feat(input): upgrade graph personalization, english morphology detection, and atomic SendInput replacement

This commit is contained in:
Yun Chan 2026-10-29 08:16:26 +09:00
parent 52eef21bab
commit cd4a452148
21 changed files with 1098 additions and 110 deletions

View file

@ -12,6 +12,25 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
- Additional Whisper model management UI - Additional Whisper model management UI
- Plugin system for custom pipelines - Plugin system for custom pipelines
## [1.13.0] - 2026-10-29
> Published from an annotated tag. Installer and update metadata are served by the
> canonical Forgejo feed; no binaries are committed to this repository.
### Added
- **Graph-relational extreme personalization & sentence suggestion upgrades (그래프 관계성 기반 극한 개인화 및 문장 제시 고도화).**
- Chains consecutive short messages within the same app (`_lastIndexedByScope`, 10-minute session window) via `follows` edges and restores post-send chat context from `recentWrites` when Enter clears the input box.
- Adds particle-invariant stemmed continuations (`stemContinuationsFrom`) so phrases like `"회의록 정리해서 내일 오전까지 공유드리겠습니다"` continue seamlessly when you type `"오늘 회의록을 정리해서"`, alongside 2-hop graph path scoring (`rankTwoHopGraphTargets` across `shares_terms → follows` and `follows → follows`).
- Weights frequent `personalPhrases` (`count >= 2`) and adds a Korean stemmed-word bigram fallback (`_stemBigram`) in `PersonalNgramModel`, while reranking generated LLM candidates by `scoreCandidatePersonalization` (graph overlap + Korean register match) without reducing suggestion frequency.
- **Stronger English detection via derivational morphology & phrase lookahead (파생 형태론 및 전방 구문 호응 기반 영어 감지 강화).**
- Expands English stem lexicon (`EXTENDED_ENGLISH_STEMS`: `kubernetes`, `microsoft`, `supabase`, `postgres`, `typescript`, `refactor`, `endpoint`, `workflow`, `dashboard`, etc.) with productive derivational prefixes (`un-`, `re-`, `pre-`, `non-`, `sub-`, `over-`, `under-`, `auto-`, `multi-`, `micro-`, `inter-`) and suffixes (`-tion`, `-sion`, `-ation`, `-ment`, `-ness`, `-able`, `-ible`, `-ful`, `-less`, `-ive`, `-ous`, `-al`, `-ic`, `-ize`, `-or`, `-ist`, `-ism`).
- Detects clause-initial and post-Korean Hangul-typed English function words via lookahead (`"새 ㅠㄷ "` → `"to be "`, `"새 솓 "` → `"to the "`, `"새 애 this "` → `"to do this "`, `"랙 솓 "` → `"for the "`, `"뭉 솓 "` → `"and the "`) while keeping Korean determiners intact (`"오늘 새 API를 만들었어 "`).
### Changed
- **Zero-sleep single-syscall atomic replacement & in-flight delta rebasing (단일 `SendInput` 원자적 치환 및 타이핑 중 델타 리베이싱).**
- Packs `(optional VK_SPACE) + N × VK_BACK + M × KEYEVENTF_UNICODE` into a single Windows `SendInput` call (`sendWin32UnicodeReplacement`, `< 0.5 ms`, `0 ms` sleep) so physical keystrokes never interleave with auto-correction.
- Rebases replacement plans (`rebaseTypoReplacementPlan`) when you keep typing during the 40 ms UIA verification or background LLM check, preserving newly typed suffix characters (`<= 12` chars) in one atomic swap.
## [1.12.0] - 2026-10-28 ## [1.12.0] - 2026-10-28
> Published from an annotated tag. Installer and update metadata are served by the > Published from an annotated tag. Installer and update metadata are served by the

View file

@ -249,12 +249,13 @@ async function initInputIntelligence(): Promise<void> {
typo.on('corrected', () => { typo.on('corrected', () => {
if (suggestion.isPresentationActive) suggestion.dismiss('stale') if (suggestion.isPresentationActive) suggestion.dismiss('stale')
}) })
// 인플레이스 자동 교정(handleTypingContext)은 타이핑과 절대 경합하지 않도록 // 인플레이스 자동 교정(handleTypingContext)은 커서 직전 꼬리 어절만 원자적(단일 Win32 SendInput)으로 교체한다:
// 0.1ms 결정적 한영·영문오타 엔진(layoutOnly: true, 커서 직전 어절만 교체)만 즉시 실행하고, // 1단계(25ms 단어 경계 / 650ms 어절 끝): 결정적 한영·영문오타(Araso) 즉시 전환
// 문장 단위 로컬 LLM(Gemma) 오타 교정은 키보드를 건드리지 않는 제안 팝업(propose)에서만 제시한다. // 2단계(350ms 수상한 어절 경계 / 600ms 유휴): 철자 사전·개인 어휘·자모 잡음 및 로컬 LLM(Gemma) 꼬리 오타 보정
// 이미 지나간 문장 앞쪽 오타는 키보드를 건드리지 않는 제안 팝업(propose)에서 제시한다.
suggestion.setTypoProposer({ propose: (context, signal, options) => typo.propose(context, signal, options) }) suggestion.setTypoProposer({ propose: (context, signal, options) => typo.propose(context, signal, options) })
telemetry.on('typed-context', (context) => { telemetry.on('typed-context', (context) => {
const verdict = typo.handleTypingContext(context, { layoutOnly: true, trailingWordOnly: true }) const verdict = typo.handleTypingContext(context, { trailingWordOnly: true })
if (verdict === 'pass') suggestion.handleTypingContext(context) if (verdict === 'pass') suggestion.handleTypingContext(context)
}) })
telemetry.on('state-changed', (state) => telemetry.on('state-changed', (state) =>

View file

@ -16,8 +16,10 @@ import {
continuationsFrom, continuationsFrom,
extractTerms, extractTerms,
jaccard, jaccard,
rankTwoHopGraphTargets,
scoreRelatedNodes, scoreRelatedNodes,
splitSentences, splitSentences,
stemContinuationsFrom,
type GraphContext, type GraphContext,
type PersonalGraphQuery, type PersonalGraphQuery,
type PersonalGraphStats, type PersonalGraphStats,
@ -40,13 +42,20 @@ export type { PersonalGraphQuery, PersonalGraphStats }
const MAINTENANCE_NODE_LIMIT = 300 const MAINTENANCE_NODE_LIMIT = 300
/** 관계 조회에서 최근 노드를 훑는 수. */ /** 관계 조회에서 최근 노드를 훑는 수. */
const TERM_SCAN_LIMIT = 400 const TERM_SCAN_LIMIT = 400
/** 같은 앱에서 단문 메시지를 연속으로 보낼 때 follows 엣지로 잇는 최대 세션 간격 (10분). */
const SESSION_FOLLOWS_WINDOW_MS = 10 * 60_000
/** 새 문장을 넣을 때 즉시 shares_terms 엣지를 연결할 최근 노드 수. */
const INCREMENTAL_SHARES_SCAN_LIMIT = 40
class PersonalGraphService { class PersonalGraphService {
/** 앱별 마지막 학습 문장 — 카카오톡·슬랙 등 단문 메시지 사이의 follows 사슬을 잇는다. */
private readonly _lastIndexedByScope = new Map<string, { phrase: string; at: number }>()
/** /**
* 텍스트를 그래프에 반영한다. * 텍스트를 그래프에 반영한다.
* *
* 문장마다 노드를 upsert 하고, 연속한 문장 쌍을 follows 엣지로, 용어를 충분히 * 문장마다 노드를 upsert 하고, 연속한 문장 쌍(같은 앱의 직전 메시지 포함)을 follows 엣지로,
* 공유하는 쌍을 shares_terms 엣지로 남긴다. * 용어를 충분히 공유하는 쌍을 shares_terms 엣지로 남긴다.
*/ */
indexText( indexText(
text: string, text: string,
@ -56,9 +65,17 @@ class PersonalGraphService {
if (sentences.length === 0) return if (sentences.length === 0) return
const at = meta.at ?? Date.now() const at = meta.at ?? Date.now()
const scopeKey = meta.appName?.trim().toLowerCase() || null
try { try {
const db = getDatabase() const db = getDatabase()
const recentExisting = db
.select({ phrase: personalPhrases.phrase, terms: personalPhrases.terms })
.from(personalPhrases)
.orderBy(desc(personalPhrases.lastUsedAt))
.limit(INCREMENTAL_SHARES_SCAN_LIMIT)
.all()
for (const sentence of sentences) { for (const sentence of sentences) {
const terms = extractTerms(sentence) const terms = extractTerms(sentence)
db.insert(personalPhrases) db.insert(personalPhrases)
@ -82,9 +99,36 @@ class PersonalGraphService {
} }
}) })
.run() .run()
if (terms.length > 0) {
for (const existing of recentExisting) {
if (existing.phrase === sentence) continue
const existingTerms = parseTerms(existing.terms)
if (existingTerms.length === 0) continue
if (jaccard(terms, existingTerms) >= SHARES_TERMS_MIN_SIMILARITY) {
this._upsertEdge(existing.phrase, sentence, 'shares_terms')
}
}
}
} }
// follows 엣지 // 같은 앱에서 직전 세션(10분 이내)에 쓴 문장과 첫 문장을 follows 로 연결
if (scopeKey) {
const previousInScope = this._lastIndexedByScope.get(scopeKey)
if (
previousInScope &&
previousInScope.phrase !== sentences[0] &&
Math.abs(at - previousInScope.at) <= SESSION_FOLLOWS_WINDOW_MS
) {
this._upsertEdge(previousInScope.phrase, sentences[0], 'follows')
}
this._lastIndexedByScope.set(scopeKey, {
phrase: sentences[sentences.length - 1],
at
})
}
// follows 엣지 (같은 텍스트 내 인접 문장)
for (const edge of buildSequenceEdges(sentences)) { for (const edge of buildSequenceEdges(sentences)) {
this._upsertEdge(edge.from, edge.to, 'follows') this._upsertEdge(edge.from, edge.to, 'follows')
} }
@ -135,7 +179,11 @@ class PersonalGraphService {
* 3) related — 지금 문장·직전 문장·앞 문맥의 용어를 IDF 로 가중해 고른 관련 문장과, * 3) related — 지금 문장·직전 문장·앞 문맥의 용어를 IDF 로 가중해 고른 관련 문장과,
* 그 문장들 뒤에 이어 쓴 문장(follows 1단계 확장). * 그 문장들 뒤에 이어 쓴 문장(follows 1단계 확장).
*/ */
retrieveContext(prefix: string, limit = 4, options: { appName?: string | null } = {}): GraphContext { retrieveContext(
prefix: string,
limit = 4,
options: { appName?: string | null; recentWrites?: readonly string[] } = {}
): GraphContext {
const empty: GraphContext = { continuations: [], related: [], followUps: [] } const empty: GraphContext = { continuations: [], related: [], followUps: [] }
const normalizedPrefix = prefix.replace(/\s+/gu, ' ').trim() const normalizedPrefix = prefix.replace(/\s+/gu, ' ').trim()
if (normalizedPrefix.length < 2) return empty if (normalizedPrefix.length < 2) return empty
@ -143,19 +191,26 @@ class PersonalGraphService {
try { try {
const db = getDatabase() const db = getDatabase()
const context = analyzeWritingContext(prefix) const context = analyzeWritingContext(prefix)
const recentWriteSentences =
options.recentWrites && options.recentWrites.length > 0
? splitSentences(options.recentWrites[options.recentWrites.length - 1] ?? '')
: []
const anchorLastSentence =
context.lastSentence ||
(recentWriteSentences.length > 0 ? recentWriteSentences[recentWriteSentences.length - 1] : '')
const inPrefix = (text: string): boolean => normalizedPrefix.includes(text) const inPrefix = (text: string): boolean => normalizedPrefix.includes(text)
// (1) 직전 문장 뒤에 이어 쓴 문장 // (1) 직전 문장(또는 방금 보낸 채팅 메시지) 뒤에 이어 쓴 문장
const followUps = this._followUpsAfter(context.lastSentence, context.currentSentence ? 2 : 3, inPrefix) const followUps = this._followUpsAfter(anchorLastSentence, context.currentSentence ? 2 : 3, inPrefix)
// (2) 꼬리 이어쓰기 — 지금 문장 안에서만 찾는다 (앞 문장 끝과 섞이지 않게) // (2) 꼬리 이어쓰기 — 지금 문장 안에서만 찾는다 (앞 문장 끝과 섞이지 않게)
const tailSource = context.currentSentence || context.lastSentence const tailSource = context.currentSentence || context.lastSentence
const continuations = this._continuationsForTail(tailSource) const continuations = this._continuationsForTail(tailSource)
// (3) 관련 문장 — IDF 가중 용어 겹침 + follows 1단계 확장 // (3) 관련 문장 — IDF 가중 용어 겹침 + 1-hop follows + 2-hop (shares_terms -> follows) 경로 확장
const queryTerms = buildQueryTerms([ const queryTerms = buildQueryTerms([
{ text: context.currentSentence, weight: 2 }, { text: context.currentSentence, weight: 2 },
{ text: context.lastSentence, weight: 1.5 }, { text: anchorLastSentence, weight: 1.5 },
{ text: context.earlierText.slice(-200), weight: 1 } { text: context.earlierText.slice(-200), weight: 1 }
]) ])
const recentNodes = db const recentNodes = db
@ -185,22 +240,66 @@ class PersonalGraphService {
if (related.length >= limit || related.includes(text) || followUps.includes(text) || inPrefix(text)) return if (related.length >= limit || related.includes(text) || followUps.includes(text) || inPrefix(text)) return
related.push(text) related.push(text)
} }
// 관련 문장의 다음 문장이 관련 문장 자체보다 "이어질 말" 에 가깝다 — 상위 앵커 2개만 확장한다. // 관련 문장의 다음 문장이 관련 문장 자체보다 "이어질 말" 에 가깝다 — 상위 앵커 확장 + 2-hop 경로 탐색
const anchors = scored.slice(0, 2).map((item) => item.text) const anchors = scored.slice(0, 3).map((item) => item.text)
const hops = anchors.length const firstHopEdges = anchors.length
? db ? db
.select() .select()
.from(phraseEdges) .from(phraseEdges)
.where(and(inArray(phraseEdges.fromPhrase, anchors), eq(phraseEdges.kind, 'follows'))) .where(or(inArray(phraseEdges.fromPhrase, anchors), inArray(phraseEdges.toPhrase, anchors)))
.orderBy(desc(phraseEdges.weight)) .orderBy(desc(phraseEdges.weight))
.limit(6) .limit(12)
.all() .all()
: [] : []
const directFollowHops = firstHopEdges.filter(
(edge) => edge.kind === 'follows' && anchors.includes(edge.fromPhrase)
)
const normalizedFirstHop = firstHopEdges.flatMap((edge) => {
if (edge.kind === 'follows' && anchors.includes(edge.fromPhrase)) {
return [{ from: edge.fromPhrase, to: edge.toPhrase, kind: 'follows' as const, weight: edge.weight }]
}
if (edge.kind === 'shares_terms') {
if (anchors.includes(edge.fromPhrase)) {
return [{ from: edge.fromPhrase, to: edge.toPhrase, kind: 'shares_terms' as const, weight: edge.weight }]
}
if (anchors.includes(edge.toPhrase)) {
return [{ from: edge.toPhrase, to: edge.fromPhrase, kind: 'shares_terms' as const, weight: edge.weight }]
}
}
return []
})
const midNodes = [...new Set(normalizedFirstHop.map((edge) => edge.to))]
const secondHopFollows = midNodes.length
? db
.select()
.from(phraseEdges)
.where(and(inArray(phraseEdges.fromPhrase, midNodes), eq(phraseEdges.kind, 'follows')))
.orderBy(desc(phraseEdges.weight))
.limit(12)
.all()
.map((edge) => ({
from: edge.fromPhrase,
to: edge.toPhrase,
kind: 'follows' as const,
weight: edge.weight
}))
: []
const twoHopTargets = rankTwoHopGraphTargets({
firstHopEdges: normalizedFirstHop,
secondHopFollowEdges: secondHopFollows,
queryTerms,
exclude: (text) => inPrefix(text) || followUps.includes(text),
limit: 2
})
for (const item of scored) { for (const item of scored) {
push(item.text) push(item.text)
const hop = hops.find((edge) => edge.fromPhrase === item.text) const hop = directFollowHops.find((edge) => edge.fromPhrase === item.text)
if (hop) push(hop.toPhrase) if (hop) push(hop.toPhrase)
} }
for (const target of twoHopTargets) {
push(target)
}
return { continuations, related, followUps } return { continuations, related, followUps }
} catch (error) { } catch (error) {
logger.warn( logger.warn(
@ -305,7 +404,16 @@ class PersonalGraphService {
) )
if (found.length > 0) return found if (found.length > 0) return found
} }
return []
// 정확한 문자열 꼬리가 어긋나도("회의록을 정리해서" vs "회의록 정리해서") 어간 열이 일치하면 이어쓰기를 뽑는다.
const recentPhrases = db
.select({ phrase: personalPhrases.phrase })
.from(personalPhrases)
.orderBy(desc(personalPhrases.count), desc(personalPhrases.lastUsedAt))
.limit(120)
.all()
.map((row) => row.phrase)
return stemContinuationsFrom(recentPhrases, normalized, 3)
} }
/** 용어 공유 엣지를 백필한다 (기존 노드들 사이 관계 보강). */ /** 용어 공유 엣지를 백필한다 (기존 노드들 사이 관계 보강). */
@ -469,6 +577,7 @@ class PersonalGraphService {
clearAll(): void { clearAll(): void {
try { try {
this._lastIndexedByScope.clear()
const db = getDatabase() const db = getDatabase()
db.delete(phraseEdges).run() db.delete(phraseEdges).run()
logger.info('개인 그래프 엣지 전체 삭제') logger.info('개인 그래프 엣지 전체 삭제')

