Silo v0.4.0 release notes

Silo v0.4.0 is a consolidation release. Over the last several months the codebase carried two implementations (Go and Rust), two sets of build/setup verbs, and a status command that tried to do three unrelated jobs. v0.4.0 collapses all of that.

Go is now the only implementation

The Rust implementation is gone. Everything ships from the Go tree. Install-time footprint, test matrix, and CI complexity all go down; the CLI surface stays identical.

CLI reshape: one model per command

The install verbs used to overlap. silo install installed; silo setup installed packages into a persisted rootfs; silo rebuild re-ran the stored setup; silo pull reconciled. Users had to remember which verb did which.

v0.4.0 picks one model per job and keeps the old names as deprecated aliases (removed in 0.6.0):

• silo build <tool> -- <cmd> — persist dependencies into a rootfs layer. Absorbs the old setup and rebuild commands. --rerun, --all, --remove are flags on the same verb.
• silo sync — reconcile the environment to .siloconf. Installs missing tools and warms the rootfs cache. silo pull / silo apply still work as aliases.
• silo doctor — runtime readiness only (kernel, initfs, bootstrap state). Split out of the old status.
• silo current — effective tool config after .siloconf overrides. Also split out of status.

silo status is still there; it prints both doctor and current output for users who had it in muscle memory.

Install now refuses duplicates

silo install python used to silently re-install on top of an existing tool. That led to confusion when users wanted “use a different Python version for this project” and got “globally replace Python.” v0.4.0 refuses the second install and points at two alternatives:

• silo install python --force — explicit global reinstall.
• silo use python@3.11 — pyenv-style project pin in .siloconf, doesn’t touch the global install.

Disk reclamation

Three features landed together:

• LRU + age-based GC runs once per process at the top of silo run. Cold entries evict when the cache exceeds the policy cap (default 8 GiB) or age limit (default 60 days). You can see what would happen with silo cache gc --dry-run.
• zstd cold tier. silo cache compress demotes entries to zstd (~4× smaller). On the next cache hit, they decompress back to raw (~1–3 s for 500 MB) and the fast clonefile path resumes.
• Image deletion on uninstall. silo uninstall python now frees the rootfs cache entry and, if no other tool shares the image, removes the OCI image and orphan layer blobs.

Policy lives in .siloconf:

cache:
  rootfs:
    maxSizeMB: 8192
    maxAgeDays: 60
  tools:
    maxSizeMB: 4096
    maxAgeDays: 30
    perMount:
      rust/cargo: 8192

silo cache report summarises the ~/.silo/ footprint by bucket.

Project lifecycle

silo sync and silo clean replace an ad-hoc “nuke ~/.silo/ when things feel weird” workflow:

• silo sync reads .siloconf, installs anything missing, and warms the rootfs cache. Safe to re-run.
• silo clean reclaims rootfs cache + per-tool caches + stale VMs for the current project. --rootfs-only, --caches-only, --force narrow or widen.

Network proxy with allowlist

Per-tool outbound network can now be restricted to an explicit domain allowlist, with wildcard support:

overrides:
  node:
    network:
      hostAccess: true
      proxy:
        allow:
          - registry.npmjs.org
          - "*.github.com"
        deny:
          - "*"

This is the default posture for claude-code installs — api.anthropic.com, github.com, registry.npmjs.org, pypi.org, nothing else.

Two-phase SIGINT

silo run now intercepts Ctrl-C twice. First press forwards SIGINT into the VM (graceful shutdown). Second press stops the container and deletes it. Previously a stuck container left a corrupt silo-* directory under ~/.silo/containers/ that required manual cleanup; the reaper at the top of the next silo run also catches these, but you rarely need it now.

Smaller quality-of-life wins

• Bootstrap downloads show a real progress bar with size hints (curl + parsing).
• silo init auto-detects tools from marker files (package.json, requirements.txt, Cargo.toml, go.mod, deno.json).
• Shimless silo install --image discovers executables by exec’ing into the image.
• silo cache clean --safe removes only orphans not referenced by any installed tool.

What’s next

The two priorities for 0.5:

1. Bootstrap speedup — the ~5-minute first install is the biggest UX cliff. Prebuilt kernel + initfs artifacts should cut it to ~30 seconds.
2. Homebrew distribution — brew install rchekalov/silo/silo now works. (The fully-qualified user/tap/formula form is required because an unrelated silo cask already exists in homebrew-cask.) Release workflow bumps the formula on every tag.

Full changelog on GitHub. Bug reports welcome.