[KooYS]

Summary

split in compat breaks emoji and accented characters when the separator is ''.
It just calls String.prototype.split, so the string is split by UTF-16 code units.
lodash handles this case differently: if the string has multi-byte characters, it splits by characters instead.

split('a😀b', '');
// es-toolkit: ['a', '\ud83d', '\ude00', 'b']
// lodash:     ['a', '😀', 'b']

split('Ame\u0301lie', '');
// es-toolkit: ['A', 'm', 'e', '\u0301', 'l', 'i', 'e']
// lodash:     ['A', 'm', 'é', 'l', 'i', 'e']

input	lodash	compat (before)	compat (after)
split('a😀b', '')	['a', '😀', 'b']	['a', '\ud83d', '\ude00', 'b']	['a', '😀', 'b']
split('Ame\u0301lie', '')	6 items, 'é' kept	7 items, '\u0301' alone	6 items, 'é' kept
split('👨‍👩‍👧', '')	['👨‍👩‍👧']	8 code units	['👨‍👩‍👧']
split('🇰🇷', '')	['🇰🇷']	4 code units	['🇰🇷']
split('a😀b', '', 2)	['a', '😀']	['a', '\ud83d']	['a', '😀']

This call type-checks with @types/lodash in strict mode.
Grist uses it to highlight autocomplete matches, and the comment there says it expects diacritics to stay combined:
https://github.com/gristlabs/grist-core/blob/897d07997d0d101b7fe5ecfaca61bb5f4e3a6ce0/app/client/lib/ACIndex.ts#L273

With the current compat, highlighting the first 3 characters of "Ame\u0301lie" gives 'Ame' and '\u0301lie', so the accent ends up in the wrong part.

Changes

I reused the reUnicode pattern that is already in unicodeSize.ts (it is the same one lodash uses), and added _internal/unicodeToArray.ts that uses it, like lodash's _unicodeToArray.
split now uses it only when the separator is '' and the string has multi-byte characters. Other inputs work the same as before.

Tests

Added a test that fails before this change. The compat tests pass.
I also compared the results with lodash 4.18.1 for 308 input combinations, and they all match.

Benchmark results

Run with vitest bench on Node 22.14 (ops/sec).
The first row is the existing split.bench.ts. For the other two rows I used a temporary bench file, which is not included in this PR.

case	main	this branch	lodash
'-' separator	2,626,812	2,869,407	6,496,327
ascii, ''	2,664,083	2,845,279	5,870,041
unicode, ''	1,225,899	365,102	423,759

The ascii cases are about the same. The unicode case is slower than main, but main was returning broken strings there, and the speed is now close to lodash.

Bundle size of compat/split goes from 388 B to 1,028 B because of the reUnicode pattern. compat/size already uses the same pattern, and lodash-es split is 4,219 B.
I also tried writing it as a loop instead of a regex. It matched lodash on 2,000,000 random strings, but the bundle was bigger (1,118 B), so I kept the regex.
The existing bundle-size snapshots don't change.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
es-toolkit	Ready	Preview	Oct 30, 2026 4:00am UTC