Skip to main content

How it works

teach (before edits) → implement → hook deny?
                ↓
          know-code-teach (unless skipped) → quiz.json → know-code ask
                ↓
          grade → know-code pass → know-code commit -m "…" → push → CI

Layers​

  1. Skills — the host agent explains the design (know-code-teach) and later quizzes you (know-code).
  2. Browser quiz — know-code ask opens a local form with one textarea per answer.
  3. Receipt — know-code pass writes .know-code/gate.json keyed to a content hash of the index (empty tree → staged tree).
  4. Git hooks — pre-commit / pre-push run know-code check.
  5. Agent hooks — Claude / Cursor / Codex deny gated shell commands until the receipt matches.
  6. CI — verifies Know-Code-Verified: <hash> on the commit (HEAD trailer first).

Hash model​

The hash is SHA-256 of the patch from Git’s empty tree to the current index. Using a fixed floor (not merge-base) keeps the hash stable when origin/main catches up to HEAD, so you are not forced to re-quiz after a routine sync.

Message-only amends that do not change the tree keep the same hash.