mirror of
https://github.com/nextlevelbuilder/ui-ux-pro-max-skill.git
synced 2026-11-18 16:23:38 +00:00
* feat(data): add rtl_level to styles and products Classify every style and product row for right-to-left interfaces: full (mirrors unchanged), partial (mirrors with per-component work), caveats (layout encodes reading direction). One column is appended per row; existing fields, column order and LF line endings are untouched. Mirrors under cli/assets and .claude are synced with npm run sync:assets. * feat(search): add --rtl filter to style and product searches search() accepts rtl="all|full|partial|caveats" and projects the new rtl_level column, so filtered rows still carry their compatibility level. Without that projection the filter saw None and dropped every row, which is what broke the earlier attempt. The filter runs after BM25 scoring but before max_results truncation. Filtering the corpus before indexing shifted IDF and made calibrated searches abstain incorrectly (dashboard --domain style --rtl=partial returned nothing); scoring first keeps the corpus and thresholds stable while still preventing -n 1 from hiding a matching row. --rtl requires an explicit style/product domain, and is rejected with --design-system and --stack. relevance thresholds: runtimeFingerprint refreshed; every metric is identical to the committed baseline. * test: cover rtl data contract and filter ordering Pin the behaviour the RTL filter depends on: - rtl_level is projected, not a search signal; - every row in both CSVs carries a supported level; - known layouts keep their curated level; - filtering happens before max_results truncation (saas --domain product -n 1 with --rtl=full returns Micro SaaS, while the unfiltered top hit is a partial row); - omitting --rtl keeps previous results; - unsupported domains raise ValueError and exit non-zero through the CLI. * docs: document the rtl compatibility filter Explain the three levels, the CLI flag and its domain restriction in the README, CLAUDE.md, the generated skill content and the hand-authored Claude skill. The agent guidance tells the skill to add --rtl for Arabic, Hebrew, Persian and Urdu work and to respect the returned level instead of silently mirroring a caveats layout.