open-multi-agent-kit 0.90.4 → 0.90.5

package/dist/core/reasoning-router-v4.js

@@ -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) ->
@@ -15,60 +7,21 @@
  * 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 { DEFAULT_WEIGHTS_V4, TASK_CLASSES_V4, } from "./reasoning-router-v4-weights.js";
 export { DEFAULT_WEIGHTS_V4, TASK_CLASSES_V4 } from "./reasoning-router-v4-weights.js";
+import { resolveThinkingLevelCore, TASK_CLASS_THINKING_LEVELS } from "./reasoning-router-resolver.js";
 /** Prompts shorter than this (trimmed) fall back to trivial when no real signal scores > 0. */
 const TRIVIAL_MAX_CHARS_V4 = 40;
 /** Prompts at/above this length with no real signal fall back to plan (long prose brief). */
@@ -81,7 +34,7 @@ const COMPOUND_SPLIT_MAX_INDEX_V4 = 300;
 const COMPOUND_SECOND_CLAUSE_MIN_CHARS_V4 = 3;
 /** Bounded look-ahead cap when checking a second clause's leading intent (defense in depth; independent of regex cost). */
 const COMPOUND_SECOND_CLAUSE_SCAN_CHARS_V4 = 200;
-/** Lane fallback class, used only when every scored class is <= 0 (zero-score cascade). Identical to v1/v2/v3. */
+/** Lane fallback class, used only when every scored class is <= 0 (zero-score cascade). */
 const LANE_FALLBACK_CLASS_V4 = {
     planner: "plan",
     security: "review",
@@ -91,15 +44,13 @@ const LANE_FALLBACK_CLASS_V4 = {
     tester: "code-gen",
 };
 // ============================================================================
-// Whole-prompt / leading-clause patterns (copied verbatim from
-// reasoning-router-v3.ts; identical regex source so DEFAULT_WEIGHTS_V4
-// reproduces v3 exactly whenever no negation cue is present).
+// Whole-prompt / leading-clause patterns used by the v4 scorer.
 // ============================================================================
 const LOCAL_EDIT_OBJECT_PATTERN = /\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;
 const LOCAL_EDIT_ACTION_OBJECT_PATTERN = /\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;
 const IMPLEMENTATION_OBJECT_PATTERN = /\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;
 const HARD_DIAGNOSTIC_PATTERN = /\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;
-const BUG_OBJECT_PATTERN = /\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;
+const BUG_OBJECT_PATTERN = /\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;
 const NON_DIAGNOSTIC_DEBUG_CONTEXT_PATTERN = /\berror\s+(?:handling|messages?|budgets?)\b|에러\s*처리|오류\s*처리/i;
 const GENERIC_DIAGNOSTIC_PATTERN = /\b(errors?|broken|wrong\s+results|rolls?\s+back|rollback)\b/i;
 const PLAN_BRIEF_PATTERN = /\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;
@@ -117,18 +68,18 @@ const OPERATIONAL_RUNBOOK_SIGNAL_PATTERNS = [
 ];
 const EXPLICIT_RELEASE_RUNBOOK_PATTERN = /\bnpm\s+publish\b|\brelease\b[^\n]{0,80}\bv?\d+\.\d+\.\d+\b|릴리즈[^\n]{0,80}\d+\.\d+\.\d+/i;
 const REVIEW_SCOPE_PATTERN = /\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;
-const REFACTOR_CUE_PATTERN = /\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;
+const REFACTOR_CUE_PATTERN = /\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;
 const ADD_KEYWORD_PATTERN = /\badd\b/i;
