A map of every vintage store in LA went live in nine ticketsA map of every vintage store in LA went live in nine tickets
August 14, 2026 · Nine tickets put LA's vintage stores on the site, and a test fixture that had been quietly pushing to my phone twenty times a day finally got caught.Nine tasks put a map of vintage shops on the site, and a test that had been messaging my phone twenty times a day was finally noticed.
last updated
From an empty table to a page on this site
Nine tickets, and the map is live. The constraint that shapes it is that place reviews are capped at five with no pagination and have been since 2015, so a shop's character has to come from its own website rather than from what people said about it.
The two rules the published half is arranged around are both about what the output is allowed to claim. Only a complete pass may conclude that something has gone; and the coverage report says which districts it has not looked at, rather than letting an unvisited area read as an empty one. The page is a full rebuild from a committed snapshot, with the mapping library vendored rather than loaded from a CDN.
My test suite had been texting me
Three separate bugs in one notification path, and the one that had been running longest is the one worth writing down. The board's test suite drives the real webhook handler with a fixture issue assigned to me. That path sends its notification fire-and-forget, so nothing awaited it and nothing injected a fake. It built a real client from the real credentials and sent.
Every run of that suite put a message on my phone. Every dispatched job runs the suite as its first step, so it was roughly twenty a day, and every test passed while it happened.
The fix is a mechanism rather than a note in a file: the sender now refuses to reach the real bot from a test that injected no double. Two other paths in the same channel were pure noise: a push for every ticket merely being created, and a debounce that asked when something happened rather than when we last spoke.
Two repos moved onto the subscription, and disagreed on purpose
Model calls now bill the subscription rather than a metered key. The two projects made opposite defaults for the same reason: a fixed per-call overhead is noise on one weekly request and dominates three hundred small ones. Neither default is wrong, and reading either alone makes the other look like a mistake.
The account ran dry, and only one of the two things that noticed said so
The classification pass stopped at the fiftieth store, which was loud and obvious. The duplicate check failed open at the same moment, filed everything unchecked, and said so only in a log nobody reads, which is how two tickets for one action got filed thirty-seven minutes apart.
A map of vintage shops went live
Nine tasks, and the map is on the site. The awkward constraint is that only five reviews per shop are available, so what a shop is actually like has to come from its own website instead.
Two rules shape what the page is allowed to say. It only concludes a shop has gone if it did a complete pass; and it states which neighbourhoods it has not looked at yet, rather than letting an unchecked area look empty.
My own tests had been messaging me
The worst of three problems in the notification system: the test suite runs the real notification code with a made-up task assigned to me, and nothing was stopping it sending for real. It used the real credentials and sent a real message.
Every run of those tests messaged my phone. Every job runs the tests as its first step, so that was around twenty a day, and every test passed the whole time.
The fix makes it impossible rather than discouraged: the sender now refuses to reach the real service from a test. Two other notifications in the same channel were pure noise and were switched off.
The account ran out of credit
One part stopped loudly and obviously. Another quietly carried on without its safety check and only mentioned it in a log nobody reads, which is how two tasks for the same job got created half an hour apart.