The robots were everywhere this week.
Reviewing our code, testing our screens, and running up our search bill. Entry 12 disclosed how this shop works with AI; this week was a live demonstration from three directions at once — including one where the robots were the problem.
The milestone first: On Account checkout shipped. Credit customers can now place web orders billed to their terms — the feature this log first mentioned as a design document back in August. The build was a long chain of named decisions — who qualifies, how available credit is computed, how open orders count against exposure, what happens when eligibility changes mid-checkout — and the answers moved under us more than once: a ruling from accounts receivable superseded an earlier answer from another colleague, and rather than silently updating, the conflict got flagged and resolved on the record. A staging-versus-live data discrepancy on cash-terms accounts was chased down with direct SQL before anything was trusted. Then the part regular readers will recognize from entry 12: the AI-drafted code went through two independent review passes by a separate agent before any human tested it — and the reviews caught real defects, including a path that would have bypassed the disabled-provider check, a cross-account access hole, and stale open-order data in the credit math. Fixed, then a dozen manual tests on staging, happy path through hostile edges. All passed. The gate working exactly as advertised, two weeks after we published how the gate works.
Second robot: the one we built to squint. The stakeholder verdict on parts of the portal was blunt — “a mobile nightmare” — and the response was to make the problem measurable before making it better: a sweep tool that loads every page at fourteen widths in two browser engines and screenshots what breaks. The device list came from our real analytics, which produced the week’s best minor finding: our desktop traffic is mostly bots, so the resolutions actual humans use are phones and tablets. Then module-by-module fixes. Tables now stack into labeled blocks on narrow screens — triggered by whether the table actually fits its box, not by a hardcoded breakpoint, so the fix doesn’t rot as content changes. A Safari-only thirteen-pixel overflow was traced to one dropdown’s hidden overflow text. And the sweep paid for itself outright when it surfaced a real commerce bug: a missing attribute had some “Add” buttons silently redirecting to the product page instead of adding to the cart. Honest status: portal, homepage, and search surfaces done; the product catalog — expected to be the hardest — deliberately saved for last and still unmeasured.
Third robot: the one running up the bill. Search usage blew past the annual contract allowance — 6.89 million requests against 5.7 million — and the investigation traced the overage to the legacy site: crawlers loading pages, combined with the search library’s default behavior of firing a billable empty query on every single page load, before any human touches a keyboard. Every page view, human or bot, cost a request. Reproduced live in the browser, written up with evidence for the vendor tickets. Then the recon turned inward: the new platform’s shop pages waste a few requests of their own on load, and the fixes are designed — cache the default results server-side after each sync, and reorder the fetches so personalization data arrives before the first search fires. The lesson travels: know what your search client does on page load, because you’re paying for it whether or not anyone searches.
And the closer: loose-ends day. A deliberate audit of old pinned threads found two things the process had dropped. A database uniqueness index that an old thread flagged as missing was confirmed missing — guarded only by application code, contrary to our own standards — and shipped as a self-deduplicating migration. And a field that two features depended on turned out to have never been built at all: it fell into the gap when ownership moved between work streams and the move was never communicated to the person building the other side. Built and deployed the same day it was discovered, but the lesson earned its place in the collection: handoffs between work streams need the same discipline as handoffs between people. Entry 10 called every rule scar tissue; this one’s fresh.
Robots reviewing, robots measuring, robots billing — and humans deciding, all week, which of their findings to trust. That division of labor is the whole system now. It held.