Skip to main content
run and eval can now use OpenRouter and Vercel AI Gateway as LLM providers. This release also fails closed on incomplete OpenAI-compatible completions, tightens Crowdin upload/error handling, and speeds up planning, scoring, and several file parsers. View the GitHub release · Full changelog

​
Breaking changes

None.

​
Features

  • Add Vercel AI Gateway as an LLM provider for run and eval. Set llm.profiles.<name>.provider to ai_gateway and export AI_GATEWAY_API_KEY. Model IDs such as openai/gpt-5.6-luna pass through to https://ai-gateway.vercel.sh/v1. Image localization stays OpenAI-only.
  • Add OpenRouter as an LLM provider (openrouter). Auth via OPENROUTER_API_KEY; override the base URL with OPENROUTER_BASE_URL. Model IDs pass through unchanged.

​
Fixes

  • Reject incomplete or failed OpenAI-compatible completions instead of writing truncated translations. Fail closed on choice-level errors, finish_reason of error / length / content_filter, and tool_calls / function_call. Transient 502/503/504 errors stay retryable. (#1810)
  • Send the raw filename in the Crowdin Crowdin-API-FileName header instead of query-escaping it, so names with spaces are not mangled. (#1839)
  • Align Crowdin SourceStringsUpload attributes with API v2 (fileId, maxLen, omitempty on optional fields). (#1823)
  • Surface Crowdin 400 error payloads instead of a generic status message, and allow the documented project-level notification roles. (#1801)

​
Performance

  • Faster SHA-512 lock fingerprinting during planning (hex.EncodeToString instead of fmt.Sprintf("%x")): about 28% faster on large catalogs, 11% fewer allocations. (#1753)
  • ICU parser plain-text fast path: about 19% faster and 60% fewer allocations on keys with no placeholders. (#1775)
  • JS/TS locale module parser: about 35% faster parse and 60% fewer allocations; marshal about 30% faster. (#1799)
  • Mozilla Fluent parser: about 45% fewer allocations via capacity hints and indent/comment fast paths. (#1786)
  • Zero-allocation brace-placeholder scanner in eval scoring (replaces regex). (#1821)
  • Lower allocation cost when building the selection catalog (pre-sized maps, slices.SortFunc). (#1809)
  • Segment profile validation fast paths and slice capacity hints. (#1837)

​
Tests

  • Env loader, CLI telemetry, sync risk detection, SHA-512 fingerprint oracles, and brace-placeholder scanner coverage.
  • Parser coverage for Apple .stringsdict, CSV, Liquid target fallback, Fluent attribute indents, JS/TS locale edges, and the ICU plain-text fast path.
  • Incomplete LLM completion and locale-path helper regressions.

​
Maintenance

  • Bump Go dependencies (OpenAI SDK v3.50.0, WorkOS v10.1.1, OpenTelemetry v1.45.0). (#1751)