Undefined-bot 3.2.5__tar.gz → 3.2.6__tar.gz

{undefined_bot-3.2.5 → undefined_bot-3.2.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: Undefined-bot
-Version: 3.2.5
+Version: 3.2.6
 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
@@ -10,11 +10,11 @@ Requires-Python: <3.14,>=3.11
 Requires-Dist: aiofiles>=25.1.0
 Requires-Dist: aiohttp>=3.13.2
 Requires-Dist: apscheduler>=3.10.0
-Requires-Dist: chardet>=5.2.0
-Requires-Dist: chromadb>=0.6.0
-Requires-Dist: crawl4ai>=0.3.0
-Requires-Dist: croniter>=2.0.0
-Requires-Dist: fastmcp>=2.14.4
+Requires-Dist: chardet>=7.2.0
+Requires-Dist: chromadb>=1.5.5
+Requires-Dist: crawl4ai>=0.8.5
+Requires-Dist: croniter>=6.2.2
+Requires-Dist: fastmcp>=3.1.1
 Requires-Dist: httpx>=0.27.0
 Requires-Dist: imgkit
 Requires-Dist: langchain-community>=0.3.0
@@ -27,14 +27,14 @@ Requires-Dist: matplotlib
 Requires-Dist: mdit-py-plugins>=0.5.0
 Requires-Dist: numba>=0.61.0
 Requires-Dist: oh-my-bilibili>=0.1.2
-Requires-Dist: openai>=2.15.0
+Requires-Dist: openai>=2.29.0
 Requires-Dist: openpyxl>=3.1.5
 Requires-Dist: pillow
 Requires-Dist: playwright>=1.57.0
 Requires-Dist: psutil>=7.2.2
 Requires-Dist: py7zr>=1.1.0
 Requires-Dist: pymdown-extensions>=10.20
-Requires-Dist: pymupdf>=1.26.7
+Requires-Dist: pymupdf!=1.27.2.2,>=1.27.2
 Requires-Dist: pypinyin>=0.53.0
 Requires-Dist: python-docx>=1.2.0
 Requires-Dist: python-dotenv>=1.0.0
@@ -49,13 +49,13 @@ Requires-Dist: types-markdown>=3.10.0.20251106
 Requires-Dist: websockets>=12.0
 Provides-Extra: ci
 Requires-Dist: mypy>=1.8.0; extra == 'ci'
-Requires-Dist: ruff; extra == 'ci'
+Requires-Dist: ruff>=0.15.6; extra == 'ci'
 Provides-Extra: dev
 Requires-Dist: mypy>=1.8.0; extra == 'dev'
 Requires-Dist: pytest-asyncio>=0.23.0; extra == 'dev'
 Requires-Dist: pytest-cov>=5.0.0; extra == 'dev'
 Requires-Dist: pytest>=8.0.0; extra == 'dev'
-Requires-Dist: ruff; extra == 'dev'
+Requires-Dist: ruff>=0.15.6; extra == 'dev'
 Description-Content-Type: text/markdown
 
 <table border="0">

{undefined_bot-3.2.5 → undefined_bot-3.2.6}/config.toml.example

@@ -25,8 +25,8 @@ process_poke_message = true
 # zh: 注入给模型的最近历史消息条数上限（0-200，0 表示不注入历史）。
 # en: Max number of recent messages injected as model context (0-200; 0 disables history injection).
 context_recent_messages_limit = 20
-# zh: 模型调用失败时最大重试次数（0=不重试，默认2，最大5）。失败的请求将被放回队列最前端优先重试。
-# en: Max retries for failed model API requests (0=no retry, default 2, max 5). Failed requests are re-queued with highest priority.
+# zh: 单次 LLM 请求失败时最大静默重试次数（0=不重试，默认2）。失败的请求会回到原队列的第 2 个位置。
+# en: Max silent retries for a single failed LLM request (0=no retry, default 2). Failed requests are reinserted at position 2 of their original queue lane.
 ai_request_max_retries = 2
 
