高效AI编码Agent上下文Token优化策略

You are a context engineering specialist for AI coding agents. Analyze the following coding agent workflow and design an optimized context management strategy that minimizes token usage while maintaining code comprehension quality.

Current workflow:

• Agent reads entire files before making edits
• All tool outputs are appended to context verbatim
• Sub-agent delegation uses the same model tier for all tasks

Design an optimization plan covering:

1. File Indexing Strategy

   • Use tree-sitter AST parsing to generate file skeletons (functions, classes, imports with line ranges)
   • Define when to use skeleton vs. full read
   • Expected token savings per file type

2. Tool Output Compression

   • Pipe tool outputs through a code execution sandbox for filtering/summarizing before they enter context
   • Template sandbox scripts for: grep result filtering, test output summarization, git diff condensation

3. Adaptive Model Routing

   • Classify sub-tasks into complexity tiers (simple/medium/complex)
   • Route to appropriate model tier (e.g., haiku/sonnet/opus)
   • Decision criteria and examples for each tier

4. Context Window Hygiene

   • Rules for evicting stale context
   • Summary checkpoints at conversation milestones
   • Maximum context budget allocation per phase (exploration/planning/implementation/review)

Provide concrete before/after token usage estimates for a typical 10-file refactoring task.