First job was the loose ends from last time. There's now a script that makes a copy of the repo from the rebuild branch (not main) with the dependencies linked back to the main checkout, plus one line of Vite config so the dev server is happy running in it. The brief every agent gets is written into the repo as well, so the next batch starts from the same place instead of from whatever I remember to paste in.
Then the comparison check needed work before handing anything out. Up to now our test page nudged each component 20px in from the edge by hand to line up with the design system's preview, and one test even painted its own background in. That was never going to scale, because nine of the next previews had different backgrounds, padding or widths. So now the check just borrows each preview's own page styling and card width and applies it to ours, and it turns on reduced motion on both sides so the animated components can be compared standing still.
That's where it went sideways for a while. Components that had passed all day started failing, but only sometimes. The first culprit was the check itself: it wrote a temporary copy of each preview into src/, the dev server spotted a new file and reloaded our page halfway through capturing it. Moving the temp file out of the project fixed that and the failures carried on anyway. The second was more interesting. With reduced motion on, the design system sets every transition to 0.01ms with !important, inside a cascade layer. The check had its own "no transitions, ever" rule, also !important, but not in a layer, and with !important the layering flips round so the earliest layer wins and unlayered loses. So the padding we'd just added was still transitioning when it got measured. Claude tried waiting a couple of frames, then waiting for the transitions to finish, and both were flaky. Moving the rule into the first layer was one line and fixed it properly. Then Claude tightened one of the glass allowances a bit too eagerly and merged with a failing result, because the command piped through tail, which ate the exit code. It got put back on its own branch a few minutes later, but it's exactly the kind of thing the check is there to stop.
With the check behaving, four Sonnet agents took a group each (the surfaces like cards and panels, the post content bits, the editorial pieces and the site header, footer and hero) while Claude did the theme toggle. That's the first real component on the site. It's the design system's own toggle, but driven by the theme code we already had, so there's still no flash of the wrong theme on reload, and it cleared up a hydration warning that had been sitting there since the foundation.
Two of the agents came back clean. The other two each held one component back and reported it instead of working around it, which is what the brief asks them to do. One was a loader that fills a grid of tiles. The plan assumed every animation in the system stops for reduced motion, and this one doesn't, it just fills the tiles in order instead of scattered, so there's no still frame to compare against. The port was fine. The check now freezes the page clock on both sides, which made it identical and makes every other comparison a bit more predictable too.
The other was the little garden-stage mark that says whether a post is a seed, growing or mature, which came back 1/255 off on one or two pixels. After the night the check had had, it would have been easy to call that noise and move on. Instead Claude checked the compiled gradient and every element's position down to the fraction of a pixel on both pages, and they were identical, so it is just the GPU drawing the hard edge of the half-filled square slightly differently. That got a small named allowance like the glass ones, and so did the post list, which uses the same mark. Two glass buttons got the same treatment after a few runs to see how far they actually drift. (The social links also threw React warnings in their preview, because every link in it points at # and the component used the link as its key. Ours uses the label as well.)
So 50 components are in now, all matching the design system in both themes, at rest, hovered and focused, with seven of them allowed a handful of pixels off by 1 or 2 out of 255. The site still looks broken apart from the toggle, because nothing else is wired in yet. That's next, along with deciding whether code in posts keeps Shiki or uses the design system's own code block.