Probing the live collector to settle the delete-sequence question turned up
something else: usage.picpeak.app answers a valid usage.v2 registration with
INVALID_PACKET while the identical v1 flow is accepted. It does not speak v2
yet — which the deployment notes already require, but the consequence of
getting that order wrong was worse than "reports do not send".
Opting in to v2 against a v1-only collector left the installation stuck.
Registration was refused, so nothing existed at the collector at all; the row
sat in activation_pending, disable moved it to deletion_pending, retry was
futile forever, and enable refused because the row was not `disabled`. The
abandon hatch added earlier did not apply: it was gated on
SIGNING_KEY_UNREADABLE. So the most harmless possible failure — nothing
registered anywhere — was the one an operator could not clear.
The gate is now the property that actually matters: a participation the
collector has provably never accepted (sequence 0, no receipt) with a failing
delivery can be discarded, from activation_pending as well as
deletion_pending. Its receipt records `never-registered` rather than an
unconfirmed deletion, because nothing remote exists to be unsure about. A
participation the collector *did* accept keeps the old narrow gate and its
explicit warning — clearing local state while the collector still holds
reports must stay a deliberate, warned-about act.
A collector that rejects a registration or a deletion outright now reports
SCHEMA_NOT_ACCEPTED instead of DELIVERY_FAILED, and the settings page says the
collector does not accept this report version yet. Retrying cannot fix that,
and sending the operator to look for a network fault they do not have was
wrong.
Verified end to end against the live collector: v2 opt-in reports
SCHEMA_NOT_ACCEPTED, the exit is offered immediately, the receipt says
never-registered, and joining again on v1 registers, reports and withdraws
with a collector-confirmed deletion.
The dialog described what is sent and where it goes, but never said which way
the connection runs. That is the part an operator is actually being asked to
accept: opening an outbound path to someone else's service.
PicPeak sends and never pulls. One place in the service reaches the network,
it is a POST, and it requests exactly two paths — /api/envelopes, and
/api/participant/lookup only when an operator asks for their own export. No
scheduled job contacts the collector; the daily rollup is driven solely by an
authenticated admin hitting /activity. There is no route the collector could
call, and redirect: 'error' means it cannot even point a request somewhere
else. From a reply only the acknowledgement for the packet just sent is read,
with every field compared against that packet before it is accepted; the
stored copy drops the session token and no read path hands it back to the UI.
A requested export is streamed to the operator as a file and never
interpreted.
The consequence is why it belongs in the consent text and not only in the
docs: this channel cannot deliver code, configuration or content into an
installation, not even from a collector that has been taken over. It is a
security property by design rather than by convention.
usageOutboundOnly.test.js guards it by source inspection rather than
behaviour, because a behavioural test only proves that today's calls behave.
It fails the moment someone adds a second fetch, a poll for messages, a
scheduled pull, or a public route touching the usage service — verified by
injecting each of those.
A QA exploration of this branch against an isolated rig — own stub
collector, SQLite and PostgreSQL — turned up one dead end and a set of
signals and controls that did not hold up. This closes all of them.
Rotating JWT_SECRET, the documented response to a suspected compromise,
made the signing key unreadable. That was already named and documented,
but it left no way out: the delete packet can never be signed, so the
row stays deletion_pending forever, and enable() refuses because it is
not `disabled`. An operator who rotated precisely because the secret was
compromised cannot restore it, so the feature was bricked with no
control left. POST /usage/abandon is offered only in that state; it
drops the local identity and records the receipt as
`collector-unconfirmed` rather than claiming a deletion that did not
happen.
Every failed delivery was retried on the next admin request, and
/activity is open to any authenticated admin while the settings ticker
fires it every five minutes per open tab — 30 activity calls against a
rejecting collector produced 30 outbound requests. Migration 206 adds
attempts/next_attempt_at and the unattended sender honours the gate;
Retry and opt-out still send immediately, and the tab names the time of
the next automatic attempt.
Feedback, votes and portal sessions now share an installation-wide
budget of 30/hour. They are the only endpoints whose effect is outbound
traffic carrying operator-written free text, and the general limiter
skips authenticated requests by design. Reading status and withdrawing
stay unthrottled.
gallery_image_protection was true on a bare install with no galleries:
PicPeak ships default_protection_level='standard' and
enable_devtools_protection=true, so it reported fleet-wide 100% and
could never separate a decision from an untouched default. It now reads
only what deviates from the shipped defaults, and the devtools flag is
not read at all — being on by default, its only informative state is
off, which is the opposite of what the key claims.
Also:
- the export receipt counted every packet and called the total "usage
reports"; reports and participant operations are now counted and named
separately
- GET /usage/preview no longer persists the custom_css marker, so the
transparency view stops changing what will be sent
- the feedback route requires every field the packet schema requires,
so an API caller gets the missing field named instead of a bare
INVALID_PACKET from inside signing
- the German strings for this feature use "Sie" throughout, matching the
rest of the admin UI; the ignore hint says what ignoring will do
rather than stating it as already true
- the consent dialog returns focus to the control that opened it
- the long buttons wrap instead of running off a 390px viewport
- a deletion receipt is labelled as belonging to an earlier
participation while a new one is active
Regression tests cover each of these, including the delete packet's
reuse of the last accepted sequence, which was an unwritten assumption
about the collector rather than a defect.
Two things reported on the reformatted consent modal.
The green box is gone. Setting "what is never included" apart as a
tinted panel broke the rhythm of the sections and read as an arbitrary
highlight rather than emphasis. All six sections are uniform now; the
icon and heading are enough to tell them apart.
