fix(ai-isolate-cloudflare): port worker from unsafe_eval to worker_loader

.changeset/worker-loader-port.md

@@ -0,0 +1,24 @@
+---
+'@tanstack/ai-isolate-cloudflare': minor
+---
+
+Port the Cloudflare worker driver from `unsafe_eval` to `worker_loader` (Dynamic Workers).
+
+Cloudflare gates the `unsafe_eval` binding for all customer prod accounts; the previous driver was unusable in production and broken in `wrangler dev` on current Wrangler 4.x. The supported replacement is the `worker_loader` binding (GA-beta'd 2026-03-24).
+
+**Breaking:** the worker now requires the `LOADER` binding instead of `UNSAFE_EVAL`. Update your `wrangler.toml`:
+
+```toml
+# before
+[[unsafe.bindings]]
+name = "UNSAFE_EVAL"
+type = "unsafe_eval"
+
+# after
+[[worker_loaders]]
+binding = "LOADER"
+```
+
+The HTTP tool-callback protocol and public driver API are unchanged. Workers Paid plan is required for any edge usage (deploy or `wrangler dev --remote`); local `wrangler dev` works on the Free plan.
+
+Closes #522.

.gitignore

@@ -46,6 +46,7 @@ stats.html
 .tsup
 .vinxi
 temp
+.wrangler
 
 vite.config.js.timestamp-*
 vite.config.ts.timestamp-*

packages/typescript/ai-isolate-cloudflare/README.md

@@ -12,9 +12,9 @@ pnpm add @tanstack/ai-isolate-cloudflare
 
 ## Environment Guidance
 
-- **Local development:** supported with the package's Miniflare dev server (`pnpm dev:worker`)
-- **Remote dev:** supported with `wrangler dev --remote`
-- **Production:** supported on Cloudflare accounts with the `unsafe_eval` binding enabled. Before rollout, put the Worker behind authentication (e.g. Cloudflare Access or the `authorization` driver option), rate limiting, and CORS restrictions — running LLM-authored code is a high-trust operation.
+- **Local development:** supported with `wrangler dev` (the `worker_loader` binding works in local workerd on the Workers Free plan).
+- **Remote dev:** supported with `wrangler dev --remote` on a Workers Paid plan.
+- **Production:** supported on Cloudflare accounts on the Workers Paid plan ($5/mo). The Free plan rejects `worker_loader` deploys at the API level. Before rollout, put the Worker behind authentication (e.g. Cloudflare Access or the `authorization` driver option), rate limiting, and CORS restrictions — running LLM-authored code is a high-trust operation.
 
 If you want a self-contained host without Cloudflare infrastructure, prefer `@tanstack/ai-isolate-node` or `@tanstack/ai-isolate-quickjs`.
 
@@ -60,31 +60,31 @@ const result = await chat({
 
 ## Worker Setup
 
-### Option 1: Local Miniflare server
+### Option 1: Local dev with `wrangler dev`
 
 From this package directory:
 
 ```bash
-pnpm dev:worker
+wrangler dev
 ```
 
-This starts a local Worker endpoint (default `http://localhost:8787`) with the `UNSAFE_EVAL` binding configured in `wrangler.toml`.
+This starts a local Worker endpoint (default `http://localhost:8787`) with the `worker_loader` binding from `wrangler.toml`. Local workerd accepts the binding on the Workers Free plan, so no upgrade is needed for inner-loop iteration.
 
-### Option 3: Production deployment
+### Option 2: Wrangler remote dev
 
 ```bash
-wrangler deploy
+wrangler dev --remote
 ```
 
-The same `wrangler.toml` `[[unsafe.bindings]]` configuration applies in production. Deploying requires that your Cloudflare account has `unsafe_eval` enabled; without it, the Worker returns an `UnsafeEvalNotAvailable` error. Because this Worker executes LLM-generated code, only deploy it behind authentication, rate limiting, and an allow-listed origin.
+Runs through Cloudflare's network for validation against the hosted runtime. Requires a Workers Paid plan because `worker_loader` is gated to Paid for any edge usage.
 
-### Option 2: Wrangler remote dev
+### Option 3: Production deployment
 
 ```bash
-wrangler dev --remote
+wrangler deploy
 ```
 
-This runs through Cloudflare's network and can be useful when validating behavior against the hosted runtime.
+Requires a Workers Paid plan. The Free plan rejects deploys with `worker_loader` (`code: 10195` — "In order to use Dynamic Workers, you must switch to a paid plan."). Without the binding, the Worker returns a `WorkerLoaderNotAvailable` error. Because this Worker executes LLM-generated code, only deploy it behind authentication, rate limiting, and an allow-listed origin.
 
