DodoSSH/.github/workflows/ci.yml

name: ci

on:
  push:
    branches: [main]
    # Release tags run the whole workflow, not only the image job that gates on it. A tag
    # is the one build nobody is watching, so it is the last place to take the tests on
    # trust.
    tags: ['v*']
  pull_request:
    branches: [main]

# Actions are pinned to commit SHAs, not tags: a tag can be moved to point at new code,
# which would let a compromised action run with this workflow's permissions.
permissions:
  contents: read

concurrency:
  group: ci-${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: true

env:
  DOTNET_NOLOGO: true
  DOTNET_CLI_TELEMETRY_OPTOUT: true
  DOTNET_SKIP_FIRST_TIME_EXPERIENCE: true
  CI: true

jobs:
  build:
    name: build and test
    runs-on: [linux]
    steps:
      # act_runner runs every `uses:` action with node inside the job container, and this
      # runner's image has none — the run died on the first line of actions/checkout with
      # "Cannot find: node in PATH". A `run:` step is shell rather than node, so this one
      # can go first and unblock the rest.
      #
      # This is a workaround and the real fix is one line of the runner's own config.yaml:
      # point container.image at an image that ships node, the way Gitea's default
      # catthehacker/ubuntu:act-latest does. Kept anyway, because a pipeline that depends
      # on a runner being configured correctly somewhere else fails confusingly when it is
      # not, and because it costs nothing on a runner that is.
      #
      # git as well as node, and said in the step name rather than smuggled in: checkout
      # shells out to git the moment node has loaded it, so an image thin enough to lack
      # one usually lacks the other, and learning that costs a whole second CI round trip.
      #
      # Repeated verbatim in all three jobs, which is not laziness. It cannot be a local
      # composite action — that would need the checkout it exists to unblock — and YAML
      # anchors, which would deduplicate it, are rejected by GitHub's parser and would make
      # this file portable to nothing. Change one copy, change all three.
      - name: ensure node and git
        run: |
          set -eu
          SUDO=""
          [ "$(id -u)" -eq 0 ] || SUDO="sudo"

          missing=""
          command -v node >/dev/null 2>&1 || missing="$missing nodejs"
          command -v git  >/dev/null 2>&1 || missing="$missing git"

          if [ -z "$missing" ]; then
            echo "node $(node --version), git $(git --version)"
            exit 0
          fi

          echo "Installing:$missing"
          if command -v apt-get >/dev/null 2>&1; then
            $SUDO apt-get update -qq
            $SUDO apt-get install -y --no-install-recommends $missing
          elif command -v apk >/dev/null 2>&1; then
            $SUDO apk add --no-cache $missing
          elif command -v dnf >/dev/null 2>&1; then
            $SUDO dnf install -y $missing
          else
            echo "No apt-get, apk or dnf here, so node cannot be installed from inside the" >&2
            echo "job. Point the runner's container.image at something that ships node." >&2
            exit 1
          fi
          echo "node $(node --version), git $(git --version)"

          # Warned about rather than failed on. Distributions pin their nodejs package to
          # the release they shipped with — Ubuntu 24.04 still serves 18, which is past end
          # of life and older than the runtime these actions declare. It generally runs
          # them anyway, since act_runner uses whichever node is on PATH regardless of what
          # the action asked for, so this is a note for when one of them misbehaves in a
          # way that makes no sense, not a reason to stop a build that is probably fine.
          major="$(node --version | sed 's/^v//; s/\..*//')"
          if [ "$major" -lt 20 ]; then
            echo "::warning::node $major is older than the runtime these actions target;" \
                 "give the runner an image with node 20 or newer if actions misbehave."
          fi

      # fetch-depth 0, and it is load-bearing rather than tidy. MinVer derives the version from the
      # nearest v* tag, and checkout's default shallow clone has no tags at all — so it would not fail,
      # it would quietly answer 0.0.0-alpha.0.N and every build would ship that. Velopack decides
      # whether an installed client is out of date by comparing versions, which makes a plausible wrong
      # answer here a client that never updates.
      #
      # Repeated in all three jobs, like the node preamble above and for the same reason. Change one
      # copy, change all three.
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
        with:
          fetch-depth: 0

      - uses: actions/setup-dotnet@a98b56852c35b8e3190ac28c8c2271da59106c68 # v6.0.0
        with:
          global-json-file: global.json
          cache: true
          cache-dependency-path: '**/packages.lock.json'

      # Locked mode fails if packages.lock.json does not match the project files, so a
      # dependency cannot change without the lock file change being reviewed.
      - name: restore
        run: dotnet restore DodoSSH.slnx --locked-mode

