The broker-pool GET put a single fixed 6s deadline on the entire paginated dump. A large pool (6+ brokers) streams field-by-field over ~8s and blew the deadline, so the client falsely reported the pool as not finished even though the device sent it all (the END terminator arrives after the client gave up). Re-arm an inactivity Timer on each received frame: a steadily-streaming dump never times out; only a genuine stall (no frame for the timeout window) fails. Adds a TDD test - a slow-but-steady dump, frames under the timeout but total far over it, must complete.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- CHANGELOG.md seeded with the 1.1.2-beta.1 entry (Added / Fixed / Known issues).
- README gains a "Download & Install" section for end users (Releases link,
extract + run on Windows with the SmartScreen note, sideload the APK) — the repo
previously had only developer build-from-source steps.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The repeater "Path hash mode" dropdown showed raw mode numbers (0/1/2). Match
the companion device's selector (settings_screen.dart _PathHashSizeTile) so the
options are self-describing: "1-byte · max 64 hops" / "2-byte · max 32 hops" /
"3-byte · max 21 hops".
Underlying value (mode 0/1/2 = 1/2/3-byte) and the `set path.hash.mode N` wire
command are unchanged — labels only.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- pubspec 1.1.2+30 -> 1.1.2-beta.1+31: the app self-identifies as the public beta
heading toward 1.1.2 final.
- scripts/package-windows.ps1 (#96): builds the distributable zip and BUNDLES the
VC++ runtime DLLs (msvcp140 / vcruntime140*) that Flutter's Windows build omits,
so the zip launches on a clean machine without the redistributable installed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Platform-adaptive export on the BLE/frame log screen: the system share sheet on
mobile (the priority path), or open the logs folder in the file manager on
desktop. Web-safe (defaultTargetPlatform, no dart:io); flushes the file first so
it is current. Completes #97.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Always-on file logging so users/support can retrieve a log without enabling
anything in the UI. No-op on web (no filesystem).
- FileLogWriter: rotating file writer (size cap + max files), single + batch
writes, unit-tested against a real temp dir.
- FileLogService: singleton initialised in main(); resolves
getApplicationSupportDirectory()/logs, queues lines, periodic 2s batched
flush, bounded queue, never crashes the app on an IO error.
- AppDebugLogService.log + BleDebugLogService.logFrame write every entry to the
file (app logs always, independent of the in-app viewer's enable gate).
Location per the #97 decision: standard support dir on all platforms (Windows
%APPDATA%\app.offband.meshcore\logs). Still TODO before recompile: access UI
("Open logs folder" on desktop / "Share logs" on mobile).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Was 1.1.1+18. Tags already referenced 1.1.2 (v1.1.2-ble-test/-queuediag) and the
rc8/rc9 APKs shipped +28/+29 via --build-number, so the pubspec build number had
drifted. +30 is the next build after the +29 that shipped. Tag v1.1.2 at release.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Consume the additive OCFG_BROKERS dump fields from firmware #172/#173
(backward-compatible — absent fields read as today's enabled/disabled):
- BrokerRuntimeState (`state`) + BrokerLastError (`last_error`) + a pure
BrokerConfig.status mapping -> Connected / Connecting / Failed (reason) /
Held (no clock|low heap) / Enabled / Disabled; held shown neutral, not error.
- Brokers list shows the real status line + colour instead of just "enabled".
- Editor shows jwt_owner / iata_override resolved defaults (`jwt_owner_resolved`,
`iata_resolved`) as greyed placeholders when the raw key is blank; only writes
the raw key when the user enters a value (blank stays the source of truth).
Sourced from the pool dump (getBrokers); getBroker's per-field GET list is
unchanged pending the firmware GET-ability answer on #173. Unit-tested mapping,
parsing, and backward-compat (9 tests).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A firmware build can ACK a broker enable/disable as success without applying it
(HV4; meshcore-firmware#179). The quick toggle and the editor save previously
trusted the ACK and showed a success snackbar, so an ack-but-no-op looked like
it worked.
Both paths now SET, wait a short settle, re-read the slot, and report what the
device ACTUALLY did:
- applied -> confirm
- notApplied -> warn "possible firmware issue"; the editor stays and re-seeds to
the device's true state instead of popping a false success
- unverified -> re-read failed (e.g. an enable that rebooted the device)
Every device is treated identically: the UI reflects the re-read, never assuming
a change took that the device didn't confirm.
