Batch fix: #11-#16, #18 (untested, needs live device pass)

.github/workflows/armbian-builder.yaml

@@ -283,7 +283,14 @@ on:
 
 jobs:
   build:
-    runs-on: ubuntu-latest
+    # Native arm64 runner -- the images we build are arm64, and Packer's
+    # arm-image plugin skips its QEMU/binfmt chroot steps entirely when
+    # image_arch (set in dpx-buttonode.pkr.hcl) matches the host's actual
+    # runtime arch (pkg/builder/builder.go: `!ImageArch.IsNative()`). On
+    # ubuntu-latest (x86_64), every apt-get/dpkg inside the chroot ran
+    # through qemu-aarch64-static emulation, which is commonly 5-10x
+    # slower than native for that kind of CPU-bound work.
+    runs-on: ubuntu-24.04-arm
     permissions:
       contents: read
     outputs:
@@ -294,7 +301,15 @@ jobs:
       - name: Checkout repository
         uses: actions/checkout@v4
 
-      # ARM emulation is required to chroot into the ARM64 Armbian image
+      # Still needed even on a native arm64 runner: the plugin's Prepare()
+      # step unconditionally resolves a qemu_binary path via exec.LookPath
+      # regardless of whether it'll actually be used later -- only its
+      # separate Run()-time IsNative() check decides whether QEMU is
+      # actually invoked for the chroot. Without this installed, Prepare()
+      # falls back to an "embedded qemu" feature that's amd64-only and
+      # fails outright on arm64 (confirmed: "embedded qemu is not
+      # available - currently, embedded qemu is only available for linux
+      # amd64"). This is just a fast package install, not the slow part.
       - name: Install QEMU user-static
         run: sudo apt-get update -q && sudo apt-get install -y qemu-user-static
 

.github/workflows/artifact-sweep.yaml

@@ -0,0 +1,35 @@
+name: Sweep expired Actions artifacts
+
+# Safety net for issue #18: release-action.yaml's own artifacts are
+# deleted immediately once they land in a release, but stray builds that
+# never go through that job (feature branches, force-rebuilds, manual
+# workflow_dispatch runs someone kicked off and forgot about) just sit
+# there. GitHub's own retention-days cleanup can lag by weeks in
+# practice -- confirmed 2026-09-05, 12.1GiB of artifacts sitting around
+# up to three weeks past their own expiry -- so this sweeps anything
+# already past expires_at rather than trusting GitHub to do it.
+on:
+  schedule:
+    - cron: '0 5 * * 0'  # weekly, Sunday 05:00 UTC
+  workflow_dispatch: {}
+
+jobs:
+  sweep:
+    runs-on: ubuntu-latest
+    permissions:
+      actions: write
+    steps:
+      - name: Delete artifacts past their own expiry
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          NOW=$(date -u +%s)
+          gh api "repos/${{ github.repository }}/actions/artifacts" --paginate \
+            --jq '.artifacts[] | [.id, .expires_at] | @tsv' | while IFS=$'\t' read -r id expires_at; do
+            [[ -z "$expires_at" || "$expires_at" == "null" ]] && continue
+            expires_epoch=$(date -u -d "$expires_at" +%s 2>/dev/null || echo 0)
+            if (( expires_epoch > 0 && expires_epoch < NOW )); then
+              echo "Deleting expired artifact $id (expired $expires_at)"
+              gh api -X DELETE "repos/${{ github.repository }}/actions/artifacts/$id" || true
+            fi
+          done

.github/workflows/raspios-builder.yaml

@@ -41,7 +41,12 @@ on:
 
 jobs:
   build:
-    runs-on: ubuntu-latest
+    # See armbian-builder.yaml's build job for why: native arm64 runner
+    # lets Packer's arm-image plugin skip QEMU/binfmt entirely
+    # (image_arch in dpx-buttonode.pkr.hcl matches the host arch), instead
+    # of emulating every chroot apt-get/dpkg call via qemu-aarch64-static
+    # on an x86_64 runner.
+    runs-on: ubuntu-24.04-arm
     permissions:
       contents: read
     outputs:
@@ -52,8 +57,9 @@ jobs:
       - name: Checkout repository
         uses: actions/checkout@v4
 
