onbuzz 4.8.3 → 4.9.1

package/src/services/__tests__/aiService.core.test.js

@@ -0,0 +1,420 @@
+/**
+ * aiService.core.test.js — coverage for the AIService class core surface.
+ *
+ * Scope (deliberate, focused):
+ *   - Constructor + initial state setup
+ *   - Ollama-model detection helper
+ *   - Model reliability tracking (success/failure counters, per-model + all)
+ *   - Model-related error detection + classification
+ *   - Model suggestions (sorted by reliability)
+ *   - Circuit breaker state machine (record → open → recovery)
+ *   - Catalog-based model resolution by api_type
+ *   - setModelsService wiring
+ *
+ * Out of scope for this file (each gets its own deep test pass):
+ *   - sendMessage / sendMessageStream HTTP paths (mock SDK, big surface)
+ *   - SSE stream parsing → covered by aiService.stream.test.js
+ *   - Embeddings hookup       → covered by aiService.embed.test.js
+ *   - _extractReasoningTokens → covered by aiService.reasoningTokens.test.js
+ *
+ * Pure-method coverage strategy: no network, no SDK mocks needed. Every
+ * method we hit here is computational over local state.
+ */
+
+import { describe, test, expect, beforeEach, jest } from '@jest/globals';
+import AIService from '../aiService.js';
+
+const silentLogger = {
+  info: jest.fn(), warn: jest.fn(), error: jest.fn(), debug: jest.fn(),
+};
+
+function makeService(overrides = {}) {
+  const config = {
+    backend: { baseUrl: 'https://test.example/', timeout: 1000 },
+    ...overrides,
+  };
+  return new AIService(config, silentLogger, null, null);
+}
+
+// ──────────────────────── constructor ──────────────────────────────
+
+describe('AIService — constructor & initial state', () => {
+  test('uses backend.baseUrl from config when provided', () => {
+    const svc = makeService();
+    expect(svc.baseUrl).toBe('https://test.example/');
+    expect(svc.timeout).toBe(1000);
+  });
+
+  test('falls back to the documented Azure default when backend not configured', () => {
+    const svc = new AIService({}, silentLogger, null, null);
+    expect(svc.baseUrl).toMatch(/^https:\/\//);
+    expect(svc.timeout).toBeGreaterThan(0);
+  });
+
+  test('initial state is consistent: empty maps + closed circuit', () => {
+    const svc = makeService();
+    expect(svc.rateLimiters).toBeInstanceOf(Map);
+    expect(svc.requestQueue).toBeInstanceOf(Map);
+    expect(svc.conversationManagers).toBeInstanceOf(Map);
+    expect(svc.activeRequests).toBeInstanceOf(Map);
+    expect(svc.userAbortedRequests).toBeInstanceOf(Set);
+    expect(svc.modelReliability).toBeInstanceOf(Map);
+    expect(svc.modelReliability.size).toBe(0);
+    expect(svc.circuitBreaker.isOpen).toBe(false);
+    expect(svc.circuitBreaker.failures).toBe(0);
+    expect(svc.circuitBreaker.threshold).toBeGreaterThan(0);
+  });
+
+  test('apiKeyManager / agentPool / embeddingService start unset (LoxiaSystem wires them later)', () => {
+    const svc = makeService();
+    expect(svc.apiKeyManager).toBeNull();
+    expect(svc.agentPool).toBeNull();
+    expect(svc.embeddingService).toBeNull();
+    expect(svc._ollamaService).toBeNull();
+  });
+
+  test('modelSpecs is populated with known model entries', () => {
+    const svc = makeService();
+    expect(svc.modelSpecs).toEqual(expect.any(Object));
+    // Has SOME model entries with provider + contextWindow shape
+    const sample = Object.values(svc.modelSpecs)[0];
+    expect(sample).toMatchObject({
+      provider: expect.any(String),
+      contextWindow: expect.any(Number),
+      maxTokens: expect.any(Number),
+    });
+  });
+});
+
+// ──────────────────── _isOllamaModel ──────────────────────────────
+
+describe('AIService — _isOllamaModel', () => {
+  const svc = makeService();
+  // True paths assert strict true; false paths use truthy/falsy because the
+  // implementation short-circuits with `model && model.startsWith(...)` —
+  // a falsy `model` short-circuits to the original value (null/undef/'')
+  // rather than coercing to literal false. Either is fine for callers since
+  // the function is always used in a boolean context.
