open-multi-agent-kit 0.90.4 → 0.90.5

package/dist/core/reasoning-router-v4.d.ts

@@ -1,13 +1,5 @@
 /**
- * Reasoning-router v4 — confidence-bearing classifier (Goal 009 Wave 1 Lane A,
- * specs/008-reasoning-router-advanced-accuracy Requirement 2).
- *
- * v4 is a NEW, OPT-IN, PURE sibling module. It does not import from, and is not
- * imported by, reasoning-router.ts (v1), reasoning-router-v2.ts,
- * reasoning-router-v3.ts, reasoning-router-weights.ts, or agent-session.ts.
- * Activating v4 in the product (a `/think auto-v4` mode, settings wiring, etc.)
- * is explicitly out of scope for this lane and is left to a later,
- * single-writer integration lane (specs/008 plan.md Wave 3).
+ * Reasoning-router v4 — the sole `/think auto` classifier/resolver.
  *
  * Deterministic by construction: same (input, weights) -> same
  * ClassifierVerdictV4; same (verdict, availableLevels, laneType, bias, hint) ->
@@ -15,57 +7,17 @@
  * or state mutation anywhere in this file (ThinkingLevel is a type-only import
  * and is erased).
  *
- * ============================================================================
- * WHAT'S NEW VS v3 (reasoning-router-v3.ts)
- * ============================================================================
- * 1. VERDICT, NOT JUST A CLASS. `classifyTaskV4` returns a `ClassifierVerdictV4`
- *    carrying the full per-class score map, the runner-up class, the margin,
- *    a normalized confidence ratio, a confidence band, a `tieBreak` flag (the
- *    top class won only via precedence order over an exact score tie), a
- *    `fallbackReason` (non-null exactly when the zero-score cascade decided
- *    the class instead of a real signal), and `suppressedFeatureIds` (an
- *    audit trail of which whole-prompt signals were nulled by bounded
- *    negation). None of this carries prompt text — every field is a bounded
- *    enum, number, boolean, or a closed set of short diagnostic-id strings.
- * 2. WEIGHTS TABLE, NOT LITERALS. Every score bump v3 hard-coded as an inline
- *    integer now reads from a `RouterWeightsV4` (see
- *    reasoning-router-v4-weights.ts). `DEFAULT_WEIGHTS_V4` reproduces v3's
- *    literals exactly, so classifyTaskV4(input, DEFAULT_WEIGHTS_V4).taskClass
- *    matches classifyTaskV3(input) on the full non-holdout GOLD_SET and every
- *    focused v3 sentinel (see the 013 regression test and this lane's
- *    evidence file). Passing a different `RouterWeightsV4` recalibrates
- *    without touching this file.
- * 3. BOUNDED NEGATION. Every whole-prompt keyword/contextual pattern scan
- *    (the ones v3 ran as a bare `PATTERN.test(prompt)`) is now negation-aware:
- *    a match is only counted if no negation cue (don't/doesn't/isn't/never/
- *    avoid/skip/without/instead of/rather than/no need to|for/not a/...) appears within
- *    `weights.negationWindowChars` characters immediately before it, and the
- *    scan never crosses a `.,;!?` clause boundary. Patterns anchored to the
- *    START of the leading clause (hasLeadingDebugAction, hasLeadingReviewIntent,
- *    etc.) are untouched and need no negation handling: "don't refactor" can
- *    never match an anchored `^refactor` test in the first place. Concrete
- *    fix: "don't refactor this, just fix the crash" now classifies as debug
- *    (v3 misclassifies this as refactor; see the 013 test).
- * 4. BOUNDED COMPOUND INTENT. A short-range conjunction split (bare "then",
- *    "and then", "and also", or ";", only within the first 300 characters)
- *    looks for a second, distinct leading-verb intent after the split point.
- *    When found and distinct from the primary leading intent, it contributes
- *    a bounded `secondClauseIntent` bump and sets `compoundIntent: true` on
- *    the verdict. This is proven inert on every current GOLD_SET row (no
- *    gold-set prompt produces a non-null, distinct second-clause intent), so
- *    it never changes an existing classification under DEFAULT_WEIGHTS_V4 —
- *    it only sharpens confidence/margin metadata for genuinely compound
- *    prompts and gives future calibration a lever.
- * 5. CONFIDENCE NEVER LOWERS EFFORT. `resolveThinkingLevelV4WithUncertainty`
- *    computes the exact same base target as v2/v3's resolvers (rule table +
- *    lane step + bias + hint), then ONLY ADDS a bounded escalation step when
- *    confidence is low or the class came from the fallback cascade. A prompt
- *    like "don't think hard, just fix the crash" cannot use its own text to
- *    talk the resolver down: the escalation term is strictly non-negative, so
- *    the output is always >= what the same taskClass would resolve to via
- *    `resolveThinkingLevelV4ForAuto` (mirrors spec 008 Req 2 acceptance
- *    criterion "confidence cannot lower effort").
- * ============================================================================
+ * `classifyTaskV4` returns a confidence-bearing verdict: task class, per-class
+ * scores, runner-up, margin, confidence band, tie-break flag, fallback reason,
+ * bounded-negation audit ids, and compound-intent metadata. None of this carries
+ * prompt text — every field is a bounded enum, number, boolean, or a closed set
+ * of short diagnostic-id strings.
+ *
+ * `resolveThinkingLevelV4WithUncertainty` starts from the canonical task-class
+ * rule table, applies lane/bias/hint adjustments, and only adds non-negative
+ * confidence escalation. A prompt cannot lower its own effort by asking the
+ * router to "think less"; low confidence or fallback routing can only hold or
+ * raise the resolved level.
  */
 import type { ThinkingLevel } from "omk-agent-core";
 import { type ReasoningLaneTypeV4, type RouterWeightsV4, type TaskClassV4 } from "./reasoning-router-v4-weights.ts";
@@ -124,28 +76,31 @@ export interface ClassifierVerdictV4 {
  * class from `weights` -> argmax with `TASK_CLASSES_V4` precedence tie-break,
  * tracking the runner-up -> compute margin/confidence/confidenceBand/tieBreak
  * from the RAW scores -> if `topScore <= 0`, replace `taskClass` (and record
- * `fallbackReason`) via the same zero-score cascade v1/v2/v3 use (fence/diff
- * -> code-gen; length < 40 -> trivial; length >= 2400 -> plan; lane fallback;
- * default code-gen) — the raw scores/margin/runnerUp/tieBreak fields still
- * reflect the pre-cascade computation, for audit purposes.
+ * `fallbackReason`) via the zero-score cascade (fence/diff -> code-gen;
+ * length < 40 -> trivial; length >= 2400 -> plan; lane fallback; default
+ * code-gen) — the raw scores/margin/runnerUp/tieBreak fields still reflect the
+ * pre-cascade computation, for audit purposes.
  *
  * Pure: same (input, weights) -> same ClassifierVerdictV4.
  */
 export declare function classifyTaskV4(input: TaskClassifierInputV4, weights?: RouterWeightsV4): ClassifierVerdictV4;
-/** Static rule table: task class -> recommended ThinkingLevel. Identical to v1/v2/v3. */
-export declare const TASK_CLASS_THINKING_LEVELS_V4: Readonly<Record<TaskClassV4, ThinkingLevel>>;
+/**
+ * Static rule table: task class -> recommended ThinkingLevel. Re-exported
+ * from the shared resolver core (byte-identical values; previously a
+ * duplicated literal here — see reasoning-router-resolver.ts).
+ */
+export declare const TASK_CLASS_THINKING_LEVELS_V4: Readonly<Record<import("./reasoning-router-resolver.ts").ReasoningTaskClass, ThinkingLevel>>;
 /**
  * Thin auto-mode wrapper with no uncertainty adjustment: base rule table ->
- * lane step -> clamp to availableLevels. Equivalent to
- * `resolveThinkingLevelV3ForAuto` / `resolveThinkingLevelV2ForAuto` for the
- * same (taskClass, availableLevels, laneType) — i.e. v4's "confident" path.
+ * lane step -> clamp to availableLevels. This is v4's "confident" path and
+ * delegates to the shared resolver core with bias=0, hint=null, escalationSteps=0.
  */
 export declare function resolveThinkingLevelV4ForAuto(taskClass: TaskClassV4, availableLevels: readonly ThinkingLevel[], laneType: ReasoningLaneTypeV4 | undefined): ThinkingLevel;
 /**
  * Uncertainty-aware resolver (spec 008 Req 2 / plan.md step 5).
  *
  * Pipeline: base rule table for `verdict.taskClass` -> lane step -> bounded
- * bias [-2,+2] -> optional hint fusion (±2 bounded, same as v2/v3) -> bounded
+ * bias [-2,+2] -> optional hint fusion (±2 bounded) -> bounded
  * non-negative confidence escalation (+1 ladder step exactly when
  * `verdict.confidenceBand === "low"` or `verdict.fallbackReason !== null`) ->
  * clamp to `availableLevels`.

package/dist/core/reasoning-router-v4.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"reasoning-router-v4.d.ts","sourceRoot":"","sources":["../../src/core/reasoning-router-v4.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoEG;AAEH,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AACpD,OAAO,EAEN,KAAK,mBAAmB,EACxB,KAAK,eAAe,EAEpB,KAAK,WAAW,EAChB,MAAM,kCAAkC,CAAC;AAE1C,YAAY,EAAE,mBAAmB,EAAE,eAAe,EAAE,WAAW,EAAE,MAAM,kCAAkC,CAAC;AAC1G,OAAO,EAAE,kBAAkB,EAAE,eAAe,EAAE,MAAM,kCAAkC,CAAC;AA4gBvF;;;;GAIG;AACH,MAAM,WAAW,qBAAqB;IACrC,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,CAAC,EAAE,mBAAmB,CAAC;IAC/B,OAAO,CAAC,EAAE,SAAS,WAAW,EAAE,CAAC;IACjC,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,SAAS,CAAC,EAAE,WAAW,GAAG,IAAI,CAAC;CAC/B;AAED,gIAAgI;AAChI,MAAM,MAAM,gBAAgB,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;AAEzD,wHAAwH;AACxH,MAAM,MAAM,gBAAgB,GACzB,oBAAoB,GACpB,gBAAgB,GAChB,sBAAsB,GACtB,YAAY,GACZ,eAAe,GACf,SAAS,CAAC;AAEb;;;;;GAKG;AACH,MAAM,WAAW,mBAAmB;IACnC,qEAAqE;IACrE,QAAQ,CAAC,SAAS,EAAE,WAAW,CAAC;IAChC,mFAAmF;IACnF,QAAQ,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,WAAW,EAAE,MAAM,CAAC,CAAC,CAAC;IACvD,mGAAmG;IACnG,QAAQ,CAAC,QAAQ,EAAE,WAAW,GAAG,IAAI,CAAC;IACtC,oIAAoI;IACpI,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,gHAAgH;IAChH,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,2FAA2F;IAC3F,QAAQ,CAAC,cAAc,EAAE,gBAAgB,CAAC;IAC1C,qHAAqH;IACrH,QAAQ,CAAC,QAAQ,EAAE,OAAO,CAAC;IAC3B,qGAAqG;IACrG,QAAQ,CAAC,cAAc,EAAE,gBAAgB,GAAG,IAAI,CAAC;IACjD,oHAAoH;IACpH,QAAQ,CAAC,oBAAoB,EAAE,SAAS,MAAM,EAAE,CAAC;IACjD,0GAA0G;IAC1G,QAAQ,CAAC,cAAc,EAAE,OAAO,CAAC;IACjC,oHAAoH;IACpH,QAAQ,CAAC,kBAAkB,EAAE,WAAW,GAAG,IAAI,CAAC;CAChD;AAMD;;;;;;;;;;;;;GAaG;AACH,wBAAgB,cAAc,CAC7B,KAAK,EAAE,qBAAqB,EAC5B,OAAO,GAAE,eAAoC,GAC3C,mBAAmB,CAqErB;AASD,yFAAyF;AACzF,eAAO,MAAM,6BAA6B,EAAE,QAAQ,CAAC,MAAM,CAAC,WAAW,EAAE,aAAa,CAAC,CAQtF,CAAC;AAmCF;;;;;GAKG;AACH,wBAAgB,6BAA6B,CAC5C,SAAS,EAAE,WAAW,EACtB,eAAe,EAAE,SAAS,aAAa,EAAE,EACzC,QAAQ,EAAE,mBAAmB,GAAG,SAAS,GACvC,aAAa,CAKf;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,qCAAqC,CACpD,OAAO,EAAE,mBAAmB,EAC5B,eAAe,EAAE,SAAS,aAAa,EAAE,EACzC,QAAQ,EAAE,mBAAmB,GAAG,SAAS,EACzC,IAAI,SAAI,EACR,IAAI,GAAE;IAAE,KAAK,EAAE,aAAa,CAAC;IAAC,UAAU,EAAE,MAAM,CAAA;CAAE,GAAG,IAAW,GAC9D,aAAa,CAoBf","sourcesContent":["/**\n * Reasoning-router v4 — confidence-bearing classifier (Goal 009 Wave 1 Lane A,\n * specs/008-reasoning-router-advanced-accuracy Requirement 2).\n *\n * v4 is a NEW, OPT-IN, PURE sibling module. It does not import from, and is not\n * imported by, reasoning-router.ts (v1), reasoning-router-v2.ts,\n * reasoning-router-v3.ts, reasoning-router-weights.ts, or agent-session.ts.\n * Activating v4 in the product (a `/think auto-v4` mode, settings wiring, etc.)\n * is explicitly out of scope for this lane and is left to a later,\n * single-writer integration lane (specs/008 plan.md Wave 3).\n *\n * Deterministic by construction: same (input, weights) -> same\n * ClassifierVerdictV4; same (verdict, availableLevels, laneType, bias, hint) ->\n * same ThinkingLevel. No clock, randomness, I/O, model calls, network access,\n * or state mutation anywhere in this file (ThinkingLevel is a type-only import\n * and is erased).\n *\n * ============================================================================\n * WHAT'S NEW VS v3 (reasoning-router-v3.ts)\n * ============================================================================\n * 1. VERDICT, NOT JUST A CLASS. `classifyTaskV4` returns a `ClassifierVerdictV4`\n *    carrying the full per-class score map, the runner-up class, the margin,\n *    a normalized confidence ratio, a confidence band, a `tieBreak` flag (the\n *    top class won only via precedence order over an exact score tie), a\n *    `fallbackReason` (non-null exactly when the zero-score cascade decided\n *    the class instead of a real signal), and `suppressedFeatureIds` (an\n *    audit trail of which whole-prompt signals were nulled by bounded\n *    negation). None of this carries prompt text — every field is a bounded\n *    enum, number, boolean, or a closed set of short diagnostic-id strings.