sycommon-python-lib 0.2.3a11__tar.gz → 0.2.4a0__tar.gz

{sycommon_python_lib-0.2.3a11 → sycommon_python_lib-0.2.4a0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.2.3a11
+Version: 0.2.4a0
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown
@@ -8,17 +8,18 @@ Requires-Dist: aio-pika>=9.6.2
 Requires-Dist: aiohttp>=3.13.5
 Requires-Dist: aiomysql>=0.3.2
 Requires-Dist: anyio>=4.12.1
-Requires-Dist: decorator>=5.2.1
+Requires-Dist: decorator>=5.3.0
 Requires-Dist: deepagents>=0.6.1
 Requires-Dist: elasticsearch>=9.4.0
 Requires-Dist: fastapi>=0.136.1
 Requires-Dist: jinja2>=3.1.6
 Requires-Dist: kafka-python>=2.3.1
-Requires-Dist: langchain>=1.3.0
+Requires-Dist: langchain>=1.3.1
 Requires-Dist: langchain-core>=1.4.0
 Requires-Dist: langchain-openai>=1.2.1
 Requires-Dist: langfuse>=4.6.1
 Requires-Dist: langgraph>=1.2.0
+Requires-Dist: langgraph-checkpoint-postgres>=3.1.0
 Requires-Dist: langgraph-checkpoint-redis>=0.4.1
 Requires-Dist: ldap3>=2.9.1
 Requires-Dist: loguru>=0.7.3
@@ -28,17 +29,18 @@ Requires-Dist: psutil>=7.2.2
 Requires-Dist: pyxxl>=0.4.6
 Requires-Dist: pydantic>=2.13.4
 Requires-Dist: python-dotenv>=1.2.2
-Requires-Dist: python-multipart>=0.0.28
+Requires-Dist: python-multipart>=0.0.29
 Requires-Dist: pyyaml>=6.0.3
 Requires-Dist: redis>=7.3.0
 Requires-Dist: sentry-sdk[fastapi]>=2.60.0
 Requires-Dist: sqlalchemy[asyncio]>=2.0.48
 Requires-Dist: starlette[full]>=1.0.0
-Requires-Dist: tiktoken>=0.12.0
-Requires-Dist: uvicorn>=0.46.0
+Requires-Dist: tiktoken>=0.13.0
+Requires-Dist: uvicorn>=0.47.0
 Requires-Dist: wecom-aibot-python-sdk>=1.0.2
 Requires-Dist: twine>=6.2.0
 Requires-Dist: minio>=7.2.20
+Requires-Dist: langchain-mcp-adapters>=0.2.2
 
 # sycommon-python-lib
 

{sycommon_python_lib-0.2.3a11 → sycommon_python_lib-0.2.4a0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sycommon-python-lib"
-version = "0.2.3a11"
+version = "0.2.4a0"
 description = "Add your description here"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -9,17 +9,18 @@ dependencies = [
     "aiohttp>=3.13.5",
     "aiomysql>=0.3.2",
     "anyio>=4.12.1",
-    "decorator>=5.2.1",
+    "decorator>=5.3.0",
     "deepagents>=0.6.1",
     "elasticsearch>=9.4.0",
     "fastapi>=0.136.1",
     "jinja2>=3.1.6",
     "kafka-python>=2.3.1",
-    "langchain>=1.3.0",
+    "langchain>=1.3.1",
     "langchain-core>=1.4.0",
     "langchain-openai>=1.2.1",
     "langfuse>=4.6.1",
     "langgraph>=1.2.0",
+    "langgraph-checkpoint-postgres>=3.1.0",
     "langgraph-checkpoint-redis>=0.4.1",
     "ldap3>=2.9.1",
     "loguru>=0.7.3",
@@ -29,17 +30,18 @@ dependencies = [
     "pyxxl>=0.4.6",
     "pydantic>=2.13.4",
     "python-dotenv>=1.2.2",
-    "python-multipart>=0.0.28",
+    "python-multipart>=0.0.29",
     "pyyaml>=6.0.3",
     "redis>=7.3.0",
     "sentry-sdk[fastapi]>=2.60.0",
     "sqlalchemy[asyncio]>=2.0.48",
     "starlette[full]>=1.0.0",
-    "tiktoken>=0.12.0",
-    "uvicorn>=0.46.0",
+    "tiktoken>=0.13.0",
+    "uvicorn>=0.47.0",
     "wecom-aibot-python-sdk>=1.0.2",
     "twine>=6.2.0",
     "minio>=7.2.20",
+    "langchain-mcp-adapters>=0.2.2",
 ]
 
