Root cause fix for all broken CSS variable opacity modifiers. Converted
every hex CSS variable to space-separated RGB channels (e.g. #fda4af →
253 164 175) and updated Tailwind config to use rgb(var(...) / <alpha-value>).
This makes bg-accent-rose/10, border-border-hard/50, etc. work natively
everywhere — no more currentColor fallback. Reverted all previous explicit
rgba() workarounds back to clean Tailwind syntax. Also fixed the voice
connection wifi icon SVG (was bottom-left aligned, now centered).
Server-side: handleReactionAdd/Remove now check dm_messages table when
message isn't found in server messages, enabling DM emoji reactions
via the same unified client event path.
Client-side: replaced all remaining broken opacity modifiers on CSS
variables (border-border-soft/50, bg-accent-rose/10, bg-surface-*/N,
etc.) with explicit rgba values. Tailwind can't decompose var() for
opacity, causing fallback to currentColor. Fixes 16 files across
error boxes, table borders, surface tints, and input separators.
Consolidated two competing scroll effects into one to fix an effect
ordering bug where prevMessagesLength was updated before the initial
scroll check could read it. Channel switch now resets tracking state
so the snap-to-bottom fires reliably via requestAnimationFrame.
Increased message list bottom padding to clear the glass input bubble.
Tailwind can't decompose CSS variables for opacity modifiers, so
border-border-hard/50 fell back to white currentColor. Replaced with
border-white/[0.06] on image attachments, code blocks, and action
toolbar. Also applied frosted glass material to message action toolbar.
Replace @apply border-none with @apply border-0 on global * selector.
border-none sets border-style: none, which silently kills all Tailwind
border utilities (they only set width/color, relying on preflight's
border-style: solid). border-0 sets border-width: 0 instead, preserving
the solid style so border-r/border-l/border-b utilities render correctly.
- Add border-r border-border-hard to ChannelSidebar (both DM and server views)
- Add border-l border-border-hard to MemberSidebar and ActivityPanel
- Replace shadow-header with border-b border-border-hard on sidebar headers
- Remove duplicate mt-1 on reaction container (parent gap-1 already provides 4px)
Channel sidebar: subtle bg-surface-elevated highlight with rounded-[6px],
active channel left pill, rose unread dot on right, smaller #/voice icons
(18px), 10px horizontal padding, 24px voice user avatars at 36px indent.
Reaction bubbles: inline styles for background/border colors to bypass
Tailwind opacity modifier failures on hex CSS variables. Both mine (mint
tint) and non-mine states now render visible pill containers reliably.
PiP positioning:
- Replace hardcoded layout constants with dynamic DOM measurement system
- Obstacle elements declare themselves via data-pip-obstacle="left|bottom"
- getPipBounds() queries actual element rects at runtime via getBoundingClientRect
- MutationObserver + ResizeObserver re-clamp PiP when obstacles appear/resize
- PiP reappears after close when navigating away from voice channel
Reactions:
- Align reaction pills with Aether Drift prototype
- Own-reaction accent: purple → mint (bg, border, count color)
- Default pills: subtle white overlay bg + border-soft border
- Count text: 12px/semibold/txt-secondary per prototype spec
Bug fix:
- Deduplicate channel insertion in createChannel store action
- Prevents double-add race between REST response and WS broadcast
Avatar fallback gradients were hashed by display name alone when no
user prop was passed, causing the same person to appear in different
colors across messages, replies, member lists, voice panels, etc.
Added userId prop to Avatar and wired it through all 17 call sites
so the gradient always hashes by user ID.
Phase 5: Remove all 54 legacy Discord token definitions from tailwind.config.js,
migrate body classes in index.html to Aether Drift tokens, update ARCHITECTURE_AUDIT.md.
Phase 6: Replace flat purple fallback avatars with deterministic colorful gradients
(mint, sky, lavender, coral, rose, teal, amber) based on user/server ID hash.
Replace Discord logo SVG with Backspace "B" in server strip. Add per-server gradient
icons with hover glow. Update profile popout banner and incoming call modal to match.
Migrate all auth pages, modals, and shared UI components from Discord
visual language to Aether Drift design tokens. Zero discord-* classes
or raw Tailwind palette colors remain in components/auth/, modals/, ui/.