+  test.each([
+    ['ollama-llama3',  true],
+    ['ollama-mistral', true],
+    ['ollama-',        true], // prefix-only is technically a hit
+  ])('_isOllamaModel(%p) is true', (input, expected) => {
+    expect(svc._isOllamaModel(input)).toBe(expected);
+  });
+
+  test.each([
+    ['llama3'],
+    ['gpt-4'],
+    ['claude-sonnet'],
+    [''],
+    [null],
+    [undefined],
+  ])('_isOllamaModel(%p) is falsy', (input) => {
+    expect(svc._isOllamaModel(input)).toBeFalsy();
+  });
+});
+
+// ─────────────────── reliability tracking ─────────────────────────
+
+describe('AIService — reliability tracking', () => {
+  let svc;
+  beforeEach(() => { svc = makeService(); });
+
+  test('unknown model returns the zero-state shape', () => {
+    expect(svc.getModelReliability('never-used')).toEqual({
+      verified: false, successCount: 0, failureCount: 0,
+    });
+  });
+
+  test('recordModelSuccess increments + marks verified', () => {
+    svc.recordModelSuccess('gpt-4');
+    const r = svc.getModelReliability('gpt-4');
+    expect(r.verified).toBe(true);
+    expect(r.successCount).toBe(1);
+    expect(r.failureCount).toBe(0);
+    expect(r.lastSuccess).toBeInstanceOf(Date);
+  });
+
+  test('multiple successes accumulate', () => {
+    svc.recordModelSuccess('gpt-4');
+    svc.recordModelSuccess('gpt-4');
+    svc.recordModelSuccess('gpt-4');
+    expect(svc.getModelReliability('gpt-4').successCount).toBe(3);
+  });
+
+  test('recordModelFailure increments + records errorType', () => {
+    svc.recordModelFailure('gpt-4', 'rate_limit');
+    const r = svc.getModelReliability('gpt-4');
+    expect(r.failureCount).toBe(1);
+    expect(r.lastErrorType).toBe('rate_limit');
+    expect(r.lastFailure).toBeInstanceOf(Date);
+  });
+
+  test('failures alone do NOT verify the model', () => {
+    svc.recordModelFailure('gpt-4', 'unknown');
+    expect(svc.getModelReliability('gpt-4').verified).toBe(false);
+  });
+
+  test('errorType defaults to "unknown" when omitted', () => {
+    svc.recordModelFailure('gpt-4');
+    expect(svc.getModelReliability('gpt-4').lastErrorType).toBe('unknown');
+  });
+
+  test('per-model state stays isolated', () => {
+    svc.recordModelSuccess('gpt-4');
+    svc.recordModelFailure('claude-sonnet', 'rate_limit');
+    expect(svc.getModelReliability('gpt-4').successCount).toBe(1);
+    expect(svc.getModelReliability('gpt-4').failureCount).toBe(0);
+    expect(svc.getModelReliability('claude-sonnet').failureCount).toBe(1);
+    expect(svc.getModelReliability('claude-sonnet').successCount).toBe(0);
+  });
+
+  test('getAllModelReliability returns the public-shape map', () => {
+    svc.recordModelSuccess('gpt-4');
+    svc.recordModelFailure('gpt-4', 'rate_limit');
+    svc.recordModelFailure('claude-sonnet', 'auth_error');
+    const all = svc.getAllModelReliability();
+    expect(Object.keys(all).sort()).toEqual(['claude-sonnet', 'gpt-4']);
+    expect(all['gpt-4']).toEqual({
+      verified: true, successCount: 1, failureCount: 1,
+    });
+    expect(all['claude-sonnet']).toEqual({
+      verified: false, successCount: 0, failureCount: 1,
+    });
+  });
+
+  test('getAllModelReliability returns {} when nothing recorded', () => {
+    expect(svc.getAllModelReliability()).toEqual({});
+  });
+});
+
+// ────────────────── setModelsService ──────────────────────────────
+
+describe('AIService — setModelsService', () => {
+  test('attaches the modelsService reference', () => {
+    const svc = makeService();
+    expect(svc.modelsService).toBeUndefined();
+    const ms = { getModels: () => [] };
+    svc.setModelsService(ms);
+    expect(svc.modelsService).toBe(ms);
+  });
+});
+
+// ──────────────── isModelRelatedError ────────────────────────────
+
+describe('AIService — isModelRelatedError', () => {
+  const svc = makeService();
+
+  test.each([
+    [{ status: 429, message: 'whatever' }, true],
+    [{ status: 503, message: 'whatever' }, true],
+    [{ code: 404,    message: 'whatever' }, true],
+    [{ message: 'Model "x" not found' },           true],
+    [{ message: 'model is unavailable' },          true],
+    [{ message: 'Model does not exist' },          true],
+    [{ message: 'invalid model name' },            true],
+    [{ message: 'Unsupported model variant' },     true],
