Reported from an iPhone, with a screenshot: tap the message box, the keyboard
comes up, and the whole app has slid off the top — header gone, only the bottom
edge of the composer left, black underneath it down to the keyboard.
That is the visual viewport being panned. WebKit shrinks it for the keyboard and
then moves it down inside the layout viewport (which it does not shrink) to
reveal the focused field. The shell had shrunk correctly, to the top --sb-vh of
the layout viewport; the window had simply moved off it. No CSS declines that
pan, and it is not a document scroll, so there is no scrollTop to put back.
So the shell is pinned to the visible rect: fixed to the layout viewport and
translated by visualViewport.offsetTop. The earlier objection to pinning is kept
where it was right — HEIGHT still moves on resize only, which is what stopped the
layout twitching under the finger, and the offset rides a transform rather than
`top`.
The first cut of this had a guard that silently disabled it: the pan was only
followed when window.innerHeight was much taller than the visual viewport, i.e.
"a keyboard is covering us". That assumes iOS never resizes the layout viewport,
which stopped being safe once WebKit started honouring interactive-widget. Only a
pinch-zoomed page switches the pan off now, by scale, and offsetTop is believed
whenever it is non-zero. The offset is also re-read across the keyboard's
animation on focusin/focusout: on iOS 26 the viewport is documented not to settle
back immediately on dismissal.
The message list stays at the newest message across a keyboard open or close,
instead of holding a scroll position from the taller layout and leaving the last
few messages under the composer.
Two smaller things on the same screen:
The QR code closes itself once the safety number appears. It was only dismissed
when both exchange flags cleared, and those stay set through verification, so a
one-time credential sat on top of the code the user was meant to be reading.
The disconnect button's label is hidden wherever the button becomes a 40px
square. Its own rule stopped at 560px while the square starts at 768px, so
between the two the word was still inside the box, pushing the icon off centre.
?vvdebug=1 paints the viewport numbers over the app — visible rect height,
offset, scale, what the shell resolved to. A phone cannot be put on a debugger
and none of those numbers show up in a screenshot of the result; this one is how
the guard above was caught. Off unless named in the URL.
Claude-Session: https://claude.ai/code/session_01R63J7dexujFQCJpDAeKxr9
Desktop 1.0.3 fixes video on a call placed after a group call — the two shared
one peer connection, the second call opened a second video section, and the two
H264 profiles collided, so the call carried sound and no picture. It also ships
the STUN and TURN list this client uses, which is what lets a browser and a
desktop agree on a path instead of one of them offering no relay at all.
The version lives in two places on this side (the landing page and the download
grid), and the consistency test also reads the README badge, the changelog and
the cryptography document, so all of them move together.
Riding along: the group status dot now takes its colour from the theme in the
call and group views, which is what the desktop's copy of groupsStore.js was
compared against.
Claude-Session: https://claude.ai/code/session_019RCW2Fm7dkAaU2zeHDksei
The palette lived as ~620 hex literals in inline styles plus a few hundred more in
the stylesheets, so there was no single thing to change. It is now 113 custom
properties in src/styles/theme.css, in two blocks.
src/scripts/theme-boot.js decides the theme before first paint — blocking, in <head>,
above the stylesheet, because a deferred script paints dark first and corrects itself.
It stores the mode ('system' | 'light' | 'dark'), never the colour it resolved to, and
stamps data-theme so an explicit choice can beat the media query. The switcher in the
header is a view onto it.
A filled accent stays the brand colour in both themes — the ink on it is near-black
either way — while an accent used as text darkens to clear 4.5:1 on white. A colour
reaches a fill by four routes (a style property, a constant, a helper argument, an SVG
source string), and tests/theme-switching.test.mjs covers all four.
The dark theme is unchanged: every colour declaration the previous build produced comes
out of this one identically once the properties are resolved.
Also: the roadmap drops its status chips on mobile, and Roadmap.jsx no longer splits a
colour with parseInt at runtime, which a var() reference cannot survive.
The bundles carried all thirteen translations at once and a page fetched them a
third time as raw source; each page now loads only its own language. Alongside
that: JavaScript is minified, the eight stylesheets are served as one file, the
QR scanner is fetched after the app is up instead of on every visit, Inter ships
once rather than five copies of the same file, and Font Awesome is subset to the
82 icons this app draws instead of all 2468.