- Avatar: status dots use status-online/idle/dnd/offline tokens
- Tooltip, ContextMenu, UserProfilePopout: z-index raised to z-[200]
- LoginPage, RegisterPage: warm bg-surface-base with lavender glow
- All modals: inputs use surface-input, buttons use accent-primary
- Toggles: bg-status-online (on), bg-surface-input (off)
- ServerSettings: slider tracks, pills, danger zone fully tokenized
- Purged raw Tailwind green-500 leak in StreamingLimitsPanel success msg
Migrate all 10 voice/video components from Discord visual language to
Aether Drift design system. Zero discord-* classes or hardcoded hex
values remain in components/voice/.
Key changes:
- Status indicators use pastel accents: mint (connected/speaking),
amber (connecting/idle), rose (muted/deafened/error)
- VoiceControlBar converted to frosted glass-bubble material
- ConnectionInfoPopover and ScreenShareSettingsPopover use glass popovers
- Speaking ring glow changed from Discord green to mint (#86efac)
- Slider accents use lavender instead of blurple
- All bg/text/border tokens mapped to surface-*/txt-*/interactive-* system
- Add Backspace-design-prototype.html: finalized "Aether Drift" design
language with warm matte surfaces and subtle frosted glass accents
- Update CLAUDE.md with DESIGN SYSTEM section and remove Discord clone references
- Rename all Opencord references to Backspace across the full codebase
- Archive outdated design experiments and Discord-specific assets
- Add science-backed accessibility fallback (prefers-reduced-transparency)
Add a server-side instance_settings table (single-row, CHECK(id=1))
that stores admin-configurable streaming bounds: bitrate min/max/step,
allowed resolutions, and allowed framerates.
Backend:
- New instance_settings schema + migrations (is_admin on users, default
settings row, first-registered-user promoted to admin)
- GET/PATCH /api/settings/streaming endpoints with admin-only writes
and full input validation including cross-field checks
Frontend:
- settingsStore fetches limits on WebSocket ready, exposes isAdmin flag
- ScreenShareSettingsPopover reads bounds from store instead of
hardcoded constants, auto-clamps stale localStorage values
- buildScreenShareOptions() clamps bitrate to server limits at build
time as enforcement backstop
- ServerSettings modal gains a "Streaming" tab (admin-only) with
bitrate range sliders, resolution/framerate toggles, and save/reset
When a viewer joined while a screen share was active, autoSubscribe:true
caused a subscribe-then-unsubscribe dance for screen share tracks,
triggering cascading renegotiations, MaxListeners warnings, and
negotiation timeouts. Now the SFU starts with no subscriptions and we
explicitly subscribe only to non-screen-share tracks (audio, camera).
Screen shares remain controlled by the watch/unwatch UI flow.
Camera was stuck at 180p because our custom <video> rendering bypassed
LiveKit's adaptive stream observer. Replaced manual srcObject binding
with track.attach()/detach() in VoiceUser, StreamTile, and PictureInPicture
so the SFU receives viewport dimensions and forwards the correct
H.264 simulcast layer.
Screen share VP9 SVC with L3T3 spatial layers failed because hardware
VP9 encoders (NVENC, QSV, VCE) don't support spatial scalability —
Chrome silently degrades to L1T1. Reverted to VP9 single-layer
(simulcast: false, no scalabilityMode). Also targets encodings[length-1]
in applyOverdrive() for correct simulcast layer targeting.
- Add customBitrateKbps to ScreenShareConfig (null = auto matrix lookup)
- Slider in Stream Settings popover: 500 kbps–20 Mbps, step 500 kbps
- "Reset to Auto" clears override back to preset-derived bitrate
- Live updates via existing applyOverdrive() pipeline on active streams
- Persist version 5 → 6 with migration
- Add MarkdownRenderer with syntax-highlighted code blocks (prism-react-renderer),
GFM tables, and Discord-style theming
- Add MentionBadge that resolves user IDs to display names with role colors
- Add MentionPopover autocomplete triggered by @ in MessageInput
- Fix mention:// URL sanitization — allowlist the scheme in urlTransform so
react-markdown v9 passes it through to the component override
- Highlight messages that mention the current user (amber border)
- Bump rate limit from 60 to 200 req/min
Remove the legacy TEXT role column ('owner'/'admin'/'member') from
server_members and make the bitwise RBAC member_roles junction table
the sole authority for role assignments. Owner detection now uses
servers.ownerId exclusively.
