feat(iOS): ReactNativeDependenciesHeaders sidecar + pure-RN ReactNativeHeaders, published to Maven by chrfalch · Pull Request #57442 · react/react-native

chrfalch · 2026-07-06T10:54:15Z

Summary

Step 2 of the prebuilt-deps roadmap: ship the deps headers as a SwiftPM-ready, self-contained artifact and make every header namespace have exactly one physical home.

New artifact: ReactNativeDependenciesHeaders.xcframework — the binary ReactNativeDependencies.xcframework is framework-type, so its root Headers/ is invisible to SwiftPM binaryTargets (HeadersPath is rejected on framework entries; verified empirically). The deps prebuild now emits a headers-only library-type sidecar (stub archives + per-slice Headers/ + HeadersPath — the exact ReactNativeHeaders recipe, factored into a shared headers-xcframework.js emitter) carrying all seven deps namespaces incl. SocketRocket, with slice parity derived from the binary artifact's Info.plist. Ships inside the deps tarball and standalone.
ReactNativeHeaders goes pure-RN — the R2 relocation of deps namespaces (and the DEPS_NAMESPACES_NOT_RELOCATED SocketRocket exclusion list) is deleted. Relocated copies are what enabled the SocketRocket dual-copy regression (duplicate @interface / poisoned module graph under use_frameworks!); that bug class is now structurally impossible. Headers gate flipped: deps namespaces must be absent from RNH; the sidecar emitter enforces set-equality with DEPS_NAMESPACES fail-closed in both directions. On the CocoaPods side, a new ReactNativeDependenciesUtils.configure_aggregate_xcconfig injects the deps pod's Headers/ globally (aggregate + every pod target), mirroring the rncore injection — this replaces the folly/glog resolution pods previously got via the flattened React-Core-prebuilt/Headers.
CI: prebuilt + dynamic-frameworks lane — the regression's exact config had no coverage (the test-ios-rntester action hard-coupled use-frameworks:true to source builds). New use-prebuilds input; test_ios_rntester's dynamic cells now consume the workflow-built prebuilt artifacts.
Maven publishing — ReactNativeHeaders and ReactNativeDependenciesHeaders publish standalone on react-native-artifacts (classifiers reactnative-headers-*, reactnative-dependencies-headers-*); verifyArtifactsAreOnMaven now HEAD-checks every classifier tarball instead of only the POM.

Stacked on #57440. The SwiftPM preview (#57332) rebases on top and wires the sidecar as its 5th binaryTarget.

Changelog:

[IOS] [CHANGED] - Prebuilt artifacts: ReactNativeHeaders is pure-RN; third-party deps headers ship in the new ReactNativeDependenciesHeaders.xcframework sidecar (and the ReactNativeDependencies pod), published standalone to Maven

Test Plan

Headers gate: include-health, structural (deps absent from RNH, byte-matched module maps), and compile smokes (React module + 14 namespace modules + Expo-shape ObjC++/Swift fixtures vs the deps include path) — ALL PASSED
jest: 33/33 (scripts/ios-prebuild/__tests__, incl. new sidecar set-equality tests)
ESLint (--max-warnings 0), Prettier, Flow (yarn flow-check): clean
E2E (locally built artifacts): rn-tester prebuilt static ✅, prebuilt USE_FRAMEWORKS=dynamic ✅ (the regression config — verified React-Core-prebuilt/Headers contains no deps namespaces and the deps pod serves all seven), helloworld static ✅, source-core + prebuilt-deps ✅ (React compiled from source resolves folly via the deps pod), source-mode control with unchanged dependency graph ✅
Sidecar inspected: per-slice HeadersPath, 7 namespaces, slice parity with the binary
Publication validated end-to-end with publishReleasePublicationToMavenLocal: all 12 files + POM land with the expected classifier names

