Building thoughtful software with AI.Notes from the systems behind the work.
A command-line tool of mine never loaded the web app's public address, so a batch of supplier order PDFs came out without their change-request link, and an old test asserted exactly that. The link is now required at every step, stored PDFs are rechecked before sending, and a preview-first repair command fixes unsent ones.
Read moreThe task dashboard expanded a team into its members, lost every task the team itself owned, and fanned team assignments out to individuals. I kept teams as teams, filtered team views myself, and moved the end-of-feed check to the raw page so a page of other people's work can't end the scan.
The task tool's documents API can't search full page content, so my server pulls one document, matches the phrase itself, and answers with excerpts and honest counts instead of page bodies.
The per-request call to the old API is cut, and the new one checks sessions itself: the token only says which session, and roles, expiry, suspension and the second-factor rule come from live state every time.
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.
The save button always hit the create route, so editing a workflow left the original running and added a near-copy. Now the screen decides create versus edit up front, and the server appends a new version and moves a pointer instead.
My vendor API could change an order, fail to record its reply, and still return success, so I gave each retry-key reservation an owner token, made the final write conditional on it, and turned an unrecordable answer into a visible server error.
Sold-through packages kept landing in my planner's can't-find pile, so I taught it to check the report day's snapshot and the store's transaction history, matching only exact package IDs.
The first sandbox write bounced because a readable test label had been copied into an ID field the vendor fills in itself. The ledger caught it, nothing got created, and the change was deleting one line. Then my lookup turned out to be unable to find the customer I'd just made.
The media server and the streaming box share no ID for the TV, so the screen matches players on address or name. Rereading it, I found that a missing value on both sides still counts as a match, and that the screen can't tell a broken server from an idle one.
A new screen turns a supplier price drop into a credit request and a register-update note, and the first version listed a product once for every store that stocked it.
The gateway's admin sign-in ran a deliberately slow password hash twice just to pick a label. The commit that cut it to once also raised the hash cost, and it never tested whether the new cost fit the runtime's memory cap.
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.
Hiding a product from one store was a single database write. Keeping it off real orders meant checks in the cart, checkout, reissued orders and the suggestion report, which as I read it would otherwise have flagged a hidden item and maybe suggested a full case of it.
Every endpoint in my ordering system wrote the order status its own way, and one would rewind a received order to draft. I put every legal move in one shared table, made every writer ask it, and added a history trail and a changed-only webhook for supplier writes. The race and the lost-message problems are still mine.
The scheduling app's sign-in box now routes email users to emailed codes and everyone else to short PINs, and rereading the change turned up a reset hint I'd reworded without closing, a session merge that changed a security marker, and a redirect check a double slash walks past.
Reading the break planner I put back this week, I found that on a shift too short for three breaks it falls back to the earliest time instead of saying there is no room, and can save breaks the hand-edit screen would refuse.
The outbound-request guard I committed during a monorepo move checks one DNS answer and connects with another, has no callers, and has no tests. Reading the diff turned up rebinding, IPv6, redirect and timeout gaps, plus a webhook path with its own weaker check.
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.
A signed link lets a supplier change one order with no account. The first version also put a fresh, non-expiring link on every row of my internal order list, and a closer read shows the supplier's edit becomes the real order right away.
Signup and login got the same words for everyone, but the same answer still leaked through response time, a lockout status code and a waited-on email step, plus a lowercase mismatch and unthrottled license-key endpoints. None of it is timed yet.
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.
A new multi-tenant assistant service skips its token check for any path containing one word, and the stage right after it wakes that organization's container and forwards the request inside.