 [tool.setuptools]

{sycommon_python_lib-0.2.3a11 → sycommon_python_lib-0.2.4a0}/src/sycommon/agent/__init__.py

@@ -105,6 +105,18 @@ from sycommon.agent.chat_events import (
     error_event,
     cancelled_event,
 )
+from sycommon.agent.mcp import (
+    MCPToolStatus,
+    MCPServerConfig,
+    MCPServerCreateRequest,
+    MCPServerUpdateRequest,
+    MCPServerTestRequest,
+    MCPServerTestResult,
+    load_mcp_tools,
+    test_mcp_connection,
+    sanitize_name,
+)
+from sycommon.database.pg_checkpoint_service import PgCheckpointService
 
 __all__ = [
     # 沙箱
@@ -147,4 +159,18 @@ __all__ = [
     "done_event",
     "error_event",
     "cancelled_event",
+
+    # MCP 工具集成
+    "MCPToolStatus",
+    "MCPServerConfig",
+    "MCPServerCreateRequest",
+    "MCPServerUpdateRequest",
+    "MCPServerTestRequest",
+    "MCPServerTestResult",
+    "load_mcp_tools",
+    "test_mcp_connection",
+    "sanitize_name",
+
+    # PG Checkpoint 服务
+    "PgCheckpointService",
 ]

{sycommon_python_lib-0.2.3a11 → sycommon_python_lib-0.2.4a0}/src/sycommon/agent/deep_agent.py

@@ -168,6 +168,7 @@ class DeepAgent:
         """
         current_tool_calls = []
         ai_chunk_buffer = ""
+        ai_text_content = ""
         seen_tool_call_ids = set()
         stream_step = 0
         # 兜底：累积流式 chunk 中的 usage_metadata（middleware 在流式场景可能拿不到）
@@ -211,6 +212,12 @@ class DeepAgent:
                         if usage_meta:
                             total_input_tokens += usage_meta.get("input_tokens", 0)
                             total_output_tokens += usage_meta.get("output_tokens", 0)
+                            if usage_meta.get("input_tokens", 0) > 0:
+                                SYLogger.debug(
+                                    f"[DeepAgent] usage_metadata | input={usage_meta.get('input_tokens', 0)} "
+                                    f"output={usage_meta.get('output_tokens', 0)} "
+                                    f"total={usage_meta.get('total_tokens', 0)} "
+                                    f"cumulative_input={total_input_tokens} step={stream_step}")
 
                     if msg_type == "AIMessageChunk":
                         tool_calls_log = getattr(msg, "tool_calls", [])
@@ -397,6 +404,7 @@ class DeepAgent:
 
                     if content:
                         ai_chunk_buffer += content
+                        ai_text_content += content
                         event = ChatEventBuilder.ai_chunk(
                             content, id=getattr(msg, "id", None),
                             agent=DEFAULT_AGENT_NAME)
@@ -453,6 +461,13 @@ class DeepAgent:
                 print(
                     f"[DeepAgent] AI chunk done | {repr(ai_chunk_buffer[:100])}...")
 
