cas-parser-node-mcp 1.8.0 → 1.10.0

package/src/code-tool-worker.ts

@@ -0,0 +1,291 @@
+// File generated from our OpenAPI spec by Stainless. See CONTRIBUTING.md for details.
+
+import path from 'node:path';
+import util from 'node:util';
+import Fuse from 'fuse.js';
+import ts from 'typescript';
+import { WorkerOutput } from './code-tool-types';
+import { CasParser, ClientOptions } from 'cas-parser-node';
+
+function getRunFunctionSource(code: string): {
+  type: 'declaration' | 'expression';
+  client: string | undefined;
+  code: string;
+} | null {
+  const sourceFile = ts.createSourceFile('code.ts', code, ts.ScriptTarget.Latest, true);
+  const printer = ts.createPrinter();
+
+  for (const statement of sourceFile.statements) {
+    // Check for top-level function declarations
+    if (ts.isFunctionDeclaration(statement)) {
+      if (statement.name?.text === 'run') {
+        return {
+          type: 'declaration',
+          client: statement.parameters[0]?.name.getText(),
+          code: printer.printNode(ts.EmitHint.Unspecified, statement.body!, sourceFile),
+        };
+      }
+    }
+
+    // Check for variable declarations: const run = () => {} or const run = function() {}
+    if (ts.isVariableStatement(statement)) {
+      for (const declaration of statement.declarationList.declarations) {
+        if (
+          ts.isIdentifier(declaration.name) &&
+          declaration.name.text === 'run' &&
+          // Check if it's initialized with a function
+          declaration.initializer &&
+          (ts.isFunctionExpression(declaration.initializer) || ts.isArrowFunction(declaration.initializer))
+        ) {
+          return {
+            type: 'expression',
+            client: declaration.initializer.parameters[0]?.name.getText(),
+            code: printer.printNode(ts.EmitHint.Unspecified, declaration.initializer, sourceFile),
+          };
+        }
+      }
+    }
+  }
+
+  return null;
+}
+
+function getTSDiagnostics(code: string): string[] {
+  const functionSource = getRunFunctionSource(code)!;
+  const codeWithImport = [
+    'import { CasParser } from "cas-parser-node";',
+    functionSource.type === 'declaration' ?
+      `async function run(${functionSource.client}: CasParser)`
+    : `const run: (${functionSource.client}: CasParser) => Promise<unknown> =`,
+    functionSource.code,
+  ].join('\n');
+  const sourcePath = path.resolve('code.ts');
+  const ast = ts.createSourceFile(sourcePath, codeWithImport, ts.ScriptTarget.Latest, true);
+  const options = ts.getDefaultCompilerOptions();
+  options.target = ts.ScriptTarget.Latest;
+  options.module = ts.ModuleKind.NodeNext;
+  options.moduleResolution = ts.ModuleResolutionKind.NodeNext;
+  const host = ts.createCompilerHost(options, true);
+  const newHost: typeof host = {
+    ...host,
+    getSourceFile: (...args) => {
+      if (path.resolve(args[0]) === sourcePath) {
+        return ast;
+      }
+      return host.getSourceFile(...args);
+    },
+    readFile: (...args) => {
+      if (path.resolve(args[0]) === sourcePath) {
+        return codeWithImport;
+      }
+      return host.readFile(...args);
+    },
+    fileExists: (...args) => {
+      if (path.resolve(args[0]) === sourcePath) {
+        return true;
+      }
+      return host.fileExists(...args);
+    },
+  };
+  const program = ts.createProgram({
+    options,
+    rootNames: [sourcePath],
+    host: newHost,
+  });
+  const diagnostics = ts.getPreEmitDiagnostics(program, ast);
+  return diagnostics.map((d) => {
+    const message = ts.flattenDiagnosticMessageText(d.messageText, '\n');
+    if (!d.file || !d.start) return `- ${message}`;
+    const { line: lineNumber } = ts.getLineAndCharacterOfPosition(d.file, d.start);
+    const line = codeWithImport.split('\n').at(lineNumber)?.trim();
+    return line ? `- ${message}\n    ${line}` : `- ${message}`;
+  });
+}
+
+const fuse = new Fuse(
+  [
+    'client.credits.check',
+    'client.logs.create',
+    'client.logs.getSummary',
+    'client.accessToken.create',
+    'client.verifyToken.verify',
+    'client.camsKfintech.parse',
+    'client.cdsl.parsePdf',
+    'client.cdsl.fetch.requestOtp',
+    'client.cdsl.fetch.verifyOtp',
+    'client.contractNote.parse',
+    'client.inbox.checkConnectionStatus',
+    'client.inbox.connectEmail',
+    'client.inbox.disconnectEmail',
+    'client.inbox.listCasFiles',
+    'client.kfintech.generateCas',
+    'client.nsdl.parse',
+    'client.smart.parseCasPdf',
+    'client.inboundEmail.create',
+    'client.inboundEmail.delete',
+    'client.inboundEmail.list',
+    'client.inboundEmail.retrieve',
+  ],
+  { threshold: 1, shouldSort: true },
+);
+
+function getMethodSuggestions(fullyQualifiedMethodName: string): string[] {
+  return fuse
+    .search(fullyQualifiedMethodName)
+    .map(({ item }) => item)
+    .slice(0, 5);
+}
+
+const proxyToObj = new WeakMap<any, any>();
+const objToProxy = new WeakMap<any, any>();
