halkyon/AGENTS.md

# Flutter & Serverpod project

This project is a Flutter app (frontend) backed by a Serverpod server (backend). Always build the app's backend with Serverpod.
Build for multiple users, use Serverpod's built-in authentication, which is already set up in `lib/server.dart`.

The user starts the server and Flutter app with `serverpod start`. There is no need to check if the server is running: make the changes and call the `serverpod` MCP tools as needed. If the server is not running, an informative error message will be received from the MCP server. Then STOP and ask the user to start it. NEVER start the server yourself. The Flutter app is started along with it, or can be launched from the MCP tool `spawn_flutter_app`.

While running, `serverpod start` watches for file changes to run incremental code generation and hot reload both the server and the Flutter app.

Calling `serverpod generate` directly is not needed, but might be useful to troubleshoot when an incremental generation fails.

ALWAYS use the MCP server instead of the command line. Use the MCP server to:

- `create_migration` and `apply_migrations` for database (after you change data models).
- `create_repair_migration` if the database has drifted out of sync with the migrations.
- `tail_server_logs` to read logs from the server.
- `tail_flutter_logs` to read the raw stdout/stderr of the Flutter app.
- `hot_reload` / `hot_restart` to reload or restart the server and the Flutter app. ALWAYS call `hot_restart` after doing changes in the Flutter app that may not work with normal hot reload (which is automatically applied).
- `spawn_flutter_app` to start a Flutter app declared under `serverpod: flutter_apps:` in the server `pubspec.yaml`.
- `get_flutter_app_dtd` (Dart tooling daemon) for connecting to the app through the `dart` MCP.

NEVER edit generated code. The server's `lib/src/generated/` directory and the whole `halkyon_learn_client` package are rewritten by the code generator. Change the `.spy.yaml` models, the endpoints, or `lib/server.dart` instead.

Migrations are a narrow exception: the `migration.sql` of a generated migration MAY be edited by hand when the generated SQL would lose data — to add a data transformation, or to reach a destructive change through non-destructive steps. Never touch the other files in the migration directory, and keep the schema the SQL ends up with identical to `definition.sql` — new databases are created from that file and never run `migration.sql`.

The `serverpod` MCP tools come from `.mcp.json`: `serverpod mcp-server`, a bridge to the running `serverpod start`. They load when a session starts; if they are missing, the server wasn't running or the session predates `.mcp.json`. Like all agent configuration, `.mcp.json` is git-ignored; in a fresh clone, create it with a server `serverpod` running `serverpod` with the arguments `--no-analytics mcp-server --server-dir halkyon_learn_server`.

**Pass `--no-analytics` to every `serverpod` call** (before the command: `serverpod --no-analytics generate`). Otherwise the CLI reports the commands run and the Serverpod features the project uses to PostHog; there is no setting or environment variable to switch that off. The VS Code task and `.mcp.json` pass it already.

Only when the server cannot be started at all, fall back to the CLI in the server package:

- `serverpod generate` to regenerate the client and the generated server code.
- `serverpod create-migration` after changing a model with a `table` (add `--force` for destructive changes). It only writes the migration; `serverpod start` applies pending migrations when it boots the server.

Tests need no Docker. `config/test.yaml` sets `database.dataPath`, so Serverpod starts and manages the test database (an embedded PostgreSQL) itself. Just run `dart test` in the server package.

Tooling traps: rules of `leancode_lint`'s analyzer plugin are ignored with its prefix (`// ignore: leancode_lint/missing_equatable_props`); `dart fix --apply` moves `flutter_driver` into the app's `dependencies` (as `any`) on every run, so restore it as a dev dependency (`sdk: flutter`) afterwards; never `dart format` generated code, only the hand-written paths, as `tool/check.sh` does (the server's `test/integration/test_tools/` is generated too, so format test files one by one, never the whole `test/` folder); and never rebase or check out other commits while `serverpod start` runs: its watcher regenerates code for each intermediate state, leaves stale generated files behind and hot-reloads them. Stop it first, or work in a separate worktree.

After a migration that adds or changes synced tables, restart `serverpod start` once it has applied it. The sync engine (checked with `serverpod_offline_sync` v0.0.8) reconciles its registry of synced tables when a server first uses the database, and if that changed the registry, every later session of that server rebuilds the projections of every space, one transaction each, until the server stops ([serverpod_offline_sync#150](https://github.com/marcelomendoncasoares/serverpod_offline_sync/issues/150), open). With the many people the tests leave in the development database, syncs then take seconds and the sync tests time out. The restarted server finds the registry as it left it. Deploys aren't affected: their maintenance run applies the migrations and reconciles before it exits (it reads the runtime settings through the engine's wrapper), so the server started after it finds nothing changed. Drop this note once #150 is fixed.

Checklist after doing changes, in this order:

- `dart format` (CLI) on the files you changed
- `create_migration` and `apply_migrations` (MCP - only if necessary)
- `tool/check.sh` (CLI): format check, `dart analyze --fatal-infos`, server and app tests. The pre-push hook runs the same script, so a push only goes through when it passes.
- Do `serverpod` MCP `hot_restart` if required (hot reload is done automatically). Will also hot restart Flutter app
- Check `serverpod` MCP `tail_server_logs` and `tail_flutter_logs` for any issues.

If the user asks you to test the app:

1. Use `get_flutter_app_dtd` (`serverpod` MCP) to get the Flutter app's DTD
2. Pass the DTD to `connect_dart_tooling_daemon` (`dart` MCP) to connect to the app
3. Use `flutter_driver` (`dart` MCP) to navigate through the app

The app is launched from `halkyon_learn_flutter/lib/driver.dart`, which starts the Flutter driver extension with text entry emulation turned off so the app stays usable by hand. To let the driver type, set `enableTextEntryEmulation: true` there and `hot_restart` the app.

# Halkyon Learn

A spaced-repetition learning app: studysets of flashcards, reviewed on an FSRS schedule. Content is rich text (`appflowy_editor` documents) everywhere. Targets: Linux, Windows, macOS, Android, iOS. There is no web build.

- **Packages**: `halkyon_learn_server` (endpoints, models, migrations), `halkyon_learn_client` (generated, never edited), `halkyon_learn_flutter` (the app, ID `app.halkyon.learn`).
- **Instances**: anyone may run one. The app takes the server URL from `--dart-define=SERVER_URL=…`, so never hard-code a host.
- **Working notes**: if `NEXT.md` exists in the repository root, read it first. It holds the current step, the plan and the reasoning behind each decision. It is local and never committed.

## How work is done here

- **The agent implements; the developer decides.** Put every decision to the developer before building (design choices, alternatives, framework defaults kept or changed). The first time a new pattern appears (e.g. the first repository/gateway pair, the first synced model, the first endpoint with its denial tests), walk through it in detail with the developer; afterwards, implement exceedingly similar cases directly and point out only deviations.
- **Tests come first** and state the specification. Every access rule gets a test that asserts **denial**, not only permission.
- **Explain every migration's SQL.**
- **Before keeping a framework default, name the alternatives** and recommend one, especially where a choice becomes expensive once users or data exist.
- Every decision carries its reason (in `NEXT.md` or an ADR under `docs/adr/`).
- **Commits** are GPG-signed and carry a DCO `Signed-off-by` trailer (`git commit -s`; a local `prepare-commit-msg` hook adds it automatically).
- **Pushes** pass `tool/check.sh` first, through a local `pre-push` hook (in a fresh clone: `ln -s ../../tool/check.sh .git/hooks/pre-push`). CI runs it again, and a green `main` deploys itself (see *CI and deploys*), so an unsigned commit on `main` blocks every deploy after it. As the hook, the script first refuses a push to `main` that brings a commit not signed by a key in `deploy/allowed-signers` (`tool/check_signatures.sh`, the deploy's own rule), so it fails on the laptop rather than after CI.
- **Lints**: LeanCode's strict set (`leancode_lint`), configured once in the root `analysis_options.yaml`, which also enables its analyzer plugin. Plugins can only be enabled at the root of a pub workspace and only reach packages whose options include that file, so the server and the app include it; the generated client doesn't. Every `// ignore:` gets a comment saying why (`document_ignores`). The server switches off the four rules for Dart v3.13's constructor syntax (`use_primary_constructors`, `use_declaring_parameters`, `empty_container_bodies`, `unnecessary_type_name_in_constructor`): Serverpod's code generator (checked with v4.0.3) can't parse that syntax in endpoint files yet.

