sycommon-python-lib 0.2.2a18__tar.gz → 0.2.3a0__tar.gz

{sycommon_python_lib-0.2.2a18 → sycommon_python_lib-0.2.3a0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.2.2a18
+Version: 0.2.3a0
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown

{sycommon_python_lib-0.2.2a18 → sycommon_python_lib-0.2.3a0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sycommon-python-lib"
-version = "0.2.2a18"
+version = "0.2.3a0"
 description = "Add your description here"
 readme = "README.md"
 requires-python = ">=3.11"

{sycommon_python_lib-0.2.2a18 → sycommon_python_lib-0.2.3a0}/src/sycommon/agent/deep_agent.py

@@ -51,6 +51,7 @@ from sycommon.agent.sandbox.sandbox_recovery import SandboxRecoveryManager
 from sycommon.agent.chat_events import ChatEvent, ChatEventBuilder, DEFAULT_AGENT_NAME
 from sycommon.middleware.background_execution import BackgroundExecutionMiddleware
 from sycommon.middleware.token_tracking import TokenTrackingMiddleware
+from sycommon.middleware.tool_result_truncation import ToolResultTruncationMiddleware
 from deepagents.middleware.summarization import create_summarization_tool_middleware  # noqa: F401 保留 re-export
 from sycommon.agent.summarization_utils import build_summarization_middleware
 
@@ -107,7 +108,7 @@ class AgentConfig(BaseModel):
 
     # 沙箱配置
     sandbox_service_name: str = "shengye-platform-sandbox"
-    sandbox_timeout: int = 60
+    sandbox_timeout: int = 300
     skills_dir: Optional[str] = None
     memory_dir: Optional[str] = None
 
@@ -489,6 +490,8 @@ class DeepAgent:
         from openai import APIConnectionError, APIError, APITimeoutError
 
         last_error = None
+        sandbox_retries = 0
+        max_sandbox_retries = 3
 
         while True:
             for attempt in range(max_retries):
@@ -511,8 +514,9 @@ class DeepAgent:
                     else:
                         raise last_error
                 except RuntimeError as e:
-                    if "沙箱服务不可用" in str(e) and self.recovery_manager:
-                        SYLogger.warning("[DeepAgent] 沙箱服务不可用，尝试恢复...")
+                    if "沙箱服务不可用" in str(e) and self.recovery_manager and sandbox_retries < max_sandbox_retries:
+                        sandbox_retries += 1
+                        SYLogger.warning(f"[DeepAgent] 沙箱服务不可用，尝试恢复 ({sandbox_retries}/{max_sandbox_retries})...")
                         recovered = await self.recovery_manager.recover()
                         if recovered:
                             SYLogger.info("[DeepAgent] 沙箱已恢复，继续执行")
@@ -598,6 +602,7 @@ async def create_deep_agent(
         "debug": config.debug,
         "middleware": [
             BackgroundExecutionMiddleware(backend=sandbox_backend),
+            ToolResultTruncationMiddleware(),
             TokenTrackingMiddleware(model_name=config.model_name, user_id=user_id),
             summarization_mw,
         ],

{sycommon_python_lib-0.2.2a18 → sycommon_python_lib-0.2.3a0}/src/sycommon/agent/multi_agent_team.py

@@ -44,6 +44,7 @@ from sycommon.agent.sandbox.sandbox_recovery import SandboxRecoveryManager
 from sycommon.agent.chat_events import ChatEvent, ChatEventBuilder, DEFAULT_AGENT_NAME
 from sycommon.middleware.background_execution import BackgroundExecutionMiddleware
 from sycommon.middleware.token_tracking import TokenTrackingMiddleware
+from sycommon.middleware.tool_result_truncation import ToolResultTruncationMiddleware
 from deepagents.middleware.summarization import create_summarization_tool_middleware  # noqa: F401 保留 re-export
 from sycommon.agent.summarization_utils import build_summarization_middleware
 
