One-click checkout: the bug Kane CLI found

The bug Kane CLI caught

The failure landed on step two of three, after 77.2 seconds of the 123 second run. The reason line in the result file is the whole finding:

• What Kane CLI observed - the agent ran out of viable actions on the page and could not make progress toward the stated objective.
• The verdict - Buy Now leaves order submission permanently pending after a null reference exception.
• The classification - application_issue, script_error, at confidence 0.98.

That classification matters more than the red cross. An application_issue verdict says the fault is in the product, not in the test, so the failure routes to an engineer rather than to whoever maintains the suite. A test framework without that distinction sends every red build to the same triage queue.

The flow under test

Nothing about the test invites a failure. It confirms the saved address and card are displayed, presses Buy Now, and expects an order confirmation with an order number and a charge line.

---
mode: testing
url: https://my-testing-repo-main.vercel.app/shop-clone-app/dp/sku_headset?reset=true
max_steps: 40
tags: [amazon, e-commerce, checkout]
---

# ShopKart 1.3: One-click checkout

Catalog objective: buy an item with saved payment details in one click.
Key assertion: order number is shown without entering card details.

## Confirm the 1-Click settings are shown
Verify the buy box shows "📍 Home - 418 Maple Street, Austin" and "💳 Visa ending in 4242".

## Buy with one click
Click "Buy Now" and verify the page shows "Order placed, thank you!" with a heading containing "Order placed".

## Confirm the order number and charge
Verify the confirmation shows an order number starting with "112-" and text reading "Charged" with "Visa ending in 4242".

Step one passed in 44 seconds, which is the important context: the saved address and the Visa ending in 4242 rendered correctly, so this is not a missing state problem. The data was there and the submission still hung. Reproduce it with:

kane-cli testmd run tests/amazon/one-click-checkout_test.md

How the failure surfaced

The result file records the failure in place, between the step that passed before it and the step that never got a chance to run.

---
test: ../one-click-checkout_test.md
status: failed
started: 2026-09-15T09:10:11.407Z
duration_s: 123
session_id: f52c8b6e-729e-49e0-9e1d-b7113c294c48
---

# ShopKart 1.3: One-click checkout - Result

## Confirm the 1-Click settings are shown ✓ passed (44s)
md5: 8e161c5cbee8dc6c1cdc4183b0f4d288
Verify the buy box shows "📍 Home - 418 Maple Street, Austin" and "💳 Visa ending in 4242".

## Buy with one click ✗ failed (77.2s)
md5: d0e061bc216ddffdbba5c975d84cc9aa
Reason: AP determined agent is stuck - no viable actions remain - bug verdict: Buy Now leaves order submission permanently pending after null-reference exception [application_issue/script_error, confidence 0.98]
Click "Buy Now" and verify the page shows "Order placed, thank you!" with a heading containing "Order placed".

## Confirm the order number and charge ✓ passed ( - )
md5: 1ff34c7274c2c8fd6bf9cea9b8f49708
Verify the confirmation shows an order number starting with "112-" and text reading "Charged" with "Visa ending in 4242".

Step three shows no duration. It was never reached, because step two never produced the confirmation page it needed. Reading the file top to bottom gives the sequence rather than a single aggregate status, which is what makes a hung submission distinguishable from a wrong order total.

Acting on the verdict

A failing run like this one is a finished bug report. It carries the reproduction steps in plain English, the point of failure, the observed behaviour, a classification and a confidence score, plus a session id that ties it to the recorded trace.

The practical move is to keep the test in the suite unchanged. When the null reference is fixed, the same file passes, and the fix is verified by the artifact that found the defect. More on working this way in evidence based QA for agentic test runs, and agent mode with NDJSON output covers parsing these verdicts in a pipeline.