elizaos--eliza/upstreams/electrobun-patches

elizaOS patches on top of Electrobun v1.18.1

Tag pinned: v1.18.1 (commit 4eba723c85b97559e1d9e13439d9a92ede0832e8).

Upstream: https://github.com/blackboardsh/electrobun.

Why we fork

The packaged elizaos.app bundle (under packages/os/linux/elizaos/artifacts/amd64/elizaos-app/) ships libNativeWrapper.so v1.0.2 from this Electrobun release. In every QEMU graphics configuration we tested, the serial log stops after:

=== ELECTROBUN NATIVE WRAPPER VERSION 1.0.2 === GTK EVENT LOOP STARTED ===

A previous strace showed bun's main thread blocking in do_sys_poll.

What we found before patching

Reading the upstream wrapper at package/src/native/linux/nativeWrapper.cpp (v1.18.1, lines 6296–6305):

void runGTKEventLoop() {
    initializeGTK();
    printf("=== ELECTROBUN NATIVE WRAPPER VERSION 1.0.2 === GTK EVENT LOOP STARTED ===\n");
    // no fflush
    gtk_main();          // blocks in poll() on the X11 fd — by design
    g_shutdownComplete.store(true);
}

The launcher (package/src/launcher/main.ts, lines 241–252) spawns a bun Worker that runs the app's bun/index.js and then synchronously calls startEventLoop() on the main thread. gtk_main() blocking in poll() is the correct, expected behavior — it is the GTK event loop waiting for X11 events. It is not a deadlock by itself.

That means the wrapper has no code that runs after gtk_main() enters until the Worker thread starts making FFI calls (createX11Window, initWebview, …) that are marshalled onto the main thread via dispatch_sync_main / dispatch_sync_main_void (lines 5673–5798).

So the actual failure has to be one of:

1. The Worker thread is wedged in JS module init and never makes its first FFI call. The wrapper would then sit forever in gtk_main() correctly polling for X11 events that nothing is going to generate, which looks identical from the outside to a "wrapper deadlock".
2. stdout is fully block-buffered (we are not on a TTY — output is going via systemd-cat to the kernel ring buffer), so the Worker has emitted output that is sitting in libc's stdout buffer indefinitely and never makes it to serial.
3. The Worker did make a first FFI call into the wrapper and g_idle_add / the GTK main loop didn't pick it up under QEMU's virtio-gpu / llvmpipe path.

(1) and (2) are bun/app-bundle problems, not wrapper problems. (3) would be a real wrapper bug. The next-run diagnostic patch below distinguishes them.

The patch (diagnostic, not a "fix that hides the bug")

package/src/native/linux/nativeWrapper.cpp:

1. Set setvbuf(stdout, nullptr, _IOLBF, 0) at the top of runGTKEventLoop so every subsequent printf("…\n") is flushed immediately. This eliminates (2) as a candidate for the silent-after-this-line behavior.
2. Add a g_timeout_add_seconds(1, …) heartbeat [wrapper-heartbeat] gtk_main alive tick=N. If this stream stops, gtk_main itself has wedged (a real wrapper bug). If it continues forever with no other wrapper output, the Worker has not made a single FFI call — (1) is the cause, and the next investigation moves into bun's Worker startup.
3. Add a one-shot [wrapper-dispatch] first dispatch_sync_main… log to both dispatch_sync_main and dispatch_sync_main_void. If the heartbeat ticks but this never appears, the Worker is alive in some sense but never reaches the point of creating its first X11 window / WebView. If it does appear and then the heartbeat stops, the bug is in whatever the first dispatched callback does (likely XOpenDisplay / webkit_web_view_new under the QEMU GL backend).

This is intentionally a diagnostic patch, not a "swallow the symptom" patch. The brief's premise that the bug is in "wrapper code after gtk_main()" was not supported by reading the source. Before we can write a real fix we need the next-run logs to say which of (1)/(2)/(3) it actually is.

