Root-cause fix for cross-file DB pollution in the vitest suite, which
surfaced as intermittent FK-violation failures once the suite hit 12
test files. tests/setup.ts previously mutated the global
process.env.DATABASE_URL and relied on src/lib/prisma.ts's singleton
picking it up, which is racy across files sharing a worker process
(process.env isn't reset between files even though Vitest resets each
file's module registry). fileParallelism: false only masked this.
Now each test file's own tests/setup.ts creates its own throwaway
Postgres database and vi.mock()s "@/lib/prisma" with a dedicated
PrismaClient constructed via the explicit datasourceUrl option, so no
mutable state is shared between files and fileParallelism can go back
to its default (true).
Verified with 4 consecutive full `npm test` runs (327/327 passing each
time) and no leftover throwaway databases after each run.
The only way to move an Assignment to a new employee was two separate calls
(PATCH return + POST create), which is not atomic and left a race window
where a third party could grab the now-in_stock AssetItem in between, plus
no rollback if the second call failed. Adds transferAssignment() which
closes the old Assignment and creates the new one inside a single
prisma.$transaction(), per docs/asset-license-management.md's "Assignment
(chốt 2026-08-02)" spec. Does not touch AssetItem.status or
License.seatsUsed — the resource never becomes free mid-transfer. Wired
into PATCH /api/assignments/[id]: an employeeId key in the body dispatches
to transfer, a returnedAt key to return, both together is rejected as
mutually exclusive.
48 tests covering the XOR validation, the AssetItem.status/active-assignment
preconditions, seat overbooking, seat/status sync across assign+return,
FK-not-found for all three FKs, double-return, 404s, auth-gating, malformed
bodies, and two explicit concurrency tests (Promise.all against a
1-remaining-seat License and against a single in_stock AssetItem) exercising
the transaction race-closing logic in src/lib/assignments.ts.
AssetItem.model was a free-text string with no referential integrity.
Now a required FK to AssetModel, mirroring the Request.employeeId and
License.vendorId retrofit precedent (no backfill needed, no real
AssetItem data existed).
28 tests covering lib-level validation/CRUD and route-level HTTP behavior:
malformed JSON, non-object body, invalid dataType, categoryId FK violation,
duplicate (categoryId, key) rejected with a clean 400 (not 500), and same
key allowed across two different categories (the "not globally unique"
case), plus auth gating.
Nullable FK from License to PurchaseOrder for supply-side traceability
(which PO a license came in through), mirroring AssetItem.purchaseOrderId's
just-merged pattern exactly. Separate commit from the earlier
License.vendorId retrofit per the project's 1-commit-per-business-change
convention.
License.vendor (free-text string) is now vendorId, a required FK to
Vendor, per docs/asset-license-management.md's "## Vendor (chốt
2026-08-02)" spec. Mirrors the Request.employeeId string->FK precedent:
Zod validates UUID format only, Prisma's own FK constraint (P2003) is
caught at insert time and translated into a clean 400 instead of a raw
500. No License data existed yet, so this is a pure column-type
migration, no backfill needed.
Also converts createLicense() to the {ok,value}|{ok,error} result
pattern used elsewhere (createRequest, createAssetItem) instead of
returning a bare LicenseRecord promise - closes the gap flagged in
CLAUDE.md now that License has its first FK-violation error case to
handle.
- createRequestSchema now requires employee_id to be a UUID-format
string (was a free-text non-empty check), matching the new FK to
Employee in schema.prisma.
- createRequest() now returns a CreateRequestResult discriminated
union and catches Prisma P2003 (FK violation), returning a friendly
ok:false error instead of a raw 500 when employee_id does not
reference an existing Employee. POST /api/requests branches on
created.ok and returns 400 in that case.
- tests/requests.test.ts: every test that used to POST a fake string
employee_id (emp-1, emp-2) now creates a real Employee first via
createTestEmployee() and uses its id, since employee_id is now a
real FK. Added coverage for a syntactically invalid UUID and for a
well-formed but non-existent UUID (expects the friendly 400 from
the P2003 catch, not a raw 500) - the exact case this retrofit was
meant to protect against.
Mirrors tests/requests.test.ts's structure and @/auth mocking
pattern. Covers: successful create (with and without managerId),
empty name/department validation failures, malformed managerId UUID,
nonexistent managerId (FK violation -> friendly 400, not a raw 500),
a self-referential manager sanity check via direct Prisma update,
list ordering, empty list, and 401-when-unauthenticated for both
routes.
Mocks @/auth (vi.mock) instead of exercising the real auth(): that
helper reads cookies via next/headers under the hood, which only
works inside Next.js own request-handling machinery, not when these
tests import and call GET()/POST() directly. beforeEach mocks a
valid session by default so the existing business-logic tests are
unaffected, and two new tests mock a null session to assert both
GET and POST return 401 with { error: Unauthorized } and that POST
does not create a row.
middleware.ts itself has no automated test - it runs on the Edge
Runtime and is not exercised by vitest calling route handlers
directly. Verified manually instead via next build + next start:
curl confirmed GET / redirects to /login (307), GET /login stays
200 with no redirect loop, GET/POST /api/requests return 401 JSON
unauthenticated, and GET /api/auth/session (needed for the login
flow) is not blocked.
toanhoang
toanhoang