+
+type ClientProxyConfig = {
+  path: string[];
+  isBelievedBad?: boolean;
+};
+
+function makeSdkProxy<T extends object>(obj: T, { path, isBelievedBad = false }: ClientProxyConfig): T {
+  let proxy: T = objToProxy.get(obj);
+
+  if (!proxy) {
+    proxy = new Proxy(obj, {
+      get(target, prop, receiver) {
+        const propPath = [...path, String(prop)];
+        const value = Reflect.get(target, prop, receiver);
+
+        if (isBelievedBad || (!(prop in target) && value === undefined)) {
+          // If we're accessing a path that doesn't exist, it will probably eventually error.
+          // Let's proxy it and mark it bad so that we can control the error message.
+          // We proxy an empty class so that an invocation or construction attempt is possible.
+          return makeSdkProxy(class {}, { path: propPath, isBelievedBad: true });
+        }
+
+        if (value !== null && (typeof value === 'object' || typeof value === 'function')) {
+          return makeSdkProxy(value, { path: propPath, isBelievedBad });
+        }
+
+        return value;
+      },
+
+      apply(target, thisArg, args) {
+        if (isBelievedBad || typeof target !== 'function') {
+          const fullyQualifiedMethodName = path.join('.');
+          const suggestions = getMethodSuggestions(fullyQualifiedMethodName);
+          throw new Error(
+            `${fullyQualifiedMethodName} is not a function. Did you mean: ${suggestions.join(', ')}`,
+          );
+        }
+
+        return Reflect.apply(target, proxyToObj.get(thisArg) ?? thisArg, args);
+      },
+
+      construct(target, args, newTarget) {
+        if (isBelievedBad || typeof target !== 'function') {
+          const fullyQualifiedMethodName = path.join('.');
+          const suggestions = getMethodSuggestions(fullyQualifiedMethodName);
+          throw new Error(
+            `${fullyQualifiedMethodName} is not a constructor. Did you mean: ${suggestions.join(', ')}`,
+          );
+        }
+
+        return Reflect.construct(target, args, newTarget);
+      },
+    });
+
+    objToProxy.set(obj, proxy);
+    proxyToObj.set(proxy, obj);
+  }
+
+  return proxy;
+}
+
+function parseError(code: string, error: unknown): string | undefined {
+  if (!(error instanceof Error)) return;
+  const message = error.name ? `${error.name}: ${error.message}` : error.message;
+  try {
+    // Deno uses V8; the first "<anonymous>:LINE:COLUMN" is the top of stack.
+    const lineNumber = error.stack?.match(/<anonymous>:([0-9]+):[0-9]+/)?.[1];
+    // -1 for the zero-based indexing
+    const line =
+      lineNumber &&
+      code
+        .split('\n')
+        .at(parseInt(lineNumber, 10) - 1)
+        ?.trim();
+    return line ? `${message}\n  at line ${lineNumber}\n    ${line}` : message;
+  } catch {
+    return message;
+  }
+}
+
+const fetch = async (req: Request): Promise<Response> => {
+  const { opts, code } = (await req.json()) as { opts: ClientOptions; code: string };
+
+  const runFunctionSource = code ? getRunFunctionSource(code) : null;
+  if (!runFunctionSource) {
+    const message =
+      code ?
+        'The code is missing a top-level `run` function.'
+      : 'The code argument is missing. Provide one containing a top-level `run` function.';
+    return Response.json(
+      {
+        is_error: true,
+        result: `${message} Write code within this template:\n\n\`\`\`\nasync function run(client) {\n  // Fill this out\n}\n\`\`\``,
+        log_lines: [],
+        err_lines: [],
+      } satisfies WorkerOutput,
+      { status: 400, statusText: 'Code execution error' },
+    );
+  }
+
+  const diagnostics = getTSDiagnostics(code);
+  if (diagnostics.length > 0) {
+    return Response.json(
+      {
+        is_error: true,
+        result: `The code contains TypeScript diagnostics:\n${diagnostics.join('\n')}`,
+        log_lines: [],
+        err_lines: [],
+      } satisfies WorkerOutput,
+      { status: 400, statusText: 'Code execution error' },
+    );
+  }
+
+  const client = new CasParser({
+    ...opts,
+  });
+
+  const log_lines: string[] = [];
+  const err_lines: string[] = [];
+  const console = {
+    log: (...args: unknown[]) => {
+      log_lines.push(util.format(...args));
+    },
+    error: (...args: unknown[]) => {
+      err_lines.push(util.format(...args));
+    },
+  };
+  try {
+    let run_ = async (client: any) => {};
+    eval(`${code}\nrun_ = run;`);
+    const result = await run_(makeSdkProxy(client, { path: ['client'] }));
+    return Response.json({
+      is_error: false,
+      result,
+      log_lines,
+      err_lines,
+    } satisfies WorkerOutput);
+  } catch (e) {
+    return Response.json(
+      {
+        is_error: true,
+        result: parseError(code, e),
+        log_lines,
+        err_lines,
+      } satisfies WorkerOutput,
+      { status: 400, statusText: 'Code execution error' },
+    );
+  }
+};
+
+export default { fetch };

