diff --git a/crates/core/tests/unit/atif_tests.rs b/crates/core/tests/unit/atif_tests.rs
index 090eae10..236aa297 100644
--- a/crates/core/tests/unit/atif_tests.rs
+++ b/crates/core/tests/unit/atif_tests.rs
@@ -900,6 +900,79 @@ fn test_exporter_anthropic_messages_lifecycle_promotes_tool_use_blocks() {
     );
 }
 
+#[test]
+fn test_exporter_openclaw_placeholder_replay_preserves_empty_user_step_and_raw_request() {
+    let exporter = AtifExporter::new("session-1".to_string(), make_agent_info());
+    let llm_uuid = Uuid::now_v7();
+    let base = base_timestamp();
+
+    let mut start = event_builder(llm_uuid, EventType::Start)
+        .name("openclaw-model-call")
+        .scope_type(ScopeType::Llm)
+        .input(json!({
+            "headers": {},
+            "content": {
+                "provider": "nvidia-inference",
+                "model": "claude-sonnet-4",
+                "prompt": "",
+                "messages": [],
+                "imagesCount": 0,
+                "placeholderRequest": true,
+                "source": "openclaw.llm_output"
+            }
+        }))
+        .model_name("claude-sonnet-4")
+        .build();
+    let mut end = event_builder(llm_uuid, EventType::End)
+        .name("openclaw-model-call")
+        .scope_type(ScopeType::Llm)
+        .output(json!({
+            "role": "assistant",
+            "content": "I will search.",
+            "assistant_texts_count": 1,
+            "openclaw": {
+                "assistant_tool_call_names": []
+            }
+        }))
+        .model_name("claude-sonnet-4")
+        .build();
+
+    for (offset, event) in [&mut start, &mut end].into_iter().enumerate() {
+        set_event_timestamp(event, base + chrono::Duration::milliseconds(offset as i64));
+    }
+
+    {
+        let mut state = exporter.state.lock().unwrap();
+        state.events.extend([start, end]);
+    }
+
+    let trajectory = exporter.export().unwrap();
+    assert_atif_v17_shape(&trajectory);
+    assert_eq!(trajectory.steps.len(), 2);
+
+    let user_step = &trajectory.steps[0];
+    assert_eq!(user_step.source, "user");
+    assert_eq!(user_step.message, json!(""));
+    let user_extra: AtifStepExtra =
+        serde_json::from_value(user_step.extra.clone().unwrap()).unwrap();
+    let llm_request = user_extra.llm_request.unwrap();
+    assert!(llm_request.get("headers").is_none());
+    assert_eq!(llm_request["placeholderRequest"], json!(true));
+    assert_eq!(llm_request["messages"], json!([]));
+    assert_eq!(llm_request["prompt"], json!(""));
+    assert_eq!(llm_request["source"], json!("openclaw.llm_output"));
+
+    let agent_step = &trajectory.steps[1];
+    assert_eq!(agent_step.source, "agent");
+    assert_eq!(agent_step.message, json!("I will search."));
+    assert_eq!(agent_step.model_name, Some("claude-sonnet-4".to_string()));
+    let agent_extra: AtifStepExtra =
+        serde_json::from_value(agent_step.extra.clone().unwrap()).unwrap();
+    let llm_response = agent_extra.llm_response.unwrap();
+    assert_eq!(llm_response["role"], json!("assistant"));
+    assert_eq!(llm_response["content"], json!("I will search."));
+}
+
 #[test]
 fn test_openai_responses_input_extracts_latest_user_content_block() {
     let message = extract_user_messages(&json!({
diff --git a/crates/core/tests/unit/observability/atof_tests.rs b/crates/core/tests/unit/observability/atof_tests.rs
index 870078f4..39827af1 100644
--- a/crates/core/tests/unit/observability/atof_tests.rs
+++ b/crates/core/tests/unit/observability/atof_tests.rs
@@ -169,6 +169,34 @@ fn openclaw_agent_scope_event(
     ))
 }
 
+fn openclaw_replay_llm_event(
+    uuid: Uuid,
+    parent_uuid: Option<Uuid>,
+    scope_category: ScopeCategory,
+    data: serde_json::Value,
+) -> Event {
+    Event::Scope(ScopeEvent::new(
+        BaseEvent::builder()
+            .uuid(uuid)
+            .parent_uuid_opt(parent_uuid)
+            .name("openclaw-model-call")
+            .data(data)
+            .metadata(json!({
+                "source": "openclaw.llm_output",
+                "hook_event_name": "llm_output"
+            }))
+            .build(),
+        scope_category,
+        Vec::new(),
+        EventCategory::llm(),
+        Some(
+            CategoryProfile::builder()
+                .model_name("claude-sonnet-4")
+                .build(),
+        ),
+    ))
+}
+
 fn read_jsonl(path: &Path) -> Vec<serde_json::Value> {
     fs::read_to_string(path)
         .unwrap()
@@ -570,6 +598,76 @@ fn openclaw_subagent_events_preserve_nested_and_fallback_parent_uuid() {
     );
 }
 
