Search kept its own name check, which rejected mixed-case names such as
"MySkill" that install happily accepts, so those skills never showed up
in results. It now uses validate.SkillName, the rule install applies.
Names install rejects, such as a leading underscore, stay hidden.
- Add repositoryMatchScore so popular repos matching the query (e.g. 'superpowers')
rank above lower-signal exact-name copies; rebalance weights to fit.
- Fetch additional skill candidates via GitHub repository search + git tree API,
surfacing SKILL.md files from well-known repos that code search misses.
- Prioritize preferred skill repos when batching star enrichment so their star
counts are always populated within the fetch cap.
- Parse one-line '--- name: ... description: ... ---' frontmatter so single-line
SKILL.md files are discoverable and render correct metadata.
- Update CHANGELOG/website changelog for v0.19.5.
Search results were dominated by low-quality repos because the scoring
formula gave name match 45% weight while stars (the strongest quality
proxy) only had 5%. Zero-star personal experiments ranked above
well-known skill repos.
Changes:
- Rebalance scoring weights: name 30%, desc 20%, stars 25%, source 25%
- Improve normalizeStars curve (log10/3.0) for better differentiation
in the 5-50 star range where most useful skills live
- Lower sourceQualityScore base from 0.35 to 0.15 for unknown repos
- Add filterLowQuality(): remove 0-star non-preferred repos, short
descriptions (<10 chars), and known spam orgs (inference-sh)
- Refactor search pipeline: extract buildGitHubCodeSearchQuery(),
processSearchItems(), convertToSearchResult() for testability
- Add preferred repo boosting (anthropics/skills, vercel-labs/skills)
- Add SKILL.md frontmatter validation to filter non-skill results
- Add equivalent skill deduplication (same name+description → keep
best source)
- Add repo-scoped query support (owner/repo[/subdir] patterns)
Closes#130 in .feature-radar/opportunities/
Search results were ranked only by GitHub stars, causing unrelated
high-star repos to outrank dedicated skill repos. Additionally,
owner/repo queries (e.g., "vercel-labs/skills") were treated as
plain text search instead of repo-scoped search, returning
completely wrong results.
Changes:
- Add multi-signal relevance scoring (name 0.45, description 0.25,
stars 0.30) with log10-normalized star counts
- Detect owner/repo[/subdir] queries and use GitHub `repo:` qualifier
- Parallelize star and description enrichment (WaitGroup + semaphore,
cap 10) — reduces wall time from 3-6s to ~200-600ms
- Extract both name and description from SKILL.md frontmatter
- Add Score field with json:"-" tag (hidden from CLI --json and API)
- Remove overly generous ContainsAny partial matching in name scoring