SWE-Pruner Pro: Coder LLMs Already Encode Which Context to Prune

arXiv·high signal

Yuhang Wang, Yuling Shi and Shaoqiu Zhang argue that long-context pruning for coding agents does not need a separate scoring model — the coder LLM's own internal signals already identify prunable context, removing the extra model call that methods like LLMLingua-style compressors require. The paper (arXiv 2607.18213, cs.CL/cs.SE) targets the exact bottleneck agent harnesses hit when trajectories exceed the window. For builders running Claude Code or similar agents on large repos, this is the cheapest lever on token cost that does not degrade the retrieved evidence.