Large language model (LLM) based agents have shown remarkable skill in automated software issue resolution, but they often stumble when tackling issues in a specific repository due to a lack of project-specific knowledge. Existing self-evolving methods either rely on historical issue-resolution signals or incur high per-issue test-time exploration costs.
To address this, researchers from Hugging Face propose SkillForge, a self-distillation framework that proactively acquires project-specific knowledge from the repository itself. Instead of waiting for real bugs to expose knowledge gaps, SkillForge synthesizes project-specific issues by re-implementing test-covered core functionalities of the repository.
By resolving these synthetic issues, SkillForge distills reusable project-specific knowledge into entity-grounded skills, associating them with relevant repository entities for future use. This enables agents to 'learn' a repository before encountering real problems, making them more effective when actual issues arise.
Extensive experiments using both open-source and closed-source models show that SkillForge consistently improves issue resolution performance over strong baselines, demonstrating that proactive knowledge acquisition substantially benefits downstream software issue resolution.