@tanstack/ai-sandbox 0.5.6 → 0.5.9

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -74,6 +74,13 @@ Pick a **provider** package for where the sandbox runs:
74
74
  | `@tanstack/ai-sandbox-daytona` | Daytona cloud sandboxes, snapshots |
75
75
  | `@tanstack/ai-sandbox-upstash-box` | Upstash Box cloud sandboxes, snapshots |
76
76
  | `@tanstack/ai-sandbox-sprites` | Sprites stateful sandboxes |
77
+ | `@tanstack/ai-sandbox-blaxel` | Blaxel cloud sandboxes and previews |
78
+
79
+ Install the provider you select separately. For Blaxel:
80
+
81
+ ```bash
82
+ npm install @tanstack/ai-sandbox-blaxel
83
+ ```
77
84
 
78
85
  **Harness adapters** are separate packages. The default path is **Grok Build** (`@tanstack/ai-grok-build`); others include `@tanstack/ai-claude-code`, `@tanstack/ai-codex`, and `@tanstack/ai-opencode`. All require `withSandbox(...)` middleware — `chat()` fails fast without it.
79
86
 
@@ -1,5 +1,6 @@
1
1
  import { randomBytes, timingSafeEqual } from "node:crypto";
2
2
  import { createServer } from "node:http";
3
+ import { once } from "node:events";
3
4
  import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
4
5
  import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/streamableHttp.js";
5
6
  import { CallToolRequestSchema, ListToolsRequestSchema } from "@modelcontextprotocol/sdk/types.js";
@@ -221,7 +222,7 @@ async function startHostToolBridge(tools, options) {
221
222
  if (!res.headersSent) res.writeHead(500).end("bridge error");
222
223
  });
223
224
  });
224
- await new Promise((resolve) => httpServer.listen(0, bindAddress, resolve));
225
+ await once(httpServer.listen(0, bindAddress), "listening");
225
226
  const port = httpServer.address().port;
226
227
  return {
227
228
  name: BRIDGED_MCP_SERVER_NAME,
@@ -1 +1 @@
1
- {"version":3,"file":"tool-bridge.js","names":[],"sources":["../../src/tool-bridge.ts"],"sourcesContent":["/**\n * MCP tool-proxy bridge, shared by all harness adapters.\n *\n * Exposes chat()-provided server tools to an in-sandbox agent as an MCP server.\n * The agent (inside the sandbox) calls `mcp__tanstack__<tool>`; the call is\n * proxied OUT to a bridge endpoint, where the tool's `execute()` runs in the\n * orchestrator process (with its closures / DB / secrets), and the result is\n * returned into the sandbox.\n *\n * The bridge is split into a transport-agnostic CORE and a TRANSPORT:\n * - {@link createToolBridgeCore} owns tool dispatch + the permission resolver\n * (no I/O). It is what makes the bridge portable.\n * - {@link startHostToolBridge} is the `node:http` transport for a long-running\n * host (laptop / CI / Docker orchestrator). It binds loopback unless the\n * sandbox must reach it via `host.docker.internal`, and authenticates with a\n * constant-time bearer check.\n * - A serverless/edge orchestrator (e.g. a Durable Object) instead serves the\n * SAME core from its own `fetch` handler — no raw TCP listener — see\n * {@link handleBridgeJsonRpc} and the Cloudflare example.\n */\nimport { createServer } from 'node:http'\nimport { randomBytes, timingSafeEqual } from 'node:crypto'\nimport { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'\nimport { StreamableHTTPServerTransport } from '@modelcontextprotocol/sdk/server/streamableHttp.js'\nimport {\n CallToolRequestSchema,\n ListToolsRequestSchema,\n} from '@modelcontextprotocol/sdk/types.js'\nimport type { AddressInfo } from 'node:net'\nimport type { AnyTool } from '@tanstack/ai'\n\n/**\n * Name of the bridged MCP server. The agent sees tools as\n * `mcp__tanstack__<tool>`; each adapter's stream translator strips this prefix\n * so tool-call events match the names the application registered.\n */\nexport const BRIDGED_MCP_SERVER_NAME = 'tanstack'\n\n/** Hostname the sandbox uses to reach the bridge endpoint, per provider. */\nexport function hostForSandbox(provider: string): string {\n return provider === 'docker' || provider === 'sbx'\n ? 'host.docker.internal'\n : '127.0.0.1'\n}\n\n/** Result of a permission decision returned to the harness's prompt tool. */\nexport interface PermissionToolResult {\n behavior: 'allow' | 'deny'\n message?: string\n updatedInput?: unknown\n}\n\nexport interface BridgePermission {\n toolName: string\n resolve: (input: {\n tool_name?: string\n input?: unknown\n }) => PermissionToolResult | Promise<PermissionToolResult>\n}\n\nexport interface ToolBridgeCoreOptions {\n /** Runtime context forwarded to each tool's `execute()`. */\n context?: unknown\n /** Abort signal forwarded to each tool's `execute()`. */\n signal?: AbortSignal\n /**\n * Forwarded to each tool's `execute()` so a bridged tool can stream progress /\n * custom events back to the client mid-execution (e.g. code mode's\n * `code_mode:console` logs). Without it those events are silently dropped — the\n * bridge runs out-of-band from the main tool executor, so the executor's own\n * `emitCustomEvent` never reaches a bridged tool. The harness adapter supplies\n * one that injects a CUSTOM chunk into its live output stream.\n */\n emitCustomEvent?: (eventName: string, value: Record<string, unknown>) => void\n /**\n * Optional permission-prompt tool (e.g. for Claude Code's\n * `--permission-prompt-tool`). When set, the bridge exposes an extra MCP tool\n * `<name>` whose handler returns the orchestrator's allow/deny decision.\n */\n permission?: BridgePermission\n}\n\n/** An MCP tool descriptor as advertised to the in-sandbox agent. */\nexport interface ToolDescriptor {\n name: string\n description?: string\n inputSchema: { type: 'object'; [key: string]: unknown }\n}\n\n/**\n * Coerce a tool's `inputSchema` into the object-schema shape MCP advertises,\n * substituting an empty object schema when it isn't already a JSON-schema object\n * (project rule: a guard, not an `as` cast).\n */\nfunction toObjectSchema(schema: unknown): {\n type: 'object'\n [key: string]: unknown\n} {\n if (\n schema !