pub struct SkillAuthoring {
pub enabled: bool,
pub min_steps: u32,
}Expand description
Whether, and how, DISCOVER may author a Skill — a reusable procedure with an applicability condition and ordered steps, derived from a trajectory that succeeded.
This exists because of a measured gap. On PAST-Bench the agent performed
the procedure correctly in the learn episode on every seed and then, asked
at session end whether there was anything to save, answered “nothing to
save” — so the store was empty at evaluation and the memory scored below
having none (crates/areev-bench/PERSIST.md). Every Skill in the memory
depended on the model volunteering one mid-task. Hermes does not depend on
that: a separate review pass writes its skills. This is Areev’s equivalent,
and it runs through the same gates as every other draft — GROUND, VERIFY,
the confidence floor, a review with a BECAUSE — and is never auto-applied.
Fields§
§enabled: boolOffer the skill proposal kind to the proposer at all (default: yes,
under LLM enrichment; no LLM, no skills).
min_steps: u32Fewer ordered steps than this is a lesson, not a procedure (default 2).