変更履歴

All notable changes to this project are written here. The format follows Keep a Changelog, and the project uses Semantic Versioning.

[Unreleased]

[2.11.0] - 2026-11-04

A minor version: nothing that 2.10.0 gave is taken away. In safe mode, ordinary words that 2.10.0 left out are back; unfiltered mode is as it was.

Changed

• Safe mode no longer leaves out ordinary words (MODULES-392; John, 2026-11-04: "we can't filter out regular good words"). 2.10.0 copied Kotoba's blunt gloss rule, which also took out a sense whose gloss only mentions sex, an erection of a building or an adulteration. Now it is Kotoba's narrowed rule (adultGloss and ADULT_FIRST_GLOSS in scripts/children-bar.mjs in Kotoba, kept in step here in scripts/words-parse.mjs): a sense is left out in safe mode only for JMdict's mark (vulgar, X-rated, sensitive, derogatory) or for a gloss that names a sex act, a sex organ or the sex trade (sexual intercourse, to have sex, prostitute, penis, porn, rape, adultery...), or whose first gloss is the mistress, the concubine or the nipple itself. Back in safe mode: 性別 "gender, (distinction of) sex", 異性 "the opposite sex", 同性 "same sex", 共学 "coeducation", 性教育 "sex education", 建立 "erection (of a building)", 組み立て, 造営, 混ぜ物 "adulteration", 混入, 不純 "adulterated", 挿入 "insertion", 菜の花 "rape blossoms", 愛人 "lover; mistress", 女将, 正室. Still out: セックス "sexual intercourse" (the sense; its "gender" sense stays), 売春, ポルノ, レイプ, 精液, 射精, 勃起, 乳首, 妾, 不倫, 性交, 娼婦 and the rest. Unfiltered mode is as it was. The data is made again from JMdict of 2026-11-04 (it was 2026-11-03) and the same Tatoeba files, so it holds that day's changes too:
  • Senses left out by the gloss rule: 1,413 in 1,387 entries to 1,276 in 1,256 (137 senses back); 109 entries that had nothing else to say are found again.
  • Rows of safe mode 263,857 to 263,987 (of 265,958; 1,971 rows left out whole, from 2,097); the common words 25,999 to 26,014.
  • The package is smaller, not bigger: 86.77 to 86.71 MB unpacked (the build as npm pack sees it) and 27.71 to 27.69 MB to download.
• Tests hold the narrowed rule to small entries (src/words-build.test.js: more than twenty ordinary glosses that stay, and the acts and organs that go) and to the real data (src/words.test.ts).

[2.10.0] - 2026-11-04

A minor version: nothing that 2.9.0 gave is taken away except, in safe mode, what children should not have been handed. Unfiltered mode is as it was.

Changed

• Safe mode also leaves out a sense whose gloss names sex or the sex trade (MODULES-390). It left out only the senses JMdict marks vulgar, X-rated, sensitive or derogatory, and JMdict does not always mark such a sense (it gives 朝雲暮雨 "sexual liaison" no mark). The rule is Kotoba's ADULT_GLOSS (scripts/children-bar.mjs in Kotoba, kept in step here in scripts/words-parse.mjs), applied to every gloss of the sense, so it holds for lookups, the common table, the readers, the whole entries and the examples alike, and it keeps what only looks like it (菜種 "rapeseed", 抹香鯨 "sperm whale", 社交 "social intercourse", 主婦 "mistress (of the house)"). The data is made again from the same JMdict (2026-11-03) and the same Tatoeba files, so nothing else in it moves:
  • 1,413 senses of 1,387 entries are left out in safe mode; 1,034 of the entries have nothing else to say (セックス, 売春, ポルノ, レイプ, 精子, 性交, 射精, 乳首, 娼婦 ...) and are not found in safe mode, the others lose the sense (倅 is "son; punk, brat", no longer also "penis").
  • Rows of safe mode 265,107 to 263,857 (1,250 fewer; 847 rows were left out whole, now 2,097); the common words 26,039 to 25,999 (40 fewer); the pieces of safe mode 318 + 71 to 316 + 70; the pieces that hold what safe mode leaves out 9 + 1 to 13 + 2 (they are fetched only unfiltered).
  • Examples of safe mode: 32,299 to 32,196 (103 fewer), for 28,730 words (from 28,813).
  • Like the rule it is copied from, it also takes out a few ordinary senses whose gloss uses one of the words: 性別 "gender, (distinction of) sex", 異性 "the opposite sex", 共学 "coeducation, mixed-sex education", 建立 "erection (of a building)", 混ぜ物 "adulteration". Unfiltered mode has them all.
  • The package is 70 kB smaller unpacked (88.00 to 87.93 MB with the build as npm pack sees it) and 45 kB smaller to download: the safe pieces shrank by about what the extra pieces grew.
• A test holds the rule to small entries (src/words-build.test.js), and others hold the real data: the words, the whole entries and the reader (src/words.test.ts, src/words-entry.test.ts, src/reader.test.ts) and the bytes of the files a safe lookup fetches (src/words-files.test.js). The guides, the README and NOTICE.md say what safe mode leaves out, in English and in Japanese.

