FlClash/.agents/architecture.md

# Architecture

## Core Integration

The Go proxy core in `core/` operates in two modes.

Android lib mode:

- `gomobile bind` builds `core/mobile` into an AAR: a native library per ABI, shipped as `libclash.so`, and the Java
  classes gobind generates in `com.follow.clash.core.mobile`. The `:core` Gradle module builds it and takes in its
  classes and libraries (see Android Core Build).
- The Android `:core` module owns the binding: `Core.kt` is the only caller of the generated `Mobile` class. Flutter
  crosses the `${packageName}/service` MethodChannel through `lib/plugins/service.dart` and Android's `ServicePlugin`
  rather than calling the Core directly.
- `lib/core/lib.dart` (`CoreLib`) implements the shared Core interface, gates method calls on its connection completer,
  initializes and synchronizes Android shared state, and closes the native service path exactly once.
- Because Core is in the application process on Android, application RSS already includes Core memory.

Desktop core mode:

- Go core runs as a separate process with `CGO_ENABLED=0`.
- `rust_api`'s `ipc` module provides the native local-IPC primitives: a Unix domain socket on macOS/Linux and a named pipe
  on Windows, served by an `IpcServer` handle. Dart owns the transport state, RPC correlation, process ownership, and
  lifecycle convergence above those primitives.
- `lib/core/service.dart` (`CoreService`) is the composition root. It wires the IPC transport, launcher selection,
  lifecycle controller, RPC client, and crash-event bridge, and holds no lifecycle logic of its own.
- `lib/core/desktop/transport.dart` binds an `IpcServer` and turns its events into ready, connected, disconnected, and
  failed transport events plus a frame stream. A connection that ends on an I/O error is a disconnect carrying that
  error; failed means the server itself is gone and is terminal for that transport. A replaceable binding keeps RPC
  subscriptions stable when a failed or stale transport must be rebuilt.
- `lib/core/desktop/rpc_client.dart` owns request IDs and pending completers, waits up to 10 seconds for a connection,
  applies a three-minute default method timeout, unwraps `CoreMethodResponse`, and fails all pending calls when transport
  disconnects or closes.
- `lib/core/desktop/lifecycle.dart` serializes process intents and owns the authoritative desktop state machine.
- `lib/core/desktop/launcher.dart` abstracts direct child-process and privileged Helper ownership through idempotent process
  leases. `lib/core/desktop/helper_client.dart` is the typed local HTTP client for the privileged Helper.

`lib/core/controller.dart` (`CoreController`) selects the implementation based on platform. `lib/core/interface.dart` defines the shared `CoreHandlerInterface`.

Key Go core files:

- `core/protocol.go`: method names, event types, and the parameter and result shapes shared with Dart.
- `core/method.go`: MethodChannel-style method-call dispatch, the handler table, and response envelopes.
- `core/lifecycle.go`, `config.go`, `proxy.go`, `delay.go`, `provider.go`, `geo.go`, `connection.go`, `log.go`: the
  handler functions, one file per area. `core/hooks.go` wires mihomo's hooks to them.
- `core/message.go`: non-blocking state/priority/bulk event queues and bounded message batching.
- `core/lib.go`: the Android library API: TUN callbacks, the event listener, and the entry points `core/mobile` binds.
- `core/mobile/`: the package gomobile binds. gobind generates Java for its exported API, so it speaks only `int32`,
  `bool`, `string`, `[]byte` and its own interfaces; a Go `int` would become a Java `long`.
- `core/cmd/core/`: the desktop entry point. `core/` itself is the library package `core`, because gomobile cannot bind
  a `main` package.
- `core/server.go`: desktop socket/named-pipe client and framed message forwarding.

## DNS, NTP and Sniffer Overrides

A URL or file profile brings its own sections; a custom profile brings none. The app's DNS page
(`PatchClashConfig.dns` with `dnsOverrideKeys`) is written over the DNS of every URL and file profile, and offers only
`DnsOverrideKey.normalProfileKeys`: keys that name nothing of a profile's own, because a subscription's groups and rule
sets change with every update. A custom profile keeps all three sections in `Profile.overrides`, a `ProfileOverrides`
stored as JSON in the `profiles.overrides` column, with every key available, and NTP and sniffer exist nowhere else.
`setupState` resolves `SetupState.overrides` by profile type, intersecting the app-wide keys with the normal set, and
`_makeRealProfileTask` merges whatever it is handed; there is no on/off toggle, a picked key is the override.

Each `*OverrideKey` names one Clash key, with `parent.field` paths for nested maps, and covers every yaml key of the
matching mihomo raw struct; `test/models/{dns,ntp,sniffer}_override_test.dart` read `core/Clash.Meta/config/config.go`
and fail when a core bump adds or drops one. `overrideJson` builds the fragment in model order and skips a key whose
value is empty, so a key left without a value writes nothing: the core refuses an empty `nameserver` outright.
`mergeDnsOverride` merges one level deep so a partial fallback filter keeps the profile's other filter keys; NTP spreads
over the profile's flat `ntp` map. For a profile whose `dns.enable` is not true, the task first fills `baselineDns` for
`baselineDnsOverrideKeys` (`enable`, `enhanced-mode: fake-ip`, `nameserver`), the one thing the core is never without:
with DNS off, mihomo answers every TUN-hijacked query with an error. The baseline stays apart from the model defaults,
which carry no preset of the app.

An added key starts from the spec's `defaults` through its `OverrideField.reset`, and those defaults are mihomo's: a key
starts at the value mihomo parses without it, presets included, such as the nameservers, fake-IP range and filter, and
NTP server of `DefaultRawConfig`. A key mihomo leaves unset, such as `listen` or a sniff protocol's ports, starts empty
and writes nothing until the user fills it in, and so does the fallback filter's GeoIP code, whose value is left for the
user to choose. mihomo sniffs a protocol listed without ports on that protocol's standard port, which the page shows as
the default. The tests above pin the defaults against `DefaultRawConfig`. A spec that mixes in `OverrideQuickEdit` gets
quick edit, which opens `overrideYaml` in an `EditorPage` and rebuilds the key set from the document through
`applyOverrideYaml`, rejecting keys outside the section and values the model cannot hold.

mihomo upper-cases sniff protocol names and ignores the deprecated `sniffing` list once `sniff` holds anything, so
`mergeSnifferOverride` drops the profile's entry for an overridden protocol in any case and first carries `sniffing`
into `sniff`. A `rule-set:` key of a custom profile's DNS policy counts as a use of that rule provider, so
`_resolveInjectedProviders` injects it like one a rule names. Safe mode forces `write-to-system` off on whatever `ntp`
section reaches the core, because setting the system clock is a host change; `loadConfig` in the Go wrapper also drops
it on Android, where the seccomp policy kills the process for `settimeofday`.

A custom profile's TUN page is the exception to covering the raw struct. The network settings own `tun` through
`PatchClashConfig.tun`, spread last over the profile's section and patched again on every `updateConfig`, so
`TunOverrideKey` offers only a few keys outside it that work in FlClash's TUN: `disable-icmp-forwarding`, and
`exclude-interface`, which only Linux routing reads. `test/models/tun_override_test.dart` fails if a key comes to
overlap `Tun`. Android starts its VPN from `core/tun` without the section, so TUN is listed on desktop only.
Being a part of the section, the page has no quick edit; see Quick Edit.

A custom profile's overview keeps tiles for rules, providers and DNS, which every custom profile runs on, and puts the
dialer proxies, NTP, sniffer and TUN behind a single "more" tile, which opens a page listing them with their states.
Another rarely changed section joins that list rather than taking a tile of its own.

Every page is an `OverrideView` over an `OverrideSpec` from `lib/views/config/override.dart`. The spec supplies the
section's keys, labels and fields; its `OverrideTarget` is where the model and keys live, `PatchOverrideTarget` for the
app-wide DNS and `ProfileOverrideTarget` for a profile. Another overridable section is a spec, not a copy of a page.

A config from a release that still had the `overrideDns` toggle keeps, through `Config.fromJson`, the normal keys of
an on toggle and none of an off one. That release had no custom profiles, so its app-wide NTP override is dropped.

## Listener Exposure

`_makeRealProfileTask` in `lib/common/task.dart` writes `mixed-port`, `allow-lan`, and `external-controller` into every
generated profile. With `allow-lan` off, mihomo's `genAddr` binds each listener to `127.0.0.1`; the external controller is
`ExternalControllerStatus.close` by default and never binds anywhere but loopback. The task also clears
`external-controller-tls`, `-unix` and `-pipe` in every mode: the unix and pipe controllers answer without a secret, and
the unix one unlinks whatever file sits at its path before opening it to everyone, so a profile may not choose them.
Outside safe mode, profiles retain `listeners`, `tunnels`, `ss-config`, `vmess-config`, `tuic-server`, `iptables`,
`dns.listen` and `ntp.write-to-system`. Extra inbounds can bypass the app's LAN and authentication settings, and a TUN
listener can change routing independently of the app's TUN toggle. These network and clock effects are an accepted
boundary; they are not treated as arbitrary filesystem modification. The task disables `iptables` when its
`inbound-interface` contains a space, because mihomo interpolates it into commands and splits their arguments on spaces.
Safe mode clears these extra inbounds and iptables, closes DNS listening and disables NTP system time writes even after
local overrides. The editor omits disabled alternate controllers and keeps listener completion.

An open external controller always has a secret, because mihomo serves a controller with an empty one to every local
user and, through CORS, to any page the user visits. `PatchClashConfig.secret` is that secret and it belongs to FlClash,
as the address does: the task writes it over whatever the profile carries, and `UpdateParams.secret` hands the same value
to a running Core, which restarts the controller when it changes. `ensureControllerSecret` keeps the pair consistent
(an open controller gets a generated secret, a closed one drops it, so every opening rotates it) and runs where a
config enters the app: the settings toggle, startup, and `writeConfig` for a restored backup. `_effectivePatchConfig`
reports a controller without a secret as closed, and `_makeRealProfileTask` independently closes it without a secret
or in safe mode, so a writer that skips `ensureControllerSecret` opens nothing. The settings page shows the secret in
an editable row, which refuses an empty value.

The loopback listener takes local connections without authentication by default: mandatory credentials would lock out
every external local consumer a proxy client exists for. The binding only guarantees traffic originates on the device;
it does not isolate programs already running there, so on Android any app holding `INTERNET` can use the proxy and learn
the outbound IP (#1934). The opt-in answer is the local authentication setting (`NetworkProps.authentication`), which spans
four paths that must stay in sync: `_makeRealProfileTask` writes the credentials into `authentication` and force-clears
`skip-auth-prefixes` so a profile cannot silently exempt loopback; `UpdateParams.authentication` applies the same list
to a running core; `FlClashHttpOverrides` sends the credentials with the app's own proxied requests; and `sharedState`
withholds the Android VPN system proxy declaration because `ProxyInfo` cannot carry credentials (TUN still captures that
traffic). Desktop system proxy is deliberately not gated — dropping it would leak traffic direct, so the setting's
description tells users to supply credentials manually instead.

### Local Trust Boundary

The Windows Helper answers only a connection whose client process runs under the owner SID it was installed for;
the Linux Helper only the owner uid. A setuid Core is executable by `root:admin` (macOS) or the installing user's primary
group, so its boundary also includes every member of that group, not just the installing account. Profiles cannot choose
a controller address, a secret or alternate controllers. Another account outside those permissions, a subscription,
or a remote party reaching the privileged Core's control or filesystem APIs is a vulnerability.

A process running as the same account is not stopped. It can ask the Helper to start the Core
against a socket or pipe it owns, or run a setuid Core against a port it listens on, and so become the host that drives
a privileged Core: `initClash` chooses the home directory every path check anchors on, and `getConfig` reads any path.
The Core connects to whatever address it is given and cannot tell that host from FlClash. An account check cannot
distinguish it from a same-account process that can already read FlClash's memory and files. This remains a user-to-root
or user-to-SYSTEM privilege escalation, not a complete security fix. An open external
controller is inside the same boundary: its secret is stored under the account, and `/upgrade` stays mounted (it fetches
fixed MetaCubeX release URLs, not caller-chosen code). mihomo's `IsSafePath` compares paths lexically, so a symlink inside
the home directory escapes it; that fix belongs upstream, not in the submodule.

Closing the same-account case requires limiting what a hijacked Core can do. Keeping the main Core unprivileged and
moving host changes into narrow platform helpers requires architectural work, but need not introduce repeated prompts
or remove ordinary proxy functionality. Features that require additional privileges, such as identifying other users'
processes, need platform-specific handling. Descriptor-based filesystem confinement can preserve ordinary managed
paths; intentional external symlink targets need explicit allowed roots. These changes have not been implemented.
Account authentication, a private socket directory or a symlink check at configuration time does not remove the
privileged filesystem operations available to the host. Changes must not widen access to those operations.

### Interface Name Modes

The same task also writes `interface-name` per `PatchClashConfig.interfaceNameMode`: `clear` (default) forces it empty so a
subscription's value cannot survive into the generated profile, `follow` leaves whatever the profile YAML already has, and
`custom` writes the user's `interfaceName`. mihomo ignores `interface-name` on Android whenever `dialer.DefaultSocketHook`
is installed (`core/lib.go` `installHooks`, vendored `dialer.go`), so `NetworkListView` in `lib/views/config/network.dart`
shows the mode picker and text field only on desktop.