The green bars across the disclosure were a focus ring, not a border.
showModal() focuses the first focusable descendant, which since the
reformat was the scrollable region I had given tabIndex={0} — so its
inset ring was drawn for every user the moment the dialog opened, and
because the dialog clips its sides a full-width inset ring appears as
two coloured bars. Focus now goes to the dialog itself, which is also
the better screen-reader behaviour: the title is announced on open, and
the region's ring appears only when someone deliberately tabs to it. It
is a thinner, softer ring for that case. The dialog suppresses its own
ring, since that focus is programmatic rather than keyboard navigation.
The collector shown in the transport sentence was never wrong: it
interpolates the configured collector, and the screenshots showing
http://127.0.0.1:9 were taken on a rig deliberately pointed at a dead
loopback port so they could not reach production. Re-checked with
USAGE_COLLECTOR_URL unset: the sentence reads
https://usage.picpeak.app and both links resolve there.
Refs #1110
It was seven anonymous paragraphs in one scrolling block, with the title
and the buttons scrolling away with them. The scroll container is
keyboard-focusable, and unstyled it drew a default focus ring, so the
disclosure also looked like a giant textarea.
Now: a fixed header carrying the icon, title and purpose; a scroll
region with six labelled sections, each with a small heading and icon so
the thing can be scanned rather than only read; and a fixed footer with
the consent checkbox and the actions, which no longer scroll out of
reach on a short screen. "What is never included" is set apart as a
tinted block, since it is the part that answers the question an operator
actually has. The focus ring is now a deliberate inset ring on a
labelled region, which is correct for keyboard use instead of an
accident that looked like a form field.
Dark mode is fixed as part of this, and it was my own doing: the dialog
used `bg-theme-surface`, which does not follow dark mode, and the
section text I added carries dark: variants. Light surface plus
near-white text is unreadable. The surface is class-driven now —
neutral-800, which is what `.card` resolves to in dark and what the rest
of the admin UI uses. Verified in both themes through the app's own
theme toggle rather than by forcing the class, which is what produced
the misleading half-state the first time I looked.
Six section headings added in EN and DE.
Refs #1110
Three items, one of which explains an error seen in the app.
"The operation could not be completed" could come from a config typo.
status() called collectorUrl() bare, and that throws on a bare hostname,
a path, a query, or http in production. The settings page renders one
generic failure when its status query errors, so a misconfigured
USAGE_COLLECTOR_URL replaced the whole tab with that sentence — no
cause, and no way to read the status or withdraw, because every control
there sits behind that call. The URL is now reported as
collector_error: 'INVALID_COLLECTOR_URL' beside the real state, the tab
says what is wrong and how to fix it, and the links are only rendered
when there is somewhere to point them.
gallery_layouts reported grid for every preset-themed install.
color_theme holds either a theme object or the NAME of a preset — the
theme picker stores names, and eventTypeService seeds them
(`theme_preset: 'corporateTimeline'`). Only reading value.galleryLayout
made masonry, timeline, mosaic and the two gallery presets invisible.
Names now resolve, and an event with no theme of its own resolves
through the global one instead of being counted as grid. Only the
name -> layout mapping is duplicated, not the presets;
frontend/src/types/theme.types.ts stays the source of truth, and an
unknown name reports `other` so a preset added later degrades to
"something else" rather than quietly inflating the grid count.
custom_css missed CSS applied through a template. An enabled
css_templates row applied via events.css_template_id is gallery styling
by the same definition as the settings fields — the Custom CSS tab is
where both are authored — but neither the snapshot nor the middleware
saw it, so those installs reported custom_css entirely false. Existence
only; template contents are never read.
Eleven tests. Reverting each fix in turn fails 3, 1 and 3 of them.
Refs #1110
Follow-up review on the previous commit, including a hole in that
commit's own fix.
The cancellation flag became a counter. Clearing a boolean needed a
write of its own, and a /disable landing between the lease and that
write was erased — the same race one level down. enable() now records
the counter it started with and claims only if it is unchanged, so no
clearing write exists to lose. It also fixes the case a boolean could
not express at all: a stale cancellation already set, and a fresh one
arriving mid-activation, are indistinguishable as flags and obvious as
counts. Migration 203, separate from 202 for the reason 202 was separate
from 201 — knex will not re-run an applied migration.
deliver() re-checks immediately before dispatch. The existing check ran
before the binding lookup, which is asynchronous, so a withdrawal that
COMPLETED during it still had its registration or report sent
afterwards. Not an already-in-flight request — a new one started after
the operator had withdrawn.
The outbox writes in tick() and command() are conditional on still being
active. /disable clears pending_packet without holding the lease, so an
unconditional write put a report — or a feedback body and name — back
into an outbox the withdrawal had just emptied, where deliver() would
then leave it, since it declines to send anything but the delete.
Per-item name consent resets with the item. `named` stayed checked after
submitting, so the next item carried the previous name automatically,
contradicting the anonymous-by-default promise the disclosure makes for
each item. The remembered name stays in preferences; attaching it is
decided again each time.
Two of these tests were worthless when first written and are noted
because the pattern keeps recurring: the pre-dispatch case passed
without the guard because an empty report payload failed schema
validation during signing, so nothing reached the collector for reasons
unrelated to the check. With a valid payload it fails without the guard
and passes with it. Same for the counter: dropping it from the claim
fails two.
