Desktop changelog (Windows / macOS)

What changed in each version of Tomoshi — a summary of the things you can actually notice, with internal engineering work left out. Version numbers match the About page in the app; if you installed from the website, new versions reach you through the app’s automatic update.

1.1.3 (2026-10-18)

Word-choice cards explain the choice better. In English-to-Japanese reverse lookup, each candidate in “Choosing between renderings” now has a row of tags (part of speech, JLPT level, common, transitivity, register) and an example sentence with a translation — tap its words to look them up or hear them. Chinese-to-Japanese lookups get the same, plus a common collocation for each candidate (服を着る, 靴を履く); Chinese words with several meanings, such as 穿, 打 or 开, have their candidates grouped by meaning, and combinations Chinese speakers often get wrong get a “Don’t say ✗ 靴を着る” line above the summary.
Save to a different deck on the fly. Right-click a star to open the deck picker directly, regardless of your “Save to” setting; right-clicking a word already saved does nothing.
The word card that pops up on hover now shows JLPT level and “common” badges next to the heading, as the browser lookup extension does.
In grammar cards, irregular formations such as する→しようとしない are now split into “base → result” tags too.
The “common words” section of the search dropdown now looks at whether the spelling or reading you typed is common: searching てい no longer puts 弟 or 鼎, which merely have that rare reading, ahead of grammar.
Fixed: on results pages with several entries, the hover card for synonyms and antonyms had a dark halo around it.
Fixed: after a rare internal error, lookups and Wordbook reads and writes could keep failing until the app was restarted.
If you use the offline browser extension downloaded from tomoshi.cn and it is older than 0.3.0, the desktop app now reminds you to download it again.
macOS only: the Safari lookup extension is updated to 0.3.1 — grammar formations show as tags, the category counts at the bottom jump to their section, adjacent cards are separated by a line, and the heading now matches the explanation above it when you select a conjugated form.

1.1.2 (2026-10-15)

Conjugated forms now tell you their dictionary form. Type たべた or 埋まらぬ into the search box and the dropdown offers the dictionary-form entry, marked “← たべた · past (ta)”. When you land on the results page — with Enter, the global shortcut or the lookup popup — a line above the entry explains it (“埋まらぬ” is the negative (nu) of “埋まる”), and the search box keeps what you typed.
The conjugation table is more complete. Verbs gain polite forms (ません / ました / ませんでした / ましょう), the classical negative ず・ぬ (解けず, 埋まらぬ; あらず / あらぬ for ある) and the causative-passive; adjectives gain the polite negative (高くありません / 静かではありません). Polite forms, ず・ぬ and たら sit as sub-rows under their main row. Forms such as 解けず and 書かずに can now be looked up as well.
The search dropdown is grouped: common words → reverse lookup → grammar → uncommon words. When there are more suggestions than fit, a category bar appears at the top so you can jump to a group, plus a “Starts with …” chip that opens the full browse list. Common words are tagged “common”, and romaji (e.g. tokoro) now finds grammar too.
Pressing Enter also switches to the Japanese spelling. If you type a spelling with simplified or traditional Chinese characters, such as ご协力, and the dictionary, reverse lookup and grammar all come up empty while the dictionary has the Japanese spelling, Enter opens ご協力 directly, with a note that the Japanese spelling was used and a one-click way to search what you typed instead.
Review cards can read themselves aloud when flipped. A new “Read aloud on flip” checkbox under “Show answer” reads the word or grammar pattern each time you turn to the back (example sentences still play when you click them). It is off by default and saved on this device only.
Listening practice shows the auto-flip countdown on “Show answer”, e.g. “Show answer (3)”, and the countdown only starts once the audio has finished; a tap that lands right after the card flips is no longer counted as a rating. The play button no longer changes width while audio plays.
In the candidate navigation on the results page, common words are tagged “common” and grammar shows its JLPT level (e.g. “N5 Grammar”).
Transitive/intransitive pair rows now list the group’s other common spellings with readings, e.g. 表れる↔表す next to 現れる↔現す.
Grammar cards: divider lines between sections; the connection formula is split into tags, with parts of speech and the Japanese to attach in different colors; in the English interface, the Japanese part-of-speech terms in connection formulas are shown in English.
The deck list in the Wordbook shows how many words in each deck are mastered, and “All” on the review card is now “learning” (learning + mastered = total).
The top-bar search dropdown no longer covers the search box, the content below is frosted while it is open so it doesn’t blend into the result cards, and it closes as soon as the search box loses focus.
Saving or removing a word from the browser extension now updates an open Wordbook page straight away; with extension 0.3.0 you can also choose the deck when you star a word.
The feedback form’s “related entry” now defaults to the result you are looking at.
Fix: switching back to the app from another window, the tray or the Dock during a review or listening session moved the cursor into the search box.
Fix: a few words with several spellings (such as 叢立ち) showed the wrong reading on the mini word card, in the Wordbook and on review cards.
Windows only: Ctrl+W now closes the window (the same as the title bar’s ×, so it goes to the tray if the tray icon is on), and Alt+← / Alt+→ no longer step through your search history — use the top-bar buttons to go back and forward.
Windows only: fixed “Delete app data” at uninstall leaving behind the folder that holds the dictionary data and your Wordbook. The fix ships in the new uninstaller this update installs, so it applies to uninstalls from now on.
macOS only: the Safari lookup extension is updated to 0.3.0 — headings show JLPT level and “common” badges, results are grouped as common words → grammar → uncommon words with counts at the bottom, grammar cards are redesigned, you can pick a deck when starring, and selecting a conjugated form tells you its dictionary form.

