Capacitor/WebView apps get suspended by iOS a few seconds after backgrounding, which
freezes the WebRTC mic + audio pipeline — so when the user minimised the app or locked
the phone during a call, no one could hear anyone. Add UIBackgroundModes=[audio] to
Info.plist (via ios-patch.sh) so iOS keeps the audio session (and the app) alive while
a call is actively playing/recording. Video rendering still pauses in the background
(unavoidable in a WebView) but voice continues. Needs a fresh Codemagic build.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Calls only notified over the chat WebSocket (CHAT.pushToUser), so a CLOSED app
(no live socket) never rang — unlike messages, which also call PUSH.sendToUser.
Add PUSH.sendToUser for both DM (startDmCall -> callee) and group (startGroupCall
-> other members) so APNs/FCM/WebPush alerts a closed device.
To make the alert actionable, add CALLS.replayActiveCalls(userId, ws), invoked
from the chat-hello handler: when a socket (re)connects, re-send any dm-call /
group-call the user is currently being rung into (the original events fire once at
call start and are missed by an app that was closed). Opening the app from the push
then re-surfaces the invite so they can answer within the ring window.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ROOT CAUSE of iOS push not working: Capacitor 7's default AppDelegate.swift
template does NOT implement application(_:didRegisterForRemoteNotificationsWithDeviceToken:)
or ...didFailToRegisterForRemoteNotificationsWithError:. So when @capacitor/push-
notifications calls registerForRemoteNotifications(), iOS fetches the APNs token
and calls the AppDelegate, but nothing posts .capacitorDidRegisterForRemoteNotifications,
so the plugin never delivers the token to JS. register() 'succeeds' yet neither the
registration nor registrationError event fires — proven by server push telemetry
(register-called logged; no token, no error; device_tokens stayed empty).
inject-push.js adds the two forwarding methods to the CI-generated AppDelegate
(idempotent, tolerant — never fails the build), wired into ios-patch.sh after the
audio patch. Verified against the real Capacitor 7 template: methods land inside
the class, braces balance, both listeners present.
This is the missing piece alongside the earlier aps-environment entitlement fix and
the server APNs config. Needs a fresh Codemagic build to take effect.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
device_tokens stays empty after reinstall+Allow, so the APNs token is never
obtained or never reaches the server, and there's no device console on Windows.
Add a /api/push-debug collector and breadcrumbs through setupNativePush (plugin
presence, permission state, register call, registration event/error, token POST
result) so the failing step is visible in server logs. Temporary — remove once
push is confirmed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The @capacitor/push-notifications plugin does not add the Push Notifications
capability to the CI-generated Xcode project (that's a manual Xcode step), and
add-share-extension.rb only merged the App Group into App.entitlements, assuming
aps-environment was already there. It never was — so on device PushNotifications.
register() failed with 'no valid aps-environment entitlement', no APNs token was
obtained, and device_tokens stayed empty (server had nothing to push to).
ios-patch.sh now creates App/App.entitlements with aps-environment=production
before the share-extension script merges the App Group in. Still requires the App
ID to have Push Notifications enabled (so the profile carries the entitlement) and
the server APNS_* key set (Step 5) for end-to-end delivery.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause of a post-deploy outage: recreating the app container can assign it a NEW
docker network IP. Nginx Proxy Manager caches the app's upstream IP at config-load,
so it kept connecting to the OLD IP (which, after adding the postgres/redis services,
was reassigned to bizgaze-postgres) → 'Connection refused' → site down until nginx
re-resolved. deploy.sh now runs 'nginx -s reload' in the NPM container after verify
(normal + rollback), non-fatal if NPM isn't detected. This makes deploys self-healing
for the app-IP-change case.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two things:
1. FIX a live regression the async conversion missed: chat.js calls repos via the
lazy repos() helper (not the R. prefix), so my sweep skipped it — effectiveStatus
/ broadcastPresence read `repos().users.byId(userId)` synchronously, but that's a
Promise now, so presence broadcasts always reported status 'active' and dropped
last_seen. Now awaited (effectiveStatus/broadcastPresence async); touchSeen is a
fire-and-forget UPDATE with .catch. Audited all non-R. repo calls — only chat.js
was affected (media.js backfill was already awaited).
2. Swappable pub/sub for multi-instance real-time fan-out (the actual blocker to
running >1 instance — not the DB). server/pubsub.js picks a backend by
PUBSUB_BACKEND (default 'memory'). Local socket delivery is UNCHANGED; publish is
additive — memory = no-op (zero hot-path cost, identical single-instance
behaviour), redis = fan-out to other instances with a self-echo guard. chat.js
pushToUser/broadcastPresence now also publish; each instance subscribes to deliver
remote events to its local sockets. Interface is tiny so Redis is one swappable
file (Postgres LISTEN/NOTIFY or NATS could drop in the same way — never hardwired,
as requested). Dormant redis service added to compose behind the 'scale' profile;
redis dep added; PUBSUB_BACKEND/REDIS_URL documented.
Validated: smoke 22/22 (memory), e2e chat delivery green. NOTE: full multi-instance
also needs distributed presence (isOnline is per-process) + meeting-signaling
sharing — chat/presence fan out via this layer; those are follow-ups.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds a dedicated Postgres 16 service (container bizgaze-postgres, own named volume
bizgaze_pg_data, healthcheck) on the shared NPM network. The app depends_on it
healthy. Inert until DB_BACKEND=pg is set in .env — default stays SQLite, so this
deploy changes nothing functionally; it just makes the engine available. Documented
POSTGRES_PASSWORD / DATABASE_URL / DB_BACKEND in .env.example.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The DM/group thread queries used `(? IS NULL OR created_at < ?)` — an all-NULL
param Postgres can't type ('could not determine data type of parameter') — and an
unaliased FROM-subquery (Postgres requires an alias). Both rewritten to add the
`created_at < ?` clause only when a cursor is given, and alias the subquery `t`.
Portable; sqlite db-smoke still 22/22.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- db/pg.js: the pg backend (prepare/exec/tx/init) — ?→$N translation, BIGINT parsed
as Number (matches sqlite; else expires_at<Date.now() compares string<number),
transactions on one pooled client, init() applies schema.pg.sql. Same interface as
db/sqlite.js, so repos are unchanged.
- repos.js: the ~7 SQLite-only queries rewritten to run on BOTH engines —
audit.add @named→positional; email lookups COLLATE NOCASE→LOWER()=LOWER();
INSERT OR IGNORE→ON CONFLICT DO NOTHING (addMember/poll vote/favorite);
mergeInto's UPDATE OR IGNORE→UPDATE…WHERE NOT EXISTS/NOT IN and INSERT OR
REPLACE→ON CONFLICT DO UPDATE. Re-validated on sqlite: db-smoke still 22/22.
- server.js: boot now `await db.init()` before listening (pg creates tables; sqlite
no-op), so the first request can't hit a missing table.
- db/migrate-sqlite-to-pg.js: one-shot row copy in FK order (bulk insert, TRUNCATE
first so re-runnable). audit_log id left to PG's identity.
- package.json: add pg ^8.13.1.