Refs #1110
It appears on the dashboard and settings only. It is an invitation, not
an alert, so it belongs on pages an admin opens deliberately rather than
on top of whatever task they are in the middle of.
The activity ticker deliberately did NOT move with it. That ticker is
what triggers the daily rollup — the backend has no scheduler — so
tying it to the banner would have stopped reporting for an admin who
works on Events and never opens the dashboard, and stopped it entirely
for a participating install, where the banner never renders at all. The
effect stays mounted on every admin page and only the visible aside is
scoped. Two tests pin exactly that, because it is the kind of thing a
later refactor would helpfully "clean up".
Highlighted like the migration banner it sits under: tinted surface,
border, icon, a title line above the body. It was previously the same
neutral surface as the page behind it and read as filler.
"Not now" is now "Ignore". The button calls dismiss(), which persists
notice_dismissed on the server — the invitation never comes back. "Not
now" promised otherwise. The label says what happens and a hint says
where to join later.
Only shown while participation is off. activation_pending,
deletion_pending and identity_conflict are in-flight states the settings
page explains properly; inviting someone to join in the middle of their
own withdrawal would be worse than saying nothing.
The first version of these tests was worthless: the negative cases
asserted absence after waiting only for the status call, so the
component was still rendering null for want of data and every one passed
with the gates removed. They now wait for the query cache to fill.
Removing the route gate fails 4; removing the status gate fails 6.
Refs #1110
Three of four findings from the follow-up review.
A withdrawal is no longer clobbered by the instance-copy check. That
update was unconditional, so an opt-out arriving while the binding
lookup was in flight was replaced by identity_conflict — and tick()
stops there, so the deletion the operator asked for was never sent. It
now carries the same whereNot('deletion_pending') guard the
collector-conflict handler beside it already had.
Scheduled database backups count as a configured backup. The middleware
records /backup/* and /database-backup/* under one capability, but
`configured` read only backup_enabled, so an install whose only backup
is the scheduled database one reported used: true, configured: false —
a contradiction in the dataset this feature exists to produce.
SIGNING_KEY_UNREADABLE gets its own message. Naming the error in the
previous commit was half the job: the settings page still showed the
generic retry/disable advice, and neither action can succeed without the
original encryption material. It now says what happened and what is
actually required, in EN and DE.
NOT fixed, and reported instead: /disable overlapping an in-flight
/enable. While activation is still doing its slow work the row still
reads `disabled`, so disable's conditional update matches nothing and
the lease conflict from its tick() is swallowed — the operator is told
participation is off while activation completes and leaves it on.
Closing it properly needs a persisted cancellation flag that enable
checks before finalising: taking the lease cannot help, since it either
conflicts immediately or would block the request for the 60s lease. That
is a schema and state-machine decision for the author, not something to
restructure underneath them.
Refs #1110
Review follow-ups on #1304.
SIGNING_KEY_UNREADABLE. USAGE_ENCRYPTION_KEY defaults to JWT_SECRET, so
rotating JWT_SECRET — the correct response to a suspected compromise —
makes the stored Ed25519 key undecryptable. That surfaced as a generic
DELIVERY_FAILED which retried forever, and it silently blocks the DELETE
packet too: an operator who withdraws has their local state cleared
while the collector keeps its copy. decrypt() now tags its own failure
and deliver() reports it under its own name, without flagging an
identity conflict — an unreadable key is not evidence of a clone. The
docs already warned that losing the key breaks deletion signing; they
now name the trigger and the error.
The collector default is no longer an inline string in the constructor.
It is a declared DEFAULT_COLLECTOR_URL, since it is a deployment choice:
self-hosters point USAGE_COLLECTOR_URL at their own collector and the UI
already derives every link from whatever is configured. schema.cjs is
deliberately untouched — it is vendored byte-identical with
picpeak-usage, and its $id is a schema identity, not a delivery address.
Links in the consent dialog. It named the collector inside prose but
never linked it, so an operator deciding whether to opt in could not
open the destination or the public schema without retyping a URL. Both
are links now, built from the configured collector.
UI standards. The tab hand-rolled its surfaces as
`<section className="rounded-xl border border-theme …">` and imported
Button from a deep path; every other settings tab uses `<Card
padding="md">` from the components/common barrel. Converted, with the
feedback <form> wrapped rather than replaced so its semantics survive,
and headings given the same colour tokens as ImageSecurityTab. The
barrel pulls ErrorBoundary -> i18n/config, so the tab's test needed the
initReactI18next shim the FaceRecognitionCard test already uses.
Not changed: the delete packet reusing the current sequence. The
collector handles delete before any sequence check — "possession proof
is sufficient for deletion, including when a restored backup has a
stale sequence" (picpeak-usage server/collector.js) — so deletion is
deliberately sequence-exempt and the client is correct as written.
Refs #1110
Two review findings, both about the warning saying things that are not
true.
The duration contradicted the rate beside it. `rate` came from the live
draft while `minutes` came from the resolution the server computed from
the SAVED rate, so any unsaved edit produced a mismatched pair: 120
recipients switched from 10/min to 1/min still claimed 12 minutes rather
than 120. Recomputed from the rate the send will actually use — queueing
persists the draft first — with the server's own formula and the same
1..10 clamp, so the two numbers cannot disagree.
The queue claim was overstated. "other email queues behind it" is only
true when the campaign saturates the queue: scheduled_at is staggered
and processEmailQueue excludes future rows before taking its batch, so a
campaign paced below the 10/minute ceiling leaves capacity for a
password reset on the next tick. Now says other email CAN be delayed
behind it, which is what the shared queue actually guarantees.
