Building thoughtful software with AI.Notes from the systems behind the work.
Each parallel test worker got its own database clone, but every worker ran several files in a row on it, so leftover store fixtures broke later files with misleading business-rule errors. Now every test file gets a fresh copy of a frozen template, behind a guard that only touches this run's clones.
Read moreA browser test accepted an unattended confirmation and cascaded a delete through a shared seed store in my local test database. Now every destructive test request has to prove the database is disposable and that its run owns the exact record it touches.
Deep test suites are now wired into deploys, scoped by a saved diff and hand-written impact tags, with every doubtful case widening to the full run, and the generated route checks turned out to claim more in their names than they actually verify.
Two name matchers in my ordering system let a product claim a different one whose name started the same way, and both of their tests asserted it as correct. Now names have to be equal, and the tests put the lookalike right next to the real thing.
Retained MQTT readings outlive the plugs that sent them, so my sampler, history overview and device list each needed an availability check; after patching them one at a time I moved the rule into one module that only counts online devices, rejects junk values, and sums power instead of averaging it.
The Users menu link did its own permission check while most page guards asked a helper, so the area rules moved into one table that the guards and the menu's standard links both read, and an empty permission list should still keep people out.
The group path in my ordering cart rounded every share up to a full case on its own, overshot the plan and ignored the supplier's real case rules; I took the rounding out of the cart and sent those lines through the same supplier policy as everything else, with explicit skips for what the cart can't hold.
The hidden browser window in my coding app checked the address I asked for against its allowlist and never the one a redirect dropped it on, so it could click, screenshot and certify a page nobody approved. Now the same check runs on the landing address before anything touches the page, though the outside page still gets fetched first.
Login emails can't be sent from the dev box, so I gave myself a short PIN that never expires or locks, then fenced it with repeated dev checks, a stamp only my script writes and a reset route that hides itself, and chased down a correct PIN that failed because a secret was read too early.
A lookup in my ordering system built its list of already-handled products from any row with the right kind of ID, ignoring setup, brand, queue state and whether repairs were applied. I rewrote the query within the half hour, and the stub-based tests had to start checking the question instead of the answer.
Collapsing repeated supplier-order lines on a cleaned-up display name would have folded distinct placeholder items into one line with the drop total still correct; the same day the key moved to the record ID, and the test that had agreed with the bug got rewritten.
Trimming a paste template down to the slots I order from made every row below a dropped slot climb up in the vendor spreadsheet, while my order check stayed green. The export now walks a map of the sheet itself, parked rows and blank dividers included.
Reading my voice assistant's test harness, I found a speed gate that skips any target with no data and then declares every checked target passed, so a broken engine, all-failed requests or mismatched config names all end in green.
Custom plans resolved to nothing and nothing meant unlimited, so unknown codes now fall to the free tier. Invites, reactivations, transfers and first sign-in all ask the same question, and a copied existence check would have blocked every workspace edit.
The script that keeps my client libraries on the right version errors when it finds no version line, but rewrites every look-alike line in a Maven file, and after that its own check has nothing left to complain about.
The agent built an encrypted, signed, two-copy backup system with no tests. Reading the restore path showed a loose success flag, a signature nobody checks, truncate-before-load, and comma-split CSV, plus what I want restore to do instead.
Reading back the commit that added a skill-writing agent, I found a demo switch that saves a confident fake finding, marks the run done and spends a real run slot, plus a keyless mock mode and two routes that say yes without doing anything.
The latency check in my agent eval harness read a timing nobody recorded, treated the blank as zero, and passed every run while its unit test stayed green.
The agent's new splitter cuts a store's group demand across the products my buyer marked, where a zero means yes and a blank means no. Rounding every share on its own breaks the total, the test signs off on an overshoot, and four different weight readers disagree about unlabeled products.
The live allocator in my ordering system never reads the stock ceiling its README promises, the random tests check five rules but not that one, and a ceiling-aware allocator sits unused in the next folder.
The licensing code I shipped checks how many devices a key already has, then inserts the new one in a separate step, so two activations racing for the last slot can both get in. The test activates devices one at a time and can't see it, a device racing itself gets a server error, and a key-only check passes for keys never bound anywhere. I found this by reading, not by running anything concurrent.