GHSA-qxc2-j82w-r537 covers @faker-js/faker <=10.4.0 and is rated high, so
`bun audit --audit-level high` fails and takes the build job with it on every
branch. The repo was on ^9.9.0; the fix landed in 10.6.0.
faker is a runtime dependency here, not a dev one: src/lib/helpers/faker.ts
ships in the console bundle and powers "generate fake data" in the databases
UI. The advisory is specific to helpers.fake(), which this repo never calls,
so the exposure was nil, but the audit gate does not read call sites.
Nothing in src/ needed changing. Every one of the 20 faker APIs used keeps its
v9 signature in the installed v10 typings, including location.latitude and
location.longitude, which still return `number` and so still serialise
correctly into the point and line column payloads.
bun audit --audit-level high clean (18 remaining are moderate/low)
bun run check 0 errors
bun run test:unit 275/275
bun run build success
The 768px crush had two causes: the full-width mobile button claiming the
whole line, and the direction breakpoint sitting at isSmallViewport (767px)
while .is-full-width-mobile applies at 768px. Dropping fullWidthMobile fixed
it on its own; moving the breakpoint to isTabletViewport (1023px) as well
just made the banner stack 256px earlier than it needs to, pushing "Try it
now" below the copy on ordinary laptop widths.
The text is ~64 characters and the button ~100px, which fits comfortably in
768px minus the shell padding, so the breakpoint goes back to 767px.
Measured at 375, 767, 768, 900, 1023 and 1440: one row above 767px, stacked
and centred below it, and main > header offset matches the stack height
exactly at every width.
Replaces the plain text modal with the cover-led layout: a real capture of
the new Console, a one line pitch, the retirement notice, and the two
choices.
The image is the generator export rather than drawn artwork, matching how
the promo card carries its cover. The retirement date is the sentence most
likely to change the answer, so it sits above the actions in Alert.Inline,
the same notice treatment the rest of the console uses, instead of below
them as fine print.
Dropped the three feature bullets. They restated what the screenshot
already showed, and their accent-tinted icon tiles were not a pattern used
anywhere else in the console.
The retirement wording is still a placeholder pending a confirmed date.
Two defects found by running it rather than reading it.
`show` starts false and the parent flips it on, so the reactive exit
handler fired on mount and recorded a dismissal before the modal had ever
appeared. It now only counts as an exit once the modal has actually been
open.
Modal also sets show = false on navigation via autoClose, which this
component read as the user choosing to stay on the old Console. Navigating
is not a choice, so autoClose is off; the only ways out are the close
button, Escape, the backdrop, and the two footer buttons.
Also moves the copy from the description slot into the body, which left a
dead gap above the footer, and drops the modal from size l to m so two
lines of text are not stranded in a very wide box.
bun audit started failing on every branch after an advisory against
js-yaml >=4.0.0 <4.3.2, reached through eslint. The existing override
allowed 4.3.1, which is vulnerable. Pinned to ^4.3.2 rather than >=4.3.2,
which resolved to 5.x and crossed a major boundary.
GHSA-2883-xcg3-v3hh
Users on the old Console get one interruptive prompt with an explicit
choice: try the new Console, or carry on with the old one. Closing it by
any route counts as choosing to stay.
Built on the existing Modal from $lib/components rather than a bespoke
full-screen surface. wizard.showCover() would be genuinely full-screen but
has never had a consumer, so it is untested in production.
Snoozes for a month with backoff via hideNotification, longer than the
banner's week because this is the most interruptive of the three surfaces.
Never shown during onboarding.
utm_medium=modal keeps it comparable with the banner and the promo card.
The fallback and its tests were about 210 of the 247 lines in this branch,
all to handle the case where an account prefs write is rejected. For a
dismissable promo the cost of not handling it is that the banner shows
again on the next load, and the user closes it again.
Reverting leaves dismissal identical to the four other hideNotification
call sites: fire and forget, account prefs as the only store. No second
source of truth, so no way for two counters to disagree.
Keeps the three fixes this branch is actually for: the 768px breakpoint
mismatch, the AlertStack resize listener, and the browserslist pin.
The local fallback recorded a flat one week with no hide count, so a run of
rejected preference writes restarted the sequence every time instead of
continuing it. That contradicted the behaviour described one line above it.
Extracts the fallback into localSnooze.ts, which reproduces the same curve
as notifications.ts: coolOffPeriod * factor ** (hideCount - 1). Repeated
failures now escalate 1w, 2w, 4w, 8w, and a successful write still clears
the local copy so account prefs remain the source of truth.
