Documents the relay TLS identity bug (cert regenerates on restart
because server_config() creates a new keypair every time, ignoring
the persisted Ed25519 seed) and the full federation design:
- YAML config with mutual peer trust (url + fingerprint)
- QUIC connections between peers, fingerprint verification
- Room bridging: media forwarding for shared room names
- Merged participant presence across relays
- Helpful log message for unconfigured peer connection attempts
- No transcoding, no re-encryption, no central coordinator
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Covers the full design for runtime codec switching based on network
conditions: 3-tier basic (GOOD/DEGRADED/CATASTROPHIC), extended
5-tier with studio levels, and bandwidth probing. Details the
existing QualityAdapter infrastructure, what's missing (report
ingestion, profile switch loop, cross-task signaling via AtomicU8),
and implementation plan for both Android and desktop engines.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. Auto codec: new "Auto" position on quality slider (JNI index 7).
When selected, the engine uses the relay's chosen_profile from
CallAnswer instead of the local preference. Slider now has 8
positions: Studio 64k → Auto → Codec2 1.2k.
2. Force ping: added refresh button (↻) in Manage Relays dialog
header. Calls pingAllServers() to re-check all relays on demand.
3. Delete relay fix: the X button was inside a Surface(onClick=...)
which swallowed the touch event. Replaced with a separate Surface
that properly intercepts the click.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The CallOffer only advertised GOOD/DEGRADED/CATASTROPHIC. When a
client uses a studio profile, the relay's choose_profile couldn't
pick it. Now advertises all 6 profiles (studio 64k/48k/32k + good +
degraded + catastrophic) in both Android engine and shared handshake.
Also: the relay MUST be rebuilt with the new CodecId variants,
otherwise it will fail to deserialize CallOffer messages containing
studio QualityProfiles in supported_profiles.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. Wire protocol: add Opus 32k/48k/64k (CodecId 6/7/8) + STUDIO
profiles with is_opus() helper. Opus enc/dec accept all Opus variants.
2. JNI bridge: expand profile_from_int to 7 levels (0-6) mapping to
GOOD, DEGRADED, CATASTROPHIC, Codec2_3200, STUDIO_32K/48K/64K.
3. Settings UI: replace radio buttons with Material3 Slider — 7 stops
from Studio 64k (green) to Codec2 1.2k (dark red), matching desktop.
4. Key-change warning: AlertDialog on connect when server fingerprint
has changed. Shows old vs new fingerprint, Accept New Key or Cancel.
Accepting saves the new fingerprint and proceeds with the call.
5. Engine recv: handle studio codec IDs in auto-switch path.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
cargo-ndk doesn't always copy libc++_shared.so into jniLibs. The
build script now finds it in the NDK and copies it manually if
missing, preventing the build check from failing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The send loop was hardcoded to 960 samples (20ms/Opus24k), causing
DEGRADED (Opus 6k, 40ms) and CATASTROPHIC (Codec2 1200, 40ms) to
fail — the encoder needed 1920 samples but only got 960.
Changes:
- capture_buf, ring read threshold, and timestamp increment are now
computed from profile.frame_duration_ms (960 for 20ms, 1920 for 40ms)
- decode_buf sized to MAX_FRAME_SAMPLES (1920) to handle any incoming codec
- recv codec switch now uses correct QualityProfile per codec (was
inheriting original profile's frame_duration_ms, breaking cross-codec)
- added ComfortNoise guard on recv path
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Both Android and Linux build scripts now send ntfy notification
when build fails, not just on success.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Same Docker image as Android build. Separate cache dirs (cache-linux/)
to avoid conflicts when running both builds simultaneously.
Builds: wzp-relay, wzp-client, wzp-client-audio, wzp-web, wzp-bench
Uploads tar.gz to rustypaste, notifies ntfy.sh/wzp.
Usage:
./scripts/build-linux-docker.sh --pull # fire and forget
./scripts/build-linux-docker.sh --pull --install # wait + download
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Relay identity:
- Stored in ~/.wzp/relay-identity (hex-encoded 32-byte seed)
- Generated on first run, reused on restart
- Fingerprint stays consistent across relay restarts
Linux build script (scripts/build-linux-notify.sh):
- Fire and forget: Hetzner VM → build all binaries → upload to rustypaste → ntfy notify → destroy VM
- Builds: wzp-relay, wzp-client, wzp-client-audio, wzp-web, wzp-bench
- Packages as tar.gz, uploads to rustypaste
- --keep flag to preserve VM
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Relay recognizes SNI "ping" and returns immediately — no handshake,
no stream accept, no timeout error logs. Client closes after QUIC
connect for RTT measurement.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Ping was a static JNI method that loaded the .so before nativeInit,
crashing jemalloc. Now ping is an instance method on WzpEngine:
- Engine is created once (nativeInit), reused for both ping and call
- pingRelay() uses same tokio runtime pattern as startCall()
- Auto-pings all servers on app launch (after engine init)
- No process restart needed
- TOFU fingerprints saved on first successful ping
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Ping button: pings all servers via native QUIC, saves RTT + fingerprint
to SharedPreferences, then exits process (System.exit)
- On restart: loads saved ping results (no native .so loading needed)
- Avoids jemalloc crash: native lib only loaded once per process lifetime
- Removed broken UDP probe (QUIC servers don't respond to it)
- SettingsRepository: savePingRtt/loadPingRtt for cached results
- PingResult: added reachable field
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- AudioPipeline.debugRecording defaults to false (was true)
- SettingsRepository: persist debug_recording preference
- CallViewModel: debugRecording StateFlow + setter, wired to AudioPipeline
- Only records PCM + RMS when explicitly enabled in settings
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace WzpEngine.pingRelay() (JNI, loads native .so, crashes jemalloc
on Android 16 MTE) with pure Kotlin DatagramSocket UDP probe.