Next: validate DB_BACKEND=pg smoke against a real Postgres on the server, then merge.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The full sync→async conversion is complete and green on the SQLite backend. Every
DB call across the app now awaits the async adapter, so the identical code runs on
Postgres at cutover.
Converted (this commit finishes Phase 3):
- session.js: currentUser/apiKeyFromReq async → 63 route awaits + WS + static.
- routes.js: all ~250 R.* awaited; DTO helpers (namesFor, avatarsFor, buildMsgDTO,
buildPollDTO, reactionsForMessage, postSystemMessage, pushGroupUpdate,
issueRefreshToken, provisionFromBizgaze) made async; every `.map(x=>buildDTO(x))`
restructured to `await Promise.all(...map(async...))` preserving order; `.filter`
predicates that hit the DB moved to an `asyncFilter` helper; chained
`R.x.y(...).length/.map/.filter` wrapped as `(await R.x.y(...)).method`; stream
upload handlers (recording/transcript/attachment) made async.
- calls.js / signaling.js: all call/meeting fns async; leaveMeeting AWAITS
persistCallHistory + finalizeTranscript BEFORE endCallByRoom (ordering matters —
fire-and-forget would race the map teardown); WS handle()/cleanup() async with
.catch guards.
- static.js: authAttachment(Raw) async (the .some carrier check became a loop),
handleGet async; server.js dispatch catches handler rejections → 500 not a hang.
- media.js backfill, push.js, reminders.js, webhooks.js await their repo calls.
Validation on DB_BACKEND=sqlite: db-smoke 22/22; legacy e2e 80 checks pass with zero
FAILs (throws only at a PRE-EXISTING WS lobby-drift assertion, unrelated). Every
server file `node --check` clean.
Still on the branch — master untouched. Next: Phase 5 (pg backend + ~7 dialect
queries + data migration + Docker Postgres + cutover), then merge.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
On the db-migration branch only; master stays clean + deployable. Foundation
(adapter, pg schema, smoke harness) is already on master and safe.
Done:
- repos.js fully async (Phase 2, validated: node --check clean, no missed transforms).
- session.js currentUser/apiKeyFromReq async.
- Mechanical `await` prefix applied across routes/static/calls/signaling/reminders/
webhooks/push.
Remaining (does NOT compile yet — deterministic to finish):
1. Async cascade: helper fns that now contain `await` must be marked async and their
callers awaited. node --check points to each (namesFor, authAttachmentRaw/
authAttachment in static, the WS handlers in calls/signaling, reminders/webhooks
loops).
2. DTO builders are the real work: namesFor, avatarsFor, buildPollDTO, buildMsgDTO,
recDTO all became async — every `.map(x => buildMsgDTO(...))` etc. must become
`await Promise.all(arr.map(async x => ...))`.
3. Chained calls `R.x.y(...).map/.length/.includes` → `(await R.x.y(...)).method`.
4. Then: node --check all green → node test/db-smoke.js green → e2e → merge to master.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
server/dbx.js selects a backend by DB_BACKEND (default sqlite; pg added at cutover).
server/db/sqlite.js wraps the synchronous node:sqlite instance in the async
interface repos will call — prepare(sql).{get,all,run}, exec(sql), tx(fn), init().
Results come back as resolved Promises so identical repo code runs on synchronous
SQLite (dev/test) and asynchronous Postgres (prod).
tx() gives multi-statement atomicity that stays correct on both engines (sqlite is
single-connection; the pg backend will run it on one pooled client) — needed for the
account-merge transaction in repos.
Verified: get/all/run/tx all work end-to-end; confirmed no code reads
.changes/.lastInsertRowid, so the repo conversion is purely sync->Promise. Unwired —
nothing requires dbx.js yet; prod path untouched.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Covers the DB-backed HTTP paths the async repo conversion touches — auth, users,
messages, attachments, conversations/groups, reactions, mentions, edit/delete,
polls, scheduled meetings (paginated), favorites, audit — asserting current API
shapes. Runs to completion with a pass/fail count and honours DB_BACKEND so it
doubles as the sqlite-vs-pg parity check at cutover. No WS/signaling (in-memory,
not the DB).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
server/db/schema.pg.sql — the full Postgres DDL, every column defined up front (no
ALTER-ordering fragility). SQLite→PG type mapping documented in-file (epoch-ms
INTEGER→BIGINT, 0/1 flags→SMALLINT kept numeric so app code is unchanged, sizes→
BIGINT, audit rowid→GENERATED IDENTITY). Mirrors the three existing FKs and adds a
new idx_messages_attachment (the /files auth scan we cached earlier becomes a keyed
lookup).
Validated against a throwaway Postgres 16: loads with no errors, 24 tables + 44
indexes created. Unwired — nothing uses it yet; the SQLite path is untouched.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Migration prep. Two real fixes surfaced while building a regression harness:
1. db.js: the `guest_emails` and `lobby` ALTER TABLEs sat at lines 241/244, BEFORE
scheduled_meetings is CREATEd (line 299). On a FRESH database the ALTER fails
(no table yet), is swallowed by the try/catch, and the columns are never added —
so a brand-new deploy is missing them and scheduling with guests crashes. Prod
escaped it only by incremental deploy history. Moved both ALTERs to after the
CREATE. (The upcoming Postgres schema defines every column up front, so this
whole class of ordering bug goes away there.)
2. test/e2e.js: /api/meetings returns paginated `{list, pastTotal, page, pageSize}`
now, not a bare array — updated three `.data.find` → `.data.list.find`.
No prod behaviour change (prod already has the columns; ALTERs are idempotent).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The top-bar buttons looked heavy against the navy bar (thick grey X-circle, bold
"Send" pill). Swapped for light SF Symbols on the navy bar: a thin `xmark` for
cancel and a `paperplane.fill` for send (semibold, enables when ≥1 chat is picked).
Icon-only send matches the Teams reference — the radio checks already show what's
selected, so the "(N)" count text is dropped. Native-only — needs a build.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two things from testing (the in-sheet picker with avatars + multi-select works):
1. Double "Send to": after the extension sent, opening the app ALSO popped the web
"Send to" modal for the same file. Cause: the extension wrote a safety-net
manifest up front, which the app then picked up. Now the manifest represents an
UNSENT share only — written solely when the extension can't send (no token) or a
send fails. A successful in-sheet send clears the staged files and leaves nothing,
so the app never re-offers it. Cancel also clears staged files (no orphans).
2. Layout aligned to the Teams reference:
- Preview strip of thumbnails for what's being shared (image → the image, video →
first frame via AVAssetImageGenerator, else a doc icon).
- Radio selectors on the right — an always-visible empty circle that fills to a
navy check when selected (clearer multi-select than an appear-on-select tick).
- "Recent chats" section header; subtitle under each name (Direct message /
Group · N members).
- Clearer branding: bold white "Share to Biz Connect" on the navy bar.
Native-only — NEEDS A NEW iOS BUILD. Balance + selectors checked.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three follow-ups on the in-sheet picker:
- Profile photos: rows now show the real avatar (fetched from the conversations
API's `avatar` field with the bearer token, or a data-URL decoded inline),
rendered as a circle; coloured initials as the fallback — matching the app.
