diff --git a/NEWS.md b/NEWS.md
index d2106035..4b3fcd6e 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -23,6 +23,12 @@
 
  * Futures failed on R (< 4.0.0). This bug was introduced in **future**
    1.68.0 (2025-11-17).
+
+ * When a cluster worker that `plan()` had started was relaunched after
+   it was canceled, interrupted or had died, `plan()` never shut the
+   relaunched worker down, because the cluster registry kept the old
+   node. The worker process and its connection were left running until
+   garbage collection.
    
 
 # Version 1.76.0 [2026-09-24]
diff --git a/R/backend_api-11.ClusterFutureBackend-class.R b/R/backend_api-11.ClusterFutureBackend-class.R
index 7d289ff7..95fffa52 100644
--- a/R/backend_api-11.ClusterFutureBackend-class.R
+++ b/R/backend_api-11.ClusterFutureBackend-class.R
@@ -1231,6 +1231,7 @@ requestNode <- local({
 
         workers[[node_idx]] <- node2
         backend[["workers"]] <- workers
+        clusterRegistry$replaceNode(workers, node_idx, node2)
         node <- node2
 
         if (debug) {
@@ -1609,6 +1610,7 @@ handleInterruptedFuture <- local({
 
         ## Update backend
         backend[["workers"]] <- workers
+        clusterRegistry$replaceNode(workers, node_idx, node2)
 
         node <- NULL
         
@@ -1743,6 +1745,18 @@ clusterRegistry <- local({
     cluster
   } ## startCluster()
 
+  ## When a worker is relaunched, the registry must track the new node,
+  ## otherwise stopCluster() closes the old, already closed, node and leaves
+  ## the relaunched worker and its connection behind
+  replaceNode <- function(workers, idx, node) {
+    if (is.null(cluster)) return(invisible(FALSE))
+    if (!identical(attr(workers, "name", exact = TRUE), attr(cluster, "name", exact = TRUE))) {
+      return(invisible(FALSE))
+    }
+    cluster[[idx]] <<- node
+    invisible(TRUE)
+  } ## replaceNode()
+
   stopCluster <- function(debug = FALSE) {
     if (debug) {
       mdebug_push("Stopping existing cluster ...")
@@ -1800,6 +1814,7 @@ clusterRegistry <- local({
   list(
       getCluster = getCluster,
     startCluster = startCluster,
+     replaceNode = replaceNode,
      stopCluster = stopCluster
   )
 }) ## clusterRegistry()
diff --git a/inst/testme/test-cancel-relaunch-cluster.R b/inst/testme/test-cancel-relaunch-cluster.R
new file mode 100644
index 00000000..b62e0100
--- /dev/null
+++ b/inst/testme/test-cancel-relaunch-cluster.R
@@ -0,0 +1,40 @@
+#' @tags cancel
+#' @tags detritus-files
+#' @tags detritus-connections
+#' @tags multisession
+
+library(future)
+
+message("Relaunched cluster workers are shut down with the plan ...")
+
+covr_testing <- ("covr" %in% loadedNamespaces())
+
+## Terminated workers can leave truncated coverage traces.
+if (!covr_testing) {
+  n0 <- nrow(showConnections())
+
+  plan(multisession, workers = I(1))
+
+  ## Cancel a running future, which terminates its worker
+  f <- future({ Sys.sleep(30); 42 })
+  Sys.sleep(1.0)
+  f <- cancel(f)
+
+  ## The next future relaunches the terminated worker
+  f2 <- future(42)
+  stopifnot(value(f2) == 42)
+
+  ## Keep a reference, so that garbage collection cannot hide the leak
+  backend <- plan("backend")
+
+  plan(sequential)
+
+  ## The relaunched worker must be closed by plan(), not by garbage collection
+  n <- nrow(showConnections())
+  message("Connections: before = ", n0, ", after = ", n)
+  stopifnot(n <= n0)
+  rm(list = "backend")
+  gc()
+}
+
+message("Relaunched cluster workers are shut down with the plan ... done")
diff --git a/tests/test-cancel-relaunch-cluster.R b/tests/test-cancel-relaunch-cluster.R
new file mode 100644
index 00000000..2d530ad3
--- /dev/null
+++ b/tests/test-cancel-relaunch-cluster.R
@@ -0,0 +1,4 @@
+#! /usr/bin/env Rscript
+## This runs testme test script inst/testme/test-cancel-relaunch-cluster.R
+## Don't edit - it was autogenerated by inst/testme/deploy.R
+future:::testme("cancel-relaunch-cluster")