v0.0.3

• The session is a chat. Each question is one conversation: you send only your follow-up, and the tutor grades the whole conversation and answers your last point. Hints and the lesson sit in the conversation, and the composer stays at the bottom.
• Waits are clear: a progress bar, the task and its seconds on the button you pressed, and placeholders. Text boxes grow with their text.
• A model for each tutor task (question, grade, hint, lesson, extract), set in Settings from ramabana's catalogue: Claude, Codex, OpenAI, Copilot, Gemini and local models. A test times one call. Install the catalogue with abhyasa[models].
• Speech: dictation stays on and cuts your speech at pauses, and the text comes in as you talk. Gemma 4 E4B is the default; Whisper, Parakeet, Qwen3-ASR and OpenAI are options. Settings finds, gets and switches speech models.
• The graph has levers: four layouts (free, path, rings, groups), spacing, size, colour, filters, find and focus. It keeps your arrangement.
• A vault viewer shows any vault document, with the resume command for a session and "Learn from this".
• Interactive demos from rupaka. A demo asks you to predict first, and it shows the result only after you lock in. The tutor makes one after two misses with a saved misconception, or for a concept with visual: true, and "Show me a demo" (⌥E) makes one on request.
• A prediction in a demo counts in the learner model: at depth 2, or at depth 3 when you write why.
• /due.json gives the count of due reviews, for the Mac app's Dock badge.
• The session key is kept in the config folder, so the app starts from any folder.