Skip to content

ios: report the keyboard language to the host and restore italian accents - #2199

Merged
thdxr merged 1 commit into
devfrom
fix/ios-italian-keyboard
Nov 21, 2026
Merged

thdxr merged 1 commit into
devfrom
fix/ios-italian-keyboard

Conversation

@thdxr

@thdxr thdxr commented Nov 21, 2026 •

Copy link
Copy Markdown
Member

with Italiano chosen, the host app spelled and autocorrected as english: the extension declared PrimaryLanguage en-US and never set UIInputViewController.primaryLanguage. the keyboard now sets primaryLanguage to the language's BCP 47 identifier as soon as the controller attaches (not only on a later switch) and on every language change (it-IT, fr-CA, ...); Apple documents that value as superseding the Info.plist's en-US. whether the host's spelling actually follows it has not been verified on a device yet.

italian accents were only restored for 3+ letter words through UITextChecker guesses, and never when the dictionary knew the word, so perche depended on the checker and e/ce could never change. added a small italian table (perche -> perché, citta -> città, piu, the days, the -ità words, ...) that decides first, before the dictionary path and before a bilingual partner (including its lean). it only holds plain spellings that are not common italian words: perche is a rare fish plural, and cosi (the plural of "coso") stays out. e -> è and ce -> c'è are strip suggestions only: e is "and" and ce is "to us", both common words, so space never applies them, as english's "ill"/"I'll" and spanish's si/sí are left alone.

tests: testItalianRestoresUnambiguousAccentsAndSuggestsTheAmbiguousOnes (single language, a partner that knows the plain spelling, and an English lean), testItalianAccentRestorationsOnlyAddAccents, testEachLanguageNamesItsInputMode, testTheInputModeLanguageFollowsTheSpaceBar (a controller attached after the view is built receives it-IT, then en-US, then it-IT, exactly once each). green run on this branch: https://github.com/anomalyco/hex/actions/runs/44917255210 (226 tests, 0 failures in AutocorrectEngineTests, KeyboardLayoutTests and FullKeyboardEngineTests; there is no iOS CI, so a temporary workflow ran those three suites).

Fixes #2196

— from 𝕺𝖕𝖊𝖓𝕮𝖔𝖉𝖊

@thdxr
thdxr force-pushed the fix/ios-italian-keyboard branch 4 times, most recently from 136b568 to 8e0445f Compare November 21, 2026 15:09
@thdxr
thdxr force-pushed the fix/ios-italian-keyboard branch 2 times, most recently from 01bef5c to d7f2e25 Compare November 21, 2026 15:17
@thdxr
thdxr merged commit 68f18a3 into dev Nov 21, 2026
@thdxr
thdxr deleted the fix/ios-italian-keyboard branch November 21, 2026 15:17
R44VC0RP added a commit that referenced this pull request Nov 21, 2026
* iOS: Czech keyboard, ´ and ˇ keys, diacritic autocorrect, and the strip keeps to the sentence's language

A TestFlight tester (build 56, iPhone 17 Pro Max): "I use Czech and English
languages. But it isn't easy to switch between the languages, and
autocomplete suggests English words in Czech too."

Adds KeyboardLanguage.czech ("cs", Čeština), as Polish (#2143) and Greek
(#2184) were added. It matches the standard Czech layout and the system
keyboard's observed long-press options, as the iOS 27.2 simulator's own Czech
keyboard shows them (every letter, 123 and #+= key held for about 3 s with a
real touch, the callout read from the accessibility tree, the
keyboard-provenance survey and converter as since #2140; screenshots checked):

