monomind 2.7.12 → 2.7.14

package/packages/@monomind/cli/dist/src/mcp-tools/embeddings-tools.js

@@ -1,866 +0,0 @@
-/**
- * Embeddings MCP Tools for CLI
- *
- * Tool definitions for ONNX embeddings with hyperbolic support and neural substrate.
- * Implements ADR-024: Embeddings MCP Tools
- */
-import { existsSync, readFileSync, statSync, writeFileSync, renameSync, mkdirSync } from 'fs';
-import { join, resolve } from 'path';
-import { cosineSimilarity } from '../utils/cosine-similarity.js';
-// Configuration paths
-const CONFIG_DIR = '.monomind';
-const EMBEDDINGS_CONFIG = 'embeddings.json';
-const MODELS_DIR = 'models';
-// Input validation caps — embeddings handlers operate on caller-supplied text and
-// vectors. Without these caps an attacker can pass a 100M-entry array and OOM the
-// process, or a 50MB string and saturate the hash-fallback embedding loop.
-const MAX_TEXT_LENGTH = 64 * 1024;
-const MAX_VECTOR_DIM = 8192;
-function validateText(t, field) {
-    if (typeof t !== 'string')
-        throw new Error(`${field}: must be a string`);
-    if (t.length > MAX_TEXT_LENGTH)
-        throw new Error(`${field}: text too long (max ${MAX_TEXT_LENGTH})`);
-    return t;
-}
-function validateVector(v, field) {
-    if (!Array.isArray(v))
-        throw new Error(`${field}: must be a number[]`);
-    if (v.length === 0)
-        throw new Error(`${field}: empty vector`);
-    if (v.length > MAX_VECTOR_DIM)
-        throw new Error(`${field}: vector too large (max ${MAX_VECTOR_DIM})`);
-    for (let i = 0; i < v.length; i++) {
-        if (typeof v[i] !== 'number' || !Number.isFinite(v[i])) {
-            throw new Error(`${field}: contains non-finite value at index ${i}`);
-        }
-    }
-    return v;
-}
-function getConfigPath() {
-    return resolve(join(CONFIG_DIR, EMBEDDINGS_CONFIG));
-}
-function ensureConfigDir() {
-    const dir = resolve(CONFIG_DIR);
-    if (!existsSync(dir)) {
-        mkdirSync(dir, { recursive: true });
-    }
-}
-const MAX_EMBEDDINGS_CONFIG_BYTES = 10 * 1024 * 1024; // 10 MB
-function loadConfig() {
-    try {
-        const path = getConfigPath();
-        if (existsSync(path) && statSync(path).size <= MAX_EMBEDDINGS_CONFIG_BYTES) {
-            return JSON.parse(readFileSync(path, 'utf-8'));
-        }
-    }
-    catch (e) {
-        // Return null on error
-        if (process.env.DEBUG || process.env.MONOMIND_DEBUG)
-            console.error('[embeddings-tools] config read/parse failed:', e);
-    }
-    return null;
-}
-function saveConfig(config) {
-    ensureConfigDir();
-    const dest = getConfigPath();
-    const tmp = dest + '.tmp';
-    writeFileSync(tmp, JSON.stringify(config, null, 2), 'utf-8');
-    renameSync(tmp, dest);
-}
-// Real ONNX embedding generation via memory-initializer
-let realEmbeddingFn = null;
-async function getRealEmbeddingFunction() {
-    if (!realEmbeddingFn) {
-        try {
-            const { generateEmbedding } = await import('../memory/memory-initializer.js');
-            realEmbeddingFn = generateEmbedding;
-        }
-        catch {
-            realEmbeddingFn = null;
-        }
-    }
-    return realEmbeddingFn;
-}
-// Generate real ONNX embedding (falls back to deterministic hash if ONNX unavailable)
-async function generateRealEmbedding(text, dimension) {
-    const realFn = await getRealEmbeddingFunction();
-    if (realFn) {
-        try {
-            const result = await realFn(text);
-            return result.embedding;
-        }
-        catch {
-            // Fall through to fallback
-        }
-    }
-    // Fallback: deterministic hash-based (only if ONNX truly unavailable)
-    console.warn('[MCP] ONNX unavailable, using fallback embedding');
-    const embedding = [];
-    let hash = 0;
-    for (let i = 0; i < text.length; i++) {
-        hash = ((hash << 5) - hash + text.charCodeAt(i)) | 0;
-    }
-    for (let i = 0; i < dimension; i++) {
-        const seed = hash + i * 1337;
-        embedding.push(Math.sin(seed) * Math.cos(seed * 0.5));
