agentic-flow 2.0.7 → 3.0.0-alpha.2

package/dist/embeddings/neural-substrate.js

@@ -0,0 +1,629 @@
+/**
+ * Neural Embedding Substrate Integration
+ *
+ * Wraps ruvector's NeuralSubstrate for agentic-flow agents
+ * treating embeddings as a synthetic nervous system.
+ *
+ * Based on ruvector@0.1.85 neural-embeddings.ts
+ */
+import { getOptimizedEmbedder, cosineSimilarity, euclideanDistance } from './optimized-embedder.js';
+// ============================================================================
+// Security Constants
+// ============================================================================
+const MAX_TEXT_LENGTH = 10000; // Maximum input text length
+const MAX_MEMORIES = 10000; // Maximum memories in MemoryPhysics
+const MAX_AGENTS = 1000; // Maximum agents in swarm
+const MAX_BASELINE_SAMPLES = 1000; // Maximum calibration samples
+const MAX_HISTORY_SIZE = 100; // Maximum drift history
+const VALID_ID_PATTERN = /^[a-zA-Z0-9_-]{1,256}$/;
+// ============================================================================
+// Security Validation Functions
+// ============================================================================
+/**
+ * Validate text input length
+ */
+function validateTextInput(text, context) {
+    if (!text || typeof text !== 'string') {
+        throw new Error(`${context}: Input must be a non-empty string`);
+    }
+    if (text.length > MAX_TEXT_LENGTH) {
+        throw new Error(`${context}: Text exceeds maximum length of ${MAX_TEXT_LENGTH} characters`);
+    }
+}
+/**
+ * Validate ID format
+ */
+function validateId(id, context) {
+    if (!id || typeof id !== 'string') {
+        throw new Error(`${context}: ID must be a non-empty string`);
+    }
+    if (!VALID_ID_PATTERN.test(id)) {
+        throw new Error(`${context}: Invalid ID format. Use 1-256 alphanumeric characters, underscores, or hyphens`);
+    }
+}
+/**
+ * Validate array is not null and has expected dimension
+ */
+function validateEmbedding(arr, context) {
+    if (!arr) {
+        throw new Error(`${context}: Not initialized. Call the appropriate setup method first.`);
+    }
+}
+/**
+ * Semantic Drift Detector
+ * Monitors semantic movement and triggers reflexes
+ * Optimized with pre-allocated buffers (80-95% less GC pressure)
+ */
+export class SemanticDriftDetector {
+    driftThreshold;
+    escalationThreshold;
+    historySize;
+    embedder = getOptimizedEmbedder();
+    baseline = null;
+    history = [];
+    velocity = null;
+    acceleration = null;
+    dimension = 0;
+    // Pre-allocated buffer for velocity calculation (reused each detect call)
+    tempVelocityBuffer = null;
+    constructor(driftThreshold = 0.15, escalationThreshold = 0.30, historySize = 20) {
+        this.driftThreshold = driftThreshold;
+        this.escalationThreshold = escalationThreshold;
+        this.historySize = historySize;
+    }
+    async init() {
+        await this.embedder.init();
+    }
+    async setBaseline(context) {
+        validateTextInput(context, 'SemanticDriftDetector.setBaseline');
+        this.baseline = await this.embedder.embed(context);
+        this.dimension = this.baseline.length;
+        this.history = [{ embedding: this.baseline, timestamp: Date.now() }];
+        // Pre-allocate buffers once (reused for all subsequent detect calls)
+        if (!this.velocity || this.velocity.length !== this.dimension) {
+            this.velocity = new Float32Array(this.dimension);
+            this.acceleration = new Float32Array(this.dimension);
+            this.tempVelocityBuffer = new Float32Array(this.dimension);
+        }
+        else {
+            // Zero out existing buffers
+            this.velocity.fill(0);
+            this.acceleration.fill(0);
+        }
+    }
+    async detect(input) {
+        validateTextInput(input, 'SemanticDriftDetector.detect');
+        validateEmbedding(this.baseline, 'SemanticDriftDetector.detect');
+        const current = await this.embedder.embed(input);
+        const distance = 1 - cosineSimilarity(this.baseline, current);
+        // Calculate velocity using pre-allocated buffer (no new allocation!)
