Native dependencies — state and what is left

Written 2026-08-14. Resume point for the work that moved the unmanaged dependencies onto vcpkg manifest mode and turned libvlc and webrtc into prebuilt downloads.

Where things are

vcpkg is pinned to c3867e714dd3a51c272826eea77267876517ed99 (tag 2026.03.18) in vcpkg.json. That commit is deliberate: it is the newest release tag that still carries zlib 1.3.1. From 2026.04.27 the zlib port renames its output to z.dll, which would break the tdjson export filter silently and the csproj copy loudly.

Setup is now: a full vcpkg clone, either beside the repository or the copy that ships with Visual Studio's vcpkg component. No VCPKG_ROOT needed, no vcpkg integrate install, no port edited by hand. Directory.Build.props finds it, Directory.Build.targets checks it is not older than the pin and prints the fix if it is.

dependency	how it arrives
ffmpeg 7.1.2	overlay port, Libraries/vcpkg-ports/ffmpeg, carrying the --enable-* flag set
libvlc 3.0.23	overlay port downloading libvlc-3.0.23-2 from UnigramDev/deps
webrtc m123	overlay port downloading webrtc-2026-08-19-1 from UnigramDev/deps
everything else	stock ports at the pinned baseline

UnigramDev/deps holds the build and packaging scripts, the webrtc patches, and the release archives. libvlc/ and webrtc/ each expose build.ps1 and pack.ps1. Its README documents publishing a new archive.

The forks: UnigramDev/vlc unigram-12.7.5 at 739b198e18, UnigramDev/webrtc-uwp m123 at 801b013618. Both are pushed and both releases name their commit. The VLC checkout that used to be a submodule is an ordinary clone at C:\Source\vlc.

Libraries/vlc and Libraries/webrtc are gone from this repository.

Verified

• Both native projects build against the ports, and a Release x64 msixbundle packages.
• The ports install, and the libvlc archive repacks byte-identically from the relocated checkout.
• The manifest resolves at the pinned commit with the overlay taking precedence.
• tdjson builds for arm64-uwp against vcpkg at the latest baseline (a separate question, asked because TDLib was considering the move).
• deps/webrtc/build.ps1 -SkipAcquire and pack.ps1, run 2026-08-19: all four configurations built clean under Visual Studio 18 and were published as webrtc-2026-08-19-1. The acquire half — fetch, sync, patch — is still unexercised.

Not verified

• A build that downloads from the releases. Every build so far resolved from archives seeded into C:\Source\vcpkg\downloads. Clearing those seven files and rebuilding is the test that a contributor can do this at all, and it takes minutes.
• The packaging wizard. The two experimental settings in 8aebeaa71 were only exercised from the command line.

Next

1. Commit and push the tgcalls changes. UwpScreenCapturer.cpp carries the try/catch from the fork review and the libyuv include fix. The second one is load-bearing: removing the hardcoded webrtc include path is what made a plain <libyuv.h> ambiguous. Until it is on the fork the submodule pin cannot move and a fresh clone will not compile Telegram.Native.Calls. This is the only thing blocking someone else from building.
2. The download test above.
3. Build.ps1 is broken twice over: it calls msbuild Telegram.sln, which was renamed to .slnx, and passes PackageCertificateThumbprint=60FFAEE6... for a certificate that expired and was never renewed. Everything is signed with Telegram.Msix_TemporaryKey.pfx now, so the thumbprint branch can go.
4. Decide on the experimental settings. The control run for the double build is: revert ShouldUnsetParentConfigurationAndPlatform in the wapproj, rebuild, and look for two Telegram -> ...Telegram.exe lines and two Generating native code passes instead of one. release.binlog in the repository root is the "after" half.
5. Libraries/tdjson still has no answer for the confidential prerelease drops. The design was $(TdjsonDir) pointing outside the repository, with the drops never inside the working tree. td_api.bak.tl is currently untracked and unignored.
6. The TDLib ABI patch to td_json_client exists only in the working tree of the submodule. Same class of problem as the wasapi change and the unpushed webrtc branch, both fixed today.

Traps worth not rediscovering

• vcpkg reads its version database from the checked-out working tree, not from the commit named by builtin-baseline. A checkout older than the pin fails with "no version database entry for <port> at <date>", which does not hint at the cause. git fetch is not enough; you must check out the pin and re-bootstrap.
• Autolink is off (VcpkgAutoLink=false). It put every .lib in the installed tree on every link line, including a 339 MB webrtc static library on projects with no reference into it. Each project names what it links in Directory.Build.targets; the lists came from what the built binaries actually import. Statics leave no import trace, so ZXing.lib was inferred.
• webrtc carries a patched libyuv whose ConvertToI420 takes three extra parameters. The webrtc port deliberately does not install include/libyuv.h; the compiled files include it by full path. See duplicated-libraries.md.
• applocal does not populate the app's own output folder, and does not run at all when the linker is skipped or a project is only queried for packaging outputs. The DLLs are declared explicitly instead, which is what let Libraries/vcpkg.patch be deleted.
• libvlc plugins cannot be flattened. Their relative paths are recorded in the generated plugins.dat, so they reach the package through Content items with Link metadata.
• *.patch must stay LF. git apply rejects a patch whose line endings were converted; this is why Libraries/vcpkg.patch had stopped applying. Pinned in .gitattributes in both repos.
• Telegram.Stub's native AOT needs vswhere on PATH — findvcvarsall.bat calls it bare, so a build from a plain shell fails where Visual Studio or a developer prompt succeeds.