Building thoughtful software with AI.Notes from the systems behind the work.
An incremental sync that reset its coverage on every shape change, and starved its own connection pool, is rebuilt around batched transactions, a serial write queue, and a separate refresh cursor.
Read moreCompilers kept getting reshuffled inside one validation run to protect their deadlines, but the bigger risk was contention between local, CI and release runs that could all start on the same workstation. A loopback port now gives each run the machine to itself, the kernel frees it on a crash, and queue time is recorded apart from run time.
The lease guarding my long supplier setup job only checked the owner's name, judged expiry by each worker's clock, and never renewed during slow phases. Now every claim bumps a fencing token, every write checks it against the database's clock, the phase-advancing writes also check the phase they expect to leave, and a heartbeat keeps slow work alive.
A supplier answer to a generic request used to live on a shared slot that every order for that supplier touched. I moved it onto a choice row owned by one version of one order, and approval now only goes through if the proposal still matches what the buyer saw.
A deleted-and-recreated cart line looked identical to my snapshot check, so reviewed checkout now trusts a database-owned version number and a one-time receipt instead of a comparison.
The gateway web service and archive worker both called the same rate-limited point-of-sale API with no shared count. I put every attempt through one priority line kept in Postgres, then fixed my own spacing math, which was leaking slots every time a timer woke late.
The scheduling app now plans a whole change set on a copy of the week, checks only the result, and replays a lost save before it asks whether the board is stale. The swap that makes it worth it still only runs in a test.
A long windowed import of store discount data now skips windows that already have an exact success record, and the route turns away a second sync while one is running, with a few gaps I still have not checked.
The button that saves a delivery as a draft in the stores' inventory system checked first and wrote its record only after a slow outside call, so a second click could slip through. Now it claims the delivery in the database first, lets a partial unique index refuse the second claim, fixes a payload hash that couldn't see nested fields, and adds a narrow admin cancel for claims that get stuck.
Full-case rounding moved from the browser to the server and still read the order before its transaction, so two overlapping requests could each add the same shortfall. Locking first and reading second makes the repeat a no-op, and I list the gaps that are still open.
The new board moves tracker tickets and planning-document tasks with the same drag, but tickets get a status change with an audit row while tasks rewrite their whole JSON document, with no history and a lost-update window I can't find a guard for.
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.
The agent gave my chore app a test runner that saves the config and puts it back after every test, then wired it to an admin button on the live server. It looks clean at the end of a run, but reading the code, visitors could get test settings mid-run, two overlapping runs could delete the real config and still report green, and a killed run loses the only saved copy.
The scheduling app's rule templates moved to the server, with the list call seeding a starter whenever a company's list is empty. Empty also means an admin deleted everything, so the starter comes back, and that same read writes to shared settings.
Sending the whole chore list on every save let a missing chore mean deleted, and then three versions of the save took different swings at what happens when an older list lands last.
Moving my scheduling app's chores onto the server, I let a list-shaped request replace the whole file with no checks, no lock and no read-back, and the browser starts from an empty list when loading fails. What I found rereading the commits, and the shape I want instead.
The break filler on the stores' schedule sheet was rewritten three times in a day: from blind arithmetic, to comparing start times, to real ranges in one shared pile. Then a fill-everything button that could race itself got serialized, and grew a redo that throws out hand-fixed breaks. It still has no stop at the end of a shift.