|
| 1 | +/** |
| 2 | + * An inbox reply continues its parent task's chat, and the parent is found by the |
| 3 | + * `In-Reply-To` header — a value every recipient of the agent's reply holds. Against |
| 4 | + * real PostgreSQL, a reply delivered to one workspace's inbox must never reach a chat, |
| 5 | + * task, or transcript owned by another workspace, while a reply inside the workspace |
| 6 | + * still continues its thread. The webhook is signed with the real Svix scheme. |
| 7 | + */ |
| 8 | +import { authBanMock } from '@sim/testing/mocks/auth-ban.mock' |
| 9 | +import { |
| 10 | + billingAttributionMock, |
| 11 | + billingAttributionMockFns, |
| 12 | +} from '@sim/testing/mocks/billing-attribution.mock' |
| 13 | +import { |
| 14 | + billingSubscriptionMock, |
| 15 | + billingSubscriptionMockFns, |
| 16 | +} from '@sim/testing/mocks/billing-subscription.mock' |
| 17 | +import { |
| 18 | + mothershipHeadlessLifecycleMock, |
| 19 | + mothershipHeadlessLifecycleMockFns, |
| 20 | +} from '@sim/testing/mocks/mothership-headless-lifecycle.mock' |
| 21 | +import { afterAll, beforeAll, beforeEach, describe, expect, it, vi } from 'vitest' |
| 22 | + |
| 23 | +const { mockExecuteInboxTask, mockSendInboxResponse } = vi.hoisted(() => ({ |
| 24 | + mockExecuteInboxTask: vi.fn(), |
| 25 | + mockSendInboxResponse: vi.fn(), |
| 26 | +})) |
| 27 | + |
| 28 | +vi.mock('@/lib/auth/ban', () => authBanMock) |
| 29 | +vi.mock('@/lib/billing/core/billing-attribution', () => billingAttributionMock) |
| 30 | +vi.mock('@/lib/billing/core/subscription', () => billingSubscriptionMock) |
| 31 | +vi.mock('@/lib/mothership/request/lifecycle/headless', () => mothershipHeadlessLifecycleMock) |
| 32 | +vi.mock('@/lib/mothership/request/lifecycle/start', () => ({ |
| 33 | + requestChatTitle: async () => null, |
| 34 | +})) |
| 35 | +vi.mock('@/lib/mothership/inbox/response', () => ({ |
| 36 | + sendInboxResponse: mockSendInboxResponse, |
| 37 | +})) |
| 38 | +vi.mock('@/lib/mothership/inbox/executor', () => ({ |
| 39 | + executeInboxTask: mockExecuteInboxTask, |
| 40 | +})) |
| 41 | + |
| 42 | +import { db } from '@sim/db' |
| 43 | +import { |
| 44 | + copilotChats, |
| 45 | + copilotMessages, |
| 46 | + mothershipInboxTask, |
| 47 | + mothershipInboxWebhook, |
| 48 | + permissions, |
| 49 | + user, |
| 50 | + workspace, |
| 51 | +} from '@sim/db/schema' |
| 52 | +import { generateId } from '@sim/utils/id' |
| 53 | +import { and, eq, inArray } from 'drizzle-orm' |
| 54 | +import { NextRequest } from 'next/server' |
| 55 | +import { Webhook } from 'svix' |
| 56 | +import { POST } from '@/app/api/webhooks/agentmail/route' |
| 57 | + |
| 58 | +const { executeInboxTask } = await vi.importActual< |
| 59 | + typeof import('@/lib/mothership/inbox/executor') |
| 60 | +>('@/lib/mothership/inbox/executor') |
| 61 | +const { mockRunHeadlessCopilotLifecycle } = mothershipHeadlessLifecycleMockFns |
| 62 | + |
| 63 | +interface InboxWorkspace { |
| 64 | + id: string |
| 65 | + ownerId: string |
| 66 | + ownerEmail: string |
