sycommon-python-lib 0.2.7a11__tar.gz → 0.2.7a13__tar.gz

{sycommon_python_lib-0.2.7a11 → sycommon_python_lib-0.2.7a13}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.2.7a11
+Version: 0.2.7a13
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown
@@ -11,14 +11,14 @@ Requires-Dist: anyio==4.14.1
 Requires-Dist: decorator==5.3.1
 Requires-Dist: deepagents==0.6.12
 Requires-Dist: elasticsearch==9.4.1
-Requires-Dist: fastapi==0.138.1
+Requires-Dist: fastapi==0.139.0
 Requires-Dist: jinja2==3.1.6
-Requires-Dist: kafka-python==3.0.6
+Requires-Dist: kafka-python==3.0.7
 Requires-Dist: langchain==1.3.11
 Requires-Dist: langchain-core==1.4.8
 Requires-Dist: langchain-openai==1.3.3
-Requires-Dist: langfuse==4.12.0
-Requires-Dist: langgraph==1.2.6
+Requires-Dist: langfuse==4.13.0
+Requires-Dist: langgraph==1.2.7
 Requires-Dist: langgraph-checkpoint-postgres==3.1.0
 Requires-Dist: langgraph-checkpoint-redis==0.5.0
 Requires-Dist: ldap3==2.9.1
@@ -32,7 +32,7 @@ Requires-Dist: python-dotenv==1.2.2
 Requires-Dist: python-multipart==0.0.32
 Requires-Dist: pyyaml==6.0.3
 Requires-Dist: redis==7.4.0
-Requires-Dist: sentry-sdk[fastapi]==2.63.0
+Requires-Dist: sentry-sdk[fastapi]==2.64.0
 Requires-Dist: sqlalchemy[asyncio]==2.0.51
 Requires-Dist: starlette[full]==1.3.1
 Requires-Dist: tiktoken==0.13.0

{sycommon_python_lib-0.2.7a11 → sycommon_python_lib-0.2.7a13}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sycommon-python-lib"
-version = "0.2.7a11"
+version = "0.2.7a13"
 description = "Add your description here"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -12,14 +12,14 @@ dependencies = [
     "decorator==5.3.1",
     "deepagents==0.6.12",
     "elasticsearch==9.4.1",
-    "fastapi==0.138.1",
+    "fastapi==0.139.0",
     "jinja2==3.1.6",
-    "kafka-python==3.0.6",
+    "kafka-python==3.0.7",
     "langchain==1.3.11",
     "langchain-core==1.4.8",
     "langchain-openai==1.3.3",
-    "langfuse==4.12.0",
-    "langgraph==1.2.6",
+    "langfuse==4.13.0",
+    "langgraph==1.2.7",
     "langgraph-checkpoint-postgres==3.1.0",
     "langgraph-checkpoint-redis==0.5.0",
     "ldap3==2.9.1",
@@ -33,7 +33,7 @@ dependencies = [
     "python-multipart==0.0.32",
     "pyyaml==6.0.3",
     "redis==7.4.0",
-    "sentry-sdk[fastapi]==2.63.0",
+    "sentry-sdk[fastapi]==2.64.0",
     "sqlalchemy[asyncio]==2.0.51",
     "starlette[full]==1.3.1",
     "tiktoken==0.13.0",

{sycommon_python_lib-0.2.7a11 → sycommon_python_lib-0.2.7a13}/src/sycommon/agent/deep_agent.py

@@ -58,6 +58,7 @@ from sycommon.agent.chat_events import ChatEvent, ChatEventBuilder, DEFAULT_AGEN
 from sycommon.middleware.token_tracking import TokenTrackingMiddleware
 from sycommon.middleware.tool_result_truncation import ToolResultTruncationMiddleware
 from sycommon.agent.middleware.skill_write_guard import SkillWriteGuardMiddleware
+from sycommon.agent.middleware.sensitive_guard import SensitiveGuardMiddleware
 from deepagents.middleware.summarization import create_summarization_tool_middleware  # noqa: F401 保留 re-export
 from sycommon.agent.summarization_utils import build_summarization_middleware
 
@@ -367,6 +368,12 @@ class AgentConfig(BaseModel):
     model_name: Optional[str] = None
     model_temperature: float = 0.7
     model_thinking: bool = False
+    model_tier: str = "external"        # "external"|"internal"；internal 时跳过敏感检测
+    sensitive_guard: bool = True        # 是否启用敏感信息检测守卫
+    # 敏感降级目标内部模型名（由应用层透传，避免基础库硬编码应用常量）
+    internal_model_name: Optional[str] = None
+    # 敏感检测提示词（应用层可覆盖）；未传(None)用基础库 DEFAULT_DETECT_PROMPT
+    sensitive_detect_prompt: Optional[str] = None
 
