Run the tests on pull requests

[greenido]

Stacked on #10 → which is stacked on #9. Stacked rather than branched from main only because it rewrites the test:* scripts in package.json that #9 and #10 each added to; branching separately would conflict with both. GitHub retargets it as they merge.

The gap

deploy.yml triggers only on push: branches: [main]. So a pull request's checks were a GitGuardian secrets scan and nothing else — a regression was caught by the deploy after the merge, not by the PR that introduced it. Both PRs in this stack show exactly one check.

ci.yml now runs lint, typecheck, test and build on every pull_request. Each step runs even when an earlier one fails (if: !cancelled()), so one push surfaces every problem instead of just the first, with an explicit gate at the end so the job can't go green with a red step. Pushes to main stay covered by deploy.yml, which now also typechecks — splitting them means no commit gets tested twice.

Both workflows read the Node version from .nvmrc rather than hardcoding '22', which had already drifted from the pinned 22.13.0.

Two things that made CI worth less than it looked

The suites were chained with &&. The first failure hid the other eight — one round trip through CI per broken suite. scripts/run-tests.js runs all of them and prints a summary:

────────────────────────────────────────────
  ✓ elevation
  ✗ zza-fails  (exit 1)
  ✓ zzb-runs-anyway
────────────────────────────────────────────
1 of 11 suite(s) failed.

Verified with throwaway probes named so the passing one sorts after the failing one — it ran, and the job still exited 1.

Each suite had to be registered by hand. Easy to forget, and two never were: tests/new-features.test.ts and tests/weather-api.test.ts had gone unrun for their entire existence. Suites are now discovered from tests/, so a *.test.ts file is picked up by existing. Network-hitting suites declare themselves with an @network marker in their header and are excluded from yarn test — CI must not go red because NOAA is having a morning.

What running those two for the first time found

weather-api.test.ts could not compile at all:

tests/weather-api.test.ts(616,5): error TS1343: The 'import.meta' meta-property is
only allowed when the '--module' option is 'es2020' | ... | 'nodenext'

It guarded its entry point with import.meta.url === \file://${process.argv[1]}`, which is invalid under the CommonJS module setting tsconfig.test.jsonuses. Note this passedyarn typecheck, because the app's tsconfig.jsonsetsmodule: esnext— it only broke when actually executed. Fixed torequire.main === module`.

All three network suites now pass: new-features, providers, weather-api (4 tests).

Script changes

Nine near-identical test:* entries collapse into one filtered entry point:

yarn test              # all 9 offline suites
yarn test scoring      # one suite, matched by substring
yarn test:network      # the 3 live-API suites
yarn typecheck         # new — tsc --noEmit, covers the test files too

README updated for all of it.

Verification

yarn lint, yarn typecheck, yarn test (9 suites, 176 assertions), yarn test:network (3 suites) and yarn build all pass locally. Exit codes confirmed in both directions: 1 with a failing suite, 0 without.

This PR validates itself — being a pull_request, ci.yml runs against its own branch, so the checks on this PR are the new workflow.

Note

ci.yml uses pull_request, not pull_request_target, so a fork's PR runs with a read-only token and no secrets. That's the right default here; it does mean a first-time contributor's run may need approval under your repo settings.

🤖 Generated with Claude Code