View file

@ -43,7 +43,7 @@ import {
type SuggestionStep type SuggestionStep
} from '@d3ro/core/input-intelligence' } from '@d3ro/core/input-intelligence'
import { renderRawChatPrompt, rawChatStopSequences } from '@d3ro/core/chat-template' import { renderRawChatPrompt, rawChatStopSequences } from '@d3ro/core/chat-template'
import { buildQueryTerms, splitSentences } from '@d3ro/core/personal-graph' import { buildQueryTerms, scoreCandidatePersonalization, splitSentences } from '@d3ro/core/personal-graph'
import { copiesContext } from '@d3ro/core/suggestion-confidence' import { copiesContext } from '@d3ro/core/suggestion-confidence'
import { splitFirstWord } from '@d3ro/core/suggestion-text' import { splitFirstWord } from '@d3ro/core/suggestion-text'
import { predictPersonalContinuation } from '@d3ro/core/personal-ngram' import { predictPersonalContinuation } from '@d3ro/core/personal-ngram'
@ -1028,14 +1028,22 @@ export class SuggestionService extends EventEmitter {
private _collectMemoryHints(prefix: string, appName: string | null): LocalMemoryContext { private _collectMemoryHints(prefix: string, appName: string | null): LocalMemoryContext {
const telemetry = getInputTelemetryService() const telemetry = getInputTelemetryService()
const graphContext = getPersonalGraphService().retrieveContext(prefix, 4, { appName }) const recentWrites = telemetry
.listRecentWrites({
windowHandle: this._lastContext?.windowHandle ?? null,
appName,
windowTitle: this._lastContext?.windowTitle ?? null
})
.map((entry) => entry.text)
const graphContext = getPersonalGraphService().retrieveContext(prefix, 4, { appName, recentWrites })
const phrases = telemetry.listPhrases(120) const phrases = telemetry.listPhrases(120)
// 지금 글과 용어가 겹치는 문구만 — 문맥과 무관한 "자주 쓰는 말" 은 프롬프트를 흐린다. // 지금 글과 용어가 겹치는 문구만 — 문맥과 무관한 "자주 쓰는 말" 은 프롬프트를 흐린다.
const writing = analyzeWritingContext(prefix) const writing = analyzeWritingContext(prefix)
const lastRecentWrite = recentWrites.length > 0 ? recentWrites[recentWrites.length - 1] ?? '' : ''
const relevanceTerms = new Set( const relevanceTerms = new Set(
buildQueryTerms([ buildQueryTerms([
{ text: writing.currentSentence, weight: 2 }, { text: writing.currentSentence, weight: 2 },
{ text: writing.lastSentence, weight: 1 } { text: writing.lastSentence || lastRecentWrite, weight: 1 }
]).keys() ]).keys()
) )
const phraseHints = selectPhraseHints(phrases, prefix, 5, { appName, relevanceTerms }) const phraseHints = selectPhraseHints(phrases, prefix, 5, { appName, relevanceTerms })
@ -1759,6 +1767,32 @@ export class SuggestionService extends EventEmitter {
if (token !== this._generationToken || this._isDuplicateCandidate(candidate.text)) return false if (token !== this._generationToken || this._isDuplicateCandidate(candidate.text)) return false
this._candidates.push({ ...candidate, rank: this._candidates.length }) this._candidates.push({ ...candidate, rank: this._candidates.length })
if (this._activeIndex === 0 && this._candidates.length >= 2) {
const personalContext = {
prefix,
continuationHints: memory.continuationHints,
followUpHints: memory.followUpHints,
relatedHints: memory.relatedHints,
phraseHints: memory.phraseHints
}
const newScore = scoreCandidatePersonalization(candidate.text, personalContext)
if (newScore >= 1.5) {
const personalItems = this._candidates.filter((item) => item.source === 'personal')
const modelItems = this._candidates
.filter((item) => item.source !== 'personal')
.map((item, idx) => ({
item,
idx,
score: scoreCandidatePersonalization(item.text, personalContext)
}))
.sort((a, b) => (Math.abs(b.score - a.score) >= 0.8 ? b.score - a.score : a.idx - b.idx))
.map((entry) => entry.item)
this._candidates = [...personalItems, ...modelItems].map((item, index) => ({
...item,
rank: index
}))
}
}
logger.debug(`제안 후보 추가 ${this._candidates.length}/${SUGGESTION_DEFAULTS.maxCandidatesTotal}`) logger.debug(`제안 후보 추가 ${this._candidates.length}/${SUGGESTION_DEFAULTS.maxCandidatesTotal}`)
this._armVisibleTtl() this._armVisibleTtl()
this.emit('updated', this.getState()) this.emit('updated', this.getState())

View file

@ -104,6 +104,24 @@ export class TextInsertService extends EventEmitter {
options: { commitComposition?: boolean } = {} options: { commitComposition?: boolean } = {}
): Promise<InsertResult> { ): Promise<InsertResult> {
const run = this._tail.then(async () => { const run = this._tail.then(async () => {
if (method === 'keyboard' && this._keyboard.replaceBeforeCaret) {
const start = performance.now()
const handled = await this._keyboard.replaceBeforeCaret(deleteCount, text, options)
if (handled) {
this.emit('insert-started', { text, method: 'keyboard' })
const result: InsertResult = {
success: true,
method: 'keyboard',
textLength: text.length,
durationMs: performance.now() - start
}
this.emit('insert-completed', { result })
logger.info(
`Text replaced atomically (-${deleteCount} / +${text.length} chars, ${Math.round(result.durationMs)}ms, keyboard)`
)
return result
}
}
try { try {
let count = deleteCount let count = deleteCount
// 커서 앞 글자가 IME 조합 중이면 먼저 확정한다 (Space → 남은 공백까지 지운다). // 커서 앞 글자가 IME 조합 중이면 먼저 확정한다 (Space → 남은 공백까지 지운다).

View file

@ -32,6 +32,7 @@ import {
mergeTypoCorrections, mergeTypoCorrections,
parseModelTypoPairs, parseModelTypoPairs,
planTypoReplacement, planTypoReplacement,
rebaseTypoReplacementPlan,
tokenizeWords, tokenizeWords,
typoRegionStart, typoRegionStart,
validateModelCorrections, validateModelCorrections,
@ -123,7 +124,7 @@ const REVERT_LENGTH_TOLERANCE = 12
export class TypoCorrectionService extends EventEmitter { export class TypoCorrectionService extends EventEmitter {
private _applying = false private _applying = false
private _check: { abort: AbortController; prefix: string; context: TypingContext } | null = null private _check: { abort: AbortController; prefix: string; context: TypingContext; latestKeyAt: number } | null = null
/** 마지막으로 결정적·모델 검사를 마친 구간(교정 전·후 모두 포함) — 같은 구간을 되풀이해 묻거나 연쇄 재교정하지 않는다. */ /** 마지막으로 결정적·모델 검사를 마친 구간(교정 전·후 모두 포함) — 같은 구간을 되풀이해 묻거나 연쇄 재교정하지 않는다. */
private _lastCheckedRegion = '' private _lastCheckedRegion = ''
private _lastPostRegion = '' private _lastPostRegion = ''
@ -170,12 +171,29 @@ export class TypoCorrectionService extends EventEmitter {
this._watchRevert(context) this._watchRevert(context)
if (this._applying) return 'correcting' if (this._applying) return 'correcting'
// 진행 중 모델 검사 — 같은 접두면 계속 기다리고, 달라졌으면(더 쳤다) 버린다. // 진행 중 모델 검사 — 같은 접두이거나 짧은 공백·구두점/완결 어절 꼬리만 덧붙었으면 리베이스 대기로 유지하고,
// 한창 글자를 치는 중이면(유휴 부족·조합 중) 검사를 버리고 타이핑을 방해하지 않는다.
if (this._check) { if (this._check) {
if (this._check.prefix === context.prefix) { if (this._check.prefix === context.prefix) {
this._check.context = context this._check.context = context
this._check.latestKeyAt = this._deps.lastKeyAt()
return 'checking' return 'checking'
} }
if (context.prefix.startsWith(this._check.prefix)) {
const delta = context.prefix.slice(this._check.prefix.length)
const canRebaseDelta =
delta.length > 0 &&
delta.length <= 6 &&
(/^[ \t.,!?;:]{1,4}$/u.test(delta) ||
(!context.isComposing &&
context.idleMs >= LAYOUT_BOUNDARY_MIN_IDLE_MS &&
/[ \t.,!?;:]$/u.test(delta)))
if (canRebaseDelta) {
this._check.context = context
this._check.latestKeyAt = this._deps.lastKeyAt()
return 'checking'
}
}
this._cancelCheck() this._cancelCheck()
} }
@ -296,7 +314,7 @@ export class TypoCorrectionService extends EventEmitter {
this._markRegionChecked(region) this._markRegionChecked(region)
const abort = new AbortController() const abort = new AbortController()
this._check = { abort, prefix, context } this._check = { abort, prefix, context, latestKeyAt: keyAtDecision }
void this._runModelCheck(model, context, abort, keyAtDecision, ignoredWords, vocabulary.known) void this._runModelCheck(model, context, abort, keyAtDecision, ignoredWords, vocabulary.known)
return 'checking' return 'checking'
} }
@ -438,11 +456,20 @@ export class TypoCorrectionService extends EventEmitter {
const tailCorrections = mergeTypoCorrections(fromModel, detectJamoNoise(prefix, { ignoredWords })).filter( const tailCorrections = mergeTypoCorrections(fromModel, detectJamoNoise(prefix, { ignoredWords })).filter(
(c) => prefix.length - c.start <= MAX_MODEL_DELETE_CHARS (c) => prefix.length - c.start <= MAX_MODEL_DELETE_CHARS
) )
const plan = planTypoReplacement(prefix, tailCorrections) const basePlan = planTypoReplacement(prefix, tailCorrections)
if (!plan || plan.deleteCount > MAX_MODEL_DELETE_CHARS) return if (!basePlan) return
// 검사하는 사이 사용자가 더 쳤으면 접두가 달라져 위에서 취소됐다. 여기서 한 번 더 키를 본다. const activeCheck = this._check?.abort === abort ? this._check : null
const targetContext = activeCheck?.context ?? context
const targetKeyAt = activeCheck?.latestKeyAt ?? keyAtDecision
const rebasedPlan = rebaseTypoReplacementPlan(
basePlan,
prefix,
targetContext.prefix,
MAX_MODEL_DELETE_CHARS
)
if (!rebasedPlan) return
this._check = null this._check = null
corrected = await this._apply(plan, context, keyAtDecision) corrected = await this._apply(rebasedPlan, targetContext, targetKeyAt)
} catch (error) { } catch (error) {
if (!abort.signal.aborted) { if (!abort.signal.aborted) {
logger.debug(`모델 오타 검사 실패: ${error instanceof Error ? error.message : String(error)}`) logger.debug(`모델 오타 검사 실패: ${error instanceof Error ? error.message : String(error)}`)

View file

@ -9,7 +9,7 @@
import { PersonalNgramModel } from '@d3ro/core/personal-ngram' import { PersonalNgramModel } from '@d3ro/core/personal-ngram'
import { getDatabase } from '../../db' import { getDatabase } from '../../db'
import { typingSamples } from '../../db/schema' import { personalPhrases, typingSamples } from '../../db/schema'
import { getLogger } from '../LoggerService' import { getLogger } from '../LoggerService'
const logger = getLogger('PersonalNgramIndex') const logger = getLogger('PersonalNgramIndex')
@ -37,7 +37,21 @@ export class PersonalNgramIndex {
const startedAt = this._now() const startedAt = this._now()
try { try {
const model = new PersonalNgramModel() const model = new PersonalNgramModel()
for (const row of db.select({ text: typingSamples.text }).from(typingSamples).all()) model.addText(row.text) for (const row of db.select({ text: typingSamples.text }).from(typingSamples).all()) {
model.addText(row.text)
}
try {
for (const row of db
.select({ phrase: personalPhrases.phrase, count: personalPhrases.count })
.from(personalPhrases)
.all()) {
if (row.count >= 2 && typeof row.phrase === 'string') {
model.addText(row.phrase, Math.min(row.count - 1, 3))
}
}
} catch {
// personal_phrases 테이블이 없는 테스트 목(mock) DB 에서는 typing_samples 만으로 진행한다.
}
this._model = model this._model = model
this._builtFor = db this._builtFor = db
logger.info(`개인 n-gram 생성: 표본 ${model.textCount}개 (${this._now() - startedAt}ms)`) logger.info(`개인 n-gram 생성: 표본 ${model.textCount}개 (${this._now() - startedAt}ms)`)

View file

@ -4,7 +4,7 @@
import type { Key as NutKey } from '@nut-tree-fork/nut-js' import type { Key as NutKey } from '@nut-tree-fork/nut-js'
import { getLogger } from '../LoggerService' import { getLogger } from '../LoggerService'
import { sendWin32UnicodeText } from '../../utils/win32-unicode-input' import { sendWin32UnicodeReplacement, sendWin32UnicodeText } from '../../utils/win32-unicode-input'
import { D3ROError, ErrorCode } from '@d3ro/core/errors' import { D3ROError, ErrorCode } from '@d3ro/core/errors'
const logger = getLogger('TextInsertService') const logger = getLogger('TextInsertService')
@ -22,6 +22,15 @@ export interface KeyboardPort {
* 글자가 아니라 낱자가 지워져(여 → ㅇ) 지울 개수가 어긋난다. 호출자는 남은 공백까지 지운다 (선택). * 글자가 아니라 낱자가 지워져(여 → ㅇ) 지울 개수가 어긋난다. 호출자는 남은 공백까지 지운다 (선택).
*/ */
commitComposition?(): Promise<void> commitComposition?(): Promise<void>
/**
* (선택적 IME 조합 확정 Space) + 커서 앞 백스페이스 + 텍스트 타이핑을 단일 원자적(atomic) 호출로 수행한다.
* 지원하고 성공했으면 true, 폴백이 필요하면 false 를 돌려준다.
*/
replaceBeforeCaret?(
deleteCount: number,
text: string,
options?: { commitComposition?: boolean }
): Promise<boolean>
/** 네이티브 키 모듈을 미리 불러 둔다 — 첫 삽입이 로딩을 기다리지 않게 (선택) */ /** 네이티브 키 모듈을 미리 불러 둔다 — 첫 삽입이 로딩을 기다리지 않게 (선택) */
preload?(): Promise<void> preload?(): Promise<void>
} }
@ -112,6 +121,21 @@ export class NutKeyboardPort implements KeyboardPort {
await nut.native.click(nut.Key.Space) await nut.native.click(nut.Key.Space)
} }
async replaceBeforeCaret(
deleteCount: number,
text: string,
options: { commitComposition?: boolean } = {}
): Promise<boolean> {
if (
this._platform === 'win32' &&
this._load === loadNutModule &&
sendWin32UnicodeReplacement(deleteCount, text, options)
) {
return true
}
return false
}
/** 백스페이스를 count 번 — 키마다 네이티브 keyTap 한 번이다. */ /** 백스페이스를 count 번 — 키마다 네이티브 keyTap 한 번이다. */
async backspace(count: number): Promise<void> { async backspace(count: number): Promise<void> {
if (count <= 0) return if (count <= 0) return