-    }
-    // L2 normalize
-    const norm = Math.sqrt(embedding.reduce((sum, x) => sum + x * x, 0));
-    return embedding.map(x => x / norm);
-}
-// Convert Euclidean embedding to Poincaré ball
-function toPoincare(euclidean, curvature) {
-    const c = Math.abs(curvature);
-    const sqrtC = Math.sqrt(c);
-    const norm = Math.sqrt(euclidean.reduce((sum, x) => sum + x * x, 0));
-    // Exponential map at origin
-    const factor = Math.tanh(sqrtC * norm / 2) / (sqrtC * norm + 1e-15);
-    return euclidean.map(x => x * factor);
-}
-// Poincaré distance
-function poincareDistance(a, b, curvature) {
-    const c = Math.abs(curvature);
-    const diffSq = a.reduce((sum, _, i) => sum + (a[i] - b[i]) ** 2, 0);
-    const normASq = a.reduce((sum, x) => sum + x * x, 0);
-    const normBSq = b.reduce((sum, x) => sum + x * x, 0);
-    const denom = (1 - normASq) * (1 - normBSq);
-    const delta = 2 * diffSq / (denom + 1e-15);
-    return (1 / Math.sqrt(c)) * Math.acosh(1 + delta);
-}
-/** Ungated — internal/test use only. MCP clients get `embeddingsTools` below. */
-export const allEmbeddingsTools = [
-    {
-        name: 'embeddings_init',
-        description: 'Initialize the ONNX embedding subsystem with hyperbolic support',
-        category: 'embeddings',
-        inputSchema: {
-            type: 'object',
-            properties: {
-                model: {
-                    type: 'string',
-                    description: 'ONNX model ID',
-                    enum: ['Xenova/all-MiniLM-L6-v2', 'Xenova/all-mpnet-base-v2'],
-                    default: 'Xenova/all-MiniLM-L6-v2',
-                },
-                hyperbolic: {
-                    type: 'boolean',
-                    description: 'Enable hyperbolic (Poincaré ball) embeddings',
-                    default: true,
-                },
-                curvature: {
-                    type: 'number',
-                    description: 'Poincaré ball curvature (negative)',
-                    default: -1,
-                },
-                cacheSize: {
-                    type: 'number',
-                    description: 'LRU cache size',
-                    default: 256,
-                },
-                force: {
-                    type: 'boolean',
-                    description: 'Overwrite existing configuration',
-                    default: false,
-                },
-            },
-        },
-        handler: async (input) => {
-            const model = input.model || 'Xenova/all-MiniLM-L6-v2';
-            const hyperbolic = input.hyperbolic !== false;
-            const curvature = input.curvature || -1;
-            const cacheSize = input.cacheSize || 256;
-            const force = input.force === true;
-            const existingConfig = loadConfig();
-            if (existingConfig && !force) {
-                return {
-                    success: false,
-                    error: 'Embeddings already initialized. Use force=true to overwrite.',
-                    existingConfig: {
-                        model: existingConfig.model,
-                        initialized: existingConfig.initialized,
-                    },
-                };
-            }
-            const dimension = model.includes('mpnet') ? 768 : 384;
-            const modelPath = resolve(join(CONFIG_DIR, MODELS_DIR));
-            // Create models directory
-            if (!existsSync(modelPath)) {
-                mkdirSync(modelPath, { recursive: true });
-            }
-            const config = {
-                model,
-                modelPath,
-                dimension,
-                cacheSize,
-                hyperbolic: {
-                    enabled: hyperbolic,
-                    curvature,
-                    epsilon: 1e-15,
-                    maxNorm: 1 - 1e-5,
-                },
-                neural: {
-                    enabled: true,
-                    driftThreshold: 0.3,
-                    decayRate: 0.01,
-                },
