feat: audio recording foundation - types, deps, IPC setup

- Add Recording types and IPC API definitions
- Install Vercel AI SDK dependencies
- Add Recordings tab to UI
- Enable media capture permissions
- Stub IPC handlers (implementation in next PR)

Author: lucasheriques

docs/prd-audio-transcription.md

@@ -0,0 +1,397 @@
+# PRD: Audio Transcription & Task Extraction
+
+## Overview
+
+Enable users to record audio (calls, meetings, voice notes), automatically transcribe using AI, and extract actionable tasks that can be created directly in PostHog.
+
+## Problem
+
+Users have valuable conversations (calls, meetings, brainstorms) where tasks and action items are discussed, but:
+- Manual note-taking is incomplete and distracting
+- Context is lost between conversation and task creation
+- Transcribing and extracting tasks manually is time-consuming
+- No seamless way to turn conversation → actionable tasks
+
+## Solution
+
+Browser-based audio recording with AI-powered transcription and intelligent task extraction:
+1. **Record** audio directly in the app
+2. **Transcribe** using OpenAI Whisper via Vercel AI SDK
+3. **Generate** concise summary of the conversation
+4. **Extract** actionable tasks automatically using GPT
+5. **Create** tasks in PostHog with one click
+
+## V1 Requirements (Ship This Week)
+
+### Functional Requirements
+
+#### 1. Audio Recording
+- ✅ Record audio from system/microphone via browser MediaRecorder API
+- ✅ Real-time duration counter during recording
+- ✅ Start/stop recording controls
+- ✅ Save recordings to local storage (`userData/recordings/`)
+- ✅ Audio format: WebM (cross-platform, compressed)
+
+#### 2. Recording Management
+- ✅ List all recordings with metadata (date, duration)
+- ✅ Play back recordings in-app
+- ✅ Delete recordings
+- ✅ Persist metadata (duration, created_at, transcription status)
+
+#### 3. Transcription
+- ✅ Transcribe recording using OpenAI Whisper (via Vercel AI SDK)
+- ✅ Show transcription status (processing/completed/error)
+- ✅ Display full transcript text
+- ✅ Generate 3-7 word summary title
+- ✅ Handle transcription errors gracefully
+
+#### 4. Task Extraction
+- ✅ Automatically extract actionable tasks from transcript
+- ✅ Use GPT-4o-mini for task identification
+- ✅ Extract: feature requests, bug reports, requirements, action items
+- ✅ Format: `{title: string, description: string}` with context
+
+#### 5. Task Creation
+- ✅ Bulk create tasks in PostHog from extracted tasks
+- ✅ One-click "Create All Tasks" button
+- ✅ Link transcript context to task descriptions
+- ✅ Success/error feedback
+
+### Non-Functional Requirements
+
+- **Performance**: Transcription completes within 2x audio duration
+- **Storage**: Audio files compressed to ~1MB per 10 minutes
+- **Reliability**: Handle network failures, large files (25MB Whisper limit)
+- **Privacy**: All recordings stored locally, API keys encrypted
+- **UX**: Clear status indicators, error messages, loading states
+
+### User Stories
+
+**As a user, I want to:**
+1. Record a call/meeting without leaving the app
+2. Get an automatic transcript without manual typing
+3. See a summary of what was discussed
+4. Get a list of action items extracted automatically
+5. Create PostHog tasks from those action items with one click
+6. Access my recording history and transcripts
+
+### UI/UX
+
+**Recordings Tab (new)**
+- Main view accessible from tab bar
+- "Record" button (prominent, red when recording)
+- Real-time duration counter during recording
+- List of recordings (newest first)
+- Each recording card shows:
+  - Date/time
+  - Duration
+  - Play button
+  - Transcribe button (if not transcribed)
+  - Delete button
+  - Transcript (if available)
+  - Summary (if available)
+  - Extracted tasks (if available)
+  - "Create Tasks" button (if tasks extracted)
+
+**Design System**
+- Use Radix UI components (Button, Card, Flex, Text, etc.)
