frontend-build-timing-audit

Frontend Build Timing Audit

The OpenOps frontend has hit production-only i18n regressions caused by a subtle interaction between Vite 8 + rolldown's chunk splitting and i18next's lazy initialization. This skill bundles the tooling we used to diagnose and prevent those bugs.

When to use

Trigger this skill proactively when modifying any of:

• packages/react-ui/** (especially main.tsx, i18n.ts, routing/lazy loading)
• packages/ui-components/** (barrel exports in src/index.ts, src/components/index.ts)
• packages/shared/** (the @openops/shared barrel)
• packages/react-ui/vite.config.ts
• Any package.json with a sideEffects field
• Adding or removing top-level import { t } from 'i18next' calls

Or reactively when a user reports:

• "Works in dev, broken in production / docker compose / staging"
• "i18n labels are missing / blank / undefined"
• "Toggle / button / nav item has no text"
• Empty strings appearing in the UI after deployment

The bug pattern (mental model)

main.tsx imports:
  1) ./i18n.ts    ← calls i18next.init() which sets this.translator
  2) ./app/app    ← the rest of the App

In dev, ./i18n.ts always evaluates before any App module, so a top-level t('Foo') inside an App-tree module gets a working translator.

In production builds, rolldown can split the App tree into static-dependency chunks of the entry chunk (import {...} from "./cloud-user-api-*.js"). Per the ESM spec, those dependency chunks' module bodies evaluate before the entry chunk body — which means before ./i18n.ts's init() call. Any module-scope t('...') in those chunks now runs while i18next.translator is undefined, and t() returns undefined. The "label" gets frozen as undefined in a top-level array/object and the UI renders nothing.

The fix has two layers:

1. Source layer: never call t() at module top level. Use useTranslation() (a hook) or a factory function instead. This is what PR #2292 did for test-step-data-viewer.tsx.
2. Bundler layer: declare "sideEffects": ["*.css", "*.scss"] on packages/ui-components/package.json so rolldown stops emitting ui-components modules as static-dependency chunks and inlines them into the entry chunk body where ./i18n always wins. This is PR #2293.

For the full investigation with measurements, see reference.md.

Workflows

Workflow 1: Quick anti-pattern scan (always run when touching FE code)

The script scripts/find-toplevel-init-calls.mjs uses TypeScript's AST to find every call to a configured init-dependent binding whose ancestor chain contains no function/arrow/method — i.e. every module-scope invocation that runs at module load.

The set of bindings to watch is declared in scripts/anti-patterns.config.mjs (data-driven). The initial entry covers the i18next t case that caused OPS-4318; adding a new pattern (Sentry, dayjs locale, etc.) is a one-liner.

node .claude/skills/frontend-build-timing-audit/scripts/find-toplevel-init-calls.mjs

Output groups results by rule, one line per offender: <path>: lines 12, 19, 25. The scan runs in ~1 second across the whole repo. Pass --json to get a machine-readable shape (used by snapshot.mjs).

Triage rule: every result is a latent OPS-4318-style regression. The risk of it manifesting today depends on which chunk the file lands in, which depends on rolldown's heuristics. Treat any hit as something to fix or explicitly justify.

When you find offenders, fix them by:

• For React components: const { t } = useTranslation() inside the component, move the constant inside the function body or wrap it in useMemo.
• For non-component files (zod schemas, filter configs, route metadata): convert the exported constant into a factory function getFooFilter(t) or a hook useFooFilter().

Workflow 2: Bundle-diff diagnostic (when actively triaging a regression)

The script scripts/bundle-diff.mjs runs two production builds (varying working-tree state), then compares:

• Chunk count and entry chunk size
• Which chunks contain a given marker string (e.g. a missing label literal)
• Static-import dependency list at the top of the entry chunk
• Critical-path size (raw / gzipped / brotli) — what the user actually downloads before the app can boot

This takes ~30s per build configuration. Use it when:

• The AST scan is clean but a build-only bug is still suspected (maybe from a vendor module, dependency upgrade, or vite config change).
• Evaluating whether a proposed sideEffects change actually helps.
• Comparing bundle behavior before/after a vite or rolldown upgrade.

Usage:

node .claude/skills/frontend-build-timing-audit/scripts/bundle-diff.mjs \
  --label-a 'baseline (main)' --ref-a main \
  --label-b 'feature branch'  --ref-b HEAD \
  --marker 'View input data'

--marker is the literal you expect to be missing or moved (e.g. a label that disappeared). It will be located in each build's chunk graph so you can see whether it's in the entry chunk (safe — runs after i18n.init()) or a statically-imported sibling chunk (dangerous — runs before).

For multi-variable experiments (e.g. toggle two patches independently to isolate which one fixes the bug), see the --patch-a / --patch-b flags documented in the script's --help.

Workflow 3: Regression check vs committed baseline

scripts/baseline.json captures the current build's chunk count, entry-chunk size, critical-path size (raw / gzipped / brotli), total JS size, and per-rule anti-pattern counts. Two scripts work against it:

# After editing FE code, verify nothing regressed past thresholds.
node .claude/skills/frontend-build-timing-audit/scripts/check-regression.mjs
# Exit codes: 0 ok, 1 warn, 2 fail (suitable as a CI gate).

# Re-anchor the baseline after an intentional change (dependency upgrade,
# refactor that legitimately moves chunks).
node .claude/skills/frontend-build-timing-audit/scripts/update-baseline.mjs
# Prints the diff vs the previous baseline before writing.

Thresholds live inside baseline.json as percent deltas so routine growth doesn't constantly require updates. Anti-pattern counts default to warnPct: 0 so any new offender (a new module-scope t(...), a new module-scope Sentry.init(...), etc.) warns immediately.

Use this workflow:

• After landing FE changes locally to catch surprise bundle bloat.
• After a Vite, rolldown, or core dependency upgrade.
• Before approving a PR that touches barrels, lazy-loading boundaries, or sideEffects declarations.

If the verdict is warn or fail, drop into Workflow 2 (bundle-diff.mjs against main) to see which chunks moved.

Workflow 4: Pre-commit verification for sideEffects changes

If editing a sideEffects field in any packages/*/package.json:

1. Identify the package's real side effects (CSS/SCSS imports, global registrations, polyfills). Grep for import '*.css', customElements.define, module-scope window.*/global.* assignments.
2. If only CSS/SCSS, use "sideEffects": ["*.css", "*.scss"] (array form).
3. Run node .../scripts/bundle-diff.mjs against the previous commit to confirm:
   • CSS files still in the build (xyflow, tailwind, app styles).
   • Critical-path size hasn't grown.
   • No expected runtime code was tree-shaken out.

Key facts to remember

• t imported from i18next is NOT reactive. It does not re-render components when language changes or resources load. Use useTranslation() from react-i18next in components.
• Before i18next.init() runs, t() returns undefined, not the key. React renders undefined as nothing — that's why labels go blank rather than showing English fallback.
• The repo's translation.json is key=key, so once init has run, missing keys fall back to the English text by accident. This masks broken usage in most cases — but only after init. Module-scope t() may run before init.
• main.tsx order matters. ./i18n MUST be imported before App. Verify with head -10 packages/react-ui/src/main.tsx when investigating.
• Vite dev mode hides this bug entirely. Always verify fixes against npx nx build react-ui output, not the dev server.

Additional resources

• reference.md: full root-cause analysis with the measured 4-way build comparison, byte-level chunk analysis, and the ESM evaluation semantics that make this bug possible.
• Related PRs: #2292 (hook fix), #2293 (sideEffects declaration).
• Related Linear: OPS-4318, OPS-4320.