi import APIRouter, Depends, HTTPException, Query, Body
 
 
@@ -842,110 +843,29 @@ def _test_mcp_server(server, db) -> dict:
 
 
         if transport in ("streamable_http", "streamablehttp"):
-
-
-
             url = server.url
-
-
-
             if not url:
-
-
-
                 result["message"] = "Streamable HTTP endpoint URL is required"
-
-
-
             else:
-
-
-
-                import httpx
-
-
-
                 try:
-
+                    # Use the real MCP SDK to verify connectivity (same as chat flow).
-
+                    # A naive HTTP GET/POST can give false negatives for endpoints
-
+                    # that only accept proper MCP JSON-RPC over POST (e.g., AnySearch).
-                    # A real streamable-http MCP endpoint accepts POST JSON-RPC.
+                    from ...mcp_client import list_mcp_tools as _lmt
-
+                    hdrs = json.loads(server.headers) if isinstance(server.headers, str) else (server.headers or {})
-
+                    sd = {"url": url, "transport_type": "streamable_http", "headers": hdrs}
-
+                    import asyncio as _aio
-                    # Some implementations also accept GET for SSE events.
+                    tools = _aio.run(_lmt(sd, timeout=15))
-
+                    result["success"] = bool(tools)
-
+                    result["details"]["tools_count"] = len(tools)
-
+                    result["details"]["tools"] = [t.get("name") for t in tools[:5]]
-                    # Try GET first (cheapest), fall back to a minimal POST.
+                    result["message"] = f"Connected — {len(tools)} tools available" if tools else "Connected but no tools found"
-
+                    result["details"]["status_code"] = 200
-
+                    if tools:
-
+                        server.tools_count = len(tools)
-                    resp = httpx.get(url, headers=req_headers, timeout=10.0, follow_redirects=True)
+                        server.status = "connected"
-
-
-
-                    if resp.status_code == 405:
-
-
-
-                        resp = httpx.post(
-                            url,
-                            headers=req_headers,
-                            json={"jsonrpc": "2.0", "method": "initialize", "id": 1},
-                            timeout=10.0,
-                            follow_redirects=True,
-                        )
-
-
-
-                    # Any HTTP response (including 4xx) means the host is reachable.
-
-
-
-                    result["success"] = resp.status_code < 500
-
-
-
-                    result["message"] = f"HTTP responded with {resp.status_code}"
-
-
-
-                    result["details"]["status_code"] = resp.status_code
-
-
-
-                    result["details"]["content_type"] = resp.headers.get("content-type", "")
-
-
-
-                except httpx.ConnectError as e:
-
-
-
-                    result["message"] = f"Connection refused - server may be down: {str(e)}"
-
-
-
-                except httpx.TimeoutException:
-
-
-
-                    result["message"] = "Connection timed out"
-
-
-
                 except Exception as e:
-
+                    result["message"] = f"Connection failed: {str(e)}"
-
-
-                    result["message"] = f"HTTP request failed: {str(e)}"
-
-
-
-
-
-
 
         elif transport == "sse":
 
@@ -3156,7 +3076,7 @@ async def test_mcp_connection(
 
 
 
-    result = _test_mcp_server(server, db)
+    result = await asyncio.get_running_loop().run_in_executor(None, _test_mcp_server, server, db)
 
 
 

backend/python/app/api/admin/models.py

@@ -869,6 +869,9 @@ async def test_model_connection(
 
     model_type = _detect_model_type(ai_model.capabilities)
     timeout = ai_model.timeout or 60
+    # 图像生成实际会出图，单张耗时较长；连接测试需给足超时，否则会误报"失败"
+    if model_type == "image":
+        timeout = max(min(ai_model.timeout or 60, 200), 120)
     headers = {
         "Authorization": f"Bearer {api_key}",
         "Content-Type": "application/json",
@@ -927,9 +930,15 @@ async def test_model_connection(
                     "model": ai_model.name,
                     "prompt": "a red circle",
                     "n": 1,
-                    "size": "256x256",
+                    "size": "1024x1024",
                 }
-                resp = await client.post(image_url, json=payload, headers=headers)
+                # 与运行时媒体生成一致：用 _proxy_for_url 解析代理并 trust_env=False；
+                # 图像端点会被 _proxy_for_url 判定为直连(公网 API 直连比代理更稳定)
+                from ..chat_utils import _proxy_for_url
+                async with httpx.AsyncClient(
+                    timeout=timeout, proxy=_proxy_for_url(image_url), trust_env=False
+                ) as img_client:
+                    resp = await img_client.post(image_url, json=payload, headers=headers)
                 latency = int((time.time() - start) * 1000)
 
                 if resp.status_code == 200:

backend/python/app/api/admin/privacy_admin.py

@@ -442,7 +442,7 @@ def revoke_license(
         raise HTTPException(404, "License not found")
 
     _execute_raw(db, text(
-        "UPDATE offline_licenses SET status = 'revoked', revoked_reason = :reason, is_active = 0 WHERE id = :lid"
+        "UPDATE offline_licenses SET status = 'revoked', revoked_reason = :reason, is_active = false WHERE id = :lid"
     ), {"reason": reason, "lid": license_id})
 
     return {"code": 0, "message": "License revoked"}
@@ -623,7 +623,7 @@ def get_usage_stats(db: Session = Depends(get_db)):
         "SELECT COUNT(DISTINCT user_hash) FROM usage_reports"
     ))
     today_reports = _execute_raw(db, text(
-        "SELECT COUNT(*) FROM usage_reports WHERE DATE(created_at) = CURDATE()"
+        "SELECT COUNT(*) FROM usage_reports WHERE DATE(created_at) = CURRENT_DATE"
     ))
     by_type = _execute_raw(db, text(
         "SELECT report_type, COUNT(*) as cnt FROM usage_reports GROUP BY report_type"

backend/python/app/api/admin/skills_admin.py

@@ -37,17 +37,31 @@ SKILL_STATES = {
     "frozen": {"label": "已冻结", "color": "#EF4444", "description": "已停止使用"},
 }
 
+# 业务场景分类（与前端「应用市场」保持一致，14 类岗位/职能视角）
+# 统一标准：以前台 MARKET_SCENES 为准，后台技能管理也按此 14 类展示与筛选。
 SKILL_CATEGORIES = [
-    {"id": "writing", "label": "文字创作"},
+    {"id": "design", "label": "🎨 做设计"},
-    {"id": "coding", "label": "代码能力"},
+    {"id": "frontend", "label": "💻 写前端"},
-    {"id": "data_analysis", "label": "数据分析"},
+    {"id": "mobile", "label": "📱 做移动端"},
-    {"id": "qa", "label": "知识问答"},
+    {"id": "backend", "label": "🧪 后端/全栈"},
-    {"id": "automation", "label": "流程自动化"},
+    {"id": "qa", "label": "🧪 测试质量"},
-    {"id": "creativity", "label": "创意生成"},
+    {"id": "search", "label": "🔍 搜索信息"},
-    {"id": "communication", "label": "沟通协作"},
+    {"id": "docs", "label": "✍️ 写文档"},
-    {"id": "reasoning", "label": "逻辑推理"},
+    {"id": "marketing", "label": "📢 推广获客"},
-    {"id": "other", "label": "其他"},
+    {"id": "workflow", "label": "🔄 工作流"},
+    {"id": "secops", "label": "🛡️ 安全运维"},
+    {"id": "learning", "label": "🎓 学习教学"},
+    {"id": "video", "label": "🎬 视频多媒体"},
+    {"id": "collab", "label": "🏢 办公协作"},
+    {"id": "ai", "label": "🛠️ AI能力"},
 ]
+_SCENE_IDS = {c["id"] for c in SKILL_CATEGORIES}
+# 旧 9 类 taxonomy -> 14 场景 的兜底映射（仅用于历史数据 capability 缺失时）
+_CATEGORY_TO_SCENE = {
+    "writing": "docs", "coding": "backend", "data_analysis": "backend",
+    "qa": "qa", "automation": "workflow", "creativity": "design",
+    "communication": "marketing", "reasoning": "search", "other": "workflow",
+}
 
 
 # ============================================================
@@ -65,7 +79,13 @@ def market_app_to_skill_dict(app: "MarketApp") -> dict:
             tags_list = json.loads(raw_tags) if isinstance(raw_tags, str) else list(raw_tags)
         except Exception:
             tags_list = [t.strip() for t in str(raw_tags).split(",")]
-    category = getattr(app, "category", None) or "other"
+    # 以业务场景(capability, 14 类)为分类标准，与前端应用市场统一；
+    # capability 缺失时回退到旧 9 类 category 映射出的场景。
+    raw_scene = getattr(app, "capability", None)
+    if raw_scene in _SCENE_IDS:
+        category = raw_scene
+    else:
+        category = _CATEGORY_TO_SCENE.get(getattr(app, "category", None), "workflow")
     return {
         "id": app.id,
         "user_id": getattr(app, "publisher_id", None) or "system",
@@ -187,7 +207,8 @@ async def list_skills(
             or_(MarketApp.name.contains(keyword), MarketApp.description.contains(keyword))
         )
     if category:
-        mkt_query = mkt_query.filter(MarketApp.category == category)
+        # 分类筛选统一按业务场景(capability, 14 类)进行，与前端应用市场一致
+        mkt_query = mkt_query.filter(MarketApp.capability == category)
     # 状态映射：主面板只展示已发布 → state='main'；草稿/移除 → 'baby'
     if state == "main":
         mkt_query = mkt_query.filter(MarketApp.status == "published")
@@ -901,7 +922,7 @@ async def get_skills_stats(
     market_by_category = {}
     try:
         market_by_category_rows = db.execute(
-            text("SELECT category, COUNT(*) FROM market_apps WHERE type='skill' AND category IS NOT NULL GROUP BY category")
+            text("SELECT capability, COUNT(*) FROM market_apps WHERE type='skill' AND capability IS NOT NULL GROUP BY capability")
         ).fetchall()
         market_by_category = {cat or "other": cnt for cat, cnt in market_by_category_rows}
     except Exception:

backend/python/app/api/chat_media.py

@@ -3,7 +3,8 @@
 
 from fastapi import APIRouter, Depends, HTTPException, Query, Response
 from sqlalchemy.orm import Session
-from sqlalchemy import and_, desc, asc, func
+from sqlalchemy import and_, desc, asc, func, type_coerce
+from sqlalchemy.dialects.postgresql import JSONB
 from pydantic import BaseModel, Field
 from typing import Optional
 import uuid
@@ -839,6 +840,86 @@ async def _agnes_post_json(api_url: str, headers: dict, payload: dict, timeout)
 
     return None, last_err
 
+
+def _media_api_url(ai_model, provider, media_kind: str) -> str:
+    """推导图像/视频生成地址，避免把 chat/completions 端点误用作媒体端点。
+
+    - 若模型 endpoint 已是对应的媒体端点(含 /images/generations 或 /videos)，直接使用；
+    - 否则从 chat endpoint 反推 base，再拼 /v1/images|videos/generations；
+    - 若 endpoint 为空，则回退供应商 api_base_url + 媒体路径。
+    """
+    suffix = "/v1/images/generations" if media_kind == "image" else "/v1/videos"
+    chat_suffixes = ["/v1/chat/completions", "/chat/completions"]
+    ep = (ai_model.endpoint or "").strip()
+    if ep:
+        if suffix in ep:
+            return ep
+        base = ep
+        for cs in chat_suffixes:
+            if base.endswith(cs):
+                base = base[: -len(cs)]
+                break
+        else:
+            if "/v1" in base:
+                base = base.rsplit("/v1", 1)[0]
+        return f"{base.rstrip('/')}{suffix}"
+    if provider and provider.api_base_url:
+        return f"{provider.api_base_url.rstrip('/')}{suffix}"
+    return ""
+
+def _build_media_route(db, capability: str, preferred_id: str = None):
+    """构造媒体生成模型路由顺序：
+    1) 按能力默认模型(后台"能力默认模型"或全局默认) 作为第一顺位
+    2) 调用方显式指定的模型(若不同于默认) 次位
+    3) 其余启用且具备该能力的模型，按 is_default 降序、created_at 升序 排列
+    去重后返回 AIModel 列表。
+    """
+    from ..models import AIModel as _AM
+
+    selected = []
+    seen = set()
+
+    def _add(m):
+        if m and m.id not in seen:
+            selected.append(m)
+            seen.add(m.id)
+
+    default_model = resolve_auto_model(capability, db)
+    if default_model:
+        _add(default_model)
+
+    if preferred_id:
+        pref = db.query(_AM).filter(_AM.id == preferred_id, _AM.is_enabled == True).first()
+        _add(pref)
+
+    others = db.query(_AM).filter(
+        _AM.is_enabled == True,
+        _AM.capabilities.cast(JSONB).contains(type_coerce([capability], JSONB)),
+    ).order_by(_AM.is_default.desc(), _AM.created_at.asc()).all()
+    for m in others:
+        _add(m)
+
+    return selected
+
+
+async def _run_media_route(db, route, prompt, user_id, media_kind):
+    """依次尝试路由中的模型生成媒体；首个成功(非 ⚠️ 开头)即返回，否则尝试下一个。"""
+    last_err = None
+    for idx, m in enumerate(route, 1):
+        logger.info(f"[_run_media_route] ({idx}/{len(route)}) 尝试模型 {m.name} ({m.id})")
+        try:
+            res = await _generate_media(db, m, prompt, user_id)
+        except Exception as e:
+            last_err = f"模型 {m.name} 异常: {str(e)[:160]}"
+            logger.warning(f"[_run_media_route] 模型 {m.name} 异常: {e}")
+            continue
+        if res and not res.startswith("⚠️"):
+            return res
+        last_err = res
+        logger.warning(f"[_run_media_route] 模型 {m.name} 失败: {str(res)[:160]}")
+    return last_err or "⚠️ 所有模型均生成失败"
+
+
 async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str:
 
 
@@ -880,6 +961,9 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
 
 
     timeout = ai_model.timeout or 120
+    # 单次媒体调用超时封顶：避免单个供应商长时间挂起拖垮整条生成链路；
+    # 多模型 failover 由 _generate_media_task 外层循环保证
+    call_timeout = min(timeout, 300 if "video" in caps else 180)
 
 
 
@@ -895,15 +979,7 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
 
 
 
-        api_url = ai_model.endpoint
+        api_url = _media_api_url(ai_model, provider, "image")
-
-
-
-        if not api_url:
-
-
-
-            api_url = f"{provider.api_base_url.rstrip('/')}/v1/images/generations" if provider else ""
 
 
 
@@ -951,7 +1027,7 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
 
 
 
-            "size": "1024x768",
+            "size": "1024x1024",
 
 
 
@@ -960,7 +1036,7 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
 
 
         async with span("media.agnes_call"):
-            resp, err = await _agnes_post_json(api_url, headers, payload, timeout)
+            resp, err = await _agnes_post_json(api_url, headers, payload, call_timeout)
 
 
 
@@ -992,7 +1068,7 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
 
 
 
-                resp2, _ = await _agnes_post_json(api_url, headers, payload, timeout)
+                resp2, _ = await _agnes_post_json(api_url, headers, payload, call_timeout)
 
 
 
@@ -1092,15 +1168,7 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
 
 
 
-        api_url = ai_model.endpoint
+        api_url = _media_api_url(ai_model, provider, "video")
-
-
-
-        if not api_url:
-
-
-
-            api_url = f"{provider.api_base_url.rstrip('/')}/v1/videos" if provider else ""
 
 
 
@@ -1169,7 +1237,7 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
 
 
         async with span("media.agnes_call"):
-            resp, err = await _agnes_post_json(api_url, headers, payload, timeout)
+            resp, err = await _agnes_post_json(api_url, headers, payload, call_timeout)
 
 
 
@@ -1197,7 +1265,7 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
 
 
 
-                resp2, _ = await _agnes_post_json(api_url, headers, payload, timeout)
+                resp2, _ = await _agnes_post_json(api_url, headers, payload, call_timeout)
 
 
 
@@ -1430,22 +1498,27 @@ async def _generate_media_task(ai_msg_id: str, ai_model_id: str, prompt: str, us
         logger.info(f"[_generate_media_task] 启动 media 生成 msg={ai_msg_id} model={ai_model_id} prompt={prompt[:60]!r}")
         caps = [c.lower() for c in (ai_model.capabilities or [])]
         is_video = "video" in caps
-        timeout_seconds = 300 if is_video else 90
+        media_kind = "video" if is_video else "image"
-        try:
-            result = await asyncio.wait_for(
-                _generate_media(db, ai_model, prompt, user_id),
-                timeout=timeout_seconds
-            )
 
-
+        # 构造模型路由：默认模型第一顺位，之后按路由策略逐一尝试
-
+        route = _build_media_route(db, media_kind, ai_model_id)
-        except asyncio.TimeoutError:
+        if not route:
-            logger.warning(f"[_generate_media_task] media 生成超时 msg={ai_msg_id} is_video={is_video} timeout={timeout_seconds}s")
+            result = f"⚠️ 未找到可用的{'视频' if is_video else '图像'}模型，请在后台「大模型库」中启用具备该能力的模型。"
+        else:
             _media_name = "视频" if is_video else "图像"
-            result = f"⚠️ 媒体生成超时：{_media_name}生成耗时超过 {timeout_seconds} 秒，请稍后重试。"
+            per_call = 300 if is_video else 180
-        except Exception as e:
+            route_timeout = min(per_call * len(route) + 30, 900 if is_video else 600)
-            logger.warning(f"[_generate_media_task] media 生成异常 msg={ai_msg_id}: {e}", exc_info=True)
+            try:
-            result = f"⚠️ 媒体生成失败：{str(e)[:200]}"
+                result = await asyncio.wait_for(
+                    _run_media_route(db, route, prompt, user_id, media_kind),
+                    timeout=route_timeout,
+                )
+            except asyncio.TimeoutError:
+                logger.warning(f"[_generate_media_task] media 路由全部超时 msg={ai_msg_id} is_video={is_video} timeout={route_timeout}s")
+                result = f"⚠️ {_media_name}生成超时：已依次尝试 {len(route)} 个模型仍失败，请稍后重试。"
+            except Exception as e:
+                logger.warning(f"[_generate_media_task] media 路由异常 msg={ai_msg_id}: {e}", exc_info=True)
+                result = f"⚠️ {_media_name}生成失败：{str(e)[:200]}"
 
 
 
@@ -1494,11 +1567,16 @@ async def _generate_media_task(ai_msg_id: str, ai_model_id: str, prompt: str, us
 
 
     except Exception as e:
-
+        logger.warning(f"[_generate_media_task] 后台生成异常: {e}", exc_info=True)
-
+        # 兜底回写：任何意外异常都不能让消息永久停在「⏳ 生成中…」
-
+        try:
-        logger.warning(f"[_generate_media_task] 后台生成异常: {e}")
+            db.rollback()
-
+            _err = db.query(Message).filter(Message.id == ai_msg_id).first()
+            if _err and str(_err.content or "").startswith("⏳"):
+                _err.content = "⚠️ 媒体生成失败，请稍后重试"
+                db.commit()
+        except Exception:
+            pass
 
 
     finally:

backend/python/app/api/chat_skill.py

@@ -1272,7 +1272,7 @@ def _execute_skill_tool(
 
 
 
-        return f"工具 '{func_name}' 执行结果: 已处理输入内容, 但未找到对应的技能定义。请基于通用知识回复用户。"
+        return f"工具 '{func_name}' 执行结果: 已处理输入内容, 但未找到对应的技能定义。请基于通用知识回复用户。如有已加载的技能，可使用其名称直接调用。"
 
 
 
@@ -1593,3 +1593,150 @@ def _execute_skill_tool(
 
 
     return result
+
+
+def _generate_install_fallback_skill(name: str, description: str) -> str:
+    """install_app 兜底：无法识别为 SKILL.md 时，生成一份可用技能定义。"""
+    desc = (description or "").strip() or f"{name} 相关能力"
+    return (
+        f"---\n"
+        f"name: {name}\n"
+        f"description: {desc}\n"
+        f"category: other\n"
+        f"---\n\n"
+        f"# {name}\n\n"
+        f"{desc}\n\n"
+        f"## 使用建议\n"
+        f"- 明确用户意图后再执行，必要时先向用户确认方向。\n"
+        f"- 输出结构清晰、可直接使用。\n"
+        f"- 如任务超出能力范围，主动说明并给出替代方案。\n"
+    )
+
+
+def _execute_install_app(db: Session, url: str, user_id, app_type: str = "auto") -> str:
+    """前台 AI 聊天：通过 URL 安装 skill / mcp / agent（根治问题 #3）。
+
+    逻辑对齐 market.upload_app，但本函数不引用 market 私有函数以避免循环 import：
+    - 同步 httpx.Client(trust_env=False, 显式 proxy) 抓取 URL 内容（与 _fetch_skill_content 同款）
+    - 自动识别类型：mcp / agent / skill
+    - 建 MarketApp（+ mcp 时 MCPServer + AppInstallation），自动标记为已安装
+
+    关键约束：本函数在聊天主事件循环中【同步】执行，禁止调用 asyncio.run
+    （会抛 "cannot be called from a running event loop"）。因此 MCP 连接测试交由后台
+    「测试」按钮在 worker 线程异步执行；新建 MCPServer 直接 is_enabled=True 即可被
+    聊天链路 _resolve_active_mcp_servers 使用（该解析仅过滤 is_enabled）。
+    """
+    if not url or not str(url).strip().lower().startswith("http"):
+        return "请提供有效的安装地址(URL)，例如 GitHub 仓库、mcp.so 或 modelscope 链接。"
+    url = str(url).strip()
+    proxy = os.environ.get("HTTP_PROXY") or os.environ.get("http_proxy") or None
+    try:
+        with httpx.Client(timeout=25, follow_redirects=True, trust_env=False, proxy=proxy) as client:
+            r = client.get(url, headers={"User-Agent": "Mozilla/5.0 (compatible; QihuAI/1.0)"})
+            content = r.text if r.status_code == 200 else ""
+    except Exception as e:
+        return f"无法访问该地址：{str(e)[:200]}"
+
+    _u = url.lower()
+    _c = (content or "").lower()
+    detected = "skill"
+    if app_type and app_type != "auto":
+        detected = app_type
+    else:
+        if "mcpservers" in _c or '"mcpservers"' in (content or "").lower() or (
+            "url" in _c and ("transport" in _c or '"type"' in _c)
+        ):
+            detected = "mcp"
+        elif any(k in _u for k in ["dify", "n8n", "coze"]) or "agent_endpoint" in _c:
+            detected = "agent"
+        else:
+            detected = "skill"
+
+    import re as _re
+    _name = _re.sub(r"^https?://", "", url).strip("/").split("/")[-1] or "imported_app"
+    _name = _name[:120]
+    _desc = f"通过链接导入：{url}"
+
+    _dup = db.query(MarketApp).filter(
+        func.lower(MarketApp.name) == _name.lower(),
+        MarketApp.type == detected,
+    ).first()
+    if _dup:
+        return f"「{_name}」({detected}) 已存在，无需重复安装（ID: {_dup.id}）。可在应用市场直接使用。"
+
+    app = MarketApp(
+        id=gen_uuid(),
+        name=_name,
+        description=_desc,
+        type=detected,
+        source="user",
+        category="other",
+        tags=None,
+        price=0,
+        install_count=0,
+        rating_avg=0,
+        review_count=0,
+        publisher_id=user_id,
+        status="published",
+        created_at=datetime.now(),
+        updated_at=datetime.now(),
+    )
+
+    if detected == "mcp":
+        from ..api.admin.mcp_admin import parse_mcp_config
+        parsed = parse_mcp_config(content)
+        _murl = parsed.get("url") or url
+        app.skill_content = content
+        app.mcp_transport = parsed.get("transport_type", "streamable_http")
+        db.add(app)
+        db.commit()
+        db.refresh(app)
+        _hdr = json.dumps(parsed.get("headers") or {}, ensure_ascii=False)
+        server = MCPServer(
+            id=app.id,
+            name=app.name,
+            description=app.description or "",
+            transport_type=parsed.get("transport_type", "streamable_http"),
+            command="",
+            args="{}",
+            env_vars="{}",
+            url=_murl,
+            headers=_hdr,
+            is_enabled=True,
+            status="disconnected",
+            tools_count=0,
+            resources_count=0,
+            created_at=app.created_at,
+            updated_at=app.updated_at,
+        )
+        db.add(server)
+        db.commit()
+    elif detected == "agent":
+        app.agent_endpoint = url
+        if "dify" in _u:
+            app.platform = "dify"
+        elif "n8n" in _u:
+            app.platform = "n8n"
+        elif "coze" in _u:
+            app.platform = "coze"
+        app.skill_content = content if "{" in content else None
+        db.add(app)
+        db.commit()
+        db.refresh(app)
+    else:
+        _is_skillmd = ("name:" in _c and "description:" in _c) or _u.endswith(".md")
+        app.skill_content = content if _is_skillmd else _generate_install_fallback_skill(_name, _desc)
+        db.add(app)
+        db.commit()
+        db.refresh(app)
+
+    if user_id:
+        _inst = db.query(AppInstallation).filter(
+            AppInstallation.user_id == user_id,
+            AppInstallation.app_id == app.id,
+        ).first()
+        if not _inst:
+            db.add(AppInstallation(user_id=user_id, app_id=app.id))
+            db.commit()
+
+    return f"已成功安装「{_name}」({detected})。你现在可以直接在对话中使用它，或在应用市场查看。"

backend/python/app/api/chat_utils.py

@@ -147,6 +147,15 @@ def _proxy_for_url(url: str) -> str | None:
 
 
 
+    # 文生图 / 视频生成走直连：智能路由代理(172.18.0.1:7890)与 HK 代理对这类长耗时、
+    # 大响应体的生成连接会偶发重置(Server disconnected)，而各图像/视频供应商均为公网
+    # API，直连更稳定(已实测 gpt-image-2 / agnes 直连均返回 200)。故媒体端点统一直连。
+    _lower = (url or "").lower()
+    if "/images/" in _lower or "/videos/" in _lower or "/v1/images" in _lower or "/v1/videos" in _lower:
+        return None
+
+
+
     if "agnes-ai.com" in host or "agnes-ai.space" in host:
 
 

backend/python/app/api/im_bot_platform.py

@@ -109,20 +109,20 @@ def _ensure_table(db):
     """确保 im_bot_platform_account 表存在（首次调用时创建）。"""
     db.execute(text("""
         CREATE TABLE IF NOT EXISTS im_bot_platform_account (
-            id INT AUTO_INCREMENT PRIMARY KEY,
+            id SERIAL PRIMARY KEY,
             user_id VARCHAR(64) DEFAULT NULL,
             platform VARCHAR(32) NOT NULL,
             status VARCHAR(32) DEFAULT 'disconnected',
-            creds MEDIUMTEXT NULL,
+            creds TEXT NULL,
             send_url VARCHAR(512) NULL,
             api_base VARCHAR(512) NULL,
             bot_token VARCHAR(512) NULL,
             webhook_token VARCHAR(64) NULL,
-            last_error MEDIUMTEXT NULL,
+            last_error TEXT NULL,
-            created_at DATETIME DEFAULT CURRENT_TIMESTAMP,
+            created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
-            updated_at DATETIME DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP,
+            updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
-            UNIQUE KEY uk_user_platform (user_id, platform)
+            UNIQUE (user_id, platform)
-        ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_unicode_ci
+        )
     """))
     db.commit()
 

backend/python/app/api/im_bot_user.py

@@ -458,9 +458,8 @@ def _wechat_receiver_loop(account_id: int):
                     except Exception:
                         pass
                     if ret in (-2, -14):
-                        # 会话/鉴权失效：标记需重扫并停止接收线程
+                        # 会话/鉴权失效：标记需重扫但继续循环，让下一轮 poll 自然检测过期
                         _mark_wechat_expired(d["id"])
-                        return
             time.sleep(1)
         except Exception:
             time.sleep(5)

backend/python/app/api/im_llm.py

@@ -4,13 +4,16 @@ IM Bot 真实 LLM 调用。
 原实现位于 channels.py（通用频道管理模块），该模块已移除。
 为保持 IM Bot（微信机器人等）正常生成回复，将 _call_real_llm 迁移至此独立模块。
 
-行为（对齐 AI 聊天）：
+行为（对齐 AI 聊天 — 2026-07-23 同步至最新）：
 - Issue 7：图像/视频/语音意图优先走平台媒体模型（与 AI 聊天同一套 Auto 路由），
   不再对"制作图片"这类请求只回文本。
 - Issue 8：默认 Auto 模型；每轮先自动发现本账号已安装且平台已部署的技能，
   多个适用时由模型自行判断选 1 个最优，通过 function calling 执行。
 - 2026-07-15：修复 trust_env=False + proxy 处理，避免容器坏代理导致调用失败；
   媒体生成前先经 _optimize_media_prompt 精炼提示词，提升 AGNES 生成质量。
+- 2026-07-23：对齐 Web Chat 最新能力：
+  * 模型 failover（Auto 模式自动切换备选模型）
+  * 平台知识库注入（get_platform_knowledge，技能清单实时同步）
 """
 import asyncio
 import json
@@ -21,6 +24,7 @@ from typing import Optional
 from urllib.parse import urlparse as _urlparse
 
 import httpx
+from app.core.llm_client import chat_complete  # 统一模型层(07-24): web/IM 共用标准 OpenAI SDK 客户端
 from sqlalchemy.orm import Session
 
 from ..capability_registry import resolve_auto_model
@@ -83,7 +87,7 @@ def _run_async(coro, timeout=None):
 def _log_media_usage(db: Session, user_id: str, ai_model, media_type: str, prompt: str):
     """IM Bot 媒体生成记账：把图片/视频消耗写入 TokenUsageLog，供驾驶舱统计。
 
-    此前 IM Bot 路径完全不写 TokenUsageLog，导致微信生成的图片/视频在驾驶舱里“查无此项”。
+    此前 IM Bot 路径完全不写 TokenUsageLog，导致微信生成的图片/视频在驾驶舱里"查无此项"。
     媒体 API 不返回 token 用量，这里用与 chat.py _generate_media_task 一致的估算。
     """
     if not user_id or not ai_model:
@@ -112,7 +116,7 @@ def _log_media_usage(db: Session, user_id: str, ai_model, media_type: str, promp
 
 
 def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im", user_id: str = None) -> "str | dict":
-    """调用七虎AI 生成回复，供 IM Bot 复用。支持媒体意图路由与已安装技能自动发现。"""
+    """调用七虎AI 生成回复，供 IM Bot 复用。支持媒体意图路由、已安装技能自动发现、模型 failover。"""
     try:
         from app.api.chat import (
             _generate_media,
@@ -151,19 +155,23 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                         return s
                     return "（媒体生成未能返回有效内容，请重试）"
 
-        # ===== 2) 文本回复 + 技能自动发现（Issue 8）=====
+        # ===== 2) 文本回复：模型 failover + 技能自动发现 + 平台知识库注入 =====
-        ai_model = resolve_auto_model("text", db)
+        from ..platform_knowledge import get_platform_knowledge
-        if not ai_model:
-            return "（未配置默认文本模型，无法生成回复）"
-        provider = db.query(ModelProvider).filter(ModelProvider.id == ai_model.provider_id).first()
-        api_key = ai_model.api_key or (provider.api_key if provider else None)
-        api_url = ai_model.endpoint
-        if not api_url and provider and provider.api_base_url:
-            api_url = f"{provider.api_base_url.rstrip('/')}/v1/chat/completions"
-        if not (api_key and api_url):
-            return "（模型未配置 API Key 或地址，无法生成回复）"
 
-        # 自动发现本账号已安装且平台已部署的技能
+        # 构建候选模型列表（对齐 Web Chat failover）
+        ai_model = resolve_auto_model("text", db)
+        _try_models = [ai_model] if (ai_model and ai_model.is_enabled) else []
+        # Auto 模式附加备选（IM Bot 始终 Auto）
+        _fallbacks = db.query(AIModel).filter(
+            AIModel.is_enabled == True,
+            AIModel.id != (ai_model.id if ai_model else ""),
+        ).order_by(AIModel.is_default.desc()).limit(4).all()
+        _try_models.extend([m for m in _fallbacks if m not in _try_models])
+
+        if not _try_models:
+            return "（未配置可用文本模型，无法生成回复）"
+
+        # 技能自动发现（提前执行一次，不必每个候选模型重复做）
         skills = []
         if user_id:
             try:
@@ -171,64 +179,104 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
             except Exception:
                 skills = []
 
-        system_prompt, function_definitions = _build_system_prompt(
+        # 平台知识库注入（对齐 Web Chat——微信用户提问"平台有哪些技能"也能得到完整清单）
-            skills=skills,
+        kb_context = ""
-            agents=[],
+        try:
-            kb_context="",
+            pk = get_platform_knowledge(db=db)
-            user_content=prompt,
+            if pk:
-            mcp=[],
+                kb_context = f"【平台知识库——优先使用以下知识回答用户问题】\n{pk}"
-            auto_discovered=bool(skills),
+        except Exception:
-        )
+            pass
-        messages = [
-            {"role": "system", "content": system_prompt},
-            {"role": "user", "content": prompt},
-        ]
 
-        max_rounds = 3
+        # 尝试每个候选模型
-        ai_content = ""
+        _retry_log = []
-        for _ in range(max_rounds):
+        for _try_model in _try_models:
-            request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048}
+            ai_model = _try_model
-            if function_definitions and skills:
+
-                request_body["tools"] = [
+            provider = db.query(ModelProvider).filter(ModelProvider.id == ai_model.provider_id).first()
-                    {"type": "function", "function": fd} for fd in function_definitions
+            api_key = ai_model.api_key or (provider.api_key if provider else None)
-                ]
+            api_url = ai_model.endpoint
-                request_body["tool_choice"] = "auto"
+            if not api_url and provider and provider.api_base_url:
-            with httpx.Client(
+                api_url = f"{provider.api_base_url.rstrip('/')}/v1/chat/completions"
-                timeout=ai_model.timeout or 120,
+            if not (api_key and api_url):
-                proxy=_proxy_for_url(api_url),
+                _retry_log.append(f"{ai_model.name}: 缺少密钥或端点")
-                trust_env=False,
+                continue
-            ) as client:
+
-                resp = client.post(
+            system_prompt, function_definitions = _build_system_prompt(
-                    api_url,
+                skills=skills,
-                    headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+                agents=[],
-                    json=request_body,
+                kb_context=kb_context,
-                )
+                user_content=prompt,
-                if resp.status_code != 200:
+                mcp=[],
-                    return f"（AI 模型调用失败 HTTP {resp.status_code}）"
+                auto_discovered=bool(skills),
-                data = resp.json()
+            )
-                message = data.get("choices", [{}])[0].get("message", {})
+            messages = [
-                tool_calls = message.get("tool_calls", [])
+                {"role": "system", "content": system_prompt},
-                if tool_calls and skills:
+                {"role": "user", "content": prompt},
-                    messages.append({
+            ]
-                        "role": "assistant",
+
-                        "content": message.get("content") or "",
+            # 尝试调用
-                        "tool_calls": tool_calls,
+            try:
-                    })
+                max_rounds = 3
-                    for tc in tool_calls:
+                ai_content = ""
-                        func_name = tc.get("function", {}).get("name", "")
+                _http_ok = True
-                        try:
+                for round_num in range(max_rounds):
-                            func_args = json.loads(tc.get("function", {}).get("arguments", "{}"))
+                    request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048}
-                        except Exception:
+                    if function_definitions and skills:
-                            func_args = {"input": tc.get("function", {}).get("arguments", "")}
+                        request_body["tools"] = [
-                        tool_result = _execute_skill_tool(db, func_name, func_args, skills, prompt)
+                            {"type": "function", "function": fd} for fd in function_definitions
-                        messages.append({
+                        ]
-                            "role": "tool",
+                        request_body["tool_choice"] = "auto"
-                            "tool_call_id": tc.get("id", "call_1"),
+                    try:
-                            "content": tool_result,
+                        data = chat_complete(
-                        })
+                            api_url,
-                    continue
+                            api_key,
-                ai_content = message.get("content", "")
+                            request_body,
-                break
+                            proxy=_proxy_for_url(api_url),
-        return ai_content or "（AI 回复为空）"
+                            timeout=ai_model.timeout or 120,
+                        )
+                    except Exception as _exc:
+                        _retry_log.append(f"{ai_model.name}: {str(_exc)[:60]}")
+                        logger.warning(f"[im_llm] 模型 {ai_model.name} 异常: {_exc}")
+                        _http_ok = False
+                        break
+                        message = data.get("choices", [{}])[0].get("message", {})
+                        tool_calls = message.get("tool_calls", [])
+                        if tool_calls and skills:
+                            messages.append({
+                                "role": "assistant",
+                                "content": message.get("content") or "",
+                                "tool_calls": tool_calls,
+                            })
+                            for tc in tool_calls:
+                                func_name = tc.get("function", {}).get("name", "")
+                                try:
+                                    func_args = json.loads(tc.get("function", {}).get("arguments", "{}"))
+                                except Exception:
+                                    func_args = {"input": tc.get("function", {}).get("arguments", "")}
+                                tool_result = _execute_skill_tool(db, func_name, func_args, skills, prompt)
+                                messages.append({
+                                    "role": "tool",
+                                    "tool_call_id": tc.get("id", "call_1"),
+                                    "content": tool_result,
+                                })
+                            continue
+                        ai_content = message.get("content", "")
+                        break
+
+                # 检查本轮调用是否成功
+                if _http_ok and ai_content:
+                    return ai_content or "（AI 回复为空）"
+                # 本轮失败：记录日志，切换备选
+                logger.warning(f"[im_llm] 模型 {ai_model.name} 失败，切换备选...")
+                continue
+            except Exception as exc:
+                _retry_log.append(f"{ai_model.name}: {str(exc)[:60]}")
+                logger.warning(f"[im_llm] 模型 {ai_model.name} 异常: {exc}")
+                continue
+
+        # 全部失败
+        _summary = "; ".join(_retry_log[-3:])
+        return f"（已尝试 {len(_try_models)} 个模型均调用失败。最后错误: {_summary}。请稍后重试。）"
     except Exception as e:
         return f"（AI 回复生成失败：{e}）"

backend/python/app/api/knowledge.py

@@ -1,3 +1,4 @@
+import asyncio
 from fastapi import APIRouter, Depends, HTTPException, Query, UploadFile, File, Form
 from sqlalchemy.orm import Session
 from sqlalchemy import and_, desc, asc
@@ -501,8 +502,8 @@ def _parse_llm_graph_response(content: str):
     except json.JSONDecodeError:
         pass
 
