feat: complexity router : adds complexity_tier CEL routing

plugins/governance/complexity/analyzer.go

@@ -0,0 +1,189 @@
+package complexity
+
+import "math"
+
+// ComplexityAnalyzer computes complexity scores from normalized text input.
+// It is stateless and safe for concurrent use.
+type ComplexityAnalyzer struct {
+	matcher *compiledKeywordMatcher
+}
+
+// NewComplexityAnalyzer creates a stateless analyzer with built-in defaults.
+func NewComplexityAnalyzer() *ComplexityAnalyzer {
+	return &ComplexityAnalyzer{
+		matcher: newCompiledKeywordMatcher(),
+	}
+}
+
+// Analyze computes complexity scores from the normalized input.
+func (a *ComplexityAnalyzer) Analyze(input ComplexityInput) *ComplexityResult {
+	// Select scan mask based on whether conversation history is present.
+	lastScanMask := lastTextBaseScanMask
+	if len(input.PriorUserTexts) > 0 {
+		lastScanMask = lastTextFullScanMask
+	}
+
+	// Extract lexical signals from last user message and system prompt.
+	lastSignals := a.matcher.analyzeText(input.LastUserText, lastScanMask)
+	systemSignals := a.matcher.analyzeText(input.SystemText, systemTextScanMask)
+
+	// Score primary message signals.
+	userCodeScore := scoreCount(lastSignals.codeCount, 3)
+	reasoningScore := scoreCount(lastSignals.reasoningCount, 2)
+	userTechnicalScore := scoreCount(lastSignals.technicalCount, 3)
+	userSimpleScore := scoreCount(lastSignals.simpleCount, 2)
+	outputScore := scoreOutputComplexity(lastSignals)
+	tokenScore := scoreTokenCount(lastSignals.wordCount)
+
+	// System prompt provides soft lexical context for code/technical/simple signals,
+	// but never drives reasoning override, token count, or output complexity.
+	systemCodeScore := scoreCount(systemSignals.codeCount, 3)
+	systemTechnicalScore := scoreCount(systemSignals.technicalCount, 3)
+	systemSimpleScore := scoreCount(systemSignals.simpleCount, 2)
+
+	codeScore := clamp(userCodeScore+(systemCodeScore*systemPromptAssistFactor), 0.0, 1.0)
+	technicalScore := clamp(userTechnicalScore+(systemTechnicalScore*systemPromptAssistFactor), 0.0, 1.0)
+	simpleScore := clamp(userSimpleScore+(systemSimpleScore*systemPromptAssistFactor), 0.0, 1.0)
+
+	// Conditional simple dampener: only apply full dampener on short, low-signal asks.
+	wordCount := lastSignals.wordCount
+	effectiveSimpleWeight := simpleWeight
+	signalCount := 0
+	if userCodeScore >= 0.3 {
+		signalCount++
+	}
+	if userTechnicalScore >= 0.3 {
+		signalCount++
+	}
+	if reasoningScore >= 0.3 {
+		signalCount++
+	}
+	if lastSignals.simpleCount > 0 && (wordCount >= 30 || signalCount >= 2) {
+		effectiveSimpleWeight = 0.01
+	}
+
+	codeContribution := codeScore * codeWeight
+	reasoningContribution := reasoningScore * reasoningWeight
+	technicalContribution := technicalScore * technicalWeight
+	simplePenalty := -(simpleScore * effectiveSimpleWeight)
+	tokenContribution := tokenScore * tokenCountWeight
+
+	// Weighted sum for last message (output complexity applied separately as a score floor).
+	lastMsgScore := codeContribution +
+		reasoningContribution +
+		technicalContribution +
+		simplePenalty +
+		tokenContribution
+	lastMsgScore = clamp(lastMsgScore, 0.0, 1.0)
+
+	// Conversation context blending (prior user turns only).
+	var blended float64
+	var convScore float64
+	if len(input.PriorUserTexts) > 0 {
+		convScore = a.scoreConversationContext(input.PriorUserTexts)
+		lastWeight := defaultLastMessageBlendWeight
+		contextWeight := defaultConversationBlendWeight
+		if isReferentialFollowup(lastSignals, lastMsgScore, convScore, wordCount) {
+			lastWeight = referentialLastMessageBlendWeight
+			contextWeight = referentialConversationBlendWeight
+		}
+
+		weightedBlend := (lastMsgScore * lastWeight) + (convScore * contextWeight)
+		blended = math.Max(lastMsgScore, weightedBlend)
+	} else {
+		blended = lastMsgScore
+	}
+
+	// Output complexity as a score floor: strong output signals set a minimum score.
+	outputFloorMinScore := 0.0
+	if outputScore > 0.5 {
+		outputFloorMinScore = outputScore * 0.5
+		if blended < outputFloorMinScore {
+			blended = outputFloorMinScore
+		}
+	}
+
+	finalScore := clamp(blended, 0.0, 1.0)
+
+	// Tier classification with reasoning override.
+	strongCount := lastSignals.strongReasoningCount
+	tier := a.classifyTier(finalScore)
+	if strongCount >= 2 {
+		tier = TierReasoning
+	} else if strongCount >= 1 && (userCodeScore > 0.5 || userTechnicalScore > 0.5) {
+		tier = TierReasoning
+	}
+
+	return &ComplexityResult{
+		Score:     finalScore,
+		Tier:      tier,
+		WordCount: wordCount,
+	}
+}
+
+func (a *ComplexityAnalyzer) scoreConversationContext(priorUserTexts []string) float64 {
+	if len(priorUserTexts) == 0 {
+		return 0.0
+	}
+
+	texts := priorUserTexts
+	if len(texts) > 10 {
+		texts = texts[len(texts)-10:]
+	}
+
+	var weightedTotal float64
+	var totalWeight float64
+	lastIdx := len(texts) - 1
+	for idx, text := range texts {
+		signals := a.matcher.analyzeText(text, contextTextScanMask)
+		code := scoreCount(signals.codeCount, 3)
+		tech := scoreCount(signals.technicalCount, 3)
+		reasoning := scoreCount(signals.reasoningCount, 2)
+		msgScore := (code*codeWeight + tech*technicalWeight + reasoning*reasoningWeight) /
+			(codeWeight + technicalWeight + reasoningWeight)
+		weight := 1.0
+		if lastIdx > 0 {
+			weight = 1.0 + (2.0 * float64(idx) / float64(lastIdx))
+		}
+		weightedTotal += msgScore * weight
+		totalWeight += weight
+	}
+
+	if totalWeight == 0 {
+		return 0.0
+	}
+
+	return math.Min(1.0, weightedTotal/totalWeight)
+}
+
+func isReferentialFollowup(signals textSignalCounts, lastMsgScore, convScore float64, wordCount int) bool {
+	if wordCount == 0 || wordCount > referentialMaxWordCount {
+		return false
+	}
+	if lastMsgScore >= referentialMaxStandaloneScore || convScore < referentialMinContextScore {
+		return false
+	}
+	if signals.taskShiftCount > 0 {
+		return false
+	}
+	if signals.referentialPhraseCount > 0 {
+		return true
+	}
+
+	hasReference := signals.referentialReferenceCount > 0
+	hasAction := signals.referentialActionCount > 0
+	return hasReference && hasAction
+}
+
+func (a *ComplexityAnalyzer) classifyTier(score float64) string {
+	switch {
+	case score < simpleMediumBoundary:
+		return TierSimple
+	case score < mediumComplexBoundary:
+		return TierMedium
+	case score < complexReasoningBoundary:
+		return TierComplex
+	default:
+		return TierReasoning
+	}
+}