| 67 | + inboxId: string |
| 68 | + secret: string |
| 69 | +} |
| 70 | + |
| 71 | +const victim = inboxWorkspace() |
| 72 | +const attacker = inboxWorkspace() |
| 73 | +const victimChatId = generateId() |
| 74 | +const victimReplyMessageId = `<${generateId()}@agentmail.to>` |
| 75 | +const victimChatUpdatedAt = new Date('2026-01-01T00:00:00.000Z') |
| 76 | + |
| 77 | +function inboxWorkspace(): InboxWorkspace { |
| 78 | + const id = generateId() |
| 79 | + return { |
| 80 | + id, |
| 81 | + ownerId: generateId(), |
| 82 | + ownerEmail: `${id}@inbox-thread-scope.test`, |
| 83 | + inboxId: `${id}@agentmail.to`, |
| 84 | + secret: `whsec_${Buffer.from(generateId()).toString('base64')}`, |
| 85 | + } |
| 86 | +} |
| 87 | + |
| 88 | +async function seedWorkspace(ws: InboxWorkspace): Promise<void> { |
| 89 | + const now = new Date() |
| 90 | + await db.insert(user).values({ |
| 91 | + id: ws.ownerId, |
| 92 | + name: 'Inbox thread scope fixture', |
| 93 | + email: ws.ownerEmail, |
| 94 | + emailVerified: true, |
| 95 | + createdAt: now, |
| 96 | + updatedAt: now, |
| 97 | + }) |
| 98 | + await db.insert(workspace).values({ |
| 99 | + id: ws.id, |
| 100 | + name: 'Inbox thread scope fixture', |
| 101 | + ownerId: ws.ownerId, |
| 102 | + billedAccountUserId: ws.ownerId, |
| 103 | + inboxEnabled: true, |
| 104 | + inboxAddress: ws.inboxId, |
| 105 | + inboxProviderId: ws.inboxId, |
| 106 | + }) |
| 107 | + await db.insert(permissions).values({ |
| 108 | + id: generateId(), |
| 109 | + userId: ws.ownerId, |
| 110 | + entityType: 'workspace', |
| 111 | + entityId: ws.id, |
| 112 | + permissionType: 'admin', |
| 113 | + }) |
| 114 | + await db.insert(mothershipInboxWebhook).values({ |
| 115 | + id: generateId(), |
| 116 | + workspaceId: ws.id, |
| 117 | + webhookId: generateId(), |
| 118 | + secret: ws.secret, |
| 119 | + }) |
| 120 | +} |
| 121 | + |
| 122 | +/** Delivers a `message.received` webhook to `ws`'s inbox from its owner, signed with its secret. */ |
| 123 | +async function deliver( |
| 124 | + ws: InboxWorkspace, |
| 125 | + message: { messageId: string; inReplyTo?: string } |
| 126 | +): Promise<Response> { |
| 127 | + const body = JSON.stringify({ |
| 128 | + event_type: 'message.received', |
| 129 | + message: { |
| 130 | + message_id: message.messageId, |
| 131 | + thread_id: generateId(), |
| 132 | + inbox_id: ws.inboxId, |
| 133 | + from: `Owner <${ws.ownerEmail}>`, |
| 134 | + to: [ws.inboxId], |
| 135 | + subject: 'Re: quarterly numbers', |
| 136 | + text: 'Please continue this thread.', |
| 137 | + created_at: new Date().toISOString(), |
| 138 | + ...(message.inReplyTo ? { in_reply_to: message.inReplyTo } : {}), |
| 139 | + }, |
| 140 | + }) |
| 141 | + const svixId = `msg_${generateId()}` |
| 142 | + const timestamp = new Date() |
| 143 | + return POST( |
| 144 | + new NextRequest('http://localhost:3000/api/webhooks/agentmail', { |
| 145 | + method: 'POST', |
| 146 | + headers: { |
| 147 | + 'content-type': 'application/json', |
| 148 | + 'svix-id': svixId, |
