open-multi-agent-kit 0.90.3 → 0.90.4

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

@@ -0,0 +1,683 @@
+/**
+ * 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).
+ *
+ * Deterministic by construction: same (input, weights) -> same
+ * ClassifierVerdictV4; same (verdict, availableLevels, laneType, bias, hint) ->
+ * same ThinkingLevel. No clock, randomness, I/O, model calls, network access,
+ * 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").
+ * ============================================================================
+ */
+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";
+/** 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). */
+const COMPLEX_PROSE_MIN_CHARS_V4 = 2400;
+/** Plan-brief prompts at/above this length count as a "long brief" even without PLAN_BRIEF_PATTERN. */
+const LONG_BRIEF_MIN_CHARS_V4 = 512;
+/** Compound-clause split only applies within this many leading characters (short direct commands, not long prose). */
+const COMPOUND_SPLIT_MAX_INDEX_V4 = 300;
+/** Minimum trimmed length of a candidate second clause to be worth checking for a leading intent. */
+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. */
+const LANE_FALLBACK_CLASS_V4 = {
+    planner: "plan",
+    security: "review",
+    explorer: "review",
+    coder: "code-gen",
+    reviewer: "review",
+    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).
+// ============================================================================
+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 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;
+const OPERATIONAL_RUNBOOK_SIGNAL_PATTERNS = [
+    { pattern: /\b(?:commit|commits|committed|committing)\b|커밋/i, critical: false },
+    { pattern: /\b(?:push|pushed|pushing)\b|푸시|푸쉬/i, critical: true },
+    {
+        pattern: /\b(?:release|releases|tag|version\s+bump|bump\s+(?:the\s+)?version)\b|릴리즈|태그|버전/i,
+        critical: true,
+    },
+    { pattern: /\bnpm\s+publish\b|\bpublish(?:ing|ed)?\b|퍼블리시|배포/i, critical: true },
+    { pattern: /\b(?:CHANGELOG\.md|changelog|release\s+notes?)\b|채널로그|체인지로그/i, critical: false },
+    { pattern: /\bREADME\.md\b|\breadme\b/i, critical: false },
+    { pattern: /\bci\s*\/\s*cd\b|\bgithub\s+actions?\b|\bworkflow\b|깃허브\s*액션|깃헙\s*액션/i, critical: false },
+];
+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 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;
+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_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". */
+const COMPOUND_SPLIT_PATTERN = /\band\s+then\b|\bthen\b|\band\s+also\b|;/i;
+/**
+ * Bounded negation-cue vocabulary (spec 008 plan.md "bounded negation window").
+ * Deliberately excludes bare "not" (too broad; verified against the current
+ * GOLD_SET that every listed cue here is either absent or occurs AFTER, not
+ * before, a tracked keyword — see this lane's evidence file for the audit).
+ */
+const NEGATION_CUE_PATTERN = /\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;
+const DIRECT_NOT_OBJECT_CUE_PATTERN_V4 = /\b(?:not|no)\s+(?:a\s+|an\s+|the\s+)?$/i;
+const DOUBLE_NEGATION_RESCUE_PATTERN_V4 = /\b(?:don't|do\s+not|doesn't|does\s+not|didn't|did\s+not|won't|will\s+not|shouldn't|should\s+not|wouldn't|would\s+not|can't|cannot)\s+(?:skip|avoid)\s+(?:the\s+|a\s+|an\s+)?$/i;
+const DELIBERATIVE_QUESTION_RESCUE_PATTERN_V4 = /\b(?:shouldn't|should\s+not|wouldn't|would\s+not|can't|cannot)\s+(?:we|i|you)\s*$/i;
+const POSTPOSITIONED_NEGATION_CUE_PATTERN_V4 = /^\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;
+/** A closed set of clause-boundary characters; a negation cue never reaches across one of these into a prior clause. */
+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. */
+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. */
+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);
+}
+// --- 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. ---
+function hasLeadingReviewIntent(text) {
+    return /^(?:review|audit(?!\s+log\b)|critique|inspect|assess|approve|double-?check|lgtm)\b/i.test(text);
+}
+function hasLeadingPlanIntent(text) {
+    return /^(?:plan|design|architect|decompose)\b|^(?:write|create)\s+(?:a\s+)?(?:technical\s+)?(?:spec|roadmap|strategy|plan)\b/i.test(text);