🤖 Generated with Claude Code

The minimal machinery to build the packaged header structures: - headers-spec.js: the executable layout contract (rules R1-R8) — which namespaces are hoisted into the React framework, which carry module maps, and how collisions are rejected. - headers-inventory.js: scans the source tree and classifies every shipped header (language surface + modularizability bucket) — the input to the spec. computeInventory() feeds the build in-memory; the CLI writes a JSON manifest. - headers-compose.js: emits the layout — writes the <React/...> headers + umbrella + module map into each React.framework slice (detected by the framework's presence), and assembles the headers-only ReactNativeHeaders.xcframework (every other namespace + deps + Hermes). Called by xcframework.js during compose. This is the alternative header source that lets consumers resolve React Native headers without a clang VFS overlay. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Emit the headers-spec layout unconditionally and delete the VFS overlay across JS, CI publish, and Ruby. Consumers resolve headers the way the SwiftPM branch does: <React/...> from the vendored React.framework, every other namespace from ReactNativeHeaders. No root Headers/ on the xcframework, no VFS. JS: - xcframework.js: always emit the React.framework spec layout and build ReactNativeHeaders.xcframework (was gated behind RN_ZERO_I_LAYOUT=1). Remove the legacy header path entirely — the podspec->root-Headers enumeration, createModuleMapFile, and copyHeaderFilesToSlices — so the published React.xcframework is a standard framework (Info.plist + per-slice React.framework/{Headers,Modules}), no root Headers/ or Modules/. Ship ReactNativeHeaders.xcframework inside the reactnative-core tarball (sibling of React.xcframework) so the prebuilt pod can vend both; keep the standalone ReactNativeHeaders.xcframework.tar.gz for the SPM path. Drop the React-VFS-template.yaml emit and the ./vfs import. - vfs.js: deleted (its only consumer was xcframework.js). - types.js: drop the now-unused VFSEntry/VFSOverlay/HeaderMapping types. - replace-rncore-version.js: drop the React-VFS.yaml preservation rationale. Ruby/CocoaPods: - React-Core-prebuilt.podspec: vend React.xcframework (its per-slice React.framework + module map serves <React/...> and @import React via FRAMEWORK_SEARCH_PATHS); flatten ReactNativeHeaders' headers into a top-level Headers/ in prepare_command and expose them via the pod header search path. Drop the VFS-era root header_mappings_dir/module_map. - rncore.rb: remove the -ivfsoverlay injection and process_vfs_overlay; add_rncore_dependency and configure_aggregate_xcconfig now add a HEADER_SEARCH_PATHS to React-Core-prebuilt/Headers for podspec, aggregate, and third-party targets. - react_native_pods.rb: drop the process_vfs_overlay post-install call. Docs: replace the "VFS Overlay System" section with the headers-spec layout; drop the obsolete "Known Issues" (pre-headers-spec) section. Verified end-to-end: prebuild compose produces a VFS-free, root-Headers-free React.xcframework; rn-tester pod install + xcodebuild (prebuilt path) BUILD SUCCEEDED with zero -ivfsoverlay. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

React.framework is a clang module; when an SPM consumer precompiles it, a modular React/ header that #imports <react/...> hit -Wnon-modular-include-in-framework-module because the lowercase react/ namespace (served from ReactNativeHeaders, per R1's Linux/Windows-safe layout) was deliberately kept out of any module. Give react/ a module where it already lives instead of relocating it (relocation would require case-folding react.framework -> React.framework, which only works on case-insensitive filesystems): - headers-spec.js: drop the react/ namespace-module exemption so its objc-modular-candidates get a module; emit that module as ReactNativeHeaders_react (a module literally named 'react' would alias the React framework module on a case-insensitive filesystem). Module names are internal; <react/...> still resolves by header path and is now modular. - headers-inventory.js: classify C++ default member initializers in aggregates (e.g. struct { NSString *family = nil; } in RCTFontProperties.h) as ObjC++ so these are not misclassified objc-modular-candidate and pulled into a plain ObjC module they cannot compile in. The unguarded ObjC/C react/ headers (e.g. JSRuntimeFactoryCAPI.h) now resolve modularly; the C++ react/renderer/* includes are #ifdef __cplusplus-guarded and skipped during the ObjC module emit, so they need no module. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

In prebuilt mode the React core pods' code + headers live entirely in React.xcframework / React-Core-prebuilt. Re-installing their SOURCE podspecs made them ship duplicate headers that shadow the prebuilt artifact and break the React framework's clang explicit-module precompile (-Wnon-modular-include-in-framework-module) under Xcode 26. Install those core pods as dependency-only FACADES instead: generated podspecs with no sources/headers, installed via :path (so nothing is fetched), each depending on React-Core-prebuilt. Version, subspecs, default_subspec and resources (e.g. the privacy manifest) are DERIVED from the real podspec so the facade stays graph- and resource-equivalent to the source pod. With the shadowing gone the React module precompiles cleanly with SWIFT_ENABLE_EXPLICIT_MODULES on, so the Xcode-26 workaround (#53457) is removed. The prebuilt header search path + ReactNativeHeaders module-map activation are consolidated into a single post-install injection site (configure_aggregate_xcconfig); add_rncore_dependency now only declares the React-Core-prebuilt dependency. rn-tester's NativeComponentExample uses the canonical <React/...> include for RCTFabricComponentsPlugins.h (resolved from the framework) so it builds against the facaded React-RCTFabric in prebuilt mode. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