+    [{ message: 'Quota exceeded for org' },        true],
+    [{ message: 'Hit a rate limit' },              true],
+    [{ message: 'Context length exceeded' },       true],
+    [{ message: 'Token limit reached' },           true],
+    [{ message: 'maximum context window' },        true],
+    [{ message: 'service capacity issue' },        true],
+    [{ message: 'model overloaded right now' },    true],
+    [{ message: 'returned 503 internal' },         true],
+    [{ message: 'got 429 from upstream' },         true],
+    [{ message: 'something completely unrelated' }, false],
+    [{ message: '' },                               false],
+    [{ status: 500, message: 'server error' },      false],
+  ])('isModelRelatedError(%p) → %p', (err, expected) => {
+    expect(svc.isModelRelatedError(err)).toBe(expected);
+  });
+});
+
+// ───────────────── classifyModelError ────────────────────────────
+
+describe('AIService — classifyModelError', () => {
+  const svc = makeService();
+
+  test.each([
+    [{ code: 404 },                                   'model_not_found'],
+    [{ message: 'model not found' },                   'model_not_found'],
+    [{ message: 'thing does not exist' },              'model_not_found'],
+    [{ code: 429 },                                    'rate_limit'],
+    [{ message: 'rate limit reached' },                'rate_limit'],
+    [{ message: 'quota was hit' },                     'rate_limit'],
+    [{ message: 'maximum context exceeded' },          'context_exceeded'],
+    [{ message: 'token limit too small' },             'context_exceeded'],
+    [{ code: 503 },                                    'model_overloaded'],
+    [{ message: 'model overload' },                    'model_overloaded'],
+    [{ message: 'capacity full right now' },           'model_overloaded'],
+    [{ code: 401 },                                    'auth_error'],
+    [{ code: 403 },                                    'auth_error'],
+    [{ code: 500, message: 'whoops' },                 'unknown'],
+    [{ message: '' },                                  'unknown'],
+  ])('classifyModelError(%p) → %p', (err, expected) => {
+    expect(svc.classifyModelError(err)).toBe(expected);
+  });
+
+  test('status field is treated equivalent to code field', () => {
+    expect(svc.classifyModelError({ status: 404 })).toBe('model_not_found');
+    expect(svc.classifyModelError({ status: 429 })).toBe('rate_limit');
+  });
+});
+
+// ───────────────── getModelSuggestions ───────────────────────────
+
+describe('AIService — getModelSuggestions', () => {
+  test('returns empty suggestions when no modelsService is attached', () => {
+    const svc = makeService();
+    const out = svc.getModelSuggestions('gpt-4', new Error('rate limit'));
+    expect(out.failedModel).toBe('gpt-4');
+    expect(out.errorType).toBe('rate_limit');
+    expect(out.errorMessage).toBe('rate limit');
+    expect(out.suggestions).toEqual([]);
+  });
+
+  test('excludes the failed model from suggestions', () => {
+    const svc = makeService();
+    svc.setModelsService({
+      getModels: () => [
+        { name: 'gpt-4',         displayName: 'GPT-4',         provider: 'openai',    contextWindow: 128000 },
+        { name: 'claude-sonnet', displayName: 'Claude Sonnet', provider: 'anthropic', contextWindow: 200000 },
+      ],
+    });
+    const out = svc.getModelSuggestions('gpt-4', new Error('whatever'));
+    expect(out.suggestions.map(s => s.name)).toEqual(['claude-sonnet']);
+  });
+
+  test('sorts verified models above unverified, then by successCount desc', () => {
+    const svc = makeService();
+    svc.setModelsService({
+      getModels: () => [
+        { name: 'A', provider: 'p', contextWindow: 1 },
+        { name: 'B', provider: 'p', contextWindow: 1 },
+        { name: 'C', provider: 'p', contextWindow: 1 },
+        { name: 'D', provider: 'p', contextWindow: 1 },
+      ],
+    });
+    // Only B and D are "verified" (have at least one success)
+    svc.recordModelSuccess('B');
+    svc.recordModelSuccess('D');
+    svc.recordModelSuccess('D'); // D has 2 successes, B has 1
+    // C has only a failure → unverified
+    svc.recordModelFailure('C', 'rate_limit');
+
+    const out = svc.getModelSuggestions('Z-not-in-list', new Error(''));
