Struct SkillDistiller 

pub struct SkillDistiller { /* private fields */ }

Decides whether a finished run is worth a skill, and drafts it.

Implementations

impl SkillDistiller

pub fn new(model: Arc<dyn Model>) -> Self

pub fn with_policy(self, policy: DistillPolicy) -> Self

pub fn with_min_tool_calls(self, n: usize) -> Self

pub fn with_similarity_threshold(self, t: f32) -> Self

pub fn policy(&self) -> &DistillPolicy

pub fn gate(&self, ep: &Episode, existing: &[ExistingSkill]) -> GateDecision

The whole trigger decision, with no model call and no I/O.

Checks run cheapest-first and return the first failure, so the reason a host logs is the most fundamental one rather than an incidental later check.

pub async fn distill( &self, ep: &Episode, existing: &[ExistingSkill], ) -> Result<Distillation, DistillError>

Gate, then (only if it passes) spend one model call to draft a skill.

Never writes to disk. existing is the duplicate-suppression corpus — use existing_skills_in to build it from a skills root, or pass the caller’s own per-tenant view.

pub async fn distill_and_write( &self, ep: &Episode, skills_root: &Path, ) -> Result<Option<PathBuf>, DistillError>

Convenience for hosts that need neither per-tenant scoping nor a human gate: scan skills_root for duplicates, distil, and write on success.

Returns the written path, or None when the gate declined.