thzxx
c727ad075c
背景：v0.5.0 发布后对工具调用链路（adapter 流式 tool_call → 引擎 PARSING → preToolHooks 管道 → ToolRegistry → 结果回填）做全链路复检，发现并修复两处问题。 安全修复: - SubAgent 中止时 pending 确认未清理（安全回归）：SubEngine 以 taskId 为 sessionId 写入 ConfirmationHook，abortSession 的 clearPending(sessionId) 清不到它们。后果：中止会话后残留弹框若被补批，孤儿工具会真实执行副作用 （v0.4.x 的全局清空反而能拦截）。修复：abortByParent 返回被中止的 taskId 列表，abortSession 一并 clearPending(taskId)。 确认弹框会话隔离（前端对齐后端 v0.5.0 语义）: - ConfirmationRequest 新增 sessionId 字段（主会话为 sessionId，SubAgent 为 taskId），弹框在会话 INIT/TERMINATED 时只清除该会话的请求 —— 修复并发 会话下任意会话结束误清其他会话等待中确认的问题 - 选中计数按当前 requests 收敛（selectedIds 残留 id 无害化） 测试（207 → 215 用例）: - 新增引擎级工具调用链路集成测试 ×6（engine-toolchain.test.ts）：真实 PermissionCheckHook + RateLimitHook + ConfirmationHook 管道 + 真实 ToolRegistry，覆盖 SAFE 工具直通 / HIGH 工具批准执行 / 拒绝以 Blocked 错误回传 LLM / 会话隔离端到端 / 工具异常不中断循环 / 策略拦截系统路径 - 新增 FTS 触发器 × NULL content 删除安全性测试 ×2：验证 truncateAfter （编辑重发/重新生成）删除 content=NULL 的 assistant 消息（模型仅发 tool_calls 的标准场景）不抛错且索引保持一致 - agent.test.ts 的 orchestrator mock 适配 abortByParent 新返回类型 验证: lint 0 problems / typecheck 双工程 0 errors / test:electron 215 全过 / build 成功
367 lines
13 KiB
TypeScript
367 lines
13 KiB
TypeScript
/**
|
||
* AgentLoopEngine 工具调用链路集成测试（v0.5.1 复检）
|
||
*
|
||
* 验证「模型工具调用能力」全链路：
|
||
* adapter 流式 tool_call 事件 → 引擎 PARSING 组装 → preToolHooks 真实管道
|
||
* （PermissionCheckHook + RateLimitHook + ConfirmationHook）→ ToolRegistry 执行
|
||
* → 工具结果回填消息历史 → 下一轮 LLM 调用 → COMPLETED。
|
||
*
|
||
* 重点覆盖 v0.5.0 会话隔离改动对工具链的影响：
|
||
* - sessionId 从 engine.currentSessionId → hooks → ConfirmationHook 记忆键的正确传递
|
||
* - 会话 A 的拒绝记忆不阻断会话 B 的工具（端到端，非 Hook 单元级）
|
||
* - 工具被 Hook 拦截时结果以 Blocked 错误回传 LLM（不中断循环）
|
||
*
|
||
* 运行：无需 SQLite，系统 Node 即可（npm test）。
|
||
*/
|
||
|
||
import { describe, it, expect, vi } from 'vitest';
|
||
|
||
vi.mock('electron-log', () => ({
|
||
default: { info: vi.fn(), warn: vi.fn(), error: vi.fn(), debug: vi.fn() },
|
||
}));
|
||
|
||
import type { BrowserWindow } from 'electron';
|
||
import { AgentLoopEngine } from '../engine';
|
||
import { TerminationReason } from '../types';
|
||
import type {
|
||
IMetonaProviderAdapter,
|
||
MetonaResponse,
|
||
MetonaStreamEvent,
|
||
MetonaToolDef,
|
||
} from '../../types';
|
||
import { MetonaStreamEventType, MetonaToolCategory, MetonaRiskLevel } from '../../types';
|
||
import { ToolRegistry } from '../../tools/registry';
|
||
import type { IMetonaTool } from '../../types/metona-tool';
|
||
import { PermissionCheckHook, RateLimitHook } from '../../hooks/pre-tool';
|
||
import { ConfirmationHook } from '../../hooks/confirmation-hook';
|
||
import { PolicyEngine } from '../../sandbox/permissions';
|
||
|
||
// ===== Mock Adapter（流式 tool_call → 文本收尾） =====
|
||
|
||
function createMockAdapter(scripts: MetonaStreamEvent[][]): IMetonaProviderAdapter {
|
||
let call = 0;
|
||
return {
|
||
providerId: 'mock',
|
||
supportedModels: ['mock-model'],