@@ -165,7 +166,7 @@ class TeamConfig(BaseModel):
     coordinator_name: str = "项目经理"
 
     sandbox_service_name: str = "shengye-platform-sandbox"
-    sandbox_timeout: int = 60
+    sandbox_timeout: int = 300
     skills_dir: Optional[str] = None
     memory_dir: Optional[str] = None
 
@@ -585,6 +586,7 @@ async def create_multi_agent_team(
 
     middleware = [
         BackgroundExecutionMiddleware(backend=sandbox_backend),
+        ToolResultTruncationMiddleware(),
         TokenTrackingMiddleware(model_name=config.model_name, user_id=user_id),
         build_summarization_middleware(model, config.model_name, sandbox_backend),
     ]
@@ -640,6 +642,7 @@ async def create_multi_agent_team(
         name=coord_name,
         middleware=[
             BackgroundExecutionMiddleware(backend=sandbox_backend),
+            ToolResultTruncationMiddleware(),
             TokenTrackingMiddleware(model_name=config.model_name, user_id=user_id),
             build_summarization_middleware(model, config.model_name, sandbox_backend),
         ],

{sycommon_python_lib-0.2.2a18 → sycommon_python_lib-0.2.3a0}/src/sycommon/agent/summarization_utils.py

@@ -23,7 +23,7 @@ def build_summarization_middleware(
     model_name: str,
     backend: "BACKEND_TYPES",
     *,
-    trigger_fraction: float = 0.85,
+    trigger_fraction: float = 0.70,
     keep_fraction: float = 0.10,
     default_max_tokens: int = 200000,
 ) -> SummarizationToolMiddleware:
@@ -35,7 +35,7 @@ def build_summarization_middleware(
         model: LLM 实例。
         model_name: 模型名称（用于从 nacos 读取配置）。
         backend: 后端实例。
-        trigger_fraction: 触发压缩的上下文窗口比例，默认 85%。
+        trigger_fraction: 触发压缩的上下文窗口比例，默认 70%。
         keep_fraction: 压缩后保留的上下文窗口比例，默认 10%。
         default_max_tokens: 无法从配置读取时的默认上下文窗口大小。
 

{sycommon_python_lib-0.2.2a18 → sycommon_python_lib-0.2.3a0}/src/sycommon/llm/get_llm.py

@@ -153,7 +153,7 @@ def get_llm(
 
     # 根据 wrap_structured 决定默认超时
     if timeout is _TIMEOUT_UNSET:
-        timeout = 180 if wrap_structured else None
+        timeout = 180 if wrap_structured else 300
 
     init_params = {
         "model_provider": llmConfig.provider,
@@ -165,7 +165,7 @@ def get_llm(
         "streaming": streaming,
         "timeout": timeout,
         "max_retries": max_retries,
-        "stream_chunk_timeout": None,
+        "stream_chunk_timeout": 180,
     }
 
     # 合并其他透传参数（包括 presence_penalty, extra_body, top_p 等）

{sycommon_python_lib-0.2.2a18 → sycommon_python_lib-0.2.3a0}/src/sycommon/llm/struct_token.py

@@ -345,7 +345,6 @@ class StructuredRunnableWithToken(Runnable):
             try:
                 with self.langfuse.start_as_current_observation(as_type="span", name="invoke") as span:
                     with propagate_attributes(session_id=trace_id, user_id=user_id):
-                        span.update_trace(user_id=user_id, session_id=trace_id)
                         return self._execute_chain(input, config, trace_id, user_id, span)
             except Exception as e:
                 # Langfuse 跟踪失败不应阻断业务，降级执行
@@ -366,7 +365,6 @@ class StructuredRunnableWithToken(Runnable):
             try:
                 with self.langfuse.start_as_current_observation(as_type="span", name="ainvoke") as span:
                     with propagate_attributes(session_id=trace_id, user_id=user_id):
-                        span.update_trace(user_id=user_id, session_id=trace_id)
                         return await self._aexecute_chain(input, config, trace_id, user_id, span)
             except Exception as e:
                 # Langfuse 跟踪失败不应阻断业务，降级执行
@@ -405,7 +403,7 @@ class StructuredRunnableWithToken(Runnable):
                 input_data = {"messages": adapted_input}
 
