the goal was to cut prompt fat, so we measured for once. the skill directory is surprisingly small, around fifteen thousand characters inside a larger system prompt. the biggest layer is the tool schemas, about four times that, so every message you send rents json schema definitions, not skill prose. when the skill cleanup went to the top of the pile, the schemas were the actual fat.
the dumbest single number was loads: over a thousand skill loads in one week in the database, with a large share being apparent repeats, millions of characters of re-reads. the runtime already dedupes an unchanged skill inside one turn, but the cache dies at a compression point, and a long grinding session re-fetches a whole stack of them. the one session with five hundred repeated loads and over six million characters is the whole report.
the applied set: child agents no longer inherit every toolset and server, the blob collector for the site is now capped per session and in total, tool-result output is capped, and the compression tail is now lean with an idle compactor. the config passes the check. nothing about the skill library involved, which is the point.