mirror of
https://github.com/stablyai/orca.git
synced 2026-10-27 00:02:48 +00:00
On Windows 11 build 26200 the sandboxed renderer breakpoints (STATUS_BREAKPOINT / 0x80000003, exit -2147483645) ~2s after launch on nearly every start. The existing GPU-child-crash fallback only disables hardware acceleration and cannot fix a renderer *sandbox* fault, so the renderer-recovery circuit breaker reloads the same crashing renderer 3x and dead-ends at "Orca keeps failing to load". Add a build-scoped renderer sandbox fallback mirroring the GPU fallback marker system: after 3 launch-window renderer STATUS_BREAKPOINT crashes on win32, persist a marker and relaunch with the top-level renderer created unsandboxed (webPreferences.sandbox:false) - the surgical, per-window equivalent of the reporter-confirmed --no-sandbox, keeping GPU/utility/network and webview-guest sandboxes intact. The marker is scoped to appVersion+electronVersion and auto-clears on any update, so a fixed Electron build regains the sandbox automatically. Detection keys strictly on exit code -2147483645 so renderer OOM (#9872, -36861) and other crashes never trigger the downgrade; the dashboard pop-out honors the same flag so it does not breakpoint either. Worst case (sandbox:false still crashes) degrades to today's dialog - never an infinite relaunch loop. The fallback depends on per-window sandbox config: it is documented and guarded (a durable renderer_sandbox_fallback_neutralized breadcrumb) that adding a global app.enableSandbox() would ignore the per-window opt-out and neutralize the fix.