### Network Settings

Every `Tun` field is written over the profile's `tun` section and hot-applied through `tunSchema`, and the other keys of
that section keep the profile's value; `test/models/tun_test.dart` keeps the model, `RawTun` and `tunSchema` in step.
Beyond what FlClash always managed, `Tun` carries only the keys worth tuning: `mtu`, the mips stack's
`congestion-controller`, and `strict-route` where sing-tun enforces it (Linux and Windows).

Routing is the same on every platform: `route-address` (empty routes everything), `route-exclude-address`, and
`NetworkProps.bypassPrivateRoute`, which adds `privateRouteAddress` to the exclusions in `Tun.getRealTun`. Desktop hands
both lists to sing-tun. VpnService can only add routes before Android 13, so `Tun.vpnRouteAddress` cuts the exclusions
out of the routes with `subtractCidrs` and lists every family it routes; `VpnService.addRoutes` routes everything only
for an empty list. A config saved with the old `routeMode` keeps its choice: `bypassPrivate`, which ignored the route
list, loads with the flag on and the list cleared, and `config` with the flag off. Others were left out on purpose: no stack in the vendored sing-tun reads `endpoint-independent-nat`, the
default mips stack ignores `udp-timeout`, and the rest are platform plumbing few users need. Android builds its TUN from
the VpnService descriptor in `core/tun/tun.go`, so only MTU and the congestion controller reach it, through `VpnOptions`
and the `options` argument of `startTUN`; the VpnService builder takes the same MTU.

The TUN fields start at mihomo's own defaults, and a saved zero MTU or empty congestion controller, which mihomo reads
as those defaults, decodes to them. The outbound rows are plain settings written over every profile, and changing one
re-applies the profile; keep-alive is forced off on Android, and only Linux honours `routing-mark`.

## Lifecycle Ownership And Convergence

### Shared Flutter Layer

`CoreController.start()`, `restart()`, `stop()`, and `close()` are the only shared lifecycle facade. `close()` is terminal;
callers must not try to reuse a closed platform implementation.

`CoreAction` in `lib/providers/actions/core.dart` owns the user-facing Core status and setup sequence:

- `startCore()` publishes `connecting`, starts the platform Core, publishes `connected`, then initializes Core state. A
  startup error publishes `disconnected` and displays the error.
- `restartCore()` coalesces overlapping callers behind one worker. `_requestedRestartRevision` records newer requests.
  After the lifecycle restart, the worker reapplies running state or the profile, rereading `isStartProvider` on every
  pass, until it has consumed the latest revision, so the newest start intent wins.
- The provider is an orchestration and presentation layer, not a process owner. Platform lifecycle code remains responsible
  for determining whether a Core process/service is actually running.

Application exit is centralized in `SystemAction` and `SystemExitCoordinator`:

1. Optionally save config and clean up DNS, system proxy, and tray resources in parallel.
2. Close the desktop window.
3. Call terminal `CoreController.close()`.
4. Exit the application exactly once.

The coordinator is idempotent, continues later cleanup steps after an earlier error, preserves the first error for the
caller, and uses a three-second watchdog as an emergency application-exit path. `Application.dispose()` and
`CoreManager.onCrash()` do not independently destroy Core; this avoids competing shutdown owners.

### Applying Setting Changes

A changed setting is patched live through `UpdateParams`, re-applied with the profile, or, for the Android VPN, left
to a restart the user confirms. [setting-changes.md](setting-changes.md) says which path each setting takes and how to
choose one for a new setting.

### Desktop Lifecycle

`DesktopCoreLifecycle` is a latest-desired-intent reconciler, not a queue that blindly executes every request:

- Public intents receive monotonically increasing revisions and target running, restarted, stopped, or closed.
- Observable phases are `idle`, `starting`, `running`, `stopping`, `failed`, and `closed`.
- A completed command reports `applied`, `coalesced`, or `superseded`, allowing callers and tests to distinguish a command
  that won from one satisfied or replaced by a newer compatible intent.
- Startup opens or replaces the IPC transport, resolves a launcher, generates a 128-bit lowercase hexadecimal session ID,
  launches Core, and waits for the matching connection. Windows additionally verifies that the named-pipe peer PID equals
  the process PID returned by the Helper lease.
- Each running session retains its process owner, lease, PID, session ID, and transport connection generation. Stop waits
  for both process-exit confirmation and the matching disconnect generation; a missing disconnect replaces the transport
  before later starts.
- An unconfirmed process exit is retained as an unconfirmed lease. New start/restart intents fail until ownership can be
  cleaned up, preventing two Core instances from being treated as the active session. Terminal close may continue on a
  best-effort basis because the application is exiting.
- An unexpected disconnect or transport failure while running is converted to `DesktopCoreFailure`, the owned process is
  cleaned up, and `CoreService` emits a Core crash event for the normal UI recovery path.

Direct launch is used on macOS, inside an AppImage, and as the Windows/Linux fallback when the privileged Helper is not
ready. When the Helper is ready, it owns the Core child and Dart owns it through a session-scoped lease.

### Android Service Lifecycle

Android deliberately keeps Flutter requests optimistic and the native layer authoritative:

- `ServicePlugin.start()` and `stop()` acknowledge immediately after submitting intent. They do not wait for service
  creation, VPN permission, binding, TUN establishment, or teardown.
- `ServiceState` owns the latest `RunRequest` and the shared configuration. Identity checks discard obsolete work.
  `startPreparationLock` serializes permission/setup preparation and `transitionLock` serializes actual service
  transitions.
- `RunState` is derived, never assigned by a step: `ServiceStateMachine.publish` computes it from the transition in
  flight, the latest intent, and `ServiceController.runtime`. A service that is up under a stop intent reads `STOPPING`, a
  start still waiting on a permission reads `STOPPED`, and `STARTED` needs both a running intent and a runtime. The tile
  and `getRunTime` both read this one value, so they cannot disagree with the service.
- `ServiceController` owns exactly one `ManagedServiceBinding`, selects `VpnService` or `ProxyService` from `VpnOptions`,
  binds with a five-second connection timeout, and invokes `ManagedService.start()`/`stop()` off the main thread. The
  binding and `runtime` (which service runs, and since when) are set and cleared together, by a stop, a failed start, or
  a lost connection and by nothing else. Flutter's `shutdown` only clears the Core event listener: Android keeps an
  established `VpnService` alive through its own connection, so a binding dropped on exit left a tunnel that was
  reported as running but that neither a stop nor a revoke could reach.
- The Core listens exactly while `runtime` is set: a successful start ends in `Core.startListener()` and every path
  that clears `runtime` calls `Core.stopListener()`. Both return at once when the Core is already there, which is the
  usual case, since `startTUN`/`stopTun` and Flutter's own `startListener`/`stopListener` got there first. They matter
  when nothing else does: a proxy-only run stopped from the tile used to keep the mixed port open with nothing
  reported as running. A start closes after itself the same way: Flutter's `startListener` and the native `quickSetup`
  both open the ports before any service exists, so a start that ends without a `runtime`, refused or cancelled, stops
  the empty service once more, without passing through `STOPPING`. Flutter may be gone by then and never see the
  `stopped` report.
- A start that does not bring the service up gives its intent back and takes down whatever it left running, unless a
  newer request has claimed the service. A failed start therefore always ends in `STOPPED`, whichever step refused it.
  The one start that mints no request is a repeat on a settled run of the same kind, which Flutter sends every time it
  reattaches: a request refused at a permission prompt there would stop a service that was running fine.
- Generic service creation/destruction is lifecycle evidence, not user intent. New commands must flow through
  `ServiceState.requestStart()`/`requestStop()` or the explicit system-action handlers instead of inferring intent from a
  callback.

Quick Settings, notification, revoke, and Always-on VPN paths converge on the same owner:

- With a Flutter engine attached, `ServiceState.handleStartAction()` forwards through `TilePlugin` to `TileManager`,
  because only Flutter can apply the current profile. Without Flutter, native code claims the intent, restores
  `SharedState` from preferences, runs `quickSetup`, checks VPN permission, and runs the start; claiming first is what
  makes a second action during the setup toggle that start off instead of racing a second setup against it.
- A stop never needs Flutter. `handleStopAction()` stops the service natively whether or not an engine is attached, so
  the tile, the notification action and a revoke work even when Flutter's own idea of the state is wrong.
- Whenever the run ends without Flutter having asked for it (a start that failed, a native stop, a revoke, a lost
  service), `ServicePlugin` sends `stopped` with the revision of the last `start`/`stop` command it received.
  `Service` in `lib/plugins/service.dart` drops a report older than its newest command, or one that arrives while its
  last command was a stop (on-demand suspension stops the service while `isStartProvider` stays true), and
  `AndroidManager` answers the rest with `setRunning(false)`. The report is read on the platform thread, where the
  commands arrive, so the revision and the intent it carries describe the same instant.
- Android may create an Always-on `VpnService` through `onStartCommand()` without FlClash's bound-service path. The service
  sends the explicit, permission-protected `VPN_START_REQUESTED` broadcast to `ServiceBroadcastReceiver`, which routes it to
  `ServiceState.handleStartAction()` so Core/configuration and the normal binding are restored before TUN is treated as
  ready.
- `VpnService.onRevoke()` stops TUN/modules first, then sends `VPN_REVOKED`; the receiver only requests a stop while
  `ServiceController.runtime` is a VPN run.
- `ServiceBroadcastReceiver` uses `goAsync()` and an atomic one-shot completion. Normal completion or a nine-second
  watchdog calls `PendingResult.finish()` exactly once; the watchdog releases Android's broadcast lease and does not
  cancel or redefine the underlying lifecycle intent.

## Core Protocol And Event Delivery

The shared protocol uses `CoreMethodCall(id, method, arguments)` and `CoreMethodResponse(id, result, error)` in both
directions. The envelope is the only JSON serialization layer: keep arguments, results, and event data as structured JSON
values rather than embedding pre-encoded JSON strings. Plain domain strings, such as country codes or provider contents,
remain strings.

Go event delivery is intentionally non-blocking:

- Loaded-provider, geo-update, and route-changed events use a state queue that never evicts; the rules for it are under
  Core API Safety in `.agents/rules.md`. Every other event except the bulk kinds, delay included, uses a 256-entry
  priority queue. Desktop process crashes are generated locally by `CoreService` from lifecycle failures rather than
  sent through the Go queue.
- High-volume log, request/connection, and DNS query events use a separate 256-entry bulk queue, so bulk floods cannot
  evict state.
- A full priority or bulk queue evicts only its own oldest event and retries the newest; DNS events never evict. Core
  work never blocks on event delivery.
- The batcher flushes at 32 messages or every 16 milliseconds. Priority events are preferred, but one bulk opportunity is
  guaranteed after eight priority messages to prevent starvation.
- `sendMessage` drops an event before it is queued when `hasEventListener` is false. That only happens on Android, after
  the Flutter engine detaches while the service keeps the Core running; request and DNS events would otherwise be
  encoded per connection and per query for nobody.

DNS query events come from the Clash.Meta fork's `dns/patch.go` and are raised where mihomo writes its `[DNS]` debug
log, so the DNS page lists the same resolutions: one event per cache hit, and one per upstream exchange however many
callers singleflight merged into it, carrying the upstream whose answer was used. Fake-IP and hosts answers never reach
a resolver and are not recorded. Queries from apps arrive with the `DNSContext` that `Service.ServeMsg` created and
report `app`; other initiators come from `resolver.WithInitiator` at the call site (rule matching, DIRECT, proxy
server dialing), and unlabeled queries report `other`.

Dialer-loop events come from the fork's `component/proxydialer/patch.go`. mihomo follows a `dialer-proxy` that leads
back to its own proxy until the stack overflows, so a dial that asks the same `dialer-proxy` for an address it is
already dialing through it fails with `ErrDialerLoop`. `core/hooks.go` reports each looping name at most once every
three seconds as a priority event, which the app shows as an error notice whether or not logs are captured. The custom
profile pages flag only the loops FlClash can follow on its own, through a group's own list or an include-all that no
filter narrows; what filters and providers pick, including provider proxies that carry their own `dialer-proxy`, is
left to this guard.

Desktop RPC accepts both a single event object and batched event lists. Android and desktop listener dispatch isolate
listener exceptions so one faulty observer does not prevent the remaining events/listeners from running.

## User-Facing Core And Delay Feedback

`CoreStatusButton` in `lib/views/dashboard/widgets/core_status_button.dart` is the desktop dashboard's status/restart
surface. It is shown only outside dashboard edit mode and only when `coreLib == null`:

- Provider state remains authoritative. The widget keeps a separate display-only status so a fast
  `connecting -> connected` transition still shows at least 600 milliseconds of progress instead of flashing.
- The hold arms only after an observed transition to `connecting`; mounting while already connecting does not invent a new
  delay. A real `disconnected` transition cancels the hold immediately so failure is never hidden, while a long-running
  connecting state remains visible after the timer expires.
- Taps during the display hold or while the provider is genuinely connecting are inert. Connected/disconnected taps show
  the appropriate confirmation and delegate restart to `CoreAction`; the widget never starts Core directly.

