agent-working-memory 0.5.4 → 0.5.6

package/src/coordination/schema.ts

@@ -0,0 +1,94 @@
+// Copyright 2026 Robert Winter / Complete Ideas
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * SQL table definitions for the coordination module.
+ * Tables are created conditionally when AWM_COORDINATION=true.
+ * Uses the same memory.db — coordination events feed the activation engine.
+ */
+
+import type Database from 'better-sqlite3';
+
+const COORDINATION_TABLES = `
+-- Coordination: agents in the hive
+CREATE TABLE IF NOT EXISTS coord_agents (
+  id           TEXT PRIMARY KEY,
+  name         TEXT NOT NULL,
+  role         TEXT NOT NULL DEFAULT 'worker',
+  status       TEXT NOT NULL DEFAULT 'idle',
+  pid          INTEGER,
+  started_at   TEXT NOT NULL DEFAULT (datetime('now')),
+  last_seen    TEXT NOT NULL DEFAULT (datetime('now')),
+  current_task TEXT,
+  metadata     TEXT,
+  capabilities TEXT,
+  workspace    TEXT
+);
+
+-- Coordination: assignments
+CREATE TABLE IF NOT EXISTS coord_assignments (
+  id           TEXT PRIMARY KEY,
+  agent_id     TEXT,
+  task         TEXT NOT NULL,
+  description  TEXT,
+  status       TEXT NOT NULL DEFAULT 'pending',
+  created_at   TEXT NOT NULL DEFAULT (datetime('now')),
+  started_at   TEXT,
+  completed_at TEXT,
+  result       TEXT,
+  workspace    TEXT,
+  FOREIGN KEY (agent_id) REFERENCES coord_agents(id)
+);
+
+-- Coordination: file locks
+CREATE TABLE IF NOT EXISTS coord_locks (
+  file_path    TEXT PRIMARY KEY,
+  agent_id     TEXT NOT NULL,
+  locked_at    TEXT NOT NULL DEFAULT (datetime('now')),
+  reason       TEXT,
+  FOREIGN KEY (agent_id) REFERENCES coord_agents(id)
+);
+
+-- Coordination: orchestrator broadcast commands
+CREATE TABLE IF NOT EXISTS coord_commands (
+  id           INTEGER PRIMARY KEY AUTOINCREMENT,
+  command      TEXT NOT NULL,
+  reason       TEXT,
+  issued_by    TEXT,
+  issued_at    TEXT NOT NULL DEFAULT (datetime('now')),
+  cleared_at   TEXT,
+  workspace    TEXT
+);
+
+-- Coordination: findings reported by agents
+CREATE TABLE IF NOT EXISTS coord_findings (
+  id           INTEGER PRIMARY KEY AUTOINCREMENT,
+  agent_id     TEXT NOT NULL,
+  category     TEXT NOT NULL,
+  severity     TEXT NOT NULL DEFAULT 'info',
+  file_path    TEXT,
+  line_number  INTEGER,
+  description  TEXT NOT NULL,
+  suggestion   TEXT,
+  status       TEXT NOT NULL DEFAULT 'open',
+  created_at   TEXT NOT NULL DEFAULT (datetime('now')),
+  resolved_at  TEXT,
+  FOREIGN KEY (agent_id) REFERENCES coord_agents(id)
+);
+
+-- Coordination: event audit trail
+CREATE TABLE IF NOT EXISTS coord_events (
+  id           INTEGER PRIMARY KEY AUTOINCREMENT,
+  agent_id     TEXT,
+  event_type   TEXT NOT NULL,
+  detail       TEXT,
+  created_at   TEXT NOT NULL DEFAULT (datetime('now'))
+);
+`;
+
+/**
+ * Create all coordination tables in the given database.
+ * Safe to call multiple times (CREATE IF NOT EXISTS).
+ */
+export function initCoordinationTables(db: Database.Database): void {
+  db.exec(COORDINATION_TABLES);
+}

package/src/coordination/schemas.ts

@@ -0,0 +1,136 @@
+// Copyright 2026 Robert Winter / Complete Ideas
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Zod validation schemas for the coordination module.
