What shipped, by version.

Current branch. The diffusion runtime finally clears four seconds end-to-end on the median Apple Silicon machine. Voice activity detection got a full rewrite.

Highlights

• +Median voice-to-deploy time on M-series: 3.8s (down from 5.2s in 0.3).
• +New lazy VAD with 40ms first-syllable latency and a 60% reduction in false wakeups.
• +Multi-cursor edits across files now stream as a single coalesced diff.
• +Gaze tracker re-trained on a 4× larger dataset; calibration time halved.
• +BYOK support for Google AI Studio and any OpenAI-compatible endpoint.

Fixes

• ·Resolved a freeze when the wakeword detector lost its lease on the audio device.
• ·Stopped sending duplicate Sentry events when a panel was opened twice in one session.
• ·Fixed Cmd-K not focusing the input on first launch after a fresh install.

First release with the iterator panel and the runtime spine split into eri_* crates. The internal team moved to dogfooding full-time on this build.

Highlights

• +New iterator panel for stepping through long-running agent runs.
• +Runtime spine refactored into five eri_* crates with clean boundaries.
• +Cmd-K can now run any registered action across the editor and the agent.
• +Local Llama backend wired up behind a feature flag for early testers.

Fixes

• ·Reduced startup memory by ~180 MB by deferring TTS warm-up.
• ·Fixed gaze calibration drift after long sessions with external monitors.
• ·Made the changelog feed survive a malformed entry instead of crashing the panel.

The first build we let strangers run. Onboarding now finishes in under sixty seconds and the editor reaches steady state on first launch.

Highlights

• +End-to-end onboarding flow with mic, gaze, and project import wired together.
• +The eri Engine now drives all multi-line edits by default.
• +Editor surface holds 120 fps through full-file diffusion on the reference machine.
• +Auto-update through the in-app channel switcher.

Fixes

• ·Closed a race where the voice pipeline could deadlock during a fast wake/sleep cycle.
• ·Stopped the agent from re-reading already-resolved errors in the same iteration.

Internal-only alpha. First end-to-end voice → diffusion → diff loop running on a single laptop.

Highlights

• +First working voice-to-diff pipeline.
• +Minimal editor surface forked from VSCodium — keybindings and extensions carry over.
• +Hand-rolled wakeword and TTS pipeline running entirely on-device.