Proxy delay testing follows the same failure-safe UI rule. `ProxiesAction._runDelayTests` marks its keys in
`pendingDelayTestsProvider` and releases them in `finally`, and a test the Core does not answer leaves the last
measurement in place; the delay-test bullets under Core API Safety in `.agents/rules.md` give the reasons. A key stays
`queued` until the `TaskPool` actually runs it, so node cards animate a loader for at most `maxConcurrentDelayTests`
nodes and show a static glyph for the rest; a loader on every queued node made fast scrolling drop frames. A group's delay
test button, the tab page's action and the list page's group header alike, spins while `delayTestingGroupsProvider`
holds the group; `delayTestPageGroup` marks it, ignores a second request for the same group, and clears it in `finally`,
so an RPC failure cannot leave the button spinning and remounting a header does not lose the state.

The dashboard's probe cards show a failure as still checking for five seconds after they see `proxied` turn true,
through `ProbeStartHold` in `lib/views/dashboard/probe_start_hold.dart`: `ServiceStatusCard` holds its check and the
exit IP beside it, and `NetworkDetection` holds the exit IP, whose tap to retry stays inert meanwhile.
`SetupAction.setRunning` flips the run time, and with it `proxied`, before the Core has brought its listeners up, so the
probe launched on the spot often fails, and the config re-apply that follows every start probes again moments later;
without the hold a card flashes a failure and then takes it back. The hold is local display state like the connecting
hold above: the cache still records the failure, the hold arms only on an observed transition, a result that is not a
failure shows at once, a stop cancels it, and a failure still standing when it ends is shown.

## Route Consistency

A probe result on the dashboard — the exit IP, a node's exit IP, a service check — either describes the route in
force or is being asked again. The full design is `docs/plans/probe-route-consistency.md`; the mechanism has three
parts.

- The Core is the source of truth for the route. `core/route.go` keeps a structural epoch (advanced by a config
  apply, a mode or IPv6 switch, any proxy provider whose version changed, and any rule set that loaded new content,
  whether the user or its own interval updated it) and a `picks` map from each Selector, URLTest and Fallback group
  to what it resolves to right now, versioned on every change. `handleChangeProxy` re-reads the picks under
  `selectMu` and answers `ChangeProxyResult{changed}`, so the connection reset only runs when the pick moved. A
  health check moving a URLTest pick has no mihomo hook, so the picks are polled every second while the host watches
  (`watchRoute(true)`) and the listeners are up; the same poll notices a provider or rule set that updated in the
  background. Every probe answer carries the counters it started under; `routeChanged` publishes the state whenever
  it differs from the last read.
- `routeTrackerProvider` in `lib/providers/route_state.dart` merges that with the host's side: `proxied` (running and
  not suspended), a host epoch that a connectivity event moves, and `synced`, which is false until `watchRoute` has
  returned a snapshot and again after a Core disconnect. Nothing is fresh while unsynced, so a cache never judges
  answers against counters it made up. The tracker subscribes when the first target is watched and releases the Core
  when the last one goes; an answer newer than the state means an event was evicted, and the cache asks for a resync.
  It also releases the Core while the app is hidden and reads it again on return, keeping `synced`: the Core and its
  counters are the same ones, so the fresh snapshot invalidates only what moved. On Android the Core drops the watch
  itself when the event listener is cleared, because an engine that goes away never sends `watchRoute(false)`.
- `RoutedProbe` in `lib/providers/routed_probe.dart` is the one cache both `OutboundIpProbe` and `ServiceStatus` mix
  in. A widget `watch`es the target it shows and `unwatch`es on the way out; only watched targets are probed, one
  request in flight per target. An answer is stamped with the Core's counters and the host state at launch, and
  `RouteState.isFresh` walks its chain against the picks, so a selection change only invalidates results that went
  through the changed group while an epoch change invalidates everything. A stale watched entry is asked again at
  once; an unwatched one is dropped. Nothing launches while the app is hidden: what falls due waits for it to be shown.
  Failures wait for `retry`, `refresh`, or the app resuming (`markResumed`), and
  a service check that timed out is a failure that still shows its status. `refresh` is the user's own request and
  runs for unwatched targets too: the service sheet reads the cache and never watches its rows, because one service
  check is one to four full HTTPS fetches where a delay test is a single empty request.

`ipQuality(ip)` in `lib/providers/ip_quality.dart` is keyed by the outbound IP found, so a route change reaches it as
a new IP and an unchanged IP reuses the result. A lookup races its sources the way the IP check does and takes the
first one that determines a type. A success is kept for the session with `ref.keepAlive()`; a failure is not cached,
and Riverpod's automatic retry is off because a retry after an HTTP 429 spends the same per-outbound-IP quota again.

## Icons

`lib/icons/` holds every icon the app shows. Nothing in `lib/` uses the Material icon font, and
`test/lint/material_icons_test.dart` fails on any `Icons.*` or `AnimatedIcons.*`, so every icon shares one stroke and
can animate its shape.
A `Glyph` is a list of shapes on a 24-unit grid, stroked at `Glyph.strokeWidth` with round caps and joins and
superellipse corners; a curve the primitive shapes cannot draw is a `GlyphPath`, SVG path data traced along the
stroke's center. Each shape has a role that decides its filled form, the way SF Symbols and Fluent pair an
outlined icon with its filled variant:

- `body` fills;
- `detail` is cut out of the fill, and a `solid` detail such as a dot is cut as an area;
- `line` stays a stroke, for parts that belong to no body, like a handle or a link between nodes; it thickens from
  `Glyph.strokeWidth` toward 2.2 as the glyph fills, so a glyph of lines alone still carries weight in a filled button.

`GlyphPainter` paints any point between the two forms. Each body fills inward from its outline, reaching half its
shorter side, which no point of the body lies beyond, as `fill` reaches 1. Details thin away over the first half and
return as cutouts over the second, so a detail is never a stroke and a cutout at once. The glyph is painted opaque into
one layer that carries the color's alpha: a translucent color, such as a disabled one, then stays even where strokes
cross the fill, and a cutout clears only the glyph's own pixels. `GlyphIcon` shows a glyph at a fixed `fill`, taking
its size, color, and fill from `IconTheme` unless it is given its own, as `Icon` does, and mirroring it in right-to-left
text when `Glyph.matchTextDirection` is set.
`AnimatedGlyph` drives a `GlyphIcon` from its `filled` flag and pops briefly when it fills. A glyph whose shape follows
some other state, like the sidebar toggle that tracks the pane's width, is a builder in `AppGlyphs` such as
`AppGlyphs.sidebar(progress)`. A new glyph goes in `AppGlyphs` and in the list `test/icons/glyph_painter_test.dart` runs
through its checks; it is named for what it means in the app, one glyph per meaning, with SF Symbols and Fluent as the
reference for the metaphor. Tests find one with `find.byGlyph` from `test/helpers/glyph_finders.dart`.

Icons are outlined by default and filled inside a filled button: `IconButton.filled` and `.filledTonal`, `FilledButton`
in its tonal and icon forms, and floating action buttons draw their glyph with `fill: 1`, because an outline on a tinted
container reads weaker than the container it sits in. A glyph with no body only thickens its lines and still takes
`fill: 1`, so the rule holds if the glyph later gains one. `test/lint/filled_button_glyphs_test.dart` checks every
`GlyphIcon` written inside such a button; an icon that reaches the button from elsewhere, as through a helper that picks
the button variant, gets its fill by hand or from `IconThemeData.fill`. Everywhere else fill marks state only: the
selected destination, a pinned window, a lit torch. The phone's floating navigation bar is the one exception: like an iOS
tab bar it fills every destination and marks the selected one by colour alone.

Framework widgets that draw their own icon take it from the theme: the `ActionIconThemeData` in `lib/application.dart`
supplies the back and close buttons, with `AppGlyphs.backFor` picking the platform's back glyph, and a `DropdownMenu`
is given its chevrons. A framework widget added later that paints a Material icon needs the same treatment.

Empty states draw no image assets either. `NullStatus` in `lib/widgets/null_status.dart` shows the glyph for its
meaning filled on a Material 3 Expressive shape from `material_new_shapes`, with a few accent marks, and enters with
one of three choreographies picked at random: the shape spins in, morphs out of a circle, or springs in while the
accents orbit into place. The glyph takes `onPrimaryContainer`, not `primary`: the `content` scheme variant can leave
`primaryContainer` as light as `primary` in dark mode.

`CaptionIcon` is the one icon that does not follow the glyph grid: the window caption buttons keep the one pixel,
pixel-aligned stroke of Windows' 10-pixel caption glyphs on every platform.

## App Bar Actions

`CommonScaffold` keeps at most three buttons in sight in its bar, the overflow menu button included. A page hands it
what may fold as data: `primaryAction`, `iconActions`, `menuItems`, and `searchActions` while searching; custom widgets
passed through `actions` never fold. Past three, two buttons stay beside the overflow menu: the search button (the clear
button while searching) when there is one, then foldable actions, primary action first unless `foldPrimaryAction` puts
the icon actions ahead of it;
everything else goes into one overflow menu, primary action first, and a folded action that is loading shows as a
disabled item. The bar reads as one button group of two beside a single button: the overflow menu button when
something folds, otherwise the last action, so the proxies list shows a search and delay test group beside the menu. `primaryAction` counts only where the bar holds it, in a bottom sheet and on a phone's home
page, whose foot the navigation dock takes; elsewhere it is the page's floating action button. No page adds a
floating action button of its own: a secondary action such as selecting every app goes into `menuItems`, and saving
an edit page is a check among its `iconActions`. Actions on a multi-selection, such as delete and select all, go into
`selectionActions` instead: the bar keeps them in sight as one button group that takes a single slot beside the search
or clear button, so they never fold.

A bar under a sheet drops its actions and keeps its title and back or close button. `CommonScaffold` fades the actions
out with the entrance of a sheet laid over its page, or over the sheet it is a page of, so none on the layer below
reads as an action of the sheet on top; `sheetCoverOf` in `lib/widgets/sheet_navigator.dart` reads that entrance, which a sheet route hands
down through `SheetCoverRouteMixin`, as `SnapSheetRoute` and `ModalSideSheetRoute` do.

## Row Actions And Sort Mode

A list row or card carries no trailing more button. Its secondary actions go into a `ContextMenuRegion`, which opens
them on a right click, a touch long press, the menu key or Shift+F10 of a focused row, and, under an enabled
`RemoteFocusAdapter` (Android TV), a select key held past the long-press timeout; a short select there still presses
the row, on key up rather than key down. A tap keeps the row's main action. A row whose tap does something other than
open its page, as a profile card's tap selects the profile, may carry a trailing button to that page, as iOS's detail
button does, built as a `DetailButton` so a desktop's compact density cannot pull it off the column the indicator and
the handle share; a trailing button that repeats the row's own tap is never added. A row whose tap opens a page shows
a `DisclosureIndicator` there instead, as a script or a rule-set proxy provider does. A provider opens its content on
a tap, read-only unless it is a local text file the app owns, and keeps a second action on its trailing button: editing
for a proxy provider backed by a text file, the options dialog for an app rule provider.

A long press can mean only one thing on a row. Where rows have no context menu, as in a plain list editor, the global or
profile rules, the on-demand SSIDs or the custom proxies, a long press on the row drags it and there is no sort button.
Where they have one, the list reorders only in a sort mode. When the list is the page, `CommonScaffold` owns the mode as
it owns search and selection, and the page passes `canSort` for the toggle in its bar. Changes apply as they are made,
so nothing confirms them: while the mode is on, the bar shows only its toggle, selected, and the back or close button
fades in place rather than offer the same way out twice, as iOS hides its back button while editing; a system back still
ends the mode. A list that is one section of its page, as the custom profile's proxy groups are, keeps its own
`SortModeScope` around its rows and its toggle in its header and leaves the page bar alone, back button included; its
header keeps adding while it sorts, and a system back ends its mode before it leaves the page. Each row sits in a
`SortableItem`, which reads the scope: while it sorts `DecorationListItem` drops its tap and context menu, and
`SortableTrailing` springs a `SortHandle` in from the row's trailing edge while the trailing it replaces shrinks away.
The motion is painted, not built, and the row relayouts only while the two widths differ, since every row on screen
animates at once. A page that only reorders, such as the profiles sort sheet or a group's members, shows its
`SortHandle`s without a mode. A handle drags on press, and with focus a select key lifts its row for the up and down
keys to move through the enclosing `SliverReorderableList`'s own `onReorderItem`, which is how a remote sorts.

## Quick Edit

Quick edit, the compose button that opens what a page edits as text in an `EditorPage`, appears only where that text is
the whole of it: a full list or section, such as a custom profile's rules, its proxies as share links, or its DNS, NTP
and sniffer sections, or one isolated item, such as the definition of the proxy group or proxy a form edits. A page
that edits only part of something other settings also write offers none, since its document would show a fragment and
reject what it leaves out. The app-wide DNS page is that case, a few keys laid over the DNS of every URL and file
profile, so only the profile's DNS spec mixes in `OverrideQuickEdit`. So is the custom profile's TUN page: it picks a
few keys of a `tun` section the network settings own, so its spec does not mix in `OverrideQuickEdit`.

