What's new in v0.15.0
Added
- First-run "Connect your AI" experience. A fresh install ships with no LLM
key, so a brand-new user's first Chat or Learn attempt used to fail with a raw
auth error and they'd give up. Now onboarding has a "Connect your AI" step
(a free online key — Groq/Gemini, no card — or fully offline Ollama with no
key), and every AI surface (Chat, Learn) shows a friendly "Connect your AI to
get started" card instead of erroring when no provider is configured, with a
one-click deep-link to Settings → AI (/settings?tab=llm). - Chat follow-up chips. After each answer, Chat streams 3 clickable follow-up
questions — turning a one-shot Q&A into a session. Best-effort, emitted after
doneso a failure can never block the answer.
Changed
- Sidebar redesigned to match the amber/copper theme: its own surface, a
gliding copper active-indicator, refined hover/focus states, and a polished
bottom cluster — across light / medium / dark. - Honest Knowledge Base stats. Removed the fake "progress %" (it was a hash
of the KB id) from the KB cards; the library filter is now All / Indexed /
Empty (real) instead of the hashed "Mastered / Review".
Fixed
- Cinematic (Manim) renders showed non-English on-screen text as boxes. A
lesson in a non-Latin language (e.g. Hindi "प्रकाश संश्लेषण") rendered every
glyph as a.notdeftofu box, while Latin ("CO2") was fine. Root cause: the
Cinematic pipeline asks the model to write the on-screenText(...)in the
lesson's language, but Manim's default font does no per-glyph fallback to a
script-covering font — even when one (e.g. Nirmala UI) is installed. The render
sidecar now picks the first installed font that covers the lesson's script
(Devanagari / CJK / Arabic) and sets it as theTextdefault before running
the generated scene, so non-Latin lessons render real glyphs. Verified end-to-end
against the real Manim env (Hindi → correct Devanagari). (Known Issue #19.) - Fabricated progress signals. The Daily Spark "streak" was the day of the
week (it showed "Day 4" on a fresh install) — now the real streak, hidden at 0;
"Knovex Noticed" no longer claims the knowledge base is "growing" on an empty
install; and "Queue review" now links to/review(was/learn).
Tests
- Chat follow-up suggestion parsing + "emits after done" stream test; Manim
sidecar script→font selection; deep E2E control coverage (chat / learn /
progress / review).
Installation
See README.md for setup instructions.
Secure · Fast · Reliable · Cost-Effective