      # The version comes from the tag, so on a tag build there are two ways to say the same number
      # and they can disagree — a moved tag, a tag on the wrong commit, or a checkout that somehow
      # still lost its history. What makes that worth a step of its own is that the disagreement is
      # silent everywhere else: MinVer answers 0.0.0-alpha.0.N rather than failing, the build goes
      # green, the package is cut, and the symptom arrives weeks later as clients that never update.
      #
      # -getProperty evaluates without building, so this costs a second and runs before the build.
      - name: the tag and the version agree
        if: startsWith(github.ref, 'refs/tags/v')
        run: |
          set -euo pipefail
          tag="${GITHUB_REF#refs/tags/v}"
          declared="$(dotnet msbuild src/DodoSSH.Client.App/DodoSSH.Client.App.csproj \
                        -getProperty:Version -nologo | tr -d '[:space:]')"

          if [ "$tag" != "$declared" ]; then
            echo "The tag says v$tag and MinVer computed $declared." >&2
            echo >&2
            echo "These come from the same place, so a mismatch means the checkout did not see the" >&2
            echo "tag it is building — most likely fetch-depth, which must stay 0 in every job here." >&2
            exit 1
          fi

          echo "v$declared"

      # No `dotnet format --verify-no-changes` step. It re-analysed the whole solution before
      # the build did, for minutes, to check something the build already checks: IDE0055 is an
      # error in .editorconfig and TreatWarningsAsErrors is on, so a misformatted file fails
      # the build step below on its own. What the separate step added was the ability to say
      # so a few minutes earlier, and it cost more than that on every run.

      # Avalonia's headless renderer is still Skia, and libSkiaSharp.so — which the layout
      # test project copies into its own output — links against libfontconfig. Without that
      # one library every test in DodoSSH.Client.App.Layout.Tests dies inside
      # HeadlessUnitTestSession before it measures anything, and 69 tests fail for a reason
      # none of their names or assertions mention.
      #
      # The library, and not fonts. Verified in a container where fc-list returns zero and
      # the suite passes anyway: the application carries Inter itself, so nothing here needs
      # a typeface installed — only the thing that would have gone looking for one.
      - name: ensure skia's native dependency
        run: |
          set -eu
          SUDO=""
          [ "$(id -u)" -eq 0 ] || SUDO="sudo"

          if ldconfig -p 2>/dev/null | grep -q 'libfontconfig\.so\.1'; then
            echo "libfontconfig present"
            exit 0
          fi

          echo "Installing fontconfig"
          if command -v apt-get >/dev/null 2>&1; then
            $SUDO apt-get update -qq
            $SUDO apt-get install -y --no-install-recommends libfontconfig1
          elif command -v apk >/dev/null 2>&1; then
            $SUDO apk add --no-cache fontconfig
          elif command -v dnf >/dev/null 2>&1; then
            $SUDO dnf install -y fontconfig
          else
            echo "No package manager here, so Skia cannot be given its dependency and the" >&2
            echo "layout suite will fail to start. Add fontconfig to the runner's image." >&2
            exit 1
          fi

      - name: build
        run: dotnet build DodoSSH.slnx --no-restore --configuration Release

      - name: test
        run: dotnet test DodoSSH.slnx --no-build --configuration Release

      # The one build shape nothing else here exercises: a self-contained RID-specific publish. Its
      # failure mode is a restore graph or a native asset that resolves for net10.0 and not for
      # net10.0/win-x64, which nobody would see until a person was halfway through cutting a release
      # on a Windows machine. vpk can pack a Windows package from Linux; only signing needs Windows,
      # and this repository signs nothing yet, so proving the publish here is worth the minutes.
      #
      # RestoreLockedMode=false for this command only, and it is not a loosened gate. The committed
      # lock files are deliberately RID-free: declaring win-x64 on the desktop head writes a
      # net10.0/win-x64 target into every project it references transitively, which includes
      # DodoSSH.Contracts and DodoSSH.Crypto — and the API's Dockerfile restores those with no RID
      # under locked mode, so the image job would fail NU1004. The gate is the locked solution
      # restore at the top of this job, which is unchanged.
      #
      # It rewrites the lock files as it goes; nothing after this step reads them, and the runner's
      # checkout is thrown away. The release script does the same thing and puts them back, because
      # there the tree is somebody's working copy.
      #
      # After the tests rather than before them, so a red suite does not first spend a hundred
      # megabytes pulling a win-x64 runtime pack. main and tags only, for the same reason: a break
      # found by the person about to release is found early enough.
      - name: the windows publish still resolves
        if: github.event_name != 'pull_request'
        run: >
          dotnet publish src/DodoSSH.Client.App/DodoSSH.Client.App.csproj
          --configuration Release --runtime win-x64 --self-contained true
          -p:RestoreLockedMode=false
          --output "$RUNNER_TEMP/win-x64-check"

      # This includes the end-to-end suite, which starts PostgreSQL, Keycloak and an OpenSSH
      # server through Testcontainers and runs the API as a child process — so it needs a
      # Docker daemon and gets one here. That is why the tests run on ubuntu rather than
      # macOS, whose runners have no daemon at all. Expect the Keycloak image pull to
      # dominate a cold run.