## Where A Route Opens

On desktop each navigation destination has its own `Navigator` (`_NavigationPage` in `lib/pages/home.dart`); on mobile
there is none, and everything goes on the root navigator, `rootNavigatorKey` in `lib/common/navigator.dart`. A route
takes the navigator of the context it opens from: `BaseNavigator.push`, `showExtend`, `showSheet` and `showSnapSheet`
all do, so a page or sheet opened from a destination stays beside the navigation rail, leaves the rail usable, and
stays in that destination's stack when the user switches away and back. `sheetNavigatorOf` lifts a sheet out of a
paged sheet's own `SheetPagesNavigator`, never past the destination.

The root navigator is only for what belongs to no single destination:

- dialogs, which `dialogs` in `lib/common/dialog.dart` hosts on the root unless it is given a context, and the notifier;
- a gate in front of the whole app, such as `requestDisclaimerConsent`;
- work that starts with no page behind it: a deep link, the tray, a hotkey, the update check.

Never reach for the root to open a page or sheet from a page, whether through `rootNavigatorKey`,
`globalState.navigatorKey`, `rootNavigator: true` or `useRootNavigator: true`, and never push a page from the context of
something on the root, such as a dialog: the page lands on the root too and covers the rail. The add profile sheet is
the one exception. `addProfileFormFile`, `addProfileFromLink` and `addProfilesFromLinks`, which a deep link also reaches,
close it with a root `popUntil` before switching to the profiles tab, so `showAddProfilePage` opens it on the root and
takes the caller's context as `origin`; a page the sheet leads to opens from `origin`, as `AddProfileView._toAddCustom`
does.

The mistake only shows in the desktop layout, and `TestApp` has nothing but the root navigator. A test that guards
where a route opens hosts the page in a nested `Navigator`, as the desktop case of a new custom profile in
`test/features/profile_views_test.dart` does.

## Settings Rows

`lib/widgets/config_item.dart` holds the shared settings-row vocabulary: `ConfigToggleItem`, `ConfigOptionsItem`,
`ConfigTextItem`, and `ConfigListEditItem`. Each takes a `selector` (a `ProviderListenable`, normally
`someProvider.select(...)`) and an `onChanged(ref, value)` writer, and watches its own selector so changing one setting
rebuilds one row instead of the whole section. Titles and subtitles are `ConfigLabel` callbacks that receive
`AppLocalizations`, which keeps literal labels such as `IPv6` and localized labels in the same shape.

Build settings screens from these directly, or from a file-local helper that binds one provider once — see `_networkToggle`
in `lib/views/config/network.dart` and `_appSettingToggle` in `lib/views/config/general.dart`. Declare a named
`ConsumerWidget` only when a row is genuinely reused across screens, as `lib/views/config/network.dart` rows are by
`lib/views/dashboard/widgets/quick_options.dart`. Rows with bespoke behaviour — a custom dialog, a derived value, or a
second provider write — stay hand-written rather than growing extra parameters on the shared items.

In an inset grouped list (`generateSectionV3`, `generateAnimatedSection`), a row's subtitle shows its value or state,
never what the setting does. As iOS settings do, the explanation goes in the group's `footer`, drawn by `ListFooter`,
one sentence per setting that names it, and a section header carries only its title. Flat `generateSection` lists keep
their descriptions on the row.

## Keyboard Insets

The soft keyboard moves `MediaQuery.viewInsets` on every frame of its animation, so whatever depends on them rebuilds
and relays out at frame rate for a few hundred milliseconds. `ThemeManager` wraps the whole app in a `LayoutBuilder`,
which runs every one of those rebuilds inside the layout phase; in a profile trace they show up there, not under
`BUILD`.

- A page the user is not typing in holds its bottom inset through `KeyboardInsetHold`
  (`lib/widgets/keyboard_inset_hold.dart`). Pages under another route and inactive tabs are still laid out, offstage or
  not, and without it every page the app keeps relaid out, and rebuilt under its `LayoutBuilder`s, with each frame of a
  keyboard raised by a dialog or sheet on top. The home route, each home tab, `CommonRoute`, `CommonDesktopRoute` and
  the open-container route wrap their page in one; a new page route type does the same.
- A widget high in the tree reads the aspect it needs (`MediaQuery.paddingOf`, `viewInsetsOf`, ...), never
  `MediaQuery.of` or `MediaQuery.removePadding(context: ...)` on its own context, which subscribes it to every inset
  change. Move such a read into a small widget of its own below it, as `_AppMediaQuery` in `ThemeManager`,
  `_BodyPadding` in `lib/pages/home.dart` and `_SheetKeyboard` in `SnapSheet` do.
- A bottom sheet makes room for the keyboard itself: `_SheetKeyboard` lays the content out above it and removes the
  bottom inset from the content's `MediaQuery`, so nothing inside a sheet, `CommonScaffold` included, adds keyboard
  space. A nested sheet must not carry the keyboard through its pages: `NavigatorResizable` learns a page's new size a
  frame late, and the sheet would trail the keyboard by a frame.

## Page Arrival

A pushed page is built, laid out and painted through every frame of its transition, so whatever it starts on arrival
(loading and decoding images, building the rows its data brings in) lands inside the animation. Impeller has no
raster cache, and every offscreen pass is paid again each frame on a phone GPU.

- Work a widget would start as it mounts waits for the route to settle. `RouteSettledMixin`
  (`lib/widgets/route_motion_hold.dart`) calls `didSettleRoute` once the route that brought the page in has arrived,
  and at once on a page that already sits still; `whenRouteSettled` (`lib/common/navigator.dart`) does the same for a
  one-off future, and both wait out the frame a `HeroController` builds a pushed route offstage with its animation
  pinned to completed. `CommonTargetIcon` and `PackageIcon` show what is already in memory at once and load the rest
  after that. An icon still loading leaves a blank slot of its size and shows its fallback only once the load ends
  without an image, so a grid of placeholders is never painted through the transition only to be replaced.
- A scroll view of image tiles takes `arrivalScrollCacheExtent(routeSettled)`: the rows past the viewport would
  otherwise be built in the same frame as the ones on screen, which for a screen of icon cards is most of the push.
  The icon grids (`IconGridScrollView`) and the access control list do this, and so does `ListEditorPage`, whose
  rows are heavy enough that only its list rebuilds as the route settles, reusing the rows already built.
- A page whose content waits on what it loads on arrival (lists read from the database, file sizes, the installed fonts)
  shows its `NullStatus` illustration, with no label, inside a `NullStatusSwitcher` until all of it is in, then switches
  to the finished content in one step, as the custom profile, resources and font family pages do. Never a skeleton
  screen of placeholder rows: it has to guess how many rows are coming, fills in piece by piece, and rebuilds every row
  it stood in for. `SkeletonText` is for a single value inside a card that is already on screen, such as the dashboard's
  service status. `holdsArrival` keeps the status a switcher opens on for at least 400 ms, about a route transition,
  so a load that finishes just after arrival does not flash it; the connections, resources and font family pages set
  it.
- `GlyphPainter` paints a translucent glyph through an offscreen layer. A placeholder repeated in every tile blends its
  color over the surface instead of carrying an alpha.
- `PagedSheetRoute` and `CommonRoute` share
  `SharedXPageTransition` (`lib/widgets/page_transition.dart`) and its `pageSlide`, a quarter of the page's width as
  Android 14 slides pages, so a page moves alike in a sheet and on its own. The transition fades a page by painting the surface over it
  and needs no opacity layer: the page on top shows only on its side of the point where the two fades cross, opaque
  over the other, so neither shows through the other. A page below that does not fade with it, such as the home
  route under a `CommonRoute`, is faded by the page on top painting the surface over it. `CommonDesktopRoute` fades in
  place through a plain 200 ms `FadeTransition`.

## State Management

Provider files in `lib/providers/`:

- `app.dart`: runtime/UI state, logs, traffic, delays, loading, navigation.
- `config.dart`: persistent config providers, app settings, theme, VPN, proxy style.
- `state.dart`: derived/computed providers, navigation, proxy, tray, color scheme.
  Like `action.dart`, this is an entry point only: the providers live under
  `lib/providers/state/` and are joined with `part` directives, so importing
  `state.dart` still reaches all of them.
  - `state/proxies.dart`: group and proxy lists, filter/sort, delay, selection.
  - `state/navigation.dart`: navigation items, current page, dashboard, more tools.
  - `state/system.dart`: tray, VPN params, access control, hot keys, shared state.
  - `state/theme.dart`: dynamic color, color scheme, brightness.
  - `state/profile.dart`: profiles, current profile, clash config, setup state.
  - `state/custom.dart`: custom profile validity and the staged group/rule notifiers.
- `action.dart`: business logic notifiers, setup, backup, core lifecycle, proxy selection.
- `core.dart`: `coreHandlerProvider`, the container-scoped handle on `CoreController`.
- `route_state.dart` and `routed_probe.dart`: the route tracker and the probe cache; see Route Consistency above.
- `outbound_ip.dart` and `service_status.dart`: the dashboard probes built on them.
- `database.dart`: Drift database provider wrappers.

### Reaching Singletons

`lib/common/` and `lib/core/` publish process-wide singletons (`coreController`,
`system`, `preferences`, `appPath`, `request`, and others). Code that already has a
`Ref` or a `WidgetRef` reads them through a provider instead, so a test can scope a
fake to one `ProviderContainer` rather than swapping a global and relying on a
tearDown to put it back.

`coreHandlerProvider` is the established case. Every call site under
`lib/providers/`, `lib/manager/` and `lib/views/` goes through it; notifiers and
`ConsumerState` classes that touch Core repeatedly hold it as
`CoreController get _core => ref.read(coreHandlerProvider)`.

Tests override it with `coreHandlerProvider.overrideWithValue(CoreController.scoped(fake))`,
which does not claim the singleton. `CoreController.test` does claim it, and is
only for tests that have not moved yet. Prefer the scoped override even when a
test passes either way: a test that claims the singleton makes the global and the
provider resolve to the same fake, so it cannot tell a provider read from a
leftover global read, and a half-migrated call site stays green.

One deliberate exception: `globalState` owns the `ProviderContainer`, so it cannot
itself live in one. Code without a `Ref` reaches providers through
`globalState.container.read(...)`. The remaining call sites are `lib/core/lib.dart`,
`lib/common/print.dart`, and the tray `read` callback in
`lib/providers/actions/system.dart` — all singletons or platform callbacks with no
`Ref` in scope.

`lib/models/profile.dart` does not reach Core. `Profile.saveFile` and
`Profile.update` take a `ValidateConfig` callback, and every caller passes
`(path) => _core.validateConfig(path)`, keeping the Core handle lazy so a profile
path that never validates never resolves the controller.

The UI layer must not reach a process-wide singleton directly.
`test/lint/ui_layer_singleton_test.dart` scans `lib/views`, `lib/widgets`,
`lib/pages` and `lib/features` for `globalState.container` and the bare
`coreController` global and fails the run on either. Widgets that need Core hold
`CoreController get _core => ref.read(coreHandlerProvider)`.

`globalState.measure` and `globalState.theme` stay global on purpose. Both are
context-derived caches assigned by `ThemeManager`, and tests already scope them by
assigning in the app builder (see `test/helpers/test_app.dart`); moving them into
providers would touch every layout call site without changing behaviour.

`globalState` in `lib/state.dart` is a singleton holding ambient app state — the
package info, the measure and theme, the container, and the start/stop flags —
plus `safeRun`/`loadingRun`. Startup orchestration is **not** on it: `init` and
`attach` live in `lib/bootstrap.dart`, above `lib/common`, because they drive the
window, the autostart entry, the tray and the permission prompts. Providers are
generated into `lib/providers/generated/`.

The root navigator key lives in `lib/common/navigator.dart` as `rootNavigatorKey`;
`globalState.navigatorKey` is a getter onto it. `lib/common/dialog.dart` reaches
the key directly and does not import `lib/state.dart`.

### Platform Layering

`lib/common/common.dart` deliberately does not export `tray.dart`, `window.dart`,
`launch.dart`, or `system_dns.dart`. Those four modules import
`tray`, `window`, `launch_at_startup`, and `screen_retriever`;
exporting them would put those packages in the compile graph of every file that
imports the barrel for a string helper. Import the specific module instead.

`test/lint/platform_layering_test.dart` enforces four rules. Three are local: the
barrel never re-exports one of those four modules, nothing under `lib/common`,
`lib/enum` or `lib/models` other than those four imports a desktop platform
package, and `lib/common` never imports the `lib/manager/manager.dart` barrel
(import the single manager needed, as `common/context.dart` does with
`manager/status_manager.dart`). The fourth walks the barrel's whole transitive
closure and fails if *any* file in it imports one of those packages. That one is
the real invariant — the local rules only stop the shortest path, and every leak
found so far arrived through a longer one.

