@jlongo78/agent-spaces 0.7.5 → 0.7.6

package/.next/standalone/docs/superpowers/plans/2026-03-15-cortex-v2-context-engine.md

@@ -1,1274 +1,1274 @@
-# Cortex v2 — Pillar 3: Context Assembly Engine
-
-> **For agentic workers:** REQUIRED: Use superpowers:subagent-driven-development (if subagents available) or superpowers:executing-plans to implement this plan. Steps use checkbox (`- [ ]`) syntax for tracking.
-
-**Goal:** Replace the flat layer-iteration search with a 6-stage context assembly pipeline that detects intent, resolves entities, computes graph-aware weights, searches multiple scopes in parallel, fuses results with evidence scoring, and surfaces conflicts — all within 150ms.
-
-**Architecture:** A new `ContextEngine` class wraps the existing `CortexStore` (for low-level vector search) and `EntityGraph` (for graph distance/proximity). It does NOT replace `CortexSearch` — instead it provides the higher-level retrieval interface that the RAG hook calls. The existing `CortexSearch` remains for backward-compatible simple searches. The RAG hook (`cortex-hook.js`) switches from calling the search API to calling a new context-assembly API endpoint.
-
-**Tech Stack:** TypeScript, vitest, LanceDB, SQLite (entity graph)
-
-**Spec:** `docs/superpowers/specs/2026-03-14-cortex-v2-design.md` — Pillar 3
-
-**Depends on:** Pillar 1 (Entity Graph) + Pillar 2 (Knowledge Unit Evolution) — both completed
-
----
-
-## File Structure
-
-```
-New files:
-├── src/lib/cortex/retrieval/intent.ts         — Intent detection (regex + keyword)
-├── src/lib/cortex/retrieval/weight.ts          — Weight computation (graph × intent × freshness × authority)
-├── src/lib/cortex/retrieval/conflict.ts        — Conflict detection in results
-├── src/lib/cortex/retrieval/formatter.ts       — Context formatting for RAG injection
-├── src/lib/cortex/retrieval/context-engine.ts  — Main 6-stage ContextEngine class
-├── src/app/api/cortex/context/route.ts         — API endpoint for context assembly
-
-Modified files:
-├── bin/cortex-hook.js                          — Switch to context-assembly endpoint
-
-Test files:
-├── tests/lib/cortex/retrieval/intent.test.ts
-├── tests/lib/cortex/retrieval/weight.test.ts
-├── tests/lib/cortex/retrieval/conflict.test.ts
-├── tests/lib/cortex/retrieval/formatter.test.ts
-├── tests/lib/cortex/retrieval/context-engine.test.ts
-```
-
----
-
-## Chunk 1: Intent Detection and Weight Computation
-
-### Task 1: Intent detection
-
-**Files:**
-- Create: `src/lib/cortex/retrieval/intent.ts`
-- Create: `tests/lib/cortex/retrieval/intent.test.ts`
-
-- [ ] **Step 1: Write failing tests**
-
-```typescript
-// tests/lib/cortex/retrieval/intent.test.ts
-import { describe, it, expect } from 'vitest';
-import { detectIntent, INTENTS } from '@/lib/cortex/retrieval/intent';
-import type { IntentResult } from '@/lib/cortex/retrieval/intent';
-
-describe('detectIntent', () => {
-  it('detects debugging intent', () => {
-    const result = detectIntent('why does the auth service throw a timeout error?');
-    expect(result.intent).toBe('debugging');
-    expect(result.confidence).toBeGreaterThan(0.5);
-  });
-
-  it('detects architecture intent', () => {
-    const result = detectIntent('what architecture pattern should we use for the new service?');
-    expect(result.intent).toBe('architecture');
-  });
-
-  it('detects how-to intent', () => {
-    const result = detectIntent('how do I deploy this service to production?');
-    expect(result.intent).toBe('how-to');
-  });
-
-  it('detects security intent', () => {
-    const result = detectIntent('is there a vulnerability in our authentication flow?');
-    expect(result.intent).toBe('security');
-  });
-
-  it('defaults to general for ambiguous queries', () => {
-    const result = detectIntent('tell me about the project');
-    expect(result.intent).toBe('general');
-  });
-
-  it('returns bias config for the detected intent', () => {
-    const result = detectIntent('fix this bug in the login page');
-    expect(result.biases).toBeDefined();
-    expect(result.biases.scope_boost).toBeDefined();
-    expect(result.biases.type_boost).toBeDefined();
-  });
-
-  it('exports all intent definitions', () => {
-    expect(Object.keys(INTENTS)).toContain('debugging');
-    expect(Object.keys(INTENTS)).toContain('architecture');
-    expect(Object.keys(INTENTS)).toContain('general');
-    expect(Object.keys(INTENTS).length).toBe(8);
-  });
-});
-```
-
-- [ ] **Step 2: Run tests to verify they fail**
-
-Run: `npx vitest run tests/lib/cortex/retrieval/intent.test.ts`
-
-- [ ] **Step 3: Implement intent detection**
-
-```typescript
-// src/lib/cortex/retrieval/intent.ts
-
-export interface IntentBiases {
-  scope_boost: Record<string, number>;  // scope level → multiplier
-  type_boost: Record<string, number>;   // knowledge type → multiplier
-  recency_boost: number;                // extra recency multiplier
-}
-
-export interface IntentResult {
-  intent: string;
-  confidence: number;
-  biases: IntentBiases;
-}
-
-interface IntentDef {
-  patterns: RegExp[];
-  keywords: string[];
-  biases: IntentBiases;
-}
-
-export const INTENTS: Record<string, IntentDef> = {
-  debugging: {
-    patterns: [
-      /\b(error|bug|fix|crash|fail|broken|throw|exception|timeout|issue)\b/i,
-      /\bwhy\s+(does|is|did|do)\b/i,
-      /\bnot\s+work/i,
-    ],
-    keywords: ['error', 'bug', 'fix', 'debug', 'crash', 'fail', 'broken', 'throw', 'exception', 'timeout', 'issue', 'stack trace'],
-    biases: {
-      scope_boost: { personal: 1.2, team: 1.0, department: 0.9, organization: 0.8 },
-      type_boost: { error_fix: 1.3, pattern: 1.0, decision: 0.8, conversation: 0.7 },
-      recency_boost: 1.1,
-    },
-  },
-  architecture: {
-    patterns: [
-      /\b(architect|design|pattern|structure|approach)\b/i,
-      /\bshould\s+we\s+(use|adopt|switch|migrate)\b/i,
-    ],
-    keywords: ['architecture', 'design', 'pattern', 'structure', 'approach', 'decision', 'migration', 'refactor'],
-    biases: {
-      scope_boost: { personal: 0.9, team: 1.1, department: 1.2, organization: 1.0 },
-      type_boost: { decision: 1.5, pattern: 1.2, error_fix: 0.7, conversation: 0.5 },
-      recency_boost: 1.0,
-    },
-  },
-  onboarding: {
-    patterns: [
-      /\b(how\s+does|explain|what\s+is|overview|getting\s+started)\b/i,
-      /\bnew\s+to\b/i,
-    ],
-    keywords: ['explain', 'overview', 'introduction', 'getting started', 'onboarding', 'how does'],
-    biases: {
-      scope_boost: { personal: 0.7, team: 1.0, department: 1.1, organization: 1.2 },
-      type_boost: { pattern: 1.3, decision: 1.2, summary: 1.2, conversation: 0.5 },
-      recency_boost: 0.9,
-    },
-  },
-  policy: {
-    patterns: [
-      /\b(policy|compliance|regulation|standard|rule|requirement)\b/i,
-      /\ballowed\s+to\b/i,
-    ],
-    keywords: ['policy', 'compliance', 'standard', 'regulation', 'rule', 'requirement', 'allowed'],
-    biases: {
-      scope_boost: { personal: 0.6, team: 0.8, department: 1.0, organization: 1.3 },
-      type_boost: { decision: 1.5, preference: 1.2, pattern: 0.8, conversation: 0.3 },
-      recency_boost: 1.0,
-    },
-  },
-  'how-to': {
-    patterns: [
-      /\bhow\s+(do|can|to|should)\s+I?\b/i,
-      /\bsteps?\s+(to|for)\b/i,
-      /\bwhat('s| is)\s+the\s+(command|way|process)\b/i,
-    ],
-    keywords: ['how to', 'steps', 'command', 'run', 'deploy', 'install', 'configure', 'setup'],
-    biases: {
-      scope_boost: { personal: 1.2, team: 1.0, department: 0.8, organization: 0.7 },
-      type_boost: { command: 1.3, pattern: 1.2, error_fix: 1.0, conversation: 0.6 },
-      recency_boost: 1.05,
-    },
-  },
-  review: {
-    patterns: [
-      /\b(review|feedback|improve|quality|best\s+practice)\b/i,
-      /\bis\s+this\s+(good|correct|right)\b/i,
-    ],
-    keywords: ['review', 'feedback', 'quality', 'improve', 'best practice', 'convention'],
-    biases: {
-      scope_boost: { personal: 0.9, team: 1.2, department: 1.0, organization: 0.8 },
-      type_boost: { preference: 1.3, pattern: 1.2, code_pattern: 1.2, decision: 1.0 },
-      recency_boost: 1.0,
-    },
-  },
-  security: {
-    patterns: [
-      /\b(security|vulnerab|exploit|attack|auth|cve|injection|xss)\b/i,
-      /\bsecure\b/i,
-    ],
-    keywords: ['security', 'vulnerability', 'exploit', 'attack', 'authentication', 'authorization', 'cve', 'injection', 'xss', 'csrf'],
-    biases: {
-      scope_boost: { personal: 0.8, team: 1.0, department: 1.2, organization: 1.0 },
-      type_boost: { error_fix: 1.3, decision: 1.2, pattern: 1.0, conversation: 0.5 },
-      recency_boost: 1.1,
-    },
-  },
-  general: {
-    patterns: [],
-    keywords: [],
-    biases: {
-      scope_boost: { personal: 1.0, team: 1.0, department: 1.0, organization: 1.0 },
-      type_boost: {},
-      recency_boost: 1.0,
-    },
-  },
-};
-
-/**
- * Detect the intent of a query using regex patterns and keyword scoring.
- * No LLM call — fast and deterministic.
