ADR-v2-010 — Gaze-Routed Dictation & Point-and-Speak¶

Status: Accepted (2026-07-02) · Wave C (experimental) Context links: [[adr-v2-000-interaction-layer]], [[adr-v2-004-context-primed-dictation]] (deixis), [[adr-011]] (frames in-RAM only)

Context¶

The AR/wearables research (internal) shows gaze is the fastest desktop pointing channel, and spatial OSes (visionOS) route input by gaze — but only inside their own UI. YazSes already has a gaze/ package (zones.py point→zone, l2cs.py webcam backend, 9-point calibration) built for Glance-Type, currently mapping gaze to a coarse screen zone. No privacy-first local daemon routes dictation output by gaze across arbitrary windows.

Decision¶

Extend the existing gaze intake so the pane you look at receives the next dictation, and resolve deictic commands ("put this here", "close that") against the gaze target + the Context-Primed accessibility signals (ADR-v2-004). Reuse gaze/zones.py (extend zone→window-handle resolution) and commands/grammar.py. Consumer-webcam gaze is ~1–2° (coarse), so keep zone granularity + a confirm for destructive routes. EXPERIMENTAL, off by default, --force to enable; frames are processed in-RAM during a hold and never stored (ADR-011).

Consequences¶

• Adds spatial routing without new base deps (gaze extra already opt-in).
• Coarse accuracy is mitigated by zones + confirm; degrades to the focused window when confidence is low.
• Pure zone→window mapping stays testable; the webcam backend remains lazy.