Reads and writes are guarded, so private mode or a corrupt value fails
towards showing the banner rather than hiding it forever.
Adds localSnooze.test.ts covering the escalation, the formula parity with
account prefs, the expiry boundary, and the corrupt-value case.
hideNotification fired updatePrefs without returning or catching it, so a
rejected write was unobservable: the banner hid locally, nothing was
recorded, and it returned on the next load with the backoff count lost.
updateNotificationPrefs and hideNotification now return the promise. This
is backwards compatible, the four existing callers ignore the return value
and keep their fire-and-forget behaviour.
The banner awaits it and, if the write is rejected, falls back to a local
snooze for a week so the dismissal is never silently dropped. Account prefs
stay the source of truth: they carry the backoff count and follow the user
across devices, and the local key is cleared as soon as a write succeeds.
Registration now goes through canShowNewConsoleBanner, which checks both.
Adds notifications.test.ts covering the returned promise, the rejection
path, the recorded expiry, and the backoff counter.
bun audit --audit-level high started failing on every branch after two
advisories were published against browserslist <=4.28.6, reached through
svelte-preprocess and @sentry/sveltekit. Neither is a direct dependency, so
this pins it in the existing overrides block alongside the other
transitive security pins. Resolves to 4.28.8.
GHSA-c83g-rgw3-j3cx, GHSA-73wf-gq98-2v4g
Three fixes found while testing on production.
Responsive: the layout switched to a column at isSmallViewport (max-width
767px) but Button's .is-full-width-mobile applies at max-width 768px. At
exactly 768px the row stayed horizontal while the button took the full
line, crushing the copy into a ~90px column seven lines tall. Switched the
direction to isTabletViewport and dropped fullWidthMobile so the button
sizes to its content. 768px goes from 182px tall to 96px.
Dismissal: closing the banner wrote a localStorage key that was never
cleared, so one click removed the message for the entire migration. Now
uses hideNotification with a one week cool-off and exponential backoff, and
shouldShowNotification to gate registration. Repeated dismissals still fade
it out, but a stray click no longer opts the user out permanently.
Resize: AlertStack owns the shell offset for alerts inside it but only
re-measured via ResizeObserver and afterNavigate. GradientBanner used to
compensate with its own window resize handler, which no longer runs when a
stack is in charge, so resizing left the header overlapping the banner.
Added the resize listener to AlertStack, which also covers HeaderAlert.
handleClose only hid the local component, leaving the store entry visible.
headerAlert.get() compares importance and breaks ties by registration
order, so the dismissed entry kept winning the importance-1 tie and
rendered an empty slot, hiding newDevUpgradePro and the backup policy
alert until a reload.
Now calls headerAlert.updateShow(id, false) so activeHeaderAlert
recomputes and the next eligible promo can take the slot. Verified: after
dismissal the AlertStack unmounts and the header offset returns to 0
rather than leaving an empty banner mounted.
Also routes NewConsoleBanner through the components barrel, per AGENTS.md.
Uses GradientBanner, the same component newDevUpgradePro uses, registered
through headerAlert at importance 1 so it sits in the promo tier and can
never outrank a payment or usage warning. Dismissal is keyed on the alert
id in localStorage, matching newDevUpgradePro.
Also makes GradientBanner stack-aware. It is position:fixed and offsets the
shell itself, but AlertStack is fixed too and measures its children to do
the same job. Inside a stack the banner measured as 0, so the stack reset
the header offset the banner had just set and the banner covered the
navigation until the next resize. It now stands down when a stack is in
charge, mirroring headerAlert.svelte. This also fixes the same first-paint
overlap for newDevUpgradePro and enterpriseTrial.
The CTA carries utm_medium=banner so this surface can be compared against
the promo card.
The console's source handler reads ref, utm_source, utm_medium and
utm_campaign and posts them to /console/sources. A bare `source` param is
not read anywhere, so it was dead weight.
utm_medium distinguishes the surface, so a banner or navbar entry added
later can be compared against the card. An explicit param is required
because the handler discards same-origin referrers, and cloud.appwrite.io
matches that check.
Follows the single-slot convention every announcement has used: the new
promo replaces its predecessor and the old asset is removed, matching
60e7ba53e which swapped the State of Appwrite survey for Init.
Adds a bottom modal alert inviting users to the new Console, following the
same shape as the Init 2026 announcement (60e7ba53e): a single
BottomModalAlertItem with a promo image, gated on isCloud.
The card is added alongside the Init promo rather than replacing it, since
Init runs August 31 - September 4. The bottom alert renders them as a
carousel. Drop the Init block to make this a straight replacement once that
campaign ends.
