merge(main): 解决 usage 展示与生命周期同步冲突

apps/aether-gateway/src/ai_serving/planner/decision_input.rs

@@ -133,6 +133,17 @@ pub(crate) fn apply_provider_request_routing_policy_to_decision(
     );
 
     let Some(context) = input.routing_context.as_ref() else {
+        // Cache identity headers are projected only at the terminal boundary. Any non-empty
+        // session headers already present here are explicit client or header-rule inputs and stay
+        // authoritative.
+        if let Some(provider_request_body) = decision.provider_request_body.as_ref() {
+            crate::ai_serving::apply_codex_openai_responses_identity_headers(
+                &mut decision.provider_request_headers,
+                provider_request_body,
+                provider_type.as_str(),
+                provider_api_format.as_str(),
+            );
+        }
         return Ok(());
     };
     let provider_body_rules = decision
@@ -182,6 +193,14 @@ pub(crate) fn apply_provider_request_routing_policy_to_decision(
     })?;
     ensure_report_context_routing_trace(input, decision, &policy);
     if policy.mutation_plan.is_empty() {
+        if let Some(provider_request_body) = decision.provider_request_body.as_ref() {
+            crate::ai_serving::apply_codex_openai_responses_identity_headers(
+                &mut decision.provider_request_headers,
+                provider_request_body,
+                provider_type.as_str(),
+                provider_api_format.as_str(),
+            );
+        }
         return Ok(());
     }
     if original_provider_request_body.is_none() && !policy.mutation_plan.body_patch.is_empty() {
@@ -255,6 +274,12 @@ pub(crate) fn apply_provider_request_routing_policy_to_decision(
             input.requested_model.as_str(),
         )
     });
+    crate::ai_serving::apply_codex_openai_responses_identity_headers(
+        &mut provider_request_headers,
+        &provider_request_body,
+        provider_type.as_str(),
+        provider_api_format.as_str(),
+    );
     crate::ai_serving::apply_codex_openai_responses_lite_header_for_request_body_with_capabilities(
         &mut provider_request_headers,
         Some(&provider_request_body),
@@ -1279,6 +1304,107 @@ mod tests {
         );
     }
 
+    #[test]
+    fn codex_prompt_cache_identity_headers_are_terminal_after_routing_mutations() {
+        let mut input = sample_decision_input();
+        input
+            .routing_context
+            .as_mut()
+            .expect("routing context")
+            .client_api_format = "openai:responses".to_string();
+        set_provider_request_rules(
+            &mut input,
+            &["gpt-5"],
+            json!([{
+                "type": "patch_headers",
+                "patch": [
+                    {"op": "remove", "name": "session-id"},
+                    {"op": "remove", "name": "thread-id"}
+                ]
+            }]),
+        );
+        let identity = "172c39e6-c0a0-5a70-8b63-e0f8e0d185a3";
+        let mut decision = sample_decision();
+        decision.provider_type = Some("codex".to_string());
+        decision.provider_api_format = Some("openai:responses".to_string());
+        decision.client_api_format = Some("openai:responses".to_string());
+        decision.provider_request_body = Some(json!({
+            "model": "gpt-5",
+            "input": [],
+            "prompt_cache_key": identity,
+            "client_metadata": {
+                "session_id": identity,
+                "thread_id": identity
+            }
+        }));
+        assert!(!decision.provider_request_headers.contains_key("session-id"));
+        assert!(!decision.provider_request_headers.contains_key("thread-id"));
+
+        apply_provider_request_routing_policy_to_decision(&input, &mut decision, None)
+            .expect("terminal Codex identity contract should be restored");
+
+        assert_eq!(
+            decision
+                .provider_request_headers
+                .get("session-id")
+                .map(String::as_str),
+            Some(identity)
+        );
+        assert_eq!(
+            decision
+                .provider_request_headers
+                .get("thread-id")
+                .map(String::as_str),
+            Some(identity)
+        );
+    }
+
+    #[test]
+    fn codex_prompt_cache_identity_headers_fail_closed_after_body_identity_removal() {
+        let mut input = sample_decision_input();
+        input
+            .routing_context
+            .as_mut()
+            .expect("routing context")
+            .client_api_format = "openai:responses".to_string();
+        set_provider_request_rules(
+            &mut input,
+            &["gpt-5"],
+            json!([{
+                "type": "json_patch_body",
+                "patch": [
+                    {"op": "remove", "path": "/prompt_cache_key"},
+                    {"op": "remove", "path": "/client_metadata"}
+                ]
+            }]),
+        );
+        let identity = "172c39e6-c0a0-5a70-8b63-e0f8e0d185a3";
+        let mut decision = sample_decision();
+        decision.provider_type = Some("codex".to_string());
+        decision.provider_api_format = Some("openai:responses".to_string());
+        decision.client_api_format = Some("openai:responses".to_string());
+        decision.provider_request_body = Some(json!({
+            "model": "gpt-5",
+            "input": [],
+            "prompt_cache_key": identity,
+            "client_metadata": {
+                "session_id": identity,
+                "thread_id": identity
+            }
+        }));
+        assert!(!decision.provider_request_headers.contains_key("session-id"));
+        assert!(!decision.provider_request_headers.contains_key("thread-id"));
+
+        apply_provider_request_routing_policy_to_decision(&input, &mut decision, None)
+            .expect("terminal Codex identity contract should fail closed");
+
+        let body = decision.provider_request_body.as_ref().expect("body");
+        assert!(body.get("prompt_cache_key").is_none());
+        assert!(body.get("client_metadata").is_none());
+        assert!(!decision.provider_request_headers.contains_key("session-id"));
+        assert!(!decision.provider_request_headers.contains_key("thread-id"));
+    }
+
     #[test]
     fn codex_compact_contract_is_terminal_after_routing_mutations() {
         let mut input = sample_decision_input();

apps/aether-gateway/src/ai_serving/planner/standard/codex.rs

@@ -3,7 +3,8 @@
 mod tests;
 
 pub(crate) use crate::ai_serving::{
-    apply_codex_openai_responses_special_body_edits, apply_codex_openai_special_headers,
+    apply_codex_openai_responses_identity_headers, apply_codex_openai_responses_special_body_edits,
+    apply_codex_openai_special_headers,
 };
 
 pub(crate) fn codex_model_capabilities_for_transport(

apps/aether-gateway/src/ai_serving/planner/standard/codex/tests.rs

@@ -1,10 +1,12 @@
 use std::collections::BTreeMap;
 
 use super::{
-    apply_codex_openai_responses_special_body_edits, apply_codex_openai_special_headers,
-    codex_model_capabilities,
+    apply_codex_openai_responses_identity_headers, apply_codex_openai_responses_special_body_edits,
+    apply_codex_openai_special_headers, codex_model_capabilities,
+};
+use crate::ai_serving::planner::standard::{
+    build_cross_format_openai_responses_request_body, build_local_openai_responses_request_body,
 };
-use crate::ai_serving::planner::standard::build_local_openai_responses_request_body;
 use http::{HeaderMap, HeaderValue};
 use serde_json::json;
 
@@ -180,11 +182,11 @@ fn does_not_synthesize_prompt_cache_key_from_api_key_identity() {
 }
 
 #[test]
-fn keeps_existing_prompt_cache_key_for_codex_requests() {
+fn adapts_generic_prompt_cache_key_to_codex_native_identity() {
     let mut body = json!({
         "model": "gpt-5",
         "input": "hello",
-        "prompt_cache_key": "existing-key",
+        "prompt_cache_key": "ltm-pc-v2-5557e02f5c9b447a97673ba330dbe77a",
     });
 
     apply_codex_openai_responses_special_body_edits(
@@ -195,11 +197,291 @@ fn keeps_existing_prompt_cache_key_for_codex_requests() {
         Some("key-123"),
     );
 
-    assert_eq!(body["prompt_cache_key"], "existing-key");
+    let expected_identity = "d9c5d122-7c1c-5fb1-ba9d-656062eda44e";
+    assert_eq!(body["prompt_cache_key"], expected_identity);
+    assert_eq!(body["client_metadata"]["session_id"], expected_identity);
+    assert_eq!(body["client_metadata"]["thread_id"], expected_identity);
 }
 
