software-mansion__argent/.github/workflows/wayland-e2e.yml

name: Wayland E2E

# Regression guard for the Linux GPU-mode + headless-window choices. Boots
# an AVD through the tool-server under headless Weston; asserts `hw.gpu.mode
# = swiftshader` and that screenshots + gesture-tap work without a display.
# A future change that re-introduces `-gpu host` (silent black-window
# failure mode) or drops `ARGENT_EMULATOR_NO_WINDOW` support fails this job.
# Manual + on-change trigger only — the job takes ~12-15 min.

on:
  workflow_dispatch:
  pull_request:
    branches:
      - main
    paths:
      - "packages/tool-server/src/tools/devices/boot-device.ts"
      - "packages/tool-server/src/utils/linux-preflight.ts"
      - "packages/tool-server/src/utils/android-binary.ts"
      - ".github/workflows/wayland-e2e.yml"

jobs:
  wayland-e2e:
    name: Boot + drive AVD on headless Weston
    runs-on: ubuntu-latest
    timeout-minutes: 30

    env:
      ANDROID_HOME: /usr/local/lib/android/sdk
      ANDROID_SDK_ROOT: /usr/local/lib/android/sdk
      # Pin AVDs to the legacy ~/.android location. Newer cmdline-tools
      # (the version on ubuntu-latest) default ANDROID_USER_HOME to
      # $XDG_CONFIG_HOME/.android = ~/.config/.android — which silently
      # diverges from the ~/.android paths this job uses for sed/grep
      # of config.ini and hardware-qemu.ini. Forcing it back keeps all
      # downstream steps pointing at the right directory.
      ANDROID_USER_HOME: /home/runner/.android
      XDG_RUNTIME_DIR: /tmp/xdg-runtime
      WAYLAND_DISPLAY: wayland-0
      # Emulator's bundled Qt has no wayland plugin; with WAYLAND_DISPLAY set
      # it'd auto-pick wayland and SIGABRT. `offscreen` keeps Qt happy without
      # a real display.
      QT_QPA_PLATFORM: offscreen
      # Opt into headless-emulator path. Required here: the runner has no
      # display so the windowed qemu-system-x86_64 SIGABRTs on the crash-
      # consent dialog.
      ARGENT_EMULATOR_NO_WINDOW: "1"

    steps:
      - name: Checkout
        uses: actions/checkout@v7

      - name: Set up Node.js
        uses: actions/setup-node@v6
        with:
          node-version: "20"

      - name: Set up JDK 17
        uses: actions/setup-java@v5
        with:
          distribution: "temurin"
          java-version: "17"

      - name: Enable KVM (nested virt on ubuntu-latest)
        # ubuntu-latest exposes /dev/kvm but the runner user isn't in the
        # `kvm` group; this matches `usermod -aG kvm $USER` from the README
        # and gives qemu the access it needs. Without it the boot falls
        # back to TCG and times out.
        run: |
          echo 'KERNEL=="kvm", GROUP="kvm", MODE="0666", OPTIONS+="static_node=kvm"' \
            | sudo tee /etc/udev/rules.d/99-kvm4all.rules
          sudo udevadm control --reload-rules
          sudo udevadm trigger --name-match=kvm
          ls -l /dev/kvm

      - name: Install Weston + emulator native deps
        # libgoogle-perftools4t64 ships libtcmalloc_minimal.so.4 which
        # qemu-system-x86_64 links against; ubuntu-latest usually has it
        # but slimmer images don't (verified blocker in ubuntu:24.04).
        run: |
          sudo apt-get update -y
          sudo apt-get install -y --no-install-recommends \
            weston xkb-data libxkbcommon0 seatd libgoogle-perftools4t64

      - name: Start Weston (headless backend)
        run: |
          mkdir -p "$XDG_RUNTIME_DIR"
          chmod 700 "$XDG_RUNTIME_DIR"
          nohup weston \
            --backend=headless \
            --socket=wayland-0 \
            --width=1920 \
            --height=1080 \
            > /tmp/weston.log 2>&1 &
          echo "weston pid=$!"
          # Wait for the wayland socket. Bounded; if Weston fails to start
          # we want a fast, clear failure rather than a hang at boot-device.
          for i in $(seq 1 30); do
            if [ -S "$XDG_RUNTIME_DIR/wayland-0" ]; then
              echo "weston socket up at t+${i}s"
              break
            fi
            sleep 1
          done
          test -S "$XDG_RUNTIME_DIR/wayland-0" || { cat /tmp/weston.log; exit 1; }

