📄 中文摘要
研究通过对比五种不同的命令行接口(CLI)在相同模型和代码库下的表现,揭示了工具在上下文使用上的显著差异。Claude Code在每次调用中发送62,600个字符的工具定义,而pi发送2,200个字符,Aider则不发送任何字符。这一发现源于对API调用的拦截,旨在明确模型行为与包装器开销之间的关系。通过保持模型不变并更换包装器,研究者能够更清晰地分析各CLI在标准化编码任务中的表现差异。
📄 English Summary
Five CLIs Walk Into a Context Window
The research compares the performance of five different command-line interfaces (CLIs) using the same model and codebase, revealing significant differences in context usage among the tools. Claude Code sends 62,600 characters of tool definitions per call, while pi sends 2,200 characters, and Aider sends none. This finding stems from intercepting API calls to clarify the relationship between model behavior and wrapper overhead. By keeping the model constant and swapping the wrapper, the researcher aims to analyze the performance differences of each CLI in a standardized coding task more clearly.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等