Security boundaries
Keep localized routing, catalog loading, CMS content, diagnostics, and AI review inside explicit security boundaries.
Treat locale values, paths, catalogs, remote content, and provider responses as untrusted input. The package validates common boundaries, but the application remains responsible for authentication, authorization, deployment policy, and safe rendering.
Routes and redirects
Use the route-map helpers for same-origin paths. They reject external and
protocol-relative inputs, malformed segments, traversal, and paths outside a
configured basePath. Redirect targets are returned as data; your framework
must enforce same-origin redirects and its own host policy before sending them.
Never use an unchecked Host header to construct canonical URLs.
Configured route diagnostics report route templates such as
/articles/:slug, not concrete parameter values. Unmatched paths use a
sentinel, and query strings and fragments are not recorded.
Catalogs and CMS content
HTTP catalog sources accept only HTTP(S), reject URL credentials and fragments, disallow redirects, bound response size and time, and isolate caches by locale and namespace. Keep authorization headers in trusted server code. Filesystem sources resolve symlinks and reject paths outside their configured root.
Catalog and CMS values are data, not HTML. Translation components render
strings as text; use RichTranslation with explicitly supplied React elements
for supported ICU tags. Validate CMS records with an application schema and
render rich text through a safe renderer. Do not pass catalogs or content to
dangerouslySetInnerHTML.
The Sanity source uses a fixed parameterized query and an injected client. The
application owns its credentials, client perspective, and preview
authorization. A preview flag is not an authorization check.
AI suggestions and review
Call injected suggestion providers only on the server. Wrap them with
createAuthorizedSuggestionProvider to fail closed before provider access;
the application must derive the authorization decision from trusted identity.
Keep provider credentials and source content out of client bundles and routine
logs.
Persist review suggestions and audit records through a transactional store.
The persistence implementation must atomically compare the source revision,
apply approval, and append audit history. Authenticate and authorize reviewers
before calling approve or reject; a suggestion's quality score is advisory,
not an approval.
Verification
The i18n unit tests cover unsafe route inputs, HTTP source boundaries, content validation, review authorization, and optimistic updates. The browser test verifies that untrusted translation strings remain literal text and do not create elements. These checks complement, but do not replace, application penetration testing and deployment-specific policy review.