1.1.1 (2026-10-10)

Fixed: after switching to another window and back (or changing the theme), hovering the toolbar icons or a word link in an example sentence made the results below jitter up and down slightly.
Fixed: grammar and Chinese-to-Japanese result pages opened scrolled down by a couple of pixels instead of starting at the very top.

1.1.0 (2026-10-09)

Your Wordbook can now be split into decks. Click “＋ New deck”, give it a name and a color (say, “Listening” or “Anime”), then move words in with the folder button on each word, or “Select” several to move or delete them at once. Once you have decks, the Wordbook opens on the deck list; open a deck to review just that one. Words you saved before are in “Default deck”. Saving a word then asks which deck it goes to — tick “make preferred” and later saves go straight there. Right-click a deck to edit, reorder or delete it.
Anki exports can include just the decks you pick, each filed as a subdeck under Tomoshi. The export file now carries a deck column, so newly imported notes go into the “Tomoshi” deck rather than the one chosen in Anki’s import dialog; notes you imported before are updated in place and stay where they are.
Listening practice is now always available. It used to unlock only after five words reached “Consolidating” in review; now any saved word will do. Words you have reviewed well still come first, and once those are done for the day, starting practice asks whether to go on with the ones you haven’t consolidated yet.
Inside a deck, listening practice drills just that deck, and its word count and hit rate are counted for that deck alone.
Listening practice now includes example sentences. Once you can catch a word on its own (listening level “Familiar” or above), it will sometimes play in one of its example sentences instead, so you practice picking it out in context; the word is shown in bold when you flip the card. Set “Sentences” to “Off” at the bottom of the practice screen if you’d rather not.
Words in the example sentences on flipped review and listening cards can be looked up directly. Hover for a mini card, click to open the entry in a popup — your session carries on where it was.
The search box accepts Chinese characters in simplified or traditional form and suggests the Japanese spelling (汤 → 湯, 学习 → 学習), marking which character it was converted from.
Senses that carry a JMdict usage note now show it. Chinese dictionaries show the note translated into Chinese (needs the 6 October dictionary data or later, which updates automatically); the English dictionary shows the original note.
JSON backups and Anki exports now mark AI-generated content (an ai_generated field in JSON, an ai-generated tag in Anki).
Fixed: importing an older Wordbook backup with cloud sync on could overwrite review progress made later on your other devices. Importing while signed in now syncs with the cloud first, and a backup imported while signed out no longer overrides newer cloud data once you sign in. Words you deleted earlier are no longer restored by importing a backup.
The grading buttons in review and listening practice show their 1 / 2 shortcuts again.
Fixed: in dark mode the formula block under “Formation” on grammar cards was nearly the same color as the card, so its edges were hard to see.
Fixed: closing the window on the first-launch dictionary download screen made the app exit with an error.
The Privacy Policy (international version) has a new section for users in the EU, UK and Switzerland, and the Terms of Use a new section on the EU Digital Services Act. The mainland China version is unchanged.

1.0.32 (2026-09-28)

Grammar cards have a new layout that, together with the 26 September dictionary data, splits each note into senses, formulas and pitfalls. The meaning is now a list of senses, each pointing at the example sentence it applies to when there are several; “Formation” is one formula per line, with label, formula and note each in its own place; “Usage” covers only nuance and when to use the pattern, while the mistakes Chinese speakers commonly make go into a separate “Watch out” block; the examples move up ahead of usage; and the body text has a maximum width, so it no longer stretches into long lines in a wide window. The block order is now Meaning → Formation → Examples → Usage → Watch out → Easily confused with. Installs that have not yet updated to the 26 September data keep showing the old notes, unchanged.
Fix: leaving a review or listening drill for a dictionary page and pressing “Back to the wordbook” now returns you to the same card. Opening the full entry from a card, or tapping a word inside an example sentence, used to drop you on the wordbook’s front page on the way back, throwing away the whole session so you had to start over. Your progress is now kept; leaving after a round is finished still lands on the list as before.
Fix: the “Starting with…” and “Containing…” tabs on the browse page now match only spellings the entry actually displays. Browsing “ヴェ” used to pull in words like ベテラン and ベスト whose search-only variant spellings contain ヴェ. When an entry qualifies through a displayable variant spelling, that spelling is shown as the headword with the usual spelling on the reading line, and the “N words” count agrees with the list.
The preview line for grammar entries in the wordbook list now reads the current dictionary data. It used to show the sentence saved at the time you added the entry, which never changed after a data update; it now shows the first sense, or the first sentence of the meaning.
AI result sharing has been removed. The “Share with other learners” action on sentence analysis and word enrichment cards, the “Shared / Withdraw” state and the “Shared” filter on the history page are all gone.
Fix: the pronunciation button no longer changes width while it shows “…” during playback, so the centred reading and example lines on review cards no longer shift sideways.

1.0.31 (2026-09-25)

The back of cards exported to Anki has a new layout, and a pitch-accent curve after the reading. The back now follows the hierarchy of the app’s own cards: small grey section headings with thin rules, body text and secondary notes at different weights, a vertical bar beside each example sentence, hanging numbers, small grey tags for translation / register / scene / notes, a blue JLPT badge and four-colour register badges. All styling is inline in the card, so it reads well in both light and dark Anki themes. Word cards now draw a pitch-accent curve after the reading, the same shape the app uses (a hollow particle dot for odaka words); it is skipped when the reading is not kana or the mora count doesn’t line up. Register badges now follow the interface language; exports from the Traditional Chinese and English interfaces used to print the Simplified labels. Anki needs no note-type or template change: re-export once and the existing notes are updated.
Fix: removing an unrated card from the wordbook during a review now deals a replacement. The card used to vanish from the deck without counting as reviewed, leaving one slot of today’s quota unused, so as long as any word was still due the red dot lit up again after the finish screen with “1 due today”. A card you had already rated is not replaced; it already counts.
The proficiency badge on listening-drill cards now shows the listening progress itself. It used to copy the review card’s four mastery levels, so a word you never managed to catch could still wear “Mastered”. It is now derived from the drill’s own scheduling interval in five levels (first time / shaky / getting familiar / solid / heard it), with a headphones icon and a hover explanation.
On editions where AI analysis requires signing in, the AI example chip on the home page now shows before you sign in; tapping it leads to the “sign in to use AI analysis” prompt instead of doing nothing.
Fix: the pronunciation buttons in the dictionary entry header and in the AI analysis headword row sat slightly above their line; they are now vertically centred.