\n * 2. WEIGHTS TABLE, NOT LITERALS. Every score bump v3 hard-coded as an inline\n *    integer now reads from a `RouterWeightsV4` (see\n *    reasoning-router-v4-weights.ts). `DEFAULT_WEIGHTS_V4` reproduces v3's\n *    literals exactly, so classifyTaskV4(input, DEFAULT_WEIGHTS_V4).taskClass\n *    matches classifyTaskV3(input) on the full non-holdout GOLD_SET and every\n *    focused v3 sentinel (see the 013 regression test and this lane's\n *    evidence file). Passing a different `RouterWeightsV4` recalibrates\n *    without touching this file.\n * 3. BOUNDED NEGATION. Every whole-prompt keyword/contextual pattern scan\n *    (the ones v3 ran as a bare `PATTERN.test(prompt)`) is now negation-aware:\n *    a match is only counted if no negation cue (don't/doesn't/isn't/never/\n *    avoid/skip/without/instead of/rather than/no need to|for/not a/...) appears within\n *    `weights.negationWindowChars` characters immediately before it, and the\n *    scan never crosses a `.,;!?` clause boundary. Patterns anchored to the\n *    START of the leading clause (hasLeadingDebugAction, hasLeadingReviewIntent,\n *    etc.) are untouched and need no negation handling: \"don't refactor\" can\n *    never match an anchored `^refactor` test in the first place. Concrete\n *    fix: \"don't refactor this, just fix the crash\" now classifies as debug\n *    (v3 misclassifies this as refactor; see the 013 test).\n * 4. BOUNDED COMPOUND INTENT. A short-range conjunction split (bare \"then\",\n *    \"and then\", \"and also\", or \";\", only within the first 300 characters)\n *    looks for a second, distinct leading-verb intent after the split point.\n *    When found and distinct from the primary leading intent, it contributes\n *    a bounded `secondClauseIntent` bump and sets `compoundIntent: true` on\n *    the verdict. This is proven inert on every current GOLD_SET row (no\n *    gold-set prompt produces a non-null, distinct second-clause intent), so\n *    it never changes an existing classification under DEFAULT_WEIGHTS_V4 —\n *    it only sharpens confidence/margin metadata for genuinely compound\n *    prompts and gives future calibration a lever.\n * 5. CONFIDENCE NEVER LOWERS EFFORT. `resolveThinkingLevelV4WithUncertainty`\n *    computes the exact same base target as v2/v3's resolvers (rule table +\n *    lane step + bias + hint), then ONLY ADDS a bounded escalation step when\n *    confidence is low or the class came from the fallback cascade. A prompt\n *    like \"don't think hard, just fix the crash\" cannot use its own text to\n *    talk the resolver down: the escalation term is strictly non-negative, so\n *    the output is always >= what the same taskClass would resolve to via\n *    `resolveThinkingLevelV4ForAuto` (mirrors spec 008 Req 2 acceptance\n *    criterion \"confidence cannot lower effort\").\n * ============================================================================\n */\n\nimport type { ThinkingLevel } from \"omk-agent-core\";\nimport {\n\tDEFAULT_WEIGHTS_V4,\n\ttype ReasoningLaneTypeV4,\n\ttype RouterWeightsV4,\n\tTASK_CLASSES_V4,\n\ttype TaskClassV4,\n} from \"./reasoning-router-v4-weights.ts\";\n\nexport type { ReasoningLaneTypeV4, RouterWeightsV4, TaskClassV4 } from \"./reasoning-router-v4-weights.ts\";\nexport { DEFAULT_WEIGHTS_V4, TASK_CLASSES_V4 } from \"./reasoning-router-v4-weights.ts\";\n\n/** Prompts shorter than this (trimmed) fall back to trivial when no real signal scores > 0. */\nconst TRIVIAL_MAX_CHARS_V4 = 40;\n/** Prompts at/above this length with no real signal fall back to plan (long prose brief). */\nconst COMPLEX_PROSE_MIN_CHARS_V4 = 2400;\n/** Plan-brief prompts at/above this length count as a \"long brief\" even without PLAN_BRIEF_PATTERN. */\nconst LONG_BRIEF_MIN_CHARS_V4 = 512;\n/** Compound-clause split only applies within this many leading characters (short direct commands, not long prose). */\nconst COMPOUND_SPLIT_MAX_INDEX_V4 = 300;\n/** Minimum trimmed length of a candidate second clause to be worth checking for a leading intent. */\nconst COMPOUND_SECOND_CLAUSE_MIN_CHARS_V4 = 3;\n/** Bounded look-ahead cap when checking a second clause's leading intent (defense in depth; independent of regex cost). */\nconst COMPOUND_SECOND_CLAUSE_SCAN_CHARS_V4 = 200;\n\n/** Lane fallback class, used only when every scored class is <= 0 (zero-score cascade). Identical to v1/v2/v3. */\nconst LANE_FALLBACK_CLASS_V4: Readonly<Record<ReasoningLaneTypeV4, TaskClassV4>> = {\n\tplanner: \"plan\",\n\tsecurity: \"review\",\n\texplorer: \"review\",\n\tcoder: \"code-gen\",\n\treviewer: \"review\",\n\ttester: \"code-gen\",\n};\n\n// ============================================================================\n// Whole-prompt / leading-clause patterns (copied verbatim from\n// reasoning-router-v3.ts; identical regex source so DEFAULT_WEIGHTS_V4\n// reproduces v3 exactly whenever no negation cue is present).\n// ============================================================================\n\nconst LOCAL_EDIT_OBJECT_PATTERN =\n\t/\\b(spelling|grammar|capitalization|date\\s+format|author\\s+e-?mail|copyright\\s+year|headline|title|tooltip|placeholder|punctuation|comma|period|semicolon|closing\\s+html\\s+tag|closing\\s+tag|double\\s+space|whitespace|indentation|table\\s+alignment|typos?|one-?liner?|single\\s+line|sentence|stray|trailing)\\b|오타|맞춤법|띄어쓰기|문구|제목/i;\nconst LOCAL_EDIT_ACTION_OBJECT_PATTERN =\n\t/\\b(update|change|swap|remove|correct|fix|add|adjust|trim)\\s+(?:the\\s+|a\\s+|an\\s+)?(?:missing\\s+|stray\\s+|author\\s+|two\\s+)?(?:e-?mail|comma|period|semicolon|tag|word|words|headline|title|copyright|spelling|grammar)\\b/i;\nconst IMPLEMENTATION_OBJECT_PATTERN =\n\t/\\b(error\\s+handling|input\\s+validation|validation|auth(?:entication)?|retry|endpoint|migration|component|function|utility|helper|module|service|database|schema|column|middleware|rate\\s+limiter|cache|caching|oauth|jwt|webhook)\\b|에러\\s*처리|오류\\s*처리|입력\\s*검증|인증|재시도|엔드포인트|마이그레이션|컴포넌트|함수|유틸|헬퍼|모듈|서비스|데이터베이스|스키마|미들웨어|레이트\\s*리미터|캐시|웹훅|기능/i;\nconst HARD_DIAGNOSTIC_PATTERN =\n\t/\\b(stack\\s*trace|traceback|panic|segfault|crash(?:es|ing|ed)?|throws?|fails?|failing|flaky|hangs?|timeout|500|EADDRINUSE|silently\\s+fails?)\\b|\\b(?:TypeError|ReferenceError|RangeError|Error):|크래시|스택\\s*트레이스|세그폴트|먹통|멈춰|타임아웃/i;\nconst BUG_OBJECT_PATTERN =\n\t/\\b(null\\s+pointer|null\\s+deref|race(?:\\s+condition)?|heap(?:\\s+overflow)?|use-after-free|memory\\s+leak|deadlock|data\\s+corruption|stale\\s+data|off-by-one|encoding\\s+bug|regression\\s+(?:was\\s+)?introduced|exceptions?|assertion\\s+error|bugs?)\\b|버그|예외|메모리\\s*누수|데드락|무한\\s*루프/i;\nconst NON_DIAGNOSTIC_DEBUG_CONTEXT_PATTERN = /\\berror\\s+(?:handling|messages?|budgets?)\\b|에러\\s*처리|오류\\s*처리/i;\nconst GENERIC_DIAGNOSTIC_PATTERN = /\\b(errors?|broken|wrong\\s+results|rolls?\\s+back|rollback)\\b/i;\nconst PLAN_BRIEF_PATTERN =\n\t/\\b(context\\s+and\\s+constraints|starting\\s+state|target\\s+state|cross-cutting|deliver:|deliverables?|roadmap|migration\\s+wave|top\\s+(?:ten\\s+)?risks|component\\s+diagram|data\\s+model|architecture|go\\/no-go|phased\\s+(?:delivery|rollout)|bounded\\s+contexts?|strangler\\s+fig|event-driven|quarter-by-quarter|milestone)\\b/i;\nconst OPERATIONAL_RUNBOOK_SIGNAL_PATTERNS: readonly { pattern: RegExp; critical: boolean }[] = [\n\t{ pattern: /\\b(?:commit|commits|committed|committing)\\b|커밋/i, critical: false },\n\t{ pattern: /\\b(?:push|pushed|pushing)\\b|푸시|푸쉬/i, critical: true },\n\t{\n\t\tpattern: /\\b(?:release|releases|tag|version\\s+bump|bump\\s+(?:the\\s+)?version)\\b|릴리즈|태그|버전/i,\n\t\tcritical: true,\n\t},\n\t{ pattern: /\\bnpm\\s+publish\\b|\\bpublish(?:ing|ed)?\\b|퍼블리시|배포/i, critical: true },\n\t{ pattern: /\\b(?:CHANGELOG\\.md|changelog|release\\s+notes?)\\b|채널로그|체인지로그/i, critical: false },\n\t{ pattern: /\\bREADME\\.md\\b|\\breadme\\b/i, critical: false },\n\t{ pattern: /\\bci\\s*\\/\\s*cd\\b|\\bgithub\\s+actions?\\b|\\bworkflow\\b|깃허브\\s*액션|깃헙\\s*액션/i, critical: false },\n];\nconst EXPLICIT_RELEASE_RUNBOOK_PATTERN =\n\t/\\bnpm\\s+publish\\b|\\brelease\\b[^\\n]{0,80}\\bv?\\d+\\.\\d+\\.\\d+\\b|릴리즈[^\\n]{0,80}\\d+\\.\\d+\\.\\d+/i;\nconst REVIEW_SCOPE_PATTERN =\n\t/\\b(pr|pull\\s+request|diff|codebase|strategy|plan|spec|schema|design|api|coverage|security\\s+posture|licensing|risks?|edge\\s+cases|clarity|consistency|dependencies|third-party|ci\\s+pipeline|threat\\s+model|retry\\s+logic|error\\s+handling\\s+strategy|error\\s+messages?)\\b/i;\nconst REFACTOR_CUE_PATTERN =\n\t/\\b(refactor(?:ing|ed)?|extract|rename|deduplicate|consolidate|modularize|reorganize|simplify|clean\\s*up|restructure|split\\s+module|move\\s+logic|untangle|merge\\s+duplicate)\\b|리팩토링(?!하지\\s*마|하지\\s*말)|리팩터링(?!하지\\s*마|하지\\s*말)|구조\\s*개선/i;\nconst ADD_KEYWORD_PATTERN = /\\badd\\b/i;\nconst LOW_RISK_EDIT_ACTION_PATTERN = /^(?:correct|update|swap|remove|reword|tweak|bump|trim)\\b/i;\n\n/** Whole-prompt keyword-family patterns (mirrors v3's KEYWORD_PATTERNS_V3; refactor/simple-edit reuse the patterns above). */\nconst DEBUG_KEYWORD_FAMILY_PATTERN =\n\t/^(?:debug|investigate\\s+why|reproduce|trace\\b)|\\bfix\\s+this\\s+(?:traceback|panic|error)\\b|디버깅|디버그|재현|원인\\s*분석/i;\nconst REVIEW_KEYWORD_FAMILY_PATTERN =\n\t/\\b(review|critique|assess|inspect|approve|lgtm|double-?check|audit(?!\\s+log\\b))\\b|리뷰|검토|점검/i;\nconst PLAN_KEYWORD_FAMILY_PATTERN =\n\t/\\b(plan|design|architect|architecture|roadmap|spec(?:ification)?|strategy|decompose|milestones?|write\\s+a\\s+(?:technical\\s+)?spec|create\\s+a\\s+(?:roadmap|strategy|plan))\\b|설계|로드맵|아키텍처|계획\\s*세워|기획/i;\nconst CODE_GEN_KEYWORD_FAMILY_PATTERN =\n\t/\\b(implement|write|create|build|generate|scaffold|prototype)\\b|구현|만들어\\s*줘|생성|작성|추가해/i;\nconst KO_SHORT_DEBUG_SIGNAL_PATTERN_V4 =\n\t/오류|에러|실패|안\\s*돼|안\\s*됨|깨졌|깨짐|고장|디버깅|디버그|재현|원인\\s*(?:분석|파악)?/i;\nconst KO_SHORT_SIMPLE_EDIT_SIGNAL_PATTERN_V4 = /오타|맞춤법|띄어쓰기|문구|제목/i;\nconst KO_SHORT_REVIEW_SIGNAL_PATTERN_V4 = /리뷰|검토|점검/i;\nconst KO_SHORT_PLAN_SIGNAL_PATTERN_V4 = /설계|로드맵|아키텍처|계획\\s*세워|기획/i;\nconst KO_SHORT_REFACTOR_SIGNAL_PATTERN_V4 = /리팩토링|리팩터링|구조\\s*개선/i;\nconst KO_SHORT_CODE_GEN_SIGNAL_PATTERN_V4 =\n\t/구현|만들|생성|작성|추가|테스트|삭제|지워|지우|제거|고쳐|고치|수정|변경|바꿔|바꾸|옮겨|옮기|넣어/i;\n\n/** Short-range compound-clause conjunction boundary. Only bare/compound \"then\"/\"and also\"/\";\" — never bare \"and\". */\nconst COMPOUND_SPLIT_PATTERN = /\\band\\s+then\\b|\\bthen\\b|\\band\\s+also\\b|;/i;\n\n/**\n * Bounded negation-cue vocabulary (spec 008 plan.md \"bounded negation window\").\n * Deliberately excludes bare \"not\" (too broad; verified against the current\n * GOLD_SET that every listed cue here is either absent or occurs AFTER, not\n * before, a tracked keyword — see this lane's evidence file for the audit).\n */\nconst NEGATION_CUE_PATTERN =\n\t/\\b(?:don't|do\\s+not|doesn't|does\\s+not|didn't|did\\s+not|won't|will\\s+not|shouldn't|should\\s+not|wouldn't|would\\s+not|can't|cannot|isn't|is\\s+not|aren't|are\\s+not|never|avoid|skip|without|instead\\s+of|rather\\s+than|no\\s+need\\s+(?:to|for)|no\\s+longer\\s+need(?:\\s+(?:to|for))?)\\b/i;\nconst DIRECT_NOT_OBJECT_CUE_PATTERN_V4 = /\\b(?:not|no)\\s+(?:a\\s+|an\\s+|the\\s+)?$/i;\nconst DOUBLE_NEGATION_RESCUE_PATTERN_V4 =\n\t/\\b(?:don't|do\\s+not|doesn't|does\\s+not|didn't|did\\s+not|won't|will\\s+not|shouldn't|should\\s+not|wouldn't|would\\s+not|can't|cannot)\\s+(?:skip|avoid)\\s+(?:the\\s+|a\\s+|an\\s+)?$/i;\nconst DELIBERATIVE_QUESTION_RESCUE_PATTERN_V4 =\n\t/\\b(?:shouldn't|should\\s+not|wouldn't|would\\s+not|can't|cannot)\\s+(?:we|i|you)\\s*$/i;\nconst POSTPOSITIONED_NEGATION_CUE_PATTERN_V4 =\n\t/^\\s*(?:-\\s*free\\b|(?:is|are|was|were)?\\s*(?:not\\s+(?:needed|required|desired)|unnecessary|not\\s+necessary)|(?:[은는이가을를도만]\\s*)?(?:하지\\s*(?:마(?:라|세요|십시오)?|말(?:고|아|라)?|않(?:아|는|고|을|게)?)|말고|금지))/i;\n\n/** A closed set of clause-boundary characters; a negation cue never reaches across one of these into a prior clause. */\nconst CLAUSE_BOUNDARY_CHARS = [\".