+ * Ported from AgentSynapse packages/coordinator/src/schemas.ts.
+ */
+
+import { z } from 'zod';
+
+// ─── Enums ──────────────────────────────────────────────────────
+
+export const agentRoleEnum = z.enum(['worker', 'orchestrator', 'dev-lead']);
+export const agentStatusEnum = z.enum(['idle', 'working', 'dead']);
+export const assignmentStatusEnum = z.enum(['in_progress', 'completed', 'failed', 'blocked']);
+export const commandEnum = z.enum(['BUILD_FREEZE', 'PAUSE', 'RESUME', 'SHUTDOWN']);
+export const findingSeverityEnum = z.enum(['critical', 'error', 'warn', 'info']);
+export const findingCategoryEnum = z.enum([
+  'typecheck', 'lint', 'test-failure', 'security', 'performance',
+  'dead-code', 'todo', 'bug', 'ux', 'a11y', 'sql', 'convention',
+  'freshdesk', 'data-quality', 'other',
+]);
+export const findingStatusEnum = z.enum(['open', 'resolved']);
+
+// ─── Checkin ────────────────────────────────────────────────────
+
+export const checkinSchema = z.object({
+  name: z.string().min(1).max(50),
+  role: agentRoleEnum.default('worker'),
+  pid: z.number().int().positive().optional(),
+  metadata: z.record(z.string(), z.unknown()).optional(),
+  capabilities: z.array(z.string().max(50)).max(20).optional(),
+  workspace: z.string().max(50).optional(),
+});
+
+export const checkoutSchema = z.object({
+  agentId: z.string().uuid(),
+});
+
+// ─── Assignments ────────────────────────────────────────────────
+
+export const assignCreateSchema = z.object({
+  agentId: z.string().uuid().optional(),
+  task: z.string().min(1).max(1000),
+  description: z.string().max(5000).optional(),
+  workspace: z.string().max(50).optional(),
+});
+
+export const assignmentQuerySchema = z.object({
+  agentId: z.string().uuid().optional(),
+  workspace: z.string().max(50).optional(),
+});
+
+export const assignmentClaimSchema = z.object({
+  agentId: z.string().uuid(),
+});
+
+export const assignmentUpdateSchema = z.object({
+  status: assignmentStatusEnum,
+  result: z.string().max(10000).optional(),
+});
+
+// ─── Locks ──────────────────────────────────────────────────────
+
+export const lockAcquireSchema = z.object({
+  agentId: z.string().uuid(),
+  filePath: z.string().min(1).max(500),
+  reason: z.string().max(500).optional(),
+});
+
+export const lockReleaseSchema = z.object({
+  agentId: z.string().uuid(),
+  filePath: z.string().min(1).max(500),
+});
+
+// ─── Commands ───────────────────────────────────────────────────
+
+export const commandCreateSchema = z.object({
+  command: commandEnum,
+  reason: z.string().max(1000).optional(),
+  issuedBy: z.string().max(50).optional(),
+  workspace: z.string().max(50).optional(),
+});
+
+export const commandWaitQuerySchema = z.object({
+  status: z.string().max(20).default('idle'),
+  timeout: z.coerce.number().int().min(0).max(30).optional(),
+  agentId: z.string().optional(),
+  workspace: z.string().max(50).optional(),
+});
+
+// ─── Findings ───────────────────────────────────────────────────
+
+export const findingCreateSchema = z.object({
+  agentId: z.string().uuid(),
+  category: findingCategoryEnum,
+  severity: findingSeverityEnum.default('info'),
+  filePath: z.string().max(500).optional(),
+  lineNumber: z.number().int().positive().optional(),
+  description: z.string().min(1).max(5000),
+  suggestion: z.string().max(5000).optional(),
+});
+
+export const findingsQuerySchema = z.object({
+  category: findingCategoryEnum.optional(),
+  severity: findingSeverityEnum.optional(),
+  status: findingStatusEnum.optional(),
+  limit: z.coerce.number().int().min(1).max(200).default(50),
+});
+
+// ─── Param Schemas ─────────────────────────────────────────────
+
+export const assignmentIdParamSchema = z.object({ id: z.string().uuid() });
+export const findingIdParamSchema = z.object({ id: z.coerce.number().int().positive() });
+
+// ─── Pulse ─────────────────────────────────────────────────────
+
+export const pulseSchema = z.object({
+  agentId: z.string().uuid(),
+});
+
+// ─── Status / Events ────────────────────────────────────────────
+
+export const eventsQuerySchema = z.object({
+  limit: z.coerce.number().int().min(1).max(200).default(50),
+});
+
+export const staleQuerySchema = z.object({
+  seconds: z.coerce.number().int().min(1).max(86400).default(120),
+  cleanup: z.enum(['0', '1', 'true', 'false']).optional(),
+});
+
+export const workersQuerySchema = z.object({
+  capability: z.string().max(50).optional(),
+  status: agentStatusEnum.optional(),
+  workspace: z.string().max(50).optional(),
+});

package/src/coordination/stale.ts

@@ -0,0 +1,89 @@
+// Copyright 2026 Robert Winter / Complete Ideas
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Stale agent detection and cleanup for the coordination module.