1.0.30 (2026-09-23)

Cards exported to Anki now carry the full explanation, the same one the app’s review card shows. The back used to hold only the reading, the first three senses and one example sentence; a grammar point was reduced to the single line saved with it, and a saved AI analysis to a one-line summary. Now a word lists every sense (each with its part-of-speech and usage tags and its own example sentences); a grammar point brings its meaning, formation, usage notes, examples and the contrast with easily confused points; a saved AI analysis is laid out by type (senses, examples and near-synonyms, or the translation, grammar notes, Japanese renderings and question answers). AI-generated content is marked as such at the foot of the back. Reverse-lookup cards were already a single line and are unchanged. The meaning column the file carries for custom note types used to hold a blob of internal data on AI cards; it is now the same one-line summary the wordbook list shows.
Example sentences on review cards, listening-drill cards and the back of cards exported to Anki now carry furigana over their kanji. They used to be plain text in all three places, so seeing a reading meant opening “Show details”. The furigana is sized as on the dictionary result page; review cards still don’t make words tappable (a word card would interrupt recall). Anki needs no note-type or template change: re-export once and the existing notes are updated.
Export to Anki from the wordbook page. “Export to Anki” now sits next to “Sync now” at the top of the wordbook, so you no longer have to dig through Settings; the entry in Settings stays where it was. Each export from here ends with a short note: how many entries went out, how to import the file into Anki (and to sync it over through AnkiWeb if you also use Anki on your phone), that exporting again updates the same notes, and that review progress does not travel with the file.
Fix: romaji spelled by sound now finds the particle. Typing korewa or konnichiwa in the English interface — the particle は spelled as wa — used to find nothing and fall through to AI analysis. The search now also tries わ → は inside the word, so これは and こんにちは match directly. Words that really contain わ, such as かわ (川), are unaffected.
The quick-reference index now lists the JLPT levels group (grammar / vocabulary / kanji) at the top.

1.0.29 (2026-09-21)

Listening practice now queues words by due date instead of repeating the same set. Until now the order came only from your lifetime hit rate: words you had missed stayed at the front forever, and words you got right almost never came back — hence “today’s first set is identical to yesterday’s”. A word you recognise now moves back along a 1 / 3 / 7 / 16 / 35 / 60 day ladder, and one you miss returns to the first step. Each set only draws words that are due today, and “Done for today” appears once the queue is empty. Your existing practice counts and statistics are unchanged.
Review and listening cards show the full entry once flipped. All senses are laid out at once, each with an example sentence (the Japanese, a play button for the whole sentence and one translation); saved AI word analyses and grammar points are also shown in full, so there is no need to open “Show details” for them. Cards now start right below the header, and “Show answer” and the two rating buttons stay pinned to the bottom of the window, so they no longer scroll away on a long card. “Mastered” is now a button, labelled “Mark as mastered”.
Extra practice shows how much is left for today: once the review button reads “Keep practising”, the wordbook adds a line with today’s total reviews and how many more you can still do — both counts merged across your devices.
Fixed: looking up a rare spelling no longer puts a different word first. 風 used to return 振り first (it lists 風 as a rare spelling and is the more frequent word), so hovering 風 in Japanese reading showed ふり. Entries matched only through a rare, old-style or irregular spelling are now ranked below the rest — in results, in autocomplete and in the browser lookup extension.
Wordbook sync holds up better on large batches and on failure. The first upload after signing in is now throttled into batches with a progress indication and resumes where it left off (20,000 entries take about 8 minutes, down from 50); a failed sync now backs off and retries on a schedule, and an account with many saved words can no longer get stuck retrying the same batch.
Fixed: the play buttons next to example sentences and headwords sat slightly low against Japanese text, and are now aligned to its visual centre line.

1.0.28 (2026-09-19)

Downloaded updates now install themselves the next time you open the app (Windows). Until now a downloaded update was only installed if you clicked “Restart to update”; quit without clicking and it had to be downloaded again next time. The download is now kept on your computer, and the next launch shows a short installation progress bar before opening the new version. “Restart to update” still works if you want it right away. The macOS version is unchanged.
Safer upgrades. Before each upgrade to a new version, the app keeps a local backup of your wordbook and search history (the two most recent are kept). The settings file is also saved more reliably, so a power cut or crash can no longer reset all your settings.
A diagnostic report is sent when an automatic update fails. It follows the same switch as “Send crash reports automatically”: the report contains only the step that failed, version numbers and the error message — never anything you looked up or any personal information. Turn that switch off and nothing is sent.
Japanese reading: more words in the text can be opened, and more accurately — verbs and adjectives written in kana, and words that only make sense as a whole such as おかあさん or 王さま. Fixed The Little Prince and Wild Pear having a first chapter of a single sentence. おりこうハンス has been removed from the selection.
Quick reference trimmed: the verb collocations, sentence connectives and onomatopoeia tables are no longer in the app (verb collocations remain available on dict.tomoshi.app); attachment forms now sit in the renamed “Basics” group.