      - name: Install Android SDK packages
        run: |
          # No `pipefail` here: `yes | sdkmanager --licenses` SIGPIPEs
          # `yes` (rc 141) as soon as sdkmanager closes stdin, which is
          # normal — pipefail would turn that into a spurious step failure.
          set -e
          yes | $ANDROID_HOME/cmdline-tools/latest/bin/sdkmanager --licenses >/dev/null
          $ANDROID_HOME/cmdline-tools/latest/bin/sdkmanager \
            "emulator" \
            "platforms;android-34" \
            "system-images;android-34;google_apis;x86_64" >/dev/null
          # sdkmanager exits 0 even when a package fails to install, so
          # assert the system image files actually landed before avdmanager
          # would silently fail on a missing -k target.
          SYSIMG="$ANDROID_HOME/system-images/android-34/google_apis/x86_64"
          test -f "$SYSIMG/system.img" || test -f "$SYSIMG/userdata.img" || {
            echo "::error::system image not installed under $SYSIMG"
            ls -la "$ANDROID_HOME/system-images/android-34" 2>&1 || true
            exit 1
          }
          echo "$ANDROID_HOME/cmdline-tools/latest/bin" >> $GITHUB_PATH
          echo "$ANDROID_HOME/platform-tools" >> $GITHUB_PATH
          echo "$ANDROID_HOME/emulator" >> $GITHUB_PATH

      - name: Create AVD (4 GB / 2 vCPUs, sized for the 4-vCPU runner)
        run: |
          set -e
          # avdmanager has a long-standing bug where it can write errors to
          # stderr but still exit 0 (e.g. unrecognised device, missing parent
          # dirs). Pre-create the prefs dir, capture stderr, and verify BOTH
          # the exit code and config.ini before continuing — otherwise the
          # next step fails with a confusing `sed: can't read` instead of
          # telling us what avdmanager actually complained about. We
          # deliberately do not set pipefail here: we want the post-check
          # to run regardless of how avdmanager exited, so the diagnostic
          # dump always fires when something is wrong.
          mkdir -p "$HOME/.android"
          echo no | avdmanager -v create avd \
            -n argent-test \
            -k "system-images;android-34;google_apis;x86_64" \
            -d "pixel_6" 2>&1 | tee /tmp/avdmanager.log
          AVD_RC=${PIPESTATUS[1]}
          CFG=$HOME/.android/avd/argent-test.avd/config.ini
          if [ "$AVD_RC" -ne 0 ] || [ ! -f "$CFG" ]; then
            echo "::error::avdmanager exit=$AVD_RC, config.ini=$([ -f "$CFG" ] && echo present || echo missing)"
            echo "--- avdmanager.log ---"; cat /tmp/avdmanager.log
            echo "--- avdmanager list avd ---"; avdmanager list avd || true
            echo "--- $HOME/.android tree ---"; find "$HOME/.android" -maxdepth 4 || true
            exit 1
          fi
          sed -i 's/^hw.ramSize.*$/hw.ramSize = 4096/' "$CFG" || echo "hw.ramSize = 4096" >> "$CFG"
          sed -i 's/^hw.cpu.ncore.*$/hw.cpu.ncore = 2/' "$CFG" || echo "hw.cpu.ncore = 2" >> "$CFG"
          grep -E "^(hw\.ramSize|hw\.cpu\.ncore|hw\.gpu)" "$CFG"

      - name: Install argent dependencies + build
        run: |
          npm ci
          npm install @rollup/rollup-linux-x64-gnu --no-save
          npm run build

      - name: Download signed simulator-server binary
        # The screenshot/gesture-tap tools spawn the simulator-server out
        # of packages/native-devtools-ios/bin/<platform>/; without this
        # download the linux binary is missing and the boot succeeds but
        # every downstream tool call fails.
        #
        # Pinned to `radon-daily` because `simulator-server-argent-linux`
        # only ships there until radon#118 lands a release into the default
        # channel (`radon-main`). Drop the `radon-daily` arg once that lands.
        run: bash scripts/download-simulator-server.sh radon-daily
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}

      - name: Start tool-server (ts-node from src)
        run: |
          mkdir -p ~/.argent
          : > ~/.argent/tool-server.log
          cd packages/tool-server
          nohup env ARGENT_PORT=3033 npx ts-node src/index.ts start \
            > ~/.argent/tool-server.log 2>&1 &
          echo "tool-server pid=$!"
          # Wait for /tools to respond.
          for i in $(seq 1 30); do
            curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:3033/tools | grep -q 200 \
              && { echo "tool-server up t+${i}s"; exit 0; }
            sleep 1
          done
          echo "tool-server failed to start"; cat ~/.argent/tool-server.log; exit 1