+- Follow existing app styling patterns
+- Responsive layout
+- Keyboard shortcuts (future)
+
+## Technical Architecture
+
+### Core Technologies
+
+**Vercel AI SDK**
+- **Why**: Provider-agnostic abstraction, makes swapping providers trivial
+- **Usage**: OpenAI provider for Whisper transcription and GPT task extraction
+- **Future**: Swap to Anthropic, AssemblyAI, or local models by changing config
+
+**Browser MediaRecorder API**
+- **Why**: No native dependencies, cross-platform, works today
+- **Format**: WebM with Opus audio codec
+- **Fallback**: Error handling for unsupported browsers
+
+**React Query**
+- **Why**: Handles data fetching, caching, optimistic updates
+- **Usage**: Recording list, transcription status polling
+- **Benefit**: No new Zustand store needed
+
+### Architecture Diagram
+
+```
+┌─────────────────────────────────────────────────┐
+│                  Renderer Process                │
+│                                                   │
+│  RecordingsView.tsx                              │
+│    ↓                                             │
+│  useRecordings.ts (React Query)                  │
+│    ↓                                             │
+│  window.electronAPI.recording*()                 │
+└──────────────────┬──────────────────────────────┘
+                   │ IPC
+┌──────────────────┴──────────────────────────────┐
+│                  Main Process                     │
+│                                                   │
+│  recording.ts (Service)                          │
+│    ├─ MediaRecorder handling                     │
+│    ├─ File system storage                        │
+│    ├─ Vercel AI SDK                              │
+│    │   ├─ OpenAI Whisper (transcription)         │
+│    │   └─ GPT-4o-mini (summary + extraction)     │
+│    └─ IPC handlers                               │
+│                                                   │
+│  transcription-prompts.ts                        │
+│    ├─ SUMMARY_PROMPT                             │
+│    └─ TASK_EXTRACTION_PROMPT                     │
+└───────────────────────────────────────────────────┘
+```
+
+### File Structure
+
+**New Files:**
+```
+src/
+├── main/
+│   └── services/
+│       ├── recording.ts                 # Core recording service
+│       └── transcription-prompts.ts     # Prompt configuration
+└── renderer/
+    ├── components/
+    │   └── RecordingsView.tsx           # Main UI
+    └── hooks/
+        └── useRecordings.ts             # Recording state hook
+```
+
+**Modified Files:**
+```
+src/
+├── main/
+│   ├── index.ts                         # Register IPC
+│   └── preload.ts                       # Expose API
+├── renderer/
+│   └── components/
+│       └── MainLayout.tsx               # Add tab
+└── shared/
+    └── types.ts                         # Recording types
+```
+
+### Data Models
+
+```typescript
+interface Recording {
+  id: string                   // Filename
+  filename: string
+  duration: number             // Seconds
+  created_at: string          // ISO 8601
+  file_path: string           // Absolute path
+  transcription?: {
+    status: 'processing' | 'completed' | 'error'
+    text: string
+    summary?: string
+    extracted_tasks?: Array<{
+      title: string
+      description: string
+    }>
+    error?: string
+  }
+}
+```
+
+### IPC API
+
+```typescript
+// Recording lifecycle
+window.electronAPI.recordingStart()
+  → Promise<{recordingId: string, startTime: string}>
+
+window.electronAPI.recordingStop(recordingId, audioData, duration)
+  → Promise<Recording>
+
+// Recording management
+window.electronAPI.recordingList()
+  → Promise<Recording[]>
+
+window.electronAPI.recordingDelete(recordingId)
+  → Promise<boolean>
+
+window.electronAPI.recordingGetFile(recordingId)
+  → Promise<ArrayBuffer>
+
+// Transcription
+window.electronAPI.recordingTranscribe(recordingId, openaiApiKey)
+  → Promise<TranscriptionResult>
+```
+
+### Key Technical Decisions
+
+| Decision | Rationale |
+|----------|-----------|
+| **Vercel AI SDK** | Provider-agnostic, production-ready, easy provider swaps |
+| **Browser Recording** | No native deps, cross-platform, works today |
+| **Prompts in Code** | Ship fast, easy to move to editable file later |
+| **React Query** | Standard data fetching, no new state layer |
+| **Local Storage** | Privacy-first, no upload costs, instant access |
+| **WebM Format** | Cross-platform, compressed, MediaRecorder default |
+
+## V1 Out of Scope
+
+❌ Meeting auto-detection (Zoom/Meet)
+❌ Multiple transcription provider UI
+❌ User-editable prompts UI
+❌ Real-time streaming transcription
+❌ Recording-only mode (no PostHog)
+❌ Cloud storage/sync
+❌ Speaker diarization
+❌ Custom vocabulary/terminology
+
+## Future Improvements (Priority Order)
+
+### P0: Based on User Feedback
+1. **Meeting Auto-Detection**
+   - Detect active Zoom/Google Meet sessions
+   - Show notification to start recording
+   - Auto-name recordings by meeting title
