limit automatic compactions per turn

codex-rs/core/src/session/turn.rs

@@ -114,6 +114,13 @@ use tracing::trace;
 use tracing::trace_span;
 use tracing::warn;
 
+const MAX_AUTO_COMPACTIONS_PER_TURN: usize = 3;
+
+pub(crate) enum RunTurnResult {
+    Continue(Option<String>),
+    Terminal,
+}
+
 /// Takes a user message as input and runs a loop where, at each sampling request, the model
 /// replies with either:
 ///
@@ -132,33 +139,44 @@ pub(crate) async fn run_turn(
     sess: Arc<Session>,
     turn_context: Arc<TurnContext>,
     turn_extension_data: Arc<codex_extension_api::ExtensionData>,
+    auto_compact_limiter: &mut AutoCompactTurnLimiter,
     input: Vec<UserInput>,
     prewarmed_client_session: Option<ModelClientSession>,
     cancellation_token: CancellationToken,
-) -> Option<String> {
+) -> RunTurnResult {
     let mut client_session =
         prewarmed_client_session.unwrap_or_else(|| sess.services.model_client.new_session());
     // TODO(ccunningham): Pre-turn compaction runs before context updates and the
     // new user message are recorded. Estimate pending incoming items (context
     // diffs/full reinjection + user input) and trigger compaction preemptively
     // when they would push the thread over the compaction threshold.
-    let pre_sampling_compact =
-        match run_pre_sampling_compact(&sess, &turn_context, &mut client_session).await {
-            Ok(pre_sampling_compact) => pre_sampling_compact,
-            Err(err) => {
-                if err.to_codex_protocol_error() == CodexErrorInfo::UsageLimitExceeded
-                    && let Err(err) = sess
-                        .goal_runtime_apply(GoalRuntimeEvent::UsageLimitReached {
-                            turn_context: turn_context.as_ref(),
-                        })
-                        .await
-                {
-                    warn!("failed to usage-limit active goal after usage-limit error: {err}");
-                }
-                error!("Failed to run pre-sampling compact");
-                return None;
+    let pre_sampling_compact = match run_pre_sampling_compact(
+        &sess,
+        &turn_context,
+        &mut client_session,
+        auto_compact_limiter,
+    )
+    .await
+    {
+        Ok(pre_sampling_compact) => pre_sampling_compact,
+        Err(err) => {
+            if err.to_codex_protocol_error() == CodexErrorInfo::UsageLimitExceeded
+                && let Err(err) = sess
+                    .goal_runtime_apply(GoalRuntimeEvent::UsageLimitReached {
+                        turn_context: turn_context.as_ref(),
+                    })
+                    .await
+            {
+                warn!("failed to usage-limit active goal after usage-limit error: {err}");
             }
-        };
+            error!("Failed to run pre-sampling compact");
+            return if err.to_codex_protocol_error() == CodexErrorInfo::ContextWindowExceeded {
+                RunTurnResult::Terminal
+            } else {
+                RunTurnResult::Continue(None)
+            };
+        }
+    };
     if pre_sampling_compact.reset_client_session {
         client_session.reset_websocket_session();
     }
@@ -166,11 +184,14 @@ pub(crate) async fn run_turn(
     sess.record_context_updates_and_set_reference_context_item(turn_context.as_ref())
         .await;
 
-    let (injection_items, explicitly_enabled_connectors) =
-        build_skills_and_plugins(&sess, turn_context.as_ref(), &input, &cancellation_token).await?;
+    let Some((injection_items, explicitly_enabled_connectors)) =
+        build_skills_and_plugins(&sess, turn_context.as_ref(), &input, &cancellation_token).await
+    else {
+        return RunTurnResult::Continue(None);
+    };
 
