dev-server

dev-server is the bash and Ansible that build my machines. ./devbox converge creates or reuses a Hetzner VPS, waits for Tailscale, locks down public SSH, runs Ansible, and runs a doctor; ./workstation converge does macOS and Arch. The README scopes it in one line: "fast, malleable coding machines, not a compliance framework." Two problems made it exist: a fresh cloud VM is exposed during bootstrap, and running work and personal subscriptions of the same AI CLI on one machine mixes credential state unless something separates it.

For the first: cloud-init opens SSH only from my current /32, enrolls Tailscale, deletes the auth key from disk, and enables UFW on first boot; converge then deletes the bootstrap firewall rule and verifies the closure — it attempts an SSH to the public IP and dies if the connection succeeds. The alternatives were leaving port 22 open or maintaining a bastion.

For the second: one router script, symlinked under each tool-and-account name, derives the tool and account from its own argv[0], glob-matches the target path against work-directory patterns, and execs the real binary with CODEX_HOME or CLAUDE_CONFIG_DIR pointed at that account's state. Separate per-tool router scripts existed; the refactor deleted them.

The third decision is a refusal. Converge installs desired state and does not clean unrelated drift; doctor runs positive checks and does not audit for absence; destroying and recreating the server is manual on purpose. Full configuration management was the alternative; the README's guardrails section names it as a non-goal.

The doctor is the only verification, and it checks resolved state rather than files: for OpenCode it runs opencode debug config and asserts the resolved model, limits, and permissions with jq, against a managed guardrail layer that disables session sharing, binds the server to localhost, denies reads and edits of env files and secrets, and makes every bash command ask.

Numbers, by wc -l over tracked files: 22 commits, 2026-05-13 to 2026-08-04; 4,854 lines, of which 1,691 is a vendored p10k.zsh; zero tests, zero CI.

What is wrong with it. Nothing verifies the scripts but the scripts. My SSH key, username, and personal project-dir globs are hardcoded — nobody else can run this without editing source, which is the stated scope but still a fact. Firewall idempotency substring-matches hcloud's error text and breaks silently if the wording changes. Claude installs by piping an unpinned remote script to bash. I wrote a 662-line walkthrough and later deleted it; the README is now the only document.

One user, three kinds of machine, one command each. The scope is the feature.