Files
flynn/src/backends/native/orchestrator.test.ts
T
2026-02-13 21:19:02 -08:00

795 lines
25 KiB
TypeScript

import { describe, it, expect, vi, beforeEach } from 'vitest';
import { AgentOrchestrator } from './orchestrator.js';
import { ModelRouter } from '../../models/router.js';
import type { ChatResponse, ModelClient } from '../../models/types.js';
import { ToolRegistry, ToolExecutor } from '../../tools/index.js';
import { HookEngine } from '../../hooks/engine.js';
import { MemoryStore } from '../../memory/store.js';
import { mkdtempSync, rmSync } from 'fs';
import { tmpdir } from 'os';
import { join } from 'path';
describe('AgentOrchestrator', () => {
let mockDefaultClient: ModelClient;
let mockFastClient: ModelClient;
let mockComplexClient: ModelClient;
let mockRouter: ModelRouter;
const createMockClient = (name: string, inputTokens = 100, outputTokens = 50): ModelClient => ({
chat: vi.fn().mockResolvedValue({
content: `${name} response`,
stopReason: 'end_turn',
usage: { inputTokens, outputTokens },
}),
});
beforeEach(() => {
mockDefaultClient = createMockClient('default', 100, 50);
mockFastClient = createMockClient('fast', 50, 25);
mockComplexClient = createMockClient('complex', 200, 100);
mockRouter = new ModelRouter({
default: mockDefaultClient,
fast: mockFastClient,
complex: mockComplexClient,
fallbackChain: [],
});
});
const requireClient = (tier: 'default' | 'fast' | 'complex'): ModelClient => {
const client = mockRouter.getClient(tier);
if (!client) {
throw new Error(`Expected ${tier} model client to exist in test router`);
}
return client;
};
describe('delegate()', () => {
it('routes to the correct tier when specified', async () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are a helpful assistant.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
const result = await orchestrator.delegate({
tier: 'fast',
systemPrompt: 'Summarize this text',
message: 'This is a test message',
maxTokens: 1000,
});
expect(result.content).toBe('fast response');
expect(result.tier).toBe('fast');
});
it('includes tools when requested', async () => {
const mockToolRegistry = new ToolRegistry();
const hooks = new HookEngine({
confirm: ['*'],
log: [],
silent: [],
});
const mockToolExecutor = new ToolExecutor(mockToolRegistry, hooks);
const mockFastChatClient = requireClient('fast');
const mockFastChatFn = vi.fn().mockResolvedValue({
content: 'response with tools',
stopReason: 'end_turn',
usage: { inputTokens: 100, outputTokens: 50 },
} as ChatResponse);
Object.assign(mockFastChatClient, { chat: mockFastChatFn });
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
toolRegistry: mockToolRegistry,
toolExecutor: mockToolExecutor,
});
await orchestrator.delegate({
tier: 'fast',
systemPrompt: 'Use available tools',
message: 'Help me analyze data',
tools: true,
});
expect(mockFastChatFn).toHaveBeenCalled();
});
it('falls back to default tier when requested tier is unavailable', async () => {
const routerWithoutComplex = new ModelRouter({
default: mockDefaultClient,
fast: mockFastClient,
fallbackChain: [],
});
const orchestrator = new AgentOrchestrator({
modelRouter: routerWithoutComplex,
systemPrompt: 'You are a helpful assistant.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
const result = await orchestrator.delegate({
tier: 'complex',
systemPrompt: 'Analyze deeply',
message: 'This is complex',
});
expect(result.content).toBe('default response');
expect(result.tier).toBe('default');
});
it('tracks cumulative usage after delegate calls', async () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
await orchestrator.delegate({
tier: 'fast',
systemPrompt: 'Fast task',
message: 'Fast message',
});
await orchestrator.delegate({
tier: 'complex',
systemPrompt: 'Complex task',
message: 'Complex message',
});
await orchestrator.delegate({
tier: 'fast',
systemPrompt: 'Another fast task',
message: 'Another fast message',
});
const usage = orchestrator.getDelegationUsage();
expect(usage.fast).toEqual({
inputTokens: 100,
outputTokens: 50,
calls: 2,
});
expect(usage.complex).toEqual({
inputTokens: 200,
outputTokens: 100,
calls: 1,
});
expect(usage.default).toBeUndefined();
});
it('tracks usage across tiers correctly', async () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
await orchestrator.delegate({
tier: 'fast',
systemPrompt: 'Fast task',
