v0.1.16

Rewrites keep borrowed words in mixed-language text more often (#1).

Mixed-language text

Improve writing, Make professional, Make friendly, Simplify and Expand used to translate the words you borrow from another language: el deadline came back as el plazo, mañana as tomorrow. ProofKey now finds those words itself and names them to the model as words to keep exactly as written:

• English words inside Spanish, French or Portuguese text (deadline, kickoff meeting, rollback), found with an English word list (SCOWL, 1.2 MB in the package) and, per language, a mask of the English words that language does not have. The word list is read only for text in one of those three languages.
• Spanish, French, Portuguese or German words inside English text (mañana, pero, merci), found by letters English spells without and a short hand-written list. No word list is needed.

Fix grammar, the live check, Summarize, Convert to bullet points and custom actions do not get the list: a typo that happens to be an English word (boy for Spanish voy) would be kept as written, and Fix grammar exists to fix exactly that. German as the base language is not covered: its dictionary already lists Meeting, Team and Bug. Until then, Terms never to flag (Profile) keeps any word you add.

API keys

The local copy of your keys (since 0.1.14) is now also kept when settings arriving from sync carry the mark a later release will write when it stops putting keys in sync. This prepares that step so that no computer loses its key. It changes nothing about where keys travel today.

How it is known

tools/action-eval.ts, the five rewording actions over the eight mixed-language fixtures, each time with the version before as a control:

• gemini-2.5-flash through OpenRouter, 10 runs, 2026-10-14: 246/400 → 321/400 checks kept the mixture (the two Spanish fixtures 21/100 → 96/100).
• The same, 2026-10-15, for the English-base direction: 311/400 → 352/400 (I told her mañana works better, pero… 3/50 → 33/50).
• Qwen3-4B on llama.cpp, temperature 0: 66/120 → 79/120 (3 runs, 2026-10-14). The English-base change, measured over one run on 2026-10-15, made no difference: 27/40 before and after.

On text in one language with typos, the shipped rule named one misspelling that happens to be an English word (boy). On 2,532 paragraphs of English prose, the English-base rule named words in one paragraph, and they were names. On the released tree, npm run verify reported 658 passing checks and no failures. Details and limits are in MODELS.md.