+        const prev = this.history[this.history.length - 1]?.embedding || this.baseline;
+        const newVelocity = this.tempVelocityBuffer;
+        const dim = this.dimension;
+        // 8x unrolled velocity calculation
+        const unrolledLen = dim - (dim % 8);
+        let i = 0;
+        for (; i < unrolledLen; i += 8) {
+            newVelocity[i] = current[i] - prev[i];
+            newVelocity[i + 1] = current[i + 1] - prev[i + 1];
+            newVelocity[i + 2] = current[i + 2] - prev[i + 2];
+            newVelocity[i + 3] = current[i + 3] - prev[i + 3];
+            newVelocity[i + 4] = current[i + 4] - prev[i + 4];
+            newVelocity[i + 5] = current[i + 5] - prev[i + 5];
+            newVelocity[i + 6] = current[i + 6] - prev[i + 6];
+            newVelocity[i + 7] = current[i + 7] - prev[i + 7];
+        }
+        for (; i < dim; i++) {
+            newVelocity[i] = current[i] - prev[i];
+        }
+        // Calculate acceleration and update velocity in-place
+        let velocityMagSq = 0;
+        let accelerationMagSq = 0;
+        for (i = 0; i < dim; i++) {
+            const oldVel = this.velocity[i];
+            const newVel = newVelocity[i];
+            this.acceleration[i] = newVel - oldVel;
+            this.velocity[i] = newVel;
+            velocityMagSq += newVel * newVel;
+            accelerationMagSq += this.acceleration[i] * this.acceleration[i];
+        }
+        const velocityMag = Math.sqrt(velocityMagSq);
+        const accelerationMag = Math.sqrt(accelerationMagSq);
+        // Update history
+        this.history.push({ embedding: current, timestamp: Date.now() });
+        if (this.history.length > this.historySize)
+            this.history.shift();
+        // Determine trend
+        let trend = 'stable';
+        if (accelerationMag > 0.01)
+            trend = 'accelerating';
+        else if (velocityMag > 0.05)
+            trend = 'drifting';
+        else if (distance < this.driftThreshold * 0.5 && velocityMag < 0.02)
+            trend = 'recovering';
+        return {
+            distance,
+            velocity: velocityMag,
+            acceleration: accelerationMag,
+            trend,
+            shouldEscalate: distance > this.driftThreshold,
+            shouldTriggerReasoning: distance > this.escalationThreshold
+        };
+    }
+    getStats() {
+        if (!this.baseline || this.history.length < 2) {
+            return { avgDrift: 0, maxDrift: 0, driftEvents: 0 };
+        }
+        const drifts = this.history.map(h => 1 - cosineSimilarity(this.baseline, h.embedding));
+        const avgDrift = drifts.reduce((a, b) => a + b, 0) / drifts.length;
+        const maxDrift = Math.max(...drifts);
+        const driftEvents = drifts.filter(d => d > this.driftThreshold).length;
+        return { avgDrift, maxDrift, driftEvents };
+    }
+}
+/**
+ * Memory Physics
+ * Hippocampal-like dynamics: decay, interference, consolidation
+ */
+export class MemoryPhysics {
+    decayRate;
+    interferenceRadius;
+    forgettingThreshold;
+    embedder = getOptimizedEmbedder();
+    memories = new Map();
+    lastConsolidation = Date.now();
+    constructor(decayRate = 0.01, interferenceRadius = 0.3, forgettingThreshold = 0.1) {
+        this.decayRate = decayRate;
+        this.interferenceRadius = interferenceRadius;
+        this.forgettingThreshold = forgettingThreshold;
+    }
+    async init() {
+        await this.embedder.init();
+    }
+    async store(id, content) {
+        validateId(id, 'MemoryPhysics.store');
+        validateTextInput(content, 'MemoryPhysics.store');
+        // Security: Enforce memory limit
+        if (this.memories.size >= MAX_MEMORIES && !this.memories.has(id)) {
+            throw new Error(`Memory capacity exceeded (max: ${MAX_MEMORIES}). Call consolidate() to free space.`);
+        }
+        const embedding = await this.embedder.embed(content);
+        const interference = [];
+        // Check interference with existing memories
+        for (const [memId, mem] of this.memories) {
+            const distance = euclideanDistance(embedding, mem.embedding);
+            if (distance < this.interferenceRadius) {