Regenerating bun.lock pulled in unrelated dependency upgrades, including
prettier 3.8.3 -> 3.9.6, which reformatted 11 untouched files and failed
the lint job.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The button-state cases reimplemented the card's `btnDisabled` expression
inside the test file, so they asserted against a copy of the logic. The
guard could be dropped from the component and every test naming the
reported bug would still pass.
Fold the derivation into `usersLimitState()`, render the card from it,
and drive the tests through it. Verified by mutation: dropping the error
guard, dropping the limited-mode guard, and ignoring mode in the
changed-check now fail 1, 1 and 2 cases; against the old tests all three
passed clean.
The user-limit policy reserves `total: 0` for "no limit", but the card
treated 0 as a value you could type. On a project with no limit,
selecting Limited and entering 0 left `policy.total === newLimit`, so
Update stayed disabled with nothing explaining why.
The same check had three more holes: entering 0 on an already-limited
project saved `total: 0` and silently turned the limit off while the UI
still read "Limited 0"; clearing the field bound `undefined` and made
the SDK throw `Missing required parameter: "total"`; and with no `min`
on the input a negative value reached the API as a 400.
Move the range and changed-vs-saved checks into usersLimit.ts, put
min/max/step on the input, and surface a rejected value as an inline
error instead of a dead button. 0 can no longer be submitted as a
limit, so the card cannot drift out of sync with the saved policy.
Also drop `maxUsersInputField` and its focus effect: nothing ever bound
it, so it was always null and never ran. Wiring it up would have
auto-focused and scrolled to the card on every page load.
* fix(storage): correct allowed file extensions input copy
The bucket settings Allowed file extensions field reused Labels copy
(id, label, and placeholder). Use extension-specific wording and clear
the placeholder once tags are present so it does not sit beside chips.
Co-authored-by: chiragaggarwal5k <chiragaggarwal5k@gmail.com>
* chore: retrigger CI after flaky onboarding e2e
Co-authored-by: chiragaggarwal5k <chiragaggarwal5k@gmail.com>
* chore: retrigger flaky onboarding e2e (2)
Co-authored-by: chiragaggarwal5k <chiragaggarwal5k@gmail.com>
* chore: retrigger flaky onboarding e2e (3)
Co-authored-by: chiragaggarwal5k <chiragaggarwal5k@gmail.com>
* fix(auth): navigate after cloud signup so onboarding e2e can proceed
Cloud register only called goto when a pricing type was present, so CI
stayed on /register until waitForURL timed out.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(e2e): drop verify-email skip after register
Address review feedback: keep waiting for onboarding create-project
and remove the temporary CI nudge comment on the extensions field.
Co-authored-by: chiragaggarwal5k <chiragaggarwal5k@gmail.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: chiragaggarwal5k <chiragaggarwal5k@gmail.com>
D5's answer to the customer in the support thread was explicit: Pro
doesn't include SOC-2, "you need the custom plan" — not Scale. The
previous gate OR'd in `planHasGroup(Scale)`, which was inferred from
planComparisonBox's marketing copy rather than any real entitlement,
and would have wrongly let a self-service Scale org through.
Checked plans.php in the cloud repo: every self-service plan (Pro,
Scale) sets `selfService: true`, and every negotiated Custom/Enterprise
contract plan sets it `false` — `supportedAddons` doesn't distinguish
them for anything SOC-2-shaped (it only carries the paid, self-service
BAA/premiumGeoDB addons). `selfService` is already a real field the
API returns per plan today, so gating on `selfService === false` alone
needs no backend change and matches support's stated policy exactly,
instead of guessing at a group boundary.
Also swapped the ineligible-state CTA from the self-service change-plan
wizard to the existing Enterprise contact-sales link, since there's no
self-service upgrade path onto a plan that actually unlocks this.
The SOC-2 card rendered for every cloud organization owner, so Free and
Pro users could fill in the whole request form and only learn it was
unavailable from a generic "please try again later" error. BAA, the card
directly above it, has always been gated on `supportedAddons.baa`.
No SOC-2 entitlement exists on the plan model, so infer it: the Scale
group is the lowest one that advertises SOC-2 (the Enterprise column on
the pricing page), and contract plans sit above it. Ineligible
organizations now get a short explanation and an upgrade CTA instead of
a form that cannot succeed, and the modal is no longer mounted for them.
Also send the request as multipart form data, the shape the support
wizard uses. SOC-2 was the last caller posting JSON to growth's
/support: the wizard moved to FormData in e446f7ff8 and the BAA request
modal, the other JSON caller, went away in bf2166bd3, so if that path
has rotted the request was failing on every plan including Enterprise.
