concurrency-auditor C-2: every queue.add(...) site previously enqueued
without a stable jobId, so a double-dispatch (webhook retry, double-
click on Send, scheduler tick collision) would create two queue jobs
and the downstream worker would deliver twice. BullMQ rejects a
duplicate jobId while the original is still queued or active, so a
stable per-entity key gives at-most-once semantics naturally.
Added jobIds across all 10 enqueue sites:
- email send-invoice → `send-invoice:<invoiceId>`
- notifications invoice-overdue-notify → keyed per UTC day so dupes
collapse intra-day but tomorrow's run can re-notify if unpaid
- export gdpr-export → keyed on the exportId (unique per request)
- webhooks deliver (3 sites: dispatch, retry, test) → keyed on the
webhook_deliveries row UUID
- maintenance expense-dedup-scan → keyed on expenseId
- notifications send-notification-email → keyed on notification id
- email send-inquiry-confirmation → keyed on interestId (1 per
submission)
- email send-inquiry-sales-notification → keyed on interestId+email
(1 per recipient per submission)
- reports generate-report → keyed on the generated_reports row id
Pure refactor — no UX impact. Closes the BullMQ dedup gap that was
the second half of the concurrency-auditor's CRITICAL-tier findings.
Test fixture update: gdpr-export integration test now asserts the
jobId option on the queue.add call.
Tests 1315/1315.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
**asset-auditor C1+C2+H1+H3 — image normalization**
Add `src/lib/services/image-normalize.ts` and wire it into
`uploadFile()` so every accepted image is re-encoded via sharp before
hitting storage:
- Strips EXIF (GPS coords, device serial, photographer) so uploaded
photos don't leak per-pixel PII to anyone with a download URL (C1).
- Caps dimensions at 4096px via `resize({fit:'inside',withoutEnlargement:true})`
so a 30000×30000 palette PNG can't decompression-bomb a downstream
sharp decode (C2).
- Re-encode drops polyglot trailers (PDF+JPEG sandwiches that beat
the prefix-only magic-byte check) (H1).
- Freezes animated GIFs to first frame (H3).
Avatar route already funnels through uploadFile so it's covered by
the single change.
**asset-auditor M2 — sanitizeFilename strips RTL/zero-width**
Add Unicode NFC + a strip of bidi-control (U+202A-U+202E, U+2066-U+2069)
+ zero-width chars (U+200B-U+200F, U+FEFF) to `sanitizeFilename`.
Closes the classic Windows-icon-spoof vector
(`invoice_fdp.exe` displaying as `invoice_exe.pdf`) plus folder-listing
collision spoofs.
**datetime-auditor C1 — reminder dueAt drift on every save**
The `<input type="datetime-local">` round-trip in reminder-form.tsx
used `iso.slice(0,16)` (load) and `new Date(value).toISOString()`
(submit). The slice drops the `Z` so a UTC instant is mis-interpreted
as local on load, then converted back to UTC on save — every save
of an existing Warsaw reminder drifted backwards by 2h (CEST). After
two saves the reminder appears at 06:00 instead of 10:00.
Add `toLocalDatetimeLocal(d: Date)` helper that builds the local
YYYY-MM-DDTHH:MM string from getter methods so the round-trip is
TZ-safe. snooze-dialog already did this correctly; the contact-log
dialog also uses the correct localIsoString pattern.
**datetime-auditor C2 — BullMQ cron in UTC, not port-local**
`upsertJobScheduler` defaulted `tz` to UTC. Patterns like
`0 8 * * *` were intended as "8 AM Warsaw" but fired at 09:00 winter
/ 10:00 summer. Pass `tz: process.env.SCHEDULER_TZ ?? 'Europe/Warsaw'`.
Sub-hourly / hourly patterns are TZ-invariant and stay UTC.
**datetime-auditor C3 — report-scheduler never advanced next_run_at**
The minutely scheduler selected `nextRunAt <= now()` and enqueued
generate-report — but never bumped nextRunAt. For weekly/monthly
reports this meant the job re-fired every single minute until a
human zeroed the row out, flooding recipients with dupes.
Now uses `cron-parser` (added as a dep) to compute the next fire
from `report.schedule` and UPDATEs the row BEFORE the enqueue.
Malformed cron expressions disable the row instead of re-attempting
every minute.
Tests 1315/1315. Migration 0058 applied via psql.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Builds on the audit infra split (severity/source) by emitting events
from every place a security or operations review would want to see:
Sensitive data views (severity=warning):
- GDPR export download URL issued
- Audit log page opened (watch-the-watchers; first page only)
- CSV export of expenses
- Webhook secret regenerated
Authentication abuse (severity=warning, source=auth):
- Portal sign-in: success + failed-credentials + portal-disabled
- Portal password reset: unknown email + portal-disabled + bad token
- Portal activation: bad/expired token
Inbound webhook hardening:
- Documenso webhook with invalid X-Documenso-Secret now writes
webhook_failed instead of being silently logged
Background work (source=cron / job):
- New attachWorkerAudit() helper wires every BullMQ worker to emit
job_failed (severity=error) on .on('failed') and cron_run on
.on('completed') for any job whose name matches the recurring
scheduler list. Applied across all 10 workers.
1175/1175 vitest passing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three audit-pass-#3 findings, all in the "wakes you at 3am" category.
- /api/public/health now runs DB SELECT 1 + Redis PING in parallel and
returns 503 + a degraded payload when either fails. Anonymous probes
(no X-Intake-Secret) still get a flat {status:'ok'} so generic uptime
monitors keep working; authenticated probes see the dep results.
- All worker entrypoints (ai, bulk, documents, email, export, import,
maintenance, notifications, reports, webhooks) and src/lib/redis.ts
now use env.REDIS_URL (Zod-validated at boot) instead of
process.env.REDIS_URL!. Previously a missing env let the app start
silently and fail at first job pickup.
- maintenance worker gains an `error-events-retention` case that
delete()s rows older than 90 days from error_events. scheduler.ts
registers it at 06:00 daily. Closes the contract from migration
0040 which declared the table "pruned at 90 days" but had no
implementation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