1.0.27 (2026-09-18)

Anki exports now work as cards out of the box. With 1.0.26, Anki’s default Basic note type could show only one exported field on the back; the file now carries Front / Back columns — the word on the front; the reading, the first three meanings with part of speech, one example sentence with its translation and your notes on the back — so choose Basic when importing and you’re done. The separate fields are still there for custom note types, and exporting again updates notes you imported earlier to the new layout.

1.0.26 (2026-09-18)

Review scheduling now uses FSRS — intervals are computed by the FSRS algorithm (the same open-source implementation Anki uses) from each card’s memory stability and difficulty; existing progress carries over, no need to start again. The two grade buttons and graduation rules are unchanged.
Export to Anki — Account & Sync → Wordbook data gains “Export to Anki”, a CSV Anki imports directly (headword, reading, furigana, meaning, notes, JLPT and mastery tags); exporting again updates the same notes instead of duplicating them.
Keep reviewing after the daily goal. Once today’s goal is met, the Wordbook’s review button turns into “Keep practising”, and the review page shows today’s total and offers “Review N more” to keep going through due words, up to a new “Daily review limit” (default 200, Account & Sync → Wordbook review). The old “Daily review limit” is renamed “Daily review goal” (default 20); it still sets how many cards are dealt each day and drives the reminder dot and streak. Both values sync with your account.
Redesigned review cards: the back of the card groups meanings by sense (number, part of speech, meaning), keeps a short part of speech on the same line as its meaning, and lets you expand the senses that don’t fit; the top of the card now shows the JLPT level, common-word, transitivity and register tags; saved AI word analyses render like word cards, with reading, pronunciation button and grouped meanings.
Listening practice stats — after you’ve practiced, the listening card shows how many words you’ve drilled, your recall rate and when you last practiced.
Reading progress syncs with your account — once signed in, your place in each text follows you across devices. The word card that pops up on hover while reading now shows the part of speech, lists up to 3 meanings and tells you how many more there are; the graded sections are now simply called “N2 short stories” and “N1 short stories”.
Clearer part-of-speech labels — labels follow the terms learners know from textbooks and are shorter; in the English interface they are no longer JMdict’s long descriptions but short names such as godan / ichidan.
The Wordbook opens faster — with a large Wordbook it no longer takes two or three seconds to appear.
Fix: words you already reviewed on another device are no longer reviewed twice here. A session deals its cards when you open it, and closing the window only hides Tomoshi to the tray — reopen it and the old deck is still there, even though you may have finished it on your phone since. Those cards are now skipped with a note (“Already reviewed on another device”): they no longer count towards today’s tally, and they cannot push a word’s mastery level an extra step.
Fix: the daily review allowance is now counted per account, not once per device. A session re-checks today’s remaining allowance after every card (reviews on your other devices count towards it) and wraps up once it runs out.
Fix: a device that hadn’t synced in a long time could reset your review streak. The streak now merges the review dates from all your devices, so a device coming back online can only add days, never erase them.
Fix: when a background sync failed, the Wordbook page stayed on “Syncing…”. It now shows “Sync failed” with a short reason until the next successful sync.
Privacy policy updated: while signed in, the dates your account is used are recorded, solely to count how many accounts are active each day, and are deleted along with the account. The mainland China channel edition shows the consent notice once more on first launch.

1.0.25 (2026-09-14)

Interface tweaks: on the Wordbook page, “Sync now” moved into the header right after the last-synced time (“Sign in to sync” when signed out); in a narrow window that line wraps below the title, and the search box now sits the same distance from the blocks above and below it. On the Account & Sync page, the signed-in account card gains a line saying what syncs (wordbook, review progress and daily review limit; lookup history stays on this device), and the duplicate hint below the card is gone.

1.0.24 (2026-09-14)

Fix: a daily review limit set before 1.0.23 did not sync to other devices. On sign-in the value already in the cloud now wins, and only if there is none yet is this device’s setting uploaded; devices that are already signed in upload their setting on the next sync after updating. After that, a change on any device still follows to the others as before.

1.0.23 (2026-09-14)

Your Wordbook now syncs across devices. Sign in under Settings → Account and your Wordbook and review progress stay in sync across every device you sign in on (desktop and mobile): once after launch, a few seconds after each change, and whenever the window returns to the foreground; the account panel also has a Sync now button. The international service region signs in with an email verification code — no password; the mainland China region keeps phone sign-in, and the same account is used for AI analysis. Without signing in nothing changes — everything stays on this device.
History never goes to the cloud. Sync covers the Wordbook alone — saved entries, grammar points, AI analyses and your notes, plus review progress; History and the AI analysis cache stay on this device only.
Changes from several devices merge instead of overwriting each other. Edit a note on one device and review offline for days on another, and both survive the sync; the same word saved separately on two devices becomes one entry with the newer review progress; a word deleted on one device disappears from the others.
Your daily review allowance and streak are per account too. Cards reviewed today on each device add up: review 20 on your phone and the desktop has only your limit minus 20 left for today; finish today’s review on one device and the badge on the other clears after the next sync. The daily review limit in Settings is changed once and applies everywhere.
The account panel shows sync status: last sync time, items waiting to upload, and the reason for any failure. Sign out only stops syncing and keeps the local Wordbook; Delete account also removes the cloud copy.
Privacy Policy and Terms of Service updated: the Wordbook moves from “stays entirely on your device” to “local by default, synced once you sign in”, with new sections on accounts and cloud sync. The mainland China build therefore shows the consent screen once more on its next launch.
Interface tweaks: Settings gains an “Account & Sync” page holding the Wordbook review and Wordbook data groups, with an account card at the top of the sidebar; the Wordbook header shows the last sync time; the model badge on AI result cards now shows a short model name.

