locale-ai 0.6.1 → 0.6.2

package/README.md

@@ -31,7 +31,7 @@ Download 577 MB? (Y/n)
 - **Layouts**: `locales/en/*.json` (nested folders too) → `locales/fr/...`, or `messages/en.json` → `messages/fr.json`. `.json`, `.yaml`, `.yml`, `.po` (gettext), `.arb` (Flutter) and `.strings` (Apple) are all supported, mixed freely in the same project. Common folders (`locales`, `src/locales`, `public/locales`, `src/i18n`, `messages`, `lang`, ...) are found automatically, or pass `--input`.
 - **Placeholders are never translated or lost**: `{name}`, `{{count}}`, `%s`, `%1$s`, `%(name)s`, `%{name}`, `${name}`, `:attribute`, `$t(key)`, `@:key`, `<b>…</b>`, `<0>…</0>`, `&nbsp;`, URLs, e-mails.
 - **ICU MessageFormat**: `{count, plural, one {# file} other {# files}}`, `select`, nesting. Only the text is translated.
-- **i18next plurals**: adds the forms the target language needs (`_few`, `_many` for Arabic, Russian, Polish...).
+- **i18next plurals**: adds the forms the target language needs (`_few`, `_many` for Arabic, Russian, Polish...), so the app doesn't fall back to the English string — each added form copies the `_other` translation verbatim rather than inflecting it, so review and adjust the wording of `_zero`/`_two`/`_few`/`_many` by hand for languages where those forms use a different noun/adjective form.
 - **Incremental**: only missing keys are translated, existing translations are kept, re-runs are no-ops. `--overwrite` re-translates everything.
 - **Structure**: key order, numbers, booleans, arrays, indentation and trailing newline are preserved.
 
@@ -157,3 +157,4 @@ npm run build       # → dist/
 ## Next steps
 
 1. **More languages**: ja, ko have no usable OPUS-MT model yet; a different model family would be needed.
+2. **Rare batch-dependent mistranslation**: translating a short string (e.g. "© 2026 Acme Inc. All rights reserved.") can occasionally produce an unrelated/garbled result depending on which *other* strings share its translation batch — the same string translates correctly alone or in a different batch. Root cause: the built-in models are dynamically int8-quantized per-tensor, so the quantization scale for a batched activation tensor is computed from the whole batch, not per row; most of the time this is just quantization noise, but occasionally it's enough to tip beam search into a bad output with no warning. A fix would mean re-quantizing the models with static per-channel scales (or giving up CPU-batching speed) — out of scope for a quick patch; flagging here so it's a known, understood limitation rather than a silent one.

package/dist/engine.js

@@ -1,4 +1,5 @@
 import { extractIcu } from "./icu.js";
+import { extractPipePlural } from "./vue-plural.js";
 import { fixTagSpacing, hasTranslatableText, mapOutsideTokens, mask, numericStyleSafe, splitOnTokens, stripMarkers, unmask, } from "./placeholders.js";
 import { splitSentences } from "./segment.js";
 const letters = (s) => stripMarkers(s).replace(/[^\p{L}]/gu, "");
@@ -207,8 +208,8 @@ export async function translateStrings(strings, hops, opts) {
             remaining.push(s);
     }
     const plans = remaining.map((s) => {
-        const icu = extractIcu(s);
-        return { source: s, icu, units: (icu?.units ?? [s]).map((text) => ({ text, owner: s })) };
+        const split = extractIcu(s) ?? extractPipePlural(s);
+        return { source: s, split, units: (split?.units ?? [s]).map((text) => ({ text, owner: s })) };
     });
     const units = plans.flatMap((p) => p.units);
     let pending = units;
@@ -220,7 +221,7 @@ export async function translateStrings(strings, hops, opts) {
         await splitFallback(pending, hops, opts, progress, issues);
     for (const p of plans) {
         const parts = p.units.map((u) => u.result ?? u.text);
-        translations.set(p.source, p.icu ? p.icu.rebuild(parts) : parts[0]);
+        translations.set(p.source, p.split ? p.split.rebuild(parts) : parts[0]);
     }
     return { translations, issues };
 }

package/dist/locales.js

@@ -51,7 +51,7 @@ export function detectLayout(cwd, from, input) {
             return { kind: "file", root: path.dirname(abs), ext: directExt };
         const l = layoutAt(abs, from);
         if (!l)
-            throw new Error(`No "${from}" locale found in ${input} (looked for ${from}/*.json and ${from}.json/.yaml/.yml).`);
+            throw new Error(`No "${from}" locale found in ${input} (looked for ${from}/* and ${from}${EXTENSIONS.join("/")}).`);
         return l;
     }
     for (const c of CANDIDATES) {
@@ -59,7 +59,7 @@ export function detectLayout(cwd, from, input) {
         if (l)
             return l;
     }
-    throw new Error(`Could not find "${from}" locale files. Use --input <dir> to point at the folder containing ${from}/ or ${from}.json/.yaml/.yml.`);
+    throw new Error(`Could not find "${from}" locale files. Use --input <dir> to point at the folder containing ${from}/ or ${from}${EXTENSIONS.join("/")}.`);
 }
 function listLocaleFiles(dir) {
     const out = [];

package/dist/vue-plural.js

@@ -0,0 +1,38 @@
+/**
+ * vue-i18n's plural syntax: forms separated by a literal `|`, chosen by count at
+ * runtime (`"no items | one item | {count} items"`). The `|` carries no meaning to
+ * a translation model — it reads as stray punctuation and gets dropped or garbles
+ * the sentence around it — so each form is translated on its own and the `|`
+ * separators are kept untouched, exactly where they were.
+ */
+/** Returns undefined when the string has no top-level `|` (ignoring one inside `{...}`). */
+export function extractPipePlural(s) {
+    const segments = [];
+    let depth = 0;
+    let start = 0;
+    const push = (end) => {
+        const seg = s.slice(start, end);
+        const m = /^(\s*)([\s\S]*?)(\s*)$/.exec(seg);
+        segments.push({ lead: m[1], core: m[2], trail: m[3] });
+    };
+    for (let i = 0; i < s.length; i++) {
+        const c = s[i];
+        if (c === "{")
+            depth++;
+        else if (c === "}")
+            depth = Math.max(0, depth - 1);
+        else if (c === "|" && depth === 0) {
+            push(i);
+            start = i + 1;
+        }
+    }
+    if (start === 0)
+        return undefined; // no top-level '|'
+    push(s.length);
+    if (segments.some((p) => !p.core))
+        return undefined; // "a||b" or a leading/trailing '|': not a plural list
+    return {
+        units: segments.map((p) => p.core),
+        rebuild: (translations) => segments.map((p, i) => p.lead + translations[i] + p.trail).join("|"),
+    };
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "locale-ai",
-  "version": "0.6.1",
+  "version": "0.6.2",
   "description": "Translate your i18n locale files locally on CPU. Free, offline, no API keys.",
   "author": "Taha KHABOUSS",
   "repository": {