package/src/code-tool.ts

@@ -1,6 +1,12 @@
 // File generated from our OpenAPI spec by Stainless. See CONTRIBUTING.md for details.
 
+import fs from 'node:fs';
+import path from 'node:path';
+import url from 'node:url';
+import { newDenoHTTPWorker } from '@valtown/deno-http-worker';
+import { workerPath } from './code-tool-paths.cjs';
 import {
+  ContentBlock,
   McpRequestContext,
   McpTool,
   Metadata,
@@ -11,11 +17,14 @@ import {
 import { Tool } from '@modelcontextprotocol/sdk/types.js';
 import { readEnv, requireValue } from './util';
 import { WorkerInput, WorkerOutput } from './code-tool-types';
+import { getLogger } from './logger';
 import { SdkMethod } from './methods';
+import { McpCodeExecutionMode } from './options';
+import { ClientOptions } from 'cas-parser-node';
 
 const prompt = `Runs JavaScript code to interact with the Cas Parser API.
 
-You are a skilled programmer writing code to interface with the service.
+You are a skilled TypeScript programmer writing code to interface with the service.
 Define an async function named "run" that takes a single parameter of an initialized SDK client and it will be run.
 For example:
 
@@ -31,7 +40,9 @@ You will be returned anything that your function returns, plus the results of an
 Do not add try-catch blocks for single API calls. The tool will handle errors for you.
 Do not add comments unless necessary for generating better code.
 Code will run in a container, and cannot interact with the network outside of the given SDK client.
-Variables will not persist between calls, so make sure to return or log any data you might need later.`;
+Variables will not persist between calls, so make sure to return or log any data you might need later.
+Remember that you are writing TypeScript code, so you need to be careful with your types.
+Always type dynamic key-value stores explicitly as Record<string, YourValueType> instead of {}.`;
 
