Repository navigation
ci: build Apple lanes with Xcode 27.1 - #1266
Merged
Merged
Conversation
This was referenced Oct 25, 2026
AnthonyRonning
force-pushed
the
research/ios-scene-lifecycle
branch
from
October 25, 2026 00:40
df76b74 to
7df88b2
Compare
AnthonyRonning
force-pushed
the
ci/xcode-27.1
branch
3 times, most recently
from
October 25, 2026 01:50
8cd7644 to
386571d
Compare
Contributor
Author
|
Heads-up for whoever picks this up: the Agent's macOS builds don't need to move to GitHub's Evidence from the experiment in #1308 (draft, based on this branch, not for merge), run 40415895366:
Suggested shape when this rebases:
The workflow is Not urgent: CI is being rewritten in a single PR first, and this PR should rebase after that lands. |
AnthonyRonning
force-pushed
the
research/ios-scene-lifecycle
branch
from
October 29, 2026 03:12
7df88b2 to
ddd929b
Compare
AnthonyRonning
force-pushed
the
ci/xcode-27.1
branch
from
October 29, 2026 03:12
386571d to
257b80d
Compare
AnthonyRonning
force-pushed
the
research/ios-scene-lifecycle
branch
from
October 29, 2026 18:31
ddd929b to
f15adfe
Compare
AnthonyRonning
force-pushed
the
ci/xcode-27.1
branch
from
October 29, 2026 18:31
14c9407 to
6856569
Compare
AnthonyRonning
force-pushed
the
research/ios-scene-lifecycle
branch
from
October 30, 2026 04:56
f15adfe to
c03e9ad
Compare
Blacksmith's macOS 27 image mounts 27.1 (27A9275) as Xcode_27.1_Release_Candidate.app, while GitHub's xcode-27 image installs it as Xcode_27.1.app. Consider Xcode_<version>_*.app installs too; selection still requires the exact pinned build and still skips betas. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
These are no-ops on Xcode 26.5 and needed before the Xcode 27.1 pin.
- Agent shell: link Rust with Xcode's clang as well. nixpkgs' ld64
cannot read the macOS 27 SDK's .tbd stubs (they list arm64e.x1), so
build scripts failed to link (_strerror_r undefined).
- Research and Agent: stop stripping build scripts and proc-macros in
release builds. rustc strips Apple outputs with rust-objcopy, which
doesn't pad __LINKEDIT entries, so chained-fixup images (deployment
target 12+) can end up with a 4-byte-aligned string table. macOS 27's
dyld refuses such dylibs once they're linked against the macOS 27 SDK,
so rustc could not load zerofrom-derive ("can't find crate").
Executables aren't checked, so shipped binaries keep their strip
settings. Upstream fix: llvm/llvm-project#203680.
- test-canonical-macho: Xcode 27's ld sorts local absolute symbols, so
both fixtures came out in the same order. Swap the entries to recreate
the input-order variant older linkers emitted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Xcode 27.1 (27A9275) adds the iOS 27.1 SDK and iPhone Duo support, and App Store Connect requires the iOS 27 SDK from April 2027. The macOS 26 images have no Xcode 27, so every job that selects the pin moves: - Blacksmith blacksmith-6vcpu-macos-26 -> blacksmith-6vcpu-macos-27: lanes.MACOS and ci.yml's special jobs in lockstep, plus the signed and release builders (agent-desktop-build, desktop-build, mobile-build, ios-dev-testflight, release, proxy-rust). - GitHub macos-26 -> xcode-27: artifact verifiers, signing, TestFlight and the nightly provider suite. - actionlint knows neither label yet. - iOS ONNX Runtime builds for iOS 16.0 (the app's minimum) instead of 13.0, because Xcode 27 only targets iOS 15.0-27.1. The hash pins move to 27A9275, from a local Xcode 27.1 build that reproduces them. - Docs, the outage advice and the StoreKit harness name the new versions; the stale project.pbxproj.backup goes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Blacksmith's macOS 27 image has no Metal Toolchain for Xcode 27.1 and cannot download it (the MobileAsset catalog fetch fails), while GitHub's xcode-27 image ships it. - A new export runs on xcode-27 and uploads the toolchain as a one-day artifact (.github/actions/export-metal-toolchain). The jobs that compile gpui's shaders import it (import-metal-toolchain): the signed macOS build in agent-desktop-build.yml and, in ci.yml, the macOS packaging rehearsal. - Lanes gain `requires`. The planner adds a required lane to every plan that selects the lane requiring it, dispatches included, so metal-toolchain runs exactly when agent-package-macos does. lint checks that requirements exist, form no cycle, select nothing on their own, and that the requiring job needs them. - agent-macos (PR Clippy) and agent-full-macos (the native suite) build gpui with runtime_shaders and need no Metal. Compiled shaders are proven by master's signed Prod build on every Agent push, the nightly Dev build and ci:package. Toolchain changes now also select agent-full-macos on master. - Cache keys include the Xcode pin, so an upgrade never restores objects built against the previous SDK: the macOS lanes' cache_files, the packaging rehearsal's key prefix, and the signed build's release key (cargo_graph_key.py takes optional pin files; C4 expects the pin for macos-build only, so the Windows and Linux keys are unchanged). - The policy tests know GitHub's per-Xcode labels are GitHub-hosted, so the attesting signing job may run on xcode-27. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The UIScene manifest (Info.ios.plist), the plugins' Swift sources and the vendored tao's iOS platform layer only compile into the iOS app, yet no path of research-ios covered them, so a change there ran no iOS build unless someone added ci:ios. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ci.yml sets CARGO_PROFILE_DEV_DEBUG=0, and Cargo strips build scripts and proc-macros whenever their profile has no debug info, so dev builds hit the same rust-objcopy alignment bug as release builds: research-dev-macos (tauri build --debug, deployment target 13.4) failed with "can't find crate for `zerofrom_derive`" on Xcode 27.1. Reproduced locally with CI's settings; the dev build-override fixes it. The Agent gets the same setting, because its dev lanes only pass today on rustc's default deployment target. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
agent-desktop-build.yml's macOS build now runs the export and import actions, but AGENT_DESKTOP did not list them: a master push that changed only them skipped the signed build, and no PR lane ran them without ci:package, so a broken edit could pass ci-ok and break the next Prod build. They join AGENT_DESKTOP and agent-package-macos's paths, a test derives every local action the signed builds run from the workflow, and the skill says when ci:package is the PR proof (Metal actions, gpui/zed bumps, Xcode pin changes). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Verify downloaded macOS was the only job that launches the signed Agent (verify-release.sh runs --version, the duplicate-runtime check and the GUI smoke test), and the Xcode 27 move put it on macOS 27, the same macOS the build links against. It goes back to macos-26, where most users run the app, and drops the pinned-Xcode step: it builds nothing and the image's default Xcode suffices. Research's two artifact verifiers never select Xcode, so they return to macos-26 as well instead of depending on the preview xcode-27 image. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The export uploads with overwrite, so "Re-run all jobs" can upload the same artifact name again in the run. - The import notes that the artifact lives one day: a later re-run needs the export job re-run too. It also captures `xcodebuild -showComponent` before matching, so grep -q cannot fail the pipeline under pipefail. - metal-toolchain is active only where agent-package-macos is (PR side and nightly), so a master fallback plan never runs the export alone. lint now requires a required lane's events to stay within its requirers', and rejects an OpenSecret selector on it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The policy tests that keep pull requests from saving caches or reading another run's artifacts walked only workflow steps, and the Metal Toolchain import is the first local composite action with an artifact step that runs in a PR job: a `run-id` added to its download, or a cache save inside it, passed every test. Both tests now walk each local composite action's steps in the job that uses it, downloads may not pass a `github-token` either, and a fixture test proves a violation inside a composite is seen. The GITHUB_HOSTED comment no longer contradicts its pattern. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
research-android did not select the vendored tao's Android layer, which carries Maple's null-intent hardening, nor the vendored plugins' Kotlin sources: research-rust never compiles cfg(target_os = "android") code, so an edit there ran no Android build without ci:android. They join its paths, like the iOS-only sources in research-ios, which gains tao's platform/ios.rs. The research-ios comment notes the iOS 27.1 Simulator download each run makes on Blacksmith's macOS 27 image. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…code-27 Comments and docs move with behaviour: - The CI design, the debugging guide, repository-workflows.md and the macOS builder's restore comment say the macOS release key also hashes scripts/ci/apple-toolchain.json, so an Xcode pin change makes the next macOS Prod build cold. - The Metal Toolchain artifact lives one day: the re-run guidance and the Agent build doc say to re-run the export job with a macOS build retried more than a day later. - repository-workflows.md and the CI runner table name the jobs that depend on GitHub's preview xcode-27 label, and what to switch to if it is renamed or retired. - changing-ci.md states the full rule for required lanes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Agent builds with Rust 1.98.0, whose stripped proc-macros keep an 8-byte-aligned string table with the macOS 27 SDK at deployment targets 13.4 and 15.0. Research's Rust 1.94.1 leaves the same dylibs 4-byte aligned, which is what macOS 27's dyld refuses. A release check of the whole Agent app at 15.0 passes on Xcode 27.1 without the override, so the Agent never needed it. It wasn't free either. The Agent's manifest is a virtual workspace manifest, which rust-cache does not hash, so the override changed no cache key. Linux and Windows Agent lanes restored master's cache as a full match and rebuilt everything downstream of a proc-macro: agent-test compiled 604 crates instead of 10, and agent-windows 196 instead of 3, taking 11 min. A full match also skips the save, so master would never have refreshed those caches. Research keeps both overrides. Its comment now says they can go once the root flake pins Rust 1.98 or newer. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fixture test calls rustc directly, so it linked with `cc`, which in the Agent's Nix shell is nixpkgs' clang wrapper. Its ld64 cannot read the macOS 27 SDK's .tbd stubs, so `just ci` failed on a Mac with Xcode 27.1 selected. The test now passes the host's CARGO_TARGET_<host>_LINKER, which the shell sets to Xcode's clang, as Cargo would. CI runs this suite only on Linux, where the test skips. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
AnthonyRonning
force-pushed
the
ci/xcode-27.1
branch
from
October 30, 2026 05:06
1d97989 to
0929f37
Compare
Contributor
|
Maple development preview: https://e7caa10f.maple-ca8.pages.dev Commit: Uses development API, billing, flags and PCR configuration. Cloudflare Access applies. |
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
Move every Apple job to Xcode 27.1 (27A9275), the release with the iOS 27.1 SDK and iPhone Duo support. App Store Connect requires the iOS 27 SDK from April 2027. Top of the stack: it needs PR #1264 (Tauri 2.12) and PR #1265 (UIScene adoption), because iOS 27 SDK builds abort at launch without scenes.
Rebuilt on the new
ci.ymland lane planner; the old per-workflow changes are gone. The CI review's fixes are the six commits from "ci: rebuild the signed Agent when the Metal Toolchain actions change" through the docs commit. The two after them drop the Agent's strip override and fix a local Agent test for the macOS 27 SDK.What
select-xcode.py): also accepts suffixed installs such asXcode_27.1_Release_Candidate.app, which is where Blacksmith's macOS 27 image mounts 27A9275 (GitHub'sxcode-27hasXcode_27.1.app). The exact build is still required and betas are still skipped..tbdstubs (arm64e.x1).strip = falsein its dev and releasebuild-override). Its Rust 1.94.1 strips them with arust-objcopythat can leave a dylib's string table 4-byte aligned once the deployment target is 12 or later (Research's is 13.4), and macOS 27's dyld refuses such dylibs when they're linked against its SDK. CI'sCARGO_PROFILE_DEV_DEBUG=0makes dev builds strip too. Executables aren't checked, so shipped binaries keep their strip settings. Upstream fix: llvm/llvm-project#203680.test-canonical-macho.pyhandles Xcode 27's ld sorting local absolute symbols.apple-toolchain.jsonmoves to 27.1/27A9275.lanes.MACOSandci.ymlmove toblacksmith-6vcpu-macos-27in lockstep. So do the signed and release builders (agent-desktop-build, desktop-build, mobile-build, ios-dev-testflight, release, proxy-rust).macos-26toxcode-27: Agent signing, both TestFlight submissions, release.yml's two macOS verifiers and the nightly provider suite.xcode-27is GitHub's preview label without an SLA;repository-workflows.mdlists these jobs and says what to switch to if it's renamed or retired.macos-26, where most users run the apps. "Verify downloaded macOS" is the only job that launches the signed Agent (--version, the duplicate-runtime check and the GUI smoke test), so the first macOS 27 SDK build is launched on macOS 26 before it can be published.xcode-27job exports it as a one-day artifact (.github/actions/export-metal-toolchain).agent-desktop-build.ymland theagent-package-macosrehearsal. The import skips when the toolchain is already installed.ci.ymlthe export is the lanemetal-toolchain. Lanes now takerequires, so the planner runs it exactly whenagent-package-macosruns, dispatches included, andci-okjudges it.lintchecks that requirements