[2.9.0] - 2026-11-04

A minor version: nothing that 2.8.0 gave is taken away. The table has 2,882 more pairs, so an answer of lookAlikes may now hold a kanji it did not, and each answer has a shape.

Added

• Look-alike pairs for the kanji Lars Yencken's measures never scored (MODULES-57). 2,882 pairs, scored by shape from the stroke paths of KanjiVG by the new scripts/build-shapes.mjs (pnpm shapes): a stroke edit distance over the strokes as written. The table has 10,093 pairs, from 7,211. 187 of the 196 kanji added to the jōyō list in 2010 now have a look-alike (9 have none: 臼, 瓦, 乞, 嫉, 遡, 遜, 溺, 藤, 鬱), 794 of the 858 kanji for names that his data did not score do, and pairs his measures ranked too low are in: 己 and 巳, 崎 and 埼.
• lookAlikes gives each answer a shape score beside strokes and parts, and its score is the highest of the three. A pair of Yencken's keeps its scores and has shape 0; a shape pair has 0 for the other two. PackedLookAlikes pairs may carry a fourth number, the shape score.
• The guide "How the look-alike pairs are measured" says how the shape score is made and what it misses, in English and Japanese.

Changed

• look-alike is under CC BY-SA 3.0 as a whole (it now holds scores made from KanjiVG beside Yencken's CC BY 3.0 ones); NOTICE.md, data-sources/sources.json and the data file's header say so. The package is 66.5 kB larger unpacked (86.19 to 86.26 MB); the look-alike entry point goes from 35.4 to 50.7 kB gzipped.

[2.8.0] - 2026-11-04

A minor version: nothing that 2.7.0 gave is taken away. The package itself is unchanged; the demo gains the family's embed panel.

Added

• The demo's "Put it in your page": the family's <kazoku-embed-panel> (kazoku-ui ^0.15.1; MODULES-359). Choose what to look up (a reading or a word, an English meaning or a kanji), the text, how many rows (3, 5 or 10) and the audience (safe or unfiltered); the panel writes the code for each way a page is made (plain HTML from jsDelivr, an ES module, React, Vue, Angular, Svelte and an iframe of the embed page) at three widths (280, 360 and 480), shows the embed page itself beside the code, and has Copy code, Open the embed page, Read the docs and the credit line (JMdict, KANJIDIC2). The tab, the width and the choices stay in the address (embed, size, lookup, for, count, audience). It is in English and Japanese. The code is written by demo/embed-code.js; scripts/embed-code.test.mjs runs the lookup of every way against the package and checks the pieces of each. The embed page itself is unchanged.

Changed

• scripts/check-standard.mjs is kazoku-ui 0.15's (it has the embed-panel item); the demo moves to @johnmorrisdotca/kazoku-ui ^0.15.1.