- Multi-select: tap toggles a checkmark instead of sending immediately; a "Send (N)"
button in the nav bar sends to every selected chat. Each file is uploaded ONCE and
its attachment id reused across all targets (the server allows the uploader to
reattach the same id), so multi-send doesn't re-upload.
- Branding: navy (#1F3B73) navigation bar with a white "Biz Connect" prompt over the
"Send to…" title and white controls.
Native-only — NEEDS A NEW iOS BUILD. Balance + selectors checked.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replaces the wrong approach (stage + try to bounce into the app, which iOS blocks)
with the one Teams/WhatsApp actually use: the picker and the send happen INSIDE the
share extension, so there's no app-open at all. Tap Share → Biz Connect → pick a
chat → it uploads and sends, right there in the sheet.
How the extension can send without the app: it's a separate process that can't see
the web app's HttpOnly cookie, so:
- server: GET /api/share/token mints a bearer token for the logged-in user.
- web: on every launch the app fetches that token and hands it to the extension via
the App Group (ShareInbox.setAuth writes token+base to the shared UserDefaults).
- extension: reads the token and calls the SAME API the native client uses —
GET /api/messages/conversations to list chats, POST /api/messages/upload for each
file, POST /api/messages to send. Native UITableView picker with search.
Robustness: it still stages the files + writes a manifest first, so if there's no
token yet (user never signed in) or the send fails, the file isn't lost — the app
collects it on next open, exactly as before. On success the manifest is cleared so
the app doesn't re-offer it.
Server + web are live now; the token endpoint is harmless until a build ships the
extension. NEEDS A NEW iOS BUILD for the picker itself.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A Share Extension opening its host app is unsupported on modern iOS (restricted
~iOS 14), so the programmatic bizconnect://share open is silently blocked and the
user just saw a blank flash back to Photos — looking broken even though the files
staged fine.
The extension now shows a small native card after staging: "✓ Ready to send — Open
Biz Connect to choose a chat", with an "Open Biz Connect" button (user-initiated
open has the best chance of working) and a Done button. It still attempts the
auto-open first. Either way the app collects the staged files when next opened, so
the manual path that already works is unchanged — this just removes the "did it
even work?" confusion.
Renamed the local `staged` array to `collected` to free `staged` for the state
flag. Balance + selectors checked. NEEDS A NEW iOS BUILD (native change).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Follow-ups from testing the share flow (which now works end-to-end):
- Send-to picker showed only initials — now shows the real profile photo when the
chat has one (matching the sidebar/forward avatars), coloured initials otherwise.
- During send it only said "Uploading 60%" with no idea WHO to — now the header and
progress name the target ("Sending to Manasa Rapolu · 60%").
- After sending, a video bubble sat blank (just a timestamp) for a second or two
while the poster generated on first view. media.js now warms the poster thumbnail
at UPLOAD (temp-then-rename), and the on-demand /thumbs handler also writes via a
temp, so the two can't serve a half-written JPEG. The bubble shows its poster
right away.
Web + server only — live on deploy. Does NOT address the share extension failing to
auto-open the app (an iOS limitation, handled next in the native build).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sharing a photo into the app: the "Send to…" picker appeared and staging worked,
but picking a chat said "no file found". Cause: the app loads its UI from the
REMOTE origin (remote.bizgaze.com), and I read the staged bytes with
fetch(convertFileSrc(uri)). Capacitor's local `_capacitor_file_` serving isn't on
the remote origin, so that fetch is CORS-blocked / hits the remote server → 404.
Read through the native bridge instead: Filesystem.readFile returns base64 in
native code, never touching the webview network stack, so it reads the App Group
file the app has entitlement access to. Falls back to it.path, then to the old
webview path for local-asset builds. Web-only fix — no rebuild.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The archive keeps failing with only "Failed to archive" and exit 65 — no reason
shown. Cause: `xcode-project build-ipa` prints a prettified summary and swallows
xcodebuild's raw "error:" lines, which only survive in the /tmp/xcodebuild_logs
artifact. On failure we now grep that log for the signing/entitlement/compile
error and print it inline, so the next red build states WHY instead of just 65.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Codemagic log: "xcodeproj not found at /Users/builder/clone/ios/App/App.xcodeproj"
— missing the `mobile/` segment. The script lives in mobile/scripts, so __dir__ is
mobile/scripts and `File.expand_path('../..', __dir__)` resolved to the REPO ROOT,
not mobile/. The project is at mobile/ios/App/App.xcodeproj. Changed to '..' so ROOT
= mobile, which fixes PROJECT, SRC_DIR, APP_DIR and the entitlements path together.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds the reverse direction: share FROM Photos/Files/Safari INTO a Biz Connect
conversation. An app can only appear in the iOS share sheet as an app-extension
target, so this is real native work, not a web change.
Pieces:
- mobile/ios-share/ShareViewController.swift: a UI-less Share Extension. It stages
the shared items into the App Group container and opens bizconnect://share. It
deliberately does NOT reimplement the chat picker — that lives in the app, which
already has the chat list, search and upload progress. Appends to the manifest
(never overwrites), so sharing twice before opening the app loses nothing.
- mobile/scripts/add-share-extension.rb: injects the extension target into the
Capacitor-generated Xcode project on every CI build (Codemagic checks out fresh),
using the xcodeproj gem that ships with CocoaPods. Embeds it, sets the bundle id
<app>.share, and MERGES the App Group into the app's entitlements rather than
clobbering them (push's aps-environment must survive). Idempotent.
- mobile/plugins/share-inbox: getPending()/clear() to read that manifest — the App
Group container isn't one of Filesystem's known directories, so it needs a bridge.
- home.html: on bizconnect://share (and every resume, and cold-launch), read the
inbox and show a "Send to…" picker over the chat list; chosen files run the SAME
upload + /api/messages send as an in-app attachment. Reuses convertFileSrc to read
the staged bytes with no base64 marshalling.
- ios-patch.sh registers the bizconnect URL scheme; codemagic.yaml fetches a profile
for the .share bundle id too.
One-time manual gate (CI cannot toggle App capabilities): the App Group
group.com.bizgaze.connect must be created and enabled on both App IDs in the Apple
portal — documented in mobile/IOS_SETUP.md. Without it the two processes can't see
each other's files and sharing silently no-ops; everything else still works.
Validated cross-file: pod-name/jsName/method wiring for all three plugins, App
Group id identical in all 4 files, URL scheme consistent across extension/plist/web,
entitlement-merge preserves push. Needs a new iOS build (new targets + plugins).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Files folder is the app's own copy — it powers offline playback and Manage
storage — but Files is not where anyone looks for photos and videos. The Photos app
is, and only PhotoKit can write there; @capacitor/filesystem cannot, because an
app sandbox and the photo library are separate stores. So this adds a small native
plugin, mirroring the existing audio-route one.
- mobile/plugins/media-library: saveToAlbum({path, album, kind}) finds or creates
the album and adds the asset. Uses addResource(with:fileURL:), which is uniform
for photo and video and non-optional, unlike creationRequestForAssetFrom*, which
can silently no-op.
- Requests .readWrite, NOT .addOnly: addOnly can add an asset but cannot look up or
create an ALBUM, which is the whole point here. Both photo-library usage strings
are already set by ios-patch.sh.
