Files
nucleic/BUILD.md
T

21 KiB

Build & deploy strategy

Nucleic ships on five channels, each fed by a dedicated branch. Code flows in one direction only — feature → dev → canary → staging → rc → main — so nothing reaches a more stable channel without passing through the one below it.

Branches

Branch Channel Merges in only from Purpose
dev dev feature branches Day-to-day integration. Default target for all feature work.
canary canary dev Bleeding-edge distributed builds for early adopters.
staging beta canary Stabilized candidates for beta.
rc rc staging Release candidates — final validation before shipping.
main stable rc Shipping / production.

Rules:

  • Always merge feature work into dev — never directly into a more stable branch.
  • canary only ever merges from dev.
  • staging only ever merges from canary.
  • rc only ever merges from staging.
  • main only ever merges from rc.

A change therefore reaches production as: feature → dev → canary → staging → rc → main.

Channels & builds

The channel is chosen at build time by NUCLEIC_CHANNEL, which Package.swift reads to set both the executable's product name (and hence the OS process name shown in Activity Monitor / ps) and a compile-time define the app reads for its build identity + warning banner.

Channel NUCLEIC_CHANNEL Process name Banner Build from Command
dev dev (default) nucleic-local red "Local Build" dev make dev / make run
canary canary nucleic-canary yellow "Canary" canary make canary
beta beta nucleic-beta blue "Beta" staging make beta
rc rc nucleic-rc gold "Release Candidate" rc make rc
stable stable nucleic none main make stable

The banner also shows the commit the binary was built from (embedded at build time by the EmbedGitCommit prebuild plugin), so even a beta/rc/stable build reports its exact source revision.

Building

make run      # build + launch the dev channel (nucleic-local)
make dev      # build the dev channel
make canary   # build the canary channel, release-optimized (nucleic-canary)
make beta     # build the beta channel, release-optimized (nucleic-beta)
make rc       # build the release-candidate channel, release-optimized (nucleic-rc)
make stable   # build the stable/prod channel, release-optimized (nucleic)
make test     # run the test suite

Or directly:

NUCLEIC_CHANNEL=rc swift build -c release --product nucleic-rc

These builds use SwiftPM's default (SwiftBuild) backend. This repo used to force --build-system native to work around SwiftBuild's in-build codesign rejecting the com.apple.provenance xattrs iCloud stamped on files; the repo no longer lives in iCloud, so that workaround (and the now-deprecated native flag) is no longer needed.

Cross-environment builds — keep VM artifacts off virtiofs

The macOS VM mounts the repository at the same absolute /Users/… path as the host, eliminating the old mixed-/Users versus /Volumes/My Shared Files build-cache poisoning. Build products still should not live in the shared checkout: virtiofs is slower than the guest disk and does not support every clonefile/framework-symlink operation used by SwiftPM binary-target extraction.

scripts/lib/build-scratch.sh is the single source of truth:

  • Host/container builds use the default in-repository .build.
  • macOS VM builds are detected by the enclosing virtiofs mount and use $HOME/.nucleic-scratch/<hash> on the guest-local disk.
  • NUCLEIC_SCRATCH_PATH remains the explicit override.

Before Swift targets build, scripts/lib/guard-swift-cache.sh fingerprints the active Swift executable/version, Xcode, macOS SDK, and package manifests. If that identity changed, it removes only the scratch directory's generated out tree. This prevents stale explicit-module scan records (commonly surfaced as missing _AtomicsShims or CNIO* modules after changing Swift snapshots or package graphs) while preserving repositories, dependency checkouts, and downloaded binary artifacts. App packaging also recognizes that exact scanner failure, performs the same targeted cleanup, and retries once in case SwiftBuild corrupts an otherwise unchanged cache.

The Makefile and macOS build/package scripts already consume this helper. Direct Swift invocations should pass $(bash scripts/lib/build-scratch.sh).

Tailscale (Tailnet transport) — optional

The Remote settings' Tailnet transport embeds a Tailscale (tsnet) node via TailscaleKit, which has no SwiftPM distribution — it's built from a pinned libtailscale commit into an untracked local binary artifact:

scripts/build-tailscalekit.sh   # needs Xcode + a Go toolchain (brew install go)

Output: third_party/TailscaleKit/TailscaleKit.xcframework (macOS + iOS + simulator slices, ~100 MB, gitignored). Package.swift links it only when it exists; without it everything still builds and the transport picker reports Tailscale support as not built in. make dev / make run therefore do not generate this large optional artifact on a fresh checkout; run make tailscalekit once when developing the Tailnet transport. Shipped/package targets require it and embed the framework in the .app; the iOS app links it through the NucleicTailnet package product.

