Shopify Canvas and Cloudflare Clef: October 2 dispatch

Shopify's new store editor lets you request design changes in ordinary language, but publishing still needs your attention. Today's other releases tackle misrouted inquiries and interrupted AI jobs. Start with a 30-minute store-copy test, then give whoever maintains your automations a short acceptance checklist before changing the systems your customers depend on.

Shopify Canvas puts store changes in a conversation

Shopify announced Canvas on October 1, with access rolling out over the coming days. Shopify runs online stores; Canvas is its new editing workspace. Sidekick, the built-in AI assistant, can change the store's underlying files while you inspect working pages together. You can also click individual elements to edit them.

For a business owner, the useful test is one awkward product-page change that normally requires outside help. Skip a full redesign. A theme is the package of layouts and styling behind your store, and Canvas has limits: third-party themes aren't supported, and themes edited there don't receive theme updates. Keep your current store while you evaluate a copy.

Follow Shopify's Canvas editing instructions on a desktop computer. In your admin, open Online Store, find a supported theme on the Themes page, then click its three-dot menu and Duplicate in Canvas. Select the product-page block you want to change. Enter the request below in Work with Sidekick and submit it. If the option hasn't reached your account, wait for access rather than rebuilding elsewhere.

Copy this request:

On this product page, make the existing delivery information easier
to find near the purchase button. Preserve its exact wording.
Do not invent delivery dates, discounts, or guarantees.
Show me the changed section so I can review it before publishing.

The part that breaks: saving and publishing are different actions. Canvas saves automatically. On an unpublished copy, Publish, followed by Publish changes, makes it visible to customers. Test the purchase button and review the copy using your AI content approval workflow before taking that step.

Cloudflare Clef separates sorting from writing

Cloudflare released Clef and Clef-flash on October 1 to make decisions from a defined set of answers. A decision model is AI that selects categories instead of composing an open-ended reply. Cloudflare's Workers AI service runs these models, which return answers with probability estimates, including which team should receive an inquiry.

That gives you a focused alternative when your current AI spends its time deciding whether a message concerns sales or an existing booking. Keep the writing step separate. A confident category is still a prediction, so don't let it promise a service appointment or dismiss a complaint. Start with the routing stage of your lead-to-customer-record build.

This connection needs a technical maintainer. Give that person a small set of past inquiries with names removed and your correct destination beside each. Ask them to test Clef's questions setting using type: "choice", the source's setting for choosing among named categories. Include a manual-review destination. Keep the results in a test sheet before connecting them to customer records.

These are proposed test cases, not measured results. Compare wrong destinations and actual usage charges with the current setup before switching. If a simple form dropdown already sorts inquiries correctly, keep it.

Pi Durable adds recovery for interrupted AI jobs

Earendil introduced experimental Pi Durable on October 1, alongside Pi 1.0. It's software building material for AI agents, meaning programs that use tools to complete tasks. It records progress so unfinished work can resume after the program stops unexpectedly. This matters when an overnight research or reporting job otherwise needs someone to start it again.

Ask your developer to test recovery on a copy of your weekly reporting routine. Have them interrupt it halfway through collecting information, restart it with the same stored progress, and show you the completed draft. The implementation uses harness.resume() to continue interrupted work; a stable requestId, the job's identifying label, prevents a repeated submission from becoming a second job.

Don't confuse that protection with a guarantee against duplicate emails. Interrupted tool actions rerun only when marked safe to repeat; others are reported as interrupted. Require a test that proves customer messages cannot be sent twice before allowing any sending step. Owners without a developer should leave this experimental release on the watchlist.

On the bench

• Canvas: test whether your existing store extras still work before considering a published change.
• Clef-flash: compare routing mistakes against Clef using the same inquiries before favoring speed.
• Pi Durable: repeat the recovery test after a developer changes the connected actions.

Get the next build: set up a guarded lead qualifier before giving AI control of customer follow-up.