[2.7.0] - 2026-11-03

A minor version: nothing that 2.6.0 gave is taken away. The package is 2.2 MB larger to download (26.4 to 28.6 MB), all of it the strokes, which an entry point of their own holds and fetches only when asked for.

Added

• Stroke order. @johnmorrisdotca/jisho/strokes: loadStrokes(text) fetches the strokes of the kanji in a text (a piece of about 20 kB gzipped for each place they are in), strokesOf(table, kanji) reads one kanji's paths, in the order they are written, and strokeOrderSvg(strokes, options) draws them as an SVG that writes itself stroke by stroke, with CSS the SVG carries, so it needs no script and shows every stroke at once for a reader who asks for less motion. It takes size, numbers, color, label, strokeMs and animate. The strokes are KanjiVG's (release r20260714, CC BY-SA 3.0): the paths of 2,967 of the 2,999 kanji, in 57 pieces made by node scripts/build-strokes.mjs, each file crediting KanjiVG and linking its website as its licence asks. NOTICE.md says what the licence asks of a page that shows them.
• Search by drawing a kanji. findByDrawing(table, strokes, { limit }) takes the strokes as drawn (lists of points in any units, in the order drawn) and gives the nearest kanji first, each with a distance. It compares the strokes one by one, in the order drawn, with a stroke more or fewer allowed at a cost, whatever the size and place of the drawing. On KanjiVG's own strokes drawn at random sizes and places with noise, the right kanji is first for 98 in 100 and among the first five for all; it is not handwriting recognition. loadAllStrokes() fetches every piece (1.1 MB gzipped) for it, and pathPoints(path) turns a path into points.
• Typo-tolerant search. suggestWords(text, { limit, mode, block, allow }), in words, gives the common words whose reading is one slip away from romaji or kana that found nothing: a letter or kana too many, too few, wrong or swapped (arigatuo gives ありがとう, nihonngo 日本語, にほんこ 日本語). It reads the table of the common words, which it fetches once, so it covers those 26,039 words and no others, only by reading.
• The demo's kanji card writes the kanji's strokes out, with Replay and Numbers; the kanji view has a pad to draw a kanji on, with Undo and Clear, which lists the kanji nearest to it; and a word lookup that finds nothing offers "Did you mean". Guides for stroke order and for typos, in English and Japanese.

Changed

• package.json's license adds CC-BY-SA-3.0 for the strokes. The strokes keep that licence, not CC BY-SA 4.0, and say so in each file.

Not added

• Pitch accent. NHK's dictionary is never used. Kanjium's accent list says everything in its repository is CC BY-SA 4.0, but credits its accents to a database of "Uros O." that it names no source or terms for, and says the EDRDG files in it are the Group's without saying whether the accents are among them; Wiktionary has accents on some entries only and needs its whole dump read; OJAD is for non-commercial use. Using a list whose origin is not named is not something an MIT package can rely on, so the choice is John's (MODULES-340).

[2.6.0] - 2026-11-03

A minor version: nothing that 2.5.0 gave is taken away. The package is 12.6 MB larger to download (13.8 to 26.4 MB, 49 to 83 MB unpacked) for the whole entries and the example sentences, which two new entry points hold and fetch only when asked for.

