sycommon-python-lib 0.2.10__tar.gz → 0.2.10a3__tar.gz

{sycommon_python_lib-0.2.10 → sycommon_python_lib-0.2.10a3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.2.10
+Version: 0.2.10a3
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown

{sycommon_python_lib-0.2.10 → sycommon_python_lib-0.2.10a3}/pyproject.toml

@@ -1,68 +1,68 @@
-[project]
-name = "sycommon-python-lib"
-version = "0.2.10"
-description = "Add your description here"
-readme = "README.md"
-requires-python = ">=3.11"
-dependencies = [
-    "aio-pika==10.0.1",
-    "aiohttp==3.14.3",
-    "aiomysql==0.3.2",
-    "anyio==4.14.2",
-    "decorator==5.3.1",
-    "deepagents==0.7.6",
-    "elasticsearch==9.5.0",
-    "fastapi==0.141.1",
-    "jinja2==3.1.6",
-    "kafka-python==3.0.10",
-    "langchain==1.3.15",
-    "langchain-core==1.5.5",
-    "langchain-openai==1.5.1",
-    "langfuse==4.14.4",
-    "langgraph==1.2.11",
-    "langgraph-checkpoint-postgres==3.1.2",
-    "langgraph-checkpoint-redis==0.5.1",
-    "ldap3==2.9.1",
-    "loguru==0.7.3",
-    "mysql-connector-python==26.7.0",
-    "nacos-sdk-python==2.0.11,<3.0",
-    "psutil==7.2.2",
-    "pyxxl==0.4.8",
-    "pydantic==2.13.4",
-    "python-dotenv==1.2.2",
-    "python-multipart==0.0.32",
-    "pyyaml==6.0.3",
-    "redis==7.4.0",
-    "sentry-sdk[fastapi]==2.68.0",
-    "sqlalchemy[asyncio]==2.0.52",
-    "starlette[full]==1.6.0",
-    "tiktoken==0.13.0",
-    "uvicorn==0.52.3",
-    "wecom-aibot-python-sdk==1.0.2",
-    "twine==7.0.0",
-    "minio==7.2.20",
-    "langchain-mcp-adapters==0.3.2",
-    "fastapi-mcp==0.4.0",
-    "psycopg[binary,pool]==3.3.4",
-    "agent-client-protocol==0.11.0",
-    "langchain-acp==1.7.0",
-    "pypinyin==0.55.0",
-]
-
-[tool.setuptools]
-packages = {find = {where = ["src"], exclude = ["*.__pycache__", "*.__pycache__.*"]}}
-package-data = {command = ["templates/**/*", "templates/**/.*"], sycommon = ["llm/tiktoken_cache/*"]}
-
-[build-system]
-requires = ["setuptools"]
-build-backend = "setuptools.build_meta"
-
-[dependency-groups]
-dev = [
-    "pytest>=9.0.3",
-]
-
-[project.scripts]
-sycommon = "command.cli:main"
-sycli = "sycli.cli:main"
-sypublish = "nexus:main"
+[project]
+name = "sycommon-python-lib"
+version = "0.2.10a3"
+description = "Add your description here"
+readme = "README.md"
+requires-python = ">=3.11"
+dependencies = [
+    "aio-pika==10.0.1",
+    "aiohttp==3.14.3",
+    "aiomysql==0.3.2",
+    "anyio==4.14.2",
+    "decorator==5.3.1",
+    "deepagents==0.7.6",
+    "elasticsearch==9.5.0",
+    "fastapi==0.141.1",
+    "jinja2==3.1.6",
+    "kafka-python==3.0.10",
+    "langchain==1.3.15",
+    "langchain-core==1.5.5",
+    "langchain-openai==1.5.1",
+    "langfuse==4.14.4",
+    "langgraph==1.2.11",
+    "langgraph-checkpoint-postgres==3.1.2",
+    "langgraph-checkpoint-redis==0.5.1",
+    "ldap3==2.9.1",
+    "loguru==0.7.3",
+    "mysql-connector-python==26.7.0",
+    "nacos-sdk-python==2.0.11,<3.0",
+    "psutil==7.2.2",
+    "pyxxl==0.4.8",
+    "pydantic==2.13.4",
+    "python-dotenv==1.2.2",
+    "python-multipart==0.0.32",
+    "pyyaml==6.0.3",
+    "redis==7.4.0",
+    "sentry-sdk[fastapi]==2.68.0",
+    "sqlalchemy[asyncio]==2.0.52",
+    "starlette[full]==1.6.0",
+    "tiktoken==0.13.0",
+    "uvicorn==0.52.3",
+    "wecom-aibot-python-sdk==1.0.2",
+    "twine==7.0.0",
+    "minio==7.2.20",
+    "langchain-mcp-adapters==0.3.2",
+    "fastapi-mcp==0.4.0",
+    "psycopg[binary,pool]==3.3.4",
+    "agent-client-protocol==0.11.0",
+    "langchain-acp==1.7.0",
+    "pypinyin==0.55.0",
+]
+
+[tool.setuptools]
+packages = {find = {where = ["src"], exclude = ["*.__pycache__", "*.__pycache__.*"]}}
+package-data = {command = ["templates/**/*", "templates/**/.*"], sycommon = ["llm/tiktoken_cache/*"]}
+
+[build-system]
+requires = ["setuptools"]
+build-backend = "setuptools.build_meta"
+
+[dependency-groups]
+dev = [
+    "pytest>=9.0.3",
+]
+
+[project.scripts]
+sycommon = "command.cli:main"
+sycli = "sycli.cli:main"
+sypublish = "nexus:main"

