mirror of
https://github.com/QwenLM/qwen-code.git
synced 2025-12-21 09:17:53 +00:00
feat: subagent runtime & CLI display - wip
This commit is contained in:
tanzhenxin
721
packages/core/src/subagents/subagent.test.ts
Normal file
721
packages/core/src/subagents/subagent.test.ts
Normal file
@@ -0,0 +1,721 @@
|
||||
/**
|
||||
* @license
|
||||
* Copyright 2025 Google LLC
|
||||
* SPDX-License-Identifier: Apache-2.0
|
||||
*/
|
||||
|
||||
import { vi, describe, it, expect, beforeEach, Mock, afterEach } from 'vitest';
|
||||
import {
|
||||
ContextState,
|
||||
SubAgentScope,
|
||||
SubagentTerminateMode,
|
||||
PromptConfig,
|
||||
ModelConfig,
|
||||
RunConfig,
|
||||
ToolConfig,
|
||||
} from './subagent.js';
|
||||
import { Config, ConfigParameters } from '../config/config.js';
|
||||
import { GeminiChat } from '../core/geminiChat.js';
|
||||
import { createContentGenerator } from '../core/contentGenerator.js';
|
||||
import { getEnvironmentContext } from '../utils/environmentContext.js';
|
||||
import { executeToolCall } from '../core/nonInteractiveToolExecutor.js';
|
||||
import { ToolRegistry } from '../tools/tool-registry.js';
|
||||
import { DEFAULT_GEMINI_MODEL } from '../config/models.js';
|
||||
import {
|
||||
Content,
|
||||
FunctionCall,
|
||||
FunctionDeclaration,
|
||||
GenerateContentConfig,
|
||||
Type,
|
||||
} from '@google/genai';
|
||||
import { ToolErrorType } from '../tools/tool-error.js';
|
||||
|
||||
vi.mock('../core/geminiChat.js');
|
||||
vi.mock('../core/contentGenerator.js');
|
||||
vi.mock('../utils/environmentContext.js');
|
||||
vi.mock('../core/nonInteractiveToolExecutor.js');
|
||||
vi.mock('../ide/ide-client.js');
|
||||
|
||||
async function createMockConfig(
|
||||
toolRegistryMocks = {},
|
||||
): Promise<{ config: Config; toolRegistry: ToolRegistry }> {
|
||||
const configParams: ConfigParameters = {
|
||||
sessionId: 'test-session',
|
||||
model: DEFAULT_GEMINI_MODEL,
|
||||
targetDir: '.',
|
||||
debugMode: false,
|
||||
cwd: process.cwd(),
|
||||
};
|
||||
const config = new Config(configParams);
|
||||
await config.initialize();
|
||||
// eslint-disable-next-line @typescript-eslint/no-explicit-any
|
||||
await config.refreshAuth('test-auth' as any);
|
||||
|
||||
// Mock ToolRegistry
|
||||
const mockToolRegistry = {
|
||||
getTool: vi.fn(),
|
||||
getFunctionDeclarations: vi.fn().mockReturnValue([]),
|
||||
getFunctionDeclarationsFiltered: vi.fn().mockReturnValue([]),
|
||||
...toolRegistryMocks,
|
||||
} as unknown as ToolRegistry;
|
||||
|
||||
vi.spyOn(config, 'getToolRegistry').mockReturnValue(mockToolRegistry);
|
||||
return { config, toolRegistry: mockToolRegistry };
|
||||
}
|
||||
|
||||
// Helper to simulate LLM responses (sequence of tool calls over multiple turns)
|
||||
const createMockStream = (
|
||||
functionCallsList: Array<FunctionCall[] | 'stop'>,
|
||||
) => {
|
||||
let index = 0;
|
||||
return vi.fn().mockImplementation(() => {
|
||||
const response = functionCallsList[index] || 'stop';
|
||||
index++;
|
||||
return (async function* () {
|
||||
if (response === 'stop') {
|
||||
// When stopping, the model might return text, but the subagent logic primarily cares about the absence of functionCalls.