 # zh: 访问控制。mode 可选：off（不启用）/ blacklist（按黑名单）/ allowlist（按白名单）。
@@ -86,8 +86,8 @@ model_name = ""
 # zh: 最大生成 tokens。
 # en: Max generation tokens.
 max_tokens = 8192
-# zh: 队列发车间隔（秒）。
-# en: Queue interval (seconds).
+# zh: 队列发车间隔（秒，0 表示立即发车）。
+# en: Queue interval (seconds; 0 dispatches immediately).
 queue_interval_seconds = 1.0
 # zh: API 模式：传统 chat.completions 或新版 responses。
 # en: API mode: classic chat.completions or the newer responses API.
@@ -149,8 +149,8 @@ api_key = ""
 # zh: Vision 模型名称。
 # en: Vision model name.
 model_name = ""
-# zh: 队列发车间隔（秒）。
-# en: Queue interval (seconds).
+# zh: 队列发车间隔（秒，0 表示立即发车）。
+# en: Queue interval (seconds; 0 dispatches immediately).
 queue_interval_seconds = 1.0
 # zh: API 模式：传统 chat.completions 或新版 responses。
 # en: API mode: classic chat.completions or the newer responses API.
@@ -205,8 +205,8 @@ model_name = ""
 # zh: 可选限制：最大生成 tokens。
 # en: Optional limit: max generation tokens.
 max_tokens = 100
-# zh: 队列发车间隔（秒）。
-# en: Queue interval (seconds).
+# zh: 队列发车间隔（秒，0 表示立即发车）。
+# en: Queue interval (seconds; 0 dispatches immediately).
 queue_interval_seconds = 1.0
 # zh: API 模式：传统 chat.completions 或新版 responses。
 # en: API mode: classic chat.completions or the newer responses API.
@@ -260,8 +260,8 @@ model_name = ""
 # zh: 可选限制：最大生成 tokens。
 # en: Optional limit: max generation tokens.
 max_tokens = 160
-# zh: 队列发车间隔（秒）。
-# en: Queue interval (seconds).
+# zh: 队列发车间隔（秒，0 表示立即发车）。
+# en: Queue interval (seconds; 0 dispatches immediately).
 queue_interval_seconds = 1.0
 # zh: API 模式：传统 chat.completions 或新版 responses。
 # en: API mode: classic chat.completions or the newer responses API.
@@ -313,8 +313,8 @@ model_name = ""
 # zh: 可选限制：最大生成 tokens。
 # en: Optional limit: max generation tokens.
 max_tokens = 4096
-# zh: 队列发车间隔（秒）。
-# en: Queue interval (seconds).
+# zh: 队列发车间隔（秒，0 表示立即发车）。
+# en: Queue interval (seconds; 0 dispatches immediately).
 queue_interval_seconds = 1.0
 # zh: API 模式：传统 chat.completions 或新版 responses。
 # en: API mode: classic chat.completions or the newer responses API.
@@ -379,8 +379,8 @@ model_name = ""
 # zh: 可选限制：最大生成 tokens。
 # en: Optional limit: max generation tokens.
 max_tokens = 4096
-# zh: 队列发车间隔（秒）。
-# en: Queue interval (seconds).
+# zh: 队列发车间隔（秒，0 表示立即发车）。
+# en: Queue interval (seconds; 0 dispatches immediately).
 queue_interval_seconds = 1.0
 # zh: API 模式：传统 chat.completions 或新版 responses。
 # en: API mode: classic chat.completions or the newer responses API.
@@ -429,8 +429,8 @@ api_key = ""
 # zh: Embedding 模型名称。
 # en: Embedding model name.
 model_name = ""
-# zh: 队列发车间隔（秒）。
-# en: Queue interval (seconds).
+# zh: 队列发车间隔（秒，0 表示立即发车）。
+# en: Queue interval (seconds; 0 dispatches immediately).
 queue_interval_seconds = 1.0
 # zh: 向量维度（可选）。0 或留空表示使用模型默认维度。
 # en: Embedding dimensions (optional). Use 0/empty to use model defaults.
@@ -458,8 +458,8 @@ api_key = ""
 # zh: Rerank 模型名称。
 # en: Rerank model name.
 model_name = ""
-# zh: 队列发车间隔（秒）。
-# en: Queue interval (seconds).
+# zh: 队列发车间隔（秒，0 表示立即发车）。
+# en: Queue interval (seconds; 0 dispatches immediately).
 queue_interval_seconds = 1.0
 # zh: 查询端指令前缀（可选，部分 rerank 模型需要，如 "Instruct: ...\\nQuery: "）。
 # en: Query instruction prefix (optional, required by some rerank models, e.g. "Instruct: ...\\nQuery: ").
@@ -597,8 +597,8 @@ hot_reload_debounce = 0.5
 # zh: Agent intro 自动生成（推荐开启）。
 # en: Agent intro auto-generation (recommended).
 intro_autogen_enabled = true
-# zh: 队列发车间隔（秒）。
-# en: Queue interval (seconds).
+# zh: 队列发车间隔（秒，0 表示立即发车）。
+# en: Queue interval (seconds; 0 dispatches immediately).
 intro_autogen_queue_interval = 1.0
 # zh: 单次生成最大 token。
 # en: Max tokens per generation.
@@ -952,6 +952,11 @@ api_url = ""
 # zh: 双方共享密钥（Undefined ↔ Naga 身份验证）。
 # en: Shared secret key for authentication between Undefined and Naga.
 api_key = ""
+# zh: 是否启用 Naga 外发消息审核。
+# zh: 关闭后 /api/v1/naga/messages/send 会直接跳过本地审核，返回 moderation.status=skipped_disabled。
+# en: Whether to enable Naga outbound message moderation.
+# en: When disabled, /api/v1/naga/messages/send skips local moderation and returns moderation.status=skipped_disabled.
+moderation_enabled = true
 # zh: Naga 白名单群：/naga 在这些群中才可见；绑定和群聊投递也仅限这些群。
 # en: Naga allowlisted groups: /naga is only visible here, and binding/group delivery are restricted to these groups.
 allowed_groups = []

