Reusing Past Repairs Through Hierarchical Trajectory Abstraction for Coding Agents
Source: https://arxiv.org/abs/2607.29658
Authors: Yisen Xu, Jiayuan Zhou, Ruiqi Pan, Tse-Hsun Chen
Published: 2026-07-31
Categories: cs.SE
Grounding: opencli arxiv metadata + Obsidian paper digest excerpt (OpenClaw定时任务/论文流水线/2026-08-04-论文流水线.md)
摘要（中文）
论文提出 STAIR，把历史修复轨迹抽象成多层可复用计划，再注入新 issue 的 prompt摘要报告在 SWE-bench Verified 上，STAIR + Lingxi 使用 MiniMax M2.5 达到 81.2% Pass@1，使用 GPT-5 达到 79.2%；同一计划还能把 mini-SWE-agent v2 从 75.8% 提升到 81.0%这篇适合和Agent 记忆到底该存什么放在一起读
Abstract
Although LLM-driven repair agents can tackle complex, repository-level issues, they treat every issue independently and discard the procedural knowledge accumulated from previous repairs. We introduce STAIR, a framework that converts historical repair trajectories into hierarchical, reusable plans that can be adapted to steer future repairs. Each past trajectory is transformed into a multi-level tree that ranges from fine-grained diagnostic actions to high-level repair strategies, encoding experience at several granularities. When a new issue arrives, STAIR selects relevant plan nodes from multiple abstraction levels, tailors them into executable, issue-specific plans, and supplies them to the agent through its prompt. On SWE-bench Verified, STAIR integrated with Lingxi reaches 81.2% Pass@1 using MiniMax M2.5 and 79.2% using GPT-5.
Intake rationale
- Category: coding
- Quality score: 5/5
- Why it matters: STAIR turns past repair trajectories into reusable plans for future coding-agent issues.
Obsidian evidence excerpt
论文提出 STAIR，把历史修复轨迹抽象成多层可复用计划，再注入新 issue 的 prompt。摘要报告在 SWE-bench Verified 上，STAIR + Lingxi 使用 MiniMax M2.5 达到 81.2% Pass@1，使用 GPT-5 达到 79.2%；同一计划还能把 mini-SWE-agent v2 从 75.8% 提升到 81.0%。这篇适合和“Agent 记忆到底该存什么”放在一起读。