      # A failing run says only that tests failed and names a log file on a machine nobody
      # has a shell on. Every diagnostic thing — exception type, message, stack — is inside
      # that file, so a red build was a filename and a guess. This prints it.
      #
      # head rather than tail, and that is the whole trick: when a suite fails wholesale it
      # writes one stack per test and they are all the same stack. The first is the one that
      # explains it, and the last two hundred lines are the same sentence repeated.
      - name: what actually failed
        if: failure()
        run: |
          set +e
          echo "=== distro ==="
          cat /etc/os-release 2>/dev/null | head -3
          id

          echo "=== what Skia needs, and whether it is here ==="
          # ldd against the copy the test project carries. Its unresolved rows are the
          # answer whenever the layout suite dies in HeadlessUnitTestSession, and asking
          # here beats inferring it from a managed TypeInitializationException.
          skia="$(find tests -name 'libSkiaSharp.so' 2>/dev/null | head -1)"
          if [ -n "$skia" ]; then
            echo "$skia"
            ldd "$skia" 2>&1 | grep -Ei 'not found|fontconfig|freetype' || echo "  all resolved"
          else
            echo "  libSkiaSharp.so was not in the test output at all"
          fi
          ldconfig -p 2>/dev/null | grep -ci fontconfig | sed 's/^/fontconfig entries in ldconfig: /'

          echo "=== docker, for the Testcontainers suites ==="
          docker version --format '{{.Server.Version}}' 2>&1 | head -2

          echo "=== test logs ==="
          find tests -path '*/TestResults/*.log' 2>/dev/null | while read -r log; do
            echo "----- $log"
            head -n 120 "$log"
          done
          exit 0
  android:
    name: android head
    runs-on: [linux]
    # Writes the nightly release at the end of the job; see that step for what the capability is and why
    # it is acceptable here and nowhere else. Job-scoped, so nothing else in this file gains it.
    permissions:
      contents: write
    steps:
      # Duplicated from the build job; see the comment there for why it cannot be factored
      # out. Any change here has to be made in all three.
      - name: ensure node and git
        run: |
          set -eu
          SUDO=""
          [ "$(id -u)" -eq 0 ] || SUDO="sudo"

          missing=""
          command -v node >/dev/null 2>&1 || missing="$missing nodejs"
          command -v git  >/dev/null 2>&1 || missing="$missing git"

          if [ -z "$missing" ]; then
            echo "node $(node --version), git $(git --version)"
            exit 0
          fi

          echo "Installing:$missing"
          if command -v apt-get >/dev/null 2>&1; then
            $SUDO apt-get update -qq
            $SUDO apt-get install -y --no-install-recommends $missing
          elif command -v apk >/dev/null 2>&1; then
            $SUDO apk add --no-cache $missing
          elif command -v dnf >/dev/null 2>&1; then
            $SUDO dnf install -y $missing
          else
            echo "No apt-get, apk or dnf here, so node cannot be installed from inside the" >&2
            echo "job. Point the runner's container.image at something that ships node." >&2
            exit 1
          fi
          echo "node $(node --version), git $(git --version)"

          # Warned about rather than failed on. Distributions pin their nodejs package to
          # the release they shipped with — Ubuntu 24.04 still serves 18, which is past end
          # of life and older than the runtime these actions declare. It generally runs
          # them anyway, since act_runner uses whichever node is on PATH regardless of what
          # the action asked for, so this is a note for when one of them misbehaves in a
          # way that makes no sense, not a reason to stop a build that is probably fine.
          major="$(node --version | sed 's/^v//; s/\..*//')"
          if [ "$major" -lt 20 ]; then
            echo "::warning::node $major is older than the runtime these actions target;" \
                 "give the runner an image with node 20 or newer if actions misbehave."
          fi

      # fetch-depth 0, and it is load-bearing rather than tidy. MinVer derives the version from the
      # nearest v* tag, and checkout's default shallow clone has no tags at all — so it would not fail,
      # it would quietly answer 0.0.0-alpha.0.N and every build would ship that. Velopack decides
      # whether an installed client is out of date by comparing versions, which makes a plausible wrong
      # answer here a client that never updates.
      #
      # Repeated in all three jobs, like the node preamble above and for the same reason. Change one
      # copy, change all three.
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
        with:
          fetch-depth: 0

      # A job of its own, because DodoSSH.Client.Android is deliberately not in DodoSSH.slnx.
      # Adding it there would make the android workload and a full Android SDK a prerequisite of
      # `dotnet build DodoSSH.slnx` for everyone — including the build job above, which needs
      # neither and would grow several minutes for a head it does not compile.
      #
      # The cost of keeping it out is that nothing in the main job would notice this head
      # breaking, which for a project sharing view models with the desktop one is a matter of
      # when rather than whether. This job is that notice.
      #
      # ============ ◆ WHY THIS HEAD IS BUILT IN A CONTAINER AND THE OTHER TWO ARE NOT ============
      #
      # .NET for Android cannot build on this runner. Not "needs setting up" — cannot. The runner is
      # Alpine, and the SDK's own MSBuild tasks load glibc shared libraries out of the workload pack
      # into the build process, which a musl-linked dotnet will not do:
      #
      #   error XARLP7000: Error relocating …/libZipSharpNative-3-3.so: __snprintf_chk: symbol not found
      #
      # There is no musl variant of that pack and no shim for a library loaded in-process — gcompat
      # gets a glibc *executable* started and is no help whatsoever here. Everything this job used to
      # do on the host to make Android work was therefore treatment of symptoms: an aapt2 reported
      # missing while being present, an "unsupported version" of a binary that had never run. All of
      # them were this one sentence in a different accent, and docs/platform-flags.md has the history
      # so the next reader does not repeat the three rounds it took to hear it.
      #
      # So the toolchain moves into an image that is glibc, and the job keeps on the host only what the
      # host is good at: checkout, git, and publishing. The daemon needs no arranging — the image job
      # builds with it and every Testcontainers suite in the build job reaches it over the socket.
      #
      # There is no setup-dotnet here any more, and its absence is the point: nothing outside the
      # container compiles anything, so a dotnet on the host would be a second toolchain with nothing
      # to do and a version nobody checks.
      - name: ensure the docker cli
        run: |
          set -eu
          SUDO=""
          [ "$(id -u)" -eq 0 ] || SUDO="sudo"