`yarn format-check` (prettier) was failing CI on PR #57285. Run prettier on the ios-prebuild headers scripts (headers-compose.js, headers-inventory.js), replace-rncore-version.js, and __docs__/README.md so format-check passes. No logic changes. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…tarball The prebuilt React core now ships two xcframeworks — React.xcframework and the headers-only ReactNativeHeaders.xcframework. React-Core-prebuilt's prepare_command flattens the latter's Headers (including module.modulemap) into the pod. The compose step only tar'd React.xcframework, so consumers got no React-Core-prebuilt/Headers/module.modulemap and failed the build with "module map file ... not found". Tar both xcframeworks. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…le (R9) The public-umbrella model (which replaced the VFS overlay) excludes `+Private` and objc-blocked headers from React.framework's module, so privileged framework consumers (e.g. Expo) that `#import <React/RCTBridge+Private.h>`, `<React/RCTMountingManager.h>`, etc. fail to compile under explicit modules even though the headers still ship in React.framework/Headers. Add R9: a curated allowlist appended to the React module map — `RCTBridge+Private.h` as a real `header` (objc-modular-candidate, reaches no C++) and the six Fabric headers as `textual header` (objc-blocked; a real member would re-trip -Wnon-modular-include, and their <react/...> C++ includes resolve at the consumer's use site). Backwards-compatible: existing `#import <React/...>` (and Swift `import React`) sites are unchanged. Fails closed if an allowlisted header is removed/renamed or drifts bucket. Note: RCTUIKit.h / RCTRootContentView.h are absent from source entirely and need restoration, not exposure — out of scope here.

The flattened ReactNativeHeaders layout ships the individual React_RCTAppDelegate/*.h headers but no per-namespace umbrella. Consumers like Expo probe `<React_RCTAppDelegate/React_RCTAppDelegate-umbrella.h>` via __has_include (RCTAppDelegateUmbrella.h); with the umbrella gone the probe fails and RCTReactNativeFactory / RCTRootViewFactory are never declared, breaking the Expo pod's clang module. Add R10: emit a per-namespace umbrella (content DERIVED from namespaceModules so it can't drift — e.g. RCTArchConfiguratorProtocol.h, gone from this branch, is correctly omitted) and add it to that namespace's module so the import stays modular under explicit modules. Targeted via UMBRELLA_NAMESPACES (currently just React_RCTAppDelegate, the only umbrella Expo imports); fails closed if a listed namespace loses its modular headers.

…ic facade Community Fabric modules quote-import "RCTFabricComponentsPlugins.h" (~47x: slider, maps, pager-view, keyboard-controller, ...). In source, React-RCTFabric vended it at header_dir "React", so it landed in dependents' CocoaPods header maps and the bare quoted name resolved. In prebuilt mode React-RCTFabric is a dependency-only facade that ships no headers — the only copy is baked angle-only into React.framework (it's objc-blocked, excluded from the framework module map), so quoted imports fail to resolve. Re-vend JUST that one header from the facade (FACADE_REEXPOSED_HEADERS), copied as a self-contained snapshot at header_dir "React", restoring dependents' header maps exactly as the source pod did. Header-only (no compiled sources, no duplicate symbols — the impl stays in React.framework). Re-exposing a single header does not put <react/...>/<yoga/...> on -I, so it does not reintroduce the non-modular-include shadowing the modular layout eliminates. Fails closed if the glob matches nothing.

…the headers compose buildReactNativeHeadersXcframework copied each declared deps namespace (folly/glog/boost/fmt/double-conversion/fast_float) from the staged ReactNativeDependencies headers, but only console.warn'd on a missing one and kept going — silently shipping a ReactNativeHeaders.xcframework without third-party header resolution (consumers then fail on <folly/...> etc.). The summary log also printed the INTENDED namespace list, masking the gap. Throw instead: a missing declared deps namespace means deps weren't staged (third-party/ReactNativeDependencies.xcframework/Headers — from a full prebuild or the cache slot), so refuse to emit an incomplete artifact.