1.0.22 (2026-09-12)

Search in romaji. Type taberu, terebi or koohii to find 食べる, テレビ or コーヒー; conjugated forms such as tabeta are resolved as usual, and autocomplete now suggests the matching kana and kanji as you type. It works in every interface language and with every dictionary pack. To stop an English word being read as romaji, wrap it in double quotes: “date”.
Hiragana and katakana look each other up. Searching たばこ finds an entry listed only as タバコ, and the other way round; in every UI language.
AI analysis is now available in the Chinese mainland service region, after signing in with a phone number. Sign in under Settings → Account with a mainland China number and the SMS code; the code is delivered via Alibaba Cloud and the message is signed 【恒创联众】. No sign-in is needed if you do not use AI analysis. You will be asked to accept the Terms of Service and Privacy Policy first; the mainland channel build shows its first-launch consent notice once more because the policy changed substantively. The note under Settings → Service region now says AI analysis requires signing in.
Complete “AI-generated” labelling on AI result cards. The badge shows a full-sentence hint on hover, the supplement card’s badge matches the analysis card’s size, and copying text from an AI card via the right-click menu appends “(The text above was generated by AI)” to the clipboard.
Right-click “Search selection” opens a word popup. It no longer jumps to the results page and interrupts your reading; the popup can be dragged, stacked and closed with Esc. When nothing is found it offers “Analyze with AI”, which expands to the full page and lets the results page run the query.
Near-synonym contrasts on the AI supplement card are now a numbered list. Headword and note flow inline with a hanging indent, the same layout as the dictionary’s own contrast lists; the previous two-column layout misaligned when headwords differed in length.
Less glow around text in dark mode. Body contrast is slightly reduced, furigana is one step brighter, and bold text uses a medium weight in dark mode; the light theme is unchanged.
English by default on non-Chinese systems. On a first install where the system language is not Chinese, the English dictionary pack is preselected and the interface language follows; both can still be changed separately in Settings. Previously such systems defaulted to Simplified Chinese.
Chinese mainland channel build: the About page shows the app’s filing number with a link to the filing system.
Fixed: on a few words, the AI supplement card could turn the whole page blank. It happened when the AI listed the same near-synonym twice; cards already saved on your device recover too, with no need to look the word up again.
No more false crash report on first install. On the first-launch download page, a background warm-up task failed silently whenever the window gained focus and was logged as a crash; nothing was actually affected, but the next launch would send me that false report.
Fixed: pressing a pronunciation button no longer plays over listening practice or shadowing audio that is already playing.
Fixed: the negative of ある was mechanically generated as あらない; it is now ない / なかった.
Fixed: reverse-lookup entries in search history and the Wordbook keep the language tag of the dictionary pack they were saved from instead of changing when you switch packs.
Fixed: the Cheat sheets and JLPT pages scroll back to the top when switching sub-pages; the Reading page’s section highlight no longer freezes and highlights the last section when scrolled to the bottom.
The spelling-variant card is no longer shown when only one visible variant remains — its closing line pointed at variants you could not see.

1.0.18 (2026-09-02)

The window can now be pinned on top. A new pin button in the top bar keeps Tomoshi floating above other windows — no more switching back and forth while you write.
The window goes narrower (minimum width 640 → 520). You can park a slim lookup window beside your document; below a certain width the result cards, grammar cards and Wordbook switch to a single-column layout.
Fixed: window size and position no longer get lost. Until now the geometry was written to disk only on a clean exit, so an installer overwrite, an End Task, a crash or a shutdown threw away everything you had adjusted and the next launch sprang back to the old size — and in-app update restarts lost it too. The window is now remembered when you close it, shortly after you stop dragging or resizing, and before an update restart.
“What’s new” is available in English. With the interface language set to English this page no longer falls back to Simplified Chinese; all four sets of release notes (desktop, mobile, browser extension, dictionary data) have English versions.
AI-generated content is now labelled everywhere it should be. Several places were unlabelled before: the Chinese translations of example sentences, the synonym comparisons, grammar cards and the Chinese readings of kanji all say so now. Content that ships with the dictionary data and content generated live in the cloud carry two visibly different labels; only the latter can be reported.
Chinese mainland edition: the data-source AI labels are gone, and the Settings entry “Speech synthesis” is now called “Pronunciation” — the “Voice” section, which only ever offered one choice, has been removed with it. Reading words and examples aloud is unchanged.
Fixed: “Which spelling to use” cards were completely missing with the English dictionary. All thirty-odd thousand of them — 使う / 作る and the rest — were lost in the English pack. They are back.
Spelling cards no longer list spellings that are absent from the headword. The card for 思う used to list 懐う・惟う・意う, variants that exist only so the entry can be found by searching.
AI lookups: asking in English how to say something in Japanese now returns a word-choice card (which of several Japanese renderings to use); it used to be treated as a plain question.
Error messages are readable again. A failed first-launch download or AI lookup used to put the server’s Simplified Chinese diagnostic string on screen as the main text, which was especially jarring in English and Traditional Chinese. The main text is now localised and the technical detail sits underneath in small print, still there for troubleshooting and bug reports.
Entry cards, grammar cards and AI result cards now share one header layout — title on one line, star and tags on the next — so nothing bunches up in a narrow window.
The dictionary language and interface language lists in Settings are now in the same order, English first. The former used to copy whatever order the dictionary data happened to use.
Text describing the dictionary contents no longer says “Chinese” outright — an English interface with a Chinese dictionary pack is a perfectly normal combination — and the pack size is given as a range, since it differs by language.
The separator between senses follows the language of the senses themselves: a half-width semicolon for English, a full-width one for Chinese.
macOS: the small tags inside the example chips on the home screen, and those in the candidate navigation in the top bar, no longer sit too high.
macOS: the Safari extension now goes by the short name “Tomoshi” in Safari’s settings, where the full name did not fit; the name shown in the system’s extension list is corrected accordingly. With “Follow the browser”, the extension’s interface now falls back to English rather than Simplified Chinese when the browser is in a language other than Chinese or English (Japanese, for instance); you can change it back on the extension’s settings page.
Assorted typographic and wording fixes in the English interface: centred wrapping in collapsible section headers, a shorter name for “Open-source licences” in the sidebar, and corrected tag sizes and alignment.
The service-region suggestion banner has been removed. It guessed which side you should use from your network location, but the region is a matter of where your account belongs, not something a probe should suggest; you can still switch it in Settings at any time.