== null &&\n typeof schema === 'object' &&\n 'type' in schema &&\n schema.type === 'object'\n ) {\n return { ...schema, type: 'object' }\n }\n return { type: 'object', properties: {} }\n}\n\n/** MCP `tools/call` result shape. */\nexport interface ToolCallResult {\n content: Array<{ type: 'text'; text: string }>\n isError?: boolean\n}\n\n/**\n * Transport-agnostic bridge logic: list tools, and dispatch a tool/permission\n * call. No sockets, no auth — a transport ({@link startHostToolBridge} or a\n * `fetch` handler) wraps this and owns I/O + the bearer check.\n */\nexport interface ToolBridgeCore {\n listTools: () => Array<ToolDescriptor>\n callTool: (name: string, args: unknown) => Promise<ToolCallResult>\n}\n\n/** Build the transport-agnostic bridge core for the given tools. */\nexport function createToolBridgeCore(\n tools: Array<AnyTool>,\n options: ToolBridgeCoreOptions = {},\n): ToolBridgeCore {\n const toolsByName = new Map(tools.map((tool) => [tool.name, tool]))\n const permission = options.permission\n\n const permissionDescriptor: ToolDescriptor | undefined = permission\n ? {\n name: permission.toolName,\n description:\n 'Permission prompt: returns {behavior:\"allow\"|\"deny\"} for a requested action.',\n inputSchema: { type: 'object', properties: {} },\n }\n : undefined\n\n return {\n listTools() {\n return [\n ...tools.map((tool) => ({\n name: tool.name,\n description: tool.description,\n inputSchema: toObjectSchema(tool.inputSchema),\n })),\n ...(permissionDescriptor ? [permissionDescriptor] : []),\n ]\n },\n\n async callTool(name, args) {\n if (permission && name === permission.toolName) {\n const result = await permission.resolve(args ?? {})\n return { content: [{ type: 'text', text: JSON.stringify(result) }] }\n }\n const tool = toolsByName.get(name)\n if (!tool?.execute) throw new Error(`Unknown tool: ${name}`)\n try {\n const result: unknown = await tool.execute(args ?? {}, {\n context: options.context,\n abortSignal: options.signal,\n // No-op default so tools that always call it (e.g. code mode) don't\n // crash when the transport didn't wire a sink.\n emitCustomEvent: options.emitCustomEvent ?? (() => {}),\n })\n const text =\n typeof result === 'string' ? result : JSON.stringify(result)\n return { content: [{ type: 'text', text }] }\n } catch (error) {\n const message = error instanceof Error ? error.message : String(error)\n return {\n isError: true,\n content: [\n { type: 'text', text: `Tool execution failed: ${message}` },\n ],\n }\n }\n },\n }\n}\n\n/**\n * Minimal JSON-RPC dispatcher over a {@link ToolBridgeCore}, so a `fetch`-based\n * transport (Worker / Durable Object) can serve MCP `initialize` / `tools/list`\n * / `tools/call` without the node-specific HTTP transport. Returns the JSON-RPC\n * response object, or `null` for a notification (no `id`).\n */\nexport async function handleBridgeJsonRpc(\n core: ToolBridgeCore,\n message: unknown,\n): Promise<unknown> {\n if (message === null || typeof message !== 'object') {\n return {\n jsonrpc: '2.0',\n id: null,\n error: { code: -32600, message: 'Invalid Request' },\n }\n }\n const rpc = message as { id?: unknown; method?: unknown; params?: unknown }\n const id = rpc.id ?? null\n const respond = (result: unknown): unknown => ({ jsonrpc: '2.0', id, result })\n switch (rpc.method) {\n case 'initialize':\n return respond({\n protocolVersion: '2024-11-05',\n capabilities: { tools: {} },\n serverInfo: { name: BRIDGED_MCP_SERVER_NAME, version: '1.0.0' },\n })\n case 'notifications/initialized':\n return null\n case 'tools/list':\n return respond({ tools: core.listTools() })\n case 'tools/call': {\n const params = (rpc.params ?? {}) as {\n name?: unknown\n arguments?: unknown\n }\n if (typeof params.name !== 'string') {\n return {\n jsonrpc: '2.0',\n id,\n error: { code: -32602, message: 'Invalid params: name' },\n }\n }\n return respond(await core.callTool(params.name, params.arguments ?? {}))\n }\n default:\n return {\n jsonrpc: '2.0',\n id,\n error: { code: -32601, message: 'Method not found' },\n }\n }\n}\n\n/**\n * Constant-time check of an `Authorization: Bearer <token>` header against the\n * expected token. Length mismatch returns false early (token length is not\n * secret); equal-length comparison is timing-safe.\n */\nexport function timingSafeBearerEqual(\n header: string | undefined,\n token: string,\n): boolean {\n if (header === undefined) return false\n const a = Buffer.from(header)\n const b = Buffer.from(`Bearer ${token}`)\n if (a.length !== b.length) return false\n return timingSafeEqual(a, b)\n}\n\nexport interface HostToolBridge {\n /** MCP server name; tools appear to the agent as `mcp__<name>__<tool>`. */\n name: string\n /** URL the SANDBOX uses to reach this bridge. */\n url: string\n /** Per-run bearer token gating the endpoint. */\n token: string\n close: () => Promise<void>\n}\n\nexport interface StartBridgeOptions extends ToolBridgeCoreOptions {\n /** Hostname the sandbox uses to reach the host (e.g. `host.docker.internal`). */\n hostForSandbox: string\n /**\n * Address to bind the listener to. Defaults to `127.0.0.1` (loopback) and is\n * widened to `0.0.0.0` only when the sandbox reaches the host via\n * `host.docker.internal` (a container can't reach the host's loopback).\n */\n bindAddress?: string\n}\n\nfunction buildMcpServer(core: ToolBridgeCore): McpServer {\n const server = new McpServer(\n { name: BRIDGED_MCP_SERVER_NAME, version: '1.0.0' },\n { capabilities: { tools: {} } },\n )\n server.server.setRequestHandler(ListToolsRequestSchema, () => ({\n tools: core.listTools(),\n }))\n server.server.setRequestHandler(CallToolRequestSchema, async (request) => {\n const result = await core.callTool(\n request.params.name,\n request.params.arguments ?? {},\n )\n return {\n content: result.content,\n ...(result.isError ? { isError: true } : {}),\n }\n })\n return server\n}\n\n/**\n * Start the `node:http` MCP tool-proxy bridge for the given tools. For a\n * long-running host (laptop / CI / Docker orchestrator). Serverless/edge\n * orchestrators serve {@link createToolBridgeCore} from their own `fetch`\n * handler instead.