project.yml

name: OpenVision
options:
  bundleIdPrefix: com.openvision
  # Bumped 16 → 18 for the on-device Gemma 4 (MLX) backend, which needs iOS 18 at runtime.
  # Revert to 16 + gate GemmaLocalService behind `if #available(iOS 18)` to keep older support.
  deploymentTarget:
    iOS: "18.0"
  xcodeVersion: "15.0"
  generateEmptyDirectories: true
  groupSortPosition: top

settings:
  base:
    MARKETING_VERSION: "2.14.0"
    CURRENT_PROJECT_VERSION: "15"
    SWIFT_VERSION: "5.9"
    DEVELOPMENT_TEAM: ${DEVELOPMENT_TEAM}
    CODE_SIGN_STYLE: Automatic
    INFOPLIST_FILE: OpenVision/Resources/Info.plist
    PRODUCT_BUNDLE_IDENTIFIER: ${PRODUCT_BUNDLE_IDENTIFIER}
    META_APP_ID: ${META_APP_ID}
    CLIENT_TOKEN: ${CLIENT_TOKEN}
    APP_LINK_URL_SCHEME: ${APP_LINK_URL_SCHEME}

configs:
  Debug:
    settings:
      DEBUG_INFORMATION_FORMAT: dwarf-with-dsym
  Release:
    settings:
      DEBUG_INFORMATION_FORMAT: dwarf-with-dsym

configFiles:
  Debug: Config.xcconfig
  Release: Config.xcconfig

targets:
  OpenVisionTests:
    type: bundle.unit-test
    platform: iOS
    sources:
      - path: OpenVisionTests
    dependencies:
      - target: OpenVision
    settings:
      base:
        GENERATE_INFOPLIST_FILE: YES

  OpenVision:
    type: application
    platform: iOS
    sources:
      - path: OpenVision
        excludes:
          - "**/*.example"
    resources:
      - path: OpenVision/Resources
        excludes:
          - "**/*.plist"
    settings:
      base:
        INFOPLIST_FILE: OpenVision/Resources/Info.plist
        PRODUCT_BUNDLE_IDENTIFIER: $(PRODUCT_BUNDLE_IDENTIFIER)
        ASSETCATALOG_COMPILER_APPICON_NAME: AppIcon
        TARGETED_DEVICE_FAMILY: "1,2"
        SUPPORTS_MACCATALYST: false
        ENABLE_PREVIEWS: true
        CODE_SIGN_ENTITLEMENTS: OpenVision/OpenVision.entitlements
        SWIFT_OBJC_BRIDGING_HEADER: OpenVision/OpenVision-Bridging-Header.h
    dependencies:
      - package: meta-wearables-dat-ios
        products:
          - MWDATCore
          - MWDATCamera
      # Acoustic voice-activity detection for end-of-turn (see SpeechActivityDetector).
      - package: FluidAudio
        products:
          - FluidAudio
      # On-device Gemma 4 (Local backend). MLX/MLXNN resolve transitively via these products.
      - package: mlx-swift-lm
        products:
          - MLXLLM           # text LLMs (Qwen 2.5, Gemma 2) via LLMModelFactory
          - MLXVLM           # vision models (Gemma 4, SmolVLM2) via VLMModelFactory
          - MLXLMCommon
          - MLXHuggingFace   # #hubDownloader() / #huggingFaceTokenizerLoader()
      # The macros above expand to HuggingFace.HubClient + Tokenizers.AutoTokenizer; mlx-swift-lm
      # deliberately doesn't depend on these, so the consumer supplies them (same as OpenGlasses).
      - package: swift-huggingface
        products:
          - HuggingFace
      - package: swift-transformers
        products:
          - Tokenizers
      # On-device neural TTS (Kokoro).
      - package: KokoroSwift
        product: KokoroSwift
    preBuildScripts:
      - name: "Check Config Files"
        script: |
          if [ ! -f "${PROJECT_DIR}/Config.xcconfig" ]; then
            echo "error: Config.xcconfig not found. Copy Config.xcconfig.example to Config.xcconfig"
            exit 1
          fi
        basedOnDependencyAnalysis: false

