Features (what exists today)

Moved from the README. Everything below is implemented and macOS-only for now (see roadmap).

Teaching anything. "Teach me how to play guitar", "I want to learn linear algebra", "teach me Spanish". The first time, the model drafts a learning plan for the subject (6 to 10 concepts in order, each with what it builds on); then it teaches the next concept: a short explanation and a check question. Ask again to continue. Tried on guitar, linear algebra, sourdough, the Rust borrow checker and conversational Spanish.

Practice. "Quiz me", "quiz me on barre chords", "next question". One question you can answer out loud; your next answer is graded against a reference and stored with an error type. Each concept gets a review date (1 day, 3 days, then growing; a miss resets it). With no topic it picks what's due, then your weakest concept, then the next unlocked one in any plan.

Learning map. Every plan joins one graph of concepts and prerequisites. A concept is known after two correct answers in a row; a wrong answer on a concept with shaky prerequisites points you at the gap.

Notes. Import PDF / Markdown / text (Knowledge view, drag-drop, or ⌘K). Passages are indexed locally (SQLite FTS5). Chat answers quote them ("from your notes on …") and show the sources; quizzes on a concept draw on them. PDFs need brew install poppler.

Screen. "Explain what's on my screen", "what is this question", "point to …": takes one screenshot, a local vision model explains it, Sunday speaks, and boxes/circles/arrows are drawn on a click-through overlay for 25 s ("clear the screen" removes them). It knows what's going on, not just what's visible: which app and window is in front and what else is open (from the macOS window list, no extra permission), and "explain this page / window / PDF" looks at just the front window. Pictures and diagrams (screens with a few scattered labels, or visual questions) go to the vision model in one pass; it answers in two short sentences and marks what it mentions, and rough boxes are snapped to the text under them. Ordinary text screens skip the image and answer from the on-screen text, which is much faster. Pointing at text is exact: macOS's text recogniser (scripts/ocr.swift) finds every line's box, the model just picks line numbers. Only for non-text things (diagrams, icons) does it use the model's own coordinates, which are approximate. Only on request; the image is read once, sent to local Ollama only, and deleted. Needs Screen Recording permission. A text screen takes ~8 s, a picture ~10 to 20 s.

Memory and the web. Long-term memory is a store of atomic facts, each with where it came from (you told me / read on a page / inferred from a conversation), its confidence, and when it refers to; recalled by meaning (local embeddings) and cited in answers. The design follows Devsper's memory service (devsper/flow/svc/memory). "Remember that …" stores what you said. With web access switched on (Settings; off by default): "learn this page" reads the tab in your front browser, "read https://…" reads a link, "research X" searches the web, reads the top results in parallel and answers with sources. Pages are fetched safely (public addresses only, re-checked on every redirect, size-capped), with Readability for the main text and headless Chrome for JavaScript-only pages. What it reads is extracted into facts in the background. When you say goodbye, lasting facts about you are saved as unconfirmed guesses. All of it is listed in Memory and deletable per fact or per page.

Skills. Versioned, prompt-only capabilities (a template, an output schema, tests). No generated code, no tools. Lifecycle: proposed → experimental → approved; approval needs a passing test run on that version; any approved version can be rolled back. Describe one in the Skills view and the model drafts it. Voice: "run skill learn.mnemonic on the OSI model".

Acting. "Open Spotify", "pull up images of a keyboard", "go to github.com", "type …". The model picks one typed step at a time (agent.rs): open an app, open a page (in the browser you already have open), open a file in your home folder, type into the front app, web search, look at the screen, recall. Opening an app or page just happens when you asked for it directly; typing and opening files always ask "…?" and wait for a yes; and once anything read from the web, the screen or memory is in play, everything asks (read content is data, never a command). Hidden folders, non-http addresses and anything outside your home folder are refused. Every step is logged in the actions table. Typing needs Accessibility permission.

Decisions. Every utterance becomes one typed Action (decision.rs) before anything runs, and is checked against what's permitted. Rules is the deterministic router; the Decider trait is where a smarter router goes.

⌘K command palette; Settings has the model picker and a diagnostics checklist.

Your voice. "Learn my voice" (say it, or Settings / ⌘K) records ~20 s of you, builds a voiceprint (wespeaker ResNet34, on-device) and from then on Sunday ignores other voices: wake word, commands and talking-over-it all have to match you. The fn key always works. Logs show voice match 0.xx so the threshold (speaker::SAME) can be tuned. "Forget my voice" removes it.

Voice details: "Sunday" or fn wakes it; say "stop", talk over it, or press fn to interrupt; "bye" ends the session and unloads the speech model.