feat: run native route middleware on in-app navigation

[shanerbaner82]

Fixes #252. Docs PR: NativePHP/nativephp.com#474

The bug

Route::native() returns a real Laravel route, so this binds exactly as it would on a web route — and looks like it works:

Route::middleware(['auth.device', 'auth.lock'])->group(function () {
    Route::native('/', Dashboard::class);
    // …13 more
});

It only ever ran for the screen the app launched into, because that screen alone arrives as an HTTP request. Every screen reached by in-app navigation resolves through NativeRouter::loop() and mounts the component directly — no request, no kernel, no pipeline.

So a v3 → v4 migration that moved routes out of a middleware group silently dropped every guard, with nothing failing and nothing logged. As the reporter notes, component tests can't catch it either, because they mount screens directly.

Worth being precise about the depth of it: the middleware wasn't merely unapplied, it was never recorded. NativeRouter::register() stored class + layout only, so resolve() couldn't have applied it even in principle.

Approach

->middleware() should work, because that's the Laravel way — so this makes it work rather than inventing a parallel API or just erroring.

Capture. Route::native() now hands the live Route object to the native registry. The instance, not a snapshot of its middleware: ->middleware() chains on after the macro returns, and group middleware is merged at registration. Holding the object means the guard reads the final list however it was assembled. Both spellings above work with no change to app code.

Execution. ScreenGuard runs that stack against a synthesized GET request for the target URI, through Laravel's own Pipeline, before the screen mounts.

The synthesized request is the deliberate trade-off, and it's documented rather than hidden. It carries the live session, the launch request's user resolver, its cookies and server bag — so auth, verified, and can: resolve the real user. It does not carry a body, query string, uploaded files, or the device's real headers and client IP. Middleware that needs those should opt out.

Skipped by default. Request-lifecycle middleware — StartSession, AuthenticateSession, VerifyCsrfToken, EncryptCookies, AddQueuedCookiesToResponse, ShareErrorsFromSession. Each ran once for the real launch request; re-running per screen push would reopen the session and rotate CSRF tokens against a request that is never sent anywhere. Apps can add their own via ScreenGuard::skip([...]) for things that should count once per launch rather than once per screen.

Denial maps onto the navigation intents that already exist:

Middleware does	Native result
passes through	screen mounts
redirects to a native route	REPLACE onto that screen
redirects elsewhere	EXIT_WEB with the full URL
throws AuthenticationException	redirect to redirectTo(), else the login route
aborts (403) / returns a response	BACK — refused, user stays put

Two guarantees

A refused screen never mounts. The guard runs before mount(), so a guarded screen performs none of its data loading — no queries, no API calls — and publishes no frame. This is why the check sits in loop() ahead of mount() rather than inside the component.

It fails closed. Middleware that throws (unresolvable alias, a bug in a guard) refuses the navigation. Failing open would silently grant access, which is precisely the bug being fixed.

Also handled: cold start is marked already-guarded so the kernel's real run isn't double-counted (this matters for anything stateful, like rate limiters); and a guard redirecting to the screen it guards is detected rather than looping the runloop forever.

Testability

Native::visit() now runs guards, so the regression class is finally catchable:

Native::visit('/dashboard')->assertReplacedWith('/login');

Native::test(Dashboard::class) still mounts directly and deliberately does not run middleware — it has no route. That distinction is documented, since a suite built only on Native::test() is exactly what let this hide.

Tests

11 new tests in NativeRouteMiddlewareTest, covering in-app navigation, mount suppression, group middleware, middleware chained after the macro, redirect mapping (native / web / self-redirect loop), the skip list, opt-out, and fail-closed.

9 of the 11 fail when the guard is stubbed out to simulate pre-fix behavior. The two that pass are the correct controls — the no-middleware route and the opted-out middleware.

Suite before: 901 tests / 3021 assertions / 0 failures. After: 912 / 3072 / 0 failures.

Review notes

The parts most worth a second opinion:

• The default skip list — whether it's the right set, and whether skipping should be opt-in rather than opt-out. Skipping too little risks side effects per navigation; skipping too much reintroduces "some of your middleware silently doesn't run", which is this bug.
• EXIT_WEB on a non-native redirect target. Reasonable for auth redirecting to a web login, but it does mean a misconfigured redirect drops the user out of the native app.
• Non-redirect denials become BACK. With nothing on the stack there is nowhere to go back to.

🤖 Generated with Claude Code

---

Two bugs found by running the demo app, not the suite

Both were in this branch and are fixed here. Recording them because each exposed a blind spot in the tests, and the tests were extended to cover them.

1. The pipeline destination returned null. Middleware routinely type-hints its return as Symfony\...\Response — Laravel's own make:middleware stub generates exactly that — so handing $next($request) a null destination throws a TypeError. The guard fails closed, so any typed middleware that allowed a navigation silently refused it instead. The destination now returns an empty 204 sentinel compared by identity. The AllowMiddleware fixture had no return type, which is why the suite was happy; it now carries : Response, and both allow-path tests fail without the fix.

2. The route was never bound to the synthesized request. SubstituteBindings calls $route->parameters(), which throws LogicException: Route is not bound. Failing closed turned that into a refusal — and since withRouting(web: routes/mobile.php) puts every native route in the web group, and SubstituteBindings is the one member of that group not on the skip list, this refused every navigation in a normally-configured app. Nothing was clickable.

Now a clone of the route is bound, so the instance in Laravel's RouteCollection — which real HTTP requests still match against — is never mutated.

The suite missed this because every test registered routes bare, with no group, so no web middleware was ever gathered. Testbench registers no web group either, so two new tests install the real one and exercise it: navigation through web both parameterised and not, plus the guarantee that the registry's route instance is left unbound. Both fail without the fix.

Suite now: 914 tests / 3083 assertions / 0 failures, against a 901 / 3021 baseline on main.

One consequence worth a reviewer's opinion

Because withRouting(web: ...) is the standard setup, this guard runs on essentially every navigation in every app, not only on routes the developer explicitly guarded. That is what makes the fail-closed behavior high-stakes: a middleware that throws for any reason takes out all navigation, as bug 2 demonstrated.

Fail-closed is still right for a security control — failing open silently drops guards, which is #252 itself. But it does mean a configuration error degrades to "the app is frozen" rather than "one screen is unreachable". Worth deciding whether that trade is acceptable, or whether a guard exception should be loud (error screen) instead of silent (refused navigation).

[simonhamp]

Is this a dup of #253?