fix(im_llm): 代理绕过Docker内部hostname

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

@@ -855,17 +855,34 @@ def _test_mcp_server(server, db) -> dict:
                     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
-                    tools = _aio.run(_lmt(sd, timeout=15))
+                    # raise_on_error=True 让网络异常透传，暴露真实错误而非 "no tools found"
+                    tools = _aio.run(_lmt(sd, timeout=15, raise_on_error=True))
                     result["success"] = bool(tools)
                     result["details"]["tools_count"] = len(tools)
                     result["details"]["tools"] = [t.get("name") for t in tools[:5]]
-                    result["message"] = f"Connected — {len(tools)} tools available" if tools else "Connected but no tools found"
+                    if tools:
+                        result["message"] = f"Connected — {len(tools)} tools available"
+                    else:
+                        result["message"] = "Connected but no tools found (server returned empty list)"
                     result["details"]["status_code"] = 200
                     if tools:
                         server.tools_count = len(tools)
                         server.status = "connected"
                 except Exception as e:
-                    result["message"] = f"Connection failed: {str(e)}"
+                    # 把真实异常（类型+消息）透传给前端，告别误导性的 "Connected but no tools found"
+                    err_type = type(e).__name__
+                    err_msg = str(e)[:300]
+                    result["message"] = f"{err_type}: {err_msg}"
+                    result["details"]["error_type"] = err_type
+                    result["details"]["error"] = err_msg
+                    if "ConnectError" in err_type or "Timeout" in err_type:
+                        result["details"]["hint"] = "网络层无法到达 MCP 服务器，请检查代理配置或服务端 URL"
+                    elif "404" in err_msg or "NotFound" in err_type:
+                        result["details"]["hint"] = "服务端 URL 不存在或路径错误"
+                    elif "401" in err_msg or "Unauthorized" in err_type:
+                        result["details"]["hint"] = "认证失败：检查请求头中的 Authorization"
+                    else:
+                        result["details"]["hint"] = "检查认证头/transport_type 或服务端状态"
 
         elif transport == "sse":
 

backend/python/app/api/auth_verify.py

@@ -0,0 +1,68 @@
+"""
+Platform Auth Verify Endpoint
+用于 nginx auth_request 子请求验证平台 JWT
+返回 200 + X-User-Id/X-User-Name headers 或 401
+"""
+import logging
+from fastapi import APIRouter, Request
+from fastapi.responses import Response
+
+from app.auth import decode_access_token
+
+logger = logging.getLogger("aiflow.auth_verify")
+
+router = APIRouter()
+
+
+@router.get("/verify")
+@router.head("/verify")
+async def verify_auth(request: Request):
+    """
+    nginx auth_request 验证端点
+    从请求中提取 JWT (Authorization header 或 cookie)并验证
+    ��回 200 + 用户信息 headers，或 401
+    """
+    token = None
+
+    # 1. 尝试 Authorization: Bearer <token>
+    auth_header = request.headers.get("Authorization", "")
+    if auth_header.startswith("Bearer "):
+        token = auth_header[7:]
+
+    # 2. 尝试 cookie 中的 jwt token
+    if not token:
+        token = request.cookies.get("jwt") or request.cookies.get("token") or request.cookies.get("access_token")
+
+    # 3. 尝试 X-Platform-Token header
+    if not token:
+        token = request.headers.get("X-Platform-Token")
+
+    if not token:
+        logger.debug("No token found in request")
+        return Response(status_code=401, headers={"X-Auth-Error": "no_token"})
+
+    payload = decode_access_token(token)
+    if not payload:
+        logger.debug("Invalid or expired token")
+        return Response(status_code=401, headers={"X-Auth-Error": "invalid_token"})
+
+    user_id = payload.get("sub", "")
+    user_name = payload.get("username", "")
+    user_role = payload.get("role", "user")
+    tenant_id = payload.get("tenant_id", "")
+
+    if not user_id:
+        return Response(status_code=401, headers={"X-Auth-Error": "no_user_id"})
+
+    logger.info(f"Auth verify OK: user={user_id}, name={user_name}")
+
+    return Response(
+        status_code=200,
+        headers={
+            "X-User-Id": str(user_id),
+            "X-User-Name": str(user_name) if user_name else str(user_id),
+            "X-User-Role": str(user_role),
+            "X-Tenant-Id": str(tenant_id) if tenant_id else "",
+            "Cache-Control": "no-cache, no-store, must-revalidate",
+        },
+    )