+    const order = out.suggestions.map(s => s.name);
+    // Verified first, sorted by success count: D (2), B (1), then unverified A & C
+    expect(order.slice(0, 2)).toEqual(['D', 'B']);
+    expect(order.slice(2).sort()).toEqual(['A', 'C']); // tail order among unverified is impl detail
+  });
+
+  test('errorType is derived via classifyModelError', () => {
+    const svc = makeService();
+    svc.setModelsService({ getModels: () => [] });
+    const e = new Error('overloaded'); e.code = 503;
+    expect(svc.getModelSuggestions('x', e).errorType).toBe('model_overloaded');
+  });
+});
+
+// ────────────────── circuit breaker ──────────────────────────────
+
+describe('AIService — circuit breaker', () => {
+  let svc;
+  beforeEach(() => { svc = makeService(); });
+
+  test('starts closed', () => {
+    expect(svc._isCircuitBreakerOpen()).toBe(false);
+  });
+
+  test('opens after `threshold` consecutive failures', () => {
+    const threshold = svc.circuitBreaker.threshold;
+    for (let i = 0; i < threshold; i++) svc._recordFailure();
+    expect(svc.circuitBreaker.isOpen).toBe(true);
+    expect(svc._isCircuitBreakerOpen()).toBe(true);
+  });
+
+  test('stays closed below threshold', () => {
+    svc._recordFailure();
+    svc._recordFailure();
+    expect(svc._isCircuitBreakerOpen()).toBe(false);
+  });
+
+  test('_isCircuitBreakerOpen auto-recovers after the timeout window', () => {
+    // Trip the breaker.
+    for (let i = 0; i < svc.circuitBreaker.threshold; i++) svc._recordFailure();
+    expect(svc.circuitBreaker.isOpen).toBe(true);
+
+    // Pretend the last failure was longer ago than the recovery window.
+    svc.circuitBreaker.lastFailureTime = Date.now() - svc.circuitBreaker.timeout - 1;
+    expect(svc._isCircuitBreakerOpen()).toBe(false);
+    // Resetting also clears failure count.
+    expect(svc.circuitBreaker.failures).toBe(0);
+  });
+
+  test('_resetCircuitBreaker zeroes counters + closes', () => {
+    svc._recordFailure();
+    svc._recordFailure();
+    svc._resetCircuitBreaker();
+    expect(svc.circuitBreaker.failures).toBe(0);
+    expect(svc.circuitBreaker.isOpen).toBe(false);
+  });
+
+  test('_resetCircuitBreaker on already-clean state is a no-op (no log spam)', () => {
+    silentLogger.info.mockClear();
+    svc._resetCircuitBreaker();
+    expect(silentLogger.info).not.toHaveBeenCalled();
+  });
+
+  test('lastFailureTime is updated on each failure', () => {
+    const before = Date.now();
+    svc._recordFailure();
+    expect(svc.circuitBreaker.lastFailureTime).toBeGreaterThanOrEqual(before);
+  });
+});
+
+// ──────────────── _resolveModelByApiType ─────────────────────────
+
+describe('AIService — _resolveModelByApiType', () => {
+  test('returns null when no modelsService is attached', () => {
+    expect(makeService()._resolveModelByApiType('image_generation')).toBeNull();
+  });
+
+  test('finds models whose api_type array includes the requested type', () => {
+    const svc = makeService();
+    svc.setModelsService({
+      getModels: () => [
+        { name: 'chat-1',  api_type: ['chat'] },
+        { name: 'image-1', api_type: ['image_generation'] },
+        { name: 'image-2', api_type: ['image_generation', 'chat'] },
+      ],
+    });
+    expect(svc._resolveModelByApiType('image_generation')).toBe('image-1');
+  });
+
+  test('falls back to modality field when api_type is absent', () => {
+    const svc = makeService();
+    svc.setModelsService({
+      getModels: () => [
+        { name: 'video-only', modality: 'video_generation' },
+      ],
+    });
+    expect(svc._resolveModelByApiType('video_generation')).toBe('video-only');
+  });
+
+  test('returns null when no model matches', () => {
+    const svc = makeService();
+    svc.setModelsService({
+      getModels: () => [{ name: 'x', api_type: ['chat'] }],
+    });
+    expect(svc._resolveModelByApiType('nonexistent')).toBeNull();
+  });
+
+  test('returns null without throwing when getModels throws', () => {
+    const svc = makeService();
+    svc.setModelsService({ getModels: () => { throw new Error('boom'); } });
+    expect(svc._resolveModelByApiType('chat')).toBeNull();
+  });
+});

package/src/services/__tests__/aiService.embed.test.js

@@ -0,0 +1,172 @@
+/**
+ * Tests for the aiService ↔ EmbeddingService wiring.