-      # ARM emulation is required to chroot into the ARM64 image, same as
-      # the Armbian pipeline -- unrelated to which OS built the base image.
+      # Still needed on a native arm64 runner -- see armbian-builder.yaml's
+      # build job for why (Prepare()-time qemu_binary resolution is
+      # unconditional, only actual usage is skipped on native arch).
       - name: Install QEMU user-static
         run: sudo apt-get update -q && sudo apt-get install -y qemu-user-static
 

.github/workflows/release-action.yaml

@@ -139,6 +139,7 @@ jobs:
     runs-on: ubuntu-latest
     permissions:
       contents: write
+      actions: write
 
     steps:
       - name: Checkout repository
@@ -180,3 +181,20 @@ jobs:
             --notes-file /tmp/release-notes.md \
             release-assets/*.img.gz \
             /tmp/buttons-version.txt
+
+      # Once the images are attached to the release as real assets, the
+      # raw CI artifacts this job downloaded from `build` serve no purpose
+      # -- delete them immediately rather than let retention-days expire
+      # them on GitHub's own timeline (issue #18: found 12.1GiB of already-
+      # expired-but-uncollected artifacts sitting around, silently
+      # billing against the account's Actions storage cap).
+      - name: Delete this run's CI artifacts (now redundant with the release)
+        if: success()
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          gh api "repos/${{ github.repository }}/actions/runs/${{ github.run_id }}/artifacts" \
+            --jq '.artifacts[].id' | while read -r id; do
+            echo "Deleting artifact $id"
+            gh api -X DELETE "repos/${{ github.repository }}/actions/artifacts/$id" || true
+          done

ACTION-PLAN.md

@@ -111,6 +111,13 @@ falling back to splash only if none is." Solving that cleanly handles both the
 the right thing should win automatically" case (#12) with one mechanism instead
 of two bolted-on fixes that could disagree with each other.
 
