sycommon-python-lib 0.2.5a28__tar.gz → 0.2.5a30__tar.gz

{sycommon_python_lib-0.2.5a28 → sycommon_python_lib-0.2.5a30}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sycommon-python-lib
-Version: 0.2.5a28
+Version: 0.2.5a30
 Summary: Add your description here
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown

{sycommon_python_lib-0.2.5a28 → sycommon_python_lib-0.2.5a30}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sycommon-python-lib"
-version = "0.2.5a28"
+version = "0.2.5a30"
 description = "Add your description here"
 readme = "README.md"
 requires-python = ">=3.11"

{sycommon_python_lib-0.2.5a28 → sycommon_python_lib-0.2.5a30}/src/sycommon/agent/deep_agent.py

@@ -153,6 +153,44 @@ def _patch_agent_streaming(agent):
         _handle_model_output = _hmo_cell.cell_contents
         _name = _name_cell.cell_contents if _name_cell is not None else None
 
+        def _clamp_max_tokens_for_model(model_, messages):
+            """在 raw_model 上执行 _clamp_max_tokens 逻辑。
+
+            由于 create_deep_agent 提取了 raw_model（绕过了 LLMWithTokenTracking 包装），
+            原本在 LLMWithTokenTracking._agenerate 中自动执行的 _clamp_max_tokens 被跳过，
+            导致 max_tokens 不随 input tokens 动态调整，当 input + max_output > context_window
+            时触发 400 错误。此函数补上这一逻辑。
+
+            注意：_get_bound_model 返回的可能是 _ChatModelBinding（RunnableBindingBase 子类），
+            需要通过 .bound 取到真正的 ChatOpenAI，否则 max_tokens 只设在 wrapper 上不生效。
+            """
+            from sycommon.config.LLMConfig import LLMConfig
+            from sycommon.config.Config import Config
+            from sycommon.llm.token_utils import clamp_max_tokens
+            try:
+                # 解包 binding wrapper，取到真正的 ChatModel
+                raw_model = getattr(model_, 'bound', model_)
+                llm_config = Config().get_llm_config(getattr(raw_model, 'model', None))
+                if not llm_config:
+                    return
+                llm_cfg = LLMConfig(**llm_config)
+                max_tokens_limit = getattr(llm_cfg, 'maxTokens', None)
+                if not max_tokens_limit:
+                    return
+                max_output = getattr(raw_model, 'max_tokens', None)
+                if not max_output:
+                    return
+                dynamic_max = clamp_max_tokens(
+                    raw_model, messages,
+                    context_window=max_tokens_limit,
+                    max_output_tokens=max_output,
+                )
+                SYLogger.debug(
+                    f"[StreamingPatch] clamp max_tokens: "
+                    f"context={max_tokens_limit}, max_output={max_output} → {dynamic_max}")
+            except Exception as e:
+                SYLogger.warning(f"[StreamingPatch] clamp max_tokens failed: {e}")
+
         async def _patched_execute_model_async(request):
             """替代 factory._execute_model_async: 用 astream + config 实现流式回调。"""
             from langchain.agents.middleware.types import ModelResponse
@@ -162,6 +200,9 @@ def _patch_agent_streaming(agent):
             if request.system_message:
                 messages = [request.system_message, *messages]
 
+            # 动态调整 max_tokens，防止 input + max_output 超过上下文窗口
+            _clamp_max_tokens_for_model(model_, messages)
+
             # 从 ContextVar 读取 LangGraph 注入的 config（含 StreamMessagesHandler）
             config = var_child_runnable_config.get()
 
@@ -278,7 +319,7 @@ class AgentConfig(BaseModel):
     debug: bool = False
 
     # Rubric 评分配置
-    rubric_max_iterations: int = 3  # rubric 评估最大迭代次数（0 = 不启用）
+    rubric_max_iterations: int = 0  # rubric 评估最大迭代次数（0 = 不启用，默认关闭）
 
     class Config:
         arbitrary_types_allowed = True