+ *
+ * Pins:
+ *   - `embeddingService` starts as null on the instance.
+ *   - `setEmbeddingService` attaches; `getEmbeddingService` reads back.
+ *   - `aiService.embed("a string")`   → embeddingService.embedQuery
+ *   - `aiService.embed(["a","b"])`    → embeddingService.embedDocuments
+ *   - `embed()` defaults kind='document' for arrays; passes `kind` through.
+ *   - Calling `embed()` without an attached service throws clearly.
+ *
+ * We do NOT exercise the full aiService class — just the embedding
+ * surface area. The constructor is tolerant enough to accept a minimal
+ * config; the heavy network code paths are never reached.
+ */
+
+import { describe, test, expect, jest, beforeEach } from '@jest/globals';
+import { default as AIService } from '../aiService.js';
+
+function makeService() {
+  // Bare-minimum config; aiService falls back to defaults for anything missing.
+  return new AIService({ backend: { baseUrl: 'https://stub.example' } }, null, null, null);
+}
+
+function makeStubEmbeddingService() {
+  return {
+    isEnabled: true,
+    embedQuery: jest.fn(async () => new Float32Array([1, 0, 0, 0])),
+    embedDocuments: jest.fn(async (texts) =>
+      texts.map(() => new Float32Array([0.5, 0.5, 0, 0]))),
+  };
+}
+
+describe('aiService — embeddings wiring', () => {
+  let aiService, stub;
+  beforeEach(() => {
+    aiService = makeService();
+    stub = makeStubEmbeddingService();
+  });
+
+  test('embeddingService starts as null', () => {
+    expect(aiService.embeddingService).toBeNull();
+    expect(aiService.getEmbeddingService()).toBeNull();
+  });
+
+  test('setEmbeddingService attaches; getEmbeddingService reads back', () => {
+    aiService.setEmbeddingService(stub);
+    expect(aiService.embeddingService).toBe(stub);
+    expect(aiService.getEmbeddingService()).toBe(stub);
+  });
+
+  test('setEmbeddingService(null) detaches cleanly', () => {
+    aiService.setEmbeddingService(stub);
+    aiService.setEmbeddingService(null);
+    expect(aiService.getEmbeddingService()).toBeNull();
+  });
+
+  test('embed("string") → embedQuery', async () => {
+    aiService.setEmbeddingService(stub);
+    const out = await aiService.embed('hello');
+    expect(stub.embedQuery).toHaveBeenCalledWith('hello');
+    expect(stub.embedDocuments).not.toHaveBeenCalled();
+    expect(out).toBeInstanceOf(Float32Array);
+  });
+
+  test('embed([a, b]) → embedDocuments with default kind=document', async () => {
+    aiService.setEmbeddingService(stub);
+    const out = await aiService.embed(['a', 'b']);
+    expect(stub.embedDocuments).toHaveBeenCalledWith(['a', 'b'], 'document');
+    expect(stub.embedQuery).not.toHaveBeenCalled();
+    expect(out).toHaveLength(2);
+  });
+
+  test('embed([a, b], {kind:"query"}) forwards kind', async () => {
+    aiService.setEmbeddingService(stub);
+    await aiService.embed(['q1', 'q2'], { kind: 'query' });
+    expect(stub.embedDocuments).toHaveBeenCalledWith(['q1', 'q2'], 'query');
+  });
+
+  test('embed() without an attached service throws clearly', async () => {
+    await expect(aiService.embed('hi')).rejects.toThrow(/EmbeddingService/i);
+  });
+});
+
+describe('aiService — augmentSystemPrompt (auto-recall integration)', () => {
+  let aiService, stub;
+  beforeEach(() => {
+    aiService = makeService();
+    stub = makeStubEmbeddingService();
+    // The embedding service needs to look minimally real for the
+    // augmentation path: `isEnabled` + `config.autoRecall.enabled`.
