Root-cause fix for 'correctness depended on post-deploy manual audits'.
Separate psql-driven suite (pnpm test:integration, own vitest config,
integration/ outside __tests__ so the hermetic pnpm test gate is
untouched; self-skips without a DB). 18 tests run against real Postgres:
- 045 class_signup_tx/class_cancel_tx: RETURNS TABLE->array shape the
actions destructure, capacity/waitlist/dedupe, and the per-session
advisory lock under TWO concurrent connections (D11 proven, never
over-enrolls).
- 042 equipment exclusion constraint (overlap 23P01; cancelled/adjacent
ok). 044 self-change trigger (role + payment cols blocked; service
bypass). 043 members_update WITH CHECK (cross-space move blocked).
Disproves the pass-52 flagged risk (signup rpc-shape). Non-runtime;
default pnpm test still 532/26 hermetic; build clean.
The webhook returned raw signature-lib / Postgres error text to the
caller (signature verify is reachable unauthenticated). Now returns
generic 'Invalid signature' / 'Webhook processing error' / 'Webhook
handler error'; the real detail is console.error'd server-side so
debuggability is kept. Stripe only acts on the HTTP status. 532 tests.
The E2 rate-limit block (f11f4b3) added a second `const h = await
headers()` in submitForm where one already existed near the insert,
breaking the build (redeclaration). Renamed the rate-limit one to
rlHdrs. Build green, 532 tests. (f11f4b3 was committed before its
build was verified — process miss; never deployed.)
Anonymous, captcha-less submitForm was an unbounded DB-bloat vector.
Added a generous per-IP+form limit (20/min) — deliberately loose
because legitimate public signups at a hackerspace commonly share one
NAT/wifi IP, so a tight limit would false-block a room; this still
bounds automated flooding. Captcha (the real control) stays the
documented deferred follow-up. 532 tests.
D10 closed the client-facing leak; this completes the defense-in-depth
so a generic-adapter connection's custom auth-param value is also
scrubbed from the operator-readable door_access_log / callDoor reason
regardless of param name. auth_param added to ConnRow +
loadEnabledConnection + the test/selfEntry selects; threaded into every
callDoor (which already redacts reason/snippet by it) and auditDoor
gained an optional authParam. Native HeatSync (e=) was already covered;
this hardens custom generic adapters. 532 tests.
Deferred deeper-audit P1. validateDoorTarget only string-checked the
host; a bare fetch did its own DNS resolution, so a hostname could
resolve to 127.0.0.1/169.254.169.254 between check and connect. The
executor now resolves once (dns.lookup all), rejects if ANY resolved
address is blocked (isBlockedDoorIp: loopback/unspecified/link-local/
metadata incl IPv4-mapped IPv6; RFC1918/LAN/ULA allowed since door
controllers live there), then connects to the validated IP literal so
fetch performs no second resolution (TOCTOU closed). Original Host kept
best-effort; redirect:'manual' retained. Dependency-free (no undici
seam). 33 door-logic tests; suite 532.
Deferred deeper-audit P1. signUpForClass (count->decide->insert) and
cancelMySignup (cancel->recount->promote) were non-atomic: concurrent
signups at the capacity boundary over-enrolled and concurrent cancels
double-promoted waitlisters (dynamic capacity can't be a DB constraint).
Migration 045 adds class_signup_tx/class_cancel_tx (SECURITY DEFINER,
pg_advisory_xact_lock per session) doing the decision+write atomically;
actions call the RPCs and keep all pre-checks. Removed the now-unused
effectiveCapacity/computeSignupStatus/pickPromotion imports +
isUniqueViolation (pure rules still live + tested in classes-logic).
528 tests; build clean.
D6's pre-truncation capped every space's bucket at PER_SPACE=5 before
the round-robin, so a single-space deployment (the common single-
hackerspace case) drained only 5 emails/min instead of 20 — a 4x
slowdown introduced by the fairness fix. Fairness must bind only under
contention: per-space queues are now bounded solely by the CANDIDATES
fetch and the round-robin itself caps any one space's share when
multiple compete; a lone space drains the full BATCH again. Caught
verifying the D1-D10 batch.
Every door action bubbled the executor's reason/e.message to the
client; redaction was password-literal + a fixed param-name regex, so a
generic-adapter connection with a custom auth param + verbose upstream
error could surface the secret client-side. Client returns are now a
generic 'controller call failed' message; the redacted detail still
goes to door_access_log for operators. redactDoorSecrets + callDoor
gained an optional authParam (scrub the connection's actual auth param
value regardless of name); threading conn.auth_param through the
door-call selects is a deferred defense-in-depth follow-up (client
leak is fully closed; native HeatSync e= is already regex-covered).
prevent_member_self_role_change covered role/status/space but not the
financial columns; RLS members_update lets a member UPDATE their own
row, so payment_status/dues_paid_until/last_paid_at/last_payment_at/
stripe_customer_id/payment_note/joined_at were self-settable via
PostgREST to forge a dues-good signal. Migration 044 adds them to the
blocked set; service-client (auth.uid() NULL) + privileged-on-another-
member writes still bypass (Stripe webhook unaffected). Schema + maps.
