Skip to content

0.8.2 (2026-10-04)

An AI chat fix: the diagram it generates now always reaches the preview card instead of staying as loose code in the chat. The chat is also fully available in English. No new features and no migrations.

⬆️ Upgrade notes

  • No migrations. Rebuild the images as usual (docker compose up -d --build).
  • Messages already saved with loose code don't change; new replies do arrive with a preview.

🤖 AI chat

  • The diagram no longer stays as code in the chat. The chat guessed from keywords whether a message asked for a diagram or asked a question, and only looked for the code in the first case. Requests such as "Can you make…?", "I need…", "Add to it…", "fix it" or "Database diagram…" were taken for questions: the model did answer with the diagram, but it was shown as text broken into fragments, with no preview or accept button, and had to be copied into the editor by hand.
  • Now the model's reply decides: if it carries the diagram, it is shown with a preview, validated and offered to accept or reject. With "Explain" only a diagram marked as a change counts (not the current diagram quoted in the explanation).
  • Measured with real Gemini calls: 7 out of 7 requests of that kind left the code in the chat; now 0 out of 7, across the 4 diagram types. Questions are still answered with text.
  • Large diagrams no longer arrive cut. Every chat reply gets the 8,192-token cap (replies that looked like questions used to get less, and were not checked for truncation). A cut reply shows a notice instead of the fragment.
  • The first guess recognises more ways to ask for changes: it ignores accents and understands Spanish verbs with an attached pronoun ("agrégale", "hazlo", "corrígelo").

🌐 Language

  • The whole chat in English: message list, preview card (Accept / Reject / Restore), session selector, relative times, phases ("Thinking…"), errors and message counts with correct plurals.
  • Chat and AI provider errors (no credits, invalid key, model unavailable, rate limit) arrive in the user's language.
  • The editor status bar ("N lines") and the AI fix comparison are translated too.

🧹 Internal

  • Removed ChatModeSelector, which no screen used.

🧪 Tests

  • Backend: 451 tests, 63.03% coverage (45% gate). Frontend: 25 Vitest tests, including es/en translation parity.
  • The new chat tests fail on the 0.8.1 code; the UI flows were checked with Playwright in Spanish and English.