- BrokerConfig.classifyApply pure classifier + BrokerApplyOutcome (unit-tested)
- ObserverConfigService.applySettleDelay
- wired into mqtt_brokers_screen quick toggle + broker_editor_screen save
- widget test for the ACK-but-not-applied path
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The BLE/USB self-info retry was an unbounded 3.5s hammer -- the Web path caps at
3 and skips while syncing; BLE had neither. On a slow device handshake it churned
the connect and starved the contact sync, hanging the contact-load bar. Inherited
from upstream zjs81/meshcore-open (git blame; zero Offband commits).
- Replace the unbounded 3.5s Timer.periodic with a self-rescheduling backoff:
3.5s -> 7 -> 14 -> 28 -> 56s, then a steady ~60s keep-alive. Never hard-stops --
a slow/recovering device still gets prompted.
- Add the Web path's skip-while-syncing guard on BLE (per-tick skip, never
permanent: the timer keeps rescheduling, so a settled sync lets the next tick
send).
- Decisions extracted as pure, tested helpers: nextAppStartRetryDelay(attempt)
and shouldSendAppStartRetry({connected, awaitingSelfInfo, syncing}).
- Disconnect teardown (Gemini's amendment) was already present (:2501 / :2374).
Gemini plan-review: Proceed, no blockers. 6 new helper tests; full suite 299 green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
On-hardware feedback: JWT owner defaults to the device pubkey and IATA override
is optional by definition, so gating a save on them is wrong -- they have
under-the-hood firmware defaults.
BrokerConfig.enableError now checks only the structural fields with no default:
URL present, port in range. The JWT field checks are gone; the firmware enforces
JWT completeness with its defaults, and a genuinely-bad enable surfaces the
firmware's reason via the failure messaging added earlier. The editor's pre-save
check and the list's quick Enable both follow.
Tests updated to the new contract (a JWT broker with blank owner/audience now
enables; a structurally-complete quick Enable confirms). Full suite green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
On-hardware test feedback:
- Long-press Enable/Disable was silent on success AND failure. It now confirms
on success and, on failure, surfaces the device's reason (lastError) instead
of nothing.
- The client now refuses to quick-Enable a broker that can't work (incomplete
required fields, e.g. JWT without audience/owner) and says why -- so it never
reports "enabled" on a broker that won't run. (Firmware should reject too.)
- The editor blocked a DISABLE on blank/invalid fields. Validation now gates
ENABLING only -- disabling a slot with partial values is fine; it won't be
active.
Validation rules shared via BrokerConfig.enableError (editor pre-save + list
quick-Enable). Full suite green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The broker pool gets its own screen, reached from a "MQTT brokers" row in the
Observer pane (replacing the read-only inline list and its stale "pending
firmware support" note).
- MqttBrokersScreen: a tile per broker (tap -> editor, long-press -> Enable /
Disable / Edit / Clear), and a + FAB that opens the next empty slot (disabled
when all 10 are full) -- matching channels_screen's FAB-add pattern. Each
action re-reads only the affected slot, never the full pool dump.
- Observer pane: the broker section is now a navigational row showing
"N configured - M enabled"; the editor is wired in end-to-end.
3/3 screen tests green; full suite green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The edit UI for one broker slot. Staged-save: controls edit local state; Save
sends only the CHANGED fields through saveBroker (enabled written last), so a
partial save can't leave a live-corrupt broker.
- Fields: url, port, transport, auth_type, username, write-only password,
topic_prefix, iata_override, jwt_* (when auth=jwt), ca_cert (when tls/wss),
plus the enabled toggle.
- Validation before any SET: port 1-65535, jwt audience+owner when auth=jwt.
- Write-only password: sent only when typed; blank keeps the stored secret.
- Refresh: single-slot re-GET (getBroker) + reseed, with a dirty discard guard.
- Back button guards unsaved edits (PopScope -> Discard? dialog) -- Gemini's
state-loss finding.
4/4 widget tests green. Brokers list screen + Observer-pane wiring next.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The write half of the broker editor, against RedCreek's confirmed field-at-a-time
contract. The existing 0xC0 key/value command already carries the broker keys --
no new wire codes -- so the codec comment at observer_config_client.dart:166 is
stale, not a gap.
- saveBroker(): disable a live slot first, write each changed field, write
`enabled` LAST. Stops at the first ERR/timeout and reports the failed field --
a partial save leaves the slot disabled, never live-corrupt (activation guard).
TDD: ordering + partial-failure tests.
- clearBroker(): mqtt.broker.<slot>.clear.
- getBroker(): single-slot re-GET via 14 per-field GETs (recovery / post-save
read), not the 84-frame pool dump. Password is write-only -> best-effort.
