HUD 0.2.0's packages from its registries in place of vendor/hud
MergedThis pull replaces vendor/hud with HUD 0.2.0's published packages (npm for next/, a Swift registry through the local ios/HUD package, and Maven for android/). I found nothing that must change.
I read every non-vendor file in the diff. The 390 vendor/** files are pure deletions, and I did not read them. I did not build any of this myself.
next/:.npmrcsits next to the lock file and Dockerfile, and the Dockerfile copies it beforenpm ci.- The lock pins 0.2.0 with integrity hashes.
- The removed aliases, tsconfig paths and
fs.allowmatch what the packages now provide, andesm-env,leafletandprotomaps-leafletcome in transitively.
- CI and deploy:
next.ymlderives the tag from the app file'shudfield and caches by tag, in a newVERSIONfile in/opt/hud.- The fetch is of
refs/tags/$tag. The first run on this pull rebuildshud, as the body says. - The
vendor/**path filters and thegit archiveand image context lines are consistent with each other.
ios/:postGenCommandruns after generation and copies the registry file and the pins into the workspace, which is what Xcode Cloud needs.android/:- The Maven registry is limited to the
place.hudgroup. - The app's catalog is trimmed to the entries the app uses.
- The MapLibre substitution is dropped on the strength of hudkit's published metadata. Your release-classpath check backs that up.
- The Maven registry is limited to the
just hud-version: the sed patterns touch each of the four files and leave the tree unchanged at 0.2.0, as you describe.
The body already lists the unsigned-archive warning, the checksum failure on a machine with a stale fingerprint, and the deploy ordering with hetzner-infra. The next, next-image and android checks are the remaining verification, and they run on the pull.
I read what changed since my approval at da3a7e4. It is cleanup that follows from vendor/ being gone. It drops vendor from .dockerignore and from the vite.config.ts lint and fmt ignore lists. It also reformats ios/HUD/.swiftpm/configuration/registries.json, with the same registry URL and the same content. None of this breaks what I approved.
7 passed, 1 skipped
- HUD 0.2.0's packages from its registries in place of vendor/hud
- main
What changes
Grain stops copying HUD's source into
vendor/hudand takes HUD 0.2.0's published packages from HUD's registries on knotless (knotless://chadtmiller.com/hud, itsdocs/sdk.md). Reading them needs no login.vendor/hud/,vendor/hud.shandjust vendor-hudare gone, so a pull no longer conflicts onvendor/hud/COMMIT.next/:
@hud/renderer,@hud/shelland@hud/shared^0.2.0 are devDependencies, from@hud:registry=…innext/.npmrc(the lock pins 0.2.0 at knotless.dev). Thevendor/hud/webaliases, tsconfig paths, dedupes andfs.allowgo, and so doleaflet,protomaps-leaflet,esm-envand@types/leaflet, which were listed only for the renderer's source and now come with@hud/renderer. Every import stays as it was (@hud/shell,@hud/shell/server,@hud/shared/result, the two stylesheets).ios/: XcodeGen can't name a registry package, so
ios/HUDis a small local package that depends on.package(id: "hud.HUDKit", exact: "0.2.0"), andproject.ymltakes it by path. Xcode reads registries only from the project's workspace, not from a local package's.swiftpm, soproject.yml'spostGenCommandcopiesios/HUD/.swiftpm/configuration/registries.jsonand the committed pins,ios/Package.resolved, intoGrain.xcodeproj/project.xcworkspace/xcshareddata/swiftpm/. Xcode Cloud resolves only from committed pins, which is why the pins are committed, as grain-ios does.ci_post_clone.shgets this throughxcodegen generate.android/:
place.hud:hudkit:0.2.0comes from HUD's Maven registry (settings.gradle.kts,includeGroup("place.hud")). The app has its owngradle/libs.versions.toml(HUD 0.2.0's versions for what the app itself names, plushud = "0.2.0"). The:hudkitinclude, HUD's catalog,-Phudkit,hudVersionand the root MapLibre substitution go: hudkit's Gradle metadata already excludesorg.maplibre.gl:android-sdkand bringsandroid-sdk-opengl, and the release runtime classpath has no Vulkanandroid-sdk.CI and deploy: the
nextcheck buildshudforhud checkfrom HUD's public repo at the tag of the app file's version ("hud": ">=0.2.0"→refs/tags/v0.2.0), cached in/opt/hudkeyed by that tag.next-image,next-deployanddeploy/next-deploy.shsend and buildnext/alone, andnext/Dockerfilecopiesnext/.npmrcbefore thenpm ciit already ran, which now fetches HUD's packages from knotless.dev.vendor/**drops out of the workflows' paths.Version bumps:
just hud-version <x.y.z>moves the three places together, next's packages (and lock),ios/HUD/Package.swiftwithios/Package.resolved, and android's cataloghud, plus the app file'shud, which is what thenextcheck builds. AGENTS.md and the iOS and Android READMEs say this.Why
Every pull that moved
vendor/hudconflicted onvendor/hud/COMMIT, and the copy was 390 files of someone else's source. HUD now publishes one version of all three kits, so Grain takes a version as it takes any other dependency.How it was checked
v0.2.0(51ffab3) have no difference underweb/renderer,web/shell,web/shared,ios/HUDKit,android/hudkitor HUD's catalog (git diff --stat). The registry's HUDKit 0.2.0 archive matchesvendor/hud/ios/HUDKitfile for file.npm ci,npm run check(0 errors, 0 warnings, 334 files),npm run build. I ran the build before and after withnode build, both againstHUD_INTERNAL_URL=https://next.grain.social(read-only, signed out), and compared the server-rendered HTML of/,/profile/manton.org/gallery/3n2w7bb54mtpkand/profile/did:plc:ypx5shor4ojgky27ncur6aei/story/3n2w7obmiid2o. Once Svelte's class hashes and the hashed asset names are normalized, the only differences are the build version, SvelteKit's__sveltekit_<hash>global and two<svelte:head>hydration markers, all build-derived. The client output is 1556 KB before and 1560 KB after, with 30 chunks each.just ios, thenxcodebuild -scheme Grain -destination 'generic/platform=iOS Simulator' build(BUILD SUCCEEDED,hud.HUDKit @ 0.2.0resolved from the registry), and a Release build forgeneric/platform=iOSwithCODE_SIGNING_ALLOWED=NO(BUILD SUCCEEDED). I checked a fresh fetch with an empty cache: it warns only that the archive is unsigned. There is nojust ios-release../gradlew :app:assembleDebug :app:assembleRelease(both APKs built, the release one unsigned locally).just hud-version 0.2.0leaves the tree as it is. Its patterns move each of the four files when given another version. A version not on the registry fails atnpm install, before anything else changes.next(the tag fetch and the firsthudbuild at v0.2.0),next-image(the kaniko build, with npm ci against knotless.dev),android.Shortcomings
hud.HUDKit0.2.0 from another registry fails on the checksum. This one had a fingerprint left by a local test registry (http://127.0.0.1:8931); I moved it to/tmp/spm-stale-backup.publishToMavenLocal, or a local Swift registry) or a version bump.-Phudkit=<checkout>is gone.nextcheck's first run after this rebuildshud(its cache was keyed by commit).grain-next-deployalready builds whatever tar it gets, sonext-deployworks with this as it stands. That pull fixesup.sh(by hand), which archivedvendor/hudand would fail on a grain without it, and it updates the notes.