Added

• The whole entry of a word. loadEntry(word), from @johnmorrisdotca/jisho/entries, fetches every sense of the word's entry of JMdict: its parts of speech, usage, field and dialect marks, cross-references, antonyms, notes, the languages a loanword comes from, every gloss with its kind (literal, explanation, figurative, trade name), the written forms and readings each sense is limited to, and the marks of the written forms and readings. A sense keeps the number JMdict gives it (Entry.senses[].number), so a cross-reference still means it when safe mode leaves a sense out; referenceOf reads a cross-reference into its word, reading and sense number. One piece of about 40 kB gzipped is fetched for an entry, and one more with mode: "unfiltered". Safe mode, the default, leaves out the senses JMdict marks vulgar, X-rated, sensitive or derogatory, with the written forms and readings that only they apply to, and fetches nothing that holds them (3 pieces hold the 857 entries it cuts).
• Example sentences. loadExamples(word), from @johnmorrisdotca/jisho/examples, gives the sentences that show a word: 32,434 of 26,377 sentences of the Tatoeba project, as the EDRDG's JMdict_e_examp links them to the senses of 28,929 words. Each has the number of the sense it shows, the Japanese with its furigana, the English, the JLPT level of its kanji and its credit, and exampleCredit(example) makes the line a page has to show with it (Tatoeba #162365, Japanese by alice, English by bob (CC BY 2.0 FR)). One piece of about 20 kB gzipped is fetched for a word. Safe mode leaves out an example of a sense it cuts, a sentence the file also links to such a sense of another word, and a sentence in which it does not find the words that unfiltered mode does.
• A Word has id, the number JMdict gives its entry. Each row of the data carries it, which is why the rows are 318 pieces of the same size and not 294.
• The demo's word list has a "Full entry" button on each word: its card shows every sense with its marks, cross-references (each looks its word up), notes and loanword sources, and under each sense its example sentences with furigana, English, level and credit. The entry open is in the address (entry=紙), and the card follows the language and the mode.
• scripts/build-examples.mjs (pnpm examples) makes the examples from JMdict_e_examp and Tatoeba's exports of Japanese and English sentences and their links; data-sources/sources.json records each file's address and SHA-256. Guides for both, in English and Japanese.

Changed

• The words entry is 1.0 MB larger gzipped as one bundle (13.8 to 14.8 MB) for the entry numbers in the rows. A lookup fetches the same one piece of about 25 kB gzipped as before.
• package.json's license adds CC-BY-2.0 for the sentences: SPDX has no identifier for the France port, CC BY 2.0 FR, which NOTICE.md names.
• The monthly refresh workflow makes the entries with the words, and the examples with them.

Not added

• JMdict's glosses in German, French, Russian, Spanish, Dutch, Hungarian, Swedish and Slovenian. The EDRDG licence statement puts only the Japanese and English parts of JMdict under CC BY-SA 4.0 and says the translations in other languages "are covered by separate copyright held by the compilers of that material". NOTICE.md says so, and the question of using them is with John (MODULES-338).

[2.5.0] - 2026-11-03

A minor version: nothing that 2.4.0 gave is taken away.

Added

• An embed page. embed.html on the demo site is made for an <iframe>: it draws what its address says (word=, meaning=, kanji=, with=, limit=, mode=, lang=, theme=) and nothing else, in safe mode unless the address says mode=unfiltered. A guide, in English and Japanese, shows the frame and every option.
• Every way to use the package, tested. examples/ has a file for each: a script tag from jsDelivr and from unpkg, an iframe, Vite, webpack, React (with a server render), Vue, Svelte, Angular, TypeScript and Node. pnpm test:usage packs the package as npm does, builds, type-checks or runs each against that tarball, opens the ones with a page in a browser, and fails when a guide does not hold an example word for word. CI runs it. Four new guides, each in English and Japanese: a plain page and a CDN, Vite and webpack, React, Vue, Svelte and Angular, TypeScript and Node, and the iframe.
• scripts/check-standard.mjs, kazoku-ui's check of the family's must-have list, copied unchanged and held to its recorded hash by a test, so pnpm check fails when a demo, an embed page, a documentation site in two languages, the family's files, a tested usage guide or a current picture goes missing.

Changed

• The demo, the API reference and the docs use the family's language picker and theme switch (kazoku-ui, a development dependency), with the family's current template, stylesheet and documentation generator. demo/index.html is now a file with the family's header and footer put in by scripts/site.mjs.
• The docs-site size test allows the few bytes of difference that a different gzip makes (a Mac and CI) and still holds the minified size exact.
• The README's pictures are made again, and show the embed page.

[2.4.0] - 2026-11-03

A minor version: nothing that 2.3.0 gave is taken away.

