EvoSkill

Why we included this project

Coding agents repeat their mistakes unless someone teaches them otherwise. EvoSkill automates that retraining for harnesses like Claude Code, Codex CLI, and OpenCode: it watches where your agent fails, proposes mutations to its skills and prompts, and evaluates the variants against a benchmark you supply, so each iteration yields a new agent program rather than a lightly edited prompt. The skills it produces are portable folders with instructions, metadata, and helper scripts, so a skill evolved in one harness can move to another, even one running a different model. Teams that already keep question/answer tasks or containerized benchmarks can point EvoSkill at them and let the improvement loop run locally, in Docker, or in a cloud sandbox. It is a good fit for anyone who wants coding agents to get measurably better over time without writing skill files by hand.