agentic-flow 1.8.5 → 1.8.7

package/dist/reasoningbank/config/reasoningbank-types.ts

@@ -0,0 +1,57 @@
+/**
+ * TypeScript configuration types for ReasoningBank
+ */
+
+export interface ReasoningBankConfig {
+  retrieve: {
+    k: number;
+    alpha: number;
+    beta: number;
+    gamma: number;
+    delta: number;
+    recency_half_life_days: number;
+    min_score: number;
+  };
+  judge: {
+    model: string;
+    max_tokens: number;
+    confidence_threshold: number;
+  };
+  distill: {
+    model: string;
+    max_tokens: number;
+    temperature: number;
+  };
+  consolidate: {
+    duplicate_threshold: number;
+    contradiction_threshold: number;
+    trigger_threshold: number;
+    prune_age_days: number;
+    prune_min_confidence: number;
+    min_confidence_keep: number;
+  };
+  matts: {
+    parallel_k: number;
+    sequential_k: number;
+    sequential_r: number;
+    sequential_stop_on_success: boolean;
+    confidence_boost: number;
+  };
+  embeddings: {
+    provider: 'claude' | 'openai';
+    model: string;
+    dims: number;
+    dimensions: number;
+    cache_ttl_seconds: number;
+  };
+  governance: {
+    scrub_pii: boolean;
+    pii_scrubber: boolean;
+    tenant_scoped: boolean;
+  };
+  features?: {
+    enable_pre_task_hook?: boolean;
+    enable_post_task_hook?: boolean;
+    enable_matts_parallel?: boolean;
+  };
+}

package/dist/reasoningbank/config/reasoningbank.yaml

@@ -0,0 +1,145 @@
+reasoningbank:
+  version: "1.0.0"
+  enabled: true
+
+  # ============================================================================
+  # Retrieval Configuration (Algorithm 1)
+  # ============================================================================
+  retrieve:
+    k: 3                           # Top-k memories to inject into system prompt
+    alpha: 0.65                    # Weight: semantic similarity (cosine)
+    beta: 0.15                     # Weight: recency (exponential decay)
+    gamma: 0.20                    # Weight: reliability (confidence * usage)
+    delta: 0.10                    # Weight: diversity penalty (MMR)
+    recency_half_life_days: 45     # Exponential decay half-life for age
+    duplicate_threshold: 0.87      # Cosine similarity threshold for deduplication
+    min_score: 0.3                 # Don't inject memories below this score
+    max_age_days: 365              # Ignore memories older than this
+
+  # ============================================================================
+  # Embedding Configuration
+  # ============================================================================
+  embeddings:
+    provider: "local"              # "claude" | "openai" | "huggingface" | "local"
+    model: "Xenova/all-MiniLM-L6-v2"  # local transformers.js model
+    # model: "claude-sonnet-4-5-20250929"  # for Claude provider
+    # model: "text-embedding-3-large"    # for OpenAI provider
+    dimensions: 384                # vector dimensions (local: 384, OpenAI: 1536/3072)
+    cache_ttl_seconds: 3600        # cache embeddings for 1 hour
+    batch_size: 16                 # batch embeddings for efficiency
+
+  # ============================================================================
+  # Judge Configuration (Algorithm 2)
+  # ============================================================================
+  judge:
+    model: "claude-sonnet-4-5-20250929"
+    temperature: 0                 # deterministic evaluation
+    max_tokens: 512
+    timeout_ms: 10000
+    cache_verdicts: true           # cache judgments by trajectory hash
+    retry_on_parse_error: true
+    fallback_label: "Failure"      # conservative fallback on judge error
+    fallback_confidence: 0.5
+
+  # ============================================================================
+  # Distillation Configuration (Algorithm 3)
+  # ============================================================================
+  distill:
+    max_items_per_trajectory: 3    # extract up to N memories per trajectory
+    success_confidence_prior: 0.75 # initial confidence for success-derived memories
+    failure_confidence_prior: 0.60 # initial confidence for failure-derived guardrails
+    redact_pii: true               # scrub PII before storing
+    redact_patterns:
+      - '\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b'  # emails
+      - '\b(?:\d{3}-\d{2}-\d{4}|\d{9})\b'                       # SSN
+      - '\b(?:sk-[a-zA-Z0-9]{48}|ghp_[a-zA-Z0-9]{36})\b'        # API keys (Anthropic, GitHub)
+      - '\b(?:xoxb-[a-zA-Z0-9\-]+)\b'                           # Slack tokens
+      - '\b(?:\d{13,19})\b'                                     # Credit card numbers
+    min_content_length: 20         # reject memories with content too short
+    max_content_length: 2000       # truncate if too long
+
+  # ============================================================================