- RelayPinger: sends QUIC Version Negotiation trigger packet, measures
RTT from response. No native lib, no JNI, zero crash risk.
- Periodic: pings all servers every 5 seconds via coroutine
- Server fingerprint: filled lazily on first real QUIC connection
(TOFU still works, just delayed)
- Lock status: OFFLINE when ping fails, NEW until first connection
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Backport from desktop client to Android:
Identicons:
- New Identicon.kt composable: deterministic 5x5 symmetric Canvas pattern
from fingerprint hash (same algorithm as desktop identicon.ts)
- Participant list shows identicon + name + tappable fingerprint
- Settings page shows identicon next to fingerprint
CopyableFingerprint:
- Tap any fingerprint text to copy to clipboard with Toast feedback
- Used in participant list and settings page
Recent rooms:
- SettingsRepository: persists last 5 (relay, room) pairs
- CallViewModel: saves on startCall, exposes as StateFlow
- InCallScreen: clickable chips that fill room + select matching server
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Previous attempt allocated DirectByteBuffer as local variables inside
runCapture/runPlayout. ART's JIT On-Stack Replacement nulled them
when recompiling the hot loop mid-execution.
Fix: allocate as class fields on AudioPipeline (captureDirectBuf,
playoutDirectBuf). Object fields live on the heap, immune to OSR
stack frame replacement.
Eliminates JNI array copies (GetShortArrayRegion/SetShortArrayRegion)
from the audio hot path, preventing ART GC SIGBUS crashes on
Android 16 with concurrent mark-compact GC.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add paths-ignore for .gitea/** so build.yml doesn't waste runner time
when only workflow files are modified.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The jni crate emits VERBOSE logs for every JNI method lookup (~10 lines
per call, 100+ calls/sec on audio threads). This floods logcat, consumes
CPU, and triggers system kills. Filter to only show INFO+ for our crates
and WARN+ for everything else.
Also fix build script: clean full Rust target to ensure libc++_shared.so
is always copied by cargo-ndk.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
DirectByteBuffer.clear() crashes with null pointer in ART's JIT OSR
compiled code on Android 16. Revert AudioPipeline to use the original
ShortArray writeAudio/readAudio path.
The DirectByteBuffer JNI functions remain in WzpEngine.kt and
jni_bridge.rs for future use once the OSR issue is resolved.
The original SIGBUS from ART GC is rare (~1 crash per 8 min call)
and doesn't warrant the DirectByteBuffer approach until we can
allocate the buffer as a class field outside the hot loop.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Filter hcloud by SERVER_NAME to avoid touching other servers
- Use rsync instead of tar (handles submodules, no macOS xattr spam)
- Default server type cx33
- Release APK failure is non-fatal (debug APK still produced)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Automatically pushes branches and tags to github.com:manawenuz/wzp.git
on every push to Forgejo. Uses GH_SSH_KEY secret for authentication.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds nativeWriteAudioDirect / nativeReadAudioDirect JNI functions
that accept a DirectByteBuffer instead of ShortArray. The buffer's
native memory is accessed directly by Rust via pointer — no
GetShortArrayRegion / SetShortArrayRegion, no GC-managed array
copies on the audio hot path.
This fixes SIGBUS crashes on Android 16 where ART's concurrent
mark-compact GC crashes when flipping thread roots during JNI
array operations on MAX_PRIORITY audio threads.
Old ShortArray methods kept for backward compatibility.
AudioPipeline switched to use Direct variants.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Android 16's concurrent mark-compact GC crashes when flipping
thread roots on our MAX_PRIORITY audio threads during JNI calls
(AudioRecord.read / AudioTrack.write). Not our code — all crash
frames are in libart.so.
Proposed fixes:
- Short term: DirectByteBuffer to reduce JNI transitions
- Long term: Oboe native audio from Rust (no JNI, no GC)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
cmake 3.28 works when ANDROID_NDK is set (not just ANDROID_NDK_HOME).