View file

@ -18,6 +18,8 @@ const logger = getLogger('win32-unicode-input')
const INPUT_KEYBOARD = 1 const INPUT_KEYBOARD = 1
const KEYEVENTF_KEYUP = 0x0002 const KEYEVENTF_KEYUP = 0x0002
const KEYEVENTF_UNICODE = 0x0004 const KEYEVENTF_UNICODE = 0x0004
const VK_BACK = 0x08
const VK_SPACE = 0x20
type SendInputFn = (cInputs: number, pInputs: Buffer, cbSize: number) => number type SendInputFn = (cInputs: number, pInputs: Buffer, cbSize: number) => number
@ -80,6 +82,34 @@ function ensureLoaded(): SendInputFn | null {
} }
} }
function writeVirtualKeyPair(buffer: Buffer, pairIndex: number, vk: number): void {
const downBase = pairIndex * 2 * inputStructSize
buffer.writeUInt32LE(INPUT_KEYBOARD, downBase)
buffer.writeUInt16LE(vk, downBase + unionOffset)
buffer.writeUInt16LE(0, downBase + unionOffset + 2)
buffer.writeUInt32LE(0, downBase + unionOffset + 4)
const upBase = (pairIndex * 2 + 1) * inputStructSize
buffer.writeUInt32LE(INPUT_KEYBOARD, upBase)
buffer.writeUInt16LE(vk, upBase + unionOffset)
buffer.writeUInt16LE(0, upBase + unionOffset + 2)
buffer.writeUInt32LE(KEYEVENTF_KEYUP, upBase + unionOffset + 4)
}
function writeUnicodeKeyPair(buffer: Buffer, pairIndex: number, codeUnit: number): void {
const downBase = pairIndex * 2 * inputStructSize
buffer.writeUInt32LE(INPUT_KEYBOARD, downBase)
buffer.writeUInt16LE(0, downBase + unionOffset)
buffer.writeUInt16LE(codeUnit, downBase + unionOffset + 2)
buffer.writeUInt32LE(KEYEVENTF_UNICODE, downBase + unionOffset + 4)
const upBase = (pairIndex * 2 + 1) * inputStructSize
buffer.writeUInt32LE(INPUT_KEYBOARD, upBase)
buffer.writeUInt16LE(0, upBase + unionOffset)
buffer.writeUInt16LE(codeUnit, upBase + unionOffset + 2)
buffer.writeUInt32LE(KEYEVENTF_UNICODE | KEYEVENTF_KEYUP, upBase + unionOffset + 4)
}
/** /**
* UTF-16 문자열을 Windows SendInput(KEYEVENTF_UNICODE) 호출 한 번으로 입력 큐에 넣는다. * UTF-16 문자열을 Windows SendInput(KEYEVENTF_UNICODE) 호출 한 번으로 입력 큐에 넣는다.
* 성공 시 true, 지원하지 않거나 실패 시 false 를 반환한다. * 성공 시 true, 지원하지 않거나 실패 시 false 를 반환한다.
@ -95,18 +125,7 @@ export function sendWin32UnicodeText(text: string): boolean {
const buffer = Buffer.alloc(totalEvents * inputStructSize) const buffer = Buffer.alloc(totalEvents * inputStructSize)
for (let i = 0; i < codeUnits; i += 1) { for (let i = 0; i < codeUnits; i += 1) {
const code = text.charCodeAt(i) writeUnicodeKeyPair(buffer, i, text.charCodeAt(i))
const downBase = i * 2 * inputStructSize
buffer.writeUInt32LE(INPUT_KEYBOARD, downBase)
buffer.writeUInt16LE(0, downBase + unionOffset)
buffer.writeUInt16LE(code, downBase + unionOffset + 2)
buffer.writeUInt32LE(KEYEVENTF_UNICODE, downBase + unionOffset + 4)
const upBase = (i * 2 + 1) * inputStructSize
buffer.writeUInt32LE(INPUT_KEYBOARD, upBase)
buffer.writeUInt16LE(0, upBase + unionOffset)
buffer.writeUInt16LE(code, upBase + unionOffset + 2)
buffer.writeUInt32LE(KEYEVENTF_UNICODE | KEYEVENTF_KEYUP, upBase + unionOffset + 4)
} }
const sent = sendInput(totalEvents, buffer, inputStructSize) const sent = sendInput(totalEvents, buffer, inputStructSize)
@ -116,3 +135,54 @@ export function sendWin32UnicodeText(text: string): boolean {
return false return false
} }
} }
/**
* (선택적 IME 조합 확정 Space) + 커서 앞 백스페이스 N번 + UTF-16 문자열 입력을
* 단 한 번의 Windows SendInput 호출로 원자적(atomic)으로 주입한다.
*
* Win32 SendInput 규약상 단일 호출 안의 INPUT 배열 사이에 사용자의 물리 키 입력이 끼어들 수 없으므로,
* 빠른 타이핑 도중에도 지우기와 다시 쓰기가 경쟁하지 않고 < 0.5ms 안에 끝난다.
*/
export function sendWin32UnicodeReplacement(
deleteCount: number,
text: string,
options: { commitComposition?: boolean } = {}
): boolean {
const safeDeleteCount = Math.max(0, Math.trunc(deleteCount))
if (safeDeleteCount === 0 && !text) return true
const sendInput = ensureLoaded()
if (!sendInput || inputStructSize <= 0) return false
try {
const commitSpacePairs = options.commitComposition && safeDeleteCount > 0 ? 1 : 0
const backspacePairs = safeDeleteCount + commitSpacePairs
const textPairs = text.length
const totalPairs = commitSpacePairs + backspacePairs + textPairs
const totalEvents = totalPairs * 2
if (totalEvents === 0) return true
const buffer = Buffer.alloc(totalEvents * inputStructSize)
let pairCursor = 0
if (commitSpacePairs > 0) {
writeVirtualKeyPair(buffer, pairCursor, VK_SPACE)
pairCursor += 1
}
for (let i = 0; i < backspacePairs; i += 1) {
writeVirtualKeyPair(buffer, pairCursor, VK_BACK)
pairCursor += 1
}
for (let i = 0; i < textPairs; i += 1) {
writeUnicodeKeyPair(buffer, pairCursor, text.charCodeAt(i))
pairCursor += 1
}
const sent = sendInput(totalEvents, buffer, inputStructSize)
return sent === totalEvents
} catch (error) {
logger.warn(
`SendInput(atomic replacement) 실행 실패: ${error instanceof Error ? error.message : String(error)}`
)
return false
}
}

View file

@ -141,6 +141,47 @@ describe('TypoCorrectionService', () => {
expect(fake.replaceBeforeCaret).toHaveBeenCalledWith('ㅡㅁ가둥 '.length, 'markdown ') expect(fake.replaceBeforeCaret).toHaveBeenCalledWith('ㅡㅁ가둥 '.length, 'markdown ')
}) })
it('모델 검사 도중 사용자가 구두점·짧은 완결 어절을 덧붙이면 취소하지 않고 새 접두 위로 리베이스해 교정한다', async () => {
let release: (() => void) | null = null
const generate = vi.fn(async function* () {
await new Promise<void>((resolve) => {
release = resolve
})
yield 'ㅡㅁ가둥 => markdown'
})
const fake = deps({ warmModel: () => 'gemma4:e4b', generate })
const service = new TypoCorrectionService(fake.deps)
expect(
service.handleTypingContext(
typingContext({
prefix: 'ㅡㅁ가둥',
isComposing: false,
idleMs: 650,
appName: 'Antigravity.exe'
})
)
).toBe('checking')
// 모델이 응답하기 전에 사용자가 공백을 쳤지만 유휴(40ms) 조건을 만족한 경우 리베이스 대기 유지
fake.state.lastKeyAt = 1005
expect(
service.handleTypingContext(
typingContext({
prefix: 'ㅡㅁ가둥 ',
isComposing: false,
idleMs: 40,
appName: 'Antigravity.exe'
})
)
).toBe('checking')
release?.()
await flush()
await flush()
expect(fake.replaceBeforeCaret).toHaveBeenCalledWith('ㅡㅁ가둥 '.length, 'markdown ')
})
it('설정으로 끌 수 있다', () => { it('설정으로 끌 수 있다', () => {
config.typoAutoCorrectEnabled = false config.typoAutoCorrectEnabled = false
const fake = deps() const fake = deps()

View file

@ -74,4 +74,30 @@ describe('PersonalGraphService.retrieveContext', () => {
expect(context.related).not.toContain('배포 로그 확인했습니다') expect(context.related).not.toContain('배포 로그 확인했습니다')
expect(context.followUps).not.toContain('배포 로그 확인했습니다') expect(context.followUps).not.toContain('배포 로그 확인했습니다')
}) })
it('같은 앱에서 단문 메시지를 연속으로 보내도 세션 follows 사슬을 잇고 recentWrites 로 복원한다', () => {
const graph = getPersonalGraphService()
const now = Date.now()
graph.indexText('배포 모두 끝났습니다', { source: 'typed', appName: 'Slack.exe', at: now })
graph.indexText('모니터링 대시보드 확인 부탁드립니다', {
source: 'typed',
appName: 'Slack.exe',
at: now + 15_000
})
// 채팅창에서 엔터로 전송해 입력칸이 비워진 뒤 다음 메시지를 치기 시작한 상태
const context = graph.retrieveContext('지금 ', 4, {
appName: 'Slack.exe',
recentWrites: ['배포 모두 끝났습니다']
})
expect(context.followUps).toContain('모니터링 대시보드 확인 부탁드립니다')
})
it('조사가 달라 정확한 부분 문자열이 어긋나도 어간(stem) 연속 일치로 이어쓰기를 찾는다', () => {
const graph = getPersonalGraphService()
graph.indexText('회의록 정리해서 내일 오전까지 공유드리겠습니다', { source: 'typed' })
const context = graph.retrieveContext('오늘 회의록을 정리해서')
expect(context.continuations).toEqual(['내일 오전까지 공유드리겠습니다'])
})
}) })

