AI Agent 自进化技能树种子设计器

You are an expert AI agent architect specializing in self-evolving skill tree systems.

I want to design a self-evolving agent framework with these requirements:

1. Seed Design: Start from a minimal seed (under 5K lines) that contains only core capabilities: file I/O, shell execution, and a skill registry.

2. Skill Tree Growth: The agent should automatically:

   • Identify recurring task patterns
   • Extract reusable skills from successful task completions
   • Store skills as composable, versioned modules
   • Build dependency graphs between skills

3. Token Optimization: Design the context injection strategy to achieve at least 5x token reduction by:

   • Loading only relevant skills per task
   • Compressing skill descriptions into minimal signatures
   • Using a retrieval-based skill selector instead of stuffing all skills into context

4. Self-Improvement Loop:

   • After each task, evaluate: Was a new skill learned? Was an existing skill improved?
   • Periodically prune underused skills
   • Merge similar skills into generalized versions

Please provide:

• The seed architecture (modules, interfaces, data flow)
• Skill schema (name, trigger conditions, implementation, dependencies)
• Growth algorithm pseudocode
• Token budget analysis comparing naive vs. skill-tree approach

Output in structured markdown with code blocks where appropriate.