\n */\nexport async function startHostToolBridge(\n tools: Array<AnyTool>,\n options: StartBridgeOptions,\n): Promise<HostToolBridge> {\n const token = randomBytes(24).toString('hex')\n const core = createToolBridgeCore(tools, options)\n // Loopback by default; widen to all interfaces only for the Docker bridge,\n // which a container reaches via host.docker.internal (host gateway).\n const bindAddress =\n options.bindAddress ??\n (options.hostForSandbox === 'host.docker.internal'\n ? '0.0.0.0'\n : '127.0.0.1')\n\n const httpServer = createServer((req, res) => {\n void (async () => {\n if (!timingSafeBearerEqual(req.headers['authorization'], token)) {\n res.writeHead(401).end('unauthorized')\n return\n }\n const server = buildMcpServer(core)\n const transport = new StreamableHTTPServerTransport({\n sessionIdGenerator: undefined,\n })\n res.on('close', () => {\n void transport.close()\n void server.close()\n })\n await server.connect(transport)\n\n let body = ''\n for await (const chunk of req) body += chunk\n let parsed: unknown\n try {\n parsed = body ? JSON.parse(body) : undefined\n } catch {\n // Malformed agent request → 400, distinct from an internal 500.\n if (!res.headersSent) res.writeHead(400).end('invalid JSON body')\n return\n }\n await transport.handleRequest(req, res, parsed)\n })().catch((error: unknown) => {\n // Log the underlying fault — on the host/Docker path there is no run-log\n // capturing it, so swallowing it leaves an operator with nothing.\n console.error('[tool-bridge] request handler failed:', error)\n if (!res.headersSent) res.writeHead(500).end('bridge error')\n })\n })\n\n await new Promise<void>((resolve) =>\n httpServer.listen(0, bindAddress, resolve),\n )\n const port = (httpServer.address() as AddressInfo).port\n const url = `http://${options.hostForSandbox}:${port}/mcp`\n\n return {\n name: BRIDGED_MCP_SERVER_NAME,\n url,\n token,\n close: () =>\n new Promise<void>((resolve) => httpServer.close(() => resolve())),\n }\n}\n\n/** A provisioned, reachable bridge endpoint (same shape as {@link HostToolBridge}). */\nexport type ProvisionedBridge = HostToolBridge\n\nexport interface ToolBridgeProvisionOptions extends ToolBridgeCoreOptions {\n /** Sandbox provider name, to derive how the sandbox reaches the bridge. */\n provider: string\n}\n\n/**\n * Stands up the tool-bridge endpoint for a run. The seam that makes the bridge\n * portable across runtimes: a harness adapter asks its capability context for a\n * provisioner and uses {@link nodeHttpBridgeProvisioner} as the default (host /\n * Docker). A serverless/edge orchestrator PROVIDES its own — e.g. a Durable\n * Object that mounts {@link createToolBridgeCore} / {@link handleBridgeJsonRpc}\n * on its `fetch` handler and returns a sandbox-reachable URL — so no raw TCP\n * listener is needed.\n */\nexport interface ToolBridgeProvisioner {\n provision: (\n tools: Array<AnyTool>,\n options: ToolBridgeProvisionOptions,\n ) => Promise<ProvisionedBridge>\n}\n\n/** Default provisioner: a `node:http` listener on the host. */\nexport const nodeHttpBridgeProvisioner: ToolBridgeProvisioner = {\n provision(tools, options) {\n const { provider, ...core } = options\n return startHostToolBridge(tools, {\n hostForSandbox: hostForSandbox(provider),\n ...core,\n })\n },\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoCA,IAAa,0BAA0B;;AAGvC,SAAgB,eAAe,UAA0B;CACvD,OAAO,aAAa,YAAY,aAAa,QACzC,yBACA;AACN;;;;;;AAmDA,SAAS,eAAe,QAGtB;CACA,IACE,WAAW,QACX,OAAO,WAAW,YAClB,UAAU,UACV,OAAO,SAAS,UAEhB,OAAO;EAAE,GAAG;EAAQ,MAAM;CAAS;CAErC,OAAO;EAAE,MAAM;EAAU,YAAY,CAAC;CAAE;AAC1C;;AAmBA,SAAgB,qBACd,OACA,UAAiC,CAAC,GAClB;CAChB,MAAM,cAAc,IAAI,IAAI,MAAM,KAAK,SAAS,CAAC,KAAK,MAAM,IAAI,CAAC,CAAC;CAClE,MAAM,aAAa,QAAQ;CAE3B,MAAM,uBAAmD,aACrD;EACE,MAAM,WAAW;EACjB,aACE;EACF,aAAa;GAAE,MAAM;GAAU,YAAY,CAAC;EAAE;CAChD,IACA,KAAA;CAEJ,OAAO;EACL,YAAY;GACV,OAAO,CACL,GAAG,MAAM,KAAK,UAAU;IACtB,MAAM,KAAK;IACX,aAAa,KAAK;IAClB,aAAa,eAAe,KAAK,WAAW;GAC9C,EAAE,GACF,GAAI,uBAAuB,CAAC,oBAAoB,IAAI,CAAC,CACvD;EACF;EAEA,MAAM,SAAS,MAAM,MAAM;GACzB,IAAI,cAAc,SAAS,WAAW,UAAU;IAC9C,MAAM,SAAS,MAAM,WAAW,QAAQ,QAAQ,CAAC,CAAC;IAClD,OAAO,EAAE,SAAS,CAAC;KAAE,MAAM;KAAQ,MAAM,KAAK,UAAU,MAAM;IAAE,CAAC,EAAE;GACrE;GACA,MAAM,OAAO,YAAY,IAAI,IAAI;GACjC,IAAI,CAAC,MAAM,SAAS,MAAM,IAAI,MAAM,iBAAiB,MAAM;GAC3D,IAAI;IACF,MAAM,SAAkB,MAAM,KAAK,QAAQ,QAAQ,CAAC,GAAG;KACrD,SAAS,QAAQ;KACjB,aAAa,QAAQ;KAGrB,iBAAiB,QAAQ,0BAA0B,CAAC;IACtD,CAAC;IAGD,OAAO,EAAE,SAAS,CAAC;KAAE,MAAM;KAAQ,MADjC,OAAO,WAAW,WAAW,SAAS,KAAK,UAAU,MAAM;IACrB,CAAC,EAAE;GAC7C,SAAS,OAAO;IAEd,OAAO;KACL,SAAS;KACT,SAAS,CACP;MAAE,MAAM;MAAQ,MAAM,0BAJV,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;KAIP,CAC5D;IACF;GACF;EACF;CACF;AACF;;;;;;;AAQA,eAAsB,oBACpB,MACA,SACkB;CAClB,IAAI,YAAY,QAAQ,OAAO,YAAY,UACzC,OAAO;EACL,SAAS;EACT,IAAI;EACJ,OAAO;GAAE,MAAM;GAAQ,SAAS;EAAkB;CACpD;CAEF,MAAM,MAAM;CACZ,MAAM,KAAK,IAAI,MAAM;CACrB,MAAM,WAAW,YAA8B;EAAE,SAAS;EAAO;EAAI;CAAO;CAC5E,QAAQ,IAAI,QAAZ;EACE,KAAK,cACH,OAAO,QAAQ;GACb,iBAAiB;GACjB,cAAc,EAAE,OAAO,CAAC,EAAE;GAC1B,YAAY;IAAE,MAAM;IAAyB,SAAS;GAAQ;EAChE,CAAC;EACH,KAAK,6BACH,OAAO;EACT,KAAK,cACH,OAAO,QAAQ,EAAE,OAAO,KAAK,UAAU,EAAE,CAAC;EAC5C,KAAK,cAAc;GACjB,MAAM,SAAU,IAAI,UAAU,CAAC;GAI/B,IAAI,OAAO,OAAO,SAAS,UACzB,OAAO;IACL,SAAS;IACT;IACA,OAAO;KAAE,MAAM;KAAQ,SAAS;IAAuB;GACzD;GAEF,OAAO,QAAQ,MAAM,KAAK,SAAS,OAAO,MAAM,OAAO,aAAa,CAAC,CAAC,CAAC;EACzE;EACA,SACE,OAAO;GACL,SAAS;GACT;GACA,OAAO;IAAE,MAAM;IAAQ,SAAS;GAAmB;EACrD;CACJ;AACF;;;;;;AAOA,SAAgB,sBACd,QACA,OACS;CACT,IAAI,WAAW,KAAA,GAAW,OAAO;CACjC,MAAM,IAAI,OAAO,KAAK,MAAM;CAC5B,MAAM,IAAI,OAAO,KAAK,UAAU,OAAO;CACvC,IAAI,EAAE,WAAW,EAAE,QAAQ,OAAO;CAClC,OAAO,gBAAgB,GAAG,CAAC;AAC7B;AAuBA,SAAS,eAAe,MAAiC;CACvD,MAAM,SAAS,IAAI,UACjB;EAAE,MAAM;EAAyB,SAAS;CAAQ,GAClD,EAAE,cAAc,EAAE,OAAO,CAAC,EAAE,EAAE,CAChC;CACA,OAAO,OAAO,kBAAkB,+BAA+B,EAC7D,OAAO,KAAK,UAAU,EACxB,EAAE;CACF,OAAO,OAAO,kBAAkB,uBAAuB,OAAO,YAAY;EACxE,MAAM,SAAS,MAAM,KAAK,SACxB,QAAQ,OAAO,MACf,QAAQ,OAAO,aAAa,CAAC,CAC/B;EACA,OAAO;GACL,SAAS,OAAO;GAChB,GAAI,OAAO,UAAU,EAAE,SAAS,KAAK,IAAI,CAAC;EAC5C;CACF,CAAC;CACD,OAAO;AACT;;;;;;;AAQA,eAAsB,oBACpB,OACA,SACyB;CACzB,MAAM,QAAQ,YAAY,EAAE,CAAC,CAAC,SAAS,KAAK;CAC5C,MAAM,OAAO,qBAAqB,OAAO,OAAO;CAGhD,MAAM,cACJ,QAAQ,gBACP,QAAQ,mBAAmB,yBACxB,YACA;CAEN,MAAM,aAAa,cAAc,KAAK,QAAQ;EAC5C,CAAM,YAAY;GAChB,IAAI,CAAC,sBAAsB,IAAI,QAAQ,kBAAkB,KAAK,GAAG;IAC/D,IAAI,UAAU,GAAG,CAAC,CAAC,IAAI,cAAc;IACrC;GACF;GACA,MAAM,SAAS,eAAe,IAAI;GAClC,MAAM,YAAY,IAAI,8BAA8B,EAClD,oBAAoB,KAAA,EACtB,CAAC;GACD,IAAI,GAAG,eAAe;IACpB,UAAe,MAAM;IACrB,OAAY,MAAM;GACpB,CAAC;GACD,MAAM,OAAO,QAAQ,SAAS;GAE9B,IAAI,OAAO;GACX,WAAW,MAAM,SAAS,KAAK,QAAQ;GACvC,IAAI;GACJ,IAAI;IACF,SAAS,OAAO,KAAK,MAAM,IAAI,IAAI,KAAA;GACrC,QAAQ;IAEN,IAAI,CAAC,IAAI,aAAa,IAAI,UAAU,GAAG,CAAC,CAAC,IAAI,mBAAmB;IAChE;GACF;GACA,MAAM,UAAU,cAAc,KAAK,KAAK,MAAM;EAChD,EAAA,CAAG,CAAC,CAAC,OAAO,UAAmB;GAG7B,QAAQ,MAAM,yCAAyC,KAAK;GAC5D,IAAI,CAAC,IAAI,aAAa,IAAI,UAAU,GAAG,CAAC,CAAC,IAAI,cAAc;EAC7D,CAAC;CACH,CAAC;CAED,MAAM,IAAI,SAAe,YACvB,WAAW,OAAO,GAAG,aAAa,OAAO,CAC3C;CACA,MAAM,OAAQ,WAAW,QAAQ,CAAC,CAAiB;CAGnD,OAAO;EACL,MAAM;EACN,eAJoB,QAAQ,eAAe,GAAG,KAAK;EAKnD;EACA,aACE,IAAI,SAAe,YAAY,WAAW,YAAY,QAAQ,CAAC,CAAC;CACpE;AACF;;AA2BA,IAAa,4BAAmD,EAC9D,UAAU,OAAO,SAAS;CACxB,MAAM,EAAE,UAAU,GAAG,SAAS;CAC9B,OAAO,oBAAoB,OAAO;EAChC,gBAAgB,eAAe,QAAQ;EACvC,GAAG;CACL,CAAC;AACH,EACF"}
1
+ {"version":3,"file":"tool-bridge.js","names":[],"sources":["../../src/tool-bridge.ts"],"sourcesContent":["/**\n * MCP tool-proxy bridge, shared by all harness adapters.\n *\n * Exposes chat()-provided server tools to an in-sandbox agent as an MCP server.\n * The agent (inside the sandbox) calls `mcp__tanstack__<tool>`; the call is\n * proxied OUT to a bridge endpoint, where the tool's `execute()` runs in the\n * orchestrator process (with its closures / DB / secrets), and the result is\n * returned into the sandbox.\n *\n * The bridge is split into a transport-agnostic CORE and a TRANSPORT:\n * - {@link createToolBridgeCore} owns tool dispatch + the permission resolver\n * (no I/O). It is what makes the bridge portable.\n * - {@link startHostToolBridge} is the `node:http` transport for a long-running\n * host (laptop / CI / Docker orchestrator). It binds loopback unless the\n * sandbox must reach it via `host.docker.internal`, and authenticates with a\n * constant-time bearer check.\n * - A serverless/edge orchestrator (e.g. a Durable Object) instead serves the\n * SAME core from its own `fetch` handler — no raw TCP listener — see\n * {@link handleBridgeJsonRpc} and the Cloudflare example.\n */\nimport { createServer } from 'node:http'\nimport { randomBytes, timingSafeEqual } from 'node:crypto'\nimport { once } from 'node:events'\nimport { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'\nimport { StreamableHTTPServerTransport } from '@modelcontextprotocol/sdk/server/streamableHttp.js'\nimport {\n CallToolRequestSchema,\n ListToolsRequestSchema,\n} from '@modelcontextprotocol/sdk/types.js'\nimport type { AddressInfo } from 'node:net'\nimport type { AnyTool } from '@tanstack/ai'\n\n/**\n * Name of the bridged MCP server. The agent sees tools as\n * `mcp__tanstack__<tool>`; each adapter's stream translator strips this prefix\n * so tool-call events match the names the application registered.\n */\nexport const BRIDGED_MCP_SERVER_NAME = 'tanstack'\n\n/** Hostname the sandbox uses to reach the bridge endpoint, per provider. */\nexport function hostForSandbox(provider: string): string {\n return provider === 'docker' || provider === 'sbx'\n ? 'host.docker.internal'\n : '127.0.0.1'\n}\n\n/** Result of a permission decision returned to the harness's prompt tool. */\nexport interface PermissionToolResult {\n behavior: 'allow' | 'deny'\n message?: string\n updatedInput?: unknown\n}\n\nexport interface BridgePermission {\n toolName: string\n resolve: (input: {\n tool_name?: string\n input?: unknown\n }) => PermissionToolResult | Promise<PermissionToolResult>\n}\n\nexport interface ToolBridgeCoreOptions {\n /** Runtime context forwarded to each tool's `execute()`. */\n context?: unknown\n /** Abort signal forwarded to each tool's `execute()`. */\n signal?: AbortSignal\n /**\n * Forwarded to each tool's `execute()` so a bridged tool can stream progress /\n * custom events back to the client mid-execution (e.g. code mode's\n * `code_mode:console` logs). Without it those events are silently dropped — the\n * bridge runs out-of-band from the main tool executor, so the executor's own\n * `emitCustomEvent` never reaches a bridged tool. The harness adapter supplies\n * one that injects a CUSTOM chunk into its live output stream.\n */\n emitCustomEvent?: (eventName: string, value: Record<string, unknown>) => void\n /**\n * Optional permission-prompt tool (e.g. for Claude Code's\n * `--permission-prompt-tool`). When set, the bridge exposes an extra MCP tool\n * `<name>` whose handler returns the orchestrator's allow/deny decision.\n */\n permission?: BridgePermission\n}\n\n/** An MCP tool descriptor as advertised to the in-sandbox agent. */\nexport interface ToolDescriptor {\n name: string\n description?: string\n inputSchema: { type: 'object'; [key: string]: unknown }\n}\n\n/**\n * Coerce a tool's `inputSchema` into the object-schema shape MCP advertises,\n * substituting an empty object schema when it isn't already a JSON-schema object\n * (project rule: a guard, not an `as` cast).\n */\nfunction toObjectSchema(schema: unknown): {\n type: 'object'\n [key: string]: unknown\n} {\n if (\n schema !== null &&\n typeof schema === 'object' &&\n 'type' in schema &&\n schema.type === 'object'\n ) {\n return { ...schema, type: 'object' }\n }\n return { type: 'object', properties: {} }\n}\n\n/** MCP `tools/call` result shape. */\nexport interface ToolCallResult {\n content: Array<{ type: 'text'; text: string }>\n isError?: boolean\n}\n\n/**\n * Transport-agnostic bridge logic: list tools, and dispatch a tool/permission\n * call. No sockets, no auth — a transport ({@link startHostToolBridge} or a\n * `fetch` handler) wraps this and owns I/O + the bearer check.