Agent Built Lfs

agent-built-lfs

A [Linux From Scratch](https://www.linuxfromscratch.org/lfs/) system built and maintained by an agent. Claude Code reads the book, turns each page into a recipe, records a reasoned decision for every command it disables or rewrites, and drives the build. The system calls itself `claude-managed` in `/etc/os-release`, hence the name.

`server` is live: i5-2500K, GTX 770 on NVIDIA 470.xx, X11 and awesome, 133 LFS steps, 218 BLFS packages, self-hosting. `laptop` is scaffolded.

Why

I ran LFS as a daily driver for years, then had kids and lost the weekends it takes to babysit a build or fix a package that broke on upgrade. That was the only reason I stopped.

Claude turned out to be good at reading a book carefully -- not clever, careful, at length, without getting bored on page 300. That is most of what LFS demands. So I pointed it at the book, with every judgment call written down where I can audit it.

The agent reads, extracts, decides individual commands, writes the tooling. I make the calls that need someone who knows the hardware and lives with the machine: `server` runs X11 rather than Wayland after the NVIDIA 470.xx dead end.

So far so good. Check back after my first real rebuild -- the test is not whether an agent can follow a book once, but whether its record is good enough to upgrade a package a year later without breaking the machine.

The part I didn't expect

It is very good at hardware. Not the big calls, but the small tweaks that are hard to find, that a general-purpose distro cannot ship, and that nobody would chase down by hand for one machine.

`defconfig` pinned every core of this i5-2500K at 1600 MHz, defaulting to a governor that waits for a process that does not exist: a measured 2.1x loss presenting as nothing but "the machine feels slow". The BIOS misreported its audio codec slots, leaving no sound card until a boot parameter force-probed them. Limitations too -- this Kepler card is stuck on NVIDIA's legacy 470.xx bran