\", \"!\", \"?\", \";\", \",\"] as const;\n\nfunction hasCodeFence(text: string): boolean {\n\treturn text.includes(\"```\");\n}\n\n/** Identical to v1/v2/v3: explicit hunk headers or `diff --git` count alone; bare +/- only count together. */\nfunction hasDiffMarkers(text: string): boolean {\n\tif (/^@@[^\\n]*@@/m.test(text) || /^diff --git /m.test(text)) return true;\n\treturn /^\\+(?!\\+)/m.test(text) && /^-(?!-)/m.test(text);\n}\n\n/** Identical to v3: first line (after stripping one polite prefix), capped at 180 chars. */\nfunction firstClause(prompt: string): string {\n\tconst firstLine =\n\t\tprompt.replace(/^(?:please|pls|can you|could you|would you|help me|i need you to)\\s+/i, \"\").split(\"\\n\")[0] ?? \"\";\n\treturn firstLine.slice(0, 180);\n}\n\n// --- Leading-clause intent tests. All are `^`-anchored against the LEADING\n// clause text only, so a negation cue at the start (\"don't refactor...\") can\n// never match one of these: none of the alternations include \"don't\"/\"never\"/\n// etc, so they are negation-immune by construction and are left unchanged\n// from v3. ---\n\nfunction hasLeadingReviewIntent(text: string): boolean {\n\treturn /^(?:review|audit(?!\\s+log\\b)|critique|inspect|assess|approve|double-?check|lgtm)\\b/i.test(text);\n}\n\nfunction hasLeadingPlanIntent(text: string): boolean {\n\treturn /^(?:plan|design|architect|decompose)\\b|^(?:write|create)\\s+(?:a\\s+)?(?:technical\\s+)?(?:spec|roadmap|strategy|plan)\\b/i.test(\n\t\ttext,\n\t);\n}\n\nfunction hasLeadingRefactorIntent(text: string): boolean {\n\treturn /^(?:refactor(?:ing|ed)?|extract|rename|deduplicate|consolidate|modularize|reorganize|simplify|restructure|untangle)\\b|^clean\\s+up\\b|^split\\s+(?:the\\s+)?module\\b|^move\\s+logic\\b/i.test(\n\t\ttext,\n\t);\n}\n\nfunction hasLeadingDebugAction(text: string): boolean {\n\treturn /^(?:debug|investigate\\s+why|reproduce|trace\\b)|^fix\\s+this\\s+(?:traceback|panic|error)\\b/i.test(text);\n}\n\nfunction hasLeadingCodeGenIntent(text: string): boolean {\n\treturn /^(?:implement|write|create|build|generate|scaffold|prototype|add)\\b/i.test(text);\n}\n\nfunction hasLeadingSimpleEditIntent(text: string): boolean {\n\treturn /^(?:correct|update|swap|remove|adjust|fix|add|change|trim|reword|tweak|bump)\\b/i.test(text);\n}\n\n// ============================================================================\n// Bounded negation gating\n// ============================================================================\n\ninterface UnnegatedMatchResultV4 {\n\t/** True when the pattern matched at least once and at least one occurrence was NOT preceded by a negation cue. */\n\treadonly matched: boolean;\n\t/** True when the pattern matched at least once but EVERY occurrence was negated (useful for suppressedFeatureIds). */\n\treadonly suppressed: boolean;\n}\n\nconst NO_MATCH_RESULT_V4: UnnegatedMatchResultV4 = { matched: false, suppressed: false };\n\nfunction matchOperationalRunbookV4(prompt: string, windowChars: number): UnnegatedMatchResultV4 {\n\tlet distinctSignals = 0;\n\tlet criticalSignals = 0;\n\tlet sawSuppressed = false;\n\tfor (const { pattern, critical } of OPERATIONAL_RUNBOOK_SIGNAL_PATTERNS) {\n\t\tconst result = matchUnnegated(prompt, pattern, windowChars);\n\t\tif (result.matched) {\n\t\t\tdistinctSignals += 1;\n\t\t\tif (critical) criticalSignals += 1;\n\t\t}\n\t\tif (result.suppressed) sawSuppressed = true;\n\t}\n\tconst explicitRelease = matchUnnegated(prompt, EXPLICIT_RELEASE_RUNBOOK_PATTERN, windowChars).matched;\n\tconst matched = explicitRelease || distinctSignals >= 3 || (criticalSignals >= 2 && distinctSignals >= 2);\n\treturn { matched, suppressed: !matched && sawSuppressed };\n}\n\n/**\n * Negation-aware replacement for v3's bare `pattern.test(prompt)`. Finds every\n * occurrence of `pattern` in `prompt`; for each, looks back up to\n * `windowChars` characters (never crossing a `.,;!?` boundary into a prior\n * clause) for a `NEGATION_CUE_PATTERN` hit, and looks forward inside the same\n * bounded window for Hangul-only post-positioned negation cues such as\n * \"하지 말고\". `matched` is true iff at least one occurrence survives\n * un-negated. Deterministic, single pass per pattern, no shared regex state (a\n * fresh global-flag RegExp is constructed per call).\n */\nfunction hasPrePositionedNegationCueV4(scoped: string): boolean {\n\tif (!NEGATION_CUE_PATTERN.test(scoped) && !DIRECT_NOT_OBJECT_CUE_PATTERN_V4.test(scoped)) return false;\n\treturn !DOUBLE_NEGATION_RESCUE_PATTERN_V4.test(scoped) && !DELIBERATIVE_QUESTION_RESCUE_PATTERN_V4.test(scoped);\n}\n\nfunction hasPostPositionedNegationCueV4(prompt: string, matchEnd: number, windowChars: number): boolean {\n\tconst scoped = prompt.slice(matchEnd, Math.min(prompt.length, matchEnd + windowChars));\n\treturn POSTPOSITIONED_NEGATION_CUE_PATTERN_V4.test(scoped);\n}\n\nfunction matchUnnegated(prompt: string, pattern: RegExp, windowChars: number): UnnegatedMatchResultV4 {\n\tconst flags = pattern.flags.includes(\"g\") ? pattern.flags : `${pattern.flags}g`;\n\tconst globalPattern = new RegExp(pattern.source, flags);\n\tlet sawMatch = false;\n\tlet sawUnnegated = false;\n\tlet match = globalPattern.exec(prompt);\n\twhile (match !== null) {\n\t\tsawMatch = true;\n\t\tconst matchStart = match.index;\n\t\tconst windowStart = Math.max(0, matchStart - windowChars);\n\t\tconst windowText = prompt.slice(windowStart, matchStart);\n\t\tlet boundary = -1;\n\t\tfor (const boundaryChar of CLAUSE_BOUNDARY_CHARS) {\n\t\t\tconst idx = windowText.lastIndexOf(boundaryChar);\n\t\t\tif (idx > boundary) boundary = idx;\n\t\t}\n\t\tconst scoped = boundary >= 0 ? windowText.slice(boundary + 1) : windowText;\n\t\tconst preNegated = hasPrePositionedNegationCueV4(scoped);\n\t\tconst postNegated =\n\t\t\t!preNegated && hasPostPositionedNegationCueV4(prompt, matchStart + match[0].length, windowChars);\n\t\tif (!preNegated && !postNegated) sawUnnegated = true;\n\t\tif (match[0].length === 0) globalPattern.lastIndex += 1;\n\t\tmatch = globalPattern.exec(prompt);\n\t}\n\treturn { matched: sawUnnegated, suppressed: sawMatch && !sawUnnegated };\n}\n\n// ============================================================================\n// Bounded compound-intent detection\n// ============================================================================\n\n/**\n * Splits off a bounded second clause after a short-range conjunction\n * (\"then\"/\"and then\"/\"and also\"/\";\"), only when the split point is within the\n * first `COMPOUND_SPLIT_MAX_INDEX_V4` characters (short direct commands, not\n * long prose briefs — verified this never fires inside the GOLD_SET's\n * >=2400-char plan entries). Returns null when no qualifying split exists.\n */\nfunction splitCompoundClauseV4(prompt: string): string | null {\n\tconst match = COMPOUND_SPLIT_PATTERN.exec(prompt);\n\tif (match === null || match.index > COMPOUND_SPLIT_MAX_INDEX_V4) return null;\n\tconst second = prompt.slice(match.index + match[0].length).trim();\n\tif (second.length < COMPOUND_SECOND_CLAUSE_MIN_CHARS_V4) return null;\n\treturn second;\n}\n\n/**\n * Leading-intent test for a compound prompt's SECOND clause. Reuses the same\n * `^`-anchored leading-intent tests as the primary clause (so it is equally\n * negation-immune), without the primary clause's diagnosticEvidence/localEdit\n * gating (a short second-clause fragment carries no such context of its own).\n */\nfunction secondClauseLeadingIntentV4(clause: string): TaskClassV4 | null {\n\tconst bounded = clause.slice(0, COMPOUND_SECOND_CLAUSE_SCAN_CHARS_V4);\n\treturn leadingIntentFromLeadingClauseV4(bounded);\n}\n\nfunction leadingIntentFromLeadingClauseV4(leading: string): TaskClassV4 | null {\n\tif (hasLeadingDebugAction(leading)) return \"debug\";\n\tif (hasLeadingReviewIntent(leading)) return \"review\";\n\tif (hasLeadingPlanIntent(leading)) return \"plan\";\n\tif (hasLeadingRefactorIntent(leading)) return \"refactor\";\n\tif (hasLeadingSimpleEditIntent(leading)) return \"simple-edit\";\n\tif (hasLeadingCodeGenIntent(leading)) return \"code-gen\";\n\treturn null;\n}\n\nfunction leadingIntentIsPostNegatedV4(leading: string, windowChars: number): boolean {\n\tconst leadingIntentMatch =\n\t\t/^(?:debug|investigate\\s+why|reproduce|trace\\b|fix\\s+this\\s+(?:traceback|panic|error)|review|audit(?!\\s+log\\b)|critique|inspect|assess|approve|double-?check|lgtm|plan|design|architect|decompose|write\\s+(?:a\\s+)?(?:technical\\s+)?(?:spec|roadmap|strategy|plan)|create\\s+(?:a\\s+)?(?:technical\\s+)?(?:spec|roadmap|strategy|plan)|refactor(?:ing|ed)?|extract|rename|deduplicate|consolidate|modularize|reorganize|simplify|restructure|untangle|clean\\s+up|split\\s+(?:the\\s+)?module|move\\s+logic|correct|update|swap|remove|adjust|fix|add|change|trim|reword|tweak|bump|implement|build|generate|scaffold|prototype)\\b/i.exec(\n\t\t\tleading,\n\t\t);\n\treturn (\n\t\tleadingIntentMatch !== null && hasPostPositionedNegationCueV4(leading, leadingIntentMatch[0].length, windowChars)\n\t);\n}\n\nfunction classifyShortKoreanZeroScoreTaskV4(prompt: string, windowChars: number): TaskClassV4 | null {\n\tif (matchUnnegated(prompt, KO_SHORT_DEBUG_SIGNAL_PATTERN_V4, windowChars).matched) return \"debug\";\n\tif (matchUnnegated(prompt, KO_SHORT_SIMPLE_EDIT_SIGNAL_PATTERN_V4, windowChars).matched) return \"simple-edit\";\n\tif (matchUnnegated(prompt, KO_SHORT_REVIEW_SIGNAL_PATTERN_V4, windowChars).matched) return \"review\";\n\tif (matchUnnegated(prompt, KO_SHORT_PLAN_SIGNAL_PATTERN_V4, windowChars).matched) return \"plan\";\n\tif (matchUnnegated(prompt, KO_SHORT_REFACTOR_SIGNAL_PATTERN_V4, windowChars).matched) return \"refactor\";\n\tif (matchUnnegated(prompt, KO_SHORT_CODE_GEN_SIGNAL_PATTERN_V4, windowChars).matched) return \"code-gen\";\n\treturn null;\n}\n\n// ============================================================================\n// Feature extraction\n// ============================================================================\n\ninterface ContextualFeaturesV4 {\n\treadonly firstClause: string;\n\treadonly codeFence: boolean;\n\treadonly diffHunk: boolean;\n\treadonly localEdit: boolean;\n\treadonly diagnosticEvidence: boolean;\n\treadonly reviewScope: boolean;\n\treadonly planBrief: boolean;\n\treadonly operationalRunbook: boolean;\n\treadonly refactorCue: boolean;\n\treadonly implementationObject: boolean;\n\treadonly leadingIntent: TaskClassV4 | null;\n\treadonly secondClauseIntent: TaskClassV4 | null;\n\treadonly compoundIntent: boolean;\n\treadonly keywordFamilyMatch: Readonly<Record<TaskClassV4, boolean>>;\n\treadonly addKeywordMatch: boolean;\n}\n\ninterface LeadingIntentInputV4 {\n\treadonly firstClause: string;\n\treadonly localEdit: boolean;\n\treadonly diagnosticEvidence: boolean;\n}\n\n/** Identical decision order to v3's `leadingIntent`. */\nfunction leadingIntentV4(input: LeadingIntentInputV4): TaskClassV4 | null {\n\tif (\n\t\tinput.diagnosticEvidence &&\n\t\t(hasLeadingDebugAction(input.firstClause) || hasLeadingSimpleEditIntent(input.firstClause))\n\t)\n\t\treturn \"debug\";\n\tif (hasLeadingReviewIntent(input.firstClause)) return \"review\";\n\tif (hasLeadingPlanIntent(input.firstClause)) return \"plan\";\n\tif (hasLeadingRefactorIntent(input.firstClause)) return \"refactor\";\n\tif (input.localEdit && hasLeadingSimpleEditIntent(input.firstClause)) return \"simple-edit\";\n\tif (!input.localEdit && hasLeadingCodeGenIntent(input.firstClause)) return \"code-gen\";\n\treturn null;\n}\n\n/** Identical decision structure to v3's `hasDiagnosticEvidence`, parameterized over precomputed negation-aware matches. */\nfunction hasDiagnosticEvidenceV4(\n\tleading: string,\n\thardDiagnosticMatched: boolean,\n\tbugObjectMatched: boolean,\n\tgenericDiagnosticMatched: boolean,\n\tnonDiagnosticContext: boolean,\n): boolean {\n\tif (hasLeadingDebugAction(leading) || hardDiagnosticMatched) return true;\n\tconst hasContextualDiagnostic = bugObjectMatched || (genericDiagnosticMatched && !nonDiagnosticContext);\n\tif (!hasContextualDiagnostic) return false;\n\treturn !hasLeadingReviewIntent(leading) && !hasLeadingPlanIntent(leading) && !hasLeadingRefactorIntent(leading);\n}\n\n/**\n * Extracts every v4 feature from one prompt, negation-gating every\n * whole-prompt scan exactly once and recording a `negation:<channel>` id in\n * `suppressed` whenever a match existed but every occurrence was negated.