     # 系统提示词
     system_prompt: str = DEFAULT_SYSTEM_PROMPT
@@ -446,6 +453,7 @@ class DeepAgent:
         sandbox_backend: HTTPSandboxBackend,
         recovery_manager: Optional[SandboxRecoveryManager] = None,
         rubric_eval_collector: Optional[_RubricEvalCollector] = None,
+        sensitive_guard: Optional[SensitiveGuardMiddleware] = None,
     ):
         self.user_id = user_id
         self.agent = agent
@@ -453,6 +461,7 @@ class DeepAgent:
         self.sandbox_backend = sandbox_backend
         self.recovery_manager = recovery_manager
         self.rubric_eval_collector = rubric_eval_collector
+        self.sensitive_guard = sensitive_guard
 
     async def cleanup(self):
         """清理资源"""
@@ -508,7 +517,8 @@ class DeepAgent:
             ):
                 # 🔑 逐 token 诊断日志：仅在 DEEP_AGENT_STREAM_DEBUG=true 时打印
                 if _STREAM_DEBUG:
-                    _chunk_len = len(chunk) if isinstance(chunk, tuple) else '?'
+                    _chunk_len = len(chunk) if isinstance(
+                        chunk, tuple) else '?'
                     if isinstance(chunk, tuple) and _chunk_len == 2:
                         # 2-tuple: either (namespace, (msg, meta)) or (msg, meta)
                         if isinstance(chunk[1], tuple) and len(chunk[1]) == 2:
@@ -533,7 +543,8 @@ class DeepAgent:
                             _msg = payload_p[0]
                             _meta = payload_p[1]
                             _msg_type = type(_msg).__name__
-                            _content = repr(str(getattr(_msg, 'content', ''))[:80])
+                            _content = repr(
+                                str(getattr(_msg, 'content', ''))[:80])
                             _node = _meta.get('langgraph_node', '?') if isinstance(
                                 _meta, dict) else '?'
                             _step = _meta.get('langgraph_step', '?') if isinstance(
@@ -1221,7 +1232,20 @@ async def create_deep_agent(
     rubric_eval_collector = None
 
     # 构建 middleware 列表
+    # 🔒 敏感信息检测守卫：插在最外层（最先执行），命中则会话内降级内部模型。
+    #    enabled=False / model_tier="internal" 时内部短路，零开销。
+    #    internal_model_name 由应用层透传（避免基础库硬编码应用常量）；
+    #    未传(None)时 middleware 内部用 get_llm() 读 nacos default 模型兜底。
+    #    sensitive_detect_prompt 由应用层透传覆盖默认检测提示词。
+    guard = SensitiveGuardMiddleware(
+        model_tier=config.model_tier,
+        enabled=config.sensitive_guard,
+        internal_model_name=config.internal_model_name,
+        detect_prompt=config.sensitive_detect_prompt,
+    )
+
     middleware_list = [
+        guard,
         ToolResultTruncationMiddleware(),
         TokenTrackingMiddleware(model_name=config.model_name, user_id=user_id),
         summarization_mw,
@@ -1245,7 +1269,8 @@ async def create_deep_agent(
         try:
             from sycommon.agent.middleware.skill_api_whitelist import (
                 SkillApiWhitelistMiddleware)
-            middleware_list.append(SkillApiWhitelistMiddleware(sandbox_backend))
+            middleware_list.append(
+                SkillApiWhitelistMiddleware(sandbox_backend))
         except Exception as _wl_err:
             SYLogger.warning(f"[DeepAgent] 技能接口白名单中间件加载失败(忽略): {_wl_err}")
 
@@ -1312,6 +1337,7 @@ async def create_deep_agent(
         sandbox_backend=sandbox_backend,
         recovery_manager=recovery_manager,
         rubric_eval_collector=rubric_eval_collector,
+        sensitive_guard=guard,
     )
 
 
@@ -1346,46 +1372,67 @@ async def _sync_skills_to_sandbox(
     版本比较只看沙箱 sandbox_dest/{name}/SKILL.md,无需 MinIO 锚点:
     - 沙箱有该技能且 version 一致 → 跳过
     - 沙箱没有 / 版本不一致 → 从本地 skills/ 全量上传到沙箱
+
+    性能: 比对与上传均绕过 backend 的懒同步(_skip_lazy_sync=True),
+    避免「比对 read SKILL.md 前先整目录上传」「删除前先上传再删」的往返浪费。
+    多技能的 SKILL.md 比对用 gather 并发; batch_upload 服务端无条件覆盖,
+    无需先删(删除仅在版本不一致时做,用 adelete 走 /delete 而非 /execute)。
     """
     if not os.path.isdir(skills_dir):
         return
 
