diff --git a/src/lib.rs b/src/lib.rs
index 42af154..158225a 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -3391,7 +3391,7 @@ impl OrchestrationContext {
     }
 
     /// Clear all keys from the KV store.
-    ///
+    /// 
     /// Emits a `KeyValuesCleared` history event. After this call,
     /// `get_kv_value(key)` returns `None` for all keys.
     pub fn clear_all_kv_values(&self) {
@@ -3408,8 +3408,13 @@ impl OrchestrationContext {
     /// turn are never prunable (they haven't been acked yet).
     ///
     /// Returns the number of keys cleared.
+        ///
+        /// The keys are cleared in sorted order so that the emitted `ClearKeyValue`
+    /// actions are deterministic: `kv_metadata` is a `HashMap`, whose iteration
+    /// order differs between runs, and replay compares actions against recorded
+    /// events in order (fixes microsoft/duroxide#57).
     pub fn prune_kv_values_updated_before(&self, updated_before_ms: u64) -> usize {
-        let keys_to_clear: Vec<String> = {
+        let mut keys_to_clear: Vec<String> = {
             let inner = self.inner.lock().unwrap();
             inner
                 .kv_metadata
@@ -3419,6 +3424,8 @@ impl OrchestrationContext {
                 .map(|(key, _)| key.clone())
                 .collect()
         };
+        // Deterministic emission order across replays (see doc comment above).
+        keys_to_clear.sort();
         let count = keys_to_clear.len();
         for key in keys_to_clear {
             self.clear_kv_value(key);
diff --git a/tests/kv_store_tests.rs b/tests/kv_store_tests.rs
index b57c86a..c76052a 100644
--- a/tests/kv_store_tests.rs
+++ b/tests/kv_store_tests.rs
@@ -13,6 +13,7 @@
 mod common;
 
 use duroxide::runtime::{self, OrchestrationStatus, registry::ActivityRegistry};
+use duroxide::providers::Provider;
 use duroxide::{ActivityContext, OrchestrationContext, OrchestrationRegistry};
 use std::sync::Arc;
 use std::time::Duration;
@@ -2573,6 +2574,80 @@ async fn kv_prune_old_values_by_timestamp() {
     rt.shutdown(None).await;
 }
 
+/// Regression test for https://github.com/microsoft/duroxide/issues/57:
+/// prune_kv_values_updated_before must emit ClearKeyValue actions in a
+/// deterministic (sorted) key order. kv_metadata is a HashMap whose iteration
+/// order varies between runs, and the replay engine compares emitted actions
+/// against recorded events in order — unsorted emission fails replays with
+/// `nondeterministic: kv clear mismatch`.
+///
+/// Strategy: exec 1 sets four keys in deliberately unsorted insertion order,
+/// then ContinueAsNew to exec 2, which prunes them all. The KeyValueCleared
+/// events recorded in exec 2's history must appear in sorted key order.
+#[tokio::test]
+async fn kv_prune_emits_clears_in_sorted_key_order() {
+    let store = Arc::new(
+        duroxide::providers::sqlite::SqliteProvider::new_in_memory()
+            .await
+            .unwrap(),
+    );
+    let activities = ActivityRegistry::builder().build();
+    let orchestrations = OrchestrationRegistry::builder()
+        .register("PruneOrder", |ctx: OrchestrationContext, input: String| async move {
+            let n: u32 = input.parse().unwrap_or(0);
+            if n == 0 {
+                // Exec 1: set keys in unsorted insertion order, then CAN.
+                for key in ["zeta", "alpha", "mike", "beta"] {
+                    ctx.set_kv_value(key, "v");
+                }
+                ctx.continue_as_new("1".to_string()).await
+            } else {
+                // Exec 2: snapshot has the four keys with their exec-1 timestamps.
+                // Sleep briefly so that "now" is definitely after those timestamps.
+                ctx.schedule_timer(Duration::from_millis(50)).await;
+                let now = ctx
+                    .utc_now()
+                    .await
+                    .map_err(|e| e.to_string())?
+                    .duration_since(std::time::UNIX_EPOCH)
+                    .unwrap()
+                    .as_millis() as u64;
+                let removed = ctx.prune_kv_values_updated_before(now);
+                Ok(format!("removed:{removed}"))
+            }
+        })
+        .build();
+
+    let rt = runtime::Runtime::start_with_store(store.clone(), activities, orchestrations).await;
+    let client = duroxide::Client::new(store.clone());
+    client.start_orchestration("kv-prune-order", "PruneOrder", "0").await.unwrap();
+
+    let status = client
+        .wait_for_orchestration("kv-prune-order", Duration::from_secs(10))
+        .await
+        .unwrap();
+    match status {
+        OrchestrationStatus::Completed { output, .. } => assert_eq!(output, "removed:4"),
+        OrchestrationStatus::Failed { details, .. } => panic!("failed: {}", details.display_message()),
+        _ => panic!("unexpected status"),
+    }
+
+    // The recorded KeyValueCleared events must be in sorted key order so that a
+    // later replay of the same turn observes the same action sequence.
+    let mgmt = store.as_management_capability().expect("ProviderAdmin required");
+    let history = mgmt.read_history("kv-prune-order").await.unwrap();
+    let cleared: Vec<String> = history
+        .iter()
+        .filter_map(|e| match &e.kind {
+            duroxide::EventKind::KeyValueCleared { key } => Some(key.clone()),
+            _ => None,
+        })
+        .collect();
+    assert_eq!(cleared, vec!["alpha", "beta", "mike", "zeta"]);
+
+    rt.shutdown(None).await;
+}
+
 /// prune_kv_values_updated_before with threshold 0 does nothing (all timestamps >= 0).
 #[tokio::test]
 async fn kv_prune_with_zero_threshold_does_nothing() {