- Remove MemberRole type and role field from shared types
- Remove role from Drizzle schema, raw SQL CREATE TABLE, and seed data
- Rewrite PATCH /members/:uid to accept { roleIds: string[] }
- Fix GET /members to populate roles array (was TODO)
- Replace member.role === 'owner' guard with isServerOwner()
- Remove getMemberRole() helper and legacy bridge code
- MemberSidebar groups by highest-positioned role instead of legacy string
- ServerSettings replaces admin/member dropdown with role checkboxes
- Message.tsx derives color from roles[] with owner fallback via ownerId
- Existing DBs keep vestigial column (Drizzle ignores it); new DBs omit it
Wire the bitwise permission engine end-to-end:
- Add sendToChannel() to ConnectionManager, filtering WS recipients by VIEW_CHANNEL
- Convert 6 channel-scoped events (messages, typing, reactions) from sendToServer to sendToChannel
- Add broadcastOverrideChange() to push channel_updated/channel_deleted per-user on override mutations
- Bridge legacy server_members.role TEXT to member_roles junction table on PATCH
- Add pushReadyPayload() to force re-sync frontend store after role changes
- Filter channels by VIEW_CHANNEL in GET /api/servers/:id to prevent initial load data leak
- Pre-compute viewers before CASCADE delete on channel_deleted
- Fix frontend channel event handlers to upsert/cleanup channelToServerMap and channelPermissions
- Add ChannelSettingsModal with Private Channel toggle and gear icon in ChannelSidebar
Replace string-based role checks (role === 'admin') with a bitwise BigInt
permission system. Adds computePermissions() resolution engine following
Discord's model: @everyone base → role union → admin shortcut → channel
overrides (role deny/allow → member deny/allow). Ready payload now filters
channels by VIEW_CHANNEL and attaches per-user myPermissions to each
server and channel. Includes channel_overrides table, @everyone role
auto-creation, migration for existing servers, and override CRUD API.
Wrap all multi-write DB operations in atomic transactions (server/channel
creation, message+attachment linking, DM creation, friend acceptance,
cascading deletes) to prevent partial-write corruption.
Batch N+1 queries in WS ready payload into O(1) bulk fetches with
chunked inArray() to respect SQLite's variable limit.
Fix chat history regression where background WS messages bypassed
channel load by switching the guard from messages.has() to hasMore.has().
Add LRU channel eviction (20 cached, evict to 15) and per-channel
message cap (200) to bound client memory growth.
Shorten WS heartbeat from 30s to 15s for aggressive proxy/NAT
environments. Clear all user-scoped stores on logout to prevent
cross-session data leaks.
Extract LiveKit internal accessors into shared livekitInternals utility.
Replace rigid SCREEN_QUALITY_MAP (6 hardcoded VideoPreset strings) with a
builder function that computes bitrate, degradation preference, and content
hint from three independent axes (height, fps, content mode). Camera is
decoupled onto a fixed 720p30 preset so screen share changes no longer
affect camera quality. New ScreenShareSettingsPopover replaces the old
VideoQualityPopover with pill-style selectors. Store migrated to v5 with
backwards-compatible migration from videoQuality string.
Extract stats polling into useTrackStats hook with per-sender/receiver
stats via RTCRtpSender.getStats(), delta-based FPS (fixes 25fps overwrite
bug with camera+screenshare), and track-to-source matching via LiveKit
TrackPublication identity. ConnectionInfoPopover becomes a pure display
component with Network/Audio/Video sections. Remove soft-launch diagnostic
loop from useLiveKit.
DM uploads silently failed because the 5-point chain (types, frontend,
POST, GET, WS broadcast) was never wired for attachments. Added
buildDmMessageWithUser/getDmMessageWithUser helpers mirroring the server
channel pattern, and plumbed attachmentIds + replyToId through all DM
code paths.