          # Only the client. The socket is already there and a daemon is already answering on it, so
          # installing an engine would start a second one beside the one in use. The image job carries
          # the same step and the same reasoning; change one, look at the other.
          if ! command -v docker >/dev/null 2>&1; then
            echo "Installing the docker cli"
            if command -v apk >/dev/null 2>&1; then
              $SUDO apk add --no-cache docker-cli
            elif command -v apt-get >/dev/null 2>&1; then
              $SUDO apt-get update -qq
              $SUDO apt-get install -y --no-install-recommends docker.io
            elif command -v dnf >/dev/null 2>&1; then
              $SUDO dnf install -y docker-cli
            else
              echo "No apt-get, apk or dnf here, so the docker client cannot be installed from" >&2
              echo "inside the job. Add it to the runner's image." >&2
              exit 1
            fi
          fi
          docker --version

      # Tagged by the digest of the Dockerfile that made it, so the tag moves exactly when the toolchain
      # does and never otherwise. This runner is persistent, so the first build pays for a JDK, an
      # Android SDK and the android workload — several minutes — and every run after it is a cache hit
      # that prints two lines. A change to that Dockerfile is what buys a new one.
      - name: the android toolchain image
        run: |
          set -eu

          tag="dodossh-android-build:$(sha256sum build/android-build.Dockerfile | cut -c1-16)"
          docker build -t "$tag" -f build/android-build.Dockerfile build

          echo "ANDROID_BUILD_IMAGE=$tag" >> "$GITHUB_ENV"
          echo "$tag"

      # The checkout is bind-mounted rather than copied in, so what the container writes — obj, bin, the
      # signed package — is on the host the moment it exits and the publishing step can read it. The
      # NuGet cache is mounted for the mirror-image reason: a container that started empty would fetch
      # every package again on every run.
      #
      # The script arrives on stdin under `bash -s` rather than as `bash -c '…'`, which is not a style
      # choice: the packaging step below has to match a versionName out of aapt2 with sed, and nesting
      # that quoting inside a shell string is how a working command becomes a silently empty variable.
      # A quoted heredoc passes the script through untouched, so what is written here is what runs.
      #
      # CI=true is handed over rather than assumed. Directory.Build.props turns ContinuousIntegrationBuild
      # on when it is set — which is what normalises the source paths baked into the PDBs — and a
      # container inherits none of the runner's environment unless it is given it.
      - name: restore and build
        run: |
          set -eu

          docker run --rm -i \
            -v "$PWD:/build" \
            -v "$HOME/.nuget/packages:/root/.nuget/packages" \
            -e CI=true \
            "$ANDROID_BUILD_IMAGE" bash -s <<'IN_CONTAINER'
          set -euo pipefail

          dotnet restore src/DodoSSH.Client.Android/DodoSSH.Client.Android.csproj --locked-mode

          # Aapt2ToolPath, so the build takes aapt2 from the Android SDK in this image rather than the
          # copy inside the workload pack. Both work here; using the SDK's makes it the same binary the
          # packaging step reads the versionName back with, so the manifest the feed publishes is read
          # by the thing that wrote it.
          dotnet build src/DodoSSH.Client.Android/DodoSSH.Client.Android.csproj \
            --no-restore --configuration Release \
            -p:Aapt2ToolPath="$ANDROID_HOME/build-tools/$ANDROID_BUILD_TOOLS"
          IN_CONTAINER

      # Packaging rather than only compiling, because the two failures this head is most exposed to are
      # both link-time: a native library with no android ABI, and a managed assembly that resolves for
      # net10.0 but has nothing to dex. Neither shows up in a compile.
      #
      # ◆ THE NIGHTLY CHANNEL, WHICH IS AN INSTALLABLE APPLICATION AND NOT THE ONE. It has its own package
      # id and is signed by a keystore committed to this repository in the open, so it can neither replace
      # nor be replaced by the release channel — see the csproj, and ADR 0014. The APK this produces is
      # meant to be installed; the release APK is cut from a v* tag by a person running
      # scripts/release-android.ps1, on a machine that holds the key ADR 0011 rule 1 keeps off runners.
      #
      # No RuntimeIdentifier, where this step used to pin android-arm64. That produced the smallest
      # possible build check and the least installable artefact: an arm64-only APK will not run on an
      # x86_64 emulator, which is what most people testing a nightly actually have. Every supported ABI
      # costs size on a package nobody ships to users.
      #
      # versionCode is the commit count, which is monotonic by construction and needs nobody to remember
      # anything. It is not a version and is never displayed; versionName carries MinVer's full answer
      # including the height, which is what tells two nightlies apart.
      - name: package the nightly
        id: nightly
        run: |
          set -euo pipefail