Added

• Read a text. readText(text, { mode, block, allow }) cuts Japanese into words. From each place it takes the longest stretch the dictionary has a word for, as it stands or as a conjugated form (食べなかった is 食べる), and weighs the few longest against the word that follows each, so that 今日はいい天気 is 今日, は and いい天気. Each ReadToken has the text as written, its kind (word, unknown, or other for spaces, punctuation, Latin letters and digits), the dictionary word, its form and the path to the text, its furigana (the text in pieces, each kanji piece with the kana that read it), its reading in hiragana, its meaning, the hardest JLPT level among its kanji (Jisho has levels for kanji and none for words), each kanji's school grade and level, and the other words the text could be (alternatives). Japanese that no word fits is unknown and is listed in result.unknown. A text is read up to MAX_READ_LENGTH (5,000) characters, and truncated says when it was cut. It fetches only Jisho's own pieces, as the text needs them: a first sentence of fifteen characters took about 30, and later text reuses them.
• Three rules keep the dictionary's rare words from taking over: a stretch counts when its word is common, or it has a kanji, or it is four characters or longer; a single katakana is never a word; and a text of kana alone is read as the word usually written in kana (は is the particle, not 葉 or 歯). A word of kanji that ends in は and begins with a word is the pair (今日は is 今日 and は).
• studyList(result, { kanjiOnly }) and studyListText(rows, "csv" | "tsv"): the words of a text once each, in order, with reading, meaning, level, the highest school grade among the kanji, how often the text has the word and the form it first came in, as rows and as CSV (a header, quoting what needs it) or TSV (tabs, no header, to paste or import).
• furiganaOf(written, reading) and furiganaOfSurface(surface, base, reading): which kana read which kanji of a word, and of a conjugated form of it (食べなかった is 食 (た) and べなかった; 来 takes the reading its ending asks for).
• An InflectedWord has form, the dictionary form as it was looked up.
• The demo has a Reader tab: paste a text and tap Read, see it as words with furigana above them, tap a word for its card (dictionary form and path, level, each kanji's grade, other readings), the words not found, and the study list with Copy, Download CSV and a switch to leave out the words with no kanji. The text is kept in the address. A picture of it in the README, a guide in English and Japanese, a troubleshooting entry and an upgrade note.

Limits

It has no grammar: it takes the longest word, so a rare long word can win, a name (ジョンモリス) is not found or is cut into small words, and a word with several readings gets the commonest one JMdict ranks, which for a single kanji can be the wrong one (本 is もと before ほん; alternatives has the others). In safe mode a word the mode leaves out is not found, so its text is listed as unknown.

[2.3.0] - 2026-11-03

A minor version: nothing that 2.2.0 gave is taken away.

Added

• Conjugated words. findInflected("食べなかった") finds 食べる and says how the text comes from it, as a path from the dictionary form, innermost first (["negative", "past"]). loadWords(texts, { inflected: true }) also loads the dictionary forms each text could be conjugated from, and inflectedOf(words, text) reads them with no waiting (and works on the table of common words with nothing more loaded). A text that is a word as it stands comes first, with an empty path; a word is given once, by the shortest path. Reading a text that was loaded without inflected: true throws a RangeError, rather than answer as if the dictionary had no word.
• deinflect(text), also at its own entry point, @johnmorrisdotca/jisho/deinflect (7.1 kB minified, 2.1 kB gzipped, no data and no fetch): the rule table that proposes the dictionary forms. It covers godan and ichidan verbs, する (and a noun with する: 勉強した is 勉強), 来る in kanji and kana, い adjectives and な adjectives with だ, です and ではない; the polite, negative, past, te, conditional (ば and たら), volitional, tari, potential, passive, causative and causative-passive forms; want-to, looks-like, too-much, easy-to, hard-to, while and the polite command; the helper verbs a te-form leads (ている, てしまう, ちゃう, ておく, とく, てみる, ていく, てくる, てくれる, てあげる, てもらう, てください); and chains of all of them (食べさせられたくなかった). It proposes from endings alone, so the dictionary decides which proposals are real, by the kind of word each must be.
• A golden set of 395 conjugated forms (src/deinflect-golden.test.ts), written out from the conjugation rules as charts of a verb or adjective to a row, each with its dictionary form, kind and path, and each carrying the note that a native reader should check it: none has been read by one yet. Every form is proposed by the rules and found in the real dictionary with its path.
• The demo shows the path under a word found from a conjugated text ("from 食べる: negative, past", and in Japanese 「食べるから：否定、過去」), a picture of it in the README, a guide in English and Japanese, a troubleshooting entry and an upgrade note.

