Can you create a repo that can reproduce this?

I don't know what you mean by "opened ./d.efg twice" marked does not do anything different if you open a file.

Here is a repo: https://github.com/CNOCTAVE/marked_reader
Press Ctrl + Shift + i to open Electron console, then you can see errors.
Click the left top button, choose a Markdown file. Do it twice, then reproduce the errors.

In Markdown, ./-style relative path is common. When in Electron(a Node.js-based framework), Electron always parse ./ path as Node path, so ./ path should be rewritten to absolute path. This repo is a Markdown reader, user should open a local Markdown file. To make that bug clearer, you can write a picture path as relative path, if picture cannot correctly shows out, that's bug!

Like:
New a xxx.md, write:

d

where this xxx.md and ./d.png is at C:/abc/

You are adding the walkTokens function everytime they open a file. You should only be adding it once:

import { marked } from "marked";
function walkTokens(token) {
    if (token.type === 'image' || token.type === 'link') {
        token.href = markdownDir + ("/" + token.href).replace("//", "/").replace("\\", "/").replace("%5C", "/").replace("./", "/"); // 修正相对路径
    }
}
marked.use({ walkTokens });

or you can use an instance to not have the extensions be global

import { Marked } from "marked";
...
ipcMain.on('render-markdown', (event, filePath) => {
    ...
    function walkTokens(token) {
        if (token.type === 'image' || token.type === 'link') {
            token.href = markdownDir + ("/" + token.href).replace("//", "/").replace("\\", "/").replace("%5C", "/").replace("./", "/"); // 修正相对路径
        }
    }
    const marked = new Marked();
    marked.use({ walkTokens });
    const htmlContent = marked.parse(markdownContent);
    event.reply('markdown-rendered', htmlContent);
});

Can you PR to https://github.com/CNOCTAVE/marked_reader to show your modified code? It doesn't work for me.

use marked.parse(...) not marked(...) still works sometimes but is not the recommended way to run markdown through marked.

Now the bug is gone after using marked.parse(...).