+}
+function hasLeadingRefactorIntent(text) {
+    return /^(?: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(text);
+}
+function hasLeadingDebugAction(text) {
+    return /^(?:debug|investigate\s+why|reproduce|trace\b)|^fix\s+this\s+(?:traceback|panic|error)\b/i.test(text);
+}
+function hasLeadingCodeGenIntent(text) {
+    return /^(?:implement|write|create|build|generate|scaffold|prototype|add)\b/i.test(text);
+}
+function hasLeadingSimpleEditIntent(text) {
+    return /^(?:correct|update|swap|remove|adjust|fix|add|change|trim|reword|tweak|bump)\b/i.test(text);
+}
+const NO_MATCH_RESULT_V4 = { matched: false, suppressed: false };
+function matchOperationalRunbookV4(prompt, windowChars) {
+    let distinctSignals = 0;
+    let criticalSignals = 0;
+    let sawSuppressed = false;
+    for (const { pattern, critical } of OPERATIONAL_RUNBOOK_SIGNAL_PATTERNS) {
+        const result = matchUnnegated(prompt, pattern, windowChars);
+        if (result.matched) {
+            distinctSignals += 1;
+            if (critical)
+                criticalSignals += 1;
+        }
+        if (result.suppressed)
+            sawSuppressed = true;
+    }
+    const explicitRelease = matchUnnegated(prompt, EXPLICIT_RELEASE_RUNBOOK_PATTERN, windowChars).matched;
+    const matched = explicitRelease || distinctSignals >= 3 || (criticalSignals >= 2 && distinctSignals >= 2);
+    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
+ * `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
+ * "하지 말고". `matched` is true iff at least one occurrence survives
+ * un-negated. Deterministic, single pass per pattern, no shared regex state (a
+ * fresh global-flag RegExp is constructed per call).
+ */
+function hasPrePositionedNegationCueV4(scoped) {
+    if (!NEGATION_CUE_PATTERN.test(scoped) && !DIRECT_NOT_OBJECT_CUE_PATTERN_V4.test(scoped))
+        return false;
+    return !DOUBLE_NEGATION_RESCUE_PATTERN_V4.test(scoped) && !DELIBERATIVE_QUESTION_RESCUE_PATTERN_V4.test(scoped);
+}
+function hasPostPositionedNegationCueV4(prompt, matchEnd, windowChars) {
+    const scoped = prompt.slice(matchEnd, Math.min(prompt.length, matchEnd + windowChars));
+    return POSTPOSITIONED_NEGATION_CUE_PATTERN_V4.test(scoped);
+}
+function matchUnnegated(prompt, pattern, windowChars) {
+    const flags = pattern.flags.includes("g") ? pattern.flags : `${pattern.flags}g`;
+    const globalPattern = new RegExp(pattern.source, flags);
+    let sawMatch = false;
+    let sawUnnegated = false;
+    let match = globalPattern.exec(prompt);
+    while (match !== null) {
+        sawMatch = true;
+        const matchStart = match.index;
+        const windowStart = Math.max(0, matchStart - windowChars);
+        const windowText = prompt.slice(windowStart, matchStart);
+        let boundary = -1;
+        for (const boundaryChar of CLAUSE_BOUNDARY_CHARS) {
+            const idx = windowText.lastIndexOf(boundaryChar);
+            if (idx > boundary)
+                boundary = idx;
+        }
+        const scoped = boundary >= 0 ? windowText.slice(boundary + 1) : windowText;
+        const preNegated = hasPrePositionedNegationCueV4(scoped);
+        const postNegated = !preNegated && hasPostPositionedNegationCueV4(prompt, matchStart + match[0].length, windowChars);
+        if (!preNegated && !postNegated)
+            sawUnnegated = true;
+        if (match[0].length === 0)
+            globalPattern.lastIndex += 1;
+        match = globalPattern.exec(prompt);
+    }
+    return { matched: sawUnnegated, suppressed: sawMatch && !sawUnnegated };
+}
+// ============================================================================
+// Bounded compound-intent detection
+// ============================================================================
+/**
+ * Splits off a bounded second clause after a short-range conjunction
+ * ("then"/"and then"/"and also"/";"), only when the split point is within the
+ * first `COMPOUND_SPLIT_MAX_INDEX_V4` characters (short direct commands, not
+ * long prose briefs — verified this never fires inside the GOLD_SET's
+ * >=2400-char plan entries). Returns null when no qualifying split exists.