 const LOW_RISK_EDIT_ACTION_PATTERN = /^(?:correct|update|swap|remove|reword|tweak|bump|trim)\b/i;
-/** Whole-prompt keyword-family patterns (mirrors v3's KEYWORD_PATTERNS_V3; refactor/simple-edit reuse the patterns above). */
-const DEBUG_KEYWORD_FAMILY_PATTERN = /^(?:debug|investigate\s+why|reproduce|trace\b)|\bfix\s+this\s+(?:traceback|panic|error)\b|디버깅|디버그|재현|원인\s*분석/i;
-const REVIEW_KEYWORD_FAMILY_PATTERN = /\b(review|critique|assess|inspect|approve|lgtm|double-?check|audit(?!\s+log\b))\b|리뷰|검토|점검/i;
-const PLAN_KEYWORD_FAMILY_PATTERN = /\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;
+/** Whole-prompt keyword-family patterns; refactor/simple-edit reuse the patterns above. */
+const DEBUG_KEYWORD_FAMILY_PATTERN = /^(?: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;
+const REVIEW_KEYWORD_FAMILY_PATTERN = /\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;
+const PLAN_KEYWORD_FAMILY_PATTERN = /\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;
 const CODE_GEN_KEYWORD_FAMILY_PATTERN = /\b(implement|write|create|build|generate|scaffold|prototype)\b|구현|만들어\s*줘|생성|작성|추가해/i;
 const KO_SHORT_DEBUG_SIGNAL_PATTERN_V4 = /오류|에러|실패|안\s*돼|안\s*됨|깨졌|깨짐|고장|디버깅|디버그|재현|원인\s*(?:분석|파악)?/i;
 const KO_SHORT_SIMPLE_EDIT_SIGNAL_PATTERN_V4 = /오타|맞춤법|띄어쓰기|문구|제목/i;
 const KO_SHORT_REVIEW_SIGNAL_PATTERN_V4 = /리뷰|검토|점검/i;
-const KO_SHORT_PLAN_SIGNAL_PATTERN_V4 = /설계|로드맵|아키텍처|계획\s*세워|기획/i;
+const KO_SHORT_PLAN_SIGNAL_PATTERN_V4 = /설계|로드맵|아키텍처|계획.{0,6}세워|기획/i;
 const KO_SHORT_REFACTOR_SIGNAL_PATTERN_V4 = /리팩토링|리팩터링|구조\s*개선/i;
 const KO_SHORT_CODE_GEN_SIGNAL_PATTERN_V4 = /구현|만들|생성|작성|추가|테스트|삭제|지워|지우|제거|고쳐|고치|수정|변경|바꿔|바꾸|옮겨|옮기|넣어/i;
 /** Short-range compound-clause conjunction boundary. Only bare/compound "then"/"and also"/";" — never bare "and". */
@@ -149,13 +100,13 @@ const CLAUSE_BOUNDARY_CHARS = [".", "!", "?", ";", ","];
 function hasCodeFence(text) {
     return text.includes("```");
 }
-/** Identical to v1/v2/v3: explicit hunk headers or `diff --git` count alone; bare +/- only count together. */
+/** Explicit hunk headers or `diff --git` count alone; bare +/- only count together. */
 function hasDiffMarkers(text) {
     if (/^@@[^\n]*@@/m.test(text) || /^diff --git /m.test(text))
         return true;
     return /^\+(?!\+)/m.test(text) && /^-(?!-)/m.test(text);
 }
-/** Identical to v3: first line (after stripping one polite prefix), capped at 180 chars. */
+/** First line (after stripping one polite prefix), capped at 180 chars. */
 function firstClause(prompt) {
     const firstLine = prompt.replace(/^(?:please|pls|can you|could you|would you|help me|i need you to)\s+/i, "").split("\n")[0] ?? "";
     return firstLine.slice(0, 180);
@@ -163,8 +114,7 @@ function firstClause(prompt) {
 // --- Leading-clause intent tests. All are `^`-anchored against the LEADING
 // clause text only, so a negation cue at the start ("don't refactor...") can
 // never match one of these: none of the alternations include "don't"/"never"/
-// etc, so they are negation-immune by construction and are left unchanged
-// from v3. ---
+// etc, so they are negation-immune by construction.