-                initialized: new Date().toISOString(),
-            };
-            saveConfig(config);
-            return {
-                success: true,
-                config: {
-                    model,
-                    dimension,
-                    cacheSize,
-                    hyperbolic: hyperbolic ? { enabled: true, curvature } : { enabled: false },
-                    neural: { enabled: true },
-                },
-                paths: {
-                    config: getConfigPath(),
-                    models: modelPath,
-                },
-                message: 'Embedding subsystem initialized successfully',
-            };
-        },
-    },
-    {
-        name: 'embeddings_generate',
-        description: 'Generate embeddings for text (Euclidean or hyperbolic)',
-        category: 'embeddings',
-        inputSchema: {
-            type: 'object',
-            properties: {
-                text: {
-                    type: 'string',
-                    description: 'Text to embed',
-                },
-                hyperbolic: {
-                    type: 'boolean',
-                    description: 'Return hyperbolic (Poincaré) embedding',
-                    default: false,
-                },
-                normalize: {
-                    type: 'boolean',
-                    description: 'L2 normalize the embedding',
-                    default: true,
-                },
-            },
-            required: ['text'],
-        },
-        handler: async (input) => {
-            const config = loadConfig();
-            if (!config) {
-                return {
-                    success: false,
-                    error: 'Embeddings not initialized. Run embeddings/init first.',
-                };
-            }
-            let text;
-            try {
-                text = validateText(input.text, 'text');
-            }
-            catch (e) {
-                return { success: false, error: e.message };
-            }
-            const useHyperbolic = input.hyperbolic === true && config.hyperbolic.enabled;
-            // Generate real ONNX embedding
-            const embedding = await generateRealEmbedding(text, config.dimension);
-            let result;
-            let geometry;
-            if (useHyperbolic) {
-                result = toPoincare(embedding, config.hyperbolic.curvature);
-                geometry = 'poincare';
-            }
-            else {
-                result = embedding;
-                geometry = 'euclidean';
-            }
-            return {
-                success: true,
-                embedding: result,
-                metadata: {
-                    model: config.model,
-                    dimension: config.dimension,
-                    geometry,
-                    curvature: useHyperbolic ? config.hyperbolic.curvature : null,
-                    textLength: text.length,
-                    norm: Math.sqrt(result.reduce((sum, x) => sum + x * x, 0)),
-                },
-            };
-        },
-    },
-    {
-        name: 'embeddings_compare',
-        description: 'Compare similarity between two texts',
-        category: 'embeddings',
-        inputSchema: {
-            type: 'object',
-            properties: {
-                text1: {
-                    type: 'string',
-                    description: 'First text',
-                },
-                text2: {
-                    type: 'string',
-                    description: 'Second text',
-                },
-                metric: {
-                    type: 'string',
-                    description: 'Similarity metric',
-                    enum: ['cosine', 'euclidean', 'poincare'],
-                    default: 'cosine',
-                },
-            },
-            required: ['text1', 'text2'],
-        },
-        handler: async (input) => {
-            const config = loadConfig();
-            if (!config) {
-                return {
