A coding agent picks its skill from a list of names and descriptions; the skill body is only loaded once it has already committed. Put a similar skill from another author beside the one you installed and the share of runs that use yours falls 19.9 points — while task completion does not move at all, because both skills satisfy most of what a completion check looks at. What goes missing is the behaviour only your skill has, and the reply names which skill ran in 0.9% of the runs where the other one won.