+#[test]
+fn subscriber_preserves_openclaw_placeholder_replay_payloads_as_raw_jsonl() {
+    let dir = temp_dir("atof-openclaw-placeholder");
+    let exporter = AtofExporter::new(
+        AtofExporterConfig::new()
+            .with_output_directory(&dir)
+            .with_filename("events.jsonl"),
+    )
+    .unwrap();
+    let subscriber = exporter.subscriber();
+
+    let uuid = Uuid::now_v7();
+    let parent_uuid = Uuid::now_v7();
+    let events = [
+        openclaw_replay_llm_event(
+            uuid,
+            Some(parent_uuid),
+            ScopeCategory::Start,
+            json!({
+                "headers": {},
+                "content": {
+                    "provider": "nvidia-inference",
+                    "model": "claude-sonnet-4",
+                    "prompt": "",
+                    "messages": [],
+                    "imagesCount": 0,
+                    "placeholderRequest": true,
+                    "source": "openclaw.llm_output"
+                }
+            }),
+        ),
+        openclaw_replay_llm_event(
+            uuid,
+            Some(parent_uuid),
+            ScopeCategory::End,
+            json!({
+                "role": "assistant",
+                "content": "I will search.",
+                "assistant_texts_count": 1,
+                "openclaw": {
+                    "assistant_tool_call_names": []
+                }
+            }),
+        ),
+    ];
+
+    for event in &events {
+        subscriber(event);
+    }
+    exporter.force_flush().unwrap();
+
+    let lines = read_jsonl(exporter.path());
+    assert_eq!(lines.len(), events.len());
+    for (line, event) in lines.iter().zip(events.iter()) {
+        assert_eq!(line, &event.try_to_json_value().unwrap());
+        assert_eq!(line["kind"], "scope");
+        assert_eq!(line["atof_version"], "0.1");
+        assert_eq!(line["parent_uuid"], parent_uuid.to_string());
+        assert_eq!(line["category"], "llm");
+        assert_eq!(line["metadata"]["source"], "openclaw.llm_output");
+        assert_eq!(line["metadata"]["hook_event_name"], "llm_output");
+    }
+
+    assert_eq!(lines[0]["scope_category"], "start");
+    assert_eq!(lines[0]["data"]["content"]["placeholderRequest"], true);
+    assert_eq!(lines[0]["data"]["content"]["messages"], json!([]));
+    assert_eq!(lines[1]["scope_category"], "end");
+    assert_eq!(lines[1]["data"]["content"], "I will search.");
+}
+
 #[test]
 fn register_deregister_flush_and_shutdown_work_with_runtime_events() {
     let _guard = crate::observability::test_mutex().lock().unwrap();
diff --git a/crates/core/tests/unit/observability/openinference_tests.rs b/crates/core/tests/unit/observability/openinference_tests.rs
index 715a3cf8..fc8d11f6 100644
--- a/crates/core/tests/unit/observability/openinference_tests.rs
+++ b/crates/core/tests/unit/observability/openinference_tests.rs
@@ -882,6 +882,80 @@ fn openclaw_subagent_scopes_preserve_nested_and_fallback_parent_linkage() {
     );
 }
 
+#[test]
+fn openclaw_placeholder_replay_falls_back_to_sanitized_json_input_value() {
+    let (provider, exporter) = make_provider();
+    let mut processor =
+        OpenInferenceEventProcessor::new(provider.clone(), "test-scope".to_string());
+    let uuid = Uuid::now_v7();
+
+    processor.process(&make_start_event(
+        uuid,
+        None,
+        "openclaw-model-call",
+        ScopeType::Llm,
+        Some(json!({
+            "headers": {"authorization": "Bearer secret-token"},
+            "content": {
+                "provider": "nvidia-inference",
+                "model": "claude-sonnet-4",
+                "prompt": "",
+                "messages": [],
+                "imagesCount": 0,
+                "placeholderRequest": true,
+                "source": "openclaw.llm_output"
+            }
+        })),
+    ));
+    processor.process(&make_end_event(
+        uuid,
+        None,
+        "openclaw-model-call",
+        ScopeType::Llm,
+        Some(json!({
+            "role": "assistant",
+            "content": "I will search.",
+            "assistant_texts_count": 1,
+            "openclaw": {
+                "assistant_tool_call_names": []
+            }
+        })),
+    ));
+
+    processor.force_flush().unwrap();
+
+    let spans = exporter.get_finished_spans().unwrap();
+    assert_eq!(spans.len(), 1);
+    let attributes = attr_map(&spans[0].attributes);
+    assert_attr(&attributes, "llm.provider", "nvidia-inference");
+    assert_attr(
+        &attributes,
+        "llm.output_messages.0.message.role",
+        "assistant",
+    );
+    assert_attr(
+        &attributes,
+        "llm.output_messages.0.message.content",
+        "I will search.",
+    );
+    assert!(!attributes.contains_key("llm.input_messages.0.message.role"));
+    assert!(!attributes.contains_key("llm.input_messages.0.message.content"));
+    assert_attr(&attributes, "input.mime_type", "application/json");
+
+    let input_value = attributes.get("input.value").expect("missing input.value");
+    let parsed_input: serde_json::Value = serde_json::from_str(input_value).unwrap();
+    assert!(parsed_input.get("headers").is_none());
+    assert_eq!(parsed_input["content"]["placeholderRequest"], json!(true));
+    assert_eq!(parsed_input["content"]["messages"], json!([]));
+    assert_eq!(parsed_input["content"]["prompt"], json!(""));
+    assert_eq!(
+        parsed_input["content"]["source"],
+        json!("openclaw.llm_output")
+    );
+    assert_no_attr_contains(&attributes, "authorization");
+    assert_no_attr_contains(&attributes, "secret-token");
+}
+
 #[test]
 fn generic_unannotated_llm_output_does_not_emit_flattened_output_message_attrs() {
     let (provider, exporter) = make_provider();