One sharp edge: SwiftPM caches the evaluated manifest by content, so Package.swift's artifact-exists check is not re-run when the xcframework appears or disappears with no manifest change. The build script clears the cache itself, but if you add or remove the artifact any other way (git clean -fdx, deleting it to reclaim space), run rm -rf ~/Library/Caches/org.swift.swiftpm/manifests before the next swift build (in Xcode: File ▸ Packages ▸ Reset Package Caches).

Running the tests

Use make test. It wraps swift test in scripts/lib/run-tests.sh, which supplies three things the bare command does not:

  • A deadline (NUCLEIC_TEST_TIMEOUT, default 1200s; 0 disables). swift test has no timeout, and a .serialized suite blocks every test queued behind a wedged one.
  • A heartbeat while the run is quiet, so slow and stuck are distinguishable.
  • A post-mortem when the deadline trips: sample stacks for the test helper, plus the names of tests that started and never finished — normally exactly the one that wedged.

It also kills the process tree. swiftpm-testing-helper is a child of swift-package; killing the parent orphans it, and an orphan keeps running.

Prefer it over swift test piped into a filter. This idiom in particular —

swift test 2>&1 | grep -vE "^objc\[" | grep -E "Test run|✘|error:" | tail -30

— is how a wedged suite went unnoticed for nine hours on the shared host on 2026-08-03 (the underlying bug: an unbounded Task.yield() spin in AppStore.awaitOpenSessionSettled, reached through a lost subscribe race in AppStore.observe; both fixed). tail prints nothing until EOF and the filter drops all progress, so a healthy run and a hung one are the same blank screen, indefinitely. make test stays safe under that same pipeline.

Two more things worth knowing:

  • Don't pass --build-path .build-host. It forces a full cold rebuild into a second tree instead of reusing .build. make test goes through scripts/lib/build-scratch.sh, which already picks the right scratch directory per environment (see Cross-environment builds).
  • Some suites are environment-coupled. Tests that create a git repo inherit your global ~/.gitconfig; with commit.gpgsign=true and an SSH/1Password signing key configured, the fixtures' git commit fails against a Nucleic-provisioned key path. That is host state, not a code regression — worth checking before chasing a wave of git commit failures.

Packaging .app bundles

The channels build as bare SwiftPM executables; scripts/package-app.sh wraps one in a proper macOS .app (Info.plist, the SwiftPM resource bundles inside Contents/Resources, ad-hoc code-signed so it launches locally). Output lands in dist/ (gitignored). The channels produce distinct, coexisting apps:

Channel Make target App bundle Bundle ID
dev make app-dev Nucleic Dev.app xyz.blakeslee.nucleic.desktop.dev
canary make app-canary Nucleic Canary.app xyz.blakeslee.nucleic.desktop.canary
beta make app-beta Nucleic Beta.app xyz.blakeslee.nucleic.desktop.beta
rc make app-rc Nucleic RC.app xyz.blakeslee.nucleic.desktop.rc
stable make app-stable Nucleic.app xyz.blakeslee.nucleic.desktop.release

make apps builds all five. The executable inside Contents/MacOS keeps the channel's product name, so the process name (nucleic / nucleic-rc / nucleic-beta / nucleic-canary / nucleic-local) is preserved when launched from the bundle.

Env overrides: NUCLEIC_VERSION sets CFBundleShortVersionString (default 0.1.0); CFBundleVersion is the commit count (git rev-list --count HEAD); NUCLEIC_SIGN_ID sets the codesign identity (default -, ad-hoc). Drop an icon at Resources/AppIcon.icon (themed Icon Composer package) or Resources/AppIcon.icns (legacy flat) to have it bundled. A channel can override the generic icon so its build is visually distinct — highest priority first: ./app-logo-<channel>.{icon,icns} (repo root) → Resources/AppIcon-<channel>.{icon,icns}. The canary build ships ./app-logo-canary.icon (yellow background, matching its banner).

Distribution & auto-update

The full distribution pipeline is wired — see signing/README.md for the one-time setup (certificates, App Store Connect API key, Sparkle key, App Store record).

  • macOS — direct (not App Store): make release-{canary,beta,rc,stable} builds → Developer-ID-signs the .app → notarizes + staples → builds + signs the DMG → notarizes + staples → regenerates the channel's Sparkle appcast, then publishes the DMG + dist/appcast-<channel>.xml to the Cloudflare R2 bucket served at updates.nucleic.blakeslee.xyz (scripts/upload-r2.sh; the GitHub release is a fallback origin). Each channel ships in-app Sparkle auto-update from its own appcast feed (baked in as SUFeedURL per channel); the dev/local channel embeds no feed, so its updater stays dormant.