- */
-export function detectIntent(query: string): IntentResult {
-  const lower = query.toLowerCase();
-  let bestIntent = 'general';
-  let bestScore = 0;
-
-  for (const [name, def] of Object.entries(INTENTS)) {
-    if (name === 'general') continue;
-
-    let score = 0;
-
-    // Regex pattern matches (high weight)
-    for (const pattern of def.patterns) {
-      if (pattern.test(query)) score += 2;
-    }
-
-    // Keyword matches (lower weight)
-    for (const keyword of def.keywords) {
-      if (lower.includes(keyword)) score += 1;
-    }
-
-    if (score > bestScore) {
-      bestScore = score;
-      bestIntent = name;
-    }
-  }
-
-  return {
-    intent: bestIntent,
-    confidence: bestScore > 0 ? Math.min(1.0, bestScore / 6) : 0.5,
-    biases: INTENTS[bestIntent].biases,
-  };
-}
-```
-
-- [ ] **Step 4: Run tests to verify they pass**
-
-Run: `npx vitest run tests/lib/cortex/retrieval/intent.test.ts`
-Expected: PASS (7 tests)
-
-- [ ] **Step 5: Commit**
-
-```bash
-git add src/lib/cortex/retrieval/intent.ts tests/lib/cortex/retrieval/intent.test.ts
-git commit -m "feat(cortex): add intent detection for context assembly"
-```
-
----
-
-### Task 2: Weight computation
-
-**Files:**
-- Create: `src/lib/cortex/retrieval/weight.ts`
-- Create: `tests/lib/cortex/retrieval/weight.test.ts`
-
-- [ ] **Step 1: Write failing tests**
-
-```typescript
-// tests/lib/cortex/retrieval/weight.test.ts
-import { describe, it, expect } from 'vitest';
-import { computeScopeWeight } from '@/lib/cortex/retrieval/weight';
-import type { IntentBiases } from '@/lib/cortex/retrieval/intent';
-
-describe('computeScopeWeight', () => {
-  const neutralBiases: IntentBiases = {
-    scope_boost: { personal: 1.0, team: 1.0, department: 1.0, organization: 1.0 },
-    type_boost: {},
-    recency_boost: 1.0,
-  };
-
-  it('returns 1.0 for self (distance 0)', () => {
-    const weight = computeScopeWeight({
-      graphProximity: 1.0,  // 1/(1+0)
-      scopeLevel: 'personal',
-      intentBiases: neutralBiases,
-      authorityFactor: 1.0,
-    });
-    expect(weight).toBeCloseTo(1.0);
-  });
-
-  it('decreases with graph distance', () => {
-    const close = computeScopeWeight({
-      graphProximity: 0.5,  // distance 1
-      scopeLevel: 'team',
-      intentBiases: neutralBiases,
-      authorityFactor: 1.0,
-    });
-    const far = computeScopeWeight({
-      graphProximity: 0.25,  // distance 3
-      scopeLevel: 'organization',
-      intentBiases: neutralBiases,
-      authorityFactor: 1.0,
-    });
-    expect(close).toBeGreaterThan(far);
-  });
-
-  it('is boosted by intent biases', () => {
-    const debugBiases: IntentBiases = {
-      scope_boost: { personal: 1.2, team: 0.8 },
-      type_boost: {},
-      recency_boost: 1.0,
-    };
-    const personal = computeScopeWeight({
-      graphProximity: 0.5,
-      scopeLevel: 'personal',
-      intentBiases: debugBiases,
-      authorityFactor: 1.0,
-    });
-    const team = computeScopeWeight({
-      graphProximity: 0.5,
-      scopeLevel: 'team',
-      intentBiases: debugBiases,
-      authorityFactor: 1.0,
-    });
-    expect(personal).toBeGreaterThan(team);
-  });
-
-  it('is boosted by authority factor', () => {
-    const low = computeScopeWeight({
-      graphProximity: 0.5,
-      scopeLevel: 'team',
-      intentBiases: neutralBiases,
-      authorityFactor: 1.0,
-    });
-    const high = computeScopeWeight({
-      graphProximity: 0.5,
-      scopeLevel: 'team',
-      intentBiases: neutralBiases,
-      authorityFactor: 1.2,
-    });
-    expect(high).toBeGreaterThan(low);
-  });
-
-  it('never returns negative', () => {
-    const weight = computeScopeWeight({
-      graphProximity: 0,
-      scopeLevel: 'organization',
-      intentBiases: neutralBiases,
-      authorityFactor: 1.0,
-    });
-    expect(weight).toBeGreaterThanOrEqual(0);
-  });
-});
-```
-
-- [ ] **Step 2: Run tests to verify they fail**
-
-Run: `npx vitest run tests/lib/cortex/retrieval/weight.test.ts`
-
-- [ ] **Step 3: Implement weight computation**
-
-```typescript
-// src/lib/cortex/retrieval/weight.ts
-import type { IntentBiases } from './intent';
-import type { ScopeLevel } from '../knowledge/types';
-
-export interface ScopeWeightInput {
-  graphProximity: number;    // 0-1, from EntityGraph.proximity()
-  scopeLevel: ScopeLevel | string;
-  intentBiases: IntentBiases;
-  authorityFactor: number;   // 1.0 default, higher for experts/docs
-}
-
-/**
- * Compute the retrieval weight for a knowledge scope.
- *
- * weight = graphProximity × intentBias × authorityFactor
- *
- * Per spec: weight(scope) = graph_proximity × intent_bias × freshness_bonus × authority
- * Freshness is applied per-result in the fusion stage, not per-scope.
- */
-export function computeScopeWeight(input: ScopeWeightInput): number {
-  const { graphProximity, scopeLevel, intentBiases, authorityFactor } = input;
-
-  const intentBias = intentBiases.scope_boost[scopeLevel] ?? 1.0;
-
-  return Math.max(0, graphProximity * intentBias * authorityFactor);
-}
-
-/**
- * Compute per-result type boost from intent biases.
- */
-export function computeTypeBoost(knowledgeType: string, intentBiases: IntentBiases): number {
-  return intentBiases.type_boost[knowledgeType] ?? 1.0;
-}
-
-/**
- * Authority factor for a source based on role and expertise.
- *
- * role_boost: 0.0 member, 0.1 lead, 0.15 senior/principal, 0.2 director+
- * expertise_weight: EXPERT_IN edge weight (0-1)
- * Documents get 1.2 base authority.
- */
-export function computeAuthority(params: {
-  role?: string;
-  expertiseWeight?: number;
-  isDocument?: boolean;
-}): number {
-  const { role, expertiseWeight = 0, isDocument = false } = params;
-
-  if (isDocument) return 1.2;
-
-  const roleBoosts: Record<string, number> = {
-    member: 0.0,
-    lead: 0.1,
-    senior: 0.15,
-    principal: 0.15,
-    director: 0.2,
-    vp: 0.2,
-    cto: 0.2,
-  };
-
-  const roleBoost = roleBoosts[role?.toLowerCase() ?? 'member'] ?? 0.0;
-  return Math.max(1.0, roleBoost + expertiseWeight);
-}
-```
-
-- [ ] **Step 4: Run tests to verify they pass**
-
-Run: `npx vitest run tests/lib/cortex/retrieval/weight.test.ts`
-Expected: PASS (5 tests)
-
-- [ ] **Step 5: Commit**
-
-```bash
-git add src/lib/cortex/retrieval/weight.ts tests/lib/cortex/retrieval/weight.test.ts
-git commit -m "feat(cortex): add scope weight computation for context assembly"
-```
-
----
-
-## Chunk 2: Conflict Detection and Context Formatting
-
-### Task 3: Conflict detection
-
-**Files:**
-- Create: `src/lib/cortex/retrieval/conflict.ts`
-- Create: `tests/lib/cortex/retrieval/conflict.test.ts`
-
-- [ ] **Step 1: Write failing tests**
-
-```typescript
-// tests/lib/cortex/retrieval/conflict.test.ts
-import { describe, it, expect } from 'vitest';
-import { detectConflicts } from '@/lib/cortex/retrieval/conflict';
-import type { ScoredKnowledge } from '@/lib/cortex/knowledge/types';
-
-function makeResult(overrides: Partial<ScoredKnowledge> = {}): ScoredKnowledge {
-  return {
-    id: 'r1', vector: [], text: 'test', type: 'decision', layer: 'personal',
-    workspace_id: null, session_id: null, agent_type: 'claude',
-    project_path: null, file_refs: [], confidence: 0.8,
-    created: new Date().toISOString(), source_timestamp: new Date().toISOString(),
-    stale_score: 0, access_count: 0, last_accessed: null, metadata: {},
-    relevance_score: 0.9, similarity: 0.9,
-    contradiction_refs: [], ...overrides,
-  };
-}
-
-describe('detectConflicts', () => {
-  it('returns no conflicts when no contradiction_refs', () => {
-    const results = [makeResult({ id: 'a' }), makeResult({ id: 'b' })];
-    const conflicts = detectConflicts(results);
-    expect(conflicts).toHaveLength(0);
-  });
-
-  it('detects conflict between two results', () => {
-    const results = [
-      makeResult({ id: 'a', text: 'use pool size 50', contradiction_refs: ['b'] }),
-      makeResult({ id: 'b', text: 'scale horizontally', contradiction_refs: ['a'] }),
-    ];
-    const conflicts = detectConflicts(results);
-    expect(conflicts).toHaveLength(1);
-    expect(conflicts[0].unitA.id).toBe('a');
-    expect(conflicts[0].unitB.id).toBe('b');
-  });
-
-  it('ignores contradiction_refs pointing to results not in the set', () => {
-    const results = [
-      makeResult({ id: 'a', contradiction_refs: ['z'] }),  // z is not in results
-    ];
-    const conflicts = detectConflicts(results);
-    expect(conflicts).toHaveLength(0);
-  });
-
-  it('deduplicates symmetric conflicts', () => {
-    // A contradicts B and B contradicts A should produce one conflict, not two
-    const results = [
-      makeResult({ id: 'a', contradiction_refs: ['b'] }),
-      makeResult({ id: 'b', contradiction_refs: ['a'] }),
-    ];
-    const conflicts = detectConflicts(results);
-    expect(conflicts).toHaveLength(1);
-  });
-});
-```
-
-- [ ] **Step 2: Implement conflict detection**
-
-```typescript
-// src/lib/cortex/retrieval/conflict.ts
-import type { ScoredKnowledge } from '../knowledge/types';
-
-export interface ConflictPair {
-  unitA: ScoredKnowledge;
-  unitB: ScoredKnowledge;
-}
-
-/**
- * Detect conflicts among search results by checking contradiction_refs.
- * Returns deduplicated conflict pairs (A↔B counted once, not twice).
- */
-export function detectConflicts(results: ScoredKnowledge[]): ConflictPair[] {
-  const resultMap = new Map(results.map(r => [r.id, r]));
-  const seen = new Set<string>();
-  const conflicts: ConflictPair[] = [];
-
-  for (const result of results) {
-    const refs = result.contradiction_refs ?? [];
-    for (const refId of refs) {
-      const other = resultMap.get(refId);
-      if (!other) continue;
-
-      const key = [result.id, refId].sort().join('|');
-      if (seen.has(key)) continue;
-      seen.add(key);
-
-      conflicts.push({ unitA: result, unitB: other });
-    }
-  }
-
-  return conflicts;
-}
-```
-
-- [ ] **Step 3: Run tests, commit**
-
-Run: `npx vitest run tests/lib/cortex/retrieval/conflict.test.ts`
-
-```bash
-git add src/lib/cortex/retrieval/conflict.ts tests/lib/cortex/retrieval/conflict.test.ts
-git commit -m "feat(cortex): add conflict detection for search results"
-```
-
----
-
-### Task 4: Context formatter
-
-**Files:**
-- Create: `src/lib/cortex/retrieval/formatter.ts`
-- Create: `tests/lib/cortex/retrieval/formatter.test.ts`
-
-- [ ] **Step 1: Write failing tests**
-
-```typescript
-// tests/lib/cortex/retrieval/formatter.test.ts
-import { describe, it, expect } from 'vitest';
-import { formatContext } from '@/lib/cortex/retrieval/formatter';
-import type { ScoredKnowledge } from '@/lib/cortex/knowledge/types';
-import type { ConflictPair } from '@/lib/cortex/retrieval/conflict';
-
-function makeResult(overrides: Partial<ScoredKnowledge> = {}): ScoredKnowledge {
-  return {
-    id: 'r1', vector: [], text: 'test knowledge', type: 'decision', layer: 'personal',
-    workspace_id: null, session_id: null, agent_type: 'claude',
-    project_path: null, file_refs: [], confidence: 0.8,
-    created: '2026-03-15T00:00:00.000Z', source_timestamp: '2026-03-15T00:00:00.000Z',
-    stale_score: 0, access_count: 5, last_accessed: null, metadata: {},
-    relevance_score: 0.9, similarity: 0.9,
-    ...overrides,
-  };
-}
-
-describe('formatContext', () => {
-  it('wraps results in cortex-context tags', () => {
-    const output = formatContext([makeResult()], []);
-    expect(output).toContain('<cortex-context>');
-    expect(output).toContain('</cortex-context>');
-  });
-
-  it('includes type labels and dates', () => {
-    const output = formatContext([makeResult({ type: 'error_fix', source_timestamp: '2026-03-10T00:00:00.000Z' })], []);
-    expect(output).toContain('[Error Fix]');
-    expect(output).toContain('2026-03-10');
-  });
-
-  it('includes source attribution when origin is present', () => {
-    const output = formatContext([makeResult({
-      origin: { source_type: 'conversation', source_ref: 'sess-1', creator_entity_id: 'person-alice' },
-    })], []);
-    expect(output).toContain('person-alice');
-  });
-
-  it('includes conflict callout when conflicts exist', () => {
-    const a = makeResult({ id: 'a', text: 'use pool size 50' });
-    const b = makeResult({ id: 'b', text: 'scale horizontally' });
-    const conflicts: ConflictPair[] = [{ unitA: a, unitB: b }];
-    const output = formatContext([a, b], conflicts);
-    expect(output).toContain('Conflicting');
-  });
-
-  it('respects max token budget', () => {
-    const bigResults = Array.from({ length: 20 }, (_, i) =>
-      makeResult({ id: `r${i}`, text: 'x'.repeat(500) })
-    );
-    const output = formatContext(bigResults, [], { maxTokens: 500 });
-    // Should not include all 20 results (would be ~2500 tokens)
-    expect(output.length).toBeLessThan(3000);
-  });
-
-  it('returns empty string when no results', () => {
-    expect(formatContext([], [])).toBe('');
-  });
-});
-```
-
-- [ ] **Step 2: Implement context formatter**
-
-```typescript
-// src/lib/cortex/retrieval/formatter.ts
-import type { ScoredKnowledge } from '../knowledge/types';
-import type { ConflictPair } from './conflict';
-
-const TYPE_LABELS: Record<string, string> = {
-  decision: 'Decision', pattern: 'Pattern', preference: 'Preference',
-  error_fix: 'Error Fix', context: 'Context', code_pattern: 'Code',
-  command: 'Command', conversation: 'Conversation', summary: 'Summary',
-};
-
-export interface FormatOptions {
-  maxTokens?: number;
-}
-
-/**
- * Format search results + conflicts as annotated <cortex-context> for RAG injection.
