diff --git a/crates/buzz-db/src/migration.rs b/crates/buzz-db/src/migration.rs
index be87faa1ac..021af5d223 100644
--- a/crates/buzz-db/src/migration.rs
+++ b/crates/buzz-db/src/migration.rs
@@ -625,7 +625,7 @@ mod tests {
         let mut migrations: Vec<_> = MIGRATOR.iter().collect();
         migrations.sort_by_key(|migration| migration.version);
 
-        assert_eq!(migrations.len(), 31);
+        assert_eq!(migrations.len(), 32);
         assert_eq!(migrations[0].version, 1);
         assert_eq!(&*migrations[0].description, "initial schema");
         assert!(migrations[0]
@@ -1057,6 +1057,25 @@ mod tests {
             .as_str()
             .contains("error_code"));
         assert!(include_str!("../../../schema/schema.sql").contains("error_code          TEXT"));
+
+        // Thread subtree pagination uses the full tenant/root/keyset tuple. Keep
+        // the covering index additive so brownfield migration checksums remain
+        // stable.
+        assert_eq!(migrations[31].version, 32);
+        assert!(migrations[31]
+            .sql
+            .as_str()
+            .contains("CREATE INDEX IF NOT EXISTS idx_thread_metadata_root_keyset"));
+        assert!(migrations[31]
+            .sql
+            .as_str()
+            .contains("(community_id, root_event_id, event_created_at, event_id)"));
+        assert!(!migrations[0]
+            .sql
+            .as_str()
+            .contains("idx_thread_metadata_root_keyset"));
+        assert!(include_str!("../../../schema/schema.sql")
+            .contains("CREATE INDEX idx_thread_metadata_root_keyset"));
     }
 
     #[test]
diff --git a/migrations/0032_thread_metadata_keyset_index.sql b/migrations/0032_thread_metadata_keyset_index.sql
new file mode 100644
index 0000000000..9142f0bccf
--- /dev/null
+++ b/migrations/0032_thread_metadata_keyset_index.sql
@@ -0,0 +1,21 @@
+-- ── Thread reply keyset pagination index ─────────────────────────────────────
+-- Thread subtree reads filter by tenant and root, then page in ascending
+-- (event_created_at, event_id) order. Include the keyset columns so PostgreSQL
+-- can satisfy both the filter and ordering from one index scan.
+--
+-- Lock note: built without CONCURRENTLY because sqlx runs each migration inside
+-- a transaction and CREATE INDEX CONCURRENTLY cannot run in one. This takes a
+-- SHARE lock on `thread_metadata` (blocking writes, not reads) for the duration
+-- of the build. `thread_metadata` is on the event insertion path and may be
+-- large on brownfield databases, so operators should pre-build it by hand before
+-- deploying this migration:
+--
+--   CREATE INDEX CONCURRENTLY idx_thread_metadata_root_keyset
+--       ON thread_metadata (community_id, root_event_id, event_created_at, event_id);
+--
+-- IF NOT EXISTS then makes this migration a no-op on that database.
+--
+-- Additive migration: previously applied files must not change checksum.
+
+CREATE INDEX IF NOT EXISTS idx_thread_metadata_root_keyset
+    ON thread_metadata (community_id, root_event_id, event_created_at, event_id);
diff --git a/schema/schema.sql b/schema/schema.sql
index 9ef7bc0a4b..44bd90bb18 100644
--- a/schema/schema.sql
+++ b/schema/schema.sql
@@ -531,6 +531,8 @@ CREATE TABLE thread_metadata (
 
 CREATE INDEX idx_thread_metadata_parent ON thread_metadata (community_id, parent_event_id);
 CREATE INDEX idx_thread_metadata_root ON thread_metadata (community_id, root_event_id);
+CREATE INDEX idx_thread_metadata_root_keyset
+    ON thread_metadata (community_id, root_event_id, event_created_at, event_id);
 CREATE INDEX idx_thread_metadata_channel_depth
     ON thread_metadata (community_id, channel_id, depth, event_created_at);
 CREATE INDEX idx_thread_metadata_event_id ON thread_metadata (community_id, event_id);