\n */\nexport interface ToolBridgeCore {\n listTools: () => Array<ToolDescriptor>\n callTool: (name: string, args: unknown) => Promise<ToolCallResult>\n}\n\n/** Build the transport-agnostic bridge core for the given tools. */\nexport function createToolBridgeCore(\n tools: Array<AnyTool>,\n options: ToolBridgeCoreOptions = {},\n): ToolBridgeCore {\n const toolsByName = new Map(tools.map((tool) => [tool.name, tool]))\n const permission = options.permission\n\n const permissionDescriptor: ToolDescriptor | undefined = permission\n ? {\n name: permission.toolName,\n description:\n 'Permission prompt: returns {behavior:\"allow\"|\"deny\"} for a requested action.',\n inputSchema: { type: 'object', properties: {} },\n }\n : undefined\n\n return {\n listTools() {\n return [\n ...tools.map((tool) => ({\n name: tool.name,\n description: tool.description,\n inputSchema: toObjectSchema(tool.inputSchema),\n })),\n ...(permissionDescriptor ? [permissionDescriptor] : []),\n ]\n },\n\n async callTool(name, args) {\n if (permission && name === permission.toolName) {\n const result = await permission.resolve(args ?? {})\n return { content: [{ type: 'text', text: JSON.stringify(result) }] }\n }\n const tool = toolsByName.get(name)\n if (!tool?.execute) throw new Error(`Unknown tool: ${name}`)\n try {\n const result: unknown = await tool.execute(args ?? {}, {\n context: options.context,\n abortSignal: options.signal,\n // No-op default so tools that always call it (e.g. code mode) don't\n // crash when the transport didn't wire a sink.\n emitCustomEvent: options.emitCustomEvent ?? (() => {}),\n })\n const text =\n typeof result === 'string' ? result : JSON.stringify(result)\n return { content: [{ type: 'text', text }] }\n } catch (error) {\n const message = error instanceof Error ? error.message : String(error)\n return {\n isError: true,\n content: [\n { type: 'text', text: `Tool execution failed: ${message}` },\n ],\n }\n }\n },\n }\n}\n\n/**\n * Minimal JSON-RPC dispatcher over a {@link ToolBridgeCore}, so a `fetch`-based\n * transport (Worker / Durable Object) can serve MCP `initialize` / `tools/list`\n * / `tools/call` without the node-specific HTTP transport. Returns the JSON-RPC\n * response object, or `null` for a notification (no `id`).\n */\nexport async function handleBridgeJsonRpc(\n core: ToolBridgeCore,\n message: unknown,\n): Promise<unknown> {\n if (message === null || typeof message !== 'object') {\n return {\n jsonrpc: '2.0',\n id: null,\n error: { code: -32600, message: 'Invalid Request' },\n }\n }\n const rpc = message as { id?: unknown; method?: unknown; params?: unknown }\n const id = rpc.id ?? null\n const respond = (result: unknown): unknown => ({ jsonrpc: '2.0', id, result })\n switch (rpc.method) {\n case 'initialize':\n return respond({\n protocolVersion: '2024-11-05',\n capabilities: { tools: {} },\n serverInfo: { name: BRIDGED_MCP_SERVER_NAME, version: '1.0.0' },\n })\n case 'notifications/initialized':\n return null\n case 'tools/list':\n return respond({ tools: core.listTools() })\n case 'tools/call': {\n const params = (rpc.params ?? {}) as {\n name?: unknown\n arguments?: unknown\n }\n if (typeof params.name !== 'string') {\n return {\n jsonrpc: '2.0',\n id,\n error: { code: -32602, message: 'Invalid params: name' },\n }\n }\n return respond(await core.callTool(params.name, params.arguments ?? {}))\n }\n default:\n return {\n jsonrpc: '2.0',\n id,\n error: { code: -32601, message: 'Method not found' },\n }\n }\n}\n\n/**\n * Constant-time check of an `Authorization: Bearer <token>` header against the\n * expected token. Length mismatch returns false early (token length is not\n * secret); equal-length comparison is timing-safe.\n */\nexport function timingSafeBearerEqual(\n header: string | undefined,\n token: string,\n): boolean {\n if (header === undefined) return false\n const a = Buffer.from(header)\n const b = Buffer.from(`Bearer ${token}`)\n if (a.length !== b.length) return false\n return timingSafeEqual(a, b)\n}\n\nexport interface HostToolBridge {\n /** MCP server name; tools appear to the agent as `mcp__<name>__<tool>`. */\n name: string\n /** URL the SANDBOX uses to reach this bridge. */\n url: string\n /** Per-run bearer token gating the endpoint. */\n token: string\n close: () => Promise<void>\n}\n\nexport interface StartBridgeOptions extends ToolBridgeCoreOptions {\n /** Hostname the sandbox uses to reach the host (e.g. `host.docker.internal`). */\n hostForSandbox: string\n /**\n * Address to bind the listener to. Defaults to `127.0.0.1` (loopback) and is\n * widened to `0.0.0.0` only when the sandbox reaches the host via\n * `host.docker.internal` (a container can't reach the host's loopback).\n */\n bindAddress?: string\n}\n\nfunction buildMcpServer(core: ToolBridgeCore): McpServer {\n const server = new McpServer(\n { name: BRIDGED_MCP_SERVER_NAME, version: '1.0.0' },\n { capabilities: { tools: {} } },\n )\n server.server.setRequestHandler(ListToolsRequestSchema, () => ({\n tools: core.listTools(),\n }))\n server.server.setRequestHandler(CallToolRequestSchema, async (request) => {\n const result = await core.callTool(\n request.params.name,\n request.params.arguments ?? {},\n )\n return {\n content: result.content,\n ...(result.isError ? { isError: true } : {}),\n }\n })\n return server\n}\n\n/**\n * Start the `node:http` MCP tool-proxy bridge for the given tools. For a\n * long-running host (laptop / CI / Docker orchestrator). Serverless/edge\n * orchestrators serve {@link createToolBridgeCore} from their own `fetch`\n * handler instead.\n */\nexport async function startHostToolBridge(\n tools: Array<AnyTool>,\n options: StartBridgeOptions,\n): Promise<HostToolBridge> {\n const token = randomBytes(24).toString('hex')\n const core = createToolBridgeCore(tools, options)\n // Loopback by default; widen to all interfaces only for the Docker bridge,\n // which a container reaches via host.docker.internal (host gateway).