\n */\nfunction extractFeaturesV4(prompt: string, weights: RouterWeightsV4, suppressed: string[]): ContextualFeaturesV4 {\n\tconst leading = firstClause(prompt);\n\tconst window = weights.negationWindowChars;\n\n\tconst implementationObjectResult = matchUnnegated(prompt, IMPLEMENTATION_OBJECT_PATTERN, window);\n\tif (implementationObjectResult.suppressed) suppressed.push(\"negation:implementation-object\");\n\tconst implementationObject = implementationObjectResult.matched;\n\n\tconst lowRiskEditAction = LOW_RISK_EDIT_ACTION_PATTERN.test(leading);\n\tconst localEditObjectResult = matchUnnegated(prompt, LOCAL_EDIT_OBJECT_PATTERN, window);\n\tif (localEditObjectResult.suppressed) suppressed.push(\"negation:local-edit-object\");\n\tconst localEditActionResult = matchUnnegated(prompt, LOCAL_EDIT_ACTION_OBJECT_PATTERN, window);\n\tif (localEditActionResult.suppressed) suppressed.push(\"negation:local-edit-action-object\");\n\tconst localEdit =\n\t\t!implementationObject && (lowRiskEditAction || localEditObjectResult.matched || localEditActionResult.matched);\n\n\tconst hardDiagnosticResult = matchUnnegated(prompt, HARD_DIAGNOSTIC_PATTERN, window);\n\tif (hardDiagnosticResult.suppressed) suppressed.push(\"negation:hard-diagnostic\");\n\tconst bugObjectResult = matchUnnegated(prompt, BUG_OBJECT_PATTERN, window);\n\tif (bugObjectResult.suppressed) suppressed.push(\"negation:bug-object\");\n\tconst genericDiagnosticResult = matchUnnegated(prompt, GENERIC_DIAGNOSTIC_PATTERN, window);\n\tconst nonDiagnosticContext = NON_DIAGNOSTIC_DEBUG_CONTEXT_PATTERN.test(prompt);\n\tif (genericDiagnosticResult.suppressed && !nonDiagnosticContext) suppressed.push(\"negation:generic-diagnostic\");\n\tconst diagnosticEvidence = hasDiagnosticEvidenceV4(\n\t\tleading,\n\t\thardDiagnosticResult.matched,\n\t\tbugObjectResult.matched,\n\t\tgenericDiagnosticResult.matched,\n\t\tnonDiagnosticContext,\n\t);\n\n\tconst reviewScopeResult = matchUnnegated(prompt, REVIEW_SCOPE_PATTERN, window);\n\tif (reviewScopeResult.suppressed) suppressed.push(\"negation:review-scope\");\n\tconst reviewScope = hasLeadingReviewIntent(leading) && reviewScopeResult.matched;\n\n\tconst planBriefResult = matchUnnegated(prompt, PLAN_BRIEF_PATTERN, window);\n\tif (planBriefResult.suppressed) suppressed.push(\"negation:plan-brief\");\n\tconst planBrief =\n\t\thasLeadingPlanIntent(leading) && (planBriefResult.matched || prompt.length >= LONG_BRIEF_MIN_CHARS_V4);\n\n\tconst operationalRunbookResult = matchOperationalRunbookV4(prompt, window);\n\tif (operationalRunbookResult.suppressed) suppressed.push(\"negation:operational-runbook\");\n\tconst operationalRunbook = operationalRunbookResult.matched;\n\n\tconst refactorCueResult = matchUnnegated(prompt, REFACTOR_CUE_PATTERN, window);\n\tif (refactorCueResult.suppressed) suppressed.push(\"negation:refactor-cue\");\n\tconst refactorCue = refactorCueResult.matched;\n\n\tconst rawPrimaryIntent = leadingIntentV4({ firstClause: leading, localEdit, diagnosticEvidence });\n\tconst primaryIntent =\n\t\trawPrimaryIntent !== null && leadingIntentIsPostNegatedV4(leading, window) ? null : rawPrimaryIntent;\n\n\tconst secondClauseText = splitCompoundClauseV4(prompt);\n\tconst secondClauseIntent = secondClauseText === null ? null : secondClauseLeadingIntentV4(secondClauseText);\n\tconst compoundIntent = secondClauseIntent !== null && secondClauseIntent !== primaryIntent;\n\n\tconst debugKeywordResult = diagnosticEvidence\n\t\t? matchUnnegated(prompt, DEBUG_KEYWORD_FAMILY_PATTERN, window)\n\t\t: NO_MATCH_RESULT_V4;\n\tif (debugKeywordResult.suppressed) suppressed.push(\"negation:keyword-debug\");\n\n\tconst reviewKeywordResult = matchUnnegated(prompt, REVIEW_KEYWORD_FAMILY_PATTERN, window);\n\tif (reviewKeywordResult.suppressed) suppressed.push(\"negation:keyword-review\");\n\n\tconst planKeywordResult = matchUnnegated(prompt, PLAN_KEYWORD_FAMILY_PATTERN, window);\n\tif (planKeywordResult.suppressed) suppressed.push(\"negation:keyword-plan\");\n\n\tconst codeGenKeywordResult = matchUnnegated(prompt, CODE_GEN_KEYWORD_FAMILY_PATTERN, window);\n\tif (codeGenKeywordResult.suppressed) suppressed.push(\"negation:keyword-code-gen\");\n\n\tconst addKeywordResult = localEdit ? NO_MATCH_RESULT_V4 : matchUnnegated(prompt, ADD_KEYWORD_PATTERN, window);\n\tif (addKeywordResult.suppressed) suppressed.push(\"negation:add-keyword\");\n\n\treturn {\n\t\tfirstClause: leading,\n\t\tcodeFence: hasCodeFence(prompt),\n\t\tdiffHunk: hasDiffMarkers(prompt),\n\t\tlocalEdit,\n\t\tdiagnosticEvidence,\n\t\treviewScope,\n\t\tplanBrief,\n\t\toperationalRunbook,\n\t\trefactorCue,\n\t\timplementationObject,\n\t\tleadingIntent: primaryIntent,\n\t\tsecondClauseIntent,\n\t\tcompoundIntent,\n\t\tkeywordFamilyMatch: {\n\t\t\ttrivial: false,\n\t\t\tdebug: debugKeywordResult.matched,\n\t\t\trefactor: refactorCue,\n\t\t\treview: reviewKeywordResult.matched,\n\t\t\tplan: planKeywordResult.matched,\n\t\t\t\"simple-edit\": localEditObjectResult.matched,\n\t\t\t\"code-gen\": codeGenKeywordResult.matched,\n\t\t},\n\t\taddKeywordMatch: addKeywordResult.matched,\n\t};\n}\n\n// ============================================================================\n// Scoring\n// ============================================================================\n\nfunction emptyScoresV4(): Record<TaskClassV4, number> {\n\treturn {\n\t\ttrivial: 0,\n\t\t\"simple-edit\": 0,\n\t\t\"code-gen\": 0,\n\t\tdebug: 0,\n\t\trefactor: 0,\n\t\treview: 0,\n\t\tplan: 0,\n\t};\n}\n\nfunction computeScoresV4(features: ContextualFeaturesV4, weights: RouterWeightsV4): Record<TaskClassV4, number> {\n\tconst scores = emptyScoresV4();\n\tif (features.codeFence || features.diffHunk) scores[\"code-gen\"] += weights.codeFenceOrDiff;\n\tif (features.leadingIntent !== null) scores[features.leadingIntent] += weights.leadingIntent;\n\tif (features.secondClauseIntent !== null) scores[features.secondClauseIntent] += weights.secondClauseIntent;\n\tif (features.localEdit) scores[\"simple-edit\"] += weights.localEdit;\n\tif (features.diagnosticEvidence) scores.debug += weights.diagnosticEvidence;\n\tif (features.reviewScope) scores.review += weights.reviewScope;\n\tif (features.planBrief) scores.plan += weights.planBrief;\n\tif (features.operationalRunbook) scores.plan += weights.operationalRunbook;\n\tif (features.refactorCue) scores.refactor += weights.refactorCue;\n\tif (features.implementationObject) scores[\"code-gen\"] += weights.implementationObject;\n\n\tfor (const taskClass of TASK_CLASSES_V4) {\n\t\tif (features.keywordFamilyMatch[taskClass]) scores[taskClass] += weights.keywordFamily[taskClass];\n\t}\n\tif (features.addKeywordMatch) scores[\"code-gen\"] += weights.addKeyword;\n\n\treturn scores;\n}\n\n/**\n * Extension signals (history / context-pressure / judge vote). Every\n * coefficient is 0 under DEFAULT_WEIGHTS_V4 (inert until calibrated by a\n * future governance-backed lane); the mechanism exists so\n * `TaskClassifierInputV4`'s optional fields are meaningfully wired rather than\n * silently accepted-and-ignored (as they are in v3's identical-shaped input).\n */\nfunction applyExtensionSignalsV4(\n\tscores: Record<TaskClassV4, number>,\n\tinput: TaskClassifierInputV4,\n\tweights: RouterWeightsV4,\n): void {\n\tconst priorClass = input.history !== undefined && input.history.length > 0 ? input.history[0] : null;\n\tif (priorClass !== null && weights.multiTurnPrior !== 0) scores[priorClass] += weights.multiTurnPrior;\n\n\tconst judgeVote = input.judgeVote ?? null;\n\tif (judgeVote !== null && weights.judgeVote !== 0) scores[judgeVote] += weights.judgeVote;\n\n\tconst pressureBucket = input.pressureBucket ?? 0;\n\tif (pressureBucket > 0 && weights.pressureBucket !== 0) {\n\t\tconst bump = pressureBucket * weights.pressureBucket;\n\t\tscores.debug += bump;\n\t\tscores.review += bump;\n\t\tscores.plan += bump;\n\t}\n}\n\n// ============================================================================\n// Public types\n// ============================================================================\n\n/**\n * Caller-supplied turn input. Identical shape to `TaskClassifierInputV3` /\n * `TaskClassifierInputV2` (prompt + optional lane/history/pressure/judge\n * slots), so existing callers can switch versions without reshaping data.\n */\nexport interface TaskClassifierInputV4 {\n\tprompt: string;\n\tlaneType?: ReasoningLaneTypeV4;\n\thistory?: readonly TaskClassV4[];\n\tpressureBucket?: number;\n\tjudgeVote?: TaskClassV4 | null;\n}\n\n/** Confidence band derived from `margin / topScore`, bucketed by `weights.lowConfidenceThreshold`/`highConfidenceThreshold`. */\nexport type ConfidenceBandV4 = \"low\" | \"medium\" | \"high\";\n\n/** Which branch of the zero-score fallback cascade decided `taskClass`; null when a real weighted signal decided it. */\nexport type FallbackReasonV4 =\n\t| \"code-fence-or-diff\"\n\t| \"trivial-length\"\n\t| \"ko-short-task-signal\"\n\t| \"long-prose\"\n\t| \"lane-fallback\"\n\t| \"default\";\n\n/**\n * The full, privacy-safe classification verdict (spec 008 Req 2). Every field\n * is a bounded enum, number, boolean, or a closed set of short diagnostic-id\n * strings (`suppressedFeatureIds`) — never raw prompt text, so this value is\n * safe to pass to evaluation, learning, and Adaptorch-advisory code.\n */\nexport interface ClassifierVerdictV4 {\n\t/** The decided task class (post zero-score-cascade if one fired). */\n\treadonly taskClass: TaskClassV4;\n\t/** Raw per-class score map from the weighted signals (pre-cascade; audit-only). */\n\treadonly scores: Readonly<Record<TaskClassV4, number>>;\n\t/** Second-highest-scoring class by the argmax, or null if every other class tied at the bottom. */\n\treadonly runnerUp: TaskClassV4 | null;\n\t/** `scores[argmaxClass] - scores[runnerUp]` (pre-cascade); 0 or negative-margin-free by construction (weights are non-negative). */\n\treadonly margin: number;\n\t/** Normalized confidence in [0, 1]: `margin / topScore`, or 0 when `topScore <= 0` (fallback cascade fired). */\n\treadonly confidence: number;\n\t/** Confidence bucketed by `weights.lowConfidenceThreshold` / `highConfidenceThreshold`. */\n\treadonly confidenceBand: ConfidenceBandV4;\n\t/** True when the argmax class won only via precedence-order tie-break over an exact score tie with the runner-up. */\n\treadonly tieBreak: boolean;\n\t/** Non-null exactly when the zero-score cascade (not a real weighted signal) decided `taskClass`. */\n\treadonly fallbackReason: FallbackReasonV4 | null;\n\t/** Audit trail of `negation:<channel>` ids for every whole-prompt signal that had a match but was fully negated. */\n\treadonly suppressedFeatureIds: readonly string[];\n\t/** True when a distinct second-clause leading intent was detected (bounded compound-intent detection). */\n\treadonly compoundIntent: boolean;\n\t/** The second clause's own leading intent, or null when no compound structure (or no distinct intent) was found. */\n\treadonly secondClauseIntent: TaskClassV4 | null;\n}\n\n// ============================================================================\n// Classifier\n// ============================================================================\n\n/**\n * Deterministic confidence-bearing classifier (spec 008 Req 2).\n *\n * Pipeline: extract negation/compound-aware contextual features -> score every\n * class from `weights` -> argmax with `TASK_CLASSES_V4` precedence tie-break,\n * tracking the runner-up -> compute margin/confidence/confidenceBand/tieBreak\n * from the RAW scores -> if `topScore <= 0`, replace `taskClass` (and record\n * `fallbackReason`) via the same zero-score cascade v1/v2/v3 use (fence/diff\n * -> code-gen; length < 40 -> trivial; length >= 2400 -> plan; lane fallback;\n * default code-gen) — the raw scores/margin/runnerUp/tieBreak fields still\n * reflect the pre-cascade computation, for audit purposes.\n *\n * Pure: same (input, weights) -> same ClassifierVerdictV4.