Files
ihasmail/.gitlab-ci.yml
jcoffey-dev 795fe43cec Fetch the registry token from the public address, not the runner's
The builder on the host's network (the last change here) didn't help: the
next publish failed exactly as before. Looking on the host showed why.
Both builders resolve git.coffeylabs.org publicly; the token isn't fetched
by the builder at all. buildx fetches registry tokens on the client side,
in the job container, and on ci-net the name git.coffeylabs.org belongs to
the gitlab container itself (172.30.0.2) -- which is how the runner clones
over plain HTTP, and which has nothing on 443. So every push asked
https://git.coffeylabs.org/jwt/auth for a token and was refused. The login
before it worked because the host's daemon does the login, and the host
resolves the name publicly.

For the publish job only, the name now points at its public address in the
job's /etc/hosts, looked up from a public resolver, as the host sees it.
/etc/hosts wins over Docker's DNS, and nothing else in the job is affected:
the checkout is done, and image layers go to the registry's own DNS-only
name, not this one. The builder goes back to the shared ci-builder; its
network was never the problem.

The lookup and the /etc/hosts write were tried in the job's own image
(docker:28-cli, same digest): it picks the first public IPv4 address and
getent then returns it.
2026-09-21 16:48:47 -07:00

251 lines
13 KiB
YAML

# CI for the self-hosted GitLab that replaced GitHub Actions when the account
# was suspended on 2026-09-20. This is a port of .github/workflows/ci.yml and
# publish.yml, which are kept in the tree for reference and for the day the
# appeal succeeds.
#
# Every `image:` here is pinned to a digest, with the tag it belonged to in the
# trailing comment. That is the direct replacement for the SHA-pinned `uses:`
# in the Actions workflows: GitLab has no equivalent of an action allowlist, so
# the only thing standing between this pipeline and whatever the publisher
# pushes to a tag next is the digest. Read the comment for the version; the
# digest is what runs. Do not "simplify" one back to a bare tag.
#
# The runner is a group runner on Web_Host with the host docker socket bound
# in, reached over the internal container network rather than
# https://git.coffeylabs.org -- that name is Cloudflare-proxied on the Free
# plan, which caps request bodies at 100 MB and would break artifact uploads.
stages: [test, build, publish, release]
variables:
# Jobs talk to the registry directly on its DNS-only name, never through the
# proxy, for the same 100 MB reason.
IMAGE: $CI_REGISTRY_IMAGE
GIT_DEPTH: "0"
default:
interruptible: true
# ---------------------------------------------------------------- test ------
node:
stage: test
image: node:26-bookworm-slim@sha256:582460f614631b59b824ac6020533b9bf339c7fdf3a6d7db31abb6b4065f0212 # 26-bookworm-slim
variables:
NPM_CONFIG_CACHE: "$CI_PROJECT_DIR/.npm"
cache:
key:
files: [package-lock.json]
paths: [.npm/]
before_script:
# version.test.ts shells out to git to resolve a build version, and the
# slim image ships without it. The clone is done by the runner's helper
# image, so nothing else here needs git and its absence is easy to miss.
- apt-get update -qq && apt-get install -y -qq --no-install-recommends git
# config.test.ts chmods a directory to 0555 and expects the write to be
# refused. Root ignores the permission bits, so as root that assertion can
# never hold. The tests run as the image's unprivileged `node` user for
# that reason; -p keeps the environment.
#
# imageproxy.test.ts needs IPv6 as well, which is not set here but on the
# runner: jobs run on the `ci-net` docker network, created with --ipv6.
# Without a non-loopback IPv6 address on the container, getaddrinfo's
# AI_ADDRCONFIG drops ::1 from the results entirely, localhost resolves to
# IPv4 only, and the test's control case connects to a port nothing is
# listening on. That is a runner property, so it cannot be fixed from this
# file -- if these tests ever fail again with ECONNREFUSED on 127.0.0.1,
# check that the runner still puts jobs on an IPv6-enabled network.
- chown -R node:node "$CI_PROJECT_DIR"
script:
- su node -p -c "npm ci --ignore-scripts"
- su node -p -c "npm run typecheck"
- su node -p -c "npm test"
- su node -p -c "npm run build"
artifacts:
paths: [dist/]
expire_in: 1 week
rules:
- if: $RELEASE_WEEKLY == "1"
when: never
- if: $CI_PIPELINE_SOURCE == "merge_request_event"
- if: $CI_COMMIT_BRANCH == $CI_DEFAULT_BRANCH
- if: $CI_COMMIT_TAG
# --------------------------------------------------------------- build ------
# Proves the Dockerfile still builds on every change, without pushing. The
# equivalent of ci.yml's final `docker build -t ihasmail:ci .` step.
#
# Not called `image`: that is a reserved keyword, and a job by that name is
# silently read as the global image: setting instead ("image name should be a
# string"). Same trap for `stages`, `cache`, `services` and `variables`.
docker-build:
stage: build
image: docker:28-cli@sha256:625d9431a9f54c5a2bc90f24f0e1c3d55b1349fd857dd85035f98c2c9acbdd4d # 28-cli
needs: [node]
script:
- docker build -t ihasmail:ci-$CI_COMMIT_SHORT_SHA .