\n const bindAddress =\n options.bindAddress ??\n (options.hostForSandbox === 'host.docker.internal'\n ? '0.0.0.0'\n : '127.0.0.1')\n\n const httpServer = createServer((req, res) => {\n void (async () => {\n if (!timingSafeBearerEqual(req.headers['authorization'], token)) {\n res.writeHead(401).end('unauthorized')\n return\n }\n const server = buildMcpServer(core)\n const transport = new StreamableHTTPServerTransport({\n sessionIdGenerator: undefined,\n })\n res.on('close', () => {\n void transport.close()\n void server.close()\n })\n await server.connect(transport)\n\n let body = ''\n for await (const chunk of req) body += chunk\n let parsed: unknown\n try {\n parsed = body ? JSON.parse(body) : undefined\n } catch {\n // Malformed agent request → 400, distinct from an internal 500.\n if (!res.headersSent) res.writeHead(400).end('invalid JSON body')\n return\n }\n await transport.handleRequest(req, res, parsed)\n })().catch((error: unknown) => {\n // Log the underlying fault — on the host/Docker path there is no run-log\n // capturing it, so swallowing it leaves an operator with nothing.\n console.error('[tool-bridge] request handler failed:', error)\n if (!res.headersSent) res.writeHead(500).end('bridge error')\n })\n })\n\n await once(httpServer.listen(0, bindAddress), 'listening')\n const port = (httpServer.address() as AddressInfo).port\n const url = `http://${options.hostForSandbox}:${port}/mcp`\n\n return {\n name: BRIDGED_MCP_SERVER_NAME,\n url,\n token,\n close: () =>\n new Promise<void>((resolve) => httpServer.close(() => resolve())),\n }\n}\n\n/** A provisioned, reachable bridge endpoint (same shape as {@link HostToolBridge}). */\nexport type ProvisionedBridge = HostToolBridge\n\nexport interface ToolBridgeProvisionOptions extends ToolBridgeCoreOptions {\n /** Sandbox provider name, to derive how the sandbox reaches the bridge. */\n provider: string\n}\n\n/**\n * Stands up the tool-bridge endpoint for a run. The seam that makes the bridge\n * portable across runtimes: a harness adapter asks its capability context for a\n * provisioner and uses {@link nodeHttpBridgeProvisioner} as the default (host /\n * Docker). A serverless/edge orchestrator PROVIDES its own — e.g. a Durable\n * Object that mounts {@link createToolBridgeCore} / {@link handleBridgeJsonRpc}\n * on its `fetch` handler and returns a sandbox-reachable URL — so no raw TCP\n * listener is needed.\n */\nexport interface ToolBridgeProvisioner {\n provision: (\n tools: Array<AnyTool>,\n options: ToolBridgeProvisionOptions,\n ) => Promise<ProvisionedBridge>\n}\n\n/** Default provisioner: a `node:http` listener on the host. */\nexport const nodeHttpBridgeProvisioner: ToolBridgeProvisioner = {\n provision(tools, options) {\n const { provider, ...core } = options\n return startHostToolBridge(tools, {\n hostForSandbox: hostForSandbox(provider),\n ...core,\n })\n },\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,IAAa,0BAA0B;;AAGvC,SAAgB,eAAe,UAA0B;CACvD,OAAO,aAAa,YAAY,aAAa,QACzC,yBACA;AACN;;;;;;AAmDA,SAAS,eAAe,QAGtB;CACA,IACE,WAAW,QACX,OAAO,WAAW,YAClB,UAAU,UACV,OAAO,SAAS,UAEhB,OAAO;EAAE,GAAG;EAAQ,MAAM;CAAS;CAErC,OAAO;EAAE,MAAM;EAAU,YAAY,CAAC;CAAE;AAC1C;;AAmBA,SAAgB,qBACd,OACA,UAAiC,CAAC,GAClB;CAChB,MAAM,cAAc,IAAI,IAAI,MAAM,KAAK,SAAS,CAAC,KAAK,MAAM,IAAI,CAAC,CAAC;CAClE,MAAM,aAAa,QAAQ;CAE3B,MAAM,uBAAmD,aACrD;EACE,MAAM,WAAW;EACjB,aACE;EACF,aAAa;GAAE,MAAM;GAAU,YAAY,CAAC;EAAE;CAChD,IACA,KAAA;CAEJ,OAAO;EACL,YAAY;GACV,OAAO,CACL,GAAG,MAAM,KAAK,UAAU;IACtB,MAAM,KAAK;IACX,aAAa,KAAK;IAClB,aAAa,eAAe,KAAK,WAAW;GAC9C,EAAE,GACF,GAAI,uBAAuB,CAAC,oBAAoB,IAAI,CAAC,CACvD;EACF;EAEA,MAAM,SAAS,MAAM,MAAM;GACzB,IAAI,cAAc,SAAS,WAAW,UAAU;IAC9C,MAAM,SAAS,MAAM,WAAW,QAAQ,QAAQ,CAAC,CAAC;IAClD,OAAO,EAAE,SAAS,CAAC;KAAE,MAAM;KAAQ,MAAM,KAAK,UAAU,MAAM;IAAE,CAAC,EAAE;GACrE;GACA,MAAM,OAAO,YAAY,IAAI,IAAI;GACjC,IAAI,CAAC,MAAM,SAAS,MAAM,IAAI,MAAM,iBAAiB,MAAM;GAC3D,IAAI;IACF,MAAM,SAAkB,MAAM,KAAK,QAAQ,QAAQ,CAAC,GAAG;KACrD,SAAS,QAAQ;KACjB,aAAa,QAAQ;KAGrB,iBAAiB,QAAQ,0BAA0B,CAAC;IACtD,CAAC;IAGD,OAAO,EAAE,SAAS,CAAC;KAAE,MAAM;KAAQ,MADjC,OAAO,WAAW,WAAW,SAAS,KAAK,UAAU,MAAM;IACrB,CAAC,EAAE;GAC7C,SAAS,OAAO;IAEd,OAAO;KACL,SAAS;KACT,SAAS,CACP;MAAE,MAAM;MAAQ,MAAM,0BAJV,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;KAIP,CAC5D;IACF;GACF;EACF;CACF;AACF;;;;;;;AAQA,eAAsB,oBACpB,MACA,SACkB;CAClB,IAAI,YAAY,QAAQ,OAAO,YAAY,UACzC,OAAO;EACL,SAAS;EACT,IAAI;EACJ,OAAO;GAAE,MAAM;GAAQ,SAAS;EAAkB;CACpD;CAEF,MAAM,MAAM;CACZ,MAAM,KAAK,IAAI,MAAM;CACrB,MAAM,WAAW,YAA8B;EAAE,SAAS;EAAO;EAAI;CAAO;CAC5E,QAAQ,IAAI,QAAZ;EACE,KAAK,cACH,OAAO,QAAQ;GACb,iBAAiB;GACjB,cAAc,EAAE,OAAO,CAAC,EAAE;GAC1B,YAAY;IAAE,MAAM;IAAyB,SAAS;GAAQ;EAChE,CAAC;EACH,KAAK,6BACH,OAAO;EACT,KAAK,cACH,OAAO,QAAQ,EAAE,OAAO,KAAK,UAAU,EAAE,CAAC;EAC5C,KAAK,cAAc;GACjB,MAAM,SAAU,IAAI,UAAU,CAAC;GAI/B,IAAI,OAAO,OAAO,SAAS,UACzB,OAAO;IACL,SAAS;IACT;IACA,OAAO;KAAE,MAAM;KAAQ,SAAS;IAAuB;GACzD;GAEF,OAAO,QAAQ,MAAM,KAAK,SAAS,OAAO,MAAM,OAAO,aAAa,CAAC,CAAC,CAAC;EACzE;EACA,SACE,OAAO;GACL,SAAS;GACT;GACA,OAAO;IAAE,MAAM;IAAQ,SAAS;GAAmB;EACrD;CACJ;AACF;;;;;;AAOA,SAAgB,sBACd,QACA,OACS;CACT,IAAI,WAAW,KAAA,GAAW,OAAO;CACjC,MAAM,IAAI,OAAO,KAAK,MAAM;CAC5B,MAAM,IAAI,OAAO,KAAK,UAAU,OAAO;CACvC,IAAI,EAAE,WAAW,EAAE,QAAQ,OAAO;CAClC,OAAO,gBAAgB,GAAG,CAAC;AAC7B;AAuBA,SAAS,eAAe,MAAiC;CACvD,MAAM,SAAS,IAAI,UACjB;EAAE,MAAM;EAAyB,SAAS;CAAQ,GAClD,EAAE,cAAc,EAAE,OAAO,CAAC,EAAE,EAAE,CAChC;CACA,OAAO,OAAO,kBAAkB,+BAA+B,EAC7D,OAAO,KAAK,UAAU,EACxB,EAAE;CACF,OAAO,OAAO,kBAAkB,uBAAuB,OAAO,YAAY;EACxE,MAAM,SAAS,MAAM,KAAK,SACxB,QAAQ,OAAO,MACf,QAAQ,OAAO,aAAa,CAAC,CAC/B;EACA,OAAO;GACL,SAAS,OAAO;GAChB,GAAI,OAAO,UAAU,EAAE,SAAS,KAAK,IAAI,CAAC;EAC5C;CACF,CAAC;CACD,OAAO;AACT;;;;;;;AAQA,eAAsB,oBACpB,OACA,SACyB;CACzB,MAAM,QAAQ,YAAY,EAAE,CAAC,CAAC,SAAS,KAAK;CAC5C,MAAM,OAAO,qBAAqB,OAAO,OAAO;CAGhD,MAAM,cACJ,QAAQ,gBACP,QAAQ,mBAAmB,yBACxB,YACA;CAEN,MAAM,aAAa,cAAc,KAAK,QAAQ;EAC5C,CAAM,YAAY;GAChB,IAAI,CAAC,sBAAsB,IAAI,QAAQ,kBAAkB,KAAK,GAAG;IAC/D,IAAI,UAAU,GAAG,CAAC,CAAC,IAAI,cAAc;IACrC;GACF;GACA,MAAM,SAAS,eAAe,IAAI;GAClC,MAAM,YAAY,IAAI,8BAA8B,EAClD,oBAAoB,KAAA,EACtB,CAAC;GACD,IAAI,GAAG,eAAe;IACpB,UAAe,MAAM;IACrB,OAAY,MAAM;GACpB,CAAC;GACD,MAAM,OAAO,QAAQ,SAAS;GAE9B,IAAI,OAAO;GACX,WAAW,MAAM,SAAS,KAAK,QAAQ;GACvC,IAAI;GACJ,IAAI;IACF,SAAS,OAAO,KAAK,MAAM,IAAI,IAAI,KAAA;GACrC,QAAQ;IAEN,IAAI,CAAC,IAAI,aAAa,IAAI,UAAU,GAAG,CAAC,CAAC,IAAI,mBAAmB;IAChE;GACF;GACA,MAAM,UAAU,cAAc,KAAK,KAAK,MAAM;EAChD,EAAA,CAAG,CAAC,CAAC,OAAO,UAAmB;GAG7B,QAAQ,MAAM,yCAAyC,KAAK;GAC5D,IAAI,CAAC,IAAI,aAAa,IAAI,UAAU,GAAG,CAAC,CAAC,IAAI,cAAc;EAC7D,CAAC;CACH,CAAC;CAED,MAAM,KAAK,WAAW,OAAO,GAAG,WAAW,GAAG,WAAW;CACzD,MAAM,OAAQ,WAAW,QAAQ,CAAC,CAAiB;CAGnD,OAAO;EACL,MAAM;EACN,eAJoB,QAAQ,eAAe,GAAG,KAAK;EAKnD;EACA,aACE,IAAI,SAAe,YAAY,WAAW,YAAY,QAAQ,CAAC,CAAC;CACpE;AACF;;AA2BA,IAAa,4BAAmD,EAC9D,UAAU,OAAO,SAAS;CACxB,MAAM,EAAE,UAAU,GAAG,SAAS;CAC9B,OAAO,oBAAoB,OAAO;EAChC,gBAAgB,eAAe,QAAQ;EACvC,GAAG;CACL,CAAC;AACH,EACF"}