1.0.16 (2026-08-30)

English added as an interface language. Set the interface language to English in Settings and the menus, buttons and messages all switch to English; AI sentence explanations switch to English too, and the privacy policy and terms of service have English versions (the Simplified Chinese versions govern). The item notes in the reference tables and in Shadowing follow suit; dictionary definitions and the introductions to the works under Reading are still in Chinese.
New reference table: Sentence connectives. The words that stitch two sentences together — だから, しかし, ちなみに and the like — sorted into nine groups by what they do (consequence, contrast, addition, comparison and choice, rephrasing and supplementing, changing tack, plus conjunctive particles that attach to the end of a clause, compound conjunctions, and nouns and adverbs pressed into service as connectives). Every word carries a register band and an example: one English “so” splits into だから / ですから / したがって / ゆえに depending on the situation, and this table exists to help you choose. 92 common words.
New reference table: Onomatopoeia. Words like ドキドキ, ふわふわ and にっこり, sorted into six groups by domain (sound; light and texture; feelings; movement and posture; bodily sensation and weather; and the ones that have settled into adverbs of degree and state). Each comes with its single most typical collocation — these words are learnt by collocation, and the collocation also shows you whether it takes と, に or する. 198 common words.
“Common attachments” is now called “Attachment forms”. Nothing about the contents changed; the new name keeps it apart from the new connectives table. One is about form (which form たい attaches to), the other about word choice.
The Reference home page has been regrouped. Nine cards under a single “Readings and expression” heading was too crowded. It is now three groups: Readings (how to read it — the kana chart, dates and times, counters), Expression (how to say it — verb collocations and the three word-choice tables) and Practice (Shadowing, Reading). Finding a table is quicker.
Reading has grown from 30 texts to 71, with a new “World classics” section. The additions are Japanese translations of foreign classics: Kafka’s The Metamorphosis, Dostoevsky’s Crime and Punishment, Gulliver’s Travels, Dickens’s A Christmas Carol, A Little Princess, Little Women, A Dog of Flanders, Stevenson’s Jekyll and Hyde, Wells’s The Invisible Man, Turgenev’s First Love and Hugo’s The Last Day of a Condemned Man.
Both graded-shorts levels have gained a dozen or so texts each. N2 gets Andersen and Grimm (The Ugly Duckling, Little Red Riding Hood, Hansel and Gretel, The Steadfast Tin Soldier…); N1 gets shorts such as Maupassant’s The Necklace, Kafka’s The Judgment, Chekhov’s Kashtanka, Lu Xun’s A Madman’s Diary and Irving’s Rip Van Winkle.
Ten of them are modern translations their translators released under a CC BY licence — The Little Prince, Alice’s Adventures in Wonderland, The Little Match Girl, The Emperor’s New Clothes and two Sherlock Holmes stories among them. Their Japanese reads far more easily than the pre-war translations; for the same Andersen story you can now read a 1950s hand and a 2000s hand at the same level. The translator and licence for each text are given under “Source and base text” at the foot of the page.
One tap back after looking a word up from a reference table or a reading page. When you tap a word on one of those pages and land on the results, the top bar now shows a link back to where you came from.
The hover card is easier to click into. Moving the pointer from the word to the card no longer makes the card vanish on the way, and the card is no longer replaced on the spot by another word the pointer happens to cross.
Senses with no Chinese definition now show the English gloss instead of leaving a gap.
The part-of-speech tag on words like 得る and あり得る used to come out as a run of raw English; it now shows the proper label.
Fixed a garbled character in the text of Night on the Galactic Railroad: the 僕んとこへ in it used to render as 〓.

1.0.15 (2026-08-23)

Two new switches on the Shadowing page: Furigana can be turned off, so you can work from reading along with the kana towards shadowing with them hidden; Slow 0.75× slows things down while natural speed is still out of reach, and stretches the pause left for you to repeat by the same factor. Either switch takes effect immediately, mid-playback. Both match the shadowing pages on the dictionary website, and the furigana switch is named the same as the one on the Reading pages.
The display switches on the reference pages have all moved to the right of the heading (where the furigana switch sits on the Reading pages): the romaji switch on the kana chart is now a single button rather than a show/hide pair — press it and romaji appears. Jump links and filters stay on the line under the heading, no longer mixed in with display switches.
Interface details: a number of slightly-too-small text sizes on entry pages, AI result cards and in the wordbook have each been bumped up a step; source tags and candidate tags are now the same size as the buttons in the same row.

