Files
PaddyandClaude Opus 5.5 b22e89001a
Test and deploy / test-and-deploy (pull_request) Successful in 2m25s
Run tests and deploy in one Gitea Actions job
Two jobs meant checking out, setting up Node and running npm ci (and in the
frontend the build) twice, about two minutes each, because every job starts
in a fresh container. Now one job tests and, for pushes to master only,
deploys. The push-to-master condition is spelled out on every deploy step
instead of an env flag compared to 'true', which would skip the deploy
silently and stay green if the boolean were rendered differently. A failed
test still stops the deploy (an if: without a status function implies
success()).

Docs: the one-job layout, and that Jenkins deployed as root - files it left
in an app folder must be chowned to the subscription user before the first
Gitea deploy (hit on the API).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-08 22:51:26 +01:00

6.9 KiB

Gotchas

Things that turned out differently than expected, so the same wrong assumption isn't made twice. Collected during the rewrite (September 2026); the progress log in rewrite-plan.md has the context for each.

The API

  • PUT replaces the whole event. Missing fields fall back to status: 'PUBLIC' and repeatFrequency: ''. Always send the full EventPayload, even to change one field.
  • "0 rows were updated" comes back as HTTP 200. api.updateEvent turns that body into an error.
  • The API returns raw DB values. wholeDay is 0/1 (observed), and a deleted event's latest version is a row with every field null except status: 'DELETED'. parseApiEvent handles both.
  • GET /admin/me answers 200 for accounts without the calendar permission. Only apps says so.
  • The admin feed ignores includes_calendars; the public/iCal feeds follow it. So an event can appear in a calendar's subscription although it belongs to another calendar.
  • PRIVATE is in the status enum but nothing uses it. The app shows it read-only and keeps it.
  • The iCal times depend on the API process's time zone (formatted with local getHours(), labelled Europe/Berlin). Inferred from the code; not checked on the server.
  • Only the admin DB pool reads DB_PORT. A local calendar test database has to listen on 3306.

Time

  • Historical DST. Until 1995 German summer time ended in September, so 1994-10-02T22:00Z is 23:00 on 2 October in Berlin, not midnight on the 3rd. The date code gets it right; hand-written test data easily doesn't.
  • DST edges in forms. A nonexistent spring time (02:30) becomes 03:30 CEST; an ambiguous autumn time resolves to the second occurrence (CET). Pinned in lib/dates.test.ts.

Next.js / hosting

  • Standalone output doesn't serve .next/static by itself. Without the copy, CSS and JS are 404s. The postbuild script copies it into .next/standalone/.next/static; there is no public/ folder to copy (the favicon lives in app/). Verified by running .next/standalone/server.js.
  • NEXT_PUBLIC_* values are inlined at build time. They must be set when npm run build runs, not only when the server starts; lib/env.ts fails the build if they're missing.
  • The build prerenders / by calling the API. If the build machine can't reach it, the build still succeeds and the events section appears after the first revalidation (≤ 5 min).
  • With revalidate, a short API outage doesn't blank the landing page's events in a running server - Next keeps serving the cached fetch.
  • The dev server must run on port 4200. That's the origin the API and the admin app trust; another port breaks sign-out and the return redirect.
  • Dev runs React strict mode, so every calendar is fetched twice on load in development.
  • A leftover Angular node_modules makes npm install fail on a vite peer conflict.
  • npm 11 blocks install scripts by default (sharp, unrs-resolver didn't run theirs). Lint and build work without them; the app doesn't use next/image.
  • Vitest 5 needs @types/node ≥ 22; the config is vitest.config.mts because package.json has no "type": "module".

Deploy (Gitea Actions)

  • ~ in an rsync -e string is not your $HOME. rsync passes it to ssh without a shell, and ssh expands ~ from the passwd entry. In a job container those can differ, so a key written to $HOME/.ssh isn't found and known_hosts is read from somewhere else. The workflow uses absolute paths (SSH_DIR) and UserKnownHostsFile. Found in a local dry run.
  • The deploy target is Plesk's own subdomain folder (calendar.nachklang.art/), so rsync --delete must exclude what Plesk or Passenger keep there (tmp/, .well-known/). The first deploy replaces the Angular build - back it up first, and switch Plesk to the Node app right after the run, because the folder no longer works as a static site.
  • Plesk's chrooted SSH shell is broken on this server: bash segfaults in the chroot copy under /var/www/vhosts/nachklang.art, so sessions close right after login (and the command of an ssh host 'cmd' never runs, exit 255). The deploy user uses plain /bin/bash instead.
  • SSH is on port 42069, not 22 (22 is closed); the deploy user is plesk-nachklang, not nachklang.
  • Jenkins deployed as root. Files it copied into an app folder belong to root:root, so the first Gitea deploy (as plesk-nachklang) fails with rsync "Operation not permitted" / "Permission denied". Before switching an app, run find . ! -user plesk-nachklang in its folder and chown -R plesk-nachklang:psacln what it lists (hit on the API, 2026-11-08; the calendar folder only held files directly in a folder the user owned, so deleting them worked).
  • Every Gitea Actions job starts in a fresh container. Splitting test and deploy into two jobs repeats checkout, Node setup and npm ci (and here the build); one job with the deploy steps conditioned on push-to-master is faster. Spell the condition out on each step instead of comparing an env flag to 'true' - a mismatch there skips the deploy silently and the run stays green.
  • Disable the Jenkins job before merging to master, or it runs ng build on the next push.

shadcn / Base UI

  • npx shadcn add imports cn from the npm package cn and adds it (plus next-themes for the toaster). This repo uses @/lib/utils like feedback, and toasts are light-only - fix both after every shadcn add.
  • Generated components have English labels ("Close"); translated to "Schließen" in dialog.tsx and sheet.tsx.
  • Base UI, not Radix: toggles expose aria-pressed (not data-state=on), triggers take a render prop (not asChild), menu radio items stay open unless closeOnClick is set.
  • A link that looks like a button is a plain <a className={buttonVariants(…)}>, not <Button render={<a …/>}>. Through Base UI's Button it warns in dev ("expected a native ", shown as Next's "2 issues" badge) and, with nativeButton={false}, gets role="button" - screen readers then announce a link as a button.
  • Base UI Select ignores synthetic click()s; automated tests need real pointer or keyboard input.
  • Closed dialogs stay in the DOM with data-closed until their exit animation ends - in a hidden tab that never happens. Query [data-open] when checking what's open.
  • In a hidden tab matchMedia change events don't fire either, so a resize there doesn't switch the sheet between bottom and side. It works in a visible window.

CSS

  • Responsive rules inside a capped panel need container queries, not viewport breakpoints (the edit sheet's footer labels use @container / @md:).
  • An implicit grid column grows to its longest unbreakable content (a URL) and defeats truncate; use grid-cols-[minmax(0,1fr)].
  • break-all splits ordinary words mid-word ("App-s"); only use it on text without spaces.