architecture

GitHub is the heavy sibling. Coding-platform APIs scrape one page or one contest API. This one walks commit diffs so language bars count the user’s patches, not the upstream tree of a Linux fork.

Live: github-stats.tashif.codes. Playground: playground. Callers never send a token. The server uses GITHUB_TOKEN. Redis is both an HTTP cache and a progress store for attribution. Full middleware walk is on Cache and limiter. Achievements and star lists are HTML scrapes: HTML scrapes.

flowchart TD
  Client["CodeTrace, portfolio, curl, README SVG"]
  MW["CacheRateLimitMiddleware platform=github"]
  Docs["GET / and /playground"]
  Card["Canonical card routes"]
  Extra["languages repos stars PRs breakdown"]
  Attr["attribution.analyze_repo_contribution"]
  REST["api.github.com REST"]
  GQL["api.github.com GraphQL"]
  Redis[("Redis or Upstash REST")]

  Client --> MW
  MW --> Docs
  MW --> Card
  MW --> Extra
  Extra --> Attr
  Card --> REST
  Card --> GQL
  Extra --> REST
  Extra --> GQL
  Attr --> REST
  Attr --> Redis
  MW --> Redis

GitHub’s language API is whole-tree bytes. A fork of Linux looks like C. A popular repo with one commit from you looks like the maintainers’ stack. Default is attributed=true: count commits author={username}, skip vendored paths, and if the walk cannot finish, return a thin own-commit mix rather than a pretty lie.

That walk is hundreds of REST calls. Vercel Hobby is about 10s. Redis stores finished repos keyed by pushed_at. The next request spends a few seconds of new work. Coverage is public: coverage, partial, status, cache_enabled.

docs, profile, heatmap, badges, summary, analytics, PRs. routes/api.py is imported as a comment only. Duplicate handlers there are leftovers and some of them forget await. Do not call them.

middleware/rate_limiter.py is a dead SlowAPI limiter (15/minute, 700/day). Nothing mounts it. Live policy is core/middleware.py.

Local bind default is port 8989. RUNNING.md pins 8001.