diff --git a/codex-rs/core/src/codex.rs b/codex-rs/core/src/codex.rs
index 6639a66db6..11d58ace3a 100644
--- a/codex-rs/core/src/codex.rs
+++ b/codex-rs/core/src/codex.rs
@@ -3759,6 +3759,7 @@ pub(crate) async fn run_turn(
                     &turn_context,
                     auto_compact_limit,
                     needs_follow_up,
+                    true,
                     "run_turn",
                 )
                 .await;
@@ -3836,6 +3837,7 @@ async fn log_post_sampling_token_usage_and_maybe_compact(
     turn_context: &Arc<TurnContext>,
     auto_compact_limit: i64,
     needs_follow_up: bool,
+    allow_auto_compact: bool,
     checkpoint: &'static str,
 ) -> bool {
     let total_usage_tokens = sess.get_total_token_usage().await;
@@ -3850,11 +3852,12 @@ async fn log_post_sampling_token_usage_and_maybe_compact(
         auto_compact_limit,
         token_limit_reached,
         needs_follow_up,
+        allow_auto_compact,
         checkpoint,
         "post sampling token usage"
     );
 
-    if token_limit_reached && needs_follow_up {
+    if allow_auto_compact && token_limit_reached && needs_follow_up {
         run_auto_compact(sess, turn_context).await;
     }
 
@@ -4500,6 +4503,7 @@ async fn drain_in_flight(
                     &turn_context,
                     auto_compact_limit,
                     true,
+                    in_flight.is_empty(),
                     "drain_in_flight",
                 )
                 .await;
@@ -4653,6 +4657,7 @@ async fn try_run_sampling_request(
                             &turn_context,
                             auto_compact_limit,
                             needs_follow_up,
+                            in_flight.is_empty(),
                             "stream_item_done_plan",
                         )
                         .await;
@@ -4682,6 +4687,7 @@ async fn try_run_sampling_request(
                     &turn_context,
                     auto_compact_limit,
                     needs_follow_up,
+                    in_flight.is_empty(),
                     "stream_item_done",
                 )
                 .await;
diff --git a/codex-rs/core/tests/suite/compact_remote.rs b/codex-rs/core/tests/suite/compact_remote.rs
index a06e46e75b..2f34290059 100644
--- a/codex-rs/core/tests/suite/compact_remote.rs
+++ b/codex-rs/core/tests/suite/compact_remote.rs
@@ -239,7 +239,7 @@ async fn remote_compact_runs_automatically() -> Result<()> {
 
 #[cfg_attr(target_os = "windows", ignore)]
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn remote_auto_compact_runs_during_multi_tool_output_drain() -> Result<()> {
+async fn remote_auto_compact_runs_after_multi_tool_output_drain() -> Result<()> {
     skip_if_no_network!(Ok(()));
 
     let first_call_id = "first-large-call";
@@ -257,7 +257,7 @@ async fn remote_auto_compact_runs_during_multi_tool_output_drain() -> Result<()>
     .await?;
     let codex = harness.test().codex.clone();
 
-    let responses_mock = responses::mount_sse_sequence(
+    responses::mount_sse_sequence(
         harness.server(),
         vec![
             sse(vec![
@@ -297,16 +297,9 @@ async fn remote_auto_compact_runs_during_multi_tool_output_drain() -> Result<()>
     );
     assert!(
         compact_request
-            .function_call_output_text(second_call_id)
-            .is_none(),
-        "expected compaction request before the second drained tool output"
-    );
-
-    assert!(
-        responses_mock
             .function_call_output_text(second_call_id)
             .is_some(),
-        "expected second tool output to be sent in follow-up sampling request"
+        "expected compaction request to include the second drained tool output"
     );
 
     Ok(())