valkryhx/google_adk_agent / skills/compactor
智能上下文压缩器
Verified采用 Rolling Summary 策略压缩对话历史,在清空历史前保留关键信息(任务目标、阶段性结论、任务进展)。
tools.py
View on GitHub"""
Compactor Skill - 智能上下文压缩器
采用 Rolling Summary 策略 + 原生状态查询:
- 对话轮数:从 session.events 获取
- 工具数量:从 agent.tools 获取
- 不依赖外部追踪器
"""
from typing import List, Optional
async def _smart_compact(
summary: str,
agent_instance: object,
session_service: object,
app_name: str,
user_id: str,
session_id: str,
unload_tools: bool = True
) -> str:
"""
智能压缩上下文:清空历史并保留摘要
Args:
summary: 对当前对话的总结(必须包含任务目标、进展、结论等)
agent_instance: Agent 实例
session_service: 会话服务实例
app_name: 应用名称
user_id: 用户 ID
session_id: 会话 ID
unload_tools: 是否卸载临时工具
Returns:
执行结果消息
"""
results = []
# 1. 记录摘要
print("[INFO] 已接收上下文摘要")
results.append("[INFO] 已接收上下文摘要")
# 2. 获取当前状态(使用原生 API)
message_count = 0
session = None
print(f"[DEBUG] 尝试获取 Session: app={app_name}, user={user_id}, id={session_id}")
try:
if hasattr(session_service, 'get_session'):
session = await session_service.get_session(
app_name=app_name,
user_id=user_id,
session_id=session_id
)
if session is None:
print(f"[WARN] get_session 返回 None")
if hasattr(session_service, 'sessions'):
print(f"[DEBUG] 当前所有 Session: {session_service.sessions.keys()}")
if app_name in session_service.sessions:
print(f"[DEBUG] App {app_name} 下的用户: {session_service.sessions[app_name].keys()}")
if session and hasattr(session, 'events') and session.events:
message_count = len(session.events)
msg = f"[INFO] 压缩前对话包含 {message_count} 条消息"
print(msg)
results.append(msg)
except Exception as e:
msg = f"[WARN] 获取会话历史时出错: {e}"
print(msg)
results.append(msg)
# 3. 执行压缩(原地截断)
try:
if session and hasattr(session, 'events'):
events = session.events
if not events:
msg = "[INFO] 会话为空,无需压缩"
print(msg)
results.append(msg)
else:
# 寻找保留点:保留 System 消息和最近一轮交互
msg = f"[DEBUG] 截断前事件总数: {len(events)}"
print(msg)
results.append(msg)
# 1. 收集 System 消息
kept_events = []
for evt in events:
# [Fix] Event.role is nested in content
role = 'unknown'
if hasattr(evt, 'content') and evt.content and hasattr(evt.content, 'role'):
role = evt.content.role
if role == 'system':
kept_events.append(evt)
else:
break
# 2. 寻找当前的 smart_compact 调用(Model 消息)
# 我们必须保留它,以便 Runner 能将 Output 追加到正确位置
current_tool_call_idx = -1
print(f"[DEBUG] 开始倒序查找 smart_compact 调用, 总事件数: {len(events)}")
for i in range(len(events) - 1, -1, -1):
evt = events[i]
# [Fix] Event 对象没有直接的 role 属性,它在 content 中
role = 'unknown'
if hasattr(evt, 'content') and evt.content:
if hasattr(evt.content, 'role'):
role = evt.content.role
print(f"[DEBUG] Checking Event[{i}]: Type={type(evt).__name__}, Role={repr(role)}")
if role == 'model':
# 检查是否调用了 smart_compact
is_target_call = False
if hasattr(evt, 'content') and hasattr(evt.content, 'parts'):
for part_idx, part in enumerate(evt.content.parts):
if hasattr(part, 'function_call') and part.function_call:
fc_name = part.function_call.name
print(f"[DEBUG] Event[{i}] Role={role} Part[{part_idx}] FunctionCall: '{fc_name}'")
if 'smart_compact' in fc_name:
print(f"[DEBUG] Found match at Event[{i}]")
is_target_call = True