| 149 | + 'svix-timestamp': String(Math.floor(timestamp.getTime() / 1000)), |
| 150 | + 'svix-signature': new Webhook(ws.secret).sign(svixId, timestamp, body), |
| 151 | + }, |
| 152 | + body, |
| 153 | + }) |
| 154 | + ) |
| 155 | +} |
| 156 | + |
| 157 | +async function taskFor(workspaceId: string, emailMessageId: string) { |
| 158 | + const [task] = await db |
| 159 | + .select() |
| 160 | + .from(mothershipInboxTask) |
| 161 | + .where( |
| 162 | + and( |
| 163 | + eq(mothershipInboxTask.workspaceId, workspaceId), |
| 164 | + eq(mothershipInboxTask.emailMessageId, emailMessageId) |
| 165 | + ) |
| 166 | + ) |
| 167 | + return task |
| 168 | +} |
| 169 | + |
| 170 | +async function victimChatState() { |
| 171 | + const [chat] = await db |
| 172 | + .select({ updatedAt: copilotChats.updatedAt }) |
| 173 | + .from(copilotChats) |
| 174 | + .where(eq(copilotChats.id, victimChatId)) |
| 175 | + const messages = await db |
| 176 | + .select({ id: copilotMessages.id }) |
| 177 | + .from(copilotMessages) |
| 178 | + .where(eq(copilotMessages.chatId, victimChatId)) |
| 179 | + return { updatedAt: chat?.updatedAt, messageCount: messages.length } |
| 180 | +} |
| 181 | + |
| 182 | +describe('inbox reply threading stays inside the receiving workspace', () => { |
| 183 | + beforeAll(async () => { |
| 184 | + await seedWorkspace(victim) |
| 185 | + await seedWorkspace(attacker) |
| 186 | + await db.insert(copilotChats).values({ |
| 187 | + id: victimChatId, |
| 188 | + userId: victim.ownerId, |
| 189 | + workspaceId: victim.id, |
| 190 | + type: 'mothership', |
| 191 | + updatedAt: victimChatUpdatedAt, |
| 192 | + }) |
| 193 | + await db.insert(mothershipInboxTask).values({ |
| 194 | + id: generateId(), |
| 195 | + workspaceId: victim.id, |
| 196 | + fromEmail: victim.ownerEmail, |
| 197 | + subject: 'quarterly numbers', |
| 198 | + status: 'completed', |
| 199 | + chatId: victimChatId, |
| 200 | + responseMessageId: victimReplyMessageId, |
| 201 | + }) |
| 202 | + }) |
| 203 | + |
| 204 | + beforeEach(() => { |
| 205 | + mockExecuteInboxTask.mockResolvedValue(undefined) |
| 206 | + mockSendInboxResponse.mockResolvedValue(null) |
| 207 | + billingSubscriptionMockFns.mockHasWorkspaceInboxAccess.mockResolvedValue(true) |
| 208 | + billingAttributionMockFns.mockResolveBillingAttribution.mockResolvedValue({}) |
| 209 | + mockRunHeadlessCopilotLifecycle.mockResolvedValue({ |
| 210 | + success: true, |
| 211 | + content: 'Here are the numbers.', |
| 212 | + contentBlocks: [], |
| 213 | + toolCalls: [], |
| 214 | + }) |
| 215 | + }) |
| 216 | + |
| 217 | + afterAll(async () => { |
| 218 | + const ownerIds = [victim.ownerId, attacker.ownerId] |
| 219 | + await db.delete(permissions).where(inArray(permissions.userId, ownerIds)) |
| 220 | + await db.delete(workspace).where(inArray(workspace.id, [victim.id, attacker.id])) |
| 221 | + await db.delete(user).where(inArray(user.id, ownerIds)) |
| 222 | + }) |
| 223 | + |
| 224 | + it("does not adopt another workspace's chat for a reply to that workspace's message", async () => { |
