ContextSniper cuts Claude Code token use 38.9% on repo repairs

A new paper cuts Claude Code's token spend by 38.9% on repository-level program repair, per arXiv 2607.01916. The method, called ContextSniper, targets the three places coding agents burn budget: reading whole files when a few lines would do, running searches too broad to be useful, and dumping long terminal output into context. On OpenClaw the cut was even bigger, 51.5% fewer tokens, translating to a 36.4% cost drop versus 27.3% for Claude Code.

The part that should worry anyone running agents at scale: resolution rates barely moved. Tested across 50 matched SWE-bench Lite tasks per condition, ContextSniper's three-level indexing of code and action memory, plus hybrid ranking for retrieval, held accuracy essentially flat while slashing the token bill. That's the claim worth testing yourself before trusting it, since a 50-task sample is thin enough that a wider benchmark could still surface a wider gap.

The most portable piece isn't the indexing scheme, it's the intention-aware filtering of tool output. Most coding agents read an entire terminal dump when they only needed the one line that changed. That's pure waste, and it's the kind of fix you could bolt onto an existing agent harness without rebuilding the memory system around it.

What this doesn't answer

The paper doesn't say whether the savings hold on tasks bigger than SWE-bench Lite's scope, or whether the indexing overhead itself eats into the win on smaller repos. If you're running agents against a codebase all day, every day, a 30-40% token cut is real money. Worth watching whether anyone ships this as a wrapper instead of a from-scratch harness.