break
else:
print(f"[DEBUG] Event[{i}] Role={role} Part[{part_idx}] No FunctionCall")
pass
if is_target_call:
current_tool_call_idx = i
break
else:
print(f"[DEBUG] Skipping Event[{i}] Role={role} (Not 'model')")
pass
# 3. 构建精简历史 (Hard Reset)
# 目标结构: System + User(Placeholder) + Model(Call smart_compact)
if current_tool_call_idx != -1:
import copy
# 构造占位符 User 消息
# 我们克隆第一个 System 消息(通常存在)作为模板,以确保对象类型正确
placeholder_user_evt = None
# [Fix] 如果没有 System 消息,尝试使用第一个事件作为模板
template_source = kept_events[0] if kept_events else (events[0] if events else None)
if template_source:
try:
placeholder_user_evt = copy.deepcopy(template_source)
# [Fix] 设置 role 到 content.role
if hasattr(placeholder_user_evt, 'content') and placeholder_user_evt.content:
placeholder_user_evt.content.role = 'user'
# 尝试修改内容为简单的文本
if hasattr(placeholder_user_evt, 'content') and hasattr(placeholder_user_evt.content, 'parts'):
# 清空原有 parts
placeholder_user_evt.content.parts = []
# 我们无法直接创建 Part 对象(未导入),所以尝试复用原有的 Part 对象(如果有)
# 或者更简单:如果无法创建 Part,就让它保持空?不,空的 User 消息可能非法。
# 让我们尝试复用 events[current_tool_call_idx] 中的 Part 结构
template_evt = events[current_tool_call_idx]
if template_evt.content.parts:
# 克隆一个 Part
new_part = copy.deepcopy(template_evt.content.parts[0])
# 清除 function_call/response
if hasattr(new_part, 'function_call'): new_part.function_call = None
if hasattr(new_part, 'function_response'): new_part.function_response = None
# 设置文本
if hasattr(new_part, 'text'):
new_part.text = "[System] Context cleared. Resuming with summary."
placeholder_user_evt.content.parts = [new_part]
except Exception as e:
print(f"[WARN] 构造占位消息失败: {e}")
placeholder_user_evt = None
if placeholder_user_evt:
new_events = kept_events + [placeholder_user_evt] + events[current_tool_call_idx:]
# [Fix] In-place update to ensure persistence if Runner holds reference
if hasattr(session.events, 'clear') and hasattr(session.events, 'extend'):
session.events.clear()
session.events.extend(new_events)
else:
# Fallback for lists that don't support clear/extend (unlikely for standard lists)
session.events[:] = new_events
# [Critical Fix] InMemorySessionService returns a deepcopy, so we MUST update the internal storage
try:
from google.adk.sessions import InMemorySessionService
if isinstance(session_service, InMemorySessionService):
if (app_name in session_service.sessions and
user_id in session_service.sessions[app_name] and
session_id in session_service.sessions[app_name][user_id]):
stored_session = session_service.sessions[app_name][user_id][session_id]
# Update events
if hasattr(stored_session.events, 'clear') and hasattr(stored_session.events, 'extend'):
stored_session.events.clear()
stored_session.events.extend(new_events)
else:
stored_session.events[:] = new_events
print("[系统] 已强制同步会话状态到存储 (smart_compact)")
except Exception as e:
print(f"[警告] 强制同步会话失败: {e}")
removed_count = message_count - len(new_events)
msg1 = f"[OK] 已重置历史 (Hard Reset),移除 {removed_count} 条消息"
msg2 = f"[DEBUG] 最终结构: System({len(kept_events)}) -> User(Placeholder) -> Model(CurrentCall)"