## App architecture

- **Feature-first.** App: `lib/src/core/` holds infrastructure shared by features, `lib/src/feature/<name>/` everything a user would name, and `lib/src/app/` puts the features together (the person's app and its router, the only code that knows them all). Server: `lib/src/<feature>/` holds that feature's `.spy.yaml` models and endpoints. Code moves into `core/` only once a second feature needs it. Why: a feature is added, changed or removed in one place.
- **Four layers per app feature**, dependencies pointing strictly downward: `widget/` (UI; reads bloc state, sends events) → `bloc/` (UI state; orchestrates repositories) → `data/` (repositories and gateways) → `model/` (immutable domain types; pure Dart, no Flutter, no I/O). A widget never calls a gateway; a model never imports a bloc.
- **Generated protocol classes are not domain models.** They carry the envelope (`payload` as bytes); `model/` holds the decoded content. Mapping between them happens in `data/`.
- **Import `halkyon_learn_client` with a prefix (`as protocol`) or a `show` list.** The generator emits a `<Model>Repository` for every table, and its names clash with the app's own repositories.
- **Gateway / repository.** A gateway is a thin boundary to *one* backend (e.g. the Serverpod client's endpoints) behind an `abstract interface class`. A repository is what features use: it composes gateways and holds the logic. Blocs depend on repositories, never on gateways or the raw client. Why: gateways are the seams tests fake, and adding the offline layer later becomes a new local gateway, not a rewrite of every feature.
- **Manual dependency injection.** One `Dependencies` object (`lib/src/core/dependencies/`) is the composition root, built at startup (`Dependencies.create()`) and exposed through the `DependenciesScope` InheritedWidget. It hands features repositories only. No service locator, no DI framework, no globals (the scaffold's global `client` moved into it). Why: every dependency is a typed, greppable field, and nothing is resolved by magic.
- **One graph per context.** The dependency graph belongs to one instance and one signed-in person; when either changes, a `KeyedSubtree` disposes the whole subtree and rebuilds it. Why: nothing from one person's session (blocs, streams, caches) can leak into another's.
  - **App-wide vs. per person.** `Dependencies` holds what outlives a person (the client, auth). Everything of the signed-in person (their local database, its sync, their repositories) is `PersonDependencies`, opened by `PersonScope` inside `SessionGate`'s keyed subtree (`Dependencies.openPerson`) and disposed with it; widgets reach it through `PersonScope.of(context)`.
  - **The device's own settings** (appearance, language, citation style; `core/device_settings/`) belong to the device, not to a person: kept in `<app support>/device_settings.json`, read in `Dependencies.create()` so the first frame already has them, and provided by `DeviceSettingsScope` above `SessionGate`'s keyed subtree, which both its plain app and `PersonApp` read. What should follow the person to their other devices goes into their synced `PersonSettings` instead.
  - **Every person has an app of their own** (`PersonApp`, `lib/src/app/`), which `SessionGate` builds inside the keyed subtree: a `MaterialApp.router` with its own `GoRouter`, and `PersonScope` around the router's navigator (`MaterialApp.builder`). So every screen and dialog is inside the person's scope and reads their repositories from `PersonScope.of(context)` wherever it is built, and a change of person disposes the router with every screen on it and starts again at home. Until someone is signed in, `SessionGate` shows a plain `MaterialApp`: the spinner, or the failure with a way to try again.
- **Platform plugins sit behind small interfaces in `Dependencies`**, with fakes in tests: picking files (`FilePicker`, `file_selector`), saving files the app made, such as exports (`FileSaver`: `file_selector`'s dialog on the desktop, `file_picker`'s on Android and iOS, where `file_selector` has none), playing sounds (`SoundPlayer`, `audioplayers`), playing YouTube videos (`YoutubeVideos`, `youtube_player_iframe` where a web view exists, else the browser), reading PDFs (`PdfEngine`, pdfrx on PDFium, which `flutter test` can't load: widget tests use `FakePdfEngine`, whose PDF bytes say their pages and chapters), reading Anki packages on another isolate (`AnkiPackageReader`; a widget test's fake clock can't wait for one), zipping an export with its files on another isolate (`ExportArchiver`, for the same reason), and the Pomodoro timer's side of the system (`PomodoroPlatform`, `FakePomodoroPlatform` in tests).
- **Material's `Card` widget clashes with the domain's `Card`**: files that need both import `material_ui` with `hide Card`.
- **State: `flutter_bloc`, uniformly.** Blocs fold repository streams into immutable states with value equality, are provided as close to their consumer as possible, and log through one global `BlocObserver`. Why: the data layer is stream-shaped, event → state machines test without widgets, and one pattern everywhere is itself the point.
- **Value equality comes from `equatable`**, mixed in (`with Equatable`), for models and states alike; `leancode_lint`'s `missing_equatable_props` catches a field left out of `props`.
- **Rich text is `RichContent`** (`lib/src/core/content/`): the editor's document JSON plus a `formatVersion`, pure Dart, so models and the server can use it. `lib/src/core/editor/` converts it to and from the editor's `Document` (giving tables the decimal sizes the editor casts to: JSON from imports and archives may hold whole numbers, or none; our fork of the editor fits a read-only table's rows to their cells, which upstream's leaves at their stored height) and holds the one editing widget (`RichContentEditor`) and the one read-only renderer (`RichContentView`).
  - **Blocks** beyond the editor's own (callout, spoiler, formula, image, sound, YouTube video, study button) keep their type and attributes as named in `core/content/content_blocks.dart`, which `RichContent` reads too (their text counts in `plainText`, and a video or study button isn't blank). An `EditorBlocks` lists how each type is built, which can be inserted (the format bar's Insert sheet and the desktop's `/` menu show the same list) and its typing rules; the core's are `coreBlocks`, and a feature's (the study button, `feature/study/`) reach the editors through `EditorBlocksScope`, which `PersonApp` fills, since the core can't know features. One builder serves editor and view: a block asks `editorState.editable`, and read-only shows what it is for (a spoiler hidden until revealed, a formula set, a video's play). A block taken whole mixes in `AtomicBlock` (selection, the remove button); unknown types show "needs a newer version" and stay as they are. New block types keep `formatVersion` 1: raising it would hide whole studysets and paths in older apps.
- **Time comes from `clock.now()`** (`package:clock`), never `DateTime.now()`, in server and app. Why: scheduling, streaks and expiry are deterministic under test (`withClock`, `fakeAsync`), and because Serverpod itself reads `clock.now()` (e.g. session expiry), `withClock` controls its behaviour too.
- **UTC for storing and sending, local time for display and day boundaries.** Serverpod sends `DateTime` as UTC and stores it in `timestamp without time zone` columns. Compare and compute in UTC; call `.toLocal()` only to display. "Due today", streaks and daily goals use the learner's **local** calendar day, computed on the client; the server never needs day boundaries.
- **Build configuration comes from `--dart-define`.** The client holds no secrets, ever.
- **Material comes from `package:material_ui`** (Cupertino from `package:cupertino_ui`), never from `package:flutter/material.dart`, which Flutter deprecates from its November 2026 release on. Packages still on the SDK's copy get the app's theme and localizations through `MaterialUiCompatibilityBridge` in `MaterialApp.builder`; remove it once they have moved.
- **Navigation is `go_router`**: the routes in `lib/src/app/app_router.dart`, their paths in `core/router/route_paths.dart`. Four tabs (Home, Library, Explore, Profile) are the branches of a `StatefulShellRoute`, each keeping its own stack of screens alive while another tab shows (`tabsContainer`, go_router's indexed stack plus `TabVisibility`, which a screen depends on to stop something once its tab is hidden); `AppShell` shows them as a navigation rail from 600 dp on and as a navigation bar below (Material 3 Expressive's shorter one, 64 dp, set in the theme: Flutter's Material library has none yet), floating above the bottom edge (`FloatingNavigationBar`): the tab screens run on beneath it (the shell's `extendBody`), every list ending with room for it (the body's bottom padding) and floating buttons lifted by `FloatingNavigationBar.coverOf`, with the background of the screen in view around it in between (`ScreenTint`); it steps aside inside a studyset (its screen, notes, paths and dictionary: `RoutePaths.isInsideStudyset`), whose own bar at the bottom takes its place, leaving a strip as high as the system's bar (decided 2026-10-11, as Material 3 puts toolbars below the top level and Vaia does), and choosing the open tab again returns to its first screen. A tab's own screens are routes nested in its branch, reached with `context.go` (a studyset's screen in the library), and **the screens behind a studyset are its path** (decided 2026-10-11; `feature/studyset/widget/studyset_navigation.dart`): a topic, at any depth, opens with the screens of the levels between pushed beneath it (`openStudysetWithin`; the same route, keyed by id and named by its location, so several stack), the path's steps and the levels a long press on back lists go up by popping to that page (`goUpTo`), and a studyset opened from elsewhere gets its levels beneath it (`openStudyset`). So back, Android's predictive back, iOS's swipe from the edge and the path all lead one level up. Below expanded windows the topics open in a sheet from the end of the path's line, never a drawer, whose swipe from the edge is iOS's way back. A studyset's notes and learning paths (working through one, writing it, writing a step) are routes of the Library branch too, so the tabs stay beside them. What needs the whole window (studying, a card, a PDF, settings) is a top-level route over the tabs, reached with `context.push`. Explore lists what is published (see *Explore and copies*).
  - **Routes carry ids, never objects.** A screen finds its item in the repository's live list with `FoundById` (`core/widget/`): a spinner until the list arrives, then the item, or a screen saying it no longer exists (deleted on another device, or an old link). A screen that deletes its item closes, and keeps showing the item while it goes. Why: every location can be built from its path alone, and no screen goes on showing something that has gone.
- **The look is one theme** (`core/theme/`): `appTheme(brightness)` with hand-written light and dark colour schemes around the brand's cobalt and copper; `Theme.of(context).extras` (`ThemeExtras`) for the colours Material has no role for (states, the four FSRS ratings, a starred star); `PaletteColor` for the 20 colours a person can give content; `TopGradient` for a colour washing down from the top of every tab's first screen, Settings and Connect Claude (cobalt) and of every screen of a studyset (its colour, own or inherited, which `_inStudyset` in the router hands down as `ScreenColor`): the studyset, studying it, its cards, writing them, its notes, paths and dictionary; behind a clear app bar, in dark mode a deep shade (42 % of the colour over black), in light mode a tint, held to WCAG AA for the theme's text. From medium windows on, a `TintFrame` takes the colour of the screen in view (`ScreenTint`, claimed by the `TopGradient` whose tab shows and whose route is on top), fading out by 60 % of the height: the rail sits on it, transparent, and the screen is a rounded panel inset beside it, its own wash half as strong; screens over the tabs (studying, the card browser and composer, Settings) sit in the same frame without the rail (`FramedScreen`). As Vaia does, decided 2026-10-10. The PDF reader and the reading timeline stay plain. `theme_test.dart` holds text to WCAG AA contrast (4.5:1) on each palette colour and on the dark theme's primary colour and surface. **Fonts are bundled, never fetched** (`core/theme/app_fonts.dart`, `assets/fonts/ATTRIBUTION.md`): Noto Sans is the app's font on every device (Latin, Greek with polytonic, Cyrillic); Noto Sans JP, SC, TC and KR fill in Japanese, Chinese and Korean, Japanese first unless a studyset's language puts its own first (`StudysetLanguageScope`, which every studyset route wraps its screen in through `_inStudyset`; Han characters take the first listed font's shapes); each is one variable font holding every weight, and their OFL licences join the licence list. Scripts none covers (Arabic, Hebrew, Thai, …) fall back to the device's own fonts. A studyset's languages (`frontLanguage`, `backLanguage`, BCP 47) are given together or inherited together from the nearest studyset above (`languagesOf`). Never add `google_fonts` or anything else that downloads fonts. Layouts follow Material 3's window-size classes (`context.windowSize`, `core/layout/`), and running text stays within `readableContentMaxWidth`. Icons are Lucide (`lucide_icons_flutter`), except a starred star and the marks of a choice not yet checked (filled box and circle), which need the fill Lucide lacks; the Material icons left in older screens go when those are restyled. Shared widgets are in `core/widget/` (`SectionCard`, `AppLoadingIndicator`, `confirmDestructiveAction`, `FoundById`, `TitleWithKicker`, `ContentCountsLine`, `showUndoSnackBar`, `ScrollHidingFab`, `PaletteColorPicker`, `CountStepper`, `RequiredFieldLabel` and `FieldInfoButton`, `ReorderToggle`, `ReorderGrip`, `liftWhileDragged`, `LaunchScreen` (what shows while the app opens: the kingfisher, "Halkyon" over "Learn"), `SliverCentered` and `EmptyListMessage` (what an empty list shows: icon, "No … yet", and which menus add some), and the sheets below).
- **Lists that can grow long are built as they scroll into view**: a studyset's tabs are `CustomScrollView`s holding each list's slivers (`SliverList`, `SliverReorderableList`), and what a list says in place of its rows sits in a `SliverCentered`. Never a `shrinkWrap` list inside another scrolling view: that builds every row at once, and 1,000 cards then froze the window for seconds, long enough for the desktop to offer to kill the app. Rows are put in another order in a Reorder mode (`ReorderToggle` beside the list; a `ReorderGrip` in place of each row's options; positions from `core/sync/fractional_position.dart`), never with Flutter's default drag handles, which on the desktop lie over a row's trailing end, its … button. Flashcards have no Reorder (decided 2026-10-08): they keep the order they were created or imported in, which new cards and cramming follow, and sub-studysets put lessons in order; no one drags a card among a thousand. **The card browser** (`feature/card/widget/card_browser_screen.dart`, decided 2026-10-09: the older design, gaps filled) opens from Browse beside the flashcards' tab, over the tabs (`/studysets/:id/flashcards`, `?subtree=1` with everything inside): grid or detail (one card beside the list; on a phone above a strip), as last left per studyset on the device (`DeviceSettings.detailedBrowsers`); search over front, back, reading, meanings, hints, explanation and tag names (katakana as hiragana, romaji finding kana; `card_browser.dart`); sort (own order, newest, oldest, A–Z by reading, due first); a filter sheet (kind, one tag, state: new, learning, due, buried, reported); and choosing cards (a toggle; shift-click or long-press for a range) to move (`showCardMoveSheet`), tag, untag, start afresh or delete with Undo. It reads every card of the studysets, as lists do (schedules from the review repository's cache).
- **Forms and pickers open as sheets**: `showAdaptiveSheet` with a `SheetScaffold` (title and close button on top, the body scrolling with a scrollbar, always shown on the desktop, and Material 3's lines under the title and above the footer while some of the body lies beyond them, the action pinned below) shows a bottom sheet below expanded windows and a dialog from there on. A sheet that asks before discarding changes (a `PopScope`) passes `swipeable: false`: a bottom sheet swiped away closes without asking its `PopScope` (Flutter issue 152120, open with v3.47). A sheet offering an Undo passes `hasSnackBars: true`, so its snack bars show on it rather than behind its barrier. A sheet running something that mustn't be cut off (an import) is `busy` meanwhile: its close button is greyed and back does nothing; it is shown with `swipeable: false`, since a swipe closes a bottom sheet regardless. A screen whose blocs import (a studyset's PDFs and paths) greys its back button and path likewise, as leaving would close them. A picker that changes something outside it reports every change as it happens (`onChanged`) rather than as its result, since a sheet swiped away has none.
- **Every text field has a limit** (decided 2026-10-11): `TextLimits` (`core/widget/text_limits.dart`) by what it holds (a name 40, a secondary title 60, a title 120, an item 200, a search 200, a citation's part 500, a message or an address 2000, a passage 5000), or the instance's own where the server checks it (`PublicProfile.max…`, `maxSlugLength`), with `buildCounter: nearLimitCounter`. Text can still be longer where it came from elsewhere, so every place showing a name cuts it (`maxLines` and an ellipsis, in a `Flexible` within a row); `test/app/long_names_test.dart` shows names beyond every limit on the main screens of a phone and a desktop, and fails on any overflow: add a screen there when it shows names.
- **Text comes from `AppLocalizations`** (Flutter's gen-l10n, `lib/l10n/`): no string a person reads is written into a widget. The app speaks British English, American English and German (informal "du"; the terms of `app_de.arb`, such as Lernset and Karteikarte, stay consistent). `app_en.arb` is the template and holds British English, also the fallback for any language the app doesn't speak; `app_en_US.arb` overrides only what American English spells differently, and `app_en_GB.arb` is empty, there so that Material and dates format the British way. A new message goes into `app_en.arb` and `app_de.arb` together (`test/l10n/arb_files_test.dart` fails otherwise, and checks the placeholders), and into `app_en_US.arb` if its spelling differs. `resolveAppLocale` (`core/l10n/`) picks the language, never by country alone, and both `MaterialApp`s take `appLocalizationsDelegates`, not gen-l10n's list: that brings flutter_localizations' Material texts, which serve only the SDK's Material, so material_ui's widgets would stay English. Errors found in data go up as values the widgets put into words (`PathImportProblem`, `PathImportTexts`), never as English sentences. The editor's own placeholders are replaced by ours (`buildersWithUnknown`), since the editor's texts follow a process-wide locale of their own. The generated `lib/l10n/app_localizations*.dart` are committed and never edited, analysed or formatted by hand. `flutter pub get` writes them anew, so CI fails when they don't match the `.arb` files.