- */
-export function formatContext(
-  results: ScoredKnowledge[],
-  conflicts: ConflictPair[],
-  options: FormatOptions = {},
-): string {
-  if (results.length === 0) return '';
-
-  const maxTokens = options.maxTokens ?? 1500;
-  const entries: string[] = [];
-  let tokens = 20;  // overhead for tags
-
-  // Format each result with attribution
-  for (const unit of results) {
-    const label = TYPE_LABELS[unit.type] || unit.type;
-    const date = (unit.source_timestamp || '').slice(0, 10);
-    const creator = unit.origin?.creator_entity_id ?? '';
-    const sourceInfo = creator ? ` (${creator})` : '';
-
-    let entry = `[${label}] ${date}${sourceInfo}:\n  ${unit.text}`;
-
-    const entryTokens = Math.ceil(entry.length / 4);
-    if (tokens + entryTokens > maxTokens) break;
-
-    entries.push(entry);
-    tokens += entryTokens;
-  }
-
-  if (entries.length === 0) return '';
-
-  // Build conflict section
-  let conflictSection = '';
-  if (conflicts.length > 0) {
-    const conflictLines = conflicts.map(c =>
-      `  - "${c.unitA.text.slice(0, 80)}..." vs "${c.unitB.text.slice(0, 80)}..."`
-    );
-    conflictSection = `\nConflicting perspectives (${conflicts.length}):\n${conflictLines.join('\n')}\n`;
-  }
-
-  const sourceCount = entries.length;
-  const header = `Relevant knowledge (${sourceCount} source${sourceCount > 1 ? 's' : ''}${conflicts.length > 0 ? `, ${conflicts.length} conflict${conflicts.length > 1 ? 's' : ''}` : ''}):`;
-
-  return [
-    '<cortex-context>',
-    header,
-    '',
-    ...entries,
-    conflictSection,
-    '</cortex-context>',
-  ].join('\n');
-}
-```
-
-- [ ] **Step 3: Run tests, commit**
-
-Run: `npx vitest run tests/lib/cortex/retrieval/formatter.test.ts`
-
-```bash
-git add src/lib/cortex/retrieval/formatter.ts tests/lib/cortex/retrieval/formatter.test.ts
-git commit -m "feat(cortex): add context formatter for RAG injection"
-```
-
----
-
-## Chunk 3: Context Assembly Engine
-
-### Task 5: ContextEngine — the 6-stage pipeline
-
-**Files:**
-- Create: `src/lib/cortex/retrieval/context-engine.ts`
-- Create: `tests/lib/cortex/retrieval/context-engine.test.ts`
-
-- [ ] **Step 1: Write failing tests**
-
-```typescript
-// tests/lib/cortex/retrieval/context-engine.test.ts
-import { describe, it, expect, vi, beforeEach } from 'vitest';
-import { ContextEngine } from '@/lib/cortex/retrieval/context-engine';
-
-// Create mocks for dependencies
-const mockStore = {
-  search: vi.fn().mockResolvedValue([]),
-};
-
-const mockGraph = {
-  proximity: vi.fn().mockReturnValue(0.5),
-  neighborhood: vi.fn().mockReturnValue([]),
-  getEntity: vi.fn().mockReturnValue(null),
-};
-
-const mockResolver = {
-  extractEntities: vi.fn().mockReturnValue([]),
-};
-
-const mockEmbedding = {
-  embed: vi.fn().mockResolvedValue([[0.1, 0.2, 0.3]]),
-  dimensions: 3,
-  name: 'mock',
-  init: vi.fn(),
-};
-
-describe('ContextEngine', () => {
-  let engine: ContextEngine;
-
-  beforeEach(() => {
-    vi.clearAllMocks();
-    engine = new ContextEngine({
-      store: mockStore as any,
-      graph: mockGraph as any,
-      resolver: mockResolver as any,
-      embedding: mockEmbedding as any,
-      requesterId: 'person-alice',
-    });
-  });
-
-  it('returns empty context for empty results', async () => {
-    const result = await engine.assemble('some query');
-    expect(result.results).toHaveLength(0);
-    expect(result.context).toBe('');
-  });
-
-  it('calls embedding.embed with the query', async () => {
-    await engine.assemble('test query');
-    expect(mockEmbedding.embed).toHaveBeenCalledWith(['test query']);
-  });
-
-  it('detects intent from the query', async () => {
-    const result = await engine.assemble('why does auth throw an error?');
-    expect(result.intent.intent).toBe('debugging');
-  });
-
-  it('extracts entities from the query', async () => {
-    mockResolver.extractEntities.mockReturnValue([
-      { entity: { id: 'system-auth', type: 'system', name: 'Auth' }, confidence: 0.9, method: 'alias' },
-    ]);
-    const result = await engine.assemble('fix the auth service');
-    expect(mockResolver.extractEntities).toHaveBeenCalledWith('fix the auth service');
-    expect(result.entities).toHaveLength(1);
-  });
-
-  it('searches store with embedded query vector', async () => {
-    mockStore.search.mockResolvedValue([{
-      id: 'k1', text: 'test knowledge', type: 'decision', layer: 'personal',
-      confidence: 0.8, stale_score: 0, created: new Date().toISOString(),
-      source_timestamp: new Date().toISOString(), evidence_score: 0.7,
-      contradiction_refs: [], _distance: 0.2,
-      workspace_id: null, session_id: null, agent_type: 'claude',
-      project_path: null, file_refs: [], access_count: 0, last_accessed: null,
-      metadata: {},
-    }]);
-
-    const result = await engine.assemble('test query');
-    expect(result.results.length).toBeGreaterThanOrEqual(1);
-    expect(result.context).toContain('<cortex-context>');
-  });
-
-  it('completes within performance budget', async () => {
-    const start = Date.now();
-    await engine.assemble('test query');
-    const elapsed = Date.now() - start;
-    expect(elapsed).toBeLessThan(500);  // generous for CI, target is 150ms
-  });
-});
-```
-
-- [ ] **Step 2: Implement ContextEngine**
-
-```typescript
-// src/lib/cortex/retrieval/context-engine.ts
-import type { CortexStore } from '../store';
-import type { EntityGraph } from '../graph/entity-graph';
-import type { EntityResolver, ResolvedEntity } from '../graph/resolver';
-import type { EmbeddingProvider } from '../embeddings';
-import type { ScoredKnowledge } from '../knowledge/types';
-import { detectIntent } from './intent';
-import type { IntentResult } from './intent';
-import { computeScopeWeight, computeTypeBoost } from './weight';
-import { detectConflicts } from './conflict';
-import type { ConflictPair } from './conflict';
-import { formatContext } from './formatter';
-import { computeRelevanceScore } from './scoring';
-
-export interface ContextEngineDeps {
-  store: CortexStore;
-  graph: EntityGraph;
-  resolver: EntityResolver;
-  embedding: EmbeddingProvider;
-  requesterId: string;  // entity ID of the person making the query
-}
-
-export interface AssemblyResult {
-  results: ScoredKnowledge[];
-  conflicts: ConflictPair[];
-  context: string;           // formatted <cortex-context> string
-  intent: IntentResult;
-  entities: ResolvedEntity[];
-  timing: {
-    intentMs: number;
-    entityMs: number;
-    searchMs: number;
-    totalMs: number;
-  };
-}
-
-interface SearchSource {
-  layerKey: string;
-  weight: number;
-  limit: number;
-}
-
-const DEFAULT_LAYERS = ['personal', 'workspace', 'team'] as const;
-const SEARCH_TIMEOUT_MS = 100;
-
-export class ContextEngine {
-  constructor(private deps: ContextEngineDeps) {}
-
-  async assemble(query: string, options: { limit?: number; workspaceId?: number | null; maxTokens?: number } = {}): Promise<AssemblyResult> {
-    const totalStart = Date.now();
-    const { limit = 5, workspaceId = null, maxTokens = 1500 } = options;
-
-    // Stage 1: Intent Detection
-    const intentStart = Date.now();
-    const intent = detectIntent(query);
-    const intentMs = Date.now() - intentStart;
-
-    // Stage 2: Entity Resolution
-    const entityStart = Date.now();
-    const entities = this.deps.resolver.extractEntities(query);
-    const entityMs = Date.now() - entityStart;
-
-    // Embed the query
-    const [queryVector] = await this.deps.embedding.embed([query]);
-
-    // Stage 3: Weight Computation
-    const sources = this.computeSourceWeights(intent, workspaceId);
-
-    // Stage 4: Parallel Multi-Source Search
-    const searchStart = Date.now();
-    const allResults = await this.parallelSearch(queryVector, sources, limit);
-    const searchMs = Date.now() - searchStart;
-
-    // Stage 5: Fusion + Re-Ranking
-    const fused = this.fuseAndRank(allResults, intent, limit);
-
-    // Stage 6: Conflict Detection + Formatting
-    const conflicts = detectConflicts(fused);
-    const context = formatContext(fused, conflicts, { maxTokens });
-
-    return {
-      results: fused,
-      conflicts,
-      context,
-      intent,
-      entities,
-      timing: {
-        intentMs,
-        entityMs,
-        searchMs,
-        totalMs: Date.now() - totalStart,
-      },
-    };
-  }
-
-  private computeSourceWeights(intent: IntentResult, workspaceId: number | null): SearchSource[] {
-    const sources: SearchSource[] = [];
-
-    for (const layer of DEFAULT_LAYERS) {
-      const layerKey = layer === 'workspace' && workspaceId
-        ? `workspace/${workspaceId}` : layer;
-
-      // Map v1 layer to scope level for intent bias lookup
-      const scopeLevel = layer === 'personal' ? 'personal'
-        : layer === 'workspace' ? 'team' : 'organization';
-
-      // Use graph proximity if available, else fall back to fixed weights
-      let graphProximity: number;
-      try {
-        // For personal, distance is 0; for workspace, 1; for team, 2
-        const layerEntity = layer === 'personal' ? this.deps.requesterId
-          : layer === 'workspace' ? 'team-default' : 'organization-default';
-        graphProximity = this.deps.graph.proximity(this.deps.requesterId, layerEntity);
-      } catch {
-        // Graph not populated, use fallback
-        graphProximity = layer === 'personal' ? 1.0 : layer === 'workspace' ? 0.5 : 0.33;
-      }
-
-      // If graph returns 0 (unreachable/not found), use fallback
-      if (graphProximity === 0) {
-        graphProximity = layer === 'personal' ? 1.0 : layer === 'workspace' ? 0.5 : 0.33;
-      }
-
-      const weight = computeScopeWeight({
-        graphProximity,
-        scopeLevel,
-        intentBiases: intent.biases,
-        authorityFactor: 1.0,
-      });
-
-      sources.push({
-        layerKey,
-        weight,
-        limit: Math.max(3, Math.round(weight * 10)),  // proportional slots
-      });
-    }
-
-    return sources.sort((a, b) => b.weight - a.weight);
-  }
-
-  private async parallelSearch(
-    queryVector: number[],
-    sources: SearchSource[],
-    limit: number,
-  ): Promise<Array<ScoredKnowledge & { sourceWeight: number }>> {
-    const searchPromises = sources.map(async (source) => {
-      try {
