ADR-v2-004 — Context-Primed Dictation & Commanding¶
Status: Accepted (2026-07-02) · Wave A Context links: [[adr-v2-000-interaction-layer]], [[adr-v04-002-lsp-context]] (LSP context), [[adr-011]] (offline/no-storage)
Context¶
Whisper accuracy improves markedly when its initial_prompt carries domain terms. YazSes already primes the app name + personal vocab + LSP symbols via the _effective_initial_prompt chokepoint. The ambient research (internal) shows proactive context capture (ContextAgent) works, and crucially that the "when to act" trigger can be non-LLM and driven by cheap, already-consented desktop signals (active window title, current selection, clipboard) — avoiding both sensors and the cloud.
Decision¶
Automatically compose per-burst context, opt-in and never stored: 1. Context sources (all local, read at decode time, discarded after): active window title, current text selection, clipboard, and (when the LSP bridge is active) nearby editor symbols. Merged into initial_prompt through the existing chokepoint. 2. Deictic command resolution: commands referencing "this/that/here/above" are resolved against the same context (e.g. selection or LSP symbol under cursor) before dispatch. 3. Privacy controls: each source individually toggleable; content is used transiently and never written to the corpus or logs; redaction patterns apply if it ever were.
Config: [context] enabled=false, use_window_title, use_selection, use_clipboard, use_lsp. Per-OS accessors in platform/*; logic in commands/context.py.
Consequences¶
- + Domain terms transcribed correctly with zero user effort; context-aware commands.
- + Reuses the
initial_promptchokepoint + LSP provider; non-LLM trigger keeps it light. - − Reading window/selection/clipboard is sensitive → strictly local, opt-in per source, transient, never stored; clear documentation of exactly what is read.
- − Per-OS accessors vary (Wayland restricts window/selection reads) → degrade gracefully to whatever the platform allows.