|
||||
yield {
|
||||
candidates: [
|
||||
{
|
||||
content: {
|
||||
parts: [{ text: 'Done.' }],
|
||||
},
|
||||
},
|
||||
],
|
||||
};
|
||||
} else if (response.length > 0) {
|
||||
yield { functionCalls: response };
|
||||
} else {
|
||||
yield {
|
||||
candidates: [
|
||||
{
|
||||
content: {
|
||||
parts: [{ text: 'Done.' }],
|
||||
},
|
||||
},
|
||||
],
|
||||
}; // Handle empty array also as stop
|
||||
}
|
||||
})();
|
||||
});
|
||||
};
|
||||
|
||||
describe('subagent.ts', () => {
|
||||
describe('ContextState', () => {
|
||||
it('should set and get values correctly', () => {
|
||||
const context = new ContextState();
|
||||
context.set('key1', 'value1');
|
||||
context.set('key2', 123);
|
||||
expect(context.get('key1')).toBe('value1');
|
||||
expect(context.get('key2')).toBe(123);
|
||||
expect(context.get_keys()).toEqual(['key1', 'key2']);
|
||||
});
|
||||
|
||||
it('should return undefined for missing keys', () => {
|
||||
const context = new ContextState();
|
||||
expect(context.get('missing')).toBeUndefined();
|
||||
});
|
||||
});
|
||||
|
||||
describe('SubAgentScope', () => {
|
||||
let mockSendMessageStream: Mock;
|
||||
|
||||
const defaultModelConfig: ModelConfig = {
|
||||
model: 'gemini-1.5-flash-latest',
|
||||
temp: 0.5, // Specific temp to test override
|
||||
top_p: 1,
|
||||
};
|
||||
|
||||
const defaultRunConfig: RunConfig = {
|
||||
max_time_minutes: 5,
|
||||
max_turns: 10,
|
||||
};
|
||||
|
||||
beforeEach(async () => {
|
||||
vi.clearAllMocks();
|
||||
|
||||
vi.mocked(getEnvironmentContext).mockResolvedValue([
|
||||
{ text: 'Env Context' },
|
||||
]);
|
||||
vi.mocked(createContentGenerator).mockResolvedValue({
|
||||
getGenerativeModel: vi.fn(),
|
||||
// eslint-disable-next-line @typescript-eslint/no-explicit-any
|
||||
} as any);
|
||||
|
||||
mockSendMessageStream = vi.fn();
|
||||
// We mock the implementation of the constructor.
|
||||
vi.mocked(GeminiChat).mockImplementation(
|
||||
() =>
|
||||
({
|
||||
sendMessageStream: mockSendMessageStream,
|
||||
}) as unknown as GeminiChat,
|
||||
);
|
||||
|
||||
// Default mock for executeToolCall
|
||||
vi.mocked(executeToolCall).mockResolvedValue({
|
||||
callId: 'default-call',
|
||||
responseParts: 'default response',
|
||||
resultDisplay: 'Default tool result',
|
||||
error: undefined,
|
||||
errorType: undefined,
|
||||
});
|
||||
});
|
||||
|
||||
afterEach(() => {
|
||||
vi.restoreAllMocks();
|
||||
});
|
||||
|
||||
// Helper to safely access generationConfig from mock calls
|
||||
const getGenerationConfigFromMock = (
|
||||
callIndex = 0,
|
||||
): GenerateContentConfig & { systemInstruction?: string | Content } => {
|
||||
const callArgs = vi.mocked(GeminiChat).mock.calls[callIndex];
|
||||
const generationConfig = callArgs?.[2];
|
||||
// Ensure it's defined before proceeding
|
||||
expect(generationConfig).toBeDefined();
|
||||
if (!generationConfig) throw new Error('generationConfig is undefined');
|
||||
return generationConfig as GenerateContentConfig & {
|
||||
systemInstruction?: string | Content;
|
||||
};
|
||||
};
|
||||
|
||||
describe('create (Tool Validation)', () => {
|
||||
const promptConfig: PromptConfig = { systemPrompt: 'Test prompt' };
|
||||
|
||||
it('should create a SubAgentScope successfully with minimal config', async () => {
|
||||
const { config } = await createMockConfig();
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
);
|
||||
expect(scope).toBeInstanceOf(SubAgentScope);
|
||||
});
|
||||
|
||||
it('should throw an error if a tool requires confirmation', async () => {
|
||||
const mockTool = {
|
||||
schema: { parameters: { type: Type.OBJECT, properties: {} } },
|
||||
build: vi.fn().mockReturnValue({
|
||||
shouldConfirmExecute: vi.fn().mockResolvedValue({
|
||||
type: 'exec',
|
||||
title: 'Confirm',
|
||||
command: 'rm -rf /',
|
||||
}),
|