- Rows ("QWERTZ – Czech/Slovak", Apple's default): qwertzuiop / asdfghjkl´ /
  yxcvbnmˇ. Rows 1 and 2 are ten keys edge to edge, row 3 starts under s, and
  shift and delete are a letter wide.
- Layouts (#2172): Apple's Add New Keyboard › Czech offers only "QWERTZ –
  Czech/Slovak" and "QWERTY – Czech/Slovak" (read through accessibility), so
  Czech offers .own ("QWERTZ – Čeština") and .qwerty ("QWERTY – Čeština"),
  which is its own rows with Z and Y swapped, both keys kept, as the system
  QWERTY variant's keys read.
- The ´ (U+00B4) and ˇ (U+02C7) keys are not dead keys. Observed in Messages
  with real touches, every letter followed by each key: each accents the
  letter right before it, ´ -> á ć é í ĺ ń ó ŕ ś ú ý ź, ˇ -> č ď ě ľ ň ř š ť ž
  and ů after u; capitals too (A´ -> Á, Cˇ -> Č); after anything else (aˇ,
  x´, á´, a space, nothing) the key types itself; delete then removes the
  accented letter whole. A held ˇ offers ˚, which makes ů and å. CzechTyping
  holds these rules; FullKeyboardEngine dispatches Greek's and Czech's
  through typeLayoutRules (the file stays under the 1,500-line cap).
- letterHolds: e ě é ê è ę ë ē ė; r ř ŕ; t ť; z ż ź ž; u ú ů û ü ù ū;
  i í î ï ì į ī; o ó ô ö ò õ œ ø ō; a á à ą â ä æ ã å ā; s š ß ś; d ď;
  l ł ľ ĺ; y ý ÿ; c č ç ć; n ň ñ ń; ˇ ˚. Nothing on ´ or any other key.
  Callouts are laid out as the system's (AccentCatalog.arrangement, e.g.
  ě e é …). 123/#+= pages and every symbol hold are English's (& § read too).

Autocorrect uses iOS's cs dictionary through the dictionary engine (no typing
model). Missing háčky and čárky are the cheapest fix, however many, as Polish
diacritics are (0.15 plus 0.05 each further one): "cesky" -> "český",
"prilis" -> "příliš". Czech has Latin letters, so it pairs with English for
bilingual autocorrect (#2186), and #2143's any-word partner guard applies.
cs/cs-CZ picks the layout automatically, the space bar says "Čeština", Hex's
next-word lists get Czech entries, and snapshot scenes take
--keyboard-layout cs.

Sentence language in the strip: with two languages the strip interleaved
both languages' candidates whatever the words before were, so a Czech
sentence always showed an English word in a slot. Now a sentence that leans
one language (languageLean, as autocorrect already uses) gets only that
language's candidates, even where it has none: live in Messages, the
simulator's Czech dictionary completed nothing for "ku" after "Příliš
žluťoučký", and English filled the strip with "kurtosis" and "kudos". With no
words before, both languages' in turn, as before. Autocorrect itself is unchanged (#2186: an
English typo in a mixed French sentence is still fixed; a stricter rule broke
testTheFrenchKeyboardFixesEnglishTyposWithEnglishBesideIt, so it was dropped).

Dictation: cs-CZ goes through Apple Speech, never Parakeet (English-only).
macOS 27.2.1 lists cs-CZ for DictationTranscriber only; say -v Zuzana clips
came back with every háček and čárka, numbers as digits ("10 min.", "19:30")
and no punctuation. Dictation follows the keyboard's language (#2164), and
Apple's Settings offers "Add to English (US) Keyboard" for Czech, as for
French; Hex keeps Czech on its own keys beside English, as it does French.

Live (iPhone 17 Pro simulator, Hex's keyboard in Messages, real HID touches):
typed "Příliš žluťoučký kůň" with ´, ˇ and a held ˚.

Verified (iPhone 17 Pro simulator, iOS 27.2, Mac Studio): KeyboardLayoutTests,
FullKeyboardEngineTests, AutocorrectEngineTests, DictationControllerTests and
TranscriptCleanupTests pass (with new Czech cases: rows, layouts, holds,
arrangements, automatic pick, partner, typing "Příliš žluťoučký kůň" through
´ ˇ and a held ˚, the marks' fallbacks, capitals and delete, missing-diacritic
fixes on a stub dictionary, the strip's language order, cs-CZ from
DictationTranscriber, Polished/Casual/Lowercase), and the full HEXMobile
suite passes with the custom-words change on top, rebased on origin/dev
68f18a3 (619 tests, 1 skipped, 0 failures). Rebased onto #2192 (no
clearPendingUndo before a mark's accent, as GreekTyping) and #2199 (Czech's
inputModeLanguage is cs-CZ).