The new test fails against the previous pairing.
Both found by review against the correct base, and both are cases the
second stripRemoteCssUrls pass had been catching before this PR removed
it. Verified against the real functions before and after.
An escaped quote outside a string. `\'` is an escaped identifier
character, not a string opener, but the scanner stepped onto the
apostrophe, entered string mode and copied the rest of the stylesheet
unexamined — so `.hero{--marker:\';background:url(https://evil/p.gif)}`
kept a live remote URL. Escapes are now consumed as a unit outside
strings.
An unterminated quote. Trusting one meant a single stray apostrophe
disabled scanning for everything after it. An unclosed quote is a parse
error, so the safe reading is to emit it as an ordinary character and
keep scanning; a newline also ends a string, as it does in CSS.
The entity mismatch behind the second case. sanitize-html writes `"`
inside an attribute as `"`, so the scanner and the recipient's
browser disagreed about where strings begin: in
`style="font-family:"don't";background:url(...)"` the browser
decodes first, reads the apostrophe as ordinary text inside a real
string, and fetches the background — a tracking pixel by another name.
Style attributes are now decoded before scanning and re-encoded after,
which also stops the old code silently deleting quotes from the value.
Also detaches the image handlers before releasing the canvas source.
That one did NOT reproduce: measured in both Chromium and WebKit,
neither fires `error` when the attribute is removed after a successful
load. Applied anyway because the ordering is free and the failure it
would cause is silent — canvasFailed set, the canvas swapped for an
<img>, and the image decoded a second time, the exact opposite of what
the release is for.
Refs #1264, #1287
Closes#1300.
Fragmentation was configurable, stored per event, served to the gallery
client, and consumed by nothing. It was not unbuilt scaffolding — both
halves exist and are individually coherent — but they were never
connected, and they disagree: the server cut a fixed 3x3 grid while the
client reassembled a 4x4 one, so wiring them together as they stood
would have produced scrambled images rather than protection.
Removed rather than finished, because finishing it buys nothing. The
client fetches the whole image and then redraws it in pieces on a
canvas, so the full original has already crossed the wire before any
"protection" is applied — that is obfuscation, not a control. The
per-fragment canvas work also lands on mobile, which is the memory
profile under investigation in #1287.
Goes: secureImageService.fragmentImageBuffer and its branch, the
?fragment=N delivery path and handleFragmentedImage in secureImages,
the fragmented-JSON response in protectedImages, fragmentation_level in
the gallery payload, the default_fragmentation_level setting, the PUT
validator, the ProtectedImage fragment renderer, and the operator
control with its strings in all eight locales.
No migration. `events.fragmentation_level` and the app_settings row stay
— dropping a column is irreversible and the stored values are harmless
once nothing reads them. If they should go, that is a deliberate data
decision and its own migration.
`fragmentGrid` on AuthenticatedImage and the layouts is deliberately
untouched: #1299 already removes it as part of the inert prop surface,
and doing it here would only collide.
The composer already paces sends — the rate is clamped to what the queue
can actually drain and `scheduled_at` is staggered — and the help text
said so. But pacing answers the wrong risk. Spam filtering reacts to a
domain's volume and reputation, not to the interval between messages, so
a throttled send of several hundred near-identical mails from a domain
that normally emits only gallery notifications is exactly the shape that
gets junked or blocked. Nothing told the operator that.
Adds a warning above the queue button once the resolved recipient count
reaches 50, covering what actually goes wrong: the reputation hit, that
it damages delivery of transactional email too, the SPF/DKIM/DMARC
prerequisite, and the suggestion to split a first campaign.
It also states the cost of the send in the operator's terms — the real
duration at the chosen rate, and that gallery invitations and password
resets queue behind it, because the email queue is global rather than
per-campaign.
50 is deliberately low: the operators who most need this are the ones
sending their first campaign.
The existing rate hint now mentions looking like spam, not only being
rate-limited.
Review follow-up. The release only ran from the effect cleanup, so it
fired on unmount or a src change — while the commit message and the test
header both explained that grid tiles never unmount, which is the whole
reason the decode piles up. For the case the change exists for, it never
ran at all.
drawToCanvas now reports whether it drew, and the source Image is
released as soon as the pixels are on the canvas. Nothing redraws from
imageRef afterwards; drawToCanvas has exactly one caller. The cleanup
stays as the fallback for the paths onload cannot cover: the draw
failed, or the source changed before onload fired.
The new test pins release while still mounted, on the same src. It fails
against the previous version.
Refs #1287
AuthenticatedImage accepted the whole image-protection prop surface and
discarded it in a `void unusedProps` block. Callers computed those props
from the event's protection level and passed them in good faith, so
raising the level produced canvas rendering (via the layouts' own OR on
`protectionLevel === 'maximum'`) and nothing else the level implies.
Removes them from the interface and from every call site, so the props
state what the component actually does. Two survive because they are
real: `useCanvasRendering`, and `onProtectionViolation` — which #1297
listed as inert but which does fire, from the canvas context-menu
handler. `useWatermark` is removed as well; #1297 did not list it (it sat
outside the `unusedProps` block) but it was equally dead.
Removal rather than implementation is deliberate. The implementation
these props describe already exists in `ProtectedImage`, which is
exported from the barrel and rendered nowhere. Wiring it in is a product
decision about what protection level should mean, not a side effect of a
cleanup.
Analytics payloads inside the surviving onProtectionViolation handlers
keep their photoId/protectionLevel fields.