…est imports Two CI fixes for the prebuild-ios-core workflow: - prebuild-ios-core.yml: the compose-xcframework job downloaded the build slices and React headers but never staged third-party/ReactNativeDependencies.xcframework. buildReactNativeHeadersXcframework folds the third-party deps namespaces (folly/glog/boost/...) into ReactNativeHeaders.xcframework, so it fail-closed with "deps namespace 'folly' missing ... refusing to ship an incomplete ReactNativeHeaders.xcframework". Add the Download + Extract ReactNativeDependencies steps (mirroring build-slices) so the deps headers are present before composing. - headers-spec-test.js: reorder requires so the `../headers-spec` import sorts before `fs`, fixing the @react-native/monorepo/sort-imports warning that failed `eslint --max-warnings 0` in test_js. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- copy facade resources into the facade pod dir: CocoaPods file accessors cannot glob past the pod root, so the ..-escaping globs shipped the privacy-manifest / i18n resource bundles empty - quote -fmodule-map-file so a PODS_ROOT containing spaces stays a single clang argument (matches the quoted HEADER_SEARCH_PATHS beside it) - fail closed in React-Core-prebuilt's prepare_command and in replace-rncore-version.js when the tarball lacks ReactNativeHeaders.xcframework, instead of silently leaving an empty or deleted Headers/ with the module-map flag dangling - thread rnRoot through planFromInventory/isUmbrellaSafe instead of the hardcoded hosting-package root (the one spot that didn't take the inventory's root) - drop the dead ios-prebuild templates/ files (their only consumers were removed by the headers-spec compose) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…1, verify, ratchet) Hardens the header layout against new/changed headers so consumer-facing regressions fail the PREBUILD instead of a downstream (rn-tester/Expo/ community) build: - headers-verify.js (new; runs in the prebuild compose CI job): include-health ratchet against a committed baseline (notShipped/unresolved/quoted- unresolvable includes in shipped headers — 27 baselined today); structural byte-compare of the composed module maps/umbrellas against the spec render; compile smokes — an ObjC TU precompiling the React module (every umbrella header) + all 14 R5 namespace modules + the R10 umbrella + __has_include asserts, an Expo-shaped ObjC++ TU (the R9 textual Fabric surface), and a Swift TU (import React + RCTBridge.moduleRegistry). - R11: one source, one content location. 116 sources ship under multiple spellings (React/X.h + CoreModules/RCTImage/RCTAnimation/... forms, bare React_RCTAppDelegate aliases). The flattened layout duplicated their declarations, so any -fmodules consumer touching two spellings (even transitively via a legacy spelling) hit redefinition errors — found by the gate's first run. The module-owned spelling keeps the content; every other spelling is emitted as a one-line redirect shim. - Single source of truth for third-party namespaces: the inventory's include classifier now derives from DEPS_NAMESPACES, and compose enforces set-equality with the deps artifact in BOTH directions (missing OR undeclared namespace fails). The undeclared direction immediately surfaced SocketRocket, which the deps artifact ships but was never relocated. - R5 exemption assert: an invalid-module-identifier namespace gaining a modular-candidate header now fails the plan instead of silently shipping a non-modular header. - Inventory records quoted includes that don't resolve to a shipped header (quotedNotShipped) instead of dropping them. Docs: __docs__/headers-rules.md documents rules R1-R11, the emission pipeline, and the resilience model.

…int MD040)

…x-warnings 0)

@interface

…iveHeaders Declaring SocketRocket in DEPS_NAMESPACES (added to satisfy the deps set-equality guard) relocated textual copies of its headers into ReactNativeHeaders and thus into the flattened React-Core-prebuilt/Headers, which sits on every pod's HEADER_SEARCH_PATHS. The real SocketRocket pod still exists in consumer graphs and vends the same headers, so the copies collide: duplicate @interface definitions when the SocketRocket pod compiles (rn-tester, prebuilt + USE_FRAMEWORKS=dynamic), and a poisoned module graph in use_frameworks/explicit-modules apps that don't compile SocketRocket sources (Expo: ReactCodegen/RCTModulesConformingToProtocolsProvider.h "file not found"). Bisect confirmed removing the SocketRocket dir flips red->green and the R11 shims are not implicated. Add DEPS_NAMESPACES_NOT_RELOCATED: such namespaces count as declared for the set-equality guard (a genuinely new dep still fails closed), the include classifier still recognizes them, and the headers gate now asserts they stay ABSENT from the artifact. Note: no CI lane covers prebuilt + use_frameworks (the dynamic-frameworks lane builds from source), which is why CI stayed green.

