Self-Improvement (wcore-evolve)
Skills do not have to stay fixed. Wayland can improve them over time through wcore-evolve, an evolutionary loop that takes a skill, produces variations, scores them, and keeps the ones that score better. The aim is a skill library that gets sharper with use rather than one you must hand-tune forever.
Where it sits in the lifecycle
Section titled “Where it sits in the lifecycle”Evolution is one stage in a longer skills pipeline. First a draft stage proposes new skills from observed tool sequences; a curate stage decides which drafts are worth keeping; then the evolve stage improves a kept skill generation by generation; and a final curate stage promotes the winners and retires the losers. Evolution refines an existing skill rather than inventing one from nothing.
The loop
Section titled “The loop”Starting from a parent skill, the loop produces a set of children by applying mutators in rotation: rephrasing the text, reordering steps, swapping a synonym, or adjusting a precondition. Each child is scored, the best child is retained, and the rest are archived. The loop repeats until it hits a generation ceiling, exhausts its budget, or detects a plateau where recent generations stop improving. The losing variants are not deleted; they are written to a graveyard so the run stays auditable.
The scoring trust boundary
Section titled “The scoring trust boundary”The part that makes this trustworthy is what does the scoring. Children are scored by a deterministic scorer with fixed constants, and that scorer never calls a model. This is the trust boundary: a model proposes variations, but a model never decides which variation is better. Without it, a system could talk itself into preferring its own output. Because the scorer is deterministic, the same skill scored twice gives the same number, so improvements are real rather than noise.
Online evolution
Section titled “Online evolution”Beyond the offline loop, the engine can run online evolution against live sessions. With --online-evolution enabled, the engine emits an evolution event at the end of a session, applies a rephrasing mutator to successful trajectories, and persists the results under the Wayland home directory. This lets the library learn from the work you actually do, while the same scoring boundary keeps the changes honest.