+ */
+function splitCompoundClauseV4(prompt) {
+    const match = COMPOUND_SPLIT_PATTERN.exec(prompt);
+    if (match === null || match.index > COMPOUND_SPLIT_MAX_INDEX_V4)
+        return null;
+    const second = prompt.slice(match.index + match[0].length).trim();
+    if (second.length < COMPOUND_SECOND_CLAUSE_MIN_CHARS_V4)
+        return null;
+    return second;
+}
+/**
+ * Leading-intent test for a compound prompt's SECOND clause. Reuses the same
+ * `^`-anchored leading-intent tests as the primary clause (so it is equally
+ * negation-immune), without the primary clause's diagnosticEvidence/localEdit
+ * gating (a short second-clause fragment carries no such context of its own).
+ */
+function secondClauseLeadingIntentV4(clause) {
+    const bounded = clause.slice(0, COMPOUND_SECOND_CLAUSE_SCAN_CHARS_V4);
+    return leadingIntentFromLeadingClauseV4(bounded);
+}
+function leadingIntentFromLeadingClauseV4(leading) {
+    if (hasLeadingDebugAction(leading))
+        return "debug";
+    if (hasLeadingReviewIntent(leading))
+        return "review";
+    if (hasLeadingPlanIntent(leading))
+        return "plan";
+    if (hasLeadingRefactorIntent(leading))
+        return "refactor";
+    if (hasLeadingSimpleEditIntent(leading))
+        return "simple-edit";
+    if (hasLeadingCodeGenIntent(leading))
+        return "code-gen";
+    return null;
+}
+function leadingIntentIsPostNegatedV4(leading, windowChars) {
+    const leadingIntentMatch = /^(?: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(leading);
+    return (leadingIntentMatch !== null && hasPostPositionedNegationCueV4(leading, leadingIntentMatch[0].length, windowChars));
+}
+function classifyShortKoreanZeroScoreTaskV4(prompt, windowChars) {
+    if (matchUnnegated(prompt, KO_SHORT_DEBUG_SIGNAL_PATTERN_V4, windowChars).matched)
+        return "debug";
+    if (matchUnnegated(prompt, KO_SHORT_SIMPLE_EDIT_SIGNAL_PATTERN_V4, windowChars).matched)
+        return "simple-edit";
+    if (matchUnnegated(prompt, KO_SHORT_REVIEW_SIGNAL_PATTERN_V4, windowChars).matched)
+        return "review";
+    if (matchUnnegated(prompt, KO_SHORT_PLAN_SIGNAL_PATTERN_V4, windowChars).matched)
+        return "plan";
+    if (matchUnnegated(prompt, KO_SHORT_REFACTOR_SIGNAL_PATTERN_V4, windowChars).matched)
+        return "refactor";
+    if (matchUnnegated(prompt, KO_SHORT_CODE_GEN_SIGNAL_PATTERN_V4, windowChars).matched)
+        return "code-gen";
+    return null;
+}
+/** Identical decision order to v3's `leadingIntent`. */
+function leadingIntentV4(input) {
+    if (input.diagnosticEvidence &&
+        (hasLeadingDebugAction(input.firstClause) || hasLeadingSimpleEditIntent(input.firstClause)))
+        return "debug";
+    if (hasLeadingReviewIntent(input.firstClause))
+        return "review";
+    if (hasLeadingPlanIntent(input.firstClause))
+        return "plan";
+    if (hasLeadingRefactorIntent(input.firstClause))
+        return "refactor";
+    if (input.localEdit && hasLeadingSimpleEditIntent(input.firstClause))
+        return "simple-edit";
+    if (!input.localEdit && hasLeadingCodeGenIntent(input.firstClause))
+        return "code-gen";
+    return null;
+}
+/** Identical decision structure to v3's `hasDiagnosticEvidence`, parameterized over precomputed negation-aware matches. */
+function hasDiagnosticEvidenceV4(leading, hardDiagnosticMatched, bugObjectMatched, genericDiagnosticMatched, nonDiagnosticContext) {
+    if (hasLeadingDebugAction(leading) || hardDiagnosticMatched)
+        return true;
+    const hasContextualDiagnostic = bugObjectMatched || (genericDiagnosticMatched && !nonDiagnosticContext);
+    if (!hasContextualDiagnostic)
+        return false;
+    return !hasLeadingReviewIntent(leading) && !hasLeadingPlanIntent(leading) && !hasLeadingRefactorIntent(leading);
+}
+/**
+ * Extracts every v4 feature from one prompt, negation-gating every
+ * whole-prompt scan exactly once and recording a `negation:<channel>` id in
+ * `suppressed` whenever a match existed but every occurrence was negated.
