Recipes

Open a pull request after each run

- id: translate
  uses: IEvangelist/resource-translator@v3
  with:
    sourceLocale: en
    subscriptionKey: ${{ secrets.TRANSLATOR_KEY }}
    endpoint: ${{ secrets.TRANSLATOR_ENDPOINT }}
    region: ${{ secrets.TRANSLATOR_REGION }}

- if: steps.translate.outputs.has-new-translations == 'true'
  uses: peter-evans/create-pull-request@v7
  with:
    branch: machine-translation
    title: ${{ steps.translate.outputs.summary-title }}
    body: ${{ steps.translate.outputs.summary-details }}
    labels: localization

Monorepos: scope translation per package

with:
  sourceLocale: en
  include: |
    apps/web/**/*.en.resx
    packages/shared/**/*.en.json
  exclude: |
    **/dist/**
    **/__tests__/**

Glossary for brand names and product terms

Drop a .github/resource-translator.yml into the repo:

glossary:
  Acme: Contoso
  Octocat: Octocat
  ".NET": ".NET"
  "C++": "C++"

Glossary terms are applied after translation, so the machine-translated output keeps your preferred branding even when the model swapped it for a synonym.

Custom Translator category (industry / domain)

with:
  sourceLocale: en
  subscriptionKey: ${{ secrets.TRANSLATOR_KEY }}
  endpoint: ${{ secrets.TRANSLATOR_ENDPOINT }}
  categoryId: 1234abcd-5678-90ef-...

Train an Azure Custom Translator category on a corpus that captures your industry vocabulary and tone (legal, medical, marketing, internal-engineering, etc.) and pass its ID via categoryId. This is the strongest tone lever Azure offers.

Tone control + strict fallback

# .github/resource-translator.yml
# Industry / domain — choose your tone via a Custom Translator category
categoryId: legal-en

# Treat profanity strictly for content aimed at younger audiences
profanityAction: Marked
profanityMarker: Tag

# Fail the run rather than silently fall back to the general model
allowFallback: false

Combine a Custom Translator categoryId with profanityAction and allowFallback: false to enforce a consistent voice and fail the workflow if the chosen category does not have a deployment for one of your target locales.

Dry run before turning the workflow on

with:
  dryRun: true
  failOnError: false

The action still emits summary outputs and a step summary, but no files are written. Combine with failOnError: false to surface every issue as a warning while you iterate.