     if run_pending_session_start_hooks(&sess, &turn_context).await {
-        return None;
+        return RunTurnResult::Continue(None);
     }
     if !input.is_empty() {
         let initial_turn_input = TurnInput::UserInput(input.clone());
@@ -183,7 +204,7 @@ pub(crate) async fn run_turn(
                 user_prompt_submit_outcome.additional_contexts,
             )
             .await;
-            return None;
+            return RunTurnResult::Continue(None);
         }
         record_pending_input(
             &sess,
@@ -249,6 +270,7 @@ pub(crate) async fn run_turn(
         let mut blocked_pending_input = false;
         let mut accepted_pending_input = false;
         for pending_input_item in pending_input {
+            let is_user_input = matches!(&pending_input_item, TurnInput::UserInput(_));
             let hook_outcome =
                 inspect_pending_input(&sess, &turn_context, &pending_input_item).await;
             if hook_outcome.should_stop {
@@ -264,6 +286,9 @@ pub(crate) async fn run_turn(
                     hook_outcome.additional_contexts,
                 )
                 .await;
+                if is_user_input {
+                    auto_compact_limiter.reset_after_user_input();
+                }
             }
         }
         if blocked_pending_input && !accepted_pending_input {
@@ -329,6 +354,7 @@ pub(crate) async fn run_turn(
                         &sess,
                         &turn_context,
                         &mut client_session,
+                        auto_compact_limiter,
                         InitialContextInjection::BeforeLastUserMessage,
                         CompactionReason::ContextLimit,
                         CompactionPhase::MidTurn,
@@ -348,7 +374,13 @@ pub(crate) async fn run_turn(
                                     "failed to usage-limit active goal after usage-limit error: {err}"
                                 );
                             }
-                            return None;
+                            return if err.to_codex_protocol_error()
+                                == CodexErrorInfo::ContextWindowExceeded
+                            {
+                                RunTurnResult::Terminal
+                            } else {
+                                RunTurnResult::Continue(None)
+                            };
                         }
                     };
                     if reset_client_session {
@@ -399,7 +431,7 @@ pub(crate) async fn run_turn(
                     )
                     .await
                     {
-                        return None;
+                        return RunTurnResult::Continue(None);
                     }
                     break;
                 }
@@ -447,7 +479,7 @@ pub(crate) async fn run_turn(
         }
     }
 
-    last_agent_message
+    RunTurnResult::Continue(last_agent_message)
 }
 
 #[expect(
@@ -652,6 +684,56 @@ struct AutoCompactTokenStatus {
     token_limit_reached: bool,
 }
 