Relaxed version check from <=3.26 to <=3.30.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Documents WHY each version is pinned:
- cmake 3.25: 3.27+ rewrote Android-Determine.cmake with bugs
- NDK 26.1: NDK 27 scudo crashes on MTE devices (Nothing A059)
- JDK 17: Gradle 8.5 + AGP 8.2.0 official support
- ANDROID_NDK: cmake checks this, not ANDROID_NDK_HOME
Idempotent, works from clone or existing tree.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- AudioRing: use vec![].into_boxed_slice() instead of Box::new([]) to
avoid 32KB stack allocation that crashes scudo on Android
- JNI bridge: wrap tracing_subscriber init in catch_unwind to survive
sharded_slab allocation failures on some devices
- Engine: per-step encode profiling (avg_agc_us, avg_opus_us, avg_fec_us,
avg_send_us) logged every 5 seconds in send stats
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds average microsecond timings for each encode step:
- avg_agc_us: AGC processing
- avg_opus_us: Opus encoding
- avg_fec_us: FEC encode + repair generation
- avg_send_us: QUIC send_media
- avg_total_us: sum of above
Logged every 5 seconds in send stats. Resets each interval.
Use to identify which step is bottlenecking the encode loop
on devices where fps drops below 50.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
SIGSEGV on hangup: capture thread calls writeAudio() via JNI after
teardown() has freed the native engine handle. TOCTOU race between
the nativeHandle==0L check and destroy() on the ViewModel thread.
Fix: CountDownLatch(2) — audio threads count down after exiting loops,
teardown() awaits before destroy(). 2 Kotlin files, no Rust changes.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Complete spec for fixing the playout ring buffer cursor race that
causes 12-16s bidirectional silence mid-call. Includes exact code,
memory ordering rationale, unit tests, and verification steps.
Any agent can implement from this document alone.
See also: debug/INCIDENT-2026-04-06-playout-ring-desync.md
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Rust tracing subscriber was never initialized — all info!/warn!/error!
calls in the engine went to /dev/null. This meant our send/recv health
logging was invisible and we couldn't confirm the congestion fix was
active.
Now initializes tracing-android layer on first nativeInit(), routing
all Rust logs to logcat under tag "wzp_android". Also expanded logcat
filter in DebugReporter to capture engine-level log lines.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Root cause: send_media() returns Err(Blocked) when QUIC congestion
window is full. The send task treated ANY send error as fatal (break),
killing the entire call. Now send errors drop the packet and continue.
Also hardened recv task to survive transient errors and added health
logging (recv gap tracking, periodic stats) to both send and recv.
Relay: added comprehensive debug logging — recv gaps, lock contention,
forward latency, send errors — all per-participant with 5s stats.
Other changes:
- AEC toggle in Settings (persisted, applied on next call)
- Debug report: records call audio (WAV), RMS histogram (CSV), logcat,
stats. Emailed as zip via Android share intent after call ends.
- Replaced LinearProgressIndicator with Box (compose version compat)
- FileProvider for sharing debug zip attachments
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Mic mute: the send loop now zeros the capture buffer when muted instead
of relying on write_audio() to skip writes. Previously stale ring data
and AGC amplification of near-silence caused crackling artifacts.
AEC: attach Android's hardware AcousticEchoCanceler to the AudioRecord
session. Also attach NoiseSuppressor when available. Both are released
on capture stop.
Room UI: deduplicate participants by fingerprint so ghost entries from
stale relay state don't show duplicate names.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
UI: deduplicate room participants by fingerprint so ghost entries from
stale relay state don't show duplicates.
Engine: after select! ends, call close_now() + connection.closed() with
500ms timeout to wait for the relay to acknowledge the CONNECTION_CLOSE.
Previously the close frame was queued but the runtime died before quinn
could retransmit if the first packet was lost.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
shutdown_background() killed the tokio runtime before quinn could send the
CONNECTION_CLOSE frame on the wire, so the relay never knew the client left.
Now use shutdown_timeout(500ms) to give quinn time to flush the close frame,
matching the desktop client pattern (which uses 2s timeout).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
o.optString("alias", null) returns the string "null" when the JSON value
is JSON null. Use o.isNull() check first. Also handle empty fingerprint
edge case with "unknown" fallback.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
stop_call() now calls close_now() on the stored transport handle before
killing the tokio runtime. This sends a QUIC CONNECTION_CLOSE frame so
the relay's recv loop breaks immediately, triggering leave() + RoomUpdate
broadcast. Previously the runtime was killed first, so transport.close()
never ran and the relay kept stale participants until idle timeout.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Settings now uses draft state — changes only persist on explicit Save
- Back button discards unsaved changes
- Added applyServers() for batch server updates
- Added missing alias field to CallOffer in featherchat tests
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