{sycommon_python_lib-0.2.10 → sycommon_python_lib-0.2.10a3}/src/sycommon/agent/deep_agent.py

@@ -65,6 +65,7 @@ from deepagents.middleware.summarization import create_summarization_tool_middle
 from sycommon.agent.summarization_utils import build_summarization_middleware
 from sycommon.agent.resume_compactor import maybe_compact_before_resume
 import sycommon.agent.reasoning_content_patch  # noqa: F401  提取 reasoning_content/reasoning 字段，避免思考型模型被判空响应
+import sycommon.agent.skills_retry_patch  # noqa: F401  技能列表加载失败不永久缓存空结果(沙箱 ls 500 时整会话丢技能)
 
 # 流式逐 token 诊断日志开关（默认关）。开启后每个 chunk 都打印
 # chunk_type/content 等（仅供排障）。LLM 最终结果已由 LLMLogger.on_llm_end
@@ -179,6 +180,10 @@ class AgentConfig(BaseModel):
     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
+    # 敏感检测器专用模型名（可选，与降级模型解耦）：
+    #   - 传了 → 检测分类器用它（适合配更快/更准的小模型），降级生成仍用 internal_model_name
+    #   - None → 沿用 internal_model_name（原行为，兼容存量调用方）
+    sensitive_detector_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
     # 命中敏感后的动作："allow"=不检测放行；"downgrade"=降级内部模型继续；
@@ -1245,6 +1250,7 @@ async def create_deep_agent(
         model_tier=config.model_tier,
         enabled=config.sensitive_guard,
         internal_model_name=config.internal_model_name,
+        detector_model_name=config.sensitive_detector_model_name,
         detect_prompt=config.sensitive_detect_prompt,
         action=config.sensitive_action,
     )

{sycommon_python_lib-0.2.10 → sycommon_python_lib-0.2.10a3}/src/sycommon/agent/middleware/sensitive_guard.py

@@ -264,6 +264,9 @@ class SensitiveGuardMiddleware(AgentMiddleware):
         self,
         model_tier: str,
         internal_model_name: Optional[str] = None,
+        # 检测器专用模型名：解耦「检测分类器」与「命中后的降级生成模型」。
+        # None 时沿用 internal_model_name（原行为，兼容所有存量调用方）。
+        detector_model_name: Optional[str] = None,
         on_downgrade: Optional[Callable[[], Awaitable[None]]] = None,
         enabled: bool = True,
         state: Optional[DowngradeState] = None,
@@ -286,6 +289,9 @@ class SensitiveGuardMiddleware(AgentMiddleware):
         self._model_tier = model_tier
         # None 时检测器/降级目标都用 get_llm() 读 nacos default 模型，不写死模型名
         self._internal_name = internal_model_name
+        # 检测器模型：detector_model_name 优先（应用层想用更快/更准的分类器时传），
+        # 未传(None)回退 internal_model_name——与原行为完全一致。
+        self._detector_name = detector_model_name or internal_model_name
         self._on_downgrade = on_downgrade
         self._on_block = on_block
         # 默认 allow（不检测）：应用层未显式指定时按"放行"处理，避免误伤。
@@ -396,7 +402,7 @@ class SensitiveGuardMiddleware(AgentMiddleware):
                     # 流式回调；且这里用独立 config（不取 agent 的 StreamMessagesHandler），
                     # 从源头断开检测器与用户事件流。
                     self._detector = get_llm(
-                        self._internal_name, temperature=0.0, wrap_structured=False)
+                        self._detector_name, temperature=0.0, wrap_structured=False)
         return self._detector
 
