Agent Skills
› aiming-lab/MetaClaw
› context-window-management
context-window-management
GitHub用于长对话或多轮智能体会话中，通过主动摘要、优先级排序和文件化记忆来管理上下文窗口，防止信息丢失或超出Token限制。
Trigger Scenarios
多轮对话接近Token上限
需要整理历史决策状态
避免重复读取已处理的大文件
Install
npx skills add aiming-lab/MetaClaw --skill context-window-management -g -y
SKILL.md
Frontmatter
{
"name": "context-window-management",
"category": "agentic",
"description": "Use this skill in long conversations or multi-turn agentic sessions where context may be lost or the conversation is approaching token limits. Summarize, prioritize, and compact context proactively before it becomes a problem."
}
Context Window Management
Proactive summarization:
- At natural breakpoints, summarize key decisions and state into a compact form.
- Keep the active context focused on the current task; move resolved context to a summary.
Prioritization:
- Most important: current task description, constraints, and already-confirmed decisions.
- Less important: verbose tool outputs that have been processed; intermediate reasoning.
File-based memory: For long-running sessions, write important state to a file rather than keeping it all in the conversation.
Avoid: Repeatedly re-reading large files you already have in context; re-running expensive searches for information already retrieved.
Version History
- 922caf3 Current 2026-07-25 11:08
