What's in the repo

Tauri 2 app: a Rust backend (src-tauri) and a React 19 + Tailwind 4 frontend (src). Everything runs locally; the only network calls are to Ollama on localhost:11434 and, if you switch web access on, the pages you ask it to read.

Data flow

mic ─▶ VAD (Silero) ─▶ wake word / speaker check ─▶ STT (Parakeet)
                                                        │
                                                        ▼
                                         decision.rs: text ─▶ one typed Action
                                                        │
        ┌──────────┬──────────┬──────────┬─────────────┼───────────┬───────────┐
        ▼          ▼          ▼          ▼             ▼           ▼           ▼
      Chat       Teach/     Screen     Research/     Remember/    Do        Skill
      (llm.rs)   Quiz       (screen.rs) LearnPage    Notes        (agent.rs) (skills.rs)
                 (learn.rs)             (web.rs)     (memory.rs,
                                                      notes.rs)
                                                        │
                                                        ▼
                                  reply text ─▶ TTS (Kokoro) ─▶ speakers
                                  HUD / overlay events ─▶ React UI

Every utterance becomes exactly one Action before anything runs. The Decider trait is the seam for a smarter router; Rules is the deterministic default and the fallback.

Backend (src-tauri/src)

File	What it does
main.rs	App setup, Tauri commands, hotkeys, and wiring of the audio, speech and model threads
llm.rs	The turn loop: routes an Action, builds the prompt with memory and note context, calls Ollama (chat, vision, embeddings, JSON-schema output), runs skills, ends sessions. Currently Ollama-only (see the provider issues)
decision.rs	Action enum, Decider trait, Rules router
agent.rs	Acting on the computer: one typed Step at a time, approvals, path/URL safety checks, audit log. Tools: open app / page / file, type, click, web search, look at screen, recall
learn.rs	Teaching: learning plans, concepts, prerequisites, quizzing, grading, spaced review
memory.rs	Long-term facts with provenance, confidence, and embeddings; recall by meaning
notes.rs	Import PDF / Markdown / text, chunk and index in SQLite FTS5
skills.rs	Versioned prompt-only skills: proposed → experimental → approved, tests, rollback
web.rs	Safe fetching (public addresses only, redirect re-checks, size caps), Readability extraction, headless Chrome fallback, search
platform/	Everything OS-specific behind the Platform trait: open app/page/file, type, click, screenshot, window list, OCR, browser address, permissions, HUD window behaviour. macos.rs is the real one; other OSes use unsupported.rs, which fails with a spoken message
screen.rs	Screenshot, OCR boxes, window list, vision call, overlay drawing, click preview
stt.rs	Speech to text (Parakeet via ONNX), wake-word spotting
tts.rs	Text to speech (Kokoro)
speaker.rs	Voiceprint (wespeaker ResNet34), "is this the owner" check
bargein.rs	Interrupt handling: stop words, talking over Sunday
mic.rs, models.rs	Audio capture, model file locations
db.rs	SQLite schema and queries

Frontend (src)

File	View
Workspace.tsx	Main window: chat and typed input
Hud.tsx	Floating status HUD
Overlay.tsx	Click-through drawing layer for boxes, circles, arrows
Palette.tsx	⌘K command palette
Knowledge.tsx	Notes import and search
Memory.tsx	Browse and delete facts and pages
Skills.tsx	Draft, test, approve, roll back skills
Settings.tsx	Model picker, web switch, diagnostics
ui.tsx, index.css	Shared components and styles

Scripts

File	Purpose
scripts/fetch-models.sh	Download speech models into ~/.cache/sunday/models
scripts/build-ocr.sh, scripts/ocr.swift	macOS Vision OCR helper: exact text line boxes, window list, cropping

Database (SQLite, one file)

messages, kv, concepts, prereqs, attempts, reviews (learning); documents + FTS index (notes); facts (memory); skills, skill_runs (skills); actions (audit log of everything the agent did).

Safety model today

• Opening apps and pages you asked for just happens. Typing, opening files and clicking ask for a yes.
• Once anything read from the web, the screen or memory is in play, every step asks. Read content is data, never a command.
• File access: your home folder only, no hidden folders. Web: http(s) only, public addresses, redirects re-checked.
• Every step lands in the actions table.

The roadmap hardens this into tiers (#10) with an injection red-team suite (#11).

Not here yet

Evals, a persistent multi-step task loop, non-Ollama model providers, an MCP/tool SDK, Windows and Linux. All tracked in the v1 issue.