chore: untrack generated Rust and iOS outputs, generate on demand - #551
Draft
pgherveou wants to merge 7 commits into
Draft
chore: untrack generated Rust and iOS outputs, generate on demand#551pgherveou wants to merge 7 commits into
pgherveou wants to merge 7 commits into
Conversation
…pi-server The untracking commits fixed only ci.yml's rust job. truapi-server declares `pub mod generated;` unconditionally, so every other job that builds it fails at parse time: ios-bindings, ios-swift, ci-android, release-android and release.yml's publish-ios. Also: - ios-swift never ran sync-bindings.sh or the js/container build, so the package had no Swift sources and no .copy resource to resolve. - the ios-changes filter keyed on an ios/ diff that generated bindings no longer produce; it now names the crates the bindings come from. - release.yml's post-rebuild `git diff --exit-code` could no longer fail. - .gitignore's truapi*.swift glob matched the hand-written TrUAPIHost.swift under core.ignorecase, which git sets by default on macOS. - check-generated is a prerequisite of every target that compiles the crate, names the missing file, and covers mod.rs and wire_table.rs. - sync-bindings.sh --check and provider-swift-check compared against files that are no longer committed.
SwiftPM resolves source targets from the git checkout, so a branch that ignores the generated bindings cannot be consumed. A release is now two artifacts: the xcframework as a release asset, and a plain semver tag whose commit carries the generated sources plus a manifest pointing at that asset. Consumers keep the same URL and the same exact() pin. publish-ios moves into release-ios.yml so it is callable from release.yml and dispatchable for a pre-release, matching release-android.yml. The tag is verified by cloning and building it before the push. tag-release.sh reads the required paths out of Package.swift so the file set cannot drift from what SwiftPM looks for. Also fixes the manifest guard: 9b708fb replaced 'let useLocalBinary = false' with the ProcessInfo form, but release.yml still grepped for the literal, so the follow-up commit step could never pass.
Also compiles the ios-swift CI gate from a clean clone of the tag rather than in place, so a generated file the tag does not carry fails the PR instead of the next consumer.
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.
Generated outputs are not tracked. The Rust dispatcher, wire table and wasm bridge, the UniFFI Swift bindings, the FFI headers and modulemaps, and the lockdown container bundle are git-ignored build outputs produced from the Rust crates on demand.
Producing the outputs
scripts/codegen.shemits the Rust outputs, including themod.rsthattruapi-serverdeclares withpub mod generated;.make setupruns it,make codegenregenerates on its own. Every Make target that compilestruapi-serverdepends oncheck-generated, which names the missing file and points atmake codegeninstead of letting the failure surface aserror[E0583]: file not found for module 'generated'.In CI the
Codegenjob produces them once and uploads thecodegen-outputartifact; every job that compilestruapi-serverdepends on that job and downloads it.ci-android,release-androidandrelease-iosare separate workflows with no such job, so they runscripts/codegen.shin place withTRUAPI_SKIP_PACKAGE_BUILD=1.The
Codegenjob also gates that nothing generated is committed, withgit ls-files -ci --exclude-standardover the whole repo.How iOS consumers get the sources
SwiftPM resolves a package's source targets from the git checkout and has no way to fetch them from a release asset:
binaryTargetaccepts only.xcframeworkand.artifactbundle. So an iOS release is two artifacts.@parity/ios-host@<version>release assettruapi_server.xcframework.zippublishedBinaryURL+ checksum<version>Consumers pin the semver tag, unchanged from today:
ios/truapi-host/scripts/tag-release.shbuilds that commit. It reads the required paths out ofPackage.swiftrather than repeating them, so the file set cannot drift from what SwiftPM looks for, and it refuses to tag an xcframework into git.publish-iosmoves out ofrelease.ymlintorelease-ios.yml, callable from the release flow and dispatchable on its own, matchingrelease-android.yml. Dispatching it from a branch with a version like0.12.0-beta.1cuts an asset and a tag an app can pin, which is how an unmerged host change gets tested in the app. A dispatched run passes nomanifest_branchand so rewrites no branch.The job clones the tag and compiles it against the published asset before pushing, so a tag that cannot be resolved is never published.
What CI proves
iOS package (swift compile)now builds the tag rather than the working tree: it runstag-release.sh, clones the result, diffs the cloned file set against the working tree under every declared path, and compiles the package and its test target from that clean clone. Compiling in place would pass on generated files the tag does not carry, which is the exact failure being guarded.iOS bindings (uniffi)regenerates the Swift and provider bindings. With nothing committed to diff against, what it gates is that bindgen still produces a binding for every UniFFI-exposed type.ios-swift's path filter names every crate the bindings are generated from (truapi,truapi-platform,truapi-server,truapi-provider) plusjs/container/, because a protocol change no longer leaves anios/diff to key on.Notes
release.yml's manifest follow-up step guarded on^let useLocalBinary = false$, but9b708fb5had replaced that line with theProcessInfoform. Underset -euo pipefailthe step could never pass, which is whymain's manifest still names ios-host@0.7.0 while 0.11.0 is published. The guard now asserts the manifest reads the environment variable, which is the invariant that actually matters.Package.swiftdeclaresTrUAPIProvideras a source target, so its bindings have to exist even for a consumer that only importsTrUAPIHost.ios/truapi-provider/scripts/sync-bindings.shproduces them without Xcode or the iOS targets, andrebuild.shdelegates to it..gitignorelists the three generated Swift filenames rather thantruapi*.swift. Undercore.ignorecase, which git sets by default on macOS, the glob also matched the hand-writtenTrUAPIHost.swift.sync-bindings.sh --checkand theprovider-swift-checktarget compared against committed files and are gone.mod.rsemitted alongsidedispatcher.rsandwire_table.rshas a golden fixture. It is the file that makestruapi-serverparse, so losing it fails a test rather than the next build.Verifying the release path
release-ios.ymlis new, and GitHub only exposesworkflow_dispatchfor workflows present on the default branch, so it cannot be dispatched from this branch. The tag's content and buildability are covered on the PR by theiOS package (swift compile)job above. Cutting0.12.0-beta.1to exercise the publish half (asset upload, tag push, resolution from GitHub by the app) needs this merged first.