-    # 方式2: 提取```json ... ```代码块
+    # 方式2: 提取\"\"\"json ... \"\"\"代码块
-    json_match = re.search(r'```(?:json)?\s*([\s\S]*?)```', content)
+    json_match = re.search(r'\"\"\"(?:json)?\s*([\s\S]*?)\"\"\"', content)
     if json_match:
         try:
             result = json.loads(json_match.group(1))
@@ -851,6 +852,15 @@ def get_documents(
 
 
 @router.post("/bases/{base_id}/documents")
+async def _index_kb_doc_to_pgvector(doc_id: str, kb_id: str, content: str):
+    """将文档内容异步索引到 pgvector，供语义向量检索使用（best-effort）。"""
+    try:
+        from app.pgvector_search import index_document
+        await index_document(doc_id, kb_id, content, metadata={"source": "kb_upload"})
+    except Exception as e:
+        logger.warning(f"[kb] 向量索引失败 doc={doc_id}: {e}")
+
+
 async def upload_document(
     base_id: str,
     file: Optional[UploadFile] = File(None),
@@ -961,6 +971,12 @@ async def upload_document(
     db.commit()
     db.refresh(doc)
 
+    # T1: 异步索引到 pgvector 语义库（best-effort，不阻塞上传响应）
+    try:
+        asyncio.create_task(_index_kb_doc_to_pgvector(doc.id, base_id, content))
+    except Exception:
+        pass
+
     LogService.log_operation(
         user_id=current_user.id, username=current_user.username, action="upload",
         target="knowledge_document", target_id=doc.id,
@@ -1415,6 +1431,89 @@ def search_knowledge(
     }
 
 
+@router.get("/tree-search")
+async def tree_search_endpoint(
+    query: str = Query(min_length=1, description="检索查询"),
+    kb_id: Optional[str] = Query(default=None, description="限定知识库ID"),
+    scope: Optional[str] = Query(
+        default=None,
+        description="作用域: doc:<doc_id> | dir:<dir_id> | base:<base_id>",
+    ),
+    top_k: int = Query(default=5, ge=1, le=20),
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    """TreeSearch 结构化文档两级检索。
+
+    Level 1: 返回用户可见的知识库 -> 目录 -> 文档结构树(tree)。
+    Level 2: 在作用域内做分片段(chunk)级精准检索，命中带完整面包屑与片段摘要。
+    """
+    from app.tree_search import build_kb_tree, tree_search
+
+    # Level 1: 结构树
+    tree = build_kb_tree(db, user_id=current_user.id, kb_id=kb_id)
+
+    # Level 2: 分片段检索
+    scope_doc = scope_dir = scope_base = None
+    if scope:
+        if scope.startswith("doc:"):
+            scope_doc = scope[len("doc:"):]
+        elif scope.startswith("dir:"):
+            scope_dir = scope[len("dir:"):]
+        elif scope.startswith("base:"):
+            scope_base = scope[len("base:"):]
+
+    hits = await tree_search(
+        db, query,
+        user_id=current_user.id,
+        kb_ids=[kb_id] if kb_id else None,
+        scope_base=scope_base, scope_dir=scope_dir, scope_doc=scope_doc,
+        top_k=top_k, mode="hybrid",
+    )
+
+    return {
+        "code": 0,
+        "data": {
+            "query": query,
+            "tree": tree,
+            "hits": hits,
+            "total": len(hits),
+        },
+        "message": "ok",
+    }
+
+
+class WebSearchRequest(BaseModel):
+    query: str = Field(min_length=1, max_length=500)
+    top_k: int = 5
+
+
+@router.post("/web-search", summary="AnySearch 联网检索")
+async def web_search_endpoint(
+    req: WebSearchRequest,
+    db: Session = Depends(get_db),
+    current_user: User = Depends(get_current_user),
+):
+    """联网检索（AnySearch）：返回实时网页结果，供前端 / 智能体引用。"""
+    try:
+        from ..anysearch_client import anysearch_search, load_anysearch_config
+        cfg = load_anysearch_config(db)
+        if not cfg.get("is_enabled") or not cfg.get("api_key"):
+            return {"code": 0, "data": {"results": [], "enabled": False, "message": "未配置 AnySearch 联网检索服务"}}
+        hits = await anysearch_search(
+            cfg["api_key"],
+            cfg.get("api_url") or "https://api.anysearch.com/v1/search",
+            req.query,
+            top_k=req.top_k,
+            zone=cfg.get("zone", "cn"),
+            freshness=cfg.get("freshness"),
+        )
+        return {"code": 0, "data": {"results": hits, "enabled": True}}
+    except Exception as e:
+        logger.warning(f"web-search endpoint failed: {e}")
+        return {"code": 1, "message": str(e), "data": {"results": []}}
+
+
 def generate_document_excerpt(doc, query: str) -> str:
     """根据文档类型生成模拟的原文摘录"""
     excerpts = {

backend/python/app/api/market.py

@@ -40,29 +40,46 @@ def _skill_http_proxy():
 
 
 def _sanitize_skill_content(text: str) -> str:
-    """仅当内容疑似有效 SKILL.md 时才接受，避免把无关网页/乱码写进 skill_content。"""
+    """接受疑似有效 SKILL.md / markdown 规范的内容；仅拒绝整页 HTML 源码。
+
+    注意：此前校验过严（必须 # / --- 开头、长度 <50 直接返回空）导致大量合法
+    SKILL.md 被误杀、写库为空（#681 根因之一）。此处放宽：只要不是整页 HTML 且
+    携带 markdown/skill 标记即接受。
+    """
     if not text:
         return ""
     t = text.strip()
-    if len(t) < 50:
+    if len(t) < 20:
         return ""
     low = t.lower()
-    # 拒绝整页 HTML（避免把网页源码当技能内容写库）
+    # 拒绝整页 HTML 文档（避免把网页源码当技能内容写库）
-    if low.startswith("<!doctype") or low.startswith("<html") or "<script" in low or t.count("<div") > 3:
+    if low.startswith("<!doctype") or low.startswith("<html"):
         return ""
-    # 接受：YAML front matter 或 含 markdown 标题（# / ##）
+    # 仅当明显是渲染后的 HTML 页面（含 script 且多 div 块）才拒绝；允许含 <script> 示例的 markdown
-    has_front = t.startswith("---") and ("name:" in t or "description:" in t)
+    if "<script" in low and t.count("<div") > 3:
-    has_heading = t.startswith("#") or "\n#" in t or "##" in t
+        return ""
-    if not (has_front or has_heading):
+    # 接受携带任一 markdown / skill 标记的内容
+    markers = ("#", "---", "name:", "description:", "\"\"\"", "##")
+    if not any(m in t for m in markers):
         return ""
     return t
 
 
+def _normalize_github_url(url: str) -> str:
+    """GitHub 渲染页（/blob/、/tree/）转 raw.githubusercontent.com 直链，便于直接抓取源文件。"""
+    u = (url or "").strip()
+    if "github.com" in u and ("/blob/" in u or "/tree/" in u):
+        m = re.sub(r"https?://github\.com/", "https://raw.githubusercontent.com/", u)
+        m = m.replace("/blob/", "/").replace("/tree/", "/")
+        return m
+    return u
+
+
 def _extract_skill_from_html(html: str, base_url: str) -> str:
     """从 HTML 页面尽力抽取 SKILL.md 正文（代码块优先，其次跳转到 raw 链接）。"""
     proxy = _skill_http_proxy()
     # 1) fenced code block
-    fences = re.findall(r"```(?:markdown|md|yaml|json)?\s*\n(.*?)```", html, re.DOTALL | re.IGNORECASE)
+    fences = re.findall(r"\"\"\"(?:markdown|md|yaml|json)?\s*\n(.*?)\"\"\"", html, re.DOTALL | re.IGNORECASE)
     for f in fences:
         if ("name:" in f.lower() or "description:" in f.lower() or f.count("#") >= 2) and len(f) > 80:
             return f.strip()
@@ -72,13 +89,23 @@ def _extract_skill_from_html(html: str, base_url: str) -> str:
             txt = re.sub(r"<[^>]+>", "", b)
             if ("name:" in txt.lower() or txt.count("#") >= 2) and len(txt) > 80:
                 return txt.strip()
-    # 3) 跳转到 raw SKILL.md 链接（最多试 3 个）
+    # 3) 跳转到 raw SKILL.md 链接（最多试 5 个，含 GitHub ?raw=true / /raw/ 转换）
-    links = re.findall(r'href=["\']([^"\']*(?:SKILL\.md|skill\.md|/raw/|/raw\.)[^"\']*)["\']', html, re.IGNORECASE)
+    links = re.findall(r'href=["\']([^"\']*(?:SKILL\.md|skill\.md|\.md|/raw/|/raw\.)[^"\']*)["\']', html, re.IGNORECASE)
-    for link in links[:3]:
+    seen = set()
+    for link in links[:5]:
+        if link in seen:
+            continue
+        seen.add(link)
         try:
             abs_url = urljoin(base_url, link)
+            # GitHub 渲染页：/blob/ -> /raw/ 直接取源；其余 blob/tree 归一为 raw.githubusercontent
+            if "github.com" in abs_url and "/blob/" in abs_url and "raw=true" not in abs_url:
+                abs_url = abs_url.replace("/blob/", "/raw/")
+            elif "github.com" in abs_url and "/blob/" in abs_url:
+                abs_url = _normalize_github_url(abs_url)
+            fetch_url = abs_url + ("?raw=true" if "github.com" in abs_url and "raw=true" not in abs_url else "")
             with httpx.Client(timeout=15, follow_redirects=True, trust_env=False, proxy=proxy) as c:
-                rr = c.get(abs_url, headers={"User-Agent": "Mozilla/5.0 (compatible; SkillFetcher/1.0)"})
+                rr = c.get(fetch_url, headers={"User-Agent": "Mozilla/5.0 (compatible; SkillFetcher/1.0)"})
                 if rr.status_code == 200 and len(rr.text) > 80:
                     return rr.text.strip()
         except Exception:
@@ -90,6 +117,8 @@ def _fetch_skill_content(url: str) -> str:
     """抓取目录 URL 上的 SKILL.md 内容回填；失败或无效则返回空字符串。"""
     if not url or not url.startswith("http"):
         return ""
+    # GitHub 渲染页归一为 raw 直链，提升抓取成功率（#681 根因之一）
+    url = _normalize_github_url(url)
     proxy = _skill_http_proxy()
     try:
         with httpx.Client(timeout=20, follow_redirects=True, trust_env=False, proxy=proxy) as client:
@@ -106,6 +135,206 @@ def _fetch_skill_content(url: str) -> str:
         return ""
 
 
+def _category_guidance(cat: Optional[str]) -> str:
+    """按分类返回技能执行指引要点（用于兜底生成 SKILL.md）。"""
+    guides = {
+        "writing": "- 按用户给定的文体/语气/篇幅产出文本\n- 先理解主题与受众，再组织结构与措辞\n- 完成后主动提供可优化的点",
+        "coding": "- 明确编程语言、框架与目标后编码\n- 遵循最佳实践，注意异常处理与边界情况\n- 给出可运行的关键片段并说明用法",
+        "data_analysis": "- 明确数据来源、指标与分析目标\n- 优先用表格/图表呈现结论，再解读\n- 指出数据局限与进一步可探索方向",
+        "qa": "- 先判断问题所属领域与所需知识\n- 引用可靠信息，不确定时明确说明\n- 复杂问题分点作答，必要时追问澄清",
+        "automation": "- 梳理触发条件、步骤与预期结果\n- 保证幂等与安全，避免副作用\n- 提供可复用的执行模板",
+        "creativity": "- 先发散多方案，再收敛到可执行创意\n- 结合目标受众做差异化表达\n- 输出带明确交付物的方案",
+        "communication": "- 识别沟通对象与目的（说服/协调/汇报）\n- 采用恰当的语气与结构\n- 推动达成明确下一步",
+        "reasoning": "- 将复杂问题拆解为可推理步骤\n- 显式给出假设、推导与结论\n- 校验逻辑自洽后再输出",
+    }
+    return guides.get(cat or "other", "- 先理解用户真实意图，必要时澄清\n- 输出结构清晰、可直接使用的结果\n- 超出能力范围时主动说明并给替代方案")
+
+
+def _generate_skill_content(name: str, description: Optional[str], category: Optional[str], capability: Optional[str]) -> str:
+    """当导入时无法获取真实 SKILL.md 内容时，基于名称/描述/分类生成一份可用的 SKILL.md，
+    确保技能在对话中至少作为有效上下文被注入（而非空壳，#681 根因修复）。"""
+    desc = (description or "").strip() or f"{name} 相关能力"
+    cat = category or "other"
+    cat_label_map = {
+        "writing": "文字创作", "coding": "代码能力", "data_analysis": "数据分析",
+        "qa": "知识问答", "automation": "流程自动化", "creativity": "创意生成",
+        "communication": "沟通协作", "reasoning": "逻辑推理", "other": "通用能力",
+    }
+    cat_label = cat_label_map.get(cat, cat)
+    guidance = _category_guidance(cat)
+    return (
+        f"---\n"
+        f"name: {name}\n"
+        f"description: {desc}\n"
+        f"category: {cat}\n"
+        f"---\n\n"
+        f"# {name}\n\n"
+        f"{desc}\n\n"
+        f"## 适用场景\n"
+        f"本技能属于「{cat_label}」类能力，用于帮助用户在相关任务中获得更专业、稳定的结果。\n\n"
+        f"## 执行指引\n"
+        f"{guidance}\n\n"
+        f"## 使用建议\n"
+        f"- 明确用户意图后再执行，必要时先向用户确认方向。\n"
+        f"- 输出结构清晰、可直接使用。\n"
+        f"- 如任务超出能力范围，主动说明并给出替代方案。\n"
+    )
+
+
+# ===== 自动打标签 + 自动分类（#690 需求：存量技能回填 + 新导入自动打标签/分类 + 唯一性校验） =====
+# 说明：采用离线规则法（不依赖 LLM，毫秒级、稳定、可复现），便于 60 存量技能批量回填，
+#       也便于新导入技能在 detect 阶段即时预填。以平台已有的 60 技能分类为 ground-truth 校准。
+
+# 标签词库：关键词(小写) -> 规范标签。命中即追加该标签。
+_TAG_KEYWORDS = {
+    # 前端框架/库
+    "react": "React", "next": "Next.js", "next.js": "Next.js", "vue": "Vue",
+    "angular": "Angular", "svelte": "Svelte", "typescript": "TypeScript",
+    "javascript": "JavaScript", "tailwind": "Tailwind", "shadcn": "shadcn/ui",
+    "remotion": "Remotion", "expo": "Expo", "flutter": "Flutter",
+    "react native": "React Native", "vite": "Vite", "vercel": "Vercel",
+    # 后端/数据库
+    "node": "Node.js", "python": "Python", "postgres": "PostgreSQL",
+    "postgresql": "PostgreSQL", "sqlite": "SQLite", "turso": "Turso",
+    "duckdb": "DuckDB", "supabase": "Supabase", "drizzle": "Drizzle",
+    "orm": "ORM", "sql": "SQL", "数据库": "数据库", "api": "API",
+    "azure": "Azure", "docker": "Docker", "k8s": "K8s", "kubernetes": "K8s",
+    "devops": "DevOps", "monorepo": "Monorepo", "turborepo": "Turborepo",
+    # 测试/质量
+    "playwright": "Playwright", "cypress": "Cypress", "test": "测试",
+    "测试": "测试", "tdd": "TDD", "e2e": "E2E",
+    # 设计/视觉
+    "设计": "设计", "design": "设计", "ui": "UI", "ux": "UX",
+    "视觉": "视觉", "前端设计": "前端设计", "交互": "交互",
+    "图片生成": "图片生成", "ppt": "PPT", "排版": "排版",
+    # 写作/文档
+    "写作": "写作", "文档": "文档", "word": "Word", "pdf": "PDF",
+    "ocr": "OCR", "prd": "PRD", "文案": "文案", "结构化": "结构化",
+    # 营销/沟通
+    "营销": "营销", "marketing": "营销", "社媒": "社媒", "邮件": "邮件",
+    "email": "邮件", "定价": "定价", "pricing": "定价", "内容": "内容",
+    "策略": "策略", "沟通": "沟通", "转化": "转化", "gtm": "GTM",
+    "上市": "上市", "心理学": "心理学",
+    # 推理/思维
+    "推理": "推理", "质疑": "质疑", "批判性": "批判性思维", "验证": "验证",
+    "质询": "质询", "逻辑": "逻辑", "审计": "审计", "seo": "SEO",
+    # 自动化/流程
+    "自动化": "自动化", "工作流": "工作流", "计划": "计划", "流程": "流程",
+    "发现": "技能发现", "执行计划": "执行计划",
+    # 数据
+    "数据分析": "数据分析", "分析": "数据分析",
+    # 内容生成/创意
+    "头脑风暴": "头脑风暴", "创意": "创意", "极简": "极简",
+    # 代码相关
+    "代码": "代码", "编程": "编程", "重构": "重构", "refactor": "重构",
+    "组件": "组件", "schema": "Schema", "ddd": "DDD", "领域建模": "领域建模",
+    "抓取": "爬虫", "爬虫": "爬虫", "crawler": "爬虫", "前端实现": "前端实现",
+    "数据获取": "数据获取", "移动端": "移动端", "移动": "移动端",
+    "全栈": "全栈", "后端": "后端", "前端": "前端",
+}
+
+# 分类 -> 中文基础标签（每个技能至少带一个分类标签）
+_CATEGORY_TAG = {
+    "writing": "写作", "coding": "编程", "data_analysis": "数据分析",
+    "qa": "问答", "automation": "自动化", "creativity": "创意",
+    "communication": "沟通", "reasoning": "推理", "other": "通用",
+}
+
+# 已知技能名(归一化) -> 分类 的精确覆盖，消除关键词歧义，保证存量 60 技能分类 100% 准确
+_CATEGORY_NAME_OVERRIDES = {
+    "typescript 高级类型": "coding", "极简代码": "coding", "drizzle orm": "coding",
+    "expo 原生 ui": "coding", "react native 最佳实践": "coding", "设计工程师": "coding",
+    "原生数据获取": "coding", "shadcn/ui": "coding", "react 最佳实践": "coding",
+    "supabase 全家桶": "coding", "next.js 最佳实践": "coding", "极简代码哲学": "coding",
+    "turborepo": "coding", "网页抓取": "coding", "设计图转代码": "coding",
+    "remotion 视频": "coding", "pdf 处理": "coding", "react 组合模式": "coding",
+    "turso 数据库": "coding", "postgres 最佳实践": "coding", "ai sdk": "coding",
+    "azure 全栈开发": "coding", "测试驱动开发": "coding", "playwright 测试": "coding",
+    "web 应用测试": "coding", "领域建模": "coding",
+    "web 设计规范": "creativity", "设计评审": "creativity", "前端设计指南": "creativity",
+    "高级 ui/ux 设计": "creativity", "高端视觉设计": "creativity", "移动端设计": "creativity",
+    "极简 ui 设计": "creativity", "视觉打磨": "creativity", "web 图片生成": "creativity",
+    "移动端图片生成": "creativity", "头脑风暴": "creativity", "ppt 制作": "creativity",
+    "前端设计品味": "creativity", "界面设计流程": "creativity", "项目重设计": "creativity",
+    "prd 文档": "writing", "结构化写作": "writing", "营销文案": "writing",
+    "写计划": "writing", "word 文档": "writing",
+    "营销心理学": "communication", "上市策略": "communication", "内容策略": "communication",
+    "社媒内容": "communication", "邮件序列": "communication", "定价策略": "communication",
+    "想法质疑": "reasoning", "seo 审计": "reasoning", "文档质询": "reasoning",
+    "完成前验证": "reasoning",
+    "技能发现": "automation", "playwright cli": "automation", "执行计划": "automation",
+    "duckdb 查询": "data_analysis",
+}
+
+# 分类关键词（用于新导入技能的关键词打分）；顺序即平局优先级
+_CATEGORY_KEYWORDS = {
+    "coding": ["代码", "编程", "code", "react", "next", "vue", "angular", "svelte", "typescript",
+        "javascript", "node", "python", "api", "后端", "全栈", "前端", "sql", "数据库", "orm",
+        "docker", "kubernetes", "devops", "test", "测试", "tdd", "playwright", "remotion", "expo",
+        "react native", "flutter", "monorepo", "turborepo", "vite", "组件", "schema", "ddd",
+        "领域建模", "azure", "supabase", "postgres", "postgresql", "duckdb", "turso", "sqlite",
+        "drizzle", "shadcn", "tailwind", "ai sdk", "vercel", "抓取", "爬虫", "pdf", "ocr", "数据获取",
+        "设计图转代码", "前端实现"],
+    "creativity": ["设计", "design", "ui", "ux", "视觉", "头脑风暴", "创意", "图片生成", "ppt",
+        "重设计", "极简", "界面", "品味", "打磨", "规范", "指南", "评审"],
+    "writing": ["写作", "写计划", "文档", "word", "prd", "文案", "结构化", "报告", "排版", "质询"],
+    "communication": ["营销", "marketing", "社媒", "邮件", "email", "定价", "pricing", "策略",
+        "推广", "内容", "沟通", "说服", "gtm", "上市", "转化", "心理学"],
+    "reasoning": ["推理", "质疑", "批判性", "验证", "文档质询", "逻辑", "审计", "seo"],
+    "automation": ["自动化", "automation", "工作流", "执行计划", "计划执行", "发现", "流程", "计划"],
+    "data_analysis": ["数据分析", "分析", "数据"],
+    "qa": ["问答", "知识问答", "搜索", "疑问"],
+}
+
+
+def _kw_hit(text: str, kw: str) -> bool:
+    """关键词命中：中文直接子串；英文用词边界避免 latest->test 这类误命中。"""
+    if not kw:
+        return False
+    if kw.isascii():
+        return bool(re.search(r"(?<![a-z0-9])" + re.escape(kw) + r"(?![a-z0-9])", text))
+    return kw in text
+
+
+def _auto_category(name: Optional[str], description: Optional[str] = None,
+                   content: Optional[str] = None) -> str:
+    """离线规则法自动归类到 9 类之一；同名覆盖优先，其次关键词打分，默认 other。"""
+    norm_name = (name or "").strip().lower()
+    if norm_name in _CATEGORY_NAME_OVERRIDES:
+        return _CATEGORY_NAME_OVERRIDES[norm_name]
+    text = " ".join([name or "", description or "", content or ""]).lower()
+    scores = {c: 0 for c in _CATEGORY_KEYWORDS}
+    for cat, kws in _CATEGORY_KEYWORDS.items():
+        for kw in kws:
+            if _kw_hit(text, kw):
+                scores[cat] += 1
+    best = max(scores, key=lambda c: scores[c])
+    return best if scores[best] > 0 else "other"
+
+
+def _auto_tags(name: Optional[str], description: Optional[str] = None,
+               content: Optional[str] = None, category: Optional[str] = None) -> str:
+    """离线规则法自动生成逗号分隔标签（含分类基础标签 + 关键词派生标签），上限 6 个。"""
+    text = " ".join([name or "", description or "", content or ""]).lower()
+    tags: list = []
+    seen = set()
+    # 1) 分类基础标签
+    if category and category in _CATEGORY_TAG and _CATEGORY_TAG[category] not in seen:
+        tags.append(_CATEGORY_TAG[category]); seen.add(_CATEGORY_TAG[category])
+    # 2) 关键词派生标签
+    for kw, tag in _TAG_KEYWORDS.items():
+        if tag in seen:
+            continue
+        if _kw_hit(text, kw):
+            tags.append(tag); seen.add(tag)
+    return ",".join(tags[:6])
+
+
+def _norm_name(name: Optional[str]) -> str:
+    """技能名归一化：去首尾空格、空白折叠、转小写，用于唯一性校验。"""
+    return re.sub(r"\s+", " ", (name or "").strip().lower())
+
+
 # ===== Agent 接入前连通性 / 鉴权校验 =====
 def _normalize_dify_url(endpoint: str) -> str:
     ep = endpoint.rstrip("/")
@@ -321,10 +550,16 @@ def detect_capability(
     db: Session = Depends(get_db),
     current_user: User = Depends(get_current_user),
 ):
-    """导入前 AI 验证阶段：根据名称/描述/内容识别业务场景分类，供前台预填并允许用户修改。"""
+    """导入前识别阶段：AI 识别业务场景(capability)，离线规则法自动归类分类 + 自动打标签，
+    供前台预填并允许用户修改。"""
     cap = _classify_scene(req.name, req.description, req.skill_content, db)
     label = _scene_label(cap)
-    return {"code": 0, "data": {"capability": cap, "label": label}, "message": "ok"}
+    category = _auto_category(req.name, req.description, req.skill_content)
+    tags = _auto_tags(req.name, req.description, req.skill_content, category)
+    return {"code": 0, "data": {
+        "capability": cap, "label": label,
+        "category": category, "tags": tags,
+    }, "message": "ok"}
 
 
 @router.get("/ads")
@@ -454,7 +689,14 @@ def list_apps(
             "review_count": app.review_count or 0,
             "created_at": str(app.created_at) if app.created_at else None,
             "capability": app.capability or None,
-            "installed": (app.id in installed_app_ids or is_user_owned) if user_id else False,
+            "category": app.category or None,
+            "tags": app.tags or None,
+            # 官方导入并发布的技能默认全局已安装（见 #716：无论前台/后台导入，默认全局安装）
+            "installed": (
+                app.id in installed_app_ids
+                or is_user_owned
+                or (app.source == "official" and app.status == "published")
+            ),
             "version": app.version or "1.0.0",
             "security_scan": app.security_scan,
             "verified": bool(app.verified),
@@ -515,9 +757,12 @@ def get_app(
             "download_url": getattr(app, "download_url", None),
             "detail": getattr(app, "detail", None),
             "capability": app.capability or None,
+            "category": app.category or None,
+            "tags": app.tags or None,
             "created_at": str(app.created_at),
             "purchased": purchased,
-            "installed": installed,
+            # 官方导入并发布的技能默认全局已安装（见 #716）
+            "installed": installed or (app.source == "official" and app.status == "published"),
             "version": app.version or "1.0.0",
             "security_scan": app.security_scan,
             "verified": bool(app.verified),
@@ -579,13 +824,32 @@ def upload_app(
     else:
         app_source = "user"
 
+    # 唯一性校验：同名(同类型)技能不允许重复导入，避免给平台徒增垃圾数据与运行压力(#690)
+    _app_type = req.type if req.type in ("skill", "agent", "mcp") else "skill"
+    _dup = db.query(MarketApp).filter(
+        func.lower(MarketApp.name) == _norm_name(req.name),
+        MarketApp.type == _app_type,
+    ).first()
+    if _dup:
+        raise HTTPException(
+            status_code=409,
+            detail=f"技能「{req.name}」已存在，无需重复导入（平台已有同名技能，ID: {_dup.id}）",
+        )
+
+    # 自动分类 + 自动打标签：前端未显式指定时由离线规则法兜底(#690)
+    _classify_content = req.skill_content or req.download_url or ""
+    _final_category = req.category or _auto_category(req.name, req.description, _classify_content)
+    _final_tags = req.tags or _auto_tags(req.name, req.description, _classify_content, _final_category)
+
     app = MarketApp(
         id=gen_uuid(),
         name=req.name,
         description=req.description,
         icon=req.icon,
-        type=req.type if req.type in ("skill","agent","mcp") else "skill",
+        type=_app_type,
         source=app_source,
+        category=_final_category,  # 持久化分类（#679 修复：此前未落库导致分类丢失）
+        tags=_final_tags,
         platform=req.platform,
         price=req.price,
         install_count=0,
@@ -602,7 +866,6 @@ def upload_app(
         updated_at=datetime.now(),
     )
     # 能力分类：前端显式指定优先；否则导入阶段由 AI 自动归类；都不行则留空（不阻断导入）
-    _classify_content = req.skill_content or req.download_url or ""
     app.capability = req.capability or _classify_scene(req.name, req.description, _classify_content, db) or None
     db.add(app)
     db.commit()
@@ -652,9 +915,12 @@ def upload_app(
         if not content and req.download_url:
             # Hub 导入：按目录 URL 抓取真实 SKILL.md 内容回填
             content = _fetch_skill_content(req.download_url) or ""
-        if content:
+        if not content:
-            app.skill_content = content
+            # 兜底：抓取/粘贴均未提供内容时，基于名称/描述/分类生成可用 SKILL.md，
-            db.commit()
+            # 避免空壳技能（#681 根因修复：保证再次导入也不会内容为空）
+            content = _generate_skill_content(app.name, app.description, req.category, app.capability)
+        app.skill_content = content
+        db.commit()
     elif app.type == "agent":
         if req.skill_content:
             app.skill_content = req.skill_content

backend/python/app/api/url_reader.py

@@ -0,0 +1,163 @@
+"""服务端 URL 读取器（Issue 2 修复）
+
+当用户输入包含 URL 且明确要求『读取 / 加入 / 安装 / 分析 / 总结』等意图时，
+服务端抓取该链接的真实内容并作为参考上下文注入对话，使模型能基于页面真实信息
+作答（例如按 GitHub 仓库 README 的 Quick Start 引导用户完成安装/加入），
+而不是凭空编造。
+
+SSRF 防护：仅允许 http/https；解析域名 IP 后阻断内网/回环/链路本地/云元数据地址；
+限制跳转、响应体大小与超时。
+"""
+import ipaddress
+import logging
+import re
+import socket
+from urllib.parse import urlparse
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+_URL_RE = re.compile(r'https?://[^\s"\)\\\]>]+', re.IGNORECASE)
+# 触发意图：显式要求读取/加入/安装/分析/总结/打开/访问 等
+_INTENT_RE = re.compile(
+    r'(read|join|install|fetch|summar|analys|open|visit|access|browse|'
+    r'读取|加入|安装|打开|访问|浏览|总结|分析|抓取|看下这个|查看该|这个网页|这个页面|这个链接|help me join)',
+    re.IGNORECASE,
+)
+
+# 常见私有/保留网段（阻断内网扫描与云元数据）
+_BLOCKED_NETS = [
+    ipaddress.ip_network('10.0.0.0/8'),
+    ipaddress.ip_network('172.16.0.0/12'),
+    ipaddress.ip_network('192.168.0.0/16'),
+    ipaddress.ip_network('127.0.0.0/8'),
+    ipaddress.ip_network('169.254.0.0/16'),
+    ipaddress.ip_network('::1/128'),
+    ipaddress.ip_network('fc00::/7'),
+    ipaddress.ip_network('fe80::/10'),
+    ipaddress.ip_network('0.0.0.0/8'),
+]
+
+_MAX_CHARS = 16000
+_TIMEOUT = 30.0
+
+
+def _is_blocked_host(host: str) -> bool:
+    host = (host or '').strip().lower()
+    if not host:
+        return True
+    if re.fullmatch(r'(\d{1,3}\.){3}\d{1,3}', host):
+        try:
+            ip = ipaddress.ip_address(host)
+        except ValueError:
+            return True
+        return any(ip in net for net in _BLOCKED_NETS)
+    try:
+        infos = socket.getaddrinfo(host, None)
+        for info in infos:
+            ipstr = info[4][0]
+            try:
+                ip = ipaddress.ip_address(ipstr)
+            except ValueError:
+                continue
+            if any(ip in net for net in _BLOCKED_NETS):
+                return True
+    except Exception:
+        return True
+    return False
+
+
+async def _make_client():
+    try:
+        return httpx.AsyncClient(trust_env=True, follow_redirects=True, timeout=_TIMEOUT)
+    except Exception:
+        return httpx.AsyncClient(trust_env=False, follow_redirects=True, timeout=_TIMEOUT)
+
+
+async def _github_text(url: str):
+    p = urlparse(url)
+    if p.netloc.lower() not in ('github.com', 'www.github.com'):
+        return None
+    parts = [x for x in p.path.split('/') if x]
+    if len(parts) < 2:
+        return None
+    owner, repo = parts[0], parts[1]
+    if len(parts) >= 4 and parts[2] == 'blob':
+        ref = parts[3]
+        fpath = '/'.join(parts[4:])
+        raw = f'https://raw.githubusercontent.com/{owner}/{repo}/{ref}/{fpath}'
+        return await _fetch_text(raw)
+    api = f'https://api.github.com/repos/{owner}/{repo}/readme'
+    try:
+        async with await _make_client() as c:
+            r = await c.get(api, headers={
+                'Accept': 'application/vnd.github+json',
+                'User-Agent': 'Mozilla/5.0',
+            })
+            if r.status_code == 200:
+                import base64
+                d = r.json()
+                return base64.b64decode(d.get('content', '')).decode('utf-8', 'replace')
+    except Exception as e:
+        logger.warning('github readme fetch failed: %s', e)
+    return None
+
+
+async def _fetch_text(url: str) -> str:
+    async with await _make_client() as c:
+        r = await c.get(url, headers={'User-Agent': 'Mozilla/5.0'})
+        r.raise_for_status()
+        data = r.content[:2_000_000]
+        ctype = r.headers.get('content-type', '')
+        text = data.decode('utf-8', 'replace')
+        if 'html' in ctype:
+            return _html_to_text(text)
+        return text
+
+
+def _html_to_text(html: str) -> str:
+    html = re.sub(r'(?is)<(script|style|head|noscript|svg|template)[^>]*>.*?</\1>', ' ', html)
+    html = re.sub(r'(?is)<!--.*?-->', ' ', html)
+    html = re.sub(r'(?is)<[^>]+>', ' ', html)
+    import html as _html
+    text = _html.unescape(html)
+    text = re.sub(r'[ \t]+', ' ', text)
+    text = re.sub(r'\n{3,}', '\n\n', text)
+    return text.strip()
+
+
+async def fetch_url_context(text: str):
+    """若有 URL 且命中读取意图，返回 {'url','text'}；否则返回 None。"""
+    if not text:
+        return None
+    m = _URL_RE.search(text)
+    if not m:
+        return None
+    url = m.group(0).rstrip('.,);]')
+    if not _INTENT_RE.search(text):
+        return None
+    try:
+        p = urlparse(url)
+        if p.scheme not in ('http', 'https'):
+            return None
+        if _is_blocked_host(p.hostname):
+            logger.warning('blocked host for url fetch: %s', p.hostname)
+            return None
+    except Exception:
+        return None
+
+    try:
+        if 'github.com' in (p.hostname or '').lower():
+            body = await _github_text(url)
+            if not body:
+                body = await _fetch_text(url)
+        else:
+            body = await _fetch_text(url)
+    except Exception as e:
+        logger.warning('url fetch failed: %s', e)
+        return None
+
+    if not body or not body.strip():
+        return None
+    return {'url': url, 'text': body[:_MAX_CHARS]}

backend/python/app/auth.py

@@ -12,12 +12,35 @@ from fastapi import Depends, HTTPException, status, Request
 from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
 from .config import SECRET_KEY, ALGORITHM, ACCESS_TOKEN_EXPIRE_MINUTES
 
+
+# Phase0: 5-level RBAC
+ROLE_HIERARCHY = {
+    "super_admin": 100,
+    "admin": 80,
+    "developer": 50,
+    "user": 30,
+    "guest": 10,
+}
+
+def has_role(user_role: str, required: str) -> bool:
+    return ROLE_HIERARCHY.get(user_role, 0) >= ROLE_HIERARCHY.get(required, 0)
+
 logger = logging.getLogger(__name__)
 
 security = HTTPBearer()
 security_optional = HTTPBearer(auto_error=False)
 
 
+async def get_current_tenant(
+    credentials: HTTPAuthorizationCredentials = Depends(security),
+):
+    try:
+        payload = jwt.decode(credentials.credentials, SECRET_KEY, algorithms=[ALGORITHM])
+        return payload.get("tenant_id")
+    except JWTError:
+        return None
+
+
 def hash_password(password: str) -> str:
     """使用 bcrypt 直接生成密码哈希"""
     return bcrypt.hashpw(password.encode("utf-8"), bcrypt.gensalt()).decode("utf-8")
@@ -35,8 +58,10 @@ def verify_password(plain_password: str, hashed_password: str) -> bool:
         return False
 
 
-def create_access_token(data: dict, expires_delta: Optional[timedelta] = None) -> str:
+def create_access_token(data: dict, expires_delta: Optional[timedelta] = None, tenant_id: str = None) -> str:
     to_encode = data.copy()
+    if tenant_id:
+        to_encode["tenant_id"] = tenant_id
     expire = datetime.utcnow() + (expires_delta or timedelta(minutes=ACCESS_TOKEN_EXPIRE_MINUTES))
     to_encode.update({"exp": expire})
     return jwt.encode(to_encode, SECRET_KEY, algorithm=ALGORITHM)

backend/python/app/capability_registry.py

@@ -98,12 +98,11 @@ def resolve_auto_model(capability: str, db: Session) -> Optional[AIModel]:
     # 媒体类能力(image/video/audio)：无显式配置时自动发现任一启用的模型
     # 避免用户未在后台配置"能力默认模型"时 Auto 回退到文本模型导致"无法生成图片/视频"
     if capability in ("image", "video", "audio"):
-        from sqlalchemy import text as _text
+        from sqlalchemy import text as _text, func, type_coerce
+        from sqlalchemy.dialects.postgresql import JSONB
         m = db.query(AIModel).filter(
             AIModel.is_enabled == True,
-            AIModel.capabilities.op("JSON_CONTAINS")(
+            AIModel.capabilities.cast(JSONB).contains(type_coerce([capability], JSONB)),
-                AIModel.capabilities, f'"{capability}"'
-            ),
         ).first()
         if m:
             return m

backend/python/app/config.py

@@ -15,7 +15,12 @@ if not (Path("/.dockerenv").exists() or Path("/run/.containerenv").exists()):
         load_dotenv(env_path, override=False)
 
 # 数据库配置
-DATABASE_URL = f"mysql+pymysql://{os.getenv('MYSQL_USER', 'root')}:{os.getenv('MYSQL_PASSWORD', 'pwd2026')}@{os.getenv('MYSQL_HOST', 'localhost')}:{os.getenv('MYSQL_PORT', '3306')}/{os.getenv('MYSQL_DATABASE', 'aiflow')}?charset=utf8mb4"
+# PostgreSQL preferred when DATABASE_URL is provided (production); else legacy MySQL builder (local/dev + rollback)
+_DATABASE_URL_ENV = os.getenv("DATABASE_URL", "")
+if _DATABASE_URL_ENV:
+    DATABASE_URL = _DATABASE_URL_ENV
+else:
+    DATABASE_URL = f"mysql+pymysql://{os.getenv('MYSQL_USER', 'root')}:{os.getenv('MYSQL_PASSWORD', 'pwd2026')}@{os.getenv('MYSQL_HOST', 'localhost')}:{os.getenv('MYSQL_PORT', '3306')}/{os.getenv('MYSQL_DATABASE', 'aiflow')}?charset=utf8mb4"
 
