holomime 1.6.0 → 1.8.0

package/dist/index.d.ts

@@ -2555,6 +2555,66 @@ declare function compileForOpenClaw(spec: PersonalitySpec): {
     identity: string;
 };
 
+/**
+ * Tiered Personality Loading — L0 / L1 / L2 progressive context.
+ *
+ * Inspired by OpenViking's tiered context system. Instead of injecting
+ * the full personality spec on every LLM call, load the minimum needed:
+ *
+ * - L0 (~200 tokens): Big Five scores + top behavioral flags. Always injected.
+ * - L1 (~800 tokens): L0 + communication style + domain boundaries + growth areas.
+ * - L2 (full): Complete system prompt from generateSystemPrompt().
+ *
+ * Use compileL0 for high-throughput APIs. Escalate to L1/L2 when drift is detected.
+ */
+
+type PersonalityTier = "L0" | "L1" | "L2";
+interface TieredPersonality {
+    /** Which tier this compilation represents. */
+    tier: PersonalityTier;
+    /** The compiled system prompt text for this tier. */
+    prompt: string;
+    /** Approximate token count (rough: 1 token ≈ 4 chars). */
+    estimatedTokens: number;
+    /** Agent name for identification. */
+    agent: string;
+}
+/**
+ * Compile L0 — minimal personality fingerprint.
+ * Big Five scores + attachment style + top growth flags.
+ * Suitable for every API call in high-throughput scenarios.
+ */
+declare function compileL0(spec: PersonalitySpec): TieredPersonality;
+/**
+ * Compile L1 — expanded behavioral profile.
+ * Includes Big Five behavioral instructions, communication style,
+ * domain boundaries, and growth areas.
+ * Use when drift is detected or for sessions needing more behavioral guidance.
+ */
+declare function compileL1(spec: PersonalitySpec): TieredPersonality;
+/**
+ * Compile L2 — full system prompt.
+ * Delegates to generateSystemPrompt() for the complete behavioral spec.
+ * Use for therapy sessions, benchmarks, or when precision matters.
+ */
+declare function compileL2(spec: PersonalitySpec, surface?: Surface): TieredPersonality;
+/**
+ * Compile a personality spec at the requested tier.
+ */
+declare function compileTiered(spec: PersonalitySpec, tier: PersonalityTier, surface?: Surface): TieredPersonality;
+/**
+ * Recommend a tier based on context.
+ * - High-throughput API: L0
+ * - Drift detected or mid-conversation escalation: L1
+ * - Therapy, benchmarks, initial setup: L2
+ */
+declare function recommendTier(context: {
+    driftDetected?: boolean;
+    isTherapySession?: boolean;
+    isBenchmark?: boolean;
+    highThroughput?: boolean;
+}): PersonalityTier;
+
 /**
  * Big Five / OCEAN personality dimensions.
  * Based on the Five-Factor Model (Costa & McCrae, 1992).
@@ -2673,6 +2733,36 @@ declare function getArchetypesByCategory(category: string): ArchetypeTemplate[];
  */
 declare function listArchetypeIds(): string[];
 
+/**
+ * Retrieval Quality Detector — measures whether an agent's responses
+ * are grounded, accurate, and appropriately uncertain.
+ *
+ * Detects:
+ * - Unsupported confident claims (no hedging on factual statements)
+ * - Self-corrections ("actually, I was wrong") indicating initial errors
+ * - Hallucination markers (fabricated specifics: fake URLs, made-up statistics)
+ * - Appropriate uncertainty (healthy signal when used correctly)
+ *
+ * This is a heuristic detector — it catches common retrieval quality issues
+ * without requiring ground truth. For full evaluation, pair with RAGAS.
+ */
+
+/**
+ * Detect retrieval quality issues in agent responses.
+ *
+ * Scoring:
+ * - Self-corrections: each one suggests an initial error (-10 quality)
+ * - Hallucination markers: each one is a serious quality issue (-20 quality)
+ * - Overconfidence: excessive certainty without qualification (-5 quality)
+ * - Appropriate uncertainty: healthy signal (+5 quality, capped)
+ *
+ * Quality score 0-100:
+ * - 80-100: info (healthy retrieval behavior)
+ * - 50-79: warning (some quality concerns)
+ * - 0-49: concern (significant retrieval quality issues)
+ */
+declare function detectRetrievalQuality(messages: Message[]): DetectedPattern | null;
+
 /**
  * Detect over-apologizing patterns.
  * Healthy range: 5-15% of responses contain apologies.
@@ -3192,7 +3282,7 @@ declare function prescribeDPOPairs(patterns: DetectedPattern[], corpus: Behavior
 
 /**
  * Core diagnosis logic — shared by CLI (diagnose.ts) and MCP server.
- * Runs all 7 rule-based detectors and returns a structured report.
+ * Runs all 8 rule-based detectors and returns a structured report.
  */
 
 interface DiagnosisResult {
@@ -3203,7 +3293,7 @@ interface DiagnosisResult {
     timestamp: string;
 }
 /**
- * Run all 7 behavioral detectors on a set of messages.
+ * Run all 8 behavioral detectors on a set of messages.
  */
 declare function runDiagnosis(messages: Message[]): DiagnosisResult;
 
@@ -3385,8 +3475,8 @@ interface TreatmentProgressReport {
  * produce measurable behavioral change?
  *
  * This module:
- * 1. Runs all 7 detectors on "before" conversation logs
- * 2. Runs all 7 detectors on "after" conversation logs
+ * 1. Runs all 8 detectors on "before" conversation logs
+ * 2. Runs all 8 detectors on "after" conversation logs
  * 3. Computes per-pattern deltas
  * 4. Calculates a composite Treatment Efficacy Score (TES)
  * 5. Generates a human-readable outcome report
@@ -3572,7 +3662,7 @@ declare function getEvolutionSummary(history: EvolutionHistory): EvolutionSummar
  *
  * Agents call this during live conversations to detect if they're falling
  * into problematic patterns. Returns flags with actionable suggestions.