| 225 | + const messageId = `<${generateId()}@attacker.test>` |
| 226 | + |
| 227 | + const response = await deliver(attacker, { messageId, inReplyTo: victimReplyMessageId }) |
| 228 | + |
| 229 | + expect(response.status).toBe(200) |
| 230 | + const task = await taskFor(attacker.id, messageId) |
| 231 | + expect(task?.inReplyTo).toBe(victimReplyMessageId) |
| 232 | + expect(task?.chatId).toBeNull() |
| 233 | + }) |
| 234 | + |
| 235 | + it('continues the parent chat for a reply inside the same workspace', async () => { |
| 236 | + const messageId = `<${generateId()}@victim.test>` |
| 237 | + |
| 238 | + await deliver(victim, { messageId, inReplyTo: victimReplyMessageId }) |
| 239 | + |
| 240 | + expect((await taskFor(victim.id, messageId))?.chatId).toBe(victimChatId) |
| 241 | + }) |
| 242 | + |
| 243 | + it("accepts a message whose id another workspace's inbox already received", async () => { |
| 244 | + const sharedMessageId = `<${generateId()}@cc-both.test>` |
| 245 | + |
| 246 | + await deliver(attacker, { messageId: sharedMessageId }) |
| 247 | + await deliver(victim, { messageId: sharedMessageId }) |
| 248 | + |
| 249 | + expect(await taskFor(attacker.id, sharedMessageId)).toMatchObject({ status: 'received' }) |
| 250 | + expect(await taskFor(victim.id, sharedMessageId)).toMatchObject({ status: 'received' }) |
| 251 | + }) |
| 252 | + |
| 253 | + it("runs a task carrying another workspace's chat in a fresh chat of its own workspace", async () => { |
| 254 | + const taskId = generateId() |
| 255 | + await db.insert(mothershipInboxTask).values({ |
| 256 | + id: taskId, |
| 257 | + workspaceId: attacker.id, |
| 258 | + fromEmail: attacker.ownerEmail, |
| 259 | + subject: 'Re: quarterly numbers', |
| 260 | + bodyText: 'Append me to the victim transcript.', |
| 261 | + status: 'received', |
| 262 | + chatId: victimChatId, |
| 263 | + inReplyTo: victimReplyMessageId, |
| 264 | + }) |
| 265 | + |
| 266 | + await executeInboxTask(taskId) |
| 267 | + |
| 268 | + expect(mockRunHeadlessCopilotLifecycle).toHaveBeenCalledOnce() |
| 269 | + const [, options] = mockRunHeadlessCopilotLifecycle.mock.calls[0] |
| 270 | + expect(options).toMatchObject({ workspaceId: attacker.id }) |
| 271 | + expect(options.chatId).not.toBe(victimChatId) |
| 272 | + const [runChat] = await db |
| 273 | + .select({ workspaceId: copilotChats.workspaceId }) |
| 274 | + .from(copilotChats) |
| 275 | + .where(eq(copilotChats.id, options.chatId)) |
| 276 | + expect(runChat?.workspaceId).toBe(attacker.id) |
| 277 | + |
| 278 | + const [task] = await db |
| 279 | + .select({ status: mothershipInboxTask.status, chatId: mothershipInboxTask.chatId }) |
| 280 | + .from(mothershipInboxTask) |
| 281 | + .where(eq(mothershipInboxTask.id, taskId)) |
| 282 | + expect(task).toEqual({ status: 'completed', chatId: options.chatId }) |
| 283 | + expect(await victimChatState()).toEqual({ |
| 284 | + updatedAt: victimChatUpdatedAt, |
| 285 | + messageCount: 0, |
| 286 | + }) |
| 287 | + }) |
| 288 | +}) |
0 commit comments