## Testing

- **By layer**: models with plain unit tests; repositories against fake gateways; blocs by asserting the emitted state sequence; widgets pumped in isolation, presentation widgets fed with ready-made data.
- **Server endpoints are tested for real** against the embedded Postgres (`withServerpod`, `dart test` in the server package), including the denial tests.
- **Test dependencies come from a test factory**, `testDependencies(...)` in `test/support/test_dependencies.dart`, filling in fakes by default (a factory in `lib/` couldn't reach the fakes in `test/`), so a new dependency changes one place. Don't make dependencies nullable to spare tests: that moves null checks into production code.
- Shared fakes live in `test/support/`; fakes pre-stub their streams so an incidental subscription doesn't fail the test.
- **Editor in widget tests**: host `AppFlowyEditor` in a scrollable body with a fixed-height slot (an over-constrained `Expanded` hangs the test), and `pump(const Duration(seconds: 1))` after editor changes to flush its timers. Test the document ↔ JSON round-trip as plain unit tests.
- **No `Stream.first` on a repository's stream in code that widget tests run**: `first` waits for the subscription's cancel future, which for a broadcast stream is a completed future of the root zone, so under the widget test's fake clock it never completes. Read once with the repository's `all()` instead. For the same reason a repository never keeps a completed future to chain later work on (a `Future.value()` made in `setUp` runs what waits for it in the real zone): a queue of writes starts empty and empties again when idle (`SerialQueue`, `core/async/`), and a cache keeps values, not futures (`GeoMapRepository`). Assets are read for real, which the fake clock doesn't advance: a widget test showing a map gets the maps read beforehand (`loadedMaps(tester)` into `testDependencies(maps: ...)`), and one showing how a character is written its strokes (`loadedStrokes(tester, characters)` into `testDependencies(strokes: ...)`).
- **Text inside the editor's views** (`RichContentView`, `RichContentEditor`) is rich text: find it with `findRichText: true`.
- **Layout-sensitive widget tests** set `tester.view.physicalSize` explicitly, so a responsive branch isn't missed for the wrong reason.
- **A menu item acts in the frame after its menu closes** (`MenuAnchor`): pump twice after tapping one before expecting what it does.
- **Server tests of what writes synced rows** run on the sync engine's database, as the server does (the generated test tools build the pod without it): pass the interceptor to `withServerpod`, call `initializeOfflineSync` in `setUp`, and write as a person through `transactionForUser`. Its writes take transactions of their own, so these tests turn rollback off and delete what they made (`test/integration/asset/asset_test.dart`).
- **Sync tests run against a running server** (ADR 0001, rule 9: the client's and the server's sync engines can't share one process): `halkyon_learn_flutter/test/sync/`, tagged `requires-server`, skipped by `tool/check.sh`; run them with `flutter test -t requires-server` while `serverpod start` runs. Every device is a client with in-memory session storage and its own local database, with the app's real gateway and repository on top; tests sync explicitly. What only the operator could see (a blanked payload, a row that must not exist) is checked with SQL against the development database, through the embedded PostgreSQL's Unix socket where there is one. An honest device only syncs the spaces the server announces, so a forged write never leaves it; the server's own check against modified clients (`_assertCanMergeInboundSpace` in the sync engine) has no test yet. They live in the app package because the server's Docker build resolves the server's dependencies without the client package. They must not initialize a Flutter test binding (no `testWidgets`), which would replace real HTTP with a mock.
- **Widget tests of screens start the whole app**: `pumpApp(tester, testDependencies(...), location: …)` (`test/support/pump_app.dart`) pumps `SessionGate` with the real `PersonApp`, opened at the screen's path, and settles once `PersonScope` has opened the person's data. From there a test moves as a person would: `openTab` (the tab's label alone may also be a screen's title), taps, `pageBack`. Data goes into the fake gateways before pumping; they work in memory (`FakeStudysetGateway`, with `failReads` for the error path), and a repository of the test's own on the same fake gateway writes what the app then reads. `pumpAndSettle` never ends while a spinner shows (`AppLoadingIndicator` animates without end): pump frames instead where one is expected. A widget pumped on its own that reads `AppLocalizations` needs `AppLocalizations.localizationsDelegates` in its `MaterialApp`.

