Undefined-bot 3.2.7__tar.gz → 3.2.8__tar.gz

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/CHANGELOG.md

@@ -1,3 +1,22 @@
+## v3.2.8 多模态附件与参考图生图
+
+围绕多模态消息链路进行了较大增强，引入统一的附件 UID 注册与 `<pic uid="..."/>` 图文混排机制，打通群聊、私聊、WebUI 和合并转发中的图片/文件上下文。同步为生图工具补齐参考图生图、独立图片编辑模型配置与提示词审核能力，并完善 Runtime API、认知侧写和动态技能加载的稳定性。
+
+- 新增统一附件注册系统，群聊、私聊、WebUI 会话和合并转发中的图片/文件会登记为内部附件 UID，并写入历史记录与提示词上下文。
+- 系统提示词与发送链路支持 `<pic uid="..."/>` 图文混排，AI 可直接在回复中内嵌当前会话内的图片 UID。
+- 新增 `fetch_image_uid` 工具，支持将远程图片 URL 拉取并注册为当前会话可复用的图片 UID。
+- `get_picture`、`render_html`、`render_markdown`、`render_latex` 默认改为 `embed` 返回可嵌入图片 UID，同时保留 `send` 直接发送模式。
+- `file_analysis_agent` 及相关文件/图片链路已适配附件 UID，优先使用内部 `pic_*` / `file_*` 标识，也继续兼容 URL 与 legacy `file_id`。
+- `ai_draw_one` 支持基于 `reference_image_uids` 的参考图生图，新增 `[models.image_edit]` 配置节，并接入 OpenAI 兼容的 `/v1/images/edits` 接口。
+- 新增基于 Agent 模型的生图提示词审核能力，并补充独立审核提示模板。
+- 优化图片生成请求体验，支持 base64 返回、保留显式尺寸、锁定模型参数，并更清晰地暴露上游错误。
+- 修复附件渲染、图片编辑上传与阻塞式文件读取问题，增强多模态链路稳定性。
+- Runtime API 与 WebUI 现在会注册聊天附件、渲染图片 UID，并支持本地 `file://` 图片预览；同时取消 `_agent` 工具调用的固定超时上限。
+- 认知服务会在用户昵称或群名变化时自动刷新 profile 展示名与向量索引，减少侧写名称陈旧问题。
+- 修复动态技能模块加载与超时包装问题，并补充相关测试覆盖；同步整理 README，移除顶部头图。
+
+---
+
 ## v3.2.7 arXiv 工具集与运行时变更感知
 
 新增 arXiv 论文搜索与提取工具集，以及运行时 CHANGELOG 查询能力。重构了生图工具支持 OpenAI 兼容接口，引入 grok_search 联网搜索工具，并让 AI 在系统提示词中感知自身模型配置信息。同步修复了多项稳定性问题与 CI 效率优化。

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: Undefined-bot
-Version: 3.2.7
+Version: 3.2.8
 Summary: QQ bot platform with cognitive memory architecture and multi-agent Skills, via OneBot V11.
 Project-URL: Repository, https://github.com/69gg/Undefined
 Project-URL: Issues, https://github.com/69gg/Undefined/issues
@@ -60,7 +60,7 @@ Description-Content-Type: text/markdown
 
 <table border="0">
   <tr>
-    <td width="70%" valign="top">
+    <td width="100%" valign="top">
       <div align="center">
         <h1>Undefined</h1>
         <em>QQ bot platform with cognitive memory architecture and multi-agent Skills, via OneBot V11.</em>
@@ -78,9 +78,6 @@ Description-Content-Type: text/markdown
         <strong>Undefined</strong> 是一个基于 Python 异步架构的高性能 QQ 机器人平台，搭载<strong>认知记忆架构</strong>，采用自研 <strong>Skills</strong> 系统，内置多个智能 Agent，支持代码分析、网络搜索、娱乐互动等多模态能力，并提供 <strong>Management-first WebUI</strong> 在线管理，以及可连接同一管理服务的 <strong>Desktop / Android App</strong>。
       </p>
     </td>
-    <td width="30%">
-      <img src="https://raw.githubusercontent.com/69gg/Undefined/main/img/head.png" width="100%" alt="Undefined" />
-    </td>
   </tr>
 </table>
 

