ci: prebuild the Arch environment as a container image

All six nightly jobs ran the same pacman transaction before doing any work.
The toolchain now lives in an image published to the Gitea registry as
git.bjphoster.com/aur/autoupdate:latest, rebuilt monthly since Arch is rolling
and a month-old image is a month behind.

The build job itself runs on plain ubuntu-latest -- it only drives docker. Only
the image is Arch.

Based on archlinux:base rather than base-devel: the autoupdater parses
PKGBUILDs and never compiles, so the toolchain is dead weight, 822 MB against
1.48 GB. nodejs is included because actions/checkout is a JavaScript action the
runner executes with node from inside the image.

Slimming to base initially dropped diffutils, and apply_pkgrel_rule used
diff -q. A missing binary makes the `if` fail, which reads as "changed", so
pkgrel ratcheted upward on every run -- the same silent-wrong-default class as
the empty .SRCINFO this rewrite was about, and invisible to the host test run.
The comparison is now a shell string equality with no external dependency, and
is verified against an image with no diff at all. require_tools() additionally
fails on startup for any missing binary rather than midway through.

.dockerignore

@@ -0,0 +1,4 @@
+# The image is a toolchain only -- the dockerfile COPYs nothing, and the
+# scripts arrive via actions/checkout at run time. Keep the build context empty.
+*
+!dockerfile

.gitea/workflows/build-image.yaml

@@ -0,0 +1,67 @@
+---
+name: Build Autoupdater Image
+
+concurrency:
+  group: build-image
+  cancel-in-progress: false
+
+on:
+  schedule:
+    # Monthly. Arch is a rolling release, so a month-old image is a month
+    # behind on packages; :latest is overwritten each time.
+    - cron: "0 3 1 * *"
+  push:
+    branches:
+      - main
+    paths:
+      - dockerfile
+      - .dockerignore
+      - .gitea/workflows/build-image.yaml
+  workflow_dispatch:
+
+defaults:
+  run:
+    shell: bash
+
+jobs:
+  build:
+    # No Arch needed here -- this only drives docker. The *image* is Arch.
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Log in to the Gitea container registry
+        env:
+          GIT_HOST: ${{ vars.GIT_HOST }}
+          REGISTRY_USER: ${{ vars.REGISTRY_USER }}
+          GIT_TOKEN: ${{ secrets.GIT_TOKEN }}
+        run: |
+          set -euo pipefail
+          # NB: this is the Gitea *username*, not vars.GIT_NAME -- that one is
+          # the display name used for git commit authorship and contains spaces.
+          printf '%s' "${GIT_TOKEN}" \
+            | docker login "${GIT_HOST}" \
+                --username "${REGISTRY_USER}" \
+                --password-stdin
+
+      - name: Build and push
+        env:
+          GIT_HOST: ${{ vars.GIT_HOST }}
+        run: |
+          set -euo pipefail
+          # The repo is `.autoupdate`, but OCI names cannot start with a dot.
+          image="${GIT_HOST}/aur/autoupdate"
+
+          # --pull so a monthly rebuild actually picks up a new archlinux base
+          # rather than reusing a cached one.
+          docker build --pull --tag "${image}:latest" .
+          docker push "${image}:latest"
+
+          echo "pushed ${image}:latest"
+
+      - name: Log out
+        if: always()
+        env:
+          GIT_HOST: ${{ vars.GIT_HOST }}
+        run: docker logout "${GIT_HOST}" || true

.gitea/workflows/update.yaml

@@ -37,8 +37,14 @@ jobs:
     # emits an empty .SRCINFO, which the AUR then rejects -- that is what broke
     # reflex-appimage, nethlink-appimage and pman-helper. regen_srcinfo() in
     # common.sh fails loudly if this ever regresses.
+    #
+    # Prebuilt by build-image.yaml (see dockerfile) so the toolchain is not
+    # reinstalled in all six jobs every night.
     container:
-      image: archlinux:base-devel
+      image: git.bjphoster.com/aur/autoupdate:latest
+      credentials:
+        username: ${{ vars.REGISTRY_USER }}
+        password: ${{ secrets.GIT_TOKEN }}
 
     strategy:
       # One broken package must not cancel the rest of the matrix.
@@ -53,15 +59,10 @@ jobs:
           - tsparams
 
     steps:
-      # Must come BEFORE actions/checkout. That action is JavaScript, and the
-      # archlinux image ships no node, so checkout dies with
-      # `exec: "node": executable file not found in $PATH`. A `run:` step needs
-      # no checkout, so the toolchain can be installed first.
-      - name: Bootstrap container
-        run: |
-          pacman -Syu --noconfirm --needed \
-            nodejs git openssh jq curl pacman-contrib namcap fuse2 python
-
+      # No bootstrap step: node, git, openssh, jq, curl and the makepkg
+      # toolchain all come from the prebuilt image above. node in particular
+      # has to be present before this action runs, since actions/checkout is
+      # JavaScript and the runner executes it with `node` from the image.
       - uses: actions/checkout@v4
 