1.0.14 (2026-08-23)

New in the reference tables: Shadowing. Over three hundred short lines and dialogues from everyday situations, written and graded N5–N1, from greetings and small talk through ordering and shopping to part-time work and workplace keigo. Every line has furigana, a Chinese meaning and synthesised audio, and many carry a note about a mistake Chinese speakers commonly make. Play line by line, loop a single line, or run a whole set as “play + shadow” — each line is followed by a beat’s pause, just long enough to repeat it. The audio is downloaded on demand the first time you play it, so it adds nothing to the install size.
New in the reference tables: Reading. Thirty Japanese texts: ten classics (I Am a Cat, No Longer Human, Rashōmon, Kokoro, Botchan…) plus twenty graded shorts, ten each at N2 and N1. Each has a Chinese introduction saying what it is about and what reading it is good practice for, and the text carries furigana; content words are marked in blue — hover for a small card, click for the full entry, the same set as the example sentences on a results page. Everything comes from Aozora Bunko and is out of copyright, and all ten classics are complete texts: long works are split into the chapters of the original, and whichever chapter and paragraph you reached is remembered for next time.
Tomoshi can now also be installed from the Microsoft Store. The Store build keeps its own separate data, so it downloads the dictionary again on first launch, and your wordbook and history do not come across; if the build downloaded from the website is already installed on the same computer, the Store build picks up its existing data instead. The two cannot run at the same time, and updates to the Store build are handled by the Microsoft Store, so it no longer shows the in-app update prompt.
Fixed the first AI lookup of the day sometimes taking twenty-odd seconds before anything appeared: the connection is now made ahead of time at launch and when the app returns to the foreground, and establishing it no longer waits indefinitely on a flaky network. The same kind of stall on the first cloud read-aloud after switching networks is fixed along with it.
Interface details: in example sentences, grammar cards and the small card that pops up when you tap a word on a reading page, longer headwords no longer break across two lines mid-word; the AI waiting message now describes what is actually happening, instead of always claiming a reasoning model is in use.

1.0.13 (2026-08-18)

AI sentence explanations and entry supplement cards now have a Report control: if the content is inappropriate, harmful or seriously misleading, pick a reason (a note is optional) and submit it for review by a person. The feedback form in the bottom right has the same category, so you can still report after leaving the card.
“Regenerate”, “Share” and “Report” on result cards now open and close consistently: opening one collapses the other two, and tapping it again collapses itself. The button no longer disappears after you tap “Share”.
Interface details: the content column on results, reference tables and the home page is now the same width as on the dictionary website; settings content is centred to the right of the sidebar; the note at the foot of the home page has been trimmed and a “View all open-source licences” link added.
Fixed the review dot lighting up again — and the review streak possibly being counted short — after un-favouriting a word you had already reviewed today.
In the Traditional Chinese interface, the “詞形變化” collapsible heading on entry pages was showing in Simplified Chinese; the translation has been added.
Privacy policy update: notes covering the dictionary website dict.tomoshi.app have been added, and the effective date is now 2026-08-18.

1.0.10 (2026-08-15)

If the network drops while the dictionary is downloading on first launch, the download now carries on from what has already been fetched instead of needing you to tap “Retry”. It only reports failure after several attempts in a row make no progress, and the message then says what actually went wrong (a timeout, a dropped connection and so on) rather than showing an opaque technical error.

1.0.9 (2026-08-14)

If the Microsoft Edge WebView2 runtime is missing from the computer (the app’s interface depends on it), startup now gives a clear explanation and a download address, instead of flashing past or showing an unreadable error.

1.0.8 (2026-08-11)

New Service region setting: choose whether the cloud services you use sit in the International or the Chinese mainland region. Cloud speech, AI analysis, dictionary data downloads and app updates all connect to the servers of the region you pick. You are asked once on first launch, with a preselection based on your current network; anything installed before this version is treated as International, and if the international servers turn out to be hard to reach you get a one-off suggestion to switch.
The two regions are separate services: the daily AI quota, shared content and cloud records are not shared between them, so switching amounts to starting again in the new region — which is why nothing is switched for you automatically. Local data such as your wordbook and history is unaffected. You can change it at any time in Settings, where the operator for the current region is also stated.
The privacy policy and terms of service inside the app follow the service region: under Chinese mainland they state the mainland operator, the applicable law and the competent court, along with the note on access logs retained as required by law.
Fixed the window chrome (title bar and window background) not following the system when the theme is set to System and the system switches between light and dark — the page itself switched correctly, but the window needed a restart.

1.0.7 (2026-08-06)

Wordbook export and import. Export the whole wordbook to a file in one tap and import it back after a reinstall or on a new machine, review progress included. Import merges by word and never overwrites — where both sides have a word, the one further along wins, and a note from the file is only taken if there is none locally. The control sits beside the search box on the wordbook page; the pair in Settings stays as well.
Search within the wordbook. No need to page through hundreds of entries: filter by word, reading or definition as you type. Katakana input finds hiragana readings too (typing コーヒー finds it all the same).
Ctrl+F no longer opens the browser’s own find bar — that one only searches text already drawn on screen, which is no help for looking words up. On the wordbook page it now focuses the search box above instead.
The three JLPT tables have moved into the reference tables: one fewer button in the top bar, and the Reference home page is split into “General reference” and “By JLPT level”, each table recognisable at a glance.
New Speech cache entry in the speech synthesis section of Settings: it shows how many clips are cached and how much disk they take, and clears them in one tap. When a reading was wrong and has since been fixed in the cloud, clearing the cache lets you hear the correction straight away rather than waiting for the cache to expire on its own.
The wordbook and history pages have been brought into line with the mobile layout: export/import moved beside the search box, the delete control is now always visible (it used to appear only on hover), and the left edge of each word lines up with the other blocks on the page.
Fixed variant spellings turning up at the head of completion candidates: looking up 据 now lists the usual 据え置き rather than 据置き. The entry you land on is unchanged.