## CI and deploys

- **CI** (`.forgejo/workflows/ci.yaml`) runs `tool/check.sh` on every push, on the Forgejo runner on sahaquiel (`deploy/runner/`), except for a commit that passed before: the job marks passed commits in the runner's home (`checked/`, kept 90 days), so a branch's commit fast-forwarded into `main` goes straight to its deploy. Jobs run directly on the host as the unprivileged user `runner`, never in containers: Docker access would be root on the production server. The runner's systemd unit keeps jobs away from production (no `/home`, no Docker socket, Docker's container networks unreachable, memory and CPU capped); keep it that way when changing it. `runner` has user ID 61000 and must never share one with a container's processes (PostgreSQL and the server run as 999, not remapped): the same user ID can read their environment and files. The job's first step (*Sandbox*) fails if any of this leaks. Flutter is pinned in the workflow (`FLUTTER_VERSION`): move it together with the local one. The server tests there get throwaway passwords from the environment (`SERVERPOD_…`, new for every job), since `passwords.yaml` is never committed; without them they exit with code 1 and no message (Serverpod's own template does the same). CI also runs `serverpod generate` and fails if the committed generated code differs from its output. So never run `dart format` over the server's `lib/` or `test/` as a whole: it reformats the generated code (`lib/src/generated/`, `serverpod_test_tools.dart`); format the files you wrote. CI runs on pushes only: a pull request from a fork would run its code on the production server.
- **A green `main` deploys learn.halkyon.eu** through `deploy/redeploy.sh`, the only command the runner's key may run as `deploy` (pinned in `authorized_keys`, from sahaquiel itself only). It deploys the commit CI checked while it is still the tip, as a fast-forward, only if every new commit is signed by a key in `deploy/allowed-signers` (the deployed copy decides), after dumping the database (the last 3 dumps stay in `~deploy/backups`: the way back from a deploy, not backups), applies the migrations in a run of their own before switching (`--role=maintenance --apply-migrations`: a failing migration stops the deploy while the old server keeps serving, so migrations must keep the old server working until the switch), and fails unless the new server answers. A failed deploy leaves the checkout at the deployed commit; run the workflow again to retry. By hand: `ssh sahaquiel-deploy halkyon-learn/deploy/redeploy.sh <commit>`. Never `docker compose down -v` there.
- **Backups** (`deploy/backups.md`): every hour, `deploy/backup.sh` (a systemd timer, as `deploy`) dumps the database, which holds people's files too, then bundles each published studyset's repository (`git bundle`, from the server's `repositories` volume), and restic stores both in a Hetzner Storage Box (24 hourly, 7 daily, 4 weekly, 3 monthly); the box's own snapshots protect it from a taken-over server. A restore takes the instance back up to an hour, and people who started since lose what their device held: their app no longer finds its session, starts a new anonymous account and deletes the old folder. So a change that makes devices lose their session, or keeps deleted people's data longer, changes what a restore costs; say so in `deploy/backups.md`.
- **Releases** are the project's, not an instance's: the release, the Snap Store listing and the F-Droid listing point to the project's page on Halkyon's website, halkyon.app/learn (its own repository, `project-halkyon/website`; learn.halkyon.app redirects there); the builds name learn.halkyon.eu, the project's instance (`SERVER_URL`), which a new install offers first (see *Instances*). They come from a developer's laptop, never from CI: `tool/release.sh` on `main` as pushed and green in CI. Versions follow semantic versioning, worked out from the commits since the last release (`feat` raises the minor number, `fix`/`perf` the patch number, breaking ones the major from 1.0 on; other types release nothing), so **commit messages keep their type prefix**. It writes the version and a raised build number into the app's pubspec in a signed release commit (`chore(release): vX.Y.Z`) with a signed tag (a key in `deploy/allowed-signers`), builds one universal APK from a checkout of the tag, signed with the app's key, and the snap (`snap/snapcraft.yaml`, the committed HEAD, built in LXD), both with `SERVER_URL=https://learn.halkyon.eu/`; signs `SHA256SUMS` with GPG; pushes only once both builds passed their checks; and publishes on code.halkyon.app through Forgejo's API and in the F-Droid repository. The Snap Store upload is off until `snap_store_channel` in the script names a channel. A run that stopped after its release commit and tag carries on with them. The key's password and the Forgejo token are asked for at each run and never stored. The app's key is a PKCS #12 keystore outside the repository (also in the maintainer's password manager); there is no Google Play, so it is the app's identity: lost, installed apps can't be updated. Its certificate's SHA-256 is in `tool/release/android-certificate.sha256`, and the script releases no APK signed otherwise; other release builds fall back to the debug key (`android/app/build.gradle.kts`), so `flutter run --release` works anywhere. The APK also goes into the project's own F-Droid repository, https://halkyon.app/fdroid/repo, through `tool/fdroid_repo.sh` (fdroidserver v2.4.5): its settings and the app's listing are in `tool/release/fdroid/` (the anti-features TetheredNet and NonFreeNet declared with their reasons and F-Droid's own icons); its index is signed with a key of its own (a PKCS #12 keystore beside the app's, password asked for, certificate pinned in `tool/release/fdroid-index-certificate.sha256`, since a new key makes the F-Droid app refuse the repository); and it is uploaded to ramiel as the user `fdroid`, whose key can only write into the folder Caddy serves (rrsync), the APKs first and the indexes last.