+    stub.config = { autoRecall: { enabled: true, threshold: 0.5, topK: 3, maxTokens: 400, surfaces: ['memory', 'reminisce'] } };
+  });
+
+  test('returns the prompt unchanged when no service attached', async () => {
+    const out = await aiService.augmentSystemPrompt('SYSTEM', 'agent-1', 'hello world');
+    expect(out).toBe('SYSTEM');
+  });
+
+  test('returns the prompt unchanged when service is disabled', async () => {
+    stub.isEnabled = false;
+    aiService.setEmbeddingService(stub);
+    const out = await aiService.augmentSystemPrompt('SYSTEM', 'agent-1', 'hello world');
+    expect(out).toBe('SYSTEM');
+  });
+
+  test('returns the prompt unchanged when recentText is empty', async () => {
+    aiService.setEmbeddingService(stub);
+    expect(await aiService.augmentSystemPrompt('SYSTEM', 'agent-1', '')).toBe('SYSTEM');
+    expect(await aiService.augmentSystemPrompt('SYSTEM', 'agent-1', '   ')).toBe('SYSTEM');
+    expect(await aiService.augmentSystemPrompt('SYSTEM', 'agent-1', null)).toBe('SYSTEM');
+  });
+
+  test('non-string systemPrompt is treated as empty (defensive)', async () => {
+    aiService.setEmbeddingService(stub);
+    const out = await aiService.augmentSystemPrompt(null, 'agent-1', 'text');
+    // No items below indexer stubs → empty section → empty string.
+    expect(out).toBe('');
+  });
+
+  test('never throws even on internal failures', async () => {
+    aiService.setEmbeddingService(stub);
+    // No indexers wired through embeddings service factory — should still resolve.
+    await expect(aiService.augmentSystemPrompt('SYSTEM', 'agent-1', 'hello')).resolves.toBeDefined();
+  });
+
+  test('appends a labelled section when items cross the threshold', async () => {
+    // Inject a stub that pretends to be a fully-working indexer.
+    aiService.setEmbeddingService(stub);
+    aiService._memoryIndexerForRecall = {
+      isEnabled: true,
+      search: jest.fn(async () => ({
+        enabled: true,
+        hits: [{ id: 'm1', score: 0.9, metadata: { text: 'a remembered thing', title: 'Note', createdAt: '2026-05-25T00:00:00Z' } }],
+      })),
+    };
+    aiService._reminisceIndexerForRecall = {
+      isEnabled: true,
+      search: jest.fn(async () => ({ enabled: true, hits: [] })),
+    };
+    const out = await aiService.augmentSystemPrompt('SYSTEM', 'agent-1', 'tell me about that');
+    expect(out).toContain('SYSTEM');
+    expect(out).toContain('## RELEVANT CONTEXT (auto-recalled)');
+    expect(out).toContain('a remembered thing');
+  });
+
+  test('appends nothing when indexers return no items above threshold', async () => {
+    aiService.setEmbeddingService(stub);
+    aiService._memoryIndexerForRecall = {
+      isEnabled: true,
+      search: jest.fn(async () => ({ enabled: true, hits: [{ id: 'm1', score: 0.1, metadata: { text: 'irrelevant' } }] })),
+    };
+    aiService._reminisceIndexerForRecall = { isEnabled: true, search: jest.fn(async () => ({ enabled: true, hits: [] })) };
+    const out = await aiService.augmentSystemPrompt('SYSTEM', 'agent-1', 'query');
+    expect(out).toBe('SYSTEM');
+  });
+
+  test('per-turn opts override config (e.g. higher threshold)', async () => {
+    aiService.setEmbeddingService(stub);
+    aiService._memoryIndexerForRecall = {
+      isEnabled: true,
+      search: jest.fn(async () => ({
+        enabled: true,
+        hits: [{ id: 'm1', score: 0.7, metadata: { text: 'borderline', title: 'note' } }],
+      })),
+    };
+    aiService._reminisceIndexerForRecall = { isEnabled: true, search: jest.fn(async () => ({ enabled: true, hits: [] })) };
+    // Per-call threshold 0.8 should drop the 0.7 hit.