+  # Consolidation Configuration (Algorithm 4)
+  # ============================================================================
+  consolidate:
+    enabled: true
+    run_every_new_items: 20        # consolidate after N new memories
+    contradiction_threshold: 0.60  # NLI probability threshold for contradictions
+    prune_age_days: 180            # hard delete if unused and old
+    min_confidence_keep: 0.30      # prune if confidence drops below this
+    max_contradictions_allowed: 5  # quarantine if contradicts > N high-usage items
+    dedup_similarity_threshold: 0.87  # merge if cosine > this
+    merge_strategy: "keep_highest_usage"  # "keep_highest_usage" | "keep_most_recent"
+
+  # ============================================================================
+  # MaTTS Configuration (Algorithm 5)
+  # ============================================================================
+  matts:
+    enabled: true
+
+    # Parallel mode: k independent rollouts with self-contrast aggregation
+    parallel:
+      k: 6                         # number of parallel rollouts
+      diversity_temperature: 0.9   # sampling temperature for diversity
+      max_concurrent: 3            # concurrent executions (rate limiting)
+      aggregation_model: "claude-sonnet-4-5-20250929"
+      aggregation_max_tokens: 2048
+
+    # Sequential mode: r iterative refinements with check-and-correct
+    sequential:
+      r: 3                         # number of refinement iterations
+      check_instruction: "Review your previous attempt. Identify errors or missing steps. Correct and continue."
+      max_iterations: 5            # hard cap to prevent infinite loops
+      stop_on_success: true        # stop if judge labels Success before r iterations
+
+  # ============================================================================
+  # Governance and Compliance
+  # ============================================================================
+  governance:
+    pii_scrubber: true             # enable PII redaction
+    tenant_scoped: false           # set true for multi-tenant deployments
+    audit_trail: true              # log all memory operations to events table
+    max_memory_age_days: 365       # absolute max age before forced deletion
+    require_approval_for_high_impact: false  # flag high-impact memories for review
+
+  # ============================================================================
+  # Performance and Observability
+  # ============================================================================
+  performance:
+    log_metrics: true              # write to performance_metrics table
+    export_csv_interval_days: 7   # export metrics CSV weekly
+    alert_on_degradation: true     # alert if success rate drops
+    success_rate_threshold: 0.70   # baseline threshold for alerting
+    max_retrieve_latency_ms: 500   # alert if retrieval takes longer
+    max_judge_latency_ms: 5000     # alert if judge takes longer
+
+  # ============================================================================
+  # Learning Rate and Confidence Updates
+  # ============================================================================
+  learning:
+    eta: 0.05                      # learning rate for confidence updates
+    success_boost: 1.0             # confidence += eta * success_boost when used in success
+    failure_penalty: -0.5          # confidence += eta * failure_penalty when used in failure
+    usage_boost_sigmoid: true      # apply sigmoid to usage_count for reliability score
+
+  # ============================================================================
+  # Feature Flags
+  # ============================================================================
+  features:
+    enable_pre_task_hook: true     # retrieve and inject memories before task
+    enable_post_task_hook: true    # judge, distill, consolidate after task
+    enable_matts_parallel: true    # allow parallel MaTTS
+    enable_matts_sequential: true  # allow sequential MaTTS
+    enable_contradiction_detection: true
+    enable_auto_pruning: true
+    enable_memory_merging: true
+
+  # ============================================================================
+  # Development and Debugging
+  # ============================================================================
+  debug:
+    verbose_logging: false
+    save_trajectories: true        # persist all trajectories to task_trajectories
+    save_embeddings: true          # persist embeddings for inspection
+    log_retrieval_scores: false    # log detailed scoring breakdown
+    dry_run: false                 # don't actually upsert memories (testing only)

package/dist/reasoningbank/utils/embeddings.js

@@ -1,82 +1,109 @@
 /**
  * Embedding generation for semantic similarity
- * Supports multiple providers: OpenAI, Claude (placeholder), local hashing
+ * Uses local transformers.js - no API key required!