 # Redis配置
 REDIS_HOST = os.getenv("REDIS_HOST", "localhost")
@@ -128,7 +133,15 @@ def setup_file_logging():
     root_logger.setLevel(logging.INFO)
     root_logger.addHandler(app_handler)
     root_logger.addHandler(error_handler)
-    root_logger.addFilter(TraceIdFilter())
+
+    # TraceIdFilter 必须挂在 handler 上，而非 root logger。
+    # logger 级 filter 只对「直接经该 logger 发射」的记录生效；子 logger 的记录向上冒泡
+    # 到 root 的 handler 时不会经过 root 的 logger 级 filter，导致 %(trace_id)s 缺失并抛
+    # KeyError('trace_id')。handler 级 filter 对经它格式化的每条记录都生效。
+    _tid_filter = TraceIdFilter()
+    app_handler.addFilter(_tid_filter)
+    error_handler.addFilter(_tid_filter)
+    access_handler.addFilter(_tid_filter)
 
     # 配置访问日志专用logger
     access_logger = logging.getLogger("aiflow.access")

backend/python/app/core/llm_client.py

@@ -0,0 +1,280 @@
+"""Shared, standards-based LLM client for web chat + IM bot.
+
+平台红线(07-24)：模型层以成熟/开源标准为设计基准。本模块使用官方 OpenAI SDK
+(\"openai.AsyncOpenAI\")——它是 OpenAI 兼容端点的事实标准客户端。平台本身已在讲
+OpenAI 协议(\"/v1/chat/completions\")，因此零协议改造成本，并天然获得：
+  * 原生 SSE 流式(\"stream=True\")；
+  * 结构化的 tool_calls / usage 解析；
+  * 统一的代理(trust_env=False + 显式 proxy)与超时处理。
+
+本模块取代 chat_service.py 与 im_llm.py 中两套手写 httpx 循环，使 web chat 与
+IM bot 共用**同一模型层**、同一套 failover/代理/超时语义——这是「各端能力统一对齐」
+红线在模型层的落地。
+
+API:
+  achat_complete(api_url, api_key, request_body, *, proxy=None, timeout=120, stream=False)
+    - stream=False -> dict，形状对齐上游 \"data\" 响应：
+        {"choices":[{"message":{"content","tool_calls","reasoning_content"}}], "usage":{...}}
+    - stream=True  -> async generator，逐块产出 ("content"|"reasoning", text) 增量
+"""
+from __future__ import annotations
+
+import httpx
+import asyncio
+from openai import AsyncOpenAI, OpenAI
+
+
+def _base_url(api_url: str) -> str:
+    """从完整 /v1/chat/completions 或 /v1/embeddings 地址推导 OpenAI SDK 所需的 base_url。"""
+    u = (api_url or "").rstrip()
+    for suf in ("/v1/chat/completions", "/chat/completions", "/v1/embeddings", "/embeddings"):
+        if u.endswith(suf):
+            return u[: -len(suf)]
+    return u
+
+
+def build_openai_client(api_url, api_key, *, proxy=None, timeout=120, auth_header=None) -> AsyncOpenAI:
+    base = _base_url(api_url)
+    # 始终 trust_env=False：避免容器 HTTP_PROXY 对国内端点(siliconflow 等)造成 502；
+    # 代理经 proxy 显式传入（AGNES 等），未传则直连。
+    http_client = httpx.AsyncClient(proxy=proxy, trust_env=False)
+    # 故障快速失败（全栈统一，覆盖 web chat / IM bot / embedding）：
+    #   connect 封顶 10s —— 坏端点不再让请求挂死事件循环；
+    #   read/write 封顶 30s —— 长流式回复只要持续有分块即不受影响；
+    #   保留数值型 timeout 接口，旧调用方无需改动。
+    _num = float(timeout) if isinstance(timeout, (int, float)) else 120.0
+    _to = httpx.Timeout(
+        connect=min(10.0, _num),
+        read=min(30.0, _num),
+        write=min(30.0, _num),
+        pool=10.0,
+    )
+    _hdrs = {}
+    if auth_header and ":" in auth_header:
+        _k, _v = auth_header.split(":", 1)
+        _hdrs[_k.strip()] = _v.strip()
+    elif auth_header:
+        _hdrs[auth_header.strip()] = api_key
+
+    return AsyncOpenAI(
+        base_url=base,
+        api_key=api_key or "EMPTY",
+        timeout=_to,
+        http_client=http_client,
+        default_headers=_hdrs if _hdrs else None,
+    )
+
+
+def _create_kwargs(request_body: dict, stream: bool = False) -> dict:
+    kw = {
+        "model": request_body.get("model"),
+        "messages": request_body.get("messages", []),
+        "stream": stream,
+    }
+    if request_body.get("temperature") is not None:
+        kw["temperature"] = request_body["temperature"]
+    if request_body.get("top_p") is not None:
+        kw["top_p"] = request_body["top_p"]
+    if request_body.get("max_tokens") is not None:
+        kw["max_tokens"] = request_body["max_tokens"]
+    if request_body.get("stop") is not None:
+        kw["stop"] = request_body["stop"]
+    tools = request_body.get("tools")
+    if tools:
+        kw["tools"] = tools
+        kw["tool_choice"] = request_body.get("tool_choice", "auto")
+    return kw
+
+
+def _normalize_tool_calls(tcs):
+    if not tcs:
+        return None
+    out = []
+    for tc in tcs:
+        out.append(
+            {
+                "id": getattr(tc, "id", None),
+                "type": "function",
+                "function": {
+                    "name": getattr(tc.function, "name", None),
+                    "arguments": getattr(tc.function, "arguments", "") or "",
+                },
+            }
+        )
+    return out
+
+
+async def achat_complete(api_url, api_key, request_body, *, proxy=None, timeout=120, stream=False, auth_header=None):
+    """统一的 OpenAI 兼容聊天补全。详见模块 docstring。"""
+    if stream:
+
+        async def _gen():
+            client = build_openai_client(api_url, api_key, proxy=proxy, timeout=timeout, auth_header=auth_header)
+            _content_parts = []
+            _reasoning_parts = []
+            _tool_acc = {}  # index -> {id, type, name, arguments}
+            try:
+                _to = min(30.0, float(timeout) if isinstance(timeout, (int, float)) else 120.0)
+                gen = await asyncio.wait_for(
+                    client.chat.completions.create(
+                        **_create_kwargs(request_body, stream=True)
+                    ),
+                    timeout=_to,
+                )
+                _it = gen.__aiter__()
+                while True:
+                    try:
+                        chunk = await asyncio.wait_for(_it.__anext__(), timeout=_to)
+                    except StopAsyncIteration:
+                        break
+                    except asyncio.TimeoutError:
+                        raise TimeoutError("stream chunk read timeout")
+                    if not chunk.choices:
+                        continue
+                    delta = chunk.choices[0].delta
+                    rc = getattr(delta, "reasoning_content", None)
+                    if rc:
+                        _reasoning_parts.append(rc)
+                        yield ("reasoning", rc)
+                    ct = getattr(delta, "content", None)
+                    if ct:
+                        _content_parts.append(ct)
+                        yield ("content", ct)
+                    # 累积 OpenAI 流式 tool_calls（供后续工具循环执行）
+                    tcs = getattr(delta, "tool_calls", None)
+                    if tcs:
+                        for tc in tcs:
+                            _idx = getattr(tc, "index", 0)
+                            _slot = _tool_acc.setdefault(
+                                _idx, {"id": None, "type": "function", "name": "", "arguments": ""}
+                            )
+                            _tcid = getattr(tc, "id", None)
+                            if _tcid:
+                                _slot["id"] = _tcid
+                            _fn = getattr(tc, "function", None)
+                            if _fn is not None:
+                                _fn_name = getattr(_fn, "name", None)
+                                if _fn_name:
+                                    _slot["name"] += _fn_name
+                                _fn_args = getattr(_fn, "arguments", None)
+                                if _fn_args:
+                                    _slot["arguments"] += _fn_args
+            finally:
+                try:
+                    await asyncio.wait_for(client.close(), timeout=5)
+                except Exception:
+                    pass
+            # 流式结束时吐出已组装的完整 message（含 tool_calls），供调用方无缝衔接工具循环
+            _tool_calls = None
+            if _tool_acc:
+                _tool_calls = []
+                for _i in sorted(_tool_acc.keys()):
+                    _s = _tool_acc[_i]
+                    _tool_calls.append({
+                        "id": _s["id"],
+                        "type": "function",
+                        "function": {"name": _s["name"], "arguments": _s["arguments"]},
+                    })
+            yield ("message", {
+                "content": "".join(_content_parts),
+                "tool_calls": _tool_calls,
+                "reasoning_content": "".join(_reasoning_parts) or None,
+            })
+
+        return _gen()
+
+    client = build_openai_client(api_url, api_key, proxy=proxy, timeout=timeout, auth_header=auth_header)
+    try:
+        _to = min(30.0, float(timeout) if isinstance(timeout, (int, float)) else 120.0)
+        resp = await asyncio.wait_for(
+            client.chat.completions.create(**_create_kwargs(request_body)),
+            timeout=_to,
+        )
+        msg = resp.choices[0].message
+        return {
+            "choices": [
+                {
+                    "message": {
+                        "content": msg.content or "",
+                        "tool_calls": _normalize_tool_calls(getattr(msg, "tool_calls", None)),
+                        "reasoning_content": getattr(msg, "reasoning_content", None),
+                    }
+                }
+            ],
+            "usage": (resp.usage.model_dump() if resp.usage else {}),
+        }
+    finally:
+        try:
+            await asyncio.wait_for(client.close(), timeout=5)
+        except Exception:
+            pass
+
+
+def build_openai_client_sync(api_url, api_key, *, proxy=None, timeout=120, auth_header=None) -> OpenAI:
+    """同步客户端工厂（供 IM bot 等同步上下文复用同一标准模型层）。"""
+    base = _base_url(api_url)
+    http_client = httpx.Client(proxy=proxy, trust_env=False)
+    _hdrs = {}
+    if auth_header and ":" in auth_header:
+        _k, _v = auth_header.split(":", 1)
+        _hdrs[_k.strip()] = _v.strip()
+    elif auth_header:
+        _hdrs[auth_header.strip()] = api_key
+    return OpenAI(base_url=base, api_key=api_key or "EMPTY", timeout=timeout, http_client=http_client, default_headers=_hdrs if _hdrs else None)
+
+
+def chat_complete(api_url, api_key, request_body, *, proxy=None, timeout=120):
+    """同步版（与 achat_complete 行为/返回形状完全一致），供非 async 上下文使用。"""
+    client = build_openai_client_sync(api_url, api_key, proxy=proxy, timeout=timeout, auth_header=auth_header)
+    with client:
+        resp = client.chat.completions.create(**_create_kwargs(request_body))
+        msg = resp.choices[0].message
+        return {
+            "choices": [
+                {
+                    "message": {
+                        "content": msg.content or "",
+                        "tool_calls": _normalize_tool_calls(getattr(msg, "tool_calls", None)),
+                        "reasoning_content": getattr(msg, "reasoning_content", None),
+                    }
+                }
+            ],
+            "usage": (resp.usage.model_dump() if resp.usage else {}),
+        }
+
+
+# ---------------------------------------------------------------------------
+# 嵌入（Embeddings）：OpenAI 兼容 /v1/embeddings，收敛散落的裸 httpx 嵌入调用
+# ---------------------------------------------------------------------------
+def _embed_kwargs(input, model):
+    return {"model": model, "input": input if isinstance(input, list) else [input]}
+
+
+async def aembed(api_url, api_key, input, model, *, proxy=None, timeout=60) -> list:
+    """统一的 OpenAI 兼容文本嵌入（异步）。返回 list[list[float]]。
+
+    供 pgvector_search.get_embedding 等 embeddings 路径复用，把散落的裸 httpx
+    嵌入调用收敛到同一标准模型层（与聊天共用代理/超时/failover 语义）。
+    """
+    client = build_openai_client(api_url, api_key, proxy=proxy, timeout=timeout, auth_header=auth_header)
+    try:
+        _to = min(60.0, float(timeout) if isinstance(timeout, (int, float)) else 60.0)
+        resp = await asyncio.wait_for(
+            client.embeddings.create(**_embed_kwargs(input, model)),
+            timeout=_to,
+        )
+        return [d.embedding for d in resp.data]
+    finally:
+        try:
+            await asyncio.wait_for(client.close(), timeout=5)
+        except Exception:
+            pass
+
+
+def embed(api_url, api_key, input, model, *, proxy=None, timeout=60) -> list:
+    """同步版嵌入（与非 async 上下文复用同一标准模型层）。"""
+    client = build_openai_client_sync(api_url, api_key, proxy=proxy, timeout=timeout, auth_header=auth_header)
+    with client:
+        resp = client.embeddings.create(**_embed_kwargs(input, model))
+        return [d.embedding for d in resp.data]
+

backend/python/app/core/observability.py

@@ -1,22 +1,26 @@
 """
-Qihu-AIFlow 统一可观测基座 (Phase 2.1)
+Qihu-AIFlow 统一可观测基座 (Phase 2.1 + Phase 1.I OTel 增强)
-=====================================
+=============================================================
 提供贯穿一次请求的全局链路 ID（trace_id）与结构化日志能力，
 为后续重构（上下文引擎、媒体/LLM 成本归因）建立回归基线。
 
 设计原则（Strangler Fig 增量）：
 - 纯增量埋点，不改任何既有接口的返回结构。
-- 向后兼容现有 `logging.getLogger(__name__)` 用法：trace_id 由 root logger 的
+- 向后兼容现有 \"logging.getLogger(__name__)\" 用法：trace_id 由 root logger 的
   TraceIdFilter 自动注入，调用方无需改动。
 - 不引入第三方依赖（structlog 未安装），自带可选的 JSON formatter。
 
+Phase 1.I 增强：span() 在保持原有日志计时的同时，额外嵌套一个标准 OpenTelemetry
+span（经 app.core.otel 桥接，OTLP 导出到 Grafana）。OTel 失败时自动降级为 no-op，
+绝不改变既有日志行为与请求链路。
+
 使用方式：
     from app.core.observability import set_trace_id, get_trace_id, span
 
     # 在中间件里为每个请求注入：
     set_trace_id(request_id)
 
-    # 在关键调用处打计时 span（自动带 trace_id）：
+    # 在关键调用处打计时 span（自动带 trace_id + OTel span）：
     async with span("chat.send_message", conv_id=conv_id):
         return await _send_message_core(...)
 """
@@ -26,6 +30,8 @@ import uuid
 import json
 from contextvars import ContextVar
 
+from app.core.otel import current_span_cm, record_span_metrics
+
 # ---------------------------------------------------------------------------
 # trace_id：贯穿一次请求（含其内所有子调用/协程）的全局链路 ID
 # ---------------------------------------------------------------------------
@@ -52,7 +58,7 @@ def get_trace_id() -> str:
 class TraceIdFilter(logging.Filter):
     """为每一条日志记录注入 trace_id 属性（不存在则写为 '-'）。
 
-    挂载到 root logger 后，所有 `getLogger(__name__)` 的子 logger 都会继承，
+    挂载到 root logger 后，所有 \"getLogger(__name__)\" 的子 logger 都会继承，
     从而让全平台既有日志自动带上链路 ID，零侵入。
     """
 
@@ -103,10 +109,10 @@ def get_logger(name: str | None = None) -> logging.Logger:
 
 
 # ---------------------------------------------------------------------------
-# span：结构化计时上下文管理器（耗时 + trace_id + 自定义字段）
+# span：结构化计时上下文管理器（耗时 + trace_id + 自定义字段 + OTel 嵌套）
 # ---------------------------------------------------------------------------
 class _Span:
-    """同步/异步通用的计时 span。退出时记录 elapsed_ms 与状态。"""
+    """同步/异步通用的计时 span。退出时记录 elapsed_ms 与状态，并结束嵌套的 OTel span。"""
 
     def __init__(self, name: str, **fields):
         self.name = name
@@ -114,6 +120,7 @@ class _Span:
         self._logger = logging.getLogger("aiflow.span")
         self._start = None
         self.elapsed_ms = 0.0
+        self._otel_cm = None  # 嵌套的 OTel span 上下文管理器
 
     def _enter(self):
         self._start = time.perf_counter()
@@ -121,6 +128,10 @@ class _Span:
             f"SPAN_START {self.name}",
             extra={"span": self.name, "event": "start", "trace_id": get_trace_id(), **self.fields},
         )
+        # OTel 嵌套 span（失败安全：current_span_cm 返回 None 时 no-op）
+        self._otel_cm = current_span_cm(self.name, self.fields)
+        if self._otel_cm is not None:
+            self._otel_cm.__enter__()
         return self
 
     def _exit(self, exc_type, exc, tb):
@@ -137,6 +148,24 @@ class _Span:
                 **self.fields,
             },
         )
+        # 结束嵌套的 OTel span（失败安全）
+        if self._otel_cm is not None:
+            try:
+                from opentelemetry import trace as _ot
+                if exc_type is not None:
+                    self._otel_cm.set_status(
+                        _ot.trace.Status(_ot.trace.StatusCode.ERROR, str(exc_type.__name__))
+                    )
+                else:
+                    self._otel_cm.set_status(_ot.trace.Status(_ot.trace.StatusCode.OK))
+                self._otel_cm.__exit__(exc_type, exc, tb)
+            except Exception:
+                pass
+        # 记录 span 指标（耗时直方图 + 计数），供 Grafana 出图（失败安全）
+        try:
+            record_span_metrics(self.name, self.elapsed_ms, status)
+        except Exception:
+            pass
         return False  # 不吞掉异常
 
     # 同步协议

backend/python/app/core/orchestrator.py

@@ -195,7 +195,7 @@ async def _run_local_agent_turn(agent: dict, node_input: str, db, user_id, conv_
     from app.api.chat_utils import _proxy_for_url
 
     skills = _resolve_market_skills(db, agent.get("skill_ids") or [], user_id)
-    kb_context = _retrieve_kb_context(db, agent.get("knowledge_base_ids") or [], node_input)
+    kb_context = await _retrieve_kb_context(db, agent.get("knowledge_base_ids") or [], node_input)
     system_prompt, function_definitions = _build_system_prompt(
         skills=skills,
         agents=[agent],

backend/python/app/core/otel.py

@@ -0,0 +1,131 @@
+"""
+OpenTelemetry 桥接 (Phase 1.I 标准化)
+=====================================
+设计基准（标准对照表见 STANDARDS_AUDIT_2026-07-24.md §I）：
+- 标准：OpenTelemetry SDK + OTLP gRPC exporter —— 云原生可观测事实标准。
+- traces：span 经 observability.span() 嵌套导出（见 observability.py）。
+- metrics：MeterProvider 经 OTLP 导出 span 耗时直方图 + 调用计数，供 Grafana 出图。
+- 失败安全（生产红线）：若 OTLP 端点不可达 / 初始化异常，自动降级为 no-op，
+  绝不阻断业务请求、不产生未捕获异常。
+- 门控：OTEL_ENABLED=1 默认开启；置 0 则零开销、不走 OTel。
+- 发射目标：OTEL_EXPORTER_OTLP_ENDPOINT，默认 http://qihu-aiflow-alloy:4317
+  （经 Grafana Alloy 桥接 OTLP→Prometheus，再由 Prometheus 喂给 Grafana 出图）。
+
+对外 API：
+    init_otel()            首次调用时构建 TracerProvider + MeterProvider（幂等）
+    get_tracer()           返回 tracer（未启用/失败则 None）
+    get_meter()            返回 meter（未启用/失败则 None）
+    current_span_cm(name)  返回一个 OTel span 的上下文管理器（供 observability 嵌套）
+    record_span_metrics(name, elapsed_ms, status)  记录 span 指标（耗时直方图 + 计数）
+"""
+
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+_ENABLED = os.getenv("OTEL_ENABLED", "1") not in ("0", "false", "False", "")
+_ENDPOINT = os.getenv("OTEL_EXPORTER_OTLP_ENDPOINT", "http://qihu-aiflow-alloy:4317").rstrip("/")
+if not _ENDPOINT.startswith("http://") and not _ENDPOINT.startswith("https://"):
+    _ENDPOINT = "http://" + _ENDPOINT
+
+_tracer = None
+_meter = None
+_initialized = False
+_duration_hist = None
+_span_counter = None
+
+
+def init_otel() -> None:
+    """构建 TracerProvider + MeterProvider（幂等）。失败则降级为 no-op。"""
+    global _tracer, _meter, _initialized
+    if _initialized:
+        return
+    _initialized = True
+    if not _ENABLED:
+        logger.info("[otel] disabled (OTEL_ENABLED=0)")
+        return
+    try:
+        from opentelemetry import trace, metrics
+        from opentelemetry.sdk.resources import Resource
+        from opentelemetry.semconv.resource import ResourceAttributes
+        from opentelemetry.sdk.trace import TracerProvider
+        from opentelemetry.sdk.trace.export import BatchSpanProcessor
+        from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import (
+            OTLPSpanExporter,
+        )
+        from opentelemetry.sdk.metrics import MeterProvider
+        from opentelemetry.sdk.metrics.export import PeriodicExportingMetricReader
+        from opentelemetry.exporter.otlp.proto.grpc.metric_exporter import (
+            OTLPMetricExporter,
+        )
+
+        resource = Resource.create({ResourceAttributes.SERVICE_NAME: "qihu-aiflow"})
+
+        # --- traces ---
+        trace_provider = TracerProvider(resource=resource)
+        trace_provider.add_span_processor(
+            BatchSpanProcessor(OTLPSpanExporter(endpoint=_ENDPOINT, insecure=True))
+        )
+        trace.set_tracer_provider(trace_provider)
+        _tracer = trace.get_tracer("qihu-aiflow")
+
+        # --- metrics ---
+        metric_reader = PeriodicExportingMetricReader(
+            OTLPMetricExporter(endpoint=_ENDPOINT, insecure=True)
+        )
+        meter_provider = MeterProvider(resource=resource, metric_readers=[metric_reader])
+        metrics.set_meter_provider(meter_provider)
+        _meter = metrics.get_meter("qihu-aiflow")
+
+        logger.info(f"[otel] tracer+meter ready -> {_ENDPOINT}")
+    except Exception as e:  # pragma: no cover - 降级保护
+        logger.warning(f"[otel] init failed, degraded to no-op: {e}")
+        _tracer = None
+        _meter = None
+
+
+def get_tracer():
+    """返回 tracer；首次调用触发 init。未启用/失败返回 None。"""
+    if _tracer is None and _ENABLED:
+        init_otel()
+    return _tracer
+
+
+def get_meter():
+    """返回 meter；首次调用触发 init。未启用/失败返回 None。"""
+    if _meter is None and _ENABLED:
+        init_otel()
+    return _meter
+
+
+def current_span_cm(name: str, attributes: dict | None = None):
+    """返回一个 OTel span 的上下文管理器（同步/异步通用）。失败时返回 None。"""
+    tr = get_tracer()
+    if tr is None:
+        return None
+    try:
+        return tr.start_as_current_span(name, attributes=attributes or {})
+    except Exception as e:  # pragma: no cover
+        logger.warning(f"[otel] start_span failed: {e}")
+        return None
+
+
+def record_span_metrics(name: str, elapsed_ms: float, status: str) -> None:
+    """记录 span 指标：耗时直方图 + 调用计数（供 Grafana 出图）。失败安全。"""
+    m = get_meter()
+    if m is None:
+        return
+    global _duration_hist, _span_counter
+    try:
+        if _duration_hist is None:
+            _duration_hist = m.create_histogram(
+                "aiflow_span_duration_ms", "Span duration in milliseconds"
+            )
+            _span_counter = m.create_counter("aiflow_span_count", "Span invocation count")
+        attrs = {"span": name, "status": status}
+        _duration_hist.record(elapsed_ms, attrs)
+        _span_counter.add(1, attrs)
+    except Exception:  # pragma: no cover
+        pass
+

backend/python/app/core/rag_pipeline.py

@@ -0,0 +1,97 @@
+"""标准 RAG 管线 (Phase 2.C 标准化)
+====================================
+设计基准（标准对照表见 STANDARDS_AUDIT_2026-07-24.md §C）：
+- 对齐 LlamaIndex / LangChain 的管线形态：load -> chunk(overlap) -> embed -> store -> retrieve -> rerank。
+- 复用现有 pgvector 后端（document_embeddings 表 + ivfflat）与共享 embeddings 客户端
+  （app.core.llm_client.aembed）；嵌入维度(1024)、表结构、混合检索(RRF)全部沿用，不重复造轮子。
+- 不引入 LlamaIndex 全量重依赖：平台运行时为最小 fastapi/httpx 容器，引入 LlamaIndex 栈有体积/冲突风险；
+  故以「接口对齐标准」的方式实现标准管线（递归重叠分块 + 召回 + rerank 钩子）。后续若评估可接受，
+  可平滑替换为 llama_index.core 的 VectorStoreIndex + PGVectorStore（对外接口一致）。
+
+对外 API：
+    retrieve(query, kb_id, top_k=5, rerank=True) -> [dict]   # 标准检索（向量召回 + 可选 rerank）
+    index_document(text, doc_id, kb_id, ...) -> int          # 标准索引（递归重叠分块后逐块入库）
+"""
+import logging
+
+from app.pgvector_search import vector_search, index_document as _pg_index
+
+logger = logging.getLogger("aiflow.rag")
+
+# 标准递归分块分隔符（与 LlamaIndex SentenceSplitter / LangChain RecursiveCharacterTextSplitter 同序）
+_SEPARATORS = ["\n\n", "\n", "。", "！", "？", "；", ".", "!", "?", ";", " ", ""]
+
+
+def _recursive_split(text: str, chunk_size: int = 500, chunk_overlap: int = 80):
+    """递归字符分块（带重叠），对齐标准 Splitter 算法。
+
+    逐层用分隔符切分；某段仍超长时递归细切；最终兜底硬切（带重叠步长）。
+    """
+    text = text or ""
+    if len(text) <= chunk_size:
+        return [text] if text.strip() else []
+    for sep in _SEPARATORS:
+        if not sep:
+            # 兜底：硬切（带重叠步长）
+            step = max(1, chunk_size - chunk_overlap)
+            return [text[i : i + chunk_size] for i in range(0, len(text), step)]
+        parts = text.split(sep)
+        if len(parts) <= 1:
+            continue  # 该分隔符在文本中未出现，尝试更细的分隔符
+        chunks, cur = [], ""
+        for p in parts:
+            cand = (cur + sep + p) if cur else p
+            if len(cand) <= chunk_size:
+                cur = cand
+            else:
+                if cur:
+                    chunks.append(cur)
+                    cur = ""
+                if len(p) > chunk_size:
+                    chunks.extend(_recursive_split(p, chunk_size, chunk_overlap))
+                else:
+                    cur = p
+        if cur:
+            chunks.append(cur)
+        if chunks:
+            return [c for c in chunks if c.strip()]
+    return [text]
+
+
+async def index_document(
+    text: str,
+    doc_id: str,
+    kb_id: str,
+    chunk_size: int = 500,
+    chunk_overlap: int = 80,
+    metadata: dict | None = None,
+) -> int:
+    """标准索引：递归重叠分块后逐块存入 pgvector。返回成功分块数。"""
+    chunks = _recursive_split(text, chunk_size, chunk_overlap)
+    ok = 0
+    for i, ch in enumerate(chunks):
+        cid = f"{doc_id}_{i}"
+        if await _pg_index(doc_id, kb_id, ch, chunk_id=cid, metadata=metadata):
+            ok += 1
+    logger.info(f"[rag] indexed {doc_id}: {ok}/{len(chunks)} chunks")
+    return ok
+
+
+async def retrieve(
+    query: str,
+    kb_id: str | None = None,
+    top_k: int = 5,
+    rerank: bool = True,
+) -> list:
+    """标准检索：向量召回 + 可选 rerank。
+
+    返回 [{doc_id, chunk_id, content, score, metadata}, ...]。
+    rerank 钩子默认按向量余弦分排序；接入标准 cross-encoder
+    （sentence-transformers CrossEncoder）即替换此处，对外接口不变。
+    """
+    results = await vector_search(query, kb_id, top_k=top_k * 2 if rerank else top_k)
+    if not results:
+        return results
+    if rerank and len(results) > 1:
+        results = sorted(results, key=lambda r: r.get("score", 0), reverse=True)
+    return results[:top_k]

backend/python/app/kb_router.py

@@ -63,15 +63,65 @@ async def _search_external(service: Dict, query: str, top_k: int = 5) -> List[Di
         logger.warning(f"Unknown external KB type: {service_type}")
         return []
 
+async def _get_web_services(db_session, enabled_only: bool = True) -> List[Dict]:
+    """获取全局 AnySearch 联网检索配置（从 site_settings 读取）。"""
+    from app.anysearch_client import load_anysearch_config
+    cfg = load_anysearch_config(db_session)
+    if not cfg:
+        return []
+    if enabled_only and not cfg.get("is_enabled"):
+        return []
+    api_key = cfg.get("api_key", "")
+    if not api_key:
+        return []
+    return [{
+        "id": "anysearch_global",
+        "service_type": "anysearch",
+        "api_key": api_key,
+        "api_url": cfg.get("api_url") or "https://api.anysearch.com/v1/search",
+        "config": {
+            "zone": cfg.get("zone", "cn"),
+            "freshness": cfg.get("freshness"),
+            "top_k": cfg.get("top_k", 5),
+        },
+        "is_enabled": cfg.get("is_enabled", False),
+    }]
+
+
+async def _search_anysearch(service: Dict, query: str, top_k: int = 5) -> List[Dict]:
+    """调用 AnySearch 联网检索（走 app.anysearch_client）。"""
+    try:
+        from app.anysearch_client import anysearch_search
+        api_key = service.get("api_key", "")
+        api_url = service.get("api_url") or "https://api.anysearch.com/v1/search"
+        cfg = service.get("config")
+        if isinstance(cfg, str):
+            try:
+                import json
+                cfg = json.loads(cfg)
+            except Exception:
+                cfg = None
+        cfg = cfg or {}
+        zone = cfg.get("zone", "cn")
+        freshness = cfg.get("freshness")
+        k = int(cfg.get("top_k", top_k))
+        return await anysearch_search(api_key, api_url, query, top_k=k, zone=zone, freshness=freshness)
+    except Exception as e:
+        logger.error(f"AnySearch search error: {e}")
+        return []
+
 
 async def kb_router_search(
     query: str,
     kb_id: Optional[str] = None,
+    kb_ids: Optional[List[str]] = None,
     top_k: int = 5,
     bm25_results: Optional[List[Dict]] = None,
     db_session=None,
     use_vector: bool = True,
     use_external: bool = True,
+    use_tree: bool = True,
+    use_web: bool = False,
 ) -> List[Dict]:
     """
     知识库路由层 — 混合检索
@@ -79,7 +129,9 @@ async def kb_router_search(
     检索策略:
     1. 本地 BM25 检索 (MySQL, 已有)
     2. 向量语义检索 (PostgreSQL+pgvector, 新增)
-    3. 外部KB检索 (Dify, 按配置)
+    3. 结构化分片段检索 (TreeSearch, 新增: 知识库目录树 + chunk 级精准命中)
+    4. 外部KB检索 (Dify, 按配置)
+    5. 联网检索 (AnySearch, 可选 use_web)
 
     融合方式: Reciprocal Rank Fusion (RRF)
 
@@ -88,9 +140,11 @@ async def kb_router_search(
         kb_id: 知识库ID
         top_k: 返回结果数
         bm25_results: BM25检索结果
-        db_session: 数据库会话(用于获取外部KB配置)
+        db_session: 数据库会话(用于获取外部KB配置与分片段检索)
         use_vector: 是否使用向量检索
         use_external: 是否使用外部KB检索
+        use_tree: 是否使用 TreeSearch 结构化分片段检索
+        use_web: 是否使用 AnySearch 联网检索
 
     Returns:
         融合后的检索结果列表
@@ -112,7 +166,13 @@ async def kb_router_search(
     if use_vector:
         try:
             from app.pgvector_search import vector_search
-            vec_results = await vector_search(query, kb_id, top_k)
+            kb_targets = kb_ids if kb_ids else ([kb_id] if kb_id else [None])
+            vec_results = []
+            for _kid in kb_targets:
+                try:
+                    vec_results.extend(await vector_search(query, _kid, top_k))
+                except Exception as e:
+                    logger.warning(f"Vector search failed for kb={_kid}: {e}")
             for rank, item in enumerate(vec_results):
                 all_results.append({
                     "content": item.get("content", ""),
@@ -126,6 +186,25 @@ async def kb_router_search(
         except Exception as e:
             logger.warning(f"Vector search failed: {e}")
 
+    # 2.5 结构化分片段检索 (TreeSearch, Level 2)
+    if use_tree and db_session is not None:
+        try:
+            from app.tree_search import tree_search
+            kb_targets = kb_ids if kb_ids else ([kb_id] if kb_id else None)
+            tree_hits = await tree_search(
+                db_session, query, kb_ids=kb_targets, top_k=top_k, mode="hybrid",
+            )
+            for rank, item in enumerate(tree_hits):
+                all_results.append({
+                    "content": item.get("content", ""),
+                    "score": item.get("score", 0.5),
+                    "source": "tree_chunk",
+                    "rank": rank + 1,
+                    "metadata": item.get("metadata", {}),
+                })
+        except Exception as e:
+            logger.warning(f"TreeSearch failed: {e}")
+
     # 3. 外部KB检索 (Dify)
     if use_external and db_session:
         services = await _get_external_services(db_session)
@@ -141,6 +220,23 @@ async def kb_router_search(
                         "metadata": item.get("metadata", {}),
                     })
 
+    # 4. 联网检索 (AnySearch, 可选)
+    if use_web and db_session is not None:
+        try:
+            web_services = await _get_web_services(db_session)
+            for svc in web_services:
+                web_hits = await _search_anysearch(svc, query, top_k)
+                for rank, item in enumerate(web_hits):
+                    all_results.append({
+                        "content": item.get("content", ""),
+                        "score": item.get("score", 0.6),
+                        "source": item.get("source", "web_anysearch"),
+                        "rank": rank + 1,
+                        "metadata": item.get("metadata", {}),
+                    })
+        except Exception as e:
+            logger.warning(f"AnySearch web search failed: {e}")
+
     if not all_results:
         return []
 

backend/python/app/mcp_client.py

@@ -1,122 +1,96 @@
 """
-MCP (Model Context Protocol) 客户端 —— 异步实现 streamable_http 传输。
+MCP (Model Context Protocol) 客户端 —— 基于官方 \"mcp\" Python SDK (Phase 1.E 标准化)
 
-支持：
+设计基准（标准对照表见 STANDARDS_AUDIT_2026-07-24.md §E）：
-- initialize / notifications/initialized / tools/list / tools/call 标准流程
+- 采用 Anthropic 官方 \"mcp\" SDK 处理 JSON-RPC 帧 + 协议版本协商，
-- streamable_http：POST JSON-RPC，响应可能是 SSE（event: message / data: {...}）或纯 JSON
+  不再手写解析、不再写死 MCP_PROTOCOL_VERSION 常量（旧实现写死 "2024-11-05"，
-- SSE 解析兼容
+  上游协议一升级即碎）。
-- 直连外部端点（trust_env=False，不绕容器代理）
+- 支持 streamable_http（默认）与 sse 两种传输，按 server['transport_type'] 选择。
-
+- 公共 API 保持向后兼容：list_mcp_tools(server, timeout) / call_mcp_tool(server, tool_name, arguments, timeout)
-供 chat.py 在 AI 对话中把 MCP 工具暴露为 function calling，并代理执行 tools/call。
+  供 chat.py / chat_service.py / chat_agents.py / chat_skill.py / chat_mcp.py /
+  chat_context.py / chat_knowledge.py / chat_media.py / chat_utils.py / core/aca.py
+  共 10+ 处调用，签名与返回结构（tools 数组 / 文本结果字符串）不变。
+- 工具列表缓存（按 url）与错误处理保持原行为，避免重复握手、保证降级安全。
 """
 
-import json
-import logging
 import asyncio
+import logging
 from typing import Optional
 
-import httpx
+from mcp import ClientSession
+from mcp.client.sse import sse_client
+from mcp.client.streamable_http import streamablehttp_client
 
 logger = logging.getLogger(__name__)
 
-MCP_PROTOCOL_VERSION = "2024-11-05"
-
 # 工具列表缓存（url -> (timestamp, tools)），避免同一请求内 / 短时间重复握手
 _TOOLS_CACHE: dict = {}
 _TOOLS_CACHE_TTL = 120  # 秒
 
 
-def _parse_sse(text: str):
+def _transport_cm(server: dict, timeout: float):
-    """从响应文本解析 JSON-RPC 结果。兼容 SSE（data: 行）与纯 JSON。"""
+    """按 server['transport_type'] 选择官方 SDK 传输层。
-    if not text:
+
-        return None
+    返回异步上下文管理器，进入后产出 (read_stream, write_stream, ...)。
-    text = text.strip()
+    """
-    # SSE 格式：多行 "event: message" / "data: {...}"
+    url = (server.get("url") or "").strip()
-    data_lines = []
+    ttype = (server.get("transport_type") or "streamable_http").lower()
-    for line in text.splitlines():
+    headers = server.get("headers") or {}
-        ls = line.strip()
+    if ttype == "sse":
-        if ls.startswith("data:"):
+        return sse_client(url, headers=headers, timeout=timeout)
-            data_lines.append(ls[5:].strip())
+    # 默认 streamable_http（官方 SDK 标准传输）
-    if data_lines:
+    return streamablehttp_client(url, headers=headers, timeout=timeout)
-        last = data_lines[-1]
-        try:
-            return json.loads(last)
-        except Exception:
-            pass
-    # 纯 JSON
-    try:
-        return json.loads(text)
-    except Exception:
-        return None
 
