refactor(sandbox): move sandbox state management server-side (ENG-4287) (#381)

Author: huv1k

src/app/dashboard/[teamSlug]/sandboxes/[sandboxId]/filesystem/page.tsx

@@ -4,7 +4,6 @@ import { COOKIE_KEYS } from '@/configs/cookies'
 import { AUTH_URLS, PROTECTED_URLS } from '@/configs/urls'
 import { getAuthContext } from '@/core/server/auth'
 import { getTeamIdFromSlug } from '@/core/server/functions/team/get-team-id-from-slug'
-import { createSandboxManagementAuth } from '@/core/shared/sandbox-management-auth.server'
 import SandboxInspectView from '@/features/dashboard/sandbox/inspect/view'
 
 const DEFAULT_ROOT_PATH = '/home/user'
@@ -41,13 +40,5 @@ export default async function SandboxInspectPage({
     cookieStore.get(COOKIE_KEYS.SANDBOX_INSPECT_ROOT_PATH)?.value ||
     DEFAULT_ROOT_PATH
 
-  return (
-    <SandboxInspectView
-      rootPath={rootPath}
-      sandboxManagementAuth={createSandboxManagementAuth(
-        authContext,
-        teamId.data
-      )}
-    />
-  )
+  return <SandboxInspectView rootPath={rootPath} />
 }

src/app/dashboard/[teamSlug]/sandboxes/[sandboxId]/terminal/page.tsx

@@ -2,7 +2,6 @@ import { redirect } from 'next/navigation'
 import { AUTH_URLS, PROTECTED_URLS } from '@/configs/urls'
 import { getAuthContext } from '@/core/server/auth'
 import { getTeamIdFromSlug } from '@/core/server/functions/team/get-team-id-from-slug'
-import { createSandboxManagementAuth } from '@/core/shared/sandbox-management-auth.server'
 import SandboxTerminalView from '@/features/dashboard/sandbox/terminal/view'
 
 interface SandboxTerminalPageProps {
@@ -36,13 +35,5 @@ export default async function SandboxTerminalPage({
     redirect(PROTECTED_URLS.DASHBOARD)
   }
 
-  return (
-    <SandboxTerminalView
-      command={command}
-      sandboxManagementAuth={createSandboxManagementAuth(
-        authContext,
-        teamId.data
-      )}
-    />
-  )
+  return <SandboxTerminalView command={command} userId={authContext.user.id} />
 }

src/app/dashboard/[teamSlug]/terminal/page.tsx

@@ -10,7 +10,6 @@ import {
 import { getAuthContext } from '@/core/server/auth'
 import { infra } from '@/core/shared/clients/api'
 import type { components as InfraComponents } from '@/core/shared/contracts/infra-api.types'
-import { createSandboxManagementAuth } from '@/core/shared/sandbox-management-auth.server'
 import { SandboxIdSchema } from '@/core/shared/schemas/api'
 import DashboardTerminal from '@/features/dashboard/terminal/dashboard-terminal'
 import { normalizeTerminalTemplate } from '@/features/dashboard/terminal/template'
@@ -132,11 +131,8 @@ export default async function TeamTerminalPage({
           sandboxId: terminalSandboxId,
           template: terminalTemplate,
         }}
-        sandboxManagementAuth={createSandboxManagementAuth(
-          authContext,
-          team.id
-        )}
         teamSlug={team.slug}
+        userId={authContext.user.id}
       />
     </div>
   )

src/core/server/api/routers/sandbox.ts

@@ -1,5 +1,6 @@
+import { TRPCError } from '@trpc/server'
 import { millisecondsInDay } from 'date-fns/constants'
-import { Sandbox } from 'e2b'
+import { Sandbox, TimeoutError } from 'e2b'
 import { z } from 'zod'
 import { authHeaders } from '@/configs/api'
 import {
@@ -17,8 +18,10 @@ import { withTeamAuthedRequestRepository } from '@/core/server/api/middlewares/r
 import { createTRPCRouter } from '@/core/server/trpc/init'
 import { protectedTeamProcedure } from '@/core/server/trpc/procedures'
 import { SandboxIdSchema } from '@/core/shared/schemas/api'
+import { SANDBOX_RESUME_TIMEOUT_MS } from '@/features/dashboard/sandbox/inspect/constants'
 import { SANDBOX_MONITORING_METRICS_RETENTION_MS } from '@/features/dashboard/sandbox/monitoring/utils/constants'
 import { TERMINAL_SANDBOX_TIMEOUT_MS } from '@/features/dashboard/terminal/constants'
+import { normalizeTerminalTemplate } from '@/features/dashboard/terminal/template'
 
 const sandboxRepositoryProcedure = protectedTeamProcedure.use(
   withTeamAuthedRequestRepository(
@@ -209,6 +212,205 @@ export const sandboxRouter = createTRPCRouter({
 
