6.1 KiB
Codebase audit — ts3-client
First audit: 2026-09-11. Revisited and worked through: 2026-09-25. Every finding below was either fixed (commit given) or closed with a reason. Logging (§3) was left out on purpose. The structural audit of 2026-10-11 is AUDIT-STRUCTURE.md.
State after the fixes: mvn -o test passes (177 tests: 25 in ts3j, 135 core, 11 desktop,
6 swing), on Java 25 and 26. The Android release build (R8 over core and ts3j) passes. The
Swing client was run against the test server on Java 25: it connected, loaded the tree and
chat logs, and wrote its profile owner-only.
1. Fixed
Tests and build
| Finding | Commit |
|---|---|
The importer test failed on every run after the first. The cause was not the real profile: the POM already points user.home at target/test-home, but that folder outlives a run, so the first run's identity was still there |
ba5de5a |
The ts3j DNS test resolved voice.teamspeak.com for real and failed when DNS did not answer |
ts3j Test SRV lookup…, acb3efa |
| README architecture and features out of date | 51f19c2 |
| Java 26 required for no reason; it now needs 25 (LTS), and builds, tests and runs on it | 85c3ae6 |
| Missing tests: channel tree, event handler, settings round-trip and migration | 71ac2c8 |
Connection lifecycle and threading
| Finding | Commit |
|---|---|
An action racing a disconnect showed NullPointerException as the error; 20 raw threads. Actions now capture the socket and run on a pool that lives as long as the connection |
f350b19 |
connect() on a live instance leaked socket and audio; fields shared across threads were not volatile |
d28bb57 |
| Permission errors were recognised by substring; now by the server's error id | d074ffa |
HotkeyEngine locked on another class's list; Hotkeys.all() now returns a snapshot |
5b1809c |
| Two quick tab switches could interleave the microphone hand-over | 665316d |
MainFrame was also the app controller: microphone ownership, push-to-talk, away and nickname fan-out moved to core/session/Sessions |
bf07f0b |
Model and events
| Finding | Commit |
|---|---|
buildTree() refilled lists shared with the UI; node fields not volatile. Lists are now published whole and unmodifiable. Whether a channel is full is counted from the model |
67f8f46 |
selfPermissionValue scanned every permission; primaryServerGroupName unused |
8c735a0 |
| A cleared avatar, description or away message was ignored | 7cb4f47 |
Voice
| Finding | Commit |
|---|---|
Fixed 60 ms jitter delay instead of TS3's adaptive buffer. The stashed libspeex port is now in VoiceStream, driven as TS3 drives it. A simulated 0–60 ms jitter stops concealing after about 3 s |
0083dc8 |
| A playback line kept open for every client that ever spoke | 6234ec6 |
| A dead playback line silently muted its speaker for the whole session | 12ca585 |
CaptureVoiceInput.stop() tore down while the capture loop was still running |
8e9c461 |
Chat logs
| Finding | Commit |
|---|---|
| Multi-line messages lost their continuation lines in history | f9b2f4d |
| History read whole, ever-growing logs; now the last 4 MiB | a5c4ba1 |
appendPrivateMessage duplicated appendMessage |
1a3a3e3 |
| Logs could not be turned off or moved (new Options → Chat page) | d7261c9 |
Chat logs only worked with admin rights: the server id came from serverinfo, which guests may not use. It turned out to be base64(SHA-1(the server's handshake key)), as the official client derives it. Found while checking the "serverUniqueId fallbacks can go" finding |
ts3j Derive the server's unique id…, eeccb49 |
Persistence, file transfers, UI
| Finding | Commit |
|---|---|
| Profile files (private keys, server passwords) readable by everyone, and truncated by a crash mid-write. Now written privately and atomically, and the profile folder is owner-only | 3cdade9 |
Three copies of the enum parser in Settings |
48a0032 |
| A stalled file server hung a download forever | ae8d1f7 |
| Collapsed channels reopened on every tree rebuild | f30cecf |
2. Closed without a change, or still open
-
Chat-log folders "may contain
/": they cannot. The value encoded is itself Base64 text, and no 6-bit group of such bytes reaches 62 or 63, the codes for+and/. -
rootMessage()hides the command: every caller names the action in front of it. -
Not tested:
TeamspeakConnection's lifecycle against a fake socket (the smoke test above covers connect only); the idle release and dead-line replacement inPerSpeakerPlayout; the jitter buffer with real voice from another client. -
Open (found 2026-10-14): an upload counts as completed before the server has the file.
FileTransferManager.runUploadmarks the transfer completed once the last bytes are flushed to the file socket; the server may still be receiving them. A listing taken right then shows a partial size (seen on the local 3.13.7 server: 241,664 of 300,000 bytes; a moment later the full size, and a download got every byte). The file browser's refresh after an upload can hit this. Fix: beforemarkCompleted, wait for the server to confirm, either- the
notifystatusfiletransferfor the transfer'sclientftfidwith the "transfer complete" status (believed to be 2065, not yet seen here);handleFileTransferEventalready parses it but only routes it topendingInitsduring negotiation, so keep a waiter registered until the upload ends; or - the file server closing the connection:
shutdownOutput(), then read to EOF;
whichever the server is shown to do first (watch what it sends after an upload). Bound the wait (a few seconds), then mark completed anyway, so a server that never confirms does not leave uploads stuck at 100 %. About 15–30 lines in core; both frontends follow. A check that lists the file from a second client right after the upload completes reproduces it.
- the
-
git stash@{0}("jitter buffer test") is now fully ported, except FEC decoding (TS3 does not use FEC either). It can be dropped.
3. Left out on purpose
- Logging. There are still 60 swallowed exceptions and no log output.