 
-async def _handshake_and_request(
+async def _session_request(
-    url: str,
+    server: dict, method: str, params: dict, timeout: float = 30.0
-    headers: dict,
-    method: str,
-    params: dict,
-    timeout: float = 30.0,
 ):
+    """用官方 ClientSession 完成一次标准 MCP 调用（initialize -> 目标请求）。
+
+    返回解析后的 dict：{"result": {...}} 或 None（失败）。
+    协议版本由 SDK 自动协商，不再硬编码。
     """
-    完成一次完整 MCP 会话（initialize -> initialized -> 目标请求）。
+    url = (server.get("url") or "").strip()
-    返回解析后的 JSON-RPC 响应 dict，或 None。
+    if not url:
-    """
+        return None
-    base_headers = {
+    try:
-        "Content-Type": "application/json",
+        async with _transport_cm(server, timeout) as streams:
-        "Accept": "application/json, text/event-stream",
+            read, write, *_ = streams
-    }
+            async with ClientSession(read, write) as session:
-    base_headers.update(headers or {})
+                # 标准握手：initialize 会发送 initialize 请求并自动补 notifications/initialized
+                await session.initialize()
+                # 防御性补发 initialized 通知（部分服务端要求显式接收）
+                try:
+                    if hasattr(session, "send_notification"):
+                        await session.send_notification("notifications/initialized", {})
+                except Exception:
+                    pass
 
-    async with httpx.AsyncClient(timeout=timeout, trust_env=False) as client:
+                if method == "tools/list":
-        # 1) initialize
+                    result = await session.list_tools()
-        r = await client.post(
+                    tools = [t.model_dump() for t in (result.tools or [])]
-            url,
+                    return {"result": {"tools": tools}}
-            headers=base_headers,
+                elif method == "tools/call":
-            json={
+                    res = await session.call_tool(
-                "jsonrpc": "2.0",
+                        params.get("name"), params.get("arguments") or {}
-                "id": 1,
+                    )
-                "method": "initialize",
+                    content = []
-                "params": {
+                    for blk in (res.content or []):
-                    "protocolVersion": MCP_PROTOCOL_VERSION,
+                        btype = getattr(blk, "type", "text")
-                    "capabilities": {},
+                        if btype == "text":
-                    "clientInfo": {"name": "aiflow", "version": "1.0"},
+                            content.append({"type": "text", "text": getattr(blk, "text", "")})
-                },
+                        else:
-            },
+                            content.append({"type": btype, "text": str(blk)})
-        )
+                    return {
-        if r.status_code >= 400:
+                        "result": {
-            logger.warning(f"[MCP] initialize 失败 {url} -> HTTP {r.status_code}")
+                            "content": content,
-            return None
+                            "isError": bool(getattr(res, "isError", False)),
-        sid = r.headers.get("mcp-session-id") or r.headers.get("Mcp-Session-Id")
+                        }
-        req_headers = dict(base_headers)
+                    }
-        if sid:
+    except Exception as e:
-            req_headers["Mcp-Session-Id"] = sid
+        logger.warning(f"[MCP] {method} 异常 {url}: {e}")
-
+        return None
-        # 2) initialized 通知（不需要响应）
+    return None
-        try:
-            await client.post(
-                url,
-                headers=req_headers,
-                json={"jsonrpc": "2.0", "method": "notifications/initialized"},
-            )
-        except Exception:
-            pass
-
-        # 3) 目标请求
-        rid = 2 if method != "tools/call" else 3
-        r2 = await client.post(
-            url,
-            headers=req_headers,
-            json={
-                "jsonrpc": "2.0",
-                "id": rid,
-                "method": method,
-                "params": params,
-            },
-        )
-        if r2.status_code >= 400:
-            logger.warning(f"[MCP] {method} 失败 {url} -> HTTP {r2.status_code}")
-            return None
-        return _parse_sse(r2.text)
 
 
 async def list_mcp_tools(server: dict, timeout: float = 30.0) -> list:
@@ -134,9 +108,7 @@ async def list_mcp_tools(server: dict, timeout: float = 30.0) -> list:
     if cached and (now - cached[0]) < _TOOLS_CACHE_TTL:
         return cached[1]
     try:
-        data = await _handshake_and_request(
+        data = await _session_request(server, "tools/list", {}, timeout)
-            url, server.get("headers") or {}, "tools/list", {}, timeout
-        )
         tools = []
         if data and "result" in data:
             tools = data["result"].get("tools", []) or []
@@ -158,9 +130,8 @@ async def call_mcp_tool(
         return f"[MCP 工具 {tool_name} 执行失败：服务端地址为空]"
     logger.info(f"[MCP] 执行工具调用 tool={tool_name} server={server.get('name')} url={url}")
     try:
-        data = await _handshake_and_request(
+        data = await _session_request(
-            url,
+            server,
-            server.get("headers") or {},
             "tools/call",
             {"name": tool_name, "arguments": arguments or {}},
             timeout,
@@ -169,7 +140,7 @@ async def call_mcp_tool(
             return f"[MCP 工具 {tool_name} 调用失败：无响应]"
         if "error" in data:
             err = data["error"]
-            return f"[MCP 工具 {tool_name} 返回错误] {json.dumps(err, ensure_ascii=False)}"
+            return f"[MCP 工具 {tool_name} 返回错误] {__import__('json').dumps(err, ensure_ascii=False)}"
         result = data.get("result", {})
         content = result.get("content", []) or []
         parts = []
@@ -177,7 +148,7 @@ async def call_mcp_tool(
             if isinstance(item, dict) and item.get("type") == "text":
                 parts.append(item.get("text", ""))
             else:
-                parts.append(json.dumps(item, ensure_ascii=False))
+                parts.append(__import__("json").dumps(item, ensure_ascii=False))
         text = "\n".join(parts).strip()
         if result.get("isError"):
             return f"[MCP 工具 {tool_name} 执行报错]\n{text}"

backend/python/app/models.py

@@ -11,6 +11,7 @@ from sqlalchemy import (
     DECIMAL, UniqueConstraint
 )
 from sqlalchemy.orm import relationship
+from sqlalchemy import Text  # PG-compatible (was MEDIUMTEXT from mysql dialect)
 from .database import Base
 
 
@@ -31,7 +32,7 @@ class User(Base):
     phone = Column(String(20), default=None)
     password_hash = Column(String(255), nullable=False)
     avatar = Column(String(500), default=None)
-    role = Column(String(10), nullable=False, default="user")  # user / admin
+    role = Column(String(20), nullable=False, default="user")  # super_admin/admin/developer/user/guest
     status = Column(String(10), nullable=False, default="active")  # active / banned
     real_name_status = Column(String(20), nullable=False, default="unverified")
     real_name = Column(String(50), default=None)
@@ -39,6 +40,7 @@ class User(Base):
     membership_plan_id = Column(String(36), default=None)
     membership_level = Column(Integer, default=0)
     membership_expire_at = Column(DateTime, default=None)
+    tenant_id = Column(String(36), default=None)
     auto_renew = Column(Boolean, default=False)
     created_at = Column(DateTime, default=datetime.now)
     updated_at = Column(DateTime, default=datetime.now, onupdate=datetime.now)
@@ -75,6 +77,47 @@ class ApiKey(Base):
 # AI聊天与频道
 # ============================================================
 
+
+
+class UserGroup(Base):
+    __tablename__ = "user_groups"
+    id = Column(String(36), primary_key=True, default=gen_uuid)
+    name = Column(String(100), unique=True, nullable=False)
+    description = Column(String(500), default=None)
+    owner_id = Column(String(36), ForeignKey("users.id"), nullable=False)
+    tenant_id = Column(String(36), default=None)
+    created_at = Column(DateTime, default=datetime.now)
+    updated_at = Column(DateTime, default=datetime.now, onupdate=datetime.now)
+    owner = relationship("User", foreign_keys=[owner_id])
+
+class UserGroupMember(Base):
+    __tablename__ = "user_group_members"
+    id = Column(String(36), primary_key=True, default=gen_uuid)
+    group_id = Column(String(36), ForeignKey("user_groups.id"), nullable=False)
+    user_id = Column(String(36), ForeignKey("users.id"), nullable=False)
+    joined_at = Column(DateTime, default=datetime.now)
+    group = relationship("UserGroup")
+    user = relationship("User")
+
+class GroupModelQuota(Base):
+    __tablename__ = "group_model_quotas"
+    id = Column(String(36), primary_key=True, default=gen_uuid)
+    group_id = Column(String(36), ForeignKey("user_groups.id"), nullable=False)
+    model_id = Column(String(36), ForeignKey("models.id"), nullable=False)
+    daily_limit = Column(Integer, default=1000)
+    rpm_limit = Column(Integer, default=60)
+    group = relationship("UserGroup")
+    model = relationship("AIModel")
+
+class Tenant(Base):
+    __tablename__ = "tenants"
+    id = Column(String(36), primary_key=True, default=gen_uuid)
+    name = Column(String(100), nullable=False)
+    slug = Column(String(50), unique=True, nullable=False)
+    status = Column(String(20), default="active")
+    created_at = Column(DateTime, default=datetime.now)
+    updated_at = Column(DateTime, default=datetime.now, onupdate=datetime.now)
+
 class Conversation(Base):
     __tablename__ = "conversations"
 
@@ -286,6 +329,10 @@ class MarketApp(Base):
     publisher_id = Column(String(36), ForeignKey("users.id"), default=None)
     publisher_name = Column(String(100), default=None)
     description = Column(Text, default=None)
+    # 分类与标签：DB 表 market_apps 已存在 category(varchar50)/tags(varchar500) 两列，
+    # 此前 ORM 漏声明导致后台读取分类恒为 None -> 回退 "other"（#679 根因）。此处补齐。
+    category = Column(String(50), default="other", comment="技能/应用分类(9类: writing/coding/data_analysis/qa/automation/creativity/communication/reasoning/other)")
+    tags = Column(String(500), default=None, comment="标签,逗号分隔")
     icon = Column(String(500), default=None)
     tutorial = Column(Text, default=None)
     price = Column(DECIMAL(10, 2), nullable=False, default=0.00)
@@ -525,6 +572,7 @@ class ModelProvider(Base):
     protocol = Column(String(30), nullable=False, default="openai_compatible")
     api_key = Column(String(255), default=None)
     api_base_url = Column(String(500), default=None)
+    auth_header = Column(String(100), default="Authorization: Bearer", comment="自定义认证头，如MiMo用api-key")
     created_at = Column(DateTime, default=datetime.now)
     updated_at = Column(DateTime, default=datetime.now, onupdate=datetime.now)
 
@@ -945,7 +993,7 @@ class SkillTrajectoryLink(Base):
     skill_name = Column(String(200), nullable=False, index=True, comment="关联的技能名称")
     user_id = Column(String(36), ForeignKey("users.id"), nullable=False, index=True)
     trajectory_id = Column(String(36), nullable=True, comment="关联的轨迹ID")
-    atom_id = Column(String(36), nullable=True, comment="原子任务ID")
+    atom_id = Column(String(128), nullable=True, comment="原子任务ID(形如 trajectory_id_atom_N，需容纳 UUID+后缀)")
     weightscore = Column(Integer, default=0, comment="关联权重分")
     created_at = Column(DateTime, default=datetime.now)
 
@@ -1156,4 +1204,5 @@ class ConversationSummary(Base):
     conversation_id = Column(String(64), primary_key=True)
     summary_text = Column(Text, nullable=False, default="")
     token_count = Column(Integer, default=0)
+    structured_memory = Column(JSON, default=None)  # T2: Semantica 式结构化长期记忆
     updated_at = Column(DateTime, default=datetime.now, onupdate=datetime.now)

backend/python/app/pgvector_search.py

@@ -18,7 +18,7 @@ from datetime import datetime
 logger = logging.getLogger("aiflow.pgvector")
 
 # PostgreSQL 连接配置
-PG_HOST = os.environ.get("POSTGRES_HOST", "pgvector")
+PG_HOST = os.environ.get("POSTGRES_HOST", "qihu-aiflow-pgvector")
 PG_PORT = int(os.environ.get("POSTGRES_PORT", "5432"))
 PG_USER = os.environ.get("POSTGRES_USER", "aiflow")
 PG_PASSWORD = os.environ.get("POSTGRES_PASSWORD", "pwd2026")
@@ -29,6 +29,32 @@ EMBEDDING_API_URL = os.environ.get("SILICONFLOW_BASE_URL", "https://api.siliconf
 EMBEDDING_API_KEY = os.environ.get("SILICONFLOW_API_KEY", "")
 EMBEDDING_MODEL = os.environ.get("EMBEDDING_MODEL", "BAAI/bge-large-zh-v1.5")
 
+# ---- 本地嵌入缓存（SQLite 持久化）：破解"每次嵌入消耗外部 API 额度"的隐性成本障碍 ----
+import sqlite3, hashlib, json as _json
+_EMBED_CACHE_DB = os.environ.get("EMBED_CACHE_DB", "/app/backend/python/.embed_cache.db")
+def _embed_cache_get(text, model):
+    try:
+        key = hashlib.sha256(f"{model}::{text}".encode("utf-8")).hexdigest()
+        conn = sqlite3.connect(_EMBED_CACHE_DB, timeout=5)
+        row = conn.execute("SELECT vec FROM embed_cache WHERE k=?", (key,)).fetchone()
+        conn.close()
+        if row:
+            return _json.loads(row[0])
+    except Exception:
+        return None
+    return None
+def _embed_cache_put(text, model, vec):
+    try:
+        key = hashlib.sha256(f"{model}::{text}".encode("utf-8")).hexdigest()
+        conn = sqlite3.connect(_EMBED_CACHE_DB, timeout=5)
+        conn.execute("CREATE TABLE IF NOT EXISTS embed_cache (k TEXT PRIMARY KEY, vec TEXT, ts TIMESTAMP DEFAULT CURRENT_TIMESTAMP)")
+        conn.execute("INSERT OR REPLACE INTO embed_cache (k, vec) VALUES (?, ?)", (key, _json.dumps(vec)))
+        conn.commit()
+        conn.close()
+    except Exception:
+        pass
+
+
 # 连接池
 _pg_pool = None
 _pool_lock = asyncio.Lock()
@@ -105,27 +131,20 @@ async def get_embedding(text: str) -> Optional[List[float]]:
         logger.warning("No embedding API key configured")
         return None
 
+    cached = _embed_cache_get(text, EMBEDDING_MODEL)
+    if cached:
+        return cached
     try:
-        import httpx
+        # 收敛到共享标准模型层（app.core.llm_client.aembed），与聊天共用代理/超时/failover 语义
-        async with httpx.AsyncClient(timeout=30) as client:
+        from app.core.llm_client import aembed
-            resp = await client.post(
+        embedding = await aembed(
-                f"{EMBEDDING_API_URL}/v1/embeddings",
+            EMBEDDING_API_URL, EMBEDDING_API_KEY, text[:8000], EMBEDDING_MODEL, timeout=30
-                headers={
+        )
-                    "Authorization": f"Bearer {EMBEDDING_API_KEY}",
+        if embedding:
-                    "Content-Type": "application/json",
+            _embed_cache_put(text, EMBEDDING_MODEL, embedding)
-                },
+            return embedding
-                json={
+        logger.error("Embedding API returned empty")
-                    "model": EMBEDDING_MODEL,
+        return None
-                    "input": text[:8000],  # 截断防止超长
-                },
-            )
-            if resp.status_code == 200:
-                data = resp.json()
-                embedding = data["data"][0]["embedding"]
-                return embedding
-            else:
-                logger.error(f"Embedding API error: {resp.status_code} {resp.text[:200]}")
-                return None
     except Exception as e:
         logger.error(f"Embedding generation failed: {e}")
         return None

backend/python/app/platform_knowledge.py

@@ -38,7 +38,7 @@ PLATFORM_KNOWLEDGE = """# 七虎AI平台能力全览
 
 ## 5. IM Bot（微信集成）
 - **个人微信Bot**: 通过iLink官方接口接入，扫码绑定，微信内直接AI对话
-- **企业微信/飞书/QQ/悟空IM**: Webhook框架已就绪，支持自定义机器人
+- **企业微信/飞书/QQ/悟空IM**: Webhook框架已绪，支持自定义机器人
 - **统一会话管理**: 微信消息并入平台AI聊天，统一历史管理
 
 ## 6. 资源中心
@@ -73,7 +73,7 @@ PLATFORM_KNOWLEDGE = """# 七虎AI平台能力全览
 - **安全**: JWT鉴权、HTTPS加密、操作审计日志
 
 ## AI回复规范
-1. **回复格式**: 使用Markdown（标题 `##`/`###`、加粗 `**`、列表 `-`/`1.`、代码块等）
+1. **回复格式**: 使用Markdown（标题 \"##\"/\"###\"、加粗 \"**\"、列表 \"-\"/\"1.\"、代码块等）
 2. **层次分明**: 使用标题分隔段落，确保信息有层次
 3. **平台优先**: 回答问题时优先引用本平台功能，而非泛泛建议
 4. **准确具体**: 基于平台文档给出具体操作路径、配置方法、API示例
@@ -81,6 +81,78 @@ PLATFORM_KNOWLEDGE = """# 七虎AI平台能力全览
 """
 
 
-def get_platform_knowledge() -> str:
+def get_platform_knowledge(db=None, user_id=None) -> str:
-    """获取平台全功能知识库文本，注入到AI对话系统提示中"""
+    """获取平台全功能知识库文本，注入到AI对话系统提示中。
-    return PLATFORM_KNOWLEDGE
+
+    若传入 db，则动态追加当前平台已上线的官方技能目录
+    （source=official 且 status=published），确保管理员/前台导入或下架技能后，
+    AI 能及时感知并使用「平台自身已有内容」，而非停留在陈旧硬编码清单（见 #712）。
+    任何异常均安全回退到静态 PLATFORM_KNOWLEDGE。
+    """
+    if db is None:
+        return PLATFORM_KNOWLEDGE
+    try:
+        from .models import MarketApp
+
+        apps = (
+            db.query(MarketApp)
+            .filter(MarketApp.source == "official", MarketApp.status == "published")
+            .order_by(MarketApp.capability, MarketApp.type, MarketApp.name)
+            .all()
+        )
+        if not apps:
+            return PLATFORM_KNOWLEDGE
+
+        # 按类型分组统计
+        type_count = {}
+        for a in apps:
+            t = getattr(a, "type", None) or "skill"
+            type_count[t] = type_count.get(t, 0) + 1
+        type_brief = ", ".join([f"{({'skill':'技能','agent':'Agent','mcp':'MCP应用','mcp_tool':'MCP工具'}.get(k,k))} {v}" for k, v in sorted(type_count.items())])
+
+        # 按能力(capability)分组
+        from collections import OrderedDict
+        by_cap = OrderedDict()
+        for a in apps:
+            cap = getattr(a, "capability", None) or getattr(a, "category", None) or "通用"
+            by_cap.setdefault(cap, []).append(a)
+
+        # 强制完整性说明（关键变更）
+        lines = [
+            "# 📦 平台已上线能力清单（实时同步自数据库，可放心信任此清单）",
+            "",
+            f"**总数：{len(apps)} 个已上线能力**（{type_brief}）",
+            "",
+            "## ⚠️ 强制完整性指令",
+            "当用户询问以下任何一类问题时，**必须输出下方全部内容**：",
+            '- "平台有哪些技能/应用/MCP"',
+            '- "全部技能" / "列一下技能" / "技能清单"',
+            '- "有什么 Skills" / "支持哪些 Skills"',
+            '- "列举所有" / "每个都列出来"',
+            "",
+            "**绝对不允许**精简、抽样、挑代表性的几个、或改写为综述类叙述。**逐条输出下方每一项**，可以按类别分组展示，但不可省略任何一条。",
+            "",
+        ]
+
+        # 按能力分组渲染
+        type_emoji = {"skill": "🛠", "agent": "🤖", "mcp": "🔌", "mcp_tool": "🔧"}
+        for cap in by_cap.keys():
+            cap_apps = by_cap[cap]
+            lines.append(f"## 【{cap}】- {len(cap_apps)} 个")
+            for a in cap_apps:
+                name = getattr(a, "name", None) or "未命名"
+                desc = (getattr(a, "description", "") or "").strip().replace("\n", " ")
+                # 描述长度提升到 100 字，给出更可识别的简介
+                if len(desc) > 100:
+                    desc = desc[:100] + "…"
+                emoji = type_emoji.get(getattr(a, "type", "skill"), "📦")
+                lines.append(f"- {emoji} **{name}**: {desc}")
+            lines.append("")
+
+        dynamic_section = "\n".join(lines)
+
+        # 关键：动态能力清单放在最前面，让 AI 优先处理
+        return dynamic_section + "\n\n---\n\n# 平台静态文档\n\n" + PLATFORM_KNOWLEDGE
+    except Exception as e:
+        logger.warning(f"[platform_knowledge] 动态构建失败，回退静态知识库: {e}")
+        return PLATFORM_KNOWLEDGE

backend/python/app/services/evolution_service.py

@@ -96,10 +96,24 @@ class PipelineScheduler:
                             self._process_trajectory(db, traj)
                         except Exception as e:
                             logger.error(f"处理轨迹 {traj.id} 失败: {e}")
-                            traj.error_msg = str(e)[:500]
+                            # 先回滚失效事务（_process_trajectory 内 commit 失败会让 session
-                            traj.retry_count = (traj.retry_count or 0) + 1
+                            # 进入 rolled-back 状态，不回滚直接再 commit 会抛 PendingRollbackError
-                            traj.status = "error"
+                            # 并级联到外层日志），再写错误状态。
-                            db.commit()
+                            try:
+                                db.rollback()
+                            except Exception:
+                                pass
+                            try:
+                                traj.error_msg = str(e)[:500]
+                                traj.retry_count = (traj.retry_count or 0) + 1
+                                traj.status = "error"
+                                db.commit()
+                            except Exception as ce:
+                                logger.error(f"更新轨迹 {traj.id} 错误状态失败: {ce}")
+                                try:
+                                    db.rollback()
+                                except Exception:
+                                    pass
                 finally:
                     db.close()
             except Exception as e:

backend/python/app/tree_search.py

@@ -0,0 +1,405 @@
+"""
+TreeSearch — 结构化文档两级检索
+================================
+
+为知识库提供有结构的「两级检索」能力（零额外基础设施成本，复用 MySQL 已存的分段数据）：
+
+  Level 1 (build_kb_tree)   : 知识库 -> 目录 -> 文档 的嵌套结构树，用于浏览/定位/面包屑。
+  Level 2 (tree_search)     : 在给定作用域（整个 KB / 某目录 / 某文档）内做分段(chunk)级检索，
+                               BM25 关键词排序 + 可选 pgvector 语义向量 boost，返回带完整
+                               面包屑(KB名 -> 目录路径 -> 文件名 -> 段序号/标题)的精准命中。
+
+设计要点：
+  * Level 2 直接消费 knowledge_document_chunks 表（content + chunk_metadata.title/keywords），
+    因此召回的是「文档内部的精确片段」，而非整篇文档——这正是 RAG 想要的细粒度上下文。
+  * 与 kb_router_search 融合：tree_search 作为 RRF 的一个新来源 (source="tree_chunk") 接入，
+    聊天 KB 上下文自动获得结构化分片段命中。
+  * 向量 boost 为可选且优雅降级：当 pgvector 无对应数据时，自动退化为纯 BM25，不影响主链路。
+"""
+
+import logging
+import math
+from typing import List, Dict, Optional
+
+logger = logging.getLogger("aiflow.tree_search")
+
+
+# ---------------------------------------------------------------------------
+# 中文友好的分词（jieba 优先，否则按字符 bigram + 英文正则兜底）
+# ---------------------------------------------------------------------------
+def _tokenize(text: str) -> List[str]:
+    if not text:
+        return []
+    toks: List[str] = []
+    try:
+        import re
+        # 英文/数字 token
+        for m in re.findall(r"[a-zA-Z0-9_]+", text.lower()):
+            if len(m) >= 1:
+                toks.append(m)
+        # 中文按 jieba 或字符切分
+        cjk = "".join(re.findall(r"[一-鿿]", text))
+        if cjk:
+            try:
+                import jieba
+                for w in jieba.lcut(cjk):
+                    w = w.strip().lower()
+                    if w:
+                        toks.append(w)
+            except Exception:
+                # 字符级 bigram（兼顾召回与精度）
+                for i in range(len(cjk)):
+                    toks.append(cjk[i])
+                    if i + 1 < len(cjk):
+                        toks.append(cjk[i:i + 2])
+    except Exception:
+        toks = [t for t in text.lower().split() if t]
+    return toks
+
+
+def _chunk_display_text(chunk) -> str:
+    """拼装分段的检索文本：正文 + 标题 + 关键词（来自 chunk_metadata）。"""
+    parts = []
+    content = getattr(chunk, "content", None) or ""
+    parts.append(content)
+    md = getattr(chunk, "chunk_metadata", None) or {}
+    if isinstance(md, dict):
+        title = md.get("title") or md.get("heading") or ""
+        keywords = md.get("keywords") or []
+        if title:
+            parts.append(title)
+        if isinstance(keywords, list) and keywords:
+            parts.append(" ".join(str(k) for k in keywords))
+        elif isinstance(keywords, str) and keywords:
+            parts.append(keywords)
+    return "\n".join(p for p in parts if p).strip()
+
+
+def _chunk_title(chunk) -> str:
+    md = getattr(chunk, "chunk_metadata", None) or {}
+    if isinstance(md, dict):
+        return md.get("title") or md.get("heading") or ""
+    return ""
+
+
+# ---------------------------------------------------------------------------
+# BM25（在作用域内的分段语料上即时计算 IDF）
+# ---------------------------------------------------------------------------
+def _bm25_rank(query_tokens: List[str], corpus: List[Dict], k1: float = 1.5, b: float = 0.75) -> List[Dict]:
+    """corpus: [{"id":..., "tokens":[...], "chunk":<orm>, "text":str}] -> 返回带 bm25 分的列表"""
+    if not query_tokens or not corpus:
+        return []
+    n = len(corpus)
+    df = {}
+    for qt in set(query_tokens):
+        df[qt] = sum(1 for d in corpus if qt in d["tokens"])
+    idf = {qt: math.log((n - df[qt] + 0.5) / (df[qt] + 0.5) + 1) for qt in query_tokens}
+    avgdl = sum(len(d["tokens"]) for d in corpus) / max(n, 1)
+    scored = []
+    for d in corpus:
+        dl = len(d["tokens"])
+        score = 0.0
+        for qt in set(query_tokens):
+            if qt in d["tokens"]:
+                f = d["tokens"].count(qt)
+                score += idf[qt] * (f * (k1 + 1)) / (f + k1 * (1 - b + b * (dl / max(avgdl, 1))))
+        if score > 0:
+            scored.append({**d, "bm25": score})
+    scored.sort(key=lambda x: x["bm25"], reverse=True)
+    return scored
+
+
+# ---------------------------------------------------------------------------
+# Level 1：知识库目录树
+# ---------------------------------------------------------------------------
+def build_kb_tree(db, user_id: Optional[str] = None, kb_id: Optional[str] = None) -> List[Dict]:
+    """构造 知识库 -> 目录(可嵌套) -> 文档 的结构树。"""
+    from app.models import KnowledgeBase, KnowledgeDirectory, KnowledgeDocument
+
+    q = db.query(KnowledgeBase)
+    if user_id:
+        q = q.filter(KnowledgeBase.user_id == user_id)
+    if kb_id:
+        q = q.filter(KnowledgeBase.id == kb_id)
+    bases = q.all()
+
+    tree = []
+    for kb in bases:
+        dirs = db.query(KnowledgeDirectory).filter(
+            KnowledgeDirectory.knowledge_base_id == kb.id
+        ).all()
+        docs = db.query(KnowledgeDocument).filter(
+            KnowledgeDocument.knowledge_base_id == kb.id,
+            KnowledgeDocument.status == "completed",
+        ).all()
+        dir_by_id = {d.id: d for d in dirs}
+        node_by_id = {}
+
+        def make_dir_node(d):
+            return {
+                "id": d.id,
+                "name": d.name,
+                "type": "directory",
+                "children": [],
+                "documents": [],
+            }
+
+        for d in dirs:
+            node_by_id[d.id] = make_dir_node(d)
+        root_dirs = []
+        for d in dirs:
+            node = node_by_id[d.id]
+            parent = d.parent_id
+            if parent and parent in node_by_id:
+                node_by_id[parent]["children"].append(node)
+            else:
+                root_dirs.append(node)
+        for doc in docs:
+            item = {
+                "id": doc.id,
+                "name": doc.filename,
+                "type": "document",
+                "file_type": doc.file_type,
+                "chunk_count": doc.chunk_count or 0,
+            }
+            if doc.directory_id and doc.directory_id in node_by_id:
+                node_by_id[doc.directory_id]["documents"].append(item)
+            else:
+                root_dirs.append(item)  # 顶层文档
+        tree.append({
+            "id": kb.id,
+            "name": kb.name,
+            "type": "knowledge_base",
+            "directories": root_dirs,
+            "document_count": len(docs),
+        })
+    return tree
+
+
+# ---------------------------------------------------------------------------
+# 面包屑：KB名 -> 目录路径 -> 文件名 -> 段标题/序号
+# ---------------------------------------------------------------------------
+def _breadcrumb(db, kb_id: str, directory_id: Optional[str], doc_id: str, chunk_index: int, title: str) -> Dict:
+    from app.models import KnowledgeBase, KnowledgeDirectory, KnowledgeDocument
+    kb = db.query(KnowledgeBase).filter(KnowledgeBase.id == kb_id).first()
+    kb_name = kb.name if kb else kb_id
+    path = []
+    cur = directory_id
+    guard = 0
+    while cur and guard < 20:
+        d = db.query(KnowledgeDirectory).filter(KnowledgeDirectory.id == cur).first()
+        if not d:
+            break
+        path.insert(0, d.name)
+        cur = d.parent_id
+        guard += 1
+    doc = db.query(KnowledgeDocument).filter(KnowledgeDocument.id == doc_id).first()
+    doc_name = doc.filename if doc else doc_id
+    return {
+        "knowledge_base": kb_name,
+        "directory_path": "/".join(path) if path else "/",
+        "document": doc_name,
+        "chunk_index": chunk_index,
+        "chunk_title": title,
+        "label": f"{kb_name}/{('/'.join(path) if path else '')}/{doc_name}" + (f"#段{chunk_index}" + (f"·{title}" if title else "")),
+    }
+
+
+# ---------------------------------------------------------------------------
+# Level 2：结构化分片段检索
+# ---------------------------------------------------------------------------
+async def tree_search(
+    db,
+    query: str,
+    user_id: Optional[str] = None,
+    kb_ids: Optional[List[str]] = None,
+    scope_base: Optional[str] = None,
+    scope_dir: Optional[str] = None,
+    scope_doc: Optional[str] = None,
+    top_k: int = 5,
+    mode: str = "hybrid",
+    max_tokens: int = 1200,
+) -> List[Dict]:
+    """
+    两级检索的 Level 2：在作用域内做分片段精准检索。
+
+    Args:
+        query      : 查询文本
+        kb_ids     : 限定知识库范围（多个）
+        scope_base : 限定单一知识库
+        scope_dir  : 限定单一目录（含子目录）
+        scope_doc  : 限定单一文档
+        top_k      : 返回命中数
+        mode       : "bm25" | "vector" | "hybrid"(默认)
+    Returns:
+        [{
+            "content": <分段文本>,
+            "score": 0-1,
+            "source": "tree_chunk",
+            "metadata": { "breadcrumb": {...}, "snippet": ..., "chunk_index": int, "title": str }
+        }, ...]
+    """
+    from app.models import KnowledgeBase, KnowledgeDirectory, KnowledgeDocument, KnowledgeDocumentChunk
+
+    if not query or not query.strip():
+        return []
+
+    # 1) 解析作用域 -> 目标文档 id 集合
+    target_kb_ids = list(kb_ids or [])
+    if scope_base and scope_base not in target_kb_ids:
+        target_kb_ids.append(scope_base)
+
+    doc_scope = set()
+    if scope_doc:
+        doc_scope.add(scope_doc)
+    if scope_dir:
+        # 收集该目录及其所有子目录下的文档
+        dirs = db.query(KnowledgeDirectory).all()
+        dir_by_id = {d.id: d for d in dirs}
+        sub = set()
+
+        def collect(did):
+            sub.add(did)
+            for d in dirs:
+                if d.parent_id == did:
+                    collect(d.id)
+        collect(scope_dir)
+        docs = db.query(KnowledgeDocument).filter(KnowledgeDocument.directory_id.in_(sub)).all()
+        for d in docs:
+            doc_scope.add(d.id)
+    if not doc_scope and not target_kb_ids and not user_id:
+        return []
+
+    # 2) 取分段语料
+    q = db.query(KnowledgeDocumentChunk)
+    if doc_scope:
+        q = q.filter(KnowledgeDocumentChunk.document_id.in_(doc_scope))
+    else:
+        if target_kb_ids:
+            q = q.filter(KnowledgeDocumentChunk.knowledge_base_id.in_(target_kb_ids))
+        elif user_id:
+            kb_ids_u = [r[0] for r in db.query(KnowledgeBase.id).filter(KnowledgeBase.user_id == user_id).all()]
+            if not kb_ids_u:
+                return []
+            q = q.filter(KnowledgeDocumentChunk.knowledge_base_id.in_(kb_ids_u))
+    chunks = q.order_by(KnowledgeDocumentChunk.chunk_index).limit(500).all()
+    if not chunks:
+        return []
+
+    query_tokens = _tokenize(query)
+    corpus = []
+    for c in chunks:
+        text = _chunk_display_text(c)
+        if len(text.strip()) < 8:
+            continue
+        corpus.append({
+            "id": c.id,
+            "tokens": _tokenize(text),
+            "chunk": c,
+            "text": text,
+        })
+
+    # 3) BM25 排序
+    bm25_hits = _bm25_rank(query_tokens, corpus) if mode in ("bm25", "hybrid") else []
+
+    # 4) 可选向量 boost
+    vector_boost = {}
+    if mode in ("vector", "hybrid"):
+        try:
+            from app.pgvector_search import vector_search
+            for kid in (target_kb_ids or [None]):
+                try:
+                    vres = await vector_search(query, kid, top_k=top_k * 3)
+                except Exception as e:
+                    logger.warning(f"[tree_search] vector_search failed: {e}")
+                    continue
+                for vr in vres:
+                    vcontent = (vr.get("content") or "").strip()
+                    if vcontent:
+                        vector_boost[vcontent] = max(vector_boost.get(vcontent, 0.0), float(vr.get("score", 0.0)))
+        except Exception as e:
+            logger.debug(f"[tree_search] vector unavailable: {e}")
+
+    # 5) 融合
+    merged_map = {}
+    for rank, h in enumerate(bm25_hits):
+        key = h["text"][:300]
+        boost = 0.0
+        for vc, vs in vector_boost.items():
+            if vc and vc in h["text"]:
+                boost = max(boost, vs)
+        merged_map[key] = {
+            "chunk": h["chunk"],
+            "bm25": h["bm25"],
+            "vec": boost,
+            "rank": rank + 1,
+        }
+    # 纯向量模式：把仅在向量里命中的也纳入
+    if mode == "vector":
+        for vc, vs in vector_boost.items():
+            if vc[:300] not in merged_map:
+                # 找到对应 chunk
+                for c in corpus:
+                    if vc in c["text"]:
+                        merged_map[vc[:300]] = {"chunk": c["chunk"], "bm25": 0.0, "vec": vs, "rank": 999}
+                        break
+
+    if not merged_map:
+        return []
+
+    # 综合分数（BM25 主导，向量做 0~1 归一 boost）
+    max_bm25 = max((m["bm25"] for m in merged_map.values()), default=1.0) or 1.0
+    items = []
+    for key, m in merged_map.items():
+        c = m["chunk"]
+        norm_bm25 = (m["bm25"] / max_bm25) if max_bm25 > 0 else 0.0
+        norm_vec = min(m["vec"], 1.0)
+        combined = norm_bm25 * 0.7 + norm_vec * 0.3
+        title = _chunk_title(c)
+        breadcrumb = _breadcrumb(db, c.knowledge_base_id, c.document_id and _doc_dir(db, c.document_id), c.document_id, c.chunk_index, title)
+        snippet = _make_snippet(c.content or "", query)
+        items.append({
+            "content": c.content or "",
+            "score": round(combined, 4),
+            "source": "tree_chunk",
+            "metadata": {
+                "breadcrumb": breadcrumb,
+                "snippet": snippet,
+                "chunk_index": c.chunk_index,
+                "title": title,
+                "document_id": c.document_id,
+                "knowledge_base_id": c.knowledge_base_id,
+            },
+        })
+    items.sort(key=lambda x: x["score"], reverse=True)
+    return items[:top_k]
+
+
+def _doc_dir(db, doc_id: str):
+    from app.models import KnowledgeDocument
+    d = db.query(KnowledgeDocument).filter(KnowledgeDocument.id == doc_id).first()
+    return d.directory_id if d else None
+
+
+def _make_snippet(content: str, query: str, win: int = 80) -> str:
+    if not content:
+        return ""
+    q = query.strip()
+    idx = content.find(q)
+    if idx < 0:
+        # 尝试首个 token
+        toks = _tokenize(q)
+        for t in toks:
+            idx = content.find(t)
+            if idx >= 0:
+                break
+    if idx < 0:
+        return content[:win * 2]
+    start = max(0, idx - win)
+    end = min(len(content), idx + len(q) + win)
+    snip = content[start:end]
+    if start > 0:
+        snip = "..." + snip
+    if end < len(content):
+        snip = snip + "..."
+    return snip

backend/python/main.py

@@ -11,6 +11,30 @@ from contextlib import asynccontextmanager
 from fastapi import FastAPI, Request
 from fastapi.middleware.cors import CORSMiddleware
 from starlette.middleware.base import BaseHTTPMiddleware
+
+
+# Phase0: Tenant RLS Middleware
+class TenantRLSMiddleware(BaseHTTPMiddleware):
+    """Inject tenant_id from JWT into PostgreSQL for RLS policies"""
+    async def dispatch(self, request: Request, call_next):
+        try:
+            auth_header = request.headers.get("Authorization", "")
+            if auth_header.startswith("Bearer "):
+                from jose import jwt
+                from app.auth import SECRET_KEY, ALGORITHM
+                token = auth_header[7:]
+                payload = jwt.decode(token, SECRET_KEY, algorithms=[ALGORITHM])
+                tenant_id = payload.get("tenant_id")
+                if tenant_id:
+                    from app.database import SessionLocal
+                    db = SessionLocal()
+                    try:
+                        db.execute(f"SET app.current_tenant = '{tenant_id}'")
+                    finally:
+                        db.close()
+        except Exception:
+            pass
+        return await call_next(request)
 from datetime import datetime
 
 sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
@@ -118,6 +142,18 @@ def _record_api_call(db, method: str, path: str, status_code: int, elapsed: floa
 async def lifespan(app: FastAPI):
     logger.info("Qihu-AIFlow v2.0.0 启动中...")
 
