Isolation and IPC transport

NetWiv supports the pinned TAURI® isolation pattern through its Wry backend. Use matching SDK and native package versions; older development packages may not provide this transport. Full security parity remains under validation.

{
  "app": {
    "security": {
      "pattern": { "use": "isolation", "options": { "dir": "isolation" } },
      "csp": "default-src 'self'; connect-src 'self' ipc: http://ipc.localhost"
    }
  }
}

The directory is relative to the configuration file and must contain index.html and a script setting window.__TAURI_ISOLATION_HOOK__. NetWiv.Build embeds these assets before NativeAOT. CLI dev watches the isolation directory even when additional watch folders are overridden. Effective config relocation passes the original absolute asset directory to MSBuild via NetWivIsolationDist.

The sandboxed iframe receives application payloads before AES-GCM encryption. Its key is generated per application instance and never sent to the parent. Core authenticates and decrypts the envelope before capability and command dispatch. Bundled asset CSP hashes and nonces use an HTML5 parser. The iframe protocol rejects cross-origin reads and serves only its secure index route. Every isolation invoke requires an authenticated envelope, including commands without arguments; an empty protocol body cannot bypass the hook.

Wry registers an asynchronous IPC protocol. Fetch uses native request headers, WebView identity and document generation. Request injection exposes its JSON or raw body and headers. CSP/browser rejection switches to postMessage for the subsequent calls in that document, retaining the pinned fallback’s JSON byte-array behavior. The failing call itself rejects without replay because native execution may already have begun. This deliberately differs from the pinned upstream retry. The differential suite selects the fallback with a side-effect-free probe before comparing subsequent requests; it does not count that initial transition as parity. Headers instances fail structured cloning at the isolation boundary as in Tauri; use a plain header record for isolation invokes.

Automatic differential suite

tests/NetWiv.Isolation.Smoke and its locked Tauri reference build the same frontend. After building both native executables:

node tests/NetWiv.Isolation.Smoke/run.mjs /path/to/reference /path/to/NetWiv.Isolation.Smoke
node --test tests/NetWiv.Isolation.Smoke/compare.test.mjs

On WSL, reference/build-linux.sh stages reference sources on a Linux filesystem because upstream codegen changes permissions unsupported on some shared Windows files. Set NETWIV_REFERENCE_TARGET to the Linux Cargo output directory. Run the Node comparison runner under scripts/run-linux-smoke-desktop.sh for X11 or Wayland. Build both hosts from the same stable source before the pair. The runner consumes functional assertions from stdout in memory; it does not save reports, logs or source inventories. Missing results and identical public failures fail comparison.

Twenty-one cases exercise public behavior: queued startup, nested IPC values, three binary input shapes, header records/clone errors, structured errors, Channel ordering, hook blocking, origin boundaries, reload recovery and CSP fallback. The added combinations check fallback Channel ordering, hook blocking, opaque-frame denial and successful IPC after a second reload. Three additional NetWiv security regressions inject forged private messages or reuse old private callback IDs after protocol and fallback requests. They must pass on NetWiv. Pinned Tauri’s failures appear as observations, not counted as public conformance failures or passes. Direct mutation of __TAURI_INTERNALS__ is outside the repository’s supported public API surface.

The Core assertion harness separately covers isolation on/off, protocol/forced fallback and replacement/close combinations. Each combination starts with an allowed local invoke, denies a remote source carrying spoofed local payload and fallback headers, then verifies invocation cancellation and document-owned resource disposal. A late handler deliberately ignores cancellation to check fallback callback suppression. The fake runtime cannot establish browser frame attestation or delivery of protocol responses after navigation; those require the desktop fixture. These checks are NetWiv security regressions rather than additional upstream business-test comparisons.

Windows x64 managed and NativeAOT desktop comparisons passed all twenty-one public cases and the three NetWiv security regressions. The integrated Linux X11 NativeAOT candidate also passed all twenty-four assertions under Xvfb with a tray host. That headless execution is not a paired reference comparison or physical Linux desktop acceptance. The integrated macOS arm64 NativeAOT candidate also passed all twenty-four assertions; its reference comparison and manual interaction remain unverified.

This suite does not establish full API coverage, package-install acceptance or manual GUI acceptance. The upstream business-test parity gate remains open.

macOS source-linked smoke

After building the native adapter, run from the repository root in a logged-in Mac desktop session:

dotnet publish tests/NetWiv.Isolation.Smoke -c Release -r osx-arm64
tests/NetWiv.Isolation.Smoke/bin/Release/net8.0/osx-arm64/publish/NetWiv.Isolation.Smoke

Physical Apple Silicon and hosted arm64 runs passed the NetWiv encrypted IPC, Channel, forged/opaque-frame, reload and fallback smoke. These NetWiv-only runs are not the paired Tauri differential suite above; macOS reference comparison remains open. Do not apply WebKitGTK frame/CORS observations to WKWebView without a matching macOS result. See macOS validation.

---

TAURI® is a trademark of The Tauri Programme within the Commons Conservancy. WRY is a trademark of The Tauri Programme within the Commons Conservancy.

NetWiv is an independent project. Its framework, plugins and templates are not affiliated with, endorsed by, or officially approved by The Tauri Programme within the Commons Conservancy.