PA cockpit — tests¶
The Public Affairs cockpit is the most heavily tested surface in the product, and the only board with its own end-to-end suite.
These tests moved into a package
As of v2026.08.27 the cockpit lives in
@ronl/pa-cockpit, and its tests travelled
with it — they are no longer part of the packages/frontend suite. That is
why the frontend's own totals fell between v2026.08.23 and v2026.08.33
without anything being deleted.
Package: 43 files · 476 tests. Backend: 16 files · 578 tests in
src/pa-monitoring — the largest single area in the repository.
E2E: 2 specs · 7 tests, still in the frontend package.
Measured with npm test --workspace=@ronl/pa-cockpit on 24 September 2026
against main at 86af73e (v2026.09.11), after a clean npm ci in a separate
clone on Node 24.14.1 / npm 11.11.0: 476 of 476 passing, Duration 37.58s.
Coverage 90.11 % statements · 88.52 % branches · 86.52 % functions ·
91.33 % lines — branches up 12.97 points from v2026.08.36 under the per-file
80% floor adopted in v2026.09.2, which v2026.09.6 wrote into this package's
vitest.config.ts as thresholds: { branches: 80, perFile: true }. The run
passed it without naming a file.
Four releases, four identical counts
43 files and 476 tests at v2026.09.7, v2026.09.9 and v2026.09.11, and the
only src/ change in the latest window was to src/scaffold.test.ts, whose
three tests did not change in number. The coverage figures above are the
12 September ones re-measured: the 20 September pass read 90.07 statements
and 86.39 functions, and this one reads what 12 September did. That is the
second-decimal noise described on
Coverage, not a change in the package.
This package holds 8 of the 26 files an 80% functions floor would fail — the second-largest share after the frontend — while none of its 38 source files is below the 80% branch floor that is actually configured.
Until v2026.09.6, these 476 tests ran nowhere in CI
@ronl/pa-cockpit is a library: it has no deploy workflow of its own, and
no other workflow ran its suite. The most heavily tested surface in the
product was therefore covered on developer machines and nowhere else, while
every deployable package had had a pipeline running its tests since
20 August. Both frontend workflows now run it, as a step placed before
the frontend's own — the frontend imports the package, so there is no point
testing the consumer while the library is broken. The pa-demo workflows
consume it too and watch packages/pa-cockpit/** in their path filters,
but run only pa-demo's own suite.
The E2E figures above are older than the rest
The package counts and coverage, and the src/pa-monitoring backend figure,
are all from 24 September 2026. The two E2E specs were last measured on
29–30 August and were not re-run in this pass — running them needs the full
local stack. See E2E & live smoke.
The package suite¶
| File | Tests | Covers |
|---|---|---|
services/pa.api.test.ts |
66 | The cockpit's API client — the single largest test file in the package |
pages/public-affairs-v2/PaDataProvider.test.tsx |
21 | The context every cockpit screen consumes |
pages/public-affairs-v2/NotificationsPanel.test.tsx |
19 | Previously 0% — notifications were hardcoded to empty |
services/dossierbeheer.api.test.ts |
17 | The dossier API client |
components/PADashboardV2/dossierbeheer/Dossierbeheer.test.tsx |
17 | Dossier list, filters, status transitions |
pages/PADashboardV2.test.tsx |
15 | The page container and the host contract it requires |
components/PADashboardV2/PaSectionsRouter.test.tsx |
15 | The section-id grammar that replaced a fourteen-export surface |
pages/public-affairs-v2/Monitoring.test.tsx |
14 | Monitoring view |
pages/public-affairs-v2/Issuekaart.test.tsx |
14 | The issue map |
components/PADashboardV2/dossierbeheer/DossierEditor.test.tsx |
14 | Authoring a dossier |
services/mock-demo.store.test.ts |
11 | The mock store both hosts drive |
pages/public-affairs-v2/Kompas.test.tsx + kompas.test.ts |
20 | The kompas view and its pure data module |
components/PADashboardV2/dossierbeheer/MdEditor.test.tsx |
9 | The Markdown editor |
components/PADashboardV2/ZoekcriteriaSection.test.tsx |
9 | Saved-search criteria |
pages/public-affairs-v2/Vandaag.test.tsx |
8 | The 30-second start screen |
pages/public-affairs-v2/FeitenCijfers.test.tsx |
8 | Facts and figures |
components/PADashboardV2/dossierbeheer/DossierRow.test.tsx |
8 | Row rendering and actions |
pages/public-affairs-v2/AgendaView.test.tsx |
7 | The agenda |
components/PADashboardV2/PACommandPalette.test.tsx |
7 | The ⌘K palette |
By directory: pages/public-affairs-v2 119, services 94,
components/PADashboardV2/dossierbeheer 67, components/PADashboardV2
53, pages 15, modes 7, package root 10, test 3.
