Testing and verification
Distinguish artifact integrity, project metadata, selected native tasks, and actual browser or deployment behavior.
Choose the evidence you need
Check formatting, lint and types
Starters with a biome.jsonc file include the commands below. Run them from the application root after installing dependencies. For an older Starter, check its package.json scripts first; updating Framework packages does not add scripts to your application.
bun run check
# Run an individual check:
bun run format:check
bun run lint
bun run typecheckcheck runs formatting, lint and type checks without rewriting source. It does not run tests or build the application. Lint warnings fail the command. The Starter includes its own pinned Biome dependency and configuration, so these commands work outside a Framework checkout and without Git.
# These commands modify source files:
bun run format
bun run lint:fixformat changes formatting only. lint:fix applies Biome's safe fixes; review the diff and rerun check afterward. Source, tests and tools are checked; installed dependencies, generated output, caches and vendor archives are excluded. Require check in CI alongside your test and build commands.
UI Starters with whisk-lint.json also run lint:ui as part of lint and check. It checks that JSX uses Whisk components and approved integrations. Use Box or Flex for layout, Text for text and Button for actions. For semantic layout, use Box as="main"; an interactive control needs its dedicated component, not Box as="button".
Configure all UI source roots in whisk-lint.json. A third-party canvas or router component needs an explicit exception for its file/element or module/export, with a reason. Keep exceptions narrow. lint:ui does not rewrite source or validate CSS, style values, accessibility or runtime-generated content; keep type and browser tests too.
Discover, then run selected tasks
smoothie project verify --format json
smoothie project verify --task typecheck --task test --task build --run --format jsonThe first command inspects supported root package scripts without probing tools or starting the application. The second invokes only the selected scripts. Missing scripts, unavailable tools, and failed checks must remain visible. Native tools still own the test runner and compiler; you can invoke your scripts directly when needed.
Project scripts are your application's code: the CLI runs them without a sandbox, and a passing script can still skip tests internally. A successful task is not evidence for an unconfigured database, live model, or production authentication. Keep deterministic checks separate from explicitly configured service/provider tests and report which lane actually ran.
Exercise the actual browser application
./node_modules/.bin/playwright install chromium
smoothie project verify --task test:e2e --runRun these from Smoothie Starter after installing dependencies. Chromium installation may download browser assets. The Starter's test:e2e script builds and manages its own test server; stop a conflicting local process first. Custom applications must provide and own their own browser test setup.
Check an already running target
smoothie project verify --url http://127.0.0.1:42741
smoothie project verify --url http://127.0.0.1:42741 --runStart the built Starter separately before execution. The first command plans without a request; the second selects the declared test:deployment script and supplies the origin through SMOOTHIE_DEPLOYMENT_URL. It does not build or start a server. Use a reviewed HTTPS origin for remote checks, and ensure your application's script honors it. Paths, query strings, credentials, and fragments are not accepted as an origin.
A passing script does not prove which artifact was served. Attempt-aware verification can additionally query recorded Cloudflare version/traffic evidence; see Build and deployment. Review tests before authorizing remote execution because they can perform application side effects.
Keep failures and limits in the result
Human and JSON output describe the same operation. In automation, inspect status, diagnostics, checks, and effects, not just a friendly summary. Missing output or a skipped prerequisite is not success. After cancellation, inspect native/external state before retrying operations that could have produced side effects.