Key spellings (mqtt.broker.N.enabled / .clear, scalar field GET) are inferred
from the existing scheme; they're the one thing the joint firmware test confirms.
12/12 service tests green. UI (brokers screen + editor) next.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The firmware answers RESP_CODE_ERR for an empty/invalid channel index, but
_handleErrorFrame never advanced the in-flight channel sync. So every empty slot
waited out the 2s timeout + 3 retries (~8s each): a device with a few populated
channels out of a large capacity took minutes to sync (~40 slots requested, ~29
empty -> ~4 min) before the real channels (which answer CHANNEL_INFO) finally
synced in the last couple seconds.
Now an ERR that lands while a channel GET is in flight (and no generic-ack
command is waiting -- that one is order-correlated to the ERR first) advances to
the next slot immediately, mirroring the CHANNEL_INFO success path minus adding
a channel. The decision is pinned as a pure static predicate and tested,
matching the connector's existing decision-helper test convention.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three observer-delta fixes (per the #80 sync diagnosis / G2):
- A: settings screen gates the Observer category via context.select on the
supported flag, not a connector-wide context.watch -- the watch rebuilt the
whole settings screen on every sync frame, thrashing the UI thread.
- B: a flat Save re-reads only the flat settings (refresh includeBrokers:false),
never the 84-frame broker pool -- no BLE re-flood for a display toggle.
- C: the initial observer read is deferred until the device's channel/contact
sync settles, so it never competes with that sync for BLE.
TDD: B (no OCFG_BROKERS on a flat refresh) + C (refresh deferred while syncing,
fires once idle). Full observer suite green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The G2 BLE log showed the config protocol works (GETs return values; the broker
dump starts) but the firmware's broker dump stalls without BROKERS_END, so
getBrokers() times out. The client was treating that as a TOTAL failure --
discarding the 7 flat settings it read cleanly and blanking the pane.
Now the broker pool loads independently: refresh() builds + shows the flat
config even when getBrokers fails, flags brokersUnavailable, and the pane shows
the broker section as "unavailable" instead of going stale. Flat settings stay
usable. TDD red->green; full observer suite (39) green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Closes the 2 MINOR findings from the G1 re-review (#76):
- MINOR-A: ObserverConfigService.refresh() no longer wipes lastError / clears
stale on a PARTIAL read. A null from any getFlat is a failed read; the
snapshot is marked stale and the error kept (error-visibility).
- MINOR-B: the settings pane validates status_interval against the firmware
range (10-3600) before sending; an out-of-range value is not put on the wire
and the user is told.
TDD: both tests RED first (current sent '5'; refresh wiped the error), GREEN
after. Full observer suite (37 tests) green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Hardened the staged-save pane (settings_screen capability gate already correct):
- Normalize an unexpected device rotation to a valid SegmentedButton segment
{0,180} so the control always has a valid selection and never echoes an
out-of-range value back on save.
- Disable Save during a refresh (no save/refresh race on the controllers).
- digitsOnly formatter on status-interval (firmware still range-checks).
- Keys on the SSID/password fields for robust widget-test targeting.
4 widget tests pin the security-sensitive behavior: saves send ONLY changed
keys, a blank password keeps the stored secret, an entered password is sent as
wifi.pwd then the field clears, rotation is normalized, stale/error surfaced
(SAFELANE §6). Full observer suite (33 tests) green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
main.dart wiring verified correct (no change): the singleton MeshCoreConnector
(main.dart:38 -> field -> provided by value) is the SAME instance passed to
ObserverConfigService, so the service watches the live transport. Plain
ChangeNotifierProvider (not a proxy) is right -- the connector instance is
stable; only its state changes, which the service reads live. Lazy-safe: the
constructor has no side effects.
This test pins that contract: the service is reachable through Provider and a
mutation to the shared connector's caps flips the Observer gate (supported).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Gemini MAJOR #4 (offset unverified / short-frame robustness):
- Offset VERIFIED against firmware MyMesh.cpp: the device-info tail is three
consecutive out_frame[i++] writes -- client_repeat@80, path_hash_mode@81,
offband_caps@82. Bytes 80/81 are shipping (path-hash), so caps@82 is correct
by adjacency, not a guess.
- Extracted MeshCoreConnector.parseOffbandCaps (pure, bounds-checked) so the
guard + offset are unit-testable; _handleDeviceInfo delegates to it.
- 4 tests pin: pre-v14 short frame -> null, len-83 reads byte 82 verbatim,
observer bit round-trips supportsConfig, hostile truncated frames never OOB.
Live observer node confirms end-to-end at G2.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The models are the existing draft (1983ed9); Gemini flagged nothing in them.