exist and form no cycle, that a required lane selects nothing on its own and runs on no event its requirers skip, and that the requiring job needs it.AGENT_DESKTOP) and ofagent-package-macos, so a master push that edits them rebuilds the signed Agent. A test derives every local action the signed builds run from the workflow.overwrite, so "Re-run all jobs" works. A macOS build retried more than a day later needs the export job re-run too; the re-run guide and the Agent build doc say so.github-token.agent-macos(PR Clippy) andagent-full-macosbuild gpui withruntime_shaders. Compiled shaders are still proven by master's signed Prod build on every Agent push, which "Verify downloaded macOS" then runs, plus the nightly Dev build andci:package. The CI skill now says whenci:packageis the PR proof: edits to the Metal actions, gpui/zed bumps and Xcode pin changes.cache_filesand leadsagent-package-macos's key prefix.cargo_graph_key.pytakes optional pin files, and C4 expects the pin formacos-buildonly, so the Windows and Linux keys are unchanged. The design and debugging docs list the pin as a cold trigger.research-ios:Info.ios.plist, the plugins' Swift code and the vendored tao's iOS layer.research-android: the vendored tao's Android layer, which carries Maple's null-intent hardening, and the plugins' Kotlin code.research-rustnever compiles Android-only code, so an edit there ran no Android build before.changing-ci.md,design.md,debugging.md, outage advice),docs/ci.md,repository-workflows.mdand the Agent'sdesktop-builds.md.Validation
plan.py lint, the CI script suite (551 tests), the Agent script suite in its Nix shell (259 tests) andjust lane repo(every root flake check) pass.CARGO_PROFILE_DEV_DEBUG=0and no strip override:just lane agent-macos(72 s) andjust lane agent-full-macos(793 s cold; 770 + 96 + 81 and smaller suites, 0 failures) pass with runtime shaders. That includes the debug app. A releasecargo checkof the whole app at its 15.0 deployment target passes too.ci:package, 40810871028), before the review fixes: every lane passed exceptresearch-dev-macos.metal-toolchainexported the toolchain onxcode-27, andagent-package-macosimported it on Blacksmith macOS 27 and compiled the shaders.research-ios(Xcode 27.1),research-desktop-macos/linux/windows,agent-package-linux/windowsandagent-macos, plus every gate lane.research-dev-macosexposed the dev-profile case:ci.ymlsetsCARGO_PROFILE_DEV_DEBUG=0, so dev builds strip proc-macros too. Reproduced locally and fixed.ci:research-dev-macos, 40812656344):ci-okpasses, includingresearch-dev-macos.ci-okpasses.agent-package-macoson this head (40938899306): with no Agent strip override,metal-toolchainexported the toolchain (0.5 min), and the unsigned Dev package built, imported Metal, compiled the shaders and verified on Blacksmith macOS 27 (11.6 min).ci-oktook 18.2 min, almost all of it the coldresearch-ios(17.4 min), the owner-accepted exception: until master saves the 27A9275 ONNX Runtime artifact, it rebuilds ONNX Runtime and downloads the iOS 27.1 Simulator runtime. The slowest of the other lanes wereresearch-android(6.8 min),agent-macos(5.8, cold by design because the Xcode pin is in its key),os-sdk-e2e(5.0) andagent-windows(4.6). The Agent lanes are warm again: before the last commits,agent-windowscompiled 196 crates and took 11.1 min, and now it compiles 3 (4.6 min).agent-testdropped from 604 crates to 10 (5.0 → 3.3 min).Not testable before merge
Signing, TestFlight and the master-only signed builds, including the Metal export and import in
agent-desktop-build.yml. The first master Agent push is the real check.Merging
Merge PR #1264, then PR #1265, then this one, each into master with a merge commit rather than a squash. After this one merges, watch:
agent-full-macoson the merge push. It'sci:full-only on PRs, so this is its first CI run on macOS 27. If it goes red, fix forward or revert.research-iostakes about 18 min on PRs, so avoid merging other iOS or CI-infra PRs in that window.Heads-up
tauri ios buildignoresDEVELOPER_DIR.xcodebuild -downloadComponent MetalToolchain). It's installed per Xcode version.select-xcode.pyfails every Apple lane and signed build untilapple-toolchain.jsonand the ONNX pins move.[profile]change there changes no cache key: Linux and Windows lanes restore master's cache as a full match and rebuild everything below a proc-macro, and master never saves a fresh cache, because a full match skips the save. An earlier version of this PR did exactly that (agent-windowstook 11 min); the last commit removes it.hdiutilwarns thatcreate,attachandconvertare deprecated in favor ofdiskutil image. The DMG steps still work; moving them is a follow-up.🤖 Generated with Claude Code