+    const out = await aiService.augmentSystemPrompt('SYSTEM', 'agent-1', 'query', { threshold: 0.8 });
+    expect(out).toBe('SYSTEM');
+  });
+});

package/src/services/__tests__/apiKeyManager.getEmbeddingApiKey.test.js

@@ -0,0 +1,86 @@
+/**
+ * Tests for ApiKeyManager.getEmbeddingApiKey — the per-provider key
+ * accessor wired into EmbeddingService.
+ *
+ * Pins the wiring contract:
+ *   - azure   → loxiaApiKey (Loxia-backend proxy auth)
+ *   - openai  → vendorKeys.openai (BYOK direct-to-OpenAI)
+ *   - local   → null (no key needed)
+ *   - disabled→ null
+ *   - unknown → null (defensive; no throw)
+ */
+
+import { describe, test, expect, beforeEach, jest } from '@jest/globals';
+
+// Avoid touching disk; mock persistence helpers so apiKeyManager can be
+// instantiated in pure-memory mode.
+jest.unstable_mockModule('../../utilities/userDataDir.js', () => ({
+  ensureUserDataDirs: jest.fn(async () => ({ settings: '/tmp/fake' })),
+  getUserDataPaths: jest.fn(() => ({ settings: '/tmp/fake' })),
+}));
+
+const { default: ApiKeyManager } = await import('../apiKeyManager.js');
+
+describe('ApiKeyManager.getEmbeddingApiKey', () => {
+  let mgr;
+  beforeEach(() => {
+    mgr = new ApiKeyManager(null);
+    mgr.keys = {
+      loxiaApiKey: 'loxia-abc',
+      vendorKeys: { openai: 'sk-openai-xyz' },
+    };
+  });
+
+  test('azure → loxiaApiKey', () => {
+    expect(mgr.getEmbeddingApiKey('azure')).toBe('loxia-abc');
+  });
+
+  test('openai → vendorKeys.openai', () => {
+    expect(mgr.getEmbeddingApiKey('openai')).toBe('sk-openai-xyz');
+  });
+
+  test('azure-custom → vendorKeys.azureCustom', () => {
+    mgr.keys.vendorKeys.azureCustom = 'sk-azure-custom-123';
+    expect(mgr.getEmbeddingApiKey('azure-custom')).toBe('sk-azure-custom-123');
+  });
+
+  test('azure-custom with no key configured → null', () => {
+    expect(mgr.getEmbeddingApiKey('azure-custom')).toBeNull();
+  });
+
+  test('local → null (no key needed)', () => {
+    expect(mgr.getEmbeddingApiKey('local')).toBeNull();
+  });
+
+  test('disabled → null', () => {
+    expect(mgr.getEmbeddingApiKey('disabled')).toBeNull();
+  });
+
+  test('unknown provider → null (defensive, no throw)', () => {
+    expect(mgr.getEmbeddingApiKey('cohere')).toBeNull();
+    expect(mgr.getEmbeddingApiKey(null)).toBeNull();
+    expect(mgr.getEmbeddingApiKey(undefined)).toBeNull();
+  });
+
+  test('missing loxiaApiKey → azure returns null (not undefined)', () => {
+    mgr.keys.loxiaApiKey = null;
+    expect(mgr.getEmbeddingApiKey('azure')).toBeNull();
+  });
+
+  test('missing openai vendor key → openai returns null', () => {
+    mgr.keys.vendorKeys = {};
+    expect(mgr.getEmbeddingApiKey('openai')).toBeNull();
+  });
+
+  test('missing vendorKeys map entirely → openai returns null', () => {
+    mgr.keys.vendorKeys = undefined;
+    expect(mgr.getEmbeddingApiKey('openai')).toBeNull();
+  });
+
+  test('keys object entirely missing → all providers return null', () => {
+    mgr.keys = null;
+    for (const p of ['azure', 'openai', 'local', 'disabled']) {
+      expect(mgr.getEmbeddingApiKey(p)).toBeNull();
+    }
+  });
+});