+**Implemented 2026-09-05**, in `scripts/install-deck-splash.sh`:
+- `OnFailure=dpx-deck-splash.service` drop-ins (`/etc/systemd/system/<mode>.service.d/dpx-recovery.conf`) on all three mode units. Drop-ins, not direct edits, since all three ship from vendor `.deb`s, not this repo — survives a package upgrade.
+- `dpx-mode-select.service` (new oneshot, `WantedBy=multi-user.target`): reads `/etc/dpx-mode` at boot and starts exactly that one mode service, falling back to `dpx-deck-splash.service` if nothing's persisted or the target refuses to start.
+- `dpx-deck-splash.service` no longer auto-enabled — it's only ever started by the fallback above or by an `OnFailure` recovery, never racing the mode service for `multi-user.target` on its own.
+- **Not yet live-verified** (no device access this pass) — needs a real boot-cycle test: confirm the persisted mode wins every time, and force a mode service into permanent failure (e.g. `systemctl kill` past its restart burst) to confirm splash actually comes back.
+- Also worth re-checking against this fix once live: the reported "device is already in a mode but not started, hitting GO does not start the thing" symptom. `execute_staged()`'s `mode_dead` check in `dpx-deck-splash.py` already looks correct on paper (re-applies if the persisted mode's service isn't actually active) — this may already have been a downstream effect of the same boot race rather than a separate bug. Confirm rather than assume once testable.
+
 Dashboard's own boot-time auto-start (the "and dashboard on/off" half of #12) is
 simpler and independent of the above — it's just "should `dpx-dashboard.service`
 be enabled or not," already a persisted systemd state via `set_dashboard_enabled()`,
@@ -128,3 +135,41 @@ new for it.
 5. #11 + #12 together (the real design work — biggest single piece here)
 6. #17's "doesn't launch" half — verify once a fresh build exists (falls out of #11/#12 work naturally, since that's a rebuild anyway)
 7. #10 and #17's "doesn't reflect state" half — both need live device access, batch them into one SSH session once available
+
+---
+
+## #18 — GitHub Actions artifact storage cleanup (housekeeping)
+
+**2026-09-05: found and fixed once.** `gh api repos/.../actions/artifacts` showed
+10 artifacts totaling 12.1 GiB, ALL already past their `retention-days: 3`
+expiration (the oldest by three weeks) but never garbage-collected by GitHub —
+they were still billing against the 2GB storage cap the whole time. Deleted
+manually via `gh api -X DELETE .../actions/artifacts/<id>`, storage now at 0.
+
+Not a one-time cleanup — this will silently refill: `release-action.yaml`'s
+nightly cron (`0 6 * * *`) builds new images whenever the Buttons mirror has an
+unreleased version, `armbian-builder.yaml`/`raspios-builder.yaml` upload
+1-1.7GB artifacts per board/variant, and the `release` job downloads them into
+a GitHub Release but never deletes the source CI artifacts afterward — nothing
+sweeps them once `retention-days` lapses, they just sit there until someone
+notices.
+
+**Squash it down properly, don't just re-delete manually next time:**
+- Add a step at the end of `release-action.yaml`'s `release` job (after the
+  release is successfully created) that deletes the just-downloaded build
+  artifacts immediately via `gh api -X DELETE` — once they're in the
+  release as `.img.gz` assets, the raw CI artifacts serve no purpose.
+- Consider a small separate scheduled workflow (e.g. weekly) that lists and
+  deletes any artifact past its `expires_at`, as a safety net for stray
+  manual/debug-branch builds (feature branches, force-rebuilds) that don't
+  go through the release job at all.
+- Low priority relative to #10-#17, but cheap to build once — fold into
+  the work whenever convenient, or do it standalone.
+
+**Implemented 2026-09-05**: `release-action.yaml`'s `release` job now deletes
+its own run's CI artifacts immediately after the release is created
+(`actions: write` added to its permissions). New `artifact-sweep.yaml`
+workflow runs weekly (Sunday 05:00 UTC) plus `workflow_dispatch`, deleting
+any artifact anywhere in the repo already past its own `expires_at` — the
+same category of already-expired-but-uncollected artifact found and
+manually cleared this session.

dpx-buttonode.pkr.hcl

@@ -73,6 +73,7 @@ source "arm-image" "base" {
   iso_url           = var.url
   target_image_size = var.variant == "full" ? 8000000000 : 5000000000
   output_filename   = "output-dpx-buttonode/dpx-buttonode.img"
+  image_arch      = "arm64"
   qemu_binary     = "qemu-aarch64-static"
   image_mounts    = var.image_mounts
 

scripts/install-dashboard.sh

@@ -86,7 +86,7 @@ xset s off
 xset s noblank
 unclutter -idle 0.5 -root &
 openbox-session &
-exec companion-dashboard --kiosk --no-sandbox
+exec companion-dashboard --kiosk-mode --no-sandbox
 XINITRC
 chmod +x "$DASH_HOME/.xinitrc"
 
@@ -113,7 +113,7 @@ chown -R dpx-dashboard:dpx-dashboard "$DASH_HOME"
 cat > /etc/systemd/system/dpx-dashboard.service << 'UNIT'
 [Unit]
 Description=Companion Dashboard Display Service
-After=network-online.target graphical.target
+After=network-online.target
 Wants=network-online.target
 
