Reproduce clean-package user tasks and responsive production-site behavior.

Start a timer in a clean checkout. Do not use repository node_modules, dist, Python caches, or prior outputs as evidence.

  1. Build and install the NPM tarball in a temporary ESM consumer.
  2. Find Chat from the README; generate three offline tool trajectories; validate them; inspect one row.
  3. Find Embeddings from the README; validate two retrieval rows; estimate the fixture; train; inspect checkpoint status; resume; evaluate; compare; inspect; export; verify the manifest and expected tree.
  4. Typecheck and run the stable TypeScript SDK example from the installed tarball.
  5. Build wheel and sdist; install the wheel without dependencies in a fresh environment; run Python help and protocol imports.
  6. Explain which steps were offline and which production/provider/GPU actions remain gated.
  7. Record elapsed time, unclear wording, wrong assumptions, failures, and recovery steps without editing commands during the run.

Automation reproduces these mechanics and reports command counts, rows, pages, artifacts, locks, and elapsed duration. Production browser acceptance also exercises search, navigation, focus, code overflow, links, and desktop/mobile layouts; maintainers still review clarity whenever workflows change.