          # Under artifacts/ rather than RUNNER_TEMP, and that is forced rather than preferred: the
          # package is produced inside a container and read outside one, so it has to land somewhere
          # both can see, which means somewhere under the bind-mounted checkout. .gitignore already
          # excludes artifacts/* with two named exceptions, neither of which is this.
          staged="artifacts/android-nightly"
          rm -rf "$staged"

          docker run --rm -i \
            -v "$PWD:/build" \
            -v "$HOME/.nuget/packages:/root/.nuget/packages" \
            -e CI=true \
            "$ANDROID_BUILD_IMAGE" bash -s <<'IN_CONTAINER'
          set -euo pipefail

          staged="artifacts/android-nightly"
          mkdir -p "$staged"

          code="$(git rev-list --count HEAD)"

          dotnet build src/DodoSSH.Client.Android/DodoSSH.Client.Android.csproj \
            --no-restore --configuration Release \
            -t:SignAndroidPackage \
            -p:DodoChannel=nightly \
            -p:DodoNightlyVersionCode="$code" \
            -p:Aapt2ToolPath="$ANDROID_HOME/build-tools/$ANDROID_BUILD_TOOLS"

          apk="$(find src/DodoSSH.Client.Android/bin/Release -name '*-Signed.apk' | head -1)"
          if [ -z "$apk" ]; then
            echo "The package step produced no signed APK." >&2
            exit 1
          fi

          # Read back out of the APK rather than recomputed, so what the feed advertises is what the
          # bytes say. A versionName derived a second time in shell is a second implementation of the
          # csproj's target, and the two would drift on the first change to either.
          badging="$("$ANDROID_HOME/build-tools/$ANDROID_BUILD_TOOLS/aapt2" dump badging "$apk")"
          name="$(printf '%s' "$badging" | sed -n "s/.*versionName='\([^']*\)'.*/\1/p" | head -1)"

          if [ -z "$name" ]; then
            echo "aapt2 reported no versionName for $apk." >&2
            exit 1
          fi

          cp "$apk" "$staged/DodoSSH-nightly-$name.apk"

          # The channel manifest, which is what the client reads and the whole reason the feed is
          # machine-readable at all. versionCode is the comparison — it is the number Android itself uses
          # to accept or refuse an install, so comparing anything else would let the client offer an
          # update the platform then rejects. versionName is for the person reading the banner.
          printf '{"versionCode":%s,"versionName":"%s","apk":"DodoSSH-nightly-%s.apk"}' \
            "$code" "$name" "$name" > "$staged/android-nightly.json"
          IN_CONTAINER

          # Read back out of the manifest the container just wrote, rather than passed out of it. A
          # container's stdout is the build log as well as its return value, so anything parsed from it
          # is one stray MSBuild line away from being wrong.
          name="$(sed -n 's/.*"versionName":"\([^"]*\)".*/\1/p' "$staged/android-nightly.json")"
          if [ -z "$name" ]; then
            echo "The container produced no usable manifest." >&2
            exit 1
          fi

          echo "staged=$staged" >> "$GITHUB_OUTPUT"
          echo "version=$name" >> "$GITHUB_OUTPUT"
          ls -la "$staged"


      # ◆ PUBLISHING IT, AND WHAT THAT CAPABILITY IS.
      #
      # Whoever can write a release here can put a build on every nightly phone, because the client fetches
      # from this feed and Android's only check is that the signature matches — and this channel's key is
      # in the repository for everybody. That is the same capability as the signing key, reached through a
      # different door, and it is exactly what ADR 0011 rule 1 keeps off runners.
      #
      # It is acceptable here for one reason: this is not that channel. A nightly is signed by a key with
      # no secrecy to lose, installs under its own package id, and cannot update the application anybody
      # is trusting with their credentials. The release channel has none of this — no job, no token, no
      # key on a runner — and the two are separate applications so that no mistake here can reach it.
      #
      # main only. A tag build must not touch this: a v* tag is the release channel's, and cutting it is a
      # person's job.
      - name: publish the nightly
        if: github.ref == 'refs/heads/main'
        env:
          FORGE: https://git.dodotech.cloud
          REPO: DodoTech/DodoSSH
          TOKEN: ${{ secrets.GITHUB_TOKEN }}
          STAGED: ${{ steps.nightly.outputs.staged }}
          VERSION: ${{ steps.nightly.outputs.version }}
        run: |
          set -euo pipefail

          if [ -z "${TOKEN:-}" ]; then
            echo "No token, so the nightly was built and not published." >&2
            echo "GITHUB_TOKEN is provided by the runner; an empty one means Actions is configured" >&2
            echo "without it, and the job's contents: write permission is what asks for it." >&2
            exit 1
          fi

          api="$FORGE/api/v1/repos/$REPO"
          auth="Authorization: token $TOKEN"