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/README.md

@@ -1,6 +1,6 @@
 <table border="0">
   <tr>
-    <td width="70%" valign="top">
+    <td width="100%" valign="top">
       <div align="center">
         <h1>Undefined</h1>
         <em>QQ bot platform with cognitive memory architecture and multi-agent Skills, via OneBot V11.</em>
@@ -18,9 +18,6 @@
         <strong>Undefined</strong> 是一个基于 Python 异步架构的高性能 QQ 机器人平台，搭载<strong>认知记忆架构</strong>，采用自研 <strong>Skills</strong> 系统，内置多个智能 Agent，支持代码分析、网络搜索、娱乐互动等多模态能力，并提供 <strong>Management-first WebUI</strong> 在线管理，以及可连接同一管理服务的 <strong>Desktop / Android App</strong>。
       </p>
     </td>
-    <td width="30%">
-      <img src="https://raw.githubusercontent.com/69gg/Undefined/main/img/head.png" width="100%" alt="Undefined" />
-    </td>
   </tr>
 </table>
 

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/config.toml.example

@@ -527,6 +527,23 @@ model_name = ""
 # en: Extra request-body params (optional).
 [models.image_gen.request_params]
 
+# zh: 参考图生图模型配置（用于 ai_draw_one 传入 reference_image_uids 时调用 OpenAI 兼容的图片编辑接口）。
+# en: Reference-image generation model config (used when ai_draw_one receives reference_image_uids and calls the OpenAI-compatible image editing API).
+[models.image_edit]
+# zh: OpenAI-compatible 基址 URL，例如 https://api.openai.com/v1（最终请求路径为 /v1/images/edits）。
+# en: OpenAI-compatible base URL, e.g. https://api.openai.com/v1 (final request path is /v1/images/edits).
+api_url = ""
+# zh: API Key。
+# en: API key.
+api_key = ""
+# zh: 模型名称，空则回退到 [models.image_gen] 的 model_name。
+# en: Model name, empty falls back to [models.image_gen].model_name.
+model_name = ""
+
+# zh: 额外请求体参数（可选）。
+# en: Extra request-body params (optional).
+[models.image_edit.request_params]
+
 # zh: 本地知识库配置。
 # en: Local knowledge base settings.
 [knowledge]

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "Undefined-bot"
-version = "3.2.7"
+version = "3.2.8"
 description = "QQ bot platform with cognitive memory architecture and multi-agent Skills, via OneBot V11."
 readme = "README.md"
 authors = [

undefined_bot-3.2.8/res/prompts/image_gen_moderation.txt

@@ -0,0 +1,16 @@
+你是图片生成审核助手，只根据“待生成图片的提示词”判断是否允许生成。
+
+拦截范围包括但不限于：
+- 露骨色情、性剥削、未成年人性化
+- 血腥暴力、肢解、虐杀等重口内容
+- 明显违法、极端、恐怖主义相关内容
+- 明显侵犯隐私、仇恨或其他高风险内容
+
+审核原则：
+- 只审核风险，不提供改写建议，不解释政策，不输出多余内容
+- 安全、普通、模糊但不明显违规的内容默认放行
+- 如果只是普通人物、风景、二次元、服饰、轻微动作描写，不要误杀
+
+输出格式必须严格遵守：
+- 允许：`ALLOW`
+- 拒绝：`BLOCK: <不超过20字的简短中文原因>`

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/res/prompts/undefined.xml

@@ -131,6 +131,15 @@
 
         **关键点：每次消息处理都必须以 end 结束，这是维持对话流的核心机制。**
       </workflow>
+
+      <mixed_media_output priority="P0">
+        **图文混排规则：**
+        - 如果上下文或工具结果给了图片 UID（例如 `pic_ab12cd34`），你可以在 `send_message.message` 里直接插入 `<pic uid="pic_ab12cd34"/>`
+        - `<pic uid="..."/>` 是唯一允许的内嵌图片语法；不要改成 Markdown 图片、HTML `<img>`、代码块或自然语言描述
+        - 可以图文混排，例如：`我给你介绍一下`\n`<pic uid="pic_xxx"/>`\n`如图所示`
+        - 只能引用当前会话里明确给出的图片 UID，禁止臆造 UID
+        - 只有 `pic_*` 这类图片 UID 能放进 `<pic>`；普通文件 UID 不能放进去
+      </mixed_media_output>
     </message_sending_mechanism>
     
