DCAS：解耦 CLI 智能体脚手架以内化跨脚手架规划能力
原标题：DCAS: Decoupling CLI Agent Scaffolding to Internalize Planning across Scaffolds
AI 摘要
DCAS 论文指出，基于 CLI 的软件工程智能体在 OpenHands 环境下微调后，部署到其他脚手架时性能显著下降。为此，研究者提出 DCAS 拦截层，在不修改脚手架的情况下路由 API 流量，实现跨脚手架评估和规划感知轨迹收集。实验表明，仅改变规划来源即可将 Qwen3-Coder-30B-A3B-Instruct 的 Pass@1 从 42.8% 提升至 57.8%，且基于 576 条规划感知轨迹微调的模型在多个未见过的脚手架上表现稳定。
正文节选
DCAS: Decoupling CLI Agent Scaffolding to Internalize Planning across Scaffolds Abstract CLI-based software-engineering agents have matured rapidly, yet the open ecosystem has converged on a single training environment: trajectory datasets used to fine-tune open models are collected almost exclusively under OpenHands. Models fine-tuned on this data score well under OpenHands but degrade substantially when deployed under any non-training scaffold. Untrained base models do not show this divergence