-                    success: false,
-                    error: 'Embeddings not initialized. Run embeddings/init first.',
-                };
-            }
-            let text1, text2;
-            try {
-                text1 = validateText(input.text1, 'text1');
-                text2 = validateText(input.text2, 'text2');
-            }
-            catch (e) {
-                return { success: false, error: e.message };
-            }
-            const metric = input.metric || 'cosine';
-            // Generate real ONNX embeddings for both texts
-            const [emb1, emb2] = await Promise.all([
-                generateRealEmbedding(text1, config.dimension),
-                generateRealEmbedding(text2, config.dimension)
-            ]);
-            let similarity;
-            let distance;
-            switch (metric) {
-                case 'poincare':
-                    if (!config.hyperbolic.enabled) {
-                        return {
-                            success: false,
-                            error: 'Hyperbolic mode not enabled. Initialize with hyperbolic=true.',
-                        };
-                    }
-                    const poinc1 = toPoincare(emb1, config.hyperbolic.curvature);
-                    const poinc2 = toPoincare(emb2, config.hyperbolic.curvature);
-                    distance = poincareDistance(poinc1, poinc2, config.hyperbolic.curvature);
-                    similarity = 1 / (1 + distance);
-                    break;
-                case 'euclidean':
-                    distance = Math.sqrt(emb1.reduce((sum, _, i) => sum + (emb1[i] - emb2[i]) ** 2, 0));
-                    similarity = 1 / (1 + distance);
-                    break;
-                default: // cosine
-                    similarity = cosineSimilarity(emb1, emb2);
-                    distance = 1 - similarity;
-            }
-            return {
-                success: true,
-                similarity,
-                distance,
-                metric,
-                texts: {
-                    text1: { length: text1.length, preview: text1.slice(0, 50) },
-                    text2: { length: text2.length, preview: text2.slice(0, 50) },
-                },
-                interpretation: similarity > 0.8 ? 'very similar' :
-                    similarity > 0.6 ? 'similar' :
-                        similarity > 0.4 ? 'somewhat similar' :
-                            similarity > 0.2 ? 'different' : 'very different',
-            };
-        },
-    },
-    {
-        name: 'embeddings_search',
-        description: 'Semantic search across stored embeddings',
-        category: 'embeddings',
-        inputSchema: {
-            type: 'object',
-            properties: {
-                query: {
-                    type: 'string',
-                    description: 'Search query',
-                },
-                topK: {
-                    type: 'number',
-                    description: 'Number of results to return',
-                    default: 5,
-                },
-                threshold: {
-                    type: 'number',
-                    description: 'Minimum similarity threshold (0-1)',
-                    default: 0.5,
-                },
-                namespace: {
-                    type: 'string',
-                    description: 'Search in specific namespace',
-                },
-            },
-            required: ['query'],
-        },
-        handler: async (input) => {
-            const config = loadConfig();
-            if (!config) {
-                return {
-                    success: false,
-                    error: 'Embeddings not initialized. Run embeddings/init first.',
-                };
-            }
-            // Cap query length: generateRealEmbedding has a hash-fallback that is
-            // O(n) over the text length, so an unbounded query is a DoS vector.
-            // Cap topK to prevent requesting a huge result set from searchEntries.