+#[derive(Default)]
+pub(crate) struct AutoCompactTurnLimiter {
+    attempts: usize,
+}
+
+impl AutoCompactTurnLimiter {
+    fn reset_after_user_input(&mut self) {
+        self.attempts = 0;
+    }
+
+    async fn acquire(
+        &mut self,
+        sess: &Session,
+        turn_context: &TurnContext,
+        reason: CompactionReason,
+        phase: CompactionPhase,
+    ) -> CodexResult<()> {
+        if self.attempts >= MAX_AUTO_COMPACTIONS_PER_TURN {
+            let token_status = auto_compact_token_status(sess, turn_context).await;
+            warn!(
+                turn_id = %turn_context.sub_id,
+                auto_compactions_attempted = self.attempts,
+                auto_compaction_limit = MAX_AUTO_COMPACTIONS_PER_TURN,
+                reason = ?reason,
+                phase = ?phase,
+                active_context_tokens = token_status.active_context_tokens,
+                auto_compact_scope_tokens = token_status.auto_compact_scope_tokens,
+                auto_compact_scope_limit = token_status.auto_compact_scope_limit,
+                full_context_window_limit = ?token_status.full_context_window_limit,
+                full_context_window_limit_reached = token_status.full_context_window_limit_reached,
+                "automatic compaction limit reached; stopping turn"
+            );
+            sess.send_event(
+                turn_context,
+                EventMsg::Error(ErrorEvent {
+                    message: format!(
+                        "Codex stopped after {MAX_AUTO_COMPACTIONS_PER_TURN} automatic compactions in this turn because the context remained too large. Queued input for this turn was not processed; start a new thread or reduce earlier history before retrying."
+                    ),
+                    codex_error_info: Some(CodexErrorInfo::ContextWindowExceeded),
+                }),
+            )
+            .await;
+            return Err(CodexErr::ContextWindowExceeded);
+        }
+
+        self.attempts += 1;
+        Ok(())
+    }
+}
+
 async fn auto_compact_token_status(
     sess: &Session,
     turn_context: &TurnContext,
@@ -708,9 +790,15 @@ async fn run_pre_sampling_compact(
     sess: &Arc<Session>,
     turn_context: &Arc<TurnContext>,
     client_session: &mut ModelClientSession,
+    auto_compact_limiter: &mut AutoCompactTurnLimiter,
 ) -> CodexResult<PreSamplingCompactResult> {
-    let mut pre_sampling_compacted =
-        maybe_run_previous_model_inline_compact(sess, turn_context, client_session).await?;
+    let mut pre_sampling_compacted = maybe_run_previous_model_inline_compact(
+        sess,
+        turn_context,
+        client_session,
+        auto_compact_limiter,
+    )
+    .await?;
     let mut reset_client_session = pre_sampling_compacted;
     let token_status = auto_compact_token_status(sess.as_ref(), turn_context.as_ref()).await;
     // Compact if the configured auto-compaction budget or usable context window is exhausted.
@@ -719,6 +807,7 @@ async fn run_pre_sampling_compact(
             sess,
             turn_context,
             client_session,
+            auto_compact_limiter,
             InitialContextInjection::DoNotInject,
             CompactionReason::ContextLimit,
             CompactionPhase::PreTurn,
@@ -741,6 +830,7 @@ async fn maybe_run_previous_model_inline_compact(
     sess: &Arc<Session>,
     turn_context: &Arc<TurnContext>,
     client_session: &mut ModelClientSession,
+    auto_compact_limiter: &mut AutoCompactTurnLimiter,
 ) -> CodexResult<bool> {
     let Some(previous_turn_settings) = sess.previous_turn_settings().await else {
         return Ok(false);
@@ -780,6 +870,7 @@ async fn maybe_run_previous_model_inline_compact(
             sess,
             &previous_model_turn_context,
             client_session,
+            auto_compact_limiter,
             InitialContextInjection::DoNotInject,
             CompactionReason::ModelDownshift,
             CompactionPhase::PreTurn,
@@ -794,10 +885,15 @@ async fn run_auto_compact(
     sess: &Arc<Session>,
     turn_context: &Arc<TurnContext>,
     client_session: &mut ModelClientSession,
+    auto_compact_limiter: &mut AutoCompactTurnLimiter,
     initial_context_injection: InitialContextInjection,
     reason: CompactionReason,
     phase: CompactionPhase,
 ) -> CodexResult<bool> {
+    auto_compact_limiter
+        .acquire(sess.as_ref(), turn_context.as_ref(), reason, phase)
+        .await?;
+
     if should_use_remote_compact_task(turn_context.provider.info()) {
         if turn_context.features.enabled(Feature::RemoteCompactionV2) {
             run_inline_remote_auto_compact_task_v2(

codex-rs/core/src/tasks/regular.rs

@@ -2,6 +2,8 @@ use std::sync::Arc;
 
 use tokio_util::sync::CancellationToken;
 
