ContextSniper Boosts LLM Code Repair Efficiency with Token-Saving Memory

Large language model agents are increasingly capable of repairing real-world software issues at the repository level. However, their efficiency is often hampered by excessive token consumption, as they frequently read entire files, perform broad searches, and process lengthy terminal outputs that contain much irrelevant information alongside useful evidence. To address this, researchers have developed ContextSniper, a token-efficient code memory layer specifically designed for repository-level program repair within AntTrail's broader agent memory engine. ContextSniper employs a "Sniper" feature that precisely selects evidence. It retrieves candidate code and runtime data, ranks it using hybrid signals, and then filters long outputs through an intention-aware context gate. This process results in compact evidence packets being fed to the LLM, while still preserving recoverable source context outside the immediate prompt. Evaluations on SWE-bench Lite show ContextSniper reducing total token use by over 50% and estimated cost by over 36% for OpenClaw, and similar reductions for Claude Code, with only a slight decrease in resolution rates. The pilot testing scripts are open-sourced.