       - name: Configure git and SSH
@@ -91,11 +92,11 @@ jobs:
           EOF
           chmod 600 ~/.ssh/config
 
-      # makepkg refuses to run as root, and CI containers are root by default.
-      - name: Create build user
+      # The `builder` user is baked into the image; this only hands it the
+      # workspace and the credentials written by the previous step.
+      - name: Hand over to the build user
         run: |
           set -euo pipefail
-          useradd --create-home --shell /bin/bash builder
           chown -R builder:builder "${GITHUB_WORKSPACE}"
           cp -r /root/.ssh /root/.gitconfig /root/.git-credentials /home/builder/ 2>/dev/null || true
           chown -R builder:builder /home/builder

README.md

@@ -103,9 +103,48 @@ The trap the previous implementation fell into was giving that same silent
 
 Needs a genuine Arch `makepkg`. Ubuntu's build environment emits an empty
 `.SRCINFO`, which the AUR rejects — this silently broke three packages. The CI
-job therefore runs in an `archlinux:base-devel` container as a non-root user
-(`makepkg` refuses to run as root), and `regen_srcinfo` in `common.sh` fails
-loudly if the output is ever empty again.
+job therefore runs in an Arch container as a non-root user (`makepkg` refuses to
+run as root), and `regen_srcinfo` in `common.sh` fails loudly if the output is
+ever empty again.
+
+`autoupdate.sh` calls `require_tools` on startup, so a missing binary fails
+immediately instead of taking a silently wrong branch further in.
+
+## Container image
+
+`dockerfile` builds the environment the nightly job runs in, published to the
+Gitea registry as `git.bjphoster.com/aur/autoupdate:latest`. Without it, all six
+matrix jobs would run a full `pacman` transaction every night.
+
+`.gitea/workflows/build-image.yaml` rebuilds and overwrites `:latest` monthly
+(Arch is rolling, so a month-old image is a month behind), on any push touching
+the `dockerfile`, and on demand via `workflow_dispatch`. That job runs on plain
+`ubuntu-latest` — it only drives `docker`; the *image* is the Arch part.
+
+Based on `archlinux:base`, not `base-devel`: the autoupdater only ever parses
+PKGBUILDs and never compiles, so the toolchain is dead weight — 822 MB against
+1.48 GB. Add it back if a real build or verify step is ever introduced.
+
+Building locally:
+
+```sh
+docker build -t git.bjphoster.com/aur/autoupdate:latest .
+docker run --rm -v "$PWD:/src:ro" git.bjphoster.com/aur/autoupdate:latest \
+  bash -c 'cp -r /src /w && chown -R builder:builder /w && cd /w &&
+           runuser -u builder -- ./test.sh'
+```
+
+Running the suite inside the image is worth doing after any `dockerfile` change.
+Trimming to `archlinux:base` originally dropped `diffutils`, and the missing
+`diff` made every run look like a packaging change and ratchet `pkgrel` upward —
+the host tests passed throughout. `pkgrel` no longer depends on `diff` at all,
+but the lesson stands.
+
+### Registry credentials
+
+Needs `vars.REGISTRY_USER` — your Gitea **username**, not `vars.GIT_NAME`, which
+is the display name used for commit authorship and contains spaces. The token is
+the existing `secrets.GIT_TOKEN`, which needs `write:package` scope.
 