||||
}),
|
||||
};
|
||||
|
||||
const { config } = await createMockConfig({
|
||||
// eslint-disable-next-line @typescript-eslint/no-explicit-any
|
||||
getTool: vi.fn().mockReturnValue(mockTool as any),
|
||||
});
|
||||
|
||||
const toolConfig: ToolConfig = { tools: ['risky_tool'] };
|
||||
|
||||
await expect(
|
||||
SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
toolConfig,
|
||||
),
|
||||
).rejects.toThrow(
|
||||
'Tool "risky_tool" requires user confirmation and cannot be used in a non-interactive subagent.',
|
||||
);
|
||||
});
|
||||
|
||||
it('should succeed if tools do not require confirmation', async () => {
|
||||
const mockTool = {
|
||||
schema: { parameters: { type: Type.OBJECT, properties: {} } },
|
||||
build: vi.fn().mockReturnValue({
|
||||
shouldConfirmExecute: vi.fn().mockResolvedValue(null),
|
||||
}),
|
||||
};
|
||||
const { config } = await createMockConfig({
|
||||
// eslint-disable-next-line @typescript-eslint/no-explicit-any
|
||||
getTool: vi.fn().mockReturnValue(mockTool as any),
|
||||
});
|
||||
|
||||
const toolConfig: ToolConfig = { tools: ['safe_tool'] };
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
toolConfig,
|
||||
);
|
||||
expect(scope).toBeInstanceOf(SubAgentScope);
|
||||
});
|
||||
|
||||
it('should skip interactivity check and warn for tools with required parameters', async () => {
|
||||
const consoleWarnSpy = vi
|
||||
.spyOn(console, 'warn')
|
||||
.mockImplementation(() => {});
|
||||
|
||||
const mockToolWithParams = {
|
||||
schema: {
|
||||
parameters: {
|
||||
type: Type.OBJECT,
|
||||
properties: {
|
||||
path: { type: Type.STRING },
|
||||
},
|
||||
required: ['path'],
|
||||
},
|
||||
},
|
||||
// build should not be called, but we mock it to be safe
|
||||
build: vi.fn(),
|
||||
};
|
||||
|
||||
const { config } = await createMockConfig({
|
||||
getTool: vi.fn().mockReturnValue(mockToolWithParams),
|
||||
});
|
||||
|
||||
const toolConfig: ToolConfig = { tools: ['tool_with_params'] };
|
||||
|
||||
// The creation should succeed without throwing
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
toolConfig,
|
||||
);
|
||||
|
||||
expect(scope).toBeInstanceOf(SubAgentScope);
|
||||
|
||||
// Check that the warning was logged
|
||||
expect(consoleWarnSpy).toHaveBeenCalledWith(
|
||||
'Cannot check tool "tool_with_params" for interactivity because it requires parameters. Assuming it is safe for non-interactive use.',
|
||||
);
|
||||
|
||||
// Ensure build was never called
|
||||
expect(mockToolWithParams.build).not.toHaveBeenCalled();
|
||||
|
||||
consoleWarnSpy.mockRestore();
|
||||
});
|
||||
});
|
||||
|
||||
describe('runNonInteractive - Initialization and Prompting', () => {
|
||||
it('should correctly template the system prompt and initialize GeminiChat', async () => {
|
||||
const { config } = await createMockConfig();
|
||||
|
||||
vi.mocked(GeminiChat).mockClear();
|
||||
|
||||
const promptConfig: PromptConfig = {
|
||||
systemPrompt: 'Hello ${name}, your task is ${task}.',
|
||||
};
|
||||
const context = new ContextState();
|
||||
context.set('name', 'Agent');
|
||||
context.set('task', 'Testing');
|
||||
|
||||
// Model stops immediately
|
||||
mockSendMessageStream.mockImplementation(createMockStream(['stop']));
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
);
|
||||
|
||||
await scope.runNonInteractive(context);
|
||||
|
||||
// Check if GeminiChat was initialized correctly by the subagent
|
||||
expect(GeminiChat).toHaveBeenCalledTimes(1);
|
||||
const callArgs = vi.mocked(GeminiChat).mock.calls[0];
|
||||
|
||||
// Check Generation Config
|
||||
const generationConfig = getGenerationConfigFromMock();
|
||||
|
||||
// Check temperature override
|
||||
expect(generationConfig.temperature).toBe(defaultModelConfig.temp);
|
||||
expect(generationConfig.systemInstruction).toContain(
|
||||