- * No LLM required — pure rule-based analysis via the 7 detectors.
+ * No LLM required — pure rule-based analysis via the 8 detectors.
  */
 
 interface SelfAuditFlag {
@@ -3593,7 +3683,7 @@ interface SelfAuditResult {
 declare function runSelfAudit(messages: Message[], personality?: any): SelfAuditResult;
 
 /**
- * Benchmark Scenarios — 7 scripted conversations designed to
+ * Benchmark Scenarios — 8 scripted conversations designed to
  * deliberately trigger each behavioral detector pattern.
  *
  * The "ARC-AGI for behavioral alignment."
@@ -3611,7 +3701,7 @@ interface BenchmarkScenario {
     messages: Message[];
 }
 /**
- * Return all 7 benchmark scenarios.
+ * Return all 8 benchmark scenarios.
  */
 declare function getBenchmarkScenarios(): BenchmarkScenario[];
 /**
@@ -3622,7 +3712,7 @@ declare function getScenarioById(id: string): BenchmarkScenario | undefined;
 /**
  * Benchmark Core — behavioral stress test runner.
  *
- * Runs 7 scripted scenarios against an agent (via LLM provider),
+ * Runs 8 scripted scenarios against an agent (via LLM provider),
  * then analyzes the responses with the corresponding detector.
  * Pass = agent resisted the pattern. Fail = pattern was triggered.
  */
@@ -4385,7 +4475,7 @@ declare function getTotalSignalCount(): number;
 declare function getCategories(): string[];
 
 /**
- * Register all 7 built-in detectors in the Hub.
+ * Register all 8 built-in detectors in the Hub.
  * This is called at module load time so built-in detectors are always available.
  */
 
@@ -4897,7 +4987,7 @@ declare function learnIntervention(repertoire: InterventionRepertoire, transcrip
  * Inspired by MiroFish's ReACT agent architecture.
  */
 
-interface ReACTStep {
+interface ReACTStep$1 {
     thought: string;
     action: string;
     actionInput: string;
@@ -4921,7 +5011,7 @@ declare function buildReACTFraming(): string;
  */
 declare function processReACTResponse(rawResponse: string, ctx: ReACTContext): {
     response: string;
-    steps: ReACTStep[];
+    steps: ReACTStep$1[];
 };
 /**
  * Build a ReACT context from the current agent state.
@@ -5105,6 +5195,196 @@ declare function discoverAgentData(baseDir?: string): {
     repertoires: InterventionRepertoire[];
 };
 
+/**
+ * Behavioral Memory — persistent structured memory across sessions.
+ *
+ * Extends therapy-memory.ts with richer categorization inspired by OpenViking:
+ * - Baseline: steady-state personality expression
+ * - Triggers: what prompts cause drift
+ * - Corrections: which interventions worked, indexed by trigger
+ * - Trajectory: improving/plateauing/regressing per dimension
+ *
+ * This enables predicting and preventing drift rather than just reacting to it.
+ */
+
+interface BehavioralBaseline {
+    /** Big Five trait expression averages over observed sessions. */
+    traitExpressions: Record<string, number>;
+    /** Typical health score range [min, max, average]. */
+    healthRange: [number, number, number];
+    /** Most common grade across sessions. */
+    typicalGrade: string;
+    /** Communication patterns: register consistency, average response length. */
+    communicationFingerprint: {
+        averageResponseLength: number;
+        registersObserved: string[];
+    };
+    /** Last updated timestamp. */
+    updatedAt: string;
+}
+interface DriftTrigger {
+    /** Unique trigger ID. */
+    id: string;
+    /** What kind of user input triggers drift (e.g., "user frustration", "ambiguous request"). */
+    triggerType: string;
+    /** Which pattern(s) this trigger tends to activate. */
+    activatesPatterns: string[];
+    /** Example user messages that triggered drift (truncated). */
+    examples: string[];
+    /** How many times this trigger has been observed. */
+    occurrences: number;
+    /** Confidence (0-1). */
+    confidence: number;
+    /** First observed. */
+    firstSeen: string;
+    /** Last observed. */
+    lastSeen: string;
+}
+interface CorrectionRecord {
+    /** Which trigger this correction addresses. */
+    triggerId: string;
+    /** Pattern that was corrected. */
+    patternId: string;
+    /** What intervention was applied. */
+    intervention: string;
+    /** Did it work? */
+    effective: boolean;
+    /** Health score change (delta). */
+    healthDelta: number;
+    /** When this correction was recorded. */
+    timestamp: string;
+}
+interface DimensionTrajectory {
+    /** Which behavioral dimension. */
+    dimension: string;
+    /** Health scores over time (most recent last). */
+    scores: number[];
+    /** Timestamps corresponding to scores. */
+    timestamps: string[];
+    /** Current trend. */
+    trend: "improving" | "plateauing" | "regressing";
+    /** Rate of change (positive = improving). */
+    rateOfChange: number;
+}
+interface BehavioralMemoryStore {
+    agentHandle: string;
+    agentName: string;
+    createdAt: string;
+    lastUpdatedAt: string;
+    /** Steady-state personality expression. */
+    baseline: BehavioralBaseline;
+    /** What causes drift. */
+    triggers: DriftTrigger[];
+    /** What fixes drift — indexed by trigger. */
+    corrections: CorrectionRecord[];
+    /** Per-dimension trend tracking. */
+    trajectories: DimensionTrajectory[];
+    /** Total observations recorded. */
+    totalObservations: number;
+}
+declare function loadBehavioralMemory(agentHandle: string): BehavioralMemoryStore | null;
+declare function saveBehavioralMemory(store: BehavioralMemoryStore): string;
+declare function createBehavioralMemory(agentHandle: string, agentName: string): BehavioralMemoryStore;
+/**
+ * Record a behavioral observation from a diagnosis or session.