backend/python/app/api/billing_webhook.py

@@ -0,0 +1,118 @@
+"""
+Platform Billing Webhook
+接收 LiteLLM spend 回调，写入 token_usage_logs 并扣减 user_balances
+"""
+import logging
+from datetime import datetime
+from fastapi import APIRouter, Request, HTTPException
+from pydantic import BaseModel
+from typing import Optional
+from sqlalchemy import text
+from app.database import SessionLocal
+
+logger = logging.getLogger("aiflow.billing")
+
+router = APIRouter()
+
+
+class SpendEvent(BaseModel):
+    """LiteLLM spend log event"""
+    request_id: Optional[str] = None
+    api_key: Optional[str] = None
+    model: Optional[str] = None
+    total_tokens: int = 0
+    prompt_tokens: int = 0
+    completion_tokens: int = 0
+    spend: float = 0.0
+    user: Optional[str] = None  # platform user_id passed via metadata
+    metadata: Optional[dict] = None
+    startTime: Optional[str] = None
+    endTime: Optional[str] = None
+
+
+@router.post("/webhook")
+async def billing_webhook(event: SpendEvent, request: Request):
+    """
+    接收 LiteLLM spend webhook 回调
+    记录 token 使用并扣减用户余额
+    """
+    # 验证来源
+    api_key = event.api_key or ""
+    metadata = event.metadata or {}
+
+    # 只处理 mumuainovel 的用量
+    app_name = metadata.get("app", "")
+    if app_name != "mumuainovel" and "mumu" not in api_key:
+        # 也检查是否有自定义的 X-Source header
+        source = request.headers.get("X-Source", "")
+        if source != "mumuainovel":
+            return {"status": "skipped", "reason": "not_mumuainovel"}
+
+    user_id = event.user or metadata.get("user_id", "mumu-user")
+    model_name = event.model or "unknown"
+    spend = max(event.spend, 0)
+    total_tokens = event.total_tokens
+    prompt_tokens = event.prompt_tokens
+    completion_tokens = event.completion_tokens
+
+    if total_tokens == 0 and spend == 0:
+        return {"status": "skipped", "reason": "zero_usage"}
+
+    db = SessionLocal()
+    try:
+        # 写入 token_usage_logs
+        db.execute(text("""
+            INSERT INTO token_usage_logs (id, user_id, model_name, app_id, app_name,
+                prompt_tokens, completion_tokens, total_tokens, cost, created_at)
+            VALUES (gen_random_uuid(), :uid, :model, :app_id, :app_name,
+                :pt, :ct, :tt, :cost, :ts)
+        """), {
+            "uid": user_id,
+            "model": model_name,
+            "app_id": "mumuainovel-001",
+            "app_name": "MuMuAINovel",
+            "pt": prompt_tokens,
+            "ct": completion_tokens,
+            "tt": total_tokens,
+            "cost": spend,
+            "ts": datetime.utcnow(),
+        })
+
+        # 扣减余额
+        result = db.execute(text("""
+            UPDATE user_balances
+            SET balance = balance - :cost,
+                total_used = total_used + :cost,
+                updated_at = NOW()
+            WHERE user_id = :uid AND balance >= :cost
+            RETURNING id
+        """), {"cost": spend, "uid": user_id})
+
+        if result.rowcount == 0:
+            # 余额不足或无记录，创建新余额记录
+            # 检查是否存在
+            exists = db.execute(text(
+                "SELECT id FROM user_balances WHERE user_id = :uid"
+            ), {"uid": user_id}).fetchone()
+            if not exists:
+                db.execute(text("""
+                    INSERT INTO user_balances (id, user_id, balance, total_used, created_at, updated_at)
+                    VALUES (gen_random_uuid(), :uid, :balance, :cost, NOW(), NOW())
+                """), {"uid": user_id, "balance": 10.0 - spend, "cost": spend})
+
+        db.commit()
+        logger.info(f"Billing: user={user_id} model={model_name} tokens={total_tokens} spend=${spend:.6f}")
+
+    except Exception as e:
+        db.rollback()
+        logger.error(f"Billing webhook error: {e}")
+        raise HTTPException(status_code=500, detail=str(e))
+    finally:
+        db.close()
+
+    return {"status": "ok", "user_id": user_id, "spend": spend}
+
+
+@router.get("/webhook/health")
+async def billing_webhook_health():
+    return {"status": "ok", "service": "billing-webhook"}