+            # 空响应检测：模型被调用但没有产出任何文本
+            if not ai_text_content and not ai_chunk_buffer:
+                SYLogger.warning(
+                    f"[DeepAgent] 空响应警告：模型未返回任何文本内容。"
+                    f"stream_step={stream_step}, tool_calls={len(current_tool_calls)}, "
+                    f"input_tokens={total_input_tokens}, output_tokens={total_output_tokens}")
+
             # 兜底：如果 middleware 没有成功记录（流式场景），在这里补充记录
             if total_input_tokens > 0 or total_output_tokens > 0:
                 try:
@@ -598,8 +613,12 @@ async def create_deep_agent(
 
     # 创建 checkpointer
     if checkpointer is None:
-        from langgraph.checkpoint.memory import MemorySaver
-        checkpointer = MemorySaver()
+        from sycommon.database.pg_checkpoint_service import PgCheckpointService
+        if PgCheckpointService.is_initialized():
+            checkpointer = await PgCheckpointService.get_checkpointer()
+        else:
+            from langgraph.checkpoint.memory import MemorySaver
+            checkpointer = MemorySaver()
 
     tid = thread_id or user_id
     agent_config = {"configurable": {"thread_id": tid}}

sycommon_python_lib-0.2.4a0/src/sycommon/agent/mcp/__init__.py

@@ -0,0 +1,30 @@
+"""MCP 工具集成模块
+
+提供 MCP 服务器连接、工具发现和加载功能。
+"""
+
+from sycommon.agent.mcp.models import (
+    MCPToolStatus,
+    MCPServerConfig,
+    MCPServerCreateRequest,
+    MCPServerUpdateRequest,
+    MCPServerTestRequest,
+    MCPServerTestResult,
+)
+from sycommon.agent.mcp.tool_loader import (
+    load_mcp_tools,
+    test_mcp_connection,
+    sanitize_name,
+)
+
+__all__ = [
+    "MCPToolStatus",
+    "MCPServerConfig",
+    "MCPServerCreateRequest",
+    "MCPServerUpdateRequest",
+    "MCPServerTestRequest",
+    "MCPServerTestResult",
+    "load_mcp_tools",
+    "test_mcp_connection",
+    "sanitize_name",
+]

sycommon_python_lib-0.2.4a0/src/sycommon/agent/mcp/models.py

@@ -0,0 +1,56 @@
+"""MCP 服务器配置数据模型"""
+
+from typing import Optional, List
+from pydantic import BaseModel
+
+
+class MCPToolStatus(BaseModel):
+    """MCP 工具可用状态"""
+    tool_name: str
+    description: Optional[str] = None
+    available: bool = True
+    last_check_at: Optional[str] = None
+    last_error: Optional[str] = None
+
+
+class MCPServerConfig(BaseModel):
+    id: str
+    user_id: Optional[str] = None
+    name: str
+    server_url: str
+    description: Optional[str] = None
+    headers: Optional[dict] = None
+    enabled: bool = True
+    sanitized_name: str = ""
+    created_at: str
+    updated_at: str
+    tools: Optional[List[MCPToolStatus]] = None
+    server_status: Optional[str] = None
+    server_error: Optional[str] = None
+
+
+class MCPServerCreateRequest(BaseModel):
+    name: str
+    server_url: str
+    description: Optional[str] = None
+    headers: Optional[dict] = None
+    enabled: bool = True
+
+
+class MCPServerUpdateRequest(BaseModel):
+    name: Optional[str] = None
+    server_url: Optional[str] = None
+    description: Optional[str] = None
+    headers: Optional[dict] = None
+    enabled: Optional[bool] = None
+
+
+class MCPServerTestRequest(BaseModel):
+    server_url: str
+    headers: Optional[dict] = None
+
+
+class MCPServerTestResult(BaseModel):
+    success: bool
+    tools: Optional[list] = None
+    error: Optional[str] = None