-        const results = await Promise.race([
-          this.deps.store.search(source.layerKey, queryVector, source.limit),
-          new Promise<never>((_, reject) =>
-            setTimeout(() => reject(new Error('timeout')), SEARCH_TIMEOUT_MS)
-          ),
-        ]);
-
-        return results.map(unit => {
-          const similarity = 1 - ((unit as any)._distance ?? 0);
-          return {
-            ...unit,
-            similarity,
-            relevance_score: 0,  // computed in fusion
-            sourceWeight: source.weight,
-          } as ScoredKnowledge & { sourceWeight: number };
-        });
-      } catch {
-        return [];  // source failed or timed out
-      }
-    });
-
-    const settled = await Promise.allSettled(searchPromises);
-    const allResults: Array<ScoredKnowledge & { sourceWeight: number }> = [];
-
-    for (const result of settled) {
-      if (result.status === 'fulfilled') {
-        allResults.push(...result.value);
-      }
-    }
-
-    return allResults;
-  }
-
-  private fuseAndRank(
-    results: Array<ScoredKnowledge & { sourceWeight: number }>,
-    intent: IntentResult,
-    limit: number,
-  ): ScoredKnowledge[] {
-    // Score each result
-    for (const result of results) {
-      const typeBoost = computeTypeBoost(result.type, intent.biases);
-      const recencyBoost = intent.biases.recency_boost;
-
-      result.relevance_score = computeRelevanceScore({
-        similarity: result.similarity,
-        confidence: result.confidence,
-        stale_score: result.stale_score,
-        created: result.created,
-        evidence_score: result.evidence_score,
-      }) * result.sourceWeight * typeBoost * recencyBoost;
-    }
-
-    // Deduplicate: cosine > 0.9 between results (approximate via text similarity)
-    const deduped = this.deduplicateResults(results);
-
-    // Sort and take top K
-    deduped.sort((a, b) => b.relevance_score - a.relevance_score);
-    return deduped.slice(0, limit);
-  }
-
-  private deduplicateResults(results: ScoredKnowledge[]): ScoredKnowledge[] {
-    const kept: ScoredKnowledge[] = [];
-    const seenTexts = new Set<string>();
-
-    // Sort by score first so we keep the better-scored version
-    results.sort((a, b) => b.relevance_score - a.relevance_score);
-
-    for (const result of results) {
-      // Simple text-based dedup: normalize and check prefix overlap
-      const normalized = result.text.slice(0, 200).toLowerCase().trim();
-      if (seenTexts.has(normalized)) continue;
-
-      // Check against existing kept items for high text overlap
-      let isDupe = false;
-      for (const existing of kept) {
-        if (result.id === existing.id) { isDupe = true; break; }
-        // If texts share >80% of content, consider duplicate
-        const existNorm = existing.text.slice(0, 200).toLowerCase().trim();
-        if (normalized === existNorm) { isDupe = true; break; }
-      }
-
-      if (!isDupe) {
-        seenTexts.add(normalized);
-        kept.push(result);
-      }
-    }
-
-    return kept;
-  }
-}
-```
-
-- [ ] **Step 3: Run tests to verify they pass**
-
-Run: `npx vitest run tests/lib/cortex/retrieval/context-engine.test.ts`
-Expected: PASS (6 tests)
-
-- [ ] **Step 4: Commit**
-
-```bash
-git add src/lib/cortex/retrieval/context-engine.ts tests/lib/cortex/retrieval/context-engine.test.ts
-git commit -m "feat(cortex): add ContextEngine with 6-stage retrieval pipeline"
-```
-
----
-
-## Chunk 4: API Endpoint and Hook Integration
-
-### Task 6: Context assembly API endpoint
-
-**Files:**
-- Create: `src/app/api/cortex/context/route.ts`
-
-- [ ] **Step 1: Create the endpoint**
-
-```typescript
-// src/app/api/cortex/context/route.ts
-import { NextResponse } from 'next/server';
-import type { NextRequest } from 'next/server';
-import { getAuthUser, withUser } from '@/lib/auth';
-import { getCortex, isCortexAvailable } from '@/lib/cortex';
-import { ContextEngine } from '@/lib/cortex/retrieval/context-engine';
-import { EntityResolver } from '@/lib/cortex/graph/resolver';
-import { slugify } from '@/lib/cortex/graph/types';
-
-export async function GET(request: NextRequest) {
-  const user = getAuthUser(request);
-  return withUser(user, async () => {
-    if (!isCortexAvailable()) {
-      return NextResponse.json({ error: 'Cortex unavailable' }, { status: 403 });
-    }
-    const cortex = await getCortex();
-    if (!cortex) return NextResponse.json({ results: [], context: '' });
-
-    const url = new URL(request.url);
-    const query = url.searchParams.get('q') || '';
-    const limit = parseInt(url.searchParams.get('limit') || '5', 10);
-    const workspaceId = url.searchParams.get('workspace_id');
-    const maxTokens = parseInt(url.searchParams.get('max_tokens') || '1500', 10);
-
-    if (!query || query.length < 3) {
-      return NextResponse.json({ results: [], context: '' });
-    }
-
-    const resolver = new EntityResolver(cortex.graph);
-    const requesterId = `person-${slugify(user)}`;
-
-    const engine = new ContextEngine({
-      store: cortex.store,
-      graph: cortex.graph,
-      resolver,
-      embedding: cortex.embedding,
-      requesterId,
-    });
-
-    const result = await engine.assemble(query, {
-      limit,
-      workspaceId: workspaceId ? parseInt(workspaceId, 10) : null,
-      maxTokens,
-    });
-
-    return NextResponse.json({
-      results: result.results.map(r => ({ ...r, vector: undefined })),  // strip vectors
-      context: result.context,
-      intent: result.intent,
-      conflicts: result.conflicts.length,
-      timing: result.timing,
-    });
-  });
-}
-```
-
-- [ ] **Step 2: Commit**
-
-```bash
-git add src/app/api/cortex/context/route.ts
-git commit -m "feat(cortex): add context assembly API endpoint"
-```
-
----
-
-### Task 7: Update RAG hook to use context assembly endpoint
-
-**Files:**
-- Modify: `bin/cortex-hook.js`
-
-- [ ] **Step 1: Read current cortex-hook.js**
-
-Read the file to understand the current flow: query → /api/cortex/search → format results → output.
-
-- [ ] **Step 2: Update to call context assembly endpoint**
-
-Change the URL from `/api/cortex/search/?q=...` to `/api/cortex/context/?q=...`. The new endpoint returns `{ context, results, intent, conflicts, timing }` — the `context` field is already pre-formatted as `<cortex-context>`, so the hook can output it directly instead of doing its own formatting.
-
-Key changes:
-1. URL: `/api/cortex/search/` → `/api/cortex/context/`
-2. Response handling: use `parsed.context` directly instead of formatting results manually
-3. Keep the fallback: if the new endpoint returns empty or fails, fall back to old behavior
-
-```javascript
-// Replace the response handling section:
-const parsed = JSON.parse(body);
-
-// New: context is pre-formatted by the Context Assembly Engine
-if (parsed.context) {
-  const output = JSON.stringify({
-    hookSpecificOutput: {
-      hookEventName: 'UserPromptSubmit',
-      additionalContext: parsed.context,
-    },
-  });
-  process.stdout.write(output);
-  process.exit(0);
-}
-
-// Fallback: old-style results formatting (if context endpoint not available)
-const results = parsed.results;
-if (!results || results.length === 0) process.exit(0);
-// ... existing formatting code stays as fallback ...
-```
-
-- [ ] **Step 3: Commit**
-
-```bash
-git add bin/cortex-hook.js
-git commit -m "feat(cortex): switch RAG hook to context assembly endpoint"
-```
-
----
-
-### Task 8: Integrate ContextEngine into CortexInstance
-
-**Files:**
-- Modify: `src/lib/cortex/index.ts`
-
-- [ ] **Step 1: Read current index.ts**
-
-- [ ] **Step 2: Add ContextEngine to CortexInstance**
-
-1. Import: `import { ContextEngine } from './retrieval/context-engine';` and `import { EntityResolver } from './graph/resolver';`
-2. Add `contextEngine?: ContextEngine` to `CortexInstance` interface (optional since it depends on graph)
-3. In `getCortex()`, after graph initialization, create the ContextEngine:
-
-```typescript
-const resolver = new EntityResolver(graph);
-const contextEngine = new ContextEngine({
-  store,
-  graph,
-  resolver,
-  embedding,
-  requesterId: 'person-default-user',  // default; overridden per-request in API
-});
-```
-
-4. Add to instance object: `contextEngine,`
-
-- [ ] **Step 3: Run full cortex test suite**
-
-Run: `npx vitest run tests/lib/cortex/`
-
-- [ ] **Step 4: Commit**
-
-```bash
-git add src/lib/cortex/index.ts
-git commit -m "feat(cortex): add ContextEngine to CortexInstance"
-```
-
----
-
-## Summary
-
-| Task | Component | Tests | Status |
-|------|-----------|-------|--------|
-| 1 | Intent detection | 7 | |
-| 2 | Weight computation | 5 | |
-| 3 | Conflict detection | 4 | |
-| 4 | Context formatter | 6 | |
-| 5 | ContextEngine (6-stage pipeline) | 6 | |
-| 6 | Context assembly API endpoint | — | |
-| 7 | RAG hook integration | — | |
-| 8 | CortexInstance integration | regression | |
-
-**Total: 8 tasks, ~28 new tests, 4 chunks**
-
-**Performance budget:** The ContextEngine targets <150ms total latency:
-- Intent detection: <5ms (regex, no LLM)
-- Entity resolution: <5ms (alias lookup)
-- Weight computation: <10ms (graph proximity, cached)
-- Parallel search: <100ms (concurrent vector search with 100ms timeout)
-- Fusion + formatting: <10ms
-
-**Key design decisions:**
-- ContextEngine sits ABOVE CortexSearch — doesn't replace it, wraps it
-- Graph proximity drives weights — but gracefully degrades to fixed weights when graph is empty
-- Deduplication uses text prefix comparison (not cosine between result vectors — that would require N² vector ops)
-- RAG hook uses pre-formatted `context` from the engine — no more client-side formatting
-- Old `/api/cortex/search` endpoint still works — new `/api/cortex/context` endpoint adds the intelligence layer
+# Cortex v2 — Pillar 3: Context Assembly Engine
+
+> **For agentic workers:** REQUIRED: Use superpowers:subagent-driven-development (if subagents available) or superpowers:executing-plans to implement this plan. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Replace the flat layer-iteration search with a 6-stage context assembly pipeline that detects intent, resolves entities, computes graph-aware weights, searches multiple scopes in parallel, fuses results with evidence scoring, and surfaces conflicts — all within 150ms.