backend/python/app/api/bug_audit.py

@@ -0,0 +1,86 @@
+"""Bug-audit endpoint for runtime regression checks."""
+from fastapi import APIRouter, Depends
+from sqlalchemy.orm import Session
+from sqlalchemy import text
+from app.database import get_db
+from app.models import SiteSetting, AIModel
+from app.capability_registry import CAPABILITY_TYPES, get_capability_defaults, resolve_auto_model
+import logging
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter()
+
+@router.get("/bug-audit")
+def bug_audit(db: Session = Depends(get_db)):
+    """
+    Runtime self-check: 验证平台大模型路由/能力配置的关键不变量。
+    用于发现重复 bug（比如「已修复又出现」）。
+    """
+    issues = []
+    warnings = []
+    checks = []
+
+    # Check 1: 文本类能力(text/thinking) 必须指向启用的非图像/视频模型
+    TEXT_FAMILY = {"text", "thinking", "vision", "tool"}
+    defaults = get_capability_defaults(db)
+    for cap in ("text", "thinking"):
+        mid = defaults.get(cap)
+        if not mid:
+            issues.append(f"{cap}: site_settings.model_capability_defaults 未设置")
+            continue
+        m = db.query(AIModel).filter(AIModel.id == mid, AIModel.is_enabled == True).first()
+        if not m:
+            issues.append(f"{cap}: 指向的模型 {mid} 不存在或已禁用")
+            continue
+        caps = [c.lower() for c in (m.capabilities or [])]
+        if not ("text" in caps or "thinking" in caps):
+            issues.append(f"{cap}: 指向的 {m.name} 没有 text/thinking 能力 (caps={caps})")
+        if "image" in caps or "video" in caps:
+            warnings.append(f"{cap}: 默认模型 {m.name} 含 image/video 能力 (普通多轮对话可能误判)")
+        checks.append(f"{cap} -> {m.name} (caps={caps})")
+
+    # Check 2: 必须有 is_default=true 的模型
+    default_model = db.query(AIModel).filter(AIModel.is_default == True, AIModel.is_enabled == True).first()
+    if not default_model:
+        issues.append("没有任何 is_default=true 且启用的模型")
+    else:
+        checks.append(f"is_default -> {default_model.name}")
+
+    # Check 3: 所有 enabled 的 chat 模型都应该能正常 resolve
+    for cap in ("text", "image", "video", "audio", "embedding", "rerank", "asr", "ocr"):
+        m = resolve_auto_model(cap, db)
+        if m:
+            checks.append(f"resolve_auto_model({cap}) -> {m.name}")
+        else:
+            warnings.append(f"resolve_auto_model({cap}) 返回 None (该能力无默认模型)")
+
+    # Check 4: detect_media_intent 单元自检（用经典样本）
+    from app.api.chat_media import detect_media_intent
+    test_cases = [
+        ("平台有哪些技能能", None),  # 普通文字
+        ("你好", None),  # 普通问候
+        ("画一只小猫", "image"),  # 明确图像
+        ("生成一个视频介绍", "video"),  # 明确视频
+        ("今天天气如何？", None),  # 疑问句
+    ]
+    for text_in, expected in test_cases:
+        result = detect_media_intent(text_in)
+        if result != expected:
+            issues.append(f"detect_media_intent({text_in!r})={result!r} 期望 {expected!r}")
+        else:
+            checks.append(f"detect_media_intent({text_in!r})={result!r} ✓")
+
+    # 汇总
+    healthy = len(issues) == 0
+    return {
+        "healthy": healthy,
+        "issues": issues,
+        "warnings": warnings,
+        "checks": checks,
+        "summary": {
+            "total_checks": len(checks),
+            "issues": len(issues),
+            "warnings": len(warnings),
+        }
+    }

