These are the only figures here that a single API call cannot produce. A star count is one request; "grew 37.7% in 90 days" and "overtook gemini-cli" are not. They come from dated Wayback Machine captures of each repository's GitHub page, matched against today's measured reading.

Nothing here is interpolated. A value "90 days ago" is the last capture at or before that date - not a line drawn between two captures - and every row prints the capture date it actually used and how many days early it was. Captures are irregular because archiving is not scheduled, so some repositories have gaps and simply do not appear in a window.

Forks fall, and that is published rather than suppressed. This is not the star rule: a star count only ever drifts down by a handful, but a fork count can lose a third of its value in three weeks while everything else rises - cline/cline went from 7,661 forks to 5,328 between 2025-11-09 and 2025-12-01 while its stars climbed from 52,197 to 53,858. That capture was re-fetched and re-parsed by hand to confirm GitHub really printed 7,661, so it is a fork cleanup on GitHub's side, not a measurement error. 12 negative movements are published in this series; 0 were excluded as more likely broken captures than real events.

Rank is among repositories that have a usable snapshot in the window, not among all tracked tools.

A crossover can only be dated to the window between two captures, never to a specific day - the archives do not exist for the days in between.

Source. Wayback Machine captures of github.com/<owner>/<repo>, reading the exact count GitHub prints in the star counter. Each figure's archive URL is stored alongside the data so any number can be re-checked by hand.

Why not the GitHub API. Per-star starred_at timestamps would be exact, but stargazer pagination stops at 40,000 stars - so the fastest-growing repositories, the interesting ones, are exactly the ones it cannot walk. GH Archive was tested and rejected: its event totals cover only 11-52% of real stars, so any curve from it is an estimate.

Star counts fall as well as rise. Users unstar. Across this dataset there are 17 decreases, all between 1 and 85 stars. A drop beyond 3% is treated as a suspect capture and excluded rather than published as negative growth; currently there are 0 such rows.

This is reconstructed data. It is kept separate from the measured readings and is never averaged with them. A backfilled point and a measured point are different evidence.

Mixed list: the first rows are 30-day growth, then 180-day.

Every figure on this page is computed from a downloadable series of dated captures — one row per repository per metric per capture, with the exact archive URL used, so any number can be re-checked by hand.

reconstructed-history.csv · reconstructed-history.json · index of all exports · field dictionary and JSON Schema

Licensed CC-BY. The file carries its own method and limits, and states that it is reconstructed rather than measured. Use it, but do not merge it into a series of measured readings without saying so — a backfilled point and a measured point are different evidence.