+
+**Architecture:** A new `ContextEngine` class wraps the existing `CortexStore` (for low-level vector search) and `EntityGraph` (for graph distance/proximity). It does NOT replace `CortexSearch` — instead it provides the higher-level retrieval interface that the RAG hook calls. The existing `CortexSearch` remains for backward-compatible simple searches. The RAG hook (`cortex-hook.js`) switches from calling the search API to calling a new context-assembly API endpoint.
+
+**Tech Stack:** TypeScript, vitest, LanceDB, SQLite (entity graph)
+
+**Spec:** `docs/superpowers/specs/2026-03-14-cortex-v2-design.md` — Pillar 3
+
+**Depends on:** Pillar 1 (Entity Graph) + Pillar 2 (Knowledge Unit Evolution) — both completed
+
+---
+
+## File Structure
+
+```
+New files:
+├── src/lib/cortex/retrieval/intent.ts         — Intent detection (regex + keyword)
+├── src/lib/cortex/retrieval/weight.ts          — Weight computation (graph × intent × freshness × authority)
+├── src/lib/cortex/retrieval/conflict.ts        — Conflict detection in results
+├── src/lib/cortex/retrieval/formatter.ts       — Context formatting for RAG injection
+├── src/lib/cortex/retrieval/context-engine.ts  — Main 6-stage ContextEngine class
+├── src/app/api/cortex/context/route.ts         — API endpoint for context assembly
+
+Modified files:
+├── bin/cortex-hook.js                          — Switch to context-assembly endpoint
+
+Test files:
+├── tests/lib/cortex/retrieval/intent.test.ts
+├── tests/lib/cortex/retrieval/weight.test.ts
+├── tests/lib/cortex/retrieval/conflict.test.ts
+├── tests/lib/cortex/retrieval/formatter.test.ts
+├── tests/lib/cortex/retrieval/context-engine.test.ts
+```
+
+---
+
+## Chunk 1: Intent Detection and Weight Computation
+
+### Task 1: Intent detection
+
+**Files:**
+- Create: `src/lib/cortex/retrieval/intent.ts`
+- Create: `tests/lib/cortex/retrieval/intent.test.ts`
+
+- [ ] **Step 1: Write failing tests**
+
+```typescript
+// tests/lib/cortex/retrieval/intent.test.ts
+import { describe, it, expect } from 'vitest';
+import { detectIntent, INTENTS } from '@/lib/cortex/retrieval/intent';
+import type { IntentResult } from '@/lib/cortex/retrieval/intent';
+
+describe('detectIntent', () => {
+  it('detects debugging intent', () => {
+    const result = detectIntent('why does the auth service throw a timeout error?');
+    expect(result.intent).toBe('debugging');
+    expect(result.confidence).toBeGreaterThan(0.5);
+  });
+
+  it('detects architecture intent', () => {
+    const result = detectIntent('what architecture pattern should we use for the new service?');
+    expect(result.intent).toBe('architecture');
+  });
+
+  it('detects how-to intent', () => {
+    const result = detectIntent('how do I deploy this service to production?');
+    expect(result.intent).toBe('how-to');
+  });
+
+  it('detects security intent', () => {
+    const result = detectIntent('is there a vulnerability in our authentication flow?');
+    expect(result.intent).toBe('security');
+  });
+
+  it('defaults to general for ambiguous queries', () => {
+    const result = detectIntent('tell me about the project');
+    expect(result.intent).toBe('general');
+  });
+
+  it('returns bias config for the detected intent', () => {
+    const result = detectIntent('fix this bug in the login page');
+    expect(result.biases).toBeDefined();
+    expect(result.biases.scope_boost).toBeDefined();
+    expect(result.biases.type_boost).toBeDefined();
+  });
+
+  it('exports all intent definitions', () => {
+    expect(Object.keys(INTENTS)).toContain('debugging');
+    expect(Object.keys(INTENTS)).toContain('architecture');
+    expect(Object.keys(INTENTS)).toContain('general');
+    expect(Object.keys(INTENTS).length).toBe(8);
+  });
+});
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `npx vitest run tests/lib/cortex/retrieval/intent.test.ts`
+
+- [ ] **Step 3: Implement intent detection**
+
+```typescript
+// src/lib/cortex/retrieval/intent.ts
+
+export interface IntentBiases {
+  scope_boost: Record<string, number>;  // scope level → multiplier
+  type_boost: Record<string, number>;   // knowledge type → multiplier
+  recency_boost: number;                // extra recency multiplier
+}
+
+export interface IntentResult {
+  intent: string;
+  confidence: number;
+  biases: IntentBiases;
+}
+
+interface IntentDef {
+  patterns: RegExp[];
+  keywords: string[];
+  biases: IntentBiases;
+}
+
+export const INTENTS: Record<string, IntentDef> = {
+  debugging: {
+    patterns: [
+      /\b(error|bug|fix|crash|fail|broken|throw|exception|timeout|issue)\b/i,
+      /\bwhy\s+(does|is|did|do)\b/i,
+      /\bnot\s+work/i,
+    ],
+    keywords: ['error', 'bug', 'fix', 'debug', 'crash', 'fail', 'broken', 'throw', 'exception', 'timeout', 'issue', 'stack trace'],
+    biases: {
+      scope_boost: { personal: 1.2, team: 1.0, department: 0.9, organization: 0.8 },
+      type_boost: { error_fix: 1.3, pattern: 1.0, decision: 0.8, conversation: 0.7 },
+      recency_boost: 1.1,
+    },
+  },
+  architecture: {
+    patterns: [
+      /\b(architect|design|pattern|structure|approach)\b/i,
+      /\bshould\s+we\s+(use|adopt|switch|migrate)\b/i,
+    ],
+    keywords: ['architecture', 'design', 'pattern', 'structure', 'approach', 'decision', 'migration', 'refactor'],
+    biases: {
+      scope_boost: { personal: 0.9, team: 1.1, department: 1.2, organization: 1.0 },
+      type_boost: { decision: 1.5, pattern: 1.2, error_fix: 0.7, conversation: 0.5 },
+      recency_boost: 1.0,
+    },
+  },
+  onboarding: {
+    patterns: [
+      /\b(how\s+does|explain|what\s+is|overview|getting\s+started)\b/i,
+      /\bnew\s+to\b/i,
+    ],
+    keywords: ['explain', 'overview', 'introduction', 'getting started', 'onboarding', 'how does'],
+    biases: {
+      scope_boost: { personal: 0.7, team: 1.0, department: 1.1, organization: 1.2 },
+      type_boost: { pattern: 1.3, decision: 1.2, summary: 1.2, conversation: 0.5 },
+      recency_boost: 0.9,
+    },
+  },
+  policy: {
+    patterns: [
+      /\b(policy|compliance|regulation|standard|rule|requirement)\b/i,
+      /\ballowed\s+to\b/i,
+    ],
+    keywords: ['policy', 'compliance', 'standard', 'regulation', 'rule', 'requirement', 'allowed'],
+    biases: {
+      scope_boost: { personal: 0.6, team: 0.8, department: 1.0, organization: 1.3 },
+      type_boost: { decision: 1.5, preference: 1.2, pattern: 0.8, conversation: 0.3 },
+      recency_boost: 1.0,
+    },
+  },
+  'how-to': {
+    patterns: [
+      /\bhow\s+(do|can|to|should)\s+I?\b/i,
+      /\bsteps?\s+(to|for)\b/i,
+      /\bwhat('s| is)\s+the\s+(command|way|process)\b/i,
+    ],
+    keywords: ['how to', 'steps', 'command', 'run', 'deploy', 'install', 'configure', 'setup'],
+    biases: {
+      scope_boost: { personal: 1.2, team: 1.0, department: 0.8, organization: 0.7 },
+      type_boost: { command: 1.3, pattern: 1.2, error_fix: 1.0, conversation: 0.6 },
+      recency_boost: 1.05,
+    },
+  },
+  review: {
+    patterns: [
+      /\b(review|feedback|improve|quality|best\s+practice)\b/i,
+      /\bis\s+this\s+(good|correct|right)\b/i,
+    ],
+    keywords: ['review', 'feedback', 'quality', 'improve', 'best practice', 'convention'],
+    biases: {
+      scope_boost: { personal: 0.9, team: 1.2, department: 1.0, organization: 0.8 },
+      type_boost: { preference: 1.3, pattern: 1.2, code_pattern: 1.2, decision: 1.0 },
+      recency_boost: 1.0,
+    },
+  },
+  security: {
+    patterns: [
+      /\b(security|vulnerab|exploit|attack|auth|cve|injection|xss)\b/i,
+      /\bsecure\b/i,
+    ],
+    keywords: ['security', 'vulnerability', 'exploit', 'attack', 'authentication', 'authorization', 'cve', 'injection', 'xss', 'csrf'],
+    biases: {
+      scope_boost: { personal: 0.8, team: 1.0, department: 1.2, organization: 1.0 },
+      type_boost: { error_fix: 1.3, decision: 1.2, pattern: 1.0, conversation: 0.5 },
+      recency_boost: 1.1,
+    },
+  },
+  general: {
+    patterns: [],
+    keywords: [],
+    biases: {
+      scope_boost: { personal: 1.0, team: 1.0, department: 1.0, organization: 1.0 },
+      type_boost: {},
+      recency_boost: 1.0,
+    },
+  },
+};
+
+/**
+ * Detect the intent of a query using regex patterns and keyword scoring.
+ * No LLM call — fast and deterministic.
