fix(agent): add shutdown signal to response cache cleanup loop and annotate TUI forwarder handle

CHANGELOG.md

@@ -14,6 +14,12 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 - `zeph-plugins`: replaced blocking `std::fs::write` / `std::fs::read_to_string` calls in the
   async `update_one_plugin` with `tokio::fs` equivalents, preventing Tokio thread starvation
   during concurrent auto-update checks (closes #4560).
+- `apply_response_cache`: hourly cleanup background task now exits cleanly when the session shuts
+  down. The loop uses `tokio::select!` on a `CancellationToken` child of `mem_cancel`, which is
+  already cancelled via `shutdown_rx` at session teardown. Closes #4572.
+- `runner`: TUI early-status-forwarder `JoinHandle` annotated as intentionally dropped at block
+  end (self-terminating when the channel closes). Removes misleading bare `let _` and satisfies
+  `clippy::let_underscore_future`. Closes #4571.
 
 ### Added
 

src/agent_setup.rs

@@ -668,6 +668,7 @@ pub(crate) fn apply_response_cache<C: Channel>(
     ttl_secs: u64,
     semantic_cache_enabled: bool,
     embed_model: String,
+    cancel: tokio_util::sync::CancellationToken,
 ) -> Agent<C> {
     if !enabled {
         if semantic_cache_enabled {
@@ -681,11 +682,18 @@ pub(crate) fn apply_response_cache<C: Channel>(
         let mut interval = tokio::time::interval(std::time::Duration::from_hours(1));
         interval.tick().await; // skip immediate first tick
         loop {
-            interval.tick().await;
-            match cache_clone.cleanup(&embed_model).await {
-                Ok(n) if n > 0 => tracing::debug!("cleaned up {n} cache entries"),
-                Ok(_) => {}
-                Err(e) => tracing::warn!("response cache cleanup failed: {e:#}"),
+            tokio::select! {
+                () = cancel.cancelled() => {
+                    tracing::debug!("response cache cleanup loop: shutting down");
+                    break;
+                }
+                _ = interval.tick() => {
+                    match cache_clone.cleanup(&embed_model).await {
+                        Ok(n) if n > 0 => tracing::debug!("cleaned up {n} cache entries"),
+                        Ok(_) => {}
+                        Err(e) => tracing::warn!("response cache cleanup failed: {e:#}"),
+                    }
+                }
             }
         }
     });
@@ -1702,7 +1710,9 @@ mod tests {
         let db_url = format!("sqlite:{}", tmp.path().display());
         let pool = zeph_db::sqlx::SqlitePool::connect(&db_url).await.unwrap();
         let agent = make_agent();
-        let result = apply_response_cache(agent, false, pool, 300, false, "embed-model".into());
+        let cancel = tokio_util::sync::CancellationToken::new();
+        let result =
+            apply_response_cache(agent, false, pool, 300, false, "embed-model".into(), cancel);
         drop(result);
     }
 
@@ -1712,7 +1722,9 @@ mod tests {
         let db_url = format!("sqlite:{}", tmp.path().display());
         let pool = zeph_db::sqlx::SqlitePool::connect(&db_url).await.unwrap();
         let agent = make_agent();
-        let result = apply_response_cache(agent, true, pool, 300, false, "embed-model".into());
+        let cancel = tokio_util::sync::CancellationToken::new();
+        let result =
+            apply_response_cache(agent, true, pool, 300, false, "embed-model".into(), cancel);
         drop(result);
     }
 

src/runner.rs

@@ -954,6 +954,7 @@ pub(crate) async fn run(cli: Cli) -> anyhow::Result<()> {
             // dropped at the end of bootstrap. The TUI thread observes the channel close and
             // shuts down independently, so explicit abort is not needed. Dropping the handle
             // is intentional — we have no cleanup to do on the bootstrap error path here.
+            // EXEMPT: self-terminating on channel close — handle dropped intentionally at block end
             let _early_status_forwarder = tokio::spawn(crate::tui_bridge::forward_status_to_tui(
                 status_rx,
                 early.agent_tx.clone(),
@@ -2459,6 +2460,7 @@ pub(crate) async fn run(cli: Cli) -> anyhow::Result<()> {
         config.llm.response_cache_ttl_secs,
         config.llm.semantic_cache_enabled,
         crate::bootstrap::effective_embedding_model(config),
+        mem_cancel.child_token(),
     );
     let agent = agent_setup::apply_cost_tracker(agent, config);
     let agent = agent_setup::apply_summary_provider(agent, summary_provider);