A saved SIMBK test run reported 778 tests. Of those, 677 passed and 101 were skipped. There were no failures, but calling that "778 passing tests" would have overstated the evidence.
SIMBK is a school counseling platform I maintain in Banjar, Indonesia, using Next.js 16 and PostgreSQL on a VPS. This maintenance pass covered audit logging, deployment checks, and backup procedures, alongside a separate review of shared VPS services. AI agents assisted with implementation, code review, command execution, and drafting this article. I set the scope and approval boundaries; the work described here was not all written or verified manually by me.
These are six gaps from that work. Some were configuration defects; others were limits in what a check could establish. One was caught by the test suite itself.
1. SQL that works outside the migration runner
The audit viewer needed an index on created_at DESC. Migration 025 initially used CREATE INDEX CONCURRENTLY IF NOT EXISTS.
The project's staging migration runner executes pending migrations inside a transaction. PostgreSQL does not allow CREATE INDEX CONCURRENTLY there. A successful standalone SQL command would not establish that the migration could run through this tool.
The corrected migration uses plain CREATE INDEX IF NOT EXISTS. That is compatible with the runner, but it has a different locking cost: a plain index build blocks writes to the table while it runs. For a large or busy table, a separately managed non-transactional migration would need its own rollout plan. Removing CONCURRENTLY is not a universal production fix.
The acceptance check must exercise the runner and its transaction boundary, not just the SQL statement in isolation.
2. Tests outside the discovery glob
The default test command selects tests/*.test.mjs before running lint. That glob covers the top level of tests/; it does not include the integration tests under tests/ruang-bk/.
The saved run's 677 passes and 101 skips describe only the tests it discovered. The unmatched subdirectory is a separate gap: its files do not appear in the skipped count because the runner did not select them.
The package scripts now expose test:db for that directory and test:all for the default command followed by it. Adding those commands makes the intended scope visible. It does not prove the database-backed tests passed. They require their configured isolated database, and I am not presenting the saved default-suite result as a complete integration run.
3. Redis authentication needs a rejection check
The shared VPS review also covered Redis. This was an infrastructure check, not evidence that SIMBK itself depends on Redis.
Redis ACL rules can leave the default user in nopass mode even when an operator expects requirepass to enforce authentication. The effective ACL state matters more than the presence of a password directive in a configuration file.
For this password-required setup, the useful acceptance pair is an unauthenticated request rejected with NOAUTH and a correctly authenticated request returning PONG. A successful authenticated request alone cannot establish that anonymous access is blocked. Nor does this pair establish network isolation or a complete Redis security audit.
4. CSP must be checked on the application response
SIMBK's Next.js configuration owns its Content Security Policy (CSP), including an img-src allowance for the Supabase origin used for student photos. A second, more restrictive policy at nginx would not replace that policy. Browsers enforce both, so an image source allowed by the application could still be blocked by the proxy.
The same review also exposed a probe-selection problem: a redirect response from an apex host is not the final application response. Checking the wrong host can miss the application's headers and encourage an unnecessary proxy-level addition.
The practical check is to follow the intended public route, inspect the final response, and confirm that required resources still load. The risk described here is an image-loading regression; I am not claiming that a production photo outage occurred.
5. Agent reports still need diff review and tests
The audit-logging changes added imports to server actions. The consultation-scheduling tests use a hand-written module loader, so their mocks also needed entries for the database pool and audit helper. The fix is visible in the same change set as the new audit calls.
This is a case where running tests did catch an integration problem. The earlier draft's claim that none of these issues were caught by tests was wrong.
Agent summaries help locate changes, but they do not establish acceptance. The review needs the diff, the test output, and the identity of the tree that produced it. That last part matters when a workspace contains unrelated uncommitted edits: a passing local build can include code that is absent from the branch being reviewed.
6. The restore process needs access to the dump
The backup scripts cover a local PostgreSQL dump, an encrypted offsite copy, and restore drills into disposable databases. The offsite drill downloads and decrypts the backup before invoking pg_restore as the postgres operating-system user.
A temporary directory with mode 0700, owned by the invoking user, prevents postgres from opening a file inside it by pathname. The script instead opens the dump in the invoking shell and passes it through standard input:
sudo -u postgres pg_restore -d "$DRILL_DB" --no-owner --no-privileges < "$DUMP"That addresses file access without widening the temporary directory's permissions. The drill checks core tables and basic account/profile counts after the restore. Those checks are narrower than proving every workflow works against the restored database, and a successful drill for one backup would not validate future backups automatically.
What is implemented, and what it proves
- The audit guard checks for audit-call text in explicitly listed source files. It can detect a missing call in those files, but it does not prove every mutation path writes the right event or automatically discover new action files.
- Audit-failure and backup-freshness checks retain state to suppress repeated alerts. The daily journal error digest is different: it summarizes a moving 24-hour window and has no deduplication. Overlapping runs can report an error again.
- The deploy wrapper builds in a scratch directory, saves a build log and rollback receipt, and cleans up the scratch directory on exit. Its checks cover service state, HTTP statuses for health/login and unauthenticated protected routes, plus error-level journal entries. A redirect check does not exercise the authenticated page.
- The Telegram alert script reads its credentials from an external environment file. That specific implementation is not proof that every credential has been removed from the repository or its history.
Verification still open
The outstanding SIMBK acceptance check is to perform an authorized action through the UI and match it to the resulting audit row. That check remains deferred; this article does not claim end-to-end audit coverage.
The evidence here combines a saved test log with inspection of code and configuration from the maintenance work. It is not a fresh replay of every operational procedure. The database-backed suite, authenticated UI behavior, and recovery of application workflows need their own evidence before they can be called verified.