+ */
+function extractFeaturesV4(prompt, weights, suppressed) {
+    const leading = firstClause(prompt);
+    const window = weights.negationWindowChars;
+    const implementationObjectResult = matchUnnegated(prompt, IMPLEMENTATION_OBJECT_PATTERN, window);
+    if (implementationObjectResult.suppressed)
+        suppressed.push("negation:implementation-object");
+    const implementationObject = implementationObjectResult.matched;
+    const lowRiskEditAction = LOW_RISK_EDIT_ACTION_PATTERN.test(leading);
+    const localEditObjectResult = matchUnnegated(prompt, LOCAL_EDIT_OBJECT_PATTERN, window);
+    if (localEditObjectResult.suppressed)
+        suppressed.push("negation:local-edit-object");
+    const localEditActionResult = matchUnnegated(prompt, LOCAL_EDIT_ACTION_OBJECT_PATTERN, window);
+    if (localEditActionResult.suppressed)
+        suppressed.push("negation:local-edit-action-object");
+    const localEdit = !implementationObject && (lowRiskEditAction || localEditObjectResult.matched || localEditActionResult.matched);
+    const hardDiagnosticResult = matchUnnegated(prompt, HARD_DIAGNOSTIC_PATTERN, window);
+    if (hardDiagnosticResult.suppressed)
+        suppressed.push("negation:hard-diagnostic");
+    const bugObjectResult = matchUnnegated(prompt, BUG_OBJECT_PATTERN, window);
+    if (bugObjectResult.suppressed)
+        suppressed.push("negation:bug-object");
+    const genericDiagnosticResult = matchUnnegated(prompt, GENERIC_DIAGNOSTIC_PATTERN, window);
+    const nonDiagnosticContext = NON_DIAGNOSTIC_DEBUG_CONTEXT_PATTERN.test(prompt);
+    if (genericDiagnosticResult.suppressed && !nonDiagnosticContext)
+        suppressed.push("negation:generic-diagnostic");
+    const diagnosticEvidence = hasDiagnosticEvidenceV4(leading, hardDiagnosticResult.matched, bugObjectResult.matched, genericDiagnosticResult.matched, nonDiagnosticContext);
+    const reviewScopeResult = matchUnnegated(prompt, REVIEW_SCOPE_PATTERN, window);
+    if (reviewScopeResult.suppressed)
+        suppressed.push("negation:review-scope");
+    const reviewScope = hasLeadingReviewIntent(leading) && reviewScopeResult.matched;
+    const planBriefResult = matchUnnegated(prompt, PLAN_BRIEF_PATTERN, window);
+    if (planBriefResult.suppressed)
+        suppressed.push("negation:plan-brief");
+    const planBrief = hasLeadingPlanIntent(leading) && (planBriefResult.matched || prompt.length >= LONG_BRIEF_MIN_CHARS_V4);
+    const operationalRunbookResult = matchOperationalRunbookV4(prompt, window);
+    if (operationalRunbookResult.suppressed)
+        suppressed.push("negation:operational-runbook");
+    const operationalRunbook = operationalRunbookResult.matched;
+    const refactorCueResult = matchUnnegated(prompt, REFACTOR_CUE_PATTERN, window);
+    if (refactorCueResult.suppressed)
+        suppressed.push("negation:refactor-cue");
+    const refactorCue = refactorCueResult.matched;
+    const rawPrimaryIntent = leadingIntentV4({ firstClause: leading, localEdit, diagnosticEvidence });
+    const primaryIntent = rawPrimaryIntent !== null && leadingIntentIsPostNegatedV4(leading, window) ? null : rawPrimaryIntent;
+    const secondClauseText = splitCompoundClauseV4(prompt);
+    const secondClauseIntent = secondClauseText === null ? null : secondClauseLeadingIntentV4(secondClauseText);
+    const compoundIntent = secondClauseIntent !== null && secondClauseIntent !== primaryIntent;
+    const debugKeywordResult = diagnosticEvidence
+        ? matchUnnegated(prompt, DEBUG_KEYWORD_FAMILY_PATTERN, window)
