Skills

What it does

hyodo skills treats a skill (a Markdown file with - rule bullets) as a lens rather than a menu. ingest records a source and compiles its rules into .hyodo/skills/manifest.json; lens re-parses every ingested source live and reports, per pillar, an integer observed / expected coverage count and a passed / observed coherence count; propose renders a tailored skill made of the rules that currently pass. Everything is pure stdlib, re-derived from the source file each run — no model, no RAG, no embeddings, no vector store anywhere in this feature.

hyodo skills ingest my-team/SKILL.md
hyodo skills lens

Only four rule-text prefixes compile into a mechanical, executed check (require file:, forbid pattern:, require pattern:, ruff:); anything else is advisory and listed under unobserved.

What is stored

Compiled rule ids, pillar assignment, and pass/fail state in .hyodo/skills/manifest.json.
A content_digest and status ("unreadable" when a source cannot be read) for every ingested source.

What is never stored

A skill’s raw body, unless --store-body is passed on ingest.
A url: source’s contents — only its domain is recorded; HyoDo never fetches a URL over the network in this package.
A percentage, probability, or confidence value anywhere in a lens report.
A vector or embedding from a --from-node research-node hand-off — only rule text, digests, and an ordinal score_rank are accepted; a score or probability field on a retrieved item is rejected outright.

Exit codes

Command	DENY	UNOBSERVED	ASK
`skills ingest`	1	2	3
`skills lens`	n/a	2 (malformed manifest)	n/a
`skills propose`	n/a	n/a	n/a

Ingesting a skill is an unconditional external variable, like a web fetch — it stays ASK at trust level 2 and only clears automatically at trust level 3.

Full reference

docs/SKILLS.md in the repository covers rule ids, pillar keyword mapping, and the research-node contract in full.