[native_assets_cli] iOS/MacOS privacy manifest

[dcharkes]

trafficstars

When using something like the camera on iOS/MacOS, the dylib needs to be bundled with a privacy manifest about why the permission is needed.

We need some way to communicate a privacy manifest from build.dart to Flutter.

This could possibly be a new asset type.

We need to check if we need a privacy manifest per dylib or just per Dart package.

• https://github.com/flutter/flutter/issues/140966

[dcharkes]

Discussing with @mosuem we have multiple options that we can pursue with adding new asset types that are connected to existing asset types:

API options

1. NativeCodeAsset.privacyManifest Uri? field

Introduce a NativeCodeAsset.privacyManifest Uri? field

output.addAssets([
  NativeCodeAsset(
    package: 'native_add',
    name: 'native_add',
    linkMode: DynamicLoadingBundled(),
    os: OS.current,
    architecture: Architecture.current,
    file: builtDylibUri,
    privacyManifestFile:
      (OS.current == OS.macOS || OS.current == OS.iOS) ?
      config.packageRoot.resolve('assets/PrivacyInfo.xcprivacy') :
      null,
  ),
]);

• Pro: concise
• Con: doesn't communicate that privacy manifests are not supported on various embedders. (We have BuildConfig.supportedAssetTypes, but it's not used this way.)
• Con: pollutes the NativeCodeAsset API for other OSes that don't have privacy manifests.

2. PrivacyManifestAsset with dylib field

Introduce a PrivacyManifestAsset which implements Asset and has a field dylib which points to a NativeCodeAsset.

(Note the class should not implement/extend DataAsset because the bytes should not be bundled in Dart/Flutter apps and accessible with data assets at runtime.)

output.addAssets([
  NativeCodeAsset(
    package: 'native_add',
    name: 'native_add',
    linkMode: DynamicLoadingBundled(),
    os: OS.current,
    architecture: Architecture.current,
    file: builtDylibUri,
  ),
  if (OS.current == OS.macOS || OS.current == OS.iOS)
    PrivacyManifestAsset(
      // Asset IDs are mandatory, and required to be unique at the moment.
      package: 'native_add',
      name: 'native_add/PrivacyInfo.xcprivacy',
      // Together with `package` defines the asset ID that this privacy manifest belongs to.
      nativeCodeAssetName: 'native_add',
      file: config.packageRoot.resolve('assets/PrivacyInfo.xcprivacy'),
    ),
]);

• pro: completely separate from NativeCodeAsset, no API pollution.
• pro: whether it's supported can be expressed with BuildConfig.supportedAssetTypes.
• con: Embedders need to first scan all assets for bundling for PrivacyManifest and NativeCodeAssets to connect the ones that belong together.
• con: verbose in the API because the PrivacyManifest needs to repeat the NativeCodeAsset asset name/id.

3. Asset.metadata

Introduce Asset.metadata List<Asset> field. (Name to be discussed) This avoids the need for a PrivacyManifestAsset.dylib field as the PrivacyManifestAsset can be nested in the dylib

output.addAssets([
  NativeCodeAsset(
    package: 'native_add',
    name: 'native_add',
    linkMode: DynamicLoadingBundled(),
    os: OS.current,
    architecture: Architecture.current,
    file: builtDylibUri,
    metadata: [
      if (OS.current == OS.macOS || OS.current == OS.iOS)
        PrivacyManifestAsset(
          // Asset IDs are mandatory, and required to be unique at the moment.
          package: 'native_add',
          name: 'native_add/PrivacyInfo.xcprivacy',
          file: config.packageRoot.resolve('assets/native_add/PrivacyInfo.xcprivacy'),
        ),
    ],
  ),
]);

• Pro: connection between dylib and privacy manifest is more obvious.
• Pro: can still be controlled by BuildConfig.supportedAssetTypes.
• Con: Asset.children is quite generic, it might look like arbitrary children are supported in arbitrary asset types.
• Con: more complexity in the API and concepts.

@mosuem I think I'm leaning towards option 2 now as well.

Class location

We could opt to not specify this asset type in package:hook/package:native_assets_cli. It would enable us to battle-test the extensibility of the protocol.

However, I would also be weary of introducing a new package for every new asset type. It would create churn specifying all the imports for users. (We can't be putting it in package:hook_flutter because once we start supporting it in Dart or another embedder it would have to be moved, and moving classes between packages is a real pain, I tried it once.)

[dcharkes]

cc @HosseinYousefi as well for discussion on how to deal with "assets-that-belong-to-other-assets" or "assets-that-augment-other-assets". ProguardRuleAssets augment JarAssets later on.

[mosuem]

Option2, but in it's own package.

I would try to keep things simple, and have a new asset type for the PrivacyManifestAsset defined in package:privacy_manifest_asset, which also exposes helper methods for creating the asset. Then all that has to be done is implement support for PrivacyManifestAsset in Flutter.

class PrivacyManifestAsset implements Asset{
  final Sting nativeCodeAssetId;
  ...
}

Asset buildPrivacyManifest(NativeCodeAsset library, PrivacyManifestOptions options) { ... }

It would create churn specifying all the imports for users.

What do you mean?

[HosseinYousefi]

In the most general form, the thing we're building could be a graph, not a tree. So using children-like API will not scale. So I like option 2 as well.

and moving classes between packages is a real pain

What if hook_flutter has a dependency on the privacy_manifest_asset but just exports it. The user will transitively depend on privacy_manifest_asset and won't have to import another thing. Later on if the same thing is available on dart standalone, hook itself can export the same class. Would this work?

[dcharkes]

Note from discussion with @mkustermann

• Privacy manifests belong to a NativeCodeAssets. So that leans towards option 1. This not only important for embedders consuming assets, but also for if you're writing a link hook helper package that links native code static libs together.

Deprioritizing because the list of required reason APIs is quite small.