 function hasLeadingReviewIntent(text) {
     return /^(?:review|audit(?!\s+log\b)|critique|inspect|assess|approve|double-?check|lgtm)\b/i.test(text);
 }
@@ -203,8 +153,8 @@ function matchOperationalRunbookV4(prompt, windowChars) {
     return { matched, suppressed: !matched && sawSuppressed };
 }
 /**
- * Negation-aware replacement for v3's bare `pattern.test(prompt)`. Finds every
- * occurrence of `pattern` in `prompt`; for each, looks back up to
+ * Negation-aware whole-prompt pattern scan. Finds every occurrence of `pattern`
+ * in `prompt`; for each, looks back up to
  * `windowChars` characters (never crossing a `.,;!?` boundary into a prior
  * clause) for a `NEGATION_CUE_PATTERN` hit, and looks forward inside the same
  * bounded window for Hangul-only post-positioned negation cues such as
@@ -312,7 +262,7 @@ function classifyShortKoreanZeroScoreTaskV4(prompt, windowChars) {
         return "code-gen";
     return null;
 }
-/** Identical decision order to v3's `leadingIntent`. */
+/** Decision order for the leading intent classifier. */
 function leadingIntentV4(input) {
     if (input.diagnosticEvidence &&
         (hasLeadingDebugAction(input.firstClause) || hasLeadingSimpleEditIntent(input.firstClause)))
@@ -329,7 +279,7 @@ function leadingIntentV4(input) {
         return "code-gen";
     return null;
 }
-/** Identical decision structure to v3's `hasDiagnosticEvidence`, parameterized over precomputed negation-aware matches. */
+/** Diagnostic-evidence decision structure, parameterized over precomputed negation-aware matches. */
 function hasDiagnosticEvidenceV4(leading, hardDiagnosticMatched, bugObjectMatched, genericDiagnosticMatched, nonDiagnosticContext) {
     if (hasLeadingDebugAction(leading) || hardDiagnosticMatched)
         return true;
@@ -482,7 +432,7 @@ function computeScoresV4(features, weights) {
  * coefficient is 0 under DEFAULT_WEIGHTS_V4 (inert until calibrated by a
  * future governance-backed lane); the mechanism exists so
  * `TaskClassifierInputV4`'s optional fields are meaningfully wired rather than
- * silently accepted-and-ignored (as they are in v3's identical-shaped input).
+ * silently accepted-and-ignored.
  */
 function applyExtensionSignalsV4(scores, input, weights) {
     const priorClass = input.history !== undefined && input.history.length > 0 ? input.history[0] : null;
@@ -509,10 +459,10 @@ function applyExtensionSignalsV4(scores, input, weights) {
  * 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.
  */
@@ -592,28 +542,12 @@ export function classifyTaskV4(input, weights = DEFAULT_WEIGHTS_V4) {
 // ============================================================================
 // Resolver
 // ============================================================================
-/** Reasoning ladder used for targets and clamping. Intentionally excludes "off". Identical to v1/v2/v3. */
-const REASONING_LADDER_V4 = ["minimal", "low", "medium", "high", "xhigh", "max"];
-/** Static rule table: task class -> recommended ThinkingLevel. Identical to v1/v2/v3. */
-export const TASK_CLASS_THINKING_LEVELS_V4 = {
-    trivial: "minimal",
-    "simple-edit": "low",
-    "code-gen": "medium",
-    debug: "high",
-    refactor: "high",
-    review: "high",
-    plan: "xhigh",
-};
-/** Lane ladder adjustment. Identical to v1/v2/v3: planner/security escalate one step, explorer de-escalates. */
-const LANE_STEP_V4 = {
-    planner: 1,
-    security: 1,
-    explorer: -1,
-};
-/** Maximum magnitude of the learning/consult bias (ladder steps). Identical bound to v2/v3. */
-const BIAS_MAX_V4 = 2;
-/** Confidence floor at which an override hint is fused into the target. Identical to v2/v3. */
-const HINT_CONFIDENCE_THRESHOLD_V4 = 0.7;
+/**
+ * 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 const TASK_CLASS_THINKING_LEVELS_V4 = TASK_CLASS_THINKING_LEVELS;
 /**
  * Bounded, strictly non-negative escalation applied when the verdict's own
  * confidence is low or a fallback (not a real signal) decided the class. This
@@ -621,37 +555,19 @@ const HINT_CONFIDENCE_THRESHOLD_V4 = 0.7;
  * confidence can never lower effort (spec 008 Req 2 acceptance criterion).