Review feedback on embedding React.framework's non-header resources: - Privacy-manifest merge no longer fabricates an empty NSPrivacyAccessedAPITypeReasons key when a source omitted it (tracks per-category key presence), so a single manifest truly passes through unchanged. - NSPrivacyCollectedDataTypes dedup canonicalizes (recursively sorts) keys before hashing, so two pods declaring the same dict in different key order dedup. - RCTI18nStrings.bundle is now built ONCE into a temp stage and cloned into each slice (mirrors the privacy manifest) instead of rebuilt inside the slice loop; i18nLocales is computed once. - i18n bundle Info.plist gains CFBundleShortVersionString / CFBundleVersion so Apple validation tooling doesn't warn on a version-less bundle. - Fixed stale "see" pointers ios-prebuild/{i18n,privacy}.js -> framework-resources.js (React-Core.podspec + rncore_facades.rb). - Tests: serialize->readPrivacyManifest round-trip, order-insensitive collected-type dedup, omitted-reasons-key regression, CFBundleVersion presence, and an emitReactFrameworkHeaders integration test asserting PrivacyInfo.xcprivacy + RCTI18nStrings.bundle land in EVERY slice. Note: resources are written after _CodeSignature is stripped (R7 — signed after compose), so no signature invalidation; readdirSync({recursive})/cpSync are safe on the repo's Node >=22 engine. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

mixed is now a deprecated utility type (use unknown), and the NSPrivacyAccessedAPITypeReasons refinement was invalidated by the intervening Map.get() call — hoist it into a local first. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…dle clone Same macOS-only clonefile issue as the other copy sites; this one is on the resource-landing path the jest integration test exercises on Linux. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The resources shipped inside the prebuilt React.framework fail SILENTLY if the framework is ever consumed without being embedded: every localized string falls back to its untranslated default and the privacy manifest drops out of the app's aggregated privacy report, with no build error. Emit a dev-only, once-per-process RCTLogWarn when the bundle resolves nil so an embed regression is observable in every dev console instead of shipping quietly. No behavior change in release (RCT_DEV compiled out) or on the happy path. Not unit-testable (ObjC runtime diagnostics, no jest surface); both RCT_DEV and NDEBUG preprocessing paths compile-verified via clang -fsyntax-only. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

… manifests collectReactPrivacyManifestPaths merged ANY PrivacyInfo.xcprivacy found recursively under the React privacy roots — a latent trap: a future pod under those roots that ships as its OWN framework would silently have its manifest folded into React.framework's aggregate (over-declaration in the app's privacy report). The scan is now validated against an explicit REACT_PRIVACY_MANIFESTS allowlist and FAILS the prebuild on an unlisted manifest, with instructions for both resolutions. Listed-but-absent stays legal: partial fixture trees keep working, an upstream deletion under-declares exactly as a source build would, and a MOVED manifest cannot slip through since its new path is unlisted. Also: only create the i18n temp stage when there are .lproj dirs to bundle, and document that the dedup key's `?? ''` is Flow appeasement (JSON.stringify is typed `string | void`), not dead code. 56 ios-prebuild tests green (drift-gate cases red-first); Flow + prettier clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

