Agent Skills: Library Knowledge

Library Knowledge

The external-knowledge shelf. shelf_index (feature-workflow) inventories the substrate you built; this inventories the substrate you depend on. It exists because a skill is static text and rots on the training clock — the very problem it solves — so this skill is not the knowledge. It is a librarian: a version-stamped, queryable store plus scripts that refresh it against live sources and serve it cheaply.

The value: confirm once (expensive, effectful), cache, serve many times cheap, refresh only on drift.

The store entry (the interface)

lib-knowledge.jsonl at the repo root — one self-describing record per line (JSONL):

• confirmed_version + confirmed_on — provenance. An entry behind the installed version is unparsed knowledge; trusting it blind is the stale-config bug.
• key_facts — decision-relevant only, not documentation.
• delta_from_prior — refresh is a diff from the recorded version, not a relearn.
• capability — what problem it solves, so the VENDOR rung can ask "already solved?"
• source_url — where the fact was confirmed.

Consult (cheap, frequent)

All reads go through store.py (the storage port) and are token-cheap by construction — the agent's context only ever holds what it asked for:

• python scripts/lib_lookup.py <name> → ONE entry + staleness, streamed from the JSONL with an early exit at the matching line (never the whole store).
• python scripts/lib_lookup.py → the index: name + version + date per line only.
• python scripts/lib_lookup.py --search "<capability terms>" → ranked library names + one-line capabilities (not full entries) — the path the VENDOR rung uses to ask "is this already solved?".

Refresh (effectful, rare) — confirm, then record

Freshness only comes from going to the world, so refresh has two halves:

• Judgment + effect (the agent): search the library's official docs/changelog for the installed version and extract decision-relevant facts. See references/confirming.md. This is "observe before you change" aimed at documentation — verify against the live source instead of memory.
• Mechanical (the script): python scripts/lib_refresh.py --set <name> --from-json <entry.json> merges the entry and stamps confirmed_on.

Trigger refresh on: a miss (no entry), a stale verdict (installed drifted past confirmed), or at seed (framework-defining deps only).

Staleness is the gate

python scripts/lib_refresh.py --check compares every entry's confirmed_version to the installed version and exits non-zero if any is STALE. Wire it as a verify-gate check so generated code can't ship against facts the lockfile has already moved past — the check that would have caught a stale config.

Storage is behind a port (token-efficient, swappable)

store.py is the ONLY module that touches the store; the scripts and the harness call it, never the file. Backend today is JSONL: one record per line, read by streaming, so lookup cost is one record and the store can grow to hundreds of libraries without becoming a context hoarder. It is also grep-friendly and imports into SQLite trivially (one row per line).

Because every consumer sees only the port's interface, the backend can change with zero caller changes. SQLite / FTS5 is the documented swap — adopt it only when capability-search across a large store becomes a hot path (lookup and index don't need it; a filesystem/stream already serves those in one record). Until then, adding a database would fail the VENDOR cost test we apply to every dependency: it solves a token cost we don't have rather than a search-at-scale cost we haven't hit.

How the harness consults it (the four seams)

• harness-setup seeds the framework-defining deps at boot.
• feature-workflow shelf-check (the VENDOR rung) queries capability to ask "is this already solved by a library?" before hand-rolling.
• verify runs lib_refresh --check as a freshness gate.
• On a miss or stale verdict, the refresh path fires: confirm against docs, record, proceed.

Seeding policy

Eager only for the framework-defining deps (framework, styling system, language runtime) at setup; lazy for the rest — confirm a library the first time a feature reaches for it. Evidence, not anticipation; the store never bloats with libraries no feature touches.

Files

• scripts/store.py — the storage port (JSONL today; the one place to swap in SQLite). The only module that touches the store.
• scripts/lib_lookup.py — cheap reads: one entry, the index, or --search over capabilities.
• scripts/lib_refresh.py — record a confirmed entry (--set), or check staleness (--check).
• references/confirming.md — how to confirm a fact against live sources (the judgment half).