* fix(tests): wait for registry readiness in global setup instead of per-test sleep
The first e2e test was flaky on CI because `start-server-and-test` only
checks that the root URL (http://localhost:4000) responds before running
tests, not the /r registry endpoint. The existing 2-second hardcoded sleep
in the first test was unreliable on slower CI runners.
Move the readiness check into the vitest globalSetup so all tests wait for
the registry /r endpoint to actually be reachable before any test starts.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(tests): fix race condition in global setup - poll correct URL and CLI binary
Two issues caused the previous fix to fail:
1. Was polling `http://localhost:4000/r` which is a directory → always 404.
Now polls `{REGISTRY_URL}/index.json`, a real static file that returns 200.
2. The v4 dev script (`pnpm --filter=shadcn build && pnpm icons:dev & next dev`)
runs the shadcn CLI build in the background while next dev starts immediately.
On fast CI runs start-server-and-test can detect the server as ready before
the CLI binary (packages/shadcn/dist/index.js) has been built, causing the
first test to fail when it tries to invoke the CLI.
Now explicitly waits for the binary to exist before any test runs.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(tests): warm up /init route in global setup to prevent first-test timeout
The CLI's first request during `shadcn init` hits the dynamic Next.js /init
route. On a cold dev server this route takes ~1.8s to compile. Combined with
the rest of what init does (pnpm install, file writes), this pushes the first
test over the 30s CLI timeout on CI. Subsequent tests pass because the route
is already warm.
Polling /init in global setup ensures the route is compiled before any test
runs, making the first test's CLI invocation as fast as all subsequent ones.
Also replaced the /r/index.json poll (a static file that responds immediately
and doesn't reflect real route readiness) with the actual /init route poll,
which also naturally verifies the registry server is up.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(tests): warm up 404 route and increase default CLI timeout
Two more issues found in CI logs:
1. The CLI requests font files that don't exist (e.g. /r/styles/new-york-v4/
font-geist.json), causing Next.js to compile /_not-found/page on the first
404 response. That compilation takes ~4-5s on a cold dev server and is
another hidden cost on the first test. Now triggering a 404 in global setup
so the not-found page is compiled before any test runs.
2. The default CLI timeout of 30s is too tight for CI. Even with the /init and
404 routes pre-warmed, pnpm install inside the fixture takes ~25s, leaving
only ~5s of headroom. Increasing the default from 30s to 60s gives a
comfortable buffer without masking real hangs.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
All monorepo templates hardcoded `packageManager: "pnpm"` which
meant running `bunx --bun shadcn@latest init --monorepo` would
still shell out to `pnpm install`, triggering Corepack and crashing
under Bun because `process.mainModule` is readonly there.
This removes the hardcoded pnpm override from every monorepo template
config so `getPackageManager()` can actually detect what the user is
running. The scaffold step now adapts the cloned template on the fly:
- strips the `packageManager` field from package.json (avoids Corepack)
- converts pnpm-workspace.yaml to a `"workspaces"` array in package.json
- removes pnpm-lock.yaml
- rewrites `workspace:*` refs to `"*"` when the detected PM is npm
(npm doesn't support the workspace: protocol)
- picks the right install flags per PM (`--no-frozen-lockfile` for pnpm,
nothing extra for bun/npm/yarn)
pnpm behavior is completely unchanged — `adaptWorkspaceConfig` early-
returns when the detected PM is pnpm.