1.0.6 (2026-08-05)

Dictionary data updates are now incremental: only the difference from the copy on your machine is fetched, which typically brings one update down to a few per cent of the previous download. If the incremental update does not apply, it falls back to a full download automatically — nothing for you to do.
Fixed the target card often showing only a sliver after you tap through from a search suggestion: dictionary cards above it were still loading content and pushing it out of view. The position is now corrected until the layout settles, and control is handed straight back to you the moment you scroll.
Completion candidates and the browse page now show the spelling that actually matched: looking up 明 lists 明く / 明くる rather than the seemingly unrelated 開く / あくる. The entry you land on is unchanged.
The review-complete screen now shows an “N-day review streak”.
The form filter on the “Common attachments” reference table now greys options out rather than hiding them, so the filter bar no longer changes height while you type and the page no longer jumps up and down.

1.0.5 (2026-08-02)

The wording around sharing an AI explanation has gone back to being about other people: the button now reads “Share with other learners”, the explanation leads with who the content will be kept for, and the quota bonus is mentioned only as a thank-you. Sharing still works normally once today’s bonus is maxed out.
A share can be withdrawn within 24 hours; after that the content stays in the shared pool and the withdraw control is no longer shown. The confirmation dialog states this period.
Results you have shared no longer show “Regenerate” and no longer keep urging you to share; they show “Shared · Withdraw” instead. To regenerate, withdraw first — both controls come back together.
When another learner has already shared a result, you are told so honestly after a restart, instead of tapping share for nothing.
Cloud read-aloud now has a local cache: playing the same audio again does not go over the network, so it sounds the moment you tap.
Hover card fixes: closing the card cuts off the read-aloud it started, and the favourite button is back on the card.

1.0.3 (2026-08-02)

New reference table: Common attachments. 462 common patterns sorted into 8 groups by the form they attach to (ない form, ます form / continuative, て form, た form, plain / dictionary form, ば form / conditional, volitional, imperative), each group listing the patterns that can follow that form, with examples. A pattern opens the matching JLPT grammar details; the page has its own search, so you can jump straight to one. Form explanations and slot notes are collapsed by default — “Show explanation” expands them all at once.
The Reference home page is now a card index, each table recognisable at a glance, with updated navigation icons.
Reference search fixes: results no longer include dead links that open an empty page, and the matching rules are now consistent across tables, so a difference in spelling no longer decides whether something is found.
Removed the outdated note on the read-aloud button of AI results saying speech uses the local engine — since 1.0.2 all speech is synthesised in the cloud.

1.0.2 (2026-07-31)

Speech has moved entirely to cloud synthesis: the built-in local speech engine is gone, and read-aloud in AI results (words, examples and whole sentences) now uses the same cloud voice as dictionary entries, which is more consistent. The cost is that read-aloud is unavailable offline for now (the FAQ has been updated to match). The installer is noticeably smaller as a result.
“AI lookup” in Settings is now “AI service”, keeping only the service name and today’s quota; the explanatory text that no longer applied has been removed.
New Verb collocations tab in the reference tables: 133 collocations that Chinese speakers regularly stumble over, organised into 8 everyday domains (putting clothes on and off, accessories and things you carry, appliances, around the house, the kitchen, washing and grooming, getting about, communication) — 帽子をかぶる / 帽子を脱ぐ and the like. Tap to hear the whole sentence.
The “contains” query on the browse page now covers kana words and mixed kanji-and-kana words, not just single kanji.
Fixed “See all words containing X” occasionally going to the wrong page.
Cleaned up a batch of leftover English wording in Chinese definitions and spelling notes.

1.0.1 (2026-07-30)

Honest refusals from the AI. When the AI cannot establish what a query really means, it no longer invents a speculative definition; it says plainly that it cannot confirm, and offers a one-tap “Ask a person to explain it”. Looking the same thing up a second time shows the confirmed no-answer result straight away, with no waiting.
Search suggestions: when there are too many candidates and the list is truncated, a new “See all” link opens the full list, browsable as “starts with…” and “contains…”. Candidate ordering has also been improved to sit closer to real-world frequency, and fixes have gone in for grammar and Chinese candidate rows sometimes landing on the wrong result, and for homographic candidates being swallowed.
The heading of AI word results was missing its reading key, which left katakana entries unable to be read aloud; fixed.
Fixed a white edge down the right of the window in dark mode on macOS. The application menu’s “About” now opens the in-app About page directly.
Dark mode polish: category badges (AI / grammar / Chinese / JLPT and so on), the underline used by select-to-look-up and text selection highlighting have all had their dark colours and legibility tuned consistently.
Fixed the grading buttons in review going dead after un-favouriting a word.
Fixed the top of a second line of text occasionally showing through after the top-bar search box loses focus and collapses; the scrollbar no longer runs through the top bar.
Headings on AI grammar results can now be opened for details directly, without the extra “Details” button. Settings gained a note saying that read-aloud in AI results uses the local engine.
Overly long entry titles in the wordbook and history are no longer squashed and truncated; after picking a grammar item from the completion dropdown, that lookup is now marked correctly in the history and takes you back to the right grammar point.
Corrected register labels in some word-comparison cards and the readings of a few entries (流量 among them), and cleared out a batch of mismatched transitive/intransitive example pairs.

1.0.0 (2026-07-27)

Dark theme. Settings → General → Appearance offers System / Light / Dark. It follows the system by default and switches the moment the system does; starting up in dark mode no longer flashes a light interface first.
Read-aloud filled in for AI results: Japanese examples can be played as whole sentences, and word results now have a play button beside the kana reading.
After the dictionary data finishes downloading on a fresh install, the app no longer needs restarting — you can start looking words up right away.
Fixed people who already had the dictionary installed being sent back to the first-launch download page in certain cases.
Fixed the interface staying blank after startup in rare cases: the app now reloads the page itself to recover.
Corrected a batch of misread furigana on personal names, including the furigana generated on the fly during a lookup.
Word-comparison cards no longer mention words that are not shown in the results.