/** * ConfirmationHook 单元测试(v0.4.1 测试补齐) * 覆盖:自动执行放行、会话内记忆(批准/拒绝)、拒绝记忆 TTL 过期、 * 超时拒绝、用户批准、批量审批、pending 管理、恢复询问接口 */ import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest'; import type { BrowserWindow } from 'electron'; import { ConfirmationHook } from '../confirmation-hook'; import type { MetonaToolCall, MetonaToolDef } from '../../types'; import { MetonaToolCategory, MetonaRiskLevel } from '../../types'; /** 需要确认的高风险工具定义 */ 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: 1_000, }; /** 低风险工具定义(无需确认) */ const SAFE_DEF: MetonaToolDef = { name: 'read_file', description: 'Read file (test fixture)', parameters: { type: 'object', properties: {}, required: [] }, category: MetonaToolCategory.FILE_SYSTEM, riskLevel: MetonaRiskLevel.SAFE, requiresPermission: false, timeoutMs: 1_000, }; /** 第二个需确认的高风险工具(用于记忆互不干扰的测试) */ const HIGH_RISK_DEF_2: MetonaToolDef = { name: 'delete_file', description: 'Delete file (test fixture)', parameters: { type: 'object', properties: {}, required: [] }, category: MetonaToolCategory.FILE_SYSTEM, riskLevel: MetonaRiskLevel.HIGH, requiresPermission: true, timeoutMs: 1_000, }; let idCounter = 0; function makeToolCall(name = 'run_command'): MetonaToolCall { return { id: `tc_${++idCounter}`, name, args: { command: 'ls' }, iteration: 1, timestamp: Date.now(), }; } function makeMockWindow(): BrowserWindow { return { isDestroyed: () => false, webContents: { send: vi.fn() }, } as unknown as BrowserWindow; } describe('ConfirmationHook — 免确认路径', () => { it('未注册工具定义时放行(由 ToolRegistry 处理未知工具错误)', async () => { const hook = new ConfirmationHook(null, null); const result = await hook.beforeExecute(makeToolCall('unknown_tool'), 'sess'); expect(result.blocked).toBe(false); }); it('无需确认的工具直接放行', async () => { const hook = new ConfirmationHook(null, null); hook.setToolDefs([SAFE_DEF]); const result = await hook.beforeExecute(makeToolCall('read_file'), 'sess'); expect(result.blocked).toBe(false); }); it('持久化自动执行(autoExecute)的工具放行', async () => { const hook = new ConfirmationHook(null, null); hook.setToolDefs([HIGH_RISK_DEF]); hook.setAutoExecute('run_command', true); const result = await hook.beforeExecute(makeToolCall(), 'sess'); expect(result.blocked).toBe(false); expect(hook.getAutoExecuteList()).toContain('run_command'); }); it('记住批准(remember approved)后同会话放行', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); // 第一次调用 → 等待确认 → 用户批准并记住 const p1 = hook.beforeExecute(makeToolCall(), 'sess'); const pending = hook.getPendingConfirmations(); expect(pending).toHaveLength(1); hook.resolveConfirmation(pending[0].toolCallId, true, true, false); expect((await p1).blocked).toBe(false); // 第二次调用 — 记住的批准直接放行 const result = await hook.beforeExecute(makeToolCall(), 'sess'); expect(result.blocked).toBe(false); }); }); describe('ConfirmationHook — 拒绝与阻断', () => { it('记住拒绝后同会话阻断(reason 含 previously denied)', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); const p1 = hook.beforeExecute(makeToolCall(), 'sess'); const pending = hook.getPendingConfirmations(); hook.resolveConfirmation(pending[0].toolCallId, false, true, false); expect((await p1).blocked).toBe(true); const result = await hook.beforeExecute(makeToolCall(), 'sess'); expect(result.blocked).toBe(true); expect(result.reason).toContain('previously denied'); }); it('拒绝记忆 TTL 过期后恢复询问(v0.4.1)', async () => { vi.useFakeTimers(); try { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); // 记住拒绝 const p1 = hook.beforeExecute(makeToolCall(), 'sess'); const pending1 = hook.getPendingConfirmations(); hook.resolveConfirmation(pending1[0].toolCallId, false, true, false); await p1; // 拒绝记忆立即生效 const blockedNow = await hook.beforeExecute(makeToolCall(), 'sess'); expect(blockedNow.blocked).toBe(true); expect(blockedNow.reason).toContain('previously denied'); // 快进 11 分钟(TTL 10 分钟)→ 拒绝记忆过期,恢复询问流程 vi.setSystemTime(Date.now() + 11 * 60 * 1000); // 撤掉主窗口,使询问流程以 'no main window' 阻断(证明走到了询问分支而非记忆分支) hook.setMainWindow(null as unknown as BrowserWindow); const result = await hook.beforeExecute(makeToolCall(), 'sess'); expect(result.blocked).toBe(true); expect(result.reason).toContain('no main window available'); // 过期记忆已被清理 expect(hook.getRememberedDenials()).toHaveLength(0); } finally { vi.useRealTimers(); } }); it('无主窗口时安全阻断(fail-closed)', async () => { const hook = new ConfirmationHook(null, null); hook.setToolDefs([HIGH_RISK_DEF]); const result = await hook.beforeExecute(makeToolCall(), 'sess'); expect(result.blocked).toBe(true); expect(result.reason).toContain('no main window available'); }); it('用户拒绝单次调用 → blocked 且 reason 含 User denied', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); const p = hook.beforeExecute(makeToolCall(), 'sess'); const pending = hook.getPendingConfirmations(); hook.resolveConfirmation(pending[0].toolCallId, false, false, false); const