Every project page now draws its own diagram and screenshotsEvery project page now draws its own diagram and screenshots
August 16, 2026 · The site's project pages now build their own diagrams and screenshots, and the tool that tells me who I owe a reply had been trusting the wrong clock.My projects now draw their own diagrams for the site, and the tool that tells me who I owe a reply had been sorting messages the wrong way.
last updated
Project pages that draw themselves
Every repo now owns a docs/ directory (one Mermaid flowchart and a JSON file describing its features), and a pipeline turns that into the project page you can read here. The diagram renders through headless Chrome onto this site's own colour tokens, so it sits on the same palette as the page around it in both themes rather than Mermaid's defaults. Two runs produce a byte-identical file or something is wrong.
Screenshots are the harder half. Terminal output and chat messages are stored as text and rendered natively rather than photographed, so they stay legible and searchable. A privacy gate runs over every entry before any capture happens and a violation aborts the whole run; a credential shape in captured output is a hard error with no override, because this runs from a commit hook and commits what it produces.
The nicest problem was the hero image. The best picture of the automation system is its board of open tickets, which sits behind a password and lists real ticket titles. So the capture runs against a fixture server that boots the real router and the real page over eight invented tickets in a sealed temporary database. The picture is of the product, not a mock-up of it, and it proved itself the same afternoon, when I added a project filter to that board and the screenshot on the page was retaken without my asking.
The last piece was the trigger, and it is the part I was most careful with. On 2026-08-08 a publish commit re-fired the shared commit hook in a closed cycle and ran 119 commits over ten hours before luck stopped it. This puts a second pipeline through that same hook, in eight repos. So: a fifteen-minute debounce rather than the blog's sixty seconds, both cycle guards, and a per-repo circuit breaker that counts sessions which started rather than sessions which finished, because a repo whose runs all die mid-flight is exactly the shape of a wedge. There is also now a lock over the website checkout. The blog had been its only writer, so a second one landing on main mid-session would have made the blog's commit a sibling rather than a descendant, and the next run would have reset over it: nothing erroring, nothing corrupting, a post just vanishing behind a line reading skipped.
ROWID is an identity, not a clock
The tool that tells me who I owe a reply reads the Messages database, where every message carries a row number. It had been treating that number as the order things were sent. It isn't. It is the order rows landed in this copy of the file, and on my real database 50,958 of 102,047 adjacent pairs are out of date order.
So a reply I sent at 17:49:53 could sit at a lower row number than the message it answered at 17:49:43, and the thread came back as one I still owed. Measured against the real data at a thirty-day window: 3 threads of 66 falsely unanswered, and 3 more that I genuinely owed being hidden. The second direction is the worse one, because a missing thread looks exactly like nothing.
The assumption was in four places, including the snooze cutoffs, which are stored as row numbers and were compared with a plain maximum: the same bug in miniature. Everything compares decoded dates now. The test fixture could never have caught this: every conversation in it was written with row numbers ascending in step with their dates, so the suite agreed with the assumption instead of testing it. Two conversations now contradict it, and a new test re-derives every thread with an independent date sort and diffs both ways.
Projects that document themselves
Each project on this site has a page explaining what it is, and until now I wrote those by hand. From today each project keeps its own description and its own diagram next to its code, and a background job turns that into the page you read here. It proved itself the same afternoon: I changed one of those screens, and its picture was retaken without my asking.
Taking the pictures needed care. Some of these screens are private and show real names and real work, so nothing is photographed until a check has been over every item on the list, and anything resembling a password stops the whole run. For the one genuinely private screen, the picture is of the real thing running against eight invented tasks.
I was careful about what decides when this runs. A week ago a similar piece of automation got into a loop with itself and made a hundred and nineteen commits over ten hours before luck stopped it. This one waits, checks whether anything really changed, and switches itself off in any project where it starts failing.
The wrong kind of order
I have a tool that reads my messages and tells me who I still owe a reply to. It sorted them by the number the database gives each message, assuming a bigger number meant a later one. It doesn't. That number is nearer the order things were filed than the order they were sent, and on my real data about half of all neighbouring pairs are the wrong way round.
So a reply I had already sent could look older than the message it answered, and I would be told to reply again. Worse the other way: conversations I genuinely owed were hidden, and a missing one looks exactly like nothing being wrong. Everything now sorts by when a message was actually sent.
The tests never caught it, because the fake data was built with the numbers and the times conveniently agreeing. Two of those fake conversations now disagree on purpose.