  */
 const LOW_CONFIDENCE_ESCALATION_STEPS_V4 = 1;
-function clampToLadderIndexV4(index) {
-    return Math.max(0, Math.min(index, REASONING_LADDER_V4.length - 1));
-}
-function clampToAvailableV4(targetIndex, availableLevels) {
-    const availableOnLadder = REASONING_LADDER_V4.filter((level) => availableLevels.includes(level));
-    if (availableOnLadder.length === 0)
-        return availableLevels[0] ?? "off";
-    for (let i = targetIndex; i >= 0; i--) {
-        const candidate = REASONING_LADDER_V4[i];
-        if (availableOnLadder.includes(candidate))
-            return candidate;
-    }
-    return availableOnLadder[0];
-}
 /**
  * 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 function resolveThinkingLevelV4ForAuto(taskClass, availableLevels, laneType) {
-    const baseIndex = REASONING_LADDER_V4.indexOf(TASK_CLASS_THINKING_LEVELS_V4[taskClass]);
-    const laneStep = laneType ? (LANE_STEP_V4[laneType] ?? 0) : 0;
-    const targetIndex = clampToLadderIndexV4(baseIndex + laneStep);
-    return clampToAvailableV4(targetIndex, availableLevels);
+    return resolveThinkingLevelCore(taskClass, availableLevels, laneType, 0, null, 0);
 }
 /**
  * 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`.
@@ -663,21 +579,7 @@ export function resolveThinkingLevelV4ForAuto(taskClass, availableLevels, laneTy
  * hint=null) — text alone cannot talk the resolver down.
  */
 export function resolveThinkingLevelV4WithUncertainty(verdict, availableLevels, laneType, bias = 0, hint = null) {
-    const baseIndex = REASONING_LADDER_V4.indexOf(TASK_CLASS_THINKING_LEVELS_V4[verdict.taskClass]);
-    const laneStep = laneType ? (LANE_STEP_V4[laneType] ?? 0) : 0;
-    const biasClamped = Math.max(-BIAS_MAX_V4, Math.min(BIAS_MAX_V4, bias));
-    let targetIndex = clampToLadderIndexV4(baseIndex + laneStep + biasClamped);
-    if (hint !== null && hint.confidence >= HINT_CONFIDENCE_THRESHOLD_V4) {
-        const hintIndex = REASONING_LADDER_V4.indexOf(hint.level);
-        if (hintIndex >= 0) {
-            const delta = hintIndex - targetIndex;
-            const step = Math.max(-BIAS_MAX_V4, Math.min(BIAS_MAX_V4, delta));
-            targetIndex = clampToLadderIndexV4(targetIndex + step);
-        }
-    }
-    if (verdict.confidenceBand === "low" || verdict.fallbackReason !== null) {
-        targetIndex = clampToLadderIndexV4(targetIndex + LOW_CONFIDENCE_ESCALATION_STEPS_V4);
-    }
-    return clampToAvailableV4(targetIndex, availableLevels);
+    const escalationSteps = verdict.confidenceBand === "low" || verdict.fallbackReason !== null ? LOW_CONFIDENCE_ESCALATION_STEPS_V4 : 0;
+    return resolveThinkingLevelCore(verdict.taskClass, availableLevels, laneType, bias, hint, escalationSteps);
 }
 //# sourceMappingURL=reasoning-router-v4.js.map