      - name: Boot AVD via tool-server
        id: boot
        # Bigger budget than the local-box numbers because CI runners are
        # smaller (4 vCPU instead of 16) and KVM is sometimes slower under
        # GitHub's nested-virt setup.
        run: |
          set -e
          T0=$(date +%s)
          RESP=$(curl -sS -m 900 -X POST http://127.0.0.1:3033/tools/boot-device \
            -H 'Content-Type: application/json' \
            -d '{"avdName":"argent-test","bootTimeoutMs":840000}')
          T1=$(date +%s)
          echo "boot-device: $RESP"
          echo "elapsed: $((T1 - T0))s"
          echo "$RESP" | grep -q '"booted":true' || { echo "boot did not report booted:true"; exit 1; }
          # Capture the hardware-qemu.ini GPU mode for the artifact step.
          GPU=$(grep "^hw.gpu.mode" $HOME/.android/avd/argent-test.avd/hardware-qemu.ini | tr -d '[:space:]')
          echo "$GPU" > /tmp/gpu-mode.txt
          echo "gpu_mode=$GPU" >> $GITHUB_OUTPUT

      - name: Verify GPU mode is swiftshader
        # Pins the selectGpuMode() decision: any future change that
        # re-introduces `-gpu host` / `auto` will fail here.
        run: |
          GPU=$(cat /tmp/gpu-mode.txt)
          echo "Observed: $GPU"
          echo "$GPU" | grep -qi swiftshader || { echo "expected swiftshader"; exit 1; }

      - name: Screenshot returns real pixels (not all-zero framebuffer)
        run: |
          set -euo pipefail
          # After boot_completed=1 two things still need to settle before a
          # screenshot carries real pixels: SurfaceFlinger has to paint the
          # lockscreen UI, and the simulator-server's screenshot stream only
          # starts warming on the first capture request. An immediate screencap
          # can therefore return either a "no image to export" error (cold stream,
          # response has no .data envelope) or an all-zero framebuffer that
          # PNG-compresses to a few KB. A single fixed sleep races both of these,
          # which is flaky on slow/contended runners. Poll instead: each request
          # also warms the stream, so we retry until we get a real frame or time
          # out. All-zero framebuffer compresses to ~3-7 KB; the real lockscreen
          # UI (clock + battery + nav pill) is reliably > 20 KB.
          # A healthy capture returns sub-second (and a cold stream errors
          # immediately), so a tight per-attempt timeout keeps the cadence near
          # the 4s interval and bounds the worst case (~6 min if the tool-server
          # wedges) instead of 15x60s.
          attempts=15
          for i in $(seq 1 "$attempts"); do
            curl -sS -m 20 -X POST http://127.0.0.1:3033/tools/screenshot \
              -H 'Content-Type: application/json' \
              -d '{"udid":"emulator-5554"}' > /tmp/shot.json || true
            # The screenshot tool returns an ArtifactHandle ({ data: { image:
            # { hostPath, ... } } }); the job runs co-located with the tool-server
            # so hostPath is directly readable. A cold stream returns
            # { error: "...no image to export" } with no .data — tolerate that and
            # retry rather than aborting the step.
            PATH_PNG=$(python3 -c "import json;print(json.load(open('/tmp/shot.json'))['data']['image']['hostPath'])" 2>/dev/null || true)
            if [ -n "$PATH_PNG" ] && [ -f "$PATH_PNG" ]; then
              cp "$PATH_PNG" /tmp/wayland-cold-boot.png
              SZ=$(stat -c%s /tmp/wayland-cold-boot.png)
              echo "attempt ${i}/${attempts}: size=${SZ}B"
              if [ "$SZ" -gt 20000 ]; then
                echo "got a real frame after ${i} attempt(s)"
                exit 0
              fi
            else
              echo "attempt ${i}/${attempts}: no image yet ($(head -c 160 /tmp/shot.json))"
            fi
            # No need to wait after the final attempt before failing.
            if [ "$i" -lt "$attempts" ]; then sleep 4; fi
          done
          echo "screenshot never returned a real frame after ${attempts} attempts"
          exit 1

      - name: gesture-tap round-trips
        run: |
          OUT=$(curl -sS -m 30 -X POST http://127.0.0.1:3033/tools/gesture-tap \
            -H 'Content-Type: application/json' \
            -d '{"udid":"emulator-5554","x":0.5,"y":0.5}')
          echo "$OUT" | grep -q '"tapped":true' || { echo "gesture-tap failed"; exit 1; }

      - name: Upload artifacts on failure or success
        if: always()
        uses: actions/upload-artifact@v7
        with:
          name: wayland-e2e-artifacts
          path: |
            /tmp/wayland-cold-boot.png
            /tmp/weston.log
            /tmp/shot.json
            /tmp/gpu-mode.txt
            /tmp/avdmanager.log
            ~/.argent/tool-server.log
            ~/.android/avd/argent-test.avd/hardware-qemu.ini
          if-no-files-found: warn
          retention-days: 7