# ❄️ Freezmo

**Turn photos of your study notes into a swipeable feed of mini study games.**

Freezmo is a TikTok/Gizmo-style study app. You paste or upload photos of your notes, an AI reads
them and builds a deck of 8–15 interactive activities — flashcards, quizzes, matching games and
more — that you swipe through vertically like a feed. Every answer gives instant feedback, XP and
streaks, and the deck is saved in your browser so you can retry it later.

All AI runs through **your own Ollama setup** (a local daemon or the Ollama cloud API). No accounts,
no database, no data leaving your machine except the requests you send to Ollama.

---

## Table of contents

- [Features](#features)
- [Quick start](#quick-start)
- [Supported input formats](#supported-input-formats)
- [Ollama setup](#ollama-setup)
- [Rate limits](#rate-limits)
  - [Mode 1 — local daemon (recommended)](#mode-1--local-daemon-proxying-cloud-models-recommended)
  - [Mode 2 — direct Ollama cloud API](#mode-2--direct-ollama-cloud-api-api-key)
  - [Changing models](#changing-models)
  - [Checking the connection](#checking-the-connection)
- [Running on your network / from your phone](#running-on-your-network--from-your-phone)
  - [Keep it running as a service](#keep-it-running-as-a-service-systemd-user-service)
- [Environments and deploy](#environments-and-deploy)
- [Folders](#folders)
- [Install as an app (PWA)](#install-as-an-app-pwa)
- [Security notes](#security-notes)
- [How it works](#how-it-works)
- [Chat](#chat)
- [Project structure](#project-structure)
- [API routes](#api-routes)
- [Scripts](#scripts)
- [Data, storage and privacy](#data-storage-and-privacy)
- [Troubleshooting](#troubleshooting)
- [Decisions and notes](#decisions-and-notes)

---

## Features

**Capture & generate**

- Drag & drop, file picker, **mobile camera capture** and **clipboard paste** (Ctrl/⌘+V)
- **Phones get a 3-step flow** — notes → options → summary — with a thin progress bar and
  “1 / 3”. Step 1 leads with the step name as the headline, a formats/limits hint, one big
  *Add photos or files* area plus *Camera* and *Paste text instead*, and a “nothing added
  yet” placeholder that becomes the thumbnail/chip list. Next stays disabled until something
  is added. Step 2 keeps folder, difficulty and count at their defaults with a helper under
  each group and tucks the topic hint and pasted notes behind one collapsed row. Step 3 is a
  real summary — thumbnails/chips, files, difficulty, activities and folder, each with an
  *Edit* link back to its step. A flat Back/Cancel + Next/Freeze bar sits at the bottom above
  the safe area (the tab bar hides inside the wizard), and it is portalled to `<body>` so the
  screen entrance animation can never displace it. While a run is generating, the wizard's
  **mode switch hides** — a loading screen can't be swapped to the other wizard mid-flight —
  and leaving the page anyway aborts the request instead of saving a deck in the background.
  The desktop page is unchanged.
- **Native-feel touches (phones)**: the Android Back gesture closes the topmost layer first
  (sheet, actions sheet, then one wizard step); deleting a deck shows an *Undo*
  toast instead of a confirm; the deck menu can share a deck as JSON through the system share
  sheet (clipboard fallback); content shared from other apps opens a chooser and lands
  pre-filled in either wizard; haptics (Settings → Haptics, on by default) tick on tab
  switches, toggles, correct/wrong answers and completion; the screen stays awake during
  study sessions and homework guides; wizard drafts survive a reload or an OS kill; re-tapping
  the active tab scrolls to the top; an offline pill appears and generation is disabled with
  a reason until the connection is back.
- **Chat with your notes**: an LLM-style conversation grounded in your decks. Ask from the
  Library, a deck's menu, or a result; pick the scope (all notes, folders/decks, or none) and the
  model; photos and documents are transcribed first, so any model works. Answers stream with stop,
  copy, regenerate and edit-and-resend, render safe markdown + math, and hide reasoning behind a
  **Thinking** row. See [Chat](#chat).
- **Learning engine**: every answer is recorded (deck, activity, topic, correct/wrong, time) in an
  IndexedDB log, and each activity gets a light **SM-2 schedule** — correct answers come back after
  1 day, then 3, then interval × ease; wrong ones after ten minutes. The Library shows a **Smart
  review** card with the due count, a **Mistakes** count (activities whose latest answer was wrong —
  they clear themselves once answered correctly) and today's **daily goal** progress. Tapping
  either one shows a **loading screen instantly** (the link reports its own pending state, the
  route has a loading boundary, and the payload is prefetched), so even a slow first request never
  looks like a dead tap. A **Progress**
  tab has the day streak, a five-week calendar, accuracy per subject and per deck and
  the weakest topics. **Exam mode** (its own tab) takes one or more decks, a question count and an
  optional timer, hides hints and shows a per-deck/per-topic breakdown at the end. **Daily
  reminders** appear in Settings only when the server has VAPID keys configured; subscriptions are
  stored as endpoint + keys only, and `node scripts/send-reminders.mjs` (cron/systemd timer) does
  the sending.
- **Backup, sync and limits**: *Settings → Data & backup* exports **everything** (decks, folders,
  homework, progress, safe settings — never the API key) to one JSON file and imports a backup with
  a preview and a merge/replace choice; a gentle, dismissible nudge appears after 5 new decks or 14
  days without a backup, and the storage-persistence state (`navigator.storage.persist()`) is shown.
  *Settings → Sync* is **off by default** and opt-in: the data is encrypted in the browser
  (PBKDF2-SHA256 → AES-GCM, key kept as a non-extractable `CryptoKey`) before it is uploaded, the
  server stores only an opaque blob under a random 128-bit sync ID, and conflicts are surfaced with
  an explicit *keep this device / use synced copy* choice. When Ollama's **free monthly usage** runs
  out, every flow that calls the model shows the same clear message with **Open Settings** (jumps to
  the key field) and **Open Ollama** — never a silent fallback to another model.
- **Photos and documents**: JPG/PNG/WEBP/HEIC images plus **PDF, DOCX, PPTX, ODT/ODP/ODS, TXT, MD, CSV**
- Documents are read **in the browser**: PDF text layers are extracted, scanned PDF pages are rendered
  as images for the vision model, DOCX/PPTX text, speaker notes and embedded images are pulled out.
  The raw file never leaves your device.
- Up to 8 photos + 6 documents per deck (max 12 images reach the model); photos are resized to max
  1600 px and re-encoded as JPEG before upload
- Optional extra text (pasted notes) and a topic hint
- Easy / Medium / Hard difficulty and 8 / 10 / 12 / 15 activities per deck
- Fun staged loading state (`Reading your notes… → Finding the key ideas… → Building your activities…`)
- Friendly error cards with a **Retry** button that point at Settings when a key is needed

**Language**

- The notes' language is detected and **preserved end to end**: transcription, deck title, questions,
  options, explanations, short-answer grading and later *regenerate* / *add more* all stay in that
  language. Czech notes produce a Czech deck; the app never translates them to English.
- **Language guard:** after the vision step, the transcription is checked (script + language markers).
  If it clearly came back in another language, `TEXT_MODEL` (which is much stronger multilingually)
  restores it before any activities are written — the UI shows *“Restoring the notes' language…”*.
  On good transcriptions the guard stays silent and costs nothing.
- The deck stores its language, so short answers are graded and feedback written in the same language.
- The **Freezmo interface** has its own language switch: **English (default)** and a full **Czech**
  translation, picked in **Settings → Language** or from the `EN`/`CS` button on the key gate. The
  chosen locale is remembered per browser and also decides the language of server messages (API
  errors, health notes), so nothing falls back to English mid-sentence.

**Study feed**

- Full-screen vertical scroll-snap feed, one activity per screen
- Arrow-key / PageUp / PageDown / Home / End navigation, plus swipe on touch devices
- Progress bar, activity counter, XP counter and answer streak in the chrome
- Instant feedback with a short explanation for every activity, confetti when you finish
- Results screen: **Retry missed ones** (fresh, answerable items), **Study the whole deck again**,
  and **Regenerate deck** — rebuilds the activities from the stored notes so you can keep drilling
  until you're ready

**Make my homework**

- Same intake as a deck — paste the assignment, photograph it, or add a PDF/DOCX/PPTX/ODT/text
  file; photos go through the same vision model and language guard that notes do
- Three modes: **Just make it** returns the finished homework (the answer only, ready to copy),
  **Explain how to make this homework** turns it into a guided walkthrough, and **Check my work**
  marks the student's own attempt — what's right, what's wrong and why, plus a corrected version
  (the attempt is typed or pasted; photos of handwriting are not supported yet, and the UI says so)
- **Follow-ups on a result**: *Shorter*, *More detail* and back to the *Original* rewrite the
  finished text in place; inside a guide every step has *Explain step N differently*
- **Turn into a study deck** on any result opens New deck with the content (the answer, the
  corrected version or the assembled guide) and the options step pre-filled
- On phones this is a **2-step flow**: the assignment first (same simplified input as the
  deck flow, with its own step headline, formats hint, big add area and empty placeholder),
  then the mode cards under a short preview — tapping a card starts it immediately (the check
  card first reveals the solution field), so there is no separate button. Step 1 starts with
  *Cancel* and the flat action bar matches the deck wizard. Desktop keeps the select-then-confirm
  layout.
- The guide is one step per screen in the same snap-feed feel as the study feed: what to do and
  why, **three progressive hints** per step (nudge → hint → the full answer), and a **checkpoint**
  built from the normal activity components and graded through `/api/grade`
- The last screen assembles the solution from the step answers, with a copy button
- Results live in `freezmo.homework.v1` — the homework screen's history and the library's
  **Homework** tab — where you can reopen or delete them; **Redo** re-runs from the
  stored assignment text (and the stored attempt for checks), and original files are never saved
- The answer and the guide are always written in the **language of the assignment**, never
  translated into the interface language

**Activity types** (each one is its own React component, rendered from JSON by a `type` field)

| Type | Interaction |
| --- | --- |
| `flashcard` | Tap to flip, then self-assess *I knew it / Not yet* |
| `multiple_choice` | 4 options, keys 1–4, correct answer + explanation |
| `true_false` | Tap the buttons **or** swipe the card right/left |
| `fill_blank` | Inline input (one typo forgiven) or a word bank |
| `matching` | Tap a term, then its definition; perfect match for full XP |
| `ordering` | Drag with the mouse/finger, or use ▲▼ buttons, then check |
| `short_answer` | Typed answer graded by the AI via `/api/grade`, with a self-check fallback |

**First run**

- A full-screen **API-key gate** replaces the app until a key is saved (in `user-key` mode): one
  masked input with a show/hide toggle, a single **Continue**, and a "?" sheet with the four steps
  to a free key. The key is validated with a real 1-token probe before it is accepted — invalid
  keys and network trouble get different inline messages. Removing the key in Settings brings the
  gate back; `server` and `local` installs skip it entirely.

**Library & folders**

- **Two pools, never mixed**: a **Study Decks** tab (the folder + card library) and a **Homework**
  tab with saved solutions and step-by-step guides — each with its own empty state and call to
  action, and the active tab lives in the URL (`/?tab=homework`)
- **Phone library, decluttered**: no stats line, the folder chips and “Unorganized” heading only
  appear once folders exist, and a **⋯** button next to the pool switcher opens a small sheet with
  **Import JSON** and **New folder**. Deck cards shrink to title + one meta line (activities · difficulty · best
  score) + a one-line description that expands on tap, with the whole card tappable and a small
  **Study** pill; homework cards match.
- Decks saved locally with title, subject, folder, activity count, date, play count and best score
- **Folders**: create named groups ("History", "Biology", …) from the sidebar or the create screen.
  New decks freeze straight into the selected folder; anything unfiled lives in **Unorganized**.
  Rename, delete (decks survive — they return to Unorganized) and move decks between folders
- Rename, delete (with confirm), regenerate (from the library or the results screen), and
  **add more activities** to an existing deck
- Export a deck as JSON and import it back — the folder name travels with the file and is recreated
  on import

**Design**

- **ChatGPT-style shell on desktop, five tabs on phones**: desktop keeps the persistent left
  **sidebar** (brand, a quiet "New deck" row, a *Folders* list with per-folder counts,
  Settings/Dark mode pinned to the bottom), which **collapses** to a 68px icon rail (remembered in
  `localStorage`). Phones have no drawer and no hamburger: a frosted header (logo + theme chip) sits
  above an iOS-style bottom tab bar — **Library · Progress · New deck · Exam · Settings** — with the
  New deck button as the prominent centre action and Settings opening the sheet. The wizards keep
  their own bottom action bar instead of the tab bar, and the in-page segmented switch moves between
  the deck and homework wizards.
- **Frosted ice glass, both themes**: light page `#eef5fd` (pale ice) with white glass panels;
  dark page `#0a1120` with `rgba(255,255,255,.06)` glass. Every large surface shares one recipe —
  translucent tint, 24px `backdrop-blur` (16px on phones), a bright 1px rim, an inner top highlight
  and a blue-tinted shadow — over a fixed ice scene (navy/pale base, drifting glacier blobs, glossy
  ice orbs and a whisper of grain).
  Text tiers are documented AA ratios (see below), and the cards fill the row with
  `repeat(auto-fill, minmax(280px, 1fr))` — no stranded cards, no dead space
- Deck cards: title → muted metadata → neutral tags (difficulty carries the only colour:
  green/amber/red) → summary → a single **Study** button, with everything else in the `⋯` menu
- **Frosted UI**: the sticky header, tab bar, sheets and panels use real `backdrop-filter` blur over
  the fixed ice scene (drifting glacier blobs, glossy orbs, a whisper of grain) — depth without noise
- **Pill language**: primary CTAs are near-black in light mode and white in dark mode; secondaries
  are outlined or ghost. Colour is reserved for meaning — activity tints (sky/lilac/mint/peach/rose),
  correct/wrong feedback, key status
- Micro-interactions: 160–200ms colour/transform transitions, `active:scale-[0.985]`, skeleton
  shimmer, hairline hover washes on every row and card
- **Motion (phones)**: one shared language (`lib/motion.ts`) — every button, card, chip, tab and
  nav row scales to 0.96 on press with a spring release; the tab bar, segmented controls and
  difficulty/size pickers glide their highlight with a shared `layoutId`; wizard steps slide
  forward and backward and page content enters direction-aware; the Settings sheet and the library
  actions sheet follow the finger and settle on release velocity; library cards fade up in a capped
  stagger and animate out on delete; score and XP count up; correct answers pulse, wrong ones shake;
  `prefers-reduced-motion` collapses everything to short fades. Only transform and opacity
  animate — never blur.
- **iOS-flavoured PWA**: standalone mode, safe-area insets, bottom tab bar, sheet-style dialogs with
  a grabber, inset grouped lists, segmented controls, "Add to Home Screen" hint — see
  [Install as an app](#install-as-an-app-pwa)
- Mobile-first, dark mode, keyboard accessible, focus rings, ARIA labels, `prefers-reduced-motion` support

---

## Quick start

Requirements: Node.js 20+ (tested on 24) and npm.

```bash
cd freezmo
./setup.sh                     # install + .env.local + production build
npm start                      # http://0.0.0.0:3207
```

Or do it by hand (and use the dev server):

```bash
cd freezmo
npm install
cp .env.example .env.local     # first time only
npm run dev                    # http://0.0.0.0:3207
```

`setup.sh` flags: `--dev` installs and prepares `.env.local` without building, `--no-build` skips the
build too.