Refs #1297
Grid was the only layout passing `lazy` without an `inViewRootMargin`, so
PhotoCard ran its observer at the IntersectionObserver default of `0px`
with `threshold: 0.1`. A tile could not begin loading until a tenth of it
was already on screen — there was no lead at all.
The gallery owner's account of the symptom is that defect's exact shape:
spinning the scroll wheel outran loading by roughly 50 images, then it
caught up. Outrun-then-recover is what a zero-width pre-load band looks
like from a chair.
This is the one thing in that investigation that does not rest on the
reporter's instrumented runs, which they have since withdrawn after
finding their automation harness ran in a hidden pane — `innerHeight: 0`,
so nothing could intersect and no tile could ever load. The missing
margin is visible in the source regardless.
Percent, not vh. `rootMargin` accepts only px and percentages, and a `vh`
value throws SyntaxError at construction, which would have taken down
every Grid gallery. Verified in Chrome:
'100% 0px' → accepted
'100px 0px' → accepted
'100vh 0px' → SyntaxError: rootMargin must be specified in pixels or percent
A percentage resolves against the root's own box, so 100% is one viewport
height of lead in each direction — viewport-relative, which a fixed 100px
like Justified's is not. A phone and a 4K desktop scroll past very
different amounts of grid per gesture.
Deliberately NOT included: a sweep for cards left un-loaded after
scrolling settles. That was aimed at permanent loss from `triggerOnce`,
and the owner's observation that tiles do come back on desktop argues
against it. Complexity chasing a symptom nobody has reproduced outside a
broken harness.
Three guard tests, including one on the unit, since the failure mode of
getting that wrong is a gallery that does not render at all.
Two follow-ups to yesterday's merges. Both were already known; neither
depends on the open question in #1287.
1. Canvas mode pinned every decoded image for the component's lifetime.
`AuthenticatedImage` keeps a detached Image in `imageRef` so drawToCanvas
can read it. The effect cleanup nulled onload/onerror and never cleared
that ref, so the Image — and the decode behind it — stayed held by a live
JS reference. A decoded <img> in the document is evictable under memory
pressure; one held by a ref is not.
That is not academic at gallery scale. The photo grid is NOT virtualised,
so a 546-photo event mounts 546 of these and none ever unmount — nothing
was ever released. The ref is cleared and the src dropped, so the browser
can reclaim without waiting for GC.
This is NOT presented as the fix for #1287. That investigation is still
open: the reporter has since shown the backend idle during a stall and
the renderer itself unresponsive for 45s, which rules out the theories
tried so far. This is a real leak on the same path, worth fixing on its
own terms while that question is settled.
2. newsletterService no longer carries its own remote-url() stripper.
It was added because the shared sanitizeCSS "blocked" remote URLs with a
CSS comment that parsers discard. #1290 replaced that with a lexer, so
the local copy is dead weight — and two definitions of "disallowed" would
drift apart. Verified the shared function covers every case the local one
did, including the quoted-paren and CSS-escape forms found in review.
Three tests on the release path, two of which fail without the fix:
unmount clears the ref and drops the src, the blob URL is revoked, and a
src change releases the previous image rather than accumulating one
pinned decode per photo a recycled tile has shown.
Part B of #1264. Flag off by default, so an install that never enables it
gains no route, no nav entry and no way to mass-mail.
A campaign is a body plus a recipient rule. Queueing one writes ordinary
email_queue rows (email_type 'newsletter', origin 'campaign', new
campaign_id), so retry, rendered_html, sent_at and error_message all come
from the existing processor rather than a parallel sender. Throttling
staggers scheduled_at; the processor loop is untouched.
Two rules the service enforces: no raw HTML is ever stored (sanitized on
write and again on render, idempotently), and opt-out is checked at queue
time AND again at send time.
Migration 199 adds email_campaigns, email_campaign_recipients,
email_queue.campaign_id, customer_accounts.marketing_opt_out(_at), and
the newsletters.view / newsletters.send permissions.
Three rounds of external review are folded in, including several that
would otherwise have shipped broken:
- Campaign rows never came due on SQLite. queueEmail writes a Date, which
the sqlite3 binding stores as epoch ms; ISO text in the same column
compares as TEXT against an INTEGER, and SQLite orders every INTEGER
below every TEXT. The feature silently sent nothing there.
- The flag had no Settings card and no sidebar entry, so it could not be
enabled through the UI at all.
- Consent is per ADDRESS, not per row: two accounts sharing an inbox meant
unsubscribing stopped one and not the other, at both queue and send time.
- The unsubscribe GET mutated consent, so a mail-security scanner walking
a campaign could have unsubscribed much of the list. GET now confirms,
POST acts.
- The rate ceiling is clamped to the queue's real throughput (10/min), so
the composer's estimate stops being wrong by up to 12x.
Closes#1264
sanitizeCSS "blocked" a remote url() by prefixing it with a
/* BLOCKED URL */ COMMENT and leaving the URL in place. CSS comments are
discarded during tokenization, so the declaration a browser parsed still
carried the live URL — while adminCssTemplates returned
sanitization_warnings claiming it had been stopped. Protection that
reports success is worse than none, which is why it survived review.
Scope is narrow: sanitizeCss (lowercase, the public-site path) never
included the pattern and permits remote URLs by design — a test now pins
that. Only sanitizeCSS (uppercase) was affected; outside this repo's
newsletter branch its sole caller is adminCssTemplates.js.
Migration 200 is required, not cosmetic: gallery.js serves
css_templates.css_content VERBATIM as text/css and does not re-sanitize on
read, so fixing the write path alone would leave every existing template
serving its URL forever.