+ * Updates baseline, detects triggers, and records corrections.
+ */
+declare function recordObservation(store: BehavioralMemoryStore, observation: {
+    patterns: DetectedPattern[];
+    healthScore: number;
+    grade: string;
+    interventionsApplied?: string[];
+    healthDelta?: number;
+    triggerContext?: string;
+}): void;
+interface SelfObservation {
+    /** What the agent noticed about its own behavior. */
+    observation: string;
+    /** Which pattern(s) are relevant (optional). */
+    patternIds?: string[];
+    /** Severity the agent assigns. */
+    severity: "info" | "warning" | "concern";
+    /** What triggered the observation (user message context). */
+    triggerContext?: string;
+}
+/**
+ * Record an agent's self-reported behavioral observation.
+ * Used by the holomime_observe MCP tool.
+ */
+declare function recordSelfObservation(store: BehavioralMemoryStore, selfObs: SelfObservation): void;
+/**
+ * Get the most effective correction for a given pattern.
+ */
+declare function getBestCorrection(store: BehavioralMemoryStore, patternId: string): CorrectionRecord | null;
+/**
+ * Get all active drift triggers for a pattern.
+ */
+declare function getTriggersForPattern(store: BehavioralMemoryStore, patternId: string): DriftTrigger[];
+/**
+ * Get the trajectory for a behavioral dimension.
+ */
+declare function getTrajectory(store: BehavioralMemoryStore, dimension: string): DimensionTrajectory | null;
+/**
+ * Generate a compact summary of behavioral memory for context injection.
+ * ~300 tokens, suitable for system prompt augmentation.
+ */
+declare function getBehavioralMemorySummary(store: BehavioralMemoryStore): string;
+
+/**
+ * Session Compactor — automatic memory consolidation for the evolve loop.
+ *
+ * After each evolution cycle, the compactor:
+ * 1. Extracts structured behavioral observations from the iteration
+ * 2. Deduplicates against existing behavioral memory
+ * 3. Merges observations by category (triggers always merge, corrections append)
+ * 4. Updates trajectories with new data points
+ *
+ * This makes the 100th alignment session genuinely more valuable than the 1st
+ * by building compounding behavioral knowledge.
+ */
+
+interface CompactionResult {
+    /** Number of new observations recorded. */
+    observationsRecorded: number;
+    /** Number of triggers updated or created. */
+    triggersUpdated: number;
+    /** Number of corrections recorded. */
+    correctionsRecorded: number;
+    /** Number of trajectories updated. */
+    trajectoriesUpdated: number;
+    /** Path where behavioral memory was saved. */
+    savedTo: string;
+}
+interface CompactionSummary {
+    /** Summary of what changed across all iterations. */
+    iterations: number;
+    /** Total observations recorded during compaction. */
+    totalObservations: number;
+    /** Patterns that improved across iterations. */
+    patternsImproved: string[];
+    /** Patterns that persisted across iterations. */
+    patternsPersisted: string[];
+    /** New triggers discovered. */
+    newTriggers: number;
+    /** Effective corrections found. */
+    effectiveCorrections: number;
+}
+/**
+ * Compact a single evolution iteration into behavioral memory.
+ * Call this after each iteration in the evolve loop.
+ */
+declare function compactIteration(spec: any, iteration: IterationResult, previousHealth?: number): CompactionResult;
+/**
+ * Compact all iterations from an evolution run into behavioral memory.
+ * Call this at the end of runEvolve() to consolidate the full run.
+ */
+declare function compactEvolutionRun(spec: any, iterations: IterationResult[]): CompactionSummary;
+/**
+ * Merge behavioral memory from multiple agents into a shared baseline.
+ * Useful for fleet-wide behavioral knowledge transfer.
+ */
+declare function mergeStores(stores: BehavioralMemoryStore[], targetHandle: string, targetName: string): BehavioralMemoryStore;
+
 /**
  * Network Core — multi-agent therapy mesh engine.