+ */
+
+import type Database from 'better-sqlite3';
+
+interface StaleAgent {
+  id: string;
+  name: string;
+  role: string;
+  status: string;
+  last_seen: string;
+  seconds_since_seen: number;
+}
+
+/** Detect agents that haven't checked in within the threshold. */
+export function detectStale(db: Database.Database, thresholdSeconds: number): StaleAgent[] {
+  return db.prepare(
+    `SELECT id, name, role, status, last_seen,
+            ROUND((julianday('now') - julianday(last_seen)) * 86400) AS seconds_since_seen
+     FROM coord_agents
+     WHERE status NOT IN ('dead')
+       AND (julianday('now') - julianday(last_seen)) * 86400 > ?`
+  ).all(thresholdSeconds) as StaleAgent[];
+}
+
+/** Clean up stale agents: fail assignments, release locks, mark dead. */
+export function cleanupStale(db: Database.Database, thresholdSeconds: number): { stale: StaleAgent[]; cleaned: number } {
+  const stale = detectStale(db, thresholdSeconds);
+  let cleaned = 0;
+
+  for (const agent of stale) {
+    // Fail active assignments
+    const orphaned = db.prepare(
+      `UPDATE coord_assignments SET status = 'failed', result = 'agent disconnected (stale)', completed_at = datetime('now')
+       WHERE agent_id = ? AND status IN ('assigned', 'in_progress')`
+    ).run(agent.id);
+
+    if (orphaned.changes > 0) {
+      db.prepare(
+        `INSERT INTO coord_events (agent_id, event_type, detail) VALUES (?, 'assignment_failed', ?)`
+      ).run(agent.id, `auto-failed ${orphaned.changes} orphaned assignment(s) — agent stale`);
+    }
+
+    // Release locks
+    const locks = db.prepare(`DELETE FROM coord_locks WHERE agent_id = ?`).run(agent.id);
+
+    // Mark dead
+    db.prepare(`UPDATE coord_agents SET status = 'dead', current_task = NULL WHERE id = ?`).run(agent.id);
+
+    cleaned += orphaned.changes + locks.changes;
+
+    if (orphaned.changes > 0 || locks.changes > 0) {
+      db.prepare(
+        `INSERT INTO coord_events (agent_id, event_type, detail) VALUES (?, 'stale_cleanup', ?)`
+      ).run(agent.id, `failed ${orphaned.changes} assignment(s), released ${locks.changes} lock(s)`);
+    }
+  }
+
+  return { stale, cleaned };
+}
+
+/** Prune heartbeat events older than 1 hour. Keeps assignment, registered, and command events permanently. */
+export function pruneOldHeartbeats(db: Database.Database): number {
+  const result = db.prepare(
+    `DELETE FROM coord_events WHERE event_type = 'heartbeat' AND created_at < datetime('now', '-1 hour')`
+  ).run();
+  return result.changes;
+}
+
+/** Clean slate on startup: mark all live agents dead, release locks, clear commands. */
+export function cleanSlate(db: Database.Database): void {
+  const alive = db.prepare(
+    `SELECT id, name FROM coord_agents WHERE status != 'dead'`
+  ).all() as Array<{ id: string; name: string }>;
+
+  if (alive.length === 0) return;
+
+  for (const agent of alive) {
+    db.prepare(`UPDATE coord_agents SET status = 'dead', current_task = NULL WHERE id = ?`).run(agent.id);
+    db.prepare(`DELETE FROM coord_locks WHERE agent_id = ?`).run(agent.id);
+  }
+
+  db.prepare(`UPDATE coord_commands SET cleared_at = datetime('now') WHERE cleared_at IS NULL`).run();
+
+  console.log(`  Coordination clean slate: marked ${alive.length} agent(s) from previous session as dead`);
+}

package/src/core/decay.ts

@@ -1,63 +1,63 @@
-// Copyright 2026 Robert Winter / Complete Ideas
-// SPDX-License-Identifier: Apache-2.0
-/**
- * ACT-R Base-Level Activation
- *
- * Based on Anderson's ACT-R cognitive architecture (1993).