- If the album can't be resolved (e.g. "limited" access), the asset is still saved
to the camera roll — landing somewhere beats failing outright. A racing create
from two simultaneous downloads re-looks-up instead of erroring.
- Podspec named MediaLibrary.podspec with s.name = 'MediaLibrary' to match
PascalCase of the package name — the same trap that broke the AudioRoute build.
Checked by a script: pod name, jsName and declared-vs-implemented methods.
Entirely best-effort from the web side: a denied permission or an older app build
never fails a download that is already safe in the app folder. Added a Settings
toggle since this does keep a second copy of the file.
Needs a new iOS build — new native plugin.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The download button was a small corner overlay, which landed on top of the native
control bar's speaker icon. That was not the design asked for either: it should be
a single control in the CENTRE that turns into a play button once downloaded.
The tile is now the masked poster plus one centred button:
not downloaded → download icon (the element has no src at all, so nothing is
fetched until it is tapped)
downloading → live % inside that same button, in place
downloaded → play icon; tapping plays the LOCAL file
The native control bar is switched on only when playback starts, so there is
nothing for the control to collide with. Progress reports into the button rather
than the floating chip, so a video download no longer shows two indicators.
Also gives the tile a min-height so the thread doesn't jump while the poster loads.
Web/PWA is untouched — it has nowhere to download to, so it keeps streaming with
native controls.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three connected pieces, so downloads stop being write-only.
1. A folder the user can actually find. ios-patch.sh now sets UIFileSharingEnabled
and LSSupportsOpeningDocumentsInPlace, and downloads go to the app's Documents
folder in typed subfolders. iOS shows it as:
Files -> On My iPhone -> Biz Connect -> Images | Videos | Files
Previously everything was written to CACHE (private, and iOS purges it whenever
it likes) and pushed straight at the share sheet, so nothing was ever really
"kept" by the app.
2. Manage storage (Settings -> Storage). Lists what this device has downloaded,
grouped by type with per-group and total sizes; each row can be shared to the OS
sheet (this is where "Save to Photos" now lives) or deleted. Plus Delete all.
Deleting removes ONLY the local copy — the attachment stays on the server, so
anything deleted can be downloaded again from the chat.
3. A downloaded video never downloads twice. Images and files have their own
download link, but a video's tile IS the player, so it had no control at all and
re-streamed on every play. It now carries a download button; once saved, the
button becomes a tick and the tile plays from the local file — no network.
The index is treated as a cache of the filesystem, never as truth, because the user
can delete these from the Files app behind our back: every listing re-stats and
forgets what is gone, the library is reconciled at startup, and a local file that
has vanished by play time falls straight back to streaming instead of showing a
broken player.
Unit-checked the path allocator: collisions between different attachments with the
same filename resolve to "name (2)", re-downloading the SAME attachment reuses its
path, and path traversal / illegal characters are neutralised.
Note: the folder and the save location need a new iOS build to take effect. The web
side degrades cleanly — none of this UI appears outside the native app.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Uploads showed a bar + %; downloads showed nothing. On the native app that gap is
worse than on web, because there is no browser download UI behind it — the app
fetches the bytes itself, so a large video looked like a frozen tap.
- Adds the same chip an upload uses (name, bar, %) driven by Content-Length.
Falls back to an indeterminate "…" when the server sends no length.
- Streams the response and appends to the file in 3-byte-aligned blocks rather
than holding blob + base64 simultaneously. The old path peaked around 250 MB of
memory for a 75 MB video, which is enough to get a WebView killed on a phone.
Verified byte-exact against empty / 1 B / 2 B (base64 padding edges) / ragged
chunk sizes / 27 MB, reassembling identical bytes every time.
- Older WebViews without streams keep the previous one-shot path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reported: "the loading buffer is not the spinner, it still loads at the left of
the timer." Correct — and it was our bug, not a cosmetic preference.
The spinner was driven by a hand-picked event list (waiting/seeking/stalled). But
a cold start from preload="none" NEVER fires `waiting`: it runs
loadstart -> loadedmetadata -> loadeddata -> canplay -> playing straight through.
Since the bandwidth fix, that cold start is the only stall left — so the spinner
sat out the exact moment it existed for, leaving just the OS control bar's own
small indicator where the play button sits, i.e. left of the timer.
Now the spinner is derived from the element's real state rather than guessed from
events: busy = seeking || (!paused && !ended && readyState < HAVE_FUTURE_DATA),
recomputed on every relevant media event. Simulated against the real event
sequences before shipping — cold start, mid-stream stall and seek all spin;
paused/ended/idle never do.
Also dim the frame to 72% brightness while buffering so the spinner reads
instantly against a bright poster, and give it a dark backing disc.
Note: the small indicator inside the native control bar belongs to the OS's own
video controls and cannot be suppressed while we use them. Ours is now the loud,
central one; removing the OS indicator entirely would mean custom controls.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
THE ANSWER to "why does an already-downloaded video still buffer?" — it was never
the download, and it was not the server. Probing the real uploads on the box:
d0e49e58… 1920x1080 19.4 Mbps 75 MB / 31 s
ad929d0b… 1920x1080 19.0 Mbps 27 MB / 11 s
9f4e0865… 720x1584 3.6 Mbps 14 MB / 31 s
To play a 19 Mbps file the client has to SUSTAIN a 19 Mbps download for the whole
clip. No mobile link does, so the <video> buffer drains every few seconds: buffers,
plays, buffers, plays. Server-side disk read was instant and load was 1.7 on 20
cores throughout — the bottleneck is the media itself, not the delivery path.
Second, independent defect: phone MP4s store `moov` AFTER `mdat` (verified on two
uploads), so the player must fetch the file's tail before it can start at all.
Fix — keep the original bytes untouched (that is what the download button serves,
full quality) and build <id>.web.mp4 beside it: longest side capped at 1280,
~2.5 Mbps ceiling, +faststart. Measured on the 19 Mbps file:
27.3 MB @ 19.0 Mbps -> 2.55 MB @ 1.78 Mbps (10.7x less bandwidth)
transcode took 2.4 s for an 11.5 s clip
- server/media.js (new): probe, decide, 2-at-a-time background queue. Already
light + correctly sized + faststart => no rendition at all. Light but wrong atom
order => remux -c copy (seconds, no re-encode). Otherwise re-encode. A rendition
that lands bigger than the original is discarded. MP4 box-walker for the
faststart test is unit-checked against known fast/slow files, both directions.
- /stream/<id> serves the rendition, falling back to the original while it is still
transcoding, so a video is never unplayable. /files/<id> is unchanged and still
serves the pristine original for download.
- Renditions are queued at upload, and backfilled 15 s after boot for the videos
that predate this. Range serving is now one shared helper for both routes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause of a downloaded/streaming video buffering repeatedly: every /files Range request
(a playing video fires dozens) re-ran the full attachment authorization, which scans the
messages table by attachment_id (un-indexed) — a per-chunk table scan = stutter. Now the
auth decision is cached per user+attachment for 60s (module-level, bounded), so range
requests after the first are ~free.