… handling - rndependencies.rb: the previous "||= [] << path" only added the deps header search path when HEADER_SEARCH_PATHS was unset — with an existing value the path was silently dropped. Normalize string values and always append. Also point the search path at the pod-local flattened Headers/ (single header home) instead of the artifact root. - ReactNativeDependencies.podspec: prepare_command exited 0 when XCFRAMEWORK_PATH/HEADERS_PATH resolved empty, producing a silent no-link pod; fail closed with exit 1. - reactNativeDependencies.js: an artifact without a version marker logged "we are going to use it anyway" but then fell through to rmSync + re-download, destroying locally staged deps builds; honor the message and use it as-is. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…uilt-deps mode In prebuilt-deps mode the real source pods (RCT-Folly, glog, boost, DoubleConversion, fmt, fast_float, SocketRocket) are not declared in the Podfile, so a community podspec's hardcoded 's.dependency "RCT-Folly"' would resolve from the CocoaPods trunk and compile from source next to the prebuilt binary — the dual-copy bug class behind the 2026-07-03 SocketRocket regression. RNDepsFacades generates dependency-only local facade podspecs (build/rndeps-facades/<Name>/): no sources, no headers, a single dependency on ReactNativeDependencies. Versions and subspecs are derived from the real podspecs in third-party-podspecs/ (RCT-Folly keeps /Default + /Fabric with default_subspecs = ["Default"]; SocketRocket is synthesized from socket_rocket_config, fail-closed if absent). Declared as :path pods so Podfile-local resolution beats trunk and nothing is fetched. ReactNativeDependencies remains the single header authority. Docs: scripts/cocoapods/__docs__/prebuilt-deps.md describes the full prebuilt-deps header/facade contract and the mode-by-mode supplier table. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ayout - Embed an RN-authored, accurate-empty PrivacyInfo.xcprivacy for SocketRocket in the deps prebuild (upstream ships no privacy manifest; the deps artifact historically carried resource bundles only for boost/folly/glog). SocketRocket uses no Required Reason APIs. - Xcode 26's SwiftPM rejects a public-headers directory where the umbrella header has sibling directories ("target 'SocketRocket' has invalid header layout"): stage the flat public headers into include/ via prepareScript and point publicHeaderFiles there. Xcode 16 accepts both layouts; the artifact's Headers/SocketRocket namespace is unaffected. - Rename the dependency 'socket-rocket' -> 'SocketRocket' so the bundle and header-namespace naming line up (ReactNativeDependencies_SocketRocket.bundle). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Pure formatting — no content change. Fixes the format-check lint lane. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Review feedback on the prebuilt-deps facades: - rndeps_facades.rb: soften the "graph-equivalent" comment — facades match the source pod's spec/subspec shape but route every derived subspec's dep to ReactNativeDependencies, so intra-pod subspec deps (RCT-Folly/Fabric -> RCT-Folly/Default) are not reproduced (harmless; the deps are declared explicitly in react_native_pods.rb). - rndeps_facades.rb: add `require_relative './helpers'` so Helpers::Constants is self-contained against require ordering (the defined? guard stays as backstop). - reactNativeDependencies.js: document that the no-version-marker branch returns before the version.txt write, so a locally-staged artifact never gains a marker and later runs re-hit the branch (intentional; downstream must not assume it). - rndependencies.rb: uniq the HEADER_SEARCH_PATHS so a second call can't duplicate. - prebuilt-deps.md: drop the stray space before SocketRocket. Note: the Ruby-tests request is not actioned — the repo's cocoapods Ruby test harness is not wired into CI, so the tests would not run. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…ODS hint distinct Two review follow-ups on the deps facades: - The prebuilt branch of add_rn_third_party_dependencies appended the deps Headers search path unconditionally; a second call on the same spec would duplicate the entry. Now .uniq, mirroring the from-source branch. - load_real_spec's method-level rescue also caught the tailored not-found raise and re-wrapped it generically, burying the "Update FACADE_PODS" hint. The rescue is now scoped to Pod::Specification.from_file only, so the two failure messages stay distinct. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…idecar The binary ReactNativeDependencies.xcframework is FRAMEWORK-type: it cannot carry a HeadersPath key, so its root Headers/ dir is invisible to SwiftPM binaryTargets (verified empirically 2026-07-04 - xcodebuild rejects -headers without -library, and SwiftPM hard-rejects HeadersPath on framework entries). This adds a headers-only LIBRARY-type sidecar, ReactNativeDependenciesHeaders.xcframework, that SwiftPM auto-serves with zero flags: stub static archives paired with per-slice Headers/ holding the seven third-party namespaces (folly, glog, boost, fmt, double-conversion, fast_float, SocketRocket). - headers-xcframework.js (new): dependency-light shared emitter - composeHeadersOnlyXcframework (the stub-archive recipe factored out of buildReactNativeHeadersXcframework), stubSlicesFromXcframework (slice parity derived from the binary artifact's Info.plist, so the sidecar resolves for every platform the binary does), and buildDepsHeadersXcframework with a set-equality gate that fails closed in both directions (declared-but-missing and undeclared namespaces). - compose-framework.js: the deps compose emits and signs the sidecar right after assembling the root Headers/ (slice-uniform content). - prebuild-ios-dependencies.yml: the sidecar ships INSIDE the deps tarball (self-contained for CocoaPods) and as a standalone tarball; cache key bumped and now also hashes compose-framework.js. - reactNativeDependencies.js: consumers extract the sidecar alongside the binary; pre-sidecar tarballs (pinned RN_DEP_VERSION) warn and continue - CocoaPods is unaffected, SwiftPM recomposes it locally. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