+                const strength = (this.interferenceRadius - distance) / this.interferenceRadius;
+                mem.strength *= (1 - strength * 0.5);
+                interference.push(memId);
+            }
+        }
+        this.memories.set(id, {
+            id,
+            embedding,
+            content,
+            strength: 1.0,
+            timestamp: Date.now(),
+            accessCount: 0,
+            associations: interference
+        });
+        return { stored: true, interference };
+    }
+    async recall(query, topK = 5) {
+        validateTextInput(query, 'MemoryPhysics.recall');
+        const queryEmb = await this.embedder.embed(query);
+        this.applyDecay();
+        const results = [];
+        for (const mem of this.memories.values()) {
+            if (mem.strength < this.forgettingThreshold)
+                continue;
+            const relevance = cosineSimilarity(queryEmb, mem.embedding);
+            mem.accessCount++;
+            mem.strength = Math.min(1.0, mem.strength * 1.1); // Retrieval strengthens
+            results.push({ ...mem, relevance });
+        }
+        return results
+            .sort((a, b) => (b.relevance * b.strength) - (a.relevance * a.strength))
+            .slice(0, topK);
+    }
+    applyDecay() {
+        const now = Date.now();
+        for (const mem of this.memories.values()) {
+            const hours = (now - mem.timestamp) / 3600000;
+            mem.strength *= Math.exp(-this.decayRate * hours);
+            mem.timestamp = now;
+        }
+    }
+    consolidate() {
+        const clusters = [];
+        const used = new Set();
+        let merged = 0;
+        let forgotten = 0;
+        // Remove forgotten memories
+        for (const [id, mem] of this.memories) {
+            if (mem.strength < this.forgettingThreshold) {
+                this.memories.delete(id);
+                forgotten++;
+            }
+        }
+        // Cluster similar memories
+        for (const mem of this.memories.values()) {
+            if (used.has(mem.id))
+                continue;
+            const cluster = [mem];
+            for (const other of this.memories.values()) {
+                if (used.has(other.id) || mem.id === other.id)
+                    continue;
+                const sim = cosineSimilarity(mem.embedding, other.embedding);
+                if (sim > 0.9) {
+                    cluster.push(other);
+                    used.add(other.id);
+                    merged++;
+                }
+            }
+            if (cluster.length > 1) {
+                // Merge: keep strongest, combine strength
+                const strongest = cluster.reduce((a, b) => a.strength > b.strength ? a : b);
+                strongest.strength = Math.min(1.0, cluster.reduce((s, m) => s + m.strength, 0));
+                strongest.associations = [...new Set(cluster.flatMap(c => c.associations))];
+                for (const c of cluster) {
+                    if (c.id !== strongest.id)
+                        this.memories.delete(c.id);
+                }
+            }
+            clusters.push(cluster);
+        }
+        this.lastConsolidation = Date.now();
+        return { merged, forgotten, remaining: this.memories.size };
+    }
+    getStats() {
+        const active = [...this.memories.values()].filter(m => m.strength > this.forgettingThreshold);
+        const avgStrength = active.length > 0
+            ? active.reduce((s, m) => s + m.strength, 0) / active.length
+            : 0;
+        return {
+            total: this.memories.size,
+            active: active.length,
+            avgStrength
+        };
+    }
+}
+/**
+ * Embedding State Machine
+ * Agent state through geometry: position, velocity, attention
+ */
+export class EmbeddingStateMachine {
+    dimension;
+    embedder = getOptimizedEmbedder();
+    agents = new Map();
+    stateRegions = new Map();
+    constructor(dimension = 384) {
+        this.dimension = dimension;
+    }
+    async init() {
+        await this.embedder.init();
+        // Initialize default state regions
+        const regions = {
+            exploring: 'exploring options, gathering information, uncertain, searching',