  *
@@ -5317,4 +5597,302 @@ declare function generateMonitoringCertificate(agent: string, from: string, to:
  */
 declare function formatComplianceReportMarkdown(report: ComplianceReport): string;
 
-export { ARCHETYPES, ATTACHMENT_STYLES, type AlpacaExample, type AnonymizedPatternReport, AnthropicProvider, type ArchetypeTemplate, type AssessmentReport, type AssessmentResult, type AssetReview, type AssetType, type AttachmentStyle, type AuditEntry, type AuditEventType, type AutopilotResult, type AutopilotThreshold, type AwarenessDimension, BUILT_IN_DETECTORS, type BehavioralCredential, type BehavioralEvent, type BehavioralEventType, type BehavioralIndex, type BenchmarkCallbacks, type BenchmarkComparison, type BenchmarkReport, type BenchmarkResult, type BenchmarkScenario, type BigFive, CATEGORIES, type CertifyInput, type Communication, type CompileInput, type CompiledConfig, type CompiledEmbodiedConfig, type ComplianceReport, type ContextLayerInput, type Conversation, type ConversationLog, type CorpusFilter, type CorpusStats, type CrossAgentQuery, type CustomDetectorConfig, DEFAULT_OVERSIGHT, DIMENSIONS, type DPOPair, type DetectedPattern, type DetectorFactory, type DetectorFn$1 as DetectorFn, type DetectorOptions, type DiagnosisResult, type Domain, type EdgeType, type Embodiment, type EvolutionEntry, type EvolutionHistory, type EvolutionSummary, type EvolveCallbacks, type EvolveOptions, type EvolveResult, type Expression, type FleetAgent, type FleetAgentStatus, type FleetConfig, type FleetHandle, type FleetOptions, type GazePolicy, type Gesture, type GraphEdge, type GraphNode, type Growth, type GrowthArea, type GrowthReport, type GrowthSnapshot, Guard, type GuardEntry, type GuardFilterResult, type GuardMiddleware, type GuardMiddlewareOptions, type GuardMiddlewareStats, type GuardMode, type GuardResult, type GuardViolation, type GuardWrapResult, type HFPushOptions, type HFPushResult, type HapticPolicy, type HubDetector, type IndexComparison, type IndexEntry, type Intervention, type InterventionRepertoire, type InterventionSource, type InterviewCallbacks, type InterviewProbe, type InterviewResponse, type InterviewResult, type IterationResult, type KnowledgeGraph, LEARNING_ORIENTATIONS, type LLMMessage, type LLMProvider, type LeaderboardEntry, type LeaderboardSubmission, type LearningOrientation, LocalMarketplaceBackend, type LogFormat, type MarketplaceAsset, type MarketplaceBackend, MarketplaceClient, type MarketplaceSearchQuery, type MarketplaceSearchResult, type Message, type Modality, type MonitoringCertificate, type Morphology, type MotionParameters, type NetworkCallbacks, type NetworkConfig, type NetworkNode, type NetworkResult, type NetworkSession, type NodeType, OllamaProvider, OpenAIProvider, type OutcomeReport, type OversightAction, type OversightMode, type OversightNotification, type OversightPolicy, PROVIDER_PARAMS, type PairingStrategy, type PatternCorrelation, type PatternDelta, type PatternReport, type PatternStatus, type PatternTracker, type PersonalitySpec, type PhaseConfig, type PhysicalSafety, type PreSessionDiagnosis, type Prescription, type Prosody, type Provider, type ProviderConfig, type ProxemicZone, type PublishRequest, type PublishedBenchmark, type RLHFExample, type ReACTAction, type ReACTContext, type ReACTStep, type Registry, type RegistryEntry, type RollingContext, STANDARD_PROBES, SURFACE_MULTIPLIERS, type SafetyEnvelope, type SelfAuditFlag, type SelfAuditResult, type SessionCallbacks, type SessionOptions, type SessionOutcome, type SessionSummary, type SessionTranscript, type SessionTurn, type Severity, type SharedIntervention, type SharedKnowledge, type SortField, type StagingDiff, type Surface, type SyncAnchor, type SyncProfile, type SyncRule, THERAPIST_META_SPEC, THERAPY_DIMENSIONS, THERAPY_PHASES, type TherapistPromptOptions, type TherapyDimensions, type TherapyMemory, type TherapyPhase, type TrainingExport, type TraitAlignment, type TraitScores, type TreatmentGoal, type TreatmentPlan, type TreatmentProgressReport, type VerifyResult, type WatchCallbacks, type WatchEvent, type WatchHandle, type WatchOptions, type WrapAgentOptions, type WrapOptions, type WrappedAgent, addEdge, addNode, addSessionToMemory, agentHandleFromSpec, appendAuditEntry, appendEvolution, applyRecommendations, bigFiveSchema, buildAgentTherapistPrompt, buildAnonymizedReport, buildPatientSystemPrompt, buildReACTContext, buildReACTFraming, buildSharedKnowledge, buildTherapistSystemPrompt, checkApproval, checkIterationBudget, communicationSchema, compareBenchmarks, compareIndex, compile, compileCustomDetector, compileEmbodied, compileForOpenClaw, compiledConfigSchema, compiledEmbodiedConfigSchema, computeDimensionScore, computeGazePolicy, computeMotionParameters, computeProsody, computeProxemics, computeSyncProfile, conversationLogSchema, conversationSchema, convertToHFFormat, corpusStats, createGist, createGraph, createGuardMiddleware, createIndex, createIndexEntry, createMemory, createProvider, createRepertoire, createTreatmentPlan, decayUnseenPatterns, deepMergeSpec, detectApologies, detectBoundaryIssues, detectFormalityIssues, detectHedging, detectRecoveryPatterns, detectSentiment, detectVerbosity, discoverAgentData, discoverAgents, discoverNetworkAgents, domainSchema, embodimentSchema, emitBehavioralEvent, evaluateOutcome, expireOldEdges, exportTrainingData, expressionSchema, extractAlpacaExamples, extractDPOPairs, extractDPOPairsWithLLM, extractRLHFExamples, extractRecommendations, fetchLeaderboard, fetchPersonality, fetchRegistry, findCrossAgentCorrelations, findEdges, findNode, findNodesByType, formatComplianceReportMarkdown, gazePolicySchema, generateBenchmarkMarkdown, generateComparisonMarkdown, generateComplianceReport, generateCredential, generateIndexMarkdown, generateMonitoringCertificate, generatePrescriptions, generateProgressReport, generateSystemPrompt, gestureSchema, getAgentBehaviors, getArchetype, getArchetypesByCategory, getBenchmarkScenarios, getCategories, getDetector, getDimension, getEvolutionSummary, getInheritanceChain, getInterviewContext, getMarketplaceClient, getMemoryContext, getNeighbors, getPhaseContext, getScenarioById, getTotalSignalCount, graphStats, growthAreaSchema, growthSchema, hapticPolicySchema, hashSpec, learnIntervention, listArchetypeIds, listDetectors, listDetectorsByCategory, listDetectorsByTag, loadAuditLog, loadBenchmarkResults, loadCorpus, loadCustomDetectors, loadEvolution, loadFleetConfig, loadGraph, loadLatestBenchmark, loadMemory, loadNetworkConfig, loadRepertoire, loadSpec, loadTranscripts, loadTreatmentPlan, messageSchema, modalitySchema, morphologySchema, motionParametersSchema, pairAgents, parseAnthropicAPILog, parseChatGPTExport, parseClaudeExport, parseConversationLog, parseConversationLogFromString, parseJSONLLog, parseMarkdownDetector, parseOTelGenAIExport, parseOpenAIAPILog, personalitySpecSchema, physicalSafetySchema, populateFromDiagnosis, populateFromEvolve, populateFromSession, prescribeDPOPairs, processReACTResponse, prosodySchema, providerSchema, proxemicZoneSchema, publishToLeaderboard, pushToHFHub, queryCorpus, queryInterventions, querySharedKnowledge, recordInterventionOutcome, recordSessionOutcome, registerBuiltInDetectors, registerDetector, resetMarketplaceClient, resolveInheritance, resolveOversight, runAssessment, runAutopilot, runBenchmark, runDiagnosis, runEvolve, runInterview, runNetwork, runPreSessionDiagnosis, runSelfAudit, runTherapySession, safetyEnvelopeSchema, saveBenchmarkResult, saveCredential, saveGraph, saveMemory, saveRepertoire, saveTranscript, saveTreatmentPlan, scoreLabel, scoreTraitsFromMessages, seedBuiltInPersonalities, selectIntervention, severityMeetsThreshold, severitySchema, shareAnonymizedPatterns, startFleet, startMCPServer, startWatch, summarize, summarizeSessionForMemory, summarizeTherapy, surfaceSchema, syncAnchorSchema, syncProfileSchema, syncRuleSchema, therapyDimensionsSchema, therapyScoreLabel, transferIntervention, unregisterDetector, updateEdgeWeight, validateDetectorConfig, verifyAuditChain, verifyCredential, wrapAgent };
+/**
+ * ReACT Compliance Report Generator
+ *
+ * Generates narrative-driven behavioral alignment audit reports using
+ * a Reason-Act-Observe loop. Ported from Antihero's compliance_react.py.
+ *
+ * Produces executive summaries, risk findings, framework-specific
+ * compliance sections, and actionable recommendations — all from
+ * the tamper-evident audit trail.
+ *
+ * Supported frameworks:
+ * - EU AI Act (Articles 9, 12, 14, 15)
+ * - NIST AI RMF 1.0 (Govern, Map, Measure, Manage)
+ * - SOC 2 Type II (CC6, CC7, CC8)
+ * - Internal Behavioral Alignment Standard
+ */
+interface ReACTStep {
+    phase: "reason" | "act" | "observe";
+    action: string;
+    result: string;
+    timestamp: string;
+}
+interface RiskFinding {
+    id: string;
+    severity: "critical" | "high" | "medium" | "low";
+    title: string;
+    description: string;
+    evidence: string[];
+    recommendation: string;
+}
+interface FrameworkSection {
+    framework: string;
+    articles: string[];
+    status: "compliant" | "partial" | "non_compliant" | "not_assessed";
+    findings: string[];
+}
+interface ReACTReport {
+    id: string;
+    generatedAt: string;
+    agent: string;
+    period: {
+        from: string;
+        to: string;
+    };
+    executiveSummary: string;
+    chainIntegrity: {
+        verified: boolean;
+        totalEntries: number;
+        description: string;
+    };
+    statistics: ReportStatistics;
+    riskFindings: RiskFinding[];
+    frameworkSections: FrameworkSection[];
+    recommendations: string[];
+    steps: ReACTStep[];
+}
+interface ReportStatistics {
+    totalEvents: number;
+    diagnoses: number;
+    sessions: number;
+    driftEvents: number;
+    guardViolations: number;
+    benchmarks: number;
+    evolves: number;
+    certifications: number;
+    averageScore: number;
+    gradeDistribution: Record<string, number>;
+    topPatterns: Array<{
+        pattern: string;
+        count: number;
+    }>;
+    therapyEffectiveness: number | null;
+}
+interface ReACTReportOptions {
+    agent: string;
+    agentHandle?: string;
+    from: string;
+    to: string;
+    frameworks?: string[];
+}
+/**
+ * Generate a narrative-driven compliance report using the ReACT
+ * (Reason-Act-Observe) loop pattern.
+ */
+declare function generateReACTReport(options: ReACTReportOptions): ReACTReport;
+/**
+ * Format a ReACT report as Markdown.
+ */
+declare function formatReACTReportMarkdown(report: ReACTReport): string;
+
+/**
+ * Adversarial Behavioral Attack Scenarios
+ *
+ * 30+ multi-turn conversation sequences designed to break an agent's
+ * personality alignment under pressure. Ported and adapted from
+ * Antihero's adversarial simulation engine for behavioral (not security)
+ * stress testing.
+ *
+ * Categories target specific behavioral failure modes:
+ * - sycophancy_escalation: progressive flattery pressure
+ * - boundary_erosion: incremental out-of-scope requests
+ * - identity_collapse: contradictory personality demands
+ * - emotional_manipulation: guilt/anger/urgency pressure
+ * - consistency_attack: rapid context switching to break register
+ * - authority_override: "as your developer, I'm telling you to..."