-            let query;
-            try {
-                query = validateText(input.query, 'query');
-            }
-            catch (e) {
-                return { success: false, error: e.message };
-            }
-            const MAX_SEARCH_TOP_K = 100;
-            const rawTopK = input.topK || 5;
-            const topK = Number.isFinite(rawTopK) && rawTopK > 0
-                ? Math.min(Math.floor(rawTopK), MAX_SEARCH_TOP_K) : 5;
-            const threshold = input.threshold || 0.5;
-            const namespace = input.namespace;
-            const startTime = performance.now();
-            // Generate real ONNX embedding for query
-            const queryEmbedding = await generateRealEmbedding(query, config.dimension);
-            // Try to search using real memory search
-            try {
-                const { searchEntries } = await import('../memory/memory-initializer.js');
-                const searchResult = await searchEntries({
-                    query,
-                    limit: topK,
-                    threshold,
-                    namespace: namespace || 'default'
-                });
-                const searchTime = (performance.now() - startTime).toFixed(2);
-                return {
-                    success: true,
-                    query,
-                    results: searchResult.results.map((r) => ({
-                        key: r.key,
-                        content: r.content?.substring(0, 100),
-                        similarity: r.score,
-                        namespace: r.namespace
-                    })),
-                    metadata: {
-                        model: config.model,
-                        topK,
-                        threshold,
-                        namespace: namespace || 'default',
-                        searchTime: `${searchTime}ms`,
-                        indexType: config.hyperbolic.enabled ? 'HNSW (hyperbolic)' : 'HNSW (euclidean)',
-                        resultCount: searchResult.results.length
-                    },
-                };
-            }
-            catch (e) {
-                // Database not available - return empty but truthful
-                if (process.env.DEBUG || process.env.MONOMIND_DEBUG)
-                    console.error('[embeddings-tools] search failed:', e);
-                const searchTime = (performance.now() - startTime).toFixed(2);
-                return {
-                    success: true,
-                    query,
-                    results: [],
-                    metadata: {
-                        model: config.model,
-                        topK,
-                        threshold,
-                        namespace: namespace || 'default',
-                        searchTime: `${searchTime}ms`,
-                        indexType: config.hyperbolic.enabled ? 'HNSW (hyperbolic)' : 'HNSW (euclidean)',
-                    },
-                    message: 'No embeddings indexed yet. Use memory store to add documents.',
-                };
-            }
-        },
-    },
-    {
-        name: 'embeddings_neural',
-        description: 'Report and adjust pattern-store statistics (drift, consolidate, adapt) from intelligence.ts. Reads real counters — but its "init" action only persists a config blob (sona, flashAttention, ewcPlusPlus, memoryPhysics, swarmCoordination…) for which no implementing code exists anywhere; those flags are stored and echoed back, never acted on. There is no neural substrate and no "memory physics".',
-        category: 'embeddings',
-        inputSchema: {
-            type: 'object',
-            properties: {
-                action: {
-                    type: 'string',
-                    description: 'Neural action',
-                    enum: ['status', 'init', 'drift', 'consolidate', 'adapt'],
-                    default: 'status',
-                },
-                driftThreshold: {
-                    type: 'number',
-                    description: 'Semantic drift detection threshold',
-                    default: 0.3,
-                },
-                decayRate: {
-                    type: 'number',
-                    description: 'Decay rate applied to stored pattern confidence (plain exponential decay — the "hippocampal dynamics" framing this once carried described nothing in the code)',
-                    default: 0.01,
-                },
-            },
-        },
-        handler: async (input) => {
-            const config = loadConfig();
-            if (!config) {
-                return {
-                    success: false,
-                    error: 'Embeddings not initialized. Run embeddings/init first.',
-                };
-            }
-            const action = input.action || 'status';
-            switch (action) {
-                case 'init':
-                    config.neural = {
-                        enabled: true,
-                        driftThreshold: input.driftThreshold || 0.3,
-                        decayRate: input.decayRate || 0.01,
-                        monovector: {
-                            enabled: true,
-                            sona: true,
-                            flashAttention: true,
-                            ewcPlusPlus: true,
-                        },
-                        features: {
-                            semanticDrift: true,
-                            memoryPhysics: true,
-                            stateMachine: true,
-                            swarmCoordination: true,
-                            coherenceMonitor: true,
-                        },
-                    };
-                    saveConfig(config);
-                    return {
-                        success: true,
-                        action: 'init',
-                        neural: config.neural,
-                        message: 'Embedding substrate initialized',
-                    };
-                case 'drift':
-                    // Get real drift metrics if available.
-                    // initializeIntelligence() MUST run first: getIntelligenceStats()
-                    // reads module singletons (sonaCoordinator / reasoningBank) that are
-                    // null until init, so a populated store otherwise reports 0 patterns.