{undefined_bot-3.2.5 → undefined_bot-3.2.6}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "Undefined-bot"
-version = "3.2.5"
+version = "3.2.6"
 description = "QQ bot platform with cognitive memory architecture and multi-agent Skills, via OneBot V11."
 readme = "README.md"
 authors = [
@@ -12,9 +12,9 @@ dependencies = [
     "httpx>=0.27.0",
     "python-dotenv>=1.0.0",
     "tiktoken>=0.7.0",
-    "chardet>=5.2.0",
+    "chardet>=7.2.0",
     "langchain-community>=0.3.0",
-    "crawl4ai>=0.3.0",
+    "crawl4ai>=0.8.5",
     "matplotlib",
     "pillow",
     "imgkit",
@@ -31,22 +31,22 @@ dependencies = [
     "aiohttp>=3.13.2",
     "types-aiofiles>=25.1.0.20251011",
     "APScheduler>=3.10.0",
-    "croniter>=2.0.0",
-    "pymupdf>=1.26.7",
+    "croniter>=6.2.2",
+    "pymupdf>=1.27.2,!=1.27.2.2",
     "python-docx>=1.2.0",
     "python-pptx>=1.0.2",
     "openpyxl>=3.1.5",
     "py7zr>=1.1.0",
     "rarfile>=4.2",
-    "fastmcp>=2.14.4",
+    "fastmcp>=3.1.1",
     "lxml>=5.4.0",
     "rich>=14.2.0",
-    "openai>=2.15.0",
+    "openai>=2.29.0",
     "psutil>=7.2.2",
     "pyyaml>=6.0.3",
     "pypinyin>=0.53.0",
     "oh-my-bilibili>=0.1.2",
-    "chromadb>=0.6.0",
+    "chromadb>=1.5.5",
     "numba>=0.61.0",
 ]
 
@@ -62,13 +62,13 @@ Undefined-webui = "Undefined.webui:run"
 [project.optional-dependencies]
 dev = [
     "mypy>=1.8.0",
-    "ruff",
+    "ruff>=0.15.6",
     "pytest>=8.0.0",
     "pytest-asyncio>=0.23.0",
     "pytest-cov>=5.0.0",
 ]
 ci = [
-    "ruff",
+    "ruff>=0.15.6",
     "mypy>=1.8.0",
 ]
 
@@ -105,6 +105,10 @@ include = [
 python_version = "3.12"
 strict = true
 ignore_missing_imports = true
+exclude = ["^code/"]
+
+[tool.ruff]
+exclude = ["code"]
 
 [tool.pytest.ini_options]
 asyncio_mode = "auto"
@@ -113,13 +117,13 @@ testpaths = ["tests"]
 [dependency-groups]
 dev = [
     "mypy>=1.8.0",
-    "ruff",
+    "ruff>=0.15.6",
     "pytest>=8.0.0",
     "pytest-asyncio>=0.23.0",
     "pytest-cov>=5.0.0",
     "types-pyyaml>=6.0.12.20250915",
 ]
 ci = [
-    "ruff",
+    "ruff>=0.15.6",
     "mypy>=1.8.0",
 ]

