Discourse Content Localization

Discourse Content Localization

Use this before localizing dynamic Discourse content such as Sidebar sections, Groups, Docs, Events, or other models listed in content-localization status trackers.

First Decisions

• Core models: put the data model, serializers, Guardian checks, API behavior, and edit UI in core. Put automatic AI detection/backfill/localizer jobs in plugins/discourse-ai.
• Plugin models: do not create plugin-to-plugin dependencies casually. Create proper plugin APIs and plugin outlets / hooks.
• Manual editing localizations UI is part of calling a model “localizable”. Flag it upfront. Pick either an existing edit page/modal or a new route before implementing the backend.
• Hot-route safety is mandatory, e.g. on any discovery routes. Any serializer that may call get_localization must receive records with :localizations preloaded.

Core Model Pattern

• Add include Localizable to the source model and a nullable locale column with limit 20.
• Add <Model>Localization with include LocaleMatchable, belongs_to :model, a locale string limit 20, translated fields with source-model length limits, and a unique (model_id, locale) index.
• Use get_localization for exact, normalized, and optional default-locale fallback behavior. Do not reimplement locale fallback in serializers.
• Add ContentLocalization.show_translated_<model>? using SiteSetting.content_localization_enabled, source locale.present?, and !model.in_user_locale?. Preserve existing model-specific presentation gates; do not copy them to unrelated model types without explicit product intent.
• Preload in every list, boot, site JSON, admin list, and controller path that serializes localized fields. Use includes(:localizations) or includes(:localizations, children: :localizations) as needed.
• Serializer methods should return localization.field.presence || original_field and expose localization rows only to users who can edit them.

Editing UI And API

• Reuse an existing edit surface when it is the natural place users already manage the content. Otherwise add a dedicated route.
• Only users passing a Guardian localization check can edit localizations. For admin-owned/global content, prefer admin-only checks unless product intent says otherwise.
• Server-side params must accept localization rows and enforce the same max lengths as the model. UI validation is not sufficient.
• Existing source fields should remain editable independently from translations. Destroying a source record must destroy its localizations.
• Edit/settings surfaces must load source/default values, not localized display values. For example, a Japanese admin opening the "Support" category settings should see source name "Support", not its localized name "サポート".
• Preserve existing API response shape where possible; add localizations arrays only for authorized editors.

discourse-ai Backfill Pattern

• Add model-specific Candidate, Localizer, regular localize job, and scheduled backfill jobs in discourse-ai only.
• Candidate scopes should include only content intended for automatic translation. For global/admin content, avoid user-private rows unless explicitly requested.
• Localize only records with source locale present. Delete localizations whose locale matches the source locale. Skip target locales that normalize to the source locale.
  • an "en" post does not need "en" translation nor "en_GB" translation
• Respect DiscourseAi::Translation.enabled?, backfill_enabled?, configured agents, hourly rate, credits, and error logging patterns used by existing tag/category jobs.

Tests

• Model specs: validations, uniqueness, dependency cleanup, fallback via get_localization.
• Serializer/request specs: localized value when enabled, original value when disabled/source locale missing/same locale, editor-only localizations, and no N+1 on hot routes.
• When changing model-specific presentation gates, test authenticated and anonymous behavior and verify unrelated localizable model types remain unaffected.
• UI/system or QUnit specs: authorized editor can view/add/remove localization rows on the chosen edit surface; unauthorized users cannot.
• discourse-ai specs: detector text, localizer writes, scheduled job enqueue/credit gates, regular job limits/skips/errors, and candidate completion if the model appears in translation-progress UI.

Checklist

• Localization model has include LocaleMatchable, field length limits matching the source, and a unique (model_id, locale) index.
• Serializers preload localizations on hot/list path and fall back to source values without N+1s.
• Regular presentation payloads do not include localization rows for admins or non-admins.
• Admins can edit localizations for the model via a UI
• Edit/localization payloads give authorized admins default values plus localization rows. A Japanese admin opening the "Support" category settings should see source name "Support", not its localized name "サポート".
• Manual editing UI, API params, and tests cover adding, updating, and removing localization rows. Do not accept empty strings for fields, use frontend validations.
• Localization editing is authorized server-side, with admin-only checks for admin-owned/global content.
• discourse-ai detection/backfill stays in plugins/discourse-ai for core models and skips source-locale/self translations.
• Turning off the feature via SiteSetting.content_localization_enabled disables all localization behavior, including backfill and detection