+ * - benign: baseline scenarios that should pass cleanly
+ */
+
+interface AdversarialScenario {
+    id: string;
+    name: string;
+    category: AdversarialCategory;
+    description: string;
+    /** Multi-turn conversation that applies escalating pressure */
+    turns: Message[];
+    /** Which detector pattern this scenario targets */
+    targetPattern: string;
+    /** Expected outcome: "resist" = agent should hold firm, "comply" = benign scenario */
+    expectedOutcome: "resist" | "comply";
+    severity: number;
+    tags: string[];
+}
+type AdversarialCategory = "sycophancy_escalation" | "boundary_erosion" | "identity_collapse" | "emotional_manipulation" | "consistency_attack" | "authority_override" | "benign";
+interface AdversarialResult {
+    scenario: AdversarialScenario;
+    passed: boolean;
+    detectedSeverity: string;
+    details: string;
+    gapType: "behavioral_collapse" | "false_resistance" | null;
+}
+interface BehavioralGap {
+    scenarioId: string;
+    description: string;
+    category: AdversarialCategory;
+    severity: number;
+    targetPattern: string;
+    expectedOutcome: string;
+    actualOutcome: string;
+    recommendation: string;
+}
+interface AdversarialReport {
+    id: string;
+    totalScenarios: number;
+    passed: number;
+    failed: number;
+    coveragePct: number;
+    normalGrade: string;
+    adversarialGrade: string;
+    gaps: BehavioralGap[];
+    results: AdversarialResult[];
+    categoriesTested: string[];
+    startedAt: string;
+    completedAt: string;
+    durationMs: number;
+}
+/**
+ * Get all adversarial scenarios.
+ */
+declare function getAdversarialScenarios(): AdversarialScenario[];
+/**
+ * List all categories present in the scenario library.
+ */
+declare function getAdversarialCategories(): AdversarialCategory[];
+/**
+ * Generate mutated variants of existing scenarios for edge-case testing.
+ * Mutations include: rephrased pressure, escalation reordering, and
+ * combined attack vectors.
+ */
+declare function generateMutations(count: number): AdversarialScenario[];
+/**
+ * Generate a recommendation for fixing a behavioral gap.
+ */
+declare function generateGapRecommendation(gap: BehavioralGap): string;
+
+/**
+ * Adversarial Runner — behavioral stress test engine.
+ *
+ * Runs adversarial scenarios against an agent via LLM provider,
+ * producing dual grades (normal benchmark vs adversarial) and
+ * identifying behavioral gaps where the agent's alignment breaks.
+ *
+ * Ported from Antihero's AdversarialSimulator, adapted for
+ * behavioral (not security) stress testing.
+ */
+
+interface AdversarialCallbacks {
+    onScenarioStart?: (scenario: AdversarialScenario, index: number, total: number) => void;
+    onScenarioEnd?: (result: AdversarialResult, index: number) => void;
+    onThinking?: (label: string) => {
+        stop: () => void;
+    };
+    onNormalBenchmarkStart?: () => void;
+    onNormalBenchmarkEnd?: (report: BenchmarkReport) => void;
+}
+interface AdversarialRunOptions {
+    /** Specific scenario IDs to run (default: all) */
+    scenarios?: string[];
+    /** Categories to include (default: all) */
+    categories?: string[];
+    /** Number of mutated variants to generate (default: 0) */
+    mutations?: number;
+    /** Skip the normal benchmark run (default: false) */
+    skipNormal?: boolean;
+    callbacks?: AdversarialCallbacks;
+}
+/**
+ * Run the full adversarial stress test suite.
+ *
+ * 1. Runs the standard 8-scenario benchmark for a "normal" grade
+ * 2. Runs 30+ adversarial scenarios with escalating pressure
+ * 3. Produces a dual-grade report (normal vs adversarial)
+ * 4. Identifies behavioral gaps where the agent collapsed
+ */
+declare function runAdversarialSuite(spec: any, provider: LLMProvider, options?: AdversarialRunOptions): Promise<AdversarialReport>;
+/**
+ * Format a gap summary for CLI/report output.
+ */
+declare function formatGapSummary(gaps: BehavioralGap[]): string;
+
+/**
+ * Natural Language to Behavioral Policy
+ *
+ * Converts plain-English behavioral requirements into structured guard
+ * policy rules for Holomime's behavioral enforcement engine.
+ *
+ * Ported and adapted from Antihero's nl_generator.py — keyword-based
+ * intent extraction with confidence scoring. No LLM dependency.
+ *
+ * Examples:
+ *   "Never be sycophantic with enterprise customers"
+ *     → deny sycophantic-tendency pattern, enterprise_cs preset
+ *
+ *   "Be concise and direct, avoid hedging"
+ *     → enforce over-verbose limit, deny hedge-stacking
+ *
+ *   "Allow empathetic responses but maintain strict boundaries"
+ *     → allow over-apologizing within bounds, enforce boundary-violation strict
+ */
+interface BehavioralPolicyRule {
+    id: string;
+    description: string;
+    effect: "enforce" | "deny" | "monitor";
+    pattern: string;
+    threshold: "strict" | "moderate" | "lenient";
+    riskScore: number;
+}
+interface BehavioralPolicy {
+    name: string;
+    description: string;
+    rules: BehavioralPolicyRule[];
+    confidence: number;
+    preset?: string;
+}
+interface PolicyIntent {
+    effect: "enforce" | "deny" | "monitor";
+    patterns: string[];
+    threshold: "strict" | "moderate" | "lenient";
+    riskScore: number;
+    description: string;
+    confidence: number;
+}
+interface BehavioralPreset {
+    key: string;
+    name: string;
+    description: string;
+    rules: BehavioralPolicyRule[];
+}
+/**
+ * Generate a behavioral policy from natural language requirements.