Open <http://localhost:3207>. The app runs without any Ollama configuration — you'll just need a
key (or a signed-in daemon) before it can generate a deck. Complete one of the Ollama setups below,
then paste the key on the gate (or later in **Settings**).

`npm run dev` and `npm run build` also copy the pdf.js worker into `public/` automatically (see
`scripts/copy-pdf-worker.mjs`), which is what lets the browser parse PDFs.

Production:

```bash
npm run build
npm start                      # http://0.0.0.0:3207
```

Production is also where the PWA comes alive: the service worker registers, the app can be installed
to the home screen and the `/offline` fallback works. In `npm run dev` none of that runs, so you
won't fight a stale cache while developing.

---

## Ollama setup

Freezmo talks to **`POST /api/chat`** and **`GET /api/tags`** of whatever URL is in
`OLLAMA_BASE_URL`. All calls happen in Next.js route handlers on the server — the browser never
talks to Ollama directly.

Who provides the API key is controlled by **`OLLAMA_AUTH`**:

| `OLLAMA_AUTH` | Key comes from | Use it when |
| --- | --- | --- |
| **`user-key`** (default) | each browser, pasted on the key gate (or in Settings) | several people/devices share the app, or you don't want a key on the server |
| `server` | `OLLAMA_API_KEY` in `.env.local` | a single-user install |
| `local` | nothing — a daemon signed in with `ollama signin` | you run Ollama locally |

### Mode 1 (default) — bring your own key

```bash
OLLAMA_AUTH=user-key
OLLAMA_BASE_URL=https://ollama.com
VISION_MODEL=gemma4:31b-cloud
TEXT_MODEL=gpt-oss:120b-cloud
```

1. Open the app — the **key gate** asks for a key on first visit.
2. Create one at <https://ollama.com/settings/keys>, paste it in and press **Continue**.
3. The gate validates it with a real 1-token request, then drops you into the icebox.
   **Settings → Test connection** re-checks any time and shows the model list.

Where the key lives:

- **In the browser only.** It is stored in this browser's `localStorage` under `freezmo.settings.v1`
  and sent with each API call in the `x-ollama-key` header.
- **Never on the server:** the route handlers read the header, use it for that one request and drop
  it. It is not written to `.env.local`, a database, a cookie, a log line or a JS bundle. Restarting
  the server does not remember any key, and every browser/device needs its own (or the same) key.
- You can remove it any time with **Remove key from this browser**.
- Because the browser hands the key to your own server, use HTTPS if the LAN is not trusted — on a
  plain-HTTP LAN somebody sniffing traffic could read it. The key never goes to any third party.

Keys are verified with a real (tiny, 1-token, 5-minute-cached) chat request, because model listing
on `ollama.com` is public and would happily accept a bogus key.

### Mode 2 — local daemon (no key at all)

Runs cloud models through your local Ollama daemon, signed in once with `ollama signin`.

```bash
ollama                              # start the app/daemon
ollama signin                       # approve the printed link in any browser
ollama pull gemma4:31b-cloud        # vision model, reads the photos of your notes
ollama pull gpt-oss:120b-cloud      # text model, builds the activities and grades
```

`.env.local`:

```bash
OLLAMA_AUTH=local
OLLAMA_BASE_URL=http://localhost:11434
VISION_MODEL=gemma4:31b-cloud
TEXT_MODEL=gpt-oss:120b-cloud
```

Restart `npm run dev` after editing `.env.local`. Verify with the **⚙️ settings dialog** in the
app or `curl localhost:3207/api/health`.

**Free tier.** `gemma4:31b-cloud` and `gpt-oss:120b-cloud` are included in the Ollama **Free** plan
(verified October 2026). Other free models: `gpt-oss:20b-cloud`, `nemotron-3-super:cloud`,
`nemotron-3-nano:30b-cloud`. Models such as `glm-5.3-flash:cloud`, `deepseek-v4.1-flash`,
`minimax-m2.7` and `mistral-large-3` reply with *"this model is not included in your free
usage…"* — Freezmo detects that (HTTP 402) and says so instead of pretending the key is broken.

### Mode 3 — one key for the whole server

```bash
OLLAMA_AUTH=server
OLLAMA_BASE_URL=https://ollama.com
OLLAMA_API_KEY=your_key_here
```

The key stays in `.env.local` (gitignored) and is never sent to the browser. Requests from any
device use it.

#### Installing without root (user-space)

The official installer needs root. If there is no `sudo` on the machine (like the box Freezmo was
built on), Ollama can be installed entirely under your home directory:

```bash
cd ~ && mkdir -p .local/ollama && cd .local/ollama
# grab the tarball for your arch from https://github.com/ollama/ollama/releases/latest
# (ollama-linux-amd64.tar.zst) and extract it here, so you end up with
#   ~/.local/ollama/bin/ollama  and  ~/.local/ollama/lib/ollama/
# No zstd CLI on Debian? Stream it with Python's zstandard module instead of tar --zstd.

export PATH="$HOME/.local/ollama/bin:$PATH"   # add to ~/.profile too
ollama serve &                                 # daemon on 127.0.0.1:11434
ollama signin
ollama pull gemma4:31b-cloud && ollama pull gpt-oss:120b-cloud
```

On the development machine the daemon runs as a **systemd user service**
(`~/.config/systemd/user/ollama.service`):

```bash
systemctl --user status ollama      # is it up?
systemctl --user restart ollama     # restart it
journalctl --user -u ollama -f      # logs
```

If the user has lingering enabled (`loginctl show-user $USER | grep Linger`) the service also
starts at boot with nobody logged in; otherwise it starts when you log in. Use the system-wide
installer if you want a boot service without a login session.

### Changing models

Any model that your Ollama instance exposes works. Edit `.env.local`:

```bash
VISION_MODEL=<model used to transcribe the photos>
TEXT_MODEL=<model used to build activities and grade short answers>
```

Current cloud models that work well (verified against the Ollama cloud in October 2026):

| Purpose | Model | Free tier? | Notes |
| --- | --- | --- | --- |
| Vision (default) | `gemma4:31b-cloud` | ✅ | Native multimodal, 256K context, current generation |
| Text (default) | `gpt-oss:120b-cloud` | ✅ | OpenAI open-weights, excellent at strict JSON, quick |
| Text, light | `gpt-oss:20b-cloud` | ✅ | Smaller/faster, slightly weaker |
| Text, cheap | `nemotron-3-super:cloud` | ✅ | 120B MoE with 12B active, very inexpensive |
| Text, smallest | `nemotron-3-nano:30b-cloud` | ✅ | Lowest cost option |
| Text, faster | `glm-5.3-flash:cloud` | 💳 credits | 18B active, 1M context, very fast |
| Text, best quality | `kimi-k3:cloud` | 💳 credits | Strongest open-weight model, slowest/most expensive |
| Vision, larger | `mistral-large-3:675b`, `deepseek-v4.1-flash` | 💳 credits | Multimodal alternatives |

> The previously suggested `qwen3-vl:235b-cloud` has been retired. The current defaults —
> `gemma4:31b-cloud` (vision) and `gpt-oss:120b-cloud` (text) — are both free-tier friendly and
> were verified end-to-end on a real photo of notes. The full list is one click away in the
> settings dialog or at <https://ollama.com/search?c=cloud>.

If your vision model is also good at generation you can point both variables at the same model —
the two pipeline steps stay separate either way.

### Checking the connection

- **In the app:** click the ⚙️ button — it shows mode, base URL, API key status, latency and
  whether each configured model is available, with a hint for what to fix. On phones the sheet
  is simplified: a **Connection** card with one status row (dot + “Connected · 2 models ready”)
  and the key field, the raw health details live behind a **Details** expander, and the key can
  be removed from a small danger button at the very bottom.
- **From the shell:** `curl http://localhost:3207/api/health`
- **Directly against Ollama:** `curl http://localhost:11434/api/tags`

---

## Rate limits

Generation is expensive (one to two model calls per request), so the API routes are limited per
client IP with a sliding window. Everything is configurable in `.env.local`:

| Route | Default | Env var |
| --- | --- | --- |
| `POST /api/generate` | 12 per 5 minutes | `RATE_LIMIT_GENERATE` |
| `POST /api/homework` | 12 per 5 minutes | `RATE_LIMIT_HOMEWORK` |
| `POST /api/grade` | 40 per 5 minutes | `RATE_LIMIT_GRADE` |
| `POST /api/homework/refine` | 30 per 5 minutes | `RATE_LIMIT_REFINE` |
| `POST /api/chat`, `POST /api/chat/extract` | 60 per 5 minutes | `RATE_LIMIT_CHAT` |
| `GET /api/health` | 90 per 5 minutes | `RATE_LIMIT_HEALTH` |
| `POST/GET/DELETE /api/sync` | 60 per 5 minutes | `RATE_LIMIT_SYNC` |
| `GET/POST/DELETE /api/push`, `POST /api/push/test` | 30 per 5 minutes | `RATE_LIMIT_PUSH` |
| window for all of them | 5 minutes | `RATE_LIMIT_WINDOW_MS` |

On top of the per-route windows there is one **account-level guard**, so no flood — from one client
or from many — can get the Ollama key throttled or banned:

| Guard | Default | Env var |
| --- | --- | --- |
| all model calls together (generate, homework, refine, chat, extract, grade) | 60 per window | `RATE_LIMIT_MODEL_TOTAL` |
| model calls in flight per client IP | 3 | `RATE_LIMIT_CONCURRENT` |
| model calls in flight across all clients | 6 | `RATE_LIMIT_CONCURRENT_TOTAL` |

A request over the concurrency cap gets a **429 `kind: "busy"`** (“Freezmo is already working on
something — try again in a moment”), not a punishment: the slot frees as soon as the model call
finishes, and the client's Retry button works. Slots are taken *after* every cheap rejection, so a
malformed request can never hold one.

Two more disk guards: `REPORTS_MAX_BYTES` (default 8 MB) stops appending to `reports.jsonl` once it
is full, and `PUSH_MAX_SUBSCRIPTIONS` (default 50) refuses new reminder devices over the cap.
Reminder endpoints must also be a real push service — `fcm.googleapis.com`,
`updates.push.services.mozilla.com`, `*.push.apple.com`, `*.notify.windows.com` — or a host you add
with `PUSH_ALLOWED_HOSTS`, because the server POSTs to whatever is stored.

When a limit is hit the route answers **429** with a `Retry-After` header and a friendly message
(“Too many requests — try again in 42s”); the UI shows that message and keeps the Retry button. The
client IP is taken from the **last** `x-forwarded-for` hop (the one your reverse proxy appended) or
from `x-real-ip`; taking the first hop would let a client spoof its own identity and dodge every
per-IP limit. Requests without those headers share one bucket.

Caveats, stated honestly: counters live **in memory**, so they reset on restart and are not shared
between multiple instances or Node workers. That is fine for a self-hosted single-instance app; if
you put Freezmo behind a reverse proxy, keep the proxy's own limits in place too.

Chat has its own allowlist: `CHAT_MODELS` (comma-separated) lists what the model chip may offer,
with the first entry as the default; when unset it falls back to `TEXT_MODEL` and `VISION_MODEL`.
The server rejects any model name outside that list.

Model names are normalised automatically: when the base URL points at `ollama.com`, Freezmo strips
the local routing suffix (`gemma4:31b-cloud` → `gemma4:31b`, `gpt-oss:120b-cloud` → `gpt-oss:120b`)
because the cloud API expects plain library names, while a local daemon wants the `-cloud` form.
The same `.env` therefore works in every mode.

## Running on your network / from your phone

The dev and start scripts already bind to `0.0.0.0` on port **3207**:

```bash
npm run dev     # next dev -H 0.0.0.0 -p 3207
npm start       # next start -H 0.0.0.0 -p 3207
```

1. Find your machine's LAN IP: `hostname -I` (Linux) or `ipconfig getifaddr en0` (macOS).
2. On your phone open `http://<your-lan-ip>:3207`.
3. Allow port 3207 through the firewall if needed
   (e.g. `sudo ufw allow 3207/tcp`, or the equivalent Windows rule).
4. Ollama only needs to be reachable from the Next.js server, not from the phone.

Notes for phones:

- **The camera button works over plain HTTP** because it uses `<input type="file" capture="environment">`,
  not `getUserMedia` (which would require HTTPS). Tapping **Camera** opens the phone's camera app
  and the photo comes back as a file.
- Clipboard paste works best on desktop; on mobile use the camera or photo picker.
- Add the page to your home screen for a full-screen, app-like experience.

### Keep it running as a service (systemd user services)

Two units keep Freezmo alive — **production** (`deploy/freezmo.service`, the
release on :3207) and **dev** (`deploy/freezmo-dev.service`, hot reload on
:3217). Both come back after a reboot and restart if they crash:

```bash
mkdir -p ~/.config/systemd/user
cp deploy/freezmo.service deploy/freezmo-dev.service ~/.config/systemd/user/
systemctl --user daemon-reload
systemctl --user enable --now freezmo.service freezmo-dev.service
loginctl enable-linger "$USER"                # start at boot without logging in
```

| Command | What it does |
| --- | --- |
| `systemctl --user status freezmo freezmo-dev` | are they up? |
| `journalctl --user -u freezmo-dev -f` | live dev logs (hot reload included) |
| `journalctl --user -u freezmo -f` | live production logs |
| `systemctl --user restart freezmo-dev` | after changing the dev unit |

Production is **never** rebuilt or restarted by hand — `scripts/deploy.sh`
does that, and only with a pushed, verified commit. The dev instance reads
`~/freezmo/.env.local`, builds into `.next-dev` and keeps its data in
`.freezmo-dev/`, so nothing it does can touch production. See
[Environments and deploy](#environments-and-deploy) and
[AGENTS.md](AGENTS.md) for the rules.

---

## Environments and deploy

Freezmo runs as two separate instances so that live work never touches what
you actually use:

| | **Dev** | **Production** |
| --- | --- | --- |
| Path | `~/freezmo` (the working tree) | `~/freezmo-prod/current` → `releases/<timestamp>-<sha>` |
| Service | `freezmo-dev.service` | `freezmo.service` |
| Port | **3217** — LAN only, hot reload | **3207** — the public URL |
| Build | `.next-dev` (dev server) | standalone server inside the release |
| Data | `~/freezmo/.freezmo-dev/` | `~/freezmo-prod/shared/freezmo/` (sync, push) |
| Env | `~/freezmo/.env.local` | `~/freezmo-prod/shared/env.local` |
| Look | DEV badge, dev favicon/theme, installs as **Freezmo Dev** | normal |

On your phone: **dev = `http://<lan-ip>:3217`**, production =
`http://<lan-ip>:3207` (and the public URL). The dev instance is bound to the
LAN only — nothing routes port 3217 through the public proxy — so source maps
and the HMR socket never leave the network. It uses a different origin, which
means separate storage and a separate service worker: the production worker is
never registered there and both PWAs can be installed side by side.