message: 'Fast message',
});
await orchestrator.delegate({
tier: 'fast',
systemPrompt: 'Another fast task',
message: 'Another fast message',
});
const usage = orchestrator.getDelegationUsage();
expect(usage.fast.inputTokens).toBe(100);
expect(usage.fast.outputTokens).toBe(50);
expect(usage.fast.calls).toBe(2);
});
it('logs delegation details with tier and token counts', async () => {
const consoleSpy = vi.spyOn(console, 'log').mockImplementation(() => {});
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
await orchestrator.delegate({
tier: 'fast',
systemPrompt: 'Fast task',
message: 'Fast message',
});
expect(consoleSpy).toHaveBeenCalledWith(
'[Flynn:delegate] tier=fast tokens=50+25',
);
consoleSpy.mockRestore();
});
});
describe('getDelegationTier()', () => {
it('returns correct tier for each task type', () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
expect(orchestrator.getDelegationTier('compaction')).toBe('fast');
expect(orchestrator.getDelegationTier('memory_extraction')).toBe('default');
expect(orchestrator.getDelegationTier('classification')).toBe('complex');
expect(orchestrator.getDelegationTier('tool_summarisation')).toBe('default');
expect(orchestrator.getDelegationTier('complex_reasoning')).toBe('complex');
});
it('returns tier that was explicitly configured', () => {
const customDelegation = {
compaction: 'local' as const,
memory_extraction: 'fast' as const,
classification: 'complex' as const,
tool_summarisation: 'default' as const,
complex_reasoning: 'local' as const,
};
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: customDelegation,
maxDelegationDepth: 10,
});
expect(orchestrator.getDelegationTier('compaction')).toBe('local');
expect(orchestrator.getDelegationTier('memory_extraction')).toBe('fast');
expect(orchestrator.getDelegationTier('complex_reasoning')).toBe('local');
});
});
describe('process()', () => {
it('proxies to NativeAgent for user messages', async () => {
const mockDefaultChatClient = requireClient('default');
const mockDefaultChatFn = vi.fn().mockResolvedValue({
content: 'Agent response',
stopReason: 'end_turn',
usage: { inputTokens: 150, outputTokens: 75 },
} as ChatResponse);
Object.assign(mockDefaultChatClient, { chat: mockDefaultChatFn });
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are a helpful agent.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
const response = await orchestrator.process('Hello, agent!');
expect(response).toBe('Agent response');
});
it('maintains conversation history through process()', async () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are a helpful agent.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
await orchestrator.process('Hello');
await orchestrator.process('How are you?');
await orchestrator.process('Tell me about yourself');
const history = orchestrator.getHistory();
expect(history).toHaveLength(6);
expect(history[0]).toEqual({ role: 'user', content: 'Hello' });
expect(history[1]).toEqual({ role: 'assistant', content: 'default response' });
expect(history[2]).toEqual({ role: 'user', content: 'How are you?' });
expect(history[3]).toEqual({ role: 'assistant', content: 'default response' });
expect(history[4]).toEqual({ role: 'user', content: 'Tell me about yourself' });
expect(history[5]).toEqual({ role: 'assistant', content: 'default response' });
});
it('uses adaptive memory injection strategy when configured', async () => {
const tempDir = mkdtempSync(join(tmpdir(), 'flynn-orchestrator-memory-'));
const memoryStore = new MemoryStore({ dir: tempDir, maxContextTokens: 2000 });
memoryStore.writeCategory('user', 'preferences', 'User prefers concise output.', 'replace');
const mockDefaultChatClient = requireClient('default');
const mockDefaultChatFn = vi.fn().mockResolvedValue({
content: 'Agent response',
stopReason: 'end_turn',
usage: { inputTokens: 50, outputTokens: 25 },
} as ChatResponse);
Object.assign(mockDefaultChatClient, { chat: mockDefaultChatFn });
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are a helpful agent.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
memoryStore,
memoryInjectionStrategy: 'adaptive',
memoryMaxInjectionTokens: 100,
});
await orchestrator.process('Keep this concise please');
expect(mockDefaultChatFn).toHaveBeenCalled();
const callArgs = mockDefaultChatFn.mock.calls[0][0];