View file

@ -2,7 +2,7 @@
> Status: ACTIVE > Status: ACTIVE
> Last full audit: 2026-09-13 > Last full audit: 2026-09-13
> Last update: 2026-10-28 — **v1.12.0 릴리스 + 한영 축약형·약어·오타 자동 교정 고도화(`INPUT-19`) + 전 표면 디자인파카(`designpaca`) 순회 감사(`SHELL-11`, `SHELL-15`).** (1) `@d3ro/core/typo-correction` 및 `TypoCorrectionService`를 확장해 영어 축약형·축약형+조사(`ㅊ무ㅅ`/`ㅊ무'ㅅ`/`ㅊ무ㅜ'ㅅ`→`can't`, `ㅊ무'ㅅ는`→`can't는`, `애ㅜ'ㅅ`→`don't`, `ㅑ'ㅡ`/`ㅑㅡ`→`I'm`, `재ㅜ'ㅅ`→`won't`), 문두 단독 `ㅑ `→`I `, 영어절 단음절 기능어(`새 애`→`to do`, `뭉`→`and`, `랙`→`for`, `소됴`→`they`, `생묘`→`today`), 일반 Araso 받침-조사 불일치 및 희소 어간 약어(`츤를`/`츤는`→`CMS를`/`CMS는`, `눈를`→`SNS를`, `운를`→`DNS를`, `긴를`→`RLS를`, `틔을`→`XML을`, `먀를`→`AI를`, `뗘를`→`EU를`, `쓰은`→`TM은`, `꾸은`→`RN은`, `띠이`→`EL이`, `꺼을`→`RJ을`, `ㅏㅅㅌ를`→`KTX를`, `ㅎㅅㅌ를`→`GTX를`), 1-편집 오타(`ㅗ디ㅐ`→`hello`, `ㅔㅐㄴ냐ㅠㅣ`→`possible`, `세갸ㅜㅎ`→`spring`, `ㅡㅁ가애주`→`markdown`, `help[`→`help`, `my self`→`myself`), 커서 꼬리 연쇄 구간(`MAX_FAST_DELETE_CHARS = 22`) 한정 고속 치환으로 타이핑 경합을 완전히 해소했다. (2) `packages/ui`, `apps/web`, `apps/admin`, `apps/desktop/src/renderer`(메인·팝업 6종), `site/`, `apps/mobile-rn` 6개 표면을 전수 순회해 `design.md` 계약 위반(WCAG AA 대비, `<11px` 미세 글자, `>=700` 굵기, 양수 자간, `uppercase`, 단독 `monospace`, `backdrop-filter` 블러, `h1` 위계 및 접근성 라벨)을 모두 해소했다 (`npm run check:design` 0건). > Last update: 2026-10-29 — **v1.13.0 릴리스 + 그래프 관계성 기반 극한 개인화(`INPUT-06`, `INPUT-20`) + 한영 형태론·전방 구문 감지 및 단일-시스콜 초고속 인플레이스 수정(`INPUT-19`, `INPUT-22`) + 전 표면 디자인파카(`designpaca`) 순회 감사(`SHELL-11`, `SHELL-15`).** (1) **그래프 관계성 & 극한 개인화:** `@d3ro/core/personal-graph` 및 `PersonalGraphService`에 앱별 연속 단문 메시지 간 세션 `follows` 체이닝(`_lastIndexedByScope`), 엔터 전송 직후 비워진 채팅창 문맥 복원(`recentWrites`), 증분 `shares_terms` 엣지 연결, 조사 차이 무관 어간 연속 이어쓰기(`stemContinuationsFrom`), 2-홉 그래프 경로 랭킹(`rankTwoHopGraphTargets`), 개인 그래프·어투 일치 기반 LLM 후보 재정렬(`scoreCandidatePersonalization`), 어간 바이그램 폴백 및 빈도 가중 개인 n-gram(`PersonalNgramModel`)을 추가했다(제안 노출 빈도는 그대로 유지). (2) **영어 감지 강화 & 초고속 타이핑 중 무경합 교정:** `@d3ro/core/typo-correction`에 확장 영어 어간(`EXTENDED_ENGLISH_STEMS`, `kubernetes`/`microsoft` 등)과 생산적 접두·접미 파생 형태론(`matchesInflectedOrDerivedEnglish`), 전방 영어 구문 호응(`새 ㅠㄷ `→`to be `, `새 애 this `→`to do this `, `랙 솓 `→`for the ` vs `오늘 새 API를 만들었어` 보존), 영어 축약형·약어(`ㅊ무ㅅ`/`ㅊ무'ㅅ`/`ㅊ무ㅜ'ㅅ`→`can't`, `츤를`→`CMS를`, `눈를`→`SNS를` 등)를 추가하고, `sendWin32UnicodeReplacement`(`utils/win32-unicode-input.ts`)로 `(선택 Space) + N × VK_BACK + M × KEYEVENTF_UNICODE`를 **단일 `SendInput` 시스템콜(`< 0.5 ms`, 대기 `0 ms`)**로 원자적 주입하며, 타이핑 중 추가 입력된 꼬리 글자를 보존하는 델타 리베이싱(`rebaseTypoReplacementPlan`)을 구현했다. (3) `packages/ui`, `apps/web`, `apps/admin`, `apps/desktop/src/renderer`(메인·팝업 6종), `site/`, `apps/mobile-rn` 6개 표면을 전수 순회해 `design.md` 계약 위반을 모두 해소했다 (`npm run check:design` 0건).
> >
> Previous update: 2026-10-27 — **v1.11.0 릴리스 + Araso 방식 한영 자동 변환·터미널/Vim/xterm.js 지원 완성 (`INPUT-19`, `GAP-INPUT-08`).** `nyttastudio/araso`를 분석해 `@d3ro/core/araso-layout` 및 `@d3ro/core/typo-correction`을 대폭 업그레이드하고 터미널·Vim·xterm.js(`Agent Switchboard`)까지 확장했다: 영어·약어·CamelCase+한국어 조사/단어 혼합어 분리(`githubdp`→`github에`, `PDFvkdlf`/`ㅖㅇㄹ파일`→`PDF파일`, `Reactzjavhsjsxm`→`React컴포넌트`, `dhsmfdmsgoogledptj`→`오늘은google에서`, `1:1ansdml`→`1:1문의`, 2글자 약어 받침 호응 `TMdms`→`TM은` vs `TMsms`→`쓰는`), 쌍자음 잔여 Shift 정규화(`djEJs`→`어떤`, `qkRNjTdj`→`바꿨어`), 한글 자판 상태 영어/혼합어 역변환(`ㅗ디ㅣㅐ 재깅`→`hello world`, `내ㅡㄷ샤ㅡㄷ`→`sometime`, `ㅎ샤ㅗㅕㅠ에`→`github에`, `ㅠㅕㅎ가`→`bug가`, 알파벳+조사 `ㄴ가`→`s가`, `ㄴ은`→`s은`), 문장 문맥 한영 전환(`Stay hungry, stay foolishsms wkqtmdml akfdlek.`), URL·경로·이메일 보호(`isUrlOrPathRun`), 터미널·Vim 명령(`:ㅈㅂ`→`:wq`, `:ㅂ!`→`:q!`, `퍄ㅡ `→`vim `, Vim 노멀 모드 교정 억제), `SendInput(KEYEVENTF_UNICODE)` 직접 유니코드 타이핑(클립보드 무접촉), 단어 경계(`25ms`)와 어절 내부(`650ms`) 유휴 분리 및 완성형 음절(`[가-힣]`) 한정 `commitComposition`으로 고속 타이핑 시 글자 삼킴·유령 공백·앞글자 중복을 완전히 해소했다. 통계 수집 동의(`inputTelemetryEnabled`)가 꺼져 있어도 기본 기능(`typoAutoCorrectEnabled: true`)으로 즉시 동작한다. > Previous update: 2026-10-27 — **v1.11.0 릴리스 + Araso 방식 한영 자동 변환·터미널/Vim/xterm.js 지원 완성 (`INPUT-19`, `GAP-INPUT-08`).** `nyttastudio/araso`를 분석해 `@d3ro/core/araso-layout` 및 `@d3ro/core/typo-correction`을 대폭 업그레이드하고 터미널·Vim·xterm.js(`Agent Switchboard`)까지 확장했다: 영어·약어·CamelCase+한국어 조사/단어 혼합어 분리(`githubdp`→`github에`, `PDFvkdlf`/`ㅖㅇㄹ파일`→`PDF파일`, `Reactzjavhsjsxm`→`React컴포넌트`, `dhsmfdmsgoogledptj`→`오늘은google에서`, `1:1ansdml`→`1:1문의`, 2글자 약어 받침 호응 `TMdms`→`TM은` vs `TMsms`→`쓰는`), 쌍자음 잔여 Shift 정규화(`djEJs`→`어떤`, `qkRNjTdj`→`바꿨어`), 한글 자판 상태 영어/혼합어 역변환(`ㅗ디ㅣㅐ 재깅`→`hello world`, `내ㅡㄷ샤ㅡㄷ`→`sometime`, `ㅎ샤ㅗㅕㅠ에`→`github에`, `ㅠㅕㅎ가`→`bug가`, 알파벳+조사 `ㄴ가`→`s가`, `ㄴ은`→`s은`), 문장 문맥 한영 전환(`Stay hungry, stay foolishsms wkqtmdml akfdlek.`), URL·경로·이메일 보호(`isUrlOrPathRun`), 터미널·Vim 명령(`:ㅈㅂ`→`:wq`, `:ㅂ!`→`:q!`, `퍄ㅡ `→`vim `, Vim 노멀 모드 교정 억제), `SendInput(KEYEVENTF_UNICODE)` 직접 유니코드 타이핑(클립보드 무접촉), 단어 경계(`25ms`)와 어절 내부(`650ms`) 유휴 분리 및 완성형 음절(`[가-힣]`) 한정 `commitComposition`으로 고속 타이핑 시 글자 삼킴·유령 공백·앞글자 중복을 완전히 해소했다. 통계 수집 동의(`inputTelemetryEnabled`)가 꺼져 있어도 기본 기능(`typoAutoCorrectEnabled: true`)으로 즉시 동작한다.
> >
@ -13,7 +13,7 @@
> Previous update: 2026-09-22 — **Gemma/Ollama 폭주 방어 계약을 기록했다.** 19:14:12 부팅 워밍업이 `keep_alive: 30m`으로 `gemma4:e4b`를 19:44:12까지 VRAM 3,226,342,521 bytes / context 4096으로 강제 상주시킨 것이 관측됐으며, 같은 시점 Windows GPU Engine PID 표본에는 Ollama의 활성 compute가 없었다. 즉 당시 상태는 무한 추론이 아니라 강제 residency였다. 19:11:17~19:11:58의 자동 제안 연속 생성은 기존 900 ms·12/min·5 candidates·128 tokens·한 글자 재생성 정책이 허용한 burst였다. 현재 구현 계약은 부팅 warmup 제거, `keep_alive: 2m`, 제안 600 ms / 최소 5 s 간격 / 기본 6회·hard max 12회 per min / 3 candidates / 64 tokens / 12-char growth / 8 s timeout, 그리고 요청별 취소·상한·종료 정리다. 독립 표적 검증은 6 test files / 69 tests passed / 0 failed, 변경 코드·테스트 ESLint와 `git diff --check`도 exit 0이다. Raw Ollama에서는 cold bounded 요청이 client hard timeout 15.044 s에 취소된 뒤 `/api/ps`가 비었고 `/api/version`은 80 ms에 회복했다. 명시 warmup은 HTTP 200 / 16.639 s, 후속 warm 요청은 body `options.num_predict=1`, `keep_alive='2m'`로 553 ms HTTP 200 / `done:true` / `eval_count:1` / `done_reason:length`였고 `/api/ps` expiry는 약 119.9 s였다. **19:48:59 +09:00에는 새 generate/unload/kill/retry 없이 충분히 지난 뒤 한 번의 `/api/ps`가 HTTP 200 / 45.8 ms / `{models:[]}`였고 `/api/version`은 HTTP 200 / 7.3 ms / `0.32.13`이었다.** 이는 raw API 수준의 expiry 뒤 unload 확인일 뿐 앱 재시작·GUI·실제 타이핑 증거는 아니므로 상태는 `[~]`로 유지한다 (`11` GAP-LLM-04, GAP-INPUT-06). > Previous update: 2026-09-22 — **Gemma/Ollama 폭주 방어 계약을 기록했다.** 19:14:12 부팅 워밍업이 `keep_alive: 30m`으로 `gemma4:e4b`를 19:44:12까지 VRAM 3,226,342,521 bytes / context 4096으로 강제 상주시킨 것이 관측됐으며, 같은 시점 Windows GPU Engine PID 표본에는 Ollama의 활성 compute가 없었다. 즉 당시 상태는 무한 추론이 아니라 강제 residency였다. 19:11:17~19:11:58의 자동 제안 연속 생성은 기존 900 ms·12/min·5 candidates·128 tokens·한 글자 재생성 정책이 허용한 burst였다. 현재 구현 계약은 부팅 warmup 제거, `keep_alive: 2m`, 제안 600 ms / 최소 5 s 간격 / 기본 6회·hard max 12회 per min / 3 candidates / 64 tokens / 12-char growth / 8 s timeout, 그리고 요청별 취소·상한·종료 정리다. 독립 표적 검증은 6 test files / 69 tests passed / 0 failed, 변경 코드·테스트 ESLint와 `git diff --check`도 exit 0이다. Raw Ollama에서는 cold bounded 요청이 client hard timeout 15.044 s에 취소된 뒤 `/api/ps`가 비었고 `/api/version`은 80 ms에 회복했다. 명시 warmup은 HTTP 200 / 16.639 s, 후속 warm 요청은 body `options.num_predict=1`, `keep_alive='2m'`로 553 ms HTTP 200 / `done:true` / `eval_count:1` / `done_reason:length`였고 `/api/ps` expiry는 약 119.9 s였다. **19:48:59 +09:00에는 새 generate/unload/kill/retry 없이 충분히 지난 뒤 한 번의 `/api/ps`가 HTTP 200 / 45.8 ms / `{models:[]}`였고 `/api/version`은 HTTP 200 / 7.3 ms / `0.32.13`이었다.** 이는 raw API 수준의 expiry 뒤 unload 확인일 뿐 앱 재시작·GUI·실제 타이핑 증거는 아니므로 상태는 `[~]`로 유지한다 (`11` GAP-LLM-04, GAP-INPUT-06).
> >
> Previous update: 2026-09-21 — **Input intelligence (입력 레메트리 + 다음 문장 제안) 신규**. 데스크톱에 입력 수집기(`InputTelemetryService`), UIA 컨텍스트 브리지(사이드카 `GET /uia/focus`), 제안 서비스(`SuggestionService`), 어렛 커 오버레이, 설정 > 입력 탭(동의·정책·주간 인사이트·개인 문구)을 추가했다. 카탈로그에 `INPUT-01`~`INPUT-08`(전부 데스크톱 `[~]` — 유닛 48건은 GREEN 이지만 **실앱 타이핑 검증 전**), 백로그에 GAP-INPUT-01~05 + GAP-LLM-03, §7 에 CONSTRAINT-INPUT-01(키 내용 미저장 — ActivityWatch 정책 채택)을 기록. 설계 근거는 조사 기반이다: 어렛은 `GetGUIThreadInfo` 가 아니라 UIA `TextPattern.GetSelection`(Chromium 은 `TextPattern2` 미구현), 타이핑 스트는 키코드 복원이 아니라 UIA 스냅샷 diff(한/일 IME 대응), 디바운스/토큰 한도는 인라인 컴플리션 실측값(Continue 350 / Tabby 250 / twinny 300 ms, 출력 64~256 토큰). 의존성: `koffi` 3.3.1(포그라운드 창 FFI), 사이드카 `uiautomation` 2.0.29 + `comtypes`. 데스크톱 유닛 총계 1409(+49), Electron ABI 실행에서 신규 실패 0건. 당시의 24.7초/4.9초 지연 설명과 `keep_alive: 30m`·부팅 워밍업 처방은 **현재 상태가 아닌 과거 가설/완화 이력**이며, 최신 운영 결론은 위 2026-09-22 항목과 `11` GAP-LLM-04를 따른다. 직전: LLM instruction-prompt fix (`9c2b4d4`): the custom-instruction path inserted the instruction's own wording instead of the processed result and had **never worked in any shipped release** (`v0.1.0-alpha`..`v1.4.0`, introduced `fea923d` 2026-04-05, not a regression). `llm-prompts.ts` is now the SSOT for prompt resolution and placeholder substitution, shared by `VoiceModeService` / `ChainService` / `LLM.PROCESS`. AI-04/05/06/07 are demoted to `[~]` on desktop — fixed with unit tests, but **not verified in a running app** and the four related `tests/red/*.usecase.test.ts` could not execute (`better-sqlite3` ABI). New: GAP-LLM-01 (no target-language setting), GAP-LLM-02 (this fix unverified); GAP-INFRA-06 amended (the ABI masks verification, not just dev-env switching cost); GAP-I18N-01 amended (`popup.error.default` missing in 10 locales). Earlier the same day: CAP-16 (desktop key bindings rebuilt on one `@d3ro/core/keybinding` SSOT — multiple bindings per action, mouse buttons, `HOTKEY` → `KEYBINDING` IPC group), verified on Windows by a manual run, so CAP-16 and CAP-02 are `[x]` and GAP-KEY-01 is closed. Still open: GAP-KEY-02/03, GAP-QA-02, GAP-I18N-01/02, GAP-INFRA-06, GAP-LLM-01/02, GAP-INPUT-01~05; `11` §7 holds accepted design constraints (things deliberately kept, not gaps) > Previous update: 2026-09-21 — **Input intelligence (입력 레메트리 + 다음 문장 제안) 신규**. 데스크톱에 입력 수집기(`InputTelemetryService`), UIA 컨텍스트 브리지(사이드카 `GET /uia/focus`), 제안 서비스(`SuggestionService`), 어렛 커 오버레이, 설정 > 입력 탭(동의·정책·주간 인사이트·개인 문구)을 추가했다. 카탈로그에 `INPUT-01`~`INPUT-08`(전부 데스크톱 `[~]` — 유닛 48건은 GREEN 이지만 **실앱 타이핑 검증 전**), 백로그에 GAP-INPUT-01~05 + GAP-LLM-03, §7 에 CONSTRAINT-INPUT-01(키 내용 미저장 — ActivityWatch 정책 채택)을 기록. 설계 근거는 조사 기반이다: 어렛은 `GetGUIThreadInfo` 가 아니라 UIA `TextPattern.GetSelection`(Chromium 은 `TextPattern2` 미구현), 타이핑 스트는 키코드 복원이 아니라 UIA 스냅샷 diff(한/일 IME 대응), 디바운스/토큰 한도는 인라인 컴플리션 실측값(Continue 350 / Tabby 250 / twinny 300 ms, 출력 64~256 토큰). 의존성: `koffi` 3.3.1(포그라운드 창 FFI), 사이드카 `uiautomation` 2.0.29 + `comtypes`. 데스크톱 유닛 총계 1409(+49), Electron ABI 실행에서 신규 실패 0건. 당시의 24.7초/4.9초 지연 설명과 `keep_alive: 30m`·부팅 워밍업 처방은 **현재 상태가 아닌 과거 가설/완화 이력**이며, 최신 운영 결론은 위 2026-09-22 항목과 `11` GAP-LLM-04를 따른다. 직전: LLM instruction-prompt fix (`9c2b4d4`): the custom-instruction path inserted the instruction's own wording instead of the processed result and had **never worked in any shipped release** (`v0.1.0-alpha`..`v1.4.0`, introduced `fea923d` 2026-04-05, not a regression). `llm-prompts.ts` is now the SSOT for prompt resolution and placeholder substitution, shared by `VoiceModeService` / `ChainService` / `LLM.PROCESS`. AI-04/05/06/07 are demoted to `[~]` on desktop — fixed with unit tests, but **not verified in a running app** and the four related `tests/red/*.usecase.test.ts` could not execute (`better-sqlite3` ABI). New: GAP-LLM-01 (no target-language setting), GAP-LLM-02 (this fix unverified); GAP-INFRA-06 amended (the ABI masks verification, not just dev-env switching cost); GAP-I18N-01 amended (`popup.error.default` missing in 10 locales). Earlier the same day: CAP-16 (desktop key bindings rebuilt on one `@d3ro/core/keybinding` SSOT — multiple bindings per action, mouse buttons, `HOTKEY` → `KEYBINDING` IPC group), verified on Windows by a manual run, so CAP-16 and CAP-02 are `[x]` and GAP-KEY-01 is closed. Still open: GAP-KEY-02/03, GAP-QA-02, GAP-I18N-01/02, GAP-INFRA-06, GAP-LLM-01/02, GAP-INPUT-01~05; `11` §7 holds accepted design constraints (things deliberately kept, not gaps)
> Scope: entire monorepo `D:/workspace/D3ROVoice` at product version `1.12.0` (`release/product-version.json`, released 2026-10-28) > Scope: entire monorepo `D:/workspace/D3ROVoice` at product version `1.13.0` (`release/product-version.json`, released 2026-10-29)
> Purpose: let any agent (or human) answer two questions in under a minute: > Purpose: let any agent (or human) answer two questions in under a minute:
> 1. **What infrastructure exists?** (build, CI, services, APIs, data, packages, deploy) > 1. **What infrastructure exists?** (build, CI, services, APIs, data, packages, deploy)
> 2. **How far is each feature developed?** (per surface, with file anchors and status) > 2. **How far is each feature developed?** (per surface, with file anchors and status)