Changed

• A Word's kinds can hold n, for a な adjective (静か, 好き, 元気), which JMdict marks adj-na and which kinds left out; the dictionary check needs it. A word that had no letters because it was only a な adjective now has n. The words are cut into pieces at a few slightly different places because their rows grew by a letter; they are from the same JMdict of 2026-11-03.
• The README's Limits no longer send a reader to Hikidashi for a conjugated word, and the Hikidashi recipe is gone from the README and the guide. Hikidashi is no longer a development dependency.

Not covered

The imperative, honorific (keigo) forms, 〜ず, 〜なければならない as a whole, the stem on its own (食べ), and a noun with だ: JMdict marks the な adjectives and not the nouns, so only な adjectives are read. 愛する and its kin, which conjugate their own way (愛さない), are found as the godan 愛す.

[2.2.0] - 2026-11-03

A minor version: the one change of behaviour is that a mistake which used to be quietly accepted now throws, and each of those is a call that had a typo or a wrong type in it.

Fixed

• A mode other than "safe" or "unfiltered" no longer takes the filter off. loadWords("あげちん", { mode: "Safe" }) returned the unfiltered words, and so did "SAFE", "", "unsafe" and "bogus", in loadWords, loadCommonWords and readCommonWords, while findByMeaning and wordsWithKanji read it the other way round. One check (modeOf) now runs at the top of every entry point and throws a TypeError for anything else; no mode at all is still safe.
• findByMeaning finds the obvious word. The index kept the commonest twenty words that held a term, so drink never gave 飲む, walk gave 歩く eighth and eat gave 食べる fourth. It now keeps up to ten words with an exact gloss first (a word whose first gloss it is before one that has it later, aside text such as "(esp. cool or cold)" left out), then the commonest that merely hold the term, and searchMeanings ranks the same way. A test holds 36 common English words to the word each must give in its top three. The meaning index is 172 pieces (it was 173) and the one for what safe mode leaves out 9 (it was 10).
• gradeOf and levelOf take exactly one character. gradeOf(GRADES, "") was 1, levelOf(JLPT, "") was "N5" and gradeOf(GRADES, "一右") was 1, because the lookup was a substring search; anything else than one kanji is now null.
• A caller can no longer change what the next caller sees: word.senses[0].glosses.push("X") showed up in every later lookup of that word. A row and every list in it, and the lists in a Words, are frozen, and a change throws a TypeError. withoutMarks, withoutWords, onlyWords and commonestFirst still return new arrays.
• block or allow given as a text (block: "神") threw "list.map is not a function" from deep inside a lookup. They, and the lists the tools take, are checked first and throw a TypeError that names the option.

Changed

• The pieces a process has read are kept up to 64 at a time (cacheLimit in words-pieces.ts), the one asked for longest ago let go. Looking up every piece once took a process from 49 MB to 266 MB. Words you have loaded are yours and stay.
• The documentation says what a lookup costs: a kana guess is one piece of about 25 kB gzipped, a written form two in a row (about 57 kB), findByMeaning("water") about 8 files and 180 kB in three steps, wordsWithKanji("山") 9 files and 210 kB; a piece of rows is about 74 kB and a piece of the written-form index about 120 kB. It no longer says "a few kilobytes" or "at most about 100 kB". The main entry is 0.9 kB gzipped (it was 0.8).
• NOTICE.md says that package.json's CC-BY-4.0 for Waller's JLPT lists is there because SPDX has no versionless "BY": his page names no version.
• The README opens with the package's thumbnail, and the demo carries it as its link preview (og:image) and publishes the card, a dark card and an icon on Pages at thumbnail.png, thumbnail-dark.png and icon.png.