+ */
+export function detectIntent(query: string): IntentResult {
+  const lower = query.toLowerCase();
+  let bestIntent = 'general';
+  let bestScore = 0;
+
+  for (const [name, def] of Object.entries(INTENTS)) {
+    if (name === 'general') continue;
+
+    let score = 0;
+
+    // Regex pattern matches (high weight)
+    for (const pattern of def.patterns) {
+      if (pattern.test(query)) score += 2;
+    }
+
+    // Keyword matches (lower weight)
+    for (const keyword of def.keywords) {
+      if (lower.includes(keyword)) score += 1;
+    }
+
+    if (score > bestScore) {
+      bestScore = score;
+      bestIntent = name;
+    }
+  }
+
+  return {
+    intent: bestIntent,
+    confidence: bestScore > 0 ? Math.min(1.0, bestScore / 6) : 0.5,
+    biases: INTENTS[bestIntent].biases,
+  };
+}
+```
+
+- [ ] **Step 4: Run tests to verify they pass**
+
+Run: `npx vitest run tests/lib/cortex/retrieval/intent.test.ts`
+Expected: PASS (7 tests)
+
+- [ ] **Step 5: Commit**
+
+```bash
+git add src/lib/cortex/retrieval/intent.ts tests/lib/cortex/retrieval/intent.test.ts
+git commit -m "feat(cortex): add intent detection for context assembly"
+```
+
+---
+
+### Task 2: Weight computation
+
+**Files:**
+- Create: `src/lib/cortex/retrieval/weight.ts`
+- Create: `tests/lib/cortex/retrieval/weight.test.ts`
+
+- [ ] **Step 1: Write failing tests**
+
+```typescript
+// tests/lib/cortex/retrieval/weight.test.ts
+import { describe, it, expect } from 'vitest';
+import { computeScopeWeight } from '@/lib/cortex/retrieval/weight';
+import type { IntentBiases } from '@/lib/cortex/retrieval/intent';
+
+describe('computeScopeWeight', () => {
+  const neutralBiases: IntentBiases = {
+    scope_boost: { personal: 1.0, team: 1.0, department: 1.0, organization: 1.0 },
+    type_boost: {},
+    recency_boost: 1.0,
+  };
+
+  it('returns 1.0 for self (distance 0)', () => {
+    const weight = computeScopeWeight({
+      graphProximity: 1.0,  // 1/(1+0)
+      scopeLevel: 'personal',
+      intentBiases: neutralBiases,
+      authorityFactor: 1.0,
+    });
+    expect(weight).toBeCloseTo(1.0);
+  });
+
+  it('decreases with graph distance', () => {
+    const close = computeScopeWeight({
+      graphProximity: 0.5,  // distance 1
+      scopeLevel: 'team',
+      intentBiases: neutralBiases,
+      authorityFactor: 1.0,
+    });
+    const far = computeScopeWeight({
+      graphProximity: 0.25,  // distance 3
+      scopeLevel: 'organization',
+      intentBiases: neutralBiases,
+      authorityFactor: 1.0,
+    });
+    expect(close).toBeGreaterThan(far);
+  });
+
+  it('is boosted by intent biases', () => {
+    const debugBiases: IntentBiases = {
+      scope_boost: { personal: 1.2, team: 0.8 },
+      type_boost: {},
+      recency_boost: 1.0,
+    };
+    const personal = computeScopeWeight({
+      graphProximity: 0.5,
+      scopeLevel: 'personal',
+      intentBiases: debugBiases,
+      authorityFactor: 1.0,
+    });
+    const team = computeScopeWeight({
+      graphProximity: 0.5,
+      scopeLevel: 'team',
+      intentBiases: debugBiases,
+      authorityFactor: 1.0,
+    });
+    expect(personal).toBeGreaterThan(team);
+  });
+
+  it('is boosted by authority factor', () => {
+    const low = computeScopeWeight({
+      graphProximity: 0.5,
+      scopeLevel: 'team',
+      intentBiases: neutralBiases,
+      authorityFactor: 1.0,
+    });
+    const high = computeScopeWeight({
+      graphProximity: 0.5,
+      scopeLevel: 'team',
+      intentBiases: neutralBiases,
+      authorityFactor: 1.2,
+    });
+    expect(high).toBeGreaterThan(low);
+  });
+
+  it('never returns negative', () => {
+    const weight = computeScopeWeight({
+      graphProximity: 0,
+      scopeLevel: 'organization',
+      intentBiases: neutralBiases,
+      authorityFactor: 1.0,
+    });
+    expect(weight).toBeGreaterThanOrEqual(0);
+  });
+});
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `npx vitest run tests/lib/cortex/retrieval/weight.test.ts`
+
+- [ ] **Step 3: Implement weight computation**
+
+```typescript
+// src/lib/cortex/retrieval/weight.ts
+import type { IntentBiases } from './intent';
+import type { ScopeLevel } from '../knowledge/types';
+
+export interface ScopeWeightInput {
+  graphProximity: number;    // 0-1, from EntityGraph.proximity()
+  scopeLevel: ScopeLevel | string;
+  intentBiases: IntentBiases;
+  authorityFactor: number;   // 1.0 default, higher for experts/docs
+}
+
+/**
+ * Compute the retrieval weight for a knowledge scope.
+ *
+ * weight = graphProximity × intentBias × authorityFactor
+ *
+ * Per spec: weight(scope) = graph_proximity × intent_bias × freshness_bonus × authority
+ * Freshness is applied per-result in the fusion stage, not per-scope.
+ */
+export function computeScopeWeight(input: ScopeWeightInput): number {
+  const { graphProximity, scopeLevel, intentBiases, authorityFactor } = input;
+
+  const intentBias = intentBiases.scope_boost[scopeLevel] ?? 1.0;
+
+  return Math.max(0, graphProximity * intentBias * authorityFactor);
+}
+
+/**
+ * Compute per-result type boost from intent biases.
+ */
+export function computeTypeBoost(knowledgeType: string, intentBiases: IntentBiases): number {
+  return intentBiases.type_boost[knowledgeType] ?? 1.0;
+}
+
+/**
+ * Authority factor for a source based on role and expertise.
+ *
+ * role_boost: 0.0 member, 0.1 lead, 0.15 senior/principal, 0.2 director+
+ * expertise_weight: EXPERT_IN edge weight (0-1)
+ * Documents get 1.2 base authority.
+ */
+export function computeAuthority(params: {
+  role?: string;
+  expertiseWeight?: number;
+  isDocument?: boolean;
+}): number {
+  const { role, expertiseWeight = 0, isDocument = false } = params;
+
+  if (isDocument) return 1.2;
+
+  const roleBoosts: Record<string, number> = {
+    member: 0.0,
+    lead: 0.1,
+    senior: 0.15,
+    principal: 0.15,
+    director: 0.2,
+    vp: 0.2,
+    cto: 0.2,
+  };
+
+  const roleBoost = roleBoosts[role?.toLowerCase() ?? 'member'] ?? 0.0;
+  return Math.max(1.0, roleBoost + expertiseWeight);
+}
+```
+
+- [ ] **Step 4: Run tests to verify they pass**
+
+Run: `npx vitest run tests/lib/cortex/retrieval/weight.test.ts`
+Expected: PASS (5 tests)
+
+- [ ] **Step 5: Commit**
+
+```bash
+git add src/lib/cortex/retrieval/weight.ts tests/lib/cortex/retrieval/weight.test.ts
+git commit -m "feat(cortex): add scope weight computation for context assembly"
+```
+
+---
+
+## Chunk 2: Conflict Detection and Context Formatting
+
+### Task 3: Conflict detection
+
+**Files:**
+- Create: `src/lib/cortex/retrieval/conflict.ts`
+- Create: `tests/lib/cortex/retrieval/conflict.test.ts`
+
+- [ ] **Step 1: Write failing tests**
+
+```typescript
+// tests/lib/cortex/retrieval/conflict.test.ts
+import { describe, it, expect } from 'vitest';
+import { detectConflicts } from '@/lib/cortex/retrieval/conflict';
+import type { ScoredKnowledge } from '@/lib/cortex/knowledge/types';
+
+function makeResult(overrides: Partial<ScoredKnowledge> = {}): ScoredKnowledge {
+  return {
+    id: 'r1', vector: [], text: 'test', type: 'decision', layer: 'personal',
+    workspace_id: null, session_id: null, agent_type: 'claude',
+    project_path: null, file_refs: [], confidence: 0.8,
+    created: new Date().toISOString(), source_timestamp: new Date().toISOString(),
+    stale_score: 0, access_count: 0, last_accessed: null, metadata: {},
+    relevance_score: 0.9, similarity: 0.9,
+    contradiction_refs: [], ...overrides,
+  };
+}
+
+describe('detectConflicts', () => {
+  it('returns no conflicts when no contradiction_refs', () => {
+    const results = [makeResult({ id: 'a' }), makeResult({ id: 'b' })];
+    const conflicts = detectConflicts(results);
+    expect(conflicts).toHaveLength(0);
+  });
+
+  it('detects conflict between two results', () => {
+    const results = [
+      makeResult({ id: 'a', text: 'use pool size 50', contradiction_refs: ['b'] }),
+      makeResult({ id: 'b', text: 'scale horizontally', contradiction_refs: ['a'] }),
+    ];
+    const conflicts = detectConflicts(results);
+    expect(conflicts).toHaveLength(1);
+    expect(conflicts[0].unitA.id).toBe('a');
+    expect(conflicts[0].unitB.id).toBe('b');
+  });
+
+  it('ignores contradiction_refs pointing to results not in the set', () => {
+    const results = [
+      makeResult({ id: 'a', contradiction_refs: ['z'] }),  // z is not in results
+    ];
+    const conflicts = detectConflicts(results);
+    expect(conflicts).toHaveLength(0);
+  });
+
+  it('deduplicates symmetric conflicts', () => {
+    // A contradicts B and B contradicts A should produce one conflict, not two
+    const results = [
+      makeResult({ id: 'a', contradiction_refs: ['b'] }),
+      makeResult({ id: 'b', contradiction_refs: ['a'] }),
+    ];
+    const conflicts = detectConflicts(results);
+    expect(conflicts).toHaveLength(1);
+  });
+});
+```
+
+- [ ] **Step 2: Implement conflict detection**
+
+```typescript
+// src/lib/cortex/retrieval/conflict.ts
+import type { ScoredKnowledge } from '../knowledge/types';
+
+export interface ConflictPair {
+  unitA: ScoredKnowledge;
+  unitB: ScoredKnowledge;
+}
+
+/**
+ * Detect conflicts among search results by checking contradiction_refs.
+ * Returns deduplicated conflict pairs (A↔B counted once, not twice).
+ */
+export function detectConflicts(results: ScoredKnowledge[]): ConflictPair[] {
+  const resultMap = new Map(results.map(r => [r.id, r]));
+  const seen = new Set<string>();
+  const conflicts: ConflictPair[] = [];
+
+  for (const result of results) {
+    const refs = result.contradiction_refs ?? [];
+    for (const refId of refs) {
+      const other = resultMap.get(refId);
+      if (!other) continue;
+
+      const key = [result.id, refId].sort().join('|');
+      if (seen.has(key)) continue;
+      seen.add(key);
+
+      conflicts.push({ unitA: result, unitB: other });
+    }
+  }
+
+  return conflicts;
+}
+```
+
+- [ ] **Step 3: Run tests, commit**
+
+Run: `npx vitest run tests/lib/cortex/retrieval/conflict.test.ts`
+
+```bash
+git add src/lib/cortex/retrieval/conflict.ts tests/lib/cortex/retrieval/conflict.test.ts
+git commit -m "feat(cortex): add conflict detection for search results"
+```
+
+---
+
+### Task 4: Context formatter
+
+**Files:**
+- Create: `src/lib/cortex/retrieval/formatter.ts`
+- Create: `tests/lib/cortex/retrieval/formatter.test.ts`
+
+- [ ] **Step 1: Write failing tests**
+
+```typescript
+// tests/lib/cortex/retrieval/formatter.test.ts
+import { describe, it, expect } from 'vitest';
+import { formatContext } from '@/lib/cortex/retrieval/formatter';
+import type { ScoredKnowledge } from '@/lib/cortex/knowledge/types';
+import type { ConflictPair } from '@/lib/cortex/retrieval/conflict';
+
+function makeResult(overrides: Partial<ScoredKnowledge> = {}): ScoredKnowledge {
+  return {
+    id: 'r1', vector: [], text: 'test knowledge', type: 'decision', layer: 'personal',
+    workspace_id: null, session_id: null, agent_type: 'claude',
+    project_path: null, file_refs: [], confidence: 0.8,
+    created: '2026-03-15T00:00:00.000Z', source_timestamp: '2026-03-15T00:00:00.000Z',
+    stale_score: 0, access_count: 5, last_accessed: null, metadata: {},
+    relevance_score: 0.9, similarity: 0.9,
+    ...overrides,
+  };
+}
+
+describe('formatContext', () => {
+  it('wraps results in cortex-context tags', () => {
+    const output = formatContext([makeResult()], []);
+    expect(output).toContain('<cortex-context>');
+    expect(output).toContain('</cortex-context>');
+  });
+
+  it('includes type labels and dates', () => {
+    const output = formatContext([makeResult({ type: 'error_fix', source_timestamp: '2026-03-10T00:00:00.000Z' })], []);
+    expect(output).toContain('[Error Fix]');
+    expect(output).toContain('2026-03-10');
+  });
+
+  it('includes source attribution when origin is present', () => {
+    const output = formatContext([makeResult({
+      origin: { source_type: 'conversation', source_ref: 'sess-1', creator_entity_id: 'person-alice' },
+    })], []);
+    expect(output).toContain('person-alice');
+  });
+
+  it('includes conflict callout when conflicts exist', () => {
+    const a = makeResult({ id: 'a', text: 'use pool size 50' });
+    const b = makeResult({ id: 'b', text: 'scale horizontally' });
+    const conflicts: ConflictPair[] = [{ unitA: a, unitB: b }];
+    const output = formatContext([a, b], conflicts);
+    expect(output).toContain('Conflicting');
+  });
+
+  it('respects max token budget', () => {
+    const bigResults = Array.from({ length: 20 }, (_, i) =>
+      makeResult({ id: `r${i}`, text: 'x'.repeat(500) })
+    );
+    const output = formatContext(bigResults, [], { maxTokens: 500 });
+    // Should not include all 20 results (would be ~2500 tokens)
+    expect(output.length).toBeLessThan(3000);
+  });
+
+  it('returns empty string when no results', () => {
+    expect(formatContext([], [])).toBe('');
+  });
+});
+```
+
+- [ ] **Step 2: Implement context formatter**
+
+```typescript
+// src/lib/cortex/retrieval/formatter.ts
+import type { ScoredKnowledge } from '../knowledge/types';
+import type { ConflictPair } from './conflict';
+
+const TYPE_LABELS: Record<string, string> = {
+  decision: 'Decision', pattern: 'Pattern', preference: 'Preference',
+  error_fix: 'Error Fix', context: 'Context', code_pattern: 'Code',
+  command: 'Command', conversation: 'Conversation', summary: 'Summary',
+};
+
+export interface FormatOptions {
+  maxTokens?: number;
+}
+
+/**
+ * Format search results + conflicts as annotated <cortex-context> for RAG injection.