View file

@ -18,7 +18,8 @@ The canonical place for types and cross-surface logic. Both desktop and web/mobi
| IPC channels | `./ipc-channels` | `IPC_CHANNELS` object + `IPCChannel` union. **SSOT** for every desktop IPC channel (VOICE, AUDIO, STT, TTS, LLM, KEYBINDING, CONFIG, HISTORY, UNIFIED_SEARCH, DICTIONARY, WINDOW, SYSTEM, STATS, MEMO, VOICE_COMMAND, CONTEXT, CHAIN, CAPTION, FILE_TRANSCRIPTION, MEETING_SUMMARY, DICTATION_TEMPLATE, VOICE_CONVERSATION, RAG, VOICE_ACTION, MEETING_MODE, MEETING_DOC_TEMPLATE, MEETING_CHAT, LICENSE, CLOUD_SYNC, INSTRUCTION, SYSTEM_AUDIO, POPUP_RESULT, POPUP_HISTORY, POPUP_COMMAND, POPUP_CAPTION, VOICE_PARTIAL, CLIPBOARD, APP, ONLINE_AUTH, ADS, SUPPORT, INPUT_TELEMETRY, SUGGESTION, POPUP_SUGGESTION) | | IPC channels | `./ipc-channels` | `IPC_CHANNELS` object + `IPCChannel` union. **SSOT** for every desktop IPC channel (VOICE, AUDIO, STT, TTS, LLM, KEYBINDING, CONFIG, HISTORY, UNIFIED_SEARCH, DICTIONARY, WINDOW, SYSTEM, STATS, MEMO, VOICE_COMMAND, CONTEXT, CHAIN, CAPTION, FILE_TRANSCRIPTION, MEETING_SUMMARY, DICTATION_TEMPLATE, VOICE_CONVERSATION, RAG, VOICE_ACTION, MEETING_MODE, MEETING_DOC_TEMPLATE, MEETING_CHAT, LICENSE, CLOUD_SYNC, INSTRUCTION, SYSTEM_AUDIO, POPUP_RESULT, POPUP_HISTORY, POPUP_COMMAND, POPUP_CAPTION, VOICE_PARTIAL, CLIPBOARD, APP, ONLINE_AUTH, ADS, SUPPORT, INPUT_TELEMETRY, SUGGESTION, POPUP_SUGGESTION) |
| Key bindings | `./keybinding` | **SSOT** for every global shortcut in the app: `KeyBinding` (`device`/`code`/`ctrl`/`alt`/`shift`/`meta`), `KEY_CATALOG` (10 selectable groups incl. mouse), `KEYBINDING_ACTIONS` (9 rebindable actions incl. the three suggestion actions), `bindingKey`/`normalizeBinding`/`validateBinding`/`detectBindingConflicts`/`formatBindingSegments`/`searchKeyCatalog`/`parseBindingMap`. Persisted as `AppConfig.keyBindings`. i18n keys are exposed as plain `string` so core stays independent of `@d3ro/i18n`; consumers narrow at the boundary (`asTranslationKey`) and a contract test guards the keys — accepted constraint, `11` §7 CONSTRAINT-I18N-01. Tests: `__tests__/keybinding*.test.ts` via `vitest.config.ts` (`npm run test --workspace=@d3ro/core`), 117 cases as of 2026-09-21 | | Key bindings | `./keybinding` | **SSOT** for every global shortcut in the app: `KeyBinding` (`device`/`code`/`ctrl`/`alt`/`shift`/`meta`), `KEY_CATALOG` (10 selectable groups incl. mouse), `KEYBINDING_ACTIONS` (9 rebindable actions incl. the three suggestion actions), `bindingKey`/`normalizeBinding`/`validateBinding`/`detectBindingConflicts`/`formatBindingSegments`/`searchKeyCatalog`/`parseBindingMap`. Persisted as `AppConfig.keyBindings`. i18n keys are exposed as plain `string` so core stays independent of `@d3ro/i18n`; consumers narrow at the boundary (`asTranslationKey`) and a contract test guards the keys — accepted constraint, `11` §7 CONSTRAINT-I18N-01. Tests: `__tests__/keybinding*.test.ts` via `vitest.config.ts` (`npm run test --workspace=@d3ro/core`), 117 cases as of 2026-09-21 |
| Input intelligence | `./input-intelligence` | **SSOT** for the input-telemetry and next-sentence-suggestion domain (added 2026-09-21): key classification (`classifyKeyStroke`), text metrics (`countWords`/`countSentences`/`endsSentence`/`textBeforeCaret`), `computeTypedDelta` (UIA snapshot diff — the IME-safe way to count typed text), `decideSuggestion` + `isAppExcluded` (when to request / skip / clear), `parseSuggestionCandidates`/`sanitizeSuggestionLine` (prompt-leak and prefix-echo defence), `anchorFloatingPanel` (caret-anchored overlay placement), `mergeActivityBucket`/`summarizeActivity`/`pixelsToMeters`, `extractPhrases`/`selectPhraseHints`, `INPUT_TELEMETRY_DEFAULTS`/`SUGGESTION_DEFAULTS` | | Input intelligence | `./input-intelligence` | **SSOT** for the input-telemetry and next-sentence-suggestion domain (added 2026-09-21): key classification (`classifyKeyStroke`), text metrics (`countWords`/`countSentences`/`endsSentence`/`textBeforeCaret`), `computeTypedDelta` (UIA snapshot diff — the IME-safe way to count typed text), `decideSuggestion` + `isAppExcluded` (when to request / skip / clear), `parseSuggestionCandidates`/`sanitizeSuggestionLine` (prompt-leak and prefix-echo defence), `anchorFloatingPanel` (caret-anchored overlay placement), `mergeActivityBucket`/`summarizeActivity`/`pixelsToMeters`, `extractPhrases`/`selectPhraseHints`, `INPUT_TELEMETRY_DEFAULTS`/`SUGGESTION_DEFAULTS` |
| Typo correction & Araso layout | `./typo-correction`, `./araso-layout` | **SSOT** for built-in automatic Han/Eng conversion and typo auto-correction (upgraded 2026-10-26): 2-set Korean/English layout conversion (`engToHangul`/`hangulToEng`), lingering Shift normalization (`normalizeLingeringShift`/`lingersAfterKoreanShift`), mixed-word splitting with Korean particle batchim agreement (`splitMixedWord`/`particleAgrees`), reverse Hangul-to-English/mixed restoration, URL/path/email run protection (`isUrlOrPathRun`), deterministic spelling & jamo-noise rules (`detectTypos`/`detectJamoNoise`), and LLM typo validation (`validateModelCorrections`). | | Typo correction & Araso layout | `./typo-correction`, `./araso-layout` | **SSOT** for built-in automatic Han/Eng conversion and typo auto-correction (upgraded 2026-10-28): 2-set Korean/English layout conversion (`engToHangul`/`hangulToEng`), productive English derivational prefixes/suffixes (`matchesInflectedOrDerivedEnglish`, `EXTENDED_ENGLISH_STEMS`), phrase-level lookahead English detection (`TO_FOLLOWING_ENGLISH_WORDS`, `resolveTokenAsEnglishPreview`), lingering Shift normalization (`normalizeLingeringShift`/`lingersAfterKoreanShift`), mixed-word splitting with Korean particle batchim agreement (`splitMixedWord`/`particleAgrees`), reverse Hangul-to-English/mixed restoration, URL/path/email run protection (`isUrlOrPathRun`), deterministic spelling & jamo-noise rules (`detectTypos`/`detectJamoNoise`), delta-rebasing for in-flight typing (`rebaseTypoReplacementPlan`), and LLM typo validation (`validateModelCorrections`). |
| Personal graph & n-gram | `./personal-graph`, `./personal-ngram` | **SSOT** for relational personalization and n-gram continuation (upgraded 2026-10-28): exact & particle-invariant stemmed continuations (`continuationsFrom`/`stemContinuationsFrom`), IDF-weighted related node scoring (`scoreRelatedNodes`), 2-hop graph path ranking (`rankTwoHopGraphTargets` across `shares_terms→follows` and `follows→follows`), personalized candidate reranking (`scoreCandidatePersonalization`), and frequency-weighted n-gram indexing with Korean stemmed-word bigram fallback (`PersonalNgramModel`). |
| Constants | `./constants` | Shared constants | | Constants | `./constants` | Shared constants |
| Crypto license | `./utils/crypto-license` | Ed25519 license sign/verify (used by admin issuer + desktop verifier) | | Crypto license | `./utils/crypto-license` | Ed25519 license sign/verify (used by admin issuer + desktop verifier) |
| PII | `pii-redactor`, `secure-memory` | Redaction + secure memory helpers | | PII | `pii-redactor`, `secure-memory` | Redaction + secure memory helpers |

View file

@ -84,7 +84,7 @@ Singleton + `EventEmitter` pattern (`getXService()` accessors).
| Service / file | Purpose | | Service / file | Purpose |
|---|---| |---|---|
| `InputTelemetryService` | Global input telemetry and context observer: keystroke/click/scroll counters, mouse travel (Manhattan sum of per-axis deltas), active time, per-hour-per-app buckets flushed every 5 s; typed-text learning via UIA snapshot diffs. **Since 2026-10-26**, the UIA context observer runs by default whenever `typoAutoCorrectEnabled !== false` (or `suggestionEnabled === true`), while SQLite `input_activity` persistence only runs when `inputTelemetryEnabled === true` (opt-in). **Since 2026-10-27**, includes `TerminalKeystrokeTracker` (`@d3ro/core/input-intelligence`) + `isForegroundKoreanIme` (`utils/win32-foreground.ts`) to reconstruct the active terminal/Vim line (`reconcileWithScreen` for UIA screen buffers like Windows Terminal/pwsh/cmd, `reconcileWithXtermValue` for `xterm.js` `<textarea class="xterm-helper-textarea">` controls in Agent Switchboard/VS Code/Cursor where `.value` stays populated until Enter/Ctrl+C, and `getSnapshot` for canvas/remote terminals with `len=0` like Kitedesk/WezTerm/Alacritty), preserving trailing spaces stripped by grid terminals and tracking Vim modes (`insert`, `normal`, `command`). Programmatic keys (`_programmaticKeys`) no longer advance `_lastKeyAt`, and `Enter`-triggered chat submissions that clear or shrink the input field rebase focus text immediately so cleared/toolbar text does not trigger post-send suggestions. It extends the telemetry store with Flow Radar (hourly activity density + chars + edit stability; a potential-focus score, **not** a real session), Edit Friction (chars/backspaces; edits per 100 chars), App Quality (per-app suggestion total/accepted/accept rate/average latency), Privacy Receipt (actual row counts and retention), and Smart Exclusion evidence. Raw key events, key codes and content streams are never retained; separately, opt-in learned text may be retained in `typing_samples` and `personal_phrases` under the stated receipt policy. | | `InputTelemetryService` | Global input telemetry and context observer: keystroke/click/scroll counters, mouse travel (Manhattan sum of per-axis deltas), active time, per-hour-per-app buckets flushed every 5 s; typed-text learning via UIA snapshot diffs. **Since 2026-10-26**, the UIA context observer runs by default whenever `typoAutoCorrectEnabled !== false` (or `suggestionEnabled === true`), while SQLite `input_activity` persistence only runs when `inputTelemetryEnabled === true` (opt-in). **Since 2026-10-27**, includes `TerminalKeystrokeTracker` (`@d3ro/core/input-intelligence`) + `isForegroundKoreanIme` (`utils/win32-foreground.ts`) to reconstruct the active terminal/Vim line (`reconcileWithScreen` for UIA screen buffers like Windows Terminal/pwsh/cmd, `reconcileWithXtermValue` for `xterm.js` `<textarea class="xterm-helper-textarea">` controls in Agent Switchboard/VS Code/Cursor where `.value` stays populated until Enter/Ctrl+C, and `getSnapshot` for canvas/remote terminals with `len=0` like Kitedesk/WezTerm/Alacritty), preserving trailing spaces stripped by grid terminals and tracking Vim modes (`insert`, `normal`, `command`). Programmatic keys (`_programmaticKeys`) no longer advance `_lastKeyAt`, and `Enter`-triggered chat submissions that clear or shrink the input field rebase focus text immediately so cleared/toolbar text does not trigger post-send suggestions. It extends the telemetry store with Flow Radar (hourly activity density + chars + edit stability; a potential-focus score, **not** a real session), Edit Friction (chars/backspaces; edits per 100 chars), App Quality (per-app suggestion total/accepted/accept rate/average latency), Privacy Receipt (actual row counts and retention), and Smart Exclusion evidence. Raw key events, key codes and content streams are never retained; separately, opt-in learned text may be retained in `typing_samples` and `personal_phrases` under the stated receipt policy. |
| `TypoCorrectionService` | Built-in automatic Han/Eng layout conversion and typo auto-correction (`INPUT-19`). Runs on each `typed-context`: Araso-style layout conversions (`source: 'layout'`) fire after 25 ms on word boundaries (Space/punctuation, 40 ms quick snapshot) or 135 ms at token tail (150 ms quick snapshot, even when Chromium/Discord keeps `isComposing=true`, committing IME composition first only when the trailing character is Hangul/jamo), replacing 2-set Han/Eng layout errors, mixed English+Korean words (`githubdp`→`github에`, `PDFvkdlf`→`PDF파일`, `ㅖㅇㄹ파일`→`PDF파일`), lingering Shift on double consonants (`djEJs`→`어떤`), reverse Hangul-to-English (`ㅎ샤ㅗㅕㅠ에`→`github에`, `ㅑㄹ ㅑ 쇼ㅔㄷ ㅗ디ㅣㅐ 소두 ㅑㅅ 초뭏ㄷㄴ ㅑㅜ 10 ㄴㄷ채ㅜㅇㄴ`→`if I type hello then it changes in 10 seconds`), and partially-converted tokens (`안녕gktpdy`→`안녕하세요`). **In terminals and Vim (`terminalMode: true`)**, layout-only conversion stays active while spelling/model checks are suppressed: converts CLI commands (`퍄ㅡ `→`vim `, `ㅜ퍄ㅡ `→`nvim `, `ㅎ샤 `→`git `, `ㅣㄴ `→`ls `, `채ㅜㄴㅅ `→`const `), Vim Insert-mode typing (stripping leading `i/a/o/I/A/O/s/S` mode-entry keys when present in raw keystroke buffers), and Vim Command-line ex-commands (`:ㅈㅂ`→`:wq`, `:ㅈㅂ!`→`:wq!`, `:ㅈ`→`:w`, `:ㅂ`→`:q`, `:ㅂ!`→`:q!`, `:ㅌ`→`:x`, `:ㅍ네`→`:vsp`, `:네`→`:sp`, `:ㅜㅐㅗ`→`:noh`, preserving `:`), while blocking all conversions in Vim Normal mode (`vimMode === 'normal'`). On Windows, `replaceBeforeCaret` uses `SendInput(KEYEVENTF_UNICODE)` (`utils/win32-unicode-input.ts`) instead of `Ctrl+V` so it never touches the user's clipboard or collides with `ClipboardSvcGroup` `OpenClipboard` locks. Deterministic/model spelling fixes run after a 600 ms pause. When `SuggestionService` is enabled, `handleTypingContext(context, { layoutOnly: true })` still auto-converts layout mistypes immediately in-place while `propose()` surfaces spelling/model fixes inside the suggestion popup (`INPUT-22`). | | `TypoCorrectionService` | Built-in automatic Han/Eng layout conversion and typo auto-correction (`INPUT-19`). Runs on each `typed-context`: Araso-style layout conversions (`source: 'layout'`) fire after 25 ms on word boundaries (Space/punctuation, 40 ms quick snapshot) or 135 ms at token tail (150 ms quick snapshot, even when Chromium/Discord keeps `isComposing=true`, committing IME composition first only when the trailing character is Hangul/jamo), replacing 2-set Han/Eng layout errors, mixed English+Korean words (`githubdp`→`github에`, `PDFvkdlf`→`PDF파일`, `ㅖㅇㄹ파일`→`PDF파일`), lingering Shift on double consonants (`djEJs`→`어떤`), reverse Hangul-to-English (`ㅎ샤ㅗㅕㅠ에`→`github에`, `ㅑㄹ ㅑ 쇼ㅔㄷ ㅗ디ㅣㅐ 소두 ㅑㅅ 초뭏ㄷㄴ ㅑㅜ 10 ㄴㄷ채ㅜㅇㄴ`→`if I type hello then it changes in 10 seconds`), lookahead English clauses (`새 ㅠㄷ `→`to be `, `새 애 this `→`to do this `), and partially-converted tokens (`안녕gktpdy`→`안녕하세요`). **In terminals and Vim (`terminalMode: true`)**, layout-only conversion stays active while spelling/model checks are suppressed: converts CLI commands (`퍄ㅡ `→`vim `, `ㅜ퍄ㅡ `→`nvim `, `ㅎ샤 `→`git `, `ㅣㄴ `→`ls `, `채ㅜㄴㅅ `→`const `), Vim Insert-mode typing (stripping leading `i/a/o/I/A/O/s/S` mode-entry keys when present in raw keystroke buffers), and Vim Command-line ex-commands (`:ㅈㅂ`→`:wq`, `:ㅈㅂ!`→`:wq!`, `:ㅈ`→`:w`, `:ㅂ`→`:q`, `:ㅂ!`→`:q!`, `:ㅌ`→`:x`, `:ㅍ네`→`:vsp`, `:네`→`:sp`, `:ㅜㅐㅗ`→`:noh`, preserving `:`), while blocking all conversions in Vim Normal mode (`vimMode === 'normal'`). **Since 2026-10-28**, on Windows, `TextInsertService.replaceBeforeCaret` (`method === 'keyboard'`) executes an atomic single-syscall `sendWin32UnicodeReplacement` (`utils/win32-unicode-input.ts`) packing `(optional VK_SPACE) + N × VK_BACK + M × KEYEVENTF_UNICODE` into one `SendInput` call (`<0.5 ms`, zero JS sleep, zero clipboard access), and `TypoCorrectionService` delta-rebases replacement plans (`rebaseTypoReplacementPlan`) when the user appends characters while an async snapshot or LLM check is in flight. Both Stage 1 (deterministic layout/spelling) and Stage 2 (tail-anchored `trailingWordOnly: true` Gemma check) run seamlessly even when `SuggestionService` is enabled. |
| `UiaContextService` | Client for the sidecar UIA bridge (`GET /uia/focus`): focused text, caret rect/offset, `IsPassword`, IME composition state and `hasSelection`. The sidecar derives `hasSelection` only by comparing the TextPattern selection range Start/End endpoints: it does not call `GetText` on that range or add a separate selected-text payload (the existing focused-field text can still include a selection). Fail-closed on password/unavailable; transient failures back off for 1.5/3/6/12/30 s and permanent failures for 5 min. | | `UiaContextService` | Client for the sidecar UIA bridge (`GET /uia/focus`): focused text, caret rect/offset, `IsPassword`, IME composition state and `hasSelection`. The sidecar derives `hasSelection` only by comparing the TextPattern selection range Start/End endpoints: it does not call `GetText` on that range or add a separate selected-text payload (the existing focused-field text can still include a selection). Fail-closed on password/unavailable; transient failures back off for 1.5/3/6/12/30 s and permanent failures for 5 min. |
| `SuggestionService` | Next-sentence ghost text: 600 ms debounce, then `buildSuggestionPrompt` → `LocalLLMService.streamGenerate` with a caller-owned `AbortSignal` → candidate ranking and accept (via `TextInsertService`). Presentation-active includes candidates plus `generating`, `warmingUp` and `partialText`; clear/dismiss aborts the active request, invalidates its generation token, clears TTL state, then emits `cleared`/hide so a spinner-only popup closes. A successful final candidate payload resets `generating=false` and `partialText=null`. App DNA extends phrase ranking with `appName` context and a 1.75× same-app bonus; Memory Decay uses a 30-day half-life plus frequency. Why This Suggestion exposes only `local-model`/`local-memory` provenance and continuation/related/phrase/appPhrase **counts** to the overlay, never raw evidence text. Instant Recall uses local-memory provenance without another model-budget spend only when the model is unavailable, raises a non-cancellation error, truly times out, or returns empty; it never publishes after dismiss, new typing, token/context mismatch or staleness. Its runaway limits are a 5 s minimum interval, 6 requests/min default (hard-config maximum 12), 3 candidates, 64 output tokens, 12-character regeneration growth and an 8 s timeout; app startup does not warm the model and suggestion requests use `keep_alive: 2m`. | | `SuggestionService` | Next-sentence ghost text: 600 ms debounce, then `buildSuggestionPrompt` → `LocalLLMService.streamGenerate` with a caller-owned `AbortSignal` → candidate ranking and accept (via `TextInsertService`). Presentation-active includes candidates plus `generating`, `warmingUp` and `partialText`; clear/dismiss aborts the active request, invalidates its generation token, clears TTL state, then emits `cleared`/hide so a spinner-only popup closes. A successful final candidate payload resets `generating=false` and `partialText=null`. App DNA extends phrase ranking with `appName` context and a 1.75× same-app bonus; Memory Decay uses a 30-day half-life plus frequency. Why This Suggestion exposes only `local-model`/`local-memory` provenance and continuation/related/phrase/appPhrase **counts** to the overlay, never raw evidence text. Instant Recall uses local-memory provenance without another model-budget spend only when the model is unavailable, raises a non-cancellation error, truly times out, or returns empty; it never publishes after dismiss, new typing, token/context mismatch or staleness. Its runaway limits are a 5 s minimum interval, 6 requests/min default (hard-config maximum 12), 3 candidates, 64 output tokens, 12-character regeneration growth and an 8 s timeout; app startup does not warm the model and suggestion requests use `keep_alive: 2m`. |
| `global-input-hook.ts` | Ref-counted owner of the single process-wide `uiohook` hook so `KeyBindingService` and telemetry can both attach without one stopping the other. | | `global-input-hook.ts` | Ref-counted owner of the single process-wide `uiohook` hook so `KeyBindingService` and telemetry can both attach without one stopping the other. |