`NEXT_DIST_DIR` keeps the build folders apart (dev uses `.next-dev`, a plain
`npm run build` keeps producing `.next` for the test suites), so no build can
clobber a running server.

### Deploying

```bash
scripts/deploy.sh                  # deploy the latest origin/main
scripts/deploy.sh <commit>         # a specific pushed commit
scripts/deploy.sh --skip-tests     # fast path (lint + build only)
scripts/rollback.sh                # back to the previous release
```

`deploy.sh` refuses a dirty working tree and commits that aren't on
`origin/main`, exports the commit into a fresh release directory, runs
`npm ci`, lint and the production build there, starts the new release on a
temporary port and runs the **full browser suite against it** (the running
production is untouched), then links the shared env + data, writes
`release.json`, flips the `current` symlink atomically, restarts the service
and polls `/api/healthz` until the new commit answers. Any failure after the
switch rolls back to the previous release automatically. The last 5 releases
are kept. The running version (short commit + build time) is shown in
Settings → Details.

Data that must survive deploys lives in `~/freezmo-prod/shared/`:
`env.local`, and `freezmo/{sync,push}` (symlinked into every release as
`.env.local` and `.freezmo`).

The day-to-day rules — only dev updates during a task, production changes only
through the deploy script at the end — are written down in
[AGENTS.md](AGENTS.md) so every future session follows them.

---

## Folders

Decks are organised into folders — the same way you'd file paper notes.

- **Unorganized** is where every deck starts. It is virtual: it simply means "no folder yet".
- Create a folder from the library's folder bar (**New folder**) or from the create screen's
  **Freeze into folder** row. Names are unique, case-insensitive, and can be renamed later.
- **Freezing a deck into a folder:** pick the folder before generating (create screen) or with the
  folder row on the create screen; or use a deck's **⋯ menu → Folder** chips afterwards. The library
  shows which folder each deck is in, and groups the list by folder.
- **Filtering:** the folder bar filters the library to one folder; *All decks* shows every folder as
  its own section, Unorganized last, with per-folder counts on each chip.
- **Deleting a folder never deletes decks** — they move back to Unorganized. The confirmation says
  exactly how many decks will be re-homed.
- Folders live beside the decks in `localStorage` (`freezmo.folders.v1`), so they are per-browser
  and need no account. Exporting a deck records its folder name, and importing recreates or reuses
  that folder.

## Install as an app (PWA)

Freezmo ships as an installable Progressive Web App with iOS in mind.

Phones and tablets get an **"Install the app" bottom sheet** about a second after load (never on
desktop, never when already installed, and at most once per 7 days after a dismissal). It is a
native-feeling sheet with the Freezmo card, numbered steps and chips for the exact buttons to tap,
plus an accent confirmation pill. Close it with the X, the backdrop, Escape, or a swipe down — the
site keeps working normally. On iOS the steps adapt to the browser (Safari, Chrome, in-app), and
on Android Chrome a captured `beforeinstallprompt` turns into a real **Install** button. The sheet
can be reopened any time from **Settings → Install app** (mobile only).

It launches full-screen in standalone mode (no browser chrome), respects the notch/home-indicator
safe areas, and has an iOS-style bottom tab bar (Library · Progress · New deck · Exam · Settings)
with sheet-style dialogs. Android Chrome also offers "Install app" from the ⋮ menu; Freezmo works offline for the
parts that don't need a model.

What's included:

| Piece | Detail |
| --- | --- |
| `app/manifest.ts` | `id`, `display: standalone`, portrait, theme/background colours, `any` + `maskable` + `monochrome` icons, three app shortcuts (New deck, Make my homework, Library) and a **share target** (text/URL via GET, files via POST) |
| Icons | 192/512 px, a maskable 512 px icon, a monochrome SVG and 180/167/152 px `apple-touch-icon`s, all rendered from `assets/app-icon-source.jpeg` |
| iOS meta | `apple-mobile-web-app-capable`, `-title`, `-status-bar-style`, `viewport-fit=cover`, plus **startup images** for the common iPhone sizes so launch has no white flash |
| Service worker | `public/sw.js`, **production only**: cached app shell, stale-while-revalidate for static assets, network-first navigations with an `/offline` fallback. `/api/*` is never cached |
| Offline page | `/offline` — "your decks are still here, reconnect to generate" |
| Safe areas | `pb-safe` / `pt-safe` utilities plus `pb-tabbar` so nothing hides behind the tab bar |
| Share target | `/share` chooser (New deck / Make my homework); POSTed files wait in a short-lived in-memory store, are handed to the wizard through `sessionStorage`, and are never written to disk |

The icons are generated by `scripts/make-icons.mjs` from `assets/app-icon-source.jpeg` (headless
Chromium); it is a one-off tool and is not part of the app's dependencies. It also writes
`app/favicon.ico`, the Next.js file convention for the browser tab.

## Security notes

An honest review of what Freezmo does and does not defend against. It is a single-user,
self-hosted study app with no accounts, no database and no server-side state.

**What is hardened**

- **No secrets on the client.** The API key lives only in the browser's `localStorage` and travels
  in the `x-ollama-key` header per request. Route handlers use it for that one call and discard it —
  no disk, no log, no cookie, no environment write. Verified by grepping the whole project and the
  production `.next` output for a test key, and by restarting the server (which remembers nothing).
- **No server-side injection surface.** The Ollama endpoint and model names come only from the
  server's `.env`; the client cannot point the server at another host. Model output is validated with
  zod and rendered as text — there is no `dangerouslySetInnerHTML` except two static inline scripts
  (the pre-paint theme choice and the `beforeinstallprompt` capture), so a deck cannot inject markup
  or script (React escapes everything).
- **Key-header sanitisation.** Keys from the client are stripped of control characters/whitespace
  and capped at 512 chars before they are used in an outbound header, so a crafted key cannot inject
  headers. Verified with a raw-socket CRLF attempt.
- **Cross-origin requests are refused** on `/api/generate` and `/api/grade` (`Origin` must match the
  host), so a random website in your browser cannot make your Freezmo spend your rate limit or, in
  `server` auth mode, your credits. Origin-less clients (curl) still work.
- **Body-size limits** (32 MB for generation, 64 KB for grading) are rejected with 413 before any
  parsing, on top of the zod limits on image count, text length and field sizes.
- **Rate limiting** per client IP; see [Rate limits](#rate-limits). Auth is checked *before* parsing
  request bodies, so unauthenticated traffic costs almost nothing.
- **An account-level model budget and concurrency caps** sit in front of every route that spends an
  Ollama call (`lib/model-guard.ts`): 60 calls per window across all clients, 3 in flight per IP, 6
  in flight in total. A flood cannot pile up parallel calls or burn the account — the thing that
  gets a key throttled or banned. Slots are released when the call finishes, including a client that
  disconnects mid-stream.
- **Images must be base64 data URLs** (`dataUrlSchema`). A plain URL in an `images` field would make
  Ollama fetch an arbitrary address on our behalf, so the shape is part of the contract.
- **Reminder endpoints must be real push services** (an allowlist, `PUSH_ALLOWED_HOSTS`): the server
  POSTs to whatever is stored, so an arbitrary URL there would be an SSRF primitive.
- **Security headers** on every response: `X-Frame-Options: DENY`, `Content-Security-Policy:
  frame-ancestors 'none'`, `X-Content-Type-Options: nosniff`, a referrer policy and a
  Permissions-Policy that turns off camera/microphone/geolocation/payment (the camera button is a
  file input with `capture`, which is not gated by that policy).
- **No CSRF surface**: no cookies, no sessions, no server-side mutations.
- **Prompt injection is low impact by design.** Your notes could contain text that tries to
  manipulate the model, but the model can only return a JSON deck (validated per activity, dropped if
  invalid) — it has no tools, no file access and no network access beyond the model call.

**Known limitations (by design)**

- The key travels from your browser to your own Freezmo server, so on an untrusted network it needs
  HTTPS to stay private. On a plain-HTTP LAN someone sniffing traffic could read it, along with your
  notes.
- Anything stored in `localStorage` is readable by JavaScript on the same origin. Freezmo has no XSS
  vector today, but that is the standard risk of browser-stored credentials.
- Rate limiting keeps counters in memory: single instance only, reset on restart. The client IP is
  read from the **last** `x-forwarded-for` hop, so a client cannot spoof its identity — but behind
  more than one proxy the "last hop" is the proxy closest to Freezmo, which means those clients
  share one bucket. Point `x-real-ip` at the real client if you need per-client buckets there.
- Ollama cloud usage is billed to whoever's key is used, so treat a shared Freezmo URL as shared key
  access in `server` mode.
- No accounts means no per-user isolation: anyone who can reach the URL can use the app (and, in
  `user-key` mode, only with their own key).

## Supported input formats

Everything is parsed **client-side**, so documents are never uploaded — only the extracted text and
page images are sent to `/api/generate`.

| Format | What Freezmo reads | Sent to the model as |
| --- | --- | --- |
| JPG / PNG / WEBP / HEIC / GIF | the photo itself, resized to 1600 px JPEG | images (vision model) |
| **PDF, text layer** | text of every page (up to 80 pages) | text |
| **PDF, scanned / image-only** | pages with little or no text are rendered to images | images (vision model) |
| **DOCX** | document text + embedded images | text + images |
| **PPTX** | slide text + speaker notes + slide images | text + images |
| **ODT / ODP / ODS** | text from the OpenDocument `content.xml` | text |
| TXT / MD / CSV / TSV / LOG | the file's text | text |

Limits and behaviour worth knowing:

- up to 6 documents and 8 photos per deck, 40 MB per document, 20 MB per photo
- at most 12 images are sent (photos, rendered scanned pages and images embedded in documents)
- each document is capped at 60 000 characters and the merged text at 100 000; longer documents show
  a “truncated” note in the UI
- the activity generator reads up to 48 000 characters of notes per deck — plenty for lecture notes,
  but a whole book will only be sampled
- password-protected or corrupted files report a clear error on their card and can be removed

## How it works

```
 photos + text                                   /api/generate
 ─────────────┐
              ▼
   ┌──────────────────────────┐   step 1 (if images)   ┌───────────────────────────┐
   │ browser: resize/compress │ ─────────────────────► │ VISION_MODEL              │
   │ max 1600px JPEG data URL │                        │ transcribe + extract      │
   └──────────────────────────┘                        │ notes/concepts/formulas   │
              │                                        └────────────┬──────────────┘
              │ NDJSON progress stream                              │ clean structured text
              ▼                                                     ▼
   ┌──────────────────────────┐                        ┌───────────────────────────┐
   │ browser: study feed      │ ◄───────────────────── │ TEXT_MODEL                │
   │ JSON validated by zod    │   deck JSON            │ design 8–15 activities    │
   └──────────────────────────┘                        └───────────────────────────┘
```

- The prompts live in [`lib/prompts.ts`](lib/prompts.ts). They instruct the model to use **only the
  student's notes**, never invent facts, vary the activity types, write a short explanation for each
  activity and answer with JSON only.
- Every response is parsed (markdown ``` fences stripped), validated with **zod**
  ([`lib/schemas.ts`](lib/schemas.ts)) and — if validation fails — automatically retried **once**
  with the validation error fed back to the model (`lib/ai.ts`).
- Individual broken activities are dropped rather than failing the whole deck; if fewer than half
  the requested activities survive, the step is retried too.
- Requests use `stream: false` with per-request timeouts and a hard overall deadline
  (two sequential model calls). Errors surface as friendly messages
  ("Ollama isn't reachable, check your settings") with a retry button.
- `POST /api/generate` answers with newline-delimited JSON (`stage` / `deck` / `error` events) so the
  UI can show real progress between the vision and generation steps.

---

## Chat

Chat mode is an LLM-style conversation that knows your notes — ask about a deck, get quizzed, or
have a step explained without leaving the app.

- **Entry points.** The Library header has **Ask** (a labelled button on desktop, an icon on
  phones), every deck's menu has **Ask about this deck** (opens scoped to that deck), the desktop
  sidebar has a **Chat** row, and study/homework results have **Ask a question** (the compact
  result travels along as extra context). On phones the chat is a full-screen route with a back
  button and **New chat** — the tab bar steps aside so the composer gets the space. Desktop keeps
  the sidebar.
- **Grounded answers, retrieved in the browser.** Chat reads each deck's stored notes text
  (`sourceText` — the same text *Regenerate* uses, already kept and size-capped per deck). The
  client chunks the scoped decks (~800 tokens per chunk), ranks the chunks against your question
  and the recent turns with a small BM25 scorer, and sends only the top chunks (~7k tokens, each
  labelled with its deck and section). The server stays stateless: the chosen context and the
  conversation window travel with every request, exactly like the API key. Decks without notes text
  fall back to their generated study content, and the chat says so.
- **Scope.** The context chip picks **All notes**, **selected folders/decks** or **No notes**
  (general chat); opening from a deck defaults to that deck. A Settings toggle (**Keep notes text
  for chat**) makes chat ignore the notes text, and each deck's menu has **Delete notes text** to
  remove it from that deck entirely.
- **Models.** The model chip lists the server's allowlist (`CHAT_MODELS`, defaulting to the
  configured text and vision models) with text/vision badges and availability from the health
  check; the last choice is remembered. The server validates every request against the same
  allowlist and never accepts a model name from the client.
- **Attachments.** Photos, PDFs and documents go through the existing vision/extraction step first
  (plain text), so any chat model works — text-only included.
- **History.** Conversations live in a local IndexedDB store: the list sits in the main sidebar on
  desktop (it takes the deck navigation's place on `/chat`), a sheet on phones, with auto titles
  from your first question, rename, delete (with Undo) and a search across titles and message
  text. Conversations remember their scope and model, reopen from `?c=<id>` (so a reload continues
  where you left off), stay readable offline, and are included in the backup export and the
  encrypted sync — still never the key.
- **From an answer.** Every answer can become a deck (**Turn into flashcards** opens New deck with
  the answer as notes, on the options step) or be appended to the scoped deck's notes (**Save to
  notes**, only offered when exactly one deck is in scope).
- **Streaming and controls.** Tokens stream with a caret and a stop button; auto-scroll pauses when
  you scroll up and a jump-to-bottom button appears; answers render safe markdown (lists, tables,
  code blocks with copy — no HTML injection) and KaTeX loads dynamically only when a message
  contains math; reasoning models' thinking is hidden behind a collapsible **Thinking** row and
  never mixed into the answer. Every answer offers copy / regenerate, the last question edit and
  resend. The composer auto-grows at 16px, sends on Enter on desktop (newline on phones), tracks
  the on-screen keyboard through `visualViewport`, respects reduced motion and keeps the shared AI
  disclaimer under the chips.
- **Privacy.** Notes text and conversations stay in this browser; nothing is stored server-side.
  The API key still only ever travels in the `x-ollama-key` header to Freezmo's own route for that
  single request.