UPDATE policy was USING-only: a privileged caller could mutate a member
row into a space they don't administer (no post-image re-check).
Migration 043 adds the mirrored WITH CHECK. Self column escalation
(space_id/role/status) remains blocked by the trigger (D9). Schema +
DB maps updated.
Stripe sends zero-decimal currencies (JPY/KRW/VND/...) already in the
major unit; the unconditional /100 recorded a JPY dues charge 100x
understated in the ledger + receipt. minorToMajor is now currency-aware
(shared ZERO_DECIMAL_CURRENCIES/isZeroDecimalCurrency in
notifications-logic); formatMoney renders no fractional part for those.
Never fires for USD; closes silent ledger corruption for such tenants.
Tests added; suite 526+.
The dispatcher drained a single global FIFO (oldest 20), so one
tenant's billing burst starved every other space's email indefinitely.
Now round-robins an oldest-first candidate window across spaces with a
per-space cap (PER_SPACE=5, BATCH=20). Status writes are guarded with
.eq('status','pending') so an overlapping run can't flip a row another
run already advanced (no false 'failed' on an already-sent row).
Stripe doesn't guarantee event order. A late customer.subscription.
updated carrying an older period overwrote a newer current_period_end,
so graceExceeded flipped a paid member to 'late' and fired a false
'dues lapsed' email + access loss. applySubscription now reads the
stored period and keeps the later of stored vs incoming
(laterPeriodEnd, pure + tested); status still tracks latest sub.status
but grace is computed from the monotonic period.
assignCustomRole inserted {member_id, custom_role_id} with no space
check on either; the RLS WITH CHECK only validated member_id's space,
so an admin of space A could attach space B's custom role to an A
member. Now both the custom role and the target member are verified
in the caller's space before insert. (Blast radius was bounded by
downstream space re-filtering; this closes the scoping defect.)
next was appended to origin unvalidated; next=//evil.com or /\evil.com
redirected off-site post-login. Now only a same-origin absolute path
(single leading slash, no //, no /\) is honored, else /dashboard.
require_approval was silently void: redeeming an admin-role invite into
a require_approval space yielded role=admin,status=unverified and
requireMemberWithRole never checked status -> instant admin pre-approval.
- PRIVILEGE_STATUSES/isPrivilegeEligible (current|late only; excludes
unverified/inactive). requireMemberWithRole now rejects non-privilege-
eligible status with 'pending approval' before the role check; Member
carries status (getAuthMember selects it). requireMember stays
permissive so unverified members still see their own /me.
- joinSpace rejects a second active membership (single-space model;
prevents the .single() self-DoS).
Deferred (HANDOFF): SECURITY DEFINER helpers user_has_role_in_space/
user_has_permission still ignore status (RLS defense-in-depth, own
careful pass). 524 tests; build clean.
reserveEquipment was check-then-insert with no DB guard: two concurrent
requests both passed the in-memory overlap check and double-booked a
machine. Migration 042 adds btree_gist + a GiST EXCLUDE constraint
(equipment_id =, tstzrange && ) WHERE status='reserved', making the DB
the arbiter; the action maps the 23P01 violation to the existing
friendly overlap message. App-side check kept as a fast pre-check.
Schema + DB maps updated.
Independent adversarial audit verdict was NOT SAFE; fixed:
- P0-1 /auth/confirm no longer trusts the query `type` (pinned to
'email_change'); a public endpoint feeding an attacker-chosen OTP
type into verifyOtp was an unintended auth entrypoint.
- P0-2 email sync uses the verifyOtp response user, not a follow-up
getUser() that may not reflect the just-verified identity; /me now
reconciles space_members.email against the authoritative auth email
on load, so a missed sync self-heals (no permanent stale email).
- P1-1 requestEmailChange builds the confirm origin from
NEXT_PUBLIC_APP_URL only, never request headers (forged Host could
redirect the confirmation link / capture token_hash).
- P2-1 isSecretConfigField also catches password/token/private_key/
*_token/*_password so non-conventionally-named credentials can't be
stored in plaintext integrations.config.
521 tests; build clean.
New getMyPayments: payments SELECT is treasurer-scoped (RLS), so the
self-view uses the validated service client strictly scoped to the
caller's space_id + member_id (same convention as getMyBilling /
getMyNotifications). Read-only list (amount, date, platform, status)
in the Membership tab. Docs updated. 521 tests; build clean.
Activity tab: confirm-dialog Cancel on registered/waitlisted class
signups (cancelMySignup) and reserved equipment reservations
(cancelReservation) via the existing actions. Reusable CancelAction
(AlertDialog + toast + refresh). Buttons only shown on cancellable
states; the actions still enforce ownership/space scoping server-side.
521 tests; build clean (UI not browser-tested from here).
Wires existing updateMyProfile + discloseAffiliations into an editable
form (name, handle, phone, bio, skills, interests, willing-to,
affiliations). Affiliations re-disclosed only when changed so the COI
timestamp is not reset on every save. Added bio to the updateMyProfile
param type (schema + column already supported it). 521 tests; build
clean (UI not browser-tested from here).