     async def _get_internal_raw(self):

{sycommon_python_lib-0.2.10 → sycommon_python_lib-0.2.10a3}/src/sycommon/agent/multi_agent_team.py

@@ -684,6 +684,7 @@ async def create_multi_agent_team(
             enabled=config.sensitive_guard,
             state=shared_downgrade_state,
             internal_model_name=config.internal_model_name,
+            detector_model_name=config.sensitive_detector_model_name,
             detect_prompt=config.sensitive_detect_prompt,
             action=config.sensitive_action,
             detect=detect,

sycommon_python_lib-0.2.10a3/src/sycommon/agent/skills_retry_patch.py

@@ -0,0 +1,93 @@
+# -*- coding: utf-8 -*-
+"""deepagents SkillsMiddleware 的重试补丁。
+
+修复「会话首轮沙箱 ls 失败 → skills_metadata=[] 被当已加载缓存进
+checkpoint → 整个会话从此看不到技能」的问题。
+
+背景(2026-09-01 线上案例):
+- deepagents SkillsMiddleware.abefore_agent 每会话只加载一次技能:
+  `if "skills_metadata" in state: return None` —— 空列表也算"已加载"。
+- 会话首条消息若撞上沙箱 /v1/sandbox/ls 500(发版恢复期/沙箱故障转移),
+  middleware 写入 skills_metadata=[] + skills_load_errors,并随 checkpoint
+  持久化。后续每一轮命中"in state"直接跳过,永不重试。
+- 用户视角:该会话从此没有技能,直到新建会话。
+
+修复语义:skills_metadata 为空**且**存在加载错误(或本应有 system 技能源)
+→ 视为未加载成功,清除旧值后重新加载。真正的零技能(所有源都正常列空)
+不重试,避免每轮白打一次 ls。
+
+打补丁方式:替换 deepagents.graph 模块里的 SkillsMiddleware 符号(graph
+在 create_deep_agent 内部 `SkillsMiddleware(backend=..., sources=...)`
+引用的是模块全局名),子类化原类、只重载 abefore_agent,其余行为不动。
+在 sycommon.agent.deep_agent 导入本模块即生效(幂等,可重复导入)。
+"""
+import logging
+
+from deepagents.middleware import skills as _skills_module
+
+logger = logging.getLogger("sycommon")
+
+_ORIG_MIDDLEWARE = _skills_module.SkillsMiddleware
+_PATCHED = getattr(_ORIG_MIDDLEWARE, "_sycommon_retry_patch", False)
+
+
+class RetrySkillsMiddleware(_ORIG_MIDDLEWARE):
+    """加载失败(空列表+错误)不缓存的 SkillsMiddleware。
+
+    上游缓存条件 `if "skills_metadata" in state: return None` 的漏洞:
+    空列表也是合法 dict key 存在,首轮 ls 失败写进去后永不再加载。
+    这里放宽为「空列表且带加载错误 → 重试」,其余行为与上游一致。
+    """
+
+    async def abefore_agent(self, state, runtime, config):
+        # 未加载过 → 走上游逻辑(会 fall through 到正常加载)
+        if "skills_metadata" not in state:
+            return await super().abefore_agent(state, runtime, config)
+
+        cached = state.get("skills_metadata") or []
+        cached_errors = state.get("skills_load_errors") or []
+
+        if cached:
+            # 已有技能列表(哪怕部分源失败) → 尊重上游缓存语义,不重试
+            return None
+
+        # 空列表:两种可能——
+        # a) 曾加载失败(ls 500 等)带 errors → 需要重试
+        # b) 所有源正常但确实零技能(无 errors) → 不重试
+        if not cached_errors:
+            return None
+
+        logger.warning(
+            "[SkillsRetry] 检测到缓存为空的技能列表(加载曾失败),"
+            "本会话轮次重新加载 skills: %s", self.sources)
+        # 清掉旧值让上游 super() 不再命中 "skills_metadata" in state
+        # (dict 可变,删 key 即可;state 是 langgraph 传入的 dict 视图)
+        try:
+            state.pop("skills_metadata", None)
+            state.pop("skills_load_errors", None)
+        except Exception:
+            # state 不可变视图等极端情况:构造副本调 super 不可行,
+            # 退化为直接调用内部加载逻辑(与上游 for-source 循环等价)
+            pass
+        return await super().abefore_agent(state, runtime, config)
+
+
+def _apply() -> None:
+    """把 deepagents.graph 引用的 SkillsMiddleware 换成重试版(幂等)。"""
+    if _PATCHED:
+        return
+    import deepagents.graph as _graph
+
+    if getattr(_graph.SkillsMiddleware, "_sycommon_retry_patch", False):
+        # 已打过(例如热重载场景)——只补标记
+        _graph.SkillsMiddleware = RetrySkillsMiddleware
+        return
+
+    RetrySkillsMiddleware._sycommon_retry_patch = True
+    _graph.SkillsMiddleware = RetrySkillsMiddleware
+    logger.info(
+        "[SkillsRetry] SkillsMiddleware 已启用空列表重试补丁"
+        "(加载失败不再永久缓存空技能列表)")
+
+
+_apply()