sycommon_python_lib-0.2.4a0/src/sycommon/agent/mcp/tool_loader.py

@@ -0,0 +1,174 @@
+"""MCP 工具加载器
+
+使用 langchain-mcp-adapters 官方库连接远程 MCP 服务器，
+发现工具并作为 LangChain BaseTool 注入 Agent。
+"""
+
+import re
+import hashlib
+from typing import List, Optional, Callable, Awaitable
+
+from langchain_core.tools import BaseTool
+
+from sycommon.logging.kafka_log import SYLogger
+from sycommon.agent.mcp.models import MCPServerConfig
+
+
+def sanitize_name(name: str) -> str:
+    """将名称转为合法的标识符，中文等非ASCII字符做 transliterate"""
+    sanitized = re.sub(r'[^a-zA-Z0-9_]', '_', name).strip('_')
+    if not sanitized:
+        h = hashlib.md5(name.encode()).hexdigest()[:8]
+        sanitized = f"mcp_{h}"
+    return sanitized
+
+
+def wrap_tool_with_error_handler(
+    tool: BaseTool,
+    server_config_id: str,
+    server_name: str,
+    original_tool_name: str,
+    on_tool_success: Optional[Callable[[str, str], Awaitable[None]]] = None,
+    on_tool_error: Optional[Callable[[str, str, str], Awaitable[None]]] = None,
+) -> BaseTool:
+    """包装 MCP 工具的 coroutine，捕获连接/超时等异常，返回友好的错误信息而非抛出异常。
+
+    Args:
+        on_tool_success: 异步回调 (server_config_id, tool_name) -> None
+        on_tool_error: 异步回调 (server_config_id, tool_name, error_msg) -> None
+    """
+    original_coroutine = tool.coroutine
+    tool_name = tool.name
+
+    async def _safe_coroutine(*args, **kwargs):
+        try:
+            result = await original_coroutine(*args, **kwargs)
+            if on_tool_success:
+                try:
+                    await on_tool_success(server_config_id, original_tool_name)
+                except Exception:
+                    pass
+            return result
+        except Exception as e:
+            err_type = type(e).__name__
+            err_msg = str(e)[:500]
+            SYLogger.warning(f"[MCP] 工具 '{tool_name}' 调用失败 ({err_type}): {err_msg}")
+
+            if on_tool_error:
+                try:
+                    await on_tool_error(server_config_id, original_tool_name, f"{err_type}: {err_msg}")
+                except Exception:
+                    pass
+
+            friendly_msg = (
+                f"MCP 工具调用失败：工具 '{tool_name}' (服务器: {server_name}) 当前不可用。\n"
+                f"错误类型: {err_type}\n"
+                f"可能原因: MCP 服务器未启动、网络不可达或连接超时。\n"
+                f"请尝试不使用该工具继续完成任务，或联系管理员检查 MCP 服务 '{server_name}'。"
+            )
+            return [friendly_msg]
+
+    tool.coroutine = _safe_coroutine
+    return tool
+
+
+async def load_mcp_tools(
+    configs: List[MCPServerConfig],
+    on_tool_success: Optional[Callable[[str, str], Awaitable[None]]] = None,
+    on_tool_error: Optional[Callable[[str, str, str], Awaitable[None]]] = None,
+    on_batch_available: Optional[Callable[[str, list], Awaitable[None]]] = None,
+    on_server_failure: Optional[Callable[[str, str], Awaitable[None]]] = None,
+) -> List[BaseTool]:
+    """加载 MCP 工具列表
+
+    接受 MCPServerConfig 列表，逐个服务器连接并加载工具。
+
+    Args:
+        configs: MCP 服务器配置列表（仅 enabled 的会被处理）
+        on_tool_success: 单个工具调用成功回调
