diff --git a/docs/4-implementation/swarm/controller.md b/docs/4-implementation/swarm/controller.md
index c5dc98ba3..94df163b7 100644
--- a/docs/4-implementation/swarm/controller.md
+++ b/docs/4-implementation/swarm/controller.md
@@ -195,7 +195,9 @@ status kept apart from the output. A task's status is one key, `<agent>.<task>`,
 `bash-task-status` bucket: its `state` (`pending`, `running`, `done`, `timed_out`,
 `interrupted` or `killed`), the command cut to 200 characters, when it began and when it
 entered its state, the exit status, and `ts`, the time of the publish. A live task republishes
-every 60 seconds, and a reader treats a live value older than three of those as stale. The output
+every 60 seconds, and a reader treats a live value older than three of those as stale. A key
+expires 72 hours after its last write, since nothing deletes a finished task's key, and a
+live task's republishing keeps its own. The output
 goes to `$SWARM.term.<agent>.bash.<task>`, with a `Swarm-Stream` header of `stdout` or
 `stderr`, into the stream `term-bash-<agent>` that swarm-controller creates for every declared
 agent, for 24 hours or 64 MiB like the subagent rows. The agent's credential may publish
diff --git a/swarm-queue-client/src/bash_task.rs b/swarm-queue-client/src/bash_task.rs
index b30593a98..910696db1 100644
--- a/swarm-queue-client/src/bash_task.rs
+++ b/swarm-queue-client/src/bash_task.rs
@@ -32,16 +32,19 @@ pub(crate) fn output_config(
 }
 
 /// Open the status bucket, creating it if nothing has yet. `history: 1`: a
-/// reader wants each task's current state, not every state it has been in.
+/// reader wants each task's current state, not every state it has been in. A
+/// key expires [`swarm_wire::bash_task::STATUS_MAX_AGE`] after its last write,
+/// since nothing deletes the key of a finished task.
 pub async fn open_status(
     client: &async_nats::Client,
 ) -> Result<async_nats::jetstream::kv::Store, crate::Error> {
     let js = async_nats::jetstream::new(client.clone());
-    crate::kv::open_or_create(
+    crate::kv::open_or_create_expiring(
         &js,
         swarm_wire::bash_task::BUCKET,
         "bash-task-status",
         "The state of each bash task, as its agent's bash MCP last published it".to_owned(),
+        Some(swarm_wire::bash_task::STATUS_MAX_AGE),
     )
     .await
 }
diff --git a/swarm-queue-client/src/kv.rs b/swarm-queue-client/src/kv.rs
index 9427a82a8..4cddcd2e6 100644
--- a/swarm-queue-client/src/kv.rs
+++ b/swarm-queue-client/src/kv.rs
@@ -9,6 +9,19 @@ pub(crate) async fn open_or_create(
     bucket: &str,
     label: &str,
     description: String,
+) -> Result<async_nats::jetstream::kv::Store, Error> {
+    open_or_create_expiring(js, bucket, label, description, None).await
+}
+
+/// [`open_or_create`] with a `max_age`: a key nobody writes for that long goes
+/// away, for a bucket whose keys come from a publisher that never deletes. The
+/// age applies to a bucket created here; one that already exists keeps its own.
+pub(crate) async fn open_or_create_expiring(
+    js: &async_nats::jetstream::Context,
+    bucket: &str,
+    label: &str,
+    description: String,
+    max_age: Option<std::time::Duration>,
 ) -> Result<async_nats::jetstream::kv::Store, Error> {
     match js.get_key_value(bucket).await {
         Ok(store) => Ok(store),
@@ -22,6 +35,7 @@ pub(crate) async fn open_or_create(
                 bucket: bucket.to_owned(),
                 description,
                 history: 1,
+                max_age: max_age.unwrap_or_default(),
                 ..Default::default()
             })
             .await
diff --git a/swarm-wire/src/bash_task.rs b/swarm-wire/src/bash_task.rs
index b2a5005ff..cdfc0dd9d 100644
--- a/swarm-wire/src/bash_task.rs
+++ b/swarm-wire/src/bash_task.rs
@@ -33,6 +33,11 @@ pub const HEARTBEAT_S: i64 = 60;
 /// A `running` status whose [`Status::ts`] is older than this is stale.
 pub const STALE_AFTER_S: i64 = 3 * HEARTBEAT_S;
 
+/// How long a task's key stays in the bucket after its last write. A live task
+/// is republished every [`HEARTBEAT_S`], which keeps its key; a finished one
+/// is only worth listing for a few days, and nothing else deletes the key.
+pub const STATUS_MAX_AGE: std::time::Duration = std::time::Duration::from_hours(72);
+
 /// The longest [`Status::cmd`], in characters; a longer command is cut.
 pub const CMD_MAX_CHARS: usize = 200;
 
@@ -218,6 +223,14 @@ mod tests {
         assert!(!status(State::Done, 10_000).is_running(10_000));
     }
 
+    /// A live task's key must outlive the gap between two heartbeats and the
+    /// stale window by a wide margin, or a running task could expire.
+    #[test]
+    fn a_key_outlives_the_stale_window_by_far() {
+        let stale = std::time::Duration::from_secs(STALE_AFTER_S.unsigned_abs());
+        assert!(STATUS_MAX_AGE >= stale * 100);
+    }
+
     #[test]
     fn a_long_command_is_cut_with_an_ellipsis() {
         assert_eq!(clip_cmd("ls"), "ls");