Tags: bwendell/gemini-desktop
Tags
fix(release): pin electron-builder to 26.8.1 for arm64 NSIS validation electron-builder 26.15.2 (pulled in for v0.12.1) regressed the Windows arm64 NSIS installer: a silent install exits 0 but no longer places the app executable at the requested install directory, so the arm64 release validation fails deterministically while x64 passes. 26.8.1 is the version that shipped v0.12.0 successfully. Pin electron-builder to an exact 26.8.1 (no caret, so a fresh install can't resolve back up to 26.15.x) and regenerate the lockfile. The electron 39.8.10 bump and all other v0.12.1 dependency updates are kept. https://claude.ai/code/session_01AzweBBdfnoZPZHEGh2AjsB
fix(linux): keep --no-v8-sandbox for text prediction; only kernel-7 p…
…ath drops it
The previous commit removed the --no-v8-sandbox js-flag entirely, which broke
the Release E2E Linux assertions and left TextPredictionIpcHandler's restart
flow permanently gated ("restart required") on kernel <= 6, where text
prediction otherwise works.
Restore sandboxInit.ts (and its unit test) to the pre-existing behavior: append
--no-v8-sandbox when textPredictionEnabled or --test-text-prediction on Linux.
The kernel-7 V8-sandbox incompatibility is handled solely by never loading
dbus-next / node-llama-cpp (the guards added in the previous commit), so PR
#333's kernel-7 --no-v8-sandbox trigger stays removed.
https://claude.ai/code/session_01MyD8Lckez5j6b5mGXv4CaF
fix(linux): disable V8 sandbox on incompatible kernels (#119, #315, #334 ) The fatal v8_ArrayBuffer_NewBackingStore crash at startup is a V8-sandbox incompatibility on kernel 7.x+ / newer glibc, not a node-llama-cpp probe issue. It reproduces with text prediction disabled (the default), where the native module is never loaded, so gating --no-v8-sandbox solely on textPredictionEnabled left affected users crashing. Add a scoped detector (shouldDisableV8Sandbox / getKernelMajorVersion) that flags kernels with major version >= 7 and disable the V8 sandbox for them in sandboxInit, in addition to the existing text-prediction triggers. The flag is applied at most once. The V8 sandbox stays enabled on compatible kernels. Covered by unit tests for the detector (kernel parsing + platform gating) and sandboxInit (kernel-triggered disabling, dedupe, compatible-kernel no-op).
PreviousNext