deploy guard: expose binary

.gitea/workflows/deploy.yml

@@ -65,10 +65,15 @@ jobs:
 
       - name: Deploy guard preflight
         run: |
+          # The gitea runner runs on muffin itself, so the /nix/store path
+          # produced by "Build muffin" is already resolvable on the target.
+          # SSH to root is still needed so the check can read agenix secrets
+          # (e.g. /run/agenix/jellyfin-api-key is mode 0400 root).
+          guard=$(readlink -f result)/sw/bin/deploy-guard-check
           ssh -i /run/agenix/ci-deploy-key \
             -o StrictHostKeyChecking=yes \
             -o UserKnownHostsFile=/etc/ci-known-hosts \
-            root@server-public deploy-guard-check
+            root@server-public "$guard"
 
       - name: Deploy via deploy-rs
         run: |

deploy.sh

@@ -13,6 +13,13 @@
 # the remote deploy via deploy-rs when explicitly named.
 #
 # DEPLOY_GUARD_FORCE=1 is equivalent to passing --force.
+#
+# The preflight builds the guard derivation locally, copies it to muffin's
+# nix store, then invokes it by /nix/store path over SSH as root (so the
+# jellyfin check can read /run/agenix/jellyfin-api-key). Building the exact
+# binary we're about to deploy avoids the bootstrap gap where
+# /run/current-system/sw/bin/deploy-guard-check may not yet exist on the target
+# (first deploy of the feature, post-rollback wiping it, etc).
 
 set -eu
 
@@ -30,11 +37,15 @@ case "$arg" in
     if [ "$force" = "1" ]; then
       echo "deploy-guard: preflight skipped (--force)"
     else
-      # Single SSH probe. Exit 255 is a connectivity failure; treat as a hard
-      # abort — without the preflight there is no other gate that prevents
-      # deploy-rs from partially activating while users are online.
+      echo "deploy-guard: building preflight binary..."
+      guard=$(nix build --no-link --print-out-paths \
+        '.#nixosConfigurations.muffin.config.system.build.deployGuardCheck')
+
+      echo "deploy-guard: copying to muffin..."
+      nix copy --to ssh-ng://root@server-public "$guard"
+
       output=$(ssh -o BatchMode=yes -o ConnectTimeout=5 \
-        root@server-public deploy-guard-check 2>&1) && rc=0 || rc=$?
+        root@server-public "$guard/bin/deploy-guard-check" 2>&1) && rc=0 || rc=$?
 
       if [ "$rc" -eq 0 ]; then
         [ -n "$output" ] && printf '%s\n' "$output"

modules/server-deploy-guard.nix

@@ -159,6 +159,13 @@ in
   config = lib.mkIf cfg.enable {
     environment.systemPackages = [ aggregator ];
 
+    # Expose the aggregator as a named system build so preflight drivers
+    # (deploy.sh, CI) can build just this derivation and invoke it by its
+    # /nix/store path — avoiding the bootstrap gap where
+    # /run/current-system/sw/bin/deploy-guard-check may not yet exist on the
+    # target (first deploy of the feature, post-rollback, etc).
+    system.build.deployGuardCheck = aggregator;
+
     assertions = [
       {
         assertion = cfg.checks != { };