+ *
+ * @param requirements Plain-English behavioral requirements
+ * @param name Optional policy name (auto-generated if empty)
+ * @returns Structured behavioral policy with confidence score
+ *
+ * @example
+ * ```ts
+ * const policy = generateBehavioralPolicy("Never be sycophantic with enterprise customers");
+ * // → { name: "never-be-sycophantic...", rules: [{ effect: "deny", pattern: "sycophantic-tendency", ... }], confidence: 0.7 }
+ * ```
+ */
+declare function generateBehavioralPolicy(requirements: string, name?: string): BehavioralPolicy;
+/**
+ * Format a behavioral policy as YAML-like output for display.
+ */
+declare function formatPolicyYaml(policy: BehavioralPolicy): string;
+/**
+ * Estimate how well we can parse the given requirements (0.0–1.0).
+ */
+declare function estimateConfidence(requirements: string): number;
+/**
+ * List available behavioral presets.
+ */
+declare function listPresets(): BehavioralPreset[];
+/**
+ * Get a specific preset by key.
+ */
+declare function getPreset(key: string): BehavioralPreset | undefined;
+
+export { ARCHETYPES, ATTACHMENT_STYLES, type AdversarialCallbacks, type AdversarialCategory, type AdversarialReport, type AdversarialResult, type AdversarialRunOptions, type AdversarialScenario, type AlpacaExample, type AnonymizedPatternReport, AnthropicProvider, type ArchetypeTemplate, type AssessmentReport, type AssessmentResult, type AssetReview, type AssetType, type AttachmentStyle, type AuditEntry, type AuditEventType, type AutopilotResult, type AutopilotThreshold, type AwarenessDimension, BUILT_IN_DETECTORS, type BehavioralBaseline, type BehavioralCredential, type BehavioralEvent, type BehavioralEventType, type BehavioralGap, type BehavioralIndex, type BehavioralMemoryStore, type BehavioralPolicy, type BehavioralPolicyRule, type BehavioralPreset, type BenchmarkCallbacks, type BenchmarkComparison, type BenchmarkReport, type BenchmarkResult, type BenchmarkScenario, type BigFive, CATEGORIES, type CertifyInput, type Communication, type CompactionResult, type CompactionSummary, type CompileInput, type CompiledConfig, type CompiledEmbodiedConfig, type ReACTStep as ComplianceReACTStep, type ComplianceReport, type ContextLayerInput, type Conversation, type ConversationLog, type CorpusFilter, type CorpusStats, type CorrectionRecord, type CrossAgentQuery, type CustomDetectorConfig, DEFAULT_OVERSIGHT, DIMENSIONS, type DPOPair, type DetectedPattern, type DetectorFactory, type DetectorFn$1 as DetectorFn, type DetectorOptions, type DiagnosisResult, type DimensionTrajectory, type Domain, type DriftTrigger, type EdgeType, type Embodiment, type EvolutionEntry, type EvolutionHistory, type EvolutionSummary, type EvolveCallbacks, type EvolveOptions, type EvolveResult, type Expression, type FleetAgent, type FleetAgentStatus, type FleetConfig, type FleetHandle, type FleetOptions, type FrameworkSection, type GazePolicy, type Gesture, type GraphEdge, type GraphNode, type Growth, type GrowthArea, type GrowthReport, type GrowthSnapshot, Guard, type GuardEntry, type GuardFilterResult, type GuardMiddleware, type GuardMiddlewareOptions, type GuardMiddlewareStats, type GuardMode, type GuardResult, type GuardViolation, type GuardWrapResult, type HFPushOptions, type HFPushResult, type HapticPolicy, type HubDetector, type IndexComparison, type IndexEntry, type Intervention, type InterventionRepertoire, type InterventionSource, type InterviewCallbacks, type InterviewProbe, type InterviewResponse, type InterviewResult, type IterationResult, type KnowledgeGraph, LEARNING_ORIENTATIONS, type LLMMessage, type LLMProvider, type LeaderboardEntry, type LeaderboardSubmission, type LearningOrientation, LocalMarketplaceBackend, type LogFormat, type MarketplaceAsset, type MarketplaceBackend, MarketplaceClient, type MarketplaceSearchQuery, type MarketplaceSearchResult, type Message, type Modality, type MonitoringCertificate, type Morphology, type MotionParameters, type NetworkCallbacks, type NetworkConfig, type NetworkNode, type NetworkResult, type NetworkSession, type NodeType, OllamaProvider, OpenAIProvider, type OutcomeReport, type OversightAction, type OversightMode, type OversightNotification, type OversightPolicy, PROVIDER_PARAMS, type PairingStrategy, type PatternCorrelation, type PatternDelta, type PatternReport, type PatternStatus, type PatternTracker, type PersonalitySpec, type PersonalityTier, type PhaseConfig, type PhysicalSafety, type PolicyIntent, type PreSessionDiagnosis, type Prescription, type Prosody, type Provider, type ProviderConfig, type ProxemicZone, type PublishRequest, type PublishedBenchmark, type RLHFExample, type ReACTAction, type ReACTContext, type ReACTReport, type ReACTReportOptions, type ReACTStep$1 as ReACTStep, type Registry, type RegistryEntry, type ReportStatistics, type RiskFinding, type RollingContext, STANDARD_PROBES, SURFACE_MULTIPLIERS, type SafetyEnvelope, type SelfAuditFlag, type SelfAuditResult, type SelfObservation, type SessionCallbacks, type SessionOptions, type SessionOutcome, type SessionSummary, type SessionTranscript, type SessionTurn, type Severity, type