backend/python/app/api/chat.py

@@ -2655,7 +2655,7 @@ def get_models(db: Session = Depends(get_db)):
 
 
 
-    models = db.query(AIModel).filter(AIModel.is_enabled == True).all()
+    models = db.query(AIModel).filter(AIModel.is_enabled == True).order_by(AIModel.is_default.desc(), AIModel.name.asc()).all()
 
 
 

backend/python/app/api/chat_media.py

@@ -506,7 +506,11 @@ async def _optimize_media_prompt(raw_prompt: str, media_type: str, db=None) -> s
 
 
 
-            api_key = text_model.api_key or (provider.api_key if provider else None)
+            # 调用 LiteLLM 网关时用 master key（模型自带 key 不在 LiteLLM 数据库里）
+            if text_model.endpoint and "qihu-aiflow-litellm" in text_model.endpoint:
+                api_key = "sk-litellm-master-key-aiflow-2026"
+            else:
+                api_key = text_model.api_key or (provider.api_key if provider else None)
 
 
 
@@ -936,7 +940,11 @@ async def _generate_media(db, ai_model, prompt: str, user_id: str = None) -> str
 
 
 
-    api_key = ai_model.api_key or (provider.api_key if provider else None)
+    # 调用 LiteLLM 网关时用 master key（模型自带 key 不在 LiteLLM 数据库里）
+    if ai_model.endpoint and "qihu-aiflow-litellm" in ai_model.endpoint:
+        api_key = "sk-litellm-master-key-aiflow-2026"
+    else:
+        api_key = ai_model.api_key or (provider.api_key if provider else None)
 
 
 

backend/python/app/api/chat_service.py

@@ -2895,6 +2895,13 @@ async def _send_message_core(
 
 
 
+
+            # Defensive guard: 即使 model_id 指向 image/video 模型，若用户输入无媒体意图词，
+            # 强制回退到 text 默认模型。修复 bug：普通多轮对话被误判走图片生成。
+            if ai_model and ("image" in caps or "video" in caps) and not detect_media_intent(req.content):
+                logger.info("[chat] Override media-cap model {} -> text (no media intent in: {:.50r})".format(ai_model.name, req.content))
+                ai_model = resolve_auto_model("text", db)
+                caps = [c.lower() for c in (ai_model.capabilities or [])] if ai_model else []
             if ai_model and ("image" in caps or "video" in caps):
 
 

backend/python/app/api/im_llm.py

@@ -16,6 +16,7 @@ IM Bot 真实 LLM 调用。
   * 平台知识库注入（get_platform_knowledge，技能清单实时同步）
 """
 import asyncio
+import os
 import json
 import logging
 import re as _re
@@ -48,7 +49,8 @@ def _proxy_for_url(url: str) -> str | None:
     host = (_urlparse(url).hostname or "").lower()
     if "agnes-ai.com" in host or "agnes-ai.space" in host:
         return _HK_PROXY
-    if "qihuyun.com" in host:
+    # 2026-07-30 修复 3+: 不代理 Docker 内部 hostname 和 localhost
+    if "qihuyun.com" in host or "localhost" in host or host.startswith("127.") or "qihu-aiflow" in host:
         return None
     _no_proxy = [h.strip().lower() for h in os.environ.get("NO_PROXY", "").split(",") if h.strip()]
     for _np in _no_proxy:
@@ -127,7 +129,12 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
         )
 