Review follow-ups replaced the regex with a small three-state lexer
(comment / string / identifier) over the RAW text, after five further
bypasses: a ")" inside a quoted url(), CSS escapes (u\72l), the HTML
comment strip JOINING tokens into a live url() after the scan, an escaped
quote desynchronising the scan, and a quote inside a comment. Escapes are
decoded only to decide, never to rewrite — a clean input now round-trips
byte-identical, which also keeps unaffected rows out of the migration's
write path.
Severity is low (writing a template needs branding.edit) but the harm is
a gallery visitor's IP reaching a third party from a page the operator
believes carries no remote requests.
tailwind.config.js had plugins: [] and @tailwindcss/typography was never
installed, so every prose / prose-neutral / dark:prose-invert class in the
app resolved to nothing. Preflight, which IS active, resets h1-h6 to
inherit size and weight and strips list-style from ul/ol — so an applied
<h2> rendered pixel-identical to the <p> it replaced.
The editor was never broken. The toolbar highlighted because
editor.isActive('heading') correctly returned true; only the CSS to show
it was missing. That also explains why pasting rendered rich text worked:
it carries inline styles.
Ten surfaces rely on these classes, including the PUBLIC CMS pages — so
impressum/datenschutz were serving unstyled headings to visitors too.
Review follow-ups: prose colours are mapped to the theme tokens wherever
.text-theme marks theme-owned text (a dark gallery preset sets
--color-text but no .dark class, so dark:prose-invert never engages and
headings would have gone near-black on dark); code blocks inherit rather
than being scaled twice; and H5/H6 get explicit rules, since the plugin
only styles h1-h4.
Closes#1288
Hardening for the large-gallery stall. The reporter could not isolate the
cause and neither could I from static reading; these are two defects that
are wrong independently of whether they are the whole story.
Gallery grids are NOT virtualized: a 546-photo event puts 546 PhotoCards
in the DOM, each mounting its own bare fetch. Two problems there: no cap,
and a cleanup that only set a flag while the request kept running.
withImageFetchSlot now holds requests to six in flight, with the BODY read
inside the slot — fetch resolves on headers, so releasing there would have
bounded header round-trips and nothing else. Teardown aborts via
AbortController.
A request queued indefinitely is PENDING, not failed, which is why the
failure left no console error, no failed request and nothing in the
backend log.
Review follow-ups: three tiers (current lightbox slide > neighbour
prefetch > grid thumbnails), because a single FIFO put the image the user
just clicked behind hundreds of thumbnails; and a synchronous throw now
releases its slot instead of permanently draining the pool.
If it recurs, capture performance.getEntriesByType('resource') for the
stalled thumbnails — a queued request shows responseStart === 0.
Relates to #1287
The business profile already carried the operator's full issuer block —
address, phone, email, website, VAT id — but none of it reached an email.
Those columns only fed the quote/invoice PDF renderer, so every outgoing
mail footer was the fixed logo + company name + copyright line.
The signature is rendered by wrapEmailHtml and nowhere else, so no
template, no per-type send path and no queue row needed a change. Two new
columns on business_profile (migration 198) carry the toggle and one
free-text legal line; everything else is read from the address fields the
operator already maintains.
Default off, with a test pinning that the disabled path is byte-identical
to a no-profile install.
Includes three rounds of external review fixes: the plain-text MIME part
also carries the signature; string booleans ('false'/'0') no longer
invert the toggle; the status line stays silent rather than asserting
"off" while unauthorised or loading; and the preview's Text tab mirrors
the send path's htmlToText fallback.
Manual Messages replies deliberately keep no signature — they bypass the
wrapper by design — and the UI copy names that exception.
Closes#1264 (Part A)
Adds docs/upload-file-types.md: the single Allowed File Types setting, every
path it governs, the extension-to-MIME table, how to enable video, and what
changed for chunked-upload/init (declared mimeType ignored, extension must
be allowed, 400 File type not allowed). README links to it, and the Settings
help text in EN and DE now says the list covers all upload paths and that
video extensions must be added explicitly.
backend: qs and body-parser (array-limit bypass, isBuffer DoS).
frontend: axios 1.17 line (formToJSON recursion DoS, prototype-pollution
gadgets, maxBodyLength bypasses), dompurify, linkify-it and the transitive
set npm audit fix resolves without a major bump.
Left out on purpose: sanitize-html 2.17.7 (its htmlparser2 12 tree is
ESM-only, which Jest 29 cannot load, and the SVG SMIL advisory needs svg
tags none of our sanitizer configs allow), and the tiptap 2->3 and
react-router 6->7 majors (open redirect via <Link> needs a user-controlled
navigation target, which the SPA has none of).
Closes#1275. Follow-up to #1263.
`matchMedia('(hover: none) and (pointer: coarse)')` answers "what is this
device's primary pointer", which on anything with both inputs is the wrong
question. A touchscreen laptop reports fine+hover, so a finger tap was handled
as a click: the photo opened with no reveal step and the tile's own actions
needed a hover a finger cannot produce. An iPad with a trackpad reports the
opposite, so a mouse click was handled as a tap and opening a photo took two of
them while hovering did nothing.