          # Deleted and recreated rather than updated in place. A rolling tag has to move, and moving one
          # through this API is two calls with no atomic form either way — so the shape with the fewest
          # states is to remove both and make them again. The window where no nightly exists is a few
          # seconds and the client's answer to it is the same as to an unreachable forge: try later.
          existing="$(curl -fsS -H "$auth" "$api/releases/tags/nightly" 2>/dev/null || true)"
          if [ -n "$existing" ]; then
            id="$(printf '%s' "$existing" | sed -n 's/.*"id":\([0-9]*\).*/\1/p' | head -1)"
            [ -n "$id" ] && curl -fsS -X DELETE -H "$auth" "$api/releases/$id" >/dev/null || true
          fi
          curl -fsS -X DELETE -H "$auth" "$api/tags/nightly" >/dev/null 2>&1 || true

          created="$(curl -fsS -X POST -H "$auth" -H 'Content-Type: application/json' \
            -d "$(printf '{"tag_name":"nightly","target_commitish":"%s","name":"Nightly %s","prerelease":true,"body":"Built from %s by CI, signed with the public nightly key. Installs beside the release build, never over it. See docs/adr/0014-android-updates.md."}' \
                    "$GITHUB_SHA" "$VERSION" "$GITHUB_SHA")" \
            "$api/releases")"

          release="$(printf '%s' "$created" | sed -n 's/.*"id":\([0-9]*\).*/\1/p' | head -1)"
          if [ -z "$release" ]; then
            echo "Gitea accepted the release call and returned no id:" >&2
            printf '%s\n' "$created" >&2
            exit 1
          fi

          # The APK first and the manifest last, which is the ordering the client depends on: it reads the
          # manifest and then fetches what the manifest names, so a manifest published before its APK is a
          # few seconds in which every phone is told to download something that is not there yet.
          for file in "$STAGED"/*.apk "$STAGED"/android-nightly.json; do
            echo "Uploading $(basename "$file")"
            curl -fsS -X POST -H "$auth" \
              -F "attachment=@$file" \
              "$api/releases/$release/assets?name=$(basename "$file")" >/dev/null
          done

          echo "Published nightly $VERSION"

  image:
    name: api image
    # Gated on the tests rather than parallel with them, which costs a few minutes of wall
    # clock on every main commit and buys the thing worth having: no image reaches the
    # registry from a commit whose tests were red. An image is not a build artefact anyone
    # inspects — it is the thing that gets deployed.
    needs: [build]
    runs-on: [linux]
    steps:
      # Duplicated from the build job; see the comment there for why it cannot be factored
      # out. Any change here has to be made in all three.
      - name: ensure node and git
        run: |
          set -eu
          SUDO=""
          [ "$(id -u)" -eq 0 ] || SUDO="sudo"

          missing=""
          command -v node >/dev/null 2>&1 || missing="$missing nodejs"
          command -v git  >/dev/null 2>&1 || missing="$missing git"

          if [ -z "$missing" ]; then
            echo "node $(node --version), git $(git --version)"
            exit 0
          fi

          echo "Installing:$missing"
          if command -v apt-get >/dev/null 2>&1; then
            $SUDO apt-get update -qq
            $SUDO apt-get install -y --no-install-recommends $missing
          elif command -v apk >/dev/null 2>&1; then
            $SUDO apk add --no-cache $missing
          elif command -v dnf >/dev/null 2>&1; then
            $SUDO dnf install -y $missing
          else
            echo "No apt-get, apk or dnf here, so node cannot be installed from inside the" >&2
            echo "job. Point the runner's container.image at something that ships node." >&2
            exit 1
          fi
          echo "node $(node --version), git $(git --version)"

          # Warned about rather than failed on. Distributions pin their nodejs package to
          # the release they shipped with — Ubuntu 24.04 still serves 18, which is past end
          # of life and older than the runtime these actions declare. It generally runs
          # them anyway, since act_runner uses whichever node is on PATH regardless of what
          # the action asked for, so this is a note for when one of them misbehaves in a
          # way that makes no sense, not a reason to stop a build that is probably fine.
          major="$(node --version | sed 's/^v//; s/\..*//')"
          if [ "$major" -lt 20 ]; then
            echo "::warning::node $major is older than the runtime these actions target;" \
                 "give the runner an image with node 20 or newer if actions misbehave."
          fi

      # fetch-depth 0, and it is load-bearing rather than tidy. MinVer derives the version from the
      # nearest v* tag, and checkout's default shallow clone has no tags at all — so it would not fail,
      # it would quietly answer 0.0.0-alpha.0.N and every build would ship that. Velopack decides
      # whether an installed client is out of date by comparing versions, which makes a plausible wrong
      # answer here a client that never updates.
      #
      # Repeated in all three jobs, like the node preamble above and for the same reason. Change one
      # copy, change all three.
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
        with:
          fetch-depth: 0