 #[test]
-fn injects_identity_headers_without_deriving_session_headers_from_body() {
+fn preserves_native_codex_cache_identity_and_metadata() {
+    let mut body = json!({
+        "model": "gpt-5",
+        "input": "hello",
+        "prompt_cache_key": "guardian:parent-thread",
+        "client_metadata": {
+            "session_id": "native-session",
+            "thread_id": "native-thread",
+            "turn_id": "native-turn"
+        }
+    });
+    let expected = body.clone();
+
+    apply_codex_openai_responses_special_body_edits(
+        &mut body,
+        "codex",
+        "openai:responses",
+        None,
+        Some("key-123"),
+    );
+
+    assert_eq!(body["prompt_cache_key"], expected["prompt_cache_key"]);
+    assert_eq!(body["client_metadata"], expected["client_metadata"]);
+}
+
+#[test]
+fn preserves_uuid_prompt_cache_key_while_completing_codex_identity() {
+    let identity = "172c39e6-c0a0-5a70-8b63-e0f8e0d185a3";
+    let mut body = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": identity
+    });
+
+    apply_codex_openai_responses_special_body_edits(
+        &mut body,
+        "codex",
+        "openai:responses",
+        None,
+        None,
+    );
+
+    assert_eq!(body["prompt_cache_key"], identity);
+    assert_eq!(body["client_metadata"]["session_id"], identity);
+    assert_eq!(body["client_metadata"]["thread_id"], identity);
+}
+
+#[test]
+fn keeps_codex_prompt_cache_domains_distinct() {
+    let mut first = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "tenant-a"
+    });
+    let mut second = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "tenant-b"
+    });
+
+    for body in [&mut first, &mut second] {
+        apply_codex_openai_responses_special_body_edits(
+            body,
+            "codex",
+            "openai:responses",
+            None,
+            None,
+        );
+    }
+
+    assert_ne!(first["prompt_cache_key"], second["prompt_cache_key"]);
+    assert_eq!(
+        first["prompt_cache_key"],
+        first["client_metadata"]["session_id"]
+    );
+    assert_eq!(
+        second["prompt_cache_key"],
+        second["client_metadata"]["session_id"]
+    );
+}
+
+#[test]
+fn completes_partial_and_null_codex_client_metadata() {
+    let mut partial = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity",
+        "client_metadata": {
+            "thread_id": "native-thread",
+            "caller": "sdk"
+        }
+    });
+    let mut null_metadata = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity",
+        "client_metadata": null
+    });
+    let mut null_session = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity",
+        "client_metadata": {
+            "session_id": null,
+            "thread_id": null,
+            "caller": "sdk"
+        }
+    });
+
+    for body in [&mut partial, &mut null_metadata, &mut null_session] {
+        apply_codex_openai_responses_special_body_edits(
+            body,
+            "codex",
+            "openai:responses",
+            None,
+            None,
+        );
+    }
+
+    assert_eq!(partial["client_metadata"]["thread_id"], "native-thread");
+    assert_eq!(partial["client_metadata"]["caller"], "sdk");
+    assert_eq!(
+        partial["client_metadata"]["session_id"],
+        partial["prompt_cache_key"]
+    );
+    assert_eq!(
+        null_metadata["client_metadata"]["session_id"],
+        null_metadata["prompt_cache_key"]
+    );
+    assert_eq!(
+        null_metadata["client_metadata"]["thread_id"],
+        null_metadata["prompt_cache_key"]
+    );
+    assert_eq!(
+        null_session["client_metadata"]["session_id"],
+        null_session["prompt_cache_key"]
+    );
+    assert_eq!(
+        null_session["client_metadata"]["thread_id"],
+        null_session["prompt_cache_key"]
+    );
+    assert_eq!(null_session["client_metadata"]["caller"], "sdk");
+}
+
+#[test]
+fn leaves_malformed_codex_client_metadata_unchanged() {
+    let mut body = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity",
+        "client_metadata": "invalid"
+    });
+    let mut malformed_fields = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity",
+        "client_metadata": {
+            "session_id": 42,
+            "thread_id": ""
+        }
+    });
+    let expected_malformed_metadata = malformed_fields["client_metadata"].clone();
+
+    for candidate in [&mut body, &mut malformed_fields] {
+        apply_codex_openai_responses_special_body_edits(
+            candidate,
+            "codex",
+            "openai:responses",
+            None,
+            None,
+        );
+    }
+
+    assert_eq!(body["prompt_cache_key"], "generic-affinity");
+    assert_eq!(body["client_metadata"], "invalid");
+    assert_eq!(malformed_fields["prompt_cache_key"], "generic-affinity");
+    assert_eq!(
+        malformed_fields["client_metadata"],
+        expected_malformed_metadata
+    );
+}
+
+#[test]
+fn limits_prompt_cache_identity_adaptation_to_codex_responses_family() {
+    let original = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity"
+    });
+    let mut standard_openai = original.clone();
+    let mut codex_compact = original.clone();
+
+    apply_codex_openai_responses_special_body_edits(
+        &mut standard_openai,
+        "openai",
+        "openai:responses",
+        None,
+        None,
+    );
+    apply_codex_openai_responses_special_body_edits(
+        &mut codex_compact,
+        "codex",
+        "openai:responses:compact",
+        None,
+        None,
+    );
+
+    assert_eq!(standard_openai, original);
+    assert_ne!(codex_compact["prompt_cache_key"], "generic-affinity");
+    assert!(codex_compact.get("client_metadata").is_none());
+}
+
+#[test]
+fn chat_to_codex_responses_adapts_prompt_cache_identity_end_to_end() {
+    let body = json!({
+        "model": "gpt-5.6-luna",
+        "messages": [{"role": "user", "content": "hello"}],
+        "prompt_cache_key": "ltm-pc-v2-5557e02f5c9b447a97673ba330dbe77a"
+    });
+
+    let provider_request_body = build_cross_format_openai_responses_request_body(
+        &body,
+        "gpt-5.6-luna",
+        "openai:chat",
+        "openai:responses",
+        true,
+        false,
+        "codex",
+        None,
+        None,
+        &HeaderMap::new(),
+        false,
+    )
+    .expect("chat to Codex Responses request should build");
+
+    let expected_identity = "d9c5d122-7c1c-5fb1-ba9d-656062eda44e";
+    assert_eq!(provider_request_body["prompt_cache_key"], expected_identity);
+    assert_eq!(
+        provider_request_body["client_metadata"]["session_id"],
+        expected_identity
+    );
+    assert_eq!(
+        provider_request_body["client_metadata"]["thread_id"],
+        expected_identity
+    );
+
+    let mut provider_request_headers = BTreeMap::new();
+    apply_codex_openai_special_headers(
+        &mut provider_request_headers,
+        &provider_request_body,
+        &HeaderMap::new(),
+        "codex",
+        "openai:responses",
+        Some("trace-codex-cache-identity"),
+        None,
+    );
+    apply_codex_openai_responses_identity_headers(
+        &mut provider_request_headers,
+        &provider_request_body,
+        "codex",
+        "openai:responses",
+    );
+    assert_eq!(
+        provider_request_headers
+            .get("session-id")
+            .map(String::as_str),
+        Some(expected_identity)
+    );
+    assert_eq!(
+        provider_request_headers
+            .get("thread-id")
+            .map(String::as_str),
+        Some(expected_identity)
+    );
+}
+
+#[test]
+fn projects_uuid_prompt_cache_identity_into_missing_session_headers() {
     let mut headers = BTreeMap::new();
     let body = json!({
         "model": "gpt-5",
@@ -215,7 +497,7 @@ fn injects_identity_headers_without_deriving_session_headers_from_body() {
         Some("trace-codex-123"),
         Some(r#"{"account_id":"acc-123","is_fedramp":true}"#),
     );
-
+    apply_codex_openai_responses_identity_headers(&mut headers, &body, "codex", "openai:responses");
     assert_eq!(
         headers.get("chatgpt-account-id"),
         Some(&"acc-123".to_string())
@@ -228,8 +510,88 @@ fn injects_identity_headers_without_deriving_session_headers_from_body() {
     assert_eq!(headers.get("originator"), Some(&"codex_cli_rs".to_string()));
     assert!(!headers.contains_key("version"));
     assert_eq!(headers.get("x-openai-fedramp"), Some(&"true".to_string()));
-    assert_eq!(headers.get("session-id"), None);
-    assert_eq!(headers.get("thread-id"), None);
+    assert_eq!(
+        headers.get("session-id").map(String::as_str),
+        Some("172c39e6-c0a0-5a70-8b63-e0f8e0d185a3")
+    );
+    assert_eq!(
+        headers.get("thread-id").map(String::as_str),
+        Some("172c39e6-c0a0-5a70-8b63-e0f8e0d185a3")
+    );
+}
+
+#[test]
+fn projects_native_codex_metadata_for_non_uuid_cache_overrides() {
+    let mut headers = BTreeMap::new();
+    let body = json!({
+        "model": "gpt-5.6-luna",
+        "prompt_cache_key": "guardian:parent-thread",
+        "client_metadata": {
+            "session_id": "019f687b-8e92-7842-9631-d5bf0dba0a3b",
+            "thread_id": "019f6d20-1111-7222-8333-444455556666"
+        }
+    });
+
+    apply_codex_openai_special_headers(
+        &mut headers,
+        &body,
+        &HeaderMap::new(),
+        "codex",
+        "openai:responses",
+        None,
+        None,
+    );
+    apply_codex_openai_responses_identity_headers(&mut headers, &body, "codex", "openai:responses");
+
+    assert_eq!(
+        headers.get("session-id").map(String::as_str),
+        Some("019f687b-8e92-7842-9631-d5bf0dba0a3b")
+    );
+    assert_eq!(
+        headers.get("thread-id").map(String::as_str),
+        Some("019f6d20-1111-7222-8333-444455556666")
+    );
+}
+
+#[test]
+fn leaves_non_native_cache_keys_out_of_identity_headers() {
+    let mut headers = BTreeMap::new();
+    let body = json!({
+        "model": "gpt-5.6-luna",
+        "prompt_cache_key": "generic-cache-key"
+    });
+
+    apply_codex_openai_special_headers(
+        &mut headers,
+        &body,
+        &HeaderMap::new(),
+        "codex",
+        "openai:responses",
+        None,
+        None,
+    );
+    apply_codex_openai_responses_identity_headers(&mut headers, &body, "codex", "openai:responses");
+
+    assert!(!headers.contains_key("session-id"));
+    assert!(!headers.contains_key("thread-id"));
+}
+
+#[test]
+fn leaves_malformed_native_metadata_out_of_identity_headers() {
+    let mut headers = BTreeMap::new();
+    let body = json!({
+        "model": "gpt-5.6-luna",
+        "prompt_cache_key": "172c39e6-c0a0-5a70-8b63-e0f8e0d185a3",
+        "client_metadata": {
+            "session_id": 42,
+            "thread_id": ""
+        }
+    });
+
+    apply_codex_openai_responses_identity_headers(&mut headers, &body, "codex", "openai:responses");
+
+    assert!(!headers.contains_key("session-id"));
+    assert!(!headers.contains_key("thread-id"));
 }
 