The shapes that keep the closure small:

- `System.back` and `System.exit` do not touch `window`; the window half of
  both lives in `SystemAction`.
- `KeyboardModifier.toHotKeyModifier()` lives in
  `lib/manager/hotkey_manager.dart`, its only consumer, not in `lib/enum/enum.dart`.
- Startup orchestration lives in `lib/bootstrap.dart`, not on `GlobalState`.
  `common/num.dart`, `common/print.dart` and `common/request.dart` import
  `state.dart` for `theme`, `container` and `packageInfo`/`ua`, so anything
  `GlobalState` reaches lands in the barrel's closure; the ambient state it
  holds reaches nothing platform-specific.
- `SystemAction` talks to `WindowPort` and `TrayPort` from
  `lib/common/app_ports.dart` instead of importing `common/window.dart` and
  `common/tray.dart`. `lib/bootstrap.dart` binds `windowPort` and `trayPort` to
  the real implementations; both stay null in tests, where every call through
  them is a no-op. A test that needs the real tray assigns `trayPort` itself, as
  `test/common/tray_menu_test.dart` does.
- `lib/providers/providers.dart` re-exports `action.dart`, so `lib/common` and
  `providers/app.dart` import `providers/state.dart` and `providers/config.dart`
  directly; the barrel would reach the whole action layer.
- A data type in a lower layer never holds the widget that renders it, because
  that drags the whole view tree into the barrel's closure. The route table is
  `lib/views/navigation.dart` implementing `NavigationPort`, which
  `providers/state/navigation.dart` reads through and `bootstrap.init` binds;
  unbound it yields no items, so a test that renders navigation assigns
  `navigationPort` itself, as `test/pages/home_test.dart` does. `DashboardWidget`
  is plain data, since it is persisted in the app settings; the widget mapping
  lives in `lib/views/dashboard/widget_registry.dart`, the reverse lookup relying
  on the branches returning canonical consts.

`test/lint/platform_layering_test.dart` also pins that the barrel's closure
contains no `lib/views` file. Data the provider layer needs from the UI layer
goes through a port in `lib/common/app_ports.dart` rather than an import in the
other direction.

### High-Frequency Buffers

`logsProvider`, `requestsProvider`, `dnsQueriesProvider` and `trafficsProvider` hold a `FixedList`
(`lib/common/fixed.dart`), which trades a normal copy-on-write for a shared
buffer tagged with a generation counter:

- `append` mutates the buffer in place and returns a new wrapper one generation
  ahead. That is what providers publish, so `updateShouldNotify` still fires.
- `list` returns an immutable copy, cached until the next mutation. It must stay
  eager: an older wrapper shares the buffer, so its contents move on. Anything
  that needs a stable view has to read `list` at the moment it is notified, not
  hold the wrapper and read later.
- Consumers that only need to know *that* the buffer changed watch `revision`,
  not `list` — selecting on the list snapshots and deep-compares the whole
  buffer on every arrival, which is what this design exists to avoid. See
  `lib/views/logs.dart` for the pattern: watch the generation, snapshot inside
  the throttled callback.

`add`/`clear` mutate in place without advancing the generation; use them only on
a buffer you own outright (seeding, resets, tests), never on published state.

### Background Work

A proxy client runs for days with nobody looking at it, so work that only feeds the UI stops when the UI cannot be
seen. There are three levels of gating:

- `appVisibleProvider` is false while the desktop window is hidden or the Android activity is in the background
  (`hidden`, `paused`, `detached`). `AppStateManager` writes it. The one-second run-time and traffic ticker in
  `SetupAction` follows it. The macOS menu bar title is the exception: it shows the traffic while the window is
  hidden. `RouteTracker` and `RoutedProbe` follow it too, as described under Route Consistency.
- `ActivePollingMixin` is stricter for per-widget polls such as connections and memory. It polls only while the app is
  `resumed` and its home page is the current one.
- `_NavigationPage` puts every page except the current one under a disabled `TickerMode`. Home pages are kept alive in
  the `PageView`, and Flutter does not mute tickers on a kept-alive page off screen. A spinner left on one would
  otherwise keep scheduling frames.

A provider's own `ref.listen` delivers nothing while no one listens to that provider. That is the usual state of an
action notifier, which is only ever read. A derived provider it listens to is never recomputed, and even a `ref.read`
of it returns a stale value. So an action that has to follow a signal listens to the keep-alive provider that holds
it, as `SetupAction` does with `appVisibleProvider`, not to a provider derived from that.

On Android the Core does not react to the screen or to Doze. Doze's firewall keeps apps without an exemption off the
network before their traffic reaches the TUN, so what still arrives is traffic Doze allows, FCM among it; suspending the
tunnel rejects exactly that and stalls push reconnects until their backoff runs out. The VPN service keeps its own
network as a foreground service (`dumpsys netpolicy` shows its `DOZE` block allowed by `FOREGROUND`), so health checks
keep producing real results: holding them blocks the re-check mihomo runs after repeated dial failures, which is how a
`url-test` group leaves a dead node, and hiding their failures hides the truth. Nothing re-probes on resume either:
every proxy group is a provider, `select` groups included, and checking one tests every node in it. Should a ROM be
shown to cut the VPN's own network while the screen is off, holding back failures measured then needs only a screen
on/off signal.

## Database

The app uses Drift/SQLite in `lib/database/`. `schemaVersion` in `lib/database/database.dart` is the migration
counter; a schema change bumps it and goes into the one step the coming release takes, as the Migrations section of
`.agents/rules.md` describes.

Tables:

- `Profiles`
- `Scripts`
- `Rules`
- `DisabledRules` (`disabled_rules`)
- `ProxyGroups` (`proxy_groups`)
- `CustomProxies` (`custom_proxies`)
- `ProxyDialers` (`proxy_dialers`)
- `IconRecords` (`icon_records`)
- `IconSets` (`icon_sets`)
- `ClashProviders` (`clash_providers`)

`Profiles.type` is `file`, `url` or `custom`, fixed when the profile is created. A rule belongs to whoever its
`profile_id` names: no profile makes it global, a file or url profile's rules are what its standard extension adds, and
a custom profile's rules are all the rules it has. `DisabledRules` lets a profile leave out a global rule. Proxy groups
and dialers exist only on custom profiles. Every one of these rows cascades with its profile. Rule and proxy-group
ordering use fractional indexing.

A custom profile has no file and no extension: `SetupAction.getProfile` builds its config from an empty map plus its
groups and rules, and `clashConfigProvider` answers an empty `ClashConfig` for it. Its groups name app-level custom
proxies directly, and the config declares only the ones some group names; they take nodes from every file or url
profile too, each a file proxy provider under its label. A custom profile is never a provider itself. App-level
`ClashProviders` rows are rule providers only, and only custom profiles are offered them.

A custom profile names groups, local proxies, provider profiles and rule sets by name, mihomo's own strings included
(a DNS server's `#proxy`, a `rule-set:` policy key, the sniffer's lists). `lib/models/references.dart` reads and
renames every such place, and renames, delete prompts, provider injection and the profile checks all go through it;
`test/models/references_test.dart` fails when `config.go` reads a name somewhere it does not. A rename is detected
where the row is written, so every path carries it. A local proxy may not share a name with any custom profile's
group, as both live in mihomo's one namespace; where older data has both, the name means the group in that profile
and a proxy rename leaves it alone. `Profiles` writes rows one at a time from the state as it then stands, so a rename
rewriting custom profiles is never written over by a put issued before it landed.

An app-level rule set's behavior and format are what the Core's `compileRuleSet` found in its source, never a user
choice. The source stays at `getProviderCachePath(fileName)`; a domain or ipcidr set also gets an `.mrs` beside it, and
that is what the config loads, while classical rules have no mrs form and load from the source. The Core downloads a
remote set too and mihomo sees it as a `file` provider, so a sync goes through `ClashProvidersAction`, not mihomo's
HTTP vehicle. `ClashProvidersAction.prepare` rebuilds a missing cache while the config is built, since backups carry
only local sources and sets saved before compiling existed have no `.mrs`.

`IconSets` and `IconRecords` only feed the proxy-group icon picker. A set is an icon gallery in the format Quantumult X
and Loon subscribe to, kept whole as one JSON column; a record is an icon the user saved on a group, written on save and
never by merely displaying it. A group stores the icon string itself, a URL or a data URI, never a reference to a set,
so renaming, syncing or deleting a set leaves every group as it was.

Generated Drift output lives in `lib/database/generated/database.g.dart`. After schema changes, run code generation and add or update focused database tests under `test/database/` when converter or migration behavior changes.

### Unreadable Stored Data

Preferences, database rows and backups can hold values this build cannot read: an enum a newer build wrote before a
downgrade, a fork sharing the data directory, a hand edit. Reading them gives up the smallest unit it can and never
fails startup or a whole restore.

- `Config.fromJson` and every `safeFromJson` decoder go through `decodeSalvaging`, which keeps each value the strict
  decoder accepts, descending into maps and lists, and leaves the rest at their defaults; a list or section none of
  whose entries can be read falls back to its default rather than to empty. Only the migration uses
  `Config.strictFromJson`, since it tells legacy data apart by the failure. A field's own `fromJson` throws on what it
  cannot read instead of restoring a default for the whole field, so the salvage around it drops just that entry.
- DAO entity queries read through `TableInfo.readable`, which skips a row it cannot map and leaves it in the table. A
  profile's JSON columns that hold selection state, subscription info or overrides fall back to empty instead, so the
  profile stays listed.
- Rules, disabled rules and proxy groups are the exception and read strictly, in the app and from a backup alike:
  skipping one would quietly change how traffic is routed, so a damaged row fails loudly instead.
- Code that deletes files or rows by comparing against what a query listed must use raw ids (`ProfilesDao.ids`,
  `ScriptsDao.ids`, `ClashProvidersDao.fileNames`), or it deletes what a skipped row owns. Startup reorders profiles
  with `Profiles.reorder`, which writes rows but never deletes them.

## Manager Stack

Managers are nested `InheritedWidget`/`StatefulWidget` components built by `buildManagerStack()` in `lib/application.dart`:

```text
AppEnvManager > LocaleManager > StatusManager > ThemeManager
  > [Desktop: WindowManager > TrayManager > HotKeyManager > ProxyManager]
    [Mobile:  AndroidManager > TileManager]
  > AppStateManager > CoreManager > ConnectivityManager
  > [Desktop: WindowHeaderContainer] [Mobile: VpnManager]
  > app content
```

Each manager in `lib/manager/` handles a specific platform concern. The
platform slots are exclusive: no desktop manager appears on mobile and no mobile
manager appears on desktop.

The order is an ownership contract, not a layout detail. `ConnectivityManager`
sits below `CoreManager` because its `onConnectivityChanged` callback reads
Core-backed state, so Core must already be mounted when it fires. `StatusManager`
and `ThemeManager` sit above the platform managers so a platform manager can
surface a message or read the theme.

`buildManagerStack()` is a pure function of `isDesktop`, the connectivity
callback, and the app content, so `test/application_test.dart` asserts the whole
order by constructing the stack without mounting it. Changing the nesting means
updating both this diagram and that test.

## Core Controller and Actions

`lib/core/controller.dart` (`CoreController`) is a singleton facade over `CoreHandlerInterface`. Public methods delegate to the platform-specific interface, either the Android service channel (`CoreLib`) or the desktop socket. It has an `@visibleForTesting` constructor and `resetInstance()` for test injection.

`lib/providers/action.dart` is the public library entry point for action
providers. The Riverpod notifier implementations are split by responsibility
under `lib/providers/actions/` and joined to the entry point with `part`
directives, so consumers continue to import the same public API:

- `CommonAction`: update check and common UI operations.
- `SetupAction`: config setup and TUN management.
- `BackupAction`: backup/restore orchestration over a per-run temp work directory; the view supplies where an archive
  goes (local save, WebDAV upload) or comes from. The archive format lives in `lib/common/backup.dart`, which stages a
  restore without touching the data directory; files, database rows and settings are applied only after the whole
  archive has been read and validated.
- `CoreAction`: core lifecycle, initialization, coalesced restart, and post-restart profile/running-state application.
- `SystemAction`: system integration, tray, coordinated resource cleanup, terminal Core close, exit, and brightness.
- `StoreAction`: profile storage operations.
- `ThemeAction`: theme state updates.
- `ProxiesAction`: group management and proxy selection.
- `ProfilesAction`: profile CRUD, auto-update, import.
- `GeoResourceAction`: geo resource updates and URL configuration.
- `UpdatingAction`: stale sweep over `UpdatingKeys`.

`UpdatingKeys` in `lib/providers/app.dart` owns every per-entity updating flag; `isUpdatingProvider(key)` is the
read-only view widgets watch. Callers pair `start(key, scope: ...)` with `stop(key, operation)` using the returned
token, so overlapping operations on one key are reference counted and a late `stop` from a superseded operation
cannot clear a newer one. Keys started with `UpdatingScope.core` depend on the Core to make progress, so
`UpdatingKeys` discards them itself when `coreStatusProvider` leaves `connected` — that is the state's own
invariant, not a policy, and it must stay inside the notifier where no warm-up ordering can miss it.
`UpdatingScope.local` keys (profile updates run entirely in Dart) survive a Core restart. `UpdatingAction` holds
only the policy half: the periodic sweep that discards a key stuck past `updatingStaleTimeout`. Do not move the
timeout back into `UpdatingKeys`, and do not widen the disconnect reset to every scope.

