1.8.5: accept the nest-server roles: string[] shape in isAdmin so the admin UI stops silently disappearing

[kaihaase]

Problem

isAdmin currently checks the singular role field only:

const isAdmin = computed<boolean>(() => user.value?.role === 'admin');

But the lt nest-server / Better-Auth user carries its roles as an array — roles: string[] — and has no role field at all. Against a real lt backend, isAdmin is therefore permanently false: the entire admin UI silently disappears for every admin, with no error and no log line.

This is not hypothetical. Verified live against a production instance — get-session returns:

{ "roles": ["admin"], "email": "…", "id": "…" }

— no role. In nest-server, roles is a core Better-Auth additional field, registered unconditionally with type: 'string[]' and defaultValue: [] (not optional, not opt-in). Every lt project with an admin UI is affected.

Fix

• LtUser gains roles?: string[] as a real, typed field.
• isAdmin accepts both shapes (singular role string and roles array), so it stays backward compatible.
• Array.isArray guard: the auth-state cookie is client-writable, so a tampered roles: 42 must yield false rather than throw inside a computed.
• roles is added to AUTHZ_KEYS. Once isAdmin reads roles, a stale cached roles: ['admin'] would keep the admin UI alive after the backend revoked the role — exactly the fail-open that AUTHZ_KEYS exists to prevent. Fail-closing role while fail-opening roles, when both feed the same isAdmin, would be incoherent.
  A backend that never sends roles loses nothing: AUTHZ_KEYS only drops a key when the cache has it and the session omits it — a roles-less backend never caches roles, so the drop is a no-op.

Tests

New test/is-admin.test.ts (8 cases): {roles:['admin']} without role → true, {role:'admin'} → true, {roles:['user']} → false, null → false, plus tampered non-array values.
Extended test/merge-session-user.test.ts with 3 roles AUTHZ cases.

Counter-checked: with the old one-liner and the old AUTHZ_KEYS, exactly 4 of the new assertions fail. A regression test that passes against the broken code would be worthless.

Verification

format:check clean · lint exit 0 · test:types exit 0 · 128/128 tests (11 files) · build exit 0.

Note

The version bump to 1.8.5 follows this repo's convention (the PR commit does the bump; the maintainer adds the changelog entry separately on main after the merge — same as #4). If another PR lands first and this collides, that single hunk is trivial to drop.