Read a notebook cell's source as text, not as characters

[Anish Mehta (anishmehta24)]

Motivation and Context

nbformat's multiline_string is a string or a list of lines, and both are valid. nbformat.validate accepts either, and nbformat.reads keeps whichever spelling the file used. The converter walked the raw value to find the title:

for line in source_lines:
    if line.startswith("# "):

With a list that iterates lines, which is what was intended. With a string it iterates characters — and no single character starts with "# " — so a notebook written that way comes out with no title at all:

notebook = {
    "nbformat": 4, "nbformat_minor": 5, "metadata": {},
    "cells": [{"cell_type": "markdown", "source": "# My Notebook\nintro text\n", "metadata": {}}],
}
IpynbConverter()._convert(notebook).title    # None

The same notebook written with "source": ["# My Notebook\n", "intro text\n"] gives "My Notebook".

The markdown body is unaffected either way, which is why this is easy to miss: "".join(...) over a string reassembles the string, so only the title is lost.

nbformat.writes normalises to lists, so notebooks saved by Jupyter itself aren't affected. Notebooks assembled as JSON and written directly — generated notebooks, tooling that builds cells programmatically and calls json.dump rather than nbformat.writes — are.

Description

Take the cell's text once, then split it for the heading scan:

source = cell.get("source", [])
source_text = source if isinstance(source, str) else "".join(source)

The body uses source_text directly, so its output is byte-identical to before for list sources.

The title scan now runs over source_text.splitlines(). Besides fixing the string case, that also picks up a heading in a list entry holding several lines (["intro\n# My Notebook\n"]), which the per-entry check couldn't see — same root cause, so I fixed it in the same place rather than leaving a second shape broken.

No behaviour changes for notebooks that already worked; the existing #hashtag title regression test still passes.

Tests

Three tests in packages/markitdown/tests/test_module_misc.py, next to the existing ipynb ones:

• test_ipynb_string_source_is_read_as_text_not_characters — a string source now yields the title
• test_ipynb_string_and_list_sources_agree — the two spellings of the same notebook produce identical title and markdown
• test_ipynb_heading_below_the_first_line_of_a_source_entry — a heading after the first line of a list entry

All three fail on main. test_module_misc.py is 84 passed, 3 skipped, 1 failed — the failure is test_speech_transcription, which fails identically on a clean checkout here (it needs audio dependencies I don't have installed).