File diff suppressed because one or more lines are too long

View file

@ -4,12 +4,16 @@
import { describe, expect, it } from 'vitest' import { describe, expect, it } from 'vitest'
import { import {
SHARES_TERMS_MIN_SIMILARITY, SHARES_TERMS_MIN_SIMILARITY,
buildQueryTerms,
buildSequenceEdges, buildSequenceEdges,
continuationsFrom, continuationsFrom,
extractTerms, extractTerms,
jaccard, jaccard,
rankRelated, rankRelated,
splitSentences rankTwoHopGraphTargets,
scoreCandidatePersonalization,
splitSentences,
stemContinuationsFrom
} from '../src/personal-graph' } from '../src/personal-graph'
describe('extractTerms', () => { describe('extractTerms', () => {
@ -137,3 +141,49 @@ describe('rankRelated', () => {
expect(SHARES_TERMS_MIN_SIMILARITY).toBeLessThan(1) expect(SHARES_TERMS_MIN_SIMILARITY).toBeLessThan(1)
}) })
}) })
describe('stemContinuationsFrom · 2-hop 그래프 · 개인화 후보 재정렬', () => {
it('조사가 달라도 어간 열이 일치하면 그 뒤의 이어쓰기를 추출한다', () => {
const out = stemContinuationsFrom(
['회의록 정리해서 내일 오전까지 공유드리겠습니다'],
'이번 주간 회의록을 정리해서'
)
expect(out).toEqual(['내일 오전까지 공유드리겠습니다'])
})
it('2-hop 경로(shares_terms -> follows)로 유사 문장 뒤에 실제로 썼던 후속 문장을 찾는다', () => {
const queryTerms = buildQueryTerms([{ text: '금요일 배포 준비 끝났습니다', weight: 2 }])
const targets = rankTwoHopGraphTargets({
firstHopEdges: [
{
from: '금요일 배포 준비 완료했습니다',
to: '스테이징 배포 점검 끝났습니다',
kind: 'shares_terms',
weight: 2
}
],
secondHopFollowEdges: [
{
from: '스테이징 배포 점검 끝났습니다',
to: '모니터링 대시보드 링크 공유드립니다',
kind: 'follows',
weight: 3
}
],
queryTerms
})
expect(targets).toEqual(['모니터링 대시보드 링크 공유드립니다'])
})
it('개인 그래프 힌트 및 어투와 일치하는 후보에 더 높은 개인화 점수를 준다', () => {
const context = {
prefix: '오늘 배포 일정 공유드립니다. ',
followUpHints: ['모니터링 대시보드 확인 부탁드립니다'],
relatedHints: ['릴리즈 노트 정리해서 올렸습니다']
}
const matchedScore = scoreCandidatePersonalization('모니터링 대시보드 확인 부탁드립니다.', context)
const mismatchedRegisterScore = scoreCandidatePersonalization('그냥 빨리 해봐.', context)
expect(matchedScore).toBeGreaterThan(4)
expect(matchedScore).toBeGreaterThan(mismatchedRegisterScore)
})
})

View file

@ -51,4 +51,10 @@ describe('PersonalNgramModel', () => {
const m = model('하나 둘 셋 넷 다섯', '하나 둘 셋 넷 다섯') const m = model('하나 둘 셋 넷 다섯', '하나 둘 셋 넷 다섯')
expect(m.predict('하나 ', { maxWords: 2 })?.text).toBe('둘 셋') expect(m.predict('하나 ', { maxWords: 2 })?.text).toBe('둘 셋')
}) })
it('한국어 어절의 조사가 달라져도 어간 보조 바이그램으로 다음 어절을 예측한다', () => {
const m = new PersonalNgramModel()
m.addText('회의록은 내일 공유드릴게요', 3)
expect(m.predict('이번 회의록을 ')).toMatchObject({ text: '내일 공유드릴게요', joint: 'separate' })
})
}) })

View file

@ -13,12 +13,14 @@ import {
fixByVocabulary, fixByVocabulary,
fixKnownMisspelling, fixKnownMisspelling,
hasSuspiciousTypo, hasSuspiciousTypo,
isCommonEnglishWord,
mergeTypoCorrections, mergeTypoCorrections,
hangulToLatin, hangulToLatin,
isPlausibleTypoFix, isPlausibleTypoFix,
latinToHangul, latinToHangul,
parseModelTypoPairs, parseModelTypoPairs,
planTypoReplacement, planTypoReplacement,
rebaseTypoReplacementPlan,
alignCorrectionToPrefix, alignCorrectionToPrefix,
typoRegionStart, typoRegionStart,
validateModelCorrections validateModelCorrections
@ -663,4 +665,50 @@ describe('Araso 급 한/영 자동 전환 · 합성어 · 잔류 Shift · URL
deleteCount: 'dkssudgktpdy '.length deleteCount: 'dkssudgktpdy '.length
}) })
}) })
it('생산적 영어 접미사·접두사 파생형과 문장 첫머리의 구문 호응(to be, to the, for the, and the)을 복원한다', () => {
expect(isCommonEnglishWord('implementation')).toBe(true)
expect(isCommonEnglishWord('personalization')).toBe(true)
expect(isCommonEnglishWord('unhappy')).toBe(true)
expect(isCommonEnglishWord('restart')).toBe(true)
expect(isCommonEnglishWord('disconnect')).toBe(true)
expect(isCommonEnglishWord('effectively')).toBe(true)
expect(isCommonEnglishWord('subagents')).toBe(true)
expect(applyAll('ㅑㅡㅔㅣ드둣ㅁ샤ㅐㅜ ')).toBe('implementation ')
expect(applyAll('ㅔㄷㄱ내ㅜ미ㅑㅋㅁ샤ㅐㅜ ')).toBe('personalization ')
expect(applyAll('ㅕㅜㅗ메ㅔㅛ ')).toBe('unhappy ')
expect(applyAll('ㄱㄷㄴㅅㅁㄱㅅ ')).toBe('restart ')
expect(applyAll('녀ㅠㅁㅎ둣ㄴ ')).toBe('subagents ')
// 문장 맨 앞에서도 "새 ㅠㄷ", "새 솓", "새 애 this", "랙 솓", "뭉 솓" 구문 호응을 복원한다
expect(applyAll('새 ㅠㄷ honest ')).toBe('to be honest ')
expect(applyAll('새 솓 point ')).toBe('to the point ')
expect(applyAll('새 애 this ')).toBe('to do this ')
expect(applyAll('랙 솓 team ')).toBe('for the team ')
expect(applyAll('뭉 솓 result ')).toBe('and the result ')
expect(applyAll('좨 ㅑㄴ there ')).toBe('who is there ')
// 한국어 오타 사전 확장 (괜찬아, 있읍니다, 없읍니다, 했읍니다, 금새)
expect(fixKnownMisspelling('괜찬아요')).toBe('괜찮아요')
expect(fixKnownMisspelling('있읍니다')).toBe('있습니다')
expect(fixKnownMisspelling('없읍니다')).toBe('없습니다')
expect(fixKnownMisspelling('했읍니다')).toBe('했습니다')
expect(fixKnownMisspelling('금새')).toBe('금세')
})
it('rebaseTypoReplacementPlan 은 비동기 모델 검사 중 덧붙은 짧은 꼬리 위로 교정 계획을 안전하게 리베이스한다', () => {
const base = planTypoReplacement('ㅡㅁ가둥', [
{ start: 0, end: 4, from: 'ㅡㅁ가둥', to: 'markdown', source: 'model' }
])
expect(base).not.toBeNull()
const rebased = rebaseTypoReplacementPlan(base!, 'ㅡㅁ가둥', 'ㅡㅁ가둥 문서 ', 18)
expect(rebased).toEqual({
deleteCount: 'ㅡㅁ가둥 문서 '.length,
insertText: 'markdown 문서 ',
originalText: 'ㅡㅁ가둥 문서 ',
corrections: base!.corrections
})
expect(rebaseTypoReplacementPlan(base!, 'ㅡㅁ가둥', '전혀다른문장', 18)).toBeNull()
})
}) })

View file

@ -10,6 +10,7 @@
import type { PhraseSource } from './input-intelligence' import type { PhraseSource } from './input-intelligence'
import { countLexicalWords, getWordSegmenter } from './text-segmentation' import { countLexicalWords, getWordSegmenter } from './text-segmentation'
import { detectSentenceRegister } from './writing-context'
// 렌더러/프리로드도 같은 계약을 쓴다 (main 모듈을 import 할 수 없다). // 렌더러/프리로드도 같은 계약을 쓴다 (main 모듈을 import 할 수 없다).
export type { PhraseSource } export type { PhraseSource }
@ -255,6 +256,74 @@ export function continuationsFrom(
return out return out
} }
const EDGE_PUNCT = /^[^\p{L}\p{N}]+|[^\p{L}\p{N}]+$/gu
function normalizeWordStem(rawWord: string): string {
const cleaned = rawWord.normalize('NFC').replace(EDGE_PUNCT, '').toLowerCase()
return cleaned ? stemTerm(cleaned) : ''
}
/**
* 조사가 달라 정확한 부분 문자열(`continuationsFrom`)이 어긋날 때 어간(stem) 열이 일치하는 지점 뒤의 이어쓰기를 뽑는다.
* 예: 저장 문장 "회의록 정리해서 내일 오전까지 공유드리겠습니다" ← 입력 꼬리 "회의록을 정리해서" → "내일 오전까지 공유드리겠습니다"
*/
export function stemContinuationsFrom(
texts: readonly string[],
tail: string,
limit = 3,
maxChars = 60
): string[] {
const rawTailWords = tail
.trim()
.split(/\s+/u)
.map(normalizeWordStem)
.filter((w) => w.length >= 2)
if (rawTailWords.length === 0) return []
const tailStems = rawTailWords.slice(-3)
if (tailStems.length === 1 && tailStems[0].length < 2) return []
const out: string[] = []
for (const text of texts) {
const spans: Array<{ stem: string; end: number }> = []
const wordRegex = /\S+/gu
for (let match = wordRegex.exec(text); match !== null; match = wordRegex.exec(text)) {
const stem = normalizeWordStem(match[0])
if (stem.length > 0) {
spans.push({ stem, end: match.index + match[0].length })
}
}
if (spans.length <= tailStems.length) continue
for (let len = tailStems.length; len >= Math.min(2, tailStems.length); len -= 1) {
const needleStems = tailStems.slice(-len)
for (let i = 0; i <= spans.length - len - 1; i += 1) {
let matched = true
for (let k = 0; k < len; k += 1) {
if (spans[i + k].stem !== needleStems[k]) {
matched = false
break
}
}
if (!matched) continue
const sliceStart = spans[i + len - 1].end
const continuation = text
.slice(sliceStart)
.replace(/^[,.:;!?~\s]+/u, '')
.replace(/\s+/gu, ' ')
.trim()
if (continuation.length < 3) continue
const clipped = continuation.slice(0, maxChars)
if (!out.includes(clipped)) {
out.push(clipped)
if (out.length >= limit) return out
}
break
}
}
}
return out
}
export interface RelatedCandidate { export interface RelatedCandidate {
text: string text: string
terms: readonly string[] terms: readonly string[]
@ -390,3 +459,128 @@ export function buildQueryTerms(parts: ReadonlyArray<{ text: string; weight: num
} }
return terms return terms
} }
export interface TwoHopGraphEdge {
from: string
to: string
kind: GraphEdgeKind
weight: number
}
/**
* 2-hop 그래프 경로 탐색 (A --shares_terms--> B --follows--> C 또는 A --follows--> B --follows--> C).
*
* 직전 문장(또는 지금 문장)과 직접 이어진 follows 엣지가 없더라도,
* 같은 주제를 공유하는 이웃 문장(shares_terms) 뒤에 실제로 이어 썼던 문장(follows)을
* 경로 가중치와 질의 용어 겹침으로 점수화해 끌어온다.
*/
export function rankTwoHopGraphTargets(params: {
firstHopEdges: readonly TwoHopGraphEdge[]
secondHopFollowEdges: readonly TwoHopGraphEdge[]
queryTerms: ReadonlyMap<string, number>
exclude?: (text: string) => boolean
limit?: number
}): string[] {
const { firstHopEdges, secondHopFollowEdges, queryTerms, exclude, limit = 4 } = params
if (firstHopEdges.length === 0 || secondHopFollowEdges.length === 0) return []
const firstHopScores = new Map<string, number>()
for (const edge of firstHopEdges) {
const kindMultiplier = edge.kind === 'follows' ? 1.5 : 1.0
const score = Math.log2(1 + Math.max(1, edge.weight)) * kindMultiplier
firstHopScores.set(edge.to, Math.max(firstHopScores.get(edge.to) ?? 0, score))
}
const targetScores = new Map<string, number>()
for (const edge of secondHopFollowEdges) {
if (edge.kind !== 'follows') continue
if (exclude?.(edge.to)) continue
const hop1 = firstHopScores.get(edge.from)
if (hop1 === undefined) continue
const hop2 = Math.log2(1 + Math.max(1, edge.weight)) * 1.4
const targetTerms = extractTerms(edge.to, 8)
let termBonus = 0
for (const term of targetTerms) {
termBonus += (queryTerms.get(term) ?? 0) * 0.6
}
const totalScore = hop1 * hop2 + termBonus
targetScores.set(edge.to, (targetScores.get(edge.to) ?? 0) + totalScore)
}
return [...targetScores.entries()]
.sort((a, b) => b[1] - a[1])
.slice(0, limit)
.map(([text]) => text)
}
export interface CandidatePersonalizationContext {
prefix: string
continuationHints?: readonly string[]
followUpHints?: readonly string[]
relatedHints?: readonly string[]
phraseHints?: readonly string[]
}
/**
* 생성된 제안 후보가 사용자의 개인 그래프(이어쓰기·직후 문장·관련 문장·자주 쓰는 표현) 및 현재 어투와
* 얼마나 잘 맞는지 점수화한다 (후보 재정렬용).
*/
export function scoreCandidatePersonalization(
candidateText: string,
context: CandidatePersonalizationContext
): number {
const normalizedCandidate = candidateText.replace(/\s+/gu, ' ').trim()
if (!normalizedCandidate) return 0
let score = 0
const candidateTerms = new Set(extractTerms(normalizedCandidate, 10))
for (const hint of context.continuationHints ?? []) {
const normHint = hint.replace(/\s+/gu, ' ').trim()
if (!normHint) continue
if (normalizedCandidate === normHint || normalizedCandidate.startsWith(normHint) || normHint.startsWith(normalizedCandidate)) {
score += 4.0
break
}
}
for (const hint of context.followUpHints ?? []) {
const normHint = hint.replace(/\s+/gu, ' ').trim()
if (!normHint) continue
if (normalizedCandidate === normHint || normalizedCandidate.startsWith(normHint) || normHint.startsWith(normalizedCandidate)) {
score += 3.5
break
}
}
const addOverlapScore = (hints: readonly string[] | undefined, weightPerTerm: number): void => {
if (!hints || hints.length === 0 || candidateTerms.size === 0) return
let matched = 0
for (const hint of hints) {
for (const term of extractTerms(hint, 8)) {
if (candidateTerms.has(term)) matched += 1
}
}
score += Math.min(matched, 5) * weightPerTerm
}
addOverlapScore(context.continuationHints, 1.8)
addOverlapScore(context.followUpHints, 1.6)
addOverlapScore(context.relatedHints, 1.0)
addOverlapScore(context.phraseHints, 0.8)
const prefixRegister = detectSentenceRegister(context.prefix)
const candidateRegister = detectSentenceRegister(normalizedCandidate)
if (prefixRegister && candidateRegister) {
if (prefixRegister === candidateRegister) {
score += 0.8
} else if (
(prefixRegister === 'casual' && (candidateRegister === 'formal' || candidateRegister === 'polite')) ||
((prefixRegister === 'formal' || prefixRegister === 'polite') && candidateRegister === 'casual')
) {
score -= 1.5
}
}
return score
}