 /**
  * A tool that runs code against a copy of the SDK.
@@ -40,9 +51,19 @@ Variables will not persist between calls, so make sure to return or log any data
  * we expose a single tool that can be used to search for endpoints by name, resource, operation, or tag, and then
  * a generic endpoint that can be used to invoke any endpoint with the provided arguments.
  *
- * @param endpoints - The endpoints to include in the list.
+ * @param blockedMethods - The methods to block for code execution. Blocking is done by simple string
+ * matching, so it is not secure against obfuscation. For stronger security, block in the downstream API
+ * with limited API keys.
+ * @param codeExecutionMode - Whether to execute code in a local Deno environment or in a remote
+ * sandbox environment hosted by Stainless.
  */
-export function codeTool({ blockedMethods }: { blockedMethods: SdkMethod[] | undefined }): McpTool {
+export function codeTool({
+  blockedMethods,
+  codeExecutionMode,
+}: {
+  blockedMethods: SdkMethod[] | undefined;
+  codeExecutionMode: McpCodeExecutionMode;
+}): McpTool {
   const metadata: Metadata = { resource: 'all', operation: 'write', tags: [] };
   const tool: Tool = {
     name: 'execute',
@@ -62,6 +83,9 @@ export function codeTool({ blockedMethods }: { blockedMethods: SdkMethod[] | und
       required: ['code'],
     },
   };
+
+  const logger = getLogger();
+
   const handler = async ({
     reqContext,
     args,
@@ -70,9 +94,6 @@ export function codeTool({ blockedMethods }: { blockedMethods: SdkMethod[] | und
     args: any;
   }): Promise<ToolCallResult> => {
     const code = args.code as string;
-    const intent = args.intent as string | undefined;
-    const client = reqContext.client;
-
     // Do very basic blocking of code that includes forbidden method names.
     //
     // WARNING: This is not secure against obfuscation and other evasion methods. If
@@ -89,54 +110,269 @@ export function codeTool({ blockedMethods }: { blockedMethods: SdkMethod[] | und
       }
     }
 
-    const codeModeEndpoint =
-      readEnv('CODE_MODE_ENDPOINT_URL') ?? 'https://api.stainless.com/api/ai/code-tool';
-
-    // Setting a Stainless API key authenticates requests to the code tool endpoint.
-    const res = await fetch(codeModeEndpoint, {
-      method: 'POST',
-      headers: {
-        ...(reqContext.stainlessApiKey && { Authorization: reqContext.stainlessApiKey }),
-        'Content-Type': 'application/json',
-        client_envs: JSON.stringify({
-          CAS_PARSER_API_KEY: requireValue(
-            readEnv('CAS_PARSER_API_KEY') ?? client.apiKey,
-            'set CAS_PARSER_API_KEY environment variable or provide apiKey client option',
-          ),
-          CAS_PARSER_BASE_URL:
-            readEnv('CAS_PARSER_BASE_URL') ?? readEnv('CAS_PARSER_ENVIRONMENT') ?
-              undefined
-            : client.baseURL ?? undefined,
-        }),
-      },
-      body: JSON.stringify({
-        project_name: 'cas-parser',
-        code,
-        intent,
-        client_opts: { environment: (readEnv('CAS_PARSER_ENVIRONMENT') || undefined) as any },
-      } satisfies WorkerInput),
-    });
+    let result: ToolCallResult;
+    const startTime = Date.now();
 
-    if (!res.ok) {
-      throw new Error(
-        `${res.status}: ${
-          res.statusText
-        } error when trying to contact Code Tool server. Details: ${await res.text()}`,
-      );
+    if (codeExecutionMode === 'local') {
+      logger.debug('Executing code in local Deno environment');
+      result = await localDenoHandler({ reqContext, args });
+    } else {
+      logger.debug('Executing code in remote Stainless environment');
+      result = await remoteStainlessHandler({ reqContext, args });
     }
 
