version: "1.0.0" name: setup-ci-host description: Prepare a new physical CI host so /data/miles_ci is canonical — a real directory on the biggest disk, or a symlink to it — before installing miles GitHub Actions runners. Triggers, "setup ci host", "/data/miles_ci not ready", "miles ci host bootstrap", or similar. Idempotent; safe to re-run. Run once per host before either the README's docker-compose runner flow or manage-gh-runners. Boundary vs manage-gh-runners, this skill only prepares the host filesystem layout (big-disk detection, mkdir, symlink); it does NOT add, remove, or talk to GitHub runners. user_invocable: true

Setup CI Host

Prepares a new physical host to run miles GitHub Actions CI runners by making /data/miles_ci canonical on the host.

The miles CI workflow (.github/workflows/_run-ci.yml) bind-mounts /data/miles_ci and its subdirectories (models, datasets, hf_cache) from the host into every CI job container, using literal paths everywhere. For that to work, every CI host must provide /data/miles_ci either as a real directory on its biggest disk OR as a symlink to wherever the real big disk is. This script automates that.

Run this once on each new CI host, before the runner-setup steps in `tests/ci/README.md` or the externally-managed flow in `manage-gh-runners`.

Usage

cd /root/miles/.claude/skills/setup-ci-host
./setup-host.sh

Re-runs are safe: the script is idempotent and a no-op on hosts that already have /data/miles_ci correctly set up.

What it does

1. Probes mounted filesystems with df and picks the mount point with the

most total bytes, excluding tmpfs/devtmpfs/overlay/squashfs and system mounts (/, /boot, /dev, /proc, /run, /sys).

1. If the biggest mount is `/data` (the canonical case, e.g. scitix-72):

ensures /data/miles_ci exists as a real directory. Done.

1. If the biggest mount is something else (e.g. /mnt/nvme0n1 on novita):

ensures <big_disk>/miles_ci exists as a real directory, then atomically replaces /data/miles_ci with a symlink pointing to <big_disk>/miles_ci (creates /data first if missing).

1. Creates /data/miles_ci/models, /data/miles_ci/datasets,

/data/miles_ci/hf_cache if they do not exist.

1. Prints a summary: chosen mount, whether /data/miles_ci is a real dir or

a symlink, free space on the backing disk.

Verification

After running, the operator can confirm the result with:

ls -la /data/miles_ci
df -h /data/miles_ci
readlink /data/miles_ci || echo "(real directory, not a symlink)"

The first command should list the four subdirs (models, datasets, hf_cache, plus any existing runner_<hostname> work dirs). df -h should report on the host's big disk regardless of whether /data/miles_ci is real or a symlink. readlink succeeds only when /data/miles_ci is a symlink.

When to re-run

• After adding a larger disk to the host (the script will re-detect; if

/data/miles_ci already exists pointing at a smaller disk, the script prompts before moving anything — see Safety).

• After noticing CI mount errors that suggest /data/miles_ci is wrong.

Safety

The script never silently overwrites a /data/miles_ci that already exists and points somewhere unexpected. When run from a terminal, it prompts the operator interactively; in non-interactive contexts (cron, CI, scripts) it refuses and exits non-zero so a human always makes the destructive call.

Two prompts can appear:

1. Existing real directory on the wrong disk — /data/miles_ci has data

but the biggest disk is elsewhere. Options:

• [m] migrate: rsync the data to the big disk, then rm -rf the

original and symlink. Preserves data.

• [w] wipe: rm -rf the original (data lost), then symlink. Requires

typing yes to confirm.

• [a] abort (default): leave everything as-is, exit non-zero.

1. Existing symlink pointing at the wrong target — /data/miles_ci

already symlinks somewhere, but not to the current biggest disk. The prompt asks [y/N] to re-point. Re-pointing does NOT migrate data from the old target; only the symlink changes.

The script requires root (or write access to /data). It will detect a non-root invocation and suggest sudo ./setup-host.sh.

Inside the container

Once the host side is prepared, the runner container's compose mount /data/miles_ci:/data/miles_ci works as an identity bind (per the docker-out- of-docker requirement explained in docker-compose.yml). The inner job container sees /data/miles_ci as a real bind from the host's canonical path, whether or not the host side is itself a symlink — Docker resolves the host symlink at bind time, so the container always sees the underlying directory. This is the desired behavior.