 [Service]
@@ -129,7 +129,7 @@ StandardOutput=journal
 StandardError=journal
 
 [Install]
-WantedBy=graphical.target
+WantedBy=multi-user.target
 UNIT
 
 systemctl daemon-reload

scripts/install-deck-splash.sh

@@ -105,8 +105,81 @@ KillMode=process
 WantedBy=multi-user.target
 UNIT
 
-systemctl enable dpx-deck-splash.service
-echo "==> dpx-deck-splash.service: enabled"
+# NOT enabled directly. dpx-mode-select.service (below) is now the only
+# thing that starts this at boot -- only as the no-persisted-mode
+# fallback -- instead of both it and the current mode service racing
+# multi-user.target with Conflicts= picking whichever happens to win
+# (dpx#12, confirmed nondeterministic on hardware). The [Install] block
+# stays so `systemctl enable dpx-deck-splash.service` still works for
+# anyone who wants the old always-auto-start behavior back.
+echo "==> dpx-deck-splash.service: installed (started via dpx-mode-select.service, not auto-enabled)"
+
+# ── Recovery: bring the splash back if a mode service dies for good ────────
+# OnFailure= only fires when a unit's ActiveState actually reaches
+# "failed" -- with Restart=on-failure, systemd holds the unit in
+# "activating (auto-restart)" between individual retry attempts, and
+# only lands in "failed" once StartLimitBurst is exhausted. So this
+# fires once per real, permanent outage, not once per transient restart
+# (dpx#11 -- "what's not clear is when the splash comes back"). Purely
+# event-driven, no polling loop.
+#
+# Drop-ins, not edits to the vendor unit files themselves -- all three
+# mode services ship from their own .deb packages (Buttons/Satellite/
+# Companion), not this repo, and a drop-in survives a package upgrade
+# that a direct edit wouldn't.
+for MODE_UNIT in bitfocus-buttons-usb-relay.service satellite.service companion.service; do
+    mkdir -p "/etc/systemd/system/${MODE_UNIT}.d"
+    cat > "/etc/systemd/system/${MODE_UNIT}.d/dpx-recovery.conf" << 'UNIT'
+[Unit]
+OnFailure=dpx-deck-splash.service
+UNIT
+done
+echo "==> OnFailure=dpx-deck-splash.service drop-ins installed for all 3 mode services"
+
+# ── Boot-time mode selection: exactly one of {persisted mode, splash} ──────
+# The other half of dpx#12/dpx#11: decide once, at boot, which single
+# thing should run instead of leaving it to a Conflicts= race. Reads
+# /etc/dpx-mode (same file switch_mode() in dpx-buttonode-ui.py writes)
+# and starts that mode's service; falls back to the splash if nothing's
+# persisted or the target service refuses to start. Mirrors
+# get_dpx_mode()'s own "buttons" default for consistency.
+cat > /usr/local/bin/dpx-mode-select.sh << 'SCRIPT'
+#!/usr/bin/env bash
+set -u
+# Existence-based toggle, same convention as /var/lib/dpx-hostname-set --
+# absent (the default on a fresh image) means autostart is ON, matching
+# the web UI's Mode tab checkbox defaulting to checked.
+if [ -e /var/lib/dpx-mode-autostart-disabled ]; then
+    systemctl start dpx-deck-splash.service
+    exit 0
+fi
+MODE="$(cat /etc/dpx-mode 2>/dev/null || echo "buttons")"
+case "$MODE" in
+    buttons)   SVC="bitfocus-buttons-usb-relay.service" ;;
+    satellite) SVC="satellite.service" ;;
+    companion) SVC="companion.service" ;;
+    *)         SVC="bitfocus-buttons-usb-relay.service" ;;
+esac
+systemctl start "$SVC" || systemctl start dpx-deck-splash.service
+SCRIPT
+chmod +x /usr/local/bin/dpx-mode-select.sh
+
+cat > /etc/systemd/system/dpx-mode-select.service << 'UNIT'
+[Unit]
+Description=Start the persisted dpx-buttonode mode (fallback: deck splash)
+Documentation=https://github.com/dubpixel/dpx_buttonode
+After=dpx-set-hostname.service
+
+[Service]
+Type=oneshot
+ExecStart=/usr/local/bin/dpx-mode-select.sh
+
+[Install]
+WantedBy=multi-user.target
+UNIT
+
+systemctl enable dpx-mode-select.service
+echo "==> dpx-mode-select.service: enabled"
 
 # ── sudoers: the ONLY door from dpx-splash (buttons group, nothing else)
 # to actually changing system state ─────────────────────────────────────────