 #[test]
@@ -247,7 +609,6 @@ fn injects_only_codex_client_headers_for_images_requests() {
         Some("trace-codex-image-123"),
         Some(r#"{"account_id":"acc-123","is_fedramp":true}"#),
     );
-
     assert_eq!(
         headers.get("chatgpt-account-id"),
         Some(&"acc-123".to_string())
@@ -330,6 +691,7 @@ fn preserves_client_context_headers_and_enforces_codex_provider_identity() {
         Some("trace-codex-123"),
         Some(r#"{"account_id":"acc-123","is_fedramp":true}"#),
     );
+    apply_codex_openai_responses_identity_headers(&mut headers, &body, "codex", "openai:responses");
 
     assert_eq!(
         headers.get("x-client-request-id"),
@@ -371,7 +733,7 @@ fn preserves_client_context_headers_and_enforces_codex_provider_identity() {
 }
 
 #[test]
-fn compact_does_not_derive_session_headers_from_body() {
+fn compact_projects_uuid_prompt_cache_identity_into_session_headers() {
     let mut headers = BTreeMap::new();
     let body = json!({
         "model": "gpt-5",
@@ -387,6 +749,12 @@ fn compact_does_not_derive_session_headers_from_body() {
         Some("trace-codex-compact-123"),
         Some(r#"{"account_id":"acc-123","is_fedramp":true}"#),
     );
+    apply_codex_openai_responses_identity_headers(
+        &mut headers,
+        &body,
+        "codex",
+        "openai:responses:compact",
+    );
 
     assert_eq!(
         headers.get("chatgpt-account-id"),
@@ -400,6 +768,12 @@ fn compact_does_not_derive_session_headers_from_body() {
     assert_eq!(headers.get("originator"), Some(&"codex_cli_rs".to_string()));
     assert!(!headers.contains_key("version"));
     assert_eq!(headers.get("x-openai-fedramp"), Some(&"true".to_string()));
-    assert_eq!(headers.get("session-id"), None);
-    assert_eq!(headers.get("thread-id"), None);
+    assert_eq!(
+        headers.get("session-id").map(String::as_str),
+        Some("172c39e6-c0a0-5a70-8b63-e0f8e0d185a3")
+    );
+    assert_eq!(
+        headers.get("thread-id").map(String::as_str),
+        Some("172c39e6-c0a0-5a70-8b63-e0f8e0d185a3")
+    );
 }

apps/aether-gateway/src/ai_serving/pure/mod.rs

@@ -4,6 +4,7 @@ pub(crate) use aether_ai_formats::api::{
     aggregate_standard_chat_stream_sync_response, aggregate_standard_cli_stream_sync_response,
     api_format_alias_matches, api_format_storage_aliases,
     apply_codex_openai_compact_terminal_headers, apply_codex_openai_responses_chat_body_edits,
+    apply_codex_openai_responses_identity_headers,
     apply_codex_openai_responses_lite_header_for_request_body_with_capabilities,
     apply_codex_openai_responses_lite_header_with_capabilities,
     apply_codex_openai_responses_special_body_edits,

apps/aether-gateway/src/execution_runtime/stream/execution.rs

@@ -131,6 +131,7 @@ const SSE_KEEPALIVE_INTERVAL: Duration = Duration::from_secs(15);
 const SSE_KEEPALIVE_BYTES: &[u8] = b": aether-keepalive\n\n";
 const SSE_CONTROL_FILTER_MAX_BUFFER_BYTES: usize = 1024 * 1024;
 const SSE_TERMINAL_DETECTOR_MAX_LINE_BYTES: usize = 1024 * 1024;
+const PROVIDER_STREAM_ERROR_INSPECTION_MAX_BYTES: usize = SSE_TERMINAL_DETECTOR_MAX_LINE_BYTES;
 const STREAM_IDLE_LOG_INTERVAL: Duration = Duration::from_secs(60);
 const STREAM_IDLE_LOG_INTERVAL_MS: u64 = 60_000;
 const REWRITTEN_STREAM_PREFETCH_TIMEOUT: Duration = Duration::from_millis(750);
@@ -140,6 +141,78 @@ const DIRECT_PASSTHROUGH_CHANNEL_CAPACITY_ENV: &str =
     "AETHER_GATEWAY_DIRECT_PASSTHROUGH_CHANNEL_CAPACITY";
 const DIRECT_PASSTHROUGH_MODE_ENV: &str = "AETHER_GATEWAY_DIRECT_PASSTHROUGH_MODE";
 
+/// Retains the incomplete tail needed to recognize provider error events split across transport
+/// chunks without retaining an unbounded copy of the stream.
+#[derive(Default)]
+struct ProviderStreamErrorInspection {
+    buffered: Vec<u8>,
+}
+
+impl ProviderStreamErrorInspection {
+    fn observe(&mut self, report_context: Option<&Value>, chunk: &[u8]) -> Option<Value> {
+        if chunk.is_empty() {
+            return None;
+        }
+        if let Some(error_body) = extract_provider_private_stream_error_body(report_context, chunk)
+        {
+            return Some(error_body);
+        }
+
+        self.append_rolling(chunk);
+        let error_body = extract_provider_private_stream_error_body(report_context, &self.buffered);
+        if error_body.is_none() {
+            self.trim_completed_sse_events();
+        }
+        error_body
+    }
+
+    fn append_rolling(&mut self, chunk: &[u8]) {
+        if chunk.len() >= PROVIDER_STREAM_ERROR_INSPECTION_MAX_BYTES {
+            self.buffered.clear();
+            self.buffered.extend_from_slice(
+                &chunk[chunk.len() - PROVIDER_STREAM_ERROR_INSPECTION_MAX_BYTES..],
+            );
+            return;
+        }
+
+        let overflow = self
+            .buffered
+            .len()
+            .saturating_add(chunk.len())
+            .saturating_sub(PROVIDER_STREAM_ERROR_INSPECTION_MAX_BYTES);
+        if overflow > 0 {
+            self.buffered.drain(..overflow);
+        }
+        self.buffered.extend_from_slice(chunk);
+    }
+
+    fn trim_completed_sse_events(&mut self) {
+        let Ok(text) = std::str::from_utf8(&self.buffered) else {
+            return;
+        };
+        if !text.lines().any(|line| {
+            let line = line.trim_start();
+            line.starts_with("event:") || line.starts_with("data:") || line.starts_with(':')
+        }) {
+            return;
+        }
+
+        let lf_end = self
+            .buffered
+            .windows(2)
+            .rposition(|window| window == b"\n\n")
+            .map(|index| index + 2);
+        let crlf_end = self
+            .buffered
+            .windows(4)
+            .rposition(|window| window == b"\r\n\r\n")
+            .map(|index| index + 4);
+        if let Some(event_end) = lf_end.into_iter().chain(crlf_end).max() {
+            self.buffered.drain(..event_end);
+        }
+    }
+}
+
 struct StageElapsedGuard {
     stage: &'static str,
     started_at: Instant,
@@ -1192,6 +1265,7 @@ struct DirectPassthroughFinalizerCore {
     stream_usage_report_context: Option<Value>,
     stream_usage_observer: Option<StreamingStandardTerminalObserver>,
     stream_usage_observer_buffered: Vec<u8>,
+    provider_error_inspection: ProviderStreamErrorInspection,
     provider_buffered_body: Vec<u8>,
     buffered_body: Vec<u8>,
     provider_body_truncated: bool,
@@ -1318,10 +1392,10 @@ impl DirectPassthroughFinalizer {
                 chunk.as_ref(),
             );
         }
-        if let Some(error_body_json) = extract_provider_private_stream_error_body(
-            core.stream_usage_report_context.as_ref(),
-            chunk.as_ref(),
-        ) {
+        if let Some(error_body_json) = core
+            .provider_error_inspection
+            .observe(core.stream_usage_report_context.as_ref(), chunk.as_ref())
+        {
             let error_status_code =
                 resolve_local_sync_error_status_code(core.status_code, &error_body_json);
             core.terminal_failure = Some(build_stream_failure_from_provider_error_body(
@@ -1486,6 +1560,7 @@ impl DirectPassthroughFinalizerCore {
             stream_usage_report_context,
             stream_usage_observer: _,
             stream_usage_observer_buffered: _,
+            provider_error_inspection: _,
             provider_buffered_body,
             buffered_body,
             provider_body_truncated,
@@ -2207,6 +2282,7 @@ async fn execute_stream_from_direct_passthrough(
             stream_usage_report_context,
             stream_usage_observer,
             stream_usage_observer_buffered: Vec::new(),
+            provider_error_inspection: ProviderStreamErrorInspection::default(),
             provider_buffered_body: Vec::new(),
             buffered_body: Vec::new(),
             provider_body_truncated: false,
@@ -2286,6 +2362,7 @@ async fn execute_stream_from_direct_passthrough(
             .as_ref()
             .map(|_| StreamingStandardTerminalObserver::default());
         let mut stream_usage_observer_buffered = Vec::new();
+        let mut provider_error_inspection = ProviderStreamErrorInspection::default();
         let mut provider_buffered_body = Vec::new();
         let mut buffered_body = Vec::new();
         let mut provider_body_truncated = false;
@@ -2458,7 +2535,7 @@ async fn execute_stream_from_direct_passthrough(
                     provider_chunk.as_ref(),
                 );
             }
-            let provider_private_error_body_json = extract_provider_private_stream_error_body(
+            let provider_private_error_body_json = provider_error_inspection.observe(
                 stream_usage_report_context.as_ref(),
                 provider_chunk.as_ref(),
             );
@@ -5007,6 +5084,7 @@ async fn execute_stream_from_frame_stream(
             .filter(|_| !sync_json_stream_bridge_active_for_report)
             .map(|_| StreamingStandardTerminalObserver::default());
         let mut stream_usage_observer_buffered = Vec::new();
+        let mut provider_error_inspection = ProviderStreamErrorInspection::default();
         append_stream_capture_bytes(
             &mut provider_buffered_body,
             &provider_prefetched_body_for_report,
@@ -5168,6 +5246,16 @@ async fn execute_stream_from_frame_stream(
             let replay_chunk = normalized_prefetched_chunk
                 .as_deref()
                 .unwrap_or(provider_prefetched_body_for_report.as_slice());
+            if let Some(error_body_json) = provider_error_inspection
+                .observe(stream_usage_report_context.as_ref(), replay_chunk)
+            {
+                let error_status_code =
+                    resolve_local_sync_error_status_code(status_code, &error_body_json);
+                terminal_failure = Some(build_stream_failure_from_provider_error_body(
+                    error_status_code,
+                    &error_body_json,
+                ));
+            }
             if let (Some(observer), Some(report_context)) = (
                 stream_usage_observer.as_mut(),
                 stream_usage_report_context.as_ref(),
@@ -5343,11 +5431,8 @@ async fn execute_stream_from_frame_stream(
                         } else {
                             chunk
                         };
-                        let provider_private_error_body_json =
-                            extract_provider_private_stream_error_body(
-                                stream_usage_report_context.as_ref(),
-                                &normalized_chunk,
-                            );
+                        let provider_private_error_body_json = provider_error_inspection
+                            .observe(stream_usage_report_context.as_ref(), &normalized_chunk);
                         if let (Some(observer), Some(report_context)) = (
                             stream_usage_observer.as_mut(),
                             stream_usage_report_context.as_ref(),
@@ -5508,11 +5593,8 @@ async fn execute_stream_from_frame_stream(
         {
             match normalizer.finish() {
                 Ok(normalized_chunk) if !normalized_chunk.is_empty() => {
-                    let provider_private_error_body_json =
-                        extract_provider_private_stream_error_body(
-                            stream_usage_report_context.as_ref(),
-                            &normalized_chunk,
-                        );
+                    let provider_private_error_body_json = provider_error_inspection
+                        .observe(stream_usage_report_context.as_ref(), &normalized_chunk);
                     if let (Some(observer), Some(report_context)) = (
                         stream_usage_observer.as_mut(),
                         stream_usage_report_context.as_ref(),
@@ -6114,7 +6196,7 @@ mod tests {
         stream_terminal_summary_missing_observed_finish,
         stream_terminal_summary_missing_observed_finish_with_requirement,
         stream_terminal_summary_represents_failure_with_requirement,
-        ClientVisibleStreamCompletionTracker, DirectPassthroughMode,
+        ClientVisibleStreamCompletionTracker, DirectPassthroughMode, ProviderStreamErrorInspection,
     };
     use crate::control::GatewayControlDecision;
     use crate::stage_metrics::RequestStageTrace;
@@ -6238,6 +6320,51 @@ mod tests {
             .observe_chunk(b"data: {\"type\":\"response.completed\",\"response\":{}}\r\n\r\n"));
     }
 