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@tanstack/ai-sandbox",
3
- "version": "0.5.6",
3
+ "version": "0.5.9",
4
4
  "description": "Provider-agnostic sandbox layer for TanStack AI — run harness adapters inside isolated sandboxes (defineSandbox, defineWorkspace, withSandbox) with a uniform SandboxHandle, workspace bootstrap, policy, and resumable lifecycle.",
5
5
  "author": "",
6
6
  "license": "MIT",
@@ -56,13 +56,13 @@
56
56
  },
57
57
  "dependencies": {
58
58
  "@modelcontextprotocol/sdk": "^1.29.0",
59
- "@tanstack/ai-skills": "^0.1.2"
59
+ "@tanstack/ai-skills": "^0.1.4"
60
60
  },
61
61
  "peerDependencies": {
62
62
  "@ngrok/ngrok": "^1.0.0",
63
63
  "vitest": "^4.1.10",
64
- "@tanstack/ai": "^0.53.0",
65
- "@tanstack/ai-persistence": "^0.5.6"
64
+ "@tanstack/ai": "^0.55.0",
65
+ "@tanstack/ai-persistence": "^0.6.0"
66
66
  },
67
67
  "peerDependenciesMeta": {
68
68
  "@ngrok/ngrok": {
@@ -79,8 +79,8 @@
79
79
  "@ngrok/ngrok": "^1.7.0",
80
80
  "@vitest/coverage-v8": "4.1.10",
81
81
  "vitest": "^4.1.10",
82
- "@tanstack/ai": "0.53.0",
83
- "@tanstack/ai-persistence": "0.5.6"
82
+ "@tanstack/ai-persistence": "0.6.0",
83
+ "@tanstack/ai": "0.55.0"
84
84
  },
85
85
  "scripts": {
86
86
  "build": "vite build",
@@ -48,9 +48,10 @@ agent CLI **inside** the sandbox and streams its events back.
48
48
  ## Setup — Claude Code in a Docker sandbox
49
49
 
50
50
  ```typescript
51
- import { chat } from '@tanstack/ai'
51
+ import { chat, toServerSentEventsResponse } from '@tanstack/ai'
52
52
  import { claudeCodeText } from '@tanstack/ai-claude-code'
53
53
  import {
54
+ createSecrets,
54
55
  defineSandbox,
55
56
  defineWorkspace,
56
57
  withSandbox,
@@ -65,17 +66,25 @@ const sandbox = defineSandbox({
65
66
  packageManager: 'pnpm',
66
67
  setup: ['corepack enable', 'pnpm install'],
67
68
  scripts: { test: 'pnpm test' },
68
- secrets: { ANTHROPIC_API_KEY: process.env.ANTHROPIC_API_KEY ?? '' },
69
+ secrets: createSecrets({
70
+ ANTHROPIC_API_KEY: process.env.ANTHROPIC_API_KEY ?? '',
71
+ }),
69
72
  }),
70
73
  lifecycle: { reuse: 'thread', snapshot: 'after-setup', keepAlive: '30m' },
71
74
  })
72
75
 
73
- const stream = chat({
74
- threadId,
75
- adapter: claudeCodeText('sonnet'),
76
- messages,
77
- middleware: [withSandbox(sandbox)],
78
- })
76
+ export async function POST(request: Request) {
77
+ const { threadId, messages } = await request.json()
78
+
79
+ const stream = chat({
80
+ threadId,
81
+ adapter: claudeCodeText('sonnet'),
82
+ messages,
83
+ middleware: [withSandbox(sandbox)],
84
+ })
85
+
86
+ return toServerSentEventsResponse(stream)
87
+ }
79
88
  ```
80
89
 
81
90
  ## Type-safe secrets
@@ -165,6 +174,8 @@ serial and parallel groups over a **persistent shell** whose cwd/env carry over
165
174
  between serial steps:
166
175
 
167
176
  ```typescript
177
+ import { githubRepo, defineWorkspace } from '@tanstack/ai-sandbox'
178
+
168
179
  defineWorkspace({
169
180
  source: githubRepo({ repo: 'owner/app' }),
170
181
  setup: ({ serial, parallel }) => {
@@ -183,10 +194,17 @@ When the provider supports snapshots, bootstrap takes one automatically after
183
194
  Override or add a TTL:
184
195
 
185
196
  ```typescript
186
- lifecycle: {
187
- snapshot: 'after-setup', // default when provider.capabilities().snapshots
188
- snapshotMaxAge: '24h', // re-create when the snapshot is older than this
189
- }
197
+ import { defineSandbox } from '@tanstack/ai-sandbox'
198
+ import { dockerSandbox } from '@tanstack/ai-sandbox-docker'
199
+
200
+ const sandbox = defineSandbox({
201
+ id: 'repo-agent',
202
+ provider: dockerSandbox({ image: 'node:22' }),
203
+ lifecycle: {
204
+ snapshot: 'after-setup', // default when provider.capabilities().snapshots
205
+ snapshotMaxAge: '24h', // re-create when the snapshot is older than this
206
+ },
207
+ })
190
208
  ```
191
209
 
192
210
  Providers without snapshot support skip the step silently.
@@ -199,8 +217,15 @@ middleware in this order, with the same persistence value in both places:
199
217
 
200
218
  ```typescript
201
219
  import { withPersistence } from '@tanstack/ai-persistence'
202
- import { memorySandboxSnapshots, withSandbox } from '@tanstack/ai-sandbox'
220
+ import {
221
+ InMemorySandboxInstanceStore,
222
+ memorySandboxSnapshots,
223
+ withSandbox,
224
+ } from '@tanstack/ai-sandbox'
225
+ // Your `defineSandbox(...)` result.
226
+ import { sandbox } from './sandbox'
203
227
 
228
+ const instances = new InMemorySandboxInstanceStore()
204
229
  const snapshots = await memorySandboxSnapshots({ sandbox, instances })
205
230
 
206
231
  const middleware = [
@@ -321,20 +346,30 @@ distributed lock: either `withLocks` from `@tanstack/ai/locks` (ordered
321
346
  **before** `withSandbox`) or the `locks` option.
322
347
 
323
348
  ```typescript
324
- import { chat } from '@tanstack/ai'
349
+ import { chat, toServerSentEventsResponse } from '@tanstack/ai'
325
350
  import { InMemoryLockStore, withLocks } from '@tanstack/ai/locks'
351
+ import { claudeCodeText } from '@tanstack/ai-claude-code'
326
352
  import { withSandbox } from '@tanstack/ai-sandbox'
353
+ // Your `defineSandbox(...)` result.