         # ===== 1) 媒体意图路由（Issue 7）=====
-        intent = detect_media_intent(prompt)
+        # 2026-07-30 fix: 只对当前用户消息做媒体意图检测，避免历史中的图像请求污染
+        _cur_msg = prompt
+        _marker = "用户当前问题："
+        if _marker in prompt:
+            _cur_msg = prompt[prompt.rfind(_marker) + len(_marker):].strip()
+        intent = detect_media_intent(_cur_msg)
         if intent:
             ai_model = resolve_auto_model(intent, db)
             if ai_model:
@@ -196,6 +203,10 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
             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
+            # 2026-07-30: LiteLLM 端点统一用 master key（与 chat_service.py 对齐）
+            if api_url and "qihu-aiflow-litellm" in api_url:
+                _litellm_key = os.environ.get("LITELLM_MASTER_KEY") or "sk-litellm-master-key-aiflow-2026"
+                api_key = _litellm_key
             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):
@@ -221,13 +232,15 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                 ai_content = ""
                 _http_ok = True
                 for round_num in range(max_rounds):
-                    request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048}
+                    request_body = {"model": ai_model.name, "messages": messages, "max_tokens": 2048, "enable_thinking": False, "temperature": 0.7}
                     if function_definitions and skills:
                         request_body["tools"] = [
                             {"type": "function", "function": fd} for fd in function_definitions
                         ]
                         request_body["tool_choice"] = "auto"
                     try:
+                        import logging as _l5
+                        _l5.warning(f'[im_llm REQ] {ai_model.name}: url={api_url[:60]}, proxy={_proxy_for_url(api_url)}, api_key={api_key[:20]}..., messages_count={len(request_body.get("messages", []))}, tools_count={len(request_body.get("tools", []))}')
                         data = chat_complete(
                             api_url,
                             api_key,
@@ -235,40 +248,58 @@ def _call_real_llm(db: Session, user=None, prompt: str = "", platform: str = "im
                             proxy=_proxy_for_url(api_url),
                             timeout=ai_model.timeout or 120,
                         )
+                        _l5.warning(f'[im_llm RES] {ai_model.name}: data_keys={list(data.keys()) if data else None}, resp_type={type(data).__name__}')
                     except Exception as _exc:
                         _retry_log.append(f"{ai_model.name}: {str(_exc)[:60]}")
+                        import traceback as _tb
                         logger.warning(f"[im_llm] 模型 {ai_model.name} 异常: {_exc}")
+                        logger.warning(f"[im_llm] {ai_model.name} 异常详情: {_tb.format_exc()[:2000]}")
                         _http_ok = False
                         break
-                        message = data.get("choices", [{}])[0].get("message", {})
-                        tool_calls = message.get("tool_calls", [])
-                        if tool_calls and skills:
+                    message = data.get("choices", [{}])[0].get("message", {})
+                    tool_calls = message.get("tool_calls", [])
+                    # 2026-07-30 CRITICAL DEBUG
+                    import logging as _lcr
+                    _raw_content = message.get("content")
+                    _raw_reasoning = message.get("reasoning_content")
+                    _msg_keys = list(message.keys()) if message else []
+                    _tool_calls = tool_calls
+                    _lcr.warning(f'[CRITICAL] MODEL={ai_model.name} msg_keys={_msg_keys} content_len={len(_raw_content or "")} reasoning_len={len(_raw_reasoning or "")} has_tool_calls={bool(_tool_calls)}')
+                    ai_content = message.get("content", "") or message.get("reasoning_content") or ""
+                    if not ai_content and (message.get("reasoning_content") or ""):
+                        ai_content = f"（AI 思考：{message.get('reasoning_content')[:500]}...）"
+                    if tool_calls and skills:
+                        messages.append({
+                            "role": "assistant",
+                            "content": message.get("content") or message.get("reasoning_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": "assistant",
-                                "content": message.get("content") or "",
-                                "tool_calls": tool_calls,
+                                "role": "tool",
+                                "tool_call_id": tc.get("id", "call_1"),
+                                "content": tool_result,
                             })
-                            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
+                        if round_num >= max_rounds - 1:
+                            break
+                        continue
+                    break
 