|
||
supportsToolCalling: true,
|
||
supportsThinking: false,
|
||
getContextWindow: () => 1_000_000,
|
||
send: vi.fn(
|
||
async (): Promise<MetonaResponse> => ({
|
||
meta: {
|
||
requestId: 'r_test',
|
||
provider: 'mock',
|
||
model: 'mock-model',
|
||
latencyMs: 1,
|
||
timestamp: Date.now(),
|
||
},
|
||
content: 'ok',
|
||
usage: { inputTokens: 10, outputTokens: 5, totalTokens: 15 },
|
||
finishReason: 'stop' as never,
|
||
}),
|
||
),
|
||
sendStream: vi.fn(async function* (): AsyncIterable<MetonaStreamEvent> {
|
||
const script = scripts[call % scripts.length];
|
||
call++;
|
||
for (const ev of script) yield ev;
|
||
}),
|
||
setAbortSignal: vi.fn(),
|
||
healthCheck: async () => true,
|
||
};
|
||
}
|
||
|
||
function toolCallEvent(name: string, args: Record<string, unknown>): MetonaStreamEvent[] {
|
||
return [
|
||
{
|
||
type: MetonaStreamEventType.TOOL_CALL_COMPLETE,
|
||
requestId: 'r1',
|
||
sessionId: 's1',
|
||
iteration: 1,
|
||
seq: 0,
|
||
timestamp: Date.now(),
|
||
toolCall: {
|
||
id: `tc_${name}_${Date.now()}_${Math.random().toString(36).slice(2, 6)}`,
|
||
name,
|
||
args,
|
||
iteration: 1,
|
||
timestamp: Date.now(),
|
||
},
|
||
},
|
||
{
|
||
type: MetonaStreamEventType.DONE,
|
||
requestId: 'r1',
|
||
sessionId: 's1',
|
||
iteration: 1,
|
||
seq: 1,
|
||
timestamp: Date.now(),
|
||
},
|
||
];
|
||
}
|
||
|
||
function textDoneEvent(text: string): MetonaStreamEvent[] {
|
||
return [
|
||
{
|
||
type: MetonaStreamEventType.TEXT_DELTA,
|
||
requestId: 'r1',
|
||
sessionId: 's1',
|
||
iteration: 1,
|
||
seq: 0,
|
||
timestamp: Date.now(),
|
||
delta: text,
|
||
},
|
||
{
|
||
type: MetonaStreamEventType.DONE,
|
||
requestId: 'r1',
|
||
sessionId: 's1',
|
||
iteration: 1,
|
||
seq: 1,
|
||
timestamp: Date.now(),
|
||
},
|
||
];
|
||
}
|
||
|
||
// ===== 真实工具（记录执行 + 返回结果） =====
|
||
|
||
const executedTools: Array<{ name: string; args: Record<string, unknown>; sessionId: string }> = [];
|
||
|
||
function makeTool(name: string): IMetonaTool {
|
||
return {
|
||
definition: {
|
||
name,
|
||
description: `${name} (test fixture)`,
|
||
parameters: { type: 'object', properties: {}, required: [] },
|
||
category: MetonaToolCategory.FILESYSTEM,
|
||
riskLevel: MetonaRiskLevel.SAFE,
|
||
requiresPermission: false,
|
||
timeoutMs: 5_000,
|
||
},
|
||
async execute(args, context) {
|
||
executedTools.push({ name, args, sessionId: context.sessionId });
|
||
return { ok: true, tool: name };
|
||
},
|
||
};
|
||
}
|
||
|
||
const userMessage = { role: 'user' as const, content: 'run the tool', timestamp: Date.now() };
|
||
const systemPrompt = { roleDefinition: '', outputConstraints: '', safetyGuidelines: '' };
|
||
|
||
/** run_command 的 HIGH 风险工具定义（供 ConfirmationHook 查询） */
|
||
const HIGH_RISK_DEF: MetonaToolDef = {
|
||
name: 'run_command',
|
||
description: 'Execute shell command (test fixture)',