This pins the defensive behavior they give untrusted wire data: enum fromWire
fallback, numeric tryParse fallback, secret-presence flag, SecretField
three-state, copyWith. 7 tests, all green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Greens A2's adversarial tests, fixing the Gemini findings:
- BLOCKER #1 (race): a chain-lock (_serialized) serializes every transport
round-trip, so overlapping refresh()/setFlat() can no longer cross responses
(firmware has no request id; responses correlate by order).
- MINOR #6 (secret leak): error messages redact secret key names
(wifi.pwd / *.password / *.pwd -> "<secret>").
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
4 tests for ObserverConfigService via a fake connector (subclass override of
the 4 members the service uses -- no production change). 2 green (capability
gate, happy-path setFlat). 2 RED BY DESIGN, documenting the Gemini findings to
be fixed by S1 (#73):
- BLOCKER #1: overlapping getFlat/setFlat cross responses (single-flight is
assumed, not enforced) -- setFlat receives the GET's value, not its ACK.
- MINOR #6: a secret key name leaks into the error string
("SET wifi.pwd failed: ...").
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
11 tests for ObserverConfigClient / BrokerListDecoder. 9 green (builders,
response parsing, happy-path decode, malformed-NUL, garbage-numeric, stray-KV).
2 RED BY DESIGN, documenting the Gemini findings to be fixed by C1 (#72):
- DoS: BrokerListDecoder retains 256 slots from a hostile BROKER_KV flood
(must bound to brokerSlotCount=10).
- out-of-range slot 200 is stored (must be dropped).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- screens/settings/observer_settings_view.dart: staged-save pane for WiFi
(ssid/pwd/enabled), MQTT region (iata/status_interval), display
(always_on/rotation 0|180), with Save (changed keys only) + Refresh +
stale/error banners; the broker pool renders read-only (enable/clear await
the firmware F-task).
- settings_screen.dart: register the Observer category, gated on
ObserverConfigService.supported (re-evaluated as the connector's offband_caps
flips on connect/disconnect).
- connector/meshcore_connector.dart: parse the offband_caps byte (device-info
byte 82, after path_hash_mode) + expose offbandCaps.
- services/observer_config_service.dart: supported computed live from the
connector (version gate + cap bit).
- main.dart: provide ObserverConfigService.
Strings are English-only pending l10n ARB keys.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reflow the systemDevices.map(...) chain after the .where() removal so
`dart format --set-exit-if-changed .` is clean. No behavior change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The scanner filtered advertisements by device-name keyword
(withKeywords: deviceNamePrefixes), so any MeshCore/Offband device
whose advertised name fell outside the hardcoded prefix list was
invisible to the scanner — even though the official MeshCore app,
which scans by the Nordic UART service UUID, found them. Switch
startScan to withServices: [Guid(service)] and drop the now-redundant
name filter on the already-service-scoped systemDevices fallback.
Device-tested on a Samsung S25: a previously-invisible node now
appears in the scan.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
CI's `dart format --set-exit-if-changed .` step is repo-wide. My A/B/C
commit hand-wrote the .catchError blocks without running the formatter,
and 2 settings views carried pre-existing format drift from #28. dart
format reflows all three (whitespace only, identical logic):
- lib/connector/meshcore_connector.dart (this PR's blocks)
- lib/screens/settings/app_settings_view.dart (pre-existing, #28)
- lib/screens/settings/message_settings_view.dart (pre-existing, #28)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The OffbandMesh org board (project #2) had no label->board sync, so issues never populated it (root cause of the 0-item board). Add sync-labels-to-board.yml (adapted from the DifferentWire/Unfocused canonical) wired to board #2's project + status/priority field and option IDs; runs on ubuntu-latest, uses repo secret PROJECT_PAT (classic PAT, project+repo scope). REQUIRES PROJECT_PAT set as a repo secret to function.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
settings_screen action handlers fire-and-forgot or awaited-without-catch several connector calls, and _loadVersionInfo called setState after an await with no mounted guard. Add a mounted guard to _loadVersionInfo; try/catch + error snackbar to the GPS-enable toggles (inline + _editLocation dialog) and getContacts; await syncTime in _syncTime so the 'synchronized' snackbar shows only on success (it previously fired before the await); and in _editLocation's save, capture the messenger before the pop and wrap the setCustomVar/setNodeLocation/refreshDeviceInfo awaits.