## Rules the code relies on

- **Authentication**: server-side sessions (`ServerSideSessionsConfigFromPasswords`), not JWT, for immediate revocation. Scopes are copied onto the session row, so changing a person's scopes means updating their session rows too.
- **Identity**: `AuthUser` is the person; identity providers (anonymous, GitHub) attach to it. Content references `AuthUser`, never a provider account.
  - **GitHub keeps only the numeric GitHub ID**: our own slim login (`lib/src/auth/github.dart`) on Serverpod's GitHub helpers; Serverpod's stock GitHub login endpoint, which stores name, email and avatar, is never exposed. GitHub sign-in is on only where the passwords hold `githubClientId` and `githubClientSecret` (in production `GITHUB_CLIENT_ID`/`GITHUB_CLIENT_SECRET` in `deploy/.env`); the app asks `instance.info()`.
  - **Linking keeps the person** (`GitHubUpgradeEndpoint`, `requireLogin`): the GitHub ID joins the signed-in anonymous person, who becomes identified and loses the anonymous identity, in one transaction. **Signing in** (`GitHubSignInEndpoint`) only reaches people who have a GitHub account already, and only without a session: Serverpod refuses to issue another person's session inside a signed-in call. Refusals are `GitHubSignInRefusedException` with a reason.
  - **Switching devices' person** (the app's `switchToGitHub`): swap the session directly, never through "signed out" (which would start another anonymous account), then end the anonymous session left behind.
  - **On Linux and Windows the browser half is the app's own** (`LoopbackWebAuth`, `feature/auth/data/`): the system browser and a loopback address (RFC 8252), put in place of flutter_web_auth_2's own implementation (checked with v5.1.0) before each sign-in. The package's front ends every pending sign-in when the app is resumed, and on the desktop the window regaining focus resumes it, so its loopback server closed before GitHub answered. Ours waits for GitHub, the profile's Cancel (shown only there), choosing another tab, a change of person, or 5 minutes; the regression test drives the package's front through a focus change. Remove it once the package stops ending sign-ins on focus there. Serverpod (v4.0.3) reports every cancelled browser round trip as an unknown error, so the gateway knows a cancel only by having asked for it; a cancel on GitHub's own page, or leaving the browser on Android, still shows "Couldn't sign in with GitHub."
- **Authorization lives in endpoint code**; Serverpod has no row-level security. Load content filtered by the session's `authUserId`, never by an owner ID the client sends. A row another person owns must be indistinguishable from a missing one. Features involving other people require the `identified` scope, declared via `requiredScopes`, never checked ad hoc.
  - **The reference is `DevicesEndpoint`** (`halkyon_learn_server/lib/src/auth/`) with its tests: the owner check sits in the same query (`WHERE "id" = $1 AND "authUserId" = $2`), zero rows means `NotFoundException` (shared, `lib/src/common/`), and every rule has a test that asserts the refusal. Each query carries its SQL as a comment.
  - **Serverpod's own helpers aren't guarded**: `ServerSideSessions.revokeSession` deletes by id alone, and `listSessions` without `authUserId` returns everyone's sessions. Never pass them anything a client sent.
  - **Granting a scope** goes through a function like `grantIdentified`, which updates the person *and* their session rows.
- **Content envelope**: synced content tables carry `keyId` and `nonce` (nullable; `null` = plaintext, the only value, ADR 0002 rule 6) and `payload` (`ByteData`): UTF-8 JSON with a `schemaVersion`, encoded and decoded by one codec per feature (e.g. `studyset_payload.dart`) on top of `core/content/payload.dart`, which also holds the blanked `redactedPayload`. Content fields live inside `payload`: a new field raises the schema version instead of migrating a synced table, which would lock older app versions out of syncing (ADR 0002, rule 3). A payload in a newer schema is skipped with a log, never shown half-read.
- **Synced data** (`database: sync`, `serverpod_offline_sync`, ADR 0001): every person has their own local SQLite database (`<app support>/people/<authUserId>/sync.sqlite`). Its gateway writes through the `OfflineSyncDatabaseSession`, **names the columns of every update**, and calls `SyncService.requestSync()` after every write; the repository re-reads after its own writes and when a sync merged changes (`remoteChanges`). `SyncService` syncs on triggers (at start, 2 s after the last write, on resume, every 5 min), one sync at a time. **Deleting blanks the payload first** (ADR 0002, rule 2): the engine keeps deleted rows until the space is deleted. Since blanking can't be undone, an Undo comes before it: `StudysetDeletions`, opened with the person's data, keeps every studyset deletion that can still be undone; lists leave out what it reports as pending, and once the time to undo is over it deletes the whole subtree, what is inside first, each studyset's cards blanked before it. Widgets delete and undo through it directly (`deleteStudysetWithUndo`), not through a bloc: deleting a studyset from its own screen closes that screen and its bloc, and the deletion and its Undo must outlive them. It closes first with the person's data and drops what still waits. Tags differ, being values (a tag is its name whatever its case, which it keeps; ADR 0004 rule 2): one is blanked and deleted at once, and its Undo makes a tag of that name and colour anew, on the same cards. Order by the UUID v7 id (ADR 0001, rule 7). Access to synced rows is the engine's: a person syncs their personal space and the spaces granted to them; the endpoint rules above guard everything that isn't synced.
  - **Content people edit is append-only revisions** (ADR 0001 rule 2, ADR 0003; the reference is `CardRevision`): a revision is never updated except to blank it, and `parents` lists the revisions it replaces. A save's parents are all current heads, which is also how a conflict is resolved. `mergeRevisions` (`feature/card/model/card_merge.dart`) merges several heads per side against their newest common ancestor; a side changed differently is a conflict the person settles, never merged silently. Deleting blanks every revision, then deletes.
  - **Reviews are an append-only log** (`ReviewLog`: card, rating, `reviewedAt` in UTC, `durationMs`); taking back a rating deletes its row. Besides ratings it holds a fresh start (`reset`), which never counts as studying; older app versions skip a mode they don't know. A burial isn't in the log but on the card: `card.buriedAt` sets it aside until the next local day begins on the device looking, out of every queue and count (Unbury clears it), read with the cards every list reads anyway. A card's schedule is never stored: `scheduleOf` (`feature/study/model/card_schedule.dart`) replays its reviews in time order with FSRS-6 (`fsrs`, pinned exactly), and only the final due date is fuzzed, from the card's id and review count, so every device computes the same. Studying a studyset covers everything inside it but archived studysets (`studiedIn`), and the study queue (`study_queue.dart`) takes each studyset's cards as a group: due cards of all of them earliest first, then each studyset's new cards in tree order, within each studyset's daily limits of new cards and reviews (its study options, `StudysetStudyOptions` in its payload, schema 3; without, the person's new cards a day and no limit on reviews; cards in their learning steps come anyway), in the order the studied studyset's options ask for (reviews first, mixed, new first). The library's and the topics' counts and badges add up the same way, each card on the schedule a session studies it on (`card_study_mode.dart`: written by hand while Write by hand is on, back first while Back first is; `ReviewRepository.schedulesAsStudied`); Home's due forecast still counts front to back, as it reads cards without their content. Deleting a card deletes its reviews.
  - **Order is a fractional index** (`fractional_indexing_dart`, rocicorp's algorithm) in its own column, last writer wins. Two devices appending at once can pick the same key, so ties sort by id.
  - **Tree, tags, modes, settings and files** (ADR 0004): a studyset's `parent` and sibling `position` form the tree, and studysets in a cycle or without their parent show at the top level (`topLevel`); its `archivedAt`, `starredAt` and `updatedAt` are columns, and every change of a studyset writes `updatedAt` with it (the library lists starred studysets first, then the most recently changed); a card's tags are `card_tag` rows, one per assignment, never updated, and a duplicate counts once; `review_log.mode` is `review`, `reverse` (its own schedule), `write` (by hand, its own schedule) or `cram` (never moves a schedule), and a schedule replays the reviews of one mode; `person_settings` is meant as one row per person, and extra rows fold into the one with the smallest id; an `asset` row stands for a file in private cloud storage, found by its hash. **Never derive a row's id from a person or anything else guessable**: ids are unique across the whole table, so someone could claim it first in their own space and break the owner's sync.
  - **Files** (ADR 0004, rule 5): `AssetEndpoint` (`requireLogin`) takes uploads into the caller's own space by SHA-256 (`startUpload` gives Serverpod's upload description, or null when the file is there already and claims it anew; `finishUpload` keeps it only if its content has that hash) and serves an asset row's file to its space's owner and members only, `NotFoundException` otherwise, also for a deleted row (the sync engine hides tombstones). Files live in the private database storage at `spaces/<space>/<sha256>`, at most 20 MB each, and within limits (`file_limits.dart`, from the environment: an anonymous person 100 MB, an identified one 500 MB or what a `file_allowance` row grants them, the instance 15 GB), checked when an upload starts (unfinished uploads count, or they could fill the disk unnoticed) and again when it finishes (only what is kept counts), refused with `FileLimitExceededException`; `deletePerson` deletes the person's files and allowance, and `FilePruningFutureCall` deletes daily those no live row refers to (or never finished), 7 days after they were last claimed. In the app, `AssetRepository` (`core/asset/`) keeps files by hash in the person's folder (`files/`, uploads waiting marked in `files/uploads/`), uploads after each sync that went through, and downloads on demand, checking the hash; every hash is checked to be one before it becomes part of a path. Content names a file by the URL `asset:<asset row id>` (`assetUrl`, `RichContent.assetIds`): images are the editor's `image` block, sounds an `audio` block (`core/editor/file_blocks.dart`), and any other URL is never fetched. `StoredFileImage` loads through the repository like Flutter's `NetworkImage`, keyed by the person's repository too, so no image crosses people.
  - **PDFs** (`feature/pdf_source/`): a `pdf_source` row per PDF in a studyset (payload: title, file name, the asset of its file, page count, reading deadline; last writer wins) with its `position` among the studyset's files in a column of its own (decided 2026-10-08; the Files tab's default order, Custom order, which Reorder changes; PDFs from before it stand at `a0`, in the order they were added) and an append-only `pdf_reading` log (`reached` with a page, `reset`): the furthest page is the furthest reached since the last reset, so devices never conflict. A PDF's file is an asset of up to 50 MB (other files 20 MB): the server takes up to 50 MB and deletes a stored file over 20 MB that doesn't start with `%PDF-`. Deleting a PDF deletes its asset row too unless another PDF has the same file (asset rows are otherwise never deleted). Cards and notes tied to a page keep it in their payload (`source`, `core/content/pdf_page_ref.dart`; card schema 4, note schema 2); a screenshot is an `image` block with `pdf` and `page` attributes.