The guards that came with the extraction¶
Seven of the smallest files are not feature tests at all — they are structural guards protecting the package boundary, and they matter out of proportion to their test counts:
| File | Tests | Guards |
|---|---|---|
host.test.ts |
3 | That a host must configure the package before render, and fails with a named error if it does not |
index.test.ts |
2 | The public surface — so a re-export added by accident is caught |
modes/no-module-scope-modes.test.ts |
2 | That the unfiltered mode helpers stay unreachable, which is what keeps the demo's curation honest |
no-tailwind.test.ts |
1 | That no Tailwind utility class returns to the package |
no-host-protocol.test.ts |
1 | That no host session-storage key, IdP literal or router path is hardcoded back in |
The last one is the residue of a real defect class. The package had previously hardcoded five facts belonging to a host application; the guard exists so the seam cannot silently close again.
Two of these guards were themselves defective when written
The modes guard matched raw text, so a single line of documentation-shaped string could disarm the only rule covering dynamic imports — and a decoy function declared in an inner scope disarmed it for a whole file while compiling with zero type and lint errors. It now parses the source, and its accusing and excusing inputs are split so one can no longer feed the other. A guard is code, and needs its own red/green like anything else.
Backend¶
src/pa-monitoring — 533 tests across 16 files:
| File | Tests |
|---|---|
pa.routes.test.ts |
131 |
pa-dossiers.routes.test.ts |
96 |
curation.service.test.ts |
61 |
rules.test.ts |
38 (pure scoring) |
pa-dossiers.db.test.ts |
30 |
pa-monitoring.db.test.ts |
10 |
pa-cache.test.ts |
9 |
notifications.service.test.ts |
8 |
query-match.test.ts |
6 |
rss.test.ts |
4 |
Plus the TK, OB, EU, agenda and media source clients under
pa-monitoring/sources. Coverage: pa-monitoring 98.32%,
pa-monitoring/sources 97.07%.
E2E¶
Two Playwright specs, run with the same playwright.config.ts as the rest of
the frontend suite.
Re-measured 30 August 2026 against acc at 15dfbf9: 7 tests, all passing,
as part of the full frontend run of 27 tests in 1.9m. The count is unchanged
since 22 August, when the two together ran in 18.9s and the live spec was
additionally run six consecutive times while chasing a flake, passing 2/2 each
time in 7.3–11.7s.
See Coverage per board for how these seven sit against the other twenty.
| Spec | Tests | Covers |
|---|---|---|
pa-mock-journey.spec.ts |
5 | Mock mode driven against the real store with no mocking: curating moves the rail badges and the move survives a reload; an ignored signal stays ignored; Reset demodata restores every source to its fixture baseline; the reset control is offered in mock mode only; every signaalbron carries a watchlist orphan that can be linked to a dossier |
pa-live-authoring.spec.ts |
2 | Authoring against the live backend and a real database — a dossier and a zoekcriterium survive a genuine cold reload; live shows authored work while mock shows fixtures, from the same screen |
Why mock mode is worth an E2E suite¶
Every mock-mode defect found by hand in this window was invisible to the unit
suites by construction. They were not logic errors — a saved-search write that
was a bare return;, a confirm that built a new object and discarded it,
notifications hardcoded to empty, a resource fetched once at mount and never
again. Component tests mock the very seam that was broken, so they cannot see
any of it, and each one passed throughout.
Mock mode makes an unmocked end-to-end run affordable: the fixtures are deterministic and nothing depends on what happens to be in the database.
Why the live spec only covers authoring¶
Curation depends on TK OData, the EU RSS feed and the media aggregator, and TK alone measured 10s and 48s for the same query minutes apart — an assertion about signals arriving would be flaky by construction.
The live spec creates everything with a run-unique stamp and removes it again in
afterEach, including when the test fails part-way. Nothing global is reset:
pa:reset-data is a feature of the product, not test tooling.
Three lessons came out of building this suite
A throttled run that looks exactly like an outage, locator.count() not
auto-waiting, and a hand-written mock that passed vacuously. All three are
written up on Writing tests, because they generalise
well beyond this board.