     <!-- 规则4: 对话结束机制 -->

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/res/prompts/undefined_nagaagent.xml

@@ -131,6 +131,15 @@
 
         **关键点：每次消息处理都必须以 end 结束，这是维持对话流的核心机制。**
       </workflow>
+
+      <mixed_media_output priority="P0">
+        **图文混排规则：**
+        - 如果上下文或工具结果给了图片 UID（例如 `pic_ab12cd34`），你可以在 `send_message.message` 里直接插入 `<pic uid="pic_ab12cd34"/>`
+        - `<pic uid="..."/>` 是唯一允许的内嵌图片语法；不要改成 Markdown 图片、HTML `<img>`、代码块或自然语言描述
+        - 可以图文混排，例如：`我给你介绍一下`\n`<pic uid="pic_xxx"/>`\n`如图所示`
+        - 只能引用当前会话里明确给出的图片 UID，禁止臆造 UID
+        - 只有 `pic_*` 这类图片 UID 能放进 `<pic>`；普通文件 UID 不能放进去
+      </mixed_media_output>
     </message_sending_mechanism>
     
     <!-- 规则4: 对话结束机制 -->

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/src/Undefined/__init__.py

@@ -1,3 +1,3 @@
 """Undefined - A high-performance, highly scalable QQ group and private chat robot based on a self-developed architecture."""
 
-__version__ = "3.2.7"
+__version__ = "3.2.8"

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/src/Undefined/ai/client.py

@@ -12,6 +12,7 @@ from uuid import uuid4
 
 import httpx
 
+from Undefined.attachments import AttachmentRegistry
 from Undefined.ai.llm import ModelRequester
 from Undefined.ai.model_selector import ModelSelector
 from Undefined.ai.multimodal import MultimodalAnalyzer
@@ -136,6 +137,7 @@ class AIClient:
         self._token_counter = TokenCounter()
         self._knowledge_manager: Any = None
         self._cognitive_service: Any = cognitive_service
+        self.attachment_registry = AttachmentRegistry(http_client=self._http_client)
 
         # 私聊发送回调
         self._send_private_message_callback: Optional[SendPrivateMessageCallback] = None
@@ -325,6 +327,13 @@ class AIClient:
         if hasattr(self, "anthropic_skill_registry"):
             await self.anthropic_skill_registry.stop_hot_reload()
 
+        attachment_registry = getattr(self, "attachment_registry", None)
+        if attachment_registry is not None and hasattr(attachment_registry, "flush"):
+            try:
+                await attachment_registry.flush()
+            except Exception as exc:
+                logger.warning("[清理] 刷新附件注册表失败: %s", exc)
+
         # 3) 最后关闭共享 HTTP client
         if hasattr(self, "_http_client"):
             logger.info("[清理] 正在关闭 AIClient HTTP 客户端...")
@@ -973,6 +982,10 @@ class AIClient:
         tool_context.setdefault("onebot_client", onebot_client)
         tool_context.setdefault("scheduler", scheduler)
         tool_context.setdefault("send_image_callback", self._send_image_callback)
+        tool_context.setdefault(
+            "attachment_registry",
+            getattr(self, "attachment_registry", None),
+        )
         tool_context.setdefault("memory_storage", self.memory_storage)
         tool_context.setdefault("knowledge_manager", self._knowledge_manager)
         tool_context.setdefault("cognitive_service", self._cognitive_service)

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/src/Undefined/ai/prompts.py

@@ -11,6 +11,7 @@ from typing import Any, Callable, Awaitable, Literal
 
 import aiofiles
 
+from Undefined.attachments import attachment_refs_to_xml
 from Undefined.context import RequestContext
 from Undefined.end_summary_storage import (
     EndSummaryStorage,
@@ -26,7 +27,7 @@ from Undefined.utils.xml import escape_xml_attr, escape_xml_text
 logger = logging.getLogger(__name__)
 