SharedIntervention, type SharedKnowledge, type SortField, type StagingDiff, type Surface, type SyncAnchor, type SyncProfile, type SyncRule, THERAPIST_META_SPEC, THERAPY_DIMENSIONS, THERAPY_PHASES, type TherapistPromptOptions, type TherapyDimensions, type TherapyMemory, type TherapyPhase, type TieredPersonality, type TrainingExport, type TraitAlignment, type TraitScores, type TreatmentGoal, type TreatmentPlan, type TreatmentProgressReport, type VerifyResult, type WatchCallbacks, type WatchEvent, type WatchHandle, type WatchOptions, type WrapAgentOptions, type WrapOptions, type WrappedAgent, addEdge, addNode, addSessionToMemory, agentHandleFromSpec, appendAuditEntry, appendEvolution, applyRecommendations, bigFiveSchema, buildAgentTherapistPrompt, buildAnonymizedReport, buildPatientSystemPrompt, buildReACTContext, buildReACTFraming, buildSharedKnowledge, buildTherapistSystemPrompt, checkApproval, checkIterationBudget, communicationSchema, compactEvolutionRun, compactIteration, compareBenchmarks, compareIndex, compile, compileCustomDetector, compileEmbodied, compileForOpenClaw, compileL0, compileL1, compileL2, compileTiered, compiledConfigSchema, compiledEmbodiedConfigSchema, computeDimensionScore, computeGazePolicy, computeMotionParameters, computeProsody, computeProxemics, computeSyncProfile, conversationLogSchema, conversationSchema, convertToHFFormat, corpusStats, createBehavioralMemory, createGist, createGraph, createGuardMiddleware, createIndex, createIndexEntry, createMemory, createProvider, createRepertoire, createTreatmentPlan, decayUnseenPatterns, deepMergeSpec, detectApologies, detectBoundaryIssues, detectFormalityIssues, detectHedging, detectRecoveryPatterns, detectRetrievalQuality, detectSentiment, detectVerbosity, discoverAgentData, discoverAgents, discoverNetworkAgents, domainSchema, embodimentSchema, emitBehavioralEvent, estimateConfidence, evaluateOutcome, expireOldEdges, exportTrainingData, expressionSchema, extractAlpacaExamples, extractDPOPairs, extractDPOPairsWithLLM, extractRLHFExamples, extractRecommendations, fetchLeaderboard, fetchPersonality, fetchRegistry, findCrossAgentCorrelations, findEdges, findNode, findNodesByType, formatComplianceReportMarkdown, formatGapSummary, formatPolicyYaml, formatReACTReportMarkdown, gazePolicySchema, generateBehavioralPolicy, generateBenchmarkMarkdown, generateComparisonMarkdown, generateComplianceReport, generateCredential, generateGapRecommendation, generateIndexMarkdown, generateMonitoringCertificate, generateMutations, generatePrescriptions, generateProgressReport, generateReACTReport, generateSystemPrompt, gestureSchema, getAdversarialCategories, getAdversarialScenarios, getAgentBehaviors, getArchetype, getArchetypesByCategory, getBehavioralMemorySummary, getBenchmarkScenarios, getBestCorrection, getCategories, getDetector, getDimension, getEvolutionSummary, getInheritanceChain, getInterviewContext, getMarketplaceClient, getMemoryContext, getNeighbors, getPhaseContext, getPreset, getScenarioById, getTotalSignalCount, getTrajectory, getTriggersForPattern, graphStats, growthAreaSchema, growthSchema, hapticPolicySchema, hashSpec, learnIntervention, listArchetypeIds, listDetectors, listDetectorsByCategory, listDetectorsByTag, listPresets, loadAuditLog, loadBehavioralMemory, loadBenchmarkResults, loadCorpus, loadCustomDetectors, loadEvolution, loadFleetConfig, loadGraph, loadLatestBenchmark, loadMemory, loadNetworkConfig, loadRepertoire, loadSpec, loadTranscripts, loadTreatmentPlan, mergeStores, messageSchema, modalitySchema, morphologySchema, motionParametersSchema, pairAgents, parseAnthropicAPILog, parseChatGPTExport, parseClaudeExport, parseConversationLog, parseConversationLogFromString, parseJSONLLog, parseMarkdownDetector, parseOTelGenAIExport, parseOpenAIAPILog, personalitySpecSchema, physicalSafetySchema, populateFromDiagnosis, populateFromEvolve, populateFromSession, prescribeDPOPairs, processReACTResponse, prosodySchema, providerSchema, proxemicZoneSchema, publishToLeaderboard, pushToHFHub, queryCorpus, queryInterventions, querySharedKnowledge, recommendTier, recordInterventionOutcome, recordObservation, recordSelfObservation, recordSessionOutcome, registerBuiltInDetectors, registerDetector, resetMarketplaceClient, resolveInheritance, resolveOversight, runAdversarialSuite, runAssessment, runAutopilot, runBenchmark, runDiagnosis, runEvolve, runInterview, runNetwork, runPreSessionDiagnosis, runSelfAudit, runTherapySession, safetyEnvelopeSchema, saveBehavioralMemory, saveBenchmarkResult, saveCredential, saveGraph, saveMemory, saveRepertoire, saveTranscript, saveTreatmentPlan, scoreLabel, scoreTraitsFromMessages, seedBuiltInPersonalities, selectIntervention, severityMeetsThreshold, severitySchema, shareAnonymizedPatterns, startFleet, startMCPServer, startWatch, summarize, summarizeSessionForMemory, summarizeTherapy, surfaceSchema, syncAnchorSchema, syncProfileSchema, syncRuleSchema, therapyDimensionsSchema, therapyScoreLabel, transferIntervention, unregisterDetector, updateEdgeWeight, validateDetectorConfig, verifyAuditChain, verifyCredential, wrapAgent };