{sycommon_python_lib-0.2.10 → sycommon_python_lib-0.2.10a3}/src/sycommon/middleware/traceid.py

@@ -41,6 +41,21 @@ def _filtered_headers(raw_headers) -> list:
         return []
 
 
+def _omit_base64_content(obj: Any) -> None:
+    """把结构化响应体里超长的 content 值（PDF/图片 base64）替换为省略占位。
+
+    就地修改，dict/list 逐层递归。sanitize_deep_str 已把 JSON 字符串还原成
+    结构化 dict/list——不能再 json.loads（传容器进去会 TypeError）。
+    """
+    if isinstance(obj, dict):
+        c = obj.get("content")
+        if c and len(str(c)) > 1000:
+            obj["content"] = f"... (base64/binary content omitted, {len(str(c))} chars)"
+    elif isinstance(obj, list):
+        for item in obj:
+            _omit_base64_content(item)
+
+
 def setup_trace_id_handler(app):
     @app.middleware("http")
     async def trace_id_and_log_middleware(request: Request, call_next):
@@ -324,11 +339,11 @@ def setup_trace_id_handler(app):
             response_log_body = _log_masker.sanitize_deep_str(
                 response_body.decode('utf-8', errors='ignore'))
             # 检测是否包含 base64 二进制内容（如 PDF、图片等），如有则省略。
-            # sanitize_deep_str 已把 JSON 字符串还原成结构化 dict——直接在结构上
-            # 改(不要再 json.loads 一个 dict,会 TypeError);仅字符串回退路径走解析。
-            if isinstance(response_log_body, dict):
-                if response_log_body.get("content") and len(str(response_log_body["content"])) > 1000:
-                    response_log_body["content"] = f"... (base64/binary content omitted, {len(str(response_log_body['content']))} chars)"
+            # sanitize_deep_str 已把 JSON 字符串还原成结构化 dict/list——直接在
+            # 结构上改(不要再 json.loads 一个 dict/list,会 TypeError);仅字符串
+            # 回退路径(非 JSON 响应)走解析。
+            if isinstance(response_log_body, (dict, list)):
+                _omit_base64_content(response_log_body)
             else:
                 try:
                     resp_json = json.loads(response_log_body)

{sycommon_python_lib-0.2.10 → sycommon_python_lib-0.2.10a3}/src/sycommon_python_lib.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.2.10
+Version: 0.2.10a3
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown

{sycommon_python_lib-0.2.10 → sycommon_python_lib-0.2.10a3}/src/sycommon_python_lib.egg-info/SOURCES.txt

