Repository navigation
v1.2.0
ChadVis 1.2.0 — v1.2.0
⚠️ READ THIS BEFORE INSTALLING ON macOS
These builds are not code-signed and not notarised. The project
carries no Developer ID certificate, so macOS Gatekeeper will refuse to
open the .app on any machine other than the one that built it. Once:
- right-click the app and choose Open, or
- run
xattr -dr com.apple.quarantine ChadVis.app
Artifacts in this release
Generated from the files actually uploaded, so this list cannot
overstate what shipped:
| File | Size |
|---|---|
chadvis-projectm-qt-1.2.0-Darwin.dmg |
103M |
chadvis-projectm-qt-1.2.0-Linux.tar.gz |
19M |
chadvis-windows-x64.zip |
83M |
The macOS build is ChadVis.app; open the .dmg and drag it to
/Applications. It is unsigned, so expect the Gatekeeper prompt above.
Windows
Unzip chadvis-windows-x64.zip anywhere and run
chadvis-projectm-qt.exe. Qt and the vcpkg runtime DLLs are inside,
so nothing has to be installed first.
What was verified before packaging
On every platform that produced an artifact: a full CMake configure and
build, then the 11 headless ctest entries (unit_tests plus 10 standalone
suites). Three further entries — integration_tests,
integration_gl_tests, test_OffscreenRenderSpike — need a real
drawable, and integration_tests also constructs a real AudioEngine,
so they are run and reported but not allowed to block a release.
Built from commit 828f0bb1aa62607dee84c4b9d2f07a1c60af8087.
What changed since the last release
Added
-
The Suno endpoint schema now covers the whole live
/mesurface, with bodies instead of
guesses. A signed-in browser walk plus the owner's 507 MB DevTools HAR
(suno.com4.har) added twenty route constants tosrc/suno/SunoEndpoints.hpp— the library
tabs (/video/hooks/user_hooks,/video/hooks/me/liked/v2,/profiles/listen-history,
/clips/trashed_v2,playlist/meandproject/mequery shapes), the Albums/Playlists feed
(POST /unified/feedwithfeed_id:"library_albums"), the Studio family
(studio/list-projects,studio/create-project,studio/project/{id},
agent/conversation/{id},plugins/catalog,plugins/{id}/presets), the image-upload chain,
contextual prompt suggestions, andstatsig/experiment/forked-onboarding. Three doc claims
were wrong and are now corrected:gen/trashwas not "unprobed" (it answers 200 with
{trash, clip_ids}→{ids, is_trashed}),{"undo":true}un-trash was not "untested" (sent
live; the scratch playlist round-tripped), and the upload failure branch is real —
initialize-clip/answeredrights_clearance_available:falsefor an ordinary upload. Speech
and Sounds turned out to share/generate/v2-web/(metadata.is_speech:true/
task:"sound"), and the sample trim has no standalone save route — its range rides the
generation body. Records:ENDPOINT-INVENTORY.md§4.12,UI-SURFACE-INVENTORY.md,
OBSERVED-LEADS.mdpromotions, andTODO.md's quick-reference. -
Suno MCP client exists.
SunoMcpClientspeaks JSON-RPC 2.0 athttps://api.suno.com/mcpwithinitialize -> notifications/initialized -> tools/list -> tools/call, plain JSON and SSE framing, and fail-closed protocol errors. (T0357) -
Suno library mutation verbs for reaction, playlist create, and legacy playlist trash.
SunoLibraryMutationsaddssetReaction,createPlaylist, andtrashPlaylistwith a 2xx-means-accepted discipline; a 204 trash parses nothing. (T0348) -
Upload consent flags are modelled.
ClipMetadataUpdateandAudioDescriptionConsentcarryisAudioUploadTosAccepted,geminiDescriptionAccepted, andaudioContentTypesas typed fields with body builders. (T0363) -
Clerk bearer tokens are bound to the session that minted them. A mismatched
sidnow fails closed instead of using a bearer from another session, andGET /v1/clientfalls back to the sole session when no selector is present. (T0340, T0342) -
Preset star ratings are keyed by path, not by preset name. Same-named presets in different banks keep separate stars, and reads/writes are mutex-guarded. Existing name-keyed ratings migrate to path keys on first non-empty preset list without overwriting newer path ratings. (T0058)
-
The documented test build now links reliably. Paired
spdlog/fmt/tomlplusplustarget include directories are prepended so headers and archives come from the same install. (T0364)
Changed
-
The file-wide
-Wdeprecated-declarationssuppression is gone from the recorder. It wrapped
every line ofVideoRecorderFFmpeg.cppfor the FFmpeg 6+ branch, which meant the next real
deprecation on an upgrade would arrive silently — the same class of trap the suppression was
originally added for. Compiling without it on libavutil 61.1.102 produced no warnings, and the
14 fast-profile ctest entries stay green, so it was hiding nothing today. -
Every long job now has a timeout, and the agent and release workflows are serialised.