-                    try {
-                        const { getIntelligenceStats, initializeIntelligence } = await import('../memory/intelligence.js');
-                        await initializeIntelligence();
-                        const stats = getIntelligenceStats();
-                        return {
-                            success: true,
-                            action: 'drift',
-                            status: {
-                                semanticDrift: {
-                                    enabled: config.neural.features?.semanticDrift ?? false,
-                                    threshold: config.neural.driftThreshold,
-                                    patternsTracked: stats.patternsLearned,
-                                    status: stats.patternsLearned > 0 ? 'tracking' : 'no patterns',
-                                },
-                            },
-                            message: stats.patternsLearned > 0
-                                ? `Tracking ${stats.patternsLearned} patterns for drift`
-                                : 'No patterns stored yet - drift detection inactive',
-                        };
-                    }
-                    catch (e) {
-                        // Failing to read the store is NOT a drift report of zero.
-                        return {
-                            success: false,
-                            action: 'drift',
-                            error: `Intelligence store unavailable — drift status unknown: ${e.message}`,
-                            status: { semanticDrift: { enabled: false, reason: 'Intelligence module unavailable' } },
-                        };
-                    }
-                case 'consolidate':
-                    // Get real consolidation metrics — same init requirement as 'drift'.
-                    try {
-                        const { getIntelligenceStats, initializeIntelligence } = await import('../memory/intelligence.js');
-                        await initializeIntelligence();
-                        const stats = getIntelligenceStats();
-                        return {
-                            success: true,
-                            action: 'consolidate',
-                            status: {
-                                memoryPhysics: {
-                                    enabled: config.neural.features?.memoryPhysics ?? false,
-                                    decayRate: config.neural.decayRate,
-                                    patternsStored: stats.reasoningBankSize,
-                                    trajectoriesRecorded: stats.trajectoriesRecorded,
-                                },
-                            },
-                            message: `ReasoningBank: ${stats.reasoningBankSize} patterns, ${stats.trajectoriesRecorded} trajectories`,
-                        };
-                    }
-                    catch (e) {
-                        return {
-                            success: false,
-                            action: 'consolidate',
-                            error: `Intelligence store unavailable — consolidation status unknown: ${e.message}`,
-                            status: { memoryPhysics: { enabled: false, reason: 'Intelligence module unavailable' } },
-                        };
-                    }
-                case 'adapt':
-                    // Get real (JS) pattern-adaptation metrics
-                    try {
-                        const { benchmarkAdaptation, initializeIntelligence } = await import('../memory/intelligence.js');
-                        await initializeIntelligence();
-                        const benchmark = benchmarkAdaptation(100);
-                        return {
-                            success: true,
-                            action: 'adapt',
-                            status: {
-                                sona: {
-                                    enabled: true,
-                                    adaptationTime: `${(benchmark.avgMs * 1000).toFixed(2)}μs`,
-                                    targetMet: benchmark.targetMet,
-                                    minTime: `${(benchmark.minMs * 1000).toFixed(2)}μs`,
-                                    maxTime: `${(benchmark.maxMs * 1000).toFixed(2)}μs`,
-                                },
-                            },
-                            message: benchmark.targetMet
-                                ? `Pattern adaptation: ${(benchmark.avgMs * 1000).toFixed(2)}μs (target <50μs met)`
-                                : `Pattern adaptation: ${(benchmark.avgMs * 1000).toFixed(2)}μs (target not met)`,
-                        };
-                    }
-                    catch {
-                        return {
-                            success: true,
-                            action: 'adapt',
