Incident: A splitter that silently dropped the rest

Module 3 · Incident file · About 35 minutes

What happened

On 24 March 2026 a GitHub user, goingforstudying-ctrl, opened issue #36186 on LangChain's repository: "ExperimentalMarkdownSyntaxTextSplitter silently discards content from unclosed code blocks". The splitter belongs to langchain-text-splitters and cuts Markdown at headings while keeping code blocks whole. The report says that when a code block's closing fence never comes, the method _resolve_code_chunk() "returns an empty string "", silently discarding all accumulated content from the unclosed block". No error is raised (LangChain issue #36186, 24 March 2026).

A second user filed the same bug on 8 June 2026 as issue #37966, with a sharper description: the splitter "silently discards the code block and every following section and its body". Its example returns one chunk from a document with two sections. The repository's triage bot closed that report the same day as "not planned", because it appeared to have been submitted programmatically and had no issue type set (LangChain issue #37966, 8 June 2026).

When we checked on 10 October 2026, #36186 was still open, labelled bug, with no maintainer comment. A fix proposed with it, pull request #36185, was closed without being merged. Several contributors offered the same one-line change. Neither report describes harm to a deployed system. The cost of a bug like this is what nobody sees: sections that never reach the index, and questions about them that quietly stop being answered. We reproduced it with langchain-text-splitters 1.1.3, the version this course pins.