      # The daemon and the client are two different things to have, and this runner had only
      # one of them. Testcontainers reaches Docker straight over /var/run/docker.sock from a
      # .NET library, so every integration suite in the build job passed while `docker` was
      # not a command here at all — which surfaced as exit 127 from the build step below,
      # after the tags had been worked out and everything looked healthy.
      #
      # Only the CLI. The socket is already there and a daemon is already answering on it;
      # installing an engine would start a second one beside the one being used.
      - name: ensure the docker cli
        run: |
          set -eu
          SUDO=""
          [ "$(id -u)" -eq 0 ] || SUDO="sudo"

          if ! command -v docker >/dev/null 2>&1; then
            echo "Installing the docker cli"
            if command -v apk >/dev/null 2>&1; then
              $SUDO apk add --no-cache docker-cli
            elif command -v apt-get >/dev/null 2>&1; then
              $SUDO apt-get update -qq
              $SUDO apt-get install -y --no-install-recommends docker.io
            elif command -v dnf >/dev/null 2>&1; then
              $SUDO dnf install -y docker-cli
            else
              echo "No apt-get, apk or dnf here, so the docker client cannot be installed from" >&2
              echo "inside the job. Add it to the runner's image." >&2
              exit 1
            fi
          fi
          docker --version

          # buildx after the client, and wanted rather than required. Without the plugin
          # `docker build` falls back to the legacy builder, which still produces the image
          # and says on every run that it will not do so forever; with it the same command
          # routes through BuildKit and the Dockerfile's independent stages stop being
          # serialised. Alpine's docker-cli package does not carry it, which is why a job
          # that had just been given a working client still built the deprecated way.
          #
          # A distribution with no package for it should get a warning and an image, not a
          # failed release — so every branch here ends in `|| true` and the check below
          # reports rather than exits.
          if ! docker buildx version >/dev/null 2>&1; then
            echo "Installing buildx"
            if command -v apk >/dev/null 2>&1; then
              $SUDO apk add --no-cache docker-cli-buildx || true
            elif command -v apt-get >/dev/null 2>&1; then
              $SUDO apt-get update -qq || true
              $SUDO apt-get install -y --no-install-recommends docker-buildx || true
            elif command -v dnf >/dev/null 2>&1; then
              $SUDO dnf install -y docker-buildx || true
            fi
          fi

          if docker buildx version >/dev/null 2>&1; then
            docker buildx version
          else
            echo "::warning::buildx is unavailable, so this image was built by the legacy" \
                 "builder Docker has deprecated. Add a buildx package to the runner image."
          fi

      # No docker/* actions here, deliberately. The build is single-architecture, so it
      # wants the daemon this runner already has, a client, and BuildKit — all of which the
      # step above arranges with two packages. What it does not want is QEMU, a builder
      # instance to create and tear down, or a third-party action whose SHA has to be
      # audited and re-pinned on a schedule. Adding linux/arm64 later is where that trade
      # changes, and where setup-buildx-action starts earning its place.
      - name: work out the tags
        id: tags
        env:
          REGISTRY: registry-docker.dodotech.cloud
          IMAGE: dodotech/dodossh-api
        run: |
          set -euo pipefail
          repo="$REGISTRY/$IMAGE"
          short="$(git rev-parse --short HEAD)"

          # sha- prefixed, because a bare hex tag is ambiguous with a digest to both a human
          # and a fair amount of tooling. This one is on every build and never moves, which
          # makes it the only tag safe to pin a deployment to.
          tags="$repo:sha-$short"
          version="$short"

          case "$GITHUB_REF" in
            refs/tags/v*)
              v="${GITHUB_REF#refs/tags/v}"
              version="$v"
              tags="$tags $repo:$v"
              # The moving major.minor tag and :latest, but only for a release proper.
              # v1.3.0-rc1 sorts after v1.2.9 and would otherwise take :latest with it,
              # which is how a release candidate ends up on somebody's server.
              case "$v" in
                *-*) ;;
                *)
                  tags="$tags $repo:${v%.*}"
                  tags="$tags $repo:latest"
                  ;;
              esac
              ;;
            refs/heads/main)
              tags="$tags $repo:main"
              version="main-$short"
              ;;
          esac

          # The version MSBuild is allowed to see, which is not the same string as the one above.
          # `version` is a docker tag and is `main-<short sha>` on a main build; handing that to
          # -p:Version fails the publish with NETSDK1018. So this is set only when it is a real
          # version, and the Dockerfile leaves the SDK default alone when it is empty.
          assembly_version=""
          case "$GITHUB_REF" in
            refs/tags/v*) assembly_version="${GITHUB_REF#refs/tags/v}" ;;
          esac

          echo "tags=$tags" >> "$GITHUB_OUTPUT"
          echo "version=$version" >> "$GITHUB_OUTPUT"
          echo "assemblyVersion=$assembly_version" >> "$GITHUB_OUTPUT"
          echo "created=$(date -u +%Y-%m-%dT%H:%M:%SZ)" >> "$GITHUB_OUTPUT"
          echo "Tagging: $tags"