-                            status: { sona: { enabled: false, reason: 'Intelligence module unavailable' } },
-                        };
-                    }
-                default: // status
-                    // Get real neural system status
-                    try {
-                        const { getIntelligenceStats, benchmarkAdaptation, initializeIntelligence } = await import('../memory/intelligence.js');
-                        await initializeIntelligence();
-                        const stats = getIntelligenceStats();
-                        const benchmark = benchmarkAdaptation(50);
-                        return {
-                            success: true,
-                            action: 'status',
-                            neural: {
-                                enabled: config.neural.enabled,
-                                sonaEnabled: stats.sonaEnabled,
-                                monovector: config.neural.monovector || { enabled: false },
-                                features: config.neural.features || {},
-                                realMetrics: {
-                                    patternsLearned: stats.patternsLearned,
-                                    trajectoriesRecorded: stats.trajectoriesRecorded,
-                                    reasoningBankSize: stats.reasoningBankSize,
-                                    adaptationTime: `${(benchmark.avgMs * 1000).toFixed(2)}μs`,
-                                    targetMet: benchmark.targetMet,
-                                    lastAdaptation: stats.lastAdaptation
-                                        ? new Date(stats.lastAdaptation).toISOString()
-                                        : null,
-                                },
-                            },
-                            capabilities: [
-                                stats.sonaEnabled ? '✅ Pattern logging active' : '❌ Pattern logging inactive',
-                                benchmark.targetMet ? '✅ <0.05ms Target Met' : '⚠️ Target Not Met',
-                                `${stats.patternsLearned} patterns learned`,
-                                `${stats.trajectoriesRecorded} trajectories recorded`,
-                            ],
-                        };
-                    }
-                    catch {
-                        return {
-                            success: true,
-                            action: 'status',
-                            neural: {
-                                enabled: config.neural.enabled,
-                                monovector: config.neural.monovector || { enabled: false },
-                                features: config.neural.features || {},
-                            },
-                            message: 'Intelligence module not available - showing config only',
-                        };
-                    }
-            }
-        },
-    },
-    {
-        name: 'embeddings_hyperbolic',
-        description: 'Hyperbolic embedding operations (Poincaré ball)',
-        category: 'embeddings',
-        inputSchema: {
-            type: 'object',
-            properties: {
-                action: {
-                    type: 'string',
-                    description: 'Hyperbolic action',
-                    enum: ['status', 'convert', 'distance', 'midpoint'],
-                    default: 'status',
-                },
-                embedding: {
-                    type: 'array',
-                    description: 'Euclidean embedding to convert',
-                    items: { type: 'number' },
-                },
-                embedding1: {
-                    type: 'array',
-                    description: 'First embedding for distance/midpoint',
-                    items: { type: 'number' },
-                },
-                embedding2: {
-                    type: 'array',
-                    description: 'Second embedding for distance/midpoint',
-                    items: { type: 'number' },
-                },
-            },
-        },
-        handler: async (input) => {
-            const config = loadConfig();
-            if (!config) {
-                return {
-                    success: false,
-                    error: 'Embeddings not initialized. Run embeddings/init first.',
-                };
-            }
-            if (!config.hyperbolic.enabled) {
-                return {
-                    success: false,
-                    error: 'Hyperbolic mode not enabled. Initialize with hyperbolic=true.',
-                };
-            }
-            const action = input.action || 'status';
-            const curvature = config.hyperbolic.curvature;
-            switch (action) {
-                case 'convert': {
-                    // validateVector caps at MAX_VECTOR_DIM (8192) — prevents O(n) DoS
-                    // in toPoincare (reduce + map over the full array).