build.yml,asan.ymlandtsan.ymlwould otherwise run to GitHub's six-hour default after a
stuck step; all four now stop at 120 minutes, matchingwindows.yml.opencode.ymlgains
timeout-minutes: 20and a per-ref concurrency group — it is started by a comment, so its cost
was unbounded by intent — andrelease.ymlgains one so two dispatches cannot race the same tag.
The harness layout is documented where it lives: rootCLAUDE.mdimportsAGENTS.mdand
.claude/skillssymlinks the canonical.opencode/skills, so Claude Code sees the rules;
SKILLS.mdgains the Kilo row, and the stale.agent/exploration prompt moved to
docs/dev/CODEBASE-EXPLORATION-PROMPT.mdwith its output routed toTODO.md, not the root. -
Lyrics are fetched when a track is opened, not for the whole library after every sync.
libraryUpdatedqueued one aligned-lyrics request per clip as each page arrived, so a 3,749-clip
first sync (measured 2026-10-26, 38 pages) left ~3,749 requests draining through the client's
polite 1 req/s limiter for most of an hour with the app otherwise idle — traffic nobody asked
for, for tracks nobody was going to play. Nothing was broken by it; it was simply a sync where
a fetch belonged. Lyrics now load onopenClipLyrics, the moment a clip is opened for playback,
and land in the SQLite cache, so a second play costs nothing and an offline replay still works. -
projectM is 4.1.8, not 4.1.6. Two releases of upstream fixes, none of which this project
could patch around from outside: an out-of-bounds read in custom waveform code, anullptr
dereference in the texture sampler, an infinite loop on self-referential shader macros, four
HLSLParser correctness bugs, and 4.1.7's "reset to default framebuffer (FBO 0) after rendering
presets and transitions" — the framebuffer this renderer deliberately draws into. The public API
delta is byte-identical (core.h,parameters.h,render_opengl.h), which the build confirms:
the same sources compile and link against the new tag with no changes. The pin is a commit SHA,
as before. (T0164) -
The Qt floor is 6.8 LTS, not the EOL 6.7. The old floor bought exactly one API —
QNetworkRequest::setTransferTimeout, whichHttpPolicystamps onto every request — and that reason stands; what changed is that 6.7 reached end of support on 2025-03-26, so a floor nobody can get security fixes for ships known bugs to whoever follows the README. 6.8 is the current LTS and also carries the backported screen-reader fixes. The CI gate has been building against 6.8.3 all along, which is what makes the floor honest rather than aspirational, andHttpPolicy.hpp's matchingstatic_assertmoved with it so an older Qt fails at compile time instead of at a missing call. (T0050) -
The first library sync asks for 100 clips a page instead of 20. The feed honours
limitup to 100 and answers400above it (measured with this app's own bearer: 20/50/100 return exactly that many clips, 200 is rejected; omitting the field yields the wire's default of 20). A 2 147-clip library was 108 round trips through the client's 1 Hz politeness limiter — about six minutes on every launch, with a spinner as the only feedback — and is now 22.has_morestill decides when to stop, so a server that caps the page lower degrades to more pages rather than to a wrong answer, andtheFirstPageRequestsAHundredClipsasserts the request body because nothing else would notice the constant reverting. (T0334) -
Qt now comes from qt.io, and the workaround for the broken Homebrew umbrella is gone. The umbrella was half-upgraded (6.11.1 while qtbase/qtdeclarative/qtmultimedia/qttools were 6.12.0), so its cmake files described tools that had moved;
scripts/qt-stitch.shpapered over that with a prefix of symlinks, and nothing stopped the stitch outliving its reason. It did:build-fastwas still configured against it long after itsQtGuiheaders went missing, and every compile failed onQtGui/qtguiglobal.hwhile the message named the wrong problem. The tree now takes Qt 6.12.0 from$HOME/Qt/<version>/macos(aqt, wired throughCMAKE_PREFIX_PATHin the shell), the stitch script is retired to.backup_graveyard/, andbuild.shresolves in a stated order — environment, newest$HOME/Qt/*/macos, MacPorts, Homebrew, distro, and.qt-stitchlast so a leftover directory cannot shadow a real install. A candidate must actually carrylib/cmake/Qt6/Qt6Config.cmake; a directory that merely exists is the failure being probed for. Two things broke silently when the install changed, and both were invisible with Homebrew's Qt: the frameworks carry@rpath/...install names where Homebrew's carried absolute ones, and this tree links everything withCMAKE_BUILD_WITH_INSTALL_RPATH ON— so test binaries, which no deploy step ever touches, had noLC_RPATHat all and died with "Library not loaded: @rpath/QtTest.framework/... no LC_RPATH's found", while the fast-profile.apppointed its only rpath at aContents/Frameworksthat deployment suppression never created. Tests now carry the Qt prefix'slibin their rpath, and so does the build-tree bundle in the fast-iteration profile only — a shipped bundle keeps@executable_path/../Frameworksand no developer path. (T0196) -
A build directory that outlived its Qt now repairs itself instead of failing forever.