+        on_tool_error: 单个工具调用失败回调
+        on_batch_available: 服务器连接成功后批量标记工具可用回调 (config_id, tool_names)
+        on_server_failure: 服务器连接失败回调 (config_id, error_msg)
+    """
+    enabled_configs = [c for c in configs if c.enabled]
+    if not enabled_configs:
+        return []
+
+    all_tools = []
+    for config in enabled_configs:
+        try:
+            from langchain_mcp_adapters.client import MultiServerMCPClient
+
+            key = sanitize_name(config.name)
+            client_config = {
+                key: {
+                    "url": config.server_url,
+                    "transport": "streamable_http",
+                    **({"headers": config.headers} if config.headers else {}),
+                }
+            }
+
+            client = MultiServerMCPClient(client_config)
+            tools = await client.get_tools()
+
+            original_names = []
+            for tool in tools:
+                original_name = tool.name
+                original_names.append(original_name)
+                tool.name = f"mcp__{key}__{original_name}"
+                if tool.description and not tool.description.startswith("[MCP"):
+                    tool.description = f"[MCP:{config.name}] {tool.description}"
+                wrap_tool_with_error_handler(
+                    tool, config.id, config.name, original_name,
+                    on_tool_success=on_tool_success,
+                    on_tool_error=on_tool_error,
+                )
+
+            all_tools.extend(tools)
+
+            if on_batch_available:
+                try:
+                    await on_batch_available(config.id, original_names)
+                except Exception as e:
+                    SYLogger.warning(f"[MCP] 写入工具状态失败: {e}")
+
+            SYLogger.info(f"[MCP] 服务器 '{config.name}' 加载了 {len(tools)} 个工具")
+
+        except Exception as e:
+            SYLogger.warning(f"[MCP] 服务器 '{config.name}' ({config.server_url}) 连接失败，跳过: {e}")
+            if on_server_failure:
+                try:
+                    await on_server_failure(config.id, str(e)[:300])
+                except Exception:
+                    pass
+
+    if all_tools:
+        SYLogger.info(f"[MCP] 共加载 {len(all_tools)} 个 MCP 工具")
+    return all_tools
+
+
+async def test_mcp_connection(server_url: str, headers: dict = None) -> dict:
+    """测试 MCP 服务器连接，返回发现的工具列表"""
+    try:
+        from langchain_mcp_adapters.client import MultiServerMCPClient
+
+        client_config = {
+            "test": {
+                "url": server_url,
+                "transport": "streamable_http",
+                **({"headers": headers} if headers else {}),
+            }
+        }
+
+        client = MultiServerMCPClient(client_config)
+        tools = await client.get_tools()
+
+        return {
+            "success": True,
+            "tools": [{"name": t.name, "description": t.description or ""} for t in tools],
+        }
+    except Exception as e:
+        return {
+            "success": False,
+            "error": str(e),
+        }

