generate App Intents SSU training assets during bundling

[markvasiv]

Generate App Intents SSU training assets during bundling

Problem

Apps built with rules_apple are invisible to Siri's App Intents natural-language matching. App Shortcut phrases ("Play something on ${applicationName}") are never recognized, and on iOS 26+ assistant schema requests (e.g. @AssistantIntent(schema: .media.playAudio)) are not routed to the app - Siri's planner reports the app as not supporting the intent and falls back to just launching it:

Play.Play@isSiriKitSupported: App <private> supports App Intent PlayAudioIntent: false, SiriKit INPlayMediaIntent: false

Root cause

Xcode runs two tools for App Intents, rules_apple only runs the first:

1. appintentsmetadataprocessor - produces Metadata.appintents/extract.actionsdata.
2. appintentsnltrainingprocessor - runs on the built product and produces the NL training assets: Metadata.appintents/root.ssu.yaml plus one compiled <locale>.lproj/nlu.appintents archive per localization.

Siri's utterance matching runs against the compiled nlu.appintents archives, so without them the metadata alone only feeds touch surfaces (Shortcuts app, Spotlight), not voice.

I found it by building Apple's sample project and inspecting produced archives:
https://developer.apple.com/documentation/appintents/integrating-your-music-app-with-apple-intelligence

Design

The tool must run against the assembled bundle - it reads the merged Info.plist and derives the locale set from the bundle's .lproj directories - so it can't be an ordinary analysis-time action over partial inputs. It runs in the existing post-assembly, pre-signing stage (where ipa_post_processor runs), so the generated assets are sealed into the code signature:

• app_intents_metadata_bundle partial: accepts bundle_id and requests SSU training via a new partial output field (app_intents_ssu_training).
• processor.bzl: collects the request (same pattern as signed_frameworks), builds the invocation with archive-relative paths, and threads it into both bundling modes:
  • Archive mode: new %ssu_training_command_lines% block in process_and_sign.sh.template, after the post-processor, before signing.
  • Tree-artifact mode: new ssu_training_commands control field executed by bundletool_experimental.py between post-processing and signing. It runs with the action's environment - the existing signing command lines run with env={}, which can't host xcrun.
• When the bundle has no Metadata.appintents or the toolchain lacks appintentsnltrainingprocessor (older Xcode), it does nothing. The tool can print error: Could not archive SSU artifacts while exiting 0, so output is scanned for error: like the metadata processor action already does.

For now, it's opt-in via --features=apple.app_intents_ssu_training and only supported in ios_application.

Follow-ups

• Pass localized AppShortcuts.strings via --source-file (Xcode's LM_STRINGS_FILE_PATH_LIST) for localized phrase training.
• Pass --deployment-postprocessing for release archives.
• Extensions and the other platform rules (macOS/tvOS/watchOS/visionOS) don't request training yet