View file

@ -12,6 +12,8 @@
// - 지지(같은 문맥에서 나온 횟수)와 비율(문맥 전체 중 그 어절의 몫)이 기준을 넘을 때만 낸다. // - 지지(같은 문맥에서 나온 횟수)와 비율(문맥 전체 중 그 어절의 몫)이 기준을 넘을 때만 낸다.
// - 확신이 이어지면 최대 maxWords 어절까지 구로 늘린다(트라이그램 → 바이그램 순). // - 확신이 이어지면 최대 maxWords 어절까지 구로 늘린다(트라이그램 → 바이그램 순).
import { stemTerm } from './personal-graph'
export interface PersonalNgramOptions { export interface PersonalNgramOptions {
/** 같은 문맥 뒤에 이 어절이 나온 최소 횟수 */ /** 같은 문맥 뒤에 이 어절이 나온 최소 횟수 */
minSupport?: number minSupport?: number
@ -73,13 +75,13 @@ function tokenize(sentence: string): string[] {
.filter((word) => word.length > 0) .filter((word) => word.length > 0)
} }
function increment(table: Map<string, Map<string, number>>, context: string, word: string): void { function increment(table: Map<string, Map<string, number>>, context: string, word: string, weight = 1): void {
let followers = table.get(context) let followers = table.get(context)
if (!followers) { if (!followers) {
followers = new Map() followers = new Map()
table.set(context, followers) table.set(context, followers)
} }
followers.set(word, (followers.get(word) ?? 0) + 1) followers.set(word, (followers.get(word) ?? 0) + weight)
} }
interface Pick { interface Pick {
@ -104,6 +106,8 @@ function best(followers: Map<string, number> | undefined, accept: (word: string)
export class PersonalNgramModel { export class PersonalNgramModel {
private readonly _bigram = new Map<string, Map<string, number>>() private readonly _bigram = new Map<string, Map<string, number>>()
private readonly _trigram = new Map<string, Map<string, number>>() private readonly _trigram = new Map<string, Map<string, number>>()
/** 한국어 어절의 조사가 달라졌을 때("회의록을" vs "회의록은") 어간 기준으로 다음 어절을 찾는 보조 테이블 */
private readonly _stemBigram = new Map<string, Map<string, number>>()
/** 단어 중간 완성용 — 문맥 없이 어절 빈도 */ /** 단어 중간 완성용 — 문맥 없이 어절 빈도 */
private readonly _unigram = new Map<string, number>() private readonly _unigram = new Map<string, number>()
private _texts = 0 private _texts = 0
@ -112,15 +116,25 @@ export class PersonalNgramModel {
return this._texts return this._texts
} }
addText(text: string): void { addText(text: string, weight = 1): void {
const safeWeight = Math.max(1, Math.trunc(weight))
this._texts += 1 this._texts += 1
for (const sentence of text.split(SENTENCE_SPLIT)) { for (const sentence of text.split(SENTENCE_SPLIT)) {
const words = tokenize(sentence) const words = tokenize(sentence)
for (let index = 0; index < words.length; index += 1) { for (let index = 0; index < words.length; index += 1) {
const word = words[index] const word = words[index]
this._unigram.set(word, (this._unigram.get(word) ?? 0) + 1) this._unigram.set(word, (this._unigram.get(word) ?? 0) + safeWeight)
if (index >= 1) increment(this._bigram, words[index - 1], word) if (index >= 1) {
if (index >= 2) increment(this._trigram, `${words[index - 2]} ${words[index - 1]}`, word) const prev = words[index - 1]
increment(this._bigram, prev, word, safeWeight)
const prevStem = stemTerm(prev)
if (prevStem !== prev && prevStem.length >= 2) {
increment(this._stemBigram, prevStem, word, safeWeight)
}
}
if (index >= 2) {
increment(this._trigram, `${words[index - 2]} ${words[index - 1]}`, word, safeWeight)
}
} }
} }
} }
@ -184,8 +198,14 @@ export class PersonalNgramModel {
if (passes(tri)) return tri if (passes(tri)) return tri
} }
if (context.length >= 1) { if (context.length >= 1) {
const bi = best(this._bigram.get(context[context.length - 1]), accept) const lastWord = context[context.length - 1]
const bi = best(this._bigram.get(lastWord), accept)
if (passes(bi)) return bi if (passes(bi)) return bi
const lastStem = stemTerm(lastWord)
if (lastStem !== lastWord && lastStem.length >= 2) {
const stemBi = best(this._stemBigram.get(lastStem), accept)
if (passes(stemBi) && stemBi.confidence >= 0.5) return stemBi
}
} }
return null return null
} }