build.shalready reconfigured whencmakeor the macOS SDK moved under a cached build directory; Qt was the third thing the cache remembers and the one nothing checked.qt_is_stale()reports and reconfigures when the cached prefix no longer carriesQt6Config.cmake, or when an explicit--qtdisagrees with the cache. Without it the symptom is a configure error that reads like Qt is not installed — "Could not find a package configuration file provided by Qt6" — on a machine where it is. (T0196) -
The repository now has the surfaces GitHub actually reads, and CI stops running its gates one test at a time.
SECURITY.md(Advisories-only private channel, secret taxonomy, 7/30-day expectations),CODEOWNERS, issue forms for bugs and features, a PR template that carries this repo's verification bar, and an.editorconfigthat matches.clang-format— none of which existed, so a security report had no private channel and a first-time contributor got a blank box. The bug form had two plain YAML scalars containing:and a leading backtick, which GitHub rejects outright; they are folded scalars now, and the forms are parsed in verification rather than assumed. Every gating and sanitizerctestrun gained runner-sized--parallel(nproc/sysctl -n hw.ncpu/$env:NUMBER_OF_PROCESSORS), while the three runtime-environment GL/audio entries stay serial on purpose: they arecontinue-on-error, their signal is already fragile, and trading evidence for seconds there is a bad trade. (T0302, T0095) -
The backlog now carries only open work. 81 items that had reached
[x]were re-verified
against the code rather than against their ownVERIFYlines, and their lines are gone:
T0010, T0012–T0047, T0074–T0075, T0095, T0103, T0153–T0155, T0160–T0163, T0166–T0168, T0172,
T0178–T0179, T0183–T0187, T0189–T0192, T0194–T0195, T0198, T0200, T0205–T0206, T0215–T0216,
T0297–T0298, T0302–T0303, T0306, T0308, T0310, T0319, T0324–T0325. Two of them needed a human
rather than a grep.T0166's fix (8750494) belongs to a loopback OAuth lane deleted wholesale
incc36dc0("Clerk never sends a loopback"), so it is reachable only from history and the item
went with the lane.T0191's ratio half was measured byscripts/check-theme-contrast.py
(16 gated pairs failed before the palettes were fixed) and its look half was answered by the only
authority that can answer it — the user, on the built app, looking at all four themes. The one
thing that look turned up is filed asT0337: ChadVis Dark'sbackgroundis pure black and reads
too dark.STATUS/<id>.logkeeps each item's evidence,git log -p TODO.mdis the ID archive,
and the fixes are the commits that named them. -
CI stopped lying to CMake and to macOS about its build configuration. The Windows lane
never named a generator, so CMake chose Visual Studio 2022 — whereCMAKE_BUILD_TYPEis inert —
and the same value was repeated as--config/-Cat every build and test call. It now
configures-G Ninja, a real single-config build atRelease, and the workarounds are gone.
The macOS gate now passes-DCHADVIS_FAST_ITERATION=ONalongside its existing Debug type, which
is the pair the suppression tests: every push was paying a measured 5–7 minutes of
macdeployqt+ codesign per app relink for a bundle no test loads.OpenGLis now a declared
Qt component (two test targets link it directly) andDBusis requested only on Linux, the one
The full history is CHANGELOG.md.