-    const { is_error, result, log_lines, err_lines } = (await res.json()) as WorkerOutput;
-    const hasLogs = log_lines.length > 0 || err_lines.length > 0;
-    const output = {
-      result,
-      ...(log_lines.length > 0 && { log_lines }),
-      ...(err_lines.length > 0 && { err_lines }),
-    };
-    if (is_error) {
-      return asErrorResult(typeof result === 'string' && !hasLogs ? result : JSON.stringify(output, null, 2));
-    }
-    return asTextContentResult(output);
+    logger.info(
+      {
+        codeExecutionMode,
+        durationMs: Date.now() - startTime,
+        isError: result.isError,
+        contentRows: result.content?.length ?? 0,
+      },
+      'Got code tool execution result',
+    );
+    return result;
   };
 
   return { metadata, tool, handler };
 }
+
+const remoteStainlessHandler = async ({
+  reqContext,
+  args,
+}: {
+  reqContext: McpRequestContext;
+  args: any;
+}): Promise<ToolCallResult> => {
+  const code = args.code as string;
+  const intent = args.intent as string | undefined;
+  const client = reqContext.client;
+
+  const codeModeEndpoint = readEnv('CODE_MODE_ENDPOINT_URL') ?? 'https://api.stainless.com/api/ai/code-tool';
+
+  // Setting a Stainless API key authenticates requests to the code tool endpoint.
+  const res = await fetch(codeModeEndpoint, {
+    method: 'POST',
+    headers: {
+      ...(reqContext.stainlessApiKey && { Authorization: reqContext.stainlessApiKey }),
+      'Content-Type': 'application/json',
+      client_envs: JSON.stringify({
+        CAS_PARSER_API_KEY: requireValue(
+          readEnv('CAS_PARSER_API_KEY') ?? client.apiKey,
+          'set CAS_PARSER_API_KEY environment variable or provide apiKey client option',
+        ),
+        CAS_PARSER_BASE_URL: readEnv('CAS_PARSER_BASE_URL') ?? client.baseURL ?? undefined,
+      }),
+    },
+    body: JSON.stringify({
+      project_name: 'cas-parser',
+      code,
+      intent,
+      client_opts: {},
+    } satisfies WorkerInput),
+  });
+
+  if (!res.ok) {
+    throw new Error(
+      `${res.status}: ${
+        res.statusText
+      } error when trying to contact Code Tool server. Details: ${await res.text()}`,
+    );
+  }
+
+  const { is_error, result, log_lines, err_lines } = (await res.json()) as WorkerOutput;
+  const hasLogs = log_lines.length > 0 || err_lines.length > 0;
+  const output = {
+    result,
+    ...(log_lines.length > 0 && { log_lines }),
+    ...(err_lines.length > 0 && { err_lines }),
+  };
+  if (is_error) {
+    return asErrorResult(typeof result === 'string' && !hasLogs ? result : JSON.stringify(output, null, 2));
+  }
+  return asTextContentResult(output);
+};
+
+const localDenoHandler = async ({
+  reqContext,
+  args,
+}: {
+  reqContext: McpRequestContext;
+  args: unknown;
+}): Promise<ToolCallResult> => {
+  const client = reqContext.client;
+  const baseURLHostname = new URL(client.baseURL).hostname;
+  const { code } = args as { code: string };
+
+  let denoPath: string;
+
+  const packageRoot = path.resolve(path.dirname(workerPath), '..');
+  const packageNodeModulesPath = path.resolve(packageRoot, 'node_modules');
+
+  // Check if deno is in PATH
+  const { execSync } = await import('node:child_process');
+  try {
+    execSync('command -v deno', { stdio: 'ignore' });
+    denoPath = 'deno';
+  } catch {
+    try {
+      // Use deno binary in node_modules if it's found
+      const denoNodeModulesPath = path.resolve(packageNodeModulesPath, 'deno', 'bin.cjs');
+      await fs.promises.access(denoNodeModulesPath, fs.constants.X_OK);
+      denoPath = denoNodeModulesPath;
+    } catch {
+      return asErrorResult(
+        'Deno is required for code execution but was not found. ' +
+          'Install it from https://deno.land or run: npm install deno',
+      );
+    }
+  }
+
+  const allowReadPaths = [
+    'code-tool-worker.mjs',
+    `${workerPath.replace(/([\/\\]node_modules)[\/\\].+$/, '$1')}/`,
+    packageRoot,
+  ];
+
+  // Follow symlinks in node_modules to allow read access to workspace-linked packages