expect(callArgs.system).toContain('# Memory Context');
expect(callArgs.system).toContain('concise');
rmSync(tempDir, { recursive: true, force: true });
});
it('falls back to default memory context when adaptive injection errors', async () => {
const tempDir = mkdtempSync(join(tmpdir(), 'flynn-orchestrator-memory-fallback-'));
const memoryStore = new MemoryStore({ dir: tempDir, maxContextTokens: 2000 });
memoryStore.write('user', 'Fallback memory content', 'replace');
const getPromptSectionsSpy = vi.spyOn(memoryStore, 'getPromptSections').mockImplementationOnce(() => {
throw new Error('boom');
});
const mockDefaultChatClient = requireClient('default');
const mockDefaultChatFn = vi.fn().mockResolvedValue({
content: 'Agent response',
stopReason: 'end_turn',
usage: { inputTokens: 50, outputTokens: 25 },
} as ChatResponse);
Object.assign(mockDefaultChatClient, { chat: mockDefaultChatFn });
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are a helpful agent.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
memoryStore,
memoryInjectionStrategy: 'adaptive',
memoryMaxInjectionTokens: 100,
});
await orchestrator.process('test message');
const callArgs = mockDefaultChatFn.mock.calls[0][0];
expect(callArgs.system).toContain('Fallback memory content');
getPromptSectionsSpy.mockRestore();
rmSync(tempDir, { recursive: true, force: true });
});
});
describe('reset()', () => {
it('clears primary agent conversation history', async () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are a helpful agent.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
await orchestrator.process('Hello');
await orchestrator.process('How are you?');
expect(orchestrator.getHistory()).toHaveLength(4);
orchestrator.reset();
expect(orchestrator.getHistory()).toHaveLength(0);
});
it('can be called multiple times', async () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are a helpful agent.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
await orchestrator.process('Hello');
orchestrator.reset();
expect(orchestrator.getHistory()).toHaveLength(0);
await orchestrator.process('World');
orchestrator.reset();
expect(orchestrator.getHistory()).toHaveLength(0);
});
});
describe('getDelegationUsage()', () => {
it('returns copy of usage stats (doesn\'t expose internal map)', async () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
await orchestrator.delegate({
tier: 'fast',
systemPrompt: 'Fast task',
message: 'Fast message',
});
const usage1 = orchestrator.getDelegationUsage();
const usage2 = orchestrator.getDelegationUsage();
expect(usage1).toEqual(usage2);
usage1.fast.inputTokens = 999;
expect(usage2.fast.inputTokens).toBe(50);
});
it('returns empty object when no usage tracked', async () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
const usage = orchestrator.getDelegationUsage();
expect(usage).toEqual({});
});
});
describe('getHistory()', () => {
it('returns conversation history from primary agent', async () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are a helpful agent.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
await orchestrator.process('Hello');
await orchestrator.process('How are you?');
const history = orchestrator.getHistory();
expect(history).toHaveLength(4);
expect(history[0]).toEqual({ role: 'user', content: 'Hello' });
expect(history[1]).toEqual({ role: 'assistant', content: 'default response' });
expect(history[2]).toEqual({ role: 'user', content: 'How are you?' });
expect(history[3]).toEqual({ role: 'assistant', content: 'default response' });
});
it('returns empty array when no history', async () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are a helpful agent.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
const history = orchestrator.getHistory();
expect(history).toEqual([]);
});
});
describe('process()', () => {
it('rolls back tool-loop provider errors, hard-trims on context overflow, and retries once', async () => {
let callCount = 0;
const mockClient: ModelClient = {
chat: vi.fn().mockImplementation(async () => {
callCount++;
if (callCount === 1) {
return {
content: '',
stopReason: 'tool_use',
usage: { inputTokens: 10, outputTokens: 5 },
toolCalls: [{ id: 'call_1', name: 'test.echo', args: { text: 'hi' } }],
} as ChatResponse;
}
if (callCount === 2) {
// Simulate llama.cpp context overflow buried inside an aggregated router error.
