diff --git a/app/Provisioning/Steps/Customer/CloneVirtualMachine.php b/app/Provisioning/Steps/Customer/CloneVirtualMachine.php
index ec7d03d..feb05fb 100644
--- a/app/Provisioning/Steps/Customer/CloneVirtualMachine.php
+++ b/app/Provisioning/Steps/Customer/CloneVirtualMachine.php
@@ -3,6 +3,7 @@
 namespace App\Provisioning\Steps\Customer;
 
 use App\Models\ProvisioningRun;
+use App\Models\RunResource;
 use App\Provisioning\StepResult;
 use App\Services\Proxmox\ProxmoxClient;
 
@@ -48,13 +49,20 @@ class CloneVirtualMachine extends CustomerStep
         // Clone once, keyed to the reserved vmid; record the task ref immediately.
         $upid = $run->context('clone_upid');
         if ($upid === null) {
-            // Crash-recovery: if the VM already exists, a prior attempt cloned it
-            // (but lost the task ref) — don't re-clone. Wait until the clone lock
-            // clears before advancing, in case the copy is still running.
+            // Crash-recovery: the VM exists but the clone task ref was lost, so we
+            // can't confirm the clone succeeded. If it's still locked, keep waiting;
+            // once unlocked, destroy the possibly-incomplete VM and re-clone cleanly.
             if ($pve->vmExists($node, $vmid)) {
-                return empty($pve->vmStatus($node, $vmid)['lock'])
-                    ? StepResult::advance()
-                    : StepResult::poll(10, 'finishing recovered clone');
+                if (! empty($pve->vmStatus($node, $vmid)['lock'])) {
+                    return StepResult::poll(10, 'finishing recovered clone');
+                }
+
+                $pve->deleteVm($node, $vmid);
+                RunResource::query()->where('run_id', $run->id)->where('kind', 'vmid')->delete();
+                $run->forgetContext('vmid');
+                $run->forgetContext('clone_upid');
+
+                return StepResult::retry(5, 're-cloning after a lost clone task reference');
             }
 
             $upid = $pve->cloneVm($node, $templateVmid, $vmid, $name);
diff --git a/app/Services/Proxmox/FakeProxmoxClient.php b/app/Services/Proxmox/FakeProxmoxClient.php
index 6d5edb0..c9e4cec 100644
--- a/app/Services/Proxmox/FakeProxmoxClient.php
+++ b/app/Services/Proxmox/FakeProxmoxClient.php
@@ -130,6 +130,16 @@ class FakeProxmoxClient implements ProxmoxClient
         return in_array($vmid, $this->clonedVmids, true) || in_array($vmid, $this->runningVmids, true);
     }
 
+    /** @var array<int, int> */
+    public array $deletedVmids = [];
+
+    public function deleteVm(string $node, int $vmid): void
+    {
+        $this->deletedVmids[] = $vmid;
+        $this->clonedVmids = array_values(array_diff($this->clonedVmids, [$vmid]));
+        $this->runningVmids = array_values(array_diff($this->runningVmids, [$vmid]));
+    }
+
     public function guestAgentPing(string $node, int $vmid): bool
     {
         return $this->guestAgentUp;
diff --git a/app/Services/Proxmox/HttpProxmoxClient.php b/app/Services/Proxmox/HttpProxmoxClient.php
index 789f637..4bc5969 100644
--- a/app/Services/Proxmox/HttpProxmoxClient.php
+++ b/app/Services/Proxmox/HttpProxmoxClient.php
@@ -89,6 +89,11 @@ class HttpProxmoxClient implements ProxmoxClient
         return $this->http()->get("/nodes/{$node}/qemu/{$vmid}/status/current")->successful();
     }
 