/me restructured into a 3-tab portal (Tabs primitive): Profile
(permissions), Membership (dues + certifications), Activity
(notifications, classes, reservations, cards, visits). Server page
keeps all data fetching; sections moved verbatim into a client portal
component. No behavior change; sets the frame for P3b-e. 521 tests.
/auth -> /auth/callback and /api/cron -> /api/cron/notifications.
Both prefixes whitelisted more than the single route that needs
anonymity; a future /auth/* page or /api/cron/* job would have been
silently public. Verified those are the only routes under each.
PayPal transaction filter + payments mapping + external-id dedupe were
inline and untested, and auth was a hand-rolled select instead of the
shared helper. Extracted to pure lib/paypal-logic.ts (also fixes a
latent throw on a missing/invalid PayPal date by falling back to now);
route now uses requireMemberWithRole. 7 tests; suite 521. Behavior
otherwise unchanged.
The webhook's money/status mapping (Basil field relocations, period
extraction, customer-id unwrap, status patch, invoice->payment row,
minor->major) lived inline in the route with zero test coverage - the
biggest untested critical path in the codebase. Extracted to a pure,
SDK-free lib/stripe/webhook-logic.ts (structural input types) and
refactored the route to only do signature verify + network + DB writes.
12 new tests cover the Basil edges and link/unlink + defaults. Behavior
unchanged; suite 514.
PayPal client_secret (and any secret-named integration field) was stored
plaintext in integrations.config, unlike Stripe/door which use the
AES-256-GCM vault. A DB read or backup leaked it in cleartext.
- lib/secrets/vault.ts: canonical generic storeSecret/readSecret +
isSecretConfigField (the proven Stripe vault logic, generalized).
- saveIntegration: secret-named fields go to the vault, only a *_ref is
kept in config; blank submit preserves the ref (write-only like
Stripe); legacy plaintext auto-migrates into the vault on next save.
- paypal/sync reads client_secret from the vault, transitional fallback
to legacy plaintext until the next save.
- Docs updated (API_REFERENCE, ARCHITECTURE 10.3). 502 tests.
Independent audit verdict was NOT SAFE; all findings fixed:
- dispatcher Idempotency-Key is now per-attempt (${id}:${attempts}),
not row id alone, so a transient-failure response cached by Resend
cannot permanently suppress every retry and silently lose an email.
- enqueueDues is fully try/catch wrapped so a notifications-table or
render failure can never throw into / wedge the Stripe money path.
- dues_lapsed dedupe falls back to the subscription id when periodEnd
is null (canceled/deleted subs), so future lapses no longer collapse
to one key and get suppressed. New test; suite 499.
- documented the dispatcher's once/min concurrency assumption.
POST /api/cron/notifications drains <=20 pending rows/min, sends via the
transport with the row id as Resend Idempotency-Key (overlap-safe),
marks sent or increments attempts (retryable -> stay pending, permanent
or budget-exhausted -> failed). Constant-time CRON_SECRET bearer check;
503 when unset. ~4.5/s spacing under Resend's 5/s limit. proxy.ts
whitelists /api/cron (no session; own shared secret). CRON_SECRET added
to .env.example.
invoice.paid -> dues_renewed; invoice.payment_failed -> dues_payment_
failed; transition to late (applySubscription) -> dues_lapsed. The
webhook only writes the outbox row (ignoreDuplicates on the
(space_id,dedupe_key) unique index) so a Stripe event replay is a no-op
and the money path stays send-free, fast, and retry-safe. Recipient +
space name resolved server-side; manage link from the request host.
lib/email/send.ts: single sendEmail() over Resend's HTTP API (no SDK
dependency). Returns a retryable flag so the dispatcher distinguishes
transient (429 / 5xx / network) from permanent (config / validation)
failures. Unset RESEND_API_KEY/EMAIL_FROM is a clean non-retryable no-op
so a fresh clone still runs. Idempotency-Key wired for retry-safe sends.
Env documented in .env.example. 6 tests (fetch mocked).
Migration 041 + schema.sql section 26: notifications outbox (service-only
writer, admin/board/treasurer SELECT, member self-view via validated
service action like member_billing; (space_id, dedupe_key) unique makes
the webhook enqueue idempotent). lib/notifications-logic.ts is pure:
deterministic dedupe keys, dues email rendering (HTML-escaped, brand-
neutral), terminal-attempt rule. 13 tests. Inert until P2c wires it.
proxy.ts auth-gates every non-public path and 307-redirects to /login.
The Stripe webhook is called without a session, so it was being
redirected; Stripe does not follow redirects, so every webhook delivery
would have failed. Add /api/stripe/webhook to PUBLIC_ROUTES (scoped to
the webhook path only) and correct the route/header comments that wrongly
claimed there was no auth middleware. Caught by smoke-testing the
deployed route (POST returned 307 instead of 400).
The droplet deploy runs pnpm install --frozen-lockfile; hand-pinning
stripe to an exact version in package.json without regenerating the
lockfile (it kept ^22.1.1) failed the install step. Lockfile now
matches. (Prod was unaffected — deploy.sh set -e aborted before any
swap.)
Moheeb Zara