Unlock group DMs by removing the 2-member assumption across the stack.
No schema migration needed — dm_members junction table already supports
N members. Includes dedup bug fix, add/leave member endpoints, late-join
call support, group-aware sidebar rendering, and AddDmMemberModal.
Replace dual voiceStates + activeCalls maps with a single VoiceRoom
system that tracks both server channels and DM calls uniformly.
Fixes four bugs:
- voice_status silently dropped for DM call participants
- DM calls not cleaned up on WebSocket disconnect
- DM call state missing from ready payload on reconnect
- No spatial tracking of DM call participants
VoiceControls was gated on server-only `currentVoiceChannelId`, now also
triggers on `activeDmCall`. Added VoiceChatPanel to DM call view matching
the server voice channel pattern.
Closure guards on all 15 handlers already isolate React state from stale
rooms — the sledgehammer removeAllListeners() was stripping the SDK's own
internal listeners (PCManager, SignalClient), causing "Tried to add a track
for a participant that's not present" errors on rapid channel switches.
LiveKit's Room.disconnect() tears down WebRTC but leaves .on() handlers
attached. Every connect() registers ~15 event handlers, which accumulate
on orphaned Room instances during rapid channel switches or HMR, causing
MaxListenersExceededWarning. Added destroyRoom() helper that calls
removeAllListeners() before disconnect() at all four teardown sites.
Merge connectDm() into connect() with isDm flag, eliminating ~145 lines
of duplicated LiveKit room setup. DM calls now inherit all event handlers
(SpeakingDetector cleanup, deafen broadcasts, metadata changes). Replace
monolithic DmCallView with shared VoiceGrid + VoiceControlBar components,
making DM calls group-DM-ready with full feature parity.
Camera and LIVE badges in the channel sidebar were only visible to users
who had joined the same LiveKit room. Widen the voice_status WS event
from {isMuted, isDeafened} to {isMuted, isDeafened, isCameraOn, isScreenSharing}
so all server members see camera/screenshare indicators without joining voice.
30s ping/pong keepalive prevents proxy/NAT from killing idle connections.
Reconnect sound now only plays if downtime exceeds 3s, suppressing phantom
audio from brief network blips.
AudioManager.setScreenShareActive(true) was firing before the browser's screen
picker, killing the mic stream. The picker suspends getUserMedia while its secure
overlay is open, trapping the mic in a dead state for 5-30s. Now the AEC rebuild
fires after the track is acquired — mic stays alive during the picker.
Root cause: degradationPreference 'maintain-resolution' forced FPS drops under CPU
pressure during fast-motion gaming. Changed to 'maintain-framerate' so the encoder
drops resolution temporarily instead of stuttering. Added contentHint='motion' to
optimize for temporal prediction (more P-frames, fewer I-frames).
Screen sharing was publishing at h360 (640x360) and never ramping to target
resolution due to stale closure in setTimeout, wrong initial quality anchor,
and 5 competing code paths with inconsistent bitrates.
- Create utils/screenShare.ts as single source of truth for all screen share ops
- Publish at target resolution from the start (not h360 → ramp)
- Read store at call time in timers (eliminates stale closure bug)
- Use maintain-resolution for screen content, maintain-framerate for camera
- Fix OS-level "Stop sharing" not resetting store or restoring AEC
- Enable dynacast for SFU quality signaling
- Reconcile QUALITY_MAP to canonical bitrates across all 7 files
LiveKit's server-side VAD has a high, non-configurable threshold that misses
conversational speech. Replace it with a SpeakingDetector singleton that uses
Web Audio AnalyserNodes to read actual RMS levels per participant (50ms poll,
0.008 threshold, 250ms hysteresis hold).
Eliminate double-state architecture (useState → useEffect bridge → store)
that lost speaking events due to React 18 batching. ActiveSpeakersChanged
now writes speakingParticipantIds directly to voiceStore; 200ms poll
catches missed SDK events. Each VoiceUser subscribes to its own identity
via fine-grained selector for minimal re-renders. Also adds connection
quality indicator and ConnectionInfoPopover.
Jannis Braun