     try:
-        # 获取沙箱 sandbox_dest 下已有的目录列表
-        existing_dirs = set()
-        tree_result = await backend.atree(sandbox_dest, max_depth=1)
-        if tree_result and tree_result.tree and tree_result.tree.children:
-            for child in tree_result.tree.children:
-                if child.is_dir:
-                    existing_dirs.add(child.name)
-
-        skill_count = 0
-        skip_count = 0
+        # 收集本地技能名 + 本地版本(纯本地读,快)
+        local_skills = []  # [(skill_name, local_version, src_path)]
         for skill_name in sorted(os.listdir(skills_dir)):
             src = os.path.join(skills_dir, skill_name)
             if not os.path.isdir(src):
                 continue
-
-            # 读取本地版本
-            local_skill_md = os.path.join(src, "SKILL.md")
             local_version = ""
+            local_skill_md = os.path.join(src, "SKILL.md")
             if os.path.isfile(local_skill_md):
                 try:
                     with open(local_skill_md, "r", encoding="utf-8") as f:
                         local_version = _parse_skill_version(f.read())
                 except Exception:
                     pass
+            local_skills.append((skill_name, local_version, src))
 
-            # 比较版本:只从沙箱 sandbox_dest/{name}/SKILL.md 读取
-            remote_version = ""
-            if skill_name in existing_dirs:
-                try:
-                    remote_content = await backend.aread(
-                        f"{sandbox_dest}/{skill_name}/SKILL.md")
-                    if remote_content:
-                        remote_version = _parse_skill_version(remote_content)
-                except Exception:
-                    pass
+        # 获取沙箱 sandbox_dest 下已有的目录列表(一次 atree)
+        existing_dirs = set()
+        tree_result = await backend.atree(sandbox_dest, max_depth=1)
+        if tree_result and tree_result.tree and tree_result.tree.children:
+            for child in tree_result.tree.children:
+                if child.is_dir:
+                    existing_dirs.add(child.name)
+
+        # 并发读取所有已存在技能的沙箱 SKILL.md 比对版本(跳过懒同步)
+        async def _read_remote_version(skill_name: str) -> str:
+            if skill_name not in existing_dirs:
+                return ""
+            try:
+                remote_content = await backend.aread(
+                    f"{sandbox_dest}/{skill_name}/SKILL.md",
+                    _skip_lazy_sync=True)
+                if remote_content and remote_content.file_data:
+                    # FileData 是 TypedDict(dict), 用键访问而非属性
+                    return _parse_skill_version(
+                        remote_content.file_data.get("content", "") or "")
+            except Exception:
+                pass
+            return ""
+
+        remote_versions = await asyncio.gather(
+            *[_read_remote_version(name) for name, _, _ in local_skills]
+        )
+        remote_version_map = {
+            name: rv for (name, _, _), rv in zip(local_skills, remote_versions)
+        }
+
+        skill_count = 0
+        skip_count = 0
+        for skill_name, local_version, src in local_skills:
+            remote_version = remote_version_map.get(skill_name, "")
 
             if remote_version and remote_version == local_version:
                 SYLogger.debug(
@@ -1394,10 +1441,13 @@ async def _sync_skills_to_sandbox(
                 skip_count += 1
                 continue
 
-            # 版本不一致或无版本,删除旧目录后全量覆盖
+            # 版本不一致或无版本,删除旧目录(走 /delete,跳过懒同步,毫秒级)
             if skill_name in existing_dirs:
                 try:
-                    await backend.aexecute(f"rm -rf {sandbox_dest}/{skill_name}")
+                    await backend.adelete(
+                        f"{sandbox_dest}/{skill_name}",
+                        recursive=True,
+                        _skip_lazy_sync=True)
                     SYLogger.info(
                         f"[DeepAgent] 技能版本不一致,覆盖: {skill_name} "
                         f"({remote_version or '无版本'} → {local_version})")
@@ -1405,7 +1455,7 @@ async def _sync_skills_to_sandbox(
                     SYLogger.warning(
                         f"[DeepAgent] 删除旧版技能失败: {skill_name}, {e}")
 
-            # 全量上传整个技能目录到沙箱
+            # 全量上传整个技能目录到沙箱(batch_upload 服务端无条件覆盖)
             await backend.async_sync_dirs([(src, f"{sandbox_dest}/{skill_name}")])
             backend._synced_skills.add(skill_name)
             skill_count += 1
@@ -1417,7 +1467,8 @@ async def _sync_skills_to_sandbox(
                 # sandbox_dest 是虚拟绝对路径(/skills/system),沙箱 shell 不重映射裸绝对路径,
                 # 须拼 $_SANDBOX_WORKSPACE 指向真实 workspace 才能命中。否则 chmod 静默失败。
                 await backend.aexecute(
-                    f'chmod -R 555 "$_SANDBOX_WORKSPACE{sandbox_dest}"')
+                    f'chmod -R 555 "$_SANDBOX_WORKSPACE{sandbox_dest}"',
+                    _skip_lazy_sync=True)
                 SYLogger.info(f"[DeepAgent] 已设只读: chmod -R 555 {sandbox_dest}")
             except Exception as e:
                 SYLogger.warning(f"[DeepAgent] chmod 只读失败(忽略): {e}")