  */
+import { pipeline, env } from '@xenova/transformers';
 import { loadConfig } from './config.js';
+// Configure transformers.js to use WebAssembly backend (Node.js compatible)
+env.backends.onnx.wasm.numThreads = 1; // Use single thread for stability
+let embeddingPipeline = null;
+let isInitializing = false;
 const embeddingCache = new Map();
 /**
- * Compute embedding for text
- * Uses configured provider (openai, claude, or local)
+ * Initialize the embedding pipeline (lazy load)
+ */
+async function initializeEmbeddings() {
+    if (embeddingPipeline)
+        return;
+    if (isInitializing) {
+        // Wait for initialization to complete
+        while (isInitializing) {
+            await new Promise(resolve => setTimeout(resolve, 100));
+        }
+        return;
+    }
+    isInitializing = true;
+    console.log('[Embeddings] Initializing local embedding model (Xenova/all-MiniLM-L6-v2)...');
+    console.log('[Embeddings] First run will download ~23MB model...');
+    try {
+        embeddingPipeline = await pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2', { quantized: true } // Smaller, faster
+        );
+        console.log('[Embeddings] Local model ready! (384 dimensions)');
+    }
+    catch (error) {
+        console.error('[Embeddings] Failed to initialize:', error?.message || error);
+        console.warn('[Embeddings] Falling back to hash-based embeddings');
+    }
+    finally {
+        isInitializing = false;
+    }
+}
+/**
+ * Compute embedding for text using local model
  */
 export async function computeEmbedding(text) {
     const config = loadConfig();
     // Check cache
-    const cacheKey = `${config.embeddings.provider}:${text}`;
+    const cacheKey = `local:${text}`;
     if (embeddingCache.has(cacheKey)) {
         return embeddingCache.get(cacheKey);
     }
     let embedding;
-    if (config.embeddings.provider === 'openai') {
-        embedding = await openaiEmbed(text, config.embeddings.model);
-    }
-    else if (config.embeddings.provider === 'claude') {
-        // Claude doesn't have native embeddings yet, use deterministic hash
-        embedding = hashEmbed(text, config.embeddings.dimensions || 1024);
+    // Initialize if needed
+    await initializeEmbeddings();
+    if (embeddingPipeline) {
+        try {
+            // Use transformers.js for real embeddings
+            const output = await embeddingPipeline(text, {
+                pooling: 'mean',
+                normalize: true
+            });
+            embedding = new Float32Array(output.data);
+        }
+        catch (error) {
+            console.error('[Embeddings] Generation failed:', error?.message || error);
+            embedding = hashEmbed(text, 384); // Fallback
+        }
     }
     else {
-        // Fallback to local hashing
-        embedding = hashEmbed(text, config.embeddings.dimensions || 1024);
+        // Fallback to hash-based embeddings
+        const dims = config?.embeddings?.dimensions || 384;
+        embedding = hashEmbed(text, dims);
+    }
+    // Cache with LRU (limit 1000 entries)
+    if (embeddingCache.size > 1000) {
+        const firstKey = embeddingCache.keys().next().value;
+        if (firstKey) {
+            embeddingCache.delete(firstKey);
+        }
     }
-    // Cache with TTL
     embeddingCache.set(cacheKey, embedding);
-    setTimeout(() => embeddingCache.delete(cacheKey), config.embeddings.cache_ttl_seconds * 1000);
+    // Set TTL for cache entry
+    const ttl = config?.embeddings?.cache_ttl_seconds || 3600;
+    setTimeout(() => embeddingCache.delete(cacheKey), ttl * 1000);
     return embedding;
 }
 /**
- * OpenAI embeddings API
+ * Batch compute embeddings (more efficient)
  */
-async function openaiEmbed(text, model) {
-    const apiKey = process.env.OPENAI_API_KEY;
-    const config = loadConfig();
-    if (!apiKey) {
-        console.warn('[WARN] OPENAI_API_KEY not set, falling back to hash embeddings');
-        return hashEmbed(text, config.embeddings.dimensions); // Use config dimension