 ## Notes on checksums
 

autoupdate.sh

@@ -24,6 +24,7 @@ _here="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 source "${_here}/common.sh"
 
 require_noninteractive
+require_tools
 
 GIT_HOST="${GIT_HOST:-git.bjphoster.com}"
 # Where the package repos live, and how the upstream version is resolved.
@@ -191,8 +192,15 @@ apply_pkgrel_rule() {
   # packaging content. An edit to a local source lands here too, because its
   # sha256sums entry was just regenerated. A README-only edit does not touch
   # the PKGBUILD and so correctly does not bump.
-  if diff -q <(grep -v '^pkgrel=' PKGBUILD) \
-             <(grep -v '^pkgrel=' "$aur_pkgbuild") >/dev/null 2>&1; then
+  #
+  # Deliberately a shell string comparison rather than diff(1): diff lives in
+  # diffutils, which archlinux:base does not ship, and a missing binary makes
+  # the `if` fail, which reads as "changed" and ratchets pkgrel upward on every
+  # single run. Same silent-wrong-default class of bug as the empty .SRCINFO.
+  local local_body aur_body
+  local_body="$(grep -v '^pkgrel=' PKGBUILD)"
+  aur_body="$(grep -v '^pkgrel=' "$aur_pkgbuild")"
+  if [ "$local_body" = "$aur_body" ]; then
     return 0  # identical packaging; nothing to do
   fi
 

common.sh

@@ -17,6 +17,19 @@ die()  { printf 'ERROR: %s\n' "$*" >&2; exit 1; }
 # The old push.sh fell back to an interactive `read` when no version was passed.
 # In CI that reads EOF and yields an empty version, which is how pman-helper
 # committed pkgver="" and the sha256 of a zero-byte download.
+# Fail up front on a missing tool rather than midway through, where an absent
+# binary tends to be read as "condition false" and silently takes the wrong
+# branch. `diff` going missing in a slimmer base image did exactly that: it made
+# every run look like a packaging change and ratcheted pkgrel upward.
+require_tools() {
+  local t missing=()
+  for t in git curl jq sha256sum awk sed grep makepkg; do
+    command -v "$t" >/dev/null 2>&1 || missing+=("$t")
+  done
+  [ "${#missing[@]}" -eq 0 ] || die "missing required tool(s): ${missing[*]}"
+  return 0
+}
+
 require_noninteractive() {
   [ -t 0 ] && die "refusing to prompt: this script must never depend on interactive input"
   return 0

dockerfile

@@ -0,0 +1,49 @@
+# Build environment for the AUR autoupdater.
+#
+# Rebuilt monthly by .gitea/workflows/build-image.yaml and pushed as :latest, so
+# the six nightly package jobs start with the toolchain already in place instead
+# of each running a full pacman transaction first.
+#
+# Arch specifically: makepkg --printsrcinfo must be the genuine Arch tool.
+# Ubuntu's build environment silently emits an empty .SRCINFO, which the AUR
+# then rejects -- that is what broke three packages before this was rewritten.
+
+# base, not base-devel: the autoupdater only ever *parses* PKGBUILDs
+# (makepkg --printsrcinfo) and never compiles anything, so the toolchain in
+# base-devel is dead weight -- 822 MB against 1.48 GB. If a real build or
+# verify step is ever added here, this needs to go back to base-devel.
+FROM archlinux:base
+
+LABEL org.opencontainers.image.title="aur-autoupdate"
+LABEL org.opencontainers.image.description="Arch build environment for the AUR package autoupdater"
+LABEL org.opencontainers.image.source="https://git.bjphoster.com/aur/.autoupdate"
+
+# nodejs        actions/checkout is a JavaScript action; the runner executes it
+#               with `node` from inside this image, and base-devel has none.
+# pacman-contrib updpkgsums/paccache and friends
+# namcap        PKGBUILD linting
+# fuse2         AppImage extraction
+# python        test.sh's local HTTP server
+RUN pacman -Syu --noconfirm --needed \
+      nodejs \
+      git \
+      openssh \
+      jq \
+      curl \
+      pacman-contrib \
+      namcap \
+      fuse2 \
+      python \
+      diffutils \
+ && rm -rf /var/cache/pacman/pkg/*
+# NB: not `pacman -Scc` -- under --noconfirm it takes the default answer, which
+# is *no* for the cache prompt, so it drops sync databases (breaking pacman -Ss
+# for debugging) without actually reclaiming the cache. The rm does the real work.
+
+# makepkg refuses to run as root and CI containers are root by default, so the
+# nightly job drops to this user. Baked in here rather than created per run.
+RUN useradd --create-home --shell /bin/bash builder
+
+# Sanity check: fail the image build rather than the nightly job if makepkg
+# ever stops working in this environment.
+RUN runuser -u builder -- bash -c 'command -v makepkg && makepkg --version | head -1'