Pointer events carry the answer per interaction. useInputMode holds one
module-level mode fed by a single window-level listener pair, so every tile
agrees and the listener count does not scale with the grid. The primary-pointer
query stays as the opening guess -- it is right for the two single-input cases
that are most of the traffic, a phone and a desktop -- and the first real
interaction corrects it on a hybrid.
pointermove matters as much as pointerdown: a mouse announces itself by
approaching, and the mode has to be right BEFORE the click, not as a
consequence of it. A pen is grouped with touch, since it taps rather than
hovers on most hardware and being wrong that way costs only a reveal step.
GalleryPremiumLayout's touch rules move off the media query onto a
data-input-mode attribute the layout sets, for the same reason: on a
touchscreen laptop the query stayed false and a finger could never reach the
checkbox or like button, and on an iPad with a trackpad it stayed true and both
were stuck on permanently.
The #1263 guarantee is unchanged and pinned by a test that walks all three
modes: a control that cannot be seen cannot be hit, whichever input is in use.
Verified in the running app under Chrome touch emulation, on a device
advertising a coarse primary pointer -- the iPad-with-trackpad case. A mouse
merely moving switched the grid to hover semantics and revealed the overlay,
and a subsequent tap switched it back; the premium layout's attribute followed,
with its checkbox reachable under touch and hidden-but-inert under mouse. The
mirror case (finger on a fine-primary device) cannot be staged in Chrome, which
couples touch emulation to a coarse primary pointer, so it rests on the jsdom
tests.
12 tests: 8 on the store, 4 more on PhotoCard. 3 of the 4 fail without the
per-interaction mode; the fourth is the #1263 no-regression guard and holds on
both sides by design.
Codex review round 4 on #1273.
Settings → Status rendered a green check for the email processor
unconditionally, against an API field that was itself the literal 'active'.
Both ends were lying and only one of them got fixed: adminSystem started
reporting the real state in an earlier commit, but StatusTab never read it, so
the second place an admin looks to find out why mail is not arriving still said
everything was fine. It now shows stopped and degraded, with the reason.
The truncation flag missed the case it most needed to cover. The loop broke on
the 200-row report cap before the flag could be set, so 201+ overdue rows came
back as exactly 200 with scanTruncated false -- a partial report presented as
complete. It is now set whenever rows were left unexamined.
The grace-window comment claimed the processor clears ~6000 rows inside the
window. It clears on the order of 100: ten rows a pass, one pass a minute. The
comment now says so, and says why the processor's own state is reported above
the list rather than inferred from it -- "running, last pass sent 10" next to a
backlog reads very differently from "not running" next to the same backlog.
One round-4 finding is NOT fixed, deliberately, and is written up at the retry
route. Clearing scheduled_at leaves created_at at the original enqueue time, so
a retried old row appears in the waiting list immediately, looking overdue,
until the processor sends it. Restarting that clock needs a timestamp written
there and no shape works: a Date matches how queueEmail writes the column and
how processEmailQueue compares it, but jest's sandbox Dates store as
"[object Object]" (CLAUDE.md) so it cannot be tested; an ISO string tests fine
but stores as TEXT, which SQLite then orders above the numeric bound in the
processor's own pickup query, leaving the row unsendable. A requeued_at column
would settle it. Cosmetic either way, and not worth risking a stuck row.
1 more test, failing before this commit.
Codex review round 3 on #1273. The first finding reverses a round-1 fix of
mine, correctly.
Retry no longer sends. Round 1 flagged that retry was a no-op for waiting rows
and offered two remedies: give them a send-now action, or stop showing them
Retry. I took the first, and round 3 showed why it is the wrong half --
processEmailQueue claims nothing before invoking the transport, so a flush
overlapping the scheduled pass has both of them sending the same email. Saving
60 seconds is not worth a duplicate landing in a customer's inbox, and a claim
protocol would need a status no query watches plus a reaper for rows abandoned
mid-send. So retry is a reset again, as it was on main.
Waiting rows now carry no actions at all, which is the other half of that
round-1 remedy and closes a worse hole the shared table opened: Dismiss DELETEs
the queue row. Those emails have not failed and still go out once the processor
recovers, so clicking the tidy-up icon on a health warning silently cancelled a
customer's mail. The section is diagnostic; what a waiting row needs is the
processor fixed, which the panel above it now says.
The grace window runs from when a row became DUE, not from when it was queued.
A split-payment invoice created three days ago and scheduled until a minute ago
has had one minute of the processor's attention, and measuring from created_at
reported every scheduled mail as unworked the instant it came due -- which is
most of what this panel would then have been showing.
A truncated scan can no longer read as an all-clear. The scan is bounded, so a
queue larger than the budget whose head is all future-scheduled can hide a due
row past the last page read; the response now says so and the UI withholds the
green check.
The test fixtures were wrong in a way worth keeping: scheduled_at also defaults
to CURRENT_TIMESTAMP, so back-dating created_at alone built rows that cannot
exist in production -- old, but scheduled for the moment the fixture ran. The
helper now back-dates both, as the database would have.
3 more tests; the two that pin new behaviour fail before this commit, and the
reverted flush is pinned by asserting the transport is NOT invoked.
Codex review round 2 on #1274. Both findings are the same shape as round 1:
a message that asserts more than the response supports, and sends the admin
somewhere that makes it worse.
A 5xx is no longer treated as a clean failure. createInvitation inserts the
customer_invitations row and only then queues the email, with no transaction
around the pair, so a 500 out of the queueing step leaves an OPEN invitation
behind. Telling the admin "no invitation went out, retry" there walks them into
a 409 that still queues nothing. 5xx now joins the no-response case as
unconfirmed; a plain 4xx keeps the clean-failure message, because that is the
one shape where nothing was written.