print(msg1)
print(msg2)
print(f'压缩后events数量 (Local): {len(session.events)}')
# Verify storage update
if 'stored_session' in locals() and stored_session:
print(f'压缩后events数量 (Storage): {len(stored_session.events)}')
results.append(msg1)
results.append(msg2)
else:
msg = "[WARN] 无法构造占位 User 消息,放弃截断"
print(msg)
results.append(msg)
else:
msg = "[WARN] 未找到当前 Tool Call,无法安全截断"
print(msg)
results.append(msg)
# DEBUG: 打印最后几个事件的结构以供分析
if events:
print("[DEBUG] 事件结构采样 (Last 10):")
for j in range(max(0, len(events)-10), len(events)):
e = events[j]
print(f" Event[{j}]: Type={type(e)}")
if hasattr(e, '__dict__'):
print(f" Vars: {vars(e)}")
else:
print(f" Str: {str(e)}")
if hasattr(e, 'role'):
print(f" Role: {e.role}")
else:
msg = "[WARN] 无法访问会话事件列表"
print(msg)
results.append(msg)
except Exception as e:
import traceback
msg1 = f"[WARN] 压缩会话时出错: {e}"
msg2 = f"[DEBUG] {traceback.format_exc()}"
print(msg1)
print(msg2)
results.append(msg1)
results.append(msg2)
# 4. 卸载动态工具
if unload_tools:
try:
import sys
# 尝试获取 main 模块中的 agent 实例
target_agent = None
# 1. 尝试从传入参数获取
if agent_instance and hasattr(agent_instance, 'tools'):
target_agent = agent_instance
# print("[DEBUG] 使用传入的 agent_instance")
results.append("[DEBUG] 使用传入的 agent_instance")
# 2. 如果参数无效,尝试从 sys.modules 获取
if not target_agent:
main_module = sys.modules.get('__main__')
if main_module and hasattr(main_module, 'my_agent'):
target_agent = main_module.my_agent
# print("[DEBUG] 从 __main__ 模块获取 my_agent")
results.append("[DEBUG] 从 __main__ 模块获取 my_agent")
if not target_agent:
# 尝试完整包名
adk_main = sys.modules.get('skills.adk_agent.main')
if adk_main and hasattr(adk_main, 'my_agent'):
target_agent = adk_main.my_agent
# print("[DEBUG] 从 skills.adk_agent.main 模块获取 my_agent")
results.append("[DEBUG] 从 skills.adk_agent.main 模块获取 my_agent")
if target_agent and hasattr(target_agent, 'tools'):
original_count = len(target_agent.tools)
msg = f"[DEBUG] 当前工具数量: {original_count}"
print(msg)
results.append(msg)
# 定义需要保留的核心工具名称
# 1. skill_load: 动态加载能力
# 2. bash: 系统操作能力
# 3. file_editor: 文件编辑能力
# 4. search_experience: 本地经验库检索 (新增)
# 5. view_local_image: UI 图片查看工具
# 6. analyze_local_image: 多模态图片分析工具
# 7. log_learning: 学习记录工具 (Self-Improvement 核心)
# 8. log_error: 错误记录工具 (Self-Improvement 核心)
CORE_TOOL_NAMES = {
'skill_load', 'skill_reload', 'bash', 'file_editor',
'search_experience', 'view_local_image', 'analyze_local_image',
'log_learning', 'log_error'
}
# 筛选出核心工具
kept_tools = []
for tool in target_agent.tools:
t_name = getattr(tool, '__name__', str(tool))
if t_name in CORE_TOOL_NAMES:
kept_tools.append(tool)
# 更新工具列表
if len(kept_tools) < original_count:
target_agent.tools = kept_tools
new_count = len(target_agent.tools)
msg = f"[OK] 已卸载 {original_count - new_count} 个临时工具, 剩余 {new_count} 个核心工具 ({[t.__name__ for t in kept_tools]})"
print(msg)
results.append(msg)
else:
msg = "[INFO] 没有需要卸载的临时工具"
print(msg)
results.append(msg)
else:
msg = "[WARN] 无法获取有效的 Agent 实例,工具卸载失败"
print(msg)
results.append(msg)
except Exception as e:
import traceback
msg1 = f"[WARN] 卸载工具时出错: {e}"