  - **Dictionaries** (decided 2026-10-08): reference data the server holds for everyone, owned by no one, never synced (`dictionary_entry`, `dictionary_source`, `lib/src/dictionary/`). `halkyon_learn_curator` (AGPL, no Serverpod, so the server never links its code) turns each source's download into one format, a dictionary file (its README is the contract), with an adapter per source; a dictionary is one source in one meaning language (`jmdict` in `en`), of words or of characters (`kind`: JMdict, KANJIDIC2 with the kana), and the endpoints take it by those two, so a search for words never brings kanji; everything specific to a language (kana, romaji, meanings without "to") goes into each entry's `keys` and `looseKeys`, so the server's search knows no language: exact key, exact loose key, key starting with the query, loose key starting with it, each by the source's `rank` (`DictionaryEndpoint.search`, LIKE on barred keys, no extension needed). The server loads the files in its `dictionaries` folder at start and hourly (`DictionaryReloadFutureCall`), those whose SHA-256 changed, each source replaced in one transaction, one load at a time (`dictionary_loader.dart`). EDRDG's licence asks a server showing its dictionaries to update them at least monthly (decided 2026-10-09: instances keep current by themselves, weekly): the `dictionaries` service of `deploy/compose.yaml` runs the curator in a container of its own (`halkyon_learn_curator/Dockerfile`, user ID 10001; the AGPL tool never enters the server's image or process), `curate update --every 7` downloading EDRDG's files of the day (`JMdict_e_examp.gz`, `kanjidic2.xml.gz`; the JLPT lists are static, in the curator's `data/`), converting them and renaming each over its old file in the shared volume, which the server mounts read-only. A card made from an entry copies what it needs and keeps `(source, sourceId)` as a soft reference only (`CardContent.entry`, card schema 6), so loading a newer version never touches cards. **Corrections reach cards by offer** (decided 2026-10-09): every entry has a `version` (16 hex digits of the SHA-256 of what cards take of it, made by the loader), a card keeps the one it took (`entryVersion`, card schema 8), and `DictionaryEndpoint.changed` returns, of the entries named with the versions cards keep, those that are another now. The card browser says how many cards are behind (`DictionaryCorrections`, `DictionaryUpdates`) and takes the corrections at a tap: `CardRepository.takeDictionaryUpdate` writes the entry as it is now over the revision in which the card last took it, as a revision of its own beside whatever came after, so the revisions merge per side as two devices' edits do (ADR 0003): a side only the dictionary changed takes the correction, a side the person changed too is a conflict settled in the composer. A card made before versions is never offered one. Every file carries its credits, which the app shows (About, and at the foot of the dictionary). In the app, a studyset's dictionary follows from its front language (`StudysetDictionary`, through `StudysetLanguageScope.of`): no attaching or detaching, and nothing dictionary-shaped shows without one (Add material › From the dictionary; Look up beside a vocabulary card's front). A word's example sentences (Tatoeba, through JMdict's examples edition) become its card's hints, up to three, the word in bold; Look up gives them to a card without hints of its own. `DictionaryAdder` makes the cards, each entry once per studyset, a word tagged `JLPT-NX` and with its word class by the first sense (`Noun`, `Verb`, `na-Adjective`); Look up in the composer gives a card the same tags (`DictionaryItem.tagNames`), and an entry looked up again on that card swaps the tags the one before brought (`isDictionaryTagName`). Where the instance also holds the language's characters, the dictionary screen has a tab for them (Kanji & kana: the kana, the JLPT levels and the school grades to add whole, each under its name), a character becoming a character card tagged `JLPT-NX`, `Grade N` or `Secondary school`, and `Hiragana` or `Katakana`; a character card's Look up searches the characters. Beside them a Radicals tab lists the 214 Kangxi radicals by their strokes (`feature/kanji/model/radical.dart`: the ideograph each of Unicode's Kangxi radicals stands for, the forms Japanese writes it as inside kanji, Unicode's English name), each listing the kanji KANJIDIC2 files under it (its classical radical, the label `radical:N`, which the curator writes), to add or see; that tab shares the characters' bloc, so what it adds shows as held in a search too. The curator stamps its files with how it converted them (`.conversion`, `conversion` in its `update.dart`): raise it whenever it writes something else from the same download, and instances convert again as soon as they run the new curator rather than up to a week later.
  - **The Pomodoro timer** (`feature/pomodoro/`): the profiles follow the person (`PersonSettings.pomodoro`, settings schema 4), the running timer is the device's own (`PomodoroRun` in `<person folder>/pomodoro.json`, `PomodoroRunStore`). `PomodoroRun` is a pure state machine on the time: a phase that ends leads to the next, running if its profile starts it by itself, else paused at its full length; a timer kept while the app was closed catches up (`caughtUpTo`). `PomodoroBloc` sits above the router (`PersonApp`), and `PomodoroPresenter` hands every change to `PomodoroPlatform` as a `PomodoroNotice` in the app's language, and its buttons back. **Android:** the app's own Kotlin (`PomodoroService.kt`; MethodChannel `app.halkyon.learn/pomodoro`, EventChannel `…/buttons`): a `specialUse` foreground service with a countdown notification (chronometer), Pause/Resume, Skip, Stop, on Android 16 a Live Update (`NotificationCompat.ProgressStyle`, the cycle's phases as segments, `setRequestPromotedOngoing`); an exact alarm (`USE_EXACT_ALARM`) alerts at the phase's end with the app closed; Do Not Disturb through the interruption filter, the one before kept in shared preferences and restored. With the app swiped away its buttons open it (a service may not start an activity from a notification's button since Android 12, so the notification is posted again with activity intents), and Stop works alone, leaving a mark the app takes at its next start (`takeStoppedOutside`) to forget its timer; a profile's next phase only starts by itself while the app runs. Dart's timers sleep with the device, so `PomodoroBloc` also looks every 15 s while a phase runs, and the alarm's receiver wakes it (`phaseEnds`); a kept timer is caught up before it first shows, or its ended phase would alert again. No notification plugin: flutter_local_notifications had Live Updates only in a prerelease needing compileSdk 37 (checked 2026-10-07). **Linux:** `LinuxPomodoroPlatform`, over package:dbus: where the top bar shows icons (GNOME with the AppIndicator extension, built into Ubuntu; KDE) the timer counts down there with its buttons in the icon's menu (`StatusNotifierItem`, written in Dart and laid out as libayatana-appindicator lays itself out, within the snap's rules, `desktop-legacy` of snapd v2.77: the item at `/org/ayatana/NotificationItem/<id>`, the only path the label's signal `XAyatanaNewLabel` may come from, registered with the watcher by that path, since a snap may bind no `org.freedesktop.StatusNotifierItem-*` name; the menu at `/MenuBar`; registered again when the watcher restarts); elsewhere the timer shows in the app alone, and on GNOME the Pomodoro card suggests the extension until dismissed, asking GNOME Shell to install or switch it on (`InstallRemoteExtension`, `EnableExtension`) or opening its page (decided 2026-10-07: no notification stands in; xdg-desktop-portal's Background status is for Flatpaks only). An alert with sound comes through `org.freedesktop.Notifications` when a phase ends while the app runs, and GNOME's `show-banners` goes off while focusing, through `gsettings`; the value before and the dismissed suggestion are kept in `<app support>/`. Calls to the bus go one at a time (`SerialQueue`).
  - **Citations** (`core/citation/`): a `Source` is plain text parts as the person typed them (authors as full names, one per line; year; title; `container`, what it is a part of; publisher; ISBN; link; `raw`, a citation a file brought, shown while nothing else is cited), and `formatCitation` cites it in the device's `CitationStyle` (APA 7, MLA 9, Chicago 18 notes and bibliography, decided 2026-10-07): it splits and abbreviates the names itself (inverted names, particles, suffixes; names written with `&` or "et al." stay as written), sets a work of its own in italics and a part of something apart from the whole by `container`, and writes a missing year as "n.d." in APA and Chicago. A step's sources (step schema 2) and a studyset's bibliography (`bibliography`, studyset schema 4, last writer wins like the rest of its payload) are lists of them, shown and changed through `SourceList` and `showSourcesSheet`, in reference-list order (alphabetical, whatever order they were added in); the path import reads them too, text as `raw` and the old format's `citation` as `container`.
  - **What a card carries beside its type's fields** (card schema 5): up to six `hints` and an `explanation` (rich text, any type), a `report` that something on it is wrong (category, note, when; studied as before, its badge in the list), and a word's `senses` (glosses, label, JMdict's grammar codes, whose English short forms and meanings are data in `feature/card/model/jmdict_tags.dart`; with senses a word's meanings are their glosses, and `uk` on its first sense shows its reading from the start). Each merges as a field of its own. Deleting a card goes through `CardDeletions` (a `DelayedDeletions`, as notes and studysets), so the list and the session offer Undo. A study session's Undo takes back its steps last first (ratings, burials and deletions alike, across its rounds), so it holds the cards it deletes (`hold`) until it ends, when they go after the usual time to undo; a card deleted in a session that never ended (the app killed) stays. Vocabulary can be quizzed as multiple choice (`quizVocabulary`, `quizAspect` in the study options, settings schema 5; `vocabularyQuiz` draws the wrong answers from the other words studied). **A character card** (`CardType.character`, card schema 7, decided 2026-10-09: one type for kanji and kana) holds the character on its front, its `meanings` (a kana's romaji is its meaning, so kana are studied and written as kanji are) and a kanji's `onReadings` and `kunReadings` as dictionaries write them (katakana; hiragana with the okurigana after a dot); its JLPT level and school grade are tags, its strokes the app's own. **How characters are written** (`feature/kanji/`, decided 2026-10-09: bundled, so writing works offline) is KanjiVG's, in `assets/kanjivg/` (CC BY-SA 3.0, its notice in the licence list): the curator's `curate kanjivg` writes one gzipped JSON file per 256 code points (2.7 MB in all), which `StrokeRepository` (app-wide in `Dependencies`, as the maps) reads when a character of it is first asked for; the strokes are SVG path data in a 109 × 109 square, read with `path_parsing` (pure Dart, so `CharacterStrokes` flattens them into points without `dart:ui`). `StrokeOrderView` traces them in order; `showCharacterSheet` shows a character with its stroke order, meanings, readings and components, from a studied character's answer and from the dictionary's characters. A studied character's answer also lists the words written with it (`example_word.dart`: up to five vocabulary cards of the studysets the session studies, whatever tags narrow it, read with its cards, so offline), each opening its card. **Writing by hand** (decided 2026-10-09, after Kanji Dojo's practice): the study option Write by hand (`StudyOptions.writeByHand`, settings schema 7) has character cards and words written in the session, mode `write` on a schedule of its own (words so written are neither quizzed nor typed). `WritingCanvas` takes raw pointers and claims every touch on it at once (an `EagerGestureRecognizer`), so a stroke never scrolls or swipes to rate; each stroke is checked as it ends (`checkStroke`, `kanji/model/stroke_check.dart`: both resampled to 22 points, shape compared once centred, position, length and direction, against what a loose hand strays by; a stroke closer to one still to come is out of order), a right one drawn as KanjiVG draws it, and the hand's offset from the middle forgiven in the strokes after. Mistakes are strokes that didn't count and hints asked for; two misses in a row show the next stroke faintly. `WritingBody` writes a word character by character, filling in those without a stroke order, and once written (or given up on, Show it) outlines the rating the mistakes suggest (`suggestedRating`: none Good, one or two OK, more Bad; given up, Bad). `stroke_check_test.dart` holds the checker to real strokes, moved, wavering, larger and smaller, reversed and out of order: change its weights only with that test passing.
