# Changelog
## [2.0.5] - 2026-03-15
### Improvements
- **Stealth Mode UI**: The Process Disguise selector is now visually disabled and locked while Undetectable mode is active, preventing accidental state mismatches.
- **State Synchronization**: Greatly improved internal state synchronization across all application windows (Settings, Launcher, Overlay).
### Fixes
- **Infinite Feedback Loops**: Completely eliminated the bug where toggling Undetectable mode would sometimes cause the app to rapidly toggle itself on and off.
- **Delayed Dock Reappearance**: Fixed a regression where the macOS dock icon would mysteriously reappear several seconds after entering stealth mode if a disguise had recently been changed.
- **Initial State Loading**: Fixed an issue where the Settings UI would briefly show incorrect toggle states when first opened.
- **macOS OS-level Events**: Hardened the app against macOS `activate` events (like clicking the app in Finder) accidentally breaking stealth mode.
### Technical
- Refactored IPC (Inter-Process Communication) listeners for `SettingsPopup` and `SettingsOverlay` to use a strict one-way (receive-only) data binding pattern.
- Added strict management and cancellation of `forceUpdate` timeouts during stealth mode transitions.
- Added explicit type safety for the new getters in `electron.d.ts`.
## [2.0.4] - 2026-03-14
### Summary
Version 2.0.4 introduces a massive architectural overhaul to the native audio pipeline, guaranteeing production-ready stability, true zero-allocation data transfer, and instantaneous STT responsiveness with WebRTC ML-based VAD.
### What's New
- **Two-Stage Silence Processing**: Replaced basic RMS noise gating with a two-stage pipeline combining an adaptive RMS threshold and WebRTC Machine Learning VAD. Rejects typing, fan noise, and non-speech sounds before they bill STT APIs.
- **Zero-Copy ABI Transfers**: Transitioned the `ThreadsafeFunction` bridging to direct `napi::Buffer` (Uint8Array) allocations, completely eliminating V8 garbage collection pressure during continuous capture.
- **Sliding-Window RAG**: Implemented a 50-token semantic overlap in `SemanticChunker.ts` to prevent conversational context loss across chunk boundaries.
### Improvements
- **Latency & Responsiveness Tuning**: Stripped redundant TS debouncing, slashed `MIN_BUFFER_BYTES`, and reduced native hangover, achieving a ~300ms reduction in end-to-end transcription latency. short utterances ("Yes", "Stop") no longer sit trapped in the buffer.
- Removed floating-point division truncation for superior downsampling from 44.1kHz external microphones.
### Fixes
- Fixed a critical bug where the native Rust monitor returned a hardcoded `16000Hz` while actually streaming 48kHz audio. Now syncs true hardware sample rates.
- Resolved the "Input missing" silent crash bug on microphone restarts by properly recreating the CPAL stream.
- Restored the 10s continuous speech backstop for REST APIs to prevent unbounded buffer growth.
- Added missing `notifySpeechEnded()` properties and cleaned up dangerous type casts.
### Technical
- Audio processing transitioned entirely to strict ABI memory bridging (`napi::Buffer`)
- Re-architected native silence_suppression state machine around WebRTC VAD inputs.
## [2.0.3] - 2026-03-13
### What's New
- **Dynamic AI Model Selection:** Replaced static model lists with dynamic dropdowns. Your preferred models synced from providers (like OpenAI, Anthropic, Google) now automatically appear across the entire app.
- **Multimodal Resilience:** Added a "Smart Dynamic Fallback" using Groq Llama 4 Scout. If default vision models fail or get rate-limited during screen analysis, Natively instantly reroutes the image to ensure uninterrupted performance.
- **Multiple Screenshot Support:** The Natively Interface can now handle and process multiple attached screenshots simultaneously instead of just one.
- **Improved Settings UX:** API keys now auto-save after 5 seconds of inactivity, and selecting a preferred model immediately updates the rest of the application without requiring a page reload.
### Architecture & Fixes
- **Better Embeddings:** Migrated from Gemini Embedding to a completely new and more robust embedding architecture.
- **Claude Fixes:** Resolved max_tokens and context limits issues specific to Anthropic Claude interactions.
- **DRY Refactoring:** Centralized model configuration strings across the codebase to ensure easier future updates.
## [2.0.2] - 2026-03-10
### Summary
v2.0.2 focuses on fixing Windows system audio capture, improving RAG stability, and resolving critical Soniox STT configuration issues.
### What's New
- Fully functional system audio capture for Windows
- Introduced system for manual transcript finalization and interim/final bridging during recordings
### Improvements
- Migrated to `app.getAppPath()` for reliable cross-platform resource discovery
- Ensured `sqlite-vec` compatibility and fixed embedding queue management
- Upgraded `@google/genai` and optimized embedding dimensionality for lower latency
### Fixes
- Improved Soniox STT streaming reliability, manual flushing, and configuration persistence
- Resolved application entry point and module resolution issues in production builds
- Fixed transcript bridging for manual recording mode
- Corrected stealth activation and window focus inconsistencies
### Technical
- Dependency updates for `@google/genai`
- Cleaned up native compiler warnings for Windows
- Fixed module resolution for internal Electron paths
## [2.0.1] - 2026-03-06
### New Features
- **Premium Profile Intelligence**: Job Description (JD) and Resume context awareness, company research, and negotiation assistance.
- **Live Meeting RAG**: Instant intelligent retrieval of context directly during a live meeting using local vectors.