+    #[test]
+    fn provider_error_inspection_detects_response_failed_at_every_chunk_boundary() {
+        let body = concat!(
+            "event: response.created\n",
+            "data: {\"type\":\"response.created\",\"response\":{\"status\":\"in_progress\"}}\n\n",
+            "event: response.failed\n",
+            "data: {\"type\":\"response.failed\",\"response\":{\"status\":\"failed\",\"error\":{\"type\":\"invalid_request\",\"message\":\"cyber policy rejected the request\",\"code\":\"cyber_policy_violation\",\"param\":\"input\"}}}\n\n",
+        )
+        .as_bytes();
+
+        for split in 1..body.len() {
+            let mut inspection = ProviderStreamErrorInspection::default();
+            let detected = inspection
+                .observe(None, &body[..split])
+                .or_else(|| inspection.observe(None, &body[split..]))
+                .unwrap_or_else(|| panic!("response.failed was missed at byte split {split}"));
+
+            assert_eq!(
+                detected.pointer("/error/code"),
+                Some(&json!("cyber_policy_violation")),
+                "string provider code changed at byte split {split}"
+            );
+            assert_eq!(
+                detected.pointer("/error/param"),
+                Some(&json!("input")),
+                "provider error fields changed at byte split {split}"
+            );
+        }
+
+        let mut inspection = ProviderStreamErrorInspection::default();
+        let mut detected = None;
+        for byte in body.chunks(1) {
+            if let Some(error_body) = inspection.observe(None, byte) {
+                detected = Some(error_body);
+                break;
+            }
+        }
+        let detected = detected.expect("byte-wise response.failed stream should be detected");
+        assert_eq!(
+            detected.pointer("/error/code"),
+            Some(&json!("cyber_policy_violation"))
+        );
+        assert_eq!(detected.pointer("/error/param"), Some(&json!("input")));
+    }
+
     fn tunnel_proxy_snapshot(base_url: String) -> aether_contracts::ProxySnapshot {
         aether_contracts::ProxySnapshot {
             enabled: Some(true),

apps/aether-gateway/src/execution_runtime/stream/execution_failures.rs

@@ -20,10 +20,9 @@ use crate::execution_runtime::submission::{
 use crate::log_ids::short_request_id;
 use crate::orchestration::{
     apply_local_execution_effect, resolve_local_failover_analysis_for_attempt,
-    trace_upstream_response_body, with_upstream_response_report_context,
-    LocalAdaptiveRateLimitEffect, LocalAttemptFailureEffect, LocalExecutionEffect,
-    LocalExecutionEffectContext, LocalHealthFailureEffect, LocalOAuthInvalidationEffect,
-    LocalPoolErrorEffect,
+    with_upstream_response_report_context, LocalAdaptiveRateLimitEffect, LocalAttemptFailureEffect,
+    LocalExecutionEffect, LocalExecutionEffectContext, LocalHealthFailureEffect,
+    LocalOAuthInvalidationEffect, LocalPoolErrorEffect,
 };
 use crate::request_candidate_runtime::record_report_request_candidate_status;
 use crate::request_diagnostics::attach_current_request_diagnostics_to_report_context;
@@ -36,6 +35,7 @@ pub(super) struct StreamFailureReport {
     pub(super) error_type: String,
     pub(super) error_message: String,
     extra_error_fields: Map<String, Value>,
+    provider_body_json: Option<Value>,
 }
 
 #[derive(Serialize)]
@@ -54,20 +54,28 @@ struct StreamFailureBodyFields<'a> {
 }
 
 impl StreamFailureReport {
-    fn into_body_json(self) -> Value {
+    fn into_body_jsons(self) -> (Value, Option<Value>) {
         let Self {
             status_code,
             error_type,
             error_message,
             mut extra_error_fields,
+            provider_body_json,
         } = self;
         extra_error_fields.insert("type".to_string(), Value::String(error_type));
         extra_error_fields.insert("message".to_string(), Value::String(error_message));
         extra_error_fields.insert("code".to_string(), Value::from(status_code));
-        Value::Object(Map::from_iter([(
+        let normalized_body = Value::Object(Map::from_iter([(
             "error".to_string(),
             Value::Object(extra_error_fields),
-        )]))
+        )]));
+        match provider_body_json {
+            Some(provider_body) if provider_body != normalized_body => {
+                (provider_body, Some(normalized_body))
+            }
+            Some(provider_body) => (provider_body, None),
+            None => (normalized_body, None),
+        }
     }
 
     pub(super) fn to_json_string(&self) -> serde_json::Result<String> {
@@ -94,6 +102,7 @@ pub(super) fn build_stream_failure_report(
         error_type,
         error_message,
         extra_error_fields: Map::new(),
+        provider_body_json: None,
     }
 }
 
@@ -124,6 +133,7 @@ pub(super) fn build_stream_failure_from_execution_error(
         error_type,
         error_message,
         extra_error_fields: error_object,
+        provider_body_json: None,
     }
 }
 
@@ -150,6 +160,7 @@ pub(super) fn build_stream_failure_from_provider_error_body(
         error_type,
         error_message,
         extra_error_fields: Map::new(),
+        provider_body_json: Some(body_json.clone()),
     }
 }
 
@@ -185,19 +196,33 @@ fn build_stream_failure_sync_payload(
     failure: StreamFailureReport,
 ) -> GatewaySyncReportRequest {
     let status_code = failure.status_code;
-    let body = trace_upstream_response_body(None, provider_buffered_body);
+    let (body, client_body) = failure.into_body_jsons();
+    headers.retain(|name, _| {
+        !name.eq_ignore_ascii_case("content-encoding")
+            && !name.eq_ignore_ascii_case("content-length")
+            && !name.eq_ignore_ascii_case("content-type")
+    });
+    headers.insert("content-type".to_string(), "application/json".to_string());
     let report_context = with_upstream_response_report_context(
         report_context.as_ref(),
         status_code,
         Some(&headers),
-        body.as_ref(),
+        Some(&body),
         None,
         None,
     )
     .or(report_context);
-    headers.remove("content-encoding");
-    headers.remove("content-length");
-    headers.insert("content-type".to_string(), "application/json".to_string());
+    let report_context = report_context.map(|mut context| {
+        if let Some(object) = context.as_object_mut() {
+            let response_headers = serde_json::to_value(&headers).unwrap_or(Value::Null);
+            object.insert(
+                "provider_response_headers".to_string(),
+                response_headers.clone(),
+            );
+            object.insert("client_response_headers".to_string(), response_headers);
+        }
+        context
+    });
 