    • Stable download link: https://updates.nucleic.blakeslee.xyz/<channel>/latest (e.g. /beta/latest) 302-redirects to the newest channel DMG — a fixed URL for the website/docs that always points at the current build. Served by the cloud/nucleic-updates/ Worker (deployed once; self-updating per release). See docs/CLOUD_INFRA.md §1.1.
  • iOS — App Store / TestFlight: two coexisting channels, each its own app, mirroring the Mac's separate beta/canary apps. make ios-release-beta uploads the public beta (bundle …nucleic-remote.beta, "Nucleic Beta", the ./app-logo-beta.icon icon, default minor bump); make ios-release-canary uploads the canary (bundle …nucleic-remote.canary, "Nucleic Canary", the ./app-logo-canary.icon icon, default build bump). Both share ios/VERSION and archive+upload to App Store Connect (→ TestFlight). The bare xyz.blakeslee.nucleic-remote / "Nucleic" identity (the Xcode project defaults) is reserved for the separate stable App Store release, so beta no longer shares stable's record. make ios-release is an alias for the beta channel.

    App Store / TestFlight reviewers have no Mac to pair with, so the app ships an in-app demo mode ("Explore a demo" on the first screen) that seeds sample data and simulates the agent locally — reviewers can exercise every feature without a host. Paste the copy from docs/APP_REVIEW_NOTES.md into App Store Connect ▸ App Review Information ▸ Notes for each submission.

(scripts/package-app.sh still signs ad-hoc by default — NUCLEIC_SIGN_ID=- — for local runs; the release targets auto-detect the Developer ID identity.)

Container runtime (sandboxing)

The session sandbox runs each agent inside a Linux VM built directly on Apple's containerization framework (the same framework Apple's container CLI is built on) — in-process, no external CLI or daemon. Consequences:

  • Platform floor: macOS 27 on Apple silicon. The app uses macOS 27-only features; the framework's in-process vmnet networking — which the approval / git-hook callback flow depends on — needs macOS 26+, comfortably below this floor. Package.swift sets the macOS deployment target to 27 and package-app.sh sets LSMinimumSystemVersion to 27. (containerization is a dependency of NucleicCore only, never of the iOS-linked NucleicProtocol, so the iPhone client's build is unaffected.)
  • Entitlements. Two, both in signing/nucleic.entitlements, applied at codesign time by package-app.sh. Neither is restricted/managed, so distribution needs no special Apple approval and no provisioning profile — a Developer ID cert + notarization is enough.
    • com.apple.security.virtualization — booting a VM requires it; honored by ad-hoc signing for local dev. (The restricted entitlement we deliberately avoid is com.apple.vm.networking.)
    • keychain-access-groups (L7UDTQ6F5W.xyz.blakeslee.nucleic) — lets Nucleic keep its own secrets in the data-protection Keychain and read them back silently (no authorization panel); see KeychainOwnedAccess. The team prefix is literal (codesign does no $(AppIdentifierPrefix) substitution). Ad-hoc local builds can't claim it and transparently fall back to the login Keychain.
  • Nothing is bundled — the kernel, vminitd, and sandbox image all download + cache automatically on first use, so the user runs no setup.
    • Kernel: published as a GHCR package — an OCI artifact carrying the single vmlinux-arm64 blob (ProjectSandbox.kernelImage) — by .github/workflows/kernel-image.yml. Like the sandbox image, the package may be public (anonymous pull) or stay private — independent of the repo's visibility, which is the point: it can be public while the repo stays private. The app downloads it on first use (≈16 MB) via the registry's distribution API, authenticating with the user's GitHub token (read:packages) only when the package is private. It first reuses a local kernel if present (NUCLEIC_KERNEL_PATH, a bundled Resources/vmlinux-arm64, or Apple container's installed kernel) — so machines with the CLI download nothing. To publish a new kernel: bump ProjectSandbox.kernelImage, then run the Publish sandbox kernel workflow (Actions tab) with the matching tag; on the first publish set the package's visibility (Public for anonymous pulls). scripts/fetch-kernel.sh + bundling remain optional (offline/dev fast-path only).
    • vminitd initfs: pulled from ghcr.io/apple/containerization/vminit (the tag in ContainerEngine.vminitReference MUST match the pinned framework version), materialized to a cached ext4. No cross-compile.
    • Sandbox image: built in CI from containers/nucleic-sandbox/Dockerfile and pushed to ghcr.io/<owner>/nucleic-sandbox by .github/workflows/sandbox-image.yml; the app pulls + unpacks it on first use and caches the rootfs. Keep the workflow's IMAGE_TAG in lockstep with ProjectSandbox.defaultImage. The GHCR package may stay private — the app authenticates the pull with the user's GitHub token (read:packages) via ContainerEngine.registryAuth (Settings → Sandbox → registry username, or NUCLEIC_REGISTRY_USER/NUCLEIC_REGISTRY_TOKEN); make it public only if you prefer anonymous pulls. Bring-your-own custom images are pulled the same way (public anonymously, private via the same credentials). Two runtime refresh paths avoid waiting for an app release when the CLIs need updating (e.g. a new Codex release unlocks new models): Settings → Control → "Check for updates" runs npm install -g …@latest for Codex + Claude Code in place inside the running container (instant, no re-pull; ContainerManager.updateAgentCLIs), and "Force re-creation" re-pulls the base image from the registry — evicting both the rootfs and image-store caches — so a tag re-pushed with a refreshed image is fetched fresh (recreateSharedremoveDefaultRootfs).
  • Ephemeral. VMs run in-process, so they're torn down when Nucleic quits and recreated on demand next launch (the per-container rootfs clone persists, so recreation skips the re-pull and re-seeding). Launch-time reconcile is on-disk GC, not orphan-VM reaping.

