skillopt

SkillOpt treats a natural-language skill document as the trainable state of an otherwise frozen LLM agent. A separate optimizer model reviews scored agent trajectories and proposes bounded add, delete, and replace edits to a single Markdown skill file. In the default paper-style workflow, candidates are accepted only when they improve a held-out validation score. The framework also provides a textual learning-rate budget, a rejected-edit buffer, and epoch-level slow or meta updates intended to make revision more stable. The output is a reusable best_skill.md artifact, described as typically compact, that can be deployed with the unchanged target model and without additional inference-time optimizer calls. The training loop covers rollout, reflection, aggregation, selection, update, and evaluation. The package supports multiple chat and execution backends, including OpenAI, Azure, Claude, Qwen, MiniMax, OpenAI-compatible services, Codex, Claude Code, Copilot, and Cursor targets. It includes six built-in benchmarks and defines adapter patterns for adding new backends and benchmark environments. An optional WebUI provides a monitoring dashboard, while the main tooling is available through Python and CLIs. Version 0.2.0 introduces SkillOpt-Sleep, a nightly offline self-evolution flow that harvests past sessions, mines recurring tasks, replays them, and consolidates validated skills behind a held-out gate. The repository also contains integration shells for Claude Code, Codex, Copilot, and Devin, plus an OpenClaw reference adaptation; some of these integration files live outside the PyPI wheel. SkillOpt requires Python 3.10 or later and is installed from PyPI. Documentation covers setup, data preparation, training and evaluation commands, configuration, backend contracts, benchmark adapters, and framework internals.