+    public function deleteVm(string $node, int $vmid): void
+    {
+        $this->http()->asForm()->delete("/nodes/{$node}/qemu/{$vmid}", ['purge' => 1])->throw();
+    }
+
     public function guestAgentPing(string $node, int $vmid): bool
     {
         return $this->http()->asForm()->post("/nodes/{$node}/qemu/{$vmid}/agent/ping")->successful();
@@ -134,6 +139,12 @@ class HttpProxmoxClient implements ProxmoxClient
 
     public function applyFirewall(string $node, int $vmid, array $rules): void
     {
+        // Clear existing rules first so a retry doesn't accumulate duplicates.
+        $existing = $this->http()->get("/nodes/{$node}/qemu/{$vmid}/firewall/rules")->throw()->json('data', []);
+        for ($pos = count($existing) - 1; $pos >= 0; $pos--) {
+            $this->http()->delete("/nodes/{$node}/qemu/{$vmid}/firewall/rules/{$pos}")->throw();
+        }
+
         foreach ($rules as $rule) {
             $this->http()->asForm()->post("/nodes/{$node}/qemu/{$vmid}/firewall/rules", $rule)->throw();
         }
diff --git a/app/Services/Proxmox/ProxmoxClient.php b/app/Services/Proxmox/ProxmoxClient.php
index 981b69d..065200d 100644
--- a/app/Services/Proxmox/ProxmoxClient.php
+++ b/app/Services/Proxmox/ProxmoxClient.php
@@ -45,6 +45,8 @@ interface ProxmoxClient
     /** Whether a VM with this id already exists on the node. */
     public function vmExists(string $node, int $vmid): bool;
 
+    public function deleteVm(string $node, int $vmid): void;
+
     public function guestAgentPing(string $node, int $vmid): bool;
 
     /** @return array{exitcode: int, out-data: string} */
diff --git a/tests/Feature/Provisioning/CustomerStepsTest.php b/tests/Feature/Provisioning/CustomerStepsTest.php
index 5c8a11c..f9c00ac 100644
--- a/tests/Feature/Provisioning/CustomerStepsTest.php
+++ b/tests/Feature/Provisioning/CustomerStepsTest.php
@@ -121,21 +121,24 @@ it('polls while cloning and fails on a clone error', function () {
     expect(app(CloneVirtualMachine::class)->execute($run2)->type)->toBe('fail');
 });
 
-it('recovers a clone whose task ref was lost (no duplicate clone)', function () {
+it('recovers a lost-task clone by waiting for the lock then re-cloning cleanly', function () {
     $s = fakeServices();
     ['run' => $run] = reservedRun([], ['vmid' => 101]);
-    // Simulate a crash after the clone but before the upid was persisted:
-    // vmid reserved + breadcrumb, no clone_upid, and the VM already exists.
+    // Crash after cloneVm() but before clone_upid was persisted: vmid reserved +
+    // breadcrumb, no clone_upid, and the VM already exists.
     RunResource::create(['run_id' => $run->id, 'kind' => 'vmid', 'external_id' => '101']);
     $run->mergeContext(['vmid' => 101]);
-    $s['pve']->clonedVmids[] = 101; // VM exists
+    $s['pve']->clonedVmids[] = 101;
 
-    expect(app(CloneVirtualMachine::class)->execute($run->fresh())->type)->toBe('advance')
-        ->and(count($s['pve']->clonedVmids))->toBe(1); // no second clone
-
-    // still locked (clone copy running) → poll, don't advance onto an incomplete VM
+    // still cloning (locked) → poll, don't touch it
     $s['pve']->vmLock = 'clone';
     expect(app(CloneVirtualMachine::class)->execute($run->fresh())->type)->toBe('poll');
+
+    // unlocked but task status unknown → destroy the maybe-incomplete VM + re-clone
+    $s['pve']->vmLock = null;
+    expect(app(CloneVirtualMachine::class)->execute($run->fresh())->type)->toBe('retry');
+    expect($s['pve']->deletedVmids)->toContain(101)
+        ->and(RunResource::where('run_id', $run->id)->where('kind', 'vmid')->exists())->toBeFalse();
 });
 
 it('fails clone when the plan has no template', function () {