\n */\nexport function classifyTaskV4(\n\tinput: TaskClassifierInputV4,\n\tweights: RouterWeightsV4 = DEFAULT_WEIGHTS_V4,\n): ClassifierVerdictV4 {\n\tconst prompt = input.prompt.trim();\n\tconst suppressed: string[] = [];\n\tconst features = extractFeaturesV4(prompt, weights, suppressed);\n\tconst scores = computeScoresV4(features, weights);\n\tapplyExtensionSignalsV4(scores, input, weights);\n\n\tlet top: TaskClassV4 = TASK_CLASSES_V4[0];\n\tfor (const c of TASK_CLASSES_V4) {\n\t\tif (scores[c] > scores[top]) top = c;\n\t}\n\tlet runnerUp: TaskClassV4 | null = null;\n\tfor (const c of TASK_CLASSES_V4) {\n\t\tif (c === top) continue;\n\t\tif (runnerUp === null || scores[c] > scores[runnerUp]) runnerUp = c;\n\t}\n\n\tconst topScore = scores[top];\n\tconst runnerUpScore = runnerUp !== null ? scores[runnerUp] : 0;\n\tconst margin = topScore - runnerUpScore;\n\tconst tieBreak = runnerUp !== null && topScore === runnerUpScore;\n\tconst confidence = topScore <= 0 ? 0 : Math.max(0, Math.min(1, margin / topScore));\n\tconst confidenceBand: ConfidenceBandV4 =\n\t\tconfidence <= weights.lowConfidenceThreshold\n\t\t\t? \"low\"\n\t\t\t: confidence >= weights.highConfidenceThreshold\n\t\t\t\t? \"high\"\n\t\t\t\t: \"medium\";\n\n\tlet taskClass: TaskClassV4 = top;\n\tlet fallbackReason: FallbackReasonV4 | null = null;\n\tif (topScore <= 0) {\n\t\tif (features.codeFence || features.diffHunk) {\n\t\t\ttaskClass = \"code-gen\";\n\t\t\tfallbackReason = \"code-fence-or-diff\";\n\t\t} else if (prompt.length < TRIVIAL_MAX_CHARS_V4) {\n\t\t\tconst shortKoreanTaskClass = classifyShortKoreanZeroScoreTaskV4(prompt, weights.negationWindowChars);\n\t\t\tif (shortKoreanTaskClass !== null) {\n\t\t\t\ttaskClass = shortKoreanTaskClass;\n\t\t\t\tfallbackReason = \"ko-short-task-signal\";\n\t\t\t} else {\n\t\t\t\ttaskClass = \"trivial\";\n\t\t\t\tfallbackReason = \"trivial-length\";\n\t\t\t}\n\t\t} else if (prompt.length >= COMPLEX_PROSE_MIN_CHARS_V4) {\n\t\t\ttaskClass = \"plan\";\n\t\t\tfallbackReason = \"long-prose\";\n\t\t} else if (input.laneType !== undefined) {\n\t\t\ttaskClass = LANE_FALLBACK_CLASS_V4[input.laneType];\n\t\t\tfallbackReason = \"lane-fallback\";\n\t\t} else {\n\t\t\ttaskClass = \"code-gen\";\n\t\t\tfallbackReason = \"default\";\n\t\t}\n\t}\n\n\treturn {\n\t\ttaskClass,\n\t\tscores,\n\t\trunnerUp,\n\t\tmargin,\n\t\tconfidence,\n\t\tconfidenceBand,\n\t\ttieBreak,\n\t\tfallbackReason,\n\t\tsuppressedFeatureIds: suppressed,\n\t\tcompoundIntent: features.compoundIntent,\n\t\tsecondClauseIntent: features.secondClauseIntent,\n\t};\n}\n\n// ============================================================================\n// Resolver\n// ============================================================================\n\n/** Reasoning ladder used for targets and clamping. Intentionally excludes \"off\". Identical to v1/v2/v3. */\nconst REASONING_LADDER_V4: readonly ThinkingLevel[] = [\"minimal\", \"low\", \"medium\", \"high\", \"xhigh\", \"max\"];\n\n/** Static rule table: task class -> recommended ThinkingLevel. Identical to v1/v2/v3. */\nexport const TASK_CLASS_THINKING_LEVELS_V4: Readonly<Record<TaskClassV4, ThinkingLevel>> = {\n\ttrivial: \"minimal\",\n\t\"simple-edit\": \"low\",\n\t\"code-gen\": \"medium\",\n\tdebug: \"high\",\n\trefactor: \"high\",\n\treview: \"high\",\n\tplan: \"xhigh\",\n};\n\n/** Lane ladder adjustment. Identical to v1/v2/v3: planner/security escalate one step, explorer de-escalates. */\nconst LANE_STEP_V4: Readonly<Partial<Record<ReasoningLaneTypeV4, 1 | -1>>> = {\n\tplanner: 1,\n\tsecurity: 1,\n\texplorer: -1,\n};\n\n/** Maximum magnitude of the learning/consult bias (ladder steps). Identical bound to v2/v3. */\nconst BIAS_MAX_V4 = 2;\n/** Confidence floor at which an override hint is fused into the target. Identical to v2/v3. */\nconst HINT_CONFIDENCE_THRESHOLD_V4 = 0.7;\n/**\n * Bounded, strictly non-negative escalation applied when the verdict's own\n * confidence is low or a fallback (not a real signal) decided the class. This\n * is the ONLY place confidence touches the target index, and it can only add:\n * confidence can never lower effort (spec 008 Req 2 acceptance criterion).\n */\nconst LOW_CONFIDENCE_ESCALATION_STEPS_V4 = 1;\n\nfunction clampToLadderIndexV4(index: number): number {\n\treturn Math.max(0, Math.min(index, REASONING_LADDER_V4.length - 1));\n}\n\nfunction clampToAvailableV4(targetIndex: number, availableLevels: readonly ThinkingLevel[]): ThinkingLevel {\n\tconst availableOnLadder = REASONING_LADDER_V4.filter((level) => availableLevels.includes(level));\n\tif (availableOnLadder.length === 0) return availableLevels[0] ?? \"off\";\n\tfor (let i = targetIndex; i >= 0; i--) {\n\t\tconst candidate = REASONING_LADDER_V4[i];\n\t\tif (availableOnLadder.includes(candidate)) return candidate;\n\t}\n\treturn availableOnLadder[0];\n}\n\n/**\n * Thin auto-mode wrapper with no uncertainty adjustment: base rule table ->\n * lane step -> clamp to availableLevels. Equivalent to\n * `resolveThinkingLevelV3ForAuto` / `resolveThinkingLevelV2ForAuto` for the\n * same (taskClass, availableLevels, laneType) — i.e. v4's \"confident\" path.\n */\nexport function resolveThinkingLevelV4ForAuto(\n\ttaskClass: TaskClassV4,\n\tavailableLevels: readonly ThinkingLevel[],\n\tlaneType: ReasoningLaneTypeV4 | undefined,\n): ThinkingLevel {\n\tconst baseIndex = REASONING_LADDER_V4.indexOf(TASK_CLASS_THINKING_LEVELS_V4[taskClass]);\n\tconst laneStep = laneType ? (LANE_STEP_V4[laneType] ?? 0) : 0;\n\tconst targetIndex = clampToLadderIndexV4(baseIndex + laneStep);\n\treturn clampToAvailableV4(targetIndex, availableLevels);\n}\n\n/**\n * Uncertainty-aware resolver (spec 008 Req 2 / plan.md step 5).\n *\n * Pipeline: base rule table for `verdict.taskClass` -> lane step -> bounded\n * bias [-2,+2] -> optional hint fusion (±2 bounded, same as v2/v3) -> bounded\n * non-negative confidence escalation (+1 ladder step exactly when\n * `verdict.confidenceBand === \"low\"` or `verdict.fallbackReason !== null`) ->\n * clamp to `availableLevels`.\n *\n * The confidence-escalation term is strictly `>= 0`: low confidence can only\n * hold the base+lane+bias+hint target or push it one step higher, never lower\n * it. This guarantees a low-confidence verdict never resolves BELOW what\n * `resolveThinkingLevelV4ForAuto` would give the same class (with bias=0,\n * hint=null) — text alone cannot talk the resolver down.\n */\nexport function resolveThinkingLevelV4WithUncertainty(\n\tverdict: ClassifierVerdictV4,\n\tavailableLevels: readonly ThinkingLevel[],\n\tlaneType: ReasoningLaneTypeV4 | undefined,\n\tbias = 0,\n\thint: { level: ThinkingLevel; confidence: number } | null = null,\n): ThinkingLevel {\n\tconst baseIndex = REASONING_LADDER_V4.indexOf(TASK_CLASS_THINKING_LEVELS_V4[verdict.taskClass]);\n\tconst laneStep = laneType ? (LANE_STEP_V4[laneType] ?? 0) : 0;\n\tconst biasClamped = Math.max(-BIAS_MAX_V4, Math.min(BIAS_MAX_V4, bias));\n\tlet targetIndex = clampToLadderIndexV4(baseIndex + laneStep + biasClamped);\n\n\tif (hint !== null && hint.confidence >= HINT_CONFIDENCE_THRESHOLD_V4) {\n\t\tconst hintIndex = REASONING_LADDER_V4.indexOf(hint.level);\n\t\tif (hintIndex >= 0) {\n\t\t\tconst delta = hintIndex - targetIndex;\n\t\t\tconst step = Math.max(-BIAS_MAX_V4, Math.min(BIAS_MAX_V4, delta));\n\t\t\ttargetIndex = clampToLadderIndexV4(targetIndex + step);\n\t\t}\n\t}\n\n\tif (verdict.confidenceBand === \"low\" || verdict.fallbackReason !== null) {\n\t\ttargetIndex = clampToLadderIndexV4(targetIndex + LOW_CONFIDENCE_ESCALATION_STEPS_V4);\n\t}\n\n\treturn clampToAvailableV4(targetIndex, availableLevels);\n}\n"]}
+{"version":3,"file":"reasoning-router-v4.d.ts","sourceRoot":"","sources":["../../src/core/reasoning-router-v4.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAEH,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AACpD,OAAO,EAEN,KAAK,mBAAmB,EACxB,KAAK,eAAe,EAEpB,KAAK,WAAW,EAChB,MAAM,kCAAkC,CAAC;AAE1C,YAAY,EAAE,mBAAmB,EAAE,eAAe,EAAE,WAAW,EAAE,MAAM,kCAAkC,CAAC;AAC1G,OAAO,EAAE,kBAAkB,EAAE,eAAe,EAAE,MAAM,kCAAkC,CAAC;AA2gBvF;;;;GAIG;AACH,MAAM,WAAW,qBAAqB;IACrC,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,CAAC,EAAE,mBAAmB,CAAC;IAC/B,OAAO,CAAC,EAAE,SAAS,WAAW,EAAE,CAAC;IACjC,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,SAAS,CAAC,EAAE,WAAW,GAAG,IAAI,CAAC;CAC/B;AAED,gIAAgI;AAChI,MAAM,MAAM,gBAAgB,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;AAEzD,wHAAwH;AACxH,MAAM,MAAM,gBAAgB,GACzB,oBAAoB,GACpB,gBAAgB,GAChB,sBAAsB,GACtB,YAAY,GACZ,eAAe,GACf,SAAS,CAAC;AAEb;;;;;GAKG;AACH,MAAM,WAAW,mBAAmB;IACnC,qEAAqE;IACrE,QAAQ,CAAC,SAAS,EAAE,WAAW,CAAC;IAChC,mFAAmF;IACnF,QAAQ,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,WAAW,EAAE,MAAM,CAAC,CAAC,CAAC;IACvD,mGAAmG;IACnG,QAAQ,CAAC,QAAQ,EAAE,WAAW,GAAG,IAAI,CAAC;IACtC,oIAAoI;IACpI,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,gHAAgH;IAChH,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,2FAA2F;IAC3F,QAAQ,CAAC,cAAc,EAAE,gBAAgB,CAAC;IAC1C,qHAAqH;IACrH,QAAQ,CAAC,QAAQ,EAAE,OAAO,CAAC;IAC3B,qGAAqG;IACrG,QAAQ,CAAC,cAAc,EAAE,gBAAgB,GAAG,IAAI,CAAC;IACjD,oHAAoH;IACpH,QAAQ,CAAC,oBAAoB,EAAE,SAAS,MAAM,EAAE,CAAC;IACjD,0GAA0G;IAC1G,QAAQ,CAAC,cAAc,EAAE,OAAO,CAAC;IACjC,oHAAoH;IACpH,QAAQ,CAAC,kBAAkB,EAAE,WAAW,GAAG,IAAI,CAAC;CAChD;AAMD;;;;;;;;;;;;;GAaG;AACH,wBAAgB,cAAc,CAC7B,KAAK,EAAE,qBAAqB,EAC5B,OAAO,GAAE,eAAoC,GAC3C,mBAAmB,CAqErB;AAMD;;;;GAIG;AACH,eAAO,MAAM,6BAA6B,8FAA6B,CAAC;AAUxE;;;;GAIG;AACH,wBAAgB,6BAA6B,CAC5C,SAAS,EAAE,WAAW,EACtB,eAAe,EAAE,SAAS,aAAa,EAAE,EACzC,QAAQ,EAAE,mBAAmB,GAAG,SAAS,GACvC,aAAa,CAEf;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,qCAAqC,CACpD,OAAO,EAAE,mBAAmB,EAC5B,eAAe,EAAE,SAAS,aAAa,EAAE,EACzC,QAAQ,EAAE,mBAAmB,GAAG,SAAS,EACzC,IAAI,SAAI,EACR,IAAI,GAAE;IAAE,KAAK,EAAE,aAAa,CAAC;IAAC,UAAU,EAAE,MAAM,CAAA;CAAE,GAAG,IAAW,GAC9D,aAAa,CAIf","sourcesContent":["/**\n * Reasoning-router v4 — the sole `/think auto` classifier/resolver.\n *\n * Deterministic by construction: same (input, weights) -> same\n * ClassifierVerdictV4; same (verdict, availableLevels, laneType, bias, hint) ->\n * same ThinkingLevel. No clock, randomness, I/O, model calls, network access,\n * or state mutation anywhere in this file (ThinkingLevel is a type-only import\n * and is erased).\n *\n * `classifyTaskV4` returns a confidence-bearing verdict: task class, per-class\n * scores, runner-up, margin, confidence band, tie-break flag, fallback reason,\n * bounded-negation audit ids, and compound-intent metadata. None of this carries\n * prompt text — every field is a bounded enum, number, boolean, or a closed set\n * of short diagnostic-id strings.\n *\n * `resolveThinkingLevelV4WithUncertainty` starts from the canonical task-class\n * rule table, applies lane/bias/hint adjustments, and only adds non-negative\n * confidence escalation. A prompt cannot lower its own effort by asking the\n * router to \"think less\"; low confidence or fallback routing can only hold or\n * raise the resolved level.\n */\n\nimport type { ThinkingLevel } from \"omk-agent-core\";\nimport {\n\tDEFAULT_WEIGHTS_V4,\n\ttype ReasoningLaneTypeV4,\n\ttype RouterWeightsV4,\n\tTASK_CLASSES_V4,\n\ttype TaskClassV4,\n} from \"./reasoning-router-v4-weights.ts\";\n\nexport type { ReasoningLaneTypeV4, RouterWeightsV4, TaskClassV4 } from \"./reasoning-router-v4-weights.ts\";\nexport { DEFAULT_WEIGHTS_V4, TASK_CLASSES_V4 } from \"./reasoning-router-v4-weights.ts\";\n\nimport { resolveThinkingLevelCore, TASK_CLASS_THINKING_LEVELS } from \"./reasoning-router-resolver.ts\";\n\n/** Prompts shorter than this (trimmed) fall back to trivial when no real signal scores > 0. */\nconst TRIVIAL_MAX_CHARS_V4 = 40;\n/** Prompts at/above this length with no real signal fall back to plan (long prose brief). */\nconst COMPLEX_PROSE_MIN_CHARS_V4 = 2400;\n/** Plan-brief prompts at/above this length count as a \"long brief\" even without PLAN_BRIEF_PATTERN. */\nconst LONG_BRIEF_MIN_CHARS_V4 = 512;\n/** Compound-clause split only applies within this many leading characters (short direct commands, not long prose). */\nconst COMPOUND_SPLIT_MAX_INDEX_V4 = 300;\n/** Minimum trimmed length of a candidate second clause to be worth checking for a leading intent. */\nconst COMPOUND_SECOND_CLAUSE_MIN_CHARS_V4 = 3;\n/** Bounded look-ahead cap when checking a second clause's leading intent (defense in depth; independent of regex cost). */\nconst COMPOUND_SECOND_CLAUSE_SCAN_CHARS_V4 = 200;\n\n/** Lane fallback class, used only when every scored class is <= 0 (zero-score cascade). */\nconst LANE_FALLBACK_CLASS_V4: Readonly<Record<ReasoningLaneTypeV4, TaskClassV4>> = {\n\tplanner: \"plan\",\n\tsecurity: \"review\",\n\texplorer: \"review\",\n\tcoder: \"code-gen\",\n\treviewer: \"review\",\n\ttester: \"code-gen\",\n};\n\n// ============================================================================\n// Whole-prompt / leading-clause patterns used by the v4 scorer.