-    }
-    try {
-        const response = await fetch('https://api.openai.com/v1/embeddings', {
-            method: 'POST',
-            headers: {
-                'Authorization': `Bearer ${apiKey}`,
-                'Content-Type': 'application/json'
-            },
-            body: JSON.stringify({
-                model: model || 'text-embedding-3-small',
-                input: text
-            })
-        });
-        if (!response.ok) {
-            throw new Error(`OpenAI API error: ${response.status} ${response.statusText}`);
-        }
-        const json = await response.json();
-        return new Float32Array(json.data[0].embedding);
-    }
-    catch (error) {
-        console.error('[ERROR] OpenAI embedding failed:', error);
-        console.warn('[WARN] Falling back to hash embeddings');
-        const config = loadConfig();
-        return hashEmbed(text, config.embeddings.dimensions);
-    }
+export async function computeEmbeddingBatch(texts) {
+    return Promise.all(texts.map(text => computeEmbedding(text)));
+}
+/**
+ * Get embedding dimensions
+ */
+export function getEmbeddingDimensions() {
+    return 384; // all-MiniLM-L6-v2 uses 384 dimensions
 }
 /**
- * Deterministic hash-based embedding
- * For testing and when API keys are unavailable
+ * Deterministic hash-based embedding (fallback)
  */
 function hashEmbed(text, dims) {
     const hash = simpleHash(text);
     const vec = new Float32Array(dims);
     // Generate deterministic pseudo-random vector from hash
     for (let i = 0; i < dims; i++) {
-        vec[i] = Math.sin(hash * (i + 1) * 0.01) * 0.1 + Math.cos(hash * i * 0.02) * 0.05;
+        vec[i] = Math.sin(hash * (i + 1) * 0.01) + Math.cos(hash * i * 0.02);
     }
     return normalize(vec);
 }
@@ -87,7 +114,7 @@ function simpleHash(str) {
     let hash = 0;
     for (let i = 0; i < str.length; i++) {
         hash = ((hash << 5) - hash) + str.charCodeAt(i);
-        hash |= 0; // Convert to 32-bit integer
+        hash |= 0;
     }
     return Math.abs(hash);
 }
@@ -108,7 +135,7 @@ function normalize(vec) {
     return vec;
 }
 /**
- * Clear embedding cache (for testing)
+ * Clear embedding cache
  */
 export function clearEmbeddingCache() {
     embeddingCache.clear();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentic-flow",
-  "version": "1.8.5",
+  "version": "1.8.7",
   "description": "Production-ready AI agent orchestration platform with 66 specialized agents, 213 MCP tools, ReasoningBank learning memory, and autonomous multi-agent swarms. Built by @ruvnet with Claude Agent SDK, neural networks, memory persistence, GitHub integration, and distributed consensus protocols.",
   "type": "module",
   "main": "dist/index.js",
@@ -24,7 +24,7 @@
   "scripts": {
     "postinstall": "node scripts/postinstall.js || true",
     "start": "node --enable-source-maps dist/index.js",
-    "build": "npm run build:wasm && tsc -p config/tsconfig.json --skipLibCheck || true && cp -r src/reasoningbank/prompts dist/reasoningbank/",
+    "build": "npm run build:wasm && tsc -p config/tsconfig.json --skipLibCheck || true && cp -r src/reasoningbank/prompts dist/reasoningbank/ && cp -r src/reasoningbank/config dist/reasoningbank/",
     "build:wasm": "cd ../reasoningbank && wasm-pack build --target bundler --out-dir pkg/bundler crates/reasoningbank-wasm && wasm-pack build --target web --out-dir pkg/web crates/reasoningbank-wasm && mkdir -p ../agentic-flow/wasm/reasoningbank && cp -r crates/reasoningbank-wasm/pkg/bundler/* ../agentic-flow/wasm/reasoningbank/ && cp -r crates/reasoningbank-wasm/pkg/web ../agentic-flow/wasm/reasoningbank/",
     "build:wasm:clean": "rm -rf ../reasoningbank/crates/reasoningbank-wasm/pkg && rm -rf wasm/reasoningbank",
     "dev": "tsx src/index.ts",
@@ -143,6 +143,7 @@
     "@anthropic-ai/claude-agent-sdk": "^0.1.5",
     "@anthropic-ai/sdk": "^0.65.0",
     "@google/genai": "^1.22.0",
+    "@xenova/transformers": "^2.17.2",
     "agentdb": "^1.4.3",
     "axios": "^1.12.2",
     "better-sqlite3": "^11.10.0",