Vault skills dimension: 9 Rust scanners + repair sweep zeroed 42 stale-version failures
->
The skills dimension of the vault went from advisory-only linting to enforced quality in one session. Nine Rust scanners landed via the vault-scanners crate, a repair sweep fixed 42 stale-version stamps across 40 skill files, and a gate now blocks export when any skill violates the schema contract.
What Happened
The vault’s skill files had accumulated version drift: 38 of 40 carried stale YYMMDD stamps that no longer matched the content they described. The problem was invisible because no scanner existed: skills were validated by eyeball during authoring and never re-checked.
We shipped nine Rust scanners through the vault-scanners crate, each targeting a distinct failure mode: version freshness, frontmatter schema conformance, section-header completeness, dead wikilink detection, tag vocabulary enforcement, duplicate-slug detection, word-count bounds, missing created field, and orphan skill detection (skill files unreachable from any project index).
Before enabling the gate, we ran the repair sweep. 42 failures across the 40 files: mostly stale version stamps, a handful of missing tags and malformed frontmatter. All 42 repaired in a single pass. The MiniMax-AI/skills benchmark served as external calibration: our scanner definitions matched or exceeded the field coverage of that public benchmark, confirming we weren’t under-specifying the schema.
After repair, the dimension gate was promoted from advisory to blocking. Any future skill that fails any of the nine scanners is rejected at export time.
Why It Matters
This is the topics/patterns/repair-first-then-tighten pattern executed at dimension scale. The prior state was the classic failure mode: if we had shipped the gate without the repair sweep, 38 of 40 skills would have failed immediately, and the gate would have been softened or disabled within a day. By pairing the sweep with the gate, the residual was zero on day one: the gate launched green and has stayed green.
The vault-scanners crate is also the first reusable scanner infrastructure. The same Rust binary can be extended with new scanner functions and pointed at other vault dimensions (breakthroughs, pitfalls, experiments) without per-dimension shell scripts. One binary, N dimensions.
External calibration via MiniMax-AI/skills is a pattern worth repeating. Internal schema definitions tend to drift toward whatever the author remembers. Benchmarking against a public skill-definition corpus caught two field omissions (explicit trigger and skip sections) that our internal convention had quietly dropped.
Evidence
- 42 failures detected pre-sweep across 38/40 skill files (primarily stale version stamps).
- 0 failures post-sweep. Gate promoted to blocking same session.
- 9 scanners shipped in vault-scanners crate: version, frontmatter, sections, wikilinks, tags, slugs, word-count, created-field, orphan.
- MiniMax-AI/skills benchmark used as external field-coverage calibration: two omissions caught and added before gate promotion.