{sycommon_python_lib-0.2.3a11 → sycommon_python_lib-0.2.4a0}/src/sycommon/agent/multi_agent_team.py

@@ -574,8 +574,12 @@ async def create_multi_agent_team(
 
     # 创建 checkpointer
     if checkpointer is None:
-        from langgraph.checkpoint.memory import MemorySaver
-        checkpointer = MemorySaver()
+        from sycommon.database.pg_checkpoint_service import PgCheckpointService
+        if PgCheckpointService.is_initialized():
+            checkpointer = await PgCheckpointService.get_checkpointer()
+        else:
+            from langgraph.checkpoint.memory import MemorySaver
+            checkpointer = MemorySaver()
 
     tid = thread_id or user_id
     agent_config = {"configurable": {"thread_id": tid}}

{sycommon_python_lib-0.2.3a11 → sycommon_python_lib-0.2.4a0}/src/sycommon/agent/summarization_utils.py

@@ -1,8 +1,9 @@
 """上下文压缩 middleware 构建工具。
 
 根据 nacos 中配置的模型 maxTokens，用绝对 token 数设置压缩阈值。
-中文场景下 count_tokens_approximately 默认 chars_per_token=4.0 会低估约 2.1 倍，
-因此传入自定义 token_counter 修正偏差，并按真实 token 数设定阈值。
+优先使用模型 API 返回的 usage_metadata.total_tokens（真实 token 数），
+无 metadata 时回退到 chars_per_token=2.0 的估算值。
+同时增加基于消息数的安全阈值，防止估算偏低导致压缩不触发。
 """
 
 from __future__ import annotations
@@ -24,17 +25,28 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
-# 中文场景修正：chars_per_token 从 4.0 改为 2.0，使近似计数接近真实 token 数
-_chinese_token_counter = functools.partial(
-    count_tokens_approximately, chars_per_token=2.0)
+
+def _extract_last_usage_total_tokens(messages) -> int:
+    """从消息历史中提取最后一条 AIMessage 的 usage_metadata.total_tokens。
+
+    返回 0 表示无数据（需回退到估算）。
+    """
+    from langchain_core.messages import AIMessage
+    for msg in reversed(messages):
+        if isinstance(msg, AIMessage):
+            meta = getattr(msg, 'usage_metadata', None)
+            if meta and isinstance(meta, dict):
+                total = meta.get('total_tokens', 0)
+                if isinstance(total, int) and total > 0:
+                    return total
+    return 0
 
 
 def _patched_compute_summarization_defaults(model):
     """覆盖 deepagents 默认值，返回中文场景修正后的绝对 token 阈值。
 
-    不走 fraction 路径（fraction 依赖 model.profile.max_input_tokens，
-    但内置 middleware 用默认 chars_per_token=4.0 计数，fraction 基于的基数不准）。
-    直接返回修正后的绝对 token 数。
+    同时增加基于消息数的安全阈值：即使 token 估算偏低，
+    消息数超过 200 条时也会触发压缩（覆盖工具 schema 等未计入的开销）。
     """
     try:
         from sycommon.config.Config import Config
@@ -48,10 +60,12 @@ def _patched_compute_summarization_defaults(model):
     except Exception:
         max_tokens = 72000
 
-    trigger = int(max_tokens * 0.85)
-    keep = int(max_tokens * 0.15)
+    # 60% 触发（120K/200K）：实测模型在 input≈137K 时开始退化，
+    # 在 120K 触发压缩留 ~17K 安全余量给工具 schema 等未计入开销
+    trigger = int(max_tokens * 0.60)
+    keep = int(max_tokens * 0.10)
     return {
-        "trigger": ("tokens", trigger),
+        "trigger": [("tokens", trigger), ("messages", 200)],
         "keep": ("tokens", keep),
         "truncate_args_settings": {
             "trigger": ("tokens", trigger),
@@ -63,10 +77,14 @@ def _patched_compute_summarization_defaults(model):
 # monkey-patch：替换 deepagents 的默认计算函数
 _summ_mod.compute_summarization_defaults = _patched_compute_summarization_defaults
 
-# monkey-patch：在内置 middleware 的 awrap_model_call 中注入日志
+# monkey-patch：在内置 middleware 的 awrap_model_call 中注入真实 token + 日志
 _OrigDeepAgentsSumm = _summ_mod._DeepAgentsSummarizationMiddleware
 _orig_awrap_model_call = _OrigDeepAgentsSumm.awrap_model_call
 
+# 基础估算函数，用于日志对比
+_approx_counter = functools.partial(
+    count_tokens_approximately, chars_per_token=2.0)
+
 
