# Close the verified quality gaps in Col MCP \(issue \#61\) · PR \#62 · screen/col

[View on GitCafe](https://git.cafe/screen/col/pull/62)

Repository: [screen/col](https://git.cafe/screen/col)

Visibility: public

State: merged

Source repository: screen/col

Source branch: feat/issue-61-quality-gaps

Target branch: main

Head commit: 81dc628b1af647de9fa4dfd57bde2f73479d3326

[Discussion and reviews](https://git.cafe/screen/col/pull/62)

[Commits](https://git.cafe/screen/col/pull/62/commits?format=markdown)

[Files changed](https://git.cafe/screen/col/pull/62/diffs?format=markdown)

## Description

---

Closes #61 — the full quality-gaps plan from the independently re-verified external review, delivered across 7 commits.

## What landed

- **Install truth.** `get_component` answers with the exact per-component command generated from each registry's documented template (`npx shadcn@latest add command`, not `add button`) plus the component's npm and registry dependencies (`cmdk`, `cn`, `dialog`).
- **Compatibility truth.** Tracked subjects derive per library (no more "Emotion: unknown" on Tailwind libraries), real ranges recorded from npm peer metadata and official docs, and a build-time rule fails when recorded facts contradict a library's own getting-started text (it caught Mantine's stale React 18+ against its v9 peers).
- **Component records.** `kind` on all 1435 components (primitives / full patterns / visual effects / layouts), a `stacks` filter on `search_components`, shared concept aliases so `modal` finds every library's dialog, and 1019 one-line summaries taken from each component's own documentation page.
- **Search precision.** Word/prefix matching with separator folding and de-spaced index variants: `cmd k` drops from 131 hits to 2, `e` from 767 to 0, and `commandpalette` / `command-palette` / `cmd+k` / ⌘K all resolve like the spaced form.
- **Ranking and attribution.** Real search UIs rank above visual effects sharing a word ("spotlight" leads with Mantine Spotlight and shadcn Command), and every hit states what matched (`component:`, `name:`, `alias:`, `tag:`, `description:`).
- **Self-explaining results.** Empty results carry an `indexStatus` hint and near-miss suggestions, `get_library` lists the library's recorded components, and shared caveats appear once per response instead of on every hit.
- **Freshness.** Verification is per field (`page` / `install` / `compatibility` / `snapshot`) and identical across tools, every record carries `verifiedAt` (all 1435 re-verified 2026-10-10), source URLs record their variants (radix/base), and `scripts/verify-sources.mjs` re-checks drift for maintainers.
- **Coverage fill.** 346 components across the five named empty shells (HeroUI, Flowbite, origin-ui/coss, Kibo UI, Cult UI), plus cmdk and kbar as first-class entries with npm-verified facts.
- **Docs corpus (item 7).** Reuse permissions recorded registry-wide from the licensing audit — 64 granted with each library's verified license source, Aceternity denied per its own license, and undecided libraries carry no entry by design. 58 documentation snapshots hold install/usage evidence with fetch dates and the git SHA each page documented; thin landing pages earn no snapshot.

## Verification

- CoVe acceptance gate: **18/18** re-verified against the live tool layer after the final commit.
- `npm test` **103/103** green, `npx tsc --noEmit` clean.
- Every component URL fetched live during ingestion: zero dead on arrival, 78 HeroUI canonical-URL drifts corrected.
- The licensing audit was re-checked at source (LICENSE files, license pages, npm peer metadata), including the MIT + Commons Clause family (react-bits, rare-ui, Nexvyn) where docs reuse with attribution is permitted while component redistribution is not.

## Remaining, deliberately optional

A handful of libraries document only on JS-rendered sites that serve nothing to a fetcher; their permission records are in place for when the content becomes reachable. Same for ~400 component one-liners. Nothing here blocks the acceptance criteria.