|
||
parameters: { type: 'object', properties: {}, required: [] },
|
||
category: MetonaToolCategory.CODE_EXECUTION,
|
||
riskLevel: MetonaRiskLevel.HIGH,
|
||
requiresPermission: true,
|
||
timeoutMs: 5_000,
|
||
};
|
||
|
||
function makeMockWindow(): BrowserWindow {
|
||
return {
|
||
isDestroyed: () => false,
|
||
webContents: { send: vi.fn() },
|
||
} as unknown as BrowserWindow;
|
||
}
|
||
|
||
/** 构造带真实 Hook 管道的引擎（PermissionCheck + RateLimit + Confirmation） */
|
||
function makeEngineWithHooks(
|
||
adapterScripts: MetonaStreamEvent[][],
|
||
hook: ConfirmationHook,
|
||
maxIterations = 5,
|
||
): AgentLoopEngine {
|
||
const registry = new ToolRegistry();
|
||
registry.registerBuiltin(makeTool('read_file'));
|
||
registry.registerBuiltin(makeTool('run_command'));
|
||
|
||
return new AgentLoopEngine(
|
||
{ maxIterations },
|
||
createMockAdapter(adapterScripts),
|
||
registry,
|
||
[new PermissionCheckHook(new PolicyEngine()), new RateLimitHook(100), hook],
|
||
[],
|
||
);
|
||
}
|
||
|
||
describe('AgentLoopEngine 工具调用链路（adapter → 引擎 → 真实 Hook 管道 → registry）', () => {
|
||
it('SAFE 工具免确认直接执行：结果回填 + 工具上下文 sessionId 正确', async () => {
|
||
executedTools.length = 0;
|
||
const hook = new ConfirmationHook(makeMockWindow(), null);
|
||
|
||
const engine = makeEngineWithHooks(
|
||
[toolCallEvent('read_file', { file_path: 'a.ts' }), textDoneEvent('done after tool')],
|
||
hook,
|
||
);
|
||
|
||
const output = await engine.runStream(userMessage, 'sess-chain-1', [], systemPrompt);
|
||
|
||
// 循环完成：工具轮 → 文本轮
|
||
expect(output.terminationReason).toBe(TerminationReason.COMPLETED);
|
||
expect(output.finalAnswer).toBe('done after tool');
|
||
expect(output.iterations).toHaveLength(2);
|
||
|
||
// 工具真实执行，且 ToolExecutionContext.sessionId 传递正确
|
||
expect(executedTools).toHaveLength(1);
|
||
expect(executedTools[0].name).toBe('read_file');
|
||
expect(executedTools[0].sessionId).toBe('sess-chain-1');
|
||
|
||
// 工具结果成功回填到迭代记录
|
||
const toolResult = output.iterations[0].toolResults?.[0];
|
||
expect(toolResult?.success).toBe(true);
|
||
expect(toolResult?.result).toEqual({ ok: true, tool: 'read_file' });
|
||
|
||
// 无确认请求产生（SAFE 工具）
|
||
expect(hook.getPendingConfirmations()).toHaveLength(0);
|
||
});
|
||
|
||
it('HIGH 风险工具经用户批准后执行（pending → approve → 工具运行）', async () => {
|
||
executedTools.length = 0;
|
||
const hook = new ConfirmationHook(makeMockWindow(), null);
|
||
hook.setToolDefs([HIGH_RISK_DEF]);
|
||
|
||
const engine = makeEngineWithHooks(
|
||
[toolCallEvent('run_command', { command: 'ls' }), textDoneEvent('approved and done')],
|
||
hook,
|
||
);
|
||
|
||
const runPromise = engine.runStream(userMessage, 'sess-chain-2', [], systemPrompt);
|
||
|
||
// 等待确认请求产生（引擎阻塞在 waitForConfirmation）
|
||
await vi.waitFor(() => expect(hook.getPendingConfirmations()).toHaveLength(1));
|
||
const pending = hook.getPendingConfirmations();
