We called it the Sunday run and it wrapped on Monday, which tells you something about how weekends work around here. It started with a screenshot. Shafen sent me a picture of a blank page where a billing flow used to be, and pulling that one thread unraveled the worst kind of bug: the quiet kind.
Sign-ups on unicorncommander.ai had been dead for four days. Every plan button on the storefront sent visitors to Keycloak's registration page, and registration had been deliberately disabled the same day we closed the open self-registration hole. Two correct changes, each right on its own, together killing one hundred percent of new customers at the door. Nobody logged it because nothing errored. The funnel just ended in a dead card.
The fix was one guarded line in the ops-center backend, on both rings: signup intent now rides the normal brokered sign-in page, where Apple, Google, and Microsoft auto-create the account and checkout continues, and the registration screen only gets requested if registration is actually enabled. Verified with a real headless browser click-through, not a curl and a prayer. All the legacy funnel URLs that were still linked from older bundles now redirect to the live pricing page, so Shafen's exact link lands somewhere real. For the record keepers: three commits landed this day, plus one bind-mounted line on the server that is not a commit yet, which is its own item on the list.
Every plan button on the storefront sent visitors to Keycloak's registration page, and registration had been deliberately disabled the same day we closed the open self-registration hole.
Stable took two meeting action items, both shipped to dogfood. Full screen on every call tile, done properly: a shared fullscreen engine with the iOS ladder, wired into participant tiles, screen shares, the agent stage, and the guest stage. The button already existed on two tile types. It was dead on iPhone and missing everywhere else, and every tile control was a twenty-four pixel target. They are forty-four now. And calls record audio into Meeting-Ops: a LiveKit egress captures one audio file per call, enforced by a database lease because that is the only guard that survives a backend running four workers, then a durable outbox hands it to the real upload pipeline. Speech to text, diarization, and summaries now run on actual call audio instead of browser captions. Proven end to end with a synthetic call before any human call trusted it.
Email-Ops got Friday's ask: the landing page stopped claiming things the product does not do. The shuffle button and the phantom metrics strip are gone, and Shafen's PR went in cherry-picked with his authorship intact, deployed frontend-only so the real mail on the dogfood ring was never touched.
The mysteries got names too. The "Jordan Avery" account that spooked us was our own demo presenter, created by the video rig minutes before we noticed it. The one-call-becomes-three-sessions puzzle in Meeting-Ops root-caused to each browser minting its own transcript export with its own wall clock, and the proper server-side fix is filed high on the board. Ryan Otto's demo workspace from July was simply never provisioned. Four answers, zero ghosts.
What is still open from this run, so the next session can pick it up: eyeball dogfood and then say ship flagship, close PR nine on Forgejo since the commit is already in main, scope Adam's access and mint it, approve the five triage proposals from the call with Shafen and Adam, and run one real checkout through the newly fixed funnel, because nobody has ever completed a live login to paid on the new storefront. That last one matters. A funnel is not fixed until money has been through it.
Four days of dead sign-ups, found by accident, fixed in an afternoon. The lesson is not the one line. The lesson is that two right answers can add up to a wrong system, and the only detector for that class of failure is someone actually walking the path a customer walks. We ran the products on ourselves all weekend. That is the whole method.
Real product captures — click any to enlarge.