Test and Cypress scripts can silently run against stale sibling dist/

[dqnykamp]

Every cross-package @doenet/* import in this repo resolves to the built dist/ of the sibling package: each one exports "." → "./dist/index.js", and of the 46 tracked vite*.config.* / vitest.config.* / cypress.config.* files exactly one maps a @doenet/* specifier to source — packages/doenetml/cypress.config.ts, which aliases @doenet/codemirror (only that package) for its component-test dev server, for an unrelated reason it documents inline. The three @doenet/* tsconfig path mappings (tsconfig.build.json, tsconfig.test.json, packages/vscode-extension/tsconfig.json) all point at dist. So any npm script that reaches a sibling package and is not a wireit script with that package's build declared can run against stale output — and it does not fail, it just quietly uses old code.

#1672 fixed this for the three generator scripts in @doenet/static-assets (build:schema, build:assets, check:docs-coverage), where the symptom was a regenerated schema losing entries a branch had just added (hit for real on #1669's Klingon catalog). The same hazard is still live in the test scripts.

Where it still bites

• Every test script whose specs reach a @doenet/* import. Most consequentially doenetml-worker-javascript's test / test:group1-4, plus doenetml, parser, utils, lsp, lsp-tools, doenetml-worker, v06-to-v07, static-assets. packages/utils/test/styleDescriptions.test.ts already carries a comment admitting the hazard and telling the reader to run npm run build -w @doenet/i18n by hand; a stale @doenet/utils is enough to make styleDescriptionLocale.test.ts fail on an otherwise-unmodified checkout.
• The Cypress runners in test-cypress, codemirror, doenetml, doenetml-iframe, and the root test:e2e-group* / test:*-cypress wrappers. This is the hazard TEST_RUN_INSTRUCTIONS_FOR_AGENTS.md's "rebuild before Cypress runs" rule exists to paper over.
• packages/static-assets' own test, which reaches the worker source through scripts/get-schema.ts — the same import path Let the schema scripts build the packages they read #1672 was about.
• packages/doenetml-to-pretext / doenetml-prototype test are partially guarded: their build:test1 is wireit and covers doenetml-worker + parser, but their src/ also imports @doenet/utils, @doenet/codemirror and @doenet/ui-components, which are not covered.

packages/memory-benchmark's benchmark is the exemplar of the right pattern already: a wireit script with ../standalone:build and ../doenetml-iframe:build declared.

Suggested fix shape: split watch and single-run entry points

The obvious move — make test a wireit script — breaks watch mode. Wireit spawns the command with piped stdio (wireit/lib/script-child-process.js passes no stdio option, so the default pipe applies), so the child sees no TTY and no stdin. These test scripts are a bare vitest, i.e. watch mode, whose interactive keypress UI would stop working — for the one command developers spend all day in. Wireit has no option to inherit stdio.

So instead:

• leave test a plain vitest (watch, interactive, no wireit in the way);
• add a wireit test:run per package: command: "vitest run", dependencies = the same sibling builds that package's own build declares, and no files/output so it always re-runs. Where a package's specs reach the worker's source (@doenet/static-assets' own test does), the dependency to declare is ../doenetml-worker-javascript:build:deps — the no-command aggregator Let the schema scripts build the packages they read #1672 added for exactly that, which avoids a vite pass over the worker's 7 MB bundle that nothing loads;
• point CI and the root aggregate scripts at test:run. The root test and test:all-no-worker-js already append -- run, so they would just call test:run instead.

Notes for whoever picks this up:

• Argument forwarding is not a blocker. npm run test:run -- <file> reaches the underlying command (wireit README, Extra arguments), and extra args are part of the fingerprint. On Node 24 passing args does make wireit emit a DEP0190 deprecation warning, because it spawns with shell: true and an args array — harmless, but it would show up on every targeted run.
• CI cost is ~nil. Both test jobs already build against the downloaded .wireit cache before testing (test-main runs build:all-no-docs; test-worker-js runs build -w packages/doenetml-worker, which reaches the same packages), so the declared dependencies would be cache hits. WIREIT_CACHE has to be set explicitly in CI, though — wireit defaults it to "none" whenever CI=true.
• Wireit's quiet logger passes the root script's output through, so test output is not swallowed.
• The Cypress runners are a bigger change than adding dependencies: packages/test-cypress/vite.config.ts hard-codes ../standalone/dist/*, so the wiring has to grow ../standalone:build and friends, and TEST_RUN_INSTRUCTIONS_FOR_AGENTS.md would want rewriting around it. Worth doing as a second step.

Not on this list

packages/i18n's check:instances (root check:i18n-instances) looks like a member of this family, since it reads packages/*/dist directly. It is not: skipping a package nobody built is deliberate and documented in scripts/check-bundle-instances.ts, the number of packages scanned is printed, and a scan that finds nothing at all is an explicit failure ("No built scripts found under packages/*/dist/ … a scan of nothing passes trivially"). Left alone here. (An earlier version of this issue listed it as a separate bug; that was wrong.)

Follow-up to #1672.

🤖 Generated with Claude Code