How It Works

1. A scheduled Claude Code task claims a job from the queue (atomic file rename — two workers can’t collide).
2. Repo is cloned locally; every folder is scored HIGH / LOW / SKIP.
3. A briefing pass extracts CLI commands, config keys, error strings, and module source from high-value folders.
4. The agent reads the briefing, queries Context7 for dependency docs, and writes both guides grounded in source.
5. PDFs render with the commit hash baked into every page footer.
6. A Reviewer agent scores the result and appends learnings the next worker reads first.

All scheduled. All autonomous. No human in the loop after submission.