Crate zeph_skills 

Skill subsystem: SKILL.md parser, registry, embedding matcher, hot-reload, and self-learning.

Overview

zeph-skills is the skills runtime for the Zeph AI agent. Skills are Markdown documents (SKILL.md) with a YAML frontmatter header that describe what an agent can do, which tools it may invoke, and how to use them. At runtime the agent matches user intent to skills via embedding similarity and BM25 lexical ranking, then injects the matching skill body as prompt context.

Module Map

Module	Responsibility
loader	Parse and validate SKILL.md frontmatter and body
registry	In-process skill index with lazy body loading
matcher	Async embedding-based skill matching with two-stage category filtering
bm25	In-memory BM25 inverted index with Reciprocal Rank Fusion
watcher	Filesystem hot-reload via notify debouncer
bundled	Compile-time embedded bundled skills and startup provisioning
manager	Install / uninstall / verify skills from git URLs or local paths
trust	Trust levels and source tracking for installed skills
trust_score	Bayesian Wilson-score re-ranking of skill match candidates
scanner	Advisory prompt-injection pattern scanner
prompt	Format skills into prompt XML blocks
generator	NL-to-SKILL.md generation via LLM
evolution	Self-learning: failure classification, step corrections, outcome tracking
erl	Experiential Reflective Learning: heuristic extraction from completed tasks
stem	STEM: automatic detection of recurring tool-use patterns
rl_head	2-layer MLP routing head trained with REINFORCE for skill re-ranking
resource	Skill-local resource file discovery and loading
qdrant_matcher	Qdrant-backed vector store for skill matching at scale (feature qdrant)

Quick Start

use zeph_skills::registry::SkillRegistry;

// Load all skills from a directory (metadata only — bodies are lazy).
let registry = SkillRegistry::load(&["/path/to/skills"]);
println!("loaded {} skills", registry.all_meta().len());

// Look up a skill body by name.
let body = registry.body("my-skill")?;
println!("{}", body);

Re-exports

pub use bundled::bundled_skill_names;

pub use error::SkillError;

pub use evaluator::EvaluationWeights;

pub use evaluator::SkillEvaluationRequest;

pub use evaluator::SkillEvaluator;

pub use evaluator::SkillQualityScore;

pub use evaluator::SkillVerdict;

pub use generator::GeneratedSkill;

pub use generator::SkillGenerationRequest;

pub use generator::SkillGenerator;

pub use matcher::IntentClassification;

pub use matcher::MatchResult;

pub use matcher::ScoredMatch;

pub use proactive::DomainLabel;

pub use proactive::ProactiveExplorer;

pub use trust::SkillSource;

pub use trust::SkillTrust;

pub use trust::compute_skill_hash;

Modules

bm25

In-memory BM25 inverted index over skill descriptions with Reciprocal Rank Fusion.

bundled

Compile-time embedded bundled skills and startup provisioning.

erl

ERL (Experiential Reflective Learning) — heuristic extraction and injection at skill match time.

error

evaluator

External-feedback skill evaluator (#3319).

evolution

Self-learning skill evolution: failure classification, outcome tracking, and prompt templates.

generator

NL-to-SKILL.md generation pipeline.

loader

SKILL.md frontmatter parser and file-system loader.

manager

Skill lifecycle management: install, remove, verify, and list installed skills.

matcher

Async embedding-based skill matcher with optional two-stage category filtering.

miner

Automated skill mining from GitHub repositories.

proactive

Proactive world-knowledge exploration (#3320).

prompt

Format matched skills into the <available_skills> XML block injected into the system prompt.

registry

In-process skill registry with lazy body loading and content fingerprinting.

resource

Skill-local resource file discovery and safe loading.

rl_head

SkillOrchestra RL routing head: 2-layer MLP for skill re-ranking.

scanner

Advisory content scanner for skill bodies.

stem

STEM (Skill Template Evolution from Mining) — automatic detection of recurring tool-use patterns and conversion to SKILL.md candidates.

trust

Skill trust levels and provenance tracking.

trust_score

Bayesian posterior weight computation for skill ranking.

watcher

Filesystem hot-reload for skill directories.

Enums

SkillTrustLevel

Trust tier controlling what a skill is allowed to do.