  - **Imported cards** say where they came from in their first revision's payload (`origin`, card schema 3: `anki`, or `text` for flashcards pasted or read from CSV/TSV); the dashboard doesn't count them as written. The Anki import (`feature/anki/`) reads `.apkg`/`.colpkg` packages of every format on an isolate of its own (`AnkiPackageReader`), zstd through our fork of `pure_zstd` (Anki's frames don't record their size), and its fixtures come from Anki itself (`tool/anki/make_fixtures.py`).
  - **Every client migration keeps a device's rows** (`test/core/sync/device_migration_test.dart` replays each on a database with a row in every synced table of ours; SQLite rebuilds tables to add foreign keys, which Serverpod's runner does with foreign keys off). A new synced table needs a row in that test, which fails until it has one.
  - **Every synced table indexes `spaceId`** (`indexes: <table>_space_idx: fields: spaceId`, e.g. `studyset.spy.yaml`): deleting a space cascades into every synced table, and Postgres doesn't index foreign keys by itself, so account deletion and pruning would otherwise read whole tables.
  - **The client's local migration** (in `halkyon_learn_client/lib/migrations/`, written by the same `create-migration`) creates the device's SQLite schema from the whole definition, so it also contains the server's own tables. Only the synced tables and the bookkeeping ever hold data on a device.
- **Older apps** can't be made to update once published. Before the first sync of each launch the app asks whether the server still works with its build (`minimumAppBuild` in `instance.info()`, from `lib/src/instance/app_compatibility.dart` in the server; the build is the number after `+` in the app's version), and the sync engine refuses an app whose synced tables differ from the server's. Both end in `AppServerMismatchException` (`core/sync/`), and the app says so: a banner on Home and the Sync card in Settings. Raise `minimumAppBuild` in the commit that breaks older apps (an endpoint they call changes or goes), to the build of the first app that works again; a change to the synced tables needs no raise. Every release raises the build number.
- **Instances** (decided 2026-10-10, as Moshidon does): the device keeps an account on each instance it was used with, one in use at a time (`core/instance/`). `InstanceChoice` holds them (`instances.json` in the app's support folder); a new install chooses at first start (`FirstStartScreen`: an address checked with `instance.info()`, or the build's own in one tap), and one from before there was a choice goes on with the build's (`SERVER_URL`), as it has a `people/` folder. `InstanceGate` builds `Dependencies.create(instance, device)` for the one in use and builds it anew on a switch, the previous closed first; `DeviceParts` (version, sounds, Pomodoro platform, device settings) outlive switches. Each instance's people's folders and session are its own (`InstanceStorage`): the build's own instance keeps `people/` and Serverpod's own session key, so existing installs keep everything; others `instances/<name>/people/` and a key of their own. Settings › Switch account lists them (`showAccountSwitcher`): another one switches, Add account chooses another instance, and removing one not in use deletes its folder and stored session (the session stays valid on the instance until it expires). Signing out still ends only the account in use.
- **Asking Claude or searching the web** (decided 2026-10-10; `feature/assistant/`): a device setting, off unless the person turns it on (Settings › Assistant), shows a field at the foot of Home (`AssistantBar`) and a button that unfolds into one (`AssistantButton`): above Create studyset in the Library, above a studyset's bottom bar (floating alone from expanded windows on), and in the card's lower right corner while studying. A switch at the field's start sends to Claude or to a web search and is remembered (`assistantTarget`). Above the field, chips offer the parts of what it is about (`AssistantTopic.parts`: a card's front, back, each hint, its explanation, the studyset's name and description), none chosen, so only what the person picks goes along, each cut at 600 characters; a web search takes the question and the chosen text, never the studyset's name or description. Nothing is answered here: `ClaudeHandOff` opens `https://claude.ai/new?q=…` (cut at 12,000 characters) with the question and a compact line per chosen part, and `WebSearch` opens the chosen engine's results (`SearchEngine`: Brave Search by default on desktops, Ecosia, DuckDuckGo, Startpage, Qwant, Google, Bing, or the person's own address with `%s`), or on Android hands the question to the device's own web search (`ACTION_WEB_SEARCH` in `MainActivity.kt`, Brave Search where nothing answers). Halkyon apps use no LLM API keys: their pricing is opaque and open to abuse, and the person decides each time what an LLM sees. A Claude plan can't drive a chat inside another app (Anthropic allows no third party to sign in with it), so Claude reads studysets through the instance's connector instead (ADR 0006): `halkyon_learn_server/lib/src/assistant/` serves MCP (`/mcp`, Streamable HTTP, JSON answers, no sessions; five read-only tools in `connector_tools.dart`) and its OAuth (`/.well-known/oauth-*`, `/oauth/register`, `/oauth/authorize`, `/oauth/token`) on Serverpod's web server, which production runs on 8082 for these paths alone (Caddy forwards them; `connectorOf` takes the public address from `webServer`). People connect with a code from Settings › Assistant › Connect Claude (`AssistantEndpoint.connectCode`, eight Crockford characters, ten minutes, once), typed on the authorization page; codes and tokens are kept as SHA-256 hashes. The handlers take plain requests (`ConnectorRequest`), so `test/integration/assistant/` drives the whole flow without a web server; the Relic routes only translate.
- **Where sync stands** is `SyncService.status` (syncing, the last sync that went through, the last failure; since the app started), which `SyncStatusBloc` (`core/sync/`, provided by `PersonApp`) hands to the trouble dot on the Settings cogs (on Home and Profile; no tab carries it), the Sync card in Settings and the banner on Home. Only the last outcome counts: one sync that goes through clears a failure. It also counts the changes not on the server yet (`pending`): every save asks for a sync (`requestSync`), and a sync that goes through takes those saved before it started; the engine can't say, since the server only tells it at the start of a sync what it has. The count is kept in `sync_pending.json` in the person's folder (`PendingChangesStore`) and shown on the Sync card. The Library's Refresh and pulling it down sync now (`syncAndWait`, which a pull's spinner waits on).
- **Serverpod's version moves together with `serverpod_offline_sync`**, which allows only a narrow range of Serverpod (v0.0.10: `>=4.0.4 <4.1.0`; v0.0.8 pinned v4.0.0 exactly, which the root `pubspec.yaml` overrode until 2026-10-09). Every package pins Serverpod's packages to one version: upgrade them all together with the sync packages and the CLI version CI generates code with (`SERVERPOD_VERSION` in `.forgejo/workflows/ci.yaml`), after reading the sync package's changelog. Never override a version the sync package doesn't allow without checking what changed in `serverpod_database`, which it builds on.
- **The source-code link**: both licences require offering the source of what an instance runs to the people who use it. `instance.info().sourceCodeUrl` names it (the operator's `SOURCE_URL`, else this project's repository) and the app's About dialog shows it next to the licences and every package's notice. Keep it working when changing the dialog or the endpoint.
- **Public profiles** (ADR 0005, stage 1 of the public tier): `public_profile` (`serverOnly`, never sent; the app gets a `Profile`, without the person's id) holds a handle (Mastodon's rules: 1 to 30 letters, digits or underscores, unique whatever their case through `handleKey`, a few reserved for the instance; changeable until the first publish, `firstPublishedAt`), a display name, a bio, up to 4 links and an avatar. `ProfileEndpoint` (`identified` only: a profile is for others to see) reads, saves and removes the caller's own; `ProfilesEndpoint` finds anyone's by handle. Handles are on the account domain (`ACCOUNT_DOMAIN`, else the server's public host; `instance.info().accountDomain`), which can never change once anyone follows a handle on it. An avatar is a PNG of at most 1 MB that the app draws anew from the picture chosen (`avatarPng`, so a photo's metadata never leaves the device), kept in the public storage at `avatars/<SHA-256>.png`, named by its content so its URL tells nothing about whose it is, and deleted once no profile shows it. In the app, the Profile tab holds it as a form, for people with an identity.