\n// ============================================================================\n\nconst LOCAL_EDIT_OBJECT_PATTERN =\n\t/\\b(spelling|grammar|capitalization|date\\s+format|author\\s+e-?mail|copyright\\s+year|headline|title|tooltip|placeholder|punctuation|comma|period|semicolon|closing\\s+html\\s+tag|closing\\s+tag|double\\s+space|whitespace|indentation|table\\s+alignment|typos?|one-?liner?|single\\s+line|sentence|stray|trailing)\\b|오타|맞춤법|띄어쓰기|문구|제목/i;\nconst LOCAL_EDIT_ACTION_OBJECT_PATTERN =\n\t/\\b(update|change|swap|remove|correct|fix|add|adjust|trim)\\s+(?:the\\s+|a\\s+|an\\s+)?(?:missing\\s+|stray\\s+|author\\s+|two\\s+)?(?:e-?mail|comma|period|semicolon|tag|word|words|headline|title|copyright|spelling|grammar)\\b/i;\nconst IMPLEMENTATION_OBJECT_PATTERN =\n\t/\\b(error\\s+handling|input\\s+validation|validation|auth(?:entication)?|retry|endpoint|migration|component|function|utility|helper|module|service|database|schema|column|middleware|rate\\s+limiter|cache|caching|oauth|jwt|webhook)\\b|에러\\s*처리|오류\\s*처리|입력\\s*검증|인증|재시도|엔드포인트|마이그레이션|컴포넌트|함수|유틸|헬퍼|모듈|서비스|데이터베이스|스키마|미들웨어|레이트\\s*리미터|캐시|웹훅|기능/i;\nconst HARD_DIAGNOSTIC_PATTERN =\n\t/\\b(stack\\s*trace|traceback|panic|segfault|crash(?:es|ing|ed)?|throws?|fails?|failing|flaky|hangs?|timeout|500|EADDRINUSE|silently\\s+fails?)\\b|\\b(?:TypeError|ReferenceError|RangeError|Error):|크래시|스택\\s*트레이스|세그폴트|먹통|멈춰|타임아웃/i;\nconst BUG_OBJECT_PATTERN =\n\t/\\b(null\\s+pointer|null\\s+deref|race(?:\\s+condition)?|heap(?:\\s+overflow)?|use-after-free|memory\\s+leak|leaks?|deadlock|data\\s+corruption|stale\\s+data|off-by-one|encoding\\s+bug|regression\\s+(?:was\\s+)?introduced|exceptions?|assertion\\s+error|bugs?)\\b|버그|예외|메모리\\s*누수|데드락|무한\\s*루프/i;\nconst NON_DIAGNOSTIC_DEBUG_CONTEXT_PATTERN = /\\berror\\s+(?:handling|messages?|budgets?)\\b|에러\\s*처리|오류\\s*처리/i;\nconst GENERIC_DIAGNOSTIC_PATTERN = /\\b(errors?|broken|wrong\\s+results|rolls?\\s+back|rollback)\\b/i;\nconst PLAN_BRIEF_PATTERN =\n\t/\\b(context\\s+and\\s+constraints|starting\\s+state|target\\s+state|cross-cutting|deliver:|deliverables?|roadmap|migration\\s+wave|top\\s+(?:ten\\s+)?risks|component\\s+diagram|data\\s+model|architecture|go\\/no-go|phased\\s+(?:delivery|rollout)|bounded\\s+contexts?|strangler\\s+fig|event-driven|quarter-by-quarter|milestone)\\b/i;\nconst OPERATIONAL_RUNBOOK_SIGNAL_PATTERNS: readonly { pattern: RegExp; critical: boolean }[] = [\n\t{ pattern: /\\b(?:commit|commits|committed|committing)\\b|커밋/i, critical: false },\n\t{ pattern: /\\b(?:push|pushed|pushing)\\b|푸시|푸쉬/i, critical: true },\n\t{\n\t\tpattern: /\\b(?:release|releases|tag|version\\s+bump|bump\\s+(?:the\\s+)?version)\\b|릴리즈|태그|버전/i,\n\t\tcritical: true,\n\t},\n\t{ pattern: /\\bnpm\\s+publish\\b|\\bpublish(?:ing|ed)?\\b|퍼블리시|배포/i, critical: true },\n\t{ pattern: /\\b(?:CHANGELOG\\.md|changelog|release\\s+notes?)\\b|채널로그|체인지로그/i, critical: false },\n\t{ pattern: /\\bREADME\\.md\\b|\\breadme\\b/i, critical: false },\n\t{ pattern: /\\bci\\s*\\/\\s*cd\\b|\\bgithub\\s+actions?\\b|\\bworkflow\\b|깃허브\\s*액션|깃헙\\s*액션/i, critical: false },\n];\nconst EXPLICIT_RELEASE_RUNBOOK_PATTERN =\n\t/\\bnpm\\s+publish\\b|\\brelease\\b[^\\n]{0,80}\\bv?\\d+\\.\\d+\\.\\d+\\b|릴리즈[^\\n]{0,80}\\d+\\.\\d+\\.\\d+/i;\nconst REVIEW_SCOPE_PATTERN =\n\t/\\b(pr|pull\\s+request|diff|codebase|strategy|plan|spec|schema|design|api|coverage|security\\s+posture|licensing|risks?|edge\\s+cases|clarity|consistency|dependencies|third-party|ci\\s+pipeline|threat\\s+model|retry\\s+logic|error\\s+handling\\s+strategy|error\\s+messages?)\\b/i;\nconst REFACTOR_CUE_PATTERN =\n\t/\\b(refactor(?:ing|ed)?|extract|rename|deduplicate|consolidate|modularize|reorganize|simplify|clean\\s*up|tidy\\s+up|restructure|split\\s+module|move\\s+logic|untangle|merge\\s+duplicate)\\b|리팩토링(?!하지\\s*마|하지\\s*말)|리팩터링(?!하지\\s*마|하지\\s*말)|구조\\s*개선/i;\nconst ADD_KEYWORD_PATTERN = /\\badd\\b/i;\nconst LOW_RISK_EDIT_ACTION_PATTERN = /^(?:correct|update|swap|remove|reword|tweak|bump|trim)\\b/i;\n\n/** Whole-prompt keyword-family patterns; refactor/simple-edit reuse the patterns above. */\nconst DEBUG_KEYWORD_FAMILY_PATTERN =\n\t/^(?:debug|investigate\\s+why|reproduce|trace\\b)|\\bfix\\s+this\\s+(?:traceback|panic|error)\\b|\\btrack\\s+down\\b|디버깅|디버그|재현|원인\\s*분석/i;\nconst REVIEW_KEYWORD_FAMILY_PATTERN =\n\t/\\b(review|critique|assess|inspect|approve|lgtm|double-?check|sanity[-\\s]?check|once-?over|eyeball|tell\\s+me\\s+(?:if|whether)[^.!?;\\n]{0,80}\\bcorrect|audit(?!\\s+log\\b))\\b|리뷰|검토|점검/i;\nconst PLAN_KEYWORD_FAMILY_PATTERN =\n\t/\\b(plan|design|architect|architecture|roadmap|spec(?:ification)?|strategy|decompose|milestones?|map\\s+out|phases|write\\s+a\\s+(?:technical\\s+)?spec|create\\s+a\\s+(?:roadmap|strategy|plan))\\b|설계|로드맵|아키텍처|계획.{0,6}세워|기획/i;\nconst CODE_GEN_KEYWORD_FAMILY_PATTERN =\n\t/\\b(implement|write|create|build|generate|scaffold|prototype)\\b|구현|만들어\\s*줘|생성|작성|추가해/i;\nconst KO_SHORT_DEBUG_SIGNAL_PATTERN_V4 =\n\t/오류|에러|실패|안\\s*돼|안\\s*됨|깨졌|깨짐|고장|디버깅|디버그|재현|원인\\s*(?:분석|파악)?/i;\nconst KO_SHORT_SIMPLE_EDIT_SIGNAL_PATTERN_V4 = /오타|맞춤법|띄어쓰기|문구|제목/i;\nconst KO_SHORT_REVIEW_SIGNAL_PATTERN_V4 = /리뷰|검토|점검/i;\nconst KO_SHORT_PLAN_SIGNAL_PATTERN_V4 = /설계|로드맵|아키텍처|계획.{0,6}세워|기획/i;\nconst KO_SHORT_REFACTOR_SIGNAL_PATTERN_V4 = /리팩토링|리팩터링|구조\\s*개선/i;\nconst KO_SHORT_CODE_GEN_SIGNAL_PATTERN_V4 =\n\t/구현|만들|생성|작성|추가|테스트|삭제|지워|지우|제거|고쳐|고치|수정|변경|바꿔|바꾸|옮겨|옮기|넣어/i;\n\n/** Short-range compound-clause conjunction boundary. Only bare/compound \"then\"/\"and also\"/\";\" — never bare \"and\". */\nconst COMPOUND_SPLIT_PATTERN = /\\band\\s+then\\b|\\bthen\\b|\\band\\s+also\\b|;/i;\n\n/**\n * Bounded negation-cue vocabulary (spec 008 plan.md \"bounded negation window\").\n * Deliberately excludes bare \"not\" (too broad; verified against the current\n * GOLD_SET that every listed cue here is either absent or occurs AFTER, not\n * before, a tracked keyword — see this lane's evidence file for the audit).\n */\nconst NEGATION_CUE_PATTERN =\n\t/\\b(?:don't|do\\s+not|doesn't|does\\s+not|didn't|did\\s+not|won't|will\\s+not|shouldn't|should\\s+not|wouldn't|would\\s+not|can't|cannot|isn't|is\\s+not|aren't|are\\s+not|never|avoid|skip|without|instead\\s+of|rather\\s+than|no\\s+need\\s+(?:to|for)|no\\s+longer\\s+need(?:\\s+(?:to|for))?)\\b/i;\nconst DIRECT_NOT_OBJECT_CUE_PATTERN_V4 = /\\b(?:not|no)\\s+(?:a\\s+|an\\s+|the\\s+)?$/i;\nconst DOUBLE_NEGATION_RESCUE_PATTERN_V4 =\n\t/\\b(?:don't|do\\s+not|doesn't|does\\s+not|didn't|did\\s+not|won't|will\\s+not|shouldn't|should\\s+not|wouldn't|would\\s+not|can't|cannot)\\s+(?:skip|avoid)\\s+(?:the\\s+|a\\s+|an\\s+)?$/i;\nconst DELIBERATIVE_QUESTION_RESCUE_PATTERN_V4 =\n\t/\\b(?:shouldn't|should\\s+not|wouldn't|would\\s+not|can't|cannot)\\s+(?:we|i|you)\\s*$/i;\nconst POSTPOSITIONED_NEGATION_CUE_PATTERN_V4 =\n\t/^\\s*(?:-\\s*free\\b|(?:is|are|was|were)?\\s*(?:not\\s+(?:needed|required|desired)|unnecessary|not\\s+necessary)|(?:[은는이가을를도만]\\s*)?(?:하지\\s*(?:마(?:라|세요|십시오)?|말(?:고|아|라)?|않(?:아|는|고|을|게)?)|말고|금지))/i;\n\n/** A closed set of clause-boundary characters; a negation cue never reaches across one of these into a prior clause. */\nconst CLAUSE_BOUNDARY_CHARS = [\".\", \"!\", \"?\", \";\", \",\"] as const;\n\nfunction hasCodeFence(text: string): boolean {\n\treturn text.includes(\"```\");\n}\n\n/** Explicit hunk headers or `diff --git` count alone; bare +/- only count together. */\nfunction hasDiffMarkers(text: string): boolean {\n\tif (/^@@[^\\n]*@@/m.test(text) || /^diff --git /m.test(text)) return true;\n\treturn /^\\+(?!\\+)/m.test(text) && /^-(?!-)/m.test(text);\n}\n\n/** First line (after stripping one polite prefix), capped at 180 chars. */\nfunction firstClause(prompt: string): string {\n\tconst firstLine =\n\t\tprompt.replace(/^(?:please|pls|can you|could you|would you|help me|i need you to)\\s+/i, \"\").split(\"\\n\")[0] ?? \"\";\n\treturn firstLine.slice(0, 180);\n}\n\n// --- Leading-clause intent tests. All are `^`-anchored against the LEADING\n// clause text only, so a negation cue at the start (\"don't refactor...\") can\n// never match one of these: none of the alternations include \"don't\"/\"never\"/\n// etc, so they are negation-immune by construction.\n\nfunction hasLeadingReviewIntent(text: string): boolean {\n\treturn /^(?:review|audit(?!\\s+log\\b)|critique|inspect|assess|approve|double-?check|lgtm)\\b/i.test(text);\n}\n\nfunction hasLeadingPlanIntent(text: string): boolean {\n\treturn /^(?:plan|design|architect|decompose)\\b|^(?:write|create)\\s+(?:a\\s+)?(?:technical\\s+)?(?:spec|roadmap|strategy|plan)\\b/i.test(\n\t\ttext,\n\t);\n}\n\nfunction hasLeadingRefactorIntent(text: string): boolean {\n\treturn /^(?:refactor(?:ing|ed)?|extract|rename|deduplicate|consolidate|modularize|reorganize|simplify|restructure|untangle)\\b|^clean\\s+up\\b|^split\\s+(?:the\\s+)?module\\b|^move\\s+logic\\b/i.test(\n\t\ttext,\n\t);\n}\n\nfunction hasLeadingDebugAction(text: string): boolean {\n\treturn /^(?:debug|investigate\\s+why|reproduce|trace\\b)|^fix\\s+this\\s+(?:traceback|panic|error)\\b/i.test(text);\n}\n\nfunction hasLeadingCodeGenIntent(text: string): boolean {\n\treturn /^(?:implement|write|create|build|generate|scaffold|prototype|add)\\b/i.test(text);\n}\n\nfunction hasLeadingSimpleEditIntent(text: string): boolean {\n\treturn /^(?:correct|update|swap|remove|adjust|fix|add|change|trim|reword|tweak|bump)\\b/i.test(text);\n}\n\n// ============================================================================\n// Bounded negation gating\n// ============================================================================\n\ninterface UnnegatedMatchResultV4 {\n\t/** True when the pattern matched at least once and at least one occurrence was NOT preceded by a negation cue. */\n\treadonly matched: boolean;\n\t/** True when the pattern matched at least once but EVERY occurrence was negated (useful for suppressedFeatureIds). */\n\treadonly suppressed: boolean;\n}\n\nconst NO_MATCH_RESULT_V4: UnnegatedMatchResultV4 = { matched: false, suppressed: false };\n\nfunction matchOperationalRunbookV4(prompt: string, windowChars: number): UnnegatedMatchResultV4 {\n\tlet distinctSignals = 0;\n\tlet criticalSignals = 0;\n\tlet sawSuppressed = false;\n\tfor (const { pattern, critical } of OPERATIONAL_RUNBOOK_SIGNAL_PATTERNS) {\n\t\tconst result = matchUnnegated(prompt, pattern, windowChars);\n\t\tif (result.matched) {\n\t\t\tdistinctSignals += 1;\n\t\t\tif (critical) criticalSignals += 1;\n\t\t}\n\t\tif (result.suppressed) sawSuppressed = true;\n\t}\n\tconst explicitRelease = matchUnnegated(prompt, EXPLICIT_RELEASE_RUNBOOK_PATTERN, windowChars).matched;\n\tconst matched = explicitRelease || distinctSignals >= 3 || (criticalSignals >= 2 && distinctSignals >= 2);\n\treturn { matched, suppressed: !matched && sawSuppressed };\n}\n\n/**\n * Negation-aware whole-prompt pattern scan. Finds every occurrence of `pattern`\n * in `prompt`; for each, looks back up to\n * `windowChars` characters (never crossing a `.,;!?` boundary into a prior\n * clause) for a `NEGATION_CUE_PATTERN` hit, and looks forward inside the same\n * bounded window for Hangul-only post-positioned negation cues such as\n * \"하지 말고\". `matched` is true iff at least one occurrence survives\n * un-negated. Deterministic, single pass per pattern, no shared regex state (a\n * fresh global-flag RegExp is constructed per call).\n */\nfunction hasPrePositionedNegationCueV4(scoped: string): boolean {\n\tif (!NEGATION_CUE_PATTERN.test(scoped) && !DIRECT_NOT_OBJECT_CUE_PATTERN_V4.test(scoped)) return false;\n\treturn !DOUBLE_NEGATION_RESCUE_PATTERN_V4.test(scoped) && !DELIBERATIVE_QUESTION_RESCUE_PATTERN_V4.test(scoped);\n}\n\nfunction hasPostPositionedNegationCueV4(prompt: string, matchEnd: number, windowChars: number): boolean {\n\tconst scoped = prompt.slice(matchEnd, Math.min(prompt.length, matchEnd + windowChars));\n\treturn POSTPOSITIONED_NEGATION_CUE_PATTERN_V4.test(scoped);\n}\n\nfunction matchUnnegated(prompt: string, pattern: RegExp, windowChars: number): UnnegatedMatchResultV4 {\n\tconst flags = pattern.flags.includes(\"g\") ? pattern.flags : `${pattern.flags}g`;\n\tconst globalPattern = new RegExp(pattern.source, flags);\n\tlet sawMatch = false;\n\tlet sawUnnegated = false;\n\tlet match = globalPattern.exec(prompt);\n\twhile (match !