---

## Project structure

```
app/
  layout.tsx                 root layout, iOS meta tags + theme bootstrap script
  manifest.ts                web app manifest (installable PWA)
  offline/page.tsx           offline fallback served by the service worker
  globals.css                Tailwind v4 theme, frosty palette, animations, reduced-motion
  page.tsx                   library ("Your icebox")
  create/page.tsx            upload screen
  deck/[id]/page.tsx         study feed for a saved deck
  homework/page.tsx          “Make my homework” (assignment in, homework or guide out)
  chat/page.tsx              chat mode (full-screen on phones, sidebar on desktop)
  api/generate/route.ts      images/text in → NDJSON progress → deck JSON out
  api/homework/route.ts      assignment in → NDJSON progress → homework/guide JSON out
  api/grade/route.ts         short-answer grading
  api/chat/route.ts          grounded chat turns → NDJSON token/thinking stream
  api/chat/extract/route.ts  chat attachments → plain text (vision step)
  api/health/route.ts        Ollama reachability + model availability
components/
  activities/                one component per activity type + shared shell
    Flashcard.tsx MultipleChoice.tsx TrueFalse.tsx FillBlank.tsx
    Matching.tsx Ordering.tsx ShortAnswer.tsx shared.tsx index.tsx
  homework/                make-my-homework screens (input, result, guide feed, history)
  chat/                    chat screen, composer, message rows, markdown + math, pickers
  UploadScreen.tsx StudyFeed.tsx ResultsScreen.tsx DeckCard.tsx ApiKeyGate.tsx
  InstallSheet.tsx TabBar.tsx ServiceWorkerRegistrar.tsx FolderBar.tsx
  LibraryScreen.tsx SettingsDialog.tsx AppHeader.tsx ThemeToggle.tsx icons.tsx ui.tsx
lib/
  ollama.ts        server-only HTTP client (timeouts, auth, error mapping)
  prompts.ts       the vision, generation, grading, homework and repair prompts
  ai.ts            every pipeline step with zod validation + retry
  schemas.ts       zod schemas and shared types
  config.ts        env config, cloud/direct model-name normalisation
  documents.ts     client-side PDF/DOCX/PPTX/ODT/text extraction
  folders.ts       folder library (create/rename/delete) in localStorage
  install.ts       install-prompt platform/browser detection + 7-day dismissal
  settings.ts      browser-side settings (the user's API key) + auth headers
  use-health.ts    client hook for /api/health, re-runs when the key changes
  rate-limit.ts    in-memory sliding-window limiter for the API routes
  http-guard.ts    same-origin check, body-size cap, API-key sanitiser
  language.ts      script/language checks that catch translated transcriptions
  ui-events.ts     tiny event bus (any screen can open the settings dialog)
  storage.ts       localStorage deck library, import/export
  homework.ts      localStorage homework results + solution assembly
  use-notes-intake.ts  shared client-side photo/document intake hook
  clipboard.ts     copy helper with a non-secure-context fallback
  image.ts         client-side validation, resize and compression
  client-api.ts    browser client for the API routes (NDJSON reader)
  chat.ts          chat types + last-model memory
  chat-context.ts  client-side chunking + BM25 retrieval over the scoped decks
  chat-client.ts   streaming chat client + attachment extraction
  chat-seed.ts     one-shot hand-off from results into the chat
  chat-trim.ts     token budgeting + the sliding history window (server + client)
  xp.ts utils.ts deck-ops.ts confetti.ts
public/
  sw.js                  service worker (production only)
  icons/                 PWA + iOS icons and splash screens
assets/
  app-icon-source.jpeg   master artwork the icons are rendered from
deploy/
  freezmo.service       systemd user unit (auto-start at boot)
scripts/
  copy-pdf-worker.mjs   copies the pdf.js worker to /public (runs on predev/prebuild)
  make-icons.mjs        regenerates the PWA/iOS icons from the master (one-off tool)
```

---

## API routes

| Route | Body | Returns |
| --- | --- | --- |
| `POST /api/generate` | `{ images?: string[], text?, difficulty, count?, focus?, language?, mode?, existing? }` | NDJSON stream: `{type:"stage"}`, `{type:"deck",deck,meta}`, `{type:"error"}` |
| `POST /api/homework` | `{ mode: "make" \| "guide" \| "check", images?: string[], text?, solution?, language? }` | NDJSON stream: `{type:"stage"}`, `{type:"homework",homework,meta}`, `{type:"error"}` |
| `POST /api/homework/refine` | `{ instruction: "shorter" \| "more" \| "explain", text, context?, language? }` | `{ text }` — one rewritten piece, plain text |
| `POST /api/grade` | `{ question, referenceAnswer, keyPoints, userAnswer, language? }` | `{ correct, score, feedback, missing }` |
| `POST /api/chat` | `{ model?, messages: [{role,content}], context?: [{deckId,deckTitle,section,text}] }` | NDJSON stream: `{type:"token"}`, `{type:"thinking"}`, `{type:"done",model}`, `{type:"error"}` |
| `POST /api/chat/extract` | `{ images?: string[], text? }` | `{ text }` — attachments reduced to plain text via the vision step |
| `POST /api/reports` | `{ deckId, deckTitle?, activityId, type, question, at }` | `{ ok: true }` — appended as one JSON line to `REPORTS_DIR/reports.jsonl` (write-only; see decision 52) |
| `GET /api/health` | – | `{ ok, reachable, mode, baseUrl, hasApiKey, vision, text, chatModels, latencyMs, modelCount, error }` |
| `POST /api/sync` | `{ id, envelope }` (opaque AES-GCM blob) | `{ updatedAt }` — 413 over `SYNC_MAX_BYTES`, 507 over `SYNC_TOTAL_BYTES` |
| `GET /api/sync?id=` | – | `{ updatedAt, envelope }` or 404 |
| `DELETE /api/sync?id=` | – | `{ ok: true }` (revokes the synced copy) |
| `GET /api/push` | – | `{ enabled, publicKey? }` — reminders are hidden when `enabled` is false |
| `POST /api/push` | `{ endpoint, keys }` | `{ ok, count }` (stored server-side, endpoint + keys only) |
| `DELETE /api/push?endpoint=` | – | `{ ok, count }` |
| `POST /api/push/test` | `{ endpoint }` | `{ ok: true }` or 502 when the push service refuses |

The model routes accept an optional **`x-ollama-key`** header (the browser's own key; required in
`user-key` auth mode). Missing keys answer **401** with `kind: "needs_key"`, a rejected key answers
`kind: "auth"` (or `kind: "no_credits"` for HTTP 402), and exceeding a rate limit answers **429**
with `Retry-After`. Keys are used for the single request and discarded.

Images are sent as base64 data URLs; the server strips the prefix before handing them to Ollama.
Uploads are validated for type (photos and the document formats above), count (max 8 photos,
6 documents, 12 images) and size (max ~8 MB per image after compression, 40 MB per document) and the
client compresses aggressively, so a typical upload is 1–3 MB. `language` is optional wherever it
appears: the vision step fills it in automatically, and the UI passes the stored deck language for
grading, regeneration and “add more” (and the assignment language for homework).

---

## Scripts

| Script | What it does |
| --- | --- |
| `npm run dev` | `next dev -H 0.0.0.0 -p 3207` |
| `npm run build` | production build (type-check + lint) |
| `npm start` | `next start -H 0.0.0.0 -p 3207` |
| `npm run lint` | ESLint |
| `npm run version:bump` | bump the app version one step (`0.1.7` → `0.1.8`, `0.1.9` → `0.2.0`); one bump per task |
| `node scripts/make-icons.mjs` | regenerate the PWA/iOS icons (needs a Playwright Chromium) |
| `node scripts/copy-pdf-worker.mjs` | copy the pdf.js worker (also runs on predev/prebuild) |

---

## Data, storage and privacy

- Decks live in **`localStorage`** under `freezmo.decks.v1`, folders under `freezmo.folders.v1` (no
  accounts, no server database).
  Stored per deck: title, subject, summary, language, difficulty, the extracted **source text**,
  activities, play count and best score.
- **Original files are never stored or uploaded** — photos and documents are parsed in the browser,
  and only the extracted text (plus page images for the vision model) is sent to your Ollama
  endpoint. That is why *regenerate* and *add more activities* work from the source text rather than
  the original files. Imported decks that contain no source text can't be regenerated; the UI says so.
- Documents are parsed locally by `pdfjs-dist` (PDF), `mammoth` (DOCX), `jszip` (PPTX/OpenDocument)
  and the browser's own `File.text()` — nothing is uploaded for parsing.
- Images and rendered PDF pages are sent to whatever Ollama endpoint you configure. In local-daemon
  mode they go to your machine first; in direct-cloud mode they are processed by Ollama's cloud.
- In the default `user-key` mode your Ollama API key is also kept in `localStorage` (key
  `freezmo.settings.v1`) and sent with each API call so the server can forward it to Ollama. It is
  never written to the server, a log, a cookie or the environment — see
  [Ollama setup](#ollama-setup) for the full story.
- Clearing browser data deletes the library and the saved key. Use **Export all data**
  (Settings → Data & backup) for anything you want to keep; the file never contains the API key.
- **Opt-in sync** (`freezmo.sync.v1` + a non-extractable `CryptoKey` in IndexedDB) uploads an
  AES-GCM envelope to `POST /api/sync`. The server writes one small JSON file per sync ID under
  `SYNC_DIR` (default `.freezmo/sync`), can't decrypt it, and never receives the passphrase or the
  key. Deleting the synced copy is one button in Settings (and works even on a device that can't
  unlock the data).
- **Question reports** (the red **!** in the study feed) keep the skip state in this browser
  (`freezmo.reports.v1`) and append one JSON line per report to `REPORTS_DIR/reports.jsonl`
  (default `.freezmo/reports`, `shared/freezmo/reports` in production). Question text is part of
  the report on purpose — it is what makes the file useful for tuning the generation prompts — and
  the route is deliberately write-only over HTTP.
- **Answer history lives in IndexedDB** (`freezmo-learning`, versioned stores `answers` + `srs`),
  because an event log would eventually fill `localStorage`; the log is capped at 20 000 answers
  (oldest pruned). Backups and sync include it, so progress survives a device move. Push
  subscriptions are the only reminder state on the server (`PUSH_DIR`, endpoint + keys only) and
  are removed when you turn reminders off — or automatically when the push service reports them
  gone.

---

## Troubleshooting

| Symptom | Fix |
| --- | --- |
| “Ollama isn't reachable at http://localhost:11434” | Start the daemon (`ollama serve`, or `systemctl --user start ollama` for a user-space install), or fix `OLLAMA_BASE_URL`. Check `curl http://localhost:11434/api/tags`. |
| A document card is stuck on “Reading page …” | The pdf.js worker may be missing: run `npm run dev`/`npm run build` (they copy it to `public/pdf.worker.min.mjs`) or `node scripts/copy-pdf-worker.mjs` manually. |
| PDF uploaded but no activities about its content | The PDF is probably image-only (a scan). Freezmo renders such pages and sends them to the vision model — check the card note (“N scanned pages were turned into images”); if it says “No readable text”, the scan may be too low quality. |
| Deck came back in English even though the notes are not | Make sure both model calls succeed (a failed vision step falls back to raw text). The prompts force the notes' language; if you still see translation, switch `TEXT_MODEL` to a stronger instruction-follower such as `gpt-oss:120b-cloud`. |
| “You've run out of your free monthly Ollama usage” | The account's free monthly usage is used up — this is not a short-term rate limit. Upgrade the account at `ollama.com/settings`, or create a new account and paste its key in Settings (the button jumps straight to the field). “Ollama is busy” is the short-term variant: just press Retry. |
| Reminders are missing from Settings | They only appear when `VAPID_PUBLIC_KEY` and `VAPID_PRIVATE_KEY` are set in `.env.local` (generate with `npx web-push generate-vapid-keys`). On iPhone, add Freezmo to the Home Screen first — Safari tabs can't receive push. |
| “Add your Ollama API key first” | Bring-your-own-key mode is active. Paste a key from <https://ollama.com/settings/keys> on the key gate (or in the ⚙️ dialog) — it is stored only in your browser. |
| “Ollama rejected this API key” | The key is invalid or revoked. Create a new one and paste it again; the check uses a real 1-token request, so a green tick means it truly works. |
| “this model is not included in your free usage” | The key works, but the model is outside your plan. Use one of the ✅ models in the table above, or add credits at <https://ollama.com/settings>. |
| “Too many requests — try again in Ns” | You hit a rate limit (12 decks / 40 grades per 5 min by default). Wait, or raise `RATE_LIMIT_*` in `.env.local`. |
| “set to OLLAMA_AUTH=local but points at https://ollama.com” | Fix the combination: use `OLLAMA_AUTH=user-key`/`server` for the cloud URL, or point `OLLAMA_BASE_URL` at `http://localhost:11434`. |
| Settings dialog says a model is “not found” | Local daemon: pull it once — `ollama pull gemma4:31b-cloud` (cloud pulls are just manifests). Direct cloud: check the model name and that your key is valid. |
| 401/403 from `https://ollama.com` | Create a key at <https://ollama.com/settings/keys> and put it in `OLLAMA_API_KEY` (no quotes), then restart. |
| Daemon logs / it keeps dying | User-space install: `journalctl --user -u ollama -f`. Manual start: `~/.local/ollama/bin/ollama serve`. |
| Generation times out | Use fewer/smaller photos, choose a faster `TEXT_MODEL` (e.g. `gpt-oss:20b-cloud`), or raise `OLLAMA_TIMEOUT_MS`. |
| Responses are slow | Thinking models spend tokens on reasoning. Set `OLLAMA_THINK=false` to send `think: false`, and/or use a smaller model. |
| “did not match the expected format” | Freezmo retried once automatically. If it still fails, the model is poor at JSON — switch to a stronger `TEXT_MODEL` (e.g. `gpt-oss:120b-cloud`). |
| Blank/boxed icons on Linux | The app chrome uses inline SVG, but activity badges use emoji. Install an emoji font (`fonts-noto-color-emoji`). |
| Port 3207 already in use | `lsof -i :3207` and stop the other process, or change the port in both scripts in `package.json`. |
| Camera button missing on mobile | It's in the *Add your notes* card; some in-app browsers hide file inputs — open the page in Chrome/Safari directly. |
| Hydration warning in dev | Usually a stale dev artifact: hard-reload. Decks load in `useEffect`/`useSyncExternalStore`, so no library content is server-rendered. |

---

## Decisions and notes

Things I chose while building, with the reasoning:

1. **Defaults follow the current Ollama catalogue and the free tier.** The suggested
   `qwen3-vl:235b-cloud` has been retired, so the defaults are `gemma4:31b-cloud` (vision) and
   `gpt-oss:120b-cloud` (text/JSON/grading) — both verified as included in the free plan on a real
   photo of notes. Paid/credit models such as `glm-5.3-flash:cloud` and `kimi-k3:cloud` are
   documented as alternatives.
2. **One `.env` for both modes.** Direct-cloud names differ from daemon names (`gemma4:31b` vs
   `gemma4:31b-cloud`), so `lib/config.ts` strips the `-cloud` / `:cloud` suffix automatically when
   `OLLAMA_BASE_URL` points at `ollama.com`.
3. **Two separable AI steps.** Photos always go through `VISION_MODEL` for transcription/extraction;
   the structured text then goes to `TEXT_MODEL`. Pointing both variables at one model works too.
   Text-only requests skip the vision step entirely.
4. **NDJSON progress stream instead of a single long request.** The two model calls are sequential
   and slow, so `/api/generate` streams `stage` events; the UI shows real stages rather than a fake
   timer. `stream: false` is still used for each Ollama call for simplicity and reliability.
5. **zod + one repair retry, plus per-activity salvage.** A single bad activity shouldn't cost you
   the whole deck: valid activities are kept, and the repair round is only used when too few survive.
6. **Flashcards count as "correct" via self-assessment** (I knew it / Not yet) so that scoring, XP
   and "retry missed ones" apply to every activity type.
7. **Fill-in-the-blank is graded locally** (case/punctuation-insensitive, one typo forgiven,
   `/` alternatives in the answer) — no round trip needed. Only short answers call `/api/grade`.
8. **Matching requires a perfect match for full XP**, otherwise it's marked as missed with an
   explanation of how many wrong attempts you made.
9. **Source text is stored with each deck** so *regenerate* and *add more* work without keeping the
   original photos (storage/privacy win). Imported decks have no source text, so those two actions
   are disabled for them.
10. **XP is assigned server-side** (easy 10 / medium 15 / hard 20, +5 for matching, ordering and
    short answer) so the model can't invent inconsistent rewards.
11. **Documents are parsed in the browser, not on the server.** pdf.js, mammoth and jszip run
    client-side, so the raw file never leaves the device and the API stays a thin text/image pipe.
    Scanned PDF pages are rendered to JPEG for the vision model, while PDFs with a text layer skip
    the vision step entirely (faster and exact).
12. **The notes' language is detected once and then carried everywhere.** The vision step reports a
    `language` field, the deck stores it, and generation, grading, regenerate and add-more all
    receive it explicitly — plus “never translate” rules in every prompt. Verified with Czech
    notes across all three input paths (text PDF, scanned PDF, DOCX + PPTX).
13. **Keys are per browser and never stored server-side.** Freezmo defaults to
    `OLLAMA_AUTH=user-key`: the key is typed into the UI, kept in `localStorage` and sent in the
    `x-ollama-key` header, and the route handlers use it for that request only — nothing is written
    to disk, logs, cookies or the environment (verified by grepping the whole project for a test key
    and by restarting the server). `server` and `local` modes remain available by env var.
14. **Key verification uses a 1-token chat probe, cached for 5 minutes.** Model listing on
    `ollama.com` is public, so a bogus key would look valid; only a real authenticated request can
    tell. HTTP 402 (“not included in your free usage”) is reported as a *plan* problem rather than a
    broken key.
15. **Rate limits are per client IP, in memory, and generous by default** (12 decks / 5 min, 40
    grades / 5 min). They exist to stop runaway loops and accidental abuse, not to be a security
    boundary — the README says so, and the counters are intentionally not persisted.
16. **A misconfiguration is caught early:** `OLLAMA_AUTH=local` combined with `https://ollama.com`
    returns an explanatory error instead of a confusing 401.
17. **Folders are a first-class, non-destructive idea.** `UNORGANIZED_ID` is virtual (decks simply
    have no `folderId`), deleting a folder re-homes its decks instead of deleting them, and folder
    names travel in exports so an imported deck lands where it came from. 13 browser checks cover
    create → freeze → move → rename → delete → reload → import.
18. **The visual language is frosted ice glass.** Glacier cyan to ice blue (`#38bdf8` → `#0ea5e9`)
    is the accent for focus rings, progress bars and selected states; the primary CTA is an icy
    gradient (deep blue in light, white-to-`#bfe3ff` in dark). The background is a fixed ice scene;
    the sidebar, cards, chrome, sheets and gate are translucent glass with a bright rim and a soft
    blue shadow. Colour appears in activity tints and semantic feedback only — pretty, but still
    readable and quiet.
19. **Custom CSS lives in `@layer components`, and glass uses Tailwind's backdrop utilities.**
    Two build-level traps, both found by looking at real screenshots: unlayered rules beat Tailwind
    utilities (which silently broke `fixed` on the feed), and a hand-written `backdrop-filter` gets
    minified down to `-webkit-` only — which painted no blur at all. `@apply backdrop-blur-*`
    emits both properties correctly.
20. **There is no demo/sample deck.** Every activity comes from the user's own notes; the empty
    state points at the create flow and Settings instead of shipping fake content. Browser tests seed
    a fixture deck into `localStorage` rather than relying on app code.
21. **Dark mode is token-driven, and contrast is audited on glass.** Text colours come from
    `--text` / `--text-muted` / `--text-subtle` (never hardcoded), and a `contrast.mjs` audit checks
    every pairing at WCAG AA against the *effective* frosted background — currently **18/18 pairs
    pass** (e.g. dark secondary `#b3c4e0` on glass `#1c2535` = 10:1, white on the light CTA
    mid-gradient `#1a5fd0` = 5.6:1). Difficulty colours were darkened until they cleared 4.5:1 on
    their pills.
    This was a real bug: an earlier edit silently failed, so `.dark` never overrode `--text`, leaving
    near-black text on a dark page.
22. **The dashboard is one inset glass rail plus glass cards.** The sidebar is a single tall
    rounded glass panel inset from the viewport edges, so the ice scene shows around it; deck cards
    are 24px-radius glass panels with a secondary Study CTA and quiet metadata. Headings are
    semibold with tighter tracking. Nested glass (chips, inputs) is more opaque but never adds a
    second blur layer. The collapse/expand arrow keeps one home: its own row directly below the
    logo in both states (labelled when expanded, icon-only and centred when collapsed), instead of
    swapping between the brand row and the middle of the icon column.
23. **One fixed ice scene instead of per-page blobs.** The decorative background used to be
    rendered inside the content column, so on wide screens it was clipped into a visible hard edge.
    It now lives in a single `position: fixed` layer (`.ice-scene`, `z-index: -1`, pointer-events
    none): base gradient, four drifting glacier blobs, five glossy ice orbs and an SVG-noise grain.
    `prefers-reduced-motion` stops the drift. The four blobs were then turned down to a quarter of
    their opacity (0.18 → 0.045 light, 0.40 → 0.10 dark): at full strength the huge circles read
    as UI rather than as backdrop.
24. **The UI was deliberately made quieter.** One accent colour (`accent-*`), a desaturated
    `frost-*` scale, hairline borders and flat surfaces replaced the saturated gradients and the
    seven-colour activity badges. Semantic colours (green/amber/red) survive only where they carry
    meaning: correct/wrong feedback, key status, rate limits.
25. **PWA instead of a native wrapper.** A manifest, iOS meta tags, generated icons, a
    production-only service worker, an offline page and a standalone-aware layout give the
    install-to-home-screen experience without adding a runtime dependency — and the app still works
    as a plain website.
26. **The API is defended like a LAN service.** Cross-origin POSTs are refused, bodies are capped
    before parsing, keys are sanitised before they reach a header, and auth is checked before any
    body work. See [Security notes](#security-notes) for the full list and the known limitations.
27. **A language guard repairs accidentally translated transcriptions.** `TEXT_MODEL` (`gpt-oss:120b`)
    is the strongest multilingual model in the free tier, so when the transcription check
    (`lib/language.ts`: expected script vs. actual script, distinctive diacritics, English stopword
    density) flags a mismatch, the text model rewrites the JSON back into the original language
    before generation. Verified both ways: it fires on a synthetic English-but-declared-Czech
    transcription and never fires on real Czech notes (no extra calls, no added latency).
28. **The first screen is the API-key gate.** In `user-key` mode a browser without a key gets a
    full-screen, monochrome gate instead of the dashboard: one masked input, one Continue button,
    and a "?" sheet with the steps to a free key. The gate reuses the cached 1-token probe through
    `/api/health`, so an invalid key and an unreachable Ollama are worded differently. The old
    dashboard banner is gone; key management stays in Settings, and removing the key brings the
    gate back. The gate covers `/`, `/create` and deck study, but never `/offline`.
29. **The install prompt is a teaching sheet, not a nag.** It appears once, phones/tablets only,
    a second after load, and never in standalone mode; dismissing it silences it for 7 days. It
    detects iOS vs Android and Safari vs Chrome vs an in-app browser so step 1 always matches the
    device, uses Chrome's deferred `beforeinstallprompt` for a one-tap install where available, and
    stays reachable from Settings. It is mounted in the root layout, so it can appear over any
    screen — including the key gate — without blocking the page underneath.
30. **Phones use touch-action, not a zoom lock.** The viewport allows pinch-zoom (accessibility),
    and double-tap zoom is disabled by `touch-action: manipulation` on the UI chrome instead. A
    mobile-only rule keeps every input and textarea at 16px so iOS Safari's focus auto-zoom never
    fires (iOS ignores `user-scalable=no` for that anyway). Desktop keeps the designed 12–14px
    field sizes.
31. **The app icon drives the palette.** The navy rounded-square artwork in
    `assets/app-icon-source.jpeg` is the source of truth: `frost-*` is navy-tinted, the accent is
    glacier cyan, the ice scene uses the artwork's blues, the primary CTA is an icy gradient, and
    every icon/splash/favicon is rendered from it by `scripts/make-icons.mjs`. The sidebar, header, key gate, empty state and install sheet all show
    the real artwork instead of the old monochrome snowflake tile.
32. **The ice scene was profiled, and the blur moved off the animated layer.** A mid-range phone
    viewport with 4x CPU throttling showed the drifting blobs were the only expensive part: a
    `filter: blur(100px)` on a huge element that is also being transformed forces a
    re-rasterisation every frame (150ms vs 33ms per frame in a software renderer). The blobs are now
    soft radial gradients with no filter — visually the same, 4-5x cheaper — phones get two blobs
    instead of four and a 16px (instead of 24px) glass blur, and nested glass never blurs. Result:
    60fps idle and ~17ms median frames while scrolling under 4x throttle.
33. **Interface copy lives in two flat dictionaries, not in a framework.** `lib/i18n-dict.ts` holds
    the `en` and `cs` tables plus `translate()` / `translatePlural()`; `lib/i18n.ts` adds the browser
    store (`freezmo.locale.v1`), the `useI18n()` hook and the `x-freezmo-locale` header that API
    routes read to answer in the same language. English is the default and the fallback for any
    missing key. Czech uses the real plural rules (1 sada / 2–4 sady / 5+ sad). Crucially, the hook's
    `t` is *bound to the locale React is rendering with* — the first client render matches the
    server HTML, then re-renders in the stored language; using the module-level `t` during render
    instead caused a hydration mismatch that silently dropped the `dark` class. Module-level `t` is
    only for event handlers and non-React helpers (documents, folders, API errors). The notes' own
    language is untouched by this — decks still come back in the language of your notes.

34. **The mobile pass was a layering *perception* fix plus a real contrast debt.** The ice orbs were
    never above the content — the scene is a fixed `pointer-events:none` layer at `z-index:-1`
    (verified with `elementFromPoint`), but it shines through the translucent page and glass. Phones
    now get a thinner scene (lower blob/orb opacity, orbs tucked to the edges, the bottom-right orb
    moved out from under the tab bar) and the tab bar paints `color-mix(page 84%)` so its labels keep
    their contrast. Helper text moved onto the tokens (`--text-subtle` is #9db1d0 in dark: 8.6:1 on
    the page, 7.2:1 on dark glass) and disabled buttons are a visible muted chip (`.btn-disabled`)
    instead of `opacity-40`. Buttons share one ice gradient in both themes — which also exposed that
    the danger/success variants never showed their colour, because the primary `background-image`
    painted over `!bg-rose-600` (`bg-none` now strips it). Light mode got a cool card rim instead of
    white-on-white. The settings sheet swipes away like the install sheet (drag only while the body
    is scrolled to the top), its body scrolls to the true bottom with a fading scroll cue, and the
    API-key note is split into a short privacy line plus a collapsible local-daemon tip. axe-core
    reports every glass surface as “incomplete”, so contrast is verified by a pixel-sampled script
    (hide the text → screenshot → worst-case background luminance per glyph rect): 0 runs below AA
    across library and settings, both themes, both languages.

35. **“Make my homework” reuses the study pipeline instead of growing its own.** The intake is the
    same client-side parsing as `/create` (lib/documents + lib/image, wrapped in `useNotesIntake`),
    the photos go through the same `extractNotes` vision step and language guard, and the guide's
    checkpoints are ordinary activity JSON rendered by the existing components — just a
    `checkpoint` field on a step. The route streams NDJSON like `/api/generate` and has its own
    rate limit. Two things are deliberately different: the finished answer and the guide are
    written in the *assignment's* language (the interface language never leaks in), and the final
    solution is assembled on the client from the step answers instead of asking the model for a
    second copy. Result records store the assignment text — the vision transcription when photos
    were used — so “Make it again” works without storing a single original file.

42. **Check my work is a checker, not a second writer.** The third homework mode sends the
    assignment and the student's attempt to one model call with its own system prompt
    (`HOMEWORK_CHECK_SYSTEM_PROMPT`): the output is a verdict, a list of what is right, a list of
    issues with *why* and *fix*, and the corrected version. Unasked-for extras are explicitly
    never errors, and the corrected version keeps everything the student got right as they wrote
    it. The attempt is typed or pasted — supporting photos of handwriting would need a second
    vision pass and a way to separate assignment from attempt in the same image, so the UI says
    plainly that it isn't supported yet. Follow-ups are a separate, cheaper route
    (`POST /api/homework/refine`, plain JSON, `RATE_LIMIT_REFINE`) shared by the result, the check
    and each guide step; they rewrite one piece and never touch the stored record. “Turn into a
    study deck” hands the content to the New deck flow through sessionStorage (one-shot, like the
    share target), with the options step pre-filled. Verified by the phase-3 suite (20 checks
    against a stub Ollama: the check payload, the missing-solution 400, refine validation, the
    phone and desktop check flows, the shorter follow-up with restore, the guide's explain action,
    and the deck hand-off).

43. **Phones navigate with five tabs, not a drawer.** The drawer, its hamburger, its
    drag-to-close gesture and its back-gesture layer are gone (`MobileSidebar`, `sidebarOpen`,
    `MenuIcon` and the three drawer strings deleted); the desktop rail is untouched and is now the
    only consumer of the shared nav content. The bottom bar is **Library · Progress · New deck (the
    prominent centre +) · Exam · Settings**, keeping the sliding pill, press feedback, re-tap-to-top
    and safe-area padding. Tab labels use dedicated short `tab.*` keys because the old names
    (“Zkouškový režim”) can't fit five-up at 360px; the matrix in the mobile suite checks every
    label against clipping at 360/390, EN/CS, dark/light and with reduced motion on and off. The
    wizards keep their own bottom action bar, so the tab bar steps aside there, exactly as before.
    Drawer-only destinations were re-homed: Progress and Exam became tabs, the New deck screen got
    a **Study deck ⇄ Make my homework** segmented switch (also on `/homework`), and the library's
    phone row gained a **⋯** sheet with Import JSON + New folder (folder chips still appear only
    once a folder exists). The mode switch is hidden while either wizard is generating
    (`lib/wizard-busy.ts`), and leaving a wizard mid-run aborts the request so a stray generation
    can never save a deck or yank the user into the study feed. Verified by the mobile suite
    (134 checks: no hamburger/drawer, tab order, the label matrix, per-tab routing, deep links,
    Back from every tab and the ⋯ sheet) plus the wizard suite (92 checks, including the
    loading-lock and abort-on-leave cases) and the other standing suites.

