Skip to main content
hl run records MT work in lock and cache identity, and writes character usage, request counts, and duration into the --output JSON report. entries --srx splits string values with the same SRX 2.0 templates as run. Starter i18n.yml and CLI docs cover Google and DeepL routing. Planning and parsers are faster for multi-locale runs, plain ICU text, Liquid files, and segment validation. View the GitHub release · Full changelog

​
Breaking changes

None.

​
Features

  • Track MT identity in lock and remote-cache keys (translation type, MT provider, and profile), with backward-compatible matching for existing LLM lock entries. MT tasks that still have a type-less 1.12.0 lock hash are re-run so LLM and MT work cannot collide.
  • Add MT usage to --output JSON: sourceChars, translatedChars, requestCount, and durationMs at run, locale (localeMTUsage), and profile (mtUsageByProfile) levels, plus translationType on failures and batches. Character usage is JSON-only (not the stdout summary). Tracing includes translation type and provider.
  • Document MT configuration and routing (Google/DeepL credentials, llm.rules → translation.rules migration, batching, retries, reporting, and LLM-only features). The starter i18n.yml from hyperlocalise init stays LLM-only, with commented mt / translation examples.
  • Add entries --srx (default, html, markdown, or an SRX file path) so hl entries can split string values with SRX 2.0 the same way hl run does, including Markdown/MDX documents. Already-segmented formats such as strict FormatJS defaultMessage objects stay unsplit.

​
Fixes

None.

​
Performance

  • Task planning reuses precomputed source-text and source-context hashes instead of recomputing SHA-512 once per target locale.
  • ICU ParseInvariant skips AST work for plain text with no {#<}'} signals; placeholder-name checks use an ASCII fast path (~112 ns, 0 allocs).
  • Liquid parsing is about 12–23% faster (~3.0–3.5 ms/op → ~2.64 ms/op) by avoiding placeholder-token heap allocs and skipping sentinel regex when none are present.
  • Segment validation is about 17–34% faster on common paths (424.5 ns → 353.2 ns/op plain; 733.5 ns → 486.3 ns/op QA) with 46–79% less heap and fewer allocs.

​
Tests

  • MT lock migration, mixed LLM/MT runs, JSON report aggregates, tracing, and secret redaction.
  • Command-level MT-only and mixed MT/LLM runs, including lock reuse, missing credentials, DeepL URL validation, and legacy LLM-only compatibility.
  • entries --srx for JSON, FormatJS skip, and Markdown documents.
  • Cache-hash reuse during task planning and ICU plain-text invariant fast-path regressions.

​
Maintenance

  • Drop protobuf scaffolding from the Go module and simplify make fmt.