== null) {\n\t\tsawMatch = true;\n\t\tconst matchStart = match.index;\n\t\tconst windowStart = Math.max(0, matchStart - windowChars);\n\t\tconst windowText = prompt.slice(windowStart, matchStart);\n\t\tlet boundary = -1;\n\t\tfor (const boundaryChar of CLAUSE_BOUNDARY_CHARS) {\n\t\t\tconst idx = windowText.lastIndexOf(boundaryChar);\n\t\t\tif (idx > boundary) boundary = idx;\n\t\t}\n\t\tconst scoped = boundary >= 0 ? windowText.slice(boundary + 1) : windowText;\n\t\tconst preNegated = hasPrePositionedNegationCueV4(scoped);\n\t\tconst postNegated =\n\t\t\t!preNegated && hasPostPositionedNegationCueV4(prompt, matchStart + match[0].length, windowChars);\n\t\tif (!preNegated && !postNegated) sawUnnegated = true;\n\t\tif (match[0].length === 0) globalPattern.lastIndex += 1;\n\t\tmatch = globalPattern.exec(prompt);\n\t}\n\treturn { matched: sawUnnegated, suppressed: sawMatch && !sawUnnegated };\n}\n\n// ============================================================================\n// Bounded compound-intent detection\n// ============================================================================\n\n/**\n * Splits off a bounded second clause after a short-range conjunction\n * (\"then\"/\"and then\"/\"and also\"/\";\"), only when the split point is within the\n * first `COMPOUND_SPLIT_MAX_INDEX_V4` characters (short direct commands, not\n * long prose briefs — verified this never fires inside the GOLD_SET's\n * >=2400-char plan entries). Returns null when no qualifying split exists.\n */\nfunction splitCompoundClauseV4(prompt: string): string | null {\n\tconst match = COMPOUND_SPLIT_PATTERN.exec(prompt);\n\tif (match === null || match.index > COMPOUND_SPLIT_MAX_INDEX_V4) return null;\n\tconst second = prompt.slice(match.index + match[0].length).trim();\n\tif (second.length < COMPOUND_SECOND_CLAUSE_MIN_CHARS_V4) return null;\n\treturn second;\n}\n\n/**\n * Leading-intent test for a compound prompt's SECOND clause. Reuses the same\n * `^`-anchored leading-intent tests as the primary clause (so it is equally\n * negation-immune), without the primary clause's diagnosticEvidence/localEdit\n * gating (a short second-clause fragment carries no such context of its own).\n */\nfunction secondClauseLeadingIntentV4(clause: string): TaskClassV4 | null {\n\tconst bounded = clause.slice(0, COMPOUND_SECOND_CLAUSE_SCAN_CHARS_V4);\n\treturn leadingIntentFromLeadingClauseV4(bounded);\n}\n\nfunction leadingIntentFromLeadingClauseV4(leading: string): TaskClassV4 | null {\n\tif (hasLeadingDebugAction(leading)) return \"debug\";\n\tif (hasLeadingReviewIntent(leading)) return \"review\";\n\tif (hasLeadingPlanIntent(leading)) return \"plan\";\n\tif (hasLeadingRefactorIntent(leading)) return \"refactor\";\n\tif (hasLeadingSimpleEditIntent(leading)) return \"simple-edit\";\n\tif (hasLeadingCodeGenIntent(leading)) return \"code-gen\";\n\treturn null;\n}\n\nfunction leadingIntentIsPostNegatedV4(leading: string, windowChars: number): boolean {\n\tconst leadingIntentMatch =\n\t\t/^(?:debug|investigate\\s+why|reproduce|trace\\b|fix\\s+this\\s+(?:traceback|panic|error)|review|audit(?!\\s+log\\b)|critique|inspect|assess|approve|double-?check|lgtm|plan|design|architect|decompose|write\\s+(?:a\\s+)?(?:technical\\s+)?(?:spec|roadmap|strategy|plan)|create\\s+(?:a\\s+)?(?:technical\\s+)?(?:spec|roadmap|strategy|plan)|refactor(?:ing|ed)?|extract|rename|deduplicate|consolidate|modularize|reorganize|simplify|restructure|untangle|clean\\s+up|split\\s+(?:the\\s+)?module|move\\s+logic|correct|update|swap|remove|adjust|fix|add|change|trim|reword|tweak|bump|implement|build|generate|scaffold|prototype)\\b/i.exec(\n\t\t\tleading,\n\t\t);\n\treturn (\n\t\tleadingIntentMatch !== null && hasPostPositionedNegationCueV4(leading, leadingIntentMatch[0].length, windowChars)\n\t);\n}\n\nfunction classifyShortKoreanZeroScoreTaskV4(prompt: string, windowChars: number): TaskClassV4 | null {\n\tif (matchUnnegated(prompt, KO_SHORT_DEBUG_SIGNAL_PATTERN_V4, windowChars).matched) return \"debug\";\n\tif (matchUnnegated(prompt, KO_SHORT_SIMPLE_EDIT_SIGNAL_PATTERN_V4, windowChars).matched) return \"simple-edit\";\n\tif (matchUnnegated(prompt, KO_SHORT_REVIEW_SIGNAL_PATTERN_V4, windowChars).matched) return \"review\";\n\tif (matchUnnegated(prompt, KO_SHORT_PLAN_SIGNAL_PATTERN_V4, windowChars).matched) return \"plan\";\n\tif (matchUnnegated(prompt, KO_SHORT_REFACTOR_SIGNAL_PATTERN_V4, windowChars).matched) return \"refactor\";\n\tif (matchUnnegated(prompt, KO_SHORT_CODE_GEN_SIGNAL_PATTERN_V4, windowChars).matched) return \"code-gen\";\n\treturn null;\n}\n\n// ============================================================================\n// Feature extraction\n// ============================================================================\n\ninterface ContextualFeaturesV4 {\n\treadonly firstClause: string;\n\treadonly codeFence: boolean;\n\treadonly diffHunk: boolean;\n\treadonly localEdit: boolean;\n\treadonly diagnosticEvidence: boolean;\n\treadonly reviewScope: boolean;\n\treadonly planBrief: boolean;\n\treadonly operationalRunbook: boolean;\n\treadonly refactorCue: boolean;\n\treadonly implementationObject: boolean;\n\treadonly leadingIntent: TaskClassV4 | null;\n\treadonly secondClauseIntent: TaskClassV4 | null;\n\treadonly compoundIntent: boolean;\n\treadonly keywordFamilyMatch: Readonly<Record<TaskClassV4, boolean>>;\n\treadonly addKeywordMatch: boolean;\n}\n\ninterface LeadingIntentInputV4 {\n\treadonly firstClause: string;\n\treadonly localEdit: boolean;\n\treadonly diagnosticEvidence: boolean;\n}\n\n/** Decision order for the leading intent classifier. */\nfunction leadingIntentV4(input: LeadingIntentInputV4): TaskClassV4 | null {\n\tif (\n\t\tinput.diagnosticEvidence &&\n\t\t(hasLeadingDebugAction(input.firstClause) || hasLeadingSimpleEditIntent(input.firstClause))\n\t)\n\t\treturn \"debug\";\n\tif (hasLeadingReviewIntent(input.firstClause)) return \"review\";\n\tif (hasLeadingPlanIntent(input.firstClause)) return \"plan\";\n\tif (hasLeadingRefactorIntent(input.firstClause)) return \"refactor\";\n\tif (input.localEdit && hasLeadingSimpleEditIntent(input.firstClause)) return \"simple-edit\";\n\tif (!input.localEdit && hasLeadingCodeGenIntent(input.firstClause)) return \"code-gen\";\n\treturn null;\n}\n\n/** Diagnostic-evidence decision structure, parameterized over precomputed negation-aware matches. */\nfunction hasDiagnosticEvidenceV4(\n\tleading: string,\n\thardDiagnosticMatched: boolean,\n\tbugObjectMatched: boolean,\n\tgenericDiagnosticMatched: boolean,\n\tnonDiagnosticContext: boolean,\n): boolean {\n\tif (hasLeadingDebugAction(leading) || hardDiagnosticMatched) return true;\n\tconst hasContextualDiagnostic = bugObjectMatched || (genericDiagnosticMatched && !nonDiagnosticContext);\n\tif (!hasContextualDiagnostic) return false;\n\treturn !hasLeadingReviewIntent(leading) && !hasLeadingPlanIntent(leading) && !hasLeadingRefactorIntent(leading);\n}\n\n/**\n * Extracts every v4 feature from one prompt, negation-gating every\n * whole-prompt scan exactly once and recording a `negation:<channel>` id in\n * `suppressed` whenever a match existed but every occurrence was negated.\n */\nfunction extractFeaturesV4(prompt: string, weights: RouterWeightsV4, suppressed: string[]): ContextualFeaturesV4 {\n\tconst leading = firstClause(prompt);\n\tconst window = weights.negationWindowChars;\n\n\tconst implementationObjectResult = matchUnnegated(prompt, IMPLEMENTATION_OBJECT_PATTERN, window);\n\tif (implementationObjectResult.suppressed) suppressed.push(\"negation:implementation-object\");\n\tconst implementationObject = implementationObjectResult.matched;\n\n\tconst lowRiskEditAction = LOW_RISK_EDIT_ACTION_PATTERN.test(leading);\n\tconst localEditObjectResult = matchUnnegated(prompt, LOCAL_EDIT_OBJECT_PATTERN, window);\n\tif (localEditObjectResult.suppressed) suppressed.push(\"negation:local-edit-object\");\n\tconst localEditActionResult = matchUnnegated(prompt, LOCAL_EDIT_ACTION_OBJECT_PATTERN, window);\n\tif (localEditActionResult.suppressed) suppressed.push(\"negation:local-edit-action-object\");\n\tconst localEdit =\n\t\t!implementationObject && (lowRiskEditAction || localEditObjectResult.matched || localEditActionResult.matched);\n\n\tconst hardDiagnosticResult = matchUnnegated(prompt, HARD_DIAGNOSTIC_PATTERN, window);\n\tif (hardDiagnosticResult.suppressed) suppressed.push(\"negation:hard-diagnostic\");\n\tconst bugObjectResult = matchUnnegated(prompt, BUG_OBJECT_PATTERN, window);\n\tif (bugObjectResult.suppressed) suppressed.push(\"negation:bug-object\");\n\tconst genericDiagnosticResult = matchUnnegated(prompt, GENERIC_DIAGNOSTIC_PATTERN, window);\n\tconst nonDiagnosticContext = NON_DIAGNOSTIC_DEBUG_CONTEXT_PATTERN.test(prompt);\n\tif (genericDiagnosticResult.suppressed && !nonDiagnosticContext) suppressed.push(\"negation:generic-diagnostic\");\n\tconst diagnosticEvidence = hasDiagnosticEvidenceV4(\n\t\tleading,\n\t\thardDiagnosticResult.matched,\n\t\tbugObjectResult.matched,\n\t\tgenericDiagnosticResult.matched,\n\t\tnonDiagnosticContext,\n\t);\n\n\tconst reviewScopeResult = matchUnnegated(prompt, REVIEW_SCOPE_PATTERN, window);\n\tif (reviewScopeResult.suppressed) suppressed.push(\"negation:review-scope\");\n\tconst reviewScope = hasLeadingReviewIntent(leading) && reviewScopeResult.matched;\n\n\tconst planBriefResult = matchUnnegated(prompt, PLAN_BRIEF_PATTERN, window);\n\tif (planBriefResult.suppressed) suppressed.push(\"negation:plan-brief\");\n\tconst planBrief =\n\t\thasLeadingPlanIntent(leading) && (planBriefResult.matched || prompt.length >= LONG_BRIEF_MIN_CHARS_V4);\n\n\tconst operationalRunbookResult = matchOperationalRunbookV4(prompt, window);\n\tif (operationalRunbookResult.suppressed) suppressed.push(\"negation:operational-runbook\");\n\tconst operationalRunbook = operationalRunbookResult.matched;\n\n\tconst refactorCueResult = matchUnnegated(prompt, REFACTOR_CUE_PATTERN, window);\n\tif (refactorCueResult.suppressed) suppressed.push(\"negation:refactor-cue\");\n\tconst refactorCue = refactorCueResult.matched;\n\n\tconst rawPrimaryIntent = leadingIntentV4({ firstClause: leading, localEdit, diagnosticEvidence });\n\tconst primaryIntent =\n\t\trawPrimaryIntent !== null && leadingIntentIsPostNegatedV4(leading, window) ? null : rawPrimaryIntent;\n\n\tconst secondClauseText = splitCompoundClauseV4(prompt);\n\tconst secondClauseIntent = secondClauseText === null ? null : secondClauseLeadingIntentV4(secondClauseText);\n\tconst compoundIntent = secondClauseIntent !== null && secondClauseIntent !== primaryIntent;\n\n\tconst debugKeywordResult = diagnosticEvidence\n\t\t? matchUnnegated(prompt, DEBUG_KEYWORD_FAMILY_PATTERN, window)\n\t\t: NO_MATCH_RESULT_V4;\n\tif (debugKeywordResult.suppressed) suppressed.push(\"negation:keyword-debug\");\n\n\tconst reviewKeywordResult = matchUnnegated(prompt, REVIEW_KEYWORD_FAMILY_PATTERN, window);\n\tif (reviewKeywordResult.suppressed) suppressed.push(\"negation:keyword-review\");\n\n\tconst planKeywordResult = matchUnnegated(prompt, PLAN_KEYWORD_FAMILY_PATTERN, window);\n\tif (planKeywordResult.suppressed) suppressed.push(\"negation:keyword-plan\");\n\n\tconst codeGenKeywordResult = matchUnnegated(prompt, CODE_GEN_KEYWORD_FAMILY_PATTERN, window);\n\tif (codeGenKeywordResult.suppressed) suppressed.push(\"negation:keyword-code-gen\");\n\n\tconst addKeywordResult = localEdit ? NO_MATCH_RESULT_V4 : matchUnnegated(prompt, ADD_KEYWORD_PATTERN, window);\n\tif (addKeywordResult.suppressed) suppressed.push(\"negation:add-keyword\");\n\n\treturn {\n\t\tfirstClause: leading,\n\t\tcodeFence: hasCodeFence(prompt),\n\t\tdiffHunk: hasDiffMarkers(prompt),\n\t\tlocalEdit,\n\t\tdiagnosticEvidence,\n\t\treviewScope,\n\t\tplanBrief,\n\t\toperationalRunbook,\n\t\trefactorCue,\n\t\timplementationObject,\n\t\tleadingIntent: primaryIntent,\n\t\tsecondClauseIntent,\n\t\tcompoundIntent,\n\t\tkeywordFamilyMatch: {\n\t\t\ttrivial: false,\n\t\t\tdebug: debugKeywordResult.matched,\n\t\t\trefactor: refactorCue,\n\t\t\treview: reviewKeywordResult.matched,\n\t\t\tplan: planKeywordResult.matched,\n\t\t\t\"simple-edit\": localEditObjectResult.matched,\n\t\t\t\"code-gen\": codeGenKeywordResult.matched,\n\t\t},\n\t\taddKeywordMatch: addKeywordResult.matched,\n\t};\n}\n\n// ============================================================================\n// Scoring\n// ============================================================================\n\nfunction emptyScoresV4(): Record<TaskClassV4, number> {\n\treturn {\n\t\ttrivial: 0,\n\t\t\"simple-edit\": 0,\n\t\t\"code-gen\": 0,\n\t\tdebug: 0,\n\t\trefactor: 0,\n\t\treview: 0,\n\t\tplan: 0,\n\t};\n}\n\nfunction computeScoresV4(features: ContextualFeaturesV4, weights: RouterWeightsV4): Record<TaskClassV4, number> {\n\tconst scores = emptyScoresV4();\n\tif (features.codeFence || features.diffHunk) scores[\"code-gen\"] += weights.codeFenceOrDiff;\n\tif (features.leadingIntent !