Verified non-issues: deleteAllPaths is void, _PrivacySection._apply() already has a try/catch, and exportGPX catches internally and returns gpxExportFailed. Completes #28; error strings English for now (separate l10n sweep).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Settings onChanged/onTap callbacks fire-and-forgot AppSettingsService Future<void> setters, so a persist failure rejected unhandled and was swallowed. Add a shared persistSetting(context, action) helper that awaits the setter in a try/catch and shows a dismissible error SnackBar, capturing the messenger before the await so it survives a picker-dialog pop. Wrap the 27 onChanged/onTap setter sites in app_settings_view + message_settings_view; setAppDebugLogEnabled and setNotificationsEnabled (which await + show their own confirmation) get an inline try/catch instead.
Covers the two onChanged panes; the settings_screen action flows (GPX export, location dialogs, version-info mounted guard) follow in a second pass, so #28 stays open. Error strings are English for now (separate settings l10n sweep).
Adds a persistSetting widget test.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
DirectRepeater.ranking added recencyScore (0..1.8M) to the SNR term (1000/dB), so recency dominated (~1 dB == 1 second), contradicting the comment that recency only breaks ties. Scale recency to a 0-999 bonus so SNR (1000/dB) is primary and recency only decides sub-dB ties.
Offset SNR by its floor (+32; SX126x SNR spans -32..+31.75 dB) so every live ranking stays >= 0, above the -1 stale sentinel; otherwise the scaled-down recency leaves live rankings negative and a stale repeater would outrank a live one.
Adds DirectRepeater.ranking tests: SNR-first, recency tie-break, stale=-1, live-always-outranks-stale.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
disconnect() never resets the channel store scope (publicKeyHex), so a channel write in the post-disconnect async gap still persists -- to the last radio's scope, and into a different radio's scope after reconnecting (cross-radio bleed). Add if (!isConnected) return; to the UI-triggered persisters setChannelUnreadCount, markChannelRead, setChannelOrder. Chosen over clearing _channels, which would risk a racing mutator persisting an empty list and wiping stored channels.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
PathHistoryService capped each contact's recentPaths at 100 but dropped any new distinct path once full, so the list froze on the first 100 ever seen and never refreshed with newer/better routes. Evict the lowest-_scorePathRecord entry to make room, keeping the strongest 100. Adds a regression test (full history evicts the weakest to admit a new path).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Query FlutterBluePlus.systemDevices on every native platform (was Linux-only)
so a radio already connected to another central -- e.g. Liam's MeshCore app --
appears in the picker and we attach to the existing link. A connected BLE radio
stops advertising, so a scan alone never finds it.
Patch increment for the #37 compact-tile feature (the bump that was owed
inside PR #47). Build +18 stays ahead of rc.3 (+17) so a 1.1.1 build
installs over the sideloaded RC.
Parse a leading @[Name] in the message body and render it as a rounded
pill that sets the name apart from the rest, matching the reference
client. Handles a missing space after the mention. Lives in
TranslatedMessageContent so channel + direct chat both get it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Clock setting (System / 12h / 24h) in App Settings; chat times follow it + locale.
- Channel tile: always-on metadata row - time + hops on inbound, time + cell_tower
heard count + status on outbound - decoupled from the message-tracing toggle. The
via-path stays behind tracing.
- Drop the auto reply-to block and the most-recent-sender guess; @mentions now
render inline in the message text.
- Direct chat: same clock-aware time formatting.
Closes#37
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Match the typed text anywhere in the shortcode name (not just the prefix),
ranked exact > prefix > substring, so `:check` surfaces the checkmarks (✅ is
`white_check_mark`). Add check/checkmark/tick aliases for ✅ in a separate,
regen-safe map the chat + channel composers consume.
Closes#45
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two more connector silent-failures from the review. requestRadioStats had an empty 'catch (_) {}' (now logs). setChannelUnreadCount fired unawaited(saveChannels) with no error path (now .catchError + logs). Verified: setContactUnreadCount calls a void debounced store method -- no call-site Future (Gemini false positive); the actual swallow is inside unread_store's debounce timer, flagged separately.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
BLE onValueReceived.listen had no onError, while USB (L1557) and TCP (L1669) both attach one that logs and gracefully disconnects; an error emission on the BLE notify stream was unhandled. Add a matching onError. Also await sendMessageWithRetry in the reaction branch of sendMessage -- the normal path already awaits it; the reaction path's unawaited call swallowed failures, contradicting its own 'same as normal messages' comment.
Verified, not assumed: Gemini's 'disconnect() not try-caught' was a FALSE POSITIVE (already wrapped at L2484-2489); 'deleteAllPaths silent void' is a sync in-memory clear (no async failure path). flutter analyze clean; Gemini-reviewed with no critique of these two changes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Strycher