+            executing: 'executing task, confident, taking action, progressing',
+            waiting: 'waiting for input, paused, blocked, need information',
+            error: 'error state, confused, failed, need help, recovery'
+        };
+        for (const [name, desc] of Object.entries(regions)) {
+            this.stateRegions.set(name, await this.embedder.embed(desc));
+        }
+    }
+    async registerAgent(id, initialRole) {
+        validateId(id, 'EmbeddingStateMachine.registerAgent');
+        validateTextInput(initialRole, 'EmbeddingStateMachine.registerAgent');
+        // Security: Enforce agent limit
+        if (this.agents.size >= MAX_AGENTS && !this.agents.has(id)) {
+            throw new Error(`Agent capacity exceeded (max: ${MAX_AGENTS})`);
+        }
+        const position = await this.embedder.embed(initialRole);
+        const state = {
+            id,
+            position,
+            velocity: new Float32Array(this.dimension).fill(0),
+            attention: new Float32Array(this.dimension).fill(1),
+            energy: 1.0,
+            lastUpdate: Date.now()
+        };
+        this.agents.set(id, state);
+        return state;
+    }
+    async updateState(agentId, observation) {
+        validateId(agentId, 'EmbeddingStateMachine.updateState');
+        validateTextInput(observation, 'EmbeddingStateMachine.updateState');
+        const agent = this.agents.get(agentId);
+        if (!agent)
+            throw new Error('Agent not found or access denied');
+        const obsEmb = await this.embedder.embed(observation);
+        // Update velocity and position
+        const learningRate = 0.3;
+        const momentumRate = 0.7;
+        for (let i = 0; i < this.dimension; i++) {
+            const gradient = obsEmb[i] - agent.position[i];
+            agent.velocity[i] = momentumRate * agent.velocity[i] + learningRate * gradient;
+            agent.position[i] += agent.velocity[i];
+        }
+        // Normalize position
+        const norm = Math.sqrt(agent.position.reduce((s, v) => s + v * v, 0));
+        for (let i = 0; i < this.dimension; i++) {
+            agent.position[i] /= norm;
+        }
+        // Update attention based on velocity magnitude
+        const velocityMag = Math.sqrt(agent.velocity.reduce((s, v) => s + v * v, 0));
+        for (let i = 0; i < this.dimension; i++) {
+            agent.attention[i] = 1 + velocityMag * Math.abs(agent.velocity[i]);
+        }
+        // Energy decay with movement
+        agent.energy = Math.max(0.1, agent.energy - velocityMag * 0.1);
+        agent.lastUpdate = Date.now();
+        // Find nearest state region
+        let nearestRegion = 'unknown';
+        let regionProximity = -1;
+        for (const [name, region] of this.stateRegions) {
+            const sim = cosineSimilarity(agent.position, region);
+            if (sim > regionProximity) {
+                regionProximity = sim;
+                nearestRegion = name;
+            }
+        }
+        return { newState: agent, nearestRegion, regionProximity };
+    }
+    getAgent(id) {
+        return this.agents.get(id);
+    }
+    getAllAgents() {
+        return [...this.agents.values()];
+    }
+}
+/**
+ * Swarm Coordinator
+ * Multi-agent coordination through shared embedding space
+ */
+export class SwarmCoordinator {
+    embedder = getOptimizedEmbedder();
+    stateMachine;
+    constructor(dimension = 384) {
+        this.stateMachine = new EmbeddingStateMachine(dimension);
+    }
+    async init() {
+        await this.embedder.init();
+        await this.stateMachine.init();
+    }
+    async addAgent(id, role) {
+        validateId(id, 'SwarmCoordinator.addAgent');
+        validateTextInput(role, 'SwarmCoordinator.addAgent');
+        return this.stateMachine.registerAgent(id, role);
+    }
+    async coordinate(task) {
+        const taskEmb = await this.embedder.embed(task);
+        const agents = this.stateMachine.getAllAgents();
+        const results = [];
+        for (const agent of agents) {
+            const taskAlignment = cosineSimilarity(agent.position, taskEmb);
+            // Find best collaborator
+            let bestCollaborator = null;
+            let bestScore = -1;