+   - *Build when*: Users manually start recordings often
+
+2. **User-Editable Prompts**
+   - Move prompts to `~/.array/prompts.json`
+   - UI to edit summary and extraction prompts
+   - Prompt templates (standup, brainstorm, bug triage)
+   - *Build when*: Users want custom task extraction logic
+
+### P1: Enhanced Functionality
+3. **Provider Selection**
+   - UI to choose transcription provider
+   - Support: OpenAI, Anthropic, AssemblyAI, local Whisper
+   - Provider-specific settings (language, model)
+   - *Build when*: Users request specific providers or cost concerns
+
+4. **Real-Time Streaming**
+   - Live transcription while recording
+   - Show transcript building in real-time
+   - Early task extraction
+   - *Build when*: Users record long sessions (>30 min)
+
+5. **Recording-Only Mode**
+   - Use Array without PostHog (pure recorder)
+   - Export tasks to Markdown/CSV
+   - Standalone app mode
+   - *Build when*: Users want tool without PostHog
+
+### P2: Power User Features
+6. **Advanced Editing**
+   - Edit transcripts before task creation
+   - Manually add/remove extracted tasks
+   - Timestamp navigation
+   - *Build when*: Transcription accuracy issues reported
+
+7. **Organization & Search**
+   - Tag recordings
+   - Search transcripts
+   - Folders/projects
+   - Export transcripts
+   - *Build when*: Users have 50+ recordings
+
+8. **Team Features**
+   - Share recordings with team
+   - Shared prompt templates
+   - Team transcription usage/costs
+   - *Build when*: Teams using Array together
+
+9. **Integration Enhancements**
+   - Auto-tag tasks by meeting type
+   - Link to calendar events
+   - Attach to GitHub issues
+   - Slack notifications
+   - *Build when*: Integration requests from users
+
+## Success Metrics
+
+### V1 Goals (First Month)
+- **Adoption**: 30% of active users record at least 1 audio
+- **Usage**: Average 3 recordings per user per week
+- **Conversion**: 70% of transcriptions → tasks created
+- **NPS**: +40 from users who transcribe recordings
+
+### Key Metrics to Track
+- Recordings created per user
+- Transcription completion rate
+- Tasks created from transcriptions
+- Time saved (estimated)
+- Feature request themes
+- Error rates (transcription failures, API issues)
+
+### User Feedback Questions
+1. How often do you record calls/meetings?
+2. Is the transcription accurate enough?
+3. Are extracted tasks relevant and actionable?
+4. What other providers would you want?
+5. What features are missing?
+
+## Open Questions
+
+### Pre-Launch
+- [ ] Should we require OpenAI API key upfront or use Array's key with usage limits?
+- [ ] How do we handle 25MB Whisper file size limit? (error message vs auto-chunking)
+- [ ] Should recordings auto-transcribe or wait for user action? (cost concern)
+
+### Post-Launch (answer based on data)
+- [ ] Do users want meeting auto-detection?
+- [ ] Do users record mostly short (<10 min) or long (>30 min) sessions?
+- [ ] Is OpenAI Whisper accuracy good enough or do we need alternatives?
+- [ ] Do users want to edit prompts or are defaults sufficient?
+- [ ] Should we add speaker diarization (who said what)?
+
+## Launch Plan
+
+### Phase 1: Internal Testing (Week 1)
+- Build and test locally
+- Record real meetings
+- Validate transcription quality
+- Iterate on UI/UX
+
+### Phase 2: Beta Users (Week 2)
+- Ship to 10-20 beta users
+- Gather feedback on accuracy and usefulness
+- Monitor error rates and API costs
+- Fix critical bugs
+
+### Phase 3: General Availability (Week 3)
+- Announce in app
+- Update docs
+- Monitor metrics
+- Plan next iteration based on feedback
+
+## Risk Mitigation
+
+| Risk | Mitigation |
+|------|------------|
+| **Poor transcription accuracy** | Use OpenAI Whisper (best in class), allow editing |
+| **High API costs** | Manual transcription trigger, warn on large files |
+| **Browser recording fails** | Clear error messages, fallback instructions |
+| **Large file handling** | Show 25MB limit error, suggest shorter recordings |
+| **Privacy concerns** | Local storage, encrypted keys, clear data policy |
+| **Low adoption** | User education, in-app tutorial, example use cases |
+
+## Appendix
+
+### Competitive Analysis
+- **Otter.ai**: Strong transcription, expensive, no task extraction
+- **Fireflies.ai**: Meeting bot, not local, privacy concerns
+- **Notion AI**: Meeting notes, requires Notion, no direct recording
+- **Array advantage**: Integrated with task management, local-first, extensible
+
+### Research & References
+- [Vercel AI SDK Docs](https://sdk.vercel.ai/docs)
+- [OpenAI Whisper API](https://platform.openai.com/docs/guides/speech-to-text)
+- [MDN MediaRecorder API](https://developer.mozilla.org/en-US/docs/Web/API/MediaRecorder)
+- [Electron safeStorage](https://www.electronjs.org/docs/latest/api/safe-storage)
+
+---
+
+**Document Status**: Draft v1.0
+**Last Updated**: 2025-10-15
+**Owner**: Product / Engineering
+**Next Review**: Post-launch + 2 weeks