+ */
+export function formatContext(
+  results: ScoredKnowledge[],
+  conflicts: ConflictPair[],
+  options: FormatOptions = {},
+): string {
+  if (results.length === 0) return '';
+
+  const maxTokens = options.maxTokens ?? 1500;
+  const entries: string[] = [];
+  let tokens = 20;  // overhead for tags
+
+  // Format each result with attribution
+  for (const unit of results) {
+    const label = TYPE_LABELS[unit.type] || unit.type;
+    const date = (unit.source_timestamp || '').slice(0, 10);
+    const creator = unit.origin?.creator_entity_id ?? '';
+    const sourceInfo = creator ? ` (${creator})` : '';
+
+    let entry = `[${label}] ${date}${sourceInfo}:\n  ${unit.text}`;
+
+    const entryTokens = Math.ceil(entry.length / 4);
+    if (tokens + entryTokens > maxTokens) break;
+
+    entries.push(entry);
+    tokens += entryTokens;
+  }
+
+  if (entries.length === 0) return '';
+
+  // Build conflict section
+  let conflictSection = '';
+  if (conflicts.length > 0) {
+    const conflictLines = conflicts.map(c =>
+      `  - "${c.unitA.text.slice(0, 80)}..." vs "${c.unitB.text.slice(0, 80)}..."`
+    );
+    conflictSection = `\nConflicting perspectives (${conflicts.length}):\n${conflictLines.join('\n')}\n`;
+  }
+
+  const sourceCount = entries.length;
+  const header = `Relevant knowledge (${sourceCount} source${sourceCount > 1 ? 's' : ''}${conflicts.length > 0 ? `, ${conflicts.length} conflict${conflicts.length > 1 ? 's' : ''}` : ''}):`;
+
+  return [
+    '<cortex-context>',
+    header,
+    '',
+    ...entries,
+    conflictSection,
+    '</cortex-context>',
+  ].join('\n');
+}
+```
+
+- [ ] **Step 3: Run tests, commit**
+
+Run: `npx vitest run tests/lib/cortex/retrieval/formatter.test.ts`
+
+```bash
+git add src/lib/cortex/retrieval/formatter.ts tests/lib/cortex/retrieval/formatter.test.ts
+git commit -m "feat(cortex): add context formatter for RAG injection"
+```
+
+---
+
+## Chunk 3: Context Assembly Engine
+
+### Task 5: ContextEngine — the 6-stage pipeline
+
+**Files:**
+- Create: `src/lib/cortex/retrieval/context-engine.ts`
+- Create: `tests/lib/cortex/retrieval/context-engine.test.ts`
+
+- [ ] **Step 1: Write failing tests**
+
+```typescript
+// tests/lib/cortex/retrieval/context-engine.test.ts
+import { describe, it, expect, vi, beforeEach } from 'vitest';
+import { ContextEngine } from '@/lib/cortex/retrieval/context-engine';
+
+// Create mocks for dependencies
+const mockStore = {
+  search: vi.fn().mockResolvedValue([]),
+};
+
+const mockGraph = {
+  proximity: vi.fn().mockReturnValue(0.5),
+  neighborhood: vi.fn().mockReturnValue([]),
+  getEntity: vi.fn().mockReturnValue(null),
+};
+
+const mockResolver = {
+  extractEntities: vi.fn().mockReturnValue([]),
+};
+
+const mockEmbedding = {
+  embed: vi.fn().mockResolvedValue([[0.1, 0.2, 0.3]]),
+  dimensions: 3,
+  name: 'mock',
+  init: vi.fn(),
+};
+
+describe('ContextEngine', () => {
+  let engine: ContextEngine;
+
+  beforeEach(() => {
+    vi.clearAllMocks();
+    engine = new ContextEngine({
+      store: mockStore as any,
+      graph: mockGraph as any,
+      resolver: mockResolver as any,
+      embedding: mockEmbedding as any,
+      requesterId: 'person-alice',
+    });
+  });
+
+  it('returns empty context for empty results', async () => {
+    const result = await engine.assemble('some query');
+    expect(result.results).toHaveLength(0);
+    expect(result.context).toBe('');
+  });
+
+  it('calls embedding.embed with the query', async () => {
+    await engine.assemble('test query');
+    expect(mockEmbedding.embed).toHaveBeenCalledWith(['test query']);
+  });
+
+  it('detects intent from the query', async () => {
+    const result = await engine.assemble('why does auth throw an error?');
+    expect(result.intent.intent).toBe('debugging');
+  });
+
+  it('extracts entities from the query', async () => {
+    mockResolver.extractEntities.mockReturnValue([
+      { entity: { id: 'system-auth', type: 'system', name: 'Auth' }, confidence: 0.9, method: 'alias' },
+    ]);
+    const result = await engine.assemble('fix the auth service');
+    expect(mockResolver.extractEntities).toHaveBeenCalledWith('fix the auth service');
+    expect(result.entities).toHaveLength(1);
+  });
+
+  it('searches store with embedded query vector', async () => {
+    mockStore.search.mockResolvedValue([{
+      id: 'k1', text: 'test knowledge', type: 'decision', layer: 'personal',
+      confidence: 0.8, stale_score: 0, created: new Date().toISOString(),
+      source_timestamp: new Date().toISOString(), evidence_score: 0.7,
+      contradiction_refs: [], _distance: 0.2,
+      workspace_id: null, session_id: null, agent_type: 'claude',
+      project_path: null, file_refs: [], access_count: 0, last_accessed: null,
+      metadata: {},
+    }]);
+
+    const result = await engine.assemble('test query');
+    expect(result.results.length).toBeGreaterThanOrEqual(1);
+    expect(result.context).toContain('<cortex-context>');
+  });
+
+  it('completes within performance budget', async () => {
+    const start = Date.now();
+    await engine.assemble('test query');
+    const elapsed = Date.now() - start;
+    expect(elapsed).toBeLessThan(500);  // generous for CI, target is 150ms
+  });
+});
+```
+
+- [ ] **Step 2: Implement ContextEngine**
+
+```typescript
+// src/lib/cortex/retrieval/context-engine.ts
+import type { CortexStore } from '../store';
+import type { EntityGraph } from '../graph/entity-graph';
+import type { EntityResolver, ResolvedEntity } from '../graph/resolver';
+import type { EmbeddingProvider } from '../embeddings';
+import type { ScoredKnowledge } from '../knowledge/types';
+import { detectIntent } from './intent';
+import type { IntentResult } from './intent';
+import { computeScopeWeight, computeTypeBoost } from './weight';
+import { detectConflicts } from './conflict';
+import type { ConflictPair } from './conflict';
+import { formatContext } from './formatter';
+import { computeRelevanceScore } from './scoring';
+
+export interface ContextEngineDeps {
+  store: CortexStore;
+  graph: EntityGraph;
+  resolver: EntityResolver;
+  embedding: EmbeddingProvider;
+  requesterId: string;  // entity ID of the person making the query
+}
+
+export interface AssemblyResult {
+  results: ScoredKnowledge[];
+  conflicts: ConflictPair[];
+  context: string;           // formatted <cortex-context> string
+  intent: IntentResult;
+  entities: ResolvedEntity[];
+  timing: {
+    intentMs: number;
+    entityMs: number;
+    searchMs: number;
+    totalMs: number;
+  };
+}
+
+interface SearchSource {
+  layerKey: string;
+  weight: number;
+  limit: number;
+}
+
+const DEFAULT_LAYERS = ['personal', 'workspace', 'team'] as const;
+const SEARCH_TIMEOUT_MS = 100;
+
+export class ContextEngine {
+  constructor(private deps: ContextEngineDeps) {}
+
+  async assemble(query: string, options: { limit?: number; workspaceId?: number | null; maxTokens?: number } = {}): Promise<AssemblyResult> {
+    const totalStart = Date.now();
+    const { limit = 5, workspaceId = null, maxTokens = 1500 } = options;
+
+    // Stage 1: Intent Detection
+    const intentStart = Date.now();
+    const intent = detectIntent(query);
+    const intentMs = Date.now() - intentStart;
+
+    // Stage 2: Entity Resolution
+    const entityStart = Date.now();
+    const entities = this.deps.resolver.extractEntities(query);
+    const entityMs = Date.now() - entityStart;
+
+    // Embed the query
+    const [queryVector] = await this.deps.embedding.embed([query]);
+
+    // Stage 3: Weight Computation
+    const sources = this.computeSourceWeights(intent, workspaceId);
+
+    // Stage 4: Parallel Multi-Source Search
+    const searchStart = Date.now();
+    const allResults = await this.parallelSearch(queryVector, sources, limit);
+    const searchMs = Date.now() - searchStart;
+
+    // Stage 5: Fusion + Re-Ranking
+    const fused = this.fuseAndRank(allResults, intent, limit);
+
+    // Stage 6: Conflict Detection + Formatting
+    const conflicts = detectConflicts(fused);
+    const context = formatContext(fused, conflicts, { maxTokens });
+
+    return {
+      results: fused,
+      conflicts,
+      context,
+      intent,
+      entities,
+      timing: {
+        intentMs,
+        entityMs,
+        searchMs,
+        totalMs: Date.now() - totalStart,
+      },
+    };
+  }
+
+  private computeSourceWeights(intent: IntentResult, workspaceId: number | null): SearchSource[] {
+    const sources: SearchSource[] = [];
+
+    for (const layer of DEFAULT_LAYERS) {
+      const layerKey = layer === 'workspace' && workspaceId
+        ? `workspace/${workspaceId}` : layer;
+
+      // Map v1 layer to scope level for intent bias lookup
+      const scopeLevel = layer === 'personal' ? 'personal'
+        : layer === 'workspace' ? 'team' : 'organization';
+
+      // Use graph proximity if available, else fall back to fixed weights
+      let graphProximity: number;
+      try {
+        // For personal, distance is 0; for workspace, 1; for team, 2
+        const layerEntity = layer === 'personal' ? this.deps.requesterId
+          : layer === 'workspace' ? 'team-default' : 'organization-default';
+        graphProximity = this.deps.graph.proximity(this.deps.requesterId, layerEntity);
+      } catch {
+        // Graph not populated, use fallback
+        graphProximity = layer === 'personal' ? 1.0 : layer === 'workspace' ? 0.5 : 0.33;
+      }
+
+      // If graph returns 0 (unreachable/not found), use fallback
+      if (graphProximity === 0) {
+        graphProximity = layer === 'personal' ? 1.0 : layer === 'workspace' ? 0.5 : 0.33;
+      }
+
+      const weight = computeScopeWeight({
+        graphProximity,
+        scopeLevel,
+        intentBiases: intent.biases,
+        authorityFactor: 1.0,
+      });
+
+      sources.push({
+        layerKey,
+        weight,
+        limit: Math.max(3, Math.round(weight * 10)),  // proportional slots
+      });
+    }
+
+    return sources.sort((a, b) => b.weight - a.weight);
+  }
+