Also: preload='none' (nothing about a video downloads until the user taps play — only the
small poster loads), per 'no auto-download'. And the buffering spinner no longer hides on
canplay/loadeddata (they fire mid-buffer), so it reliably spins whenever it's buffering.
build batch159.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The native controls' buffering indicator sits by the timer and is easy to miss. Wrap the
<video> and overlay a centered spinner shown while it buffers (waiting/seeking/stalled ->
show; playing/canplay/seeked/loadeddata -> hide). Media events don't bubble, so the
listeners run in the capture phase; the overlay is pointer-events:none so it never blocks
the native controls. build batch158.
- Dockerfile: add ffmpeg (Alpine).
- static.js: new /thumbs/<id> — ffmpeg extracts the first frame (0.5s), caches it next to
the file, serves as the video poster (cosmetic; 404s gracefully if ffmpeg unavailable).
- static.js: /files now supports HTTP Range (206 Partial Content) + Accept-Ranges, which
iOS requires to stream/seek video reliably (fixes the buffer-before-play / multi-tap);
media (image/video/audio) now served inline, other files still download. Shared
attachment auth refactored into one helper used by /files and /thumbs.
- home.html: video poster points at /thumbs/<id>. build batch157.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Swipe-back: the finger-following version is dropped (parked for the future) per request —
back to the reliable release-triggered swipe (rightward edge release runs bzcBack's slide).
Video: the custom download->play overlay caused layout 'dancing' on load and flaky
multi-tap playback. Replaced with a plain native <video controls playsinline preload=
metadata> (poster via #t=0.1) at a fixed box size — poster + OS play button, plays inline
on one tap, streams once and is cached (no re-downloads). build batch156.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Bug 1: .ptr-ind used top:8px so the pull-to-refresh spinner sat under the notch/Dynamic
Island. Now top:calc(var(--sat)+8px) clears the safe-area inset.
Bug 2: video attachments rendered as a plain download link that re-downloaded on every
tap. Server now sends isVideo/isAudio on message attachments; videos render as an in-chat
player — masked poster with a DOWNLOAD button that loads the file ONCE (preload=none ->
load on tap), then becomes a PLAY button; playing hands off to native inline controls, so
no repeat downloads. build batch155.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The conversation pane now tracks your finger from the left edge and reveals the chat
list behind it, completing the back past ~35% width or springing back otherwise. Reuses
the existing body.chat-dragging layout (already defined, identical to chat-closing that
showWelcome uses): content z-index:2 at translateX(0), list .chatcol absolute behind at
z-index:0 — so the pane starts at the correct on-screen origin (the earlier attempt's
'one screen-width off' bug was a different setup). Vertical drags still scroll; popup/
search edge-release still closes via bzcBack. build batch154.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Investigations are closed, so strip the probes: bzDbg + __perfProbe definitions, the
older/render/slide/pwaFocus call sites, and the server-side /api/dbg (MDBG) sink. Kept
the functional code around each probe (renderThread's innerHTML build, the older-page
re-anchor, the slide fade). No behavior change. build batch153.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause of 'tapping stops the Bluetooth audio': bzUnlockAudio fires on EVERY tap
during a call and called bzApplyRoute -> setSpeaker(false) -> configureDevice, whose
setCategory+setActive(true) reconfigures the AVAudioSession mid-call and interrupts
WebKit's audio unit, dropping the call audio. Fix: bzApplyRoute on iOS no longer touches
the session at all (only wires the route-icon listener); iOS keeps auto-routing. build batch152.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Telemetry confirmed the ceiling: a single speaker override holds ~1s over active BT then
WebKit reverts. So stop trying to control the route on iOS — always use the default port
(iOS auto-routes: BT/wired if connected, else loudspeaker), and make the button a live
INDICATOR of the real output; tapping shows a toast (connect/disconnect a headset to
change). Removes the temp nroute/sptap probes. Pure web change; plugin v1.1.2 already
supports it. build batch151.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
v1.1.1's route observer re-forced speaker on every change, and WebKit re-added Bluetooth
each time -> the audio flapped speaker<->BT many times/sec (telemetry: dozens of route
flips from 3 taps), which read as 'sound doesn't switch'. WKWebView won't let the app
hold the built-in speaker over an active BT device. So: one override per tap, observer
only REPORTS the output (no fighting). Also stop dimming the iOS button (it's a live
output indicator, not on/off; the dim read as 'disabled' on BT). Probe now carries the
native marker so we can confirm the binary. plugin v1.1.2-stable, web batch150.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
overrideOutputAudioPort(.speaker) alone can't beat a connected BT headset (BT is higher
priority), so 'Speaker' snapped back to BT. Now setSpeaker(true) sets category options
[.defaultToSpeaker] (no allowBluetooth) so BT isn't an eligible output and the speaker
wins; setSpeaker(false) restores [.allowBluetooth,.allowBluetoothA2DP] and uses the
default port (routes to the headset). Observer re-holds speaker if a BT connect steals it.
Adds a TEMP web probe (nroute/sptap) to verify from telemetry. plugin v1.1.1, web batch149.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
BT audio already routes on iOS, but the web UI can't see it (iOS hides audio outputs
from enumerateDevices), so the icon was stuck on speaker. Plugin v1.1.0 now exposes the
active output: getRoute() + a 'routeChange' event ('speaker'|'bluetooth'|'wired'|
'receiver'|'airplay'). Web subscribes and drives the icon/label from the real route
(bluetooth/headphones/speaker), and the iOS toggle becomes a 2-state Speaker <-> Device
cycle (JS can't enumerate outputs there). Also strips the earpiece-investigation debug
logging from the plugin. Native needs one Codemagic build; web is live (batch148).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Proven via device telemetry: inside a WKWebView, WebKit owns the WebRTC audio unit and
forces the loudspeaker; overrideOutputAudioPort(.none) is a no-op (route settles on
Speaker 1.2s later), so the built-in earpiece cannot be selected. Present only what
actually works on iOS: Speaker, and Bluetooth/wired headset when connected. bzApplyRoute
now maps only 'speaker' to the loudspeaker override; bt/headset use the default port.
Coerce any stale 'earpiece' pref to speaker on iOS. Also strips the route debug telemetry.
Other platforms keep the earpiece option. build batch147.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
1.0.3 telemetry showed mode already == .voiceChat when earpiece still landed on
Speaker, and setSpeaker read the route synchronously (stale) right after the override.
So the mode re-pin alone may be insufficient and the sync read is unreliable.
1.0.5: keeps the .voiceChat re-pin, but stamps live mode into every route-change log
line and re-reads the SETTLED port+mode at +0.4s and +1.2s after each toggle (reported
in the next toggle's trail). This definitively answers whether WebKit flips to .videoChat
and where the route truly settles. native marker 1.0.5-settle.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause (WebKit source MediaSessionManagerCocoa.mm + Apple DTS, confirmed by
telemetry): WKWebView's WebRTC re-pins the session to mode .videoChat while capture
is active, and .videoChat auto-implies .defaultToSpeaker. So override(.none) reverts
to the mode default = LOUDSPEAKER, and .none alone can never reach the earpiece once
WebKit flips the mode. Our first override won only because .voiceChat was still active.