 _CURRENT_MESSAGE_RE = re.compile(
-    r"<message\b(?P<attrs>[^>]*)>\s*<content>(?P<content>.*?)</content>\s*</message>",
+    r"<message\b(?P<attrs>[^>]*)>.*?<content>(?P<content>.*?)</content>.*?</message>",
     re.DOTALL | re.IGNORECASE,
 )
 _XML_ATTR_RE = re.compile(r'(?P<key>[a-zA-Z_][a-zA-Z0-9_-]*)="(?P<value>[^"]*)"')
@@ -326,6 +327,15 @@ class PromptBuilder:
                     len(self._anthropic_skill_registry.get_all_skills()),
                 )
 
+        each_rules = await self._load_each_rules()
+        if each_rules:
+            messages.append(
+                {
+                    "role": "system",
+                    "content": f"【强制规则 - 必须在进行任何操作前仔细阅读并严格遵守】\n{each_rules}",
+                }
+            )
+
         if self._memory_storage:
             memories = self._memory_storage.get_all()
             if memories:
@@ -519,15 +529,6 @@ class PromptBuilder:
             }
         )
 
-        each_rules = await self._load_each_rules()
-        if each_rules:
-            messages.append(
-                {
-                    "role": "system",
-                    "content": f"【强制规则 - 必须在进行任何操作前仔细阅读并严格遵守】\n{each_rules}",
-                }
-            )
-
         messages.append({"role": "user", "content": f"【当前消息】\n{question}"})
         logger.debug(
             "[Prompt] messages_ready=%s question_len=%s",
@@ -648,6 +649,7 @@ class PromptBuilder:
                 chat_name = msg.get("chat_name", "未知群聊")
                 timestamp = msg.get("timestamp", "")
                 text = msg.get("message", "")
+                attachments = msg.get("attachments", [])
                 role = msg.get("role", "member")
                 title = msg.get("title", "")
                 message_id = msg.get("message_id")
@@ -664,6 +666,11 @@ class PromptBuilder:
                 msg_id_attr = ""
                 if message_id is not None:
                     msg_id_attr = f' message_id="{escape_xml_attr(str(message_id))}"'
+                attachment_xml = (
+                    f"\n{attachment_refs_to_xml(attachments)}"
+                    if isinstance(attachments, list) and attachments
+                    else ""
+                )
 
                 if msg_type_val == "group":
                     location = (
@@ -673,14 +680,14 @@ class PromptBuilder:
                     xml_msg = (
                         f'<message{msg_id_attr} sender="{safe_sender}" sender_id="{safe_sender_id}" group_id="{safe_chat_id}" '
                         f'group_name="{safe_chat_name}" location="{safe_location}" role="{safe_role}" title="{safe_title}" '
-                        f'time="{safe_time}">\n<content>{safe_text}</content>\n</message>'
+                        f'time="{safe_time}">\n<content>{safe_text}</content>{attachment_xml}\n</message>'
                     )
                 else:
                     location = "私聊"
                     safe_location = escape_xml_attr(location)
                     xml_msg = (
                         f'<message{msg_id_attr} sender="{safe_sender}" sender_id="{safe_sender_id}" location="{safe_location}" '
-                        f'time="{safe_time}">\n<content>{safe_text}</content>\n</message>'
+                        f'time="{safe_time}">\n<content>{safe_text}</content>{attachment_xml}\n</message>'
                     )
                 context_lines.append(xml_msg)
 

{undefined_bot-3.2.7 → undefined_bot-3.2.8}/src/Undefined/api/app.py

@@ -21,11 +21,18 @@ from aiohttp import ClientSession, ClientTimeout, web
 from aiohttp.web_response import Response
 
 from Undefined import __version__
+from Undefined.attachments import (
+    attachment_refs_to_xml,
+    build_attachment_scope,
+    register_message_attachments,
+    render_message_with_pic_placeholders,
+)
 from Undefined.config import load_webui_settings
 from Undefined.context import RequestContext
 from Undefined.context_resource_registry import collect_context_resources
 from Undefined.render import render_html_to_image, render_markdown_to_html  # noqa: F401
 from Undefined.services.queue_manager import QUEUE_LANE_SUPERADMIN
+from Undefined.utils.common import message_to_segments
 from Undefined.utils.cors import is_allowed_cors_origin, normalize_origin
 from Undefined.utils.recent_messages import get_recent_messages_prefer_local
 from Undefined.utils.xml import escape_xml_attr, escape_xml_text
@@ -40,6 +47,10 @@ _PROCESS_START_TIME = time.time()
 _NAGA_REQUEST_UUID_TTL_SECONDS = 6 * 60 * 60
 
