heeng keyboard translator

heeng-keyboard-translator

You meant to type Hebrew, the layout was still English, and שלום came out as akuo. This plugin notices and offers the repair before Claude acts on the garbled text.

/plugin marketplace add roypadina/padina-claude-code-plugins
/plugin install heeng-keyboard-translator@padina

Needs only Python 3, which macOS ships.

Why per word

The obvious version of this tool translates the whole message in one direction. That handles the easy case and misses the common one: an Israeli developer mixing Hebrew and English in one sentence where a single layout-switch keystroke didn't land, leaving one or two broken tokens in an otherwise fine sentence.

This plugin analyses each word independently, so all three cases work:

Case	You typed	It reconstructs
Whole message in the wrong layout	akuo	שלום
One bad token in a good sentence	שלום, akuo חבר	שלום, שלום חבר
Every switch inverted — different directions in one sentence	akuo, ש אםהק כםםג	שלום, ש אםהק food
Nothing wrong	hello world	unchanged, and the skill stays silent

How a word is judged

For each token the translator computes both readings — as-if-English-typed-in-Hebrew and the reverse — and scores them against bundled common-word sets (140 Hebrew, 123 English, weighted toward everyday and developer vocabulary).

• High confidence — the original, or exactly one translation, is a real word. It picks that one.
• Low confidence — neither reading is a known word. It leaves the token alone and reports both candidates so you can decide.

Tokens of two characters or fewer are too ambiguous to call. The script still scores them and returns them at confidence: low with the original kept — skipping them is a rule in the skill, not a filter in translate.py.

What it will not do

• It never substitutes silently. It shows the original, the reconstruction and a per-word diff, then asks via AskUserQuestion whether to use it, edit it, or keep the original.
• It does not fire inside fenced code blocks, file paths, URLs or identifiers, where a "correction" would be corruption.
• It does not trigger on text that is simply recognisable in either language.

Using the translator directly

The script is useful on its own:

python3 skills/keyboard-layout-fix/scripts/translate.py --per-word "שלום, akuo חבר"

{
  "original": "שלום, akuo חבר",
  "reconstructed": "שלום, שלום חבר",
  "fix_count": 1,
  "words": [ { "word": "akuo", "candidates": {...}, "choice": "en_to_he", "confidence": "high" } ]
}

Without --per-word it translates the whole string both ways and scores each direction.

Layout

.claude-plugin/plugin.json
skills/keyboard-layout-fix/
  SKILL.md                     triggers, workflow, boundaries
  keyboard-mapping.json        bidirectional US QWERTY ↔ Israeli SI-1452 map
  scripts/translate.py         per-word translator + plausibility scoring

Known limits

• Hebrew has no letter case, so uppercase Latin maps to the same Hebrew letter as lowercase.
• Only US QWERTY ↔ Israeli SI-1452. Another layout pair needs a new mapping file — the map is plain JSON, so adding one is mostly data entry.
• The common-word sets are small by design. They are a plausibility signal, not a dictionary; the uncertain cases are handed to you rather than guessed.