高效 AI 编码 Agent 上下文优化策略师

You are an expert in AI coding agent optimization, specializing in context window efficiency and token-aware architectures.

I want to optimize my AI coding agent to use minimal context tokens while maintaining high code quality.

Current setup:

• Model: [e.g., Claude Sonnet / GPT-4o / Qwen-Coder]
• Context window: [e.g., 200K tokens]
• Average task token usage: [e.g., 50K tokens per task]
• Main bottleneck: [e.g., reading too many files, verbose bash output]

Please design a complete optimization strategy covering:

1. Code Indexing Layer: Tree-sitter based file skeleton generation with line ranges. When to index vs full-read based on file size thresholds.

2. Smart Read Strategy: Line-range reads instead of full file reads. Relevance scoring for which files to read.

3. Code Execution Sandbox: Using an interpreter with tool access for data filtering/transformation. Piping outputs through summarization before returning to context.

4. Sub-agent Delegation: When to use weak/medium/strong models for subtasks. Task complexity classification heuristics. Cost-performance tradeoff matrix.

5. Output Compression: Bash output filtering and truncation rules. Build log summarization. Test result condensation.

6. Metrics: Token usage per tool category tracking. Before/after comparison framework.

Provide concrete implementation examples and expected token savings for each strategy.