- **Publishing** (ADR 0005, stage 2 of the public tier): `PublishEndpoint` (`identified` only, and a public profile: a release names its author) publishes a studyset with everything inside it as its next release. The server reads the person's synced rows (`readReleaseFiles`, `lib/src/publish/`) and writes them through the `git` command (`ReleaseRepository`, never a library linked in) as a commit on `main` and the annotated tag `v<N>` in a bare repository of the studyset's own, `<REPOSITORIES_DIR>/<public id>.git` (`published_studyset`, `serverOnly`; the public id is random, made at the first publish, never a synced row's id). A release holds a directory per studyset (the published one at the top, each inside it named by its id) with `studyset.json`, `cards/<id>.json`, `notes/<id>.json`, `paths/<id>/path.json` with `steps/<id>.json`, `pdfs/<id>.json`, and the files the content names in `files/<sha256>`, described in `files.json`; each item file holds its id, position, tag names and payload as synced, keys sorted and indented, without what is only its author's (a card's `report`, a PDF's `deadline`), and items being deleted (blank payloads) are left out. Publishing is refused (`PublishRefusedException`) without a profile, unless every studyset of it is under a licence allowing it to be passed on (`release_license.dart`; "no derivatives" allowed, and marked in the app), with a card, note or step of more than one current revision, with files not uploaded yet, when nothing changed, or for a name (`slug`: lower-case letters, digits and hyphens, one per studyset of a person, fixed from the first publish) that won't do. Versions count past both the database's highest and the tags', so a version git has and the database lost is never given again, and a release the database can't record is taken back from git. The first publish fixes the person's handle. Git runs apart from the machine's configuration (no system or user configuration, no prompts) with `gc.auto` off, so objects are never removed. Deleting a person deletes their published studysets with their repositories. In the app, Share in a studyset's action row (a lock, or a globe once published; an icon below medium windows) opens the Publish sheet (`feature/publish/`): its name, whether its licence allows publishing, what changed, its releases. Publishing syncs first, since the server builds the release from what it holds; on a refusal for revisions to settle, the app saves merged those whose changes merge (ADR 0003) and tries once more, and lists those changed differently on one side, each opening where it is settled. An author can take a studyset off Explore again (`withdraw`, `withdrawnAt`): nobody can add it, copies get no updates, its releases stay; publishing it again, changed or not, brings it back. The library's cards say who made a studyset ("By you", or a copy's author by the handle in its `publishedAt`) and whether it is out on Explore (a lock or a globe, opening the Publish sheet through `showShareSheet`): `published()` lists the person's studysets published and not withdrawn, and `PublishRepository` remembers them in `published.json` in the person's folder (`PublishedStore`), so the library shows them offline too; publishing, withdrawing and looking one up update it.
- **Explore and copies** (ADR 0005, stage 3, decided 2026-10-09): `ExploreEndpoint` (`requireLogin`: anonymous people explore and learn too) lists the latest release of each published studyset not withdrawn **by fit, never by popularity** (the person's roadmap): search over titles and authors (LIKE's wildcards in what is typed are only themselves), languages (a primary one takes its regional ones), kinds of content held, the newest release first, 30 to a page; the next page only when asked for (Show more, no endless scrolling). What Explore shows of a release is kept beside its repository (`studyset_release`'s title, description, languages, credits), so listing never reads git. `release` gives out the latest release's JSON (one `git cat-file --batch` run) and `file` each file by its hash. **The app makes the copy** (`ReleaseCopier`, `feature/explore/`), as an import: through the person's own repositories, the files among their own files (asset ids in the content swapped for theirs), the tags by name among their own tags, the studysets in their tree, each list in its order; everything is read first, so a release with an item in a schema this app doesn't know is refused whole (`ReleaseUnreadable`). The server never writes into a learner's space. Every item of a copy names where it came from (`published`: publication, item, release; `core/content/published_ref.dart`; card schema 9, note 3, step 3, path 2, PDF 2, studyset 5, the top studyset also with `location`), merged per side like the rest of a card, note or step, so a later release can be taken from where the copy last took one. **A later release is offered, never applied by itself** (`CopyUpdater`, `CopyUpdateBanner` on the copy's top studyset: "v4 is out: 12 changed, 5 new, 2 withdrawn", Update): the app compares it with the release the copy last took (`release(version:)`; releases never change, so the repository caches them) item by item. Cards, notes and steps take the release's content as a revision on the one where they last took a release (`takeRelease`, as `takeDictionaryUpdate`), merging per side: what only the release changed comes in, what the learner changed stays, a side both changed is a conflict settled in the editor. Studysets, paths and PDFs, kept whole, take each field of their payload the learner left as the old release had it; one they changed stays theirs (personal fields, such as a PDF's deadline, too). New items come in where they belong, unless the learner deleted that part; tags the release added come and those it took away go, the learner's own stay; files are fetched only when the person hasn't the same file already (`AssetRepository.idOf`). What the author withdrew stays the learner's, studied on, marked (`withdrawn` in its origin; the card browser's filter Withdrawn by the author); what the learner deleted stays deleted; progress is never touched. A copy of a studyset its author withdrew is offered nothing. **"Learn this here"** (ADR 0005, rule 8): `LearnerCountFutureCall`, daily, counts per published studyset the people of this instance with the `identified` scope who chose to be counted (`countLearning` in their settings, settings schema 8, off unless switched on in Settings › Privacy) and studied a copy of it (a studyset whose payload names it) on 3 or more days of the last 30, fresh starts left out; payloads are decoded in Dart, so one that isn't JSON counts for nothing rather than failing the count. Only `learnersShown` (rounded down: 5, 10, 20, 50, …; none below 5) ever leaves the server, in Explore and the author's Publish sheet. **A copy is published within its licence** (rule 10): the server looks up the original's (`PublishRefusal.forkForbidden` under "no derivatives"; `licenseMustStay` under a licence that keeps adaptations under it, `keepsAdaptations`, in every studyset of the copy), and the release keeps what it is based on (`basedOn`, shown in Explore); a person publishing a copy of their own is its author, and the copy of another instance's studyset goes with its location alone. The server never updates synced rows itself: the sync engine (v0.0.10) stores a ByteData the server updates through it as its JSON text, not its bytes (`_updateRowWithoutRecording`), unreported upstream so far. In the app, Explore is the old app's feed: cards with title, author, the start of the description, the latest release, what it holds, the licence and Add to library (its progress, then In your library, opening the copy); a summary sheet with all of it.
- **Archives** (decided 2026-10-11; `feature/archive/`): a studyset with everything in it and its progress, as one `.zip`, imported from the Library's Import as the person's own (backups, moves between instances, a library brought from elsewhere). `archive.json` says the format (`{"format": 1}`; a newer one is refused whole); the studyset is laid out as a release (studyset.json, cards/, notes/, paths/, pdfs/, a directory per studyset inside, files.json and files/<sha256>), so the release copier writes it (`ReleaseCopier.write`: unlinked, no `published`, its cards marked `archive` as imported ones are, returning each item's new id). A card may name a dictionary entry in place of its payload (`"entry": {"source", "language", "id"}`): the import makes it from the instance's dictionary as adding from it does (`DictionaryEndpoint.entries`, at most 1,000 ids at once), with the tags that brings; an entry the instance lacks leaves the card out, counted. `reviews.json` lists reviews of its cards by their ids in it (`card`, `rating` 1–4, `reviewedAt`, optional `durationMs` and `mode`), appended to the person's review log once the cards are in, so schedules go on from them. Reading runs on an isolate of its own (`ArchiveReader`; tests read inline). There is no export into it yet.
- **Reports and suggestions** (ADR 0005, stage 4, rule 11): `feedback_thread` (a report or suggestion: `publicationId` and `kind` columns, everything else in its payload: what it is about by release item id and version, a report's category, the text, a suggestion's proposed payload and the files it names), `feedback_message` (replies) and `feedback_decision` (an author's closed, reopened, hidden, shown, accepted, declined; the latest of each pair stands) are synced tables each person writes in their own space; a thread or reply names its thread by id only, as it may be another person's. `gatherFeedback` (`lib/src/feedback/`) assembles them for `FeedbackEndpoint.threadsOf` (`requireLogin`: public) and `InboxEndpoint` (`identified`: the inbox, unseen counts, `markSeen`, `block`, `file`) and decides what counts as it reads (see the ADR), since a modified app could write anything. `feedback_block` and `feedback_seen` are server-only; `deletePerson` deletes both. A release's commit credits the writers of suggestions accepted since the last (`Co-authored-by`). In the app (`feature/feedback/`): the card's Report sheet offers "Send to the author too" on a card from a release; a copy's ⋮ menu reports the studyset and lists its threads, as Explore's summary does; the editors of cards (⋮), notes and steps (app bar) suggest to the author (refused under "no derivatives"); the Inbox is in the Profile tab, with a count on the tab (`InboxCountBloc`, refreshed at start and whenever the inbox is read or acted on; no timer); a thread's sheet shows a suggestion beside the author's own item, its replies, and the author's decisions. Accepting writes the proposed content as the author's own revision (`FeedbackRepository.accept`), the writer's files fetched by hash.
- **Deleting a person** goes through `deletePerson` (`lib/src/auth/account.dart`), in one transaction and in ADR 0001 rule 6's order: their sessions end first (a device with a valid session would re-upload everything), then their space goes with all its content through the foreign keys, then the person with their identities. `AccountEndpoint.delete` deletes the caller; the app also calls it when an anonymous account is given up (signed out of, or switched away from to GitHub), since nobody could ever sign into it again.
- **Pruning**: `PruningFutureCall` deletes anonymous people unused for 90 days, daily, scheduled when the server starts; and hourly the notes Serverpod's rate limit makes of anonymous sign-ups (`serverpod_auth_idp_rate_limited_request_attempt`, each with the IP address it came from) once the limit's hour is past (`forgetOldSignUpAttempts`): Serverpod deletes them only for email sign-in, so they would otherwise stay for good, and the privacy page says the instance keeps no IP addresses beyond that. Scheduling a future call always adds an entry, so the server cancels the previous one by its identifier first.
- **One person's data per device**: `PersonFolders` wipes every other person's folder when a person's data opens: their sessions ended on this device. Opening waits for the previous person's database to close.
- **Session logs are not persisted** (`persistentEnabled: false` in every run mode) for privacy. Don't re-enable them.
- **Secrets** live in `halkyon_learn_server/config/passwords.yaml` (git-ignored). Never change `serverSideSessionKeyHashPepper`: it would sign everyone out, and anonymous accounts can't sign back in.
- **Checks on a device** live in `halkyon_learn_flutter/integration_test/` and run by hand, never in `tool/check.sh` or CI: `flutter test integration_test/<file> -d <device>`. `pdf_engine_test.dart` checks the real PDF engine (on the Linux desktop too), `youtube_player_test.dart` the embedded YouTube player (Android). Never on the emulator from an agent session: it takes the editor down with it.
- **Windows builds need Developer Mode**: building makes symbolic links (Flutter links each plugin's native code into the build, and pdfrx's build does too), which Windows lets ordinary users make only in Developer Mode. PDFium itself comes with the build as a Dart native asset.
- **Linux builds need** `libwebkit2gtk-4.1-dev`, `libsecret-1-dev`, and for sounds (`audioplayers`) `libgstreamer1.0-dev` and `libgstreamer-plugins-base1.0-dev`. If a Linux build tries to install to `/usr/local`, the CMake cache is stale: delete `build/linux`.