+use crate::session::turn::AutoCompactTurnLimiter;
+use crate::session::turn::RunTurnResult;
 use crate::session::turn::run_turn;
 use crate::session::turn_context::TurnContext;
 use crate::session_startup_prewarm::SessionStartupPrewarmResolution;
@@ -69,17 +71,42 @@ impl SessionTask for RegularTask {
         };
         let mut next_input = input;
         let mut prewarmed_client_session = prewarmed_client_session;
+        let mut auto_compact_limiter = AutoCompactTurnLimiter::default();
         loop {
-            let last_agent_message = run_turn(
+            let turn_result = run_turn(
                 Arc::clone(&sess),
                 Arc::clone(&ctx),
                 Arc::clone(&turn_extension_data),
+                &mut auto_compact_limiter,
                 next_input,
                 prewarmed_client_session.take(),
                 cancellation_token.child_token(),
             )
             .instrument(run_turn_span.clone())
             .await;
+            let (last_agent_message, can_restart_with_pending_input) = match turn_result {
+                RunTurnResult::Continue(last_agent_message) => (last_agent_message, true),
+                RunTurnResult::Terminal => {
+                    let turn_state = {
+                        let active_turn = sess.active_turn.lock().await;
+                        active_turn
+                            .as_ref()
+                            .map(|active_turn| Arc::clone(&active_turn.turn_state))
+                    };
+                    if let Some(turn_state) = turn_state {
+                        turn_state.lock().await.clear_pending_waiters();
+                        drop(
+                            sess.input_queue
+                                .take_pending_input_for_turn_state(turn_state.as_ref())
+                                .await,
+                        );
+                    }
+                    (None, false)
+                }
+            };
+            if !can_restart_with_pending_input {
+                return last_agent_message;
+            }
             if !sess.input_queue.has_pending_input(&sess.active_turn).await {
                 return last_agent_message;
             }

codex-rs/core/tests/suite/compact.rs

@@ -13,6 +13,7 @@ use codex_protocol::models::PermissionProfile;
 use codex_protocol::openai_models::ModelInfo;
 use codex_protocol::openai_models::ModelsResponse;
 use codex_protocol::protocol::AskForApproval;
+use codex_protocol::protocol::CodexErrorInfo;
 use codex_protocol::protocol::EventMsg;
 use codex_protocol::protocol::HookEventName;
 use codex_protocol::protocol::HookRunStatus;
@@ -1490,6 +1491,119 @@ async fn multiple_auto_compact_per_task_runs_after_token_limit_hit() {
     assert_eq!(requests_payloads.len(), 7);
 }
 