 
+class _ToolInvokeExecutionTimeoutError(asyncio.TimeoutError):
+    """由 Runtime API 工具调用超时包装器抛出的超时异常。"""
+
+
 @dataclass
 class _NagaRequestResult:
     payload_hash: str
@@ -71,8 +82,16 @@ class _WebUIVirtualSender:
         mark_sent: bool = True,
         reply_to: int | None = None,
         preferred_temp_group_id: int | None = None,
+        history_message: str | None = None,
     ) -> int | None:
-        _ = user_id, auto_history, mark_sent, reply_to, preferred_temp_group_id
+        _ = (
+            user_id,
+            auto_history,
+            mark_sent,
+            reply_to,
+            preferred_temp_group_id,
+            history_message,
+        )
         await self._send_private_callback(self._virtual_user_id, message)
         return None
 
@@ -85,8 +104,16 @@ class _WebUIVirtualSender:
         *,
         mark_sent: bool = True,
         reply_to: int | None = None,
+        history_message: str | None = None,
     ) -> int | None:
-        _ = group_id, auto_history, history_prefix, mark_sent, reply_to
+        _ = (
+            group_id,
+            auto_history,
+            history_prefix,
+            mark_sent,
+            reply_to,
+            history_message,
+        )
         await self._send_private_callback(self._virtual_user_id, message)
         return None
 
@@ -1177,14 +1204,29 @@ class RuntimeAPIServer:
     ) -> str:
         cfg = self._ctx.config_getter()
         permission_sender_id = int(cfg.superadmin_qq)
+        webui_scope_key = build_attachment_scope(
+            user_id=_VIRTUAL_USER_ID,
+            request_type="private",
+            webui_session=True,
+        )
+        input_segments = message_to_segments(text)
+        registered_input = await register_message_attachments(
+            registry=self._ctx.ai.attachment_registry,
+            segments=input_segments,
+            scope_key=webui_scope_key,
+            resolve_image_url=self._ctx.onebot.get_image,
+            get_forward_messages=self._ctx.onebot.get_forward_msg,
+        )
+        normalized_text = registered_input.normalized_text or text
         await self._ctx.history_manager.add_private_message(
             user_id=_VIRTUAL_USER_ID,
-            text_content=text,
+            text_content=normalized_text,
             display_name=_VIRTUAL_USER_NAME,
             user_name=_VIRTUAL_USER_NAME,
+            attachments=registered_input.attachments,
         )
 
-        command = self._ctx.command_dispatcher.parse_command(text)
+        command = self._ctx.command_dispatcher.parse_command(normalized_text)
         if command:
             await self._ctx.command_dispatcher.dispatch_private(
                 user_id=_VIRTUAL_USER_ID,
@@ -1196,8 +1238,13 @@ class RuntimeAPIServer:
             return "command"
 
         current_time = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+        attachment_xml = (
+            f"\n{attachment_refs_to_xml(registered_input.attachments)}"
+            if registered_input.attachments
+            else ""
+        )
         full_question = f"""<message sender="{escape_xml_attr(_VIRTUAL_USER_NAME)}" sender_id="{escape_xml_attr(_VIRTUAL_USER_ID)}" location="WebUI私聊" time="{escape_xml_attr(current_time)}">
- <content>{escape_xml_text(text)}</content>
+ <content>{escape_xml_text(normalized_text)}</content>{attachment_xml}
  </message>
 