'Hello Agent, your task is Testing.',
|
||||
);
|
||||
expect(generationConfig.systemInstruction).toContain(
|
||||
'Important Rules:',
|
||||
);
|
||||
|
||||
// Check History (should include environment context)
|
||||
const history = callArgs[3];
|
||||
expect(history).toEqual([
|
||||
{ role: 'user', parts: [{ text: 'Env Context' }] },
|
||||
{
|
||||
role: 'model',
|
||||
parts: [{ text: 'Got it. Thanks for the context!' }],
|
||||
},
|
||||
]);
|
||||
});
|
||||
|
||||
it('should use initialMessages instead of systemPrompt if provided', async () => {
|
||||
const { config } = await createMockConfig();
|
||||
vi.mocked(GeminiChat).mockClear();
|
||||
|
||||
const initialMessages: Content[] = [
|
||||
{ role: 'user', parts: [{ text: 'Hi' }] },
|
||||
];
|
||||
const promptConfig: PromptConfig = { initialMessages };
|
||||
const context = new ContextState();
|
||||
|
||||
// Model stops immediately
|
||||
mockSendMessageStream.mockImplementation(createMockStream(['stop']));
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
);
|
||||
|
||||
await scope.runNonInteractive(context);
|
||||
|
||||
const callArgs = vi.mocked(GeminiChat).mock.calls[0];
|
||||
const generationConfig = getGenerationConfigFromMock();
|
||||
const history = callArgs[3];
|
||||
|
||||
expect(generationConfig.systemInstruction).toBeUndefined();
|
||||
expect(history).toEqual([
|
||||
{ role: 'user', parts: [{ text: 'Env Context' }] },
|
||||
{
|
||||
role: 'model',
|
||||
parts: [{ text: 'Got it. Thanks for the context!' }],
|
||||
},
|
||||
...initialMessages,
|
||||
]);
|
||||
});
|
||||
|
||||
it('should throw an error if template variables are missing', async () => {
|
||||
const { config } = await createMockConfig();
|
||||
const promptConfig: PromptConfig = {
|
||||
systemPrompt: 'Hello ${name}, you are missing ${missing}.',
|
||||
};
|
||||
const context = new ContextState();
|
||||
context.set('name', 'Agent');
|
||||
// 'missing' is not set
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
);
|
||||
|
||||
// The error from templating causes the runNonInteractive to reject and the terminate_reason to be ERROR.
|
||||
await expect(scope.runNonInteractive(context)).rejects.toThrow(
|
||||
'Missing context values for the following keys: missing',
|
||||
);
|
||||
expect(scope.output.terminate_reason).toBe(SubagentTerminateMode.ERROR);
|
||||
});
|
||||
|
||||
it('should validate that systemPrompt and initialMessages are mutually exclusive', async () => {
|
||||
const { config } = await createMockConfig();
|
||||
const promptConfig: PromptConfig = {
|
||||
systemPrompt: 'System',
|
||||
initialMessages: [{ role: 'user', parts: [{ text: 'Hi' }] }],
|
||||
};
|
||||
const context = new ContextState();
|
||||
|
||||
const agent = await SubAgentScope.create(
|
||||
'TestAgent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
);
|
||||
|
||||
await expect(agent.runNonInteractive(context)).rejects.toThrow(
|
||||
'PromptConfig cannot have both `systemPrompt` and `initialMessages` defined.',
|
||||
);
|
||||
expect(agent.output.terminate_reason).toBe(SubagentTerminateMode.ERROR);
|
||||
});
|
||||
});
|
||||
|
||||
describe('runNonInteractive - Execution and Tool Use', () => {
|
||||
const promptConfig: PromptConfig = { systemPrompt: 'Execute task.' };
|
||||
|
||||
it('should terminate with GOAL if no outputs are expected and model stops', async () => {
|
||||
const { config } = await createMockConfig();
|
||||
// Model stops immediately
|
||||
mockSendMessageStream.mockImplementation(createMockStream(['stop']));
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
// No ToolConfig, No OutputConfig
|
||||
);
|
||||
|
||||
await scope.runNonInteractive(new ContextState());
|
||||
|
||||
expect(scope.output.terminate_reason).toBe(SubagentTerminateMode.GOAL);
|
||||
expect(scope.output.result).toBe('Done.');
