Skip to main content
Phrase, Smartling, and Lokalise CLIs reach file-mode and discovery parity with Crowdin for many workflows. hl run can route strings through Google or DeepL MT profiles, segment text with SRX 2.0, and execute MT in bounded batches. sync push and sync pull honor storage.adapter in i18n.yml, and hl check warns on escaped-character regressions plus optional per-locale spellcheck allow-lists. View the GitHub release · Full changelog

​
Breaking changes

  • Configs that define both translation (including translation.rules) and legacy llm.rules fail validation. Move routing rules to translation.rules before enabling MT routing. (#2294)
  • TMS sync pull and sync push through storage adapters treat locale files without provenance metadata (flat JSON with no .meta.json sidecar, origin=unknown) as conflicts (unknown_provenance_mismatch) instead of silently overwriting local or remote values. Use --force-conflicts on push when you intend to override. (#2385)

​
Features

  • Add Phrase file-mode parity: phrase init, phrase push / phrase pull, phrase locales list, upload --wait with terminal-state polling, and phrase uploads cleanup --id for unmentioned keys.
  • Expand hyperlocalise smartling with translation file import (upload translations), discovery (files list / status, locales list, projects list / info), file rename/delete, download sources|translations --all with official name_locale.ext paths, optional download locales, retrieval type (pending / published / pseudo), upload directives, and glossary list/create/import/download. (#2248, #2264, #2274)
  • Add hyperlocalise lokalise: upload translations (requires --target-locale), locales list, files list, and --poll / --poll-timeout on uploads to wait for Lokalise process completion. (#2330)
  • Route sync push / sync pull through storage.adapter when Hyperlocalise Cloud sync is not configured; native Hyperlocalise sync still wins when both blocks are present.
  • Add translation routing and typed mt.profiles so hl run can select LLM or MT engines per group; MT image tasks are rejected at planning time.
  • Add an MT engine factory for Google Translate and DeepL (credentials from configured environment variables) and bounded MT batch execution with retries, cancellation, and partial-success behavior aligned with the existing LLM worker path.
  • Add native SRX 2.0 segmentation for hl run via buckets.*.files[].srx or run --srx, with span keys, lock fingerprints, and skips for already-segmented formats and ICU/printf/Fluent placeholders.
  • Add escaped_char_mismatch warnings to hl check when translations introduce literal escapes or decoded control characters not present in the source; supports --quiet, --fix, and --diff-stdin.
  • Add spellcheck.dictionary_dir in i18n.yml and hl check --dictionary-dir for per-locale Hunspell allow-list files ({locale}.txt).

​
Fixes

​
Performance

  • Java Properties parsing is about 28% faster (632 µs → 453 µs/op) with allocations dropping from 2,044 to 19 allocs/op when escapes are absent.
  • Eval scoring token counting is about 2.3× faster (41.5 µs → 18.0 µs/op) with lower heap use in BenchmarkEvaluatorEvaluate.
  • Scoring evaluator tag analysis avoids strings.ToLower and strings.Builder allocations for typical HTML/ICU tokens.
  • Android XML parsing is about 7–8% faster by lazily computing line numbers only for errors and <plurals> metadata.
  • JSONC comment parsing is about 10% faster (874 µs → 788 µs/op) with 301 fewer allocations per operation.

​
Tests

  • Phrase init/push/pull/wait/cleanup, Smartling import/discovery/download-all/glossary flows, and Lokalise upload/list/poll command suites.
  • Storage-adapter sync push / sync pull, unknown-provenance conflict regressions, and Lokalise adapter live/httptest paths.
  • MT profile validation, engine factory credential errors, bounded MT batch execution, and SRX segmentation/join round-trips.
  • escaped_char_mismatch and spellcheck dictionary directory resolution for hl check.

​
Maintenance

  • Bump google.golang.org/grpc from v1.83.1 to v1.83.2. (#2306)