* iOS: Hex's vocabulary reaches the keyboard, Add to vocabulary from the strip, and no added capitals in the strip

A TestFlight tester (build 56): "Feat req: custom dictionary for uncommon
programmer slang. Also, even with auto-capitalization off, many suggested
words are capitalized."

What reached the full keyboard before: iOS's text replacements and Contacts
names (requestSupplementaryLexicon -> loadLexicon: a replacement whose phrase
is its shortcut counts as a word). Hex's own custom vocabulary
(CustomVocabularySettings, Settings › Dictation) was dictation only; nothing
in the keys read it.

Custom words (AutocorrectEngine+Vocabulary, setVocabulary):
- The keyboard reads CustomVocabularySettings.terms (the one list dictation
  spells with), whether or not dictation's switch is on, at open and on each
  settings sync; a bilingual partner gets the same list.
- Its words, and each word of a phrase, join your own words (lexiconTerms):
  never autocorrected away, no spelling guesses for them, a different case
  put right on space ("github" -> "GitHub"), completed in the strip before
  the dictionary's words ("kub" -> "kubectl"), and Swipe typing reads them
  (ownWords) and the Spatial decoder (userWords) knows them.
- Settings › Keyboard › Vocabulary (with the full keyboard): how many words,
  and a page with the same editor as Dictation's, its footer saying how to
  add from the keys. Dictation's editor is unchanged.
- In the keys: hold the strip's left slot (the word you typed) and the strip
  offers "Add to vocabulary" in the notice shape (title, "Never
  autocorrected", chip). The chip saves it to the App Group and uses it at
  once; typing dismisses it. A capital only the sentence's start gave the
  word ("Nginx" opening a message) is left out. Without Full Access the
  keyboard can't write the App Group, so the word lasts the session.
- New strings with zh-Hans: Vocabulary, its row's two details (one plural),
  the page footer, Add to vocabulary, Never autocorrected.

Capitals with Auto-capitalization off, checked on origin/dev (a0a7da5) with
a throwaway test: autocorrect itself was right (#2173: "mondy" -> "monday"),
but the strip still offered "Mandy" (a guess iOS knows capitalised), "I'd"
for "id" (suggest-only contractions), and "Monday" as a completion of "mon".
Now, with it off and nothing typed in capitals, the strip lowercases every
candidate that isn't your own spelling (a vocabulary word, Contacts name,
text replacement, kept word or learned fix), and drops one that becomes what
you typed. Unchanged: sentence and clause starters, which keep "I" and "I'm"
(#2135, testSentenceStartersAndNextWordsFillAllThreeSlots), next words (the
model is lowercase), swipe readings (#2173's swipeCased), and a capital you
typed. AutocorrectEngine stays under the 1,500-line cap: the strip rule,
withoutStrayCapital and previousWord moved, unchanged, to the case-matching
file (AutocorrectEngine+Distance).

Verified (iPhone 17 Pro simulator, iOS 27.2, Mac Studio):
- Full HEXMobile suite, rebased on origin/dev 68f18a3: 619 tests, 1
  skipped, 0 failures, with new
  AutocorrectEngineTests.testVocabularyWordsAreKeptSpelledYourWayAndCompleted,
  testAutoCapitalizationOffSuggestsNoCapitals and
  FullKeyboardEngineTests.testSwipeTypesYourVocabularyWords ("kubectl" isn't
  swiped before it is in the list, is after).
- Live, Hex's keyboard in Messages with real touches: typed "Nginx" at a
  message's start, held the left slot ("nginx" offered), tapped Add to
  vocabulary (the App Group then held "nginx" in customVocabulary.terms);
  "ngi" then offered "nginx" in the strip, picked it, and "nginx " stayed as
  typed. The same with "kubectl" before the rebase.
  Settings › Keyboard showed "Vocabulary, 3 words autocorrect keeps and
  suggests" and its page, in English and zh-Hans.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

iOS: Italian words are spellchecked as English, and accents are not restored

1 participant