Task-augmented tool call errors are wrapped incorrectly, masking actual error messages

[punkpeye]

Summary

When a task-augmented tools/call request fails (e.g., due to input validation), the error is caught and converted to a tool error result ({content: [...], isError: true}), which then fails CreateTaskResultSchema validation. This results in a confusing error message that masks the actual underlying error.

Steps to Reproduce

1. Register a task-based tool with input validation:

server.experimental.tasks.registerToolTask(
  'batch_process',
  {
    inputSchema: {
      itemCount: z.number().min(1).max(10),
      processingTimeMs: z.number().min(500).max(5000).optional(),
    },
    // ... other config
  },
  {
    createTask: async (args, extra) => {
      // ... handler
    },
  },
);

2. Call the tool with task augmentation and invalid arguments:

{
  "method": "tools/call",
  "params": {
    "name": "batch_process",
    "arguments": { "itemCount": 5, "processingTimeMs": 100 },
    "task": { "ttl": 60000 }
  }
}

Expected Behavior

The client should receive a clear error message indicating the actual problem:

{
  "error": {
    "code": -32602,
    "message": "Input validation error: Invalid arguments for tool batch_process: Too small: expected number to be >=500"
  }
}

Actual Behavior

The client receives a confusing error that hides the actual cause:

{
  "error": {
    "code": -32602,
    "message": "MCP error -32602: Invalid task creation result: [{\"expected\":\"object\",\"code\":\"invalid_type\",\"path\":[\"task\"],\"message\":\"Invalid input: expected object, received undefined\"}]"
  }
}

Root Cause

The issue is in the error handling flow in server/mcp.js:

// In setToolRequestHandlers() - CallToolRequestSchema handler
try {
  const args = await this.validateToolInput(tool, request.params.arguments, request.params.name);
  const result = await this.executeToolHandler(tool, args, extra);
  // ...
} catch (error) {
  // ALL errors get wrapped as tool errors, including validation errors
  return this.createToolError(error instanceof Error ? error.message : String(error));
}

Then in server/index.js, the wrapped handler validates the result:

if (params.task) {
  const taskValidationResult = safeParse(CreateTaskResultSchema, result);
  if (!taskValidationResult.success) {
    // This fails because createToolError returns {content: [...], isError: true}
    // which doesn't have a 'task' property
    throw new McpError(ErrorCode.InvalidParams, `Invalid task creation result: ${errorMessage}`);
  }
}

Suggested Fix

For task-augmented requests, errors thrown before task creation (like input validation errors) should be re-thrown directly instead of being wrapped as tool errors:

// In setToolRequestHandlers() - CallToolRequestSchema handler
try {
  // ...
} catch (error) {
  if (error instanceof McpError) {
    if (error.code === ErrorCode.UrlElicitationRequired) {
      throw error;
    }
    // For task-augmented requests, also re-throw validation errors
    // instead of wrapping them as tool errors
    if (isTaskRequest && error.code === ErrorCode.InvalidParams) {
      throw error;
    }
  }
  return this.createToolError(error instanceof Error ? error.message : String(error));
}

Or alternatively, modify server/index.js to not validate non-task results against CreateTaskResultSchema:

if (params.task) {
  // Only validate if the result looks like a CreateTaskResult (has task property)
  // If it's a tool error, return it as-is or convert to a proper error response
  if ('isError' in result && result.isError) {
    throw new McpError(ErrorCode.InvalidParams, result.content[0]?.text ?? 'Tool execution failed');
  }
  // ... existing validation
}

Environment

• SDK Version: 1.25.2
• Node.js: v24.1.0

---

This bug affects the developer experience significantly as the actual error (input validation, missing taskStore, handler exceptions, etc.) is completely hidden behind the generic "Invalid task creation result" message.

[punkpeye]

you can replicate this in an inspector

https://glama.ai/mcp/inspector?servers=%5B%7B%22id%22%3A%22test%22%2C%22name%22%3A%22test%22%2C%22requestTimeout%22%3A10000%2C%22url%22%3A%22https%3A%2F%2Fmcp-test.glama.ai%2Fmcp%22%7D%5D

select batch_process and then enter invalid inputs, e.g. 1 and 1