Fix: for earpiece, setMode(.voiceChat) (its default route IS the receiver) before
override(.none) in both setSpeaker and the debounced route-change re-assert. Add an
accessory guard so a connected BT/wired headset isn't yanked to the built-in receiver.
Reconcile the launch patch: drop .defaultToSpeaker from inject-audio.js so it stops
contradicting the plugin. native marker 1.0.4-mode.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Telemetry proof: overrideOutputAudioPort(.none) routes to Receiver once, then every
later toggle latches to built-in Speaker despite opts=36 (no .defaultToSpeaker) —
WebKit's WebRTC engine re-forces the loudspeaker after our override. The prior build
ignored .override-reason route changes and never corrected it.
Now: react to ALL route changes, debounced 0.25s, and re-assert the chosen port only
on a genuine mismatch (self-terminating, capped at 6/toggle to avoid thrash). setSpeaker
returns a reason->port route-change trail so the log shows whether WebKit is one-shot
or persistent. native marker 1.0.3-reassert.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Telemetry only showed 'override didn't throw', and the web __BUILD tag can't tell
native binaries apart, so we couldn't see WHERE iOS actually routed the audio or
which plugin build ran. setSpeaker now resolves with the real currentRoute output
port (Receiver/Speaker/Bluetooth), the live AVAudioSession category/mode/options,
and a native-build marker (1.0.2-diag) so the route log is unambiguous.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Speaker worked but the earpiece was silent, and call audio flickered on speaker
then dropped. Causes: (1) the category set .defaultToSpeaker, so overrideOutputAudioPort(.none)
fell back to the loudspeaker instead of the receiver; (2) setSpeaker re-ran
setCategory+setActive on every toggle mid-call, tearing down the audio unit WebKit's
WebRTC engine was using and silencing the earpiece route.
Fix: drop .defaultToSpeaker (drive the port explicitly), make setSpeaker flip ONLY
overrideOutputAudioPort, and observe routeChangeNotification to re-assert the chosen
route when WebKit reconfigures the session at call start / device change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Capacitor derives the pod name from the npm package name (audio-route -> AudioRoute)
and writes 'pod AudioRoute, :path => ../../plugins/audio-route' into the generated
Podfile. CocoaPods then requires a file literally named AudioRoute.podspec whose
s.name is 'AudioRoute'. The old AudioRoutePlugin.podspec (s.name AudioRoutePlugin)
caused the Codemagic build to fail with 'No podspec found for AudioRoute'.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Device telemetry proved the app-embedded AudioRoute class never registered (not in
Capacitor.Plugins) — appending a CAPBridgedPlugin to AppDelegate.swift gets stripped/undiscovered
in release builds. The plugins that DO register (Share, Camera, Filesystem) are all npm packages
wired by cap sync. So AudioRoute is now a local plugin package (mobile/plugins/audio-route,
file: dep in mobile/package.json) with a podspec + CAPBridgedPlugin Swift — cap sync adds its pod
and Capacitor registers it like the others. load() sets the launch speaker default; setSpeaker({on})
overrides the output port. inject-audio.js no longer injects the plugin class (would duplicate);
it keeps only the AppDelegate launch default as a fallback.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Server streams the upload body straight to /data/uploads (a .part temp file, atomic rename on
success), backpressure-aware, so a 1 GB file never buffers in RAM. MAX_UPLOAD_MB env (default
1024 = 1 GB) controls the cap; error message reflects it.
- Client size guard raised 25 MB -> 1 GB.
- docker-compose documents MAX_UPLOAD_MB and the required Nginx Proxy Manager client_max_body_size.
NOTE: the actual bottleneck for the user's 9.7 MB reject is almost certainly NPM's client_max_body_size
(nginx default 1 MB) — that must be raised in the NPM admin; the app change alone can't lift it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- ios-patch.sh now injects an AudioRoutePlugin (CAPBridgedPlugin, Capacitor 7 auto-registers it)
into AppDelegate.swift with setSpeaker({on}) -> AVAudioSession.overrideOutputAudioPort. Tolerant/
build-safe: if it doesn't register, the web call just no-ops (can't crash or fail the build).
- web: nativeAudioRoute()/bzApplyRoute() drive the plugin; toggleSpeakerphone + the on-join/on-tap
unlock now actually switch the route on iOS (setSinkId can't). canRouteAudio() shows the toggle
when the native plugin is present. Dormant until the next Codemagic build ships the plugin.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Drag the open conversation rightward from the left edge and it follows the finger 1:1 with the
list parallaxing in underneath; release past ~35% (or a quick flick) completes the pop, else it
snaps back. Fixes the old stuck-pane bug: once a clear horizontal drag is detected we
preventDefault (passive:false) to CLAIM the gesture so iOS/scroll can't steal it and fire
touchcancel; touchcancel always resolves to a clean state. showWelcome(skipAnim) does the state
swap without re-animating. Works in native app, PWA and mobile browser.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Calls were routing to the quiet earpiece. ios-patch.sh now runs a Node helper (Node 20 is
already in the build env) that injects an AVAudioSession .playAndRecord/.voiceChat category with
.defaultToSpeaker + Bluetooth into the generated AppDelegate. Tolerant: exits 0 and no-ops if the
template differs, so it can never fail the Codemagic build. Verified locally against the Cap 7
AppDelegate template — injects correctly and is idempotent. First pass; if WebRTC re-grabs the
session mid-call on device, a follow-up plugin will re-assert .overrideOutputAudioPort(.speaker).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Suppressing the save dialog made downloads completely silent — "nothing happened" even though
the file saved to Downloads. Now on download 'done': (1) tell the web UI → branded toast
"Saved X to your Downloads folder" when the window is focused; (2) native OS notification
(click → reveal in Explorer) when the app is minimized/in the tray, so it's never double-noticed;
(3) a failure notice. preload exposes onDownloadDone; home.html shows the toast.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The app closes to tray, so it kept running on the old version and only checked on a fresh launch.
Now: (1) check for updates on EVERY window show (X-to-tray then reopen, taskbar, relaunch via
second-instance) via win.on('show'), throttled 1/10min; (2) explicit autoInstallOnAppQuit=true so
a downloaded update installs on the next real quit / PC restart even if the user never clicks
"Restart now"; (3) native notification when an update finishes downloading while hidden in the
tray. The existing 6-hour background timer is unchanged and still runs regardless of window state.
Also carries the 0.1.18 downloads-to-Downloads-folder change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Accessory bar: call Keyboard.setAccessoryBarVisible({isVisible:false}) on iOS in the native
IIFE, hiding the grey chevrons+Done strip above the keyboard. The plugin is already bundled
in the current TestFlight build, so this takes effect on a web deploy — no rebuild.
- Attach: tapping the paperclip now opens a composer-anchored Photos/Camera/Document menu
(like the emoji/mention popups) instead of firing the generic mid-screen OS chooser as the
first thing. Each option opens a type-scoped picker (image/*, capture). Web fix, all clients.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Mobile: the Return key now inserts a newline instead of sending (send via the button, like
every native chat app). Desktop keeps Enter=send / Shift+Enter=newline.
- Image preview: hide the on-screen +/- zoom buttons on mobile (pinch-to-zoom covers it).
