| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188 |
- """
- Context 压缩 — 两级压缩策略
- Level 1: GoalTree 过滤(确定性,零成本)
- - 跳过 completed/abandoned goals 的消息(信息已在 GoalTree summary 中)
- - 始终保留:system prompt、第一条 user message、当前 focus goal 的消息
- Level 2: LLM 总结(仅在 Level 1 后仍超限时触发)
- - 在消息列表末尾追加压缩 prompt → 主模型回复 → summary 存为新消息
- - summary 的 parent_sequence 跳过被压缩的范围
- 压缩不修改存储:原始消息永远保留在 messages/,通过 parent_sequence 树结构实现跳过。
- """
- from dataclasses import dataclass
- from typing import List, Dict, Any, Optional, Set
- from .goal_models import GoalTree
- from .models import Message
- # ===== 配置 =====
- @dataclass
- class CompressionConfig:
- """压缩配置"""
- max_tokens: int = 100000 # 最大 token 数
- threshold_ratio: float = 0.8 # 触发 Level 2 的阈值比例(80%)
- keep_recent_messages: int = 10 # Level 1 中始终保留最近 N 条消息
- # ===== Level 1: GoalTree 过滤 =====
- def filter_by_goal_status(
- messages: List[Message],
- goal_tree: Optional[GoalTree],
- ) -> List[Message]:
- """
- Level 1 过滤:跳过 completed/abandoned goals 的消息
- 始终保留:
- - goal_id 为 None 的消息(system prompt、初始 user message)
- - 当前 focus goal 及其祖先链上的消息
- - in_progress 和 pending goals 的消息
- 跳过:
- - completed 且不在焦点路径上的 goals 的消息
- - abandoned goals 的消息
- Args:
- messages: 主路径上的有序消息列表
- goal_tree: GoalTree 实例
- Returns:
- 过滤后的消息列表
- """
- if not goal_tree or not goal_tree.goals:
- return messages
- # 构建焦点路径(当前焦点 + 父链 + 直接子节点)
- focus_path = _get_focus_path(goal_tree)
- # 构建需要跳过的 goal IDs
- skip_goal_ids: Set[str] = set()
- for goal in goal_tree.goals:
- if goal.id in focus_path:
- continue # 焦点路径上的 goal 始终保留
- if goal.status in ("completed", "abandoned"):
- skip_goal_ids.add(goal.id)
- # 过滤消息
- result = []
- for msg in messages:
- if msg.goal_id is None:
- result.append(msg) # 无 goal 的消息始终保留
- elif msg.goal_id not in skip_goal_ids:
- result.append(msg) # 不在跳过列表中的消息保留
- return result
- def _get_focus_path(goal_tree: GoalTree) -> Set[str]:
- """获取焦点路径上的所有 goal IDs(焦点 + 父链 + 直接子节点)"""
- focus_ids: Set[str] = set()
- if not goal_tree.current_id:
- return focus_ids
- # 焦点自身
- focus_ids.add(goal_tree.current_id)
- # 父链
- goal = goal_tree.find(goal_tree.current_id)
- while goal and goal.parent_id:
- focus_ids.add(goal.parent_id)
- goal = goal_tree.find(goal.parent_id)
- # 直接子节点
- children = goal_tree.get_children(goal_tree.current_id)
- for child in children:
- focus_ids.add(child.id)
- return focus_ids
- # ===== Token 估算 =====
- def estimate_tokens(messages: List[Dict[str, Any]]) -> int:
- """
- 估算消息列表的 token 数量
- 简单估算:字符数 / 4。实际使用时应该用 tiktoken 或 API 返回的 token 数。
- """
- total_chars = 0
- for msg in messages:
- content = msg.get("content", "")
- if isinstance(content, str):
- total_chars += len(content)
- elif isinstance(content, list):
- for part in content:
- if isinstance(part, dict) and part.get("type") == "text":
- total_chars += len(part.get("text", ""))
- # tool_calls
- tool_calls = msg.get("tool_calls")
- if tool_calls and isinstance(tool_calls, list):
- for tc in tool_calls:
- if isinstance(tc, dict):
- func = tc.get("function", {})
- total_chars += len(func.get("name", ""))
- args = func.get("arguments", "")
- if isinstance(args, str):
- total_chars += len(args)
- return total_chars // 4
- def estimate_tokens_from_messages(messages: List[Message]) -> int:
- """从 Message 对象列表估算 token 数"""
- return estimate_tokens([msg.to_llm_dict() for msg in messages])
- def needs_level2_compression(
- token_count: int,
- config: CompressionConfig,
- ) -> bool:
- """判断是否需要触发 Level 2 压缩"""
- return token_count > config.max_tokens * config.threshold_ratio
- # ===== Level 2: 压缩 Prompt =====
- COMPRESSION_PROMPT = """请对以上对话历史进行压缩总结。
- 要求:
- 1. 保留关键决策、结论和产出(如创建的文件、修改的代码、得出的分析结论)
- 2. 保留重要的上下文(如用户的要求、约束条件、之前的讨论结果)
- 3. 省略中间探索过程、重复的工具调用细节
- 4. 使用结构化格式(标题 + 要点)
- 5. 控制在 2000 字以内
- 当前 GoalTree 状态(完整版,含 summary):
- {goal_tree_prompt}
- """
- REFLECT_PROMPT = """请回顾以上整个执行过程,提取有价值的经验教训。
- 关注以下方面:
- 1. **人工干预**:如果有用户中途修改了指令或纠正了方向,说明之前的决策哪里有问题
- 2. **弯路**:哪些尝试是不必要的,有没有更直接的方法
- 3. **好的决策**:哪些判断和选择是正确的,值得记住
- 4. **工具使用**:哪些工具用法是高效的,哪些可以改进
- 请以简洁的规则列表形式输出,每条规则格式为:
- - 当遇到 [条件] 时,应该 [动作](原因:[简短说明])
- """
- def build_compression_prompt(goal_tree: Optional[GoalTree]) -> str:
- """构建 Level 2 压缩 prompt"""
- goal_prompt = ""
- if goal_tree:
- goal_prompt = goal_tree.to_prompt(include_summary=True)
- return COMPRESSION_PROMPT.format(goal_tree_prompt=goal_prompt)
- def build_reflect_prompt() -> str:
- """构建反思 prompt"""
- return REFLECT_PROMPT
|