feat: voice command mapping — context-gated approve/reject over voice

While a tool-permission gate is held on the active tab, spoken confirm
phrases resolve it via the existing approve/reject WS channel — no server
change (byte-shuttle intact). Everything else stays ordinary dictation.

Safety: whole-utterance-exact matching + leading-negation guard + reject
precedence + confidence gate (approve only) + a cancellable 1.5s confirm
window whose commit re-validates gate identity/epoch/connectivity (TOCTOU
guard) + stale-gate epoch bound at PTT-start.

New pure modules public/voice-commands.ts and public/voice-confirm.ts at
100% coverage; wiring in voice/terminal-session/tabs/main. 1307 tests pass,
typecheck clean, build:web OK. Independent code + security reviews flagged a
confirm-window TOCTOU and an unwired cancel path — both fixed and covered.

Plan: docs/PLAN_VOICE_COMMANDS.md.
This commit is contained in:
Yaojia Wang
2026-07-01 10:34:51 +02:00
parent 03612323c0
commit e4c327e25e
13 changed files with 1193 additions and 19 deletions

91
public/voice-commands.ts Normal file
View File

@@ -0,0 +1,91 @@
/**
* public/voice-commands.ts — pure, DOM-free/WS-free voice command matcher.
*
* Context-gated confirm-class command mapping: while the active tab has a
* HELD tool-gate permission (`pendingApproval===true`, `gate==='tool'`),
* spoken confirm/deny phrases resolve to `approve`/`reject`; every other
* transcript (or context) falls through to ordinary dictation (`text`),
* byte-for-byte unchanged. See docs/PLAN_VOICE_COMMANDS.md §3/§7 for the
* locked decisions and security rationale — this module is intentionally
* whole-utterance-exact (no substring/token matching) so incidental
* dictation can never auto-approve a shell-granting gate.
*/
export type VoiceAction = 'approve' | 'reject' | 'text'
export interface VoiceMatchContext {
readonly pendingApproval: boolean // active session held-permission flag
readonly gate: 'tool' | 'plan' | null // v1 acts ONLY on 'tool'
readonly confidence?: number // optional ASR confidence (decision D)
}
/** Approve requires ASR confidence at or above this threshold; reject is unaffected. */
export const MIN_APPROVE_CONFIDENCE = 0.6
/** Normalized whole-utterance affirmative phrases (see normalizeTranscript). */
export const APPROVE_PHRASES: readonly string[] = [
'确认', '批准', '同意', '通过', '允许', '可以', '好', '好的', '好吧', '好啊',
'是', '是的', '对', '行', '继续', '回车',
'yes', 'yeah', 'yep', 'ok', 'okay', 'confirm', 'approve', 'accept', 'proceed', 'go ahead',
]
/** Normalized whole-utterance negative phrases (see normalizeTranscript). */
export const REJECT_PHRASES: readonly string[] = [
'拒绝', '取消', '不行', '不要', '不用', '不同意', '不批准', '不可以', '不允许', '不通过',
'中断', '停止', '算了', '别', '不',
'no', 'nope', 'reject', 'deny', 'cancel', 'abort', 'decline', 'stop',
]
/** Leading tokens that negate an utterance; a match can never approve. */
export const NEGATION_PREFIXES: readonly string[] = [
'不', '别', '没', '未', '勿',
'no', 'not', 'dont', 'cannot', 'wont', 'never',
]
/** ASCII + CJK punctuation and quote marks, plus straight/curly apostrophes. */
const PUNCTUATION_RE =
/['.,!?;:,。!?;:、「」『』“”‘’()()[\]{}【】〈〉《》…—\-_/\\|"`~@#$%^&*+=<>]/g
/**
* `trim → toLowerCase → strip punctuation (so don't→dont) → collapse ASCII
* whitespace → trim`. Keeps CJK + ASCII letters/digits.
*/
export function normalizeTranscript(raw: string): string {
return raw
.trim()
.toLowerCase()
.replace(PUNCTUATION_RE, '')
.replace(/\s+/g, ' ')
.trim()
}
/** True if a normalized transcript begins with a negation token (word-boundary for ASCII). */
export function startsWithNegation(norm: string): boolean {
return NEGATION_PREFIXES.some((prefix) => {
if (!norm.startsWith(prefix)) return false
// CJK negation particles (不/别/没/未/勿) have no word boundary, so any
// following character still counts as negation. ASCII negators require a
// word boundary after them, so "now" is NOT negated by the "no" prefix.
if (!/^[a-z0-9]/.test(prefix)) return true
const rest = norm.slice(prefix.length)
return rest === '' || !/^[a-z0-9]/.test(rest)
})
}
/**
* Whole-utterance-exact, reject-first, negation-guarded command match.
* Falls through to `'text'` (ordinary dictation) for anything not an exact,
* gate-eligible, confidence-sufficient command phrase.
*/
export function matchCommand(transcript: string, ctx: VoiceMatchContext): VoiceAction {
const norm = normalizeTranscript(transcript)
if (!ctx.pendingApproval || norm === '') return 'text'
if (ctx.gate !== 'tool') return 'text'
if (startsWithNegation(norm)) return REJECT_PHRASES.includes(norm) ? 'reject' : 'text'
if (REJECT_PHRASES.includes(norm)) return 'reject'
if (APPROVE_PHRASES.includes(norm)) {
if (ctx.confidence !== undefined && ctx.confidence < MIN_APPROVE_CONFIDENCE) return 'text'
return 'approve'
}
return 'text'
}