Problem: First `npx ruflo init` is slow and the npx cache grows to ~2GB (disk-sensitive setups) —— Trove problem library

What you will run into

The first npx ruflo init takes about five minutes, and disk usage jumps — ~/.npm/_npx grows past 1.2GB, roughly 2GB-class together with the global cache. ruflo itself is not that large: this is the npx -y cache mechanic plus native prebuild downloads stacking up.

Why this happens

Conclusion first: slow comes from the first download; large comes from the double npx cache.

Chain: npx -y ruflo ... -> pull the package from the registry -> download its native prebuilds -> everything lands in the ~/.npm/_npx cache. The amplifier: npx -y unpacks the package into the cache on every run (one copy per isolated HOME and one global: measured 1.2GB + 775MB on 2026-08-29), with native prebuild downloads dominating (baseline ~5 minutes / ~1.6GB on 2026-08-27).

Measured (2026-08-29,）: npmmirror init succeeded (v3.38.20) with near-instant downloads — the deterministic part of the slowness disappears with a mirror; the cache size remains.

Conclusion: slow -> switch to npmmirror (deterministic win); large -> clear the cache after verification (rm -rf ~/.npm/_npx, existing projects unaffected). Use both on disk-sensitive setups (CI/containers/small VPS).

Step by step

1. Install via a mirror — Required

Re-run init with the npm_config_registry=https://registry.npmmirror.com prefix; measured from ~5 minutes down to seconds.

2. Shrink the download surface — Required

--no-signup --no-skills-sh --no-codex-detect skips prompts, signup, and the full skills pull (install the 267 skills on demand later).

3. Clear the cache — By scenario

Disk-sensitive: rm -rf ~/.npm/_npx or npm cache clean --force; cache only, no effect on installed projects.

4. Check the write surface — Diagnostic

init writes workspace .claude/, .claude-flow/, .swarm/ and so on plus HOME ~/.claude/CLAUDE.md; --help starts a background daemon, so run ruflo daemon stop after testing.

How to confirm it works

• init completes quickly (near-instant downloads after switching the mirror);
• ruflo --version prints the version;
• du -sh ~/.npm/_npx is within your acceptable range (drops back after clearing).

Other environments

• CI / containers: cache size directly affects image layers; clear on install or mount a cache volume;
• Small VPS: reserve for 2GB-class usage;
• Source material: measurements in this market (2026-08-27 baseline + 2026-08-29 npmmirror/write-surface/cache trio).

Symptom

npx ruflo init takes about 5 minutes on first run and disk usage visibly grows.

Cause

Native prebuild downloads plus the double npx cache: measured under an isolated HOME on 2026-08-29, ~/.npm/_npx was 1.2GB + 775MB global (roughly 2GB-class); npx -y caches every run.

Solution

1. Speed up in China: npm_config_registry=https://registry.npmmirror.com npx -y ruflo ... (measured init downloads completing near-instantly).
2. Disk-sensitive: clear the npx cache after verification (rm -rf ~/.npm/_npx or npm cache clean --force); existing projects are unaffected.
3. Use --no-signup --no-skills-sh --no-codex-detect for init to skip prompts, signup, and the full skills pull, shrinking the download surface.