throw new Error(
'llama-server error (400): {"error":{"type":"exceedcontextsizeerror","nprompttokens":9183,"nctx":4096}}',
);
}
return {
content: 'ok',
stopReason: 'end_turn',
usage: { inputTokens: 10, outputTokens: 5 },
} as ChatResponse;
}),
};
const router = new ModelRouter({
default: mockClient,
fallbackChain: [],
});
// Minimal Session stub that supports rollback via replaceHistory().
const history: any[] = [];
const session = {
id: 'test',
addMessage: vi.fn((m: any) => { history.push(m); }),
getHistory: vi.fn(() => [...history]),
clear: vi.fn(() => { history.length = 0; }),
replaceHistory: vi.fn((msgs: any[]) => {
history.length = 0;
history.push(...msgs);
}),
getConfig: vi.fn(() => undefined),
setConfig: vi.fn(),
deleteConfig: vi.fn(),
} as any;
const registry = new ToolRegistry();
registry.register({
name: 'test.echo',
description: 'echo',
inputSchema: { type: 'object', properties: { text: { type: 'string' } }, required: ['text'] },
execute: async (args: any) => ({ success: true, output: String(args.text ?? '') }),
});
const hooks = new HookEngine({ confirm: [], log: [], silent: [] });
const executor = new ToolExecutor(registry, hooks);
const orchestrator = new AgentOrchestrator({
modelRouter: router,
systemPrompt: 'You are helpful.',
session,
toolRegistry: registry,
toolExecutor: executor,
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 3,
});
const res = await orchestrator.process('hello');
expect(res).toBe('ok');
// Ensure we didn't persist the low-level error string in history.
const textHistory = history
.map(m => (typeof m.content === 'string' ? m.content : ''))
.join('\n');
expect(textHistory).not.toContain('Error in tool loop');
});
});
describe('setModelTier()', () => {
it('sets model tier on primary agent', () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
orchestrator.setModelTier('fast');
expect(orchestrator.getModelTier()).toBe('fast');
});
it('allows tier changes after initialization', () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
expect(orchestrator.getModelTier()).toBe('default');
orchestrator.setModelTier('complex');
expect(orchestrator.getModelTier()).toBe('complex');
orchestrator.setModelTier('fast');
expect(orchestrator.getModelTier()).toBe('fast');
});
});
describe('setOnToolUse()', () => {
it('sets tool-use callback on primary agent', () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
const callback = vi.fn();
orchestrator.setOnToolUse(callback);
expect(orchestrator.getModelTier()).toBe('default');
});
it('allows callback changes', () => {
const orchestrator = new AgentOrchestrator({
modelRouter: mockRouter,
systemPrompt: 'You are helpful.',
primaryTier: 'default',
delegation: {
compaction: 'fast',
memory_extraction: 'default',
classification: 'complex',
tool_summarisation: 'default',
complex_reasoning: 'complex',
},
maxDelegationDepth: 10,
});
const callback1 = vi.fn();
const callback2 = vi.fn();
orchestrator.setOnToolUse(callback1);
expect(orchestrator.getModelTier()).toBe('default');
orchestrator.setOnToolUse(callback2);
expect(orchestrator.getModelTier()).toBe('default');
});
});
});