354
+ import { sandbox } from './sandbox'
327
355
  // Production: your BYO store — docs/sandbox/durability.md
328
356
  import { instanceStore } from './sandbox-instance-store'
329
357
 
330
- chat({
331
- adapter,
332
- messages,
333
- middleware: [
334
- withLocks(new InMemoryLockStore()), // multi-replica: distributed lock
335
- withSandbox(sandbox, { instances: instanceStore }),
336
- ],
337
- })
358
+ export async function POST(request: Request) {
359
+ const { threadId, messages } = await request.json()
360
+
361
+ const stream = chat({
362
+ threadId,
363
+ adapter: claudeCodeText('sonnet'),
364
+ messages,
365
+ middleware: [
366
+ withLocks(new InMemoryLockStore()), // multi-replica: distributed lock
367
+ withSandbox(sandbox, { instances: instanceStore }),
368
+ ],
369
+ })
370
+
371
+ return toServerSentEventsResponse(stream)
372
+ }
338
373
  ```
339
374
 
340
375
  The store option takes precedence over an ambient `SandboxInstanceStoreCapability`
@@ -360,8 +395,11 @@ middleware via the `sandbox` group (run-scoped):
360
395
  import { defineSandbox, withSandbox } from '@tanstack/ai-sandbox'
361
396
  // `defineChatMiddleware` is core's, not this package's — `@tanstack/ai-sandbox`
362
397
  // consumes it too (see its own `src/middleware.ts`).
363
- import { defineChatMiddleware } from '@tanstack/ai'
398
+ import { chat, defineChatMiddleware } from '@tanstack/ai'
399
+ import { claudeCodeText } from '@tanstack/ai-claude-code'
364
400
  import { dockerSandbox } from '@tanstack/ai-sandbox-docker'
401
+ import { db } from './db'
402
+ import { metrics } from './metrics'
365
403
 
366
404
  // Sandbox-scoped hooks (all optional):
367
405
  const sandbox = defineSandbox({
@@ -392,6 +430,13 @@ const auditMiddleware = defineChatMiddleware({
392
430
 
393
431
  // No extra middleware needed — sandbox.file CUSTOM events are emitted
394
432
  // automatically. Read them from the stream:
433
+ const stream = chat({
434
+ threadId: 'thread-1',
435
+ adapter: claudeCodeText('sonnet'),
436
+ messages: [{ role: 'user', content: 'Add a README.' }],
437
+ middleware: [auditMiddleware, withSandbox(sandbox)],
438
+ })
439
+
395
440
  for await (const chunk of stream) {
396
441
  if (chunk.type === 'CUSTOM' && chunk.name === 'sandbox.file') {
397
442
  const value = chunk.value
@@ -412,7 +457,10 @@ outside a `chat()` run:
412
457
 
413
458
  ```typescript
414
459
  import { watchWorkspace } from '@tanstack/ai-sandbox'
460
+ // Your `defineSandbox(...)` result.
461
+ import { sandbox } from './sandbox'
415
462
 
463
+ const handle = await sandbox.ensure({ threadId: 'thread-1', runId: 'run-1' })
416
464
  const watcher = await watchWorkspace(handle, {
417
465
  onEvent: (e) => console.log(e.type, e.path),
418
466
  ignore: ['.git', 'node_modules'], // default
@@ -424,8 +472,24 @@ Enable the `sandbox` debug category to log watcher start/stop, event dispatch,
424
472
  and lifecycle transitions:
425
473
 
426
474
  ```typescript
427
- chat({ threadId, adapter, messages, debug: { sandbox: true } })
428
- // or debug: true to enable all categories
475
+ import { chat, toServerSentEventsResponse } from '@tanstack/ai'
476
+ import { claudeCodeText } from '@tanstack/ai-claude-code'
477
+ import { withSandbox } from '@tanstack/ai-sandbox'
478
+ import { sandbox } from './sandbox'
479
+
480
+ export async function POST(request: Request) {
481
+ const { threadId, messages } = await request.json()
482
+
483
+ const stream = chat({
484
+ threadId,
485
+ adapter: claudeCodeText('sonnet'),
486
+ messages,
487
+ middleware: [withSandbox(sandbox)],
488
+ debug: { sandbox: true }, // or debug: true to enable all categories
489
+ })
490
+
491
+ return toServerSentEventsResponse(stream)
492
+ }
429
493
  ```
430
494
 
431
495
  ## Edge / serverless execution
@@ -551,12 +615,17 @@ file from byte 0 at any point, including after the original host has died.
551
615
 
552
616
  ```typescript
553
617
  import { spawnNdjson } from '@tanstack/ai-sandbox'
554
-
555
- for await (const event of spawnNdjson(sandbox, agentCommand, {
556
- cwd,
557
- journal: { runId }, // durability is opt-in: pass `journal` to route through it
558
- })) {
559
- // parsed NDJSON objects, translated by the harness adapter as usual
618
+ import type { SandboxHandle } from '@tanstack/ai-sandbox'
619
+
620
+ export async function runAgent(handle: SandboxHandle, runId: string) {
621
+ const agentCommand = 'claude -p --output-format stream-json'
622
+ for await (const event of spawnNdjson(handle, agentCommand, {
623
+ cwd: '/workspace',
624
+ journal: { runId }, // durability is opt-in: pass `journal` to route through it
625
+ })) {
626
+ // parsed NDJSON objects, translated by the harness adapter as usual
627
+ console.log(event)
628
+ }
560
629
  }
561
630
  ```
562
631
 
@@ -1006,6 +1075,17 @@ import {
1006
1075
  } from '@tanstack/ai-sandbox'
1007
1076
  import type { RunRecord } from '@tanstack/ai'
1008
1077
  import type { ReapResult, RunExitProbe } from '@tanstack/ai-sandbox'
1078
+ // Your distributed LockStore, the same one `withSandbox` gets.
1079
+ import { locks } from './locks'
1080
+ // Your persistence — the SAME RunStore the chat routes use.
1081
+ import { runs } from './persistence'
1082
+ // Your `defineSandbox(...)` result and the `SandboxInstanceStore` you passed to
1083
+ // `withSandbox(sandbox, { instances })`.
1084
+ import { instances, sandbox } from './sandbox'
1085
+ // The per-run log factory, resolving the SAME log the producing route wrote.
1086
+ import { durabilityFor } from './durability'
1087
+ // The same `drive` the attach route passes to `sandboxRunDriver`.
1088
+ import { driveRun } from './drive-run'
1009
1089
 
1010
1090
  async function hasFinished(record: RunRecord): Promise<RunExitProbe> {
1011
1091
  if (record.sandboxKey === undefined) return { state: 'unknown' }
@@ -20,6 +20,7 @@
20
20
  */
21
21
  import { createServer } from 'node:http'
22
22
  import { randomBytes, timingSafeEqual } from 'node:crypto'
23
+ import { once } from 'node:events'
23
24
  import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
24
25
  import { StreamableHTTPServerTransport } from '@modelcontextprotocol/sdk/server/streamableHttp.js'
25
26
  import {
@@ -350,9 +351,7 @@ export async function startHostToolBridge(
350
351
  })
351
352
  })
352
353
 
353
- await new Promise<void>((resolve) =>
354
- httpServer.listen(0, bindAddress, resolve),
355
- )
354
+ await once(httpServer.listen(0, bindAddress), 'listening')
356
355
  const port = (httpServer.address() as AddressInfo).port
357
356
  const url = `http://${options.hostForSandbox}:${port}/mcp`
358
357