+// Windows CI only: bump to 4 workers to prevent SSE/event starvation and test timeouts.
+#[cfg_attr(windows, tokio::test(flavor = "multi_thread", worker_threads = 4))]
+#[cfg_attr(not(windows), tokio::test(flavor = "multi_thread", worker_threads = 2))]
+async fn auto_compact_stops_after_per_turn_limit() {
+    skip_if_no_network!();
+
+    let server = start_mock_server().await;
+
+    let token_count_used = 270_000;
+    let token_count_used_after_compaction = 80_000;
+    let sampling_response = |idx: usize| {
+        let reasoning = ev_reasoning_item(
+            &format!("limit-reasoning-{idx}"),
+            &[&format!("step {idx}")],
+            &[],
+        );
+        sse(vec![
+            reasoning,
+            ev_shell_command_call(
+                &format!("limit-shell-{idx}"),
+                &format!("echo auto-compact-limit-{idx}"),
+            ),
+            ev_completed_with_tokens(&format!("limit-response-{idx}"), token_count_used),
+        ])
+    };
+    let compact_response = |idx: usize| {
+        sse(vec![
+            ev_assistant_message(
+                &format!("limit-compact-message-{idx}"),
+                &format!("AUTO_LIMIT_SUMMARY_{idx}"),
+            ),
+            ev_completed_with_tokens(
+                &format!("limit-compact-response-{idx}"),
+                token_count_used_after_compaction,
+            ),
+        ])
+    };
+
+    let request_log = mount_sse_sequence(
+        &server,
+        vec![
+            sampling_response(1),
+            compact_response(1),
+            sampling_response(2),
+            compact_response(2),
+            sampling_response(3),
+            compact_response(3),
+            sampling_response(4),
+        ],
+    )
+    .await;
+
+    let model_provider = non_openai_model_provider(&server);
+    let codex = test_codex()
+        .with_config(move |config| {
+            config.model_provider = model_provider;
+            set_test_compact_prompt(config);
+            config.model_auto_compact_token_limit = Some(200_000);
+        })
+        .build(&server)
+        .await
+        .expect("build codex")
+        .codex;
+
+    codex
+        .submit(Op::UserInput {
+            environments: None,
+            items: vec![UserInput::Text {
+                text: "trigger repeated auto compaction".into(),
+                text_elements: Vec::new(),
+            }],
+            final_output_json_schema: None,
+            responsesapi_client_metadata: None,
+            thread_settings: Default::default(),
+        })
+        .await
+        .expect("submit user input");
+
+    let error = wait_for_event_match(&codex, |event| match event {
+        EventMsg::Error(err)
+            if err
+                .message
+                .contains("Codex stopped after 3 automatic compactions in this turn") =>
+        {
+            Some(err.clone())
+        }
+        _ => None,
+    })
+    .await;
+
+    assert_eq!(
+        error.codex_error_info,
+        Some(CodexErrorInfo::ContextWindowExceeded)
+    );
+
+    let requests = request_log.requests();
+    let compaction_requests = requests
+        .iter()
+        .filter(|request| {
+            body_contains_text(&request.body_json().to_string(), SUMMARIZATION_PROMPT)
+        })
+        .count();
+    assert_eq!(
+        compaction_requests, 3,
+        "fourth automatic compaction should be blocked before dispatch"
+    );
+    assert_eq!(
+        requests.len(),
+        7,
+        "expected four sampling requests and three dispatched compaction requests"
+    );
+}
+
 // Windows CI only: bump to 4 workers to prevent SSE/event starvation and test timeouts.
 #[cfg_attr(windows, tokio::test(flavor = "multi_thread", worker_threads = 4))]
 #[cfg_attr(not(windows), tokio::test(flavor = "multi_thread", worker_threads = 2))]

codex-rs/core/tests/suite/pending_input.rs

@@ -20,6 +20,7 @@ use core_test_support::responses::ev_output_text_delta;
 use core_test_support::responses::ev_reasoning_item;
 use core_test_support::responses::ev_reasoning_item_added;
 use core_test_support::responses::ev_response_created;
+use core_test_support::responses::ev_shell_command_call;
 use core_test_support::streaming_sse::StreamingSseChunk;
 use core_test_support::streaming_sse::StreamingSseServer;
 use core_test_support::streaming_sse::start_streaming_sse_server;
@@ -679,6 +680,237 @@ async fn steered_user_input_follows_compact_when_only_the_steer_needs_follow_up(
     server.shutdown().await;
 }
 
