Reusing Past Repairs Through Hierarchical Trajectory Abstraction for Coding Agents

Researchers from Concordia University’s SPEAR Lab and Huawei Canada introduce STAIR, a framework that lets coding agents reuse procedural knowledge from earlier repairs instead of treating every issue independently. STAIR divides successful Lingxi repair trajectories into localization, planning, and execution-and-verification stages, then uses GPT-5 to organize each stage into a hierarchy ranging from repository-specific actions to transferable debugging strategies. For a new issue, similarity retrieval and an LLM relevance verifier select nodes across abstraction levels, while an LLM adapts them into executable, stage-specific plans. On the 500-instance SWE-bench Verified benchmark, STAIR reaches 81.2% Pass@1 with MiniMax M2.5 and 79.2% with GPT-5, outperforming same-backbone baselines including mini-SWE-agent v2, which scores 75.8% with MiniMax M2.5. The plans also transfer without code changes to mini-SWE-agent v2, raising Pass@1 to 81.0%, an absolute gain of 5.2%, while reducing average token use from 1,245k to 1,180k per instance. Ablations show why hierarchy matters: on a difficulty-stratified 125-instance subset, the full system resolves 80.0%, compared with 57.6% for raw trajectories, a 22.4% drop. The results indicate that reusable repair knowledge must combine concrete actions, intermediate investigation strategies, and high-level principles, then be adapted to the target repository rather than pasted into an agent prompt as generic advice.