Usage of AI localization with Crowdin branches

[amannn]

Hey @andrii-bodnar, I'm currently trying out the AI localization workflow!

So far, this is working for me, but one issue I see with this workflow is that translations are only generated on a push to main. This way, you don't get a preview of translations while working on a feature branch, and if a push to main means a deployment, you'll deploy with missing translations while a PR with translations is being added.

So the adjustments I'm trying to make are:

1. Trigger the action on every branch
2. Set create_pull_request to false and localization_branch_name to ${{ github.ref_name }} to get translations directly on feature branches
3. Set crowdin_branch_name to ${{ github.ref_name }} as well, so translation changes in feature branches don't interfere with production translations

This seems to be generally working for me, but the crowdin_branch_name setting seems to set up a Crowdin branch from scratch, therefore AI translation to target locales completely regenerates translations instead of using what's already available in the production translation set.

It seems like crowdin branch clone is what I'm looking for, but I'm unsure if this can be integrated with the Crowdin GitHub action.

Another aspect to consider is that I want Crowdin branches to be deleted once the branch in my VCS is deleted, I guess that would require another manual step …

Can you provide some guidance here on whether this workflow is possible currently with the Crowdin GitHub action? Or should users be encouraged to use the CLI manually once branches become necessary? The former would definitely be preferred from my perspective.

Many thanks for your help!

My current workflow file

name: Translations
on:
  push:
    paths: ['messages/**/*.json']
jobs:
  main:
    runs-on: ubuntu-latest
    permissions:
      contents: write
    steps:
      - uses: actions/checkout@v4
      - name: Upload messages
        uses: crowdin/github-action@v2
        with:
          crowdin_branch_name: ${{ github.ref_name }}
          upload_sources: true
        env:
          CROWDIN_PERSONAL_TOKEN: ${{ secrets.CROWDIN_PERSONAL_TOKEN }}
      - name: AI pre-translate
        uses: crowdin/github-action@v2
        with:
          command: 'pre-translate'
          command_args: '--method ai --ai-prompt=${{ vars.CROWDIN_PROMPT_ID }}'
        env:
          CROWDIN_PERSONAL_TOKEN: ${{ secrets.CROWDIN_PERSONAL_TOKEN }}
      - name: Download translations
        uses: crowdin/github-action@v2
        with:
          crowdin_branch_name: ${{ github.ref_name }}
          upload_sources: false
          download_translations: true
          localization_branch_name: ${{ github.ref_name }}
          create_pull_request: false
        env:
          CROWDIN_PERSONAL_TOKEN: ${{ secrets.CROWDIN_PERSONAL_TOKEN }}
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

[andrii-bodnar]

Hi @amannn!

I wanted to make this guide as simple as possible, so pushing to main was the simplest way of the trigger :). But there are a lot of possible options documented in the Triggers section.

The described strategy with the create_pull_request: false and localization_branch_name: ${{ github.ref_name }} seems an ideal setup. It will require separate workflows or jobs for uploading sources and downloading translations. To avoid the Pre-Translation of the strings pushed to the newly created branch in Crowdin, try to adjust the Duplicate Strings Management settings.

The crowdin branch clone command is only for String-based projects, and I believe it's not needed in this setup, the duplicate strings option should handle everything.

However, to automatically delete branches from Crowdin, some additional setup is indeed required. In this case, you can use the crowdin branch delete CLI command along with the command and command_args action inputs.

[amannn]

My impression was that the workflow waits for this step to finish. Is that correct @andrii-bodnar?

[andrii-bodnar]

@amannn right, this would be the simplest and most straightforward solution.

Although it is not possible to run crowdin pre-translte asynchronously, I think it would be possible to run this job with some short timeout and finish the workflow. Then, listen for the "Project if fully translated" Webhook, for example.

This will require additional setup described here. There will be two separate workflows: one to run Pre-Translate and another triggered by a Webhook to download translations and create a pull request.

[amannn]

Thanks for clarifying! If I understood the question from @jaumegui correctly, he was mostly concerned with the workflow waiting for AI translations to finish—so the out-of-the-box behavior seems to do the right thing here I think 👍

[andrii-bodnar]

I believe it makes sense to implement the Webhook approach only in the case of large projects where Pre-Translate could take significant amount of time. Otherwise, the out-of-the-box behavior should suffice 🙂

[jaumegui]

Thanks a lot for all your answers!