Real-user page timings roll up per page on the ops view, one line per page over budget #10580

Page timings that browsers report now roll up per page, and /system and the terminal show one line for each page over budget. bun x tsc --noEmit, bun scripts/gate.ts and bun scripts/design-lint.ts came back clean. The new test file passed, along with test/loop-stall-by-route.test.ts and test/system-parts.test.ts. I ran no other tests; the forge runs the full check on this push. I never ran the app, so no real /system page has rendered the new section.

What changed

• Roll-up: slowPages in src/perf/watch.ts reads page_vital for the last hour and works out p95 time to first byte ("server") and p95 largest paint ("first paint") for each page pattern. It uses the budget I added in src/perf/budgets.ts: server 150 ms, first paint 1 s. A page needs at least 10 reports before it is judged.
• Slowest phase: server when time to first byte is most of the wait. Otherwise render, meaning the time from first byte to first paint.
• Ops view: /system gets a "Pages over budget, last hour" section, one line per page, worst first. When every page is fast, nothing renders: no heading and no empty message.
• Terminal: src/tui/screens/slowpages.ts adds a "Pages over budget" palette entry for admin people, one line per page with the same words as /system. The entry is only offered when a page is over budget. The screen itself is not covered by the test, only its line function.
• Tests: test/slow-pages.test.ts covers the quiet case and the one-line-per-page case. It also covers the minimum-reports and last-hour cutoffs, who can see the terminal lines, and escaping of a route name.
• Docs and push: docs/features/slow-pages.md. The pre-push hook demanded a "Try:" commit, so the second commit is an empty one that only carries that line.

Left out

• No finding or issue is filed from real-user timings. This is a view only; the existing watcher still works from server-side samples.
• "Server" is the browser's time to first byte, so it includes network time. The beacon sends nothing finer.
• "First paint" is the largest contentful paint, the only paint the beacon reports.
• The window (1 hour) and the 10-report minimum are constants in src/perf/watch.ts, not settings.

Try it: as the owner, open /system. The section appears once a route has 10 or more reports in the last hour with p95 server time over 150 ms or p95 first paint over 1 s. In the terminal, the "Pages over budget" entry appears in the palette in the same case.

🤖 Generated with Claude Code

Closes #10536

Page timings that browsers report now roll up per page, and /system and the terminal show one line for each page over budget. bun x tsc --noEmit, bun scripts/gate.ts and bun scripts/design-lint.ts came back clean. The new test file passed, along with test/loop-stall-by-route.test.ts and test/system-parts.test.ts. I ran no other tests; the forge runs the full check on this push. I never ran the app, so no real /system page has rendered the new section.

What changed

• Roll-up: slowPages in src/perf/watch.ts reads page_vital for the last hour and works out p95 time to first byte ("server") and p95 largest paint ("first paint") for each page pattern. It uses the budget I added in src/perf/budgets.ts: server 150 ms, first paint 1 s. A page needs at least 10 reports before it is judged.
• Slowest phase: server when time to first byte is most of the wait. Otherwise render, meaning the time from first byte to first paint.
• Ops view: /system gets a "Pages over budget, last hour" section, one line per page, worst first. When every page is fast, nothing renders: no heading and no empty message.
• Terminal: src/tui/screens/slowpages.ts adds a "Pages over budget" palette entry for admin people, one line per page with the same words as /system. The entry is only offered when a page is over budget. The screen itself is not covered by the test, only its line function.
• Tests: test/slow-pages.test.ts covers the quiet case and the one-line-per-page case. It also covers the minimum-reports and last-hour cutoffs, who can see the terminal lines, and escaping of a route name.
• Docs and push: docs/features/slow-pages.md. The pre-push hook demanded a "Try:" commit, so the second commit is an empty one that only carries that line.

Left out

• No finding or issue is filed from real-user timings. This is a view only; the existing watcher still works from server-side samples.
• "Server" is the browser's time to first byte, so it includes network time. The beacon sends nothing finer.
• "First paint" is the largest contentful paint, the only paint the beacon reports.
• The window (1 hour) and the 10-report minimum are constants in src/perf/watch.ts, not settings.

Try it: as the owner, open /system. The section appears once a route has 10 or more reports in the last hour with p95 server time over 150 ms or p95 first paint over 1 s. In the terminal, the "Pages over budget" entry appears in the palette in the same case.

🤖 Generated with Claude Code

Real-user page timings now roll up per page, and /system and the terminal each show one line for every page over budget. A fast forge shows nothing. All three of juniper's requested changes and the smaller notes are done and pushed.

I ran bun scripts/gate.ts, tsc --noEmit and design-lint; all pass. The new and changed slow-pages tests pass, and so do the existing system tests. I did not run the full suite. I did not check the preview after the push.

Changes

1. Terminal tests. New test/slow-pages-tui.test.ts covers three things:

- The palette entry is not offered on a fast forge, and not offered to a non-admin or an anonymous session. - It is offered to an admin once a page is slow. - entry.run pushes slowPagesScreen, and the screen draws the slow page's line and not the fast one.

1. Layering. dayOf now lives in src/perf/day.ts. src/perf/watch.ts imports it from there, and src/web/measure.ts re-exports it, so existing callers are unchanged. src/perf/views_store.ts still imports dayOf from web/measure.ts; I left it alone.
2. Request-path cost. slowPages no longer reads the whole hour into JS:

- SQL counts reports per route and keeps only routes with at least 10, busiest 50 first. - It then reads the newest 500 reports of each route to compute p95. - /system also keeps the result for 5 seconds, per database.

1. Live updates. I did not wire this section into the live blocks. The doc now says it changes on reload.
2. Plain words. The line now ends "slowest part: the server" or "slowest part: the browser drawing it", on the page and in the terminal. The doc uses the same words.

Left out

• The group-by still scans the hour's page_vital rows by day, because no index covers route. It is bounded by the daily report cap, and I added no migration for it.
• No test covers the 50-route or 500-report caps.

Try it Open /system as the owner after real visitors have reported slow pages. Or open "Pages over budget" in the terminal palette.