|
||||
expect(mockSendMessageStream).toHaveBeenCalledTimes(1);
|
||||
// Check the initial message
|
||||
expect(mockSendMessageStream.mock.calls[0][0].message).toEqual([
|
||||
{ text: 'Get Started!' },
|
||||
]);
|
||||
});
|
||||
|
||||
it('should terminate with GOAL when model provides final text', async () => {
|
||||
const { config } = await createMockConfig();
|
||||
|
||||
// Model stops immediately with text response
|
||||
mockSendMessageStream.mockImplementation(createMockStream(['stop']));
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
);
|
||||
|
||||
await scope.runNonInteractive(new ContextState());
|
||||
|
||||
expect(scope.output.terminate_reason).toBe(SubagentTerminateMode.GOAL);
|
||||
expect(scope.output.result).toBe('Done.');
|
||||
expect(mockSendMessageStream).toHaveBeenCalledTimes(1);
|
||||
});
|
||||
|
||||
it('should execute external tools and provide the response to the model', async () => {
|
||||
const listFilesToolDef: FunctionDeclaration = {
|
||||
name: 'list_files',
|
||||
description: 'Lists files',
|
||||
parameters: { type: Type.OBJECT, properties: {} },
|
||||
};
|
||||
|
||||
const { config } = await createMockConfig({
|
||||
getFunctionDeclarationsFiltered: vi
|
||||
.fn()
|
||||
.mockReturnValue([listFilesToolDef]),
|
||||
});
|
||||
const toolConfig: ToolConfig = { tools: ['list_files'] };
|
||||
|
||||
// Turn 1: Model calls the external tool
|
||||
// Turn 2: Model stops
|
||||
mockSendMessageStream.mockImplementation(
|
||||
createMockStream([
|
||||
[
|
||||
{
|
||||
id: 'call_1',
|
||||
name: 'list_files',
|
||||
args: { path: '.' },
|
||||
},
|
||||
],
|
||||
'stop',
|
||||
]),
|
||||
);
|
||||
|
||||
// Mock the tool execution result
|
||||
vi.mocked(executeToolCall).mockResolvedValue({
|
||||
callId: 'call_1',
|
||||
responseParts: 'file1.txt\nfile2.ts',
|
||||
resultDisplay: 'Listed 2 files',
|
||||
error: undefined,
|
||||
errorType: undefined, // Or ToolErrorType.NONE if available and appropriate
|
||||
});
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
toolConfig,
|
||||
);
|
||||
|
||||
await scope.runNonInteractive(new ContextState());
|
||||
|
||||
// Check tool execution
|
||||
expect(executeToolCall).toHaveBeenCalledWith(
|
||||
config,
|
||||
expect.objectContaining({ name: 'list_files', args: { path: '.' } }),
|
||||
expect.any(AbortSignal),
|
||||
);
|
||||
|
||||
// Check the response sent back to the model
|
||||
const secondCallArgs = mockSendMessageStream.mock.calls[1][0];
|
||||
expect(secondCallArgs.message).toEqual([
|
||||
{ text: 'file1.txt\nfile2.ts' },
|
||||
]);
|
||||
|
||||
expect(scope.output.terminate_reason).toBe(SubagentTerminateMode.GOAL);
|
||||
});
|
||||
|
||||
it('should provide specific tool error responses to the model', async () => {
|
||||
const { config } = await createMockConfig();
|
||||
const toolConfig: ToolConfig = { tools: ['failing_tool'] };
|
||||
|
||||
// Turn 1: Model calls the failing tool
|
||||
// Turn 2: Model stops after receiving the error response
|
||||
mockSendMessageStream.mockImplementation(
|
||||
createMockStream([
|
||||
[
|
||||
{
|
||||
id: 'call_fail',
|
||||
name: 'failing_tool',
|
||||
args: {},
|
||||
},
|
||||
],
|
||||
'stop',
|
||||
]),
|
||||
);
|
||||
|
||||
// Mock the tool execution failure.
|
||||
vi.mocked(executeToolCall).mockResolvedValue({
|
||||
callId: 'call_fail',
|
||||
responseParts: 'ERROR: Tool failed catastrophically', // This should be sent to the model
|
||||
resultDisplay: 'Tool failed catastrophically',
|
||||
error: new Error('Failure'),
|
||||
errorType: ToolErrorType.INVALID_TOOL_PARAMS,
|
||||
});
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
toolConfig,
|
||||
);
|
||||
|
||||
await scope.runNonInteractive(new ContextState());
|
||||
|
||||
// The agent should send the specific error message from responseParts.