44. **Smart Review must never look like a dead tap.** The reported symptom was a tap on *Review
    due* sitting for many seconds with no feedback. The flow itself is local and fast (the counts
    are already hydrated on the library; measurements show <1.5s even at 6× CPU throttle with
    3,000 answers), so the visible fault was the missing feedback: `/review` is a dynamic route
    (search params), and until its payload arrived the screen simply didn't change. Three layers
    now cover it: the card links **prefetch** their payload; the link's own **`useLinkStatus`**
    pending state renders a full-screen loading screen (`LinkLoadingScreen`) the instant the tap
    happens — measured at **58 ms** even with the route request held open; and `loading.tsx`
    boundaries cover the render/stream phase (review, progress and exam). The client screen's own
    bare skeleton was replaced with the same loading screen. While in there, health probes got a
    per-key 30-second cache plus in-flight sharing (`use-health.ts`): every page mounts its own
    gate and each probe can sit for seconds while Ollama answers, so a session now fires **one**
    probe instead of one per navigation (verified). Also fixed a Czech typo on the exam heading
    (“Zkouškové režim” → “Zkouškový režim”). Verified by the review suite (15 checks:
    stalled-request loading screen at 58 ms, session arrival after release, the normal fast path,
    the mistakes path, end-to-end grading, and the probe cache).

45. **The browser suites live in the repo, not in `/tmp`.** The harness was built over several
    sessions in a scratch directory that a reboot wiped; it now lives in `tests/` and runs with
    `npm run test:e2e` against the running production build (Playwright + axe-core as dev
    dependencies, Chromium auto-detected or via `CHROME`). Ten suites cover the core app
    (regression), the phone tab bar and sheets (mobile), the wizards (wizard), PWA/offline/share
    (native), i18n, the freeze loading screen, Smart Review, and phases 1–3 (backup/sync/usage
    limits, the learning engine, homework check/follow-ups). Suites that need a stub Ollama or
    VAPID keys start their own `next start` instance on a spare port, verify the port is free and
    kill the process group on exit; test scratch stays in `tests/.tmp/` (gitignored). Twelve suites,
    513 checks.

46. **The ice orbs are rising bubbles now, composited and cheaper to read through.** The glossy
    orbs (the glacier blobs stay exactly as they were) rise from below the bottom edge to above the
    top edge, each with its own 36–68 s pass (slower on phones), a negative delay so the scene is
    already populated on load, its own horizontal position, a gentle sway and a slight scale drift;
    opacity fades in/out near the edges so the loop never pops. Only `transform` and `opacity`
    animate, each orb owns a composited layer (`will-change: transform`), there is no
    filter/backdrop-filter on them, and nothing animates size, offsets or shadow. Opacity dropped
    ~38% in both themes (0.55→0.34 light, 0.8→0.5 dark; phone overrides 0.32→0.2 / 0.42→0.26) so
    the background reads through while the gloss stays soft. Under `prefers-reduced-motion` the
    rise is removed and the orbs return to their old static placements — a paused-from-the-start
    CSS animation is not applied at all in Chrome (`getAnimations()` is empty), which would have
    parked them below the fold; while the tab is hidden a visibility listener parks the whole scene
    (`.ice-paused`), which also freezes the blob drift. Measured with a CDP performance trace on a
    4× throttled phone profile (390×844 @2x): p95 33.4 ms, zero frames over 34 ms and no long tasks
    — identical to the pre-change baseline — for ~0.6% extra main-thread CPU; the desktop scene
    stays compositor-bound from the pre-existing blob drift, exactly as before. Verified by the
    bubbles suite (34 checks: scene stacking/layout contract, keyframes only transform+opacity,
    rise/sway/scale/fade phases, per-breakpoint counts 5/3/2, phone pacing, reduced motion,
    tab-hidden parking) plus the 16-run axe contrast matrix (360/390/1280 × dark/light × EN/CS).
    `tests/bubble-perf.mjs` keeps the measurement repeatable.

47. **Chat grounds on the notes that were already there, and the server stays stateless.** The
    reconnaissance before this feature found the extracted notes text already kept per deck
    (`sourceText`, size-capped, used by Regenerate), so the "store it in IndexedDB" branch of the
    brief didn't apply — chat reads the deck record instead, with the generated study content as
    the documented fallback for old or imported decks. Retrieval is client-side and dependency-free:
    ~800-token chunks, BM25 (k1 = 1.2, b = 0.75, stopwords, diacritics folded) scored against the
    question plus the last turns, top chunks under a ~7k-token budget, each labelled with deck and
    section. The request carries the chosen context and the conversation window (like the key
    does); the server trims the window to ~6k tokens, caps the context at ~9k, validates the model
    against `CHAT_MODELS` and streams `token` / `thinking` / `done` / `error` NDJSON. Reasoning is a
    separate event, so it can be hidden behind the Thinking row and never mixed into the answer.
    Attachments reuse the vision step (`/api/chat/extract`) so text-only chat models work. Markdown
    is a small React-node parser (no HTML injection, no dependency); KaTeX is dynamically imported
    only when a message contains math and writes into its container imperatively — there is still
    no `dangerouslySetInnerHTML` anywhere in the app. Stopping mid-stream aborts the upstream call;
    the abort arrives as a reader rejection and is mapped back to a clean "stopped" state instead of
    an error card. Usage-limit failures keep the draft and show the shared message with
    Settings/Ollama actions in an inline banner. The Settings toggle "Keep notes text for chat"
    makes chat ignore notes text, and a deck's menu can delete the notes text outright. Verified by
    the chat suite (63 checks against a stub Ollama: validation and the allowlist, context travel,
    the history window, thinking separation, usage-limit mapping, attachment extraction, entry
    points, streaming, stop, markdown + KaTeX, the model and scope pickers, edit/regenerate,
    offline, attachments and reduced motion).

48. **Chat history is local, bounded and part of the backup/sync payload.** Conversations live in
    their own versioned IndexedDB (`freezmo-chat` v1, store `chats`), validated with the same zod
    schema the backup and sync use (`chatConversationSchema`), capped at 100 conversations × 300
    messages, with the in-memory mirror the other stores use. Every settled turn is persisted (the
    question always, the answer when one was produced — partial answers from a stop included); the
    URL carries `?c=<id>` so a reload continues the conversation, and that URL tweak uses
    `history.replaceState` so it never triggers a router round-trip mid-chat. The list (desktop
    column, phone sheet) has auto titles from the first question, inline rename, delete with an Undo
    toast, and diacritic-folding search over titles and message text. Chats ride along in the backup
    file (the import preview counts them) and in the encrypted sync payload — still never the key.
    Assistant answers also offer **Turn into flashcards** (opens New deck with the answer as notes,
    landing on the options step) and **Save to notes** (appends to the single scoped deck, trimming
    at 40k characters). Motion comes from `lib/motion.ts`: message fade-up, sheet springs, a 150 ms
    composer height transition — all skipped under reduced motion. Verified by the chat suite (92
    checks: everything from phase 1 plus history persistence, reload, rename/search/delete/undo,
    both actions, the export payload, a two-device sync round-trip of a chat, offline readability
    and the desktop panel).

49. **The sidebar rail never lets its content widen it.** The desktop rail is a fixed-width `flex`
    aside and its glass panel is a `flex-1` item, so the panel's automatic minimum size
    (`min-width: auto`) is its content's min-content width. When the chat list moved into the
    sidebar, its search input's intrinsic width (an `<input>` defaults to `size=20`) stretched the
    panel 89 px past the rail, where it painted over the page — the content offset itself was
    correct, which is what made it look like the page had lost its left margin. The panel now
    carries `min-w-0` so wide content truncates instead, and the chat suite asserts that the panel
    stays inside the rail and the main content starts after it, expanded and collapsed.

50. **One AI disclaimer, only where the model's words are on screen.** A shared `AiDisclaimer`
    (copy, style and translations in one place, `common.aiDisclaimer`) sits under the chat composer
    — always, since chat is pure model output — next to the generation buttons of the homework and
    deck wizards, under their results, and under AI grading feedback. It also rides along with the
    content itself once it is being consumed: under every card in the study feed and in the review,
    mistakes and exam sessions, and on their results slides. Navigation surfaces (the library, deck
    lists, chat history, settings) deliberately carry no note: the disclaimer belongs to the moment
    of generation and to the card being answered, not to browsing, and the empty states stay clean.
    It is a plain `--text-subtle` paragraph (documented at 5.6:1 light / 8.6:1 dark), centred,
    wrapping to two lines on a 360 px phone — not clickable, not dismissable.

36. **The mobile declutter pass reshapes, it doesn't fork.** Phones (≤639px) get a 3-step deck
    flow and a 2-step homework flow that share the exact state and handlers with the desktop
    cards (one `UploadScreen` / one `HomeworkInput`, two layouts switched by `useIsPhone()`);
    desktop markup is untouched. Library cards collapse to title + one meta line + one
    description line (expand on tap) with the whole card tappable and a small Study pill, the
    stats line and folder row only appear once folders exist, and the settings sheet trades the
    health cards for a Connection status row plus a collapsed Details list. Decorative orbs were
    pushed further off the phone edges so they never glow under CTAs or cards.

37. **The animation pass is one shared language, and it stays off the desktop.** `lib/motion.ts`
    owns the spring presets (snappy / gentle / sheet), durations and variants; components import
    them instead of ad-hoc numbers. Press feedback is CSS (a spring-like cubic-bezier on
    transform) scoped to `≤639px / pointer: coarse`, so every control responds without
    per-component wiring. Sliding highlights use one `layoutId` per control. Route content
    animates on entrance only — the App Router unmounts the outgoing page, so exits are left to
    in-page `AnimatePresence` (list items, sheets). Only transform and opacity animate (plus the
    explicit expander heights); `prefers-reduced-motion` collapses motion to short fades.
    LazyMotion was skipped on purpose: the sliding highlights need layout animations (domMax),
    the same feature set the current `motion` import already pulls in. Verified with a
    4x-CPU-throttled mobile trace (no long tasks, CLS 0) and a reduced-motion run.

38. **Wizard chrome lives outside the animated tree.** The step name is the headline, the
    action bar is a flat full-width bar at the bottom (the tab bar hides inside the wizard),
    and the bar is rendered through a React portal to `<body>`. That last part is a real bug
    fix: the screen entrance animation puts a transform on the content wrapper, and a
    transform on an ancestor becomes the containing block for `position: fixed`, so the bar
    used to paint inside the content and jump to the viewport bottom a second later. It now
    only fades in, pinned to the bottom from its first frame — verified by sampling its
    position across the entrance (5 frames, 60–450ms) from the tab bar, the mode switch and the
    empty state, at 360/390, both themes, both languages and with reduced motion.

39. **Native feel is layered onto the platform, never around it.** The Back gesture closes the
    topmost layer through a small history-entry stack (`lib/use-back-close.ts`): each layer pushes
    a same-URL entry, only the top of the stack reacts, and a layer closed by a navigation leaves
    its entry for the permanent listener to skip. Haptics are best-effort behind a Settings toggle
    (Vibration API; the hidden `<input switch>` tick on iOS, throttled); the wake lock is requested
    only while a study feed or guide is open and released on leave. Destructive-but-reversible
    actions (delete deck, remove file) became Undo toasts; irreversible ones keep their confirm.
    Wizard drafts and the last URL are per-browser; the share target hands POSTed files to the
    client through a ten-minute in-memory store so nothing touches disk. Verified by the native
    suite (35 checks: manifest/head/viewport, offline banner + disabled generate, share target GET
    and POST, back-gesture closing for sheet/wizard, undo toast, share fallback, tab
    re-tap, haptics persistence, draft restore) plus the six standing suites.

    **Needs a real device to verify** (emulation can't do these honestly): the iOS hidden-switch
    haptic tick, the Android/iOS **share target** entry in the system share sheet, the actual
    **wake lock** grant, the **startup images** on an installed iPhone, the iOS **edge-swipe**
    back gesture in standalone, and installability/rich-install prompts (Lighthouse covers the
    manifest side). Pull-to-refresh and the app badge were skipped on purpose — the Settings
    refresh icon already re-checks the connection, and there is no meaningful badge number.

40. **Backups are plain, sync is encrypted, and the key never leaves either.** The backup file is
    the same zod-validated shape the sync payload uses (`lib/backup.ts`), which is why the import
    preview can count added/replaced items without touching the key — the key is simply not part of
    the shape. Sync adds PBKDF2-SHA256 (310k iterations) → AES-GCM on top, keeps the derived key as
    a **non-extractable** `CryptoKey` in IndexedDB, and lets the server be a dumb blob store: it
    stamps `updatedAt` (the conflict clock), enforces `SYNC_MAX_BYTES` / `SYNC_TOTAL_BYTES` and
    writes atomically (temp file + rename). A device that lost its stored key can still *unlock*
    (passphrase re-derivation) or *revoke* (DELETE) the synced copy. Monthly-usage exhaustion and
    short-term throttling are told apart by inspecting the real 429 body
    (`classifyOllamaFailure`) and mapped to `usage_limit` / `throttled` kinds — the client renders
    the same localized message in every model flow, with a jump to the key field and a link to
    Ollama, and keeps the user's input so Retry works right after swapping the key. Verified by the
    phase-1 suite (43 checks, including the real route mapping against a stub Ollama returning a
    monthly-limit body, a two-device sync round-trip, and the generation guards below).

66. **Aero's glass is thin, and the OS dark preference can't leak into it.** Two fixes from one
    report ("the white UI should be blurry and transparent; some text is hard to read on first
    load, until the theme is toggled"). The text bug was at boot: the inline script's dark test
    was `s && !a ? s === 'dark' : matchMedia(...)`, which for a stored `aero` fell through to the
    OS preference — so on a dark-preferring machine the page loaded with **both** `dark` and
    `aero` classes and every `.dark` rule (starting with `body`'s colour and the segment labels)
    painted the dark palette over the bright wallpaper, until a theme toggle re-applied the
    classes. The stored value now decides alone and the class is toggled explicitly;
    `tests/mobile-test.mjs` pins it with a dark-`colorScheme` browser context. And the surfaces
    drop their thick white veils (a card top measured ~71% white) for thin veils over the
    pre-blurred wallpaper — the scene is clearly visible through the glass, and the
    pixel-contrast scan still passes.

83. **The chat fills the space next to the sidebar.** The conversation column was `max-w-3xl`
    centred, which on a 1280 window left a ~140px dead strip between the rail and the messages —
    "the chat cuts off weirdly". It now uses the library's container (`max-w-6xl`, `lg:px-10`) and
    the composer's content is padded to match, so the conversation starts right next to the sidebar
    (measured: column left = 268 = the rail's edge at 1280) and every row — messages, table, action
    chips, input, chips, note — shares one left edge.

    In Aero the welcome panel is the same frosted glass as the cards (blurred, not a flat tint) and
    stretches to the content column's edges — flush under the header and above the composer
    (measured: 268..1280 at a 1280 window) — a flat pane floating 40px inside the scene read as
    "cut off next to the sidebar".

