en I18n_ci

i18n CI workflow

.github/workflows/i18n.yml regenerates src/i18n/locales/auto/*.auto.json from src/i18n/locales/en.ts on every change to the English source (or the generator pipeline). Output lands as a PR — never committed direct to main — so a human eyeballs machine-translation quality before it ships.

Triggers

• Push to main that touches any of:
  • src/i18n/locales/en.ts
  • src/i18n/glossary.json
  • scripts/generate_auto_locales.py, i18n_backends.py, i18n_plurals.py, i18n_config.py, i18n_export.mjs
  • .github/workflows/i18n.yml
• workflow_dispatch — manual trigger, optional targets=ru,pl input for narrow reruns. Release-time regeneration uses this path (tags don't trigger the workflow directly; see the YAML comment).

Step-by-step

1. Checkout + setup Python 3.11 (with pip cache) + Node 20 (npm cache).
2. Install Python deps — argostranslate>=1.9 and Babel>=2.14. Full requirements.txt is intentionally skipped here because it pulls torch / ctranslate2 / cuda which would blow out the CI disk.
3. Install minimal Node deps — npm ci --ignore-scripts for tsx.
4. Restore Argos package cache from ~/.local/share/argos-translate, keyed on hashFiles('scripts/i18n_config.py'). Adding / removing a language invalidates the cache cleanly. Cold start ~10 min (14 lang packs × ~30 MB each). Warm rerun ~2 min.
5. Export en.ts → JSON via npx tsx scripts/i18n_export.mjs en.
6. Run the generator with --backend argos over the exported JSON.
7. Key-parity sanity check — every auto bundle must cover every source key (modulo plural expansion which adds _one/_few/etc.). Fails the job with ::error lines listing the first 5 missing keys per offending bundle.
8. Detect changes — skip PR creation if nothing actually diffed.
9. Create / update PR via peter-evans/create-pull-request@v6 on branch i18n/auto-update. PR body has a checklist for the reviewer.

Local reruns

You don't need the CI to regenerate locally — everything the workflow does is runnable on a dev machine:

# One-off regenerate all 14 auto-languages using the real Argos backend.
# Works only with Python 3.9+ where argostranslate installs cleanly.
pip install 'argostranslate>=1.9' 'Babel>=2.14'
npx tsx scripts/i18n_export.mjs en > /tmp/en.json
python scripts/generate_auto_locales.py \
    --source /tmp/en.json \
    --output src/i18n/locales/auto \
    --backend argos

# Single-language regenerate (e.g. after fixing a glossary entry):
python scripts/generate_auto_locales.py \
    --source /tmp/en.json \
    --output src/i18n/locales/auto \
    --targets ja \
    --backend argos \
    --force

# Dry-run (useful for pipeline debugging):
python scripts/generate_auto_locales.py --dry-run

The first real run downloads Argos language packs automatically — plan for ~10 min and ~500 MB of disk usage.

Overriding targets

To regenerate only a subset via workflow_dispatch:

1. Actions tab → "Auto-translate locales" → "Run workflow"
2. Branch: main
3. targets: comma-separated codes, e.g. ru,pl,cs
4. force: true if the source hash hasn't changed but you still want a re-run (rare — usually only when fixing a generator bug).

Debugging a failed run

Parity check failed — generator silently dropped some keys. Most likely a placeholder like {{count}} got stripped by Argos. Fix by adding the offending term / pattern to src/i18n/glossary.json and rerunning.

Cache miss on warm run — check whether scripts/i18n_config.py was touched. Every change to that file invalidates the cache intentionally.

PR not opening — check permissions: in the workflow. Needs contents: write + pull-requests: write. GitHub occasionally tightens defaults; if you see "resource not accessible by integration" errors, re-read the workflow permissions block.

Disabling / pausing

Short-term: comment out the on.push stanza and rely on workflow_dispatch only. Long-term: delete the workflow file — locale output is committed, so disabling CI doesn't break the SPA, it just stops automatic regeneration.