+        : NO_MATCH_RESULT_V4;
+    if (debugKeywordResult.suppressed)
+        suppressed.push("negation:keyword-debug");
+    const reviewKeywordResult = matchUnnegated(prompt, REVIEW_KEYWORD_FAMILY_PATTERN, window);
+    if (reviewKeywordResult.suppressed)
+        suppressed.push("negation:keyword-review");
+    const planKeywordResult = matchUnnegated(prompt, PLAN_KEYWORD_FAMILY_PATTERN, window);
+    if (planKeywordResult.suppressed)
+        suppressed.push("negation:keyword-plan");
+    const codeGenKeywordResult = matchUnnegated(prompt, CODE_GEN_KEYWORD_FAMILY_PATTERN, window);
+    if (codeGenKeywordResult.suppressed)
+        suppressed.push("negation:keyword-code-gen");
+    const addKeywordResult = localEdit ? NO_MATCH_RESULT_V4 : matchUnnegated(prompt, ADD_KEYWORD_PATTERN, window);
+    if (addKeywordResult.suppressed)
+        suppressed.push("negation:add-keyword");
+    return {
+        firstClause: leading,
+        codeFence: hasCodeFence(prompt),
+        diffHunk: hasDiffMarkers(prompt),
+        localEdit,
+        diagnosticEvidence,
+        reviewScope,
+        planBrief,
+        operationalRunbook,
+        refactorCue,
+        implementationObject,
+        leadingIntent: primaryIntent,
+        secondClauseIntent,
+        compoundIntent,
+        keywordFamilyMatch: {
+            trivial: false,
+            debug: debugKeywordResult.matched,
+            refactor: refactorCue,
+            review: reviewKeywordResult.matched,
+            plan: planKeywordResult.matched,
+            "simple-edit": localEditObjectResult.matched,
+            "code-gen": codeGenKeywordResult.matched,
+        },
+        addKeywordMatch: addKeywordResult.matched,
+    };
+}
+// ============================================================================
+// Scoring
+// ============================================================================
+function emptyScoresV4() {
+    return {
+        trivial: 0,
+        "simple-edit": 0,
+        "code-gen": 0,
+        debug: 0,
+        refactor: 0,
+        review: 0,
+        plan: 0,
+    };
+}
+function computeScoresV4(features, weights) {
+    const scores = emptyScoresV4();
+    if (features.codeFence || features.diffHunk)
+        scores["code-gen"] += weights.codeFenceOrDiff;
+    if (features.leadingIntent !== null)
+        scores[features.leadingIntent] += weights.leadingIntent;
+    if (features.secondClauseIntent !== null)
+        scores[features.secondClauseIntent] += weights.secondClauseIntent;
+    if (features.localEdit)
+        scores["simple-edit"] += weights.localEdit;
+    if (features.diagnosticEvidence)
+        scores.debug += weights.diagnosticEvidence;
+    if (features.reviewScope)
+        scores.review += weights.reviewScope;
+    if (features.planBrief)
+        scores.plan += weights.planBrief;
+    if (features.operationalRunbook)
+        scores.plan += weights.operationalRunbook;
+    if (features.refactorCue)
+        scores.refactor += weights.refactorCue;
+    if (features.implementationObject)
+        scores["code-gen"] += weights.implementationObject;
+    for (const taskClass of TASK_CLASSES_V4) {
+        if (features.keywordFamilyMatch[taskClass])
+            scores[taskClass] += weights.keywordFamily[taskClass];
+    }
+    if (features.addKeywordMatch)
+        scores["code-gen"] += weights.addKeyword;
+    return scores;
+}
+/**
+ * Extension signals (history / context-pressure / judge vote). Every
+ * 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).