- * Memories that are accessed more recently and more frequently
- * have higher activation — a well-established model of human memory.
- *
- * Formula: B(M) = ln(n + 1) - d * ln(ageDays / (n + 1))
- *
- * Where:
- *   n = access count
- *   d = decay exponent (default 0.5)
- *   ageDays = age of memory in days
- */
-
-export function baseLevelActivation(
-  accessCount: number,
-  ageDays: number,
-  decayExponent: number = 0.5
-): number {
-  const n = Math.max(accessCount, 0);
-  const age = Math.max(ageDays, 0.001); // Avoid log(0)
-  return Math.log(n + 1) - decayExponent * Math.log(age / (n + 1));
-}
-
-/**
- * Softplus — smooth approximation of ReLU.
- * Used to keep activation scores positive without hard clipping.
- */
-export function softplus(x: number): number {
-  return Math.log(1 + Math.exp(x));
-}
-
-/**
- * Composite activation score combining content match, temporal decay,
- * Hebbian boost, and confidence.
- *
- * Score = contentMatch * softplus(B(M) + scale * hebbianBoost) * confidence
- */
-export function compositeScore(params: {
-  contentMatch: number;
-  accessCount: number;
-  ageDays: number;
-  hebbianBoost: number;
-  confidence: number;
-  decayExponent?: number;
-  hebbianScale?: number;
-}): number {
-  const {
-    contentMatch,
-    accessCount,
-    ageDays,
-    hebbianBoost,
-    confidence,
-    decayExponent = 0.5,
-    hebbianScale = 1.0,
-  } = params;
-
-  const bm = baseLevelActivation(accessCount, ageDays, decayExponent);
-  return contentMatch * softplus(bm + hebbianScale * hebbianBoost) * confidence;
-}
+// Copyright 2026 Robert Winter / Complete Ideas
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * ACT-R Base-Level Activation
+ *
+ * Based on Anderson's ACT-R cognitive architecture (1993).
+ * Memories that are accessed more recently and more frequently
+ * have higher activation — a well-established model of human memory.
+ *
+ * Formula: B(M) = ln(n + 1) - d * ln(ageDays / (n + 1))
+ *
+ * Where:
+ *   n = access count
+ *   d = decay exponent (default 0.5)
+ *   ageDays = age of memory in days
+ */
+
+export function baseLevelActivation(
+  accessCount: number,
+  ageDays: number,
+  decayExponent: number = 0.5
+): number {
+  const n = Math.max(accessCount, 0);
+  const age = Math.max(ageDays, 0.001); // Avoid log(0)
+  return Math.log(n + 1) - decayExponent * Math.log(age / (n + 1));
+}
+
+/**
+ * Softplus — smooth approximation of ReLU.
+ * Used to keep activation scores positive without hard clipping.
+ */
+export function softplus(x: number): number {
+  return Math.log(1 + Math.exp(x));
+}
+
+/**
+ * Composite activation score combining content match, temporal decay,
+ * Hebbian boost, and confidence.
