What's New

• Skills fix: Auto-resolve script parameter in skills.run_script when omitted by small local models (Qwen 3.5 4B/8B). Single-script skills now work without the model needing to specify the script name.
• Memory dedup: Hardened fuzzy dedup threshold and stripped punctuation in containment checks to prevent near-duplicate entry accumulation.
• Low-memory model: Keep low-memory local model visible and honor fallback correctly.
• CI stability: Model tests are now hardware-independent.

What's New

Reliability & Stability (M.12 Phase 2)

• Fixed observer leaks — Overlay window and memory file watcher now properly clean up on deallocation
• Improved error visibility — Model download failures now show in the UI; internal errors are properly logged
• FluidVAD retry — Voice activity detection now retries after a transient failure instead of permanently falling back
• Faster long sessions — ASR audio buffer now uses O(1) circular buffer, eliminating memory spikes in extended recordings
• Cancellation fixes — Task cancellation now correctly stops in-progress operations

Architecture (M.12 Phase 3)

• Decomposed orchestration layer into focused, maintainable components
• Removed dead persistence migration code
• Centralized logging, constants, and configuration

Bug Fixes (M.12 Phase 1)

• Fixed GPU cache race condition in MLX inference
• Fixed crash-on-startup in persistence recovery
• Fixed hotkey observer leak on app termination

What's New in Ora 1.0.2

• Fixed microphone permission packaging for distribution builds by shipping the audio-input entitlement.
• Fixed release signing so notarization succeeds without get-task-allow entitlement rejection.
• Improved signing script follow-up documentation for inside-out signing hardening.
• Fixed DMG naming in release signing flow to use the app version reliably.

Requirements

• macOS 26.0 (Tahoe) or later
• Apple Silicon (M1+)
• 16GB RAM recommended

What's New in Ora 1.0.1

• Fixed Codex/OpenAI provider reliability issues across auth, model discovery, and structured response handling.
• Improved cloud model request diagnostics and guidance when a selected model rejects structured request shape.
• Stabilized onboarding model preparation, including correct reuse of already-downloaded local models.

Requirements

• macOS 26.0 (Tahoe) or later
• Apple Silicon (M1+)
• 16GB RAM recommended

Ora v1.0.0 - First Release

Privacy-first macOS voice assistant powered by local AI inference.

What's New

• ✅ Fully on-device processing (ASR, LLM, TTS)
• ✅ Calendar, Reminders, and Contacts integration
• ✅ Push-to-talk activation (⌥Space)
• ✅ Streaming pipeline for fast response
• ✅ macOS 26 (Tahoe) required

Installation

1. Download Ora-1.0.0.dmg
2. Open the DMG
3. Drag Ora.app to Applications
4. Launch and grant permissions when prompted

Requirements

• macOS: 26.0 (Tahoe) or later
• Chip: Apple Silicon (M1+)
• RAM: 16GB recommended

Note: This is a signed and notarized release - it will open without warnings.