1.85 MB across 43 requests becomes under 700 KB across 33. On mobile the page
starts drawing in 1.6 s instead of 6.3 s and is usable in 4.5 s instead of 11 s.
Pages also carry their text in the HTML now. Everything was drawn by JavaScript
into an empty div, so crawlers saw correct metadata around nothing, and twelve of
the thirteen language pages had never been shown to anyone. The documentation is
published under /docs/ with a new FAQ, and unknown addresses return a real 404.
Separately: the localized shells were served with the year-long immutable cache
header meant for static assets, which pinned anyone who opened /de/ or /ru/ to
that build. The header is fixed and the service worker refreshes what it cached.
Claude-Session: https://claude.ai/code/session_014KjzTXxrhzYoDDWChYQ4u2
The badge image is vendored rather than hotlinked from snapcraft.io. The CSP
is img-src 'self' data: and would block it anyway, but the reason to leave the
CSP alone is the page's own claim: fetching a badge from someone else's server
hands them the address of every visitor to a page that says no servers are
involved.
The README badge stays dynamic — it reports the published version, and GitHub
proxies images, so no reader is exposed by it.
1.0.1 is the first desktop build that works on Linux at all. It also carries
the desktop.callsNotHere string these dictionaries now define for it — calls
are not available on the Linux desktop yet and the app says so rather than
failing.
src/app.jsx carries its own SB_DESKTOP_VERSION for the compact download
table; src/components/ui/DownloadApps.jsx carries DESKTOP_VERSION for the
grid. Bumping one leaves the other pointing at a tag whose filenames no
longer exist — the failure tests/desktop-download-links.test.mjs was written
for, and which it caught here.
The i18n scan also learns that these dictionaries feed two clients. The
desktop client is a separate repository, so its keys can never appear in a
scan of this src/ and were being reported as dead strings to delete.
Group calls are a new capability, and 6.6.6 only landed as a patch because the
version was bumped once per deploy while the feature was being tested. The
number is the release's, not the deploy's.
No code changes: version, badge, changelog heading, docs and build stamp.
Claude-Session: https://claude.ai/code/session_01XSxAkET3hQTkYDQfbjCQwZ
A group call is N-1 ordinary 1:1 calls, one to each other member, each riding
the pairwise session that member already has — a transport a human already
authenticated by comparing the safety code. No mixer, no SFU, no point at which
two people's media meets anywhere but on a device.
Call control is separate from call media, because the two reach different sets
of people. Who opened a call, who joined and who left travels as group frames
signed with the sender's group identity key, so it reaches members currently
reachable only through a relay — and a relaying member can drop one but cannot
write one. Media flows only where a direct link exists, so a member without one
shows as connecting rather than being omitted. Frames carry a per-sender
sequence checked before the action, so a captured leave cannot end a later call,
and simultaneous calls converge on the lower random call id.
One capture is shared across every leg rather than one getUserMedia per member,
and legs answer without prompting: the flag permitting that is set only locally,
only while this user is in the call, and cleared when they leave.
UI: a gallery that sizes itself from the space it has, a spotlight view, an
active-speaker indicator read from the waveform, and the call surface in the
same visual language as the 1:1 one.
Also in this commit, the v6.5.0 language-suggestion work that had not been
pushed yet; its notes are in the changelog. And two fixes: the safety-code input
asks for digits rather than text, and starting a new chat from inside a group no
longer creates it behind the group where it cannot be seen — which had made it
impossible to connect to anyone new, or to add anyone to a group, while a group
was open.
Claude-Session: https://claude.ai/code/session_01XSxAkET3hQTkYDQfbjCQwZ
Four right-to-left languages at /ar/, /he/, /fa/ and /ur/ — thirteen in all.
The layout no longer has a left and a right, it has a start and an end: the
margins, insets and corners are CSS logical properties now, so they follow
dir on <html>. Directional glyphs flip; keys, safety codes and session
descriptors are pinned left-to-right so bidi cannot reorder what two people
compare against each other's screens.
Also fixed: the Service Worker was registered as './sw.js', which 404s from
every locale subdirectory, so twelve of the thirteen pages had no worker at
all. And the bundler ran before the dictionaries were generated, so a newly
added language could reach the page ahead of the app that renders it.
lockbitchat