Documentation menu
On this page

0043 — Seed skills and the zero-grant rule (v83-F12/F13)

Status

Accepted (planned v83).

Question

skep had a library with shelves, a card catalog, a checkout system — and zero books: the skill machinery (registry, learning loop, observer, forge, SKILL.md import) was complete and the registry empty. Hermes ships 83 skills; the field gap was content, not architecture. How does stock content enter a system whose whole posture is "nothing enters without the operator"?

Decision

Seeds ship in-repo at src/skep/seeds/skills/<name>/SKILL.md (the v44-F6 pack format) and sync into the registry at serve startup and via skep skill seed, with provenance='seed'. Three rules, ordered by who wins:

  1. Zero-grant only. A seed is procedural knowledge — instructions the Queen can read and follow with the tools the operator already governs. A seed shipping scripts or wanting any capability is SKIPPED with a message naming the deliberate lane (skep skill import-md --allow-script). Shelf space is free; permissions never are (I5, I6 — the skill_md human grant gate holds for seeds exactly as for imports).
  2. The operator wins. An existing template under a seed's name — any provenance, including an edited seed — is never overwritten. Amended v126-F2: an unedited seed or shelf copy (it still matches the fingerprint recorded when it loaded) follows the shipped text and leaves an event row; an edited one is kept and reported diverged. A pre-v126 row is refreshed only when it matches a text that once shipped; otherwise skep skill seed --refresh NAME|--all.
  3. Deletes are durable. Deleting a seed-provenance skill writes a tombstone the loader honors forever; a restart never resurrects what the operator removed (I8).

The skill index orders operator-authored skills above the stock shelf, so seeds can never evict what the operator built from the prompt; the existing entry cap and overflow line bound the block regardless of shelf size (the review item 6 measurement rides the F13 tests).

Consequences

  • A fresh install answers "research this company" / "review this PR" with a loaded recipe instead of improvisation.
  • Every seed names only tools that exist — a lockstep test parses the whole shelf against TOOL_SPECS (the v25 COMMANDS-pin lesson).
  • Seeds whose Hermes originals assumed capabilities skep routes differently are adapted, and say so in their own text — a seed never pretends to a capability skep doesn't route (I8).