88. **A custom 404, Frutiger Aero style.** Unknown URLs used to fall through to Next's default
    "This page could not be found". `app/not-found.tsx` now renders a standalone screen — the root
    layout's ice scene is the backdrop, and the glass card, the rim and the rising bubbles come from
    the shared recipes, so it is full Frutiger Aero while that theme is on and stays in the
    frosted-glass language in light and dark. It wears the scene's glossy ice orb
    (`.not-found-orb`, at full strength so the snowflake reads) with a "404" chip, "This page melted
    away", and two ways out: back to the library and into the chat (44px targets, en + cs). The
    regression suite pins the 404 status, both links and the Czech copy; the pixel-contrast scan now
    covers the screen in all three themes. The service worker also stopped caching non-OK
    navigations as the offline shell — a typo'd URL used to write the error page into the shell
    cache.

87. **Chat's chrome docks into the rail (the sidebar seam).** The chat header and composer are
    full-bleed bands inside the main pane; at desktop widths that pane starts right at the sidebar,
    so both bands ended in a hard vertical edge beside the floating rail card — the 12px gutter
    showed the raw scene, and the bands' fill + backdrop blur made it read as a color step at the
    seam, with their hairlines starting in mid-air at that edge. Both bands now reach the card's
    right edge (the rail's own 12px padding, expanded or collapsed) and their hairlines are inset
    to the content box instead of touching the seam; phones keep the plain full-width borders. The
    chat suite pins the dock (band edge === card edge, zeroed band borders, inset `::after`
    hairlines). The "Chat is set to ignore notes text" banner went away with it — the scope chip
    already says what the chat may read.

86. **Reduced motion no longer ghosts the sheets.** The in-app toggle fed framer through
    `MotionConfig reducedMotion="always"`, which only strips *transform* animations — while every
    component branched on the OS-only `useReducedMotion`. So a translucent sheet still spring-faded
    its opacity in over the page, showing the page through it for ~300 ms ("ugly visual glitch with
    the settings panel when reduced motion is on"). All 16 components now read
    `useAppReducedMotion()` (the OS preference *or* the toggle), and the settings sheet + scrim and
    the library's phone sheet go instant under reduced motion. Verified frame by frame: opacity
    1.00 from the very first frame.

85. **Aero's buttons read as buttons, and homework results can be reported.** Two reports from one
    screenshot. The recipe's thin veils let secondary/ghost buttons blend into the grass and sky
    ("why are there buttons which dont have any background"): `btn-secondary`, `btn-ghost` and
    `glass-pill` now carry a stronger pale-glass veil (the "Ask pill" look) over the blurred scene,
    with ghosts also getting a rim and full-contrast text. And the homework make/check results had
    no report control — both now carry the same red "!" as the study feeds and the guide, mapping
    the result into the existing report flow through a synthetic short-answer activity (the report
    keeps only the flattened question/type/id), pinned in phase3's result-flow checks.

84. **"Available" means proven, not just listed.** With a cloud Ollama key the picker marked every
    model from /api/tags as *available* — but a free plan can list models it cannot run, so the
    badge promised more than the account could keep. A tag listing now proves availability only on
    a local daemon (there the listing *is* the installed set); on the cloud, and for provider seeds
    before their first live /models test, the badge stays off until the model is actually used. The
    configured favourites keep their verified state — that is what the health probe really checks.

82. **First run offers the choice: free Ollama or a paid provider.** The gate (user-key installs)
    now shows Ollama with a "Free" badge as the headline option, and below it an "or · Use a paid
    provider instead" button that opens the same provider rows as Settings — with the framing that
    those are billed by the provider directly (Freezmo never sees a card) and that keys stay in the
    browser. Saving a provider key on that screen lists its models and switches study generation to
    it automatically, so the gate steps aside without an Ollama key at all — pinned end to end in
    the providers suite with a second, user-key-mode app instance.

81. **Bring-your-own keys for cloud providers.** Besides the server's Ollama, Settings → Providers takes
    keys for OpenAI, Anthropic (their official OpenAI-compatibility layer), Google Gemini (its
    OpenAI-compatible endpoint), DeepSeek, Groq, OpenRouter and any custom OpenAI-compatible
    endpoint. One adapter (`lib/openai-compat`) covers them all —
    model listing, streaming chat, JSON generation and vision — and every model flow can run on a
    provider: the chat picker lists each enabled provider's models (labelled, remembered across
    visits), and a "Study generation" choice moves decks, homework, grading and attachment
    transcription onto a provider model. Keys stay in this browser exactly like the Ollama one
    (`freezmo.providers.v1`): never stored server-side, never in a backup or the sync payload, only
    forwarded with the request that needs them — the provider test asserts the key reaches the
    provider and that the app keeps the routing header to itself. The suite (`tests/providers-test`,
    13 checks) covers listing, chat, generation and the Settings flow against a two-protocol stub.

80. **Chat UI pass: glass chrome, one composer panel, friendlier chips.** The chat header was a flat
    blue strip and the composer area floated on the wallpaper; both are glass surfaces now (the
    composer is one flat frosted panel behind input + chips + note). The disclaimer is plain
    centered text in Aero too (`AiDisclaimer plain`), the welcome block + suggestions sit on a soft
    frosted panel and are vertically centered between header and composer, the model chip shows a
    friendly name ("Gemma 4 31B") with the raw name in a tooltip, the chips are bigger with a
    stronger rim, the send button is the filled primary (blue in every theme, clearly dimmed only
    while empty/offline), and the header's "+" is a chat-plus icon so it no longer reads as the
    rail's "new deck". (The review also asked for the light/dark toggle to come back to the Aero
    header/rail — that stays out: Aero is switched from Settings by the owner's earlier rule.)

79. **Any model the server offers is selectable in chat, default `gemma4:31b`.** The picker used
    to be the server's allowlist only. The health check now appends every chat-capable model from
    the server's `/api/tags` (favourites first, embeddings filtered, 120 cap), the picker gained a
    search field when the list is long, and the chat route accepts a model that is in the
    favourites *or* in the live listing (one tags fetch for non-favourites, so nothing selectable
    is rejected). The default is `CHAT_MODEL_DEFAULT` (env) or `gemma4:31b`; when a server doesn't
    have the default, the client falls back to the first available model.

78. **Reduce motion: a Settings toggle, default off.** It mirrors the OS preference — an
    `html.reduce-motion` class (applied pre-paint by the boot script) parks the scene and collapses
    CSS animations/transitions, and `MotionPreference` wraps the app in framer-motion's
    `MotionConfig reducedMotion="always"`, so every JS animation calms too. Stored in
    `freezmo.reduce-motion.v1`; pinned in the mobile suite (126 checks).

77. **The Aero review pass (tabs, primaries, panels, tints).** The active segment keeps the blue
    gradient in Aero again (it had been glassed), and so does every `btn-primary`/`grad-ice` — the
    "Review due" action, the card's Study button (now the filled primary) and the composer's send
    button included. New Aero surfaces: a frosted backing behind the page heading, a chip behind
    the "All decks" label, a bottom-weighted tint on the rail (its lower rows sit over grass), and
    a dedicated `--track` token so progress bars read in every theme without shifting the shared
    surfaces (the first attempt moved `--surface-sunken` and broke sheet contrast — the pixel scan
    caught it). Difficulty labels got darker, the inactive segmented labels use the full text
    token, the card border/shadow are a touch stronger in light, orbs are calmer in light and
    dark, the folders list nudges "create a folder", and a fresh deck shows a "New" badge instead
    of the gray "Not played yet".

76. **A dropped connection says so.** A user's run failed with the wizard's generic "check your
    Ollama settings" card, and it could not be replicated on the owner's machine with the same key
    — because the key and model were never the problem. A raw error (a mid-run socket drop, a
    wifi/LTE switch, an app backgrounded by the phone OS) leaked out of the NDJSON reader as a
    DOMException, and the wizard can only show that generic card for unknown errors. The reader now
    classifies everything: an abort is "cancelled", a stall is "the model stopped responding", and
    anything else on the wire is "the connection dropped mid-run — try again". Pinned with a
    fetch stub that kills the stream mid-run in the wizard suite (95 checks).

75. **A quiet run can no longer load forever.** A user got stuck on the wizard's loading screen:
    a hung model meant silence, and silence was indistinguishable from work — the server's own hard
    deadline is 2x `OLLAMA_TIMEOUT_MS` (ten minutes by default), and the client waited on a dead
    stream indefinitely. Both generate routes (deck, homework) now send an NDJSON heartbeat every
    15s, and the client guards the two quiet phases: no response at all for 45s, and no bytes for
    90s mid-stream, both surfacing the regular error state with a Retry — verified with a stalled
    route and Playwright's fake clock in the wizard suite (93 checks). The heartbeat also keeps
    proxies from closing an idle connection.

74. **Phones get flat glass, not live blur.** The touch fallback from 73 (a real backdrop blur
    over the live scene) turned out to be the wrong trade: every pane re-blurs on every frame
    while the bubbles animate, and phones choke — input lag, "buttons take ages to respond". Touch
    devices now drop both the bake (which mispaints there) *and* the live blur: the panes keep the
    thin Ask-pill tint over the sharp scene (the hand-rolled activity fills keep their own tints),
    i.e. transparent glass at zero per-frame cost. Desktop keeps the baked blur — the same look as
    the phones had, without the GPU bill. `tests/mobile-test.mjs` pins the flat contract (no baked
    layer, no backdrop-filter on a card) with a touch context.

73. **Touch devices blur the live backdrop; desktop keeps the bake.** Mobile engines repaint
    `background-attachment: fixed` late (and iOS treats it as scroll), so the baked glass showed a
    stale crop of the wallpaper until a tap forced a repaint — "some of the UI looks bugged until I
    click it". On touch devices (`max-width: 639px` or `pointer: coarse`) the bake is dropped for a
    real `backdrop-filter` over the live scene (what the floating chrome always did) — the bubbles
    now blur *through* the glass as they pass, and the position is correct by construction. The
    hand-rolled frost fills keep their own tints there; desktop keeps the baked layer (measured
    cheaper there and verified correct). Pinned in `tests/mobile-test.mjs` with a touch context.

72. **More, shinier bubbles.** Aero's rising bubbles went from ten to eighteen (the extra eight
    interleaved between the others, 18px to 96px, 10-23s runs) with a tighter specular hotspot, a
    second reflected-light glint low on the sphere, a 1px glass edge and a brighter inner rim on
    both the bubbles and the drifting orbs. The bodies stay translucent and their opacities are
    capped (0.75-0.9), so text sitting on the wallpaper stays readable, and the bubble-perf
    diagnostic shows no added frame cost (18 bubbles, same p95 as ten).

71. **The chat's top edge fades instead of hard-clipping.** The conversation scrolls right up to
    the glass header, and over the busy wallpaper a line of text clipped exactly at that edge read
    as a rendering bug ("the chat looks cut off"). The scroller now carries a 14px top mask in
    Aero — shorter than the column's 16px top padding, so a fresh conversation at the top shows no
    dimming at all.

70. **Chat answers are bubbles in Aero too.** The assistant answers are deliberately bubble-less
    (full-width text on the flat themes) and the user bubble's thin tint washed out over the bright
    parts of the wallpaper, so the longest text in the app ended up sitting straight on the photo.
    In Aero both get a real pane: `.chat-answer` wraps the markdown in a bubble (plus a
    `chat-bubble` class on the user turn), with a stronger veil than the shared glass so a wall of
    text still reads comfortably over the same blurred scene, and the bright rim. Light and dark
    keep the bubble-less assistant design — these classes are styled only under `.aero`.

69. **The glass got thinner and reaches the whole app.** Two asks in one pass: "slightly more
    transparent" and "some buttons aren't themed" while doing decks/homework. The veils drop
    another notch (recipe 0.15 -> 0.12, the tokens ~0.03 less, the tab bar / wizard bar / gloss
    likewise), and the recipe now catches the flows' hand-rolled frost: the study activities'
    `bg-white/xx` chips (multiple-choice options, matching terms, ordering cards, fill-in-blank
    chips), the exam deck rows and pickers that used `bg-[var(--surface)]` — an *opaque* white in
    Aero — and everything using the surface tokens (chat action chips, homework chips, inputs,
    dropzones); attribute selectors match them, so no component needed a new class. The same shapes
    inside sheets stay flat (the exclusion list covers the new selectors — there the sheet itself is
    the glass), and the dropzones lost their press-scale (a transform on a fixed backdrop
    re-anchors it) and shade inward instead. `tests/contrast.mjs`: 0 findings.

68. **The Ask pill joins the glass, and Aero's hovers light the pane instead of moving it.** Two
    leftovers from the previous pass. The Ask button (and the two phone icon pills that share its
    styling) was a hand-styled `bg-[var(--surface-strong)]` pill that never picked up the frosted
    recipe, so it showed the sharp wallpaper through a flat tint while everything around it showed
    the blurred one — it carries a `glass-pill` class now and is part of the same recipe. And the
    hover feedback: a `transform` or `filter` on a pane with a viewport-fixed backdrop re-anchors
    the glass, so the deck card's hover lift made the card flash a wrong crop of the wallpaper
    (and the buttons' `filter: brightness()` would have done the same, and the mobile
    `:active { transform: scale(0.96) }` foundation made panes jump while pressed). Aero's
    feedback never moves or filters the glass now: the pane lights up with an inner glow, a white
    rim and a deeper shadow, transitioned over 170ms, and the pressed state shades inward.

67. **The Aero wallpaper URL is cache-busted, and every surface is one glass.** Two fixes from
    one report. The background "didn't change" because the service worker's asset cache kept
    serving the previous wallpaper bytes (stale-while-revalidate returns the cached copy first,
    and the old worker can outlive a deploy): both aero images now carry a `?v=3` query, so a
    stale cache cannot answer for them, and future wallpaper swaps get a new query too. And
    "everything should look like the Ask button" is now literal: the rail, the cards, the chips,
    the segments and the buttons all share one recipe — the pre-blurred wallpaper under a ~15%
    veil (the minimum the pixel-contrast scan needs) with the bright rim. The blur bake is
    shadow-lifted (gamma 0.86) and slightly saturated so the scene's colours stay vivid instead
    of washing out to a milky film, the wizard's phone action bar (Aero's `--page` token was the
    saturated water blue, a heavy slab over the photo) joins in, and nested chrome inside sheets
    keeps its flat fills (the sheet itself is the real glass there). `tests/contrast.mjs` is back
    to 0 findings across every screen and theme.

65. **Slide content never flex-shrinks, and the AI note appears once.** A long typed answer in a
    short-answer activity shrank the textarea inside the slide's flex column (58px instead of its
    ~82px), clipping the user's text mid-line on a phone; every direct child of a study/review/guide
    slide is now `shrink-0` and the slide scrolls instead. The same screenshot showed **two** AI
    disclaimers — the activity's own after the grading box plus the feed's row. The activity-level
    copy is gone: the feed/guide row (which also carries the report control) is the single one, and
    the phase-2 suite asserts the note sits below the graded feedback rather than right after it.