result = await p; expect(result.blocked).toBe(true); expect(result.reason).toContain('User denied'); }); }); describe('ConfirmationHook — 超时行为', () => { beforeEach(() => { vi.useFakeTimers(); }); afterEach(() => { vi.useRealTimers(); }); it('确认超时视为拒绝(blocked)', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); hook.setConfirmationTimeout(30_000); // 最小值 30s const p = hook.beforeExecute(makeToolCall(), 'sess'); // 快进超过超时时间 vi.advanceTimersByTime(31_000); const result = await p; expect(result.blocked).toBe(true); expect(result.reason).toContain('User denied'); // pending 已被超时清理 expect(hook.getPendingConfirmations()).toHaveLength(0); }); it('确认超时与用户点击的竞态:先到者赢(settled 标志)', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); hook.setConfirmationTimeout(30_000); const p = hook.beforeExecute(makeToolCall(), 'sess'); // timer 回调已入队但未执行时,用户点击批准 vi.advanceTimersByTime(30_000); // 超时已 resolve(false) — 后续 resolveConfirmation 无效果 const pending = hook.getPendingConfirmations(); expect(pending).toHaveLength(0); const result = await p; expect(result.blocked).toBe(true); }); }); describe('ConfirmationHook — 批量审批(v0.3.2)', () => { it('批量批准并行工具调用', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); const p1 = hook.beforeExecute(makeToolCall(), 'sess'); const p2 = hook.beforeExecute(makeToolCall(), 'sess'); expect(hook.getPendingConfirmations()).toHaveLength(2); const ids = hook.getPendingConfirmations().map((r) => r.toolCallId); const resolved = hook.resolveConfirmationsBatch(ids, true, false, false); expect(resolved).toHaveLength(2); expect((await p1).blocked).toBe(false); expect((await p2).blocked).toBe(false); }); it('批量拒绝 + 记住 → 同工具后续调用被记忆阻断', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); const p1 = hook.beforeExecute(makeToolCall(), 'sess'); const p2 = hook.beforeExecute(makeToolCall(), 'sess'); const ids = hook.getPendingConfirmations().map((r) => r.toolCallId); hook.resolveConfirmationsBatch(ids, false, true, false); expect((await p1).blocked).toBe(true); expect((await p2).blocked).toBe(true); const after = await hook.beforeExecute(makeToolCall(), 'sess'); expect(after.blocked).toBe(true); expect(after.reason).toContain('previously denied'); }); it('批量批准 + autoExecute → 写入持久化自动执行列表', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); const p = hook.beforeExecute(makeToolCall(), 'sess'); const ids = hook.getPendingConfirmations().map((r) => r.toolCallId); hook.resolveConfirmationsBatch(ids, true, false, true); expect((await p).blocked).toBe(false); expect(hook.getAutoExecuteList()).toContain('run_command'); }); }); describe('ConfirmationHook — 拒绝记忆管理接口(v0.4.1)', () => { it('getRememberedDenials 只返回拒绝记忆(含剩余时间)', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF, HIGH_RISK_DEF_2]); // 记住一个批准(run_command)、一个拒绝(delete_file) const pApprove = hook.beforeExecute(makeToolCall('run_command'), 'sess'); hook.resolveConfirmation(hook.getPendingConfirmations()[0].toolCallId, true, true, false); await pApprove; const pDeny = hook.beforeExecute(makeToolCall('delete_file'), 'sess'); hook.resolveConfirmation(hook.getPendingConfirmations()[0].toolCallId, false, true, false); await pDeny; const denials = hook.getRememberedDenials(); expect(denials).toHaveLength(1); expect(denials[0].toolName).toBe('delete_file'); expect(denials[0].expiresInSeconds).toBeGreaterThan(0); expect(denials[0].expiresInSeconds).toBeLessThanOrEqual(600); }); it('resetRememberedDenial 重置后恢复询问', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); const p = hook.beforeExecute(makeToolCall(), 'sess'); hook.resolveConfirmation(hook.getPendingConfirmations()[0].toolCallId, false, true, false); await p; expect(hook.getRememberedDenials()).toHaveLength(1); // 重置 → 拒绝记忆清空 expect(hook.resetRememberedDenial('run_command')).toBe(true); expect(hook.getRememberedDenials()).toHaveLength(0); // 后续调用恢复询问(有窗口 → 产生新 pending) const p2 = hook.beforeExecute(makeToolCall(), 'sess'); expect(hook.getPendingConfirmations()).toHaveLength(1); hook.resolveConfirmation(hook.getPendingConfirmations()[0].toolCallId, true, false, false); expect((await p2).blocked).toBe(false); }); it('resetRememberedDenial 对无拒绝记忆的工具返回 false', () => { const hook = new ConfirmationHook(null, null); expect(hook.resetRememberedDenial('run_command')).toBe(false); }); }); describe('ConfirmationHook — clearPending', () => { it('清空所有等待中的确认(全部视为拒绝)', async () => { const hook = new ConfirmationHook(makeMockWindow(), null); hook.setToolDefs([HIGH_RISK_DEF]); const p1 = hook.beforeExecute(makeToolCall(), 'sess'); const p2 = hook.beforeExecute(makeToolCall(), 'sess'); hook.clearPending(); expect((await p1).blocked).toBe(true); expect((await p2).blocked).toBe(true); expect(hook.getPendingConfirmations()).toHaveLength(0); }); });