- docker image rm ihasmail:ci-$CI_COMMIT_SHORT_SHA
rules:
- if: $RELEASE_WEEKLY == "1"
when: never
- if: $CI_PIPELINE_SOURCE == "merge_request_event"
- if: $CI_COMMIT_BRANCH == $CI_DEFAULT_BRANCH
# ------------------------------------------------------------- publish ------
# Tag-driven, replacing the release -> publish workflow_call chain. GitHub
# needed that dance because a release cut with GITHUB_TOKEN raises no event;
# GitLab has no such rule, so a tag pipeline is enough.
#
# arm64 is built under QEMU on this amd64 host, not on a native runner as
# GitHub's free `ubuntu-24.04-arm` did. It is slow -- tens of minutes for the
# npm install and Vite build through instruction translation -- which is
# tolerable for a weekly tag and would not be for every push. That is why this
# job is tag-only. If arm64 ever starts timing out, the fix is an arm64 runner,
# not dropping the platform: TrueNAS and Unraid users pull it.
# The version the image is built with, computed the way publish.yml did it:
# scripts/version.mjs, which needs node and the full history. The build is
# *told* the real form (IHASMAIL_VERSION, what About and /api/health report);
# the Docker tag gets the same string with '+' turned into '-', because a tag
# may not contain '+'. The first port of this job left the build arg out, so
# a tag would have shipped an image reporting itself unversioned -- which is
# exactly what version.test.ts calls looking wrong.
version:
stage: build
image: node:26-bookworm-slim@sha256:582460f614631b59b824ac6020533b9bf339c7fdf3a6d7db31abb6b4065f0212 # 26-bookworm-slim
variables:
GIT_DEPTH: "0"
before_script:
- apt-get update -qq && apt-get install -y -qq --no-install-recommends git >/dev/null
# The build directory is reused between jobs, and the node job chowns it to
# the unprivileged `node` user so its tests can run. A later job running
# git as root then finds the checkout owned by somebody else, and git
# refuses with "detected dubious ownership" (exit 128). Whether it happens
# depends on which cached directory a job lands on, so it comes and goes.
- git config --global --add safe.directory "$CI_PROJECT_DIR"
script:
- V="$(node scripts/version.mjs)"
- echo "VERSION=$V" > version.env
- echo "DOCKER_TAG=${V/+/-}" >> version.env
- cat version.env
artifacts:
reports:
dotenv: version.env
rules:
- if: $CI_COMMIT_TAG
publish:
stage: publish
image: docker:28-cli@sha256:625d9431a9f54c5a2bc90f24f0e1c3d55b1349fd857dd85035f98c2c9acbdd4d # 28-cli
needs: [node, version]
variables:
DOCKER_BUILDKIT: "1"
before_script:
- echo "$CI_REGISTRY_PASSWORD" | docker login -u "$CI_REGISTRY_USER" --password-stdin "$CI_REGISTRY"
- docker run --privileged --rm tonistiigi/binfmt --install arm64
# The registry hands out push tokens from https://git.coffeylabs.org/jwt/auth,
# and buildx fetches them here, in the job, not in its builder. On ci-net
# that name is the gitlab container itself (172.30.0.2), which serves
# plain HTTP to the runner and nothing on 443, so every push failed at the
# last step with "connection refused". The login above works because the
# host's daemon does it, and the host resolves the name publicly. So, for
# this job only, point the name at its public address the same way. Only
# the token request uses it; layers go to the registry's own DNS-only name.
- |
public="$(nslookup "$CI_SERVER_HOST" 1.1.1.1 2>/dev/null | awk '/^Address: / && $2 !~ /:/ { print $2; exit }')"
if [ -z "$public" ]; then echo "Could not resolve $CI_SERVER_HOST publicly" >&2; exit 1; fi
echo "$public $CI_SERVER_HOST" >> /etc/hosts
echo "$CI_SERVER_HOST -> $public for the registry token"
- docker buildx create --use --name ci-builder --driver docker-container || docker buildx use ci-builder
script:
- |
docker buildx build \
--platform linux/amd64,linux/arm64 \
--build-arg IHASMAIL_VERSION="$VERSION" \
--provenance=false --sbom=false \
--tag "$IMAGE:$DOCKER_TAG" \
--tag "$IMAGE:latest" \
--push .
