Pengcheng Xu

Three lines of context in the LSP reply cut follow-up file reads from 15.2 to 3.2

Pengcheng Xu's AgentConnect comparison finds semantic navigation beats grep only where text search is noisy; on clean repositories it added 16 to 19 percent to token cost for little or no accuracy gain.