CI: Build CLI step on Windows takes ~5 min due to redundant NAPI rebuild

[fengmk2]

Problem

The Build CLI step in the CLI E2E test (windows-latest) job takes ~5 minutes, significantly slower than Linux (~2 min) and macOS (~1 min).

Example run: https://github.com/voidzero-dev/vite-plus/actions/runs/22795306880/job/66128870036

Step timing comparison (CLI E2E test)

Step	Windows	Ubuntu	macOS
setup-rust	1m 5s	29s	11s
setup-node	1m 44s	25s	15s
Build with upstream	1m 12s (cache hit)	44s	16s
Build CLI	5m 12s	1m 49s	58s
Total job	~14 min	~8 min	~4 min

Root Cause Analysis

1. Redundant NAPI Rust compilation (primary bottleneck)

The Build CLI step runs pnpm build, which calls:

pnpm -F vite-plus-cli build  →  oxnode -C dev ./build.ts

build.ts runs buildNapiBinding() by default (no --skip-native flag), triggering a full cargo build --release of the NAPI binding — even though Build with upstream already compiled the same NAPI bindings (with caching via actions/cache).

The Build with upstream action correctly uses build-ts (which passes --skip-native) for the TypeScript parts and separately handles native builds with caching. But the subsequent pnpm build in Build CLI re-compiles everything.

Cargo release build of 9 workspace crates (same code, all NAPI cache hit):

Platform	Time	Relative
macOS (Apple Silicon)	49s	1×
Ubuntu (x86_64)	1m 24s	1.7×
Windows (x86_64, MSVC)	3m 10s	3.9×

2. Windows MSVC is inherently slower for Rust compilation

• MSVC linker (link.exe) is single-threaded and slower than lld/macOS linker (PDB generation overhead)
• NTFS file I/O has higher per-operation latency than ext4/APFS for the many small files cargo generates
• Windows Defender real-time scanning on GitHub Actions runners scans every file read/write, adding 30-50% overhead
• Process spawning — CreateProcess is ~10× slower than Unix fork/exec; cargo spawns many rustc processes

3. Slower toolchain setup on Windows

• setup-node (pnpm install): 1m 44s vs 25s on Linux (NTFS I/O + antivirus overhead)
• setup-rust: 1m 5s vs 29s on Linux

Proposed Solutions

1. Skip native rebuild in Build CLI step (quick fix, biggest impact)

Change the Build CLI step from:

- name: Build CLI
  run: |
    pnpm build
    pnpm bootstrap-cli:ci

To:

- name: Build CLI
  run: |
    pnpm build-ts-only   # or: pnpm -F @voidzero-dev/* -F vite-plus build && pnpm -F vite-plus-cli build-ts
    pnpm bootstrap-cli:ci

Since Build with upstream already handles the NAPI bindings with proper caching, the Build CLI step only needs to do the TypeScript build (build-ts / --skip-native).

Expected savings: ~3 min on Windows, ~1-2 min on Ubuntu

2. Disable Windows Defender real-time scanning

Add this step before any build steps on Windows:

- name: Disable Windows Defender real-time monitoring
  if: runner.os == 'Windows'
  run: Set-ExecutionPolicy Bypass -Scope Process -Force; Set-MpPreference -DisableRealtimeMonitoring $true
  shell: powershell

GitHub-hosted Windows runners allow this without admin elevation. This can reduce I/O-heavy operations (cargo, pnpm install) by 30-50%.

3. Additional optimizations (lower priority)

• Add sccache for Windows Rust builds when NAPI cache misses
• Consider cargo build --profile ci: a custom profile with fewer optimizations for faster compile time in CI

[linear]

VP-230