+ *
+ * Score = contentMatch * softplus(B(M) + scale * hebbianBoost) * confidence
+ */
+export function compositeScore(params: {
+  contentMatch: number;
+  accessCount: number;
+  ageDays: number;
+  hebbianBoost: number;
+  confidence: number;
+  decayExponent?: number;
+  hebbianScale?: number;
+}): number {
+  const {
+    contentMatch,
+    accessCount,
+    ageDays,
+    hebbianBoost,
+    confidence,
+    decayExponent = 0.5,
+    hebbianScale = 1.0,
+  } = params;
+
+  const bm = baseLevelActivation(accessCount, ageDays, decayExponent);
+  return contentMatch * softplus(bm + hebbianScale * hebbianBoost) * confidence;
+}

package/src/core/embeddings.ts

@@ -1,88 +1,88 @@
-// Copyright 2026 Robert Winter / Complete Ideas
-// SPDX-License-Identifier: Apache-2.0
-/**
- * Embedding Engine — local vector embeddings via transformers.js
- *
- * Default: bge-small-en-v1.5 (384 dimensions, ~90MB, MTEB retrieval-optimized).
- * Better short-text similarity than MiniLM for agent memory concepts.
- * Configurable via AWM_EMBED_MODEL env var.
- * Model is downloaded once on first use and cached locally.
- *
- * Singleton pattern — call getEmbedder() to get the shared instance.
- *
- * NOTE: Changing the model invalidates existing embeddings.
- * Set AWM_EMBED_MODEL=Xenova/all-MiniLM-L6-v2 for backward compatibility.
- */
-
-import { pipeline, type FeatureExtractionPipeline } from '@huggingface/transformers';
-
-const MODEL_ID = process.env.AWM_EMBED_MODEL ?? 'Xenova/bge-small-en-v1.5';
-const DIMENSIONS = parseInt(process.env.AWM_EMBED_DIMS ?? '384', 10);
-const POOLING = (process.env.AWM_EMBED_POOLING ?? 'mean') as 'cls' | 'mean';
-
-let instance: FeatureExtractionPipeline | null = null;
-let initPromise: Promise<FeatureExtractionPipeline> | null = null;
-
-/**
- * Get or initialize the embedding pipeline (singleton).
- * First call downloads the model (~22MB), subsequent calls are instant.
- */
-export async function getEmbedder(): Promise<FeatureExtractionPipeline> {
-  if (instance) return instance;
-  if (initPromise) return initPromise;
-
-  initPromise = pipeline('feature-extraction', MODEL_ID, {
-    dtype: 'fp32',
-  }).then(pipe => {
-    instance = pipe;
-    console.log(`Embedding model loaded: ${MODEL_ID} (${DIMENSIONS}d)`);
-    return pipe;
-  });
-
-  return initPromise;
-}
-
-/**
- * Generate an embedding vector for a text string.
- * Returns a normalized float32 array of length DIMENSIONS.
- */
-export async function embed(text: string): Promise<number[]> {
-  const embedder = await getEmbedder();
-  const result = await embedder(text, { pooling: POOLING, normalize: true });
-  // result is a Tensor — extract the data
-  return Array.from(result.data as Float32Array).slice(0, DIMENSIONS);
-}
-
-/**
- * Generate embeddings for multiple texts in a batch.
- * More efficient than calling embed() in a loop.
- */
-export async function embedBatch(texts: string[]): Promise<number[][]> {
-  if (texts.length === 0) return [];
-  const embedder = await getEmbedder();
-  const result = await embedder(texts, { pooling: POOLING, normalize: true });
-  const data = result.data as Float32Array;
-
-  const vectors: number[][] = [];
-  for (let i = 0; i < texts.length; i++) {
-    vectors.push(Array.from(data.slice(i * DIMENSIONS, (i + 1) * DIMENSIONS)));
-  }
-  return vectors;
-}
-
-/**
- * Cosine similarity between two normalized vectors.
- * Since vectors are pre-normalized, this is just the dot product.