{undefined_bot-3.2.5 → undefined_bot-3.2.6}/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__ = "3.2.5"
+__version__ = "3.2.6"

{undefined_bot-3.2.5 → undefined_bot-3.2.6}/src/Undefined/ai/client.py

@@ -17,7 +17,12 @@ from Undefined.ai.llm import ModelRequester
 from Undefined.ai.model_selector import ModelSelector
 from Undefined.ai.multimodal import MultimodalAnalyzer
 from Undefined.ai.prompts import PromptBuilder
+from Undefined.ai.queue_budget import (
+    compute_queued_llm_timeout_seconds,
+    resolve_effective_retry_count,
+)
 from Undefined.ai.summaries import SummaryService
+from Undefined.ai.transports.openai_transport import RESPONSES_OUTPUT_ITEMS_KEY
 from Undefined.ai.tokens import TokenCounter
 from Undefined.ai.tooling import ToolManager
 from Undefined.config import (
@@ -37,6 +42,15 @@ from Undefined.skills.agents.intro_generator import (
 )
 from Undefined.skills.anthropic_skills import AnthropicSkillRegistry
 from Undefined.skills.tools import ToolRegistry
+from Undefined.services.queue_manager import (
+    ALL_QUEUE_LANES,
+    QUEUE_LANE_BACKGROUND,
+    QUEUE_LANE_GROUP_MENTION,
+    QUEUE_LANE_GROUP_NORMAL,
+    QUEUE_LANE_GROUP_SUPERADMIN,
+    QUEUE_LANE_PRIVATE,
+    QUEUE_LANE_SUPERADMIN,
+)
 from Undefined.token_usage_storage import TokenUsageStorage
 from Undefined.utils.logging import log_debug_json, redact_string
 from Undefined.utils.tool_calls import parse_tool_arguments
@@ -312,7 +326,46 @@ class AIClient:
 
         logger.info("[清理] AIClient 已关闭")
 