     GatewaySyncReportRequest {
         trace_id: trace_id.to_string(),
@@ -205,8 +230,8 @@ fn build_stream_failure_sync_payload(
         report_context,
         status_code,
         headers,
-        body_json: Some(failure.into_body_json()),
-        client_body_json: None,
+        body_json: Some(body),
+        client_body_json: client_body,
         body_base64: (!provider_buffered_body.is_empty())
             .then(|| base64::engine::general_purpose::STANDARD.encode(provider_buffered_body)),
         telemetry,
@@ -218,8 +243,9 @@ fn stream_failure_body_field<'a>(
     field: &str,
 ) -> Option<&'a str> {
     payload
-        .body_json
+        .client_body_json
         .as_ref()
+        .or(payload.body_json.as_ref())
         .and_then(|body_json| body_json.get("error"))
         .and_then(|value| value.get(field))
         .and_then(Value::as_str)
@@ -477,3 +503,151 @@ pub(super) async fn submit_midstream_stream_failure(
         );
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+
+    use base64::Engine as _;
+    use serde_json::json;
+
+    use super::{build_stream_failure_from_provider_error_body, build_stream_failure_sync_payload};
+
+    #[test]
+    fn midstream_failure_trace_uses_terminal_error_instead_of_buffered_sse() {
+        let provider_buffered_body = concat!(
+            "event: response.created\n",
+            "data: {\"type\":\"response.created\",\"response\":{\"instructions\":\"AGENTS.md secret prompt\",\"tools\":[{\"name\":\"update_plan\"}]}}\n\n",
+            "event: response.failed\n",
+            "data: {\"type\":\"response.failed\",\"response\":{\"status\":\"failed\",\"error\":{\"type\":\"invalid_request\",\"message\":\"This content was flagged for possible cybersecurity risk.\",\"code\":\"cyber_policy_violation\",\"param\":\"input\",\"details\":{\"policy_category\":\"cybersecurity\",\"appeal_allowed\":true}}}}\n\n",
+        )
+        .as_bytes();
+        let terminal_error = aether_ai_formats::api::extract_provider_private_stream_error_body(
+            None,
+            provider_buffered_body,
+        )
+        .expect("raw upstream SSE should expose its terminal provider error JSON");
+        let failure = build_stream_failure_from_provider_error_body(400, &terminal_error);
+
+        let payload = build_stream_failure_sync_payload(
+            "trace-cyber-policy",
+            "openai_responses_sync_error".to_string(),
+            Some(json!({"request_id": "request-cyber-policy"})),
+            BTreeMap::from([
+                ("Content-Encoding".to_string(), "gzip".to_string()),
+                ("Content-Length".to_string(), "4096".to_string()),
+                ("Content-Type".to_string(), "text/event-stream".to_string()),
+                (
+                    "x-request-id".to_string(),
+                    "req_usage-cyber-risk-demo".to_string(),
+                ),
+            ]),
+            None,
+            provider_buffered_body,
+            failure,
+        );
+
+        let trace_body = payload
+            .report_context
+            .as_ref()
+            .and_then(|context| context.pointer("/upstream_response/body"))
+            .expect("candidate trace should include the terminal error body");
+        assert_eq!(
+            trace_body,
+            payload.body_json.as_ref().expect("usage error body")
+        );
+        assert_eq!(trace_body, &terminal_error);
+        assert_eq!(trace_body["error"]["type"], json!("invalid_request"));
+        assert_eq!(
+            trace_body["error"]["message"],
+            json!("This content was flagged for possible cybersecurity risk.")
+        );
+        assert_eq!(trace_body["error"]["code"], json!("cyber_policy_violation"));
+        assert_eq!(trace_body["error"]["param"], json!("input"));
+        assert_eq!(
+            trace_body["error"]["details"],
+            json!({
+                "policy_category": "cybersecurity",
+                "appeal_allowed": true
+            })
+        );
+        assert_eq!(
+            payload
+                .report_context
+                .as_ref()
+                .and_then(|context| context.pointer("/upstream_response/headers/content-type")),
+            Some(&json!("application/json"))
+        );
+        assert_eq!(
+            payload
+                .report_context
+                .as_ref()
+                .and_then(|context| context.pointer("/upstream_response/headers/x-request-id")),
+            Some(&json!("req_usage-cyber-risk-demo"))
+        );
+        assert_eq!(
+            payload
+                .report_context
+                .as_ref()
+                .and_then(|context| context.pointer("/provider_response_headers/content-type")),
+            Some(&json!("application/json"))
+        );
+        assert_eq!(
+            payload
+                .report_context
+                .as_ref()
+                .and_then(|context| context.pointer("/client_response_headers/content-type")),
+            Some(&json!("application/json"))
+        );
+        let trace_headers = payload
+            .report_context
+            .as_ref()
+            .and_then(|context| context.pointer("/upstream_response/headers"))
+            .and_then(serde_json::Value::as_object)
+            .expect("candidate trace should include terminal JSON headers");
+        assert!(!trace_headers
+            .keys()
+            .any(|name| name.eq_ignore_ascii_case("content-encoding")));
+        assert!(!trace_headers
+            .keys()
+            .any(|name| name.eq_ignore_ascii_case("content-length")));
+        assert_eq!(
+            payload.headers.get("content-type").map(String::as_str),
+            Some("application/json")
+        );
+        assert!(!payload
+            .headers
+            .keys()
+            .any(|name| name.eq_ignore_ascii_case("content-encoding")));
+        assert!(!payload
+            .headers
+            .keys()
+            .any(|name| name.eq_ignore_ascii_case("content-length")));
+        assert_eq!(
+            payload
+                .client_body_json
+                .as_ref()
+                .and_then(|body| body.pointer("/error/message")),
+            Some(&json!(
+                "This content was flagged for possible cybersecurity risk."
+            ))
+        );
+        assert_eq!(
+            payload
+                .client_body_json
+                .as_ref()
+                .and_then(|body| body.pointer("/error/code")),
+            Some(&json!(400))
+        );
+        assert_ne!(payload.client_body_json.as_ref(), Some(&terminal_error));
+        assert!(!trace_body.to_string().contains("AGENTS.md secret prompt"));
+
+        let raw_capture = payload
+            .body_base64
+            .as_deref()
+            .and_then(|body| base64::engine::general_purpose::STANDARD.decode(body).ok())
+            .expect("raw provider stream should remain available for usage auditing");
+        assert_eq!(raw_capture, provider_buffered_body);
+        assert!(String::from_utf8_lossy(&raw_capture).contains("AGENTS.md secret prompt"));
+    }
+}

apps/aether-gateway/src/handlers/admin/observability/monitoring/tests/trace.rs

@@ -733,6 +733,110 @@ async fn admin_monitoring_trace_request_exposes_failed_candidate_upstream_respon
     assert!(extra.get("provider_response").is_none());
 }
 
+#[tokio::test]
+async fn admin_monitoring_trace_request_prefers_ref_backed_usage_response_body() {
+    let mut candidate = sample_candidate(
+        "cand-used",
+        "request-ref-body",
+        0,
+        RequestCandidateStatus::Failed,
+        Some(101),
+        Some(33),
+        Some(400),
+    );
+    candidate.extra_data = Some(json!({
+        "upstream_response": {
+            "status_code": 400,
+            "headers": {
+                "content-type": "text/event-stream",
+                "x-request-id": "stale-request-like-body"
+            },
+            "body": {
+                "model": "gpt-5.6-sol",
+                "input": [{"role": "user", "content": "request prompt"}]
+            }
+        }
+    }));
+
+    let request_candidates = Arc::new(InMemoryRequestCandidateRepository::seed(vec![candidate]));
+    let provider_catalog = Arc::new(InMemoryProviderCatalogReadRepository::seed(
+        vec![sample_provider()],
+        vec![sample_endpoint()],
+        vec![sample_key()],
+    ));
+    let mut usage = sample_usage(
+        "request-ref-body",
+        "provider-1",
+        "OpenAI",
+        0,
+        0.0,
+        "failed",
+        Some(400),
+        100,
+    );
+    usage.candidate_id = Some("cand-used".to_string());
+    usage.response_headers = Some(json!({
+        "content-type": "application/json",
+        "x-request-id": "req_usage-cyber-risk-demo"
+    }));
+    usage.response_body = Some(json!({
+        "error": {
+            "type": "invalid_request",
+            "message": "This content was flagged for possible cybersecurity risk.",
+            "code": 400
+        }
+    }));
+    usage.response_body_state = Some(UsageBodyCaptureState::Reference);
+    let usage_repository = Arc::new(InMemoryUsageReadRepository::seed_with_detached_bodies(
+        vec![usage],
+    ));
+    let data_state =
+        crate::data::GatewayDataState::with_request_candidate_and_usage_repository_for_tests(
+            request_candidates,
+            usage_repository,
+        )
+        .with_provider_catalog_reader(provider_catalog);
+    let state = AppState::new()
+        .expect("state should build")
+        .with_data_state_for_tests(data_state);
+    let context = request_context(
+        http::Method::GET,
+        "/api/admin/monitoring/trace/request-ref-body",
+    );
+
+    let response = local_monitoring_response(&state, &context)
+        .await
+        .expect("handler should not error")
+        .expect("route should be handled locally");
+
+    assert_eq!(response.status(), http::StatusCode::OK);
+    let body = to_bytes(response.into_body(), usize::MAX)
+        .await
+        .expect("body should read");
+    let payload: serde_json::Value = serde_json::from_slice(&body).expect("json body should parse");
+    let upstream_response = &payload["candidates"][0]["extra_data"]["upstream_response"];
+    assert_eq!(
+        upstream_response["headers"],
+        json!({
+            "content-type": "application/json",
+            "x-request-id": "req_usage-cyber-risk-demo"
+        })
+    );
+    assert_eq!(
+        upstream_response["body"]["error"],
+        json!({
+            "type": "invalid_request",
+            "message": "This content was flagged for possible cybersecurity risk.",
+            "code": 400
+        })
+    );
+    assert!(upstream_response["body"].get("input").is_none());
+    assert_eq!(
+        upstream_response["body_ref"],
+        json!("usage://request/request-ref-body/response_body")
+    );
+}
+
 #[tokio::test]
 async fn admin_monitoring_trace_request_decodes_connect_json_response_body_refs() {
     let mut candidate = sample_candidate(

apps/aether-gateway/src/handlers/admin/observability/monitoring/trace.rs

@@ -30,6 +30,25 @@ struct ResolvedAdminMonitoringTrace {
     usage: Option<StoredRequestUsageAudit>,
 }
 
