♻️ refactor(agent): 重构 Agent 状态管理与执行器流程，优化 Token 预估与自愈反思机制 (#2150)

- 统一使用 `run_context.run.messages` 作为消息历史的单一数据源，清理 `AgentState` 冗余字段
- 将工具消息装配逻辑 `assemble_tool_message` 提取并重构至 `ToolExecutor`
- 引入 `token_drift` 动态校准偏移量，并精确计算工具与系统提示词的 Token 开销
- 重构 `ReflexionCapability` 自愈反思引擎，基于异常多态与模板字典动态生成反馈提示词
- 支持通过 `resolve_model_capabilities` 解析并合并用户自定义的模型能力覆盖
- 在执行器循环中支持 `should_reset_cycle`，以优雅处理外部干预（如用户追加指示）
- 扩展 `capabilities` 中对 `gpt-[5-9]*` 等新型号模型的能力定义与上下文限制

Co-authored-by: webjoin111 <455457521@qq.com>

zhenxun/services/ai/context/memory/compression.py

@@ -4,6 +4,7 @@ from typing import Any, Generic, TypeVar
 
 from pydantic import BaseModel, Field
 
+from zhenxun.services.ai.config import get_llm_config
 from zhenxun.services.ai.context.memory.models import MemoryConfig
 from zhenxun.services.ai.core.engine.token_counter import token_counter
 from zhenxun.services.ai.core.messages import (
@@ -15,7 +16,10 @@ from zhenxun.services.ai.core.messages import (
     TextPart,
     VideoPart,
 )
+from zhenxun.services.ai.core.models import ModelCapabilities
+from zhenxun.services.ai.llm.api import chat, generate_structured
 from zhenxun.services.ai.llm.manager import get_default_model
+from zhenxun.services.ai.llm.system.capabilities import get_model_capabilities
 from zhenxun.services.ai.utils.logger import log_memory as logger
 from zhenxun.utils.pydantic_compat import model_copy
 
@@ -396,8 +400,6 @@ class LLMSummarizerReducer(AbstractSummarizerReducer):
             prompt_text += f"[{speaker}]: {c_str}\n"
         prompt_text += "</需要合并的旧对话记录>\n"
 
-        from zhenxun.services.ai.llm.api import chat
-
         try:
             model_to_use = self.summarization_model or get_default_model("chat")
             response = await chat(
@@ -474,8 +476,6 @@ class StructuredSummaryReducer(AbstractSummarizerReducer, Generic[_T_Summary]):
             prev_summary=prev_summary, dialogue=dialogue_text
         )
 
-        from zhenxun.services.ai.llm.api import generate_structured
-
         try:
             model_to_use = self.summarization_model or get_default_model("chat")
             summary_obj = await generate_structured(
@@ -513,15 +513,19 @@ class CondenserPipeline:
 
     @classmethod
     def create_from_configs(
-        cls, memory_config: MemoryConfig | None, model_name: str
+        cls,
+        memory_config: MemoryConfig | None,
+        capabilities: ModelCapabilities | None,
+        model_name: str,
     ) -> "CondenserPipeline":
         """基于全局和局部配置组装压缩管线工厂方法"""
-        from zhenxun.services.ai.config import get_llm_config
-        from zhenxun.services.ai.llm.system.capabilities import get_model_capabilities
-
         config = get_llm_config().context_settings
         pipeline_reducers = []
-        caps = get_model_capabilities(model_name)
+        caps = (
+            capabilities
+            if capabilities is not None
+            else get_model_capabilities(model_name)
+        )
 
         vw = config.vision_window_size
         if memory_config and memory_config.compression.vision_window is not None:

zhenxun/services/ai/context/memory/engine.py

@@ -3,6 +3,7 @@ from typing import cast
 
 from zhenxun.services.ai.core.engine.context_renderer import ContextConverter
 from zhenxun.services.ai.core.messages import AgentMessage
+from zhenxun.services.ai.core.models import ModelCapabilities
 from zhenxun.services.ai.run.context import RunContext
 from zhenxun.services.ai.utils.logger import log_memory as logger
 from zhenxun.utils.pydantic_compat import model_copy
@@ -42,7 +43,9 @@ class SessionMemoryContext:
     async def read(
         self,
         model_name: str,
+        capabilities: ModelCapabilities | None = None,
         override_history: Sequence[AgentMessage] | None = None,
+        base_overhead: int = 0,
     ) -> list[AgentMessage]:
         """
         拉取短期对话历史，并执行 Token 压缩与管线修剪。
@@ -69,14 +72,16 @@ class SessionMemoryContext:
                 )
 
             pipeline = CondenserPipeline.create_from_configs(
-                self.memory_config, model_name
+                self.memory_config, capabilities, model_name
             )
             if pipeline.reducers:
                 flattened_to_reduce = ContextConverter.flatten_to_llm_messages(
                     current_history
                 )
                 new_history, changed = await pipeline.run(
-                    flattened_to_reduce, model_name=model_name, base_overhead=0
+                    flattened_to_reduce,
+                    model_name=model_name,
+                    base_overhead=base_overhead,
                 )
                 if changed:
                     await chat_context.set_messages(self.session_meta, new_history)