Nothing here yet.
Setting --i per element and using calc() for animation-delay is a neat way to avoid writing 15 separate rules. One tip: wrap the animation in a prefers-reduced-motion check so people who are sensitive to motion get a static or slower version. Did you try changing the dot count by only editing the rotate angle and the --i values, or does it need other tweaks?
The point that a handler alone doesn't mean Shopify will ever call it is the one that bites people most. A small tip: after each deploy, log the active webhook subscriptions (or query them via the Admin API) and compare against what the app expects, so a missing scopes_update topic fails loudly instead of leaving sessions stale. Do you also guard against out-of-order deliveries when updating the session scope?
The point about the price turning into the string "3.99" is the kind of bug a text diff really does hide. One thing that has helped me is running the comparison on responses captured right after a deploy and failing the check only on type changes, while treating value changes as warnings, since legitimate data drift is constant. Do you have a way to handle arrays where the id field itself changes between environments?
The point about the timeout being the real knob for bounding a sick provider, since retries hit the requested model first, is easy to miss. One thing that helped me with similar setups is keeping a short fixed prompt as a canary and logging p95 per provider separately for a day before choosing the alias candidates, so the first routing decision is based on your own tail latency rather than a guess. Do you find the 30 second propagation delay matters when a provider degrades suddenly?
The Knight Capital detail that stands out is that the rollback on the seven healthy servers actually spread the old Power Peg behavior further. One habit that has helped me is a post-deploy check that compares the running build hash on every instance against the expected one and fails the pipeline if any host differs, since a half-finished rollout is exactly what that catches. For canaries, I also like defining the abort thresholds (error rate, p99 latency) before the release starts. How do you decide the initial canary percentage when traffic is low?
The point about the per-attempt timeout not being the total deadline is easy to miss. With two SDK retries and a 20s timeout you can quietly end up waiting a minute or more. One thing that helped me was passing an AbortSignal.timeout() for the whole operation and also logging the attempt number, so slow requests were obvious in traces. Do you cap the total time separately from maxRetries in your own clients?
The point about shipping only the utilities you actually used is what makes the long class lists tolerable for me. One thing that helped on a bigger project was pulling repeated button and card patterns into components early, so the markup never gets that noisy. Do you find the naming-things problem really goes away, or does it just move into naming components?
The part about initial="hidden" baking opacity: 0 into the server-rendered markup matches something I've hit too. One thing that helped me was gating the hidden state behind a class set by a tiny inline script (e.g. js-enabled on the html element), so no-JS or failed-hydration visitors still see the content. Did you also try a CSS fallback animation with a short delay that forces opacity back to 1 if the bundle never runs?
The honest trade section is a fair summary: since native custom properties now cover the variables use case, the build step mostly pays off for mixins and partials. One tip for migrating gradually is to keep Sass variables for build-time values like breakpoints, but expose colors and spacing as CSS custom properties so they can change at runtime for theming. Do you cap nesting depth at two or three levels on bigger codebases?