+async fn hydrate_admin_monitoring_trace_response_body(
+    state: &AdminAppState<'_>,
+    mut usage: StoredRequestUsageAudit,
+) -> Result<StoredRequestUsageAudit, GatewayError> {
+    let is_error_node = !usage.status.eq_ignore_ascii_case("completed")
+        || usage
+            .status_code
+            .is_some_and(|status| !(200..300).contains(&status));
+    let response_body_ref = if is_error_node && usage.response_body.is_none() {
+        usage.response_body_ref.clone()
+    } else {
+        None
+    };
+    if let Some(body_ref) = response_body_ref.as_deref() {
+        usage.response_body = state.resolve_request_usage_body_ref(body_ref).await?;
+    }
+    Ok(usage)
+}
+
 pub(super) async fn build_admin_monitoring_trace_request_response(
     state: &AdminAppState<'_>,
     request_context: &AdminRequestContext<'_>,
@@ -92,6 +111,10 @@ async fn resolve_admin_monitoring_trace(
             .read_request_usage_audit_shallow(request_id)
             .await
             .map_err(|err| GatewayError::Internal(err.to_string()))?;
+        let usage = match usage {
+            Some(usage) => Some(hydrate_admin_monitoring_trace_response_body(state, usage).await?),
+            None => None,
+        };
         return Ok(Some(ResolvedAdminMonitoringTrace { trace, usage }));
     }
 
@@ -102,9 +125,10 @@ async fn resolve_admin_monitoring_trace(
         .await
         .map_err(|err| GatewayError::Internal(err.to_string()))?
     {
-        usage_candidates.push(usage);
+        usage_candidates.push(hydrate_admin_monitoring_trace_response_body(state, usage).await?);
     }
     if let Some(usage) = state.find_request_usage_by_id(request_id).await? {
+        let usage = hydrate_admin_monitoring_trace_response_body(state, usage).await?;
         if !usage_candidates.iter().any(|item| item.id == usage.id) {
             usage_candidates.push(usage);
         }

apps/aether-gateway/src/handlers/admin/provider/endpoint_keys/mutations/reset_cycle_stats.rs

@@ -106,7 +106,17 @@ fn reset_codex_cycle_usage_windows(status_snapshot: &mut Value, now_unix_secs: u
             .and_then(Value::as_str)
             .map(str::trim)
             .unwrap_or_default();
-        if !code.eq_ignore_ascii_case("5h") && !code.eq_ignore_ascii_case("weekly") {
+        let scope = window
+            .get("scope")
+            .and_then(Value::as_str)
+            .map(str::trim)
+            .unwrap_or("account");
+        let has_zero_window = window.get("window_minutes").and_then(Value::as_u64) == Some(0);
+        if code.is_empty()
+            || !scope.eq_ignore_ascii_case("account")
+            || code.to_ascii_lowercase().starts_with("spark_")
+            || has_zero_window
+        {
             continue;
         }
 
@@ -177,7 +187,7 @@ mod tests {
             }
         });
 
-        assert_eq!(reset_codex_cycle_usage_windows(&mut snapshot, 1_234), 2);
+        assert_eq!(reset_codex_cycle_usage_windows(&mut snapshot, 1_234), 3);
         let windows = snapshot["quota"]["windows"].as_array().expect("windows");
         assert_eq!(windows[0]["usage_reset_at"], json!(1_234));
         assert_eq!(windows[0]["usage"]["request_count"], json!(0));
@@ -187,7 +197,7 @@ mod tests {
         assert_eq!(windows[1]["usage"]["request_count"], json!(0));
         assert_eq!(windows[1]["usage"]["total_tokens"], json!(0));
         assert_eq!(windows[1]["usage"]["total_cost_usd"], json!("0.00000000"));
-        assert!(windows[2].get("usage_reset_at").is_none());
-        assert!(windows[2].get("usage").is_some());
+        assert_eq!(windows[2]["usage_reset_at"], json!(1_234));
+        assert_eq!(windows[2]["usage"]["request_count"], json!(0));
     }
 }

apps/aether-gateway/src/handlers/admin/provider/pool_admin/payloads.rs

@@ -255,12 +255,25 @@ fn admin_pool_quota_window_reset_seconds(
 
 fn admin_pool_codex_quota_part_from_window(
     quota_snapshot: &serde_json::Map<String, serde_json::Value>,
-    window_code: &str,
-    label: &str,
+    window: &serde_json::Map<String, serde_json::Value>,
     now_unix_secs: u64,
     show_reset_without_consumption: bool,
 ) -> Option<String> {
-    let window = admin_pool_quota_window(quota_snapshot, window_code)?;
+    if admin_pool_json_to_u64(window.get("window_minutes")) == Some(0) {
+        return None;
+    }
+    let label = window
+        .get("label")
+        .and_then(serde_json::Value::as_str)
+        .map(str::trim)
+        .filter(|label| !label.is_empty())
+        .or_else(|| {
+            window
+                .get("code")
+                .and_then(serde_json::Value::as_str)
+                .map(str::trim)
+                .filter(|code| !code.is_empty())
+        })?;
     let used_percent = admin_pool_quota_window_used_percent(window)?;
     let reset_seconds =
         admin_pool_quota_window_reset_seconds(quota_snapshot, window, now_unix_secs);
@@ -294,23 +307,35 @@ fn admin_pool_build_codex_account_quota_from_snapshot(
         .and_then(admin_provider_quota_pure::coerce_json_bool)
         .unwrap_or(false);
 
-    if let Some(part) = admin_pool_codex_quota_part_from_window(
-        quota_snapshot,
-        "weekly",
-        "周",
-        now_unix_secs,
-        exhausted,
-    ) {
-        parts.push(part);
-    }
-    if let Some(part) = admin_pool_codex_quota_part_from_window(
-        quota_snapshot,
-        "5h",
-        "5H",
-        now_unix_secs,
-        exhausted,
-    ) {
-        parts.push(part);
+    if let Some(windows) = quota_snapshot
+        .get("windows")
+        .and_then(serde_json::Value::as_array)
+    {
+        for window in windows.iter().filter_map(serde_json::Value::as_object) {
+            let code = window
+                .get("code")
+                .and_then(serde_json::Value::as_str)
+                .map(str::trim)
+                .unwrap_or_default();
+            let scope = window
+                .get("scope")
+                .and_then(serde_json::Value::as_str)
+                .map(str::trim)
+                .unwrap_or("account");
+            if !scope.eq_ignore_ascii_case("account")
+                || code.to_ascii_lowercase().starts_with("spark_")
+            {
+                continue;
+            }
+            if let Some(part) = admin_pool_codex_quota_part_from_window(
+                quota_snapshot,
+                window,
+                now_unix_secs,
+                exhausted,
+            ) {
+                parts.push(part);
+            }
+        }
     }
 
     if !parts.is_empty() {
@@ -345,6 +370,25 @@ fn admin_pool_current_unix_secs() -> u64 {
         .unwrap_or(0)
 }
 
+fn admin_pool_is_regular_codex_cycle_window(
+    window: &serde_json::Map<String, serde_json::Value>,
+) -> bool {
+    let code = window
+        .get("code")
+        .and_then(serde_json::Value::as_str)
+        .map(str::trim)
+        .unwrap_or_default();
+    let scope = window
+        .get("scope")
+        .and_then(serde_json::Value::as_str)
+        .map(str::trim)
+        .unwrap_or("account");
+    !code.is_empty()
+        && scope.eq_ignore_ascii_case("account")
+        && !code.to_ascii_lowercase().starts_with("spark_")
+        && admin_pool_json_to_u64(window.get("window_minutes")) != Some(0)
+}
+
 fn admin_pool_prune_expired_codex_window_usage_at(
     status_snapshot: &mut serde_json::Value,
     now_unix_secs: u64,
@@ -362,12 +406,7 @@ fn admin_pool_prune_expired_codex_window_usage_at(
         .iter_mut()
         .filter_map(serde_json::Value::as_object_mut)
     {
-        let code = window
-            .get("code")
-            .and_then(serde_json::Value::as_str)
-            .map(str::trim)
-            .unwrap_or_default();
-        if !code.eq_ignore_ascii_case("5h") && !code.eq_ignore_ascii_case("weekly") {
+        if !admin_pool_is_regular_codex_cycle_window(window) {
             continue;
         }
         let Some(reset_at) = admin_pool_json_to_u64(window.get("reset_at")) else {
@@ -1405,7 +1444,9 @@ mod tests {
             "quota": {
                 "windows": [
                     {
-                        "code": "5h",
+                        "code": "monthly",
+                        "scope": "account",
+                        "window_minutes": 43_800,
                         "reset_at": 1,
                         "usage": {
                             "request_count": 7,
@@ -1451,6 +1492,31 @@ mod tests {
         assert_eq!(usage["total_cost_usd"], json!("0.60000000"));
     }
 