- */
-export function cosineSimilarity(a: number[], b: number[]): number {
-  if (a.length !== b.length || a.length === 0) return 0;
-  let dot = 0;
-  for (let i = 0; i < a.length; i++) {
-    dot += a[i] * b[i];
-  }
-  // Clamp to [-1, 1] to handle floating point drift
-  return Math.max(-1, Math.min(1, dot));
-}
-
-/** Vector dimensions for this model */
-export const EMBEDDING_DIMENSIONS = DIMENSIONS;
+// Copyright 2026 Robert Winter / Complete Ideas
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Embedding Engine — local vector embeddings via transformers.js
+ *
+ * Default: bge-small-en-v1.5 (384 dimensions, ~90MB, MTEB retrieval-optimized).
+ * Better short-text similarity than MiniLM for agent memory concepts.
+ * Configurable via AWM_EMBED_MODEL env var.
+ * Model is downloaded once on first use and cached locally.
+ *
+ * Singleton pattern — call getEmbedder() to get the shared instance.
+ *
+ * NOTE: Changing the model invalidates existing embeddings.
+ * Set AWM_EMBED_MODEL=Xenova/all-MiniLM-L6-v2 for backward compatibility.
+ */
+
+import { pipeline, type FeatureExtractionPipeline } from '@huggingface/transformers';
+
+const MODEL_ID = process.env.AWM_EMBED_MODEL ?? 'Xenova/bge-small-en-v1.5';
+const DIMENSIONS = parseInt(process.env.AWM_EMBED_DIMS ?? '384', 10);
+const POOLING = (process.env.AWM_EMBED_POOLING ?? 'mean') as 'cls' | 'mean';
+
+let instance: FeatureExtractionPipeline | null = null;
+let initPromise: Promise<FeatureExtractionPipeline> | null = null;
+
+/**
+ * Get or initialize the embedding pipeline (singleton).
+ * First call downloads the model (~22MB), subsequent calls are instant.
+ */
+export async function getEmbedder(): Promise<FeatureExtractionPipeline> {
+  if (instance) return instance;
+  if (initPromise) return initPromise;
+
+  initPromise = pipeline('feature-extraction', MODEL_ID, {
+    dtype: 'fp32',
+  }).then(pipe => {
+    instance = pipe;
+    console.log(`Embedding model loaded: ${MODEL_ID} (${DIMENSIONS}d)`);
+    return pipe;
+  });
+
+  return initPromise;
+}
+
+/**
+ * Generate an embedding vector for a text string.
+ * Returns a normalized float32 array of length DIMENSIONS.
+ */
+export async function embed(text: string): Promise<number[]> {
+  const embedder = await getEmbedder();
+  const result = await embedder(text, { pooling: POOLING, normalize: true });
+  // result is a Tensor — extract the data
+  return Array.from(result.data as Float32Array).slice(0, DIMENSIONS);
+}
+
+/**
+ * Generate embeddings for multiple texts in a batch.
+ * More efficient than calling embed() in a loop.
+ */
+export async function embedBatch(texts: string[]): Promise<number[][]> {
+  if (texts.length === 0) return [];
+  const embedder = await getEmbedder();
+  const result = await embedder(texts, { pooling: POOLING, normalize: true });
+  const data = result.data as Float32Array;
+
+  const vectors: number[][] = [];
+  for (let i = 0; i < texts.length; i++) {
+    vectors.push(Array.from(data.slice(i * DIMENSIONS, (i + 1) * DIMENSIONS)));
+  }
+  return vectors;
+}
+
+/**
+ * Cosine similarity between two normalized vectors.
+ * Since vectors are pre-normalized, this is just the dot product.
+ */
+export function cosineSimilarity(a: number[], b: number[]): number {
+  if (a.length !== b.length || a.length === 0) return 0;
+  let dot = 0;
+  for (let i = 0; i < a.length; i++) {
+    dot += a[i] * b[i];
+  }
+  // Clamp to [-1, 1] to handle floating point drift
+  return Math.max(-1, Math.min(1, dot));
+}
+
+/** Vector dimensions for this model */
+export const EMBEDDING_DIMENSIONS = DIMENSIONS;