== null) scores[features.leadingIntent] += weights.leadingIntent;\n\tif (features.secondClauseIntent !== null) scores[features.secondClauseIntent] += weights.secondClauseIntent;\n\tif (features.localEdit) scores[\"simple-edit\"] += weights.localEdit;\n\tif (features.diagnosticEvidence) scores.debug += weights.diagnosticEvidence;\n\tif (features.reviewScope) scores.review += weights.reviewScope;\n\tif (features.planBrief) scores.plan += weights.planBrief;\n\tif (features.operationalRunbook) scores.plan += weights.operationalRunbook;\n\tif (features.refactorCue) scores.refactor += weights.refactorCue;\n\tif (features.implementationObject) scores[\"code-gen\"] += weights.implementationObject;\n\n\tfor (const taskClass of TASK_CLASSES_V4) {\n\t\tif (features.keywordFamilyMatch[taskClass]) scores[taskClass] += weights.keywordFamily[taskClass];\n\t}\n\tif (features.addKeywordMatch) scores[\"code-gen\"] += weights.addKeyword;\n\n\treturn scores;\n}\n\n/**\n * Extension signals (history / context-pressure / judge vote). Every\n * coefficient is 0 under DEFAULT_WEIGHTS_V4 (inert until calibrated by a\n * future governance-backed lane); the mechanism exists so\n * `TaskClassifierInputV4`'s optional fields are meaningfully wired rather than\n * silently accepted-and-ignored.\n */\nfunction applyExtensionSignalsV4(\n\tscores: Record<TaskClassV4, number>,\n\tinput: TaskClassifierInputV4,\n\tweights: RouterWeightsV4,\n): void {\n\tconst priorClass = input.history !== undefined && input.history.length > 0 ? input.history[0] : null;\n\tif (priorClass !== null && weights.multiTurnPrior !== 0) scores[priorClass] += weights.multiTurnPrior;\n\n\tconst judgeVote = input.judgeVote ?? null;\n\tif (judgeVote !== null && weights.judgeVote !== 0) scores[judgeVote] += weights.judgeVote;\n\n\tconst pressureBucket = input.pressureBucket ?? 0;\n\tif (pressureBucket > 0 && weights.pressureBucket !== 0) {\n\t\tconst bump = pressureBucket * weights.pressureBucket;\n\t\tscores.debug += bump;\n\t\tscores.review += bump;\n\t\tscores.plan += bump;\n\t}\n}\n\n// ============================================================================\n// Public types\n// ============================================================================\n\n/**\n * Caller-supplied turn input. Identical shape to `TaskClassifierInputV3` /\n * `TaskClassifierInputV2` (prompt + optional lane/history/pressure/judge\n * slots), so existing callers can switch versions without reshaping data.\n */\nexport interface TaskClassifierInputV4 {\n\tprompt: string;\n\tlaneType?: ReasoningLaneTypeV4;\n\thistory?: readonly TaskClassV4[];\n\tpressureBucket?: number;\n\tjudgeVote?: TaskClassV4 | null;\n}\n\n/** Confidence band derived from `margin / topScore`, bucketed by `weights.lowConfidenceThreshold`/`highConfidenceThreshold`. */\nexport type ConfidenceBandV4 = \"low\" | \"medium\" | \"high\";\n\n/** Which branch of the zero-score fallback cascade decided `taskClass`; null when a real weighted signal decided it. */\nexport type FallbackReasonV4 =\n\t| \"code-fence-or-diff\"\n\t| \"trivial-length\"\n\t| \"ko-short-task-signal\"\n\t| \"long-prose\"\n\t| \"lane-fallback\"\n\t| \"default\";\n\n/**\n * The full, privacy-safe classification verdict (spec 008 Req 2). Every field\n * is a bounded enum, number, boolean, or a closed set of short diagnostic-id\n * strings (`suppressedFeatureIds`) — never raw prompt text, so this value is\n * safe to pass to evaluation, learning, and Adaptorch-advisory code.\n */\nexport interface ClassifierVerdictV4 {\n\t/** The decided task class (post zero-score-cascade if one fired). */\n\treadonly taskClass: TaskClassV4;\n\t/** Raw per-class score map from the weighted signals (pre-cascade; audit-only). */\n\treadonly scores: Readonly<Record<TaskClassV4, number>>;\n\t/** Second-highest-scoring class by the argmax, or null if every other class tied at the bottom. */\n\treadonly runnerUp: TaskClassV4 | null;\n\t/** `scores[argmaxClass] - scores[runnerUp]` (pre-cascade); 0 or negative-margin-free by construction (weights are non-negative). */\n\treadonly margin: number;\n\t/** Normalized confidence in [0, 1]: `margin / topScore`, or 0 when `topScore <= 0` (fallback cascade fired). */\n\treadonly confidence: number;\n\t/** Confidence bucketed by `weights.lowConfidenceThreshold` / `highConfidenceThreshold`. */\n\treadonly confidenceBand: ConfidenceBandV4;\n\t/** True when the argmax class won only via precedence-order tie-break over an exact score tie with the runner-up. */\n\treadonly tieBreak: boolean;\n\t/** Non-null exactly when the zero-score cascade (not a real weighted signal) decided `taskClass`. */\n\treadonly fallbackReason: FallbackReasonV4 | null;\n\t/** Audit trail of `negation:<channel>` ids for every whole-prompt signal that had a match but was fully negated. */\n\treadonly suppressedFeatureIds: readonly string[];\n\t/** True when a distinct second-clause leading intent was detected (bounded compound-intent detection). */\n\treadonly compoundIntent: boolean;\n\t/** The second clause's own leading intent, or null when no compound structure (or no distinct intent) was found. */\n\treadonly secondClauseIntent: TaskClassV4 | null;\n}\n\n// ============================================================================\n// Classifier\n// ============================================================================\n\n/**\n * Deterministic confidence-bearing classifier (spec 008 Req 2).\n *\n * Pipeline: extract negation/compound-aware contextual features -> score every\n * class from `weights` -> argmax with `TASK_CLASSES_V4` precedence tie-break,\n * tracking the runner-up -> compute margin/confidence/confidenceBand/tieBreak\n * from the RAW scores -> if `topScore <= 0`, replace `taskClass` (and record\n * `fallbackReason`) via the zero-score cascade (fence/diff -> code-gen;\n * length < 40 -> trivial; length >= 2400 -> plan; lane fallback; default\n * code-gen) — the raw scores/margin/runnerUp/tieBreak fields still reflect the\n * pre-cascade computation, for audit purposes.\n *\n * Pure: same (input, weights) -> same ClassifierVerdictV4.\n */\nexport function classifyTaskV4(\n\tinput: TaskClassifierInputV4,\n\tweights: RouterWeightsV4 = DEFAULT_WEIGHTS_V4,\n): ClassifierVerdictV4 {\n\tconst prompt = input.prompt.trim();\n\tconst suppressed: string[] = [];\n\tconst features = extractFeaturesV4(prompt, weights, suppressed);\n\tconst scores = computeScoresV4(features, weights);\n\tapplyExtensionSignalsV4(scores, input, weights);\n\n\tlet top: TaskClassV4 = TASK_CLASSES_V4[0];\n\tfor (const c of TASK_CLASSES_V4) {\n\t\tif (scores[c] > scores[top]) top = c;\n\t}\n\tlet runnerUp: TaskClassV4 | null = null;\n\tfor (const c of TASK_CLASSES_V4) {\n\t\tif (c === top) continue;\n\t\tif (runnerUp === null || scores[c] > scores[runnerUp]) runnerUp = c;\n\t}\n\n\tconst topScore = scores[top];\n\tconst runnerUpScore = runnerUp !== null ? scores[runnerUp] : 0;\n\tconst margin = topScore - runnerUpScore;\n\tconst tieBreak = runnerUp !== null && topScore === runnerUpScore;\n\tconst confidence = topScore <= 0 ? 0 : Math.max(0, Math.min(1, margin / topScore));\n\tconst confidenceBand: ConfidenceBandV4 =\n\t\tconfidence <= weights.lowConfidenceThreshold\n\t\t\t? \"low\"\n\t\t\t: confidence >= weights.highConfidenceThreshold\n\t\t\t\t? \"high\"\n\t\t\t\t: \"medium\";\n\n\tlet taskClass: TaskClassV4 = top;\n\tlet fallbackReason: FallbackReasonV4 | null = null;\n\tif (topScore <= 0) {\n\t\tif (features.codeFence || features.diffHunk) {\n\t\t\ttaskClass = \"code-gen\";\n\t\t\tfallbackReason = \"code-fence-or-diff\";\n\t\t} else if (prompt.length < TRIVIAL_MAX_CHARS_V4) {\n\t\t\tconst shortKoreanTaskClass = classifyShortKoreanZeroScoreTaskV4(prompt, weights.negationWindowChars);\n\t\t\tif (shortKoreanTaskClass !== null) {\n\t\t\t\ttaskClass = shortKoreanTaskClass;\n\t\t\t\tfallbackReason = \"ko-short-task-signal\";\n\t\t\t} else {\n\t\t\t\ttaskClass = \"trivial\";\n\t\t\t\tfallbackReason = \"trivial-length\";\n\t\t\t}\n\t\t} else if (prompt.length >= COMPLEX_PROSE_MIN_CHARS_V4) {\n\t\t\ttaskClass = \"plan\";\n\t\t\tfallbackReason = \"long-prose\";\n\t\t} else if (input.laneType !== undefined) {\n\t\t\ttaskClass = LANE_FALLBACK_CLASS_V4[input.laneType];\n\t\t\tfallbackReason = \"lane-fallback\";\n\t\t} else {\n\t\t\ttaskClass = \"code-gen\";\n\t\t\tfallbackReason = \"default\";\n\t\t}\n\t}\n\n\treturn {\n\t\ttaskClass,\n\t\tscores,\n\t\trunnerUp,\n\t\tmargin,\n\t\tconfidence,\n\t\tconfidenceBand,\n\t\ttieBreak,\n\t\tfallbackReason,\n\t\tsuppressedFeatureIds: suppressed,\n\t\tcompoundIntent: features.compoundIntent,\n\t\tsecondClauseIntent: features.secondClauseIntent,\n\t};\n}\n\n// ============================================================================\n// Resolver\n// ============================================================================\n\n/**\n * Static rule table: task class -> recommended ThinkingLevel. Re-exported\n * from the shared resolver core (byte-identical values; previously a\n * duplicated literal here — see reasoning-router-resolver.ts).\n */\nexport const TASK_CLASS_THINKING_LEVELS_V4 = TASK_CLASS_THINKING_LEVELS;\n\n/**\n * Bounded, strictly non-negative escalation applied when the verdict's own\n * confidence is low or a fallback (not a real signal) decided the class. This\n * is the ONLY place confidence touches the target index, and it can only add:\n * confidence can never lower effort (spec 008 Req 2 acceptance criterion).\n */\nconst LOW_CONFIDENCE_ESCALATION_STEPS_V4 = 1;\n\n/**\n * Thin auto-mode wrapper with no uncertainty adjustment: base rule table ->\n * lane step -> clamp to availableLevels. This is v4's \"confident\" path and\n * delegates to the shared resolver core with bias=0, hint=null, escalationSteps=0.\n */\nexport function resolveThinkingLevelV4ForAuto(\n\ttaskClass: TaskClassV4,\n\tavailableLevels: readonly ThinkingLevel[],\n\tlaneType: ReasoningLaneTypeV4 | undefined,\n): ThinkingLevel {\n\treturn resolveThinkingLevelCore(taskClass, availableLevels, laneType, 0, null, 0);\n}\n\n/**\n * Uncertainty-aware resolver (spec 008 Req 2 / plan.md step 5).\n *\n * Pipeline: base rule table for `verdict.taskClass` -> lane step -> bounded\n * bias [-2,+2] -> optional hint fusion (±2 bounded) -> bounded\n * non-negative confidence escalation (+1 ladder step exactly when\n * `verdict.confidenceBand === \"low\"` or `verdict.fallbackReason !== null`) ->\n * clamp to `availableLevels`.\n *\n * The confidence-escalation term is strictly `>= 0`: low confidence can only\n * hold the base+lane+bias+hint target or push it one step higher, never lower\n * it. This guarantees a low-confidence verdict never resolves BELOW what\n * `resolveThinkingLevelV4ForAuto` would give the same class (with bias=0,\n * hint=null) — text alone cannot talk the resolver down.\n */\nexport function resolveThinkingLevelV4WithUncertainty(\n\tverdict: ClassifierVerdictV4,\n\tavailableLevels: readonly ThinkingLevel[],\n\tlaneType: ReasoningLaneTypeV4 | undefined,\n\tbias = 0,\n\thint: { level: ThinkingLevel; confidence: number } | null = null,\n): ThinkingLevel {\n\tconst escalationSteps =\n\t\tverdict.confidenceBand === \"low\" || verdict.fallbackReason !== null ? LOW_CONFIDENCE_ESCALATION_STEPS_V4 : 0;\n\treturn resolveThinkingLevelCore(verdict.taskClass, availableLevels, laneType, bias, hint, escalationSteps);\n}\n"]}