64. **Ordering cards only where an order exists.** "Sort these" was landing on decks whose notes
    have no sequence at all (a list of works, authors or terms), which the learner cannot reason
    about. The deck prompt now allows `ordering` only for real sequences — chronological events, the
    stages of a process, steps of a procedure, a lifecycle, a ranked scale — and points at matching,
    multiple choice or true/false when there is none, with one or two ordering cards per deck as the
    usual ceiling.

63. **Frutiger Aero is the third theme.** Settings → Theme offers Light, Dark and Frutiger Aero.
    Aero is not a dark/light variant — it replaces the palette and the scene: the wallpaper is a
    Frutiger Aero scene made with Gemini (latest 1280x714 revision re-encoded to WebP in `public/aero/`), with ten drawn
    bubbles rising over it (`.ice-bubble` — `display: none`
    outside aero so they cost nothing elsewhere, still under reduced motion). Buttons and segments
    are **liquid glass** — transparent, glossed, with a bright rim — and every surface shows a
    **pre-blurred copy of the wallpaper** (`public/aero/background-blur.webp`, 3 kB, glued to the
    viewport with `background-attachment: fixed`) instead of blurring the live scene: the bubbles
    move behind every surface, so a real `backdrop-filter` re-runs ~60 times a second (measured
    66ms/frame at 2x throttle vs 17ms with nothing blurring). A thin veil and the gloss ride on top of
    that layer, so the look is identical at zero per-frame cost. The helper lines that sit straight
    on the wallpaper (the keyboard tip, the AI note) get a small glass chip — dark text on grass
    measured 2.5:1. The light/dark
    toggles (header, rail, chat header) hide while it is on and the picker says why; the PWA status
    bar follows the wallpaper's sky (#2fa7e6). Covered by the axe and pixel-contrast scans in all
    three themes.

62. **Accessibility: axe is clean everywhere, and contrast is measured on real pixels.** axe cannot
    judge text on blurred glass — it reports those nodes as "incomplete" — so `tests/contrast.mjs`
    samples the rendered pixels (dominant non-glyph colour per element) and compares it with the
    computed text colour using the WCAG formula, skipping disabled controls (1.4.3 exempts them)
    and content behind an open sheet. Fixes from that pass: the hidden file inputs are out of the
    accessibility tree (aria-hidden + tabIndex -1; labelled buttons open them), the heading order is
    right (deck/homework/smart-review cards are h2; LearningSection takes a level so it is h2 on the
    library and h3 inside the dialog), the study and review feeds have landmarks, and the light
    settings sheet — translucent over a dark scrim, where its muted text measured ~2.8:1 — got its
    own more opaque surface (`--glass-sheet-bg`). Result: 0 axe violations and 0 below-minimum
    findings across every screen, all three themes, phone and desktop.

61. **Glass that sits still doesn't blur.** The MacBook got hot just browsing: the ice scene animates
    constantly, and every `backdrop-filter` surface re-runs its blur on every frame while the backdrop
    moves — the whole card and rail area was re-blurring the scene ~60 times a second. Measured on
    desktop (2x CPU throttle, software renderer): p50 frame 66ms with everything blurring, 33ms once
    the static surfaces stopped, 17ms with the scene's animation also off — frames over 34ms dropped
    from 128 to 14. The recipe is now split: **floating glass** (sheets, dropdowns, dialogs, the
    sticky header/chrome/tab bar — the layers that can have content scroll underneath) keeps the
    blur, now `backdrop-blur-lg` (which is what phones already used), while **static glass** (cards,
    the rail, the gate, the study prompt card) keeps the exact same tint, rim, highlight and shadow
    without the filter. A deterministic before/after pixel diff (scene motion frozen) shows the
    change is a ~1% colour shift in the card backdrops — invisible in both themes.
    `tests/bubble-perf.mjs` now reports the glass cost too, so a future blurred surface over the
    scene shows up immediately.

60. **There is exactly one loading screen.** The deck wizard had the good one — the app icon, the
    rotating stage message as the headline, the progress bar, the stage checklist, a tip and
    Cancel — while the homework flows showed a spinner with a title and the route loads a pulsing
    icon with a smaller line. All three now render `components/LoadingState`: the wizard passes its
    vision/generation stages (and keeps its deck tips), the homework passes the same shape with
    homework wording (no tips — the tips are deck-specific), and the route screens pass a single
    message with no stages or progress. Cancel stays a 44px touch target and the icon animation
    respects `prefers-reduced-motion`. Verified by the freeze-loading, review, wizard and mobile
    suites. The link overlay is portalled to `<body>`: inside a glass card `backdrop-filter` makes
    the card a containing block for `position: fixed`, so the overlay was squeezed into a 44px
    strip of the card instead of covering the viewport (the review suite now measures it).

59. **Adding a deck or a homework syncs itself.** Sync used to be push-on-demand only: a deck
    frozen on the laptop sat in `dirty` until someone opened Settings and pressed Sync now. The two
    stores whose writes mean "the user just made something" (`lib/storage`, `lib/homework`) now
    schedule `syncNow("auto")` through `lib/sync-auto` — debounced 2.5 s so a burst of writes is one
    push, one run at a time, silent on failure (the data stays dirty and Settings still offers the
    manual sync), and retried on the browser's `online` event. Deliberately *not* wired to answers
    or chat: those change every few seconds while studying. The module is imported dynamically from
    the stores, so the sync code stays out of the first load and no import cycle forms.
    `markSyncDirty()` now reports whether it actually marked, which also keeps pull/import paths
    quiet while a backup import pushes like any other local change. Verified by the phase-1 suite
    (a deck write pushes itself, then Sync now is a no-op) and the phase-3 suite (a check run pushes
    the new homework record).

58. **The smart review follows the deck list.** Deleting a deck used to leave its due schedules
    and mistakes in the library counts: `dueCount` counted raw SRS rows (no deck lookup) and both
    counts were memoised on the learning store alone, so a deck change never recomputed them. The
    counts now resolve through `dueItems`/`mistakeItems` (which skip decks and activities that no
    longer exist) inside a memo keyed on the live deck ids, so deleting a deck drops its numbers at
    once — and Undo brings them straight back, because the learning data itself is never erased.
    Verified by the review suite (two decks due → delete one → counts drop → Undo restores).

57. **"Check my work" takes photos and documents, not just text.** The attempt field used to
    apologise that handwritten work was not supported. It now carries its own intake (photos of
    pages, PDF/DOCX/PPTX/ODT/TXT…): documents are read in the browser and their text joins what was
    typed, while photos travel to `/api/homework` as `solutionImages` and are transcribed by the
    vision model before the check runs (the stage reads "Reading your solution…"). The server
    returns the finished attempt in `meta.solutionText`, which the client stores as the record's
    `attempt` — so "Redo" re-checks the same work without the photos ever being stored. Verified by
    the phase-3 suite (API: a photos-only attempt is accepted, transcribed and runs vision first;
    desktop: thumbnail appears, the payload carries `solutionImages`, the result renders).

56. **Model text keeps its own line breaks.** Every homework prompt asks for plain text with short
    lines and numbered or dashed lists (`HOMEWORK_BASE_RULES`), but the guide rendered the step
    explanation in a bare `<p>` — JSX collapses newlines, so a step arrived as one run-on paragraph
    with the whole worked solution inside it (the reported "completely broken" step). The
    explanation, the progressive hints and the verdict/wrong entries of a check now render with
    `whitespace-pre-wrap` like the other homework results always did, and the guide prompt says the
    explanation is 2-4 sentences of coaching, with the finished working left to hint 3 and
    `answer`.

55. **Popovers close the way they open.** The chat's context ("what may chat read") and model chips
    used to only ever open: the sole way out was picking something, or Escape. The chip now toggles
    its own popover (and reads pressed while open, `aria-expanded`), opening one closes the other,
    and on desktop a pointer press anywhere outside closes it — except on the trigger chips
    themselves, which the outside handler ignores precisely so the toggle cannot re-open what it
    just closed. Phones keep the sheet + backdrop they already had. Verified by the chat suite
    (toggle, pressed state, switch, outside click).

54. **The study chrome is one pill, and it sits above the app chrome.** Progress header, close
    button, title, position and streak/xp (or the exam timer) used to be two stacked rows inside a
    `rounded-full` box — on a tall box that is a stadium, so the corner curves cut into the icon
    row and the controls looked like they were falling into the corners. It is now a single 58px
    pill: round close button, title + counter over the progress bar in the middle, chips at the far
    end, everything on the centre line — the same recipe in the study feed, review/exam sessions and
    the homework guide. The immersive feeds also carry `z-[45]`: on `/exam` the app header (`z-40`)
    used to paint over the top of the chrome. Verified by the phase-2 suite (the close button's
    corners stay inside the rounded end, single row, counter present).

53. **The flashcard flip never relies on `backface-visibility`.** The card is two faces in a
    `preserve-3d` container (the back pre-rotated 180°) — the classic CSS flip — but each face is a
    glass layer with `backdrop-filter`, and on some GPU/compositing paths the front's backface keeps
    rendering, so the flipped card showed the question mirrored over the answer. The faces are now
    painted by an explicit `visibility` swap driven by the animation's own rotation (the swap lands
    at the 90° edge-on moment, where it cannot be seen), so exactly one face is ever in the render
    tree. Reduced motion skips the spring and the swap still lands via `onAnimationComplete`. The
    phase-2 suite asserts the painted-face invariant at rest, flipped and flipped back.

52. **Reporting a question is local-first, with a server copy for the operator.** The red "!" under
    the card (bottom-right of the exercise — it used to sit in the top chrome, where it read as
    chrome and was easy to mis-tap) records the question in this browser — that local list is the
    skip state, so the card leaves study, review, mistakes and exam sessions, and Undo in the toast
    brings it back (the server file is append-only). It also POSTs a copy to `/api/reports`, which
    appends one JSON line to `shared/freezmo/reports/reports.jsonl`. The app has no auth, so the
    route is write-only: the question text is readable on the server (that is the point — prompt
    tuning) but never over HTTP. A reported card counts as "not attempted" in the results rather
    than wrong, and a deck where everything is reported stays studiable so it can be regenerated.
    Verified by the phase-1 suite (append, validation, no GET) and the phase-2 suite (button
    placement under the card with a 44px target, toast, server POST, local copy, skip, hidden next
    session, Undo, exam pool).

51. **Cards with repeated wording never ship.** The model sometimes pads a matching card with one
    definition copied across several terms (a Czech literature deck arrived on a phone with three
    identical definitions, so the card could not be solved by knowing the material), or repeats an
    option. The generation pipeline now rejects any card whose options, items, terms or definitions
    repeat (case-insensitive) — the same drop-or-repair path the ordering giveaway guard already
    used, so a deck with too few cards left over gets one repair round instead. The matching and
    ordering dedupe only trims a card when it still meets the schema's minimum afterwards: a
    shorter card would fail validation on the next read and take the whole deck down with it, so
    below the minimum the original is kept and the card is rejected instead. The generation prompt
    states the uniqueness rule too. Verified by the phase-1 suite's stub-Ollama generation: the
    repeated-definition card is dropped, a salvageable card is trimmed to three pairs, the
    repeated-option card is dropped and the good cards still ship.

41. **The learning log is an event log, so it lives in IndexedDB.** Decks and homework stay in
    `localStorage` (small, always loaded), but answers + SRS schedules go to a versioned
    `freezmo-learning` database with an in-memory mirror so React can read them synchronously via
    `useSyncExternalStore`; writes update the mirror first and persist asynchronously, and the log
    is capped at 20 000 answers. The schedule is deliberately simple (1 day → 3 days → interval ×
    ease; wrong = ten minutes, ease −0.2): predictable beats clever, and the mistakes set needs no
    storage at all — it is *the latest answer per activity being wrong*, so it clears itself the
    moment the activity is answered correctly. Review and exam sessions **freeze their item list at
    start** (a bug the phase-2 suite caught: answering a mistake removed it from the live list and
    collapsed the session mid-run). Push reminders are all-or-nothing: no VAPID keys → no UI, no
    endpoints, nothing to misconfigure; `web-push` is the one added server dependency and the daily
    send is an explicit cron/timer script. Verified by the phase-2 suite (43 checks: history and
    SRS shapes, the review/mistakes sessions, stats math, exam with timer and breakdowns, the
    VAPID-less hidden state and the configured subscribe/validate/test/delete path; the suite
    runs its own server instance on a separate port and kills the process group on exit).

14. **`gpt-oss:120b` does all content work — but it cannot read images** (`this model does not
    support image input`), so `VISION_MODEL` has to stay a multimodal model (`gemma4:31b-cloud`).
    Text-only inputs — pasted text, PDFs with a text layer, DOCX/PPTX/ODT — never touch the vision
    model at all, so those decks are produced by `gpt-oss:120b` start to finish.
13. **No auth, no server storage.** Everything personal lives in `localStorage`; the only server
    state is the request in flight.
14. **`canvas-confetti` and `framer-motion` are the only UI dependencies.** Document parsing adds
    `pdfjs-dist`, `mammoth` and `jszip`, all dynamically imported (the `/create` page grew ~3 kB;
    the parsers load only when a document is added). Animations are skipped when
    `prefers-reduced-motion` is set.
15. **Tailwind v4 class-based dark mode** with a pre-paint script so there is no theme flash, and
    `suppressHydrationWarning` on `<html>`.
16. **Testing.** The pipeline was first tested against a mock Ollama HTTP server (including a
    deliberately invalid first response to exercise the repair path) with the UI driven by headless
    Chromium, then re-run against the real Ollama cloud. Current suites: 13/13 document extraction
    (one fixture per format, including a text PDF and a scanned PDF), 15/15 language preservation,
    24/24 activity/scoring/feed checks and 11/11 real-model browser checks.

## Legal pages

Freezmo ships its own plain-language documents, in both app languages, rendered by
`components/LegalPage.tsx` from `lib/legal.ts`:

| Page | What it says |
| --- | --- |
| `/privacy` | Local-first storage, the key never leaving the browser, the three opt-in server copies (sync, reminders, reports), what the server does *not* store, and the third parties (Ollama, your browser's push service). |
| `/terms` | Free software under GPL-3.0-only with no warranty, your own Ollama key and its limits, fair use of the instance, your content, availability and liability. |
| `/cookies` | No cookies at all (verified: no `Set-Cookie`, no `document.cookie`), what the strictly-necessary local storage holds, no tracking, and how to clear it. |

A cookie notice (`components/CookieConsent.tsx`) tells every new browser the same thing once and
remembers the dismissal (`freezmo.consent.v1`). It is floating glass — blurred backdrop, rim and
shadow — so it reads as a notice above the page rather than part of it. It sits below the immersive
feeds in the stacking order, so it can never cover a study session, and above the phone tab bar. Settings links all three
documents. The documents describe what the code actually does — keep them in step with it.

## License

Freezmo is free software, licensed under the **GNU General Public License, version 3 only**
(SPDX: `GPL-3.0-only`) — the full text is in [LICENSE](LICENSE), and `package.json` carries the
same identifier.

You may run, study, share and modify it under that license. It comes with **no warranty**, to the
extent permitted by law. If you redistribute it (or a modified version), the license requires you
to pass the same freedoms on and to make the corresponding source available.