 ## API
 

packages/typescript/ai-isolate-cloudflare/package.json

@@ -35,7 +35,7 @@
   "scripts": {
     "build": "vite build",
     "clean": "premove ./build ./dist",
-    "dev:worker": "node dev-server.mjs",
+    "dev:worker": "wrangler dev",
     "deploy:worker": "wrangler deploy",
     "lint:fix": "eslint ./src --fix",
     "test:build": "publint --strict",
@@ -59,8 +59,6 @@
   "devDependencies": {
     "@cloudflare/workers-types": "^4.20241230.0",
     "@vitest/coverage-v8": "4.0.14",
-    "esbuild": "^0.25.12",
-    "miniflare": "^4.20260305.0",
-    "wrangler": "^4.19.1"
+    "wrangler": "^4.88.0"
   }
 }

packages/typescript/ai-isolate-cloudflare/src/worker/index.ts

@@ -1,83 +1,150 @@
 /**
  * Cloudflare Worker for Code Mode execution
  *
- * This Worker executes JavaScript code in a V8 isolate on Cloudflare's edge network.
- * Tool calls are handled via a request/response loop with the driver.
+ * Executes JavaScript code in a fresh V8 isolate on Cloudflare's edge network
+ * using the `worker_loader` (Dynamic Workers) binding. Tool calls round-trip
+ * to the driver via the same request/response protocol as before.
  *
  * Flow:
  * 1. Receive code + tool schemas
- * 2. Execute code, collecting any tool calls
- * 3. If tool calls are needed, return them to the driver
- * 4. Driver executes tools locally, sends results back
- * 5. Re-execute with tool results injected
- * 6. Return final result
+ * 2. Wrap user code in an ES module exporting a `fetch` handler that returns
+ *    the IIFE result as JSON
+ * 3. Load the module into a child Worker via `env.LOADER.load(...)` and
+ *    invoke its entrypoint
+ * 4. If tool calls are needed, return them to the driver
+ * 5. Driver executes tools locally, sends results back
+ * 6. Re-execute with tool results injected
+ * 7. Return final result
+ *
+ * `worker_loader` replaces the previous `unsafe_eval` binding, which is gated
+ * by Cloudflare for all customer accounts and unusable in production. See
+ * https://developers.cloudflare.com/dynamic-workers/ for the supported API.
  */
 
 import { wrapCode } from './wrap-code'
 import type { ExecuteRequest, ExecuteResponse, ToolCallRequest } from '../types'
 
 /**
- * UnsafeEval binding type.
+ * Compatibility date for the loaded child Worker. Pinned at this layer so
+ * sandbox semantics don't drift with the parent Worker's compat date.
+ */
+const SANDBOX_COMPAT_DATE = '2026-05-01'
+
+/**
+ * Worker Loader binding type.
  *
- * Provides dynamic-code execution against the Worker's V8 isolate. Available
- * locally (via wrangler dev) and in production deployments where the
- * `unsafe_eval` binding has been enabled on the Cloudflare account.
+ * Provides dynamic-code execution by loading a module into a fresh V8
+ * isolate. Configure in wrangler.toml under `[[worker_loaders]]`. Requires a
+ * Workers Paid plan; see https://developers.cloudflare.com/dynamic-workers/.
  */
-interface UnsafeEval {
-  eval: (code: string) => unknown
+interface WorkerLoaderEntrypoint {
+  fetch: (request: Request) => Promise<Response>
+}
+
+interface LoadedWorker {
+  getEntrypoint: (name?: string) => WorkerLoaderEntrypoint
+}
+
+interface WorkerLoader {
+  load: (options: {
+    compatibilityDate: string
+    mainModule: string
+    modules: Record<string, string>
+    globalOutbound?: unknown
+    env?: Record<string, unknown>
+  }) => LoadedWorker
 }
 
 interface Env {
   /**
-   * UnsafeEval binding. Configured in wrangler.toml as an unsafe binding.
+   * worker_loader (Dynamic Workers) binding. Configured in wrangler.toml
+   * under `[[worker_loaders]] binding = "LOADER"`.
    */
-  UNSAFE_EVAL?: UnsafeEval
+  LOADER?: WorkerLoader
 }
 
 /**
- * Execute code in the Worker's V8 isolate
+ * Wrap the existing IIFE-returning string in an ES module that exposes a
+ * `fetch` handler. The child Worker's entrypoint runs the IIFE on each
+ * invocation and returns the structured result as JSON.
+ */
+function wrapAsSandboxModule(wrappedCode: string): string {
+  return `
+export default {
+  async fetch() {
+    const __result = await ${wrappedCode};
+    return new Response(JSON.stringify(__result), {
+      headers: { 'Content-Type': 'application/json' },
+    });
+  }
+};
+`
+}
+
+/**
+ * Execute code in a freshly loaded child Worker isolate.
  */
 async function executeCode(
   request: ExecuteRequest,
   env: Env,
 ): Promise<ExecuteResponse> {
   const { code, tools, toolResults, timeout = 30000 } = request
 
