Continuous integration#

A translation pipeline normally has three independent gates:

source files are formatted;
source files pass lint;
committed generated files match the current catalogs and generator.

Keep those checks separate enough that a failure says what changed.

Verify committed output#

When generated files are committed, regenerate them and fail on a diff:

globetrotter format --check
globetrotter lint
globetrotter
git diff --exit-code -- generated/

Replace generated/ with the output paths from your config. Running generation before git diff tests the same config developers use locally; it does not rely on a second list of expected keys.

GitHub Actions#

This job installs the published CLI. A repository that builds Globetrotter itself should use its working-tree binary instead.

name: translations
on:
  pull_request:
  push:
    branches: [main]

jobs:
  translations:
    runs-on: ubuntu-24.04
    timeout-minutes: 10
    steps:
      - uses: actions/checkout@v7
      - uses: dtolnay/rust-toolchain@stable
      - name: Install globetrotter
        run: cargo install --locked globetrotter-cli
      - name: Check formatting
        run: globetrotter format --check
      - name: Lint translations
        run: globetrotter lint
      - name: Verify generated files
        run: |
          globetrotter
          git diff --exit-code -- generated/

Pin the CLI version when output stability across tool upgrades matters:

cargo install --locked globetrotter-cli --version 0.0.12

Update the pin and committed generated files together.

Generate instead of commit#

If generated files are build artifacts, run Globetrotter immediately before compiling the consumer:

- name: Generate translations
  run: globetrotter
- name: Build
  run: cargo build --locked

Cache compiler output, not generated translations. The generation step is small and should remain visible in the build log.

LLM review in CI#

Keep --llm-judge outside a required correctness gate. It depends on an external model, is advisory, and intentionally does not fail lint for semantic findings. A separate review job can surface suggestions without making deterministic checks depend on model availability.