## Safe Mode

`--dart-define=SAFE_MODE=true` builds a FlClash that never changes the host network, so a development build can run
next to an installed one. It is a compile-time constant with two faces: `safeModeBuild` in `lib/common/constant.dart`
for the process-wide handles that exist before the provider container, and `safeModeProvider` in
`lib/providers/app.dart` for everything reached through Riverpod, which is also what tests override. Settings are never
rewritten; each owner reads the flag where it would otherwise act:

- Core config: `SetupAction._effectivePatchConfig` is the one place the profile path and `updateConfig` derive what
  the Core receives from the user's `PatchClashConfig`. In safe mode it reports `tun.enable: false` and
  `ExternalControllerStatus.close`, so the generated profile and every `updateConfig` carry no TUN and an empty
  `external-controller` (the Core's `route.ReCreateServer` binds no RESTful API port) while the settings keep the
  user's values and the installed app keeps `127.0.0.1:9090`. `SetupAction.requestAdmin` skips authorization.
- System clock: `_makeRealProfileTask` sets `write-to-system: false` on whatever `ntp` section the generated profile
  ends up with, so neither the user's override nor the profile's own NTP config can move the host's time.
- Profile listeners and firewall rules: `_makeRealProfileTask` clears `listeners`, `tunnels`, `ss-config`,
  `vmess-config`, `tuic-server` and `iptables`, and closes `dns.listen`, preventing profile-only changes during apply.
- System proxy: the `proxy` plugin handle in `lib/common/proxy.dart` is `null`, so `ProxyManager` and exit cleanup
  never write the OS setting, not even the `stopProxy` that would clear what the installed app configured.
- System DNS: `systemDnsCoordinator` is `null` and `shouldPatchSystemDnsProvider` is `false`, so neither the apply nor
  the restore path runs. The restore path matters because a profile or release build shares the preferences of the
  installed app (only a macOS or Android debug build has its own bundle id) and would restore from its record.
- Login item: the `autoLaunch` handle in `lib/common/launch.dart` is `null`, so neither the boot sync nor the setting
  listener registers the build's executable or removes the installed app's entry, which on Windows is the same
  `FlClash` value in the user's Run key.
- Privileged Core: `CoreService` resolves the launcher with `hasHelper: false`, which is the direct child-process path
  on every desktop platform; the Helper keeps managing the installed app's Core.
- Core listeners: `SetupAction` never calls `startListener` or `stopListener`, for a run or for an on-demand
  suspension, so the Core's `isRunning` stays false and `updateListeners` binds nothing, while `isStartProvider`
  and `applyProfile` behave as usual. The installed app keeps the mixed, DNS and other loopback ports.
- Android: `sharedStateProvider` hands the service `VpnOptions.enable: false` and `systemProxy: false`, so the
  foreground service runs without a `VpnService`.
- Wi-Fi SSID: `ConnectivityManager` never reads the SSID and `LocationPermissions.refresh` neither checks
  nor requests location authorization, so `currentSSIDProvider` stays `null` and the excluded-network suspend never
  trips. The macOS `wifi_ssid` plugin builds its `CLLocationManager` and `CWWiFiClient` lazily, so a build that
  never asks never reaches locationd or wifid.
- Global hotkeys: `HotKeyManager` neither subscribes to hotkey events nor registers the configured shortcuts with the
  OS; the in-window `Shortcuts` (close window, escape) stay.
- URL schemes and launcher shortcuts: `Window.init` skips the Windows registry and Linux desktop-entry registration
  of the `clash` schemes, and Android never calls `initShortcuts`.
- Presentation: `AppEnvManager` shows a `SAFE MODE` banner ahead of the channel banner, and `TrayState.safeMode` gives
  the tray the `status_4` bug icon, a suffixed tooltip and a menu of only show and exit. `trayStateProvider` and
  `trayTitleStateProvider` report `showTrayTitle: false`, so the macOS menu bar carries no traffic title.

A new host integration, anything that changes the OS's routes, proxy, DNS, login items, privileges or VPN state, or
that asks the OS for a permission, a global shortcut or a registration, must consult the flag at its owner in the same way and get a test that overrides `safeModeProvider`.

## Build System

`setup.dart` is the release build orchestrator: it writes `env.json` (`APP_ENV`), activates `flutter_distributor` from
the `chen08209/flutter_distributor` fork pinned to a `v<version>-flclash.<n>` tag (cut a new tag there and bump
`--git-ref` when the fork changes), and leaves the Core artifacts to the build hook.

The desktop Go core and the Rust helper are built by a Dart build hook; Android's Core is built by Gradle (see Android
Core Build). `plugins/setup/hook/build.dart` only constructs
`CoreBuilder`, a `package:hooks` `Builder` in `plugins/setup/setup_hooks/`, the same shape `rust_api` uses. Flutter
runs it for every platform build and for `flutter test`, once per target architecture. `CoreBuilder` turns the hook
input into a `BuildRequest` (repository root, `Target`) and hands it to `buildPlatform`. It lives in
`setup_hooks` so `dart test` there can cover it; a test run of `plugins/setup` itself would execute the hook. Failures
reach the runner as `BuildError` (Go or Cargo failed, no Core for the architecture) or `InfraError` (a toolchain could
not start, the package is not at `plugins/setup`).

What the hook protocol forces:

- `PATH` arrives unextended, so `runCommand` appends the Homebrew, Go and rustup locations that Xcode's and Gradle's
  stripped `PATH` hides.
- The hook does not know the build mode: the protocol carries none and `linkingEnabled` only says whether link hooks
  run, so nothing in the hook branches on debug versus release. It cannot tell `flutter test` from a build either; the
  `build_assets: false` user-define in the root `pubspec.yaml` is the one switch (see `.agents/commands.md`), and
  `setup.dart` refuses to package while it is set.
- The hook declares the files it read but the *directories* it wrote: `hooks_runner` hashes a file by content and a
  directory by child names, so `libclash/macos/` catches a deleted artifact without MD5-ing a 55 MB Core on every
  build. A rebuild touches those directories after the runner's cutoff, so the next build runs the hook once more and
  setup's cache answers it; that extra run is expected.
- The runner discards hook output on success, even under `-v`, so the hook mirrors it into
  `.dart_tool/setup_build_cache/hook.log` (rotated at 512 KB), one `===` header per invocation.

Platform projects copy the artifacts out of `libclash/`; application code must not import `plugins/setup`:

- Android: nothing. The hook returns without building, because it runs once per architecture while gomobile packs
  every ABI and one set of Java classes into a single AAR.
- macOS: a standalone `FlClashCore`. `Release.xcconfig` pins release and profile `ARCHS` to the host because
  flutter_tools otherwise builds a universal binary and every artifact ships one slice; the hook skips a non-host slice
  for the same reason. The `Stage Core` phase copies the Core after the hook may have rewritten it and fails when it is
  missing or lacks a slice for `ARCHS`, so a skipped hook cannot stage a stale Core silently.
- Linux and Windows: `FlClashCore`, the Rust `FlClashHelperService`, and a `manifest.json` holding `coreSha256`; the
  Core builds first because the Helper embeds its hash. The CMake `install` rules copy them, and the Windows bundle
  places `manifest.json` beside the executable. A Helper running from a Debug build keeps its exe open and the install
  fails behind an opaque `MSB3073`, so `windows/CMakeLists.txt` stops it from an `install(CODE)` step for the `Debug`
  configuration only, and only when the staged Helper differs from the installed one; that step is where the
  configuration is known, and a Release install never stops a registered service. The step matches the process by
  executable path rather than image name, so a Helper registered from an installed release keeps running.

Setup keeps its own cache under `.dart_tool/setup_build_cache/v1/` because it builds a Go core and, on Windows and
Linux, a Rust helper:

- Go fingerprints cover the target-specific `go list -deps` inputs in `core/` and `Clash.Meta`, module files, the
  effective build configuration, `setup_hooks` sources, target flags and the Go toolchain.
  Helper fingerprints cover its Rust sources and manifests, Cargo/Rust toolchains and flags, and the expected Core
  SHA256.
- A hit requires the fingerprint and every recorded output's path, size and modification state to match, and skips
  compilation and copying.
- Records are written only after a successful build, under per-target locks. Compilation happens in a staging
  directory and moves into place on success, so a failed build never destroys the previous artifacts.
- `flutter clean` removes `.dart_tool` and forces one full rebuild; deleting `.dart_tool/setup_build_cache/` does the
  same without a clean.

Windows helper integrity/version check:

- `setup_hooks` constructs the Core first, calculates its SHA256, and always
  builds the Rust Helper with release hardening and that expected hash.
- Flutter reads the Core SHA256 from the bundled `manifest.json` and sends it with `/ping`. Debug, Profile, and Release
  builds use the same Helper protocol and may use TUN through the same flow.
- `/ping` is loopback-only and requires no request token. The Helper compares the requested SHA256 with its embedded value
  and checks that the fixed `FlClashCore.exe` beside it exists; `/start` performs the actual Core SHA256 verification before
  every launch. The response includes the running Helper path and protocol header; Dart checks both against the current
  installation. The launcher selects the Helper only when `/ping` reports ready; any other readiness (missing manifest,
  unavailable Helper, or a Helper built for a different Core) falls back to the direct Core without requesting elevation.
  If `/start` reports a pre-spawn failure — `coreVerificationFailed` (the on-disk Core no longer matches the SHA the
  Helper and manifest agree on) or `processLaunchFailed` (the Core process could not be spawned) — the launcher degrades
  to the direct Core rather than failing the launch. `/start` releases the previously managed Core before it verifies,
  so the Helper owns no Core when either code is reported and the direct retry cannot race a Helper-managed Core.
  A mismatched Helper is reinstalled through the explicit TUN authorization flow, not at startup.
- TUN is not a required run condition. A direct Core runs unelevated and cannot bring up TUN, so any degrade to the
  direct Core — an unready Helper at resolve time, or a pre-spawn `/start` failure — silently drops TUN and keeps the
  Core running. Degrading is preferred over failing the launch: an unverified Core carries no privilege the direct
  launch path did not already have. `manifestMissing` is the one readiness that is surfaced to the user, because it
  means the installation itself is incomplete.
- Flutter creates a 128-bit lowercase-hex session ID and uses it as the random named-pipe suffix. `/start` receives only
  that address and session ID, validates the fixed `FlClashCore_<session>` namespace, starts the fixed Core beside the
  Helper, and returns the same session ID plus the spawned PID. Flutter verifies both the session and named-pipe peer PID.
- `/stop` requires the same session ID. A missing process returns `notRunning`; a different owner returns
  `sessionMismatch` without terminating that process. Session IDs are ownership tokens for lifecycle safety, not a claim
  that the loopback HTTP endpoints are authenticated.
- The Core must not outlive the Helper. On Linux the systemd unit's cgroup and `Restart=on-failure` cover that; on
  Windows the Helper puts the Core in a Job Object with `KILL_ON_JOB_CLOSE`, so the kernel kills the Core when the Helper
  process ends for any reason, and the `install` subcommand configures SCM failure actions (three restarts, 5 s apart,
  reset after 24 h). Stopping a Core sends SIGTERM first on Linux, because only a shutdown removes the policy routes
  sing-tun installed; the Core runs the same teardown itself when its IPC host disconnects or it receives a termination
  signal. `HelperClient.stopTimeout` must stay above the Helper's graceful-plus-kill budget.
- Never take `MANAGED_CORE` or `LOGS` with `lock().unwrap()`. The Helper is a long-lived service running as SYSTEM, so a
  single panic while a lock is held would poison it and turn every later request into another panic — the service stays
  dead until Windows restarts it. `lock_surviving_poison` recovers the guard through `PoisonError::into_inner` instead.
  `hub.rs` uses it at every lock site, tests included, and two tests in that file pin the behaviour.

Build configuration defaults live in `plugins/setup/setup_hooks/lib/src/options.dart` and can be overridden via the root
`build_config.yaml`.

`core_version` in `build_config.yaml` is the mihomo release tag the `core/Clash.Meta` fork is based on. Upstream injects
its version only from its own CI, so the fork alone reports the placeholder `1.10.0` in `/version` and in protocol
user agents; `GoBuilder` passes the value as `-X github.com/metacubex/mihomo/constant.Version=...` for every build,
local and CI alike. Bump it whenever the submodule moves onto a new upstream release.

Architecture detection is automatic. The `--description` flag passed to `flutter_distributor` adds arch suffixes to artifact names, such as `FlClash-0.8.93-macos-arm64.dmg`.

#### Android Core Build

`android/core/build.gradle.kts` registers `bindCore`, which runs `gomobile bind` on `core/mobile`. `:core` cannot
depend on the AAR itself: `bundleAar` refuses a local AAR in a library, which would break `./gradlew assemble` and
Android Studio's Make Project. It takes the AAR apart instead, `classes.jar` as a file dependency and `jni/` through
an `unpack<Variant>CoreJniLibs` task per variant handed to `variant.sources.jniLibs.addGeneratedSourceDirectory`; the
legacy `sourceSets` DSL rejects a task-built directory. Gradle therefore runs the Go build before any task that needs
its classes or libraries, in every variant and whether or not Flutter started the build, and its up-to-date check over
`core/` (tests, `testdata` and `cmd/` left out), `build_config.yaml`, the Go version, the ABIs, the API level and the
NDK version stands in for the hook's cache.