+ */
+function applyExtensionSignalsV4(scores, input, weights) {
+    const priorClass = input.history !== undefined && input.history.length > 0 ? input.history[0] : null;
+    if (priorClass !== null && weights.multiTurnPrior !== 0)
+        scores[priorClass] += weights.multiTurnPrior;
+    const judgeVote = input.judgeVote ?? null;
+    if (judgeVote !== null && weights.judgeVote !== 0)
+        scores[judgeVote] += weights.judgeVote;
+    const pressureBucket = input.pressureBucket ?? 0;
+    if (pressureBucket > 0 && weights.pressureBucket !== 0) {
+        const bump = pressureBucket * weights.pressureBucket;
+        scores.debug += bump;
+        scores.review += bump;
+        scores.plan += bump;
+    }
+}
+// ============================================================================
+// Classifier
+// ============================================================================
+/**
+ * Deterministic confidence-bearing classifier (spec 008 Req 2).
+ *
+ * Pipeline: extract negation/compound-aware contextual features -> score every
+ * 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.
+ *
+ * Pure: same (input, weights) -> same ClassifierVerdictV4.
+ */
+export function classifyTaskV4(input, weights = DEFAULT_WEIGHTS_V4) {
+    const prompt = input.prompt.trim();
+    const suppressed = [];
+    const features = extractFeaturesV4(prompt, weights, suppressed);
+    const scores = computeScoresV4(features, weights);
+    applyExtensionSignalsV4(scores, input, weights);
+    let top = TASK_CLASSES_V4[0];
+    for (const c of TASK_CLASSES_V4) {
+        if (scores[c] > scores[top])
+            top = c;
+    }
+    let runnerUp = null;
+    for (const c of TASK_CLASSES_V4) {
+        if (c === top)
+            continue;
+        if (runnerUp === null || scores[c] > scores[runnerUp])
+            runnerUp = c;
+    }
+    const topScore = scores[top];
+    const runnerUpScore = runnerUp !== null ? scores[runnerUp] : 0;
+    const margin = topScore - runnerUpScore;
+    const tieBreak = runnerUp !== null && topScore === runnerUpScore;
+    const confidence = topScore <= 0 ? 0 : Math.max(0, Math.min(1, margin / topScore));
+    const confidenceBand = confidence <= weights.lowConfidenceThreshold
+        ? "low"
+        : confidence >= weights.highConfidenceThreshold
+            ? "high"
+            : "medium";
+    let taskClass = top;
+    let fallbackReason = null;
+    if (topScore <= 0) {
+        if (features.codeFence || features.diffHunk) {
+            taskClass = "code-gen";
+            fallbackReason = "code-fence-or-diff";
+        }
+        else if (prompt.length < TRIVIAL_MAX_CHARS_V4) {
+            const shortKoreanTaskClass = classifyShortKoreanZeroScoreTaskV4(prompt, weights.negationWindowChars);
+            if (shortKoreanTaskClass !== null) {
+                taskClass = shortKoreanTaskClass;
+                fallbackReason = "ko-short-task-signal";
+            }
+            else {
+                taskClass = "trivial";
+                fallbackReason = "trivial-length";
+            }
+        }
+        else if (prompt.length >= COMPLEX_PROSE_MIN_CHARS_V4) {
+            taskClass = "plan";
+            fallbackReason = "long-prose";
+        }
+        else if (input.laneType !== undefined) {
+            taskClass = LANE_FALLBACK_CLASS_V4[input.laneType];
+            fallbackReason = "lane-fallback";
+        }
+        else {
+            taskClass = "code-gen";
+            fallbackReason = "default";
+        }
+    }
+    return {
+        taskClass,
+        scores,
+        runnerUp,
+        margin,
+        confidence,
+        confidenceBand,
+        tieBreak,
+        fallbackReason,
+        suppressedFeatureIds: suppressed,
+        compoundIntent: features.compoundIntent,
+        secondClauseIntent: features.secondClauseIntent,
+    };
+}
+// ============================================================================
+// 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;
+/**
+ * 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
+ * is the ONLY place confidence touches the target index, and it can only add:
+ * 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.
+ */
+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);
+}
+/**
+ * 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
+ * non-negative confidence escalation (+1 ladder step exactly when
+ * `verdict.confidenceBand === "low"` or `verdict.fallbackReason !== null`) ->
+ * clamp to `availableLevels`.
+ *
+ * The confidence-escalation term is strictly `>= 0`: low confidence can only
+ * hold the base+lane+bias+hint target or push it one step higher, never lower
+ * it. This guarantees a low-confidence verdict never resolves BELOW what
+ * `resolveThinkingLevelV4ForAuto` would give the same class (with bias=0,
+ * 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);
+}
+//# sourceMappingURL=reasoning-router-v4.js.map