- Desktop (Electron): a will-download handler saves straight to the OS Downloads folder with no
"where to save?" dialog, de-duping the name if it exists. NOTE: desktop code only — NOT
published to the update feed (needs an explicit desktop rebuild/publish).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The mobile viewport fixes only ever went into home.html. The other standalone pages still had
the old viewport, so in the native app they auto-zoom on input focus and (dashboard/host) run
under the notch/Dynamic Island. Bring every page's viewport to match home.html
(maximum-scale=1, user-scalable=no, viewport-fit=cover) and add safe-area top padding to the
dashboard header and the host body/indicator so nothing sits under the island now that the
viewport is cover. index/connect/share already pad for safe-area; they only needed maximum-scale.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
iOS home-screen PWAs IGNORE viewport maximum-scale (Apple disabled it for accessibility), so
the maximum-scale=1 that stops auto-zoom in the Capacitor WebView does nothing in the PWA — it
still zooms on input focus. Real cross-platform fix: make every focusable text field 16px on
mobile (the composer was .92rem). autoGrow's empty-guard keeps it one line.
Also the PWA has no Capacitor Keyboard plugin, so the composer never lifted above the keyboard
("not the same keyboard"). Added a VisualViewport-based lift for non-native clients (browser +
PWA); native still uses the plugin. >100px threshold ignores the Safari toolbar.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Bug: for ~3.4s after opening a chat, the "keep newest in view as images load" pin schedule
would yank you back to the bottom if you tried to scroll up for older history — so scrolling
up snapped back to latest, and only worked once the schedule expired (~5s). Fix: the user's
first scroll gesture now sets _openScrolled, which cancels the pending pin timers and short-
circuits _pinNewest (so the late image-load pins stop too). loadOlder/anchor then works
immediately, no yank.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Device log showed the slide motion itself is now clean 60fps (frames 0-15 all ~16ms in every
sample); the only remaining spike was at frame ~16 — AFTER the slide lands — from applying the
network thread result + pin at that instant (the "adjusts after it lands" feel). Since the open
is a compositor opacity/transform fade, main-thread work during it doesn't stall the animation
and is masked by the low opacity. So apply the reconcile/render + pins immediately when the
fetch returns (mid-fade) instead of deferring to slide-end. By the time the pane is fully opaque
the content is already settled. Removed the now-obsolete appendBubble mid-slide deferral.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The reflow that caused the end-of-slide hitch was never the content slide itself — it was
SHOWING the list column during the slide and hiding it (display:flex->none) at the end. So
keep .chatcol display:none the whole time (untouched) and animate ONLY .content: it pushes in
from translateX(12%)->0 with an opacity fade, on its already-promoted GPU layer. Real slide
motion, and nothing to reflow when it ends. (This also explains the batch126 breakage: that
made .chatcol position:absolute, disturbing the touch/layout target — not touched here.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
User chose the fade over chasing the WebView slide-teardown reflow. Open now: the cache-
rendered conversation is placed instantly and fades opacity 0->1 over 140ms. No transform,
no off-screen pane, no list-column display toggle -> nothing to reflow at the end, so the
~40ms end-of-slide settle cannot occur. Same afterOpenSlide queue flushes network reconcile/
pins/appends at fade-end. Desktop unaffected (fade is mobile-only via __freshOpen).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The list-column position:absolute change stopped conversations from opening. Revert to the
known-good display:none. Keep the other batch124-125 fixes (content promoted, reconcile,
reopen-at-latest, download interceptor). Slide back to the batch125 behaviour (opens fine,
minor end hitch) while I find a safe way to remove the reflow.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Device log proved the residual 30-78ms end-of-slide hitch is NOT the content-layer demote
(content stays promoted) and NOT the re-render (reconcile, longtasks:[]). It's the chat-list
column flipping display:flex->none at transitionend, which reflows that whole subtree. Keep
the list RENDERED behind the conversation (position:absolute; z-index:0; covered by the opaque
content pane) instead of toggling display. Teardown now only clears transforms = no reflow.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Slide: the residual ~40ms end-of-slide hitch (down from 128ms in b124) was the layer
DEMOTE + the deferred full renderThread at transitionend. Two fixes:
- Keep .content promoted for the WHOLE time a chat is open (will-change lives on
body.chat-open, not on the animating class). The open transition now only changes the
transform VALUE 100%->0; when it ends there is nothing to tear down -> no demote raster.
- Slide-end no longer rebuilds all 40 nodes: reconcileOpen() appends only the 0-2 genuinely
new tail messages (full renderThread only if the page structure diverged).
Download: routing to Safari failed auth (no login cookie). Real fix: fetch the file in the
WebView (cookie present) and hand the bytes to the OS save/share sheet via Filesystem+Share
(added to mobile deps; ships next TestFlight build). Until then, images fall back to the iOS
long-press "Save to Photos" instead of breaking the app.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Slide: the 40-128ms hitch was a compositor RE-RASTER at transitionend, not JS
(longtasks:[], renderMs 2-6ms in the device log). Removed contain:layout paint /
isolation:isolate from the .content slide layer and the filter:brightness parallax
on .chatcol (both force a full re-raster on teardown); defer the layer demotion two
frames past transitionend so it lands off the motion path. Also replaced the fragile
length+lastId slide-end skip (it almost never held once a live message grew the cache,
so the full innerHTML rebuild kept running at slide-end) with an exact per-message diff.
Reopen-at-latest: loadOlder wrote the GROWN thread (100+ msgs) into THREAD_CACHE, so
re-opening re-rendered all of it and stranded you mid-history as images shifted the
(pre-load) bottom past the 1200px pin guard. Open now renders the latest PAGE only;
added _forcePinOpen to glue to the newest through late image loads until the user scrolls.
Download hang: <a download> navigated the whole WKWebView away to the raw file (no back,
app frozen). Native-only capture-phase interceptor opens downloads in the system browser.
Lightbox close/download buttons: solid dark chips so they're visible over bright images.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Device log: slide is 60fps EXCEPT one ~60-130ms hitch at the END = the deferred renderThread rebuilding
the (identical) DOM at slide end. Skip it when the network result matches the cache render (signature).
Load-older: fired at scrollTop<120 = during the rubber-band over-scroll past the top (beforeScrollTop
was -289), so the prepend+anchor jerked momentum. Trigger earlier (<700, skip negative scrollTop) so it
loads while still scrolling and never interrupts the bounce.
Build marker -> 2026-07-19-batch123.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The perf-diagnostics script tag was never closed, so it merged with the following native/keyboard
<script> and threw 'Unexpected token <', disabling the keyboard/safe-area init. Add the missing </script>.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Re-add /api/dbg + a frame-gap/long-task probe over the chat-open slide, the innerHTML render time, and
before/after scroll for load-older — to measure on-device where the jank actually is (main-thread block
vs compositor) instead of guessing. Temporary; removed once fixed.
Build marker -> 2026-07-19-batch122.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Slide (workflow-verified): the jump was the main thread building the message DOM + scroll writes DURING
the 300ms animation. Now: render off-screen first, then arm a transform-only slide on a promoted GPU
layer (translate3d/contain:layout paint) with all innerHTML/scrollTop work deferred to transitionend
(afterOpenSlide queue). armOpenSlide runs after the sync render; appendBubble defers live writes mid-slide.
