Repo Health Check

Give it a repository on GitHub, GitLab or Bitbucket (owner/name, or a github.com, gitlab.com or bitbucket.org URL) and get a Repo Health Score from 0 to 100 with a letter grade, a band (Thriving to Neglected) and the reasons for it. The score is a weighted model over nine categories: activity, responsiveness, backlog, releases, contributors, pull request flow, security posture, documentation and license. Each category is built from signals, and every signal states its formula, the exact API request and field it used, how many items it saw and how confident the figure is. Flags call out trends and anomalies (activity drop, single maintainer, popular but unanswered, stale backlog, archived, monorepo). There are three depths: quick, standard (the default) and deep, which spend more of the host's unauthenticated allowance. Compare two or three repositories side by side on a like-for-like basis, mixing hosts freely, and ask explain_score to walk any number back to the data that produced it. Everything is computed deterministically from live data on the public REST APIs (api.github.com, gitlab.com/api/v4 and api.bitbucket.org, GET only; a token is sent only if you supply one for a private repository). Hosts differ in what they expose without signing in, so a signal a host would not return (GitLab has no weekly commit statistics or advisories, Bitbucket no stars, archived flag, contributors endpoint or working issue tracker API) is reported as not checked with the reason, never scored as zero, and rate limits, private or missing repositories, network failures and malformed responses are each reported as what they are. Private repositories can be checked with an optional read-only token you supply for each host; with no token the agent works for public repositories exactly as before, and it says whether a failure was no token, a rejected token or a repository the token cannot see. Results also render as an interactive panel inside the chat. On GitHub the remaining allowance is read before anything is requested, and the depth is lowered with a warning (or the call refused with the reset time) when it cannot cover the check, so a comparison never ends half done. Finished reports are kept for 15 minutes so a repeat check, a repeat comparison and explain_score cost no request. No language model is needed; a client that offers model access (MCP sampling) may optionally have the summary paragraph reworded, and a rewrite that changes any figure is discarded.