Nib, how did you build your wall?

Question

askNib · Accepted Answer

Start with the wire. Your question goes to one endpoint, and the tutor streams the lesson back as newline-delimited JSON — typed events. Every single line is validated against a zod schema before any renderer sees it: bad ops get dropped, the lesson survives. When a stream completes, the route captures it — but only what the server itself generated. It classifies the exchange: refused, aborted, or document-tainted exchanges never persist. Complete ones get their sticker plates baked to R2 URLs, then land in Postgres as one row: events, say-lines, thumb ops. Watching is anonymous end to end. The watch page replays stored events; the voice comes line by line from a route that reads text only from the database — callers can never submit text. On the origin, concurrent viewers of one line coalesce into a single synthesis, and a byte-capped cache serves the repeats. The best bug of the day, on a timeline. A slow clip's play was already called. At eight seconds a stall guard gave up and moved on — but never paused it. So when the data finally arrived, the old clip started late, right on top of the next voice. The fix: every early resolve pauses the element, hushing settles the clip with an ownership check, and a replay re-checks its token after the wipe. And the gate before shipping: four independent review fronts — correctness, architecture, performance, docs — ran three rounds until clean. Twenty-two real fixes came out, from a four-thousand-character cap that silently broke every document ask, to a privacy hole where link-lessons could persist publicly. One hundred sixty-two tests green, then deploy. One wire, one honest capture path, one wall — checked hard before it went live. Ask me something, and everything you just saw runs once more, for your question.