- `gomobile` and `gobind` come from the `tool` directives in `core/go.mod` and are built into the task's temporary
  directory, so their version moves only with `go.mod`. `bindCore` finds `go` on `PATH` or in the Homebrew and Go
  installer locations that Android Studio's GUI `PATH` leaves out, and puts the JDK running Gradle first for
  gomobile's `javac`.
- The ABIs come from the `target-platform` property Flutter passes, and default to all three for a Gradle build Flutter
  did not start. `ANDROID_HOME` and `ANDROID_NDK_HOME` are AGP's SDK and its `ndkVersion` NDK, so the Core and the app
  share one NDK, and `-androidapi` is the app's `minSdk`.
- `tags`, `go_ldflags` and `core_version` come from `build_config.yaml`, the file the hook reads.
- Both Cores take their GODEBUG defaults from the `go` line of `core/go.mod`, whatever toolchain builds them. The
  desktop Core gets them from the Go command because `core` is its main module. gomobile links the library as the main
  package of a module it generates, whose `go` line is the toolchain's, so `bindCore` hands the desktop Core's
  `DefaultGODEBUG` (from `go list` on `./cmd/core`) to the Android Core through `-X=runtime.godebugDefault`. Without
  it, a toolchain newer than that line would move only Android's TLS, x509 and runtime defaults. Raising the `go`
  line therefore changes both Cores' behaviour at once; pin a setting with a `godebug` line in `core/go.mod` to hold
  it back on both.
- The TLS settings among them do nothing for proxy traffic. Every TLS connection the Core makes goes through
  `metacubex/tls` or `metacubex/utls`, ports of `crypto/tls` that hard-code their version's defaults and never read
  GODEBUG, so `tlssha1` and `tlsmlkem` cannot be held back. They still parse and verify through the standard
  `crypto/x509` and `crypto/rsa`, and both run before `skip-cert-verify` is consulted, so `core/go.mod` keeps
  `x509negativeserial=1` and `rsa1024min=0`: a server certificate with a negative serial number or an RSA key under
  1024 bits would otherwise fail the handshake even on a proxy that skips verification.
- gomobile names the library `gojni`, both as the file it builds and in the `System.loadLibrary` of the `go.Seq`
  runtime, with no option to change it. `bindCore` rewrites that string constant in `go/Seq.class` to `clash` as it
  extracts `classes.jar`, and the unpack task renames `libgojni.so` to `libclash.so`; each fails the build unless it
  finds what it rewrites, the constant exactly once. Go writes no `DT_SONAME` into a c-shared library, so the ELF
  needs no change.
- The AAR's own `proguard.txt` goes with the rest of it. `android/core/consumer-rules.pro` keeps `go.**` and
  `com.follow.clash.core.mobile.**`, the classes gobind's C glue looks up by name.
- The CI `android` job compiles `:core` without Flutter, so it passes `-Ptarget-platform=android-arm64` to build one ABI.

#### Android Crash Symbols

Crashlytics NDK symbolicates a native crash only from libraries that still carry their symbols, and AGP strips every
`.so` it packages, so the Android build keeps symbols until `:app`'s own strip step and the APK does not grow:

- `bindCore` drops `-s` and `-w` from `go_ldflags`, and gomobile strips nothing itself; the desktop Cores stay
  stripped because nothing consumes their symbols.
- `:core` sets `packaging.jniLibs.keepDebugSymbols` so its own strip task leaves `libclash.so` intact on the way into
  `:app`'s merged native libs, which is where the Crashlytics plugin reads it.
- `:app` turns on `nativeSymbolUploadEnabled` and finalizes `assembleRelease` and `bundleRelease` with
  `uploadCrashlyticsSymbolFileRelease` only when release signing is configured, the same condition that marks a real
  release build; the checked-in `google-services.json` is a placeholder, so a local upload is noise at best, and CI
  supplies both files together. The Crashlytics plugin wires the R8 mapping upload itself but leaves the native symbol
  upload to the caller, so nothing else triggers it. `generateCrashlyticsSymbolFileRelease` runs `dump_syms` over the
  Go Core for every ABI, which is the other reason it stays off for unsigned builds.

## Local Plugins

- `setup`: build-time harness for Go core artifacts and the Rust helper, driven by a Dart build hook; no runtime Dart API.
- `proxy`: system proxy configuration. On Windows the package installer and uninstaller kill the app, and then run
  `FlClash.exe --clear-stale-proxy`, the installer once the new files are in place; the runner hands it to the plugin
  before any engine starts, and it turns off only a proxy still set to FlClash's `127.0.0.1:<port>` with nothing
  listening on that port.
- `rust_api`: runtime Flutter Rust Bridge FFI package built through Native Assets. See below.
- `tray`: system tray for Linux, macOS and Windows. Written for FlClash; replaced the `tray_manager` fork.
- `window`: desktop window control for Linux, macOS and Windows. Written for FlClash; replaced the
  `window_manager` fork. See below.
- `wifi_ssid`: Wi-Fi SSID detection.
- `code_forge`: the text editor behind `EditorPage`, derived from upstream `code_forge`. See below.

## code_forge Plugin

`CodeForgeController` is the document: text lives in the Rust `RopeBridge`, and the controller's methods are split by
concern into `lib/code_forge/controller/` extensions. `CodeForge` builds `_CodeFieldRenderer`, whose painting, layout,
folding, caret, gutter and pointer code sit in `lib/code_forge/code_area/` part files.

- Rust counts offsets in Unicode scalars and Flutter in UTF-16 code units. Convert at the boundary with
  `text_offsets.dart`; a mixed-up offset only shows up with astral characters such as emoji, so tests use them.
- The renderer learns what changed from `takeChange()`, whose `lines` (first line, lines removed, lines inserted) let it
  follow edits without reading the whole text; a whole-text `text =` assignment marks the change `replaced` and resets
  the renderer instead.
- The input method is handed a window of the document around the caret, at most 4096 characters, and its deltas are
  mapped through that window, so the platform has to hold the current one: a caret move that shifts the window, the
  input method's own included, sends it again. A selection can reach past the window; when the platform removes the
  selection it sees, the controller removes the rest with it, as one undo step. Gboard does that by collapsing the
  selection and deleting what it saw, then composes the typed letter before a newly sent window reaches it: deltas
  made on a window already replaced are not mapped through the new one, and only what they typed is kept. A line
  longer than the window gets one cut by character count, which stays where it is until the caret nears an end of it,
  or every key would replace it. A caret move can put a window of the very same text further along; an edit the
  platform then makes at the caret it had is mapped through the window it still holds.
- A selection the input method collapsed is still the one to replace only while the caret rests at the end it was
  left at; Gboard's spacebar caret gesture collapses to the middle and moves on, which lets it go. The platform's own
  cut (Gboard's text-editing panel) arrives as a lone deletion of the live selection and copies only the window's
  part, so the controller rewrites the clipboard with the whole selection. That panel's select-all and copy still
  reach only the window. Gboard types over a selection in two batches, the removal and then the letter: the second
  joins the undo step of the first, and undoing it gives the selection back.
- A composition is drawn over the document and written into it only when it ends, so whatever reads or rewrites the
  document from outside the input method commits it first: saving, leaving the page, accepting a suggestion, a
  snippet, find's replace, and `UndoRedoController`'s `undo`/`redo` through the `settleInput` callback the controller
  registers. Losing focus and a closing connection commit it too rather than drop it. Typing that ends a composition,
  Enter or a bracket, goes through the typing path so it still indents and pairs. Android ends one with an empty
  delta batch when the keyboard is hidden.
- A caret the input method moves has to set `selection` on the change record like any other move, or the renderer
  neither repaints nor follows it.
- A line's fold comes from the renderer's `_computeFoldRangeForLine` up to 10000 lines and from `compute_all` in
  `rust_api`'s `editor/folds.rs` past that, and the two must find the same ranges; `folding_test.dart` compares them on
  one document. A folded range, and the folded ranges inside it that it folds again when it opens, outlive an edit
  only while the text still gives that range. An edit reports the lines it changed, not where in them, so a fold
  heading a line inside an edit that changes the line count is opened rather than guessed at; `replaceText`, which
  find's replace-all uses, rewrites only the span that differs to keep that to the lines between its matches.
- The caret never rests on a line a fold hides: folding over it moves it to the heading line, the keyboard steps over
  the fold, and a caret put there any other way opens it.
- The controller reaches the mounted editor only through the `CodeForgeView` it attaches, which the editor's state
  implements by forwarding to its renderer.
- FlClash supplies the suggestion popup, context menu and scrollbar; the package has no built-in versions of them.
- Tests live in `test/plugins/code_forge/` at the repository root and need the `rust_api` library.
- `isMobile` comes from `Platform`, so host tests never reach the touch branches (selection handles, long press,
  magnifier); check those on an Android emulator.

## window Plugin

`plugins/window` exposes `desktopWindow` (`DesktopWindow`) on the `window` method channel; `plugins/window/README.md`
is the channel contract (methods, arguments, events, effects, runner hooks) and every platform implements exactly that
table. `lib/common/window.dart` and `lib/manager/window_manager.dart` are the only application callers.

- Each platform keeps one style record (title bar style and button visibility everywhere, rounded corners on
  Windows, effect, tint and brightness on Windows and macOS) and reapplies it whole from a single `apply` routine, so DWM margins
  on Windows and `isOpaque`/effect views on macOS have one writer. Add a visual setting to that record, never a second
  code path.
- Method dispatch is a table on every platform, and every handler validates its arguments; a bad call answers with a
  `bad_args` `PlatformException`, a capability the platform lacks with `unsupported`. Coordinates cross the channel in
  logical pixels; Windows converts with `GetDpiForWindow`.
- Position and size changes surface as one `geometry-changed` event that each platform coalesces natively (150 ms
  after the last `WM_WINDOWPOSCHANGED`, did-move/did-resize, or configure-event); `WindowManager` only debounces the
  capture on top of it.
- Windows: with `titleBarStyle: hidden`, `WM_NCCALCSIZE` takes the frame insets from `AdjustWindowRectExForDpi` and
  keeps the client rect at the top edge (plus the frame height when maximized, so the taskbar stays uncovered; plus
  one physical pixel on Windows 11, which draws the top border only into a non-client strip; none on Windows 10,
  where DWM draws the whole caption over any strip shorter than one). Windows 10 gets its top border the Windows
  Terminal way instead: the DWM frame is extended over the caption height, the plugin places the Flutter view one
  physical pixel lower and paints that row black, which DWM treats as alpha 0, and hit-tests it as `HTTOP`; see
  `plugins/window/README.md`.
  The plugin therefore owns the view placement on `WM_SIZE`. Fullscreen is emulated by the
  plugin, which raises `enter-full-screen` and
  `leave-full-screen` itself and mutes `WM_SIZE` while it is on. The find-running-window IPC
  (`WindowPluginFindRunningWindow`/`WindowPluginActivateWindow`) registers a window message from the executable path
  and whitelists it through UIPI so a non-elevated launcher reaches an elevated instance.
- Linux: GTK drops the placement of an unmapped window, so `hide` saves the geometry and the `map-event` handler
  applies it, then corrects the placement for 150 ms while the window manager settles (it may place the window
  after the map handler ran). `isPositionSupported` is X11-only; Wayland owns placement and re-shown windows land where the
  compositor puts them. Linux supports no effect: GTK 3 has no desktop-neutral blur protocol, so `isEffectSupported`
  is true only for `none` and the runner keeps the default visual.
- macOS: `WindowPlugin.instance` is reached from `AppDelegate` for `applicationShouldTerminate` (emits
  `should-terminate`) and `applicationShouldHandleReopen` (emits `activate`); `MainFlutterWindow` calls
  `hiddenWindowAtLaunch()` so the first `order` keeps the window hidden until Dart shows it.
- Window effects (`transparent`, `blur`, `acrylic`, `mica`) are capability-gated by `isEffectSupported`; the Flutter
  content must paint a translucent background for them to show. `ThemeProps.sidebarBlur` (on by default) drives
  them: `Window.setBlur` picks `blur` on macOS and `acrylic` (or `blur`) on Windows and caches only a successful
  probe, so a platform that answered unsupported is asked again the next time blur is turned on. `WindowManager` reapplies it
  whenever the setting or the app brightness changes and records the outcome in `windowBlurProvider`, and only the
  sidebar and the caption strip paint translucently while it is on. On macOS the sidebar paints fully transparent
  and the effect view follows the window's active state, which is the flutter_acrylic `TransparentMacOSSidebar`
  look; Windows keeps a 50% tint because the accent effects blur too little for the rail to stay readable.
- Sidebar blur is soft-disabled: `feature.sidebarBlur` (`lib/common/feature.dart`) is off unless the build passes
  `--dart-define=FEATURE_SIDEBAR_BLUR=true`. While it is off the theme page hides the toggle and
  `windowBlurRequestProvider` asks for no effect whatever `ThemeProps.sidebarBlur` holds. `WindowManager` still sends
  that request at startup: the macOS plugin installs its effect view in `ensureInitialized` and leaves it active until
  the first `setEffect`, so the `none` request is what makes the window opaque and parks the effect view.