Pre-1.0 dependency: containerization is pinned to an exact commit in Package.swift (tag 0.34.0). When bumping it, also bump ContainerEngine.vminitReference to the matching vminit:<version> (host framework and guest vminitd must speak the same vsock protocol) and re-verify the API (all framework calls are centralized in ContainerEngine).

Verified end-to-end (Step-0 spike)

Sources/container-spike is a standalone proof (run signed with signing/spike.entitlements). On macOS 27 / Apple silicon it confirmed, in-process:

  • VM boot + exec + stdout streaming — PASS. A real Linux VM boots from the kernel + the runtime-pulled vminit:0.34.0 initfs; exec runs and its stdout streams back (the agent's NDJSON path). This is the same acquisition model the app uses (pull vminit + sandbox image).
  • vmnet networking — works. The guest gets eth0 + a default route via the gateway, and NAT egress works (guest reached the public internet) with only com.apple.security.virtualizationcom.apple.vm.networking is not needed and, being a restricted entitlement, gets an ad-hoc-signed binary killed at launch, so don't add it.
  • Guest→host callback — blocked by the macOS application firewall for the ad-hoc spike binary (the approval/git-hook HTTP hits the host listener over the gateway). This is not a rework regression: the old container-CLI path used the same guest→host vmnet gateway, so the real (Developer-ID-signed, firewall-allowed) Nucleic.app receives the callback the same way it did before. The firewall-immune fallback, if ever needed, is a host↔guest vsock relay (LinuxContainer.dialVsock) instead of IP routing.

Build + run the spike:

swift build --product container-spike
codesign --force --sign - --entitlements signing/spike.entitlements .build/*/container-spike  # path varies by build system
.build/.../container-spike

Releasing

  1. Land feature branches into dev and verify on the dev channel.
  2. When dev is stable, merge dev → canary and cut a canary (make canary) — bleeding-edge builds for early adopters.
  3. Promote a validated canary: merge canary → staging and cut a beta (make beta).
  4. Promote a validated beta: merge staging → rc and cut a release candidate (make rc) for final validation.
  5. Once the rc is signed off, merge rc → main and cut a stable build (make stable).

Keeping versions in sync across branches

Each channel is released independently and bumps its own ./VERSION, so a more-stable branch can end up shipping a newer version than a less-stable one that hasn't been released in a while. Because code only flows upward, a less-stable branch must never trail a more-stable one — e.g. if staging ships 0.3.1 while dev is still stuck on 0.2.1, dev should be fast-forwarded to 0.3.1 (its build number raised too, so CFBundleVersion stays monotonic). A branch that is already ahead (say dev at 0.3.2 vs staging 0.3.1) is left alone — versions are only ever raised, never lowered.

This is handled by scripts/sync-version.sh, which walks the chain from most- to least-stable and fast-forwards each branch to the highest marketing version above it:

make sync-versions              # dry run — show which branches are behind and what they'd become
make sync-versions APPLY=1      # write the VERSION-bump commits, then push the updated branches

Each behind-branch gets a single VERSION-only commit (non-checked-out branches are updated in place via git plumbing; the current branch, if behind, is committed normally and needs a clean tree). Nothing is pushed for you. make release-{canary,beta,rc,stable} runs this automatically after committing its own bump, so releasing a channel pulls the branches below it up to match; disable that with NUCLEIC_SYNC_LOWER=0.