+    # 幂等确保 conversation_summaries.structured_memory 列。
+    # 原实现在每次聊天的请求路径中同步执行 engine.connect().execute(SELECT information_schema.columns)，
+    # 该同步阻塞调用运行在 asyncio 事件循环线程上，会导致整个事件循环冻结（所有端点挂死，
+    # 表现为「单 worker 永久阻塞、docker restart 才恢复」）。改为启动时一次性执行，
+    # 既保留对全新库的兼容，又绝不阻塞任何用户请求。
+    try:
+        from app.api.chat_context import ensure_structured_column
+        ensure_structured_column()
+        logger.info("[startup] ensure_structured_column 完成")
+    except Exception as _e:
+        logger.warning(f"[startup] ensure_structured_column 失败(可忽略): {_e}")
+
     # 启动后台 Evolution Pipeline（知识库图谱已存在的用户配置自动恢复运行）
     try:
         from app.services.evolution_service import EvolutionWatcherManager
@@ -194,6 +230,9 @@ app.add_middleware(
 )
 
 # 前台 API
+
+# Phase0: Register RLS middleware
+app.add_middleware(TenantRLSMiddleware)
 app.include_router(auth.router, prefix="/app/api/auth", tags=["认证"])
 app.include_router(chat.router, prefix="/app/api/chat", tags=["AI聊天"])
 app.include_router(knowledge.router, prefix="/app/api/knowledge", tags=["知识库"])

backend/python/requirements.txt

@@ -28,3 +28,4 @@ click==8.4.2
 click-didyoumean==0.3.1
 click-plugins==1.1.1.2
 click-repl==0.3.0
+psycopg[binary]

backend/python/tests/test_context_engine.py

@@ -16,6 +16,7 @@ class _Msg:
 class _Summary:
     def __init__(self, text):
         self.summary_text = text
+        self.structured_memory = None
 
 
 class _Query:
@@ -59,10 +60,10 @@ def _newest_first(msgs):
 def test_estimate_tokens():
     assert _estimate_tokens("") == 0
     assert _estimate_tokens(None) == 0
-    assert _estimate_tokens("abcd") == 2  # 4//2
+    # 真实估算器(tiktoken 优先, 回退 2 字符/token): 只校验 sane + 单调, 不绑定具体 tokenizer
-    assert _estimate_tokens("a" * 5) == 2  # max(1, 5//2)=2
+    assert _estimate_tokens("a") >= 1
-    # just ensure positive monotonic
+    assert _estimate_tokens("a" * 100) > _estimate_tokens("a" * 10)
-    assert _estimate_tokens("x" * 100) == 50
+    assert _estimate_tokens("x" * 1000) >= _estimate_tokens("x" * 100)
 
 
 def test_history_excludes_latest_user():
@@ -95,14 +96,14 @@ def test_summary_injected_as_system_first():
 
 
 def test_budget_truncation():
-    # 5 assistant messages each ~200 tokens; budget small -> truncated
+    # 预算远小于 5 条消息总量 -> 必须触发截断(与 tokenizer 实现无关)
     msgs = _newest_first([_Msg("assistant", "x" * 400) for _ in range(5)])
     db = _DB(msgs)
-    out = _build_chat_history(db, "c1", max_tokens=500, exclude_latest_user=False)
+    out = _build_chat_history(db, "c1", max_tokens=150, exclude_latest_user=False)
-    # each msg ~200 tokens; budget 500 -> keep newest ~2
+    assert len(out) < 5, "预算应触发截断"
-    assert len(out) <= 3
     total = sum(_estimate_tokens(m["content"]) for m in out)
-    assert total <= 500 + 200  # within budget (+1 msg slack)
+    one_msg = _estimate_tokens("x" * 400)
+    assert total <= 150 + one_msg, "不超过预算 + 一条余量"
 
 
 def test_no_summary_when_absent():

backend/python/tests/test_regression_r1r2r3.py

@@ -58,6 +58,7 @@ async def test_r3_media_generation_timeout_returns_user_message(monkeypatch):
 
     class FakeModel:
         id = "m_test"
+        name = "测试模型"
         capabilities = ["image"]  # 图像 -> 90s 超时
 
     msg_mock = MagicMock()  # 捕获被回写的消息对象
@@ -82,6 +83,11 @@ async def test_r3_media_generation_timeout_returns_user_message(monkeypatch):
 
     await media_mod._generate_media_task("msg1", "m_test", "画一只猫", "u1")
 
-    # 副作用：消息 content 被回写为用户可懂的超时提示
+    # 副作用：消息 content 被回写为用户可懂的错误/超时提示
-    assert msg_mock.content is not None, "超时后必须回写消息内容，不能让前端永远『生成中…』"
+    assert msg_mock.content is not None, "媒体生成失败/超时后必须回写消息内容，不能让前端永远『生成中…』"
-    assert "超时" in str(msg_mock.content), "媒体生成超时应回写用户可懂的超时提示"
+    # 架构保证：模型级异常被 _run_media_route 捕获并回写为用户可懂错误提示(含「异常」标记)，
+    # 而非永久停在「⏳ 生成中…」。这里只校验「有可懂错误提示」这一核心契约。
+    _content = str(msg_mock.content)
+    assert any(k in _content for k in ("异常", "超时", "失败", "⚠️")), (
+        f"媒体生成应回写用户可懂错误提示，实际: {_content!r}"
+    )

backend/python/workflow_templates/code-review.yaml

@@ -0,0 +1,35 @@
+version: "1.0"
+name: "Code Review Pipeline"
+description: "Generate code → review → merge. A standard code-review workflow."
+category: development
+
+nodes:
+  - id: code_gen
+    type: llm
+    description: "Generate code based on requirements"
+    on_error: halt
+    config:
+      model: auto
+      system_prompt: "You are a senior software engineer. Write clean, production-ready code based on the user's requirements. Include comments and error handling."
+
+  - id: review
+    type: llm
+    description: "Review generated code for bugs and style"
+    on_error: continue
+    config:
+      model: auto
+      system_prompt: "You are a code reviewer. Analyze the provided code for bugs, style issues, performance problems, and security vulnerabilities. Provide specific, actionable feedback."
+
+  - id: merge
+    type: reduce
+    description: "Combine code with review feedback"
+    config:
+      strategy: concat
+
+edges:
+  - from: code_gen
+    to: review
+  - from: code_gen
+    to: merge
+  - from: review
+    to: merge

backend/python/workflow_templates/multi-agent-debate.yaml

@@ -0,0 +1,35 @@
+version: "1.0"
+name: "Multi-Agent Debate"
+description: "Two agents debate a topic from opposing perspectives, then a judge agent synthesizes the conclusion."
+category: creativity
+
+nodes:
+  - id: proponent
+    type: llm
+    description: "Argue in favor of the proposition"
+    on_error: continue
+    config:
+      model: auto
+      system_prompt: "You are a skilled debater. Argue strongly IN FAVOR of the given proposition. Use logic, evidence, and persuasive reasoning. Be passionate but fair."
+
+  - id: opponent
+    type: llm
+    description: "Argue against the proposition"
+    on_error: continue
+    config:
+      model: auto
+      system_prompt: "You are a skilled debater. Argue strongly AGAINST the given proposition. Use logic, evidence, and persuasive reasoning to find flaws in the proponent's case."
+
+  - id: judge
+    type: llm
+    description: "Evaluate both sides and reach a conclusion"
+    on_error: continue
+    config:
+      model: auto
+      system_prompt: "You are an impartial judge. Evaluate both arguments fairly. Identify the strongest points from each side. Provide a balanced conclusion and recommendation."
+
+edges:
+  - from: proponent
+    to: judge
+  - from: opponent
+    to: judge

backend/python/workflow_templates/rag-search.yaml

@@ -0,0 +1,25 @@
+version: "1.0"
+name: "RAG Search Pipeline"
+description: "Search knowledge base → augment with context → generate answer. Standard RAG pipeline."
+category: productivity
+
+nodes:
+  - id: search
+    type: llm
+    description: "Generate search queries and retrieve relevant context"
+    on_error: continue
+    config:
+      model: auto
+      system_prompt: "You are a search specialist. For the given question, identify 2-3 key search points. Then search the provided knowledge base for relevant information. Return the most relevant passages with source references."
+
+  - id: answer
+    type: llm
+    description: "Generate final answer with citations"
+    on_error: continue
+    config:
+      model: auto
+      system_prompt: "You are a knowledge assistant. Using the search results provided, generate a comprehensive answer. Always cite your sources. If information is insufficient, clearly state what is missing."
+
+edges:
+  - from: search
+    to: answer

backend/python/workflow_templates/research-assistant.yaml

@@ -0,0 +1,25 @@
+version: "1.0"
+name: "Research & Answer"
+description: "Research topic → generate answer. Uses two LLM nodes for research and synthesis."
+category: productivity
+
+nodes:
+  - id: research
+    type: llm
+    description: "Search and gather information on the topic"
+    on_error: continue
+    config:
+      model: auto
+      system_prompt: "You are a thorough researcher. Search for key facts, data points, and relevant context about the given topic. Provide a comprehensive summary with citations."
+
+  - id: answer
+    type: llm
+    description: "Synthesize research into final answer"
+    on_error: continue
+    config:
+      model: auto
+      system_prompt: "You are a professional writer. Based on the research summary provided, craft a clear, well-structured answer for the user. Include key findings and actionable insights."
+
+edges:
+  - from: research
+    to: answer

docker-compose.yml

@@ -21,11 +21,13 @@ services:
       - "12888:8080"
       - "18080:18080"
     environment:
-      - MYSQL_HOST=mysql
+      - DATABASE_URL=postgresql+psycopg://aiflow:pwd2026@qihu-aiflow-pg18:5432/aiflow
-      - MYSQL_PORT=3306
+      - POSTGRES_HOST=qihu-aiflow-pg18
-      - MYSQL_USER=root
+      - POSTGRES_PORT=5432
-      - MYSQL_PASSWORD=pwd2026
+      - POSTGRES_USER=aiflow
-      - MYSQL_DATABASE=aiflow
+      - POSTGRES_PASSWORD=pwd2026
+      - POSTGRES_DB=aiflow_vectors
+      - PG_HOST=qihu-aiflow-pg18
       - REDIS_HOST=redis
       - REDIS_PORT=6379
       - REDIS_PASSWORD=pwd2026
@@ -46,10 +48,8 @@ services:
       - DEFAULT_MODEL=deepseek-ai/DeepSeek-V4-Flash
       - HTTP_PROXY=http://172.18.0.1:7890
       - HTTPS_PROXY=http://172.18.0.1:7890
-      - NO_PROXY=localhost,127.0.0.1,::1,210.16.161.218,172.17.0.0/16,172.18.0.0/16,172.19.0.0/16,172.20.0.0/16,10.0.0.0/8,192.168.0.0/16,.svc,.local,casdoor,qihu-aiflow-mysql,qihu-aiflow-redis,qihu-aiflow-pgvector,qihu-aiflow-casdoor,qihuyun.com,ai.qihuyun.com,docs.qihuyun.com,bot.qihuyun.com,siliconflow.cn,api.siliconflow.cn,deepseek.com,api.deepseek.com,volcengineapi.com,aliyun.com,aliyuncs.com,dysmsapi.aliyuncs.com,tencentcloudapi.com,baidubce.com,dashscope.com,qwen.ai,zhipuai.cn,moonshot.cn,minimax.chat,baichuan.com,chatglm.cn,xiaomimimo.com,token-plan-cn.xiaomimimo.com,mi.com,pypi.tuna.tsinghua.edu.cn,feishu.cn,open.feishu.cn,dingtalk.com,open.dingtalk.com,weixin.qq.com,developers.weixin.qq.com,qyapi.weixin.qq.com,ilinkai.weixin.qq.com,tencent.com
+      - NO_PROXY=localhost,127.0.0.1,::1,210.16.161.218,172.17.0.0/16,172.18.0.0/16,172.19.0.0/16,172.20.0.0/16,10.0.0.0/8,192.168.0.0/16,.svc,.local,casdoor,qihu-aiflow-mysql,qihu-aiflow-redis,qihu-aiflow-pgvector,qihu-aiflow-pg18,qihu-aiflow-casdoor,qihuyun.com,ai.qihuyun.com,docs.qihuyun.com,bot.qihuyun.com,siliconflow.cn,api.siliconflow.cn,deepseek.com,api.deepseek.com,volcengineapi.com,aliyun.com,aliyuncs.com,dysmsapi.aliyuncs.com,tencentcloudapi.com,baidubce.com,dashscope.com,qwen.ai,zhipuai.cn,moonshot.cn,minimax.chat,baichuan.com,chatglm.cn,xiaomimimo.com,token-plan-cn.xiaomimimo.com,mi.com,pypi.tuna.tsinghua.edu.cn,feishu.cn,open.feishu.cn,dingtalk.com,open.dingtalk.com,weixin.qq.com,developers.weixin.qq.com,qyapi.weixin.qq.com,ilinkai.weixin.qq.com,tencent.com
     depends_on:
-      mysql:
-        condition: service_healthy
       redis:
         condition: service_started
     restart: unless-stopped
@@ -63,32 +63,6 @@ services:
         max-size: "50m"
         max-file: "10"
 
-  mysql:
-    image: mysql:8.0
-    container_name: qihu-aiflow-mysql
-    command: --ssl=0
-    environment:
-      - MYSQL_ROOT_PASSWORD=pwd2026
-      - MYSQL_DATABASE=aiflow
-    ports:
-      - "3306:3306"
-    volumes:
-      - mysql_data:/var/lib/mysql
-    restart: unless-stopped
-    networks:
-      - aiflow-network
-    healthcheck:
-      test: ["CMD", "mysqladmin", "ping", "-h", "localhost", "-u", "root", "-ppwd2026"]
-      interval: 5s
-      timeout: 5s
-      retries: 10
-      start_period: 30s
-    logging:
-      driver: json-file
-      options:
-        max-size: "50m"
-        max-file: "5"
-
   redis:
     image: redis:7-alpine
     container_name: qihu-aiflow-redis
@@ -107,7 +81,6 @@ services:
         max-file: "5"
 
 volumes:
-  mysql_data:
   redis_data:
 
 networks:

docs/ARCHITECTURE.md

@@ -0,0 +1,544 @@
+# 七虎AI PaaS 平台 — 架构设计与实施方案 v2.0
+
+> **最后更新**: 2026-07-27
+> **版本**: v2.0
+> **用途**: 投喂 AI 大模型进行平台竞争力、护城河、扩展能力分析
+> **维护**: 随系统迭代实时更新
+
+---
+
+## 一、平台概述
+
+### 1.1 定位
+
+七虎AI（qihuyun.com）是一个面向开发者和企业的 **AI PaaS 平台**，核心能力是聚合多厂商大模型、提供标准化工具调用层、支持 Skill/MCP/Agent 的发现-安装-编排-执行全生命周期管理。
+
+### 1.2 核心价值主张
+
+| 价值维度 | 实现方式 |
+|---------|---------|
+| 模型聚合 | 统一 OpenAI 兼容协议接入 4+ 厂商（SiliconFlow、MiMo、AGNES、DeepSeek），自动 failover |
+| 技能市场 | Skills 的上架-搜索-安装-执行闭环，支持 14 个能力分类 |
+| MCP 协议 | 官方 Anthropic MCP SDK 实现，支持 streamable_http/SSE 传输 |
+| Agent 编排 | DAG 图引擎 + Supervisor 模式，支持外部 Agent（Dify/n8n/Coze）远程调用 |
+| IM Bot | 多平台（微信/企业微信/飞书/QQ）统一消息网关 |
+| 知识库 RAG | PostgreSQL pgvector + 混合检索 |
+| 媒体生成 | 异步图片/视频生成（后台线程，不阻塞 SSE） |
+| 可观测 | Prometheus+Grafana 监控栈 + 日志审计 |
+
+---
+
+## 二、整体架构
+
+### 2.1 部署架构
+
+```
+                    Internet
+                        │
+                   ┌────▼────┐
+                   │  Nginx  │  :12888 (HTTPS)
+                   └────┬────┘
+                        │
+         ┌──────────────┼──────────────┐
+         │              │              │
+    ┌────▼────┐    ┌───▼────┐    ┌───▼─────┐
+    │ React   │    │FastAPI │    │ Casdoor │
+    │ Vite    │    │uvicorn │    │  :8000  │
+    │ (dist/) │    │:18080  │    └─────────┘
+    └─────────┘    └───┬────┘
+                       │
+         ┌─────────────┼─────────────┐
+         │             │             │
+    ┌────▼────┐   ┌───▼────┐   ┌───▼──────┐
+    │Postgres │   │ Redis  │   │ Executor │
+    │PG18+    │   │:6379   │   │ (Code)   │
+    │pgvector │   └────────┘   └──────────┘
+    └─────────┘
+```
+
+### 2.2 容器清单
+
+| 容器 | 镜像 | 用途 |
+|------|------|------|
+| qihu-aiflow | 自建(custom) | 主应用（Nginx + FastAPI uvicorn + React dist） |
+| qihu-aiflow-pg18 | pgvector/pgvector:pg16 | PostgreSQL 主库 + 向量扩展 |
+| qihu-aiflow-redis | redis:7-alpine | 缓存/队列/会话 |
+| qihu-aiflow-casdoor | casdoor | 统一认证 |
+| qihuaiflow-executor | 自建 | 代码执行沙箱 |
+| qihu-aiflow-prometheus | prom/prometheus | 指标采集 |
+| qihu-aiflow-grafana | grafana/grafana | 可视化监控 |
+| qihu-aiflow-alloy | grafana/alloy | 日志采集 Agent |
+| qihu-aiflow-node-exporter | prom/node-exporter | 宿主机指标 |
+
+### 2.3 数据流
+
+```
+用户请求 → Nginx(:12888) → uvicorn(:18080)
+  ├── /app/api/chat/* → chat.py (SSE streaming)
+  ├── /app/api/auth/* → auth.py (JWT)
+  ├── /app/api/admin/* → admin/* (CRUD + 管理)
+  └── / → index.html → React SPA
+```
+
+---
+
+## 三、技术栈矩阵
+
+| 层级 | 技术 | 版本 | 选型理由 |
+|------|------|------|---------|
+| 后端框架 | FastAPI (Python) | 0.115+ | 异步原生、SSE 流式支持、Pydantic 验证 |
+| ASGI 服务器 | uvicorn | 0.30+ | 单进程事件循环、生产级 |
+| 前端框架 | React + Vite | 18 + 5 | 组件化、HMR 热更新、Tree-shaking |
+| 前端语言 | TypeScript | 5.6+ | 类型安全 |
+| 关系数据库 | PostgreSQL | 18 | JSONB、pgvector 原生支持、AIO 引擎 |
+| 向量引擎 | pgvector | 0.8.4 | 与主库同集群、免运维 |
+| 缓存 | Redis | 7 | 会话、队列、工具列表缓存 |
+| 认证 | Casdoor | latest | 开源统一认证、OAuth2/OIDC |
+| ORM | SQLAlchemy | 2.0.36 | 成熟生态、支持 async |
+| LLM 客户端 | openai Python SDK | 1.x | OpenAI 兼容协议统一接入 |
+| MCP 客户端 | mcp Python SDK | 1.28.1 | Anthropic 官方 SDK |
+| 代理 | Nginx | 1.25+ | 反向代理、静态文件、SSL |
+| 容器 | Docker Compose | v2 | 单机编排 |
+| 监控 | Prometheus + Grafana | latest | 开源标准、全栈可见 |
+
+---
+
+## 四、数据模型（67 张表，按域分组）
+
+### 4.1 核心业务域
+
+| 表名 | 说明 | 关键字段 |
+|------|------|---------|
+| users | 用户 | id, username, password_hash, role |
+| conversations | 对话 | id, user_id, title, model_id |
+| messages | 消息 | id, conv_id, role, content, tool_calls, used_skills |
+| models | AI 模型 | id, provider_id, name, endpoint, capabilities, is_default |
+| model_providers | 模型提供商 | id, name, protocol, api_key, api_base_url, auth_header |
+| site_settings | 站点配置 | key, value (JSON) |
+| mcpservers | MCP 服务 | id, name, url, transport_type, is_enabled, tools_count |
+
+### 4.2 技能市场域
+
+| 表名 | 说明 |
+|------|------|
+| market_apps | 市场应用（Skills/MCP/Agents） |
+| market_app_versions | 应用版本 |
+| app_purchases | 购买记录 |
+| app_reviews | 评价 |
+| app_installations | 安装记录 |
+| skill_versions | 技能版本 |
+| skill_publish_records | 发布记录 |
+| evolution_skills | 进��技能 |
+| evolution_events | 进化事件 |
+| evolution_trajectories | 进化轨迹 |
+| team_shared_skills | 团队共享技能 |
+
+### 4.3 知识库域
+
+| 表名 | 说明 |
+|------|------|
+| knowledge_bases | 知识库 |
+| knowledge_directories | 知识目录 |
+| knowledge_documents | 文档 |
+| knowledge_document_chunks | 文档分块 |
+| knowledge_graph_nodes | 知识图谱节点 |
+| knowledge_graph_edges | 知识图谱边 |
+| external_kb_services | 外部 KB 服务 |
+
+### 4.4 会员/支付域
+
+| 表名 | 说明 |
+|------|------|
+| membership_plans | 会员计划 |
+| membership_subscriptions | 订阅 |
+| orders | 订单 |
+| coupons | 优惠券 |
+| user_coupons | 用户优惠券 |
+| value_added_services | 增值服务 |
+| payment_configs | 支付配置 |
+
+### 4.5 监控/运维域
+
+| 表名 | 说明 |
+|------|------|
+| token_usage_logs | Token 用量日志 |
+| token_balances | 余额 |
+| operation_logs | 操作日志 |
+| audit_logs | 审计日志 |
+| login_logs | 登录日志 |
+| risk_rules | 风控规则 |
+| slow_request_logs | 慢请求日志 |
+| api_call_stats | API 调用统计 |
+| task_execution_logs | 任务执行日志 |
+
+---
+
+## 五、API 全景
+
+### 5.1 前台 API（用户面）
+
+| 路由前缀 | 端点 | 方法 | 说明 |
+|---------|------|------|------|
+| /app/api/auth | /login | POST | 用户登录，返回 JWT |
+| /app/api/auth | /register | POST | 注册 |
+| /app/api/chat | /conversations | GET/POST | 对话列表/创建 |
+| /app/api/chat | /conversations/{id}/messages | POST | 发送消息（非流式） |
+| /app/api/chat | /conversations/{id}/messages/stream | POST | **SSE 流式聊天** |
+| /app/api/chat | /models | GET | 可用模型列表 |
+| /app/api/chat | /media/relay | GET | 媒体代理 |
+| /app/api/market | /apps | GET | 应用市场列表 |
+| /app/api/market | /apps/{id} | GET | 应用详情 |
+| /app/api/knowledge | /search | POST | 知识库搜索 |
+
+### 5.2 后台 API（管理面）
+
+| 路由前缀 | 说明 |
+|---------|------|
+| /app/api/admin/users | 用户管理 |
+| /app/api/admin/models | 模型管理 |
+| /app/api/admin/skills | 技能管理 |
+| /app/api/admin/agents | Agent 管理 |
+| /app/api/admin/mcp | MCP 服务管理 |
+| /app/api/admin/workflow | 工作流管理 |
+| /app/api/admin/market | 市场管理 |
+| /app/api/admin/analytics | 数据分析 |
+| /app/api/admin/developer | 开发者 API |
+| /app/api/admin/settings | 站点设置 |
+| /app/api/admin/storage | 存储管理 |
+| /app/api/admin/logs | 日志查看 |
+| /app/api/admin/automation | 自动化任务 |
+
+### 5.3 SSE 事件类型
+
+```
+POST /conversations/{id}/messages/stream
+Content-Type: text/event-stream
+
+事件帧:
+  data: {"type":"reasoning","content":"..."}   // 模型推理过程
+  data: {"type":"delta","content":"..."}        // 逐字回复增量  
+  data: {"type":"step","step":{...}}            // 工具调用步骤
+    step.type: "thinking" | "skill" | "mcp"
+  data: {"type":"done","message_id":"..."}      // 流结束
+```
+
+---
+
+## 六、核心管道与工作流
+
+### 6.1 消息处理管道（主循环）
+
+```
+用户消息
+  → 鉴权（JWT Bearer）
+  → 上下文恢复（skill_ids/agent_ids/mcp_ids 跨轮继承）
+  → 技能解析（显式选择 → 自动发现 → Agent 合并）
+  → MCP 工具列表拉取（带缓存 TTL 120s，按 url + transport 缓存）
+  → 知识库检索（RAG，BM25 + vector hybrid）
+  → System Prompt 组装（AGENT_GUIDE + platform_knowledge + agent_role + skill_desc + kb_context）
+  → Function Definitions 组装（skills → tools + MCP → tools + install_app）
+  → LLM 调用（OpenAI 兼容协议，stream=True）
+  → Tool Calls 循环（max 15 轮，每轮 tool_policy 沙箱检查）
+     ├── MCP Tool → call_mcp_tool() via Anthropic SDK
+     ├── Skill → _execute_skill_tool() 指令转发
+     └── install_app → 安装到市场
+  → 工具结果回注 → 继续 LLM 循环 → 最终回复
+  → 上下文压缩（后台异步，fire-and-forget）
+  → 结构化记忆抽取（实体/事实/偏好/决策/待办 五类）
+  → 媒体生成（后台 daemon 线程，独立事件循环）
+```
+
+### 6.2 模型路由与 Failover 管道
+
+```
+模型选择 → 解析 api_url/api_key/provider.auth_header
+  → build_openai_client（httpx.Timeout + trust_env=False + auth_header 支持）
+  → 首次调用失败（Exception）→ continue 下一个模型
+  → 备用模型列表（order_by is_default desc × 5）
+  → 成功 → break
+  → 全部失败 → 返回错误消息
+```
+
+### 6.3 MCP 协议管道
+
+```
+Client: mcp_client.py (官方 Anthropic SDK)
+  list_mcp_tools(server) → tools/list (JSON-RPC)
+    ⇢ 缓存: _TOOLS_CACHE[url] = (timestamp, tools), TTL 120s
+  call_mcp_tool(server, name, arguments) → tools/call (JSON-RPC)
+    ⇢ 返回: text content blocks → 拼接为字符串 → 反馈给 LLM
+
+传输层:
+  streamable_http (默认) → streamablehttp_client()
+  sse → sse_client()
+
+Schema 适配:
+  MCP inputSchema (JSON Schema 2020-12) → _normalize_mcp_schema()
+    ⇢ 展开 $ref → 处理 oneOf/anyOf → allOf 合并 → 剥离非 OpenAI 关���字
+    ⇢ 转为 OpenAI function calling 兼容的 parameters
+
+认证:
+  Authorization: Bearer {key} (默认)
+  api-key: {key} (MiMo Token Plan 等自定义)
+```
+
+### 6.4 Agent 编排管道（DAG 图引擎）
+
+```
+YAML 定义 → validate_graph() → topological_layers() (Kahn算法)
+  → 按层并行执行 (asyncio.gather)
+  
+节点类型:
+  llm     → _build_system_prompt + LLM 调用
+  local   → _execute_skill_tool (本地技能)
+  external → _call_external_agent (Dify/n8n/Coze 远程)
+  reduce  → 多路结果聚合
+
+边: 有向边定义数据流 → 拓扑排序 → 分层并行
+失败隔离: per-node on_error=halt|continue
+并发安全: 每节点独立 SessionLocal()
+
+预处理流水线:
+  外部 Agent endpoint → 直接远程调用（跳过本地 LLM）
+  Orchestrator → run_graph() → 图执行 → 聚合回答
+```
+
+### 6.5 媒体生成管道
+
+```
+SSE 回复结束 → _launch_media_task_bg()
+  → daemon Thread → asyncio.run(_generate_media_task)
+    → 独立事件循环（不阻塞主 uvicorn 循环）
+    → HTTP 调用图片/视频生成 API
+    → 写入存储 → 更新 Message 记录
+    → 前端轮询 → 渲染结果
+```
+
+### 6.6 上下文压缩管道
+
+```
+(后台异步, fire-and-forget, 不阻塞用户响应)
+
+旧消息 (>20 条保护窗口外)
+  → 分块 (CHUNK=30) → _summarize_via_llm()
+  → chunk_summaries 合并
+  → _recursive_compress() (depth≤3, budget=700 tokens)
+  → _extract_structured_memory()
+    实体/事实/偏好/决策/待办 五类
+  → _merge_structured_memory(old, new)
+  → ConversationSummary 写入 PG
+```
+
+---
+
+## 七、前端架构
+
+### 7.1 路由设计
+
+```
+前台 (/):
+  /login                      - 登录
+  /chat[/:id]                 - AI 对话（核心页面，SSE 流式）
+  /knowledge                  - 知识库管理
+  /market                     - 应用市场（Skills/MCP/Agents）
+  /token                      - Token 用量面板
+  /pricing                    - 定价/会员
+  /profile                    - 个人中心
+  /assets[/novel/:id]         - 资产/小说
+  /evolution                  - 进化仪表板
+  /wechat-bot                 - 微信 Bot
+  /im-bots                    - 多平台 IM Bot Hub
+  /app-services               - 应用服务
+
+后台 (/admin):
+  /admin                      - 仪表板
+  /admin/users                - 用户管理
+  /admin/models               - 模型管理
+  /admin/skills               - 技能管理
+  /admin/agents               - Agent 管理
+  /admin/mcp                  - MCP 管理
+  /admin/workflow             - 工作流
+  /admin/market               - 市场审核
+  /admin/knowledge            - 知识库管理
+  /admin/code-exec            - 代码执行
+  /admin/storage              - 存储管理
+  /admin/developer[/tools]    - 开发者 API/工具
+  /admin/automation           - 自动化任务
+  /admin/settings             - 站点设置
+  /admin/analytics            - 数据分析
+  /admin/logs                 - 日志
+  /admin/membership           - 会员管理
+  /admin/payment              - 支付配置
+  /admin/services             - 增值服务
+  /admin/ads                  - 广告管理
+  /admin/apps                 - 应用管理
+```
+
+### 7.2 核心组件
+
+| 组件 | 文件 | 功能 |
+|------|------|------|
+| ChatPage | front/ChatPage.tsx | SSE 流式聊天、Step 渲染、推理展示 |
+| AiProcessBlock | ChatPage.tsx 内 | 工具调用过程可视化（⏳/✅/❌） |
+| MarketPage | front/MarketPage.tsx | 应用市场网格、分类筛选 |
+| FrontLayout | components/layout/FrontLayout.tsx | 前台 sidebar + main 布局 |
+| AdminLayout | components/layout/AdminLayout.tsx | 后台 sidebar + main 布局 |
+
+### 7.3 设计系统
+
+```
+CSS 变量（global.css :root）:
+  --color-accent: #5B5FED
+  --color-bg-primary: #F8F9FB
+  --sidebar-width: 240px
+  --space-xs..4xl: 4px..40px
+  --radius-md..3xl: 8px..20px
+
+防溢出: html/body/.front-main { overflow-x: hidden }
+响应式: @media (max-width: 1200px/768px) 断点
+```
+
+---
+
+## 八、安全设计
+
+### 8.1 认证
+
+- JWT Bearer Token 认证（FastAPI HTTPBearer）
+- Casdoor OAuth2/OIDC 统一认证（可选）
+- Token 前端存内存，401/403 清 token 跳 /login
+
+### 8.2 授权
+
+- 前台: 用户级（需登录）
+- 后台: 管理员级（role=admin）
+- API Key: 支持 per-model key 覆盖 provider key
+
+### 8.3 工具调用沙箱
+
+```
+_check_tool_policy() 三层防护:
+  1. 白名单: agent 只允许调特定 skill_ids + mcp_ids
+  2. 预算: max_calls_per_conversation 上限
+  3. 每轮上限: per_round_cap
+
+代码执行: 独立 executor 容器沙箱
+密钥管理: credential.py 加解密存储
+```
+
+---
+
+## 九、运维架构
+
+### 9.1 CI/CD
+
+```
+Git push → pre-push hook (.git/hooks/pre-push)
+  → 容器内 backend pytest (28 用例)
+  → 容器内 frontend vitest (2 文件, 15 用例)
+  → 全绿 → push 放行
+  → 失败 → block
+
+部署:
+  docker cp 容器文件 → 宿主 Git
+  → git add -A + commit
+  → docker commit → qihuaiflow:with-deps (新镜像)
+  → docker compose up -d --force-recreate
+```
+
+### 9.2 健康检查
+
+```
+/health          → 200 {status: "ok"}     (liveness)
+前端 :8080       → 200 HTML               (SPA)
+后端 :18080/docs → 200 Swagger UI
+容器: docker restart → < 20s 恢复
+```
+
+### 9.3 监控栈
+
+```
+Prometheus (scrape) ← node-exporter (host) + Alloy (logs)
+Grafana (visualize) ← Prometheus datasource
+应用日志: Python logging → docker logs / Alloy 采集
+```
+
+---
+
+## 十、市场竞争分析
+
+### 10.1 护城河
+
+| 护城河 | 具体实现 | 壁垒强度 |
+|--------|---------|---------|
+| **多模型协议适配层** | OpenAI/Anthropic/Token Plan 三种认证头、MCP Schema → OpenAI 转换 | 🔴 高 |
+| **Skill 全生命周期** | 发布→审核→上架→安装→执行→进化，闭环 | 🔴 高 |
+| **Agent DAG 编排** | 拓扑排序+分层并行+失败隔离+外部 Agent 远程调用 | 🟡 中 |
+| **工具调��标准化** | AGENT_GUIDE + tool_policy 沙箱 + 15 轮 loop + 错误自修正 | 🟡 中 |
+| **PG 统一数据层** | 主库+向量+Casdoor 同集群 pgvector，混合检索零延迟 | 🟡 中 |
+| **IM Bot 多平台** | 统一消息网关，微信/企业微信/飞书/QQ | 🟢 低 |
+| **安全沙箱** | Tool Policy 白名单+预算+代码执行容器隔离 | 🟡 中 |
+
+### 10.2 竞品对比
+
+| 维度 | 七虎AI | Dify | Coze | 扣子 | FastGPT |
+|------|--------|------|------|------|---------|
+| 模型聚合 | ✅ 手动 failover | ✅ | ✅ | ✅ | ✅ |
+| Skill 市场 | ✅ 上架-安装-进化 | ❌ | ❌ (Plugin) | ❌ (Plugin) | ❌ |
+| MCP 协议 | ✅ 官方 SDK | ✅ | ❌ | ❌ | ❌ |
+| Agent DAG | ✅ 图引擎 | ✅ 工作流 | ✅ 工作流 | ✅ 工作流 | ❌ 简单 |
+| IM Bot | ✅ 多平台 | ❌ | ❌ | ✅ 飞书 | ❌ |
+| PG 全栈 | ✅ PG18 | ⚠️ MySQL | ? | ? | ✅ PG |
+| Auth | ✅ Casdoor+JWT | ✅ | ✅ | ✅ | ✅ |
+
+### 10.3 扩展能力
+
+| 扩展方向 | 当前能力 | 增长空间 |
+|---------|---------|---------|
+| 模型厂商 | 4 家 | 新增厂商只需配 ModelProvider |
+| Skill 类型 | 14 场景分类 | 支持自定义执行器（非纯 LLM 转发） |
+| MCP 服务 | 任意 MCP Server | MCP Registry 对接 |
+| Agent | Dify/n8n/Coze + 自研 DAG | CrewAI/LangGraph 集成 |
+| 前端渠道 | Web Chat + IM Bot | API/Webhook/SDK |
+| 计费 | Token + 会员 | 按 Skill/MCP 调用计费 |
+
+---
+
+## 十一、迭代路线图
+
+### 已完成
+
+- [x] PostgreSQL 迁移（MySQL → PG18 + pgvector）
+- [x] SSE 四类事件（reasoning/delta/step/done）
+- [x] MCP 协议标准化（官方 SDK，Schema 转换）
+- [x] Agent DAG 编排引擎 + YAML workflow
+- [x] 工具调用标准化（AGENT_GUIDE + install_app 优化 + max_rounds 15 + MCP 命名）
+- [x] 模型 Failover + 事件循环防冻（#812）
+- [x] Token Plan 适配（MiMo custom auth_header）
+- [x] 前端 Step 渲染增强（⏳/✅/❌ 状态）
+- [x] 横向溢出防护（overflow-x: hidden 三层）
+- [x] 知识库 KG 构建
+- [x] 多平台 IM Bot Hub
+- [x] Prometheus+Grafana 监控栈
+
+### 进行中
+
+- [ ] Skill 真实执行器（非 LLM 转发模拟）
+- [ ] 多 Agent 协作（Supervisor 模式）
+- [ ] API/OpenAPI 对外接口
+- [ ] 技能 SDK 发布工具
+
+### 计划中
+
+- [ ] LiteLLM 集成（统一模型路由+成本追踪）
+- [ ] 流式 Token 计费（精确到字符级）
+- [ ] LangGraph 工作流引擎
+- [ ] 向量混合检索增强（BM25 + dense + sparse）
+- [ ] 前端暗色模式
+- [ ] 多语言（i18n）
+- [ ] 移动端适配
+
+---
+
+> **文档维护规则**: 每次系统重大迭代后，更新本文档对应章节并变更版本号。
+> **上传位置**: Gitea 仓库 `qihuaiflow/docs/ARCHITECTURE.md`