      # Every value from a step output or the event goes through env rather than being
      # interpolated into the script text. A git tag may contain a semicolon, and
      # `${{ }}` is a textual substitution performed before the shell ever sees the line —
      # so an interpolated tag name is a command the workflow agreed to run.
      - name: build the image
        env:
          TAGS: ${{ steps.tags.outputs.tags }}
          VERSION: ${{ steps.tags.outputs.version }}
          ASSEMBLY_VERSION: ${{ steps.tags.outputs.assemblyVersion }}
          REVISION: ${{ github.sha }}
          CREATED: ${{ steps.tags.outputs.created }}
        run: |
          set -euo pipefail
          args=()
          for tag in $TAGS; do
            args+=(--tag "$tag")
          done
          # --pull rather than whatever the runner happens to have cached: the base images
          # are floating tags, and a runner that has held aspnet:10.0-noble-chiseled for a
          # month is a month of unapplied CVE fixes shipping in every image built on it.
          docker build \
            --pull \
            --file src/DodoSSH.Api/Dockerfile \
            --build-arg "VERSION=$VERSION" \
            --build-arg "ASSEMBLY_VERSION=$ASSEMBLY_VERSION" \
            --build-arg "REVISION=$REVISION" \
            --build-arg "CREATED=$CREATED" \
            "${args[@]}" \
            .

      # Everything above runs on a pull request too. Building a fork's Dockerfile is safe —
      # nothing is pushed and no credential is in scope — and it means a change that breaks
      # the image fails on the PR rather than on main. Only these last two steps are held
      # back, and the condition is on the event rather than on the branch so that a PR
      # targeting main cannot reach them.
      - name: log in to the registry
        if: github.event_name != 'pull_request'
        env:
          REGISTRY_USERNAME: ${{ secrets.REGISTRY_USERNAME }}
          REGISTRY_PASSWORD: ${{ secrets.REGISTRY_PASSWORD }}
        run: |
          set -eu

          # Checked before use, because an unset secret is not an error anywhere upstream of
          # here: an expression that resolves to nothing renders as the empty string, so
          # docker is handed --username "" and answers with something about credentials,
          # which sends people to the registry to debug a value that never left the
          # settings page.
          #
          # Note for anyone editing this comment: an expression delimiter written literally
          # here is interpolated even though this is a shell comment. The runner substitutes
          # the whole script before any shell sees it, so an empty one fails the step with a
          # parse error and no line number — which is how this very block broke the release
          # it was added to protect.
          #
          # Reported by length, and never by value. Gitea masks known secret values in logs,
          # but a mask is only as good as the runner's bookkeeping and a length answers the
          # only question being asked: did anything arrive.
          missing=""
          [ -n "${REGISTRY_USERNAME:-}" ] || missing="$missing REGISTRY_USERNAME"
          [ -n "${REGISTRY_PASSWORD:-}" ] || missing="$missing REGISTRY_PASSWORD"

          if [ -n "$missing" ]; then
            echo "Empty or unset:$missing" >&2
            echo >&2
            echo "Both come from repository secrets, which in Gitea are at" >&2
            echo "  Settings -> Actions -> Secrets" >&2
            echo "and are a different page from Settings -> Actions -> Variables. A value" >&2
            echo "added as a variable is invisible to the secrets context and arrives here" >&2
            echo "as an empty string, which is exactly what this message means." >&2
            exit 1
          fi

          echo "username: ${#REGISTRY_USERNAME} characters; password: set"

          printf '%s' "$REGISTRY_PASSWORD" \
            | docker login registry-docker.dodotech.cloud \
                --username "$REGISTRY_USERNAME" --password-stdin

      - name: push
        if: github.event_name != 'pull_request'
        env:
          TAGS: ${{ steps.tags.outputs.tags }}
        run: |
          set -euo pipefail
          for tag in $TAGS; do
            docker push "$tag"
          done

      # The daemon is shared with every other job on this runner, and a credential left in
      # ~/.docker/config.json outlives the job that created it. always(), so a failed push
      # does not leave it behind.
      - name: log out
        if: always() && github.event_name != 'pull_request'
        run: docker logout registry-docker.dodotech.cloud

# There is no job here that publishes the desktop client, and there is not going to be one. Two
# independent reasons, and both need saying because someone will fix one and think they are done.
#
# The smaller one is mechanical: vpk stamps and embeds the Setup.exe and Update.exe stubs with Windows
# tooling, and every job in this file is runs-on: [linux]. A Windows runner would answer that.
#
# The larger one is that a Windows runner would not answer the other. Velopack clients fetch from the
# release feed and do not verify a package signature when they apply it, so whoever can write a release
# on this repository can publish an update that every installed client downloads and runs. That is the
# same capability as the signing key, reached through a different door — and docs/adr/0011 rule 1 puts
# that capability on a machine which is not a runner, because a workflow secret is held by everyone who
# can change a workflow file. See docs/adr/0013-desktop-distribution-and-updates.md.
#
# What cuts a release is scripts/release-windows.ps1, run by a person. What this file does is prove the
# thing still builds and packages, which is the same division of labour the android job above already
# has: it packages an APK nobody installs, so that a link-time break fails here rather than later.