lighter 0.2.0 — the network, the memory, and a second full record.
The network is now streams over vsock joined in the guest kernel: a container's TCP socket and its vsock stream are one socket pair, so bytes cross the boundary once, and UDP rides the same path. Memory goes back to the Mac three ways — free page reporting, the cache trims, and on guests of eight gigabytes or more a balloon the guest itself offers into — and an idle machine costs a third less CPU than 0.1.0 did. The file server's apply lane is shaped per machine: creates on directory-keyed lanes, and eight workers on a Mac of eight cores.
What is in the box: lighter (the CLI and daemon, Developer ID signed and notarized), gvproxy for what stays on the network device, a 6.18 kernel with twenty-two patches, and an Alpine rootfs with Docker.
Measured against OrbStack on the same machine, same fixtures, medians of three (benchmarks/RESULTS.md, generated from the CSVs in the repo; Colima and Docker Desktop in the same tables):
M5 Pro, 48 GB — ahead on every install and read case, share and own disk. On the share: npm ci 6.3 s vs 8.5, pnpm 3.7 vs 4.7, yarn 5.2 vs 7.8, ripgrep 79 ms vs 1020, find 87 vs 595, copy-tree 3.3 s vs 8.7, rm -rf 2.4 vs 3.0. Network: 98.5 Gbit/s out of a container vs 97.2, 90.1 into one vs 52.9, 97.7 into a published port vs 54.2, 17,600 connects a second vs 16,200, a GET on a published port 68 µs vs 73, DNS 55 µs vs 251. Memory a minute after an install: 862 MiB vs 2114. Idle: 5 ms/s and 98 wakeups a second vs 2 and 99.
M1, 8 GB — npm ci 11.2 s vs 11.5, pnpm 5.5 vs 5.9, yarn 10.5 vs 10.4, ripgrep 193 ms vs 1110, find 121 vs 552, copy-tree 5.3 s vs 15.2, rm -rf 3.5 vs 4.0; own disk ahead on six of seven. Network: 43.5 Gbit/s into a published port vs 29.9, DNS 141 µs vs 422, UDP 4.9 vs 3.1; behind on the two container-to-Mac paths (52.7 vs 63.6 and 54.5 vs 67.5) and on a GET's 133 µs vs 128. Settled memory on a 4 GiB guest trails OrbStack's; the worklog says why and why it is left so.
Requirements: Apple Silicon, macOS 15 or later, the Docker CLI.
Install: brew install fieldwork-ai/tap/lighter, or the one-line installer in the README.