Building the patched wrapper

The wrapper is a single C++ TU compiled with g++ against pkg-config-resolved webkit2gtk-4.1 gtk+-3.0 (and optionally ayatana-appindicator3-0.1). It does not require the full Electrobun build pipeline (no CEF, no Dawn/WGPU, no Zig launcher rebuild) when all we are replacing is libNativeWrapper.so in an already-built bundle.

Host prerequisites (Ubuntu 24.04)

sudo apt-get install -y \
    libwebkit2gtk-4.1-dev libgtk-3-dev \
    libayatana-appindicator3-dev \
    pkg-config g++

Build command (linux-x64)

cd eliza/upstreams/electrobun/package

mkdir -p src/native/linux/build src/native/build

PKG_CFLAGS=$(pkg-config --cflags webkit2gtk-4.1 gtk+-3.0 ayatana-appindicator3-0.1)
PKG_LIBS=$(pkg-config --libs   webkit2gtk-4.1 gtk+-3.0 ayatana-appindicator3-0.1)

g++ -c -std=c++20 -fPIC $PKG_CFLAGS \
    -I vendors/cef \
    -o src/native/linux/build/nativeWrapper.o \
       src/native/linux/nativeWrapper.cpp

g++ -shared \
    -o src/native/build/libNativeWrapper.so \
       src/native/linux/build/nativeWrapper.o \
       vendors/zig-asar/libasar.so \
       $PKG_LIBS \
       -ldl -lpthread

If ayatana-appindicator3-0.1 is unavailable, drop it from both pkg-config calls and add -DNO_APPINDICATOR to the compile step (matches upstream's build.ts lines 1934–1955 fallback).

If vendors/cef / vendors/zig-asar are absent (a fresh submodule clone has neither, they are downloaded by bun build.ts), the cheapest way to obtain them is one-time:

cd eliza/upstreams/electrobun/package && bun install && bun build.ts

…which produces dist/libNativeWrapper.so directly. The two-step manual g++ above is only useful for fast iteration on the wrapper itself.

Installing the patched wrapper into the bundle

cp src/native/build/libNativeWrapper.so \
   <repo-root>/packages/os/linux/elizaos/artifacts/amd64/elizaos-app/bin/libNativeWrapper.so

Then rebuild the ISO via packages/os/linux/elizaos/'s normal pipeline (build.sh / Makefile) so the new wrapper is staged onto the live image.

riscv64

The same wrapper TU has to be cross-compiled for riscv64-linux-gnu. The constraints:

• GTK3 + WebKit2GTK on riscv64 Debian. apt-cache search on a Debian unstable riscv64 chroot lists libgtk-3-dev and libwebkit2gtk-4.1-dev, but the latter has historically lagged on riscv64 (JIT support gaps in JavaScriptCore on rv64). The pragmatic path is to build in a riscv64 sysroot (debootstrap sid on riscv64) and link against whatever pkg-config --libs webkit2gtk-4.1 resolves to inside that sysroot.
• Cross toolchain. riscv64-linux-gnu-g++ (cross from amd64 host) plus a riscv64 sysroot exported via --sysroot=… and PKG_CONFIG_SYSROOT_DIR=… PKG_CONFIG_PATH=…/usr/lib/pkgconfig.
• Reuse bun-riscv64. packages/app-core/scripts/bun-riscv64/ already cross-builds bun for riscv64 in a Debian-based Dockerfile; the cleanest integration is to add a sibling stage in that same Dockerfile that apt-get installs libwebkit2gtk-4.1-dev libgtk-3-dev and runs the same two-step g++ build above, producing artifacts/riscv64/libNativeWrapper.so for the rv64 ISO to consume.
• No CEF on riscv64. CEF doesn't publish rv64 binaries; the wrapper must be built without the CEF code paths (isCEFAvailable() is a runtime check that returns false when vendors/cef is missing). We already only ship the GTK-only libNativeWrapper.so in the amd64 bundle, so this is consistent.

The Zig launcher in package/src/extractor/ is already cross-targetable via Zig's built-in riscv64-linux-gnu target — that part is the same story as the bun-riscv64 pipeline, no new infrastructure needed.