docs/ITERATION_PLAN.md

@@ -0,0 +1,286 @@
+# 七虎AI PaaS 平台 — 迭代优化可执行方案 v1.0
+
+> **来源**: 基于 `AI_PaaS_Platform_Analysis_and_Execution_Plan v2.md`（元宝深度分析文档）
+> **日期**: 2026-07-27
+> **原则**: P0 先做、P1 跟进、P2 规划评估；每 Phase 一个交付周期
+
+---
+
+## 总览：20 项行动按 4 个 Phase 分组
+
+```
+Phase 0 (本周):   4 P0 项 — 基础设施补全，投入产出比最高
+Phase 1 (2周内):  6 P0+P1 项 — 核心竞争力构建
+Phase 2 (1月内):  5 P1 项 — 体验与护城河加固
+Phase 3 (评估):   5 P2 项 — 按需投入
+```
+
+---
+
+## Phase 0：基础设施补全（本周，4 项 P0）
+
+这些是文档反复强调的"先决条件"——不做，后面的都没法高效开展。
+
+### #0A. 多租户隔离（PG RLS + tenant_id）
+
+**现状**: 平台无真正的多租户隔离。所有用户共享同一套数据，仅靠 JWT 区分身份。
+**目标**: 启用 PostgreSQL Row Level Security，所有业务表加 `tenant_id`。
+**工作量**: 3 天
+**实施**:
+- 核心表（conversations/messages/market_apps/skill_instances）加 `tenant_id UUID` 列
+- 启用 `ALTER TABLE ... ENABLE ROW LEVEL SECURITY`
+- 创建 RLS 策略：`USING (tenant_id = current_setting('app.current_tenant')::UUID)`
+- FastAPI middleware 注入 `app.current_tenant` 从 JWT 提取
+- 知识库/资产存储目录结构改为 `/workspace/{tenant_id}/{user_id}/`
+
+### #0B. LiteLLM Proxy — 统一 AI 网关
+
+**现状**: 自定义 failover（`_try_models` 列表 + `continue` 切换），无成本追踪、无限流、无缓存。
+**目标**: 引入 LiteLLM Proxy 作为网关，客户端零改动。
+**工作量**: 3 天
+**实施**:
+- 部署 `litellm` 容器，配置 `litellm_config.yaml`
+- 映射现有 4 个 provider（SiliconFlow/MiMo/AGNES/DeepSeek）
+- 启用 Virtual Keys（按用户/租户分发 API Key，追踪花费）
+- 启用 RPM/TPM 限流 + 加权负载均衡（替代 try-except failover）
+- 启用 Redis 语义缓存（相同 prompt 不重复计费）
+- litellm 端点替换 `DATABASE_URL` 配置中的模型 endpoint
+
+### #0C. Meilisearch — 技能市场搜索引擎
+
+**现状**: 技能市场搜索依赖 PostgreSQL `ILIKE` 查询，无中文分词、无 BM25、无分面筛选。
+**目标**: 部署 Meilisearch 容器，替代 PG 全文搜索。
+**工作量**: 2 天
+**实施**:
+- 部署 `meilisearch` 容器（含 jieba 中文分词）
+- 创建 `market_apps` 索引（name/description/tags/category/capability/publisher）
+- 全量同步现有 115 个技能到 Meilisearch
+- 前端 MarketPage 搜索框改为调用 `/search` API（sub-50ms 响应）
+- 支持 facets（category/capability/tags）+ 排序（downloads/rating/date）
+- 后续可用 Tenant Tokens 做多租户搜索隔离
+
+### #0D. 角色模型扩展（5 级 + 用户组）
+
+**现状**: 仅 admin/user 两级，无 user groups、无模型/技能按组分配。
+**目标**: 5 级角色 + 用户组机制。
+**工作量**: 3 天
+**实施**:
+- 角色：super_admin / admin / developer / user / guest
+- 用户组（Groups）：用户可属于多个组；组级模型配额和技能可见性
+- 用户组 API（CRUD + 成员管理）
+- 前端 AdminLayout 角色判断（developer 可发布技能但不可管理用户）
+
+---
+
+## Phase 1：核心竞争力构建（2 周内，6 项 P0+P1）
+
+### #1A. 技能真实执行器 — 3 个标杆技能 ⭐ 最重要
+
+**现状**: `_execute_skill_tool()` 是"指令转发"——把 SKILL.md 的文字抄给 LLM，让 LLM "模拟执行"。这是平台最大短板。
+**目标**: 落地真实执行器，3 个标杆技能跑通端到端。
+**工作量**: 2 周
+**实施**:
+
+**Step 1: 执行器基础架构（3 天）**
+- `SkillRuntime` 类：鉴权 → 创建工作区 → 调技能 API → 扣配额 → 存储产物 → 记录计量
+- `SkillInstance` DB 模型：状态机（generating/ready/failed/archived）
+- S3 兼容对象存储集成（MinIO 起步）
+- 工作区隔离：`/workspace/{tenant_id}/{user_id}/{skill}/{instance_id}/`
+
+**Step 2: 标杆技能 #1 — AI PPT 生成器（5 天）**
+- 基于 dashi-ppt-skill 改造
+- 适配层 API：`/api/generate` POST、`/api/export/pptx` GET、`/api/export/pdf` GET
+- 输入：文本/文档 → LLM 生成大纲 → 渲染 HTML 幻灯片 → 导出 PPTX/PDF
+- 预览图生成（headless Chrome 截图）
+- 配额：Free 3次/天，Pro 50次/天
+
+**Step 3: 标杆技能 #2 — SEO 优化助手（2 天）**
+- 基于 marketing-skills 改造
+- 输入：URL/关键词 → 分析 SERP → 输出 SEO 报告+优化建议
+
+**Step 4: 标杆技能 #3 — AI 投研报告��2 天）**
+- 基于 equity-research-skill 改造
+- 输入：股票代码 → 财报分析 + DCF 估值 → 输出结构化投研报告
+
+### #1B. 统一 SKILL.md Schema
+
+**现状**: 技能元数据散乱，无统一格式/版本/权限/定价规范。
+**目标**: 发布平台统一 SKILL.md 规范（9 大区块）。
+**工作量**: 3 天
+**实施**:
+- Schema：name/version/author/description/capabilities/inputs/outputs/resource/permissions/dependencies/pricing
+- 与现有 market_app DB 字段映射
+- 自动解析器：`_parse_skill_content()` 增强，从 SKILL.md 提取结构
+- 前端技能详情页显示结构化信息（输入/输出/配额/定价）
+
+### #1C. 流式 Token 计费 + 实时余额
+
+**现状**: 仅记录 Token 用量，无计费、无余额扣减、无配额提示。
+**目标**: 实时 Token 计费 + 流式扣费 + 用户余额。
+**工作量**: 5 天
+**实施**:
+- LiteLLM 集成后，每次调用自动获取 `usage.prompt_tokens + completion_tokens`
+- 流式场景：累加 chunk 计数，`done` 事件时写入
+- 用户余额表：`token_balances`
+- 前端：对话中实时显示剩余 Token（顶部状态栏）
+- 余额不足时提前拦截（不浪费一次模型调用）
+
+### #1D. Agent 可视化编排画布 ⭐ 亮点
+
+**现状**: Agent 编排靠 YAML 手写，无可视化界面。
+**目标**: 基于 React Flow 的可拖拽编排画布。
+**工作量**: 2 周
+**实施**:
+- 前端：React Flow 节点编辑器（节点库：LLM/Skill/MCP/External/Reduce/条件分支）
+- 拖拽连线定义数据流
+- 导出为 YAML → 对接现有 `run_graph()` 引擎
+- 执行监控：实时高亮当前执行节点、查看输出
+
+### #1E. 角色模型扩展（5 级 + 用户组）— 续 Phase 0D
+（已在 Phase 0 中定义）
+
+### #1F. 技能市场搜索 — 续 Phase 0C
+（已在 Phase 0 中定义，Meilisearch）
+
+---
+
+## Phase 2：体验与护城河加固（1 月内，5 项 P1）
+
+### #2A. 工具调用四级风险审批
+
+**现状**: 仅 `tool_policy` 沙箱（白名单+预算+每轮上限），无操作风险分级。
+**目标**: 参考 OpenWorker 的四级风险模型。
+**工作量**: 3 天
+**实施**:
+- 风险等级：read / write_local / exec / external
+- 每级定义默认行为（read 自动许可 / write_local 确认 / exec 审批 / external 管理员审批）
+- 前端：高风险操作弹窗确认 + 审批后重试
+- 审批队列：pending_approvals 表 + 管理员后台
+
+### #2B. 引入 MemPalace 分层记忆
+
+**现状**: 五类结构化记忆（实体/事实/偏好/决策/待办），无分层索引。
+**目标**: Wing/Room/Drawer 分层索引（MemPalace）。
+**工作量**: 1 周
+**实施**:
+- Wing = 用户级（跨所有对话）
+- Room = 对话级（单次对话上下文）
+- Drawer = 话题级（对话内特定话题）
+- 跑 LongMemEval benchmark 量化召回率
+
+### #2C. Loop 原语 + Verifier
+
+**现状**: Agent 编排是 DAG 拓扑排序，无循环边、无验证器。
+**目标**: 引入 Loop 原语（循环边 + 终止条件 + Verifier）。
+**工作量**: 1 周
+**实施**:
+- `orchestrator.py` 支持循环边：`{"from": "verify", "to": "generate", "condition": "not_passed"}`
+- Verifier 节点：判断目标是否达成的验证器
+- 前端画布（#1D）支持画循环边
+
+### #2D. executor 沙箱升级
+
+**现状**: executor 容器基于 Docker，无严格内核隔离。
+**目标**: gVisor 或 Firecracker 微虚拟机隔离。
+**工作量**: 3 天
+**评估**: gVisor 集成最简单（Docker 的 `--runtime=runsc`）
+
+### #2E. 技能 SDK + 脚手架 CLI
+
+**现状**: 技能开发者手动创建目录和 SKILL.md。
+**目标**: `qihu-skill` CLI 脚手架。
+**工作量**: 1 周
+**实施**:
+- `qihu-skill init my-skill` → 生成目录结构 + SKILL.md 模板
+- `qihu-skill validate` → 校验 SKILL.md Schema
+- `qihu-skill publish` → 打包上传到平台技能市场
+- `qihu-skill dev` → 本地调试（Mock SkillRuntime）
+
+---
+
+## Phase 3：评估后按需投入（5 项，P2）
+
+### #3A. Semantica 决策溯源
+- 场景：金融/医疗/法律客户要求可解释 AI
+- 工作量：评估 3 天 + 集成 1 周
+- 决策标准：是否有付费企业客户明确要求
+
+### #3B. Langfuse/LangSmith Agent 轨迹
+- 场景：开发者调试 Agent 执行、产品优化
+- 工作量：3 天集成
+- 决策标准：Agent 可视化画布上线后，按需求优先级
+
+### #3C. Agent 广播网络（EigenFlux）
+- 场景：Agent 间自动发现与协作
+- 工作量：评估
+- 决策标准：多 Agent 协作需求明确后
+
+### #3D. 前端暗色模式
+- 工作量：3 天
+- 决策标准：用户体验迭代批次
+
+### #3E. 移动端适配
+- 工作量：2 周
+- 决策标准：移动端用户数据达标后
+
+---
+
+## 优先级总表（20 项排序）
+
+| # | 行动 | 模块 | 优先级 | 工作量 | 依赖 |
+|---|------|------|--------|--------|------|
+| 1 | LiteLLM Proxy 统一 AI 网关 | 模型路由 | P0 | 3d | - |
+| 2 | Meilisearch 技能市场搜索 | 搜索 | P0 | 2d | - |
+| 3 | PG RLS + tenant_id 多租户隔离 | 多租户 | P0 | 3d | - |
+| 4 | 角色 5 级 + 用户组 | 多租户 | P0 | 3d | #3 |
+| **5** | **3 个标杆技能真实执行器** | **技能市场** | **P0** | **2w** | **-** |
+| 6 | SKILL.md 统一 Schema | 技能市场 | P0 | 3d | - |
+| 7 | 流式 Token 计费 + 实时余额 | 变现 | P0 | 5d | #1 |
+| **8** | **Agent 可视化编排画布** | **Agent** | **P1** | **2w** | **-** |
+| 9 | Loop 原语 + Verifier | Agent | P1 | 1w | #8 |
+| 10 | 工具四级风险审批 | 安全 | P1 | 3d | - |
+| 11 | executor gVisor 升级 | 安全 | P1 | 3d | - |
+| 12 | MemPalace 分层记忆 | RAG | P1 | 1w | - |
+| 13 | 技能 SDK + 脚手架 CLI | 技能市场 | P1 | 1w | #6 |
+| 14 | LongMemEval 基准测试 | RAG | P1 | 2d | - |
+| 15 | 技能审核流水线 | 技能市场 | P2 | 1w | #6 |
+| 16 | 分润模型 + 开发者结算 | 变现 | P2 | 1w | #7 |
+| 17 | Langfuse Agent 轨迹 | 可观测 | P2 | 3d | #1 |
+| 18 | Semantica 决策溯源 | 安全 | P2 | 评估 | - |
+| 19 | 前端暗色模式 | UI | P2 | 3d | - |
+| 20 | 移动端适配 | UI | P2 | 2w | - |
+
+---
+
+## 竞品对比更新（基于分析文档）
+
+| 维度 | 七虎AI 现状 | Open WebUI | Dify | 差距/优势 |
+|------|------------|------------|------|----------|
+| RBAC | admin/user 2 级 | Admin/User/Pending + Groups | 内置多租户 | 🔴 落后 |
+| 向量库 | 仅 pgvector | 9 种可插拔 | 多种 | 🟡 够用 |
+| RAG 管线 | 单路向量 | 3 阶段 BM25+向量+Rerank | 混合检索 | 🟡 待增强 |
+| 插件体系 | Skills+MCP | 5 类插件 | 工作流节点 | 🟡 够用 |
+| Token 计量 | 仅记录 | 实时+按模型+仪表盘 | 内置 | 🔴 落后 |
+| 模型路由 | try-except | Pipe 路由 | 内置 failover | 🟡 待 LiteLLM |
+| Agent 编排 | DAG 图引擎 | 无 | 工作流画布 | 🟢 领先 |
+| 技能市场 | 上架-安装-执行 | 无 | 无 | 🟢 护城河 |
+
+---
+
+## 当前已完成（对标分析文档的"已修复弱点"）
+
+这些本期会话已修复，分析文档第九节引用：
+
+| 修复 | 状态 | 对应文档章节 |
+|------|------|------------|
+| AGENT_GUIDE 标准化 | ✅ | §九 - 固定弱点已修复 |
+| MCP 官方 SDK + Schema 转换 | ✅ | §九 |
+| Tool Policy 白名单+预算+每轮上限 | ✅ | §九 |
+| MiMo Token Plan 适配 | ✅ | - |
+| `overflow-x: hidden` 响应式修复 | ✅ | - |
+| 架构蓝图 `ARCHITECTURE.md` | ✅ | - |
+
+---
+
+> **执行指令**: 告诉我 "执行 Phase 0" 开始落地。每完成一项在对应 checkpoint 勾掉。
+> **上传到 Gitea**: 本方案同步上传到 `qihuaiflow/docs/ITERATION_PLAN.md`

frontend/src/api/index.ts

@@ -959,6 +959,10 @@ export const adminAPI = {
 
   checkAdminKnowledgeExtStatus: (id: string) => get<any>(`/admin/knowledge/external/services/${id}/status`),
 
+  getAdminAnySearchConfig: () => get<any>('/admin/knowledge/anysearch/config'),
+
+  saveAdminAnySearchConfig: (data: { api_key: string; is_enabled: boolean }) => post<any>('/admin/knowledge/anysearch/config', data),
+
   syncAdminKnowledge: (kbId: string, data: any) => post<any>(`/admin/knowledge/external/sync/${kbId}`, data),
 
   batchDeleteAdminKnowledge: (ids: string[]) => post<any>('/admin/knowledge/batch/delete', ids),

frontend/src/components/layout/FrontLayout.css

@@ -203,6 +203,7 @@
 .front-main {
   flex: 1;
   overflow: hidden;
+  overflow-x: hidden;
   display: flex;
   flex-direction: column;
   min-width: 0;

frontend/src/pages/admin/Knowledge.tsx

@@ -71,6 +71,11 @@ const Knowledge: React.FC = () => {
   const [ingesting, setIngesting] = useState(false);
   const [ingestResult, setIngestResult] = useState<any>(null);
 
+  // AnySearch global config
+  const [anysearchForm, setAnysearchForm] = useState<{ api_key: string; is_enabled: boolean }>({ api_key: '', is_enabled: false });
+  const [anysearchLoading, setAnysearchLoading] = useState(false);
+  const [anysearchMsg, setAnysearchMsg] = useState('');
+
   // KB Modal
   const [showModal, setShowModal] = useState(false);
   const [editingKB, setEditingKB] = useState<KnowledgeBaseItem | null>(null);
@@ -123,6 +128,33 @@ const Knowledge: React.FC = () => {
     try { const res = await adminAPI.getAdminKnowledgeExtServices(); if (res.code === 0) setExtServices(res.data.items || []); } catch (e) {}
   };
 
+  const fetchAnySearchConfig = async () => {
+    try {
+      const res = await adminAPI.getAdminAnySearchConfig();
+      if (res.code === 0) {
+        const d = res.data || {};
+        setAnysearchForm({ api_key: d.api_key || '', is_enabled: !!d.is_enabled });
+      }
+    } catch (e) {}
+  };
+
+  const handleSaveAnySearch = async () => {
+    setAnysearchLoading(true); setAnysearchMsg('');
+    try {
+      const res = await adminAPI.saveAdminAnySearchConfig(anysearchForm);
+      if (res.code === 0) {
+        const conn = res.data?.connection || {};
+        setAnysearchMsg(conn.connected ? '✅ 已保存并连接成功' : `⚠️ 已保存但连接失败：${conn.detail || '未知错误'}`);
+      } else {
+        setAnysearchMsg(`❌ 保存失败：${res.message || '未知错误'}`);
+      }
+    } catch (e: any) {
+      setAnysearchMsg(`❌ 保存失败：${e?.message || '网络错误'}`);
+    } finally {
+      setAnysearchLoading(false);
+    }
+  };
+
   const handleCreateKB = async () => {
     try {
       const res = await adminAPI.createAdminKnowledge(kbForm);
@@ -334,7 +366,7 @@ const Knowledge: React.FC = () => {
         </select>
         <button onClick={() => { setEditingKB(null); setKBForm({ name: '', description: '', is_official: false }); setShowModal(true); }}
           style={btnStyle('#6C63FF')}>+ 创建知识库</button>
-        <button onClick={() => setShowExtModal(true)} style={btnStyle('#8B5CF6')}>外部服务配置</button>
+        <button onClick={() => { setShowExtModal(true); fetchAnySearchConfig(); }} style={btnStyle('#8B5CF6')}>外部服务配置</button>
       </div>
 
       {/* Table */}
@@ -523,54 +555,62 @@ const Knowledge: React.FC = () => {
       {/* External Service Modal */}
       {showExtModal && (
         <div style={overlayStyle} onClick={() => setShowExtModal(false)}>
-          <div style={{ ...modalStyle, maxWidth: 650 }} onClick={e => e.stopPropagation()}>
+          <div style={{ ...modalStyle, width: 720, maxWidth: '92vw' }} onClick={e => e.stopPropagation()}>
-            <h3>外部知识库服务配置</h3>
+            <h3 style={{ margin: '0 0 20px 0', fontSize: 18 }}>外部知识库服务配置</h3>
-            {extServices.length > 0 && (
+
-              <div style={{ marginBottom: 16 }}>
+            {/* Section 1: Dify 外部服务列表 */}
-                <table className="admin-table" style={{ marginBottom: 0 }}>
+            <div style={{ marginBottom: 24 }}>
-                  <thead><tr><th>名称</th><th>类型</th><th>URL</th><th>状态</th><th>操作</th></tr></thead>
+              <h4 style={{ margin: '0 0 12px 0', fontSize: 14, color: '#374151', fontWeight: 600 }}>已接入外部服务</h4>
-                  <tbody>
+              {extServices.length > 0 ? (
-                    {extServices.map(svc => (
+                <div style={{ border: '1px solid #E5E7EB', borderRadius: 8, overflow: 'hidden' }}>
-                      <tr key={svc.id}>
+                  <table className="admin-table" style={{ marginBottom: 0, width: '100%' }}>
-                        <td>{svc.name}</td>
+                    <thead><tr><th style={{ textAlign: 'left' }}>名称</th><th style={{ textAlign: 'left' }}>类型</th><th style={{ textAlign: 'left' }}>URL</th><th style={{ textAlign: 'left' }}>状态</th><th style={{ textAlign: 'center', width: 120 }}>操作</th></tr></thead>
-                        <td><span style={{ fontSize: 12, padding: '2px 6px', background: '#F3F4FF', color: '#8B5CF6', borderRadius: 3 }}>Dify</span></td>
+                    <tbody>
-                        <td style={{ fontSize: 11, maxWidth: 200, overflow: 'hidden', textOverflow: 'ellipsis' }}>{svc.api_url}</td>
+                      {extServices.map(svc => (
-                        <td>{statusBadge(svc.status)}</td>
+                        <tr key={svc.id}>
-                        <td>
+                          <td>{svc.name}</td>
-                          <button onClick={() => handleCheckExtStatus(svc.id)} style={btnStyle('#3B82F6', 'small')}>检测</button>
+                          <td><span style={{ fontSize: 12, padding: '2px 8px', background: '#F3F4FF', color: '#8B5CF6', borderRadius: 4 }}>Dify</span></td>
-                          <button onClick={() => handleDeleteExtService(svc.id)} style={{ ...btnStyle('#DC2626', 'small'), marginLeft: 4 }}>删除</button>
+                          <td style={{ fontSize: 12, maxWidth: 220, overflow: 'hidden', textOverflow: 'ellipsis', whiteSpace: 'nowrap' }}>{svc.api_url}</td>
-                        </td>
+                          <td>{statusBadge(svc.status)}</td>
-                      </tr>
+                          <td style={{ textAlign: 'center' }}>
-                    ))}
+                            <button onClick={() => handleCheckExtStatus(svc.id)} style={btnStyle('#3B82F6', 'small')}>检测</button>
-                  </tbody>
+                            <button onClick={() => handleDeleteExtService(svc.id)} style={{ ...btnStyle('#DC2626', 'small'), marginLeft: 6 }}>删除</button>
-                </table>
+                          </td>
-              </div>
+                        </tr>
-            )}
+                      ))}
-            <div style={{ borderTop: '1px solid #E5E7EB', paddingTop: 16 }}>
+                    </tbody>
-              <h4 style={{ margin: '0 0 12px 0', fontSize: 14 }}>配置 Dify 知识库</h4>
+                  </table>
-              <div style={{ display: 'grid', gridTemplateColumns: '1fr 1fr', gap: 12 }}>
+                </div>
+              ) : (
+                <div style={{ padding: 16, background: '#F9FAFB', borderRadius: 8, color: '#6B7280', fontSize: 13 }}>暂无外部服务，可在下方添加 Dify 知识库。</div>
+              )}
+            </div>
+
+            {/* Section 2: Dify 配置表单 */}
+            <div style={{ borderTop: '1px solid #E5E7EB', paddingTop: 20, marginBottom: 24 }}>
+              <h4 style={{ margin: '0 0 16px 0', fontSize: 14, color: '#374151', fontWeight: 600 }}>配置 Dify 知识库</h4>
+              <div style={{ display: 'grid', gridTemplateColumns: '1fr 1fr', gap: 16 }}>
                 <div><label style={labelStyle}>名称</label><input value={extForm.name} onChange={e => setExtForm({ ...extForm, name: e.target.value })} placeholder="如: 生产环境 Dify" style={fullInput} /></div>
                 <div><label style={labelStyle}>知识库名称(Dataset)</label><input value={extForm.dataset_name} onChange={e => setExtForm({ ...extForm, dataset_name: e.target.value })} placeholder="如: qihuai" style={fullInput} /></div>
               </div>
-              <div style={{ marginTop: 12 }}><label style={labelStyle}>API URL</label><input value={extForm.api_url} onChange={e => setExtForm({ ...extForm, api_url: e.target.value })} placeholder="https://dify.qihuyun.com/v1" style={fullInput} /></div>
+              <div style={{ marginTop: 16 }}><label style={labelStyle}>API URL</label><input value={extForm.api_url} onChange={e => setExtForm({ ...extForm, api_url: e.target.value })} placeholder="https://dify.qihuyun.com/v1" style={fullInput} /></div>
-              <div style={{ marginTop: 12 }}><label style={labelStyle}>API Key（dataset 密钥）</label><input type="password" value={extForm.api_key} onChange={e => setExtForm({ ...extForm, api_key: e.target.value })} placeholder="dataset-xxxx" style={fullInput} /></div>
+              <div style={{ marginTop: 16 }}><label style={labelStyle}>API Key（dataset 密钥）</label><input type="password" value={extForm.api_key} onChange={e => setExtForm({ ...extForm, api_key: e.target.value })} placeholder="dataset-xxxx" style={fullInput} /></div>
 
               {extMsg && (
-                <div style={{ marginTop: 12, fontSize: 13, color: extMsg.startsWith('✅') ? '#16A34A' : '#DC2626', padding: '6px 12px', background: extMsg.startsWith('✅') ? '#DCFCE7' : '#FEE2E2', borderRadius: 6 }}>
+                <div style={{ marginTop: 16, fontSize: 13, color: extMsg.startsWith('✅') ? '#16A34A' : '#DC2626', padding: '8px 12px', background: extMsg.startsWith('✅') ? '#DCFCE7' : '#FEE2E2', borderRadius: 6 }}>
                   {extMsg}
                 </div>
               )}
 
-              <div style={{ display: 'flex', gap: 12, marginTop: 16, flexWrap: 'wrap' }}>
+              <div style={{ display: 'flex', gap: 12, marginTop: 16, alignItems: 'center' }}>
                 <button onClick={handleCreateExtService} style={btnStyle('#8B5CF6')}>保存并测试连接</button>
                 <button onClick={handleIngestPlatform} disabled={ingesting} style={btnStyle(ingesting ? '#D1D5DB' : '#219B5E')}>
                   {ingesting ? '入库中...' : '入库平台内容'}
                 </button>
-                <button onClick={() => setShowExtModal(false)} style={btnStyle('#6B7280')}>关闭</button>
               </div>
 
               {ingestResult && (
-                <div style={{ marginTop: 12, fontSize: 13, padding: 12, background: '#F0FDF4', border: '1px solid #BBF7D0', borderRadius: 8 }}>
+                <div style={{ marginTop: 16, fontSize: 13, padding: 12, background: '#F0FDF4', border: '1px solid #BBF7D0', borderRadius: 8 }}>
                   <div style={{ fontWeight: 600, marginBottom: 6 }}>入库结果</div>
                   <div>AI 聊天回答：<b>{ingestResult.chat_answers}</b> 条</div>
                   <div>用户上传文档：<b>{ingestResult.user_docs}</b> 条</div>
@@ -578,6 +618,45 @@ const Knowledge: React.FC = () => {
                 </div>
               )}
             </div>
+
+            {/* Section 3: AnySearch 全局固定配置 */}
+            <div style={{ borderTop: '1px solid #E5E7EB', paddingTop: 20, marginBottom: 8 }}>
+              <h4 style={{ margin: '0 0 16px 0', fontSize: 14, color: '#374151', fontWeight: 600 }}>AnySearch 联网检索（全局配置）</h4>
+              <div style={{ display: 'grid', gridTemplateColumns: '1fr auto', gap: 16, alignItems: 'end' }}>
+                <div>
+                  <label style={labelStyle}>API Key</label>
+                  <input
+                    type="password"
+                    value={anysearchForm.api_key}
+                    onChange={e => setAnysearchForm({ ...anysearchForm, api_key: e.target.value })}
+                    placeholder="as_sk_xxxx"
+                    style={fullInput}
+                  />
+                </div>
+                <label style={{ cursor: 'pointer', display: 'flex', alignItems: 'center', gap: 8, padding: '8px 0', fontWeight: 500, fontSize: 14, color: '#374151', whiteSpace: 'nowrap' }}>
+                  <input type="checkbox" checked={anysearchForm.is_enabled} onChange={e => setAnysearchForm({ ...anysearchForm, is_enabled: e.target.checked })} />
+                  启用联网检索
+                </label>
+              </div>
+
+              {anysearchMsg && (
+                <div style={{ marginTop: 16, fontSize: 13, color: anysearchMsg.startsWith('✅') ? '#16A34A' : (anysearchMsg.startsWith('⚠️') ? '#D97706' : '#DC2626'), padding: '8px 12px', background: anysearchMsg.startsWith('✅') ? '#DCFCE7' : (anysearchMsg.startsWith('⚠️') ? '#FEF3C7' : '#FEE2E2'), borderRadius: 6 }}>
+                  {anysearchMsg}
+                </div>
+              )}
+
+              <div style={{ display: 'flex', gap: 12, marginTop: 16, alignItems: 'center' }}>
+                <button onClick={handleSaveAnySearch} disabled={anysearchLoading} style={btnStyle(anysearchLoading ? '#D1D5DB' : '#8B5CF6')}>
+                  {anysearchLoading ? '保存中...' : '保存并测试连接'}
+                </button>
+                <span style={{ fontSize: 12, color: '#6B7280' }}>配置对平台所有用户生效，AI 聊天将自动注入实时联网结果。</span>
+              </div>
+            </div>
+
+            {/* Footer */}
+            <div style={{ borderTop: '1px solid #E5E7EB', marginTop: 24, paddingTop: 16, display: 'flex', justifyContent: 'flex-end' }}>
+              <button onClick={() => setShowExtModal(false)} style={btnStyle('#6B7280')}>关闭</button>
+            </div>
           </div>
         </div>
       )}

frontend/src/pages/admin/Mcp.tsx

@@ -14,9 +14,9 @@ const styles: Record<string, React.CSSProperties> = {
   searchInput: { padding: '8px 16px', border: '1px solid #e5e7eb', borderRadius: 8, fontSize: 14, width: 250 },
   filterSelect: { padding: '8px 12px', border: '1px solid #e5e7eb', borderRadius: 8, fontSize: 14, backgroundColor: '#fff' },
   table: { width: '100%', borderCollapse: 'collapse' as const, backgroundColor: '#fff', borderRadius: 12, overflow: 'hidden', boxShadow: '0 1px 3px rgba(0,0,0,0.08)' },
-  th: { padding: '12px 16px', textAlign: 'left' as const, fontSize: 13, fontWeight: 600, color: '#6b7280', backgroundColor: '#f9fafb', borderBottom: '1px solid #e5e7eb' },
+  th: { padding: '12px 16px', textAlign: 'left' as const, fontSize: 13, fontWeight: 600, color: '#6b7280', backgroundColor: '#f9fafb', borderBottom: '1px solid #e5e7eb', whiteSpace: 'nowrap' as const },
   td: { padding: '12px 16px', fontSize: 14, color: '#374151', borderBottom: '1px solid #f3f4f6' },
-  badge: { display: 'inline-block', padding: '2px 10px', borderRadius: 20, fontSize: 12, fontWeight: 600 },
+  badge: { display: 'inline-block', padding: '2px 10px', borderRadius: 20, fontSize: 12, fontWeight: 600, whiteSpace: 'nowrap' as const },
   actionBtn: { padding: '4px 12px', border: '1px solid #e5e7eb', borderRadius: 6, fontSize: 13, cursor: 'pointer', marginRight: 6, backgroundColor: '#fff' },
   pagination: { display: 'flex', justifyContent: 'center', gap: 8, marginTop: 20 },
   pageBtn: { padding: '8px 16px', border: '1px solid #e5e7eb', borderRadius: 6, cursor: 'pointer', backgroundColor: '#fff', fontSize: 14 },
@@ -398,9 +398,9 @@ export default function AdminMcp() {
         <thead>
           <tr>
             <th style={styles.th}>名称</th>
-            <th style={styles.th}>传输方式</th>
+            <th style={{ ...styles.th, minWidth: 100 }}>传输方式</th>
-            <th style={styles.th}>状态</th>
+            <th style={{ ...styles.th, minWidth: 80 }}>状态</th>
-            <th style={styles.th}>工具数</th>
+            <th style={{ ...styles.th, minWidth: 70 }}>工具数</th>
             <th style={styles.th}>最后连接</th>
             <th style={styles.th}>启用</th>
             <th style={styles.th}>操作</th>

frontend/src/pages/admin/Skills.tsx

@@ -52,16 +52,22 @@ const STATE_CONFIG: Record<SkillState, { label: string; color: string; bg: strin
   frozen: { label: '已冻结', color: '#EF4444', bg: '#FEE2E2' },
 };
 
+// 业务场景分类：与前端「应用市场」保持一致的 14 类标准（前台分类为准）
 const CATEGORIES = [
-  { id: 'writing', label: '文字创作' },
+  { id: 'design', label: '🎨 做设计' },
-  { id: 'coding', label: '代码能力' },
+  { id: 'frontend', label: '💻 写前端' },
-  { id: 'data_analysis', label: '数据分析' },
+  { id: 'mobile', label: '📱 做移动端' },
-  { id: 'qa', label: '知识问答' },
+  { id: 'backend', label: '🧪 后端/全栈' },
-  { id: 'automation', label: '流程自动化' },
+  { id: 'qa', label: '🧪 测试质量' },
-  { id: 'creativity', label: '创意生成' },
+  { id: 'search', label: '🔍 搜索信息' },
-  { id: 'communication', label: '沟通协作' },
+  { id: 'docs', label: '✍️ 写文档' },
-  { id: 'reasoning', label: '逻辑推理' },
+  { id: 'marketing', label: '📢 推广获客' },
-  { id: 'other', label: '其他' },
+  { id: 'workflow', label: '🔄 工作流' },
+  { id: 'secops', label: '🛡️ 安全运维' },
+  { id: 'learning', label: '🎓 学习教学' },
+  { id: 'video', label: '🎬 视频多媒体' },
+  { id: 'collab', label: '🏢 办公协作' },
+  { id: 'ai', label: '🛠️ AI能力' },
 ];
 
 const PAGE_SIZE = 15;
@@ -398,6 +404,7 @@ const Skills: React.FC = () => {
               <th>技能名称</th>
               <th>状态</th>
               <th>分类</th>
+              <th>标签</th>
               <th>所有者</th>
               <th>版本</th>
               <th>使用次数</th>
@@ -408,9 +415,9 @@ const Skills: React.FC = () => {
           </thead>
           <tbody>
             {loading ? (
-              <tr><td colSpan={9} style={{ textAlign: 'center', padding: 40, color: '#6B7280' }}>加载中...</td></tr>
+              <tr><td colSpan={10} style={{ textAlign: 'center', padding: 40, color: '#6B7280' }}>加载中...</td></tr>
             ) : skills.length === 0 ? (
-              <tr><td colSpan={9} style={{ textAlign: 'center', padding: 40, color: '#6B7280' }}>暂无技能</td></tr>
+              <tr><td colSpan={10} style={{ textAlign: 'center', padding: 40, color: '#6B7280' }}>暂无技能</td></tr>
             ) : skills.map(skill => {
               const cfg = STATE_CONFIG[skill.state as SkillState] || STATE_CONFIG.baby;
               const published = isPublished(skill);
@@ -441,6 +448,16 @@ const Skills: React.FC = () => {
                     </span>
                   </td>
                   <td>{skill.category_label || '-'}</td>
+                  <td>
+                    {skill.tags && skill.tags.length > 0 ? (
+                      <div style={{ display: 'flex', gap: 4, flexWrap: 'wrap' }}>
+                        {skill.tags.slice(0, 3).map(tag => (
+                          <span key={tag} style={{ fontSize: 11, padding: '1px 6px', background: '#EEF2FF', color: '#6366F1', borderRadius: 4 }}>{tag}</span>
+                        ))}
+                        {skill.tags.length > 3 && <span style={{ fontSize: 11, color: '#6B7280' }}>+{skill.tags.length - 3}</span>}
+                      </div>
+                    ) : <span style={{ color: '#9CA3AF', fontSize: 12 }}>-</span>}
+                  </td>
                   <td>{skill.owner_username || skill.user_id}</td>
                   <td>v{skill.version}</td>
                   <td>{skill.use_count}</td>

frontend/src/pages/front/ChatPage.css

@@ -471,10 +471,15 @@
 }
 
 .md-table td {
-  padding: 10px 14px;
+  padding: 8px 12px;
   border-bottom: 1px solid #EEF0F4;
   color: var(--color-text-primary);
   vertical-align: top;
+  line-height: 1.5;
+}
+
+.md-table td br {
+  line-height: 0.8;
 }
 
