Garry's Opinionated OpenClaw/Hermes Agent Brain
TypeScript
/**
* Latin letters whose diacritic is drawn INSIDE the glyph — a stroke, bar, or
* ligature rather than a floating accent.
*
* Unicode assigns these no decomposition mapping, so the usual fold (NFD/NFKD
* decompose, then strip combining marks U+0300–U+036F) turns "é" into "e" but
* leaves "đ" exactly as it was. What happens next depends on the grammar
* downstream, and both outcomes are wrong in the same way:
*
* - a grammar that keeps only `[a-z0-9]` DELETES the letter
* ("Đăng Example" → "ang-example");
* - a grammar that keeps all letters KEEPS it unfolded
* ("Đức Example" → "đuc-example").
*
* Neither matches the ASCII form the rest of the pipeline expects, and the
* mismatch is silent: a slug nothing resolves to, or a basename lookup that
* misses.
*
* Leaf module by design — text primitives only, no engine or graph imports —
* so every grammar that folds accents can share one table.
*/
/** Keys are lowercase: callers fold after their own lowercasing pass. */
export const NON_DECOMPOSING_LATIN: Readonly<Record<string, string>> = {
đ: 'd', // U+0111 Vietnamese, Croatian, Sami
ð: 'd', // U+00F0 Icelandic eth
ø: 'o', // U+00F8 Danish, Norwegian, Faroese
ł: 'l', // U+0142 Polish
ħ: 'h', // U+0127 Maltese
ŧ: 't', // U+0167 Northern Sami
ı: 'i', // U+0131 Turkish dotless i
ß: 'ss', // U+00DF German sharp s
æ: 'ae', // U+00E6 Danish, Norwegian, Icelandic
œ: 'oe', // U+0153 French
þ: 'th', // U+00FE Icelandic thorn
};
const NON_DECOMPOSING_RE = new RegExp(`[${Object.keys(NON_DECOMPOSING_LATIN).join('')}]`, 'g');
/**
* Fold stroke/bar/ligature letters to their base ASCII letters.
*
* Call AFTER lowercasing (the table is lowercase-keyed) and AFTER the
* combining-mark strip, so composed forms reduce in a single pass:
* "ǿ" → NFD → "ø" + U+0301 → mark stripped → "ø" → "o".
*/
export function foldNonDecomposingLatin(s: string): string {
return s.replace(NON_DECOMPOSING_RE, ch => NON_DECOMPOSING_LATIN[ch]);
}