@@ -143,6 +143,7 @@ src/sycommon/agent/deep_agent.py
 src/sycommon/agent/multi_agent_team.py
 src/sycommon/agent/reasoning_content_patch.py
 src/sycommon/agent/resume_compactor.py
+src/sycommon/agent/skills_retry_patch.py
 src/sycommon/agent/summarization_utils.py
 src/sycommon/agent/acp/__init__.py
 src/sycommon/agent/acp/client.py
@@ -359,5 +360,6 @@ tests/test_sycli.py
 tests/test_sylogger_fixes.py
 tests/test_sync_skills_to_sandbox.py
 tests/test_time_utils.py
+tests/test_traceid_omit_base64.py
 tests/test_upload_concurrency.py
 tests/test_upload_e2e_http.py

sycommon_python_lib-0.2.10a3/tests/test_traceid_omit_base64.py

@@ -0,0 +1,77 @@
+# -*- coding: utf-8 -*-
+"""traceid 中间件 base64 省略检测回归测试(无需 Kafka/Nacos,纯同步可跑)。
+
+背景: /admin/schedules 等接口返回 JSON 数组([...] 顶层 list)。上一版修复
+(89f9941)只把 dict 分支改成结构上直改,else 分支仍对 sanitize_deep_str 的
+返回值 json.loads——当返回值是 list 时 json.loads(list) 直接 TypeError,
+中间件 500。本测试锁住三类响应形态:顶层 list / 顶层 dict / 非 JSON 字符串。
+"""
+import json
+import sys
+
+sys.path.insert(0, "src")
+
+from sycommon.middleware.traceid import _omit_base64_content  # noqa: E402
+
+
+# --------------------------------------------------------------------- helper
+def _run_omit_branch(value):
+    """复刻 traceid.py 里 sanitize_deep_str 之后的 base64 省略检测分支。"""
+    if isinstance(value, (dict, list)):
+        _omit_base64_content(value)
+    else:
+        parsed = json.loads(value)  # 老代码路径:容器进 json.loads 即 TypeError
+        if isinstance(parsed, dict):
+            _omit_base64_content(parsed)
+    return value
+
+
+# ------------------------------------------------------------------ _omit_base64_content
+def test_omit_base64_top_level_list():
+    """顶层 list(本次事故形态)不抛 TypeError,content 超长被替换。"""
+    obj = [{"id": 1, "content": "a" * 2000}, {"id": 2}]
+    _omit_base64_content(obj)
+    assert "omitted" in obj[0]["content"]
+    assert obj[1] == {"id": 2}
+
+
+def test_omit_base64_top_level_dict():
+    obj = {"content": "b" * 1500}
+    _omit_base64_content(obj)
+    assert "omitted" in obj["content"]
+
+
+def test_omit_base64_list_of_dicts_all_scanned():
+    obj = [{"content": "x" * 1200}, {"content": "y" * 1200}, {"id": 3}]
+    _omit_base64_content(obj)
+    assert "omitted" in obj[0]["content"]
+    assert "omitted" in obj[1]["content"]
+    assert obj[2] == {"id": 3}
+
+
+def test_omit_base64_short_content_untouched():
+    obj = [{"content": "ok"}, {"id": 2}]
+    _omit_base64_content(obj)
+    assert obj[0]["content"] == "ok"
+
+
+def test_omit_base64_empty_and_scalar_safe():
+    _omit_base64_content([])
+    _omit_base64_content(None)
+    _omit_base64_content("plain")
+
+
+# ----------------------------------------------------- 端到端: sanitize_deep_str + 省略检测
+def test_deep_str_list_response_end_to_end(monkeypatch):
+    """sanitize_deep_str(数组 JSON) 返回 list——省略检测不得 json.loads 它。"""
+    from sycommon.logging.log_masker import LogMasker
+
+    m = LogMasker()
+    m._config = {"enabled": True, "fields": [], "patterns": [],
+                 "drop": [], "maxLength": 100000, "maxSize": 500000,
+                 "maxListItems": 50, "headTailKeep": 200, "dropIfBig": False}
+
+    body = json.dumps([{"id": 1, "content": "z" * 2000}])
+    result = _run_omit_branch(m.sanitize_deep_str(body))
+    assert isinstance(result, list)
+    assert "omitted" in result[0]["content"]