Added

• pnpm test:coverage (@vitest/coverage-v8, a development dependency) fails below today's level: statements 99, branches 92, functions 98, lines 99 (measured 99.25, 92.11, 98.74 and 99.76). pnpm check runs it.
• AGENTS.md names the demo's local port block, 8500 to 8599. The demo's Playwright tests draw in software and run one worker.
• Tests for each of the above: every wrong spelling of mode, the golden set of English words, the frozen rows, the cache limit and the list options.

[2.1.0] - 2026-11-03

A minor version: nothing that 2.0.0 gave is taken away.

Added

• findByMeaning("water"): the commonest words whose English meaning holds the whole word or phrase, from an index of English terms in pieces of its own that only this lookup fetches. It takes mode, block, allow and limit. searchMeanings(words, text) is the same search over words already loaded, such as the common table, and fetches nothing.
• wordsWithKanji("山") and kanjiOfWord(word, table): the commonest words written with a kanji, and the entries of the kanji table (grade, JLPT level, readings, meanings) for the kanji in a word, so that a word can show its kanji.
• Romaji: loadWords, wordsOf, meaningOf and isLoaded take romaji: true and read Latin letters as modified Hepburn or Kunrei (kami, gakkō, kon'nichi, Tōkyō). A long vowel can read more than one way, so every way is looked up (kōhī finds コーヒー). romajiToKana and kanaFromRomaji are the conversion alone, with no dependency.
• How common a word is: a Word has priority, the tags JMdict gives it (news1, ichi1, spec2, gai1, nf05…), and rank, one number made from them, lower being commoner. commonestFirst, priorityTags and rankOfTags work on them.
• normalizeText: a text made plain for a lookup, and used by every lookup.
• WordsLoadError: what a lookup rejects with when a piece of the dictionary cannot be fetched, naming the piece (kind, index, piece) and keeping the reason in cause.
• A guide section for each, in English and Japanese, a recipe for looking up 食べました as 食べる with Hikidashi's dictionaryForms (a development dependency, run by the README's check), and a demo that finds words by reading, English meaning or kanji, leads from a word to its kanji and from a kanji to its words, and reads romaji.

Changed

• A text is made plain before it is looked up: NFKC (full-width and half-width, composed and decomposed), whitespace collapsed and removed beside Japanese, zero-width and control characters removed, a dash between kana made the long vowel mark ー. A word JMdict writes in full-width characters (１人, ＣＤ) is found by either spelling and shown as JMdict writes it. A text that is empty, not a string or longer than 100 characters is nothing to load and has no words; none of them throws.
• A piece that fails to fetch is asked for once more, and then rejects with a WordsLoadError, where it rejected with whatever the bundler or the browser threw. A failure is not remembered. Lookups at the same moment for the same pieces share one fetch.
• src/words.ts, which was 652 lines, is split by what each part does: words-load.ts, words-lookup.ts, words-common.ts, words-search.ts, words-pieces.ts, words-rows.ts, words-text.ts, words.types.ts and romaji.ts. Every export is still at @johnmorrisdotca/jisho/words.
• The words are from JMdict of 2026-11-03. The rows carry the priority tags, and the package holds the indexes by meaning and by kanji: 173 and 50 pieces of about 15 kB gzipped, and 14 more for what safe mode leaves out. A kana guess fetches the same one piece of 25 kB gzipped as before; the package is 3.7 MB larger to download (10.2 to 13.9 MB), and the pieces' JavaScript 13 MB larger (34.9 to 48.0 MB unpacked).
• The monthly refresh workflows no longer fail on purpose when new data arrives. They push the data/… branch as before and open a pull request from it (or update the one already open), titled with the month and saying to land it with a patch version, and the run ends green. They use only the workflow's own token, with pull-requests: write; opening the pull request needs the repository setting "Allow GitHub Actions to create and approve pull requests", and where it is off the run says so in a warning with the address to open it by hand.

