False positives
- Same-sender repeats need at least 3 words and 10 characters, so short
replies ("ok", "test") repeated within 10 minutes stay visible.
- Repeated characters needs a run of 20+ of one non-digit character;
"Noooooooo!", separators, a row of emoji and long numbers are chat.
Correctness
- Rules apply to stored messages: adding a rule hides matching messages
already in the chat, and removing or disabling one shows again what only
it was hiding. Other filter reasons are left alone. Also applied when
history is loaded.
- The filter runs after mesh dedup, inside _addChannelMessage, so each
message is evaluated and counted by the duplicate tracker once.
- Channels excluded from filtering no longer feed implicit clean examples
to the classifier.
- Unread counting and the unread anchor treat filtered messages as normal
when filtering is off for the channel.
Performance
- Classifier training runs in a background isolate (injectable trainer for
widget tests); app start-up no longer waits for it.
- Reply previews look up the quoted message through an id index instead
of scanning the channel per bubble.
- The latest-messages channel sort computes each channel's latest visible
message once.
Cleanup
- Use MessageFilterText.isStructuredPayload instead of a screen copy.
Docs: message-filter.md explains how each check works, including the
learned classifier, and drops the outdated "new messages only" limit.
- Channel chat: when the message filter hides or collapses loaded messages
and the list is too short to scroll (so the load-older trigger can't
fire), load older pages automatically until it can scroll, storage runs
out, or 20 pages have loaded. Past that cap an empty channel offers a
"Load older messages" button instead of a dead end.
- Notifications: incoming channel messages notify after awaiting
translation; marking one as spam meanwhile replaced the stored copy but
not the captured one, so the spam still notified. Check the stored
message's filter state before notifying.
pubspec.yaml now pins environment.flutter to 3.47.9, so pub get refuses
other SDKs. Every CI job reads that pin via flutter-version-file,
including the web deploy (was 3.41.2). The AppImage SDK pin moves to
3.47.9 and CI fails if it drifts from pubspec. The Nix shell no longer
bundles Flutter, since nixpkgs only packages 3.47.4; it warns when the
SDK on PATH is missing or the wrong version.
Flutter 3.47 is reported to fix the Impeller/Vulkan crash in
Canvas::GetLocalCoverageLimit (flutter#190919) seen on Pixel 10 and 11
with the 3.44.1 engine. 3.47.8 also updates Dart's root certificates.
pub get on 3.47 adds platform-folder excludes to flserial's analysis
options.
Client-side filter for channel messages, with no model download:
- User rules on message text and/or sender name (contains, whole word,
exact), normalised for case, accents, zero-width and look-alike letters.
- Automatic checks: repeated characters, same-sender repeats within 10
minutes, and the same sentence (>= 6 words) from 3+ senders within 30
minutes. Windows use send times, so synced history is not a flood.
- On-device n-gram classifier trained by Mark as spam / Not spam taps.
Filtered messages are kept, never notify or count as unread, are not
auto-translated or used for channel previews, and are concealed in reply
quotes. They show as a calm, collapsible placeholder that animates open
and closed, or can be hidden completely. Includes a Filtered messages
list, per-channel toggle, settings screen, translations for all locales
and documentation.
The Fix Native Asset Minimum OS phase edits each framework's Info.plist after
it was signed, so development builds fail to install (ApplicationVerificationFailed
on objective_c.framework). App Store exports re-sign and hid the problem.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Translation runs GGUF models on llama.cpp. llamadart 0.8.20+ also bundles the
LiteRT-LM runtime, whose iOS GemmaModelConstraintProvider.framework requires
iOS 26.4, so App Store Connect rejects 16.4 builds (ITMS-90208).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fill the 70 new theme, chat background and holiday strings in all 17
locales and review the holiday greeting phrasings. Prefer the longest
matching greeting so combined greetings pick the right holiday.
- Themes: name + background + accent, derived surfaces/text/bubbles; editor
with live preview; share codes and .jsonl import/export
- Theme icons: swap tab, send, settings, node avatar and map marker icons for
named icons or pictures (lossless JPEG XL)
- Chat backgrounds per channel/DM (color or JPEG XL picture with fade);
themes can set one, optionally for every chat
- Map follows custom themes (panels, accent, tile tint, markers)
- Holiday themes (Halloween, Thanksgiving, Christmas, New Year's) with
generated backgrounds, icons and map markers; automatic switching from
17 days before each holiday
- Holiday greeting animations (bats, turkeys, Santa, fireworks) matched
across all app languages via translation strings
Repeater settings:
- The Advanced and Identity Keys cards store their expanded flag in
PageStorage. Their TextFields' inner Scrollables shared that identifier,
read the bool as a scroll offset and failed to build, which showed as a
large grey box in release builds. Give each card's body its own storage
scope. Regression test added.
Discovered contacts:
- Adverts parse the two optional 16-bit feature fields before the name,
as MeshCore's AdvertDataParser does, so their bytes are no longer read
as the name (a zero byte gave an empty name).
- A nameless advert or removal no longer wipes a known name.
- The discovery tile gives the name the full width (the type chip beside
it squeezed it to nothing at large text sizes); the type moves to the
key line as plain text, and an unnamed node shows its key prefix.
- The node scan sheet's type chip can shrink instead of the name.
Contrast:
- StatusChip draws its label in the normal text color; the icon or dot
keeps the accent, adjusted to 3:1 against its tint. Avatar icons and
initials get the same adjustment. Amber on its tint in the light theme
was about 2:1.
Channel chat:
- Hashtag channels drop the redundant "Hashtag" subtitle label, and the
region action moves into the overflow menu (tapping the title still
opens it) so the channel name keeps its room.