   // MUTATIONS
 
+  // Runs the control-plane create/connect server-side so the user's
+  // account-level access token never reaches the browser. Returns only the
+  // sandbox-scoped envd credentials the client needs for PTY access.
+  openTerminal: protectedTeamProcedure
+    .input(
+      z.object({
+        template: z.string().min(1, 'Template is required'),
+        sandboxId: SandboxIdSchema.optional(),
+        requestTimeoutMs: z.number().int().positive().optional(),
+      })
+    )
+    .mutation(async ({ ctx, input }) => {
+      const { sandboxId, template, requestTimeoutMs } = input
+      const { session, teamId } = ctx
+
+      const normalizedTemplate = normalizeTerminalTemplate(template)
+      if (!normalizedTemplate) {
+        throw new TRPCError({
+          code: 'BAD_REQUEST',
+          message: 'Invalid terminal template',
+        })
+      }
+
+      const connectionOpts = {
+        apiUrl: process.env.NEXT_PUBLIC_INFRA_API_URL,
+        domain: process.env.NEXT_PUBLIC_E2B_DOMAIN,
+        sandboxUrl: process.env.NEXT_PUBLIC_E2B_SANDBOX_URL,
+        headers: authHeaders(session.access_token, teamId),
+      }
+
+      try {
+        let resolvedSandboxId: string
+        if (sandboxId) {
+          const sandbox = await Sandbox.connect(sandboxId, {
+            ...connectionOpts,
+            timeoutMs: TERMINAL_SANDBOX_TIMEOUT_MS,
+            requestTimeoutMs,
+          })
+          resolvedSandboxId = sandbox.sandboxId
+        } else {
+          const sandbox = await Sandbox.create(normalizedTemplate, {
+            ...connectionOpts,
+            timeoutMs: TERMINAL_SANDBOX_TIMEOUT_MS,
+            requestTimeoutMs,
+            lifecycle: {
+              onTimeout: 'pause',
+              autoResume: true,
+            },
+            metadata: {
+              source: 'dashboard-terminal',
+              template: normalizedTemplate,
+              userId: session.user.id,
+            },
+          })
+          resolvedSandboxId = sandbox.sandboxId
+        }
+
+        // `Sandbox.create`/`connect` build a full SDK instance but only expose
+        // the sandbox id/domain publicly; fetch the envd credentials via the
+        // public info endpoint rather than reading the SDK's internal fields.
+        // Kept inside this try (with the same requestTimeoutMs) so a stalled
+        // GET times out promptly and is normalized like the connect timeout.
+        const info = await Sandbox.getFullInfo(resolvedSandboxId, {
+          ...connectionOpts,
+          requestTimeoutMs,
+        })
+
+        // `envdAccessToken` is absent for `secure: false` sandboxes, whose envd
+        // is reachable without the `X-Access-Token` header — pass it through
+        // as-is rather than treating its absence as a failure.
+        return {
+          sandboxId: resolvedSandboxId,
+          sandboxDomain: info.sandboxDomain,
+          envdVersion: info.envdVersion,
+          envdAccessToken: info.envdAccessToken,
+        }
+      } catch (error) {
+        if (error instanceof TimeoutError) {
+          throw new TRPCError({
+            code: 'TIMEOUT',
+            message: sandboxId
+              ? 'Timed out connecting to terminal sandbox'
+              : 'Timed out creating terminal sandbox',
+            cause: error,
+          })
+        }
+
+        throw new TRPCError({
+          code: 'INTERNAL_SERVER_ERROR',
+          message: sandboxId
+            ? 'Failed to connect to terminal sandbox'
+            : 'Failed to create terminal sandbox',
+          cause: error,
+        })
+      }
+    }),
+
+  // Explicit, user-triggered resume of a paused sandbox for the inspect view.
+  // The control-plane connect (which resumes + sets TTL) runs server-side so
+  // the account token never reaches the browser; returns the sandbox-scoped
+  // envd credentials the client uses to rebuild its envd-only client.
+  resume: protectedTeamProcedure
+    .input(
+      z.object({
+        sandboxId: SandboxIdSchema,
+        requestTimeoutMs: z.number().int().positive().optional(),
+      })
+    )
+    .mutation(async ({ ctx, input }) => {
+      const { sandboxId, requestTimeoutMs } = input
+      const { session, teamId } = ctx
+
+      const connectionOpts = {
+        apiUrl: process.env.NEXT_PUBLIC_INFRA_API_URL,
+        domain: process.env.NEXT_PUBLIC_E2B_DOMAIN,
+        sandboxUrl: process.env.NEXT_PUBLIC_E2B_SANDBOX_URL,
+        headers: authHeaders(session.access_token, teamId),
+      }
+
+      try {
+        await Sandbox.connect(sandboxId, {
+          ...connectionOpts,
+          timeoutMs: SANDBOX_RESUME_TIMEOUT_MS,
+          requestTimeoutMs: requestTimeoutMs ?? SANDBOX_RESUME_TIMEOUT_MS,
+        })
+
+        const info = await Sandbox.getFullInfo(sandboxId, {
+          ...connectionOpts,
+          requestTimeoutMs: requestTimeoutMs ?? SANDBOX_RESUME_TIMEOUT_MS,
+        })
+
+        return {
+          sandboxId,
+          sandboxDomain: info.sandboxDomain,
+          envdVersion: info.envdVersion,
+          envdAccessToken: info.envdAccessToken,
+        }
+      } catch (error) {
+        if (error instanceof TimeoutError) {
+          throw new TRPCError({
+            code: 'TIMEOUT',
+            message: 'Timed out resuming sandbox',
+            cause: error,
+          })
+        }
+
+        throw new TRPCError({
+          code: 'INTERNAL_SERVER_ERROR',
+          message: 'Failed to resume sandbox',
+          cause: error,
+        })
+      }
+    }),
+
+  // Explicit, user-triggered pause of a running sandbox. Uses the SDK's
+  // control-plane pause (which snapshots and pauses) server-side so the
+  // account token never reaches the browser.
+  pause: protectedTeamProcedure
+    .input(
+      z.object({
+        sandboxId: SandboxIdSchema,
+        requestTimeoutMs: z.number().int().positive().optional(),
+      })
+    )
+    .mutation(async ({ ctx, input }) => {
+      const { sandboxId, requestTimeoutMs } = input
+      const { session, teamId } = ctx
+
+      const connectionOpts = {
+        apiUrl: process.env.NEXT_PUBLIC_INFRA_API_URL,
+        domain: process.env.NEXT_PUBLIC_E2B_DOMAIN,
+        sandboxUrl: process.env.NEXT_PUBLIC_E2B_SANDBOX_URL,
+        headers: authHeaders(session.access_token, teamId),
+      }
+
+      try {
+        // Returns false when the sandbox was already paused, which we treat
+        // as success since the desired end state is reached.
+        await Sandbox.pause(sandboxId, {
+          ...connectionOpts,
+          ...(requestTimeoutMs ? { requestTimeoutMs } : {}),
+        })
+      } catch (error) {
+        if (error instanceof TimeoutError) {
+          throw new TRPCError({
+            code: 'TIMEOUT',
+            message: 'Timed out pausing sandbox',
+            cause: error,
+          })
+        }
+
+        throw new TRPCError({
+          code: 'INTERNAL_SERVER_ERROR',
+          message: 'Failed to pause sandbox',
+          cause: error,
+        })
+      }
+    }),
+
   killTerminalPty: protectedTeamProcedure
     .input(
       z.object({