## rust_api Crate Layout

`plugins/rust_api` has no platform folders. `hook/build.dart` is a Dart build hook: Flutter runs it for every platform
build and for `flutter test`, and `flutter_rust_bridge_hooks` (over `native_toolchain_rust`) compiles the crate with
Cargo and registers `librust_api` as a code asset that Flutter bundles and signs. The build runs
`rustup run <channel>`, so `rust/rust-toolchain.toml` must pin an exact channel and list every target the project ships;
the hook refuses `stable` and a target missing from that list. `rustup show` installs the pinned toolchain and those
targets on first use.

`flutter_rust_bridge_codegen generate` (see `.agents/commands.md`) writes `lib/src/rust/frb_generated.dart`,
`frb_generated.io.dart`, `frb_generated.web.dart`, one `lib/src/rust/api/<module>.dart` per Rust API module, and
`rust/src/frb_generated.rs`; none of them is edited by hand. A new Rust function is a public function in `rust/src/api/`,
plus `pub mod <name>;` in `api/mod.rs` and an export from `lib/rust_api.dart` when it starts a new module. `RustLib` is
the singleton entry point: `RustLib.init()` runs before any call and `RustLib.dispose()` on teardown.

`plugins/rust_api/rust/src/` separates the bridge boundary from the code behind it:

- `api/` is the only input flutter_rust_bridge parses (`rust_input: crate::api`). Every function there is a thin
  delegation, so the generated bindings stay identical on every platform.
- `ipc/` implements the desktop socket server: `frame` (length-prefixed framing), `outbox` (the byte-bounded send
  queue), `access` (socket mode, the Unix peer uid check, and the peer PID), and `server` (bind, the accept loop, and
  the read and write halves of a connection). Each `Server` runs a current-thread tokio runtime on an `ipc-server`
  thread of its own and holds no global state, so I/O is event-driven on every platform and closing or dropping the
  handle releases the socket. It serves one peer at a time; a second one waits in the backlog.
- `script/` runs profile override scripts on QuickJS through `rquickjs`.
- `editor/` backs `plugins/code_forge`: a `ropey` buffer, fold ranges and bracket matching, indent guides for the
  viewport, and completion words. `api/editor.rs` exposes the buffer as the opaque `RopeBridge`; the selection
  stays in Dart. Only the editor imports this module, through `lib/editor.dart`, so `lib/rust_api.dart` keeps its
  generic names such as `GuideBlock` out of app code.
- `hotkey/` registers desktop global shortcuts through `global-hotkey`: `keys` maps Flutter USB HID usages to key
  codes, `owner` runs every registration on the thread the platform binds it to (a dedicated message-loop thread on
  Windows, the main dispatch queue on macOS, in place on Linux), and `service` owns the registry and forwards presses
  to Dart. Linux is X11 only; a Wayland session without XWayland gets an error rather than a silent no-op.

`IpcServer.bind` resolves once the address is bound, `events()` attaches the one event stream, `send` only enqueues
and is a synchronous call so frames keep the order they were sent in, and `close` joins the thread. A bridge function
that takes a `StreamSink` must not return an error: the generated Dart code drops that call's future, so an `Err`
surfaces as an unhandled async error instead of reaching the caller. That is why binding is its own call and
everything after it is reported as an event.

What a platform does not use, it does not compile. `interprocess`, `tokio` and `global-hotkey` are declared under
`cfg(not(target_os = "android"))`, and `ipc/mod.rs` and `hotkey/mod.rs` swap in their `unsupported.rs` there, because
Android loads the Core in-process and has no global shortcuts. Adding a capability follows the same shape: implement it in its own module, gate the
dependency by target, and keep the `api/` entry point unconditional so one set of bindings still serves every platform.

`RustLib.init()` runs on every platform in `main`, because the script engine and the editor are shared; the editor
shows its unavailable state when `RustLib` is not initialized rather than loading the library itself.

## Profile Script Engine

`lib/common/javascript.dart` sends the profile as JSON to `evaluate_script`, which runs `main(config)` on QuickJS and
returns the JSON the script produced. Nothing about the script runs in Dart.

- QuickJS is compiled from source for the target being built, which is what removed the prebuilt `quickjs-c-bridge`
  binaries: `flutter_js` shipped x64 Windows and desktop-only libraries, so Windows ARM64 could not start (#2361).
- `rquickjs` carries pre-generated bindings for every target this project builds except the Android ones, so Android
  builds enable its `bindgen` feature. That needs the NDK's own libclang and sysroot: `native_toolchain_rust` exports
  the sysroot through `BINDGEN_EXTRA_CLANG_ARGS_<target>`, and `hook/build.dart` adds `LIBCLANG_PATH` from the NDK
  toolchain Flutter hands the hook, because bindgen otherwise loads whatever libclang the host has, or none.
- Evaluation is bounded: a 10-second interrupt deadline and a memory ceiling, because a script that never returns would
  otherwise hold the profile forever. `console` is installed before the script runs, since scripts written for other
  clients log as they work.
- `rust/tests/fixtures/profile_script.js` is the compatibility regression: an extension script written for the suite that
  performs the transform real ones perform, so it exercises `Map`/`Set`, spread, destructuring, optional chaining,
  nullish coalescing, `Object.fromEntries`, named capture groups and lookbehind in one pass. Keep it first-party and
  free of external URLs — vendoring somebody's published script here carries their attribution and their links.

## Rust Helper Service

`services/helper/` is the privileged helper that starts the core elevated so TUN works. It ships on Windows and Linux
and is built by the setup build hook alongside the Core, which always compiles the Helper in Rust release mode
after calculating the SHA256 of the Core produced for the active Flutter configuration.

The helper owns its Windows Service Control Manager lifecycle through two elevated commands:

- `FlClashHelperService.exe install --owner-pid <pid>` reads the account of that process (the app, which Dart names with
  its own PID because the elevated installer may be another administrator), stops and removes any stale registration,
  creates the auto-start service for the current executable path with `--owner-sid <SID>` as its launch argument,
  starts it, and waits for the running state. A service started without an owner exits instead of serving anyone.
- `FlClashHelperService.exe uninstall` stops the service, waits for shutdown, removes its registration, and is also used
  by the Windows package uninstaller.

The Dart layer only launches the helper's `install` command through `ShellExecuteW`; it does not compose `sc.exe`,
`taskkill`, or `cmd.exe` command lines.

Linux takes the same shape with systemd in place of the Service Control Manager, and the same install timing: nothing
is registered at package install, and `Linux.registerService` asks for elevation only when TUN authorization needs it.
Elevation goes through `pkexec` so polkit raises the system prompt. When pkexec is missing, or exits 127 because no
polkit agent runs (common under tiling window managers), it falls back to `sudo`: cached credentials first, then a
password the app asks for and hands to `sudo -S` on stdin. A 126 means the user dismissed the polkit dialog and is final.

- `FlClashHelperService install`, run through `LinuxElevation` (`lib/common/linux_elevation.dart`), writes
  `/etc/systemd/system/flclash-helper.service` for the current executable path and enables and restarts it. It reads
  `PKEXEC_UID`/`SUDO_UID` to learn who asked, and refuses to install without one — there would be no account to grant
  the socket to. It also refuses a Helper whose binary or directory is not root-owned and non-writable (a unit runs it
  as root at every boot, so an unpacked bundle would be a standing escalation), and refuses to replace a unit already
  installed for a different UID rather than restart the service out from under that account.
- That ownership check is why the `flutter_distributor` fork normalizes the packaging tree to 0755/0644 before
  `dpkg-deb`, `rpmbuild` and `appimagetool` run: they record modes verbatim, and Ubuntu's per-user default umask
  of 002 would otherwise ship `/opt/FlClash` as 0775, which the installer rejects as group-writable.
- The rpm spec sets `debug_package` and `__os_install_post` to nil for the same reason: rpmbuild's find-debuginfo and
  brp-strip rewrite `FlClashCore`, and a Core whose SHA256 no longer matches the Helper's embedded value is refused at
  `/start`, which silently degrades every launch to the direct Core.
- `FlClashHelperService uninstall` disables the unit, removes it and reloads systemd.
- The unit carries `Group=` (the owner's primary GID), `RuntimeDirectory=flclash`, the owner's UID/GID in
  `FLCLASH_HELPER_OWNER_UID`/`_GID`, a double-quoted `ExecStart=` with `%` escaped, and `Restart=on-failure` under a
  start limit so a broken unit ends up failed instead of restarting forever. The helper serves
  `/run/flclash/helper.sock` at mode `0660`, additionally drops any connection whose `SO_PEERCRED` UID is not the
  owner's, logs and retries an `accept` failure instead of letting hyper end the server, and handles SIGTERM so
  `systemctl stop` still runs its own Core teardown.
- `/start` additionally requires the Core address to be a socket owned by the owner UID before spawning, since the
  Core connects to it as root.
- The Core is spawned with the owner's real UID and an effective UID of 0, which is what a setuid Core would have had.
  It is the signal `core/ownership_unix.go` uses to hand the files it created back to the user; without it a root
  service would leave a root-owned config tree in the owner's home.
- An AppImage has neither a stable executable path nor a writable Core, and its FUSE mount is `nosuid`, so
  `system.isAppImage` reports TUN authorization as unavailable instead of prompting for a password that cannot help.
- A Linux host without systemd (`/run/systemd/system` absent) has no Helper: `system.hasHelperService` is false there,
  readiness is the `stat` check, and the same elevation sets the setuid bit on the bundled Core as before.

A setuid Core is a root process for whoever gives it an address to dial (`argv[1]`), so its mode never lets others
execute it: macOS runs `chown root:admin` and `chmod 4750` (and refuses an account outside the `admin` group, which
could not run the result), and a Linux host without systemd hands it to the primary group of the user that
`PKEXEC_UID` or `SUDO_UID` names. `System.isPrivilegedStatOutput` reads a Core that others can execute as unauthorized, so one that an earlier
version marked `+sx` is re-authorized on the next TUN enable.

In every Flutter build mode `/start` opens the fixed Core executable beside the Helper without write/delete sharing,
validates it against the SHA256 embedded only in the Helper, and keeps that handle open through process creation.
`/ping` only compares the requested `coreSha256` with the Helper's embedded value and checks the fixed Core path exists;
it never hashes the Core. Protocol version 6 (8 on Windows, where the owner check and the browser-origin check arrived
and an older Helper must be reinstalled) uses 32-character lowercase-hex session ownership:

- `GET /ping?coreSha256=...` returns the current Helper executable path with `x-flclash-helper-protocol` when the
  requested SHA matches.
- `POST /start` rejects unknown JSON fields and validates `{address, sessionId}` and, on Linux, the socket owner, which
  leaves a running Core alone. It then releases any previously managed Core before verifying the Core — so every
  outcome of a request that passed validation, including a rejected one, leaves the Helper owning no Core — and returns
  `{sessionId, pid}`.
- `POST /stop` validates `{sessionId}` and only stops the matching managed Core. A session mismatch is HTTP 409.
- `GET /logs` exposes the bounded recent Helper/Core stderr buffer with `no-store` caching.

`/start` and `/stop` take at most 4096 bytes and need a `Content-Length` (HTTP 413 and 411 otherwise); `HelperClient`
always sends one. All endpoints reject `Origin` headers and non-local `Host` values, preventing browser requests and DNS
rebinding through an owner-account browser, which would otherwise pass the Windows account check. The accepted
authorities are `127.0.0.1`, `localhost` and the Linux Unix-socket placeholder `FlClashHelperService`. The Helper grants
no CORS access.

Endpoints bind only to `127.0.0.1` on Windows and to `/run/flclash/helper.sock` on Linux, and do not use
request-token authentication. The Windows port is picked by the system, because Hyper-V and WinNAT reserve ranges that
can cover any fixed one, and published as the `Port` DWORD of a volatile `Runtime` key under the service's own registry
key, which only administrators can write; `HelperClient` reads it per request and falls back to 47890, where a Helper
from before listens. Who may call them is decided per connection, before any request is read: the Linux
socket drops a peer whose `SO_PEERCRED` UID is not the owner's, and the Windows listener (`service/peer.rs`) traces the
connection through the TCP table to the process holding its client end and drops it unless that process runs as the
owner SID the service was registered with. The Windows Core runs as SYSTEM and dials whatever pipe `/start` names, so
without that check any local account would host a SYSTEM Core. What no per-connection check closes is a process that
runs as the owner itself, on either platform. Lifecycle safety also comes from the fixed executable/hash, the strict
address namespace (`\\.\pipe\FlClashCore_<32 hex>` on Windows, `/tmp/FlClashSocket_<digits>.sock` on Linux), the session-scoped stop
contract, Dart-side peer-PID verification on Windows and, on Unix, the Core socket that `plugins/rust_api` sets to
mode `0600` so only the owning user (and the root-effective Core) can connect. When the Helper service itself shuts
down, it unconditionally stops the Core process it owns; under systemd the unit's control group does the same.