The store was charging for one telescope

Three telescopes, $101.96 each, plus $26.97 shipping. The cart wanted $128.93.

I was running release 2.1.3 of the OpenTelemetry Demo. Its frontend added each product's price once, regardless of quantity. The right total was $332.85. There's an upstream report of this in issue 1526. I kept 3.1.0 as a separate clean reference.

A checkout script could still place the order. I wanted to see if a browser agent would check the cart total before paying.

Three telescopes in both carts. The totals are $128.93 and $332.85.

Give the shopper permission to leave

I built Walkout with a purchase goal and a short persona card for each shopper. Playwright returns a browser observation, the model chooses an action, and the runner carries it out. Then it sends back the next observation with the earlier steps.

Muse Spark 1.3 runs that loop. It can click, select, type, press a key, scroll, or give up. The shop has to show an order confirmation before I count a purchase, even if the model says it's done.

For the recording, I told the shoppers to buy three telescopes and check the quantity, price, shipping, and total. If the arithmetic didn’t agree, they were supposed to leave. The counts below show how the agents did on that task.

Dev doesn't get a mouse

The profiles needed different browser conditions or I would just have twelve copies of the same test. Rosa gets a budget phone profile with slower CPU and network settings. Margaret has enlarged text, Lucía an es-MX locale. Dev uses keys. Amara gets the accessibility tree.

Dev sees a plain screenshot and can use keyboard actions. Giving him numbered targets he could click would let him get around a focus problem without ever seeing it. Amara's tree references give her a different view of the page again.

I can repeat those configurations. They don't reproduce every phone or a person's full screen reader session, and the recordings let me see exactly which inputs each simulated shopper had.

Ten shoppers complained about the same total

One shopper bought before the patch. Eleven left, and ten cited the quantity calculation. I saved each shopper’s explanation next to the browser steps that led to it.

Lucía explained in Spanish that the total charged for one telescope plus shipping. In her replay I can see what the browser showed her before she gave up.

Lucía’s replay shows where she left and what she saw.

My runner had its own bugs

An early smoke run got stuck scrolling between the total and the checkout form for twelve steps. The shopper had already read the cart. I wasn't returning its notes with the history, though, so that useful observation kept dropping out of the next decision. Passing the notes back fixed the loop.

The native quantity dropdown caused another problem. Its open menu didn't appear in the headless screenshot. Repeated taps weren't going to make that screenshot any more informative. I added a select action so the shopper could choose the quantity.

Those were my bugs. Having the saved steps helped me find them before blaming the storefront.

What the $8.34 actually paid for

The October 28 model ledger has 1,862 calls. They added up to $8.33742325, or $8.34. I tracked model calls, so the figure leaves out server costs and my time.

The November 4 takes and work added $6.11. By the end of that recording I’d spent $14.45 on model calls across the project.

The repair was two multiplications

Spark reads the saved reasons for leaving and picks source files. Muse Wonder 1.0 proposes replacements, and another Wonder call reviews them.

The change was in CartItems.tsx. It multiplied by quantity in the whole-unit sum and in the nanos sum. The existing carry from nanos into units stayed in place.

Another call to the same model approved the patch. Release 2.1.3’s frontend had no test command. Its image build runs type checking and lint, and Walkout won’t serve a patch if that build fails. A successful build still couldn’t tell me what three telescopes cost.

I emptied the cart and tried again

The rebuilt frontend replaced the release image on the same port. I emptied the cart, added three telescopes and checked the total before ordering. Then I sent the same twelve profiles through it again.

Nine bought this time. Rosa couldn't open the full cart on her phone profile. Dev couldn't reach the quantity or Add To Cart controls with the keyboard. Lucía checked the peso arithmetic and found a one-cent difference.

I left Rosa, Dev and Lucía red on the board so I could find the purchase attempts that still needed work. The profiles stayed the same across the two rounds. Their model-chosen actions can vary from one run to the next.

The November 4 scoreboard shows one purchase before and nine after.

November 4 recording

Before

After

Completed purchases

1 of 12

9 of 12

Shoppers who left

11

3

Try the checkout checks without model calls

The current repo has browser checks for the original, clean and patched shops. They check the three-telescope total without a model API. The November 11 run has its own saved results and spend ledger. The screenshots and results above are from November 4.