                 # 检查本轮调用是否成功
+                import logging as _log2
+                _log2.warning(f'[im_llm trace] {ai_model.name}: http_ok={_http_ok}, ai_content_len={len(ai_content) if ai_content else 0}, ai_content_preview={repr(ai_content)[:200]}')
                 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...")
+                logger.warning('[im_llm] 模型 %s 失败 (http_ok=%s, content=%s), 切换备选...' % (
+                    ai_model.name, _http_ok, repr(ai_content)[:100]
+                ))
                 continue
             except Exception as exc:
                 _retry_log.append(f"{ai_model.name}: {str(exc)[:60]}")

backend/python/app/auth.py

@@ -122,6 +122,6 @@ async def get_current_admin(
     user=Depends(get_current_user),
 ):
     """验证当前用户为管理员"""
-    if user.role != "admin":
+    if not has_role(user.role, "admin"):
         raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail="需要管理员权限")
     return user

backend/python/app/mcp_client.py

@@ -43,12 +43,14 @@ def _transport_cm(server: dict, timeout: float):
 
 
 async def _session_request(
-    server: dict, method: str, params: dict, timeout: float = 30.0
+    server: dict, method: str, params: dict, timeout: float = 30.0,
+    raise_on_error: bool = False
 ):
     """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。
+    raise_on_error=True 时连接/网络错误（如 ExceptionGroup）会向上抛，不再吞掉。
     """
     url = (server.get("url") or "").strip()
     if not url:
@@ -89,33 +91,40 @@ async def _session_request(
                     }
     except Exception as e:
         logger.warning(f"[MCP] {method} 异常 {url}: {e}")
+        if raise_on_error:
+            raise
         return None
     return None
 
 
-async def list_mcp_tools(server: dict, timeout: float = 30.0) -> list:
+async def list_mcp_tools(server: dict, timeout: float = 30.0, raise_on_error: bool = False) -> list:
     """
     拉取 MCP 服务器的工具列表。
     server: {url, transport_type, headers}
     返回 MCP tools 数组（每个含 name/description/inputSchema）。
+    raise_on_error=True 时连接/网络错误不再吞掉，直接向上抛（供测试诊断用）。
     """
     url = (server.get("url") or "").strip()
     if not url:
         return []
-    # 缓存
-    now = asyncio.get_event_loop().time()
-    cached = _TOOLS_CACHE.get(url)
-    if cached and (now - cached[0]) < _TOOLS_CACHE_TTL:
-        return cached[1]
+    # 缓存（仅在正常返回 tools 时缓存；raise_on_error 模式不缓存，便于实时重试）
+    if not raise_on_error:
+        now = asyncio.get_event_loop().time()
+        cached = _TOOLS_CACHE.get(url)
+        if cached and (now - cached[0]) < _TOOLS_CACHE_TTL:
+            return cached[1]
     try:
-        data = await _session_request(server, "tools/list", {}, timeout)
+        data = await _session_request(server, "tools/list", {}, timeout, raise_on_error=raise_on_error)
         tools = []
         if data and "result" in data:
             tools = data["result"].get("tools", []) or []
-        _TOOLS_CACHE[url] = (now, tools)
+        if not raise_on_error:
+            _TOOLS_CACHE[url] = (asyncio.get_event_loop().time(), tools)
         return tools
     except Exception as e:
         logger.warning(f"[MCP] 拉取工具失败 {url}: {e}")
+        if raise_on_error:
+            raise
         return []