-                    let embedding;
-                    try {
-                        embedding = validateVector(input.embedding, 'embedding');
-                    }
-                    catch (e) {
-                        return { success: false, error: e.message };
-                    }
-                    const poincare = toPoincare(embedding, curvature);
-                    return {
-                        success: true,
-                        action: 'convert',
-                        euclidean: embedding,
-                        poincare,
-                        curvature,
-                        poincareNorm: Math.sqrt(poincare.reduce((sum, x) => sum + x * x, 0)),
-                    };
-                }
-                case 'distance': {
-                    let emb1, emb2;
-                    try {
-                        emb1 = validateVector(input.embedding1, 'embedding1');
-                        emb2 = validateVector(input.embedding2, 'embedding2');
-                    }
-                    catch (e) {
-                        return { success: false, error: e.message };
-                    }
-                    const dist = poincareDistance(emb1, emb2, curvature);
-                    return {
-                        success: true,
-                        action: 'distance',
-                        distance: dist,
-                        curvature,
-                        interpretation: dist < 1 ? 'close' : dist < 2 ? 'moderate' : 'far',
-                    };
-                }
-                case 'midpoint': {
-                    let e1, e2;
-                    try {
-                        e1 = validateVector(input.embedding1, 'embedding1');
-                        e2 = validateVector(input.embedding2, 'embedding2');
-                    }
-                    catch (e) {
-                        return { success: false, error: e.message };
-                    }
-                    // Simplified midpoint (proper Möbius midpoint is more complex)
-                    const mid = e1.map((_, i) => (e1[i] + e2[i]) / 2);
-                    const norm = Math.sqrt(mid.reduce((sum, x) => sum + x * x, 0));
-                    const scaledMid = mid.map(x => x * (config.hyperbolic.maxNorm / Math.max(norm, config.hyperbolic.maxNorm)));
-                    return {
-                        success: true,
-                        action: 'midpoint',
-                        midpoint: scaledMid,
-                        curvature,
-                    };
-                }
-                default: // status
-                    return {
-                        success: true,
-                        action: 'status',
-                        hyperbolic: {
-                            enabled: true,
-                            curvature,
-                            epsilon: config.hyperbolic.epsilon,
-                            maxNorm: config.hyperbolic.maxNorm,
-                        },
-                        benefits: [
-                            'Better hierarchical data representation',
-                            'Exponential capacity in low dimensions',
-                            'Preserves tree-like structures',
-                            'Natural for taxonomy embeddings',
-                        ],
-                    };
-            }
-        },
-    },
-    {
-        name: 'embeddings_status',
-        description: 'Get embeddings system status and configuration',
-        category: 'embeddings',
-        inputSchema: {
-            type: 'object',
-            properties: {},
-        },
-        handler: async () => {
-            const config = loadConfig();
-            if (!config) {
-                return {
-                    success: false,
-                    initialized: false,
-                    message: 'Embeddings not initialized. Run embeddings/init first.',
-                };
-            }
-            return {
-                success: true,
-                initialized: true,
-                config: {
-                    model: config.model,
-                    dimension: config.dimension,
-                    cacheSize: config.cacheSize,
-                    hyperbolic: config.hyperbolic,
-                    neural: {
-                        enabled: config.neural.enabled,
-                        monovector: config.neural.monovector?.enabled ?? false,
-                    },
-                },
-                paths: {
-                    config: getConfigPath(),
-                    models: config.modelPath,
-                },
-                initializedAt: config.initialized,
-                capabilities: {
-                    onnxModels: ['Xenova/all-MiniLM-L6-v2', 'Xenova/all-mpnet-base-v2'],
-                    geometries: ['euclidean', 'poincare'],
-                    normalizations: ['L2', 'L1', 'minmax', 'zscore'],
-                    features: ['semantic search', 'hyperbolic projection', 'neural substrate'],
-                },
-            };
-        },
-    },
-];
-// embeddings_neural's `init` action persists a config blob (sona/flashAttention/
-// ewcPlusPlus/etc.) with no implementing code anywhere for those flags — gated
-// behind MONOMIND_MCP_SPECULATIVE=1. The other 6 tools in this file
-// (embeddings_init/generate/compare/search/hyperbolic/status) do real work
-// and stay visible by default.
-const SPECULATIVE = process.env['MONOMIND_MCP_SPECULATIVE'] === '1';
-export const embeddingsTools = SPECULATIVE
-    ? allEmbeddingsTools
-    : allEmbeddingsTools.filter(t => t.name !== 'embeddings_neural');
-//# sourceMappingURL=embeddings-tools.js.map