Load-older jump: anchor on the specific oldest-loaded message element (not scrollHeight math, which broke
when prepended images loaded and shifted content) and re-anchor as those images load.
Removed all /api/dbg diagnostics (client reporter + server route + spike/diag probes).
Build marker -> 2026-07-18-batch121.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
User's insight: the jank/jumps come from rendering ALL ~500 messages on open. Load only the latest 40
(server + client PAGE), and the existing loadOlder() pages in older history on scroll-up with a scroll
anchor (no jump). Whole-thread SEARCH is a separate endpoint, unaffected. Also #1: keyboard-show only
pins to newest when already near the bottom, so replying to an OLD message no longer yanks to latest.
Build marker -> 2026-07-18-batch120.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Pager: force the staged (opposite-edge) frame to PAINT (void offsetWidth) before animating in, so the
incoming image enters from the correct side (was appearing from the exit side). Hide prev-arrow at the
first image and next-arrow at the last.
- #1: keyboard closed after reply-cancel but the --kb lift stayed (empty gap). Add keyboardDidHide reset +
a composer-blur fallback that drops --kb when focus leaves the composer.
Build marker -> 2026-07-18-batch119.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Reply quote: was showing the FULL quoted body untruncated -> DIV.quote rendered 1571px, ballooning the
bubble + leaving the chat not-at-bottom. Clamp to 2 lines (-webkit-line-clamp) + collapse newlines.
- Swipe-to-reply: translateX(+72px) on a right-edge bubble pushed past the viewport -> a horizontal scroll
bar. .convo-msgs overflow-x:clip stops it.
- Image pager (#8): replace instant src swap with a real slide — current image slides out, next slides in
from the opposite side; clamp at first/last (no wrap -> no blank slide past the end).
Build marker -> 2026-07-18-batch118.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Parallel root-cause analysis: during body.chat-opening the pane is momentarily ~0 width; word-break:
break-word / overflow-wrap:anywhere let bubbles collapse to 1-GLYPH min-content, so pre-wrap text wraps
into hundreds of lines (5684px 'balloon'); same inflates the empty/draft textarea (autoGrow->140=4 lines);
and translateX(100%) of a ~0-width pane = no travel = the 'jump'. The tall box hid in anonymous line
boxes (renderMsgBody emits bare text nodes), which is why the probe only saw the 84px timestamp span.
Fixes (all @media max-width:760px / chat-opening scoped, desktop untouched):
1. .bubble: word-break:normal + overflow-wrap:break-word (min-content = longest word, never collapses)
2. mobile bubble group: overflow-wrap:anywhere -> break-word
3. .msg-link: word-break:break-word -> overflow-wrap:break-word (URL-only msgs can't reopen the hole)
4. body.chat-opening .content: width:100vw -> definite width, no collapse frame, real slide travel
Build marker -> 2026-07-18-batch117.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
tallImg was no-img, so the balloon is a TEXT bubble stretching (even in a 17-msg chat). Add: composer/
textarea height+scrollHeight+rows+value+fontSize+lineHeight (why ~4 lines), whether chat-opening class +
.content animationName are applied (why transition jumps), and the tall bubble's text + tallest child.
Build marker -> 2026-07-18-batch116.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Spike log showed bubbles ballooning to 3576-5684px (an image loading at natural ~2778px height before
constraint). att-img is capped 240 but some image path wasn't — added .convo-msgs img/.bubble img
max-height:340px as a blanket cap so no message can stretch the thread. Enhanced spike report to name
the tall bubble's img (natural size, rendered height, computed max-height) to confirm.
Build marker -> 2026-07-18-batch115.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Revert font-size:16px !important (it made the composer ~4 lines tall). Disable iOS zoom with
viewport maximum-scale=1,user-scalable=no instead -> composer keeps its size, AND double-tap-to-zoom
(in-chat search arrows #3) is gone. touch-action:manipulation on controls as belt-and-suspenders.
Custom lightbox pinch still works (it's transform-based, not browser zoom).
- Transition: show the newest message immediately (was hidden until images loaded -> empty pane slid in
then popped = the 'jumping' feel). Pinning still prevents dance.
Build marker -> 2026-07-18-batch114.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Transition: open now slides the whole .content pane in over the list with a subtle parallax + iOS
easing (was an instant/panel-only slide). Close/back already parallax.
- scrollH blip: added a spike watcher that reports the tallest element (class + img src tail) when the
thread's per-message height balloons, so we can see what stretches then settles.
Build marker -> 2026-07-18-batch113.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Device log (post-zoom-fix): keyboard open leaves clientHeight AND visualViewport at 926 -> vv gives 0 ->
composer covered. On resize:none the plugin's keyboardHeight is the only signal; now that zoom is gone
it's clean, so --kb = keyboardHeight lifts the composer flush, and keyboardWillShow at animation start +
the CSS bottom-transition makes it slide up WITH the keyboard (smooth). Config pinned to resize:none.
Build marker -> 2026-07-18-batch112.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Debug now reports curKb/compBottom/kbTop/coveredBy on keyboard events so I can see WHY the composer is
covered (lift coming out 0) from the server log.
- Transition: force the bzConvoIn slide to replay on every chat open (CSS animation only auto-plays on a
class change, so chat->chat looked instant). Restart via reflow + inline animation.
Build marker -> 2026-07-18-batch111.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Device log confirmed zoom persisted (vvW:394,vvTop:74) because .composer-row textarea{font-size:.92rem}
out-specified the generic rule, keeping the composer <16px -> iOS still zoomed on focus. !important
forces 16px on all fields so the zoom never fires. (#7 latest + budge already verified fixed.)
Build marker -> 2026-07-18-batch110.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Device log revealed the truth:
- Keyboard: focusing the <16px composer made iOS ZOOM the page (vvW 428->394, vvTop->74), scrambling
every viewport/keyboard calc. Set all form fields to 16px on mobile -> no zoom -> keyboard math holds.
- Latest messages: on a 500-msg chat atBottom was false because late images grow the thread after the
initial scroll. Now re-pin on a schedule + on each image load while near the bottom.
- Budge already gone (device reported ovf:false).
Build marker -> 2026-07-18-batch109.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Temporary: the app POSTs its real runtime state to the server log so device-only bugs (budge source,
keyboard viewport numbers, chat-open scroll state) can be diagnosed from docker logs instead of
screenshots. Removed once mobile issues settle.
Build marker -> 2026-07-18-batch108.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Keyboard: the 60ms timer in batch106 sampled the viewport MID-animation (clientHeight/visualViewport
temporarily inconsistent) and over-lifted -> gap. Reverted to reading clientHeight-visualViewport LIVE
on every vv resize, which is correct and self-adapting (unit test: residual=46 flush on the device).
- Budge: reverted the file <input> back to display:none (the position:fixed;left:50% hack overflowed
the viewport horizontally).
- Config back to resize:native to match the installed build (works, no rebuild).
- Added test/keyboard-lift.test.js (5/5 pass) verifying the lift math against real device numbers.
Build marker -> 2026-07-18-batch107.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
SravanandClaude Opus 4.8