The two already-active conflicts now say so. The send-invite route returns
CUSTOMER_ALREADY_ACTIVE from its own check, but createInvitation rechecks
customer_accounts afterwards and threw a bare ConflictError -- code CONFLICT,
indistinguishable from the pending-invitation conflict. An invitation accepted
between the two checks therefore landed in the pending branch, telling the
admin to cancel an invitation that acceptance had just closed. Both conflicts
in the service now carry a code of their own, so the client reads the code
rather than inferring from the status.
4 more tests. One round-1 test changed with the behaviour it pinned: its 500
now asserts the unconfirmed message, and a new 400 case covers the clean
failure it used to stand for.
Codex review round 1 on #1273. One of the four is a real bug on every SQLite
deployment.
The waiting-row query compared `created_at` against a bound ISO string. On
SQLite that column does not hold a string: queueEmail writes a JS Date and the
native binding stores epoch ms, and SQLite orders INTEGER before TEXT
regardless of value -- so the comparison was true for EVERY row. Mail queued a
second ago read as ten minutes overdue, and a scheduled_at years in the future
read as already due. Confirmed directly against sqlite3: a 2026 row matches
`created_at <= '2020-01-01T00:00:00.000Z'`.
Binding a Date instead is not the fix, since knex hands sqlite3 a Date the same
way and jest's sandbox Dates stringify to "[object Object]" (CLAUDE.md). So the
engine-safe half of the predicate stays in SQL and the two time comparisons
move into JS behind a toMillis() that accepts all three shapes this column
really has -- Date from Postgres, ms-number from SQLite, ISO string from
fixtures and older rows. The scan is capped at 1000 pending rows ordered
oldest-first; everything overdue sorts into that window, and the response was
already capped at 200. The existing tests missed this because they store ISO
strings, which is what CLAUDE.md prescribes for jest -- so the new ones store
epoch ms, the production shape, and one mixes both in a single queue.
Retry was a no-op for the rows it most needed to help. It wrote pending /
retry_count 0 / no schedule, which is exactly what a waiting row already is:
the row came back unchanged while the toast said it had been re-queued. And
since the usual reason a row is waiting is that nothing is working the queue,
deferring it to the next pass is the one answer that cannot help. It now
follows the reset with the same single-row flush the project cockpit uses.
An idle pass no longer inherits the previous pass's totals -- the no-pending
early return skipped the lastResult assignment, so System Health kept
attributing an old sent/failed count to a run that did nothing.
"All clear" now means the whole queue is clear, which is what the PR claimed
and the code did not do. An empty waiting list is only reassuring when
something is working the queue: a processor stopped a minute ago has no overdue
rows yet either, and a green check there is the same false all-clear this
branch exists to remove.
7 more tests. The 5 that pin new behaviour fail before this commit; the SQLite
ones fail in the way the bug predicts rather than erroring.
Both new tests stub the webhook transport with a spy rather than pointing it at
a dead port: real connection attempts left open handles that destabilised
unrelated suites in the same jest worker.
Codex review round 1 on #1274. Three places where this branch replaced one
overclaim with another.
The badge tooltip said "Invitation sent". createInvitation inserts the
customer_invitations row and only then queues the email, with no transaction
around the pair, so an open invitation does not prove an email_queue row exists
-- and even when it does, delivery is the queue processor's business minutes
later. The tooltip now describes the invitation link itself and points at
System health, which is the same distinction #1273 draws.
Both conflicts are 409 and were being treated as one. Migration-era
send-invite returns 409 with code CUSTOMER_ALREADY_ACTIVE when the customer
already has a password, which happens if an open invitation for that address is
accepted between createDirect and sendInvite. There is then no invitation row
to cancel, so directing the admin to the Invitations tab points them at
something that does not exist. The code is now read before the message is
chosen.
A dropped connection or a timeout rejects with no `response` at all, and the
request may well have succeeded server-side. Saying "no invitation went out"
there sends the admin into a retry that then 409s, which is the same trap the
CUSTOMER_ALREADY_ACTIVE case sets. That branch is now explicitly unconfirmed
and says where the answer is.
3 more tests, all 3 failing before this commit.
Codex review round 1 on #1272. Both findings are consequences of extending the
Grid/Justified tap-to-reveal model to every layout: what those two layouts got
away with, because they were the only ones using it, becomes wrong once
Masonry, Mosaic and Timeline inherit it.
The hover variants no longer hide behind `md:`. On a fine pointer under 768px
`isTouchDevice` is false, so nothing reveals the overlay, and the `md:` prefix
disabled the only hover variants there were -- the controls stayed
`opacity-0 pointer-events-none` with no way to reach them. Grid and Justified
already behaved that way, but Masonry, Mosaic and Timeline had unprefixed
`group-hover:` and revealed at any width, so this was a regression for them.
Width was never the real question: what the breakpoint was standing in for is
that :hover latches on a touchscreen once a tile is tapped. So the variants are
emitted for pointer devices only and withheld on touch, which says that
directly.
detectCoarsePointer no longer ORs the touch fallbacks over matchMedia.
matchMedia describes the PRIMARY pointer; `ontouchstart` and `maxTouchPoints`
only say a touchscreen exists somewhere, which is equally true of a touchscreen
laptop or a docked tablet being driven by its mouse. OR-ing them classified
those as touch-only, so an ordinary click merely revealed the overlay and
opening a photo took two clicks. The fallbacks now stand in only where
matchMedia is absent, which is what the comment already claimed.
3 more tests, all 3 failing before this commit.
Paul Nothaft