@interface

Every namespace now has exactly ONE physical home: RN namespaces in ReactNativeHeaders, third-party deps namespaces in the ReactNativeDependenciesHeaders sidecar / the ReactNativeDependencies pod. Relocated deps copies are what made the 2026-07-03 SocketRocket regression possible (duplicate @interface + poisoned module graph under use_frameworks when a real pod coexisted with relocated copies) - that bug class is now structurally impossible, and the SocketRocket exclusion list is deleted rather than maintained. - headers-spec.js: R2 rewritten; DEPS_NAMESPACES now includes SocketRocket (= the sidecar contents = the artifact's Headers/ dirs, set-equality enforced); DEPS_NAMESPACES_NOT_RELOCATED deleted. - headers-compose.js: RNH stages RN headers only; ensureHeadersLayout also builds the sidecar from the cache slot's deps headers and returns it, so SwiftPM consumers need no published sidecar. - headers-verify.js: structural gate flipped - deps namespaces must be ABSENT from RNH; compile gates take the deps artifact's Headers on the include path (exactly how consumers resolve them). - CocoaPods: NEW ReactNativeDependenciesUtils.configure_aggregate_xcconfig injects $(PODS_ROOT)/ReactNativeDependencies/Headers into the aggregate and every pod target in prebuilt-deps mode, mirroring the rncore injection. Pods previously resolved folly/glog through the globally-injected React-Core-prebuilt/Headers flatten of RNH; after the flip the deps pod is the single global home, so the mirror injection is what keeps arbitrary pods compiling. Verified: headers gate green (include-health, structural, compile smokes vs the deps include path); rn-tester prebuilt static + dynamic (the regression config), helloworld static, source-core+prebuilt-deps (React compiled from source resolves folly via the deps pod), and a source-mode control with an unchanged dependency graph. BREAKING (prebuilt artifacts only): ReactNativeHeaders.xcframework no longer contains folly/glog/boost/fmt/double-conversion/fast_float; consumers reaching them through React-Core-prebuilt/Headers must use the ReactNativeDependencies pod / ReactNativeDependenciesHeaders sidecar paths instead. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

prebuilt + use_frameworks is the exact configuration of the 2026-07-03 SocketRocket dual-copy regression, and no CI lane covered it: the test-ios-rntester action hard-coupled use-frameworks:true to building from source. Add a use-prebuilds input ('auto' keeps the historical coupling for existing callers) and pass use-prebuilds:true from test_ios_rntester, so its dynamic-frameworks matrix cells now consume the prebuilt ReactCore + ReactNativeDependencies artifacts built in the same workflow run. Source-built dynamic frameworks stay covered by test_ios_rntester_dynamic_frameworks. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…eaders to Maven Publish the two headers-only xcframeworks as standalone artifacts on the react-native-artifacts coordinate so SwiftPM consumers can wire them as separate binaryTargets (they also keep shipping inside the ReactCore / ReactNativeDependencies tarballs for CocoaPods): - external-artifacts/build.gradle.kts: four new PublishArtifacts with classifiers reactnative-headers-{debug,release} and reactnative-dependencies-headers-{debug,release}. - prebuild-ios-core.yml: upload the standalone ReactNativeHeaders<flavor>.xcframework.tar.gz the compose step already produces (previously only bundled inside the ReactCore tarball); include it in the workflow cache. - build-npm-package action: download ReactNativeHeaders* artifacts (ReactNativeDependenciesHeaders* already rides the existing ReactNativeDependencies* pattern). - verifyArtifactsAreOnMaven.js: HEAD-check every classifier tarball, not just the POM - a release with a missing classifier artifact now fails verification instead of passing silently. Validated locally end-to-end with publishReleasePublicationToMavenLocal: all 12 files + POM land on com.facebook.react:react-native-artifacts with the expected classifier names. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The verifier now HEAD-checks the POM plus all eight classifier tarballs instead of GETting only the POM; the mocks/assertions follow, and a new case covers a missing classifier behind an existing POM (the exact gap the classifier checks were added to catch). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Review feedback on ReactNativeDependenciesHeaders + pure-RN ReactNativeHeaders: - prebuild-ios-dependencies.yml: the sidecar cache key now hashes the two cross-package files it is actually generated from — ios-prebuild/ headers-xcframework.js (emitter) + headers-spec.js (DEPS_NAMESPACES) — in BOTH occurrences, and bumps v4->v5. A DEPS_NAMESPACES/emitter edit can no longer serve a stale cached sidecar on main. - prebuild-ios-core.yml: bump the core xcframework cache key v2->v3 to reflect the added ReactNativeHeaders tarball in the artifact set (was relying only on the hashFiles component). - headers-xcframework.js: wrap both staging temp dirs (.stub-work-* / .deps-headers-stage-*) in try/finally so an xcrun/xcodebuild failure can't leave them behind in outDir (= third-party/). - verifyArtifactsAreOnMaven.js: clarify the comment — the 8 primary xcframework classifier tarballs are HEAD-checked; the 4 dSYM classifiers are intentionally excluded. - test-ios-rntester/action.yml: harden the use-prebuilds resolution — only true/false enable/disable; an unexpected value warns instead of silently disabling. - reactNativeDependencies.js: comment reconciling the two sidecar sources (the full-slice tarball copy vs the ensureHeadersLayout DEFAULT_STUB_SLICES fallback) — they agree on the iOS slices SPM needs; headers are slice-uniform. - Tests: stubSlicesFromXcframework known-slice mapping + unknown-slice throw (mocked plutil JSON). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

