XRepoSkill is a method for learning reusable skills for software engineering agents. It generates candidate rules by contrasting successful and failed trajectories from the same agent on the same issue, then verifies each rule by checking whether its prescribed behavior is associated with success and recurs across multiple repositories. This filters out incidental or habit-driven behaviors that often make skills fail to transfer.
Evaluated on SWE-bench Pro and DeepSWE with three different backbone LLMs, XRepoSkill achieves the highest issue resolution rate in all six benchmark-LLM combinations. On DeepSWE it improves resolution by 10.3 percentage points over the same agent without learned skills and by 5.0 points over the strongest skill-learning baseline. The key insight is that not everything in a successful trajectory causes success, so before reuse matters.