|
||
// pending 携带正确的 sessionId（v0.5.1）
|
||
expect(pending[0].sessionId).toBe('sess-chain-2');
|
||
expect(pending[0].toolName).toBe('run_command');
|
||
|
||
// 用户批准
|
||
hook.resolveConfirmation(pending[0].toolCallId, true, false, false);
|
||
const output = await runPromise;
|
||
|
||
expect(output.terminationReason).toBe(TerminationReason.COMPLETED);
|
||
// 工具在批准后执行
|
||
expect(executedTools).toHaveLength(1);
|
||
expect(executedTools[0].name).toBe('run_command');
|
||
expect(output.iterations[0].toolResults?.[0].success).toBe(true);
|
||
});
|
||
|
||
it('用户拒绝后工具以 Blocked 错误回传 LLM（循环继续，不崩溃）', async () => {
|
||
executedTools.length = 0;
|
||
const hook = new ConfirmationHook(makeMockWindow(), null);
|
||
hook.setToolDefs([HIGH_RISK_DEF]);
|
||
|
||
const engine = makeEngineWithHooks(
|
||
[toolCallEvent('run_command', { command: 'ls' }), textDoneEvent('denied, continuing')],
|
||
hook,
|
||
);
|
||
|
||
const runPromise = engine.runStream(userMessage, 'sess-chain-3', [], systemPrompt);
|
||
|
||
await vi.waitFor(() => expect(hook.getPendingConfirmations()).toHaveLength(1));
|
||
hook.resolveConfirmation(hook.getPendingConfirmations()[0].toolCallId, false, false, false);
|
||
const output = await runPromise;
|
||
|
||
expect(output.terminationReason).toBe(TerminationReason.COMPLETED);
|
||
// 工具未执行
|
||
expect(executedTools).toHaveLength(0);
|
||
// 拒绝以 Blocked 错误回传（LLM 下一轮可见，可调整策略）
|
||
const toolResult = output.iterations[0].toolResults?.[0];
|
||
expect(toolResult?.success).toBe(false);
|
||
expect(toolResult?.error).toContain('User denied');
|
||
});
|
||
|
||
it('会话隔离端到端：会话 A 的拒绝记忆阻断 A 的后续调用，不影响会话 B', async () => {
|
||
executedTools.length = 0;
|
||
const hook = new ConfirmationHook(makeMockWindow(), null);
|
||
hook.setToolDefs([HIGH_RISK_DEF]);
|
||
|
||
// --- 会话 A 第 1 次：拒绝 + remember ---
|
||
const engineA1 = makeEngineWithHooks(
|
||
[toolCallEvent('run_command', { command: 'ls' }), textDoneEvent('A1 done')],
|
||
hook,
|
||
);
|
||
const runA1 = engineA1.runStream(userMessage, 'sess-iso-a', [], systemPrompt);
|
||
await vi.waitFor(() => expect(hook.getPendingConfirmations()).toHaveLength(1));
|
||
hook.resolveConfirmation(hook.getPendingConfirmations()[0].toolCallId, false, true, false);
|
||
const outA1 = await runA1;
|
||
expect(outA1.iterations[0].toolResults?.[0].error).toContain('User denied');
|
||
|
||
// --- 会话 A 第 2 次：拒绝记忆直接阻断（无新 pending、无 LLM 等待） ---
|
||
const engineA2 = makeEngineWithHooks(
|
||
[toolCallEvent('run_command', { command: 'ls' }), textDoneEvent('A2 done')],
|
||
hook,
|
||
);
|
||
const outA2 = await engineA2.runStream(userMessage, 'sess-iso-a', [], systemPrompt);
|
||
expect(outA2.iterations[0].toolResults?.[0].error).toContain('previously denied');
|
||
// 记忆阻断不产生确认请求
|
||
expect(hook.getPendingConfirmations()).toHaveLength(0);
|
||
|
||