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn steered_user_input_resets_auto_compact_limit() {
+    let (gate_first_completed_tx, gate_first_completed_rx) = oneshot::channel();
+    let (gate_second_completed_tx, gate_second_completed_rx) = oneshot::channel();
+    let (gate_third_completed_tx, gate_third_completed_rx) = oneshot::channel();
+    let (gate_fourth_completed_tx, gate_fourth_completed_rx) = oneshot::channel();
+
+    let sampling_chunks = |idx: usize, text: &'static str, gate| {
+        vec![
+            chunk(ev_response_created(&format!("resp-{idx}"))),
+            chunk(ev_message_item_added(&format!("msg-{idx}"), "")),
+            chunk(ev_output_text_delta(text)),
+            chunk(json!({
+                "type": "response.output_item.done",
+                "item": {
+                    "type": "message",
+                    "role": "assistant",
+                    "id": format!("msg-{idx}"),
+                    "content": [{"type": "output_text", "text": text}],
+                    "phase": "commentary",
+                }
+            })),
+            gated_chunk(
+                gate,
+                vec![ev_completed_with_tokens(
+                    &format!("resp-{idx}"),
+                    /*total_tokens*/ 500,
+                )],
+            ),
+        ]
+    };
+    let compact_chunks = |idx: usize| {
+        vec![
+            chunk(ev_response_created(&format!("resp-compact-{idx}"))),
+            chunk(ev_message_item_done(
+                &format!("msg-compact-{idx}"),
+                &format!("AUTO_COMPACT_SUMMARY_{idx}"),
+            )),
+            chunk(ev_completed_with_tokens(
+                &format!("resp-compact-{idx}"),
+                /*total_tokens*/ 50,
+            )),
+        ]
+    };
+    let final_chunks = vec![
+        chunk(ev_response_created("resp-final")),
+        chunk(ev_message_item_done("msg-final", "processed fifth prompt")),
+        chunk(ev_completed_with_tokens(
+            "resp-final",
+            /*total_tokens*/ 70,
+        )),
+    ];
+
+    let (server, _completions) = start_streaming_sse_server(vec![
+        sampling_chunks(1, "first answer", gate_first_completed_rx),
+        compact_chunks(1),
+        sampling_chunks(2, "processed second prompt", gate_second_completed_rx),
+        compact_chunks(2),
+        sampling_chunks(3, "processed third prompt", gate_third_completed_rx),
+        compact_chunks(3),
+        sampling_chunks(4, "processed fourth prompt", gate_fourth_completed_rx),
+        compact_chunks(4),
+        final_chunks,
+    ])
+    .await;
+
+    let codex = test_codex()
+        .with_model("gpt-5.4")
+        .with_config(|config| {
+            config.model_provider.name = "OpenAI (test)".to_string();
+            config.model_provider.supports_websockets = false;
+            config.model_auto_compact_token_limit = Some(200);
+        })
+        .build_with_streaming_server(&server)
+        .await
+        .unwrap_or_else(|err| panic!("build streaming Codex test session: {err}"))
+        .codex;
+
+    submit_user_input(&codex, "first prompt").await;
+
+    wait_for_agent_message(&codex, "first answer").await;
+    steer_user_input(&codex, "second prompt").await;
+    let _ = gate_first_completed_tx.send(());
+
+    wait_for_agent_message(&codex, "processed second prompt").await;
+    steer_user_input(&codex, "third prompt").await;
+    let _ = gate_second_completed_tx.send(());
+
+    wait_for_agent_message(&codex, "processed third prompt").await;
+    steer_user_input(&codex, "fourth prompt").await;
+    let _ = gate_third_completed_tx.send(());
+
+    wait_for_agent_message(&codex, "processed fourth prompt").await;
+    steer_user_input(&codex, "fifth prompt").await;
+    let _ = gate_fourth_completed_tx.send(());
+
+    wait_for_agent_message(&codex, "processed fifth prompt").await;
+    wait_for_turn_complete(&codex).await;
+
+    let requests = server.requests().await;
+    assert_eq!(
+        requests.len(),
+        9,
+        "four auto compactions should be allowed when separated by accepted steered input"
+    );
+
+    let final_body: Value =
+        from_slice(&requests[8]).unwrap_or_else(|err| panic!("parse final request: {err}"));
+    let final_user_texts = message_input_texts(&final_body, "user");
+    assert!(
+        final_user_texts.iter().any(|text| text == "fifth prompt"),
+        "steered input after the fourth compaction should be processed"
