STAIR Turns Past Repair Trajectories Into Reusable Plans, Hitting 81.2% on SWE-bench Verified and Transferring Across Agents

arXiv 2607.29658·high signalXBlueskyLinkedInCopy link

Repair agents normally treat every issue independently and throw away procedural knowledge from prior fixes. STAIR converts historical repair trajectories into multi-level trees spanning fine-grained diagnostic actions up to high-level strategies, then selects and tailors plan nodes into issue-specific prompts. Integrated with Lingxi it reaches 81.2% Pass@1 on SWE-bench Verified using MiniMax M2.5 and 79.2% with GPT-5; the plans also transfer with zero code changes to the structurally different mini-SWE-agent v2, lifting it from 75.8% to 81.0%. Ablations show mixing abstraction levels beats any single level and that raw unabstracted trajectories transfer far worse.