[2.0.0] - 2026-11-03

Changed

• Breaking: forChildren is replaced by mode: "safe" (the default, which is what forChildren: true was) or "unfiltered" (which forChildren: false was). Words has mode, block and allow where it had forChildren; readCommonWords and loadCommonWords take the same options. The names are not the same, so this is a new major version.
• Every sense of a Word carries the marks JMdict gives it, in senses: usage (col, sl, arch, obs, uk, vulg…), field (food, med…) and dialect (ksb…). A row keeps its first four senses, and the meaning is made from the senses a word has when it is read. Unfiltered answers, and safe answers, are the same as 1.0.0's wherever a word has no more than four senses to leave out or keep, which a test holds to the dictionary.
• What safe mode leaves out is in pieces of its own (xr*, xf*, a manifest, and words-common-extra), cut small, so that an unfiltered lookup fetches one extra piece of about 19 kB gzipped for a reading, where 1.0.0 fetched one file of 92 kB. The common table is larger (about 770 kB gzipped) because every sense has its marks.

Added

• block and allow, the consumer's own lists, by written form or reading in either kana, on top of either mode. Applied only when passed.
• hasMarks, withoutMarks, withoutWords and onlyWords: small pure tools on what a lookup returned, which apply nothing on their own, and meaningFrom, which makes a meaning from senses. SAFE_MARKS lists the four marks safe mode leaves out. Jisho keeps no list of words of its own.
• A guide section on modes, lists and marks in English and Japanese, and a Mode row in the demo (Safe and Unfiltered) that shows each word's marks.

[1.0.0] - 2026-11-02

This package was published as @johnmorrisdotca/moji 0.1.x until now. Its entries are kept below as history.

Added

• Words (/words): loadWords, wordsOf and meaningOf look up the words of JMdict by reading (かみ: 紙 paper, 神 god, 髪 hair) or by written form (紙), each with a short English meaning, commonest first. The rows are cut into 254 pieces by reading and 71 of an index of written forms, and a lookup fetches only the pieces it needs, about 25 kB gzipped each. loadWords adds to words already loaded with previous, and wordsOf says when a text was not loaded instead of answering with no words.
• The common words (/words-common, loadCommonWords, readCommonWords): 26,039 rows in one table for a page that looks many words up.
• For children by default, and in the files: a sense JMdict marks vulgar, X-rated, sensitive or derogatory, a written form or reading that only such a sense applies to, and a word with nothing else to say, are not in any file a default lookup fetches, nor in words-common. They are in extra.data and common-extra.data (/words-common-extra), which are fetched only with forChildren: false; then every answer is what JMdict has. A word with no meaning is null and never a placeholder.
• scripts/build-words.mjs (pnpm words) writes the words from JMdict, and .github/workflows/jmdict-refresh.yml does it again on the 1st of every month and leaves a change on a branch, as the EDRDG asks of a package that uses its files. The release, source and licence are in NOTICE.md, data-sources/sources.json and the head of every data file.
• The demo can look a word up (a reading, a written form, or an example), for children or for everything, in English and Japanese.
• A guide, "Look words up", in English and Japanese, including how a server keeps the dictionary out of its bundle.

Changed

• The package is Jisho 辞書, a dictionary of Japanese words and kanji: its name, address (github.com/johnmorrisdotca/jisho), demo (johnmorrisdotca.github.io/jisho/), description, README and guides say so.

[0.1.1] - 2026-11-02

Changed

• The README's JLPT N3 picture ends after ten whole rows instead of cutting a row through, and its text says it shows the first ten rows of the 367 kanji (the grid scrolls in the demo).

[0.1.0] - 2026-11-02

Added

• School grades for 2,999 kanji (/grades), from KANJIDIC2.
• JLPT levels for 2,211 kanji (/jlpt), from Jonathan Waller's lists.
• Readings, English and French meanings, strokes and newspaper ranks for the same 2,999 (/kanji).
• 7,211 pairs of kanji that look alike (/look-alike), with two scores each, from Lars Yencken's kanji confusion datasets.
• gradeOf, kanjiOfGrade, levelOf, kanjiOfLevel, readKanji, searchKanji, kanjiIn and lookAlikes, and a demo.