             if span:
-                span.update_trace(input=input_data)
+                span.update(input=input_data)
 
             structured_result = self.retry_chain.invoke(
                 input_data,
@@ -413,7 +411,7 @@ class StructuredRunnableWithToken(Runnable):
             )
 
             if span:
-                span.update_trace(output=structured_result)
+                span.update(output=structured_result)
 
             token_usage = token_handler.usage_metadata
             structured_result._token_usage_ = token_usage
@@ -465,7 +463,7 @@ class StructuredRunnableWithToken(Runnable):
                 input_data = {"messages": adapted_input}
 
             if span:
-                span.update_trace(input=input_data)
+                span.update(input=input_data)
 
             # 3. 调用子链
             structured_result = await self.retry_chain.ainvoke(
@@ -474,7 +472,7 @@ class StructuredRunnableWithToken(Runnable):
             )
 
             if span:
-                span.update_trace(output=structured_result)
+                span.update(output=structured_result)
 
             token_usage = token_handler.usage_metadata
             structured_result._token_usage_ = token_usage

{sycommon_python_lib-0.2.2a18 → sycommon_python_lib-0.2.3a0}/src/sycommon/llm/sy_langfuse.py

@@ -3,11 +3,26 @@ from typing import Tuple, List, Optional, Any, Dict
 from langfuse import Langfuse, get_client
 from sycommon.config.Config import Config, SingletonMeta
 from sycommon.logging.kafka_log import SYLogger
-from langfuse.langchain import CallbackHandler
+from langfuse.langchain import CallbackHandler as _LangfuseCallbackHandler
 from sycommon.tools.env import get_env_var
 from sycommon.tools.merge_headers import get_header_value
 
 
+class SafeLangfuseCallbackHandler(_LangfuseCallbackHandler):
+    """Wraps Langfuse CallbackHandler to handle unpicklable error objects.
+
+    LangGraph errors (e.g. Command/Interrupt dataclasses) may hold
+    _asyncio.Task references.  When Langfuse calls str(error) → asdict() →
+    deepcopy(), this raises TypeError: cannot pickle '_asyncio.Task' object.
+    """
+
+    def on_chain_error(self, error, **kwargs):
+        try:
+            super().on_chain_error(error, **kwargs)
+        except (TypeError, ValueError):
+            pass
+
+
 class LangfuseInitializer(metaclass=SingletonMeta):
     """
     Langfuse 初始化管理器
@@ -53,7 +68,7 @@ class LangfuseInitializer(metaclass=SingletonMeta):
 
                 self._langfuse_client = get_client()
 
-                langfuse_handler = CallbackHandler()
+                langfuse_handler = SafeLangfuseCallbackHandler()
                 self._base_callbacks.append(langfuse_handler)
 
                 SYLogger.info(f"Langfuse 初始化成功 [Service: {server_name}]")