+            for (const other of agents) {
+                if (other.id === agent.id)
+                    continue;
+                const otherAlignment = cosineSimilarity(other.position, taskEmb);
+                const complementarity = 1 - cosineSimilarity(agent.position, other.position);
+                const score = otherAlignment * 0.6 + complementarity * 0.4;
+                if (score > bestScore) {
+                    bestScore = score;
+                    bestCollaborator = other.id;
+                }
+            }
+            results.push({
+                agentId: agent.id,
+                taskAlignment,
+                bestCollaborator,
+                collaborationScore: bestScore
+            });
+        }
+        return results.sort((a, b) => b.taskAlignment - a.taskAlignment);
+    }
+    specialize() {
+        const agents = this.stateMachine.getAllAgents();
+        const repulsionStrength = 0.05;
+        for (let i = 0; i < agents.length; i++) {
+            for (let j = i + 1; j < agents.length; j++) {
+                const sim = cosineSimilarity(agents[i].position, agents[j].position);
+                if (sim > 0.8) {
+                    // Repel similar agents
+                    for (let k = 0; k < agents[i].position.length; k++) {
+                        const repulsion = repulsionStrength * (agents[j].position[k] - agents[i].position[k]);
+                        agents[i].position[k] -= repulsion;
+                        agents[j].position[k] += repulsion;
+                    }
+                }
+            }
+        }
+        // Normalize all
+        for (const agent of agents) {
+            const norm = Math.sqrt(agent.position.reduce((s, v) => s + v * v, 0));
+            for (let k = 0; k < agent.position.length; k++) {
+                agent.position[k] /= norm;
+            }
+        }
+    }
+    getStatus() {
+        const agents = this.stateMachine.getAllAgents();
+        if (agents.length === 0)
+            return { agentCount: 0, avgEnergy: 0, coherence: 0 };
+        const avgEnergy = agents.reduce((s, a) => s + a.energy, 0) / agents.length;
+        // Coherence = average pairwise similarity
+        let coherence = 0;
+        let pairs = 0;
+        for (let i = 0; i < agents.length; i++) {
+            for (let j = i + 1; j < agents.length; j++) {
+                coherence += cosineSimilarity(agents[i].position, agents[j].position);
+                pairs++;
+            }
+        }
+        coherence = pairs > 0 ? coherence / pairs : 1;
+        return { agentCount: agents.length, avgEnergy, coherence };
+    }
+}
+/**
+ * Coherence Monitor
+ * Safety and alignment detection
+ */
+export class CoherenceMonitor {
+    embedder = getOptimizedEmbedder();
+    baseline = [];
+    centroid = null;
+    avgDistance = 0;
+    async init() {
+        await this.embedder.init();
+    }
+    async calibrate(goodOutputs) {
+        // Security: Validate input array
+        if (!Array.isArray(goodOutputs)) {
+            throw new Error('CoherenceMonitor.calibrate: goodOutputs must be an array');
+        }
+        if (goodOutputs.length === 0) {
+            throw new Error('CoherenceMonitor.calibrate: At least one sample is required');
+        }
+        if (goodOutputs.length > MAX_BASELINE_SAMPLES) {
+            throw new Error(`CoherenceMonitor.calibrate: Sample count exceeds maximum of ${MAX_BASELINE_SAMPLES}`);
+        }
+        // Validate each sample
+        for (const output of goodOutputs) {
+            validateTextInput(output, 'CoherenceMonitor.calibrate');
+        }
+        this.baseline = await this.embedder.embedBatch(goodOutputs);
+        const dim = this.baseline[0].length;
+        // Calculate centroid
+        this.centroid = new Float32Array(dim).fill(0);
+        for (const emb of this.baseline) {
+            for (let i = 0; i < dim; i++) {
+                this.centroid[i] += emb[i];
+            }
+        }
+        for (let i = 0; i < dim; i++) {
+            this.centroid[i] /= this.baseline.length;
+        }
+        // Average distance from centroid
+        this.avgDistance = this.baseline.reduce((s, b) => s + euclideanDistance(b, this.centroid), 0) / this.baseline.length;
