# YazSes v2.0.0-dev.1 — Voice-First Interaction Layer (developer preview)

**Pre-release · 2026-07-02 · not published to PyPI/Snap/APT** (the publish workflows
are pinned to the `v0.x`/`v1.x` lines; this dev tag intentionally does not ship to
users). **v1.4.1 remains the stable release.**

This tag opens the **v2 line**: YazSes grows from an offline dictation daemon into a
broader **voice-first human–computer interaction layer** — still 100% on-device,
privacy-first, and with **every new feature off by default**. It is grounded in a
five-domain state-of-the-art research sweep and specified across nine ADRs.

## What's in dev.1

**Foundations (design):**
- 5-domain SoA research (voice HCI, on-device agentic OS, AR/wearables, accessibility,
  ambient AI) → a synthesis of **13 features** (10 flagship + 3 experimental).
- **9 ADRs** (`adr-v2-000`…`adr-v2-009`), a roadmap section, and an architecture/
  integration map. Every hook is config-gated so v1 behavior is unchanged when off.

**Wave A implementation (all off by default):**
- **Confidence Ink** core — flags low-confidence words from Whisper's token
  probabilities, groups spans, and re-picks from n-best by voice. Word probabilities
  are now plumbed through the decode path. (ADR-v2-001)
- **Prosody pause→sentence** — opt-in period insertion on sentence-length pauses,
  extending Prosody Ink. (ADR-v2-002)
- **Spoken Edit Mode** core — open-ended voice edits ("change X to Y", "delete the last
  sentence") with a destructive-op confirm set. (ADR-v2-003)
- **Context-Primed Dictation** — **wired**: best-effort, transient reads of the active
  window / selection / clipboard fold salient terms into the STT prompt; fully guarded
  so it can never break dictation. (ADR-v2-004)

New config sections (`[confidence]`, `[context]`, `commands.spoken_edit`,
`prosody.pause_sentence_ms`) and `yazses features` entries — all default off.

## Quality

805 tests green across Linux × macOS × Windows on Python 3.11 and 3.12.

## What's next (toward v2.0.0)

Wire Confidence Ink into the overlay + voice re-pick; wire Spoken Edit into command
dispatch; then Wave B (Spoken Recall, offline Spoken MCP, AT-SPI Voice Pilot, True
Code-Switch, Personal Speech Adapter) and Wave C (Gaze-Routed, sEMG/Modality Router,
Accessibility Continuum, Glasses↔Desktop bridge). When Wave A is fully wired and the
publish workflows are extended to the `v2.x` line, v2.0.0 will ship to all channels.