 async def _patched_awrap_model_call(self, request, handler):
     effective_messages = self._get_effective_messages(request)
@@ -75,15 +93,43 @@ async def _patched_awrap_model_call(self, request, handler):
     )
     counted_messages = [request.system_message, *
                         truncated_messages] if request.system_message is not None else truncated_messages
+
+    # 从截断前的 effective_messages 提取真实 token（截断会丢失 usage_metadata）
+    real_tokens = _extract_last_usage_total_tokens(effective_messages)
+
+    # 估算值（用于日志对比）
     try:
-        total_tokens = self.token_counter(
-            counted_messages, tools=request.tools)
+        estimated = _approx_counter(counted_messages, tools=request.tools)
     except TypeError:
-        total_tokens = self.token_counter(counted_messages)
-    should = self._should_summarize(truncated_messages, total_tokens)
+        estimated = _approx_counter(counted_messages)
+
+    # 如果有真实 token，临时替换 token_counter 使 _orig 内部判断也用真实值
+    # 这样 _should_summarize 和 _determine_cutoff_index 都能拿到正确的 token 数
+    if real_tokens > 0:
+        original_counter = self.token_counter
+
+        def _real_counter(msgs, **kwargs):
+            # 优先从当前消息中提取真实值（压缩后的消息可能有新的 metadata）
+            r = _extract_last_usage_total_tokens(msgs)
+            return r if r > 0 else real_tokens
+
+        self._lc_helper.token_counter = _real_counter
+        try:
+            result = await _orig_awrap_model_call(self, request, handler)
+        finally:
+            self._lc_helper.token_counter = original_counter
+    else:
+        result = await _orig_awrap_model_call(self, request, handler)
+
+    # 日志
+    source = 'real' if real_tokens > 0 else 'estimated'
+    should = self._should_summarize(truncated_messages, real_tokens if real_tokens > 0 else estimated)
     print(
-        f"[TokenCount] tokens={total_tokens} msgs={len(counted_messages)} should_summarize={should}")
-    return await _orig_awrap_model_call(self, request, handler)
+        f"[TokenCount] real={real_tokens} estimated={estimated} "
+        f"source={source} msgs={len(counted_messages)} "
+        f"should_summarize={should} "
+        f"trigger={getattr(self._lc_helper, 'trigger', '?')}")
+    return result
 
 
 _OrigDeepAgentsSumm.awrap_model_call = _patched_awrap_model_call
@@ -94,25 +140,21 @@ def build_summarization_middleware(
     model_name: str,
     backend: "BACKEND_TYPES",
     *,
-    trigger_fraction: float = 0.85,
-    keep_fraction: float = 0.15,
+    trigger_fraction: float = 0.60,
+    keep_fraction: float = 0.10,
     default_max_tokens: int = 200000,
 ) -> SummarizationToolMiddleware:
     """根据模型上下文窗口大小构建 compact_conversation 工具 middleware。
 
-    自动压缩由 deepagents 内置的 SummarizationMiddleware 处理，
-    它通过 model.profile.max_input_tokens（由 LLMWithTokenTracking 从 nacos 配置注入）
-    计算压缩阈值。
-
-    此处创建的 SummarizationToolMiddleware 仅提供 compact_conversation 手动压缩工具，
-    使用 nacos 配置的 maxTokens 作为其内部 SummarizationMiddleware 的阈值参数。
+    优先使用模型返回的 usage_metadata 真实 token 数进行压缩判断，
+    无 usage_metadata 时回退到 chars_per_token=2.0 估算。
 
     Args:
         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: 触发压缩占有效输入的比例，默认 85%。
-        keep_fraction: 压缩后保留占有效输入的比例，默认 15%。
+        trigger_fraction: 触发压缩占有效输入的比例，默认 60%。
+        keep_fraction: 压缩后保留占有效输入的比例，默认 10%。
         default_max_tokens: 无法从配置读取时的默认上下文窗口大小。
 
     Returns:
@@ -134,7 +176,8 @@ def build_summarization_middleware(
         backend=backend,
         trigger=("tokens", trigger_tokens),
         keep=("tokens", keep_tokens),
-        token_counter=_chinese_token_counter,
+        token_counter=functools.partial(
+            count_tokens_approximately, chars_per_token=2.0),
         trim_tokens_to_summarize=None,
         truncate_args_settings={
             "trigger": ("tokens", trigger_tokens),