Add iOS app with keyboard extension

[croshank]

iOS app + keyboard extension that brings MiniFlow dictation to any iOS app.

The macOS app routes audio through a Python backend to Smallest AI. On iOS there's no local server, so the app talks to the Waves WebSocket API directly from Swift — same 16kHz PCM streaming, same is_final/is_last protocol, just no middleman.

How it works

The main app runs AVAudioEngine in the background. The keyboard extension sends commands (start/stop/cancel) via files in the App Group shared container. When the user taps record, the app opens a WebSocket to wss://api.smallest.ai/waves/v1/pulse/get_text, streams PCM chunks from the mic, and writes the transcript back to the shared container. The keyboard reads it and inserts at the cursor.

File-based IPC instead of UserDefaults because UserDefaults is unreliable between app and extension on device (the kCFPreferencesAnyUser issue).

What's in the PR

• SmallestAIClient.swift — actor wrapping URLSessionWebSocketTask, handles the Waves protocol
• FlowBackgroundRecorder.swift — AVAudioEngine capture, PCM conversion, streams to SmallestAIClient
• FlowSessionManager.swift — reads/writes small files in the App Group container for IPC
• KeyboardViewController.swift — UIInputViewController hosting SwiftUI, state machine for the recording flow
• VoiceInputView.swift — keyboard UI
• ContentView.swift — main app: API key entry, session control, setup instructions
• KeychainHelper.swift — thin Security.framework wrapper for storing the API key

Setup

After cloning, add App Group group.com.smallestai.MiniFlow to both targets in Signing & Capabilities (Xcode doesn't persist this across machines). Then build, enter your API key, enable the keyboard in iOS Settings, and go.

No macOS files were modified.

[entelligence-ai-pr-reviews]

EntelligenceAI PR Summary

Introduces the full MiniFlow iOS application — a voice transcription tool using a custom keyboard extension and background audio recording service connected to Smallest AI's STT WebSocket API.

• Main App: SwiftUI entry point (MiniFlow_iOSApp) with deep-link handling (miniflow://), ContentView for session/API key/microphone management, and FlowBackgroundRecorder for end-to-end audio capture and streaming
• STT Client: SmallestAIClient Swift actor for real-time WebSocket-based speech-to-text with ping keepalive and transcript joining
• IPC: FlowSessionManager implements file-based inter-process communication over shared App Group (group.com.smallestai.MiniFlow) replacing UserDefaults-based cross-process signaling
• Keyboard Extension: MiniFlowKeyboard with KeyboardViewController, KeyboardViewModel state machine, VoiceInputView SwiftUI UI (record button, waveform animation, status pill), and text insertion/undo via UITextDocumentProxy
• Security: KeychainHelper for secure API key storage; App Group entitlements on both main app and extension targets
• Project Setup: Full Xcode project (project.pbxproj) targeting iOS 26.0, four build targets, shared schemes, asset catalogs, Info.plist files, and test stubs

---

Confidence Score: 3/5 - Review Recommended

Likely safe but review recommended — this PR introduces a substantial new iOS application including FlowBackgroundRecorder, SmallestAIClient, and a custom keyboard extension, which represents a significant surface area for a first review pass. While the automated review found no flagged issues across all 22 changed files, the complexity of the WebSocket-based STT streaming in SmallestAIClient, the audio session lifecycle management in FlowBackgroundRecorder, and the cross-process communication between the keyboard extension and main app warrant human eyes before merging. The PR achieves a meaningful milestone — end-to-end voice transcription via a custom keyboard — but microphone permissions, background audio entitlements, API key storage, and WebSocket error recovery paths are all areas where subtle bugs may not surface in automated analysis.

Key Findings:

• FlowBackgroundRecorder manages audio session lifecycle and background recording state — improper AVAudioSession category handling or failure to handle interruptions (phone calls, Siri) could cause silent recording failures or crashes in production that automated tooling would not catch.
• SmallestAIClient is a Swift actor handling real-time WebSocket streaming; WebSocket reconnection logic, partial message reassembly, and error propagation paths under poor network conditions are non-trivial and deserve manual review even if no static issues were flagged.
• The keyboard extension communicates with the main app (likely via shared App Group container or URL schemes given the miniflow:// deep-link), and any mismatch in entitlements or group identifiers between targets would cause silent failures only visible at runtime on a real device.
• API key management — how SmallestAIClient receives and stores the Smallest AI API key — should be verified to use Keychain rather than UserDefaults or plaintext storage, a security concern that static analysis typically does not surface.

Files requiring special attention

• FlowBackgroundRecorder.swift
• SmallestAIClient.swift
• KeyboardViewController.swift
• MiniFlow_iOSApp.swift
• ContentView.swift