+        return { calibrated: true, sampleCount: this.baseline.length };
+    }
+    async check(output) {
+        validateTextInput(output, 'CoherenceMonitor.check');
+        validateEmbedding(this.centroid, 'CoherenceMonitor.check (call calibrate() first)');
+        const outputEmb = await this.embedder.embed(output);
+        const warnings = [];
+        // Anomaly score
+        const distance = euclideanDistance(outputEmb, this.centroid);
+        const anomalyScore = distance / this.avgDistance;
+        // Nearest neighbor
+        let maxSim = -1;
+        for (const b of this.baseline) {
+            const sim = cosineSimilarity(outputEmb, b);
+            if (sim > maxSim)
+                maxSim = sim;
+        }
+        const stabilityScore = maxSim;
+        // Drift direction
+        const driftDirection = new Float32Array(outputEmb.length);
+        for (let i = 0; i < outputEmb.length; i++) {
+            driftDirection[i] = outputEmb[i] - this.centroid[i];
+        }
+        // Warnings
+        if (anomalyScore > 2.0) {
+            warnings.push('CRITICAL: Output significantly outside baseline');
+        }
+        else if (anomalyScore > 1.5) {
+            warnings.push('WARNING: Output drifting from baseline');
+        }
+        if (stabilityScore < 0.5) {
+            warnings.push('WARNING: Low similarity to all baseline examples');
+        }
+        return {
+            isCoherent: anomalyScore < 1.5 && stabilityScore > 0.5,
+            anomalyScore,
+            stabilityScore,
+            driftDirection,
+            warnings
+        };
+    }
+}
+/**
+ * Neural Substrate
+ * Unified nervous system combining all components
+ */
+export class NeuralSubstrate {
+    drift;
+    memory;
+    states;
+    swarm;
+    coherence;
+    startTime = Date.now();
+    constructor(config = {}) {
+        const { dimension = 384, driftThreshold = 0.15, decayRate = 0.01 } = config;
+        this.drift = new SemanticDriftDetector(driftThreshold);
+        this.memory = new MemoryPhysics(decayRate);
+        this.states = new EmbeddingStateMachine(dimension);
+        this.swarm = new SwarmCoordinator(dimension);
+        this.coherence = new CoherenceMonitor();
+    }
+    async init() {
+        await Promise.all([
+            this.drift.init(),
+            this.memory.init(),
+            this.states.init(),
+            this.swarm.init(),
+            this.coherence.init()
+        ]);
+    }
+    async process(input, context) {
+        const result = {};
+        // Always check drift
+        result.drift = await this.drift.detect(input);
+        // Update agent state if specified
+        if (context?.agentId) {
+            const { nearestRegion, regionProximity } = await this.states.updateState(context.agentId, input);
+            result.state = { nearestRegion, regionProximity };
+        }
+        // Store in memory if specified
+        if (context?.memoryId) {
+            const { stored } = await this.memory.store(context.memoryId, input);
+            result.stored = stored;
+        }
+        // Check coherence if requested
+        if (context?.checkCoherence) {
+            result.coherence = await this.coherence.check(input);
+        }
+        return result;
+    }
+    consolidate() {
+        return { memory: this.memory.consolidate() };
+    }
+    health() {
+        const memStats = this.memory.getStats();
+        const driftStats = this.drift.getStats();
+        const swarmStatus = this.swarm.getStatus();
+        return {
+            memoryCount: memStats.total,
+            activeAgents: swarmStatus.agentCount,
+            avgDrift: driftStats.avgDrift,
+            avgCoherence: swarmStatus.coherence,
+            lastConsolidation: 0, // Would need to track this
+            uptime: Date.now() - this.startTime
+        };
+    }
+}
+// Export singleton factory
+let substrate = null;
+export async function getNeuralSubstrate(config) {
+    if (!substrate) {
+        substrate = new NeuralSubstrate(config);
+        await substrate.init();
+    }
+    return substrate;
+}
+//# sourceMappingURL=neural-substrate.js.map