Skip to content

Heliox OS v0.9.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 22 Jul 14:13
v0.9.0
5fd0569

Heliox OS v0.9.0

Download & Install (Auto-Setup Architect)

  1. Install Python 3.11+ on your system (if not already installed)
  2. Download the installer for your OS below
  3. Run the installer and launch Heliox OS
  4. The app will automatically build a Virtual Environment and install every shipped feature for you on first launch!
  5. Enter your API key (Gemini/OpenAI/Claude) in Settings

What's New

Voice

  • Kyutai Pocket TTS as the new default daemon-side voice — free, open-source, CPU-only, no API key, with automatic fallback to your OS's built-in voice
  • Local speech-to-text (Whisper) and text-to-speech now install by default alongside every other feature
  • Continuous VAD-based listening with barge-in: start talking and Heliox stops mid-sentence to listen, instead of talking over you
  • Daemon-side speech no longer talks over itself — in-progress speech is superseded rather than overlapping

Cognitive Engine (replaces Meta TRIBE v2)

  • Removed the Meta TRIBE v2 integration entirely: both its code (GitHub) and model weights (Hugging Face) are licensed CC-BY-NC-4.0 (non-commercial), incompatible with a commercial product
  • Replaced with a lightweight, dependency-free heuristic Cognitive Engine — zero external model, zero network download, zero license concerns — powering the same attention/stress/intent features (Attention-Aware UI, Stress-Aware Task Gating, JARVIS intent disambiguation, cognitive coaching)
  • Meaningfully more accurate than a drop-in replacement: an audit of every call site found real signal that was being silently discarded — OCR text + window titles, real keystroke/click cadence, and gaze region were all computed and thrown away. All three now actually feed the engine, alongside recency-decayed (smooth, not cliff-edged) history weighting and a fixed circular-feedback bug in the stress gate

Safety & control

  • Mid-flight cancellation: the Stop button now genuinely kills an in-flight shell command or PTY session, instead of only stopping the next step in a plan
  • Live Execution Narrator: narrates plan execution as it happens and can pre-emptively pause on a detected risk
  • Learned Risk Gate: a lightweight learned model augmenting the destructive-action safety pipeline's risk scoring
  • User Manual Supervision (opt-in): cognitive coaching and auditable risk-pattern warnings from your own screen/keyboard/mouse activity — privacy-first, nothing raw ever stored
  • Pre-execution target assessment for browser dry-runs

Autonomy & awareness

  • Autonomous Healing Engine: passively watches CPU/memory/disk and self-remediates, with tiered auto-execute vs. propose-and-confirm
  • Gaze tracking as a third fusion modality alongside voice and gesture

Gesture & workflows

  • Durable, pausable/resumable multi-step workflows triggered by voice or gesture, with user-configurable gesture-to-goal bindings
  • 3D wrist-relative gesture tracking (HandLandmarker) behind a config flag

Install & distribution

  • Real end-user installs now request every shipped feature by default instead of a bare/partial install
  • pilot-daemon is published to PyPI for the first time, so the install chain actually works end to end: pypi.org/project/pilot-daemon/0.9.0

You can view the full commit history in our Releases history.

Assets 11
Loading