-  // Check if UNSAFE_EVAL binding is available
-  if (!env.UNSAFE_EVAL) {
+  if (!env.LOADER) {
     return {
       status: 'error',
       error: {
-        name: 'UnsafeEvalNotAvailable',
+        name: 'WorkerLoaderNotAvailable',
         message:
-          'UNSAFE_EVAL binding is not available. ' +
-          'This Worker requires the unsafe_eval binding. ' +
-          'Declare it in wrangler.toml under [[unsafe.bindings]] ' +
-          '(works for local development and production where the ' +
-          'account has unsafe_eval enabled).',
+          'LOADER binding is not available. ' +
+          'This Worker requires the worker_loader (Dynamic Workers) binding. ' +
+          'Declare it in wrangler.toml under [[worker_loaders]] with ' +
+          'binding = "LOADER" (Workers Paid plan required).',
       },
     }
   }
 
   try {
     const wrappedCode = wrapCode(code, tools, toolResults)
+    const moduleSource = wrapAsSandboxModule(wrappedCode)
 
-    // Execute with timeout
+    // AbortController propagates into the loaded Worker via Request.signal so
+    // a timeout actually cancels the in-flight fetch instead of leaking the
+    // child isolate. The Promise.race remains as a belt-and-suspenders guard
+    // for runtimes that ignore the signal.
     const controller = new AbortController()
-    const timeoutId = setTimeout(() => controller.abort(), timeout)
+    let timeoutId: ReturnType<typeof setTimeout> | undefined
+    const TIMEOUT_SENTINEL = '__SANDBOX_TIMEOUT__'
+    const timeoutPromise = new Promise<never>((_resolve, reject) => {
+      timeoutId = setTimeout(() => {
+        controller.abort()
+        reject(new Error(TIMEOUT_SENTINEL))
+      }, timeout)
+    })
 
     try {
-      // Execute the wrapped code through the UNSAFE_EVAL binding.
-      const result = (await env.UNSAFE_EVAL.eval(wrappedCode)) as {
+      const loaded = env.LOADER.load({
+        compatibilityDate: SANDBOX_COMPAT_DATE,
+        mainModule: 'main.js',
+        modules: { 'main.js': moduleSource },
+        globalOutbound: null,
+        env: {},
+      })
+      const entrypoint = loaded.getEntrypoint()
+      const fetchPromise = entrypoint.fetch(
+        new Request('https://sandbox.invalid/', { signal: controller.signal }),
+      )
+      const response = await Promise.race([fetchPromise, timeoutPromise])
+      if (timeoutId) clearTimeout(timeoutId)
+
+      const result: {
         status: string
         success?: boolean
         value?: unknown
         error?: { name: string; message: string; stack?: string }
         logs: Array<string>
         toolCalls?: Array<ToolCallRequest>
-      }
-
-      clearTimeout(timeoutId)
+      } = await response.json()
 
       if (result.status === 'need_tools') {
         return {
@@ -96,9 +163,13 @@ async function executeCode(
         logs: result.logs,
       }
     } catch (evalError: unknown) {
-      clearTimeout(timeoutId)
+      if (timeoutId) clearTimeout(timeoutId)
+      const error = evalError as Error
 
-      if (controller.signal.aborted) {
+      // Either branch of the Promise.race may win on timeout: timeoutPromise
+      // rejects with TIMEOUT_SENTINEL, while the AbortController.abort() call
+      // can race-reject the in-flight fetch first. Treat both as a timeout.
+      if (error.message === TIMEOUT_SENTINEL || controller.signal.aborted) {
         return {
           status: 'error',
           error: {
@@ -108,7 +179,6 @@ async function executeCode(
         }
       }
 
-      const error = evalError as Error
       return {
         status: 'done',
         success: false,

packages/typescript/ai-isolate-cloudflare/src/worker/wrap-code.ts

@@ -1,6 +1,6 @@
 /**
  * Code wrapping utilities for the Cloudflare Worker.
- * Extracted for testability without UNSAFE_EVAL.
+ * Extracted for testability without a live worker_loader binding.
  */
 
 import type { ToolResultPayload, ToolSchema } from '../types'

packages/typescript/ai-isolate-cloudflare/tests/escape-attempts.test.ts

@@ -4,9 +4,9 @@ import type { ToolResultPayload, ToolSchema } from '../src/types'
 
 /**
  * The CF Worker delegates actual sandboxing to Workers' V8 isolate via the
- * UNSAFE_EVAL binding, so we can't perform a real escape attempt in Node. What
- * we verify here instead is structural — the wrapper must not let user inputs
- * break out of their intended quoting/scoping.
+ * worker_loader binding, so we can't perform a real escape attempt in Node.
+ * What we verify here instead is structural — the wrapper must not let user
+ * inputs break out of their intended quoting/scoping.
  */
 
 const benignTool: ToolSchema = {