+    #[test]
+    fn codex_monthly_quota_is_rendered_from_actual_snapshot_window() {
+        let quota_snapshot = json!({
+            "provider_type": "codex",
+            "code": "ok",
+            "exhausted": false,
+            "windows": [
+                {
+                    "code": "monthly",
+                    "label": "月",
+                    "scope": "account",
+                    "used_ratio": 0.14,
+                    "remaining_ratio": 0.86,
+                    "window_minutes": 43_800
+                }
+            ]
+        });
+        let quota_snapshot = quota_snapshot.as_object().unwrap();
+
+        assert_eq!(
+            admin_pool_build_account_quota("codex", Some(quota_snapshot)),
+            Some("月剩余 86.0%".to_string())
+        );
+    }
+
     #[test]
     fn grok_model_quota_is_rendered_for_pool_rows() {
         let quota_snapshot = json!({

apps/aether-gateway/src/handlers/admin/provider/pool_admin/read_routes/keys.rs

@@ -50,6 +50,8 @@ fn admin_pool_codex_default_window_minutes(code: &str) -> Option<u64> {
         Some(300)
     } else if code.eq_ignore_ascii_case("weekly") {
         Some(10_080)
+    } else if code.eq_ignore_ascii_case("monthly") {
+        Some(43_800)
     } else {
         None
     }
@@ -98,16 +100,27 @@ fn admin_pool_codex_cycle_usage_request(
     window: &serde_json::Map<String, serde_json::Value>,
     now_unix_secs: u64,
 ) -> Option<ProviderApiKeyWindowUsageRequest> {
+    let scope = window
+        .get("scope")
+        .and_then(serde_json::Value::as_str)
+        .map(str::trim)
+        .unwrap_or("account");
+    if !scope.eq_ignore_ascii_case("account") {
+        return None;
+    }
     let window_code = window
         .get("code")
         .and_then(serde_json::Value::as_str)
         .map(str::trim)
-        .filter(|code| code.eq_ignore_ascii_case("5h") || code.eq_ignore_ascii_case("weekly"))?
+        .filter(|code| !code.is_empty() && !code.to_ascii_lowercase().starts_with("spark_"))?
         .to_ascii_lowercase();
     let reset_at = admin_pool_json_u64(window.get("reset_at"))?;
-    let window_seconds = admin_pool_json_u64(window.get("window_minutes"))
-        .or_else(|| admin_pool_codex_default_window_minutes(&window_code))?
-        .checked_mul(60)?;
+    let window_minutes = match admin_pool_json_u64(window.get("window_minutes")) {
+        Some(0) => return None,
+        Some(value) => value,
+        None => admin_pool_codex_default_window_minutes(&window_code)?,
+    };
+    let window_seconds = window_minutes.checked_mul(60)?;
     if reset_at <= now_unix_secs {
         return None;
     }
@@ -788,6 +801,57 @@ mod tests {
         );
     }
 
+    #[test]
+    fn codex_cycle_usage_request_uses_actual_monthly_window_boundaries() {
+        let key = sample_key("oauth");
+        let reset_at = 5_000_000u64;
+        let now = 3_000_000u64;
+        let window = json!({
+            "code": "monthly",
+            "label": "月",
+            "scope": "account",
+            "reset_at": reset_at,
+            "window_minutes": 43_800u64
+        });
+
+        let request = admin_pool_codex_cycle_usage_request(
+            &key,
+            window.as_object().expect("window should be object"),
+            now,
+        )
+        .expect("monthly usage request should build");
+
+        assert_eq!(request.window_code, "monthly");
+        assert_eq!(request.start_unix_secs, reset_at - 43_800 * 60);
+        assert_eq!(request.end_unix_secs, now);
+    }
+
+    #[test]
+    fn codex_cycle_usage_request_ignores_zero_and_spark_windows() {
+        let key = sample_key("oauth");
+        for window in [
+            json!({
+                "code": "weekly",
+                "scope": "account",
+                "reset_at": 5_000_000u64,
+                "window_minutes": 0
+            }),
+            json!({
+                "code": "spark_weekly",
+                "scope": "account",
+                "reset_at": 5_000_000u64,
+                "window_minutes": 10_080
+            }),
+        ] {
+            assert!(admin_pool_codex_cycle_usage_request(
+                &key,
+                window.as_object().expect("window should be object"),
+                3_000_000,
+            )
+            .is_none());
+        }
+    }
+
     #[test]
     fn oauth_status_filter_prefers_catalog_key_expiry_over_auth_config_expiry() {
         let mut key = sample_key("oauth");

apps/aether-gateway/src/handlers/admin/users/api_keys/responses/create.rs

@@ -35,13 +35,16 @@ pub(crate) async fn build_admin_create_user_api_key_response(
     let Some(user_id) = admin_user_id_from_api_keys_path(request_context.path()) else {
         return Ok(build_admin_users_bad_request_response("缺少 user_id"));
     };
-    if state.find_user_auth_by_id(&user_id).await?.is_none() {
+    let Some(target_user) = state.find_user_auth_by_id(&user_id).await? else {
         return Ok((
             http::StatusCode::NOT_FOUND,
             Json(json!({ "detail": "用户不存在" })),
         )
             .into_response());
-    }
+    };
+    // The authenticated principal authorizes this admin operation, but ownership and inherited
+    // policies must always come from the user selected in the request path.
+    let target_user_id = target_user.id;
 
     let Some(request_body) = request_body else {
         return Ok((
@@ -148,7 +151,7 @@ pub(crate) async fn build_admin_create_user_api_key_response(
 
     let Some(created) = state
         .create_user_api_key(aether_data::repository::auth::CreateUserApiKeyRecord {
-            user_id: user_id.clone(),
+            user_id: target_user_id.clone(),
             api_key_id: uuid::Uuid::new_v4().to_string(),
             key_hash: hash_admin_user_api_key(&plaintext_key),
             key_encrypted: Some(key_encrypted),
@@ -174,7 +177,11 @@ pub(crate) async fn build_admin_create_user_api_key_response(
 
     let created = if allowed_providers.is_some() {
         match state
-            .set_user_api_key_allowed_providers(&user_id, &created.api_key_id, allowed_providers)
+            .set_user_api_key_allowed_providers(
+                &target_user_id,
+                &created.api_key_id,
+                allowed_providers,
+            )
             .await?
         {
             Some(updated) => updated,
@@ -186,7 +193,7 @@ pub(crate) async fn build_admin_create_user_api_key_response(
     let created = if feature_settings.is_some() {
         match state
             .set_user_api_key_feature_settings(
-                &user_id,
+                &target_user_id,
                 &created.api_key_id,
                 feature_settings.clone(),
             )

apps/aether-gateway/src/handlers/public/support/user_me_usage.rs

@@ -506,6 +506,9 @@ fn build_users_me_usage_record_payload(
     if let Some(reasoning_effort) = item.provider_reasoning_effort() {
         payload["reasoning_effort"] = json!(reasoning_effort);
     }
+    if let Some(requested_reasoning_effort) = item.requested_reasoning_effort() {
+        payload["requested_reasoning_effort"] = json!(requested_reasoning_effort);
+    }
     if let Some(service_tier) = item.provider_service_tier() {
         payload["service_tier"] = json!(service_tier);
     }
@@ -577,6 +580,9 @@ fn build_users_me_usage_active_payload(item: &StoredRequestUsageAudit) -> serde_
     if let Some(reasoning_effort) = item.provider_reasoning_effort() {
         payload["reasoning_effort"] = json!(reasoning_effort);
     }
+    if let Some(requested_reasoning_effort) = item.requested_reasoning_effort() {
+        payload["requested_reasoning_effort"] = json!(requested_reasoning_effort);
+    }
     if let Some(service_tier) = item.provider_service_tier() {
         payload["service_tier"] = json!(service_tier);
     }
@@ -1655,6 +1661,27 @@ mod tests {
         assert_eq!(payload["cache_creation_ephemeral_1h_input_tokens"], 6);
     }
 
+    #[test]
+    fn user_usage_payloads_expose_requested_and_provider_reasoning_mapping() {
+        let item = StoredRequestUsageAudit {
+            request_body: Some(json!({
+                "reasoning": { "effort": "xhigh" }
+            })),
+            provider_request_body: Some(json!({
+                "reasoning": { "effort": "max" }
+            })),
+            ..sample_usage("completed")
+        };
+
+        let record = build_users_me_usage_record_payload(&item, false, &BTreeMap::new(), false);
+        let active = build_users_me_usage_active_payload(&item);
+
+        assert_eq!(record["requested_reasoning_effort"], "xhigh");
+        assert_eq!(active["requested_reasoning_effort"], "xhigh");
+        assert_eq!(record["reasoning_effort"], "max");
+        assert_eq!(active["reasoning_effort"], "max");
+    }
+
     #[test]
     fn user_usage_active_override_uses_terminal_candidate_latency() {
         let candidate = sample_candidate(

apps/aether-gateway/src/handlers/shared/catalog.rs

@@ -806,6 +806,50 @@ fn codex_default_window_minutes(code: &str) -> Option<u64> {
     }
 }
 
+fn codex_quota_period_identity(window_minutes: u64) -> (String, String) {
+    const MINUTES_PER_HOUR: u64 = 60;
+    const MINUTES_PER_DAY: u64 = 24 * MINUTES_PER_HOUR;
+    const MINUTES_PER_WEEK: u64 = 7 * MINUTES_PER_DAY;
+
+    if window_minutes == 5 * MINUTES_PER_HOUR {
+        return ("5h".to_string(), "5H".to_string());
+    }
+    if window_minutes == MINUTES_PER_WEEK {
+        return ("weekly".to_string(), "周".to_string());
+    }
+    if (28 * MINUTES_PER_DAY..=31 * MINUTES_PER_DAY).contains(&window_minutes) {
+        return ("monthly".to_string(), "月".to_string());
+    }
+
+    let label = if window_minutes % MINUTES_PER_WEEK == 0 {
+        format!("{}周", window_minutes / MINUTES_PER_WEEK)
+    } else if window_minutes % MINUTES_PER_DAY == 0 {
+        format!("{}天", window_minutes / MINUTES_PER_DAY)
+    } else if window_minutes % MINUTES_PER_HOUR == 0 {
+        format!("{}H", window_minutes / MINUTES_PER_HOUR)
+    } else {
+        format!("{window_minutes}分钟")
+    };
+    (format!("window_{window_minutes}m"), label)
+}
+
+fn codex_quota_window_identity(
+    fallback_code: &str,
+    fallback_label: &str,
+    window_minutes: Option<u64>,
+) -> (String, String) {
+    let Some(window_minutes) = window_minutes else {
+        return (fallback_code.to_string(), fallback_label.to_string());
+    };
+    let is_spark = fallback_code.to_ascii_lowercase().starts_with("spark_");
+    let (code, label) = codex_quota_period_identity(window_minutes);
+    if is_spark {
+        (format!("spark_{code}"), format!("Spark {label}"))
+    } else {
+        (code, label)
+    }
+}
+
 fn codex_quota_window_snapshot(
     metadata: &Map<String, Value>,
     prefix: &str,
@@ -847,6 +891,10 @@ fn codex_quota_window_snapshot(
         .get(&window_minutes_key)
         .and_then(admin_provider_quota_pure::coerce_json_u64);
 