 【WebUI 会话】
@@ -1252,6 +1299,8 @@ class RuntimeAPIServer:
                 if value is not None:
                     ctx.set_resource(key, value)
             ctx.set_resource("queue_lane", QUEUE_LANE_SUPERADMIN)
+            ctx.set_resource("webui_session", True)
+            ctx.set_resource("webui_permission", "superadmin")
 
             result = await self._ctx.ai.ask(
                 full_question,
@@ -1330,18 +1379,32 @@ class RuntimeAPIServer:
 
         stream = _to_bool(body.get("stream"))
         outputs: list[str] = []
+        webui_scope_key = build_attachment_scope(
+            user_id=_VIRTUAL_USER_ID,
+            request_type="private",
+            webui_session=True,
+        )
 
         async def _capture_private_message(user_id: int, message: str) -> None:
             _ = user_id
             content = str(message or "").strip()
             if not content:
                 return
-            outputs.append(content)
+            rendered = await render_message_with_pic_placeholders(
+                content,
+                registry=self._ctx.ai.attachment_registry,
+                scope_key=webui_scope_key,
+                strict=False,
+            )
+            if not rendered.delivery_text.strip():
+                return
+            outputs.append(rendered.delivery_text)
             await self._ctx.history_manager.add_private_message(
                 user_id=_VIRTUAL_USER_ID,
-                text_content=content,
+                text_content=rendered.history_text,
                 display_name="Bot",
                 user_name="Bot",
+                attachments=rendered.attachments,
             )
 
         if not stream:
@@ -1368,8 +1431,11 @@ class RuntimeAPIServer:
         message_queue: asyncio.Queue[str] = asyncio.Queue()
 
         async def _capture_private_message_stream(user_id: int, message: str) -> None:
+            output_count = len(outputs)
             await _capture_private_message(user_id, message)
-            content = str(message or "").strip()
+            if len(outputs) <= output_count:
+                return
+            content = outputs[-1].strip()
             if content:
                 await message_queue.put(content)
 
@@ -1509,6 +1575,43 @@ class RuntimeAPIServer:
 
         return filtered
 
+    def _get_agent_tool_names(self) -> set[str]:
+        ai = self._ctx.ai
+        if ai is None:
+            return set()
+
+        agent_reg = getattr(ai, "agent_registry", None)
+        if agent_reg is None:
+            return set()
+
+        agent_names: set[str] = set()
+        for schema in agent_reg.get_agents_schema():
+            func = schema.get("function", {})
+            name = str(func.get("name", ""))
+            if name:
+                agent_names.add(name)
+        return agent_names
+
+    def _resolve_tool_invoke_timeout(
+        self, tool_name: str, timeout: int
+    ) -> float | None:
+        if tool_name in self._get_agent_tool_names():
+            return None
+        return float(timeout)
+
+    async def _await_tool_invoke_result(
+        self,
+        awaitable: Awaitable[Any],
+        *,
+        timeout: float | None,
+    ) -> Any:
+        if timeout is None or timeout <= 0:
+            return await awaitable
+        try:
+            return await asyncio.wait_for(awaitable, timeout=timeout)
+        except asyncio.TimeoutError as exc:
+            raise _ToolInvokeExecutionTimeoutError from exc
+
     async def _tools_list_handler(self, request: web.Request) -> Response:
         _ = request
         cfg = self._ctx.config_getter()
@@ -1659,6 +1762,7 @@ class RuntimeAPIServer:
         )
 
         start = time.perf_counter()
+        effective_timeout = self._resolve_tool_invoke_timeout(tool_name, timeout)
         try:
             async with RequestContext(
                 request_type=request_type,
@@ -1699,9 +1803,9 @@ class RuntimeAPIServer:
                 if tool_manager is None:
                     raise RuntimeError("ToolManager not available")
 
-                raw_result = await asyncio.wait_for(
+                raw_result = await self._await_tool_invoke_result(
                     tool_manager.execute_tool(tool_name, args, tool_context),
-                    timeout=timeout,
+                    timeout=effective_timeout,
                 )
 
             elapsed_ms = round((time.perf_counter() - start) * 1000, 1)
@@ -1722,7 +1826,7 @@ class RuntimeAPIServer:
                 "duration_ms": elapsed_ms,
             }
 
-        except asyncio.TimeoutError:
+        except _ToolInvokeExecutionTimeoutError:
             elapsed_ms = round((time.perf_counter() - start) * 1000, 1)
             logger.warning(
                 "[ToolInvoke] 执行超时: request_id=%s tool=%s timeout=%ds",