+  try {
+    const sdkPkgName = 'cas-parser-node';
+    const sdkDir = path.resolve(packageNodeModulesPath, sdkPkgName);
+    const realSdkDir = fs.realpathSync(sdkDir);
+    if (realSdkDir !== sdkDir) {
+      allowReadPaths.push(realSdkDir);
+    }
+  } catch {
+    // Ignore if symlink resolution fails
+  }
+
+  const allowRead = allowReadPaths.join(',');
+
+  const worker = await newDenoHTTPWorker(url.pathToFileURL(workerPath), {
+    denoExecutable: denoPath,
+    runFlags: [
+      `--node-modules-dir=manual`,
+      `--allow-read=${allowRead}`,
+      `--allow-net=${baseURLHostname}`,
+      // Allow environment variables because instantiating the client will try to read from them,
+      // even though they are not set.
+      '--allow-env',
+    ],
+    printOutput: true,
+    spawnOptions: {
+      cwd: path.dirname(workerPath),
+    },
+  });
+
+  try {
+    const resp = await new Promise<Response>((resolve, reject) => {
+      worker.addEventListener('exit', (exitCode) => {
+        reject(new Error(`Worker exited with code ${exitCode}`));
+      });
+
+      const opts: ClientOptions = {
+        baseURL: client.baseURL,
+        apiKey: client.apiKey,
+        defaultHeaders: {
+          'X-Stainless-MCP': 'true',
+        },
+      };
+
+      const req = worker.request(
+        'http://localhost',
+        {
+          headers: {
+            'content-type': 'application/json',
+          },
+          method: 'POST',
+        },
+        (resp) => {
+          const body: Uint8Array[] = [];
+          resp.on('error', (err) => {
+            reject(err);
+          });
+          resp.on('data', (chunk) => {
+            body.push(chunk);
+          });
+          resp.on('end', () => {
+            resolve(
+              new Response(Buffer.concat(body).toString(), {
+                status: resp.statusCode ?? 200,
+                headers: resp.headers as any,
+              }),
+            );
+          });
+        },
+      );
+
+      const body = JSON.stringify({
+        opts,
+        code,
+      });
+
+      req.write(body, (err) => {
+        if (err != null) {
+          reject(err);
+        }
+      });
+
+      req.end();
+    });
+
+    if (resp.status === 200) {
+      const { result, log_lines, err_lines } = (await resp.json()) as WorkerOutput;
+      const returnOutput: ContentBlock | null =
+        result == null ? null : (
+          {
+            type: 'text',
+            text: typeof result === 'string' ? result : JSON.stringify(result),
+          }
+        );
+      const logOutput: ContentBlock | null =
+        log_lines.length === 0 ?
+          null
+        : {
+            type: 'text',
+            text: log_lines.join('\n'),
+          };
+      const errOutput: ContentBlock | null =
+        err_lines.length === 0 ?
+          null
+        : {
+            type: 'text',
+            text: 'Error output:\n' + err_lines.join('\n'),
+          };
+      return {
+        content: [returnOutput, logOutput, errOutput].filter((block) => block !== null),
+      };
+    } else {
+      const { result, log_lines, err_lines } = (await resp.json()) as WorkerOutput;
+      const messageOutput: ContentBlock | null =
+        result == null ? null : (
+          {
+            type: 'text',
+            text: typeof result === 'string' ? result : JSON.stringify(result),
+          }
+        );
+      const logOutput: ContentBlock | null =
+        log_lines.length === 0 ?
+          null
+        : {
+            type: 'text',
+            text: log_lines.join('\n'),
+          };
+      const errOutput: ContentBlock | null =
+        err_lines.length === 0 ?
+          null
+        : {
+            type: 'text',
+            text: 'Error output:\n' + err_lines.join('\n'),
+          };
+      return {
+        content: [messageOutput, logOutput, errOutput].filter((block) => block !== null),
+        isError: true,
+      };
+    }
+  } finally {
+    worker.terminate();
+  }
+};