- **Soniox Speech Provider**: Added support for ultra-fast and highly accurate streaming STT with Soniox.
- **Multilingual Support**: Choose from various response languages, set speech recognition matching specific accents and dialects.
### Improvements & Fixes
- Fixed numerous issues and merged 3 community pull requests to improve overall stability.
## [1.1.8] - 2026-02-23
### Summary
Patch update addressing OpenAI GPT 5.x compatibility and increasing token output limits for all providers.
### What's New
- Replaced deprecated `max_tokens` parameter with `max_completion_tokens` required by GPT 5.x models.
- Increased max output tokens for OpenAI (GPT 5.2) and Claude (Sonnet 4.5) to 65,536.
- Increased max output tokens for Groq (Llama 3.3 70B) to 32,768.
### Improvements
- Improved response length capabilities across all text-generation AI models.
- Updated connection test model to use `gpt-5.2-chat-latest` instead of the deprecated `gpt-3.5-turbo`.
### Fixes
- Fixed 400 error when using OpenAI GPT 5.x models for text queries and toggle actions.
### Technical
- Replaced `max_tokens` with `max_completion_tokens` in `LLMHelper.ts` and `ipcHandlers.ts`.
## [1.1.7] - 2026-02-20
### Summary
Security hardening, memory optimization, and stability improvements for a more robust and reliable experience.
### What's New
- API rate limiting to prevent 429 errors on free-tier plans (Gemini, Groq, OpenAI, Claude)
- Cross-platform screenshot support (macOS, Linux, Windows)
- Official website link added to the About section
### Improvements
- Smarter transcript memory management with epoch summarization instead of hard truncation — no more losing early meeting context
- API keys are now scrubbed from memory on app quit to minimize exposure window
- Credentials manager now overwrites key data before disposal for enhanced security
- Helper process renaming for improved stealth in Activity Monitor
### Fixes
- Fixed V8/Electron entitlements crash on Intel Macs by including entitlements.mac.plist during ad-hoc signing
- Fixed process disguise not applying correctly when undetectable mode is toggled on
- Fixed usage array capping with dedicated helper method to prevent unbounded growth
### Technical
- Added `RateLimiter` service (token bucket algorithm with configurable burst and refill rates)
- Added `PRIVACY.md` and `SECURITY.md` policy documents
- Refactored ad-hoc signing script with helper renaming and proper entitlements flow
- Version bump to 1.1.7
## [1.1.6] - 2026-02-15
### New Features
- **Speech Providers**: Added support for multiple speech providers including Google, Groq, OpenAI, Deepgram, ElevenLabs, Azure, and IBM Watson.
- **Fast Response Mode**: Introduced ultra-fast text responses using Groq Llama 3.
- **Local RAG & Memory**: Full offline vector retrieval for past meetings using SQLite.
- **Custom Key Bindings**: Added ability to customize global shortcuts for easier control.
- **Stealth Mode Improvements**: Enhanced disguise modes (Terminal, Settings, Activity Monitor) for better privacy.
- **Markdown Support**: Improved Markdown rendering in the Usage section for better readability of AI responses.
- **Image Processing**: Integrated `sharp` for optimized image handling and faster analysis.
### Improvements & Fixes
- Fixed various UI bugs and focus stealing issues.
- Improved application stability and performance.
## [1.1.5] - 2026-02-13
### Summary
The Stealth & Intelligence Update: Enhances stealth capabilities, expands AI provider support, and improves local AI integration.
### What's New
- **Native Speech Provider Support:** Added Deepgram, Groq, and OpenAI speech providers.
- **Custom LLM Providers:** Connect to any OpenAI-compatible API including OpenRouter and DeepSeek.
- **Smart Local AI:** Auto-detection of available Ollama models for local AI.
- **Global Spotlight Search:** Toggle chat overlay with Cmd+K (macOS) and Ctrl+K (Windows/Linux).
- **Masquerading Mode:** Appear as system processes like Terminal or Activity Monitor.
- **Improved Stealth Mode:** Enhanced activation and window focus transitions.
### Improvements
- **Natural Responses:** Updated system prompts for more concise and natural responses.
- **Conversational Logic:** Reduced robotic preambles and unnecessary explanations.
- **Performance:** Improved UI scaling and reduced speech-to-text latency.
### Fixes
- No critical fixes reported in this release.
### Technical
- Internal logic refinements for improved conversational flow.
- Updater and background process stability improvements.
#### macOS Installation (Unsigned Build)
If you see "App is damaged":
1. Move the app to your Applications folder.
2. Open Terminal and run: `xattr -cr /Applications/Natively.app`
## [1.1.4] - 2026-02-12
### What's New in v1.1.4
- **Custom LLM Providers:** Connect to any OpenAI-compatible API (OpenRouter, DeepSeek, commercial endpoints) simply by pasting a cURL command.
- **Smart Local AI:** Enhanced Ollama integration that automatically detects and lists your available local models—no configuration required.
- **Refined Human Persona:** Major updates to system prompts (`prompts.ts`) to ensure responses are concise, conversational, and indistinguishable from a real candidate.
- **Anti-Chatbot Logic:** Specific negative constraints to prevent "AI-like" lectures, distinct "robot" preambles, and over-explanation.
- **Global Spotlight Search:** Access AI chat instantly with `Cmd+K` / `Ctrl+K`.
- **Masquerading (Undetectable Mode):** Stealth capability to disguise the app as common utility processes (Terminal, Activity Monitor) for discreet usage.