+    if explicit_window_minutes == Some(0) {
+        return None;
+    }
+
     if used_percent.is_none()
         && reset_at.is_none()
         && reset_seconds.is_none()
@@ -856,6 +904,7 @@ fn codex_quota_window_snapshot(
     }
 
     let window_minutes = explicit_window_minutes.or_else(|| codex_default_window_minutes(code));
+    let (code, label) = codex_quota_window_identity(code, label, window_minutes);
     let used_ratio = used_percent.map(|value| (value / 100.0).clamp(0.0, 1.0));
     let remaining_ratio = used_ratio.map(|value| (1.0 - value).max(0.0));
 
@@ -931,12 +980,16 @@ fn build_codex_quota_status_snapshot(
     let primary_windows = windows
         .iter()
         .filter(|window| {
-            window
+            let code = window
                 .get("code")
                 .and_then(Value::as_str)
-                .is_some_and(|code| {
-                    code.eq_ignore_ascii_case("weekly") || code.eq_ignore_ascii_case("5h")
-                })
+                .unwrap_or_default();
+            let scope = window
+                .get("scope")
+                .and_then(Value::as_str)
+                .unwrap_or_default();
+            scope.eq_ignore_ascii_case("account")
+                && !code.to_ascii_lowercase().starts_with("spark_")
         })
         .cloned()
         .collect::<Vec<_>>();
@@ -3792,6 +3845,39 @@ mod tests {
         assert_eq!(five_h.get("window_minutes"), Some(&json!(300u64)));
     }
 
+    #[test]
+    fn sync_provider_key_quota_status_snapshot_labels_actual_monthly_window() {
+        let upstream_metadata = json!({
+            "codex": {
+                "updated_at": 1_784_287_450u64,
+                "plan_type": "team",
+                "primary_used_percent": 14.0,
+                "primary_reset_at": 1_786_915_122u64,
+                "primary_window_minutes": 43_800u64,
+                "secondary_used_percent": 0.0,
+                "secondary_reset_after_seconds": 0u64,
+                "secondary_window_minutes": 0u64
+            }
+        });
+
+        let payload = sync_provider_key_quota_status_snapshot(
+            None,
+            "codex",
+            Some(&upstream_metadata),
+            "response_headers",
+        )
+        .expect("quota snapshot should sync");
+        let windows = payload["quota"]["windows"]
+            .as_array()
+            .expect("quota windows should exist");
+
+        assert_eq!(windows.len(), 1);
+        assert_eq!(windows[0]["code"], json!("monthly"));
+        assert_eq!(windows[0]["label"], json!("月"));
+        assert_eq!(windows[0]["window_minutes"], json!(43_800u64));
+        assert_eq!(windows[0]["remaining_ratio"], json!(0.86));
+    }
+
     #[test]
     fn provider_key_status_snapshot_payload_backfills_thin_ok_snapshot_from_upstream_metadata() {
         let mut key = sample_catalog_key();

apps/aether-gateway/src/tests/ai_execute/stream_cli/compact.rs

@@ -657,7 +657,8 @@ async fn gateway_executes_openai_responses_compact_as_unary_request_impl() {
     );
     assert_eq!(
         seen_execution_runtime_request.body["prompt_cache_key"],
-        json!("session:compact-e2e")
+        // Compact omits client_metadata, but keeps the same deterministic Codex cache identity.
+        json!("f3eb8726-b7b2-56c5-90b5-8789d628c8cf")
     );
     assert_eq!(
         seen_execution_runtime_request.proxy_node_id,

apps/aether-gateway/src/tests/control/admin/pool.rs

@@ -1362,6 +1362,18 @@ async fn gateway_pool_list_overrides_stale_codex_cycle_usage_from_usage_facts()
                         "total_tokens": 700,
                         "total_cost_usd": "0.70000000"
                     }
+                },
+                {
+                    "code": "monthly",
+                    "label": "月",
+                    "scope": "account",
+                    "reset_at": reset_at,
+                    "window_minutes": 43_800,
+                    "usage": {
+                        "request_count": 8,
+                        "total_tokens": 600,
+                        "total_cost_usd": "0.60000000"
+                    }
                 }
             ]
         }
@@ -1441,6 +1453,10 @@ async fn gateway_pool_list_overrides_stale_codex_cycle_usage_from_usage_facts()
         .iter()
         .find(|window| window["code"] == json!("5h"))
         .expect("5h window should exist");
+    let monthly = windows
+        .iter()
+        .find(|window| window["code"] == json!("monthly"))
+        .expect("monthly window should exist");
 
     assert_eq!(weekly["usage"]["request_count"], json!(3));
     assert_eq!(weekly["usage"]["total_tokens"], json!(2_199));
@@ -1448,6 +1464,9 @@ async fn gateway_pool_list_overrides_stale_codex_cycle_usage_from_usage_facts()
     assert_eq!(five_hour["usage"]["request_count"], json!(1));
     assert_eq!(five_hour["usage"]["total_tokens"], json!(200));
     assert_eq!(five_hour["usage"]["total_cost_usd"], json!("0.75000000"));
+    assert_eq!(monthly["usage"]["request_count"], json!(3));
+    assert_eq!(monthly["usage"]["total_tokens"], json!(2_199));
+    assert_eq!(monthly["usage"]["total_cost_usd"], json!("11.99000000"));
 }
 
 #[tokio::test]

apps/aether-gateway/src/tests/control/admin/users.rs

@@ -20,7 +20,7 @@ use http::StatusCode;
 use serde_json::json;
 
 use super::super::{
-    build_router_with_state, issue_test_admin_access_token, start_server, AppState,
+    build_router_with_state, hash_api_key, issue_test_admin_access_token, start_server, AppState,
 };
 use crate::constants::{
     GATEWAY_HEADER, TRUSTED_ADMIN_SESSION_ID_HEADER, TRUSTED_ADMIN_USER_ID_HEADER,
@@ -1672,6 +1672,192 @@ async fn gateway_handles_admin_user_api_key_routes_locally_with_trusted_admin_pr
     upstream_handle.abort();
 }
 
+#[tokio::test]
+async fn admin_created_user_key_inherits_target_user_policy_not_admin_policy() {
+    let mut admin_snapshot = sample_admin_api_key_snapshot("admin-user", "admin-seed-key");
+    admin_snapshot.user_role = "admin".to_string();
+    let auth_repository = Arc::new(InMemoryAuthApiKeySnapshotRepository::seed(vec![
+        (Some(hash_api_key("sk-admin-seed")), admin_snapshot),
+        (
+            Some(hash_api_key("sk-target-seed")),
+            sample_admin_api_key_snapshot("target-user", "target-seed-key"),
+        ),
+    ]));
+    let user_repository = Arc::new(InMemoryUserReadRepository::seed_auth_users(vec![
+        sample_admin_user_with_role("admin-user", "admin", "admin@example.com", "admin"),
+        sample_admin_user_with_role("target-user", "user", "target@example.com", "target"),
+    ]));
+
+    let admin_group = user_repository
+        .create_user_group(UpsertUserGroupRecord {
+            name: "Admin OpenAI".to_string(),
+            description: None,
+            priority: 10,
+            allowed_providers: Some(vec!["openai".to_string()]),
+            allowed_providers_mode: "specific".to_string(),
+            allowed_api_formats: Some(vec!["openai:responses".to_string()]),
+            allowed_api_formats_mode: "specific".to_string(),
+            allowed_models: Some(vec!["gpt-5.4".to_string()]),
+            allowed_models_mode: "specific".to_string(),
+            rate_limit: Some(100),
+            rate_limit_mode: "custom".to_string(),
+        })
+        .await
+        .expect("admin group should create")
+        .expect("admin group should exist");
+    user_repository
+        .add_user_to_group(&admin_group.id, "admin-user")
+        .await
+        .expect("admin group membership should create");
+
+    let target_group = user_repository
+        .create_user_group(UpsertUserGroupRecord {
+            name: "Target Claude".to_string(),
+            description: None,
+            priority: 10,
+            allowed_providers: Some(vec!["anthropic".to_string()]),
+            allowed_providers_mode: "specific".to_string(),
+            allowed_api_formats: Some(vec!["claude:messages".to_string()]),
+            allowed_api_formats_mode: "specific".to_string(),
+            allowed_models: Some(vec!["claude-sonnet-4-5".to_string()]),
+            allowed_models_mode: "specific".to_string(),
+            rate_limit: Some(30),
+            rate_limit_mode: "custom".to_string(),
+        })
+        .await
+        .expect("target group should create")
+        .expect("target group should exist");
+    user_repository
+        .add_user_to_group(&target_group.id, "target-user")
+        .await
+        .expect("target group membership should create");
+    user_repository
+        .update_user_feature_settings(
+