README

Magento 2 module that detects untranslated strings at runtime, reports them in an admin grid and can translate them automatically with Claude (Anthropic API) through Magento's message queue.

A spiritual successor to the Magento 1 extension EW_UntranslatedStrings.

How it works

1. Detect — a plugin on Magento\Framework\Phrase\Renderer\Translate checks every rendered phrase against the loaded translation dictionary (i18n CSVs, theme translations and database translations) for the current store locale. Strings without a translation are buffered per request and flushed to the database in a single query at shutdown.
2. Report — findings appear under Reports → Untranslated Strings: filterable, sortable and exportable, with occurrence counts, store view, locale and the URL where the string was first seen.
3. Translate — select rows and use the Translate with Claude mass action. The rows are queued (Magento message queue, DB transport — no RabbitMQ required) and a consumer sends them to the Anthropic API in batches. Successful translations are written to Magento's native translation database table (store-scoped) and go live after the automatically-invalidated translate / full_page caches are refreshed. Failed rows get status Error with the error message in the grid; re-select them to retry.

Installation

composer require elgentos/magento2-untranslated-strings
bin/magento module:enable Elgentos_UntranslatedStrings
bin/magento setup:upgrade

Configuration

Stores → Configuration → Advanced → Untranslated Strings

Detection adds a small overhead per rendered phrase — enable it per store view, ideally on development or staging while you browse/crawl the shop, then review the report.

The queue consumer

Translation runs asynchronously through the consumer elgentosUntranslatedStringsTranslate. With Magento's default consumers_runner cron configuration it starts automatically. To run it manually:

bin/magento queue:consumers:start elgentosUntranslatedStringsTranslate

Notes

• Translations are stored in the core translation table, so they survive deployments and can be reviewed/edited with any DB-translation tooling.
• The Claude prompt instructs the model to preserve placeholders (%1…%9, %s, %d, named placeholders), HTML and punctuation, and to keep brand/product names untranslated. Structured output (JSON schema) guarantees a parseable response.
• Detection never breaks rendering: all failure paths are caught and logged.

License

MIT