sycommon_python_lib-0.2.3a0/src/sycommon/middleware/tool_result_truncation.py

@@ -0,0 +1,136 @@
+"""工具结果截断中间件
+
+在工具执行结果返回给 LLM 之前，自动截断过长的 ToolMessage.content，
+防止 "Prompt exceeds max length" 错误。
+
+通过 wrap_tool_call / awrap_tool_call 钩子拦截每个工具调用结果，
+在结果进入对话历史之前完成截断。
+"""
+
+from collections.abc import Awaitable, Callable
+from typing import Any, FrozenSet
+
+from langchain.agents.middleware.types import (
+    AgentMiddleware,
+    AgentState,
+    ContextT,
+    ResponseT,
+)
+from langchain_core.messages import ToolMessage
+from langgraph.prebuilt.tool_node import ToolCallRequest
+from langgraph.types import Command
+
+from sycommon.logging.kafka_log import SYLogger
+
+
+# 不需要截断的工具（输出始终很短，或前端需要完整数据）
+DEFAULT_PASSTHROUGH_TOOLS: FrozenSet[str] = frozenset({
+    "get_download_url",
+    "get_current_date",
+    "memory",
+    "skill_manager",
+    "session_search",
+    "web_search",  # web_search 已在工具内部自行截断
+})
+
+# 截断提示模板
+DEFAULT_TRUNCATION_SUFFIX = (
+    "\n\n[输出已截断至前{kept}字符，原始长度{original}字符。"
+    "如需查看完整输出，可以：\n"
+    "1. 重新执行命令并添加过滤条件（如 grep、head、tail）\n"
+    "2. 将输出重定向到文件后用 read_file 分段读取]"
+)
+
+
+class ToolResultTruncationMiddleware(AgentMiddleware):
+    """截断过长的工具结果，防止超出模型上下文窗口。
+
+    Args:
+        max_content_length: 工具结果内容的最大字符长度，默认 15000（约 3750 token）。
+            模型上下文 200K token，70% 触发压缩 = 140K token，给单条工具结果留 3750 token
+            约占剩余空间的 2.6%，多条工具调用并行时也不会逼近上限。
+        passthrough_tools: 不截断的工具名集合。
+        truncation_suffix: 截断提示文本，支持 {kept} 和 {original} 占位符。
+    """
+
+    def __init__(
+        self,
+        *,
+        max_content_length: int = 15000,
+        passthrough_tools: FrozenSet[str] | None = None,
+        truncation_suffix: str = DEFAULT_TRUNCATION_SUFFIX,
+    ) -> None:
+        self._max_content_length = max_content_length
+        self._passthrough_tools = passthrough_tools or DEFAULT_PASSTHROUGH_TOOLS
+        self._truncation_suffix = truncation_suffix
+
+    @property
+    def name(self) -> str:
+        return "ToolResultTruncationMiddleware"
+
+    def _should_passthrough(self, tool_name: str) -> bool:
+        return tool_name in self._passthrough_tools
+
+    def _truncate(self, content: str, tool_name: str) -> str:
+        if self._should_passthrough(tool_name):
+            return content
+
+        original_len = len(content)
+        if original_len <= self._max_content_length:
+            return content
+
+        kept = self._max_content_length
+        suffix = self._truncation_suffix.format(kept=kept, original=original_len)
+        truncated = content[:kept] + suffix
+
+        SYLogger.info(
+            f"[ToolResultTruncation] tool='{tool_name}' truncated: "
+            f"{original_len} -> {len(truncated)} chars"
+        )
+        return truncated
+
+    def _process_result(
+        self,
+        result: Any,
+        tool_name: str,
+    ) -> Any:
+        """处理工具调用结果，截断过长的 ToolMessage.content"""
+        if isinstance(result, Command):
+            return result
+
+        if not isinstance(result, ToolMessage):
+            return result
+
+        content = result.content
+        if not isinstance(content, str):
+            return result
+
+        new_content = self._truncate(content, tool_name)
+        if new_content is content:
+            return result
+
+        return ToolMessage(
+            content=new_content,
+            tool_call_id=result.tool_call_id,
+            name=result.name,
+            status=result.status,
+            artifact=result.artifact,
+        )
+
+    def wrap_tool_call(
+        self,
+        request: ToolCallRequest,
+        handler: Callable[[ToolCallRequest], ToolMessage | Command],
+    ) -> ToolMessage | Command:
+        tool_name = request.tool_call.get("name", "")
+        result = handler(request)
+        return self._process_result(result, tool_name)
+
+    async def awrap_tool_call(
+        self,
+        request: ToolCallRequest,
+        handler: Callable[[ToolCallRequest], Awaitable[ToolMessage | Command]],
+    ) -> ToolMessage | Command:
+        tool_name = request.tool_call.get("name", "")
+        result = await handler(request)
+        return self._process_result(result, tool_name)

sycommon_python_lib-0.2.3a0/src/sycommon/notice/__init__.py

@@ -0,0 +1,5 @@
+from sycommon.notice.wecom_message import WeComMessageService
+
+__all__ = [
+    "WeComMessageService",
+]