after_script:
- docker logout "$CI_REGISTRY" || true
rules:
- if: $CI_COMMIT_TAG
# ----------------------------------------------------------- weekly release --
# Port of .github/workflows/release.yml: cut a release once a week, but only
# when there is something in it. The decision is the workflow's, unchanged --
# count the commits on main since the newest published release, and skip the
# week if there are none or if the tag already exists (the version comes from
# the commit, so an unchanged commit is an existing tag).
#
# It runs from a pipeline schedule (Mondays 09:17 UTC, the same odd minute as
# before) that sets RELEASE_WEEKLY=1. GitLab keeps schedules on the project,
# not in this file, so the schedule and this job only work as a pair. Run it by
# hand with RELEASE_WEEKLY=1, adding DRY_RUN=1 to see the decision and stop.
#
# The release -- and with it the tag -- is created with RELEASE_TOKEN, a
# project access token (protected, masked), not CI_JOB_TOKEN. A tag pushed that
# way is an ordinary push, so it starts the tag pipeline, and the version and
# publish jobs above build the image from it. That replaces release.yml's
# direct call of publish.yml, which only existed because a tag created with
# GITHUB_TOKEN raises no event. The token expires; when it does this job fails
# at the API call, loudly, and a new one goes in the same variable.
weekly-release:
stage: release
image: node:26-bookworm-slim@sha256:582460f614631b59b824ac6020533b9bf339c7fdf3a6d7db31abb6b4065f0212 # 26-bookworm-slim
# One at a time: two overlapping runs would race to create the same tag.
resource_group: weekly-release
variables:
GIT_DEPTH: "0"
before_script:
- apt-get update -qq && apt-get install -y -qq --no-install-recommends git curl jq >/dev/null
# See the version job: same shared directory, same root, same refusal.
- git config --global --add safe.directory "$CI_PROJECT_DIR"
script:
- |
set -euo pipefail
# Internal address, as for everything else CI does: never through the proxy.
API="http://gitlab/api/v4/projects/${CI_PROJECT_ID}"
auth=(--header "PRIVATE-TOKEN: ${RELEASE_TOKEN}")
# The newest published release, or empty on a project that has never had
# one -- in which case everything counts as new.
previous="$(curl -fsS "${auth[@]}" "${API}/releases?order_by=released_at&sort=desc&per_page=1" | jq -r '.[0].tag_name // ""')"
# A release can outlive its tag. Falling back to the whole history
# over-counts, which cuts a release that was due anyway; under-counting
# would skip one that was.
# Tag lookups use show-ref, which matches an exact ref and nothing else.
# `rev-parse --verify refs/tags/<name>` does not: on the git in this image
# (2.39) a name ending in -g<hex> falls back to being read as
# git-describe output, resolves to that commit, and so "exists" whether
# or not the tag does. Every commit not merged through a pull request has
# a -g<hex> version, so that check reported every such week as already
# released. Newer git (and GitHub's runners) do not fall back, which is
# why release.yml never showed it.
if [ -n "$previous" ] && git show-ref --verify --quiet "refs/tags/${previous}"; then
count="$(git rev-list --count "${previous}..HEAD")"; range="${previous}..HEAD"
else
count="$(git rev-list --count HEAD)"; range="HEAD"
fi
version="$(node scripts/version.mjs)"
# A Docker tag may not contain '+', and neither should the git tag, so
# the two always agree about what to call a build.
tag="v${version/+/-}"
title="v${version%%+*}"
sha="$(git rev-parse HEAD)"
if [ "$count" -eq 0 ]; then
echo "Nothing to release: no commits since ${previous}."; exit 0
fi
if git show-ref --verify --quiet "refs/tags/${tag}"; then
echo "Nothing to release: tag ${tag} already exists."; exit 0
fi
echo "Releasing ${tag} -- ${count} commit(s) since ${previous:-the beginning}, at ${sha}."
if [ "${DRY_RUN:-0}" = "1" ]; then echo "DRY_RUN=1: stopping here."; exit 0; fi
# Notes bounded to what is new, from the first-parent history of main --
# one line per merge, which is what GitHub's generated notes listed.
notes="$(git log --first-parent --format='- %s' "$range")"
jq -n --arg tag "$tag" --arg ref "$sha" --arg name "$title" \
--arg desc "$(printf '%s commit(s) since %s.\n\n%s' "$count" "${previous:-the beginning}" "$notes")" \
'{tag_name:$tag, ref:$ref, name:$name, description:$desc}' > release.json
curl -fsS "${auth[@]}" --header "Content-Type: application/json" \
--data @release.json "${API}/releases" | jq -r '"created release " + .tag_name'
rules:
- if: $RELEASE_WEEKLY == "1" && $CI_COMMIT_BRANCH == $CI_DEFAULT_BRANCH