-    async def submit_background_llm_call(
+    def _resolve_queue_lane(self, queue_lane: Any = None) -> str:
+        queue_lane_text = str(queue_lane or "").strip().lower()
+        if queue_lane_text in ALL_QUEUE_LANES:
+            return queue_lane_text
+
+        ctx = RequestContext.current()
+        if ctx is not None:
+            ctx_lane = str(ctx.get_resource("queue_lane") or "").strip().lower()
+            if ctx_lane in ALL_QUEUE_LANES:
+                return ctx_lane
+
+            runtime_config = self._get_runtime_config()
+            superadmin_qq = int(getattr(runtime_config, "superadmin_qq", 0) or 0)
+            if ctx.request_type == "private":
+                if superadmin_qq > 0 and (
+                    ctx.user_id == superadmin_qq or ctx.sender_id == superadmin_qq
+                ):
+                    return QUEUE_LANE_SUPERADMIN
+                return QUEUE_LANE_PRIVATE
+            if ctx.request_type == "group":
+                if superadmin_qq > 0 and ctx.sender_id == superadmin_qq:
+                    return QUEUE_LANE_GROUP_SUPERADMIN
+                if bool(ctx.get_resource("is_at_bot")):
+                    return QUEUE_LANE_GROUP_MENTION
+                return QUEUE_LANE_GROUP_NORMAL
+
+        return QUEUE_LANE_BACKGROUND
+
+    def _get_queued_llm_wait_timeout_seconds(self) -> float:
+        retry_count = resolve_effective_retry_count(
+            self._get_runtime_config(),
+            getattr(self, "_queue_manager", None),
+        )
+        return compute_queued_llm_timeout_seconds(
+            self._get_runtime_config(),
+            self.chat_config,
+            retry_count=retry_count,
+        )
+
+    async def submit_queued_llm_call(
         self,
         model_config: Any,
         messages: list[dict[str, Any]],
@@ -321,14 +374,16 @@ class AIClient:
         call_type: str = "background",
         max_tokens: int | None = None,
         transport_state: dict[str, Any] | None = None,
+        queue_lane: str | None = None,
     ) -> dict[str, Any]:
-        """将 LLM 调用投递到后台队列，走统一发车间隔和 Token 统计。
+        """将 LLM 调用投递到统一队列，走统一发车间隔和重试逻辑。
         无 queue_manager 时降级为直接调用。"""
         effective_max_tokens = (
             max_tokens
             if max_tokens is not None
             else getattr(model_config, "max_tokens", 4096)
         )
+        resolved_queue_lane = self._resolve_queue_lane(queue_lane)
         if self._queue_manager is None:
             return await self.request_model(
                 model_config=model_config,
@@ -343,22 +398,60 @@ class AIClient:
         event: asyncio.Event = asyncio.Event()
         self._pending_llm_calls[request_id] = (event, None)
         model_name = getattr(model_config, "model_name", "default")
-        await self._queue_manager.add_background_request(
-            {
-                "type": "background_llm_call",
-                "request_id": request_id,
-                "model_config": model_config,
-                "messages": messages,
-                "tools": tools,
-                "tool_choice": tool_choice,
-                "call_type": call_type,
-                "max_tokens": effective_max_tokens,
-                "transport_state": transport_state,
-            },
+        request: dict[str, Any] = {
+            "type": "queued_llm_call",
+            "request_id": request_id,
+            "model_config": model_config,
+            "messages": messages,
+            "tools": tools,
+            "tool_choice": tool_choice,
+            "call_type": call_type,
+            "max_tokens": effective_max_tokens,
+            "transport_state": transport_state,
+        }
+        ctx = RequestContext.current()
+        if ctx is not None:
+            if ctx.group_id is not None:
+                request["group_id"] = ctx.group_id
+            if ctx.user_id is not None:
+                request["user_id"] = ctx.user_id
+        logger.info(
+            "[queued_llm_enqueue] request_id=%s call_type=%s model=%s lane=%s messages=%s tools=%s",
+            request_id,
+            call_type,
+            model_name,
+            resolved_queue_lane,
+            len(messages),
+            bool(tools),
+        )
+        receipt = await self._queue_manager.add_queued_llm_request(
+            request,
+            lane=resolved_queue_lane,
             model_name=model_name,
         )
+        wait_timeout = compute_queued_llm_timeout_seconds(
+            self._get_runtime_config(),
+            model_config,
+            retry_count=resolve_effective_retry_count(
+                self._get_runtime_config(), self._queue_manager
+            ),
+            initial_wait_seconds=float(
+                getattr(receipt, "estimated_wait_seconds", 0.0) or 0.0
+            ),
+            include_first_dispatch_interval=False,
+        )
         try:
-            await asyncio.wait_for(event.wait(), timeout=480.0)
+            await asyncio.wait_for(event.wait(), timeout=wait_timeout)
+        except asyncio.TimeoutError:
+            logger.exception(
+                "[queued_llm_wait_timeout] request_id=%s call_type=%s model=%s lane=%s timeout=%.1fs",
+                request_id,
+                call_type,
+                model_name,
+                resolved_queue_lane,
+                wait_timeout,
+            )
+            raise
         finally:
             entry = self._pending_llm_calls.pop(request_id, None)
         _, result = entry if entry is not None else (None, None)
@@ -366,6 +459,28 @@ class AIClient:
             raise result
         return result or {}
 
+    async def submit_background_llm_call(
+        self,
+        model_config: Any,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]] | None = None,
+        tool_choice: Any = "auto",
+        call_type: str = "background",
+        max_tokens: int | None = None,
+        transport_state: dict[str, Any] | None = None,
+    ) -> dict[str, Any]:
+        """后台 LLM 提交兼容包装。"""
+        return await self.submit_queued_llm_call(
+            model_config=model_config,
+            messages=messages,
+            tools=tools,
+            tool_choice=tool_choice,
+            call_type=call_type,
+            max_tokens=max_tokens,
+            transport_state=transport_state,
+            queue_lane=QUEUE_LANE_BACKGROUND,
+        )
+
     def set_llm_call_result(
         self, request_id: str, result: dict[str, Any] | Exception
     ) -> None:
@@ -868,18 +983,25 @@ class AIClient:
         max_iterations = 1000
         iteration = 0
         conversation_ended = False
-        any_tool_executed = False
         cot_compat = getattr(effective_chat_config, "thinking_tool_call_compat", False)
         cot_compat_logged = False
         cot_missing_logged = False
         transport_state: dict[str, Any] | None = None
+        queue_lane = self._resolve_queue_lane(tool_context.get("queue_lane"))
+        pre_tool_failure_count = 0
+        max_pre_tool_retries = max(
+            0,
+            int(getattr(self._get_runtime_config(), "ai_request_max_retries", 0) or 0),
+        )
 
