[Nobbettt]

Summary

sharp-pen now reviews every document in its own language, detected automatically from the text, including documents that mix languages. This PR also releases VS Code extension 0.3.0.

Extension

• Language: the prompt tells every AI client to detect the language from the text, review each passage by its own language's rules, keep the spelling variant, and never translate. Each note is in the same language as the text it's about. There is no language setting.
• Protected text: quoted text, product names, technical terms, code-like text and deliberate fragments are never changed.
• Quality:
  • fewer, confident suggestions
  • Level 1 errors are always listed at Level 1, even when a Level 2 rewrite covers the sentence
  • short or common words get a neighbouring word, so their anchor is unambiguous
• Reliability: the prompt has a bilingual worked example, describes the response format once (by its schema), and repeats the output rule after the document.
• Version 0.3.0, with a CHANGELOG.md entry.

Skill

• Language: the skill identifies the draft's language or languages, reviews each passage by its own language's rules, and never translates. Its reply to the author is in the author's language.
• changes.json: gains an optional, validated language tag, such as sv or en-GB. The page sets it as the lang of its text panes.
• Wording: Level 2 is renamed to "sentence structure and clarity".

Tests and docs

• npm run eval:languages: runs the real AI CLIs on four mixed-language documents: Hungarian/Spanish in both orders, Swedish/English/Hungarian, and Swedish/Norwegian/Danish. It scores recall and flags protected text that changed, Level 1 rewrites, and notes in the wrong language. It isn't part of CI; its scoring has unit tests that are.
• READMEs: both say sharp-pen works in every language the chosen model supports.

After merging

• Skill and plugin: v0.1.6 publishes automatically.
• Extension: push the tag extension-v0.3.0 on main to publish 0.3.0 to the VS Code Marketplace.
• Still to rename: "sentence construction" is still used in the plugin manifests, the extension README and the website.

Test plan

• Unit tests: npm test, 270 tests, pass.
• Language evaluation: npm run eval:languages, 3 runs per AI tool and document. Claude, Codex and OpenCode pass all four documents. Copilot passes two; it fails both Hungarian/Spanish documents at 86% recall, because its default model misses the Hungarian comma before "aki" and some Spanish agreement. No tool changed protected text or mixed up languages.
• Skill evaluation: a throwaway test with the skill's real instructions and build.py, 2 runs per tool and document. Claude and Codex caught 100% of planted errors, Copilot 88–100%.
• build.py: validates language, and the page sets lang on its text panes.
• CI: green on all 20 jobs, covering Linux, macOS and Windows on x64 and ARM, and VS Code 1.85.0.