mixed is a deprecated utility type (use unknown), and Buffer.from needs the string+encoding overload since JSON.stringify on unknown types as string | void. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…sidecar Same macOS-only clonefile flag as headers-compose; keeps the prebuild scripts runnable under Linux CI jest. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ose freshness marker composeToolingHash() covered only three of headers-compose.js's five local sibling modules, so a local edit to headers-xcframework.js (which ensureHeadersLayout delegates the ReactNativeHeaders compose and the deps sidecar build to) or framework-resources.js (privacy manifest + i18n bundle emission) served stale composed output. Extract the list to COMPOSE_TOOLING_FILES, add the two missing files, and add a guard test that parses this file's own requires so the list can't silently drift again — the same bug class as the CI cache-key hashFiles gap, now unrepresentable. Also rewrite the stale headers-rules.md paragraph that still described the removed DEPS_NAMESPACES_NOT_RELOCATED carve-out; SocketRocket lives in DEPS_NAMESPACES with a single physical home in the deps sidecar. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…mework.js The pure-RN commit moved the stub-xcframework toolchain out of headers-compose.js (which got execFileSync arg arrays on the vfs branch) into this emitter; apply the same conversion here so the shell-quoting fix survives the move: plutil/xcrun/clang/libtool/lipo/xcodebuild/cp all take argument arrays now, with libtool's stderr suppression via stdio options. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…se guard test new Set() needed an explicit type argument under strict-local, and the exec-in-while-condition loop lost Flow's null refinement — matchAll avoids both. headers-rules.md reflowed by prettier. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

meta-codesync · 2026-07-10T14:57:05Z

@cipolleschi has imported this pull request. If you are a Meta employee, you can view this in D111449462.

meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Jul 6, 2026

facebook-github-tools Bot added p: Expo Partner: Expo Partner labels Jul 6, 2026

chrfalch force-pushed the chrfalch/deps-headers-sidecar branch from 73226c2 to c01f55a Compare July 6, 2026 12:10

chrfalch and others added 17 commits July 6, 2026 18:35

style: prettier-format headers-rules.md (format-check lints md too)

e6ab638

style: add language tags to headers-rules.md fenced blocks (markdownl…

6648f90

…int MD040)

style: sort requires in headers-inventory.js (eslint sort-imports, ma…

dcb5ad8

…x-warnings 0)

chrfalch force-pushed the chrfalch/prebuilt-deps-facades branch from 069c350 to 017ca51 Compare July 6, 2026 16:37

chrfalch force-pushed the chrfalch/deps-headers-sidecar branch from 90d8f50 to 0c44552 Compare July 6, 2026 16:37

chrfalch mentioned this pull request Jul 6, 2026

feat(iOS) swiftpm support in iOS #57332

Open