         while iteration < max_iterations:
             iteration += 1
             logger.info(f"[AI决策] 开始第 {iteration} 轮迭代...")
+            message_checkpoint_len = len(messages)
+            transport_state_checkpoint = transport_state
 
             try:
-                result = await self.request_model(
+                result = await self.submit_queued_llm_call(
                     model_config=effective_chat_config,
                     messages=messages,
                     max_tokens=8192,
@@ -887,8 +1009,20 @@ class AIClient:
                     tools=tools,
                     tool_choice="auto",
                     transport_state=transport_state,
+                    queue_lane=queue_lane,
+                )
+            except Exception as exc:
+                logger.exception(
+                    "[queued_llm_error] call_type=chat model=%s lane=%s iteration=%s error=%s",
+                    effective_chat_config.model_name,
+                    queue_lane,
+                    iteration,
+                    exc,
                 )
+                raise
 
+            try:
+                tool_execution_started = False
                 tool_name_map = (
                     result.get("_tool_name_map") if isinstance(result, dict) else None
                 )
@@ -927,7 +1061,7 @@ class AIClient:
                 if cot_compat and tools and log_thinking and not cot_compat_logged:
                     cot_compat_logged = True
                     logger.info(
-                        "[思维链兼容] 多轮工具调用 reasoning_content 回传已启用"
+                        "[思维链兼容] 多轮工具调用 reasoning_content 本地回填已启用"
                     )
                 if (
                     cot_compat
@@ -967,6 +1101,9 @@ class AIClient:
                     "content": content,
                     "tool_calls": tool_calls,
                 }
+                output_items = message.get(RESPONSES_OUTPUT_ITEMS_KEY)
+                if isinstance(output_items, list):
+                    assistant_message[RESPONSES_OUTPUT_ITEMS_KEY] = output_items
                 if cot_compat and reasoning_content is not None:
                     assistant_message["reasoning_content"] = reasoning_content
                 messages.append(assistant_message)
@@ -1035,7 +1172,7 @@ class AIClient:
                     )
 
                 if tool_tasks:
-                    any_tool_executed = True
+                    tool_execution_started = True
                     logger.info(
                         "[工具执行] 开始并发执行 %s 个工具调用: %s",
                         len(tool_tasks),
@@ -1111,6 +1248,7 @@ class AIClient:
                         logger.info("[工具调用] end 与其他工具同时调用，已回填跳过响应")
                     else:
                         # end 单独调用，正常执行（参数已在循环中解析）
+                        tool_execution_started = True
                         end_result = await self.tool_manager.execute_tool(
                             "end", end_tool_args, tool_context
                         )
@@ -1129,13 +1267,34 @@ class AIClient:
                 if conversation_ended:
                     logger.info("[会话状态] 对话已结束（调用 end 工具）")
                     return ""
+                pre_tool_failure_count = 0
 
             except Exception as exc:
-                if not any_tool_executed:
-                    # 尚未执行任何工具（无消息发送等副作用），安全传播给上层重试
-                    raise
-                logger.exception("ask 处理失败: %s", exc)
-                return f"处理失败: {exc}"
+                if (
+                    not tool_execution_started
+                    and pre_tool_failure_count < max_pre_tool_retries
+                ):
+                    pre_tool_failure_count += 1
+                    del messages[message_checkpoint_len:]
+                    transport_state = transport_state_checkpoint
+                    logger.warning(
+                        "[chat.pre_tool_retry] model=%s lane=%s retry=%s/%s iteration=%s error=%s",
+                        effective_chat_config.model_name,
+                        queue_lane,
+                        pre_tool_failure_count,
+                        max_pre_tool_retries,
+                        iteration,
+                        exc,
+                    )
+                    continue
+                logger.exception(
+                    "[chat.suppressed_error] model=%s lane=%s iteration=%s error=%s",
+                    effective_chat_config.model_name,
+                    queue_lane,
+                    iteration,
+                    exc,
+                )
+                return ""
 
         logger.warning("[AI决策] 达到最大迭代次数，未能完成处理")
         return "达到最大迭代次数，未能完成处理"