Large language model (LLM) agents increasingly rely on reusable skills to tackle complex software engineering tasks. To keep these skills effective across different repositories, they must evolve with changes in code structure, implementation conventions, and maintenance requirements. Pull requests (PRs) capture implementation and review processes, naturally providing the knowledge of repository evolution. New PRs need to follow repository‑specific patterns, keep related components consistent, and ensure existing tests pass while adding coverage for new functionality. This creates a mutually reinforcing loop: PRs supply material for skill updates, and the updated skills help future PRs implement, maintain, and respect conventions.
To enable the co‑evolution of PRs, repositories, and skills, we propose the Skill Dreaming framework. Given a repository and its PR history, the framework first extracts executable knowledge from validated PRs to build a repository‑specific skill library. As new PRs arrive, the framework continuously refines and extends the skills. The resulting skills guide coding agents not only in implementing requested features but also in maintaining associated tests, documentation, compatibility, and repository conventions.
Skill Dreaming links four essential stages:
- PR analysis – parses code changes, review comments, and CI outcomes to extract implementation steps and constraints.
- Skill management – organizes extracted steps into callable functions or scripts, tracking version information.
- Code execution – runs skills in a sandbox to verify correctness and generate test coverage.
- Continuous integration – feeds execution results back into CI, ensuring new skills pass all checks when applied to future PRs.
The end‑to‑end workflow forms a closed loop from PR to skill and back to PR, allowing each code submission to become new learning material for automated support. Over time the skill library grows richer, enabling agents to adapt faster to repository specifics and reducing manual review effort.
Review: Skill Dreaming transforms software maintenance experience into machine‑callable skills, achieving continuous knowledge accumulation and automation reuse, and offering a practical path for LLMs to collaborate effectively in real codebases.