ContextSniper cuts Claude Code's tokens 38.9% on repo-level repair with no measurable drop in resolve rate

arXiv 2607.01916·high signal

arXiv 2607.01916 targets the real budget sink in coding agents — whole-file reads, broad searches, and long terminal dumps — with three-level indexing of code and action memory, hybrid ranking for candidate retrieval, and intention-aware filtering of verbose tool output. On 50 matched SWE-bench Lite tasks per condition it cut tokens 38.9% on Claude Code and 51.5% on OpenClaw (cost down 27.3% and 36.4%) while resolution rates stayed essentially unchanged. The intention-aware output filter is the piece most worth copying standalone: most agents pay full price for terminal output they read one line of.