// --- 会话 B：同工具正常询问并批准（A 的拒绝记忆不影响 B） ---
|
||
const engineB = makeEngineWithHooks(
|
||
[toolCallEvent('run_command', { command: 'ls' }), textDoneEvent('B done')],
|
||
hook,
|
||
);
|
||
const runB = engineB.runStream(userMessage, 'sess-iso-b', [], systemPrompt);
|
||
await vi.waitFor(() => expect(hook.getPendingConfirmations()).toHaveLength(1));
|
||
const pendingB = hook.getPendingConfirmations();
|
||
expect(pendingB[0].sessionId).toBe('sess-iso-b');
|
||
hook.resolveConfirmation(pendingB[0].toolCallId, true, false, false);
|
||
const outB = await runB;
|
||
|
||
expect(outB.iterations[0].toolResults?.[0].success).toBe(true);
|
||
expect(executedTools.filter((t) => t.sessionId === 'sess-iso-b')).toHaveLength(1);
|
||
// A 会话的工具始终未执行
|
||
expect(executedTools.filter((t) => t.sessionId === 'sess-iso-a')).toHaveLength(0);
|
||
});
|
||
|
||
it('工具抛出异常时以错误结果回传（不中断引擎循环）', async () => {
|
||
executedTools.length = 0;
|
||
const registry = new ToolRegistry();
|
||
// read_file 已有 PolicyEngine 策略（SAFE，免确认）— 此处注册抛错实现验证异常路径
|
||
registry.registerBuiltin({
|
||
definition: {
|
||
name: 'read_file',
|
||
description: 'always throws (test fixture)',
|
||
parameters: { type: 'object', properties: {}, required: [] },
|
||
category: MetonaToolCategory.FILESYSTEM,
|
||
riskLevel: MetonaRiskLevel.SAFE,
|
||
requiresPermission: false,
|
||
timeoutMs: 5_000,
|
||
},
|
||
async execute() {
|
||
throw new Error('tool exploded');
|
||
},
|
||
});
|
||
|
||
const engine = new AgentLoopEngine(
|
||
{ maxIterations: 3 },
|
||
createMockAdapter([toolCallEvent('read_file', {}), textDoneEvent('recovered')]),
|
||
registry,
|
||
[new PermissionCheckHook(new PolicyEngine()), new RateLimitHook(100)],
|
||
[],
|
||
);
|
||
|
||
const output = await engine.runStream(userMessage, 'sess-chain-4', [], systemPrompt);
|
||
expect(output.terminationReason).toBe(TerminationReason.COMPLETED);
|
||
expect(output.finalAnswer).toBe('recovered');
|
||
const toolResult = output.iterations[0].toolResults?.[0];
|
||
expect(toolResult?.success).toBe(false);
|
||
expect(toolResult?.error).toContain('tool exploded');
|
||
});
|
||
|
||
it('PermissionCheckHook 拦截系统路径（read_file /etc 被策略拒绝）', async () => {
|
||
executedTools.length = 0;
|
||
const hook = new ConfirmationHook(makeMockWindow(), null);
|
||
|
||
const engine = makeEngineWithHooks(
|
||
[toolCallEvent('read_file', { file_path: '/etc/passwd' }), textDoneEvent('blocked path')],
|
||
hook,
|
||
);
|
||
|
||
const output = await engine.runStream(userMessage, 'sess-chain-5', [], systemPrompt);
|
||
expect(output.terminationReason).toBe(TerminationReason.COMPLETED);
|
||
// 工具未执行（被 deniedPatterns 拦截）
|
||
expect(executedTools).toHaveLength(0);
|
||
expect(output.iterations[0].toolResults?.[0].error).toContain('Blocked');
|
||
expect(output.iterations[0].toolResults?.[0].error).toContain('security policy');
|
||
});
|
||
});
|