+    );
+
+    server.shutdown().await;
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn steered_user_input_is_discarded_when_auto_compact_guard_fires() {
+    let (gate_fourth_completed_tx, gate_fourth_completed_rx) = oneshot::channel();
+    let token_count_used = 270_000;
+    let token_count_used_after_compaction = 80_000;
+
+    let tool_sampling_chunks = |idx: usize| {
+        vec![
+            chunk(ev_response_created(&format!("resp-{idx}"))),
+            chunk(ev_shell_command_call(
+                &format!("call-{idx}"),
+                &format!("echo step-{idx}"),
+            )),
+            chunk(ev_completed_with_tokens(
+                &format!("resp-{idx}"),
+                token_count_used,
+            )),
+        ]
+    };
+    let compact_chunks = |idx: usize| {
+        vec![
+            chunk(ev_response_created(&format!("resp-compact-{idx}"))),
+            chunk(ev_message_item_done(
+                &format!("msg-compact-{idx}"),
+                &format!("AUTO_COMPACT_SUMMARY_{idx}"),
+            )),
+            chunk(ev_completed_with_tokens(
+                &format!("resp-compact-{idx}"),
+                token_count_used_after_compaction,
+            )),
+        ]
+    };
+    let fourth_sampling_chunks = vec![
+        chunk(ev_response_created("resp-4")),
+        chunk(ev_message_item_added("msg-4", "")),
+        chunk(ev_output_text_delta("ready for late steer")),
+        chunk(ev_message_item_done("msg-4", "ready for late steer")),
+        gated_chunk(
+            gate_fourth_completed_rx,
+            vec![ev_completed_with_tokens("resp-4", token_count_used)],
+        ),
+    ];
+    let next_turn_chunks = vec![
+        chunk(ev_response_created("resp-next")),
+        chunk(ev_message_item_done("msg-next", "processed next prompt")),
+        chunk(ev_completed_with_tokens(
+            "resp-next",
+            /*total_tokens*/ 70,
+        )),
+    ];
+
+    let mut responses = Vec::new();
+    for idx in 1..=3 {
+        responses.push(tool_sampling_chunks(idx));
+        responses.push(compact_chunks(idx));
+    }
+    responses.extend([fourth_sampling_chunks, compact_chunks(4), next_turn_chunks]);
+    let (server, _completions) = start_streaming_sse_server(responses).await;
+
+    let test = test_codex()
+        .with_model("gpt-5.4")
+        .with_config(|config| {
+            config.model_provider.name = "OpenAI (test)".to_string();
+            config.model_provider.supports_websockets = false;
+            config.model_auto_compact_token_limit = Some(200_000);
+        })
+        .build_with_streaming_server(&server)
+        .await
+        .unwrap_or_else(|err| panic!("build streaming Codex test session: {err}"));
+    let codex = test.codex.clone();
+
+    submit_danger_full_access_user_turn(&test, "first prompt").await;
+    wait_for_agent_message(&codex, "ready for late steer").await;
+    steer_user_input(&codex, "late steer").await;
+    let _ = gate_fourth_completed_tx.send(());
+
+    wait_for_turn_complete(&codex).await;
+    assert_eq!(
+        server.requests().await.len(),
+        7,
+        "guarded turn should stop before dispatching a fourth automatic compaction"
+    );
+
+    submit_user_input(&codex, "next prompt").await;
+    wait_for_agent_message(&codex, "processed next prompt").await;
+    wait_for_turn_complete(&codex).await;
+
+    let requests = server.requests().await;
+    assert_eq!(
+        requests.len(),
+        9,
+        "next turn should compact and sample without replaying the discarded steer"
+    );
+
+    let all_user_texts: Vec<String> = requests
+        .iter()
+        .flat_map(|request| {
+            let body: Value =
+                from_slice(request).unwrap_or_else(|err| panic!("parse request body: {err}"));
+            message_input_texts(&body, "user")
+        })
+        .collect();
+    assert!(
+        !all_user_texts.iter().any(|text| text == "late steer"),
+        "late steer should be discarded after the terminal compaction guard"
+    );
+    assert!(
+        all_user_texts.iter().any(|text| text == "next prompt"),
+        "new input after the terminal guard should still start a normal turn"
+    );
+
+    server.shutdown().await;
+}
+
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
 async fn steered_user_input_waits_when_tool_output_triggers_compact_before_next_request() {
     let (gate_first_completed_tx, gate_first_completed_rx) = oneshot::channel();