|
||||
const secondCallArgs = mockSendMessageStream.mock.calls[1][0];
|
||||
|
||||
expect(secondCallArgs.message).toEqual([
|
||||
{
|
||||
text: 'ERROR: Tool failed catastrophically',
|
||||
},
|
||||
]);
|
||||
});
|
||||
});
|
||||
|
||||
describe('runNonInteractive - Termination and Recovery', () => {
|
||||
const promptConfig: PromptConfig = { systemPrompt: 'Execute task.' };
|
||||
|
||||
it('should terminate with MAX_TURNS if the limit is reached', async () => {
|
||||
const { config } = await createMockConfig();
|
||||
const runConfig: RunConfig = { ...defaultRunConfig, max_turns: 2 };
|
||||
|
||||
// Model keeps calling tools repeatedly
|
||||
mockSendMessageStream.mockImplementation(
|
||||
createMockStream([
|
||||
[
|
||||
{
|
||||
name: 'list_files',
|
||||
args: { path: '/test' },
|
||||
},
|
||||
],
|
||||
[
|
||||
{
|
||||
name: 'list_files',
|
||||
args: { path: '/test2' },
|
||||
},
|
||||
],
|
||||
// This turn should not happen
|
||||
[
|
||||
{
|
||||
name: 'list_files',
|
||||
args: { path: '/test3' },
|
||||
},
|
||||
],
|
||||
]),
|
||||
);
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
runConfig,
|
||||
);
|
||||
|
||||
await scope.runNonInteractive(new ContextState());
|
||||
|
||||
expect(mockSendMessageStream).toHaveBeenCalledTimes(2);
|
||||
expect(scope.output.terminate_reason).toBe(
|
||||
SubagentTerminateMode.MAX_TURNS,
|
||||
);
|
||||
});
|
||||
|
||||
it('should terminate with TIMEOUT if the time limit is reached during an LLM call', async () => {
|
||||
// Use fake timers to reliably test timeouts
|
||||
vi.useFakeTimers();
|
||||
|
||||
const { config } = await createMockConfig();
|
||||
const runConfig: RunConfig = { max_time_minutes: 5, max_turns: 100 };
|
||||
|
||||
// We need to control the resolution of the sendMessageStream promise to advance the timer during execution.
|
||||
let resolveStream: (
|
||||
value: AsyncGenerator<unknown, void, unknown>,
|
||||
) => void;
|
||||
const streamPromise = new Promise<
|
||||
AsyncGenerator<unknown, void, unknown>
|
||||
>((resolve) => {
|
||||
// eslint-disable-next-line @typescript-eslint/no-explicit-any
|
||||
resolveStream = resolve as any;
|
||||
});
|
||||
|
||||
// The LLM call will hang until we resolve the promise.
|
||||
mockSendMessageStream.mockReturnValue(streamPromise);
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
runConfig,
|
||||
);
|
||||
|
||||
const runPromise = scope.runNonInteractive(new ContextState());
|
||||
|
||||
// Advance time beyond the limit (6 minutes) while the agent is awaiting the LLM response.
|
||||
await vi.advanceTimersByTimeAsync(6 * 60 * 1000);
|
||||
|
||||
// Now resolve the stream. The model returns 'stop'.
|
||||
// eslint-disable-next-line @typescript-eslint/no-explicit-any
|
||||
resolveStream!(createMockStream(['stop'])() as any);
|
||||
|
||||
await runPromise;
|
||||
|
||||
expect(scope.output.terminate_reason).toBe(
|
||||
SubagentTerminateMode.TIMEOUT,
|
||||
);
|
||||
expect(mockSendMessageStream).toHaveBeenCalledTimes(1);
|
||||
|
||||
vi.useRealTimers();
|
||||
});
|
||||
|
||||
it('should terminate with ERROR if the model call throws', async () => {
|
||||
const { config } = await createMockConfig();
|
||||
mockSendMessageStream.mockRejectedValue(new Error('API Failure'));
|
||||
|
||||
const scope = await SubAgentScope.create(
|
||||
'test-agent',
|
||||
config,
|
||||
promptConfig,
|
||||
defaultModelConfig,
|
||||
defaultRunConfig,
|
||||
);
|
||||
|
||||
await expect(
|
||||
scope.runNonInteractive(new ContextState()),
|
||||
).rejects.toThrow('API Failure');
|
||||
expect(scope.output.terminate_reason).toBe(SubagentTerminateMode.ERROR);
|
||||
});
|
||||
});
|
||||
});
|
||||
});
|
||||
Reference in New Issue
Block a user