Skip to content

ADR-v2-000 — YazSes v2: the Voice-First Interaction Layer (umbrella)

Status: Accepted (2026-07-02) Context links: [[adr-011]] (zero telemetry / offline), [[adr-012-self-improvement-loop]] (encrypted corpus), [[adr-013-llm-cleanup]] (guarded text-transform philosophy), [[adr-014-tune-holdout-validation]] (held-out promotion gate), [[adr-v04-001-slm-inference]] (SLM tier), [[adr-v04-002-lsp-context]] (editor context)

Context

v1.4.1 is a stable, well-adopted offline dictation + voice-command daemon. The 2026-07 state-of-the-art sweep (an internal design note, five evidence-anchored domains) found that the industry shipped all the ingredients for a richer voice-first interaction layer — on-device SLMs (Apple Foundation Models, Phi Silica, Gemini Nano), the MCP tool standard, consumer sEMG/gaze wearables (Meta Neural Band, Vision Pro), atypical-speech ASR (Speech Accessibility Project), and "draft-then-polish" dictation (Fluid Dictation, Gboard Rambler) — but tied them to the cloud or a walled garden. Microsoft Recall's backlash and the Rewind→Limitless→Meta collapse show the market is retreating toward "provably never leaves your machine," which is exactly YazSes's posture.

The synthesis (an internal design note) defines 13 features (10 flagship + 3 experimental) that are novel, on-device, and genuinely useful.

Decision

Open the v2.0.0 line as an additive interaction layer over the v1 pipeline. Governing invariants (binding on every v2 feature):

  1. On-device only — no cloud, no API key, no telemetry (extends ADR-011).
  2. Off by default & dependency-isolated — pure logic is dependency-free; heavy backends import lazily behind optional extras; each feature is a yazses features capability. Experimental features refuse enable without --force.
  3. Consent-first data — captured content lives only in the encrypted machine-bound corpus (ADR-012); redaction + retention honored; biometrics never plaintext.
  4. Human-in-the-loop for side effects — any state-changing action needs spoken/visual confirmation. This also engineers around the documented sub-50-step agent-reliability cliff (OSWorld-Human): keep voice actions short, verifiable, confirmable.
  5. Cross-platform — Linux-first where OS APIs differ, but no feature is macOS-only.

Delivery in three waves, each its own ADR(s): - Wave A (P0, current stack): Confidence Ink ([[adr-v2-001-confidence-ink]]), Prosody Auto-Format ([[adr-v2-002-prosody-autoformat]]), Spoken Edit Mode ([[adr-v2-003-spoken-edit-mode]]), Context-Primed Dictation ([[adr-v2-004-context-primed-dictation]]). - Wave B (P1): Spoken Recall, Voice-to-Tool (Spoken MCP), AT-SPI Voice Pilot, True Code-Switch, Personal Speech Adapter. - Wave C (EXP): Gaze-Routed Dictation, sEMG Command Layer + Modality Router, Accessibility Continuum, Glasses↔Desktop Bridge.

The v2.0.0-dev tag is cut only when Wave A code + tests land, so a tag always points at working code (the current well-working v1.4.1 is the preserved fallback baseline).

Consequences

  • + A defensible, privacy-first position no cloud vendor can match; strong job-showcase narrative (research-grounded, ADR-disciplined, incremental).
  • + Reuses existing modules (slm_router, corpus, gaze/zones, polyglot/lid, LSP context, overlay, tts, remote/agent, personalize) — most features are integration, not greenfield.
  • − Surface-area growth and optional-dependency matrix complexity → mitigated by the off-by-default + features registry discipline and dependency isolation.
  • − Some features (Personal Adapter, AT-SPI on Wayland, webcam gaze accuracy) carry real feasibility risk → each ADR states a fallback and, where relevant, a measurement gate before shipping.