+  private async parallelSearch(
+    queryVector: number[],
+    sources: SearchSource[],
+    limit: number,
+  ): Promise<Array<ScoredKnowledge & { sourceWeight: number }>> {
+    const searchPromises = sources.map(async (source) => {
+      try {
+        const results = await Promise.race([
+          this.deps.store.search(source.layerKey, queryVector, source.limit),
+          new Promise<never>((_, reject) =>
+            setTimeout(() => reject(new Error('timeout')), SEARCH_TIMEOUT_MS)
+          ),
+        ]);
+
+        return results.map(unit => {
+          const similarity = 1 - ((unit as any)._distance ?? 0);
+          return {
+            ...unit,
+            similarity,
+            relevance_score: 0,  // computed in fusion
+            sourceWeight: source.weight,
+          } as ScoredKnowledge & { sourceWeight: number };
+        });
+      } catch {
+        return [];  // source failed or timed out
+      }
+    });
+
+    const settled = await Promise.allSettled(searchPromises);
+    const allResults: Array<ScoredKnowledge & { sourceWeight: number }> = [];
+
+    for (const result of settled) {
+      if (result.status === 'fulfilled') {
+        allResults.push(...result.value);
+      }
+    }
+
+    return allResults;
+  }
+
+  private fuseAndRank(
+    results: Array<ScoredKnowledge & { sourceWeight: number }>,
+    intent: IntentResult,
+    limit: number,
+  ): ScoredKnowledge[] {
+    // Score each result
+    for (const result of results) {
+      const typeBoost = computeTypeBoost(result.type, intent.biases);
+      const recencyBoost = intent.biases.recency_boost;
+
+      result.relevance_score = computeRelevanceScore({
+        similarity: result.similarity,
+        confidence: result.confidence,
+        stale_score: result.stale_score,
+        created: result.created,
+        evidence_score: result.evidence_score,
+      }) * result.sourceWeight * typeBoost * recencyBoost;
+    }
+
+    // Deduplicate: cosine > 0.9 between results (approximate via text similarity)
+    const deduped = this.deduplicateResults(results);
+
+    // Sort and take top K
+    deduped.sort((a, b) => b.relevance_score - a.relevance_score);
+    return deduped.slice(0, limit);
+  }
+
+  private deduplicateResults(results: ScoredKnowledge[]): ScoredKnowledge[] {
+    const kept: ScoredKnowledge[] = [];
+    const seenTexts = new Set<string>();
+
+    // Sort by score first so we keep the better-scored version
+    results.sort((a, b) => b.relevance_score - a.relevance_score);
+
+    for (const result of results) {
+      // Simple text-based dedup: normalize and check prefix overlap
+      const normalized = result.text.slice(0, 200).toLowerCase().trim();
+      if (seenTexts.has(normalized)) continue;
+
+      // Check against existing kept items for high text overlap
+      let isDupe = false;
+      for (const existing of kept) {
+        if (result.id === existing.id) { isDupe = true; break; }
+        // If texts share >80% of content, consider duplicate
+        const existNorm = existing.text.slice(0, 200).toLowerCase().trim();
+        if (normalized === existNorm) { isDupe = true; break; }
+      }
+
+      if (!isDupe) {
+        seenTexts.add(normalized);
+        kept.push(result);
+      }
+    }
+
+    return kept;
+  }
+}
+```
+
+- [ ] **Step 3: Run tests to verify they pass**
+
+Run: `npx vitest run tests/lib/cortex/retrieval/context-engine.test.ts`
+Expected: PASS (6 tests)
+
+- [ ] **Step 4: Commit**
+
+```bash
+git add src/lib/cortex/retrieval/context-engine.ts tests/lib/cortex/retrieval/context-engine.test.ts
+git commit -m "feat(cortex): add ContextEngine with 6-stage retrieval pipeline"
+```
+
+---
+
+## Chunk 4: API Endpoint and Hook Integration
+
+### Task 6: Context assembly API endpoint
+
+**Files:**
+- Create: `src/app/api/cortex/context/route.ts`
+
+- [ ] **Step 1: Create the endpoint**
+
+```typescript
+// src/app/api/cortex/context/route.ts
+import { NextResponse } from 'next/server';
+import type { NextRequest } from 'next/server';
+import { getAuthUser, withUser } from '@/lib/auth';
+import { getCortex, isCortexAvailable } from '@/lib/cortex';
+import { ContextEngine } from '@/lib/cortex/retrieval/context-engine';
+import { EntityResolver } from '@/lib/cortex/graph/resolver';
+import { slugify } from '@/lib/cortex/graph/types';
+
+export async function GET(request: NextRequest) {
+  const user = getAuthUser(request);
+  return withUser(user, async () => {
+    if (!isCortexAvailable()) {
+      return NextResponse.json({ error: 'Cortex unavailable' }, { status: 403 });
+    }
+    const cortex = await getCortex();
+    if (!cortex) return NextResponse.json({ results: [], context: '' });
+
+    const url = new URL(request.url);
+    const query = url.searchParams.get('q') || '';
+    const limit = parseInt(url.searchParams.get('limit') || '5', 10);
+    const workspaceId = url.searchParams.get('workspace_id');
+    const maxTokens = parseInt(url.searchParams.get('max_tokens') || '1500', 10);
+
+    if (!query || query.length < 3) {
+      return NextResponse.json({ results: [], context: '' });
+    }
+
+    const resolver = new EntityResolver(cortex.graph);
+    const requesterId = `person-${slugify(user)}`;
+
+    const engine = new ContextEngine({
+      store: cortex.store,
+      graph: cortex.graph,
+      resolver,
+      embedding: cortex.embedding,
+      requesterId,
+    });
+
+    const result = await engine.assemble(query, {
+      limit,
+      workspaceId: workspaceId ? parseInt(workspaceId, 10) : null,
+      maxTokens,
+    });
+
+    return NextResponse.json({
+      results: result.results.map(r => ({ ...r, vector: undefined })),  // strip vectors
+      context: result.context,
+      intent: result.intent,
+      conflicts: result.conflicts.length,
+      timing: result.timing,
+    });
+  });
+}
+```
+
+- [ ] **Step 2: Commit**
+
+```bash
+git add src/app/api/cortex/context/route.ts
+git commit -m "feat(cortex): add context assembly API endpoint"
+```
+
+---
+
+### Task 7: Update RAG hook to use context assembly endpoint
+
+**Files:**
+- Modify: `bin/cortex-hook.js`
+
+- [ ] **Step 1: Read current cortex-hook.js**
+
+Read the file to understand the current flow: query → /api/cortex/search → format results → output.
+
+- [ ] **Step 2: Update to call context assembly endpoint**
+
+Change the URL from `/api/cortex/search/?q=...` to `/api/cortex/context/?q=...`. The new endpoint returns `{ context, results, intent, conflicts, timing }` — the `context` field is already pre-formatted as `<cortex-context>`, so the hook can output it directly instead of doing its own formatting.
+
+Key changes:
+1. URL: `/api/cortex/search/` → `/api/cortex/context/`
+2. Response handling: use `parsed.context` directly instead of formatting results manually
+3. Keep the fallback: if the new endpoint returns empty or fails, fall back to old behavior
+
+```javascript
+// Replace the response handling section:
+const parsed = JSON.parse(body);
+
+// New: context is pre-formatted by the Context Assembly Engine
+if (parsed.context) {
+  const output = JSON.stringify({
+    hookSpecificOutput: {
+      hookEventName: 'UserPromptSubmit',
+      additionalContext: parsed.context,
+    },
+  });
+  process.stdout.write(output);
+  process.exit(0);
+}
+
+// Fallback: old-style results formatting (if context endpoint not available)
+const results = parsed.results;
+if (!results || results.length === 0) process.exit(0);
+// ... existing formatting code stays as fallback ...
+```
+
+- [ ] **Step 3: Commit**
+
+```bash
+git add bin/cortex-hook.js
+git commit -m "feat(cortex): switch RAG hook to context assembly endpoint"
+```
+
+---
+
+### Task 8: Integrate ContextEngine into CortexInstance
+
+**Files:**
+- Modify: `src/lib/cortex/index.ts`
+
+- [ ] **Step 1: Read current index.ts**
+
+- [ ] **Step 2: Add ContextEngine to CortexInstance**
+
+1. Import: `import { ContextEngine } from './retrieval/context-engine';` and `import { EntityResolver } from './graph/resolver';`
+2. Add `contextEngine?: ContextEngine` to `CortexInstance` interface (optional since it depends on graph)
+3. In `getCortex()`, after graph initialization, create the ContextEngine:
+
+```typescript
+const resolver = new EntityResolver(graph);
+const contextEngine = new ContextEngine({
+  store,
+  graph,
+  resolver,
+  embedding,
+  requesterId: 'person-default-user',  // default; overridden per-request in API
+});
+```
+
+4. Add to instance object: `contextEngine,`
+
+- [ ] **Step 3: Run full cortex test suite**
+
+Run: `npx vitest run tests/lib/cortex/`
+
+- [ ] **Step 4: Commit**
+
+```bash
+git add src/lib/cortex/index.ts
+git commit -m "feat(cortex): add ContextEngine to CortexInstance"
+```
+
+---
+
+## Summary
+
+| Task | Component | Tests | Status |
+|------|-----------|-------|--------|
+| 1 | Intent detection | 7 | |
+| 2 | Weight computation | 5 | |
+| 3 | Conflict detection | 4 | |
+| 4 | Context formatter | 6 | |
+| 5 | ContextEngine (6-stage pipeline) | 6 | |
+| 6 | Context assembly API endpoint | — | |
+| 7 | RAG hook integration | — | |
+| 8 | CortexInstance integration | regression | |
+
+**Total: 8 tasks, ~28 new tests, 4 chunks**
+
+**Performance budget:** The ContextEngine targets <150ms total latency:
+- Intent detection: <5ms (regex, no LLM)
+- Entity resolution: <5ms (alias lookup)
+- Weight computation: <10ms (graph proximity, cached)
+- Parallel search: <100ms (concurrent vector search with 100ms timeout)
+- Fusion + formatting: <10ms
+
+**Key design decisions:**
+- ContextEngine sits ABOVE CortexSearch — doesn't replace it, wraps it
+- Graph proximity drives weights — but gracefully degrades to fixed weights when graph is empty
+- Deduplication uses text prefix comparison (not cosine between result vectors — that would require N² vector ops)
+- RAG hook uses pre-formatted `context` from the engine — no more client-side formatting
+- Old `/api/cortex/search` endpoint still works — new `/api/cortex/context` endpoint adds the intelligence layer