The billing plan now rides along in metaFields so support can see the
tier on the ticket.
This repo has no existing Svelte component test pattern, so covering
the fix meant inventing one from scratch: jsdom localStorage/
ResizeObserver shims plus a harness component, none of which serves
anything else here. Not warranted for a bug fix. The fix itself
(selection reset on URL change + keyed row loops) stands on its own.
Selection state in MultiSelectionTable stuck around across pagination
and search, keyed only on row id. Selecting rows, then paginating to a
different page and hitting "select all" would carry the old ids into
the new batch delete, silently deleting rows the user never saw
selected. Reported from Storage and Databases in production.
Reset selection whenever the URL changes, and key the previously
unkeyed row each-blocks (files, users, memberships, platforms,
backups) so Svelte doesn't reuse row elements across renders.
* fix(git): portal branch selector to body to stop modal wiggle
Portaling the fixed-position dropdown into <dialog> made it resolve
against the dialog containing block, so opening Production branch
shifted modal contents (#2790). Append to document.body instead and
stopPropagation so outside-click dismiss still works.
Co-authored-by: chiragaggarwal5k <chiragaggarwal5k@gmail.com>
* fix(git): portal branch menu into dialog with dialog-relative coords
Body portals sit under showModal() top layer, so the dropdown was
unreachable. Keep portaling into the open dialog for stacking, but
convert getBoundingClientRect() to dialog-local coordinates — fixed
positioning inside the dialog is relative to that box, and viewport
coords were what shifted the modal (#2790).
Co-authored-by: chiragaggarwal5k <chiragaggarwal5k@gmail.com>
* fix(git): detect owning dialog via closest() for branch portal
Svelte slot projection means BranchSelector never sees pink Modal's
dialog-group context, so the dialog portal path never ran. Use
containerEl.closest('dialog') instead so create-deployment menus
still join the showModal() top layer with dialog-local coordinates.
Co-authored-by: chiragaggarwal5k <chiragaggarwal5k@gmail.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: chiragaggarwal5k <chiragaggarwal5k@gmail.com>
Back only cleared the selected user; it left the membership request in
flight and did not advance the request counter, so nothing invalidated
it. On rejection the guard still matched, and the error landed on the
user step — reporting "Could not load users." over a user list that had
loaded fine, with a Retry that re-ran the wrong request.
Back now invalidates the in-flight request and clears the error and
spinner it owns. The rejection path also checks the selected user, so it
matches the success path rather than relying on the counter alone.
A rejected request left the list undefined, which reads identically to a
successful empty response, so the picker stated that the user belongs to
no teams — and offered nothing to click, leaving Back and re-selecting
the user as the only way to try again.
The failure is now held separately from the list and rendered as its own
state with a Retry, in both steps. In the user step it is checked ahead
of the search branch, so a search that fails reports the failure rather
than claiming the term matched nothing.
Stepping back and choosing a second user, or typing another search term,
leaves the earlier request in flight. Whichever resolves last was writing
to the shared list, so one user's memberships could be listed under
another user's name — and a row selected there would have granted the
membership actually belonging to the first user.
Requests now carry a sequence number and only the newest may write, with
the membership step additionally confirming the selection has not moved
on beneath it.
A rejected request also skipped the assignment that cleared the spinner,
leaving the picker loading forever with nothing said. Both requests now
settle in a finally and surface the failure.
The member role has been in the API and the docs since 1.0, but neither
console ever offered it — the only way to set one was to know the syntax
and type it into the custom permission box.
It keys on a membership ID, which nothing in the console surfaces, so
the picker resolves it in two steps: find the user, then choose which of
their team memberships to grant. It has to run in that direction —
teams.listMemberships indexes only the membership ID and the user ID for
search, so picking a team first would leave the member step with a
search box that matches nothing. users.list indexes name, email and
phone, so the first step behaves like the existing user picker.
parsePermission only recognised user and team, so a member role fell
through to the invalid branch: it rendered as a raw truncated string in
the styling reserved for a deleted entity, and its hover card showed the
missing-entity avatar — presenting a perfectly valid grant as broken.
It also called onNotFound. Nothing passes that callback today, but
d36135f5b wired it to auto-remove stale roles before row.svelte moved to
runes turned the event into a prop and left the consumer behind. If that
wiring is restored, member roles would be silently deleted from any
resource a user merely opened.
A membership can only be read through its team or its user, and the role
string carries neither, so the ID stands in for a name. That is still an
improvement on the raw role: the badge now reads Member and the hover
card offers the membership ID to copy.