Skip to content

ADR-v2-066 — Spoken Regex Builder

Status: Accepted (2026-07-02) · Wave I Context links: [[adr-v2-029-spoken-code-mode]] (literal code vs a search pattern), [[adr-v2-030-spoken-math-latex]], [[adr-011]]

Context

Wave I research (#9) — build regexes/search patterns by voice for developers who can't type fluently: "four digits dash two digits" → \d{4}-\d{2}, "lines starting with a capital letter" → ^[A-Z], "one or more digits" → \d+. Feeds find dialogs, grep, editor search. Distinct from Spoken Code Mode (dictates literal code) and Math→LaTeX (equations).

Decision

Add an opt-in Spoken Regex Builder: [spokenregex] enabled=false. The pure core nl_to_regex(text) is a compositional grammar: number-word + unit → {N} quantifier (\d, [A-Za-z], [A-Z], \w, \s), literal words (dash, dot, slash, at sign, …), one or more/zero or more/optional → +/*/? on the following atom, and a starting with prefix → ^. Returns None when nothing parses. Dependency-free. An optional local model for free-form descriptions is deferred. OFF by default.

Consequences

  • Hands-free pattern authoring; pure grammar, no model, deterministic and inspectable.
  • Distinct from Code Mode / Math (search pattern, not literal code/equation).
  • Privacy (ADR-011): fully local; patterns never leave the machine.
  • Caveat: covers a bounded phrase grammar (counts, atoms, quantifiers, anchors) → free-form English deferred to the optional model tier; off by default.