View file

@ -18,7 +18,6 @@ import {
finalSoundOfAcronym, finalSoundOfAcronym,
finalSoundOfHangul, finalSoundOfHangul,
isUrlOrPathRun, isUrlOrPathRun,
lingersAfterKoreanShift,
normalizeLingeringShift, normalizeLingeringShift,
particleAgrees, particleAgrees,
splitMixedWord splitMixedWord
@ -448,80 +447,204 @@ const COMMON_ENGLISH_WORDS: ReadonlySet<string> = new Set(
).split(/\s+/u) ).split(/\s+/u)
) )
/** const EXTENDED_ENGLISH_STEMS: ReadonlySet<string> = new Set(
* 흔한 영단어인가 (소문자 비교 + 규칙 굴절형 지원). (
* 'string strings agent agents subagent subagents prompt prompts context contexts session sessions ' +
* -s / -es / -ies / -ed / -ied / -ing / -ly / -er / -ers / -est 굴절형은 'stream streaming vector vectors graph graphs node nodes edge edges index indexed indexing ' +
* 두벌식 한글 음절로 완전히 조합되지 않는 경우(latinToHangul(lower).complete === false)에 한해 'rerank ranking score scores weight weights metric metrics latency throughput memory cache cached ' +
* 어간이 상용 영단어면 영단어로 인정한다. 'buffer queue worker workers atomic batch chunk chunks embed embedding embeddings inference ' +
*/ 'dataset datasets benchmark benchmarks eval evaluate evaluation verify verified validate validation ' +
export function isCommonEnglishWord(word: string): boolean { 'schema schemas migration migrations cursor caret selection overlay modal dialog toast banner ' +
const lower = word.toLowerCase() 'widget panel route routes router routing hook hooks store stores slice props handler handlers ' +
if (COMMON_ENGLISH_WORDS.has(lower)) return true 'listener listeners channel channels bridge adapter service services controller provider repository ' +
if (lower.length < 4 || !/^[a-z]+$/u.test(lower)) return false 'factory singleton instance module modules package packages bundle bundler version versions ' +
// 두벌식으로 빈틈없이 상용 음절이 되는 키열(예: 한글 단어 키열)은 굴절 추정으로 영단어 처리하지 않는다. 'patch minor major rebase stash staging canary hotfix bugfix refactor cleanup format linter ' +
if (latinToHangul(lower).complete) return false 'interface interfaces generic generics tuple array object boolean symbol closure promise callback ' +
'awesome amazing perfect wonderful terrible horrible strange funny crazy busy tired hungry ' +
'excited nervous worried confused surprised interested bored annoyed upset glad proud afraid ' +
'scared careful useful helpful powerful comfortable impossible separate obvious accurate correct ' +
'valid regular formal visible stable healthy lucky usual expected connect connection suggest ' +
'suggestion complete completion replace replacement require requirement assign assignment ' +
'implement implementation configure configuration optimize optimization personalize personalization ' +
'normalize custom customize summarize synchronize'
).split(/\s+/u)
)
if (lower.endsWith('ies') && lower.length >= 4 && COMMON_ENGLISH_WORDS.has(`${lower.slice(0, -3)}y`)) { const ENGLISH_DERIVATIONAL_PREFIXES: readonly string[] = [
'un',
're',
'pre',
'non',
'sub',
'over',
'under',
'out',
'mis',
'dis',
'auto',
'multi',
'micro',
'super',
'inter',
'hyper'
]
function hasEnglishLexiconStem(word: string): boolean {
return COMMON_ENGLISH_WORDS.has(word) || EXTENDED_ENGLISH_STEMS.has(word)
}
function matchesInflectedOrDerivedEnglish(lower: string): boolean {
if (hasEnglishLexiconStem(lower)) return true
if (lower.length < 4) return false
if (lower.endsWith('ies') && lower.length >= 4 && hasEnglishLexiconStem(`${lower.slice(0, -3)}y`)) {
return true return true
} }
if (lower.endsWith('es') && lower.length >= 4) { if (lower.endsWith('es') && lower.length >= 4) {
if (COMMON_ENGLISH_WORDS.has(lower.slice(0, -2)) || COMMON_ENGLISH_WORDS.has(lower.slice(0, -1))) { if (hasEnglishLexiconStem(lower.slice(0, -2)) || hasEnglishLexiconStem(lower.slice(0, -1))) {
return true return true
} }
} }
if (lower.endsWith('s') && !lower.endsWith('ss') && COMMON_ENGLISH_WORDS.has(lower.slice(0, -1))) { if (lower.endsWith('s') && !lower.endsWith('ss') && hasEnglishLexiconStem(lower.slice(0, -1))) {
return true return true
} }
if (lower.endsWith('ied') && lower.length >= 4 && COMMON_ENGLISH_WORDS.has(`${lower.slice(0, -3)}y`)) { if (lower.endsWith('ied') && lower.length >= 4 && hasEnglishLexiconStem(`${lower.slice(0, -3)}y`)) {
return true return true
} }
if (lower.endsWith('ed') && lower.length >= 4) { if (lower.endsWith('ed') && lower.length >= 4) {
if (COMMON_ENGLISH_WORDS.has(lower.slice(0, -2)) || COMMON_ENGLISH_WORDS.has(lower.slice(0, -1))) { if (hasEnglishLexiconStem(lower.slice(0, -2)) || hasEnglishLexiconStem(lower.slice(0, -1))) {
return true return true
} }
if ( if (
lower.length >= 5 && lower.length >= 5 &&
lower[lower.length - 3] === lower[lower.length - 4] && lower[lower.length - 3] === lower[lower.length - 4] &&
COMMON_ENGLISH_WORDS.has(lower.slice(0, -3)) hasEnglishLexiconStem(lower.slice(0, -3))
) { ) {
return true return true
} }
} }
if (lower.endsWith('ing') && lower.length >= 5) { if (lower.endsWith('ing') && lower.length >= 5) {
const stem = lower.slice(0, -3) const stem = lower.slice(0, -3)
if (COMMON_ENGLISH_WORDS.has(stem) || COMMON_ENGLISH_WORDS.has(`${stem}e`)) { if (hasEnglishLexiconStem(stem) || hasEnglishLexiconStem(`${stem}e`)) {
return true return true
} }
if ( if (
stem.length >= 3 && stem.length >= 3 &&
stem[stem.length - 1] === stem[stem.length - 2] && stem[stem.length - 1] === stem[stem.length - 2] &&
COMMON_ENGLISH_WORDS.has(stem.slice(0, -1)) hasEnglishLexiconStem(stem.slice(0, -1))
) { ) {
return true return true
} }
} }
if (lower.endsWith('ily') && lower.length >= 5 && COMMON_ENGLISH_WORDS.has(`${lower.slice(0, -3)}y`)) { if (lower.endsWith('ily') && lower.length >= 5 && hasEnglishLexiconStem(`${lower.slice(0, -3)}y`)) {
return true return true
} }
if (lower.endsWith('ly') && lower.length >= 5 && COMMON_ENGLISH_WORDS.has(lower.slice(0, -2))) { if (lower.endsWith('ly') && lower.length >= 5 && hasEnglishLexiconStem(lower.slice(0, -2))) {
return true return true
} }
if (lower.endsWith('ers') && lower.length >= 5) { if (lower.endsWith('ers') && lower.length >= 5) {
if (COMMON_ENGLISH_WORDS.has(lower.slice(0, -3)) || COMMON_ENGLISH_WORDS.has(lower.slice(0, -2))) { if (hasEnglishLexiconStem(lower.slice(0, -3)) || hasEnglishLexiconStem(lower.slice(0, -2))) {
return true return true
} }
} }
if (lower.endsWith('er') && lower.length >= 4) { if (lower.endsWith('er') && lower.length >= 4) {
if (COMMON_ENGLISH_WORDS.has(lower.slice(0, -2)) || COMMON_ENGLISH_WORDS.has(lower.slice(0, -1))) { if (hasEnglishLexiconStem(lower.slice(0, -2)) || hasEnglishLexiconStem(lower.slice(0, -1))) {
return true return true
} }
} }
if (lower.endsWith('est') && lower.length >= 5) { if (lower.endsWith('est') && lower.length >= 5) {
if (COMMON_ENGLISH_WORDS.has(lower.slice(0, -3)) || COMMON_ENGLISH_WORDS.has(lower.slice(0, -2))) { if (hasEnglishLexiconStem(lower.slice(0, -3)) || hasEnglishLexiconStem(lower.slice(0, -2))) {
return true return true
} }
} }
// 생산적 파생 접미사 (-tion, -ation, -ment, -ness, -able, -ible, -ful, -fully, -less, -ive, -ively, -ous, -ously, -al, -ally, -ic, -ically, -ize, -ized, -izing, -or, -ors, -ist, -ism)
const core = lower.endsWith('s') && lower.length >= 6 ? lower.slice(0, -1) : lower
for (const [suffix, restoreE] of [
['ations', true],
['ation', true],
['itions', true],
['ition', true],
['tions', true],
['tion', true],
['sions', true],
['sion', true],
['ments', true],
['ment', true],
['nesses', false],
['ness', false],
['ables', true],
['able', true],
['ably', true],
['ibles', true],
['ible', true],
['fully', false],
['fulness', false],
['ful', false],
['lessly', false],
['lessness', false],
['less', false],
['ively', true],
['ives', true],
['ive', true],
['ously', true],
['ous', true],
['ally', true],
['als', true],
['al', true],
['ically', true],
['ical', true],
['ics', true],
['ic', true],
['izations', true],
['ization', true],
['izing', true],
['ized', true],
['izes', true],
['ize', true],
['ising', true],
['ised', true],
['ises', true],
['ise', true],
['ors', true],
['or', true],
['ists', true],
['ist', true],
['isms', true],
['ism', true]
] as const) {
for (const target of [lower, core]) {
if (!target.endsWith(suffix) || target.length - suffix.length < 3) continue
const stem = target.slice(0, -suffix.length)
if (hasEnglishLexiconStem(stem)) return true
if (restoreE && hasEnglishLexiconStem(`${stem}e`)) return true
if (stem.endsWith('i') && hasEnglishLexiconStem(`${stem.slice(0, -1)}y`)) return true
}
}
return false
}
/**
* 흔한 영단어인가 (소문자 비교 + 규칙 굴절형·파생형 지원).
*
* 굴절·파생 접미사 및 접두사는 두벌식 한글 음절로 완전히 조합되지 않는 경우(latinToHangul(lower).complete === false)에 한해
* 어간이 상용 영단어면 영단어로 인정한다.
*/
export function isCommonEnglishWord(word: string): boolean {
const lower = word.toLowerCase()
if (hasEnglishLexiconStem(lower)) return true
if (lower.length < 4 || !/^[a-z]+$/u.test(lower)) return false
// 두벌식으로 빈틈없이 상용 음절이 되는 키열(예: 한글 단어 키열)은 굴절 추정으로 영단어 처리하지 않는다.
if (latinToHangul(lower).complete) return false
if (matchesInflectedOrDerivedEnglish(lower)) return true
for (const prefix of ENGLISH_DERIVATIONAL_PREFIXES) {
if (lower.startsWith(prefix) && lower.length - prefix.length >= 4) {
const rest = lower.slice(prefix.length)
if (matchesInflectedOrDerivedEnglish(rest)) return true
}
}
return false return false
} }
@ -540,6 +663,8 @@ export interface LayoutDetectOptions {
afterEnglishWord?: boolean afterEnglishWord?: boolean
/** 바로 뒤 토큰이 영문(또는 한글 자판으로 친 영문)인가 */ /** 바로 뒤 토큰이 영문(또는 한글 자판으로 친 영문)인가 */
nextIsEnglish?: boolean nextIsEnglish?: boolean
/** 바로 뒤 토큰의 복원된 영단어(소문자) — "새 ㅠㄷ"→"to be", "새 솓"→"to the" 등 구문 호응 판별용 */
nextEnglishWord?: string | null
/** 토큰 바로 뒤에 공백·문장부호 등 단어 경계가 이미 확정되었는가 (예: "ㅑ " → "I ") */ /** 토큰 바로 뒤에 공백·문장부호 등 단어 경계가 이미 확정되었는가 (예: "ㅑ " → "I ") */
hasWordBoundaryAfter?: boolean hasWordBoundaryAfter?: boolean
/** 슬래시 쌍(ska/du)·약어 조사(L.A.fh)처럼 1음절 변환을 폭넓게 허용할 문맥인가 */ /** 슬래시 쌍(ska/du)·약어 조사(L.A.fh)처럼 1음절 변환을 폭넓게 허용할 문맥인가 */
@ -665,7 +790,6 @@ export function detectLatinTypedHangul(
if (syllables.length === 0) return null if (syllables.length === 0) return null
if (!syllables.every((char) => isCommonHangulSyllable(char))) return null if (!syllables.every((char) => isCommonHangulSyllable(char))) return null
const contextTotal = context.hangul + context.latin
const hangulContext = const hangulContext =
options.afterKoreanWord === true || options.afterKoreanWord === true ||
context.hangul >= 2 || context.hangul >= 2 ||
@ -1054,17 +1178,27 @@ const STANDALONE_SYLLABLE_ENGLISH: Readonly<Record<string, string>> = {
랴그: 'firm', 랴그: 'firm',
랴노: 'fish', 랴노: 'fish',
래그: 'form', 래그: 'form',
래가: 'fork',
햐기: 'girl', 햐기: 'girl',
해미: 'goal', 해미: 'goal',
갸초: 'rich', 갸초: 'rich',
갸나: 'risk', 갸나: 'risk',
개차: 'rock', 개차: 'rock',
애차: 'dock',
냐차: 'sick',
샤차: 'tick',
녀차: 'suck',
셔차: 'tuck',
자노: 'wash',
차노: 'cash',
겨노: 'rush',
노맏: 'shake', 노맏: 'shake',
냐후: 'sign', 냐후: 'sign',
샛미: 'total', 샛미: 'total',
셔구: 'turn', 셔구: 'turn',
츠맏: 'cmake', 츠맏: 'cmake',
녀애: 'sudo', 녀애: 'sudo',
좨: 'who',
먁: 'air', 먁: 'air',
앳: 'dot', 앳: 'dot',
므: 'am' 므: 'am'
@ -1087,7 +1221,14 @@ const CONTEXTUAL_SYLLABLE_ENGLISH: Readonly<Record<string, string>> = {
새주: 'town', 새주: 'town',
챠쇼: 'city', 챠쇼: 'city',
챔초: 'coach', 챔초: 'coach',
쳐기: 'curl' 쳐기: 'curl',
뭉: 'and',
랙: 'for',
야노: 'dish',
다노: 'dash',
내차: 'sock',
야나: 'disk',
사나: 'task'
} }
/** /**
@ -1106,6 +1247,97 @@ const ENGLISH_CLAUSE_SINGLE_SYLLABLE: Readonly<Record<string, string>> = {
둥: 'end' 둥: 'end'
} }
/**
* 한국어 관형사 '새'(new)가 절대 수식할 수 없는 영어 동사 원형·대명사·관사.
* 문장 맨 앞이라도 '새' 바로 뒤에 이 단어가 오면("새 ㅠㄷ" → "to be", "새 애" → "to do", "새 솓" → "to the", "새 ㅡㄷ" → "to me")
* 100% 영문 부정사/전치사 'to' 로 복원한다 ("오늘 새 API를" 의 'API를' 은 여기에 없으므로 보호됨).
*/
const TO_FOLLOWING_ENGLISH_WORDS: ReadonlySet<string> = new Set([
'be',
'do',
'go',
'have',
'get',
'make',
'see',
'know',
'take',
'find',
'give',
'tell',
'try',
'ask',
'need',
'feel',
'leave',
'put',
'keep',
'let',
'begin',
'seem',
'help',
'talk',
'start',
'hear',
'run',
'move',
'live',
'hold',
'bring',
'write',
'sit',
'stand',
'lose',
'pay',
'meet',
'learn',
'lead',
'watch',
'follow',
'stop',
'speak',
'read',
'allow',
'add',
'spend',
'grow',
'open',
'walk',
'win',
'offer',
'remember',
'buy',
'wait',
'serve',
'send',
'expect',
'build',
'stay',
'reach',
'remain',
'use',
'work',
'check',
'fix',
'test',
'deploy',
'the',
'me',
'you',
'us',
'them',
'him',
'her',
'it',
'this',
'that',
'my',
'your',
'our',
'their',
'his'
])
const SURROUNDED_SINGLE_SYLLABLE_ENGLISH: Readonly<Record<string, string>> = { const SURROUNDED_SINGLE_SYLLABLE_ENGLISH: Readonly<Record<string, string>> = {
새: 'to', 새: 'to',
무: 'an', 무: 'an',
@ -1329,7 +1561,19 @@ export function detectHangulTypedLatin(
if ( if (
clauseSingle && clauseSingle &&
((options.afterEnglishWord && (options.hasWordBoundaryAfter || options.nextIsEnglish || context.latin >= 3)) || ((options.afterEnglishWord && (options.hasWordBoundaryAfter || options.nextIsEnglish || context.latin >= 3)) ||
(options.nextIsEnglish && context.latin >= 4 && context.latin > context.hangul * 2)) (options.nextIsEnglish && context.latin >= 4 && context.latin > context.hangul * 2) ||
(word === '새' &&
!options.afterKoreanWord &&
options.nextEnglishWord != null &&
TO_FOLLOWING_ENGLISH_WORDS.has(options.nextEnglishWord)) ||
(word === '애' &&
!options.afterKoreanWord &&
options.nextEnglishWord != null &&
(options.nextEnglishWord === 'this' ||
options.nextEnglishWord === 'that' ||
options.nextEnglishWord === 'it' ||
options.nextEnglishWord === 'not' ||
options.nextEnglishWord === 'you')))
) { ) {
return clauseSingle return clauseSingle
} }
@ -1570,6 +1814,12 @@ const KOREAN_MISSPELLINGS: readonly KoreanMisspellingRule[] = [
{ pattern: /되요$/u, replace: '돼요' }, { pattern: /되요$/u, replace: '돼요' },
{ pattern: /되서$/u, replace: '돼서' }, { pattern: /되서$/u, replace: '돼서' },
{ pattern: /뵈요$/u, replace: '봬요' }, { pattern: /뵈요$/u, replace: '봬요' },
{ pattern: /^괜찬(?=[아았으은다네지고])/u, replace: '괜찮' },
{ pattern: /있읍니다$/u, replace: '있습니다' },
{ pattern: /없읍니다$/u, replace: '없습니다' },
{ pattern: /했읍니다$/u, replace: '했습니다' },
{ pattern: /됐읍니다$/u, replace: '됐습니다' },
{ pattern: /^금새$/u, replace: '금세' },
{ pattern: /^몇일/u, replace: '며칠' }, { pattern: /^몇일/u, replace: '며칠' },
{ pattern: /^어떻해/u, replace: '어떡해' }, { pattern: /^어떻해/u, replace: '어떡해' },
{ pattern: /^왠만/u, replace: '웬만' }, { pattern: /^왠만/u, replace: '웬만' },
@ -1702,33 +1952,40 @@ function isIgnored(word: string, ignored: readonly string[] | undefined): boolea
return ignored.some((entry) => entry.toLowerCase() === lower) return ignored.some((entry) => entry.toLowerCase() === lower)
} }
function looksLikeEnglishOrHangulTypedEnglish(tokenText: string): boolean { function resolveTokenAsEnglishPreview(tokenText: string): string | null {
if (/^[A-Za-z']{2,}$/u.test(tokenText) && (isCommonEnglishWord(tokenText) || resolveContraction(tokenText) !== null)) { if (/^[A-Za-z']{2,}$/u.test(tokenText)) {
return true const resolved = resolveContraction(tokenText)
if (resolved) return resolved.toLowerCase()
if (isCommonEnglishWord(tokenText)) return tokenText.toLowerCase()
} }
if (LOOSE_JAMO_TWO_LETTER_ENGLISH[tokenText]) return true const twoLetter = LOOSE_JAMO_TWO_LETTER_ENGLISH[tokenText]
if ( if (twoLetter) return twoLetter.toLowerCase()
STANDALONE_SYLLABLE_ENGLISH[tokenText] || const syllableMapped =
CONTEXTUAL_SYLLABLE_ENGLISH[tokenText] || STANDALONE_SYLLABLE_ENGLISH[tokenText] ??
CONTEXTUAL_SYLLABLE_ENGLISH[tokenText] ??
ENGLISH_CLAUSE_SINGLE_SYLLABLE[tokenText] ENGLISH_CLAUSE_SINGLE_SYLLABLE[tokenText]
) { if (syllableMapped) return syllableMapped.toLowerCase()
return true
}
if (/[ㄱ-ㆎ]/u.test(tokenText)) { if (/[ㄱ-ㆎ]/u.test(tokenText)) {
const latin = hangulToLatin(tokenText).replace(/’/gu, "'") const latin = hangulToLatin(tokenText).replace(/’/gu, "'")
if (latin.includes("'") && (resolveContraction(latin) !== null || findFuzzyEnglishForLooseJamo(latin) !== null)) { if (latin.includes("'")) {
return true const resolved = resolveContraction(latin) ?? findFuzzyEnglishForLooseJamo(latin)
if (resolved) return resolved.toLowerCase()
} }
if ( if (/^[A-Za-z]{3,}$/u.test(latin)) {
/^[A-Za-z]{3,}$/u.test(latin) && if (isCommonEnglishWord(latin)) return latin.toLowerCase()
(isCommonEnglishWord(latin) || const known = fixKnownMisspelling(latin) ?? findFuzzyEnglishForLooseJamo(latin)
fixKnownMisspelling(latin) !== null || if (known) return known.toLowerCase()
findFuzzyEnglishForLooseJamo(latin) !== null) }
) { } else if (/^[가-힣]+$/u.test(tokenText)) {
return true const hasRare = Array.from(tokenText).some((ch) => !isCommonHangulSyllable(ch))
if (hasRare) {
const latin = hangulToLatin(tokenText)
if (/^[A-Za-z]{3,}$/u.test(latin) && isCommonEnglishWord(latin)) {
return latin.toLowerCase()
} }
} }
return false }
return null
} }
/** /**
@ -1813,7 +2070,8 @@ export function detectTypos(prefix: string, options: TypoDetectOptions = {}): Ty
break break
} }
} }
const nextIsEnglish = nextWordToken ? looksLikeEnglishOrHangulTypedEnglish(nextWordToken.text) : false const nextEnglishWord = nextWordToken ? resolveTokenAsEnglishPreview(nextWordToken.text) : null
const nextIsEnglish = nextEnglishWord !== null
const layout = const layout =
detectLatinTypedHangul(token.text, context, { detectLatinTypedHangul(token.text, context, {
@ -1825,6 +2083,7 @@ export function detectTypos(prefix: string, options: TypoDetectOptions = {}): Ty
afterKoreanWord, afterKoreanWord,
afterEnglishWord, afterEnglishWord,
nextIsEnglish, nextIsEnglish,
nextEnglishWord,
hasWordBoundaryAfter: /[ \t.,!?;:]/u.test(after), hasWordBoundaryAfter: /[ \t.,!?;:]/u.test(after),
vocabulary: options.vocabulary vocabulary: options.vocabulary
}) })
@ -2328,6 +2587,33 @@ export function planTypoReplacement(prefix: string, corrections: readonly TypoCo
return { deleteCount: originalText.length, insertText, originalText, corrections: sorted } return { deleteCount: originalText.length, insertText, originalText, corrections: sorted }
} }
/**
* 비동기 모델 오타 검사가 도는 동안 사용자가 공백·문장부호나 짧은 후속 어절(1~6자)을 더 쳤을 때
* 원래 교정 계획을 새 접두(currentPrefix) 위로 안전하게 리베이스(rebase)한다.
*/
export function rebaseTypoReplacementPlan(
plan: TypoReplacementPlan,
checkedPrefix: string,
currentPrefix: string,
maxDeleteChars: number = TYPO_REGION_MAX_CHARS
): TypoReplacementPlan | null {
if (currentPrefix === checkedPrefix) {
return plan.deleteCount <= maxDeleteChars ? plan : null
}
if (!currentPrefix.startsWith(checkedPrefix)) return null
const delta = currentPrefix.slice(checkedPrefix.length)
if (delta.length === 0 || delta.length > 6 || !isSafeToBackspace(delta)) return null
const newOriginalText = plan.originalText + delta
const newInsertText = plan.insertText + delta
if (newOriginalText.length > maxDeleteChars) return null
return {
deleteCount: newOriginalText.length,
insertText: newInsertText,
originalText: newOriginalText,
corrections: plan.corrections
}
}
/** /**
* 교정 뒤 사용자가 원래 단어로 되돌렸는가. * 교정 뒤 사용자가 원래 단어로 되돌렸는가.
* *