src/core/shared/create-envd-sandbox.ts

@@ -0,0 +1,54 @@
+import Sandbox, { type SandboxConnectOpts } from 'e2b'
+
+/**
+ * Sandbox-scoped credentials needed to talk to a sandbox's envd daemon
+ * (filesystem + PTY). These are safe to expose to the browser: they grant
+ * access to a single sandbox only, unlike the account-level access token.
+ */
+export interface EnvdSandboxParams {
+  sandboxId: string
+  sandboxDomain?: string | null
+  envdVersion: string
+  // Optional: `secure: false` sandboxes expose envd without an access token.
+  envdAccessToken?: string
+  domain?: string
+  sandboxUrl?: string
+  trafficAccessToken?: string
+}
+
+/**
+ * The {@link Sandbox} constructor is marked `@internal` in the SDK type defs
+ * (the public entry points are `Sandbox.create`/`Sandbox.connect`, both of
+ * which make a control-plane call that requires the account-level access
+ * token). We deliberately bypass those and build an envd-only instance
+ * directly from sandbox-scoped credentials: the constructor performs NO
+ * network call and authenticates every `files.*`/`pty.*` request with the
+ * `envdAccessToken` alone.
+ *
+ * The cast is isolated here so the rest of the codebase never touches the
+ * internal API. The option object is typed against the exported
+ * `SandboxConnectOpts`, so an SDK field rename surfaces as a compile error.
+ */
+type EnvdSandboxConstructor = new (
+  opts: SandboxConnectOpts & {
+    sandboxId: string
+    sandboxDomain?: string
+    envdVersion: string
+    envdAccessToken?: string
+    trafficAccessToken?: string
+  }
+) => Sandbox
+
+export function createEnvdSandbox(params: EnvdSandboxParams): Sandbox {
+  const SandboxCtor = Sandbox as unknown as EnvdSandboxConstructor
+
+  return new SandboxCtor({
+    sandboxId: params.sandboxId,
+    sandboxDomain: params.sandboxDomain ?? undefined,
+    envdVersion: params.envdVersion,
+    envdAccessToken: params.envdAccessToken,
+    trafficAccessToken: params.trafficAccessToken,
+    domain: params.domain,
+    sandboxUrl: params.sandboxUrl,
+  })
+}

src/core/shared/sandbox-management-auth.server.ts

@@ -1,8 +1,10 @@
 import KillButton from './kill-button'
+import PauseButton from './pause-button'
 
 export default function SandboxDetailsControls() {
   return (
     <div className="flex items-center gap-2 md:pb-2">
+      <PauseButton />
       <KillButton />
     </div>
   )