packages:
  meta-wearables-dat-ios:
    url: https://github.com/facebook/meta-wearables-dat-ios
    # Pinned EXACTLY — this SDK rewrites its API most releases, so never float (`from:`).
    #
    # 0.5.0 -> 0.9.0: device support gates on SDK version. "Blayzer Optics" glasses (issue #55)
    # enumerate as 0 devices on 0.5.0, which predates that hardware; support for the newer
    # frames landed in 0.6.0 (Ray-Ban Meta Optics) and 0.8.0 (Meta Glasses). 0.9.0 is the
    # camera-consolidation API: DeviceSession.addCamera(config:) -> Camera -> camera.stream
    # (StreamSession/addStream are gone). Requires iOS 17.2+.
    # MUST be regression-tested on Gen 1/2 glasses (register + stream + HFP mic) before merge.
    exactVersion: "0.9.0"
  # PrismML fork of mlx-swift, adding 1-bit quantized Metal kernels (upstream PR pending).
  # Needed to load Bonsai-8B (Qwen3-8B at 1-bit g128, 1.28 GB vs E2B's 3.6 GB). Declared at the
  # ROOT so every consumer (mlx-swift-lm + the three vendored packages) resolves the same identity.
  # Pinned to v0.31.6_prism, NOT the fork's default `prism` branch: `prism` is 10 commits behind and
  # misses "gate NAX to gen-18+ (M5/gen-17 miscompute)" — a correctness fix for current Apple GPUs.
  # Upstream base is 0.31.6 (we were on 0.31.4); Kokoro asks for >= 0.29.1, so it satisfies both.
  mlx-swift:
    url: https://github.com/PrismML-Eng/mlx-swift.git
    branch: v0.31.6_prism
  mlx-swift-lm:
    url: https://github.com/ml-explore/mlx-swift-lm.git
    # Pinned to a main commit PAST the 3.31.4 tag that fixes Gemma 4 E2B loading. 3.31.4's VLM backbone
    # built k_proj/v_proj/k_norm for ALL layers (never passed kvSharedOnly), so E2B's shared-KV layers
    # 15-34 demanded weights the checkpoint omits → keyNotFound. This commit ("Fix MLXVLM Gemma4 loader
    # to honor num_kv_shared_layers") passes kvSharedOnly per layer + guards attention on isKVSharedLayer.
    # The previous pin stopped just before #369, which needed a newer mlx-swift API
    # (greatestFiniteMagnitudeArray) than the 0.31.4 we had then.
    #
    # Bumped to 14414441 ("Fix MLXFoundationModels compilation against the Xcode 27 beta 5 SDK", #544):
    # the commit above was written against an earlier iOS 27 beta FoundationModels API, so on the
    # Xcode 27 release SDK MLXFoundationModels fails with "extraneous argument label 'capabilities:'"
    # and "'_' can only appear in a pattern". It descends from the commit above (keeps the Gemma 4
    # kvSharedOnly fix) and includes #369, which now builds since mlx-swift moved to 0.31.6 (PrismML fork).
    revision: "14414441fa44f45eee35a61e9fa0bab577cf9734"
  # CoreML Silero VAD (Apache 2.0). Gives us ACOUSTIC end-of-speech detection: the endpoint
  # countdown can start when the mic actually goes quiet instead of when SFSpeechRecognizer stops
  # emitting partials (which it does in bursts with ~1s gaps mid-sentence, so transcript timing
  # can't distinguish "still talking" from "done"). Runs on the Neural Engine — neither the GPU
  # our MLX models use nor the CPU — so it doesn't contend with on-device inference.
  FluidAudio:
    url: https://github.com/FluidInference/FluidAudio.git
    from: "0.15.5"
  swift-huggingface:
    url: https://github.com/huggingface/swift-huggingface.git
    from: "0.1.0"
  swift-transformers:
    url: https://github.com/huggingface/swift-transformers.git
    from: "1.0.0"
  # On-device neural TTS (Kokoro-82M) via MLX Swift — VENDORED under Vendor/ for full visibility
  # and reproducibility (source frozen in-repo, not pulled live). Shares mlx-swift with mlx-swift-lm.
  KokoroSwift:
    path: Vendor/kokoro-ios

schemes:
  OpenVision:
    build:
      targets:
        OpenVision: all
    run:
      config: Debug
      commandLineArguments:
        "-FIRDebugEnabled": false
    test:
      config: Debug
      gatherCoverageData: true
      targets:
        - OpenVisionTests
    profile:
      config: Release
    analyze:
      config: Debug
    archive:
      config: Release