Changelog

Each repository page now has share buttons for X and LinkedIn, plus a button that copies the link to the clipboard. The X button pre-fills the tweet with the AI-generated headline for the repository. The same URL works on both desktop and mobile.

A short note now appears on every deep analysis card clarifying that the report is based on the README and public GitHub signals — not an inspection of the source code. This makes the scope of the analysis clear before reading.

The deep analysis card previously showed a single badge labelled ADOPT, AVOID, or MONITOR. In practice this was always ADOPT because the AI produces a conditional verdict covering all three cases. The badge has been removed. The keywords ADOPT IF, AVOID IF, and MONITOR IF now appear highlighted directly in the text, preserving the full conditional recommendation.

Deep analysis reports are now only shown publicly when staleness_type is "ok" and the repository is not marked as fossil. Previously, analyses on stale or low-signal repositories were included in public counts and surfaces. The public count of deep analyses now reflects only reports that are current and attached to actively-cataloged repositories.

When no star range filter is active, the quality sort now combines quality score with discovery value. This prevents the quality ranking from always surfacing the same well-documented mega-projects. Users who want pure quality ranking can set a star range to override this behavior.

The catalog now explicitly tracks five states: active, archived (below 1k stars), fossil (stale and low-signal), excluded (manually hidden), and deleted. Only active repositories appear on public pages. The other states are preserved for historical tracking and can transition back to active if signals improve.

Each repository AI enrichment is now tagged with a prompt version. When the enrichment prompt changes in a way that affects scores or schema, repositories are automatically re-enriched on the next sync cycle. This prevents old enrichments from mixing with new ones in rankings.

Deep analysis reports are now checked for staleness on a scheduled basis. Reports tied to repositories that have grown significantly, changed category, or been produced by an outdated analyst version are marked for regeneration. Stale reports are hidden from public surfaces until refreshed.

Previously, clicking a language pill in the directory reset other active filters (quality bar, domain, sort). Language pills now preserve all active filters and only replace the language parameter. Quality, domain, license, and time filters remain active when switching language.

Weekly growth, star rankings, and search results were sometimes reading from different fields, causing minor inconsistencies between pages. All star-based calculations now use a single canonical source. Sort order, growth figures, and repository counts should now be consistent across catalog, search, trending, and ranking pages.

The language filter now shows only the 16 most represented languages by default (those with 100+ repositories). Less common languages (10–99 repositories) are accessible via a "more +" toggle. Active filters are always shown regardless of category size.

The Recent sort (newest repositories) now only surfaces repositories that have completed at least base AI enrichment. This prevents showing newly indexed but unenriched repositories that lack quality context.