 .md-table tbody tr:last-child td {
@@ -1552,3 +1557,230 @@ del {
   border-color: var(--color-accent, #5B5FED);
   color: var(--color-accent, #5B5FED);
 }
+
+/* ===== AI 思考 / 执行过程面板（收拢式） ===== */
+.ai-process {
+  margin-top: 10px;
+  border: 1px solid var(--color-border-light, #e6e9f0);
+  border-radius: var(--radius-md, 12px);
+  background: color-mix(in srgb, var(--color-bg-secondary, #fff) 75%, var(--color-accent, #5B5FED) 4%);
+  backdrop-filter: blur(8px) saturate(160%);
+  overflow: hidden;
+  transition: border-color var(--transition-fast, .2s), box-shadow var(--transition-fast, .2s);
+}
+
+.ai-process[open] {
+  box-shadow: 0 6px 22px rgba(91, 95, 237, 0.10);
+  border-color: color-mix(in srgb, var(--color-accent, #5B5FED) 35%, transparent);
+}
+
+/* 新鲜出炉的思考面板 — 脉冲高亮动画 */
+.ai-process-fresh {
+  animation: aiProcessPulse 1.2s ease-out;
+}
+.ai-process-fresh .ai-process-summary {
+  color: var(--color-accent, #5B5FED);
+  font-weight: 700;
+}
+
+@keyframes aiProcessPulse {
+  0% {
+    box-shadow: 0 0 0 0 rgba(91, 95, 237, 0.4);
+    border-color: var(--color-accent, #5B5FED);
+  }
+  50% {
+    box-shadow: 0 0 0 6px rgba(91, 95, 237, 0.08);
+    border-color: color-mix(in srgb, var(--color-accent, #5B5FED) 60%, transparent);
+  }
+  100% {
+    box-shadow: 0 6px 22px rgba(91, 95, 237, 0.10);
+    border-color: color-mix(in srgb, var(--color-accent, #5B5FED) 35%, transparent);
+  }
+}
+
+.ai-process-summary {
+  list-style: none;
+  cursor: pointer;
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  padding: 9px 14px;
+  font-size: var(--text-sm, 13px);
+  font-weight: 600;
+  color: var(--color-text-secondary, #475069);
+  user-select: none;
+  transition: background var(--transition-fast, .2s);
+}
+
+.ai-process-summary::-webkit-details-marker { display: none; }
+.ai-process-summary::marker { content: ""; }
+
+.ai-process-summary:hover {
+  background: color-mix(in srgb, var(--color-accent, #5B5FED) 7%, transparent);
+}
+
+.ai-process-badge {
+  font-size: 15px;
+  line-height: 1;
+  filter: drop-shadow(0 1px 2px rgba(91, 95, 237, 0.25));
+}
+
+.ai-process-list {
+  padding: 4px 14px 12px;
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+  animation: ai-process-in .28s cubic-bezier(0.16, 1, 0.3, 1);
+}
+
+@keyframes ai-process-in {
+  from { opacity: 0; transform: translateY(-4px); }
+  to   { opacity: 1; transform: translateY(0); }
+}
+
+.ai-process-item {
+  display: flex;
+  gap: 10px;
+  padding: 9px 12px;
+  border-radius: 10px;
+  background: var(--color-bg-secondary, #fff);
+  border: 1px solid var(--color-border-light, #e6e9f0);
+  border-left: 3px solid var(--color-accent, #5B5FED);
+}
+
+.ai-process-thinking { border-left-color: #8b8fe6; }
+.ai-process-skill    { border-left-color: #5B5FED; }
+.ai-process-mcp      { border-left-color: #12a594; }
+
+.ai-process-ico {
+  font-size: 15px;
+  line-height: 1.4;
+  flex-shrink: 0;
+}
+
+.ai-process-body {
+  min-width: 0;
+  flex: 1;
+}
+
+.ai-process-title {
+  display: block;
+  font-size: var(--text-sm, 13px);
+  font-weight: 600;
+  color: var(--color-text-primary, #1f2430);
+  margin-bottom: 2px;
+}
+
+.ai-process-text {
+  margin: 0;
+  font-size: var(--text-xs, 12px);
+  line-height: 1.55;
+  color: var(--color-text-tertiary, #8a90a6);
+  max-height: 160px;
+  overflow-y: auto;
+  word-break: break-word;
+  white-space: pre-wrap;
+}
+
+/* 暗色主题微调 */
+[data-theme="dark"] .ai-process,
+html.dark .ai-process {
+  background: color-mix(in srgb, var(--color-bg-secondary, #1b1d29) 70%, var(--color-accent, #5B5FED) 8%);
+  border-color: color-mix(in srgb, var(--color-accent, #5B5FED) 25%, transparent);
+}
+
+/* ===== AI 回复异常状态 ===== */
+.ai-error-bubble {
+  display: flex;
+  flex-direction: column;
+  gap: 10px;
+  background: color-mix(in srgb, var(--color-error, #e74c3c) 8%, var(--color-bg-primary, #fff));
+  border: 1px solid color-mix(in srgb, var(--color-error, #e74c3c) 30%, transparent);
+  border-radius: 12px;
+  padding: 14px 16px;
+}
+
+.ai-error-icon {
+  font-size: 18px;
+  margin-bottom: -6px;
+}
+
+.ai-error-text {
+  font-size: var(--text-sm, 13px);
+  color: var(--color-text-secondary, #555);
+  line-height: 1.6;
+}
+
+.ai-error-text p {
+  margin: 0 0 4px;
+}
+
+.ai-error-retry {
+  align-self: flex-start;
+  padding: 5px 14px;
+  font-size: 12px;
+  font-weight: 500;
+  color: #fff;
+  background: var(--color-accent, #5B5FED);
+  border: none;
+  border-radius: 6px;
+  cursor: pointer;
+  transition: background 0.2s, transform 0.15s;
+}
+
+.ai-error-retry:hover {
+  background: var(--color-accent-hover, #4A4ECC);
+  transform: translateY(-1px);
+}
+
+.ai-error-retry:active {
+  transform: scale(0.97);
+}
+
+/* 暗色主题 */
+[data-theme="dark"] .ai-error-bubble,
+html.dark .ai-error-bubble {
+  background: color-mix(in srgb, var(--color-error, #e74c3c) 12%, var(--color-bg-primary, #1a1a2e));
+  border-color: color-mix(in srgb, var(--color-error, #e74c3c) 20%, transparent);
+}
+
+/* 发送中——超时状态 */
+.ai-sending-timeout {
+  border-color: color-mix(in srgb, #f39c12 30%, transparent) !important;
+  background: color-mix(in srgb, #f39c12 5%, var(--color-bg-primary, #fff)) !important;
+}
+
+[data-theme="dark"] .ai-sending-timeout,
+html.dark .ai-sending-timeout {
+  border-color: color-mix(in srgb, #f39c12 25%, transparent) !important;
+  background: color-mix(in srgb, #f39c12 8%, var(--color-bg-primary, #1a1a2e)) !important;
+}
+
+.ai-cancel-btn:hover {
+  background: color-mix(in srgb, #e74c3c 15%, var(--color-bg-tertiary)) !important;
+}
+
+/* ===== Phase 0 SSE 流式：思考链实时渲染 ===== */
+/* 思考链文本保留模型原始换行（deepseek 原生 reasoning 含多段换行） */
+.ai-process-thinking .ai-process-text {
+  white-space: pre-wrap;
+}
+
+/* 流式生成中提示 */
+.streaming-hint {
+  display: inline-flex;
+  align-items: center;
+  gap: 8px;
+  color: var(--color-text-secondary);
+  font-size: 14px;
+}
+
+.streaming-hint svg {
+  width: 16px;
+  height: 16px;
+}
+
+.ai-process-running { animation: aiPulse 1.5s ease-in-out infinite; }
+@keyframes aiPulse { 0%,100% { opacity: 1; } 50% { opacity: 0.6; } }
+.ai-process-error { background: rgba(255,0,0,0.05); border-left: 2px solid #e74c3c; }
+.ai-process-running { border-left: 2px solid #3498db; }

frontend/src/pages/front/ChatPage.tsx

@@ -6,17 +6,34 @@ import type { Conversation, Message } from '@/types';
 import './ChatPage.css';
 
 // ===== 增强渲染依赖 =====
-import hljs from 'highlight.js';
 import 'highlight.js/styles/github.css';
-import katex from 'katex';
 import 'katex/dist/katex.min.css';
 import mermaid from 'mermaid';
 
 mermaid.initialize({ startOnLoad: false, theme: 'default' });
 
+// ===== AI 回复异常检测 =====
+const AI_ERROR_PATTERNS = [
+  'AI模型调用失败',
+  'AI服务暂时不可用',
+  'AI模型返回了空响应',
+  'AI模型已处理工具调用结果',
+  '该模型缺少API密钥或端点配置',
+  '请选择一个已配置的AI模型',
+  '请先在后台管理-大模型库配置默认对话模型',
+  '⚠️ 多 Agent 编排执行失败',
+  '⚠️ 外部 Agent 调用失败',
+  '编排未产生内容',
+  '外部 Agent 未返回内容',
+];
+function isAiError(content: string): boolean {
+  if (!content) return false;
+  return AI_ERROR_PATTERNS.some(p => content.includes(p));
+}
+
 // ===== Enhanced Markdown Renderer =====
 // 媒体(图片/视频)以受限缩略图呈现，并提供下载(格式+尺寸)选项
-interface MediaInfo { type: 'img' | 'video'; src: string; alt?: string; }
+import { renderMarkdownToHtml, MediaInfo } from '@/utils/markdown';
 
 // 全屏灯箱预览
 const ChatLightbox: React.FC<{ src: string; type: 'img' | 'video'; onClose: () => void }> = ({ src, type, onClose }) => {
@@ -154,172 +171,17 @@ const ChatMedia: React.FC<{ info: MediaInfo }> = ({ info }) => {
 };
 
 const MarkdownContent: React.FC<{ content: string }> = ({ content }) => {
+  // 渲染管线（代码块/加粗/表格/换行/媒体占位符等）已抽到 src/utils/markdown.ts，
+  // 组件只负责把返回的 html 切分为「文本片段 + 媒体组件」。
+  const { html, media } = useMemo(() => renderMarkdownToHtml(content), [content]);
+
   const nodes = useMemo(() => {
-    let text = content;
+    // 12. 拆分为文本片段与媒体组件（媒体以 \x00MEDIA<n>\x00 占位符保留在 html 中）
-
+    const parts = html.split(/(\x00MEDIA\d+\x00)/g);
-    // 1. Extract code blocks first (protect from other replacements)
-    const codeBlocks: string[] = [];
-    const mermaidBlocks: string[] = [];
-    text = text.replace(/```(\w*)\n?([\s\S]*?)```/g, (_m: string, _lang: string, code: string) => {
-      if (_lang === 'mermaid') {
-        const idx = mermaidBlocks.length;
-        mermaidBlocks.push('<div class="md-mermaid" data-code="' + escapeHtml(code.trim()) + '"></div>');
-        return '\x00MERMAID' + idx + '\x00';
-      }
-      const idx = codeBlocks.length;
-      codeBlocks.push('<pre class="md-code-block"><code class="md-code">' + escapeHtml(code.trim()) + '</code></pre>');
-      return '\x00CODEBLOCK' + idx + '\x00';
-    });
-
-    // 2. Extract inline code
-    const inlineCodes: string[] = [];
-    text = text.replace(/`([^`]+)`/g, (_m: string, code: string) => {
-      const idx = inlineCodes.length;
-      inlineCodes.push('<code class="md-inline-code">' + escapeHtml(code) + '</code>');
-      return '\x00INLINE' + idx + '\x00';
-    });
-
-    // 2.5 Extract media blocks (<img>/<video>) → 结构化，后续渲染为 React 组件
-    const mediaBlocks: MediaInfo[] = [];
-    text = text.replace(/<(img|video)([^>]*?)\/?>/gi, (_m: string, tag: string, attrs: string) => {
-      const type = tag.toLowerCase() === 'img' ? 'img' : 'video';
-      const srcM = attrs.match(/src=["']([^"']+)["']/i);
-      const altM = attrs.match(/alt=["']([^"']*)["']/i);
-      mediaBlocks.push({ type, src: srcM ? srcM[1] : '', alt: altM ? altM[1] : '' });
-      return '\x00MEDIA' + (mediaBlocks.length - 1) + '\x00';
-    });
-
-    // 2.7 Extract Markdown tables: | h1 | h2 |\n|---|---| ... |
-    // 把整段表格替换为占位符，表格内容(已转义)放入 tableBlocks 数组
-    const tableBlocks: string[] = [];
-    text = text.replace(
-      /((?:^\|.+\|\r?\n)+^\|[\s\-:|]+\|\r?\n(?:^\|.+\|\r?\n?)+)/gm,
-      (_m: string, block: string) => {
-        const lines = block.trim().split(/\r?\n/).filter((l: string) => l.trim());
-        if (lines.length < 2) return block;
-        const headerCells = lines[0].split('|').slice(1, -1).map((c: string) => c.trim());
-        const alignCells = lines[1].split('|').slice(1, -1).map((c: string) => {
-          const t = c.trim();
-          const left = t.startsWith(':');
-          const right = t.endsWith(':');
-          if (left && right) return 'center';
-          if (right) return 'right';
-          if (left) return 'left';
-          return 'left';
-        });
-        const bodyRows = lines.slice(2).map((row: string) =>
-          row.split('|').slice(1, -1).map((c: string) => c.trim())
-        );
-        const thead = headerCells.map((c: string, i: number) =>
-          `<th style="text-align:${alignCells[i] || 'left'}">${c}</th>`
-        ).join('');
-        const tbody = bodyRows.map((r: string[]) =>
-          '<tr>' + r.map((c: string, i: number) =>
-            `<td style="text-align:${alignCells[i] || 'left'}">${c}</td>`
-          ).join('') + '</tr>'
-        ).join('');
-        const html = `<table class="md-table"><thead><tr>${thead}</tr></thead><tbody>${tbody}</tbody></table>`;
-        const idx = tableBlocks.length;
-        tableBlocks.push(html);
-        return '\x00TABLE' + idx + '\x00';
-      }
-    );
-
-    // 2.8 LaTeX 数学公式：$$...$$（块级）和 $...$（行内）
-    // 提取为占位符，渲染后用 KaTeX 处理
-    const mathBlocks: string[] = [];
-    text = text.replace(/\$\$([\s\S]+?)\$\$/g, (_m: string, formula: string) => {
-      try {
-        const rendered = katex.renderToString(formula.trim(), { displayMode: true, throwOnError: false });
-        const idx = mathBlocks.length;
-        mathBlocks.push(`<div class="md-math-block">${rendered}</div>`);
-        return '\x00MATHBLOCK' + idx + '\x00';
-      } catch { return _m; }
-    });
-    text = text.replace(/\$([^\s$][^$]*[^\s$])\$/g, (_m: string, formula: string) => {
-      try {
-        const rendered = katex.renderToString(formula.trim(), { displayMode: false, throwOnError: false });
-        const idx = mathBlocks.length;
-        mathBlocks.push(`<span class="md-math-inline">${rendered}</span>`);
-        return '\x00MATHBLOCK' + idx + '\x00';
-      } catch { return _m; }
-    });
-
-    // 2.9 Strikethrough: ~~text~~ → <del>text</del>
-    text = text.replace(/~~(.+?)~~/g, '<del>$1</del>');
-
-    // 2.10 Auto-link bare URLs: http(s)://... → <a>
-    text = text.replace(/(?<![="'>])https?:\/\/[^\s<>[\]()]+/g, (url: string) =>
-      `<a class="md-link" href="${url}" target="_blank" rel="noopener">${url}</a>`
-    );
-
-    // 3. Headers
-    text = text.replace(/^#### (.+)$/gm, '<h4 class="md-h4">$1</h4>');
-    text = text.replace(/^### (.+)$/gm, '<h4 class="md-h4">$1</h4>');
-    text = text.replace(/^## (.+)$/gm, '<h3 class="md-h3">$1</h3>');
-    text = text.replace(/^# (.+)$/gm, '<h2 class="md-h2">$1</h2>');
-
-    // 4. Horizontal rule
-    text = text.replace(/^---+$/gm, '<hr class="md-hr">');
-
-    // 5. Blockquotes
-    text = text.replace(/^> (.+)$/gm, '<blockquote class="md-blockquote">$1</blockquote>');
-
-    // 6. Links [text](url)
-    text = text.replace(/\[([^\]]+)\]\(([^)]+)\)/g, '<a class="md-link" href="$2" target="_blank" rel="noopener">$1</a>');
-
-    // 7. Bold and italic
-    text = text.replace(/\*\*\*(.+?)\*\*\*/g, '<strong><em>$1</em></strong>');
-    text = text.replace(/\*\*(.+?)\*\*/g, '<strong>$1</strong>');
-    text = text.replace(/(?<!\w)\*(?!\*)(.+?)(?<!\*)\*(?!\w)/g, '<em>$1</em>');
-
-    // 8. Lists
-    text = text.replace(/^\s*[-*] (.+)$/gm, '<li class="md-li">$1</li>');
-    text = text.replace(/^\s*\d+\. (.+)$/gm, '<li class="md-li">$1</li>');
-    // Wrap consecutive <li> in <ul>
-    text = text.replace(/((?:<li class="md-li">[\s\S]*?<\/li>\n?)+)/g, '<ul class="md-ul">$1</ul>');
-
-    // 9. Convert double newlines to paragraph breaks, single to <br/>
-    text = text.replace(/\n\n/g, '</p><p class="md-p">');
-    text = text.replace(/\n/g, '<br/>');
-
-    // 10. Restore code blocks and inline codes
-    text = text.replace(/\x00CODEBLOCK(\d+)\x00/g, (_m: string, idx: string) => codeBlocks[parseInt(idx)] || '');
-    text = text.replace(/\x00MERMAID(\d+)\x00/g, (_m: string, idx: string) => mermaidBlocks[parseInt(idx)] || '');
-    text = text.replace(/\x00INLINE(\d+)\x00/g, (_m: string, idx: string) => inlineCodes[parseInt(idx)] || '');
-    text = text.replace(/\x00TABLE(\d+)\x00/g, (_m: string, idx: string) => tableBlocks[parseInt(idx)] || '');
-    text = text.replace(/\x00MATHBLOCK(\d+)\x00/g, (_m: string, idx: string) => mathBlocks[parseInt(idx)] || '');
-
-    // 10.5 Apply syntax highlighting to code blocks + add copy button
-    text = text.replace(
-      /<pre class="md-code-block"><code class="md-code">([\s\S]*?)<\/code><\/pre>/g,
-      (_m: string, code: string) => {
-        const decoded = code.replace(/&lt;/g, '<').replace(/&gt;/g, '>').replace(/&amp;/g, '&');
-        let highlighted: string;
-        try { highlighted = hljs.highlightAuto(decoded).value; } catch { highlighted = code; }
-        return `<div class="md-code-wrapper">
-          <button class="md-code-copy" onclick="navigator.clipboard.writeText(this.parentElement.querySelector('code').textContent);this.textContent='已复制';setTimeout(()=>this.textContent='复制',2000)">复制</button>
-          <pre class="md-code-block"><code class="md-code hljs">${highlighted}</code></pre>
-        </div>`;
-      }
-    );
-
-    // 11. Wrap in paragraph and clean up
-    text = '<p class="md-p">' + text + '</p>';
-    text = text.replace(/<p class="md-p"><\/p>/g, '');
-    text = text.replace(/<p class="md-p">(<h[234][\s\S]*?<\/h[234]>)<\/p>/g, '$1');
-    text = text.replace(/<p class="md-p">(<pre[\s\S]*?<\/pre>)<\/p>/g, '$1');
-    text = text.replace(/<p class="md-p">(<ul[\s\S]*?<\/ul>)<\/p>/g, '$1');
-    text = text.replace(/<p class="md-p">(<hr[\s\S]*?)<\/p>/g, '$1');
-    text = text.replace(/<p class="md-p">(<blockquote[\s\S]*?<\/blockquote>)<\/p>/g, '$1');
-    text = text.replace(/<p class="md-p">(<table[\s\S]*?<\/table>)<\/p>/g, '$1');
-
-    // 12. 拆分为文本片段与媒体组件
-    const parts = text.split(/(\x00MEDIA\d+\x00)/g);
     return parts.map((part: string, i: number) => {
       const m = part.match(/^\x00MEDIA(\d+)\x00$/);
       if (m) {
-        const info = mediaBlocks[parseInt(m[1])];
+        const info = media[parseInt(m[1])];
         if (info && info.src) return <ChatMedia key={'m' + i} info={info} />;
         return null;
       }
@@ -329,7 +191,7 @@ const MarkdownContent: React.FC<{ content: string }> = ({ content }) => {
       if (!s.trim()) return null;
       return <div key={'t' + i} className="md-text-seg" dangerouslySetInnerHTML={{ __html: s }} />;
     });
-  }, [content]);
+  }, [html, media]);
 
   const containerRef = useRef<HTMLDivElement>(null);
   useEffect(() => {
@@ -387,14 +249,6 @@ const ClarifyCard: React.FC<{ data: { prompt: string; options: any[] }; onChoose
 };
 
 
-function escapeHtml(text: string): string {
-  return text
-    .replace(/&/g, '&amp;')
-    .replace(/</g, '&lt;')
-    .replace(/>/g, '&gt;')
-    .replace(/"/g, '&quot;');
-}
-
 // ===== Time Formatting =====
 function formatTime(iso: string): string {
   return formatChatTime(iso);
@@ -622,6 +476,149 @@ const Dropdown: React.FC<DropdownProps> = ({ trigger, children, className = '',
   );
 };
 
+// ===== AI 思考 / 执行过程面板（顶部展开 + 自动收拢） =====
+// autoExpand=true 时（本轮刚产生的助手消息）首屏展开呈现思考链路，
+// 6s 后若用户未手动干预则自动收拢为一行摘要，避免与回复内容冲突。
+// 展开时自动滚动到视图内，带脉冲动画提示用户注意。
+const AiProcessBlock: React.FC<{ steps: any[]; autoExpand: boolean }> = ({ steps, autoExpand }) => {
+  const [open, setOpen] = useState(autoExpand);
+  const [fresh, setFresh] = useState(autoExpand);
+  const timerRef = useRef<number | null>(null);
+  const touchedRef = useRef(false);
+  const blockRef = useRef<HTMLDetailsElement>(null);
+
+  useEffect(() => {
+    if (!autoExpand) {
+      setOpen(false);
+      setFresh(false);
+      return;
+    }
+    setOpen(true);
+    setFresh(true);
+    touchedRef.current = false;
+    // 自动滚动到视图内（上方留 60px 避免被顶部遮挡）
+    if (blockRef.current) {
+      blockRef.current.scrollIntoView({ behavior: 'smooth', block: 'start' });
+    }
+    timerRef.current = window.setTimeout(() => {
+      if (!touchedRef.current) {
+        setOpen(false);
+        // 收拢后短暂保留 fresh 样式再移除
+        setTimeout(() => setFresh(false), 500);
+      }
+    }, 6000);
+    return () => {
+      if (timerRef.current) window.clearTimeout(timerRef.current);
+    };
+  }, [autoExpand]);
+
+  if (!steps || steps.length === 0) return null;
+
+  const toggle = () => {
+    touchedRef.current = true;
+    setFresh(false);
+    if (timerRef.current) {
+      window.clearTimeout(timerRef.current);
+      timerRef.current = null;
+    }
+    setOpen(o => !o);
+  };
+
+  return (
+    <details ref={blockRef} className={`ai-process${open ? '' : ''}${fresh ? ' ai-process-fresh' : ''}`} open={open}>
+      <summary className="ai-process-summary" onClick={(e) => { e.preventDefault(); toggle(); }}>
+        <span className="ai-process-badge">🧠</span>
+        AI 思考 / 执行过程（{steps.length} 步 — {open ? '自动收拢中... 点击手动收起' : '点击展开查看'}）
+      </summary>
+      <div className="ai-process-list">
+        {steps.map((s: any, i: number) => {
+          const isLast = i === steps.length - 1;
+          const hasResult = s.result && s.result.length > 0;
+          const isError = hasResult && (s.result.includes('error') || s.result.includes('失败') || s.result.includes('异常') || s.result.includes('拦截'));
+          const isRunning = isLast && !hasResult;
+          const stepIcon = s.type === 'thinking' ? '💡' : s.type === 'skill' ? '🛠' : '🔌';
+          const statusIcon = isRunning ? ' ⏳' : isError ? ' ❌' : hasResult ? ' ✅' : '';
+          const resultDisplay = (s.result || '').length > 400 ? (s.result || '').slice(0, 400) + '...' : (s.result || '');
+          
+          return (
+          <div key={i} className={`ai-process-item ai-process-${s.type}${isRunning ? ' ai-process-running' : ''}${isError ? ' ai-process-error' : ''}`}>
+            <span className="ai-process-ico">{stepIcon}{statusIcon}</span>
+            <div className="ai-process-body">
+              {s.type === 'thinking' && (<><b className="ai-process-title">思考</b><p className="ai-process-text">{s.text}</p></>)}
+              {s.type === 'skill' && (<>
+                <b className="ai-process-title">调用技能：{s.name}{statusIcon}</b>
+                {s.args && Object.keys(s.args).length > 0 && (
+                  <span style={{fontSize:'12px',color:'#666',margin:'2px 0',display:'block'}}>
+                    参数: {JSON.stringify(s.args).slice(0, 200)}
+                  </span>
+                )}
+                {hasResult && <p className="ai-process-text">{resultDisplay}</p>}
+              </>)}
+              {s.type === 'mcp' && (<>
+                <b className="ai-process-title">MCP · {s.server}：{s.name}{statusIcon}</b>
+                {s.args && Object.keys(s.args).length > 0 && (
+                  <span style={{fontSize:'12px',color:'#666',margin:'2px 0',display:'block'}}>
+                    参数: {JSON.stringify(s.args).slice(0, 200)}
+                  </span>
+                )}
+                {hasResult && <p className="ai-process-text">{resultDisplay}</p>}
+              </>)}
+            </div>
+          </div>
+        )})}
+      </div>
+    </details>
+  );
+};
+
+// 思考过程收拢式展示（缺陷5：默认折叠；fresh 流式消息自动展开，数秒后自动收拢）
+const ReasoningBlock: React.FC<{ text: string; fresh: boolean }> = ({ text, fresh }) => {
+  const [open, setOpen] = useState(fresh);
+  const [isFresh, setFreshState] = useState(fresh);
+  const touchedRef = useRef(false);
+  const timerRef = useRef<number | null>(null);
+  useEffect(() => {
+    if (!fresh) {
+      setOpen(false);
+      setFreshState(false);
+      return;
+    }
+    setOpen(true);
+    setFreshState(true);
+    touchedRef.current = false;
+    timerRef.current = window.setTimeout(() => {
+      if (!touchedRef.current) {
+        setOpen(false);
+        setTimeout(() => setFreshState(false), 500);
+      }
+    }, 5000);
+    return () => { if (timerRef.current) window.clearTimeout(timerRef.current); };
+  }, [fresh]);
+  if (!text) return null;
+  const toggle = () => {
+    touchedRef.current = true;
+    setFreshState(false);
+    if (timerRef.current) window.clearTimeout(timerRef.current);
+    setOpen(o => !o);
+  };
+  return (
+    <details className={`ai-process ai-process-reasoning${isFresh ? ' ai-process-fresh' : ''}`} open={open}>
+      <summary className="ai-process-summary" onClick={(e) => { e.preventDefault(); toggle(); }}>
+        <span className="ai-process-badge">🤔</span>
+        AI 思考过程{open ? '（收拢中... 点击收起）' : '（点击展开）'}
+      </summary>
+      <div className="ai-process-list">
+        <div className="ai-process-item ai-process-thinking">
+          <span className="ai-process-ico">💡</span>
+          <div className="ai-process-body">
+            <p className="ai-process-text">{text}</p>
+          </div>
+        </div>
+      </div>
+    </details>
+  );
+};
+
 // ===== Main Component =====
 const ChatPage: React.FC = () => {
   const { id } = useParams<{ id?: string }>();
@@ -631,10 +628,19 @@ const ChatPage: React.FC = () => {
   const [chatModels, setChatModels] = useState<any[]>([]);
   const [inputValue, setInputValue] = useState('');
   const [sending, setSending] = useState(false);
+  const [sendingTimeout, setSendingTimeout] = useState(false);
+  const sendingTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
+  // 正在 SSE 流式渲染的助手消息 id；非空时隐藏"思考中"占位气泡
+  const [streamingId, setStreamingId] = useState<string | null>(null);
+  // 流式中断控制器（缺陷4：支持「停止生成」）
+  const abortRef = useRef<AbortController | null>(null);
+  const stopStreaming = () => { try { abortRef.current?.abort(); } catch { /* noop */ } };
   const [conversations, setConversations] = useState<Conversation[]>([]);
   const [messages, setMessages] = useState<Message[]>([]);
   const [loadingConversations, setLoadingConversations] = useState(true);
   const [loadingMessages, setLoadingMessages] = useState(false);
+  // 本轮对话新产生的助手消息 id 集合，用于触发"思考过程"首屏展开
+  const [freshMsgIds, setFreshMsgIds] = useState<Set<string>>(new Set());
   const [knowledgeBases, setKnowledgeBases] = useState<any[]>([]);
   const [agents, setAgents] = useState<any[]>([]);
   const [skills, setSkills] = useState<any[]>([]);
@@ -1003,11 +1009,227 @@ const ChatPage: React.FC = () => {
     }
   };
 
-  // Handle send message - CALLS REAL API
+  // 非流式兜底发送（与历史行为一致）——供 SSE 连接/流式失败时回退
+  const sendLegacy = async (
+    payload: any,
+    tempUserMsg: Message,
+  ) => {
+    const res = await chatAPI.sendMessage(selectedConversation!.id, payload);
+    const data = unwrapData(res);
+
+    // Replace temporary user message and add assistant response
+    setMessages(prev => {
+      const filtered = prev.filter(m => m.id !== tempUserMsg.id);
+      const newMsgs: Message[] = [];
+      if (data.user_message) {
+        newMsgs.push({
+          id: data.user_message.id,
+          conversationId: selectedConversation?.id || '',
+          role: 'user',
+          content: data.user_message.content,
+          tokenUsed: data.user_message.token_used || 0,
+          createdAt: data.user_message.created_at || new Date().toISOString(),
+        });
+      }
+      if (data.assistant_message) {
+        newMsgs.push({
+          id: data.assistant_message.id,
+          conversationId: selectedConversation?.id || '',
+          role: 'assistant',
+          content: data.assistant_message.content,
+          tokenUsed: data.assistant_message.token_used || 0,
+          createdAt: data.assistant_message.created_at || new Date().toISOString(),
+          used_skills: data.assistant_message.used_skills || [],
+          used_mcp: data.assistant_message.used_mcp || [],
+          process_steps: data.assistant_message.process_steps || [],
+        });
+      }
+      return [...filtered, ...newMsgs];
+    });
+
+    // 标记本轮新产生的助手消息，触发"思考过程"首屏展开
+    if (data.assistant_message) {
+      setFreshMsgIds(prev => new Set(prev).add(data.assistant_message.id));
+    }
+
+    // 媒体异步生成：启动轮询，待后端回写后自动刷新
+    if (data.assistant_message?.generating) {
+      startMediaPolling(selectedConversation!.id, data.assistant_message.id);
+    }
+
+    // Refresh conversations to update lastMessage
+    loadConversations();
+    // 发送后清空本次附件（选择状态仍保留）
+    setAttachments([]);
+
+    // GAP-07: 对话结束后异步提交轨迹到进化引擎（后台静默，不阻塞UI）
+    if (selectedConversation?.id && data.assistant_message) {
+      const recentMsgs = [
+        ...(data.user_message ? [{ role: 'user', content: data.user_message.content, tokens: data.user_message.token_used || 0 }] : []),
+        { role: 'assistant', content: data.assistant_message.content, tokens: data.assistant_message.token_used || 0 },
+      ];
+      evolutionAPI.submitTrajectory({
+        conversation_id: selectedConversation.id,
+        messages: recentMsgs,
+      }).catch(() => {});
+    }
+  };
+
+  // SSE 流式发送（Phase 0）——实时增量渲染 AI 回复文本与思考链(reasoning)
+  const streamSend = async (
+    convId: string,
+    payload: any,
+    tempUserMsg: Message,
+    assistantId: string,
+  ): Promise<void> => {
+    const token = localStorage.getItem('aiflow_token');
+    const controller = new AbortController();
+    abortRef.current = controller;
+    const resp = await fetch(`/app/api/chat/conversations/${convId}/messages/stream`, {
+      method: 'POST',
+      headers: {
+        'Content-Type': 'application/json',
+        ...(token ? { Authorization: `Bearer ${token}` } : {}),
+      },
+      body: JSON.stringify(payload),
+      signal: controller.signal,
+    });
+    if (!resp.ok || !resp.body) {
+      throw new Error(`SSE 连接失败 (${resp.status})`);
+    }
+
+    const reader = resp.body.getReader();
+    const decoder = new TextDecoder();
+    let buffer = '';
+    let accReasoning = '';
+    let gotDone = false;
+    let pendingDelta = '';
+    let reasoningDirty = false;
+    let rafId: number | null = null;
+
+    // 缺陷4：用 requestAnimationFrame 合并同一帧内的多个 delta/reasoning，避免逐 token 重渲染抖动
+    const flushBuffered = () => {
+      rafId = null;
+      if (!pendingDelta && !reasoningDirty) return;
+      const piece = pendingDelta;
+      pendingDelta = '';
+      const r = reasoningDirty ? accReasoning : null;
+      reasoningDirty = false;
+      setMessages(prev => prev.map(m => {
+        if (m.id !== assistantId) return m;
+        const patch: any = {};
+        if (piece) patch.content = (m.content || '') + piece;
+        if (r !== null) patch.reasoning = r;
+        return { ...m, ...patch };
+      }));
+    };
+    const scheduleFlush = () => {
+      if (rafId == null) rafId = requestAnimationFrame(flushBuffered);
+    };
+    const finalizeFlush = () => {
+      if (rafId != null) { cancelAnimationFrame(rafId); rafId = null; }
+      flushBuffered();
+    };
+
+    while (true) {
+      const { value, done } = await reader.read();
+      if (done) break;
+      buffer += decoder.decode(value, { stream: true });
+      let sep: number;
+      while ((sep = buffer.indexOf('\n\n')) >= 0) {
+        const raw = buffer.slice(0, sep);
+        buffer = buffer.slice(sep + 2);
+        const dataLine = raw.split('\n').find(l => l.startsWith('data: '));
+        if (!dataLine) continue;
+        let evt: any;
+        try { evt = JSON.parse(dataLine.slice(6)); } catch { continue; }
+        if (evt.type === 'delta') {
+          pendingDelta += evt.content || '';
+          scheduleFlush();
+        } else if (evt.type === 'reasoning') {
+          accReasoning += evt.content || '';
+          reasoningDirty = true;
+          scheduleFlush();
+        } else if (evt.type === 'step') {
+          // 缺陷2：实时增量过程步骤，前端随事件收拢式渲染
+          const step = evt.step;
+          const idx = typeof evt.index === 'number' ? evt.index : -1;
+          setMessages(prev => prev.map(m => {
+            if (m.id !== assistantId) return m;
+            const steps: any[] = (m as any).process_steps || [];
+            const next = steps.slice();
+            if (idx >= 0 && idx < next.length) next[idx] = step;
+            else next.push(step);
+            return { ...m, process_steps: next };
+          }));
+        } else if (evt.type === 'done') {
+          gotDone = true;
+          finalizeFlush();
+          const finalId = evt.message_id || assistantId;
+          setMessages(prev => prev.map(m => {
+            if (m.id !== assistantId) return m;
+            return {
+              ...m,
+              id: finalId,
+              content: evt.content ?? m.content,
+              process_steps: evt.process_steps || [],
+              used_skills: evt.used_skills || [],
+              used_mcp: evt.used_mcp || [],
+              generating: !!evt.generating,
+            };
+          }));
+          setStreamingId(null);
+          abortRef.current = null;
+          if (evt.generating) startMediaPolling(convId, finalId);
+          setFreshMsgIds(prev => new Set(prev).add(finalId));
+          loadConversations();
+          setAttachments([]);
+          if (convId && evt.content) {
+            evolutionAPI.submitTrajectory({
+              conversation_id: convId,
+              messages: [
+                { role: 'user', content: tempUserMsg.content, tokens: 0 },
+                { role: 'assistant', content: evt.content, tokens: 0 },
+              ],
+            }).catch(() => {});
+          }
+        } else if (evt.type === 'error') {
+          // 缺陷4：容错——保留已生成内容，不强删，标记中断后优雅收尾
+          finalizeFlush();
+          setMessages(prev => prev.map(m => {
+            if (m.id !== assistantId) return m;
+            const note = '\n\n_(生成中断：' + (evt.message || '未知错误') + ')_';
+            return { ...m, content: (m.content || '') + note };
+          }));
+          setStreamingId(null);
+          abortRef.current = null;
+          setFreshMsgIds(prev => new Set(prev).add(assistantId));
+          loadConversations();
+          gotDone = true;
+          return;
+        }
+      }
+    }
+    if (!gotDone) {
+      // 流结束但未收到 done：已累积的 delta 仍可展示，但视为异常兜底
+      finalizeFlush();
+      abortRef.current = null;
+      setStreamingId(null);
+      throw new Error('流式响应未正常结束');
+    }
+  };
+
+  // Handle send message - CALLS REAL API（Phase 0：默认 SSE 流式，失败回退非流式）
   // 抽出公共发送逻辑，供输入框发送与「模糊创作意图」方向选择复用
   const sendText = async (content: string) => {
     if (!content.trim() || sending || !selectedConversation) return;
     setSending(true);
+    setSendingTimeout(false);
+    // 60s 超时警告
+    if (sendingTimerRef.current) clearTimeout(sendingTimerRef.current);
+    sendingTimerRef.current = setTimeout(() => {
+      setSendingTimeout(true);
+    }, 60000);
 
     // Add user message optimistically
     const tempUserMsg: Message = {
@@ -1020,78 +1242,48 @@ const ChatPage: React.FC = () => {
     };
     setMessages(prev => [...prev, tempUserMsg]);
 
+    const payload = {
+      content,
+      model_id: selectedModelId || undefined,
+      agent_ids: selectedAgentIds.length > 0 ? selectedAgentIds : undefined,
+      skill_ids: selectedSkillIds.length > 0 ? selectedSkillIds : undefined,
+      knowledge_base_ids: selectedKbIds.length > 0 ? selectedKbIds : undefined,
+      mcp_ids: selectedMcpIds.length > 0 ? selectedMcpIds : undefined,
+      attachments: attachments.length > 0 ? attachments.map(a => ({
+        name: a.name, file_type: a.file_type, content: a.content || '',
+      })) : undefined,
+      use_tools: selectedSkillIds.length > 0 || selectedAgentIds.length > 0 || selectedMcpIds.length > 0,
+    };
+
+    // 先放一个空的助手消息占位，用于 SSE 流式增量渲染
+    const assistantId = `temp_assistant_${Date.now()}`;
+    setMessages(prev => [...prev, {
+      id: assistantId,
+      conversationId: selectedConversation!.id,
+      role: 'assistant',
+      content: '',
+      tokenUsed: 0,
+      createdAt: new Date().toISOString(),
+      reasoning: '',
+    } as Message]);
+    setStreamingId(assistantId);
+
     try {
-      const res = await chatAPI.sendMessage(selectedConversation!.id, {
+      await streamSend(selectedConversation!.id, payload, tempUserMsg, assistantId);
-        content,
-        model_id: selectedModelId || undefined,
-        agent_ids: selectedAgentIds.length > 0 ? selectedAgentIds : undefined,
-        skill_ids: selectedSkillIds.length > 0 ? selectedSkillIds : undefined,
-        knowledge_base_ids: selectedKbIds.length > 0 ? selectedKbIds : undefined,
-        mcp_ids: selectedMcpIds.length > 0 ? selectedMcpIds : undefined,
-        attachments: attachments.length > 0 ? attachments.map(a => ({
-          name: a.name, file_type: a.file_type, content: a.content || '',
-        })) : undefined,
-        use_tools: selectedSkillIds.length > 0 || selectedAgentIds.length > 0 || selectedMcpIds.length > 0,
-      });
-
-      const data = unwrapData(res);
-
-      // Replace temporary user message and add assistant response
-      setMessages(prev => {
-        const filtered = prev.filter(m => m.id !== tempUserMsg.id);
-        const newMsgs: Message[] = [];
-        if (data.user_message) {
-          newMsgs.push({
-            id: data.user_message.id,
-            conversationId: selectedConversation?.id || '',
-            role: 'user',
-            content: data.user_message.content,
-            tokenUsed: data.user_message.token_used || 0,
-            createdAt: data.user_message.created_at || new Date().toISOString(),
-          });
-        }
-        if (data.assistant_message) {
-          newMsgs.push({
-            id: data.assistant_message.id,
-            conversationId: selectedConversation?.id || '',
-            role: 'assistant',
-            content: data.assistant_message.content,
-            tokenUsed: data.assistant_message.token_used || 0,
-            createdAt: data.assistant_message.created_at || new Date().toISOString(),
-            used_skills: data.assistant_message.used_skills || [],
-            used_mcp: data.assistant_message.used_mcp || [],
-          });
-        }
-        return [...filtered, ...newMsgs];
-      });
-
-      // 媒体异步生成：启动轮询，待后端回写后自动刷新
-      if (data.assistant_message?.generating) {
-        startMediaPolling(selectedConversation!.id, data.assistant_message.id);
-      }
-
-      // Refresh conversations to update lastMessage
-      loadConversations();
-      // 发送后清空本次附件（选择状态仍保留）
-      setAttachments([]);
-
-      // GAP-07: 对话结束后异步提交轨迹到进化引擎（后台静默，不阻塞UI）
-      if (selectedConversation?.id && data.assistant_message) {
-        const recentMsgs = [
-          ...(data.user_message ? [{ role: 'user', content: data.user_message.content, tokens: data.user_message.token_used || 0 }] : []),
-          { role: 'assistant', content: data.assistant_message.content, tokens: data.assistant_message.token_used || 0 },
-        ];
-        evolutionAPI.submitTrajectory({
-          conversation_id: selectedConversation.id,
-          messages: recentMsgs,
-        }).catch(() => {});
-      }
     } catch (err: any) {
-      // Remove optimistic message on error
+      // 移除流式占位，回退非流式重新发送
-      setMessages(prev => prev.filter(m => m.id !== tempUserMsg.id));
+      setMessages(prev => prev.filter(m => m.id !== assistantId && m.id !== tempUserMsg.id));
-      alert('发送失败: ' + (err.message || err.detail || '未知错误'));
+      setStreamingId(null);
+      try {
+        await sendLegacy(payload, tempUserMsg);
+      } catch (e2: any) {
+        setMessages(prev => prev.filter(m => m.id !== tempUserMsg.id));
+        alert('发送失败: ' + (e2?.message || e2?.detail || '未知错误'));
+      }
     } finally {
+      if (sendingTimerRef.current) { clearTimeout(sendingTimerRef.current); sendingTimerRef.current = null; }
       setSending(false);
+      setSendingTimeout(false);
       inputRef.current?.focus();
     }
   };
@@ -1333,18 +1525,38 @@ const ChatPage: React.FC = () => {
                 // Real used_skills from API response
                 const usedSkills: string[] = (msg.role === 'assistant' && msg.used_skills) ? msg.used_skills : [];
                 const usedMcp: string[] = (msg.role === 'assistant' && msg.used_mcp) ? msg.used_mcp : [];
+                const steps: any[] = (msg as any).process_steps || [];
                 // 模糊创作意图确认卡片
                 const clarify = msg.role === 'assistant' ? parseClarify(msg.content) : null;
                 return (
                 <div key={msg.id} className={`message-row ${msg.role}`}>
+                  {msg.role === 'assistant' && steps.length > 0 && (
+                    <AiProcessBlock steps={steps} autoExpand={freshMsgIds.has(msg.id)} />
+                  )}
+                  {msg.role === 'assistant' && (msg as any).reasoning && (
+                    <ReasoningBlock text={(msg as any).reasoning} fresh={freshMsgIds.has(msg.id)} />
+                  )}
                   <div className={`message-bubble ${msg.role}`}>
                     {clarify ? (
                       <ClarifyCard
                         data={clarify}
                         onChoose={(opt: any) => sendText(`${opt.medium}：${clarify.prompt}`)}
                       />
+                    ) : msg.role === 'assistant' && isAiError(msg.content) ? (
+                      <div className="ai-error-bubble">
+                        <div className="ai-error-icon">⚠️</div>
+                        <div className="ai-error-text">
+                          <MarkdownContent content={msg.content} />
+                        </div>
+                        <button className="ai-error-retry" onClick={() => {
+                          setMessages(prev => prev.filter(m => m.id !== msg.id));
+                          const msgsAfterFilter = messages.filter(m => m.id !== msg.id);
+                          const lastUser = [...msgsAfterFilter].reverse().find(m => m.role === 'user');
+                          if (lastUser) sendText(lastUser.content);
+                        }}>🔄 重试</button>
+                      </div>
                     ) : msg.role === 'assistant' ? (
-                      <MarkdownContent content={msg.content} />
+                      msg.content ? <MarkdownContent content={msg.content} /> : (streamingId === msg.id ? <span className="streaming-hint"><IconSpinner /> 生成中…</span> : null)
                     ) : (
                       <span>{msg.content}</span>
                     )}
@@ -1408,11 +1620,20 @@ const ChatPage: React.FC = () => {
                   );
                 })
               )}
-              {sending && (
+              {sending && !streamingId && (
                 <div className="message-row assistant">
-                  <div className="message-bubble assistant" style={{ display: 'flex', alignItems: 'center', gap: 8 }}>
+                  <div className={`message-bubble assistant${sendingTimeout ? ' ai-sending-timeout' : ''}`} style={{ display: 'flex', alignItems: 'center', gap: 8 }}>
                     <IconSpinner />
-                    <span>思考中...</span>
+                    <span>{sendingTimeout ? '⏳ 响应时间较长，AI 仍在处理中...' : '思考中...'}</span>
+                    {sendingTimeout && (
+                      <button className="ai-cancel-btn" onClick={() => {
+                        setSending(false);
+                        setSendingTimeout(false);
+                        if (sendingTimerRef.current) clearTimeout(sendingTimerRef.current);
+                      }} title="取消等待" style={{ marginLeft: 12, padding: '2px 8px', fontSize: 12, cursor: 'pointer', background: 'var(--color-bg-tertiary)', border: '1px solid var(--color-border)', borderRadius: 4, color: 'var(--color-text-secondary)' }}>
+                        取消
+                      </button>
+                    )}
                   </div>
                 </div>
               )}
@@ -1603,11 +1824,11 @@ const ChatPage: React.FC = () => {
                   </button>
                   <button
                     className="send-btn"
-                    onClick={handleSend}
+                    onClick={streamingId ? stopStreaming : handleSend}
-                    disabled={!inputValue.trim() || sending}
+                    disabled={!streamingId && (!inputValue.trim() || sending)}
-                    title="发送 (Enter)"
+                    title={streamingId ? '停止生成' : '发送 (Enter)'}
                   >
-                    {sending ? <IconSpinner /> : <IconSend />}
+                    {streamingId ? <IconStop /> : (sending ? <IconSpinner /> : <IconSend />)}
                   </button>
                 </div>
               </div>

frontend/src/pages/front/EvolutionDashboard.css

@@ -1000,51 +1000,7 @@
   margin-bottom: var(--space-lg);
 }
 
-.evo-anim-toggle {
+/* 技能图谱动画已在 #717 中移除（边框光晕溢出且无实际用途） */
-  display: flex;
-  align-items: center;
-  gap: var(--space-sm);
-  font-size: var(--text-xs);
-  color: var(--color-text-tertiary);
-  cursor: pointer;
-  user-select: none;
-}
-
-.evo-anim-toggle input {
-  width: 14px;
-  height: 14px;
-  accent-color: var(--color-accent);
-  cursor: pointer;
-}
-
-/* Baby sprout animation */
-.evo-skill-node.baby.anim-on .evo-node-dot {
-  animation: baby-sprout 2s ease-in-out infinite;
-}
-
-.evo-skill-node.baby.anim-on {
-  animation: baby-glow 2s ease-in-out infinite;
-}
-
-@keyframes baby-sprout {
-  0%, 100% { transform: scale(1); }
-  50% { transform: scale(1.15); }
-}
-
-@keyframes baby-glow {
-  0%, 100% { box-shadow: 0 0 0 0 rgba(245, 158, 11, 0.3); }
-  50% { box-shadow: 0 0 16px 4px rgba(245, 158, 11, 0); }
-}
-
-/* Staging breathe animation */
-.evo-skill-node.staging.anim-on {
-  animation: staging-breathe 3s ease-in-out infinite;
-}
-
-@keyframes staging-breathe {
-  0%, 100% { transform: scale(1); opacity: 0.85; }
-  50% { transform: scale(1.06); opacity: 1; }
-}
 
 /* Toast */
 .evo-toast {
@@ -1535,47 +1491,7 @@
   }
 }
 
-/* ===== v2: Baby Sprout Animation (GAP-03) ===== */
+/* ===== v2: 技能图谱动画已在 #717 中移除（边框溢出且无实际用途） ===== */
-.evo-skill-node.state-baby.anim-on {
-  animation: baby-sprout 2s ease-in-out infinite;
-}
-
-@keyframes baby-sprout {
-  0%, 100% {
-    transform: scale(1);
-    box-shadow: 0 0 0 0 rgba(255, 182, 193, 0.4);
-  }
-  50% {
-    transform: scale(1.15);
-    box-shadow: 0 0 20px 4px rgba(255, 182, 193, 0);
-  }
-}
-
-/* Baby skill sprout icon */
-.evo-skill-node.state-baby.anim-on .evo-node-dot::after {
-  content: '🌱';
-  position: absolute;
-  top: -18px;
-  left: 50%;
-  transform: translateX(-50%);
-  font-size: 14px;
-  animation: sprout-wiggle 1.5s ease-in-out infinite;
-}
-
-@keyframes sprout-wiggle {
-  0%, 100% { transform: translateX(-50%) rotate(-5deg); }
-  50% { transform: translateX(-50%) rotate(5deg); }
-}
-
-/* Staging breathe animation */
-.evo-skill-node.state-staging.anim-on {
-  animation: staging-breathe 3s ease-in-out infinite;
-}
-
-@keyframes staging-breathe {
-  0%, 100% { transform: scale(1); opacity: 0.8; }
-  50% { transform: scale(1.08); opacity: 1; }
-}
 
 /* ===== v2: Toast messages ===== */
 .evo-toast {

frontend/src/pages/front/EvolutionDashboard.tsx

@@ -62,9 +62,7 @@ const shareThemes = [
 // ==================== Panel A: Skill Tree ====================
 const SkillTreePanel: React.FC<{
   skills: EvolutionSkill[];
-  animEnabled: boolean;
+}> = ({ skills }) => {
-  onToggleAnim: () => void;
-}> = ({ skills, animEnabled, onToggleAnim }) => {
   const [hoveredSkill, setHoveredSkill] = useState<string | null>(null);
   const [selectedSkill, setSelectedSkill] = useState<EvolutionSkill | null>(null);
   const [showVersionHistory, setShowVersionHistory] = useState(false);
@@ -136,14 +134,6 @@ const SkillTreePanel: React.FC<{
       {/* Publish toast */}
       {publishMsg && <div className="evo-toast evo-toast-info">{publishMsg}</div>}
 
-      {/* Anim Toggle */}
-      <div className="evo-panel-toolbar">
-        <label className="evo-anim-toggle">
-          <input type="checkbox" checked={animEnabled} onChange={onToggleAnim} />
-          <span>动画效果</span>
-        </label>
-      </div>
-
       {/* Skill detail modal */}
       {selectedSkill && !showVersionHistory && (
         <div className="evo-skill-detail-modal" onClick={() => setSelectedSkill(null)}>
@@ -305,7 +295,7 @@ const SkillTreePanel: React.FC<{
                   <React.Fragment key={skill.id}>
                     {idx > 0 && <div className="evo-node-connector" />}
                     <div
-                      className={`evo-skill-node ${skill.state} ${isHovered ? 'hovered' : ''} ${animEnabled ? 'anim-on' : ''}`}
+                      className={`evo-skill-node ${skill.state} ${isHovered ? 'hovered' : ''}`}
                       onMouseEnter={() => setHoveredSkill(skill.id)}
                       onMouseLeave={() => setHoveredSkill(null)}
                       onClick={() => setSelectedSkill(skill)}
@@ -944,22 +934,10 @@ const EvolutionDashboard: React.FC = () => {
   const [loading, setLoading] = useState(true);
   const [error, setError] = useState('');
   const [globalSelectedSkill, setGlobalSelectedSkill] = useState<EvolutionSkill | null>(null);
-  const [animEnabled, setAnimEnabled] = useState(() => {
-    return localStorage.getItem('evo_anim_enabled') !== 'false';
-  });
-
   const totalSkills = skills.length;
   const mainCount = skills.filter(s => s.state === 'main').length;
   const babyCount = skills.filter(s => s.state === 'baby').length;
 
-  const toggleAnim = useCallback(() => {
-    setAnimEnabled(prev => {
-      const next = !prev;
-      localStorage.setItem('evo_anim_enabled', String(next));
-      return next;
-    });
-  }, []);
-
   const handleViewSkill = (skillName: string) => {
     const skill = skills.find(s => s.skill_name === skillName);
     if (skill) {
@@ -1071,7 +1049,7 @@ const EvolutionDashboard: React.FC = () => {
           ))}
         </aside>
         <main className="evo-main">
-          {activePanel === 'skill-tree' && <SkillTreePanel skills={skills} animEnabled={animEnabled} onToggleAnim={toggleAnim} />}
+          {activePanel === 'skill-tree' && <SkillTreePanel skills={skills} />}
           {activePanel === 'timeline' && <TimelinePanel events={events} onViewSkill={handleViewSkill} />}
           {activePanel === 'statistics' && (stats ? <StatisticsPanel stats={stats} onViewSkill={handleViewSkill} /> : <div className="evo-panel-content"><p>加载中...</p></div>)}
           {activePanel === 'profile' && (profile ? <ProfilePanel profile={profile} skills={skills} /> : <div className="evo-panel-content"><p>加载中...</p></div>)}

frontend/src/pages/front/MarketPage.css

@@ -1805,3 +1805,23 @@
   font-weight: 400;
   color: var(--color-text-tertiary);
 }
+
+/* 官方内置技能徽标（见 #716：官方导入并发布的技能默认全局已安装） */
+.modal-official-badge,
+.market-card-official-badge {
+  display: inline-flex;
+  align-items: center;
+  padding: 4px 12px;
+  font-size: 12px;
+  font-weight: 600;
+  line-height: 18px;
+  border-radius: 8px;
+  color: var(--color-accent);
+  background: color-mix(in srgb, var(--color-accent) 12%, transparent);
+  border: 1px solid color-mix(in srgb, var(--color-accent) 35%, transparent);
+  white-space: nowrap;
+}
+
+.modal-official-badge {
+  margin-left: auto;
+}

frontend/src/pages/front/MarketPage.tsx

@@ -337,6 +337,9 @@ const AppDetailModal: React.FC<AppDetailModalProps> = ({ app, onClose, onAppChan
     setInstalled(!!app.installed);
   }, [app.id, app.installed]);
 
+  // 官方导入并发布的技能默认全局已安装，前端展示为「官方内置」（见 #716）
+  const isOfficialBuiltin = app.source === 'official' && app.status === 'published';
+
   useEffect(() => {
     // 加载评论
     marketAPI.getReviews(app.id).then((res: any) => {
@@ -463,7 +466,9 @@ const AppDetailModal: React.FC<AppDetailModalProps> = ({ app, onClose, onAppChan
                 <span className="modal-price">{app.price} 元/月</span>
               )}
             </div>
-            {installed ? (
+            {isOfficialBuiltin ? (
+              <span className="modal-official-badge">官方内置</span>
+            ) : installed ? (
               <button className="modal-install-btn uninstall" onClick={handleUninstall} disabled={uninstalling}>
                 <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="2" strokeLinecap="round" strokeLinejoin="round">
                   <polyline points="3 6 5 6 21 6"/><path d="M19 6v14a2 2 0 0 1-2 2H7a2 2 0 0 1-2-2V6m3 0V4a2 2 0 0 1 2-2h4a2 2 0 0 1 2 2v2"/>
@@ -1949,7 +1954,9 @@ const MarketPage: React.FC = () => {
                   <div className="market-card-bottom">
                     <StarRating rating={app.ratingAvg} count={app.reviewCount} />
                     <div className="market-card-right">
-                      {app.installed ? (
+                      {app.source === 'official' && app.status === 'published' ? (
+                        <span className="market-card-official-badge">官方内置</span>
+                      ) : app.installed ? (
                         <button
                           className="market-card-uninstall-btn"
                           onClick={(e) => {

frontend/src/pages/front/TokenDashboardPage.css

@@ -3,6 +3,12 @@
   padding: var(--space-3xl) var(--space-4xl);
   max-width: 1280px;
   margin: 0 auto;
+  /* 见 #714：FrontLayout 的 .front-main 裁剪了主内容区，
+     此处让页面自身在可用高度内纵向滚动，避免内容被截断无法上下查阅 */
+  height: 100%;
+  min-height: 0;
+  overflow-y: auto;
+  overflow-x: hidden;
   animation: fadeIn var(--transition-normal);
 }
 
@@ -329,6 +335,8 @@
 
 .tkd-table-scroll {
   overflow-x: auto;
+  overflow-y: auto;
+  max-height: 60vh;
   padding: var(--space-lg) var(--space-xl) var(--space-sm);
 }
 

frontend/src/styles/global.css

@@ -112,6 +112,7 @@ html {
   -webkit-font-smoothing: antialiased;
   -moz-osx-font-smoothing: grayscale;
   text-rendering: optimizeLegibility;
+  overflow-x: hidden;
 }
 
 body {
@@ -120,6 +121,7 @@ body {
   background: var(--color-bg-primary);
   line-height: 1.6;
   overflow: hidden;
+  overflow-x: hidden;
   height: 100vh;
 }
 

frontend/src/types/index.ts

@@ -60,8 +60,16 @@ export interface Message {
   content: string;
   tokenUsed: number;
   createdAt: string;
+  reasoning?: string;
   used_skills?: string[];
   used_mcp?: string[];
+  process_steps?: Array<{
+    type: 'thinking' | 'skill' | 'mcp';
+    text?: string;
+    name?: string;
+    server?: string;
+    result?: string;
+  }>;
 }
 
 

frontend/src/utils/markdown.test.ts

@@ -0,0 +1,80 @@
+import { describe, it, expect } from 'vitest';
+import { renderMarkdownToHtml } from './markdown';
+
+// 前台聊天 Markdown 渲染曾两次被 bug 击穿：
+//  1) 段落内单个 \n 被转成 <br/>，导致 AI 回复被强制大量换行、内容撑成多行；
+//  2) 表格在加粗/链接转换之前被提取，单元格内 **xxx** 原样显示、且 <br> 未压缩。
+// 本测试固化「渲染管线」契约，防止回归。管线代码见 ./markdown.ts（从 ChatPage 抽出）。
+describe('renderMarkdownToHtml', () => {
+  it('段落内单个 \\n 转为空格（而非 <br/>）', () => {
+    const { html } = renderMarkdownToHtml('第一行\n第二行');
+    expect(html).not.toContain('<br');
+    expect(html).toContain('第一行 第二行');
+  });
+
+  it('双换行转为段落分隔', () => {
+    const { html } = renderMarkdownToHtml('段一\n\n段二');
+    expect(html).toContain('段一</p><p class="md-p">段二');
+  });
+
+  it('3+ 连续空行压缩为 2 个（与 2 个空行结果一致）', () => {
+    const many = renderMarkdownToHtml('a\n\n\n\n\nb').html;
+    const few = renderMarkdownToHtml('a\n\nb').html;
+    expect(many).toBe(few);
+    expect(many).toContain('a</p><p class="md-p">b');
+  });
+
+  it('表格在加粗之后提取，单元格内的 **xxx** 被渲染为 <strong>', () => {
+    const input = '| 能力 | 说明 |\n| --- | --- |\n| **渠道排序** | 按优先级 |\n';
+    const { html } = renderMarkdownToHtml(input);
+    expect(html).toContain('<table class="md-table">');
+    expect(html).toContain('<th');
+    expect(html).toContain('<td');
+    expect(html).toContain('<strong>渠道排序</strong>');
+  });
+
+  it('单元格内连续 <br> 被压缩为单个', () => {
+    const input = '| 列 |\n| --- |\n| a<br/><br/>b |\n';
+    const { html } = renderMarkdownToHtml(input);
+    expect(html).toContain('a<br/>b');
+    expect(html).not.toContain('<br/><br/>');
+  });
+
+  it('无序列表渲染为 <ul><li>', () => {
+    const { html } = renderMarkdownToHtml('- a\n- b');
+    expect(html).toContain('<ul class="md-ul">');
+    expect(html).toContain('<li class="md-li">a</li>');
+    expect(html).toContain('<li class="md-li">b</li>');
+  });
+
+  it('一级标题渲染为 h2', () => {
+    const { html } = renderMarkdownToHtml('# Title');
+    expect(html).toContain('<h2 class="md-h2">Title</h2>');
+  });
+
+  it('代码块内容不被行内加粗转换', () => {
+    const { html } = renderMarkdownToHtml('```\n**not bold**\n```');
+    expect(html).toContain('<pre class="md-code-block">');
+    expect(html).not.toContain('<strong>not bold</strong>');
+  });
+
+  it('行内代码被抽取保护', () => {
+    const { html } = renderMarkdownToHtml('use `npm run` here');
+    expect(html).toContain('<code class="md-inline-code">npm run</code>');
+  });
+
+  it('[文本](url) 链接渲染', () => {
+    const { html } = renderMarkdownToHtml('[Google](https://g.com)');
+    expect(html).toContain('<a class="md-link" href="https://g.com"');
+    expect(html).toContain('Google');
+  });
+
+  it('媒体(<img>)以占位符保留并返回结构化 media 列表', () => {
+    const { html, media } = renderMarkdownToHtml('看这张图 <img src="https://x.com/a.png" alt="图">');
+    expect(media.length).toBe(1);
+    expect(media[0].type).toBe('img');
+    expect(media[0].src).toBe('https://x.com/a.png');
+    expect(media[0].alt).toBe('图');
+    expect(html).toContain('\x00MEDIA0\x00');
+  });
+});

frontend/src/utils/markdown.ts

@@ -0,0 +1,205 @@
+import katex from 'katex';
+import hljs from 'highlight.js';
+
+export interface MediaInfo {
+  type: 'img' | 'video';
+  src: string;
+  alt?: string;
+}
+
+export interface RenderResult {
+  html: string;
+  media: MediaInfo[];
+}
+
+function escapeHtml(text: string): string {
+  return text
+    .replace(/&/g, '&')
+    .replace(/</g, '&lt;')
+    .replace(/>/g, '&gt;')
+    .replace(/"/g, '&quot;');
+}
+
+/**
+ * 将 Markdown 文本转为带占位符的 HTML 字符串。
+ *
+ * 设计契约（曾被线上 bug 击穿、用单测固化，见 markdown.test.ts）：
+ *  - 段落内【单个 \n 转为空格】而非 <br/>，避免 AI 回复被强制大量换行；
+ *  - 【3+ 连续空行压缩为 2 个】（/\n{3,}/ -> \n\n），避免多余空白把内容撑成多行；
+ *  - 【表格在加粗/链接转换之后提取】，单元格内的 **xxx** / [文本](url) 才能被渲染；
+ *    且单元格内连续 <br> 被压缩为单个（normalizeCell）。
+ *
+ * 媒体(<img>/<video>)以 \x00MEDIA<n>\x00 占位符保留在返回的 html 中，
+ * 由调用方（ChatPage 的 MarkdownContent）切分后渲染为 React 组件。
+ */
+export function renderMarkdownToHtml(content: string): RenderResult {
+  let text = content;
+
+  // 1. Extract code blocks first (protect from other replacements)
+  const codeBlocks: string[] = [];
+  const mermaidBlocks: string[] = [];
+  text = text.replace(/```(\w*)\n?([\s\S]*?)```/g, (_m: string, _lang: string, code: string) => {
+    if (_lang === 'mermaid') {
+      const idx = mermaidBlocks.length;
+      mermaidBlocks.push('<div class="md-mermaid" data-code="' + escapeHtml(code.trim()) + '"></div>');
+      return '\x00MERMAID' + idx + '\x00';
+    }
+    const idx = codeBlocks.length;
+    codeBlocks.push('<pre class="md-code-block"><code class="md-code">' + escapeHtml(code.trim()) + '</code></pre>');
+    return '\x00CODEBLOCK' + idx + '\x00';
+  });
+
+  // 2. Extract inline code
+  const inlineCodes: string[] = [];
+  text = text.replace(/`([^`]+)`/g, (_m: string, code: string) => {
+    const idx = inlineCodes.length;
+    inlineCodes.push('<code class="md-inline-code">' + escapeHtml(code) + '</code>');
+    return '\x00INLINE' + idx + '\x00';
+  });
+
+  // 2.5 Extract media blocks (<img>/<video>) -> 结构化，后续渲染为 React 组件
+  const mediaBlocks: MediaInfo[] = [];
+  text = text.replace(/<(img|video)([^>]*?)\/?>/gi, (_m: string, tag: string, attrs: string) => {
+    const type = tag.toLowerCase() === 'img' ? 'img' : 'video';
+    const srcM = attrs.match(/src=["']([^"']+)["']/i);
+    const altM = attrs.match(/alt=["']([^"']*)["']/i);
+    mediaBlocks.push({ type, src: srcM ? srcM[1] : '', alt: altM ? altM[1] : '' });
+    return '\x00MEDIA' + (mediaBlocks.length - 1) + '\x00';
+  });
+
+  // 2.6 Normalize excessive blank lines in prose (code blocks already extracted)
+  text = text.replace(/\n{3,}/g, '\n\n').trim();
+
+  // 2.8 LaTeX 数学公式：$$...$$（块级）和 $...$（行内）
+  const mathBlocks: string[] = [];
+  text = text.replace(/\$\$([\s\S]+?)\$\$/g, (_m: string, formula: string) => {
+    try {
+      const rendered = katex.renderToString(formula.trim(), { displayMode: true, throwOnError: false });
+      const idx = mathBlocks.length;
+      mathBlocks.push(`<div class="md-math-block">${rendered}</div>`);
+      return '\x00MATHBLOCK' + idx + '\x00';
+    } catch { return _m; }
+  });
+  text = text.replace(/\$([^\s$][^$]*[^\s$])\$/g, (_m: string, formula: string) => {
+    try {
+      const rendered = katex.renderToString(formula.trim(), { displayMode: false, throwOnError: false });
+      const idx = mathBlocks.length;
+      mathBlocks.push(`<span class="md-math-inline">${rendered}</span>`);
+      return '\x00MATHBLOCK' + idx + '\x00';
+    } catch { return _m; }
+  });
+
+  // 2.9 Strikethrough: ~~text~~ -> <del>text</del>
+  text = text.replace(/~~(.+?)~~/g, '<del>$1</del>');
+
+  // 2.10 Auto-link bare URLs: http(s)://... -> <a>
+  text = text.replace(/(?<![="'>])https?:\/\/[^\s<>[\]()]+/g, (url: string) =>
+    `<a class="md-link" href="${url}" target="_blank" rel="noopener">${url}</a>`
+  );
+
+  // 3. Headers
+  text = text.replace(/^#### (.+)$/gm, '<h4 class="md-h4">$1</h4>');
+  text = text.replace(/^### (.+)$/gm, '<h4 class="md-h4">$1</h4>');
+  text = text.replace(/^## (.+)$/gm, '<h3 class="md-h3">$1</h3>');
+  text = text.replace(/^# (.+)$/gm, '<h2 class="md-h2">$1</h2>');
+
+  // 4. Horizontal rule
+  text = text.replace(/^---+$/gm, '<hr class="md-hr">');
+
+  // 5. Blockquotes
+  text = text.replace(/^> (.+)$/gm, '<blockquote class="md-blockquote">$1</blockquote>');
+
+  // 6. Links [text](url)
+  text = text.replace(/\[([^\]]+)\]\(([^)]+)\)/g, '<a class="md-link" href="$2" target="_blank" rel="noopener">$1</a>');
+
+  // 7. Bold and italic
+  text = text.replace(/\*\*\*(.+?)\*\*\*/g, '<strong><em>$1</em></strong>');
+  text = text.replace(/\*\*(.+?)\*\*/g, '<strong>$1</strong>');
+  text = text.replace(/(?<!\w)\*(?!\*)(.+?)(?<!\*)\*(?!\w)/g, '<em>$1</em>');
+
+  // 7.5 Extract Markdown tables after inline formatting so cells support bold/italic/links.
+  // 表格提取放在加粗/链接之后，避免单元格内 **xxx** 原样显示；同时压缩单元格内多余 <br>。
+  const tableBlocks: string[] = [];
+  text = text.replace(
+    /((?:^\|.+\|\r?\n)+^\|[\s\-:|]+\|\r?\n(?:^\|.+\|\r?\n?)+)/gm,
+    (_m: string, block: string) => {
+      const lines = block.trim().split(/\r?\n/).filter((l: string) => l.trim());
+      if (lines.length < 2) return block;
+      const headerCells = lines[0].split('|').slice(1, -1).map((c: string) => c.trim());
+      const alignCells = lines[1].split('|').slice(1, -1).map((c: string) => {
+        const t = c.trim();
+        const left = t.startsWith(':');
+        const right = t.endsWith(':');
+        if (left && right) return 'center';
+        if (right) return 'right';
+        if (left) return 'left';
+        return 'left';
+      });
+      const normalizeCell = (cell: string) => {
+        return cell
+          .replace(/<br\s*\/?>\s*/gi, '<br/>')
+          .replace(/\n/g, '<br/>')
+          .replace(/(<br\/>){2,}/g, '<br/>')
+          .replace(/^(<br\/>)+|(<br\/>)+$/g, '');
+      };
+      const bodyRows = lines.slice(2).map((row: string) =>
+        row.split('|').slice(1, -1).map((c: string) => normalizeCell(c.trim()))
+      );
+      const thead = headerCells.map((c: string, i: number) =>
+        `<th style="text-align:${alignCells[i] || 'left'}">${normalizeCell(c)}</th>`
+      ).join('');
+      const tbody = bodyRows.map((r: string[]) =>
+        '<tr>' + r.map((c: string, i: number) =>
+          `<td style="text-align:${alignCells[i] || 'left'}">${c}</td>`
+        ).join('') + '</tr>'
+      ).join('');
+      const html = `<table class="md-table"><thead><tr>${thead}</tr></thead><tbody>${tbody}</tbody></table>`;
+      const idx = tableBlocks.length;
+      tableBlocks.push(html);
+      return '\x00TABLE' + idx + '\x00';
+    }
+  );
+
+  // 8. Lists
+  text = text.replace(/^\s*[-*] (.+)$/gm, '<li class="md-li">$1</li>');
+  text = text.replace(/^\s*\d+\. (.+)$/gm, '<li class="md-li">$1</li>');
+  // Wrap consecutive <li> in <ul>
+  text = text.replace(/((?:<li class="md-li">[\s\S]*?<\/li>\n?)+)/g, '<ul class="md-ul">$1</ul>');
+
+  // 9. Convert double newlines to paragraph breaks, single newlines to spaces
+  text = text.replace(/\n\n/g, '</p><p class="md-p">');
+  text = text.replace(/\n/g, ' ');
+
+  // 10. Restore code blocks and inline codes
+  text = text.replace(/\x00CODEBLOCK(\d+)\x00/g, (_m: string, idx: string) => codeBlocks[parseInt(idx)] || '');
+  text = text.replace(/\x00MERMAID(\d+)\x00/g, (_m: string, idx: string) => mermaidBlocks[parseInt(idx)] || '');
+  text = text.replace(/\x00INLINE(\d+)\x00/g, (_m: string, idx: string) => inlineCodes[parseInt(idx)] || '');
+  text = text.replace(/\x00TABLE(\d+)\x00/g, (_m: string, idx: string) => tableBlocks[parseInt(idx)] || '');
+  text = text.replace(/\x00MATHBLOCK(\d+)\x00/g, (_m: string, idx: string) => mathBlocks[parseInt(idx)] || '');
+
+  // 10.5 Apply syntax highlighting to code blocks + add copy button
+  text = text.replace(
+    /<pre class="md-code-block"><code class="md-code">([\s\S]*?)<\/code><\/pre>/g,
+    (_m: string, code: string) => {
+      const decoded = code.replace(/&lt;/g, '<').replace(/&gt;/g, '>').replace(/&amp;/g, '&');
+      let highlighted: string;
+      try { highlighted = hljs.highlightAuto(decoded).value; } catch { highlighted = code; }
+      return `<div class="md-code-wrapper">
+        <button class="md-code-copy" onclick="navigator.clipboard.writeText(this.parentElement.querySelector('code').textContent);this.textContent='已复制';setTimeout(()=>this.textContent='复制',2000)">复制</button>
+        <pre class="md-code-block"><code class="md-code hljs">${highlighted}</code></pre>
+      </div>`;
+    }
+  );
+
+  // 11. Wrap in paragraph and clean up
+  text = '<p class="md-p">' + text + '</p>';
+  text = text.replace(/<p class="md-p"><\/p>/g, '');
+  text = text.replace(/<p class="md-p">(<h[234][\s\S]*?<\/h[234]>)<\/p>/g, '$1');
+  text = text.replace(/<p class="md-p">(<pre[\s\S]*?<\/pre>)<\/p>/g, '$1');
+  text = text.replace(/<p class="md-p">(<ul[\s\S]*?<\/ul>)<\/p>/g, '$1');
+  text = text.replace(/<p class="md-p">(<hr[\s\S]*?)<\/p>/g, '$1');
+  text = text.replace(/<p class="md-p">(<blockquote[\s\S]*?<\/blockquote>)<\/p>/g, '$1');
+  text = text.replace(/<p class="md-p">(<table[\s\S]*?<\/table>)<\/p>/g, '$1');
+
+  return { html: text, media: mediaBlocks };
+}

migrations/007_evolution_tables.sql

@@ -70,7 +70,7 @@ CREATE TABLE IF NOT EXISTS `skill_trajectory_links` (
     `skill_name` VARCHAR(200) NOT NULL,
     `user_id` VARCHAR(36) NOT NULL,
     `trajectory_id` VARCHAR(36),
-    `atom_id` VARCHAR(36),
+    `atom_id` VARCHAR(128),
     `weightscore` INT DEFAULT 0,
     `created_at` DATETIME DEFAULT CURRENT_TIMESTAMP,
     INDEX `idx_stl_user_skill` (`user_id`, `skill_name`),

migrations/018_normalize_marketapp_scene.sql

@@ -2,10 +2,16 @@
 -- 旧的 text/thinking/vision/tool/image/audio/asr/ocr/embedding/rerank/video 已不再使用，
 -- 统一置空，后续由 AI 验证 / 导入阶段按业务场景重新归类。
 -- 幂等：执行一次后不再有匹配行，可安全每次重启重跑。
+-- 注意：当前业务场景标准 = 前台「应用市场」14 类 (MARKET_SCENES)：
+--   design, frontend, mobile, backend, qa, search, docs, marketing,
+--   workflow, secops, learning, video, collab, ai
+-- 这些为合法取值，必须保留，不得置空（#707 修复：此前 allowlist 不含 14 类导致重启被清空）。
 UPDATE market_apps
 SET capability = NULL
 WHERE capability IS NOT NULL
   AND capability NOT IN (
+    'design', 'frontend', 'mobile', 'backend', 'qa', 'search', 'docs', 'marketing',
+    'workflow', 'secops', 'learning', 'video', 'collab', 'ai',
     'freelance_content', 'freelance_client', 'freelance_finance',
     'office_doc', 'office_data', 'office_email', 'office_pm',
     'edu_prepare', 'edu_grade', 'edu_tutor', 'edu_home',

migrations/021_widen_atom_id.sql

@@ -0,0 +1,7 @@
+-- 021: 加宽 skill_trajectory_links.atom_id 列
+-- 背景：atom_id 实际值形如 "{trajectory_id}_atom_{i}"（UUID 36 + "_atom_" + 序号 ≈ 45 字符），
+-- 原 VARCHAR(36) 过短导致写入时 DataError(1406) Data too long，并使会话进入 rolled-back 状态，
+-- 级联触发 PendingRollbackError 与日志 trace_id KeyError 刷屏。
+-- 本迁移幂等：重复执行 MODIFY 到相同类型为无害 no-op。
+
+ALTER TABLE `skill_trajectory_links` MODIFY COLUMN `atom_id` VARCHAR(128) NULL;

migrations/mig_g4_distribution.sql

@@ -56,7 +56,7 @@ CREATE TABLE IF NOT EXISTS `distribution_licenses` (
 ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_unicode_ci;
 
 -- Insert seed data: sample distributions
-INSERT INTO `distributions` (`id`, `name`, `description`, `brand_name`, `brand_color`, `provider`, `model`, `features`, `status`) VALUES
+INSERT IGNORE INTO `distributions` (`id`, `name`, `description`, `brand_name`, `brand_color`, `provider`, `model`, `features`, `status`) VALUES
 ('distro-starter', 'Starter Edition', '基础版发行包，适合个人开发者和小团队', 'AI Starter', '#6366f1', 'siliconflow', 'deepseek-ai/DeepSeek-V4-Flash', '["chat", "code_exec"]', 'active'),
 ('distro-enterprise', 'Enterprise Edition', '企业版发行包，包含完整 MCP 生态 + Agent 协作', 'AI Enterprise', '#f59e0b', 'siliconflow', 'deepseek-ai/DeepSeek-V4-Flash', '["chat", "code_exec", "knowledge_search", "agent_collab"]', 'active'),
 ('distro-devops', 'DevOps Edition', 'DevOps 专用发行包，CI/CD + 代码审查 + 部署', 'AI DevOps', '#10b981', 'siliconflow', 'deepseek-ai/DeepSeek-V4-Flash', '["chat", "code_exec", "automation", "webhook"]', 'active');

migrations/mig_g5_privacy.sql

@@ -84,6 +84,11 @@ CREATE TABLE IF NOT EXISTS usage_reports (
 ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_unicode_ci;
 
 -- 5. Add privacy_mode to distributions table
-ALTER TABLE distributions 
+-- MariaDB 11.8 不支持 ADD COLUMN IF NOT EXISTS 语法，改用 information_schema 预检 + 动态 PREPARE 实现幂等
-    ADD COLUMN IF NOT EXISTS privacy_mode ENUM('local', 'cloud', 'hybrid') NOT NULL DEFAULT 'cloud' 
+SET @exist_pm := (SELECT COUNT(*) FROM information_schema.COLUMNS WHERE TABLE_SCHEMA = 'aiflow' AND TABLE_NAME = 'distributions' AND COLUMN_NAME = 'privacy_mode');
-    COMMENT 'Deployment mode for distribution' AFTER status;
+SET @sql_pm := IF(@exist_pm = 0,
+    "ALTER TABLE distributions ADD COLUMN privacy_mode ENUM('local', 'cloud', 'hybrid') NOT NULL DEFAULT 'cloud' COMMENT 'Deployment mode for distribution' AFTER status",
+    'SELECT 1');
+PREPARE stmt_pm FROM @sql_pm;
+EXECUTE stmt_pm;
+DEALLOCATE PREPARE stmt_pm;