Changelog
@eloqnt/engine
Historic reference of all changes to @eloqnt/engine.
0.4.35
- The translation-quality eval's placeholder gate also fails an output that prints an argument through a different formatter than the source or drops its select.
inconsistent-exact-pluralsandtranslatenow accept exact plural cases a translation adds, e.g. French=0 {Aucun objectif}, and only report the source's exact cases a translation drops.translateadds one where a number the source words without the count shares its category with other numbers.- Add a
missing-plural-countlint rule for a translated plural case without the count that the locale selects for several numbers, e.g. Portugueseone {uma meta}, which also renders for 0. - Translations follow styleguide decisions over the prompt's defaults. Place and historical names now take their established name in the target locale instead of staying in English.
0.4.33
inconsistent-argsno longer reports an ICU argument as missing when the source spells it only inside plural cases the target locale never selects.
0.4.32
- Switch the hosted model to
gpt-6-luna.
0.4.31
eloqnt reviewnow retries a fix that fails the post-review checks, telling the model what was wrong, instead of dropping it. A fix that still fails after the retries is dropped as before.- Attach source outlines as context:
translateandreviewruns withsrcPathconfigured now collapse each module around its call sites into a small, deterministic code excerpt and send it alongside the strings, so the model can e.g. tell an alert title from a button label and more.
0.4.30
eloqnt reviewnow replaces other wording for a term your styleguides define, not just misspellings of it.
0.4.26
- @eloqnt/sdk@0.6.25
0.4.25
- @eloqnt/sdk@0.6.24
0.4.23
- @eloqnt/sdk@0.6.22
0.4.22
- @eloqnt/sdk@0.6.21
0.4.21
- @eloqnt/sdk@0.6.20
0.4.15
- Teach the translate prompt ICU apostrophe escaping, real line breaks, and the
previousAttemptretry field, and escalate reasoning effort on retry requests.
0.4.5
- Improve review prompt to avoid a contradiction related to brand names.
0.4.3
- Improve accuracy of handling plurals.
- Switch the hosted model to
gpt-5.6-luna.
0.4.2
- A translation no longer has to spell out every plural category its locale can select on (ref). The rule is replaced by two, since its halves answer to different authorities:
unreachable-plural-casefor a case the target locale never selects, andinconsistent-exact-pluralsfor=0-style cases that don't match the source. - Translations now pick the plural categories a string needs rather than every category the locale allows, guided by the numbers each one selects. No two cases may repeat the same wording, and trailing punctuation has to match the source in both directions.
0.4.0
eloqnt translatenow reads CLDR data sopluralandselectordinalstatements are guaranteed to be adjusted for target locales.
0.3.2
- Improve prompt caching.
0.3.0
- Added
eloqnt review source: Reviews source strings for typos and grammatical errors and fixes them in place.
0.2.2
- Broaden the translate prompt to preserve non-ICU placeholder styles (e.g.
{{name}},%d,${name}) and to handle suffix-encoded plurals like_one/_other
0.2.0
- Upgrade to AI SDK 7.
0.1.1
- Stop adding units, currency, or
%next to placeholders whose substituted value already carries the symbol.
0.1.0
- Major prompt revision based on new eval suite and switch the engine model to
openai/gpt-5.4-mini.
0.0.1
- Consume meaningful message ids as translation context. When a message id is a readable key (e.g.
settings.table.sortOrder), the engine prompt now uses it to disambiguate the source; opaque hashes are ignored.