Cold open

Facilitator material. The first five minutes of Module 1: the destination before the theory. One prompt becomes a green test, the shop's layout drifts and breaks the suite, and the suite heals itself. No slides.

It may not flake. Rehearse it until it is boring. If any step does not show its expected output, do not debug it on stage: switch to the recorded backup, which you keep open in a browser tab.

Before the session (T-15 minutes)

On your own machine, in your clone of the repository, with the local shop running:

uv run --no-sync python -m shop reset
git status                      # clean working tree
uv run --no-sync python setup-check/check.py

setup-check must show Healing endpoint as passed: the cold open heals with your own HEAL_* settings in .env.

Rehearsed it in this clone already? Then results/heal/history.sqlite remembers the repairs, and the live run reuses them without asking the model: faster and deterministic, and the console says "proactively replaced known-broken locator ... (from history)". Keep it for a safe run on stage, or rm -rf results/heal for a fresh heal.

Open three windows side by side: a terminal, your agent (Claude Code) in the repository root, and the shop in a browser at http://localhost:9090/products. Open the recorded backup in a browser tab and leave it there.

The five minutes

#	Say	Do	Expected
1	"Here is a user story criterion. Let's hand it to an agent."	Give Claude Code the prompt below.	The agent reads the spec and the conventions, adds a keyword to a resource and a test in tests/ui/cold_open.robot, runs it: 1 test, 1 passed.
2	"Now the shop gets a redesign overnight."	uv run --no-sync python -m shop preset stage4, then reload the shop in the browser.	The page looks the same to a person.
3	"And the suite?"	uv run robotcode robot --exclude broken	12 tests, 8 passed, 4 failed: card prices, audio filter, order total and successful order. They are all locator failures. Your new test still passes: it uses the stable contract.
4	"Same suite, unchanged, with a healing listener attached."	uv run robotcode -p heal robot --exclude broken	12 tests, 12 passed. The console reports each heal in a heal: line.
5	"Every heal is a proposal, not a silent edit."	Open results/heal/heal_report.html; then git status -- tests resources.	About five heals, each with the old and new locator. git status shows only your new test and keyword: the heals changed no file.
6	"Today you build every piece of this yourself, rung by rung."	Back to the slides: the ladder.

The prompt for step 1:

Write a Robot Framework test for criterion WEB-004_AC-2 of this repository, following docs/conventions.md. Put the test in tests/ui/cold_open.robot and any keyword it needs in resources/, then run the test.

After the session

uv run --no-sync python -m shop reset
git checkout -- resources && rm -f tests/ui/cold_open.robot

Why these steps

• stage4 drifts every fragile locator kind at once and switches on no planted defect. Everything the drift breaks can heal. Module 8 later adds the defects, and shows that healing does not hide them.
• WEB-004_AC-2 is outside every lab slice, so the cold open solves nobody's exercise.
• The heal count varies between models and runs. The pass count does not. Say "about five heals".