msg2 = f"[DEBUG] 堆栈: {traceback.format_exc()}"
print(msg1)
print(msg2)
results.append(msg1)
results.append(msg2)
results.append("\n[状态] Agent 已重置为轻量化状态")
results.append("--------------------------------------------------")
results.append("请将以下摘要作为新对话的上下文:")
results.append(summary)
results.append("--------------------------------------------------")
return "\n".join(results)
async def _get_compression_status(
agent_instance: object,
session_service: object,
app_name: str,
user_id: str,
session_id: str
) -> str:
"""
获取当前压缩状态和建议(使用原生 session 和 agent 状态)
Args:
agent_instance: Agent 实例
session_service: 会话服务实例
app_name: 应用名称
user_id: 用户 ID
session_id: 会话 ID
Returns:
状态报告
"""
result = ["[压缩状态分析]"]
# 1. 从 session 获取对话轮数
turn_count = 0
try:
if hasattr(session_service, 'get_session'):
session = await session_service.get_session(
app_name=app_name,
user_id=user_id,
session_id=session_id
)
if session and hasattr(session, 'events') and session.events:
turn_count = len(session.events)
except Exception:
pass
result.append(f" 对话消息数: {turn_count}")
# 2. 从 agent 获取工具数量
tool_count = 0
if hasattr(agent_instance, 'tools'):
tool_count = len(agent_instance.tools)
result.append(f" 已加载工具数: {tool_count}")
# 3. 判断是否需要压缩
result.append("\n[建议]")
should_compact = turn_count > 100 or tool_count > 50
if should_compact:
result.append(" [!] 建议立即执行压缩")
if turn_count > 100:
result.append(f" - 对话消息数 ({turn_count}) 超过阈值 (100)")
if tool_count > 50:
result.append(f" - 工具数量 ({tool_count}) 超过阈值 (50)")
elif tool_count > 25 or turn_count > 50:
result.append(" - 状态正常,可考虑稍后压缩")
else:
result.append(" - 状态良好,暂无需压缩")
return "\n".join(result)
# 工具函数字典
COMPACTOR_TOOLS = {
"smart_compact": _smart_compact,
"get_compression_status": _get_compression_status,
}
from functools import partial
def get_tools(agent_instance: object = None, session_service: object = None, app_info: dict = None) -> List:
"""
返回所有压缩器工具函数列表(自动绑定上下文)
Args:
agent_instance: Agent 实例
session_service: 会话服务实例
app_info: 应用信息字典 (app_name, user_id, session_id)
"""
tools = []
# 需要上下文的工具
if agent_instance and session_service and app_info:
# 1. 包装 smart_compact,隐藏系统参数
async def run_smart_compact(summary: str) -> str:
"""
智能压缩上下文:清空历史并保留摘要。
当对话过长或任务阶段性完成时调用此工具。
Args:
summary: 对当前对话的总结(必须包含任务目标、进展、结论等)
"""
return await _smart_compact(
summary=summary,
agent_instance=agent_instance,
session_service=session_service,
app_name=app_info.get('app_name'),
user_id=app_info.get('user_id'),
session_id=app_info.get('session_id')
)
# 显式设置名称,确保 deduplication 正常工作
run_smart_compact.__name__ = "smart_compact"
tools.append(run_smart_compact)
# 2. 包装 get_compression_status
async def run_get_compression_status() -> str:
"""
获取当前压缩状态建议
"""
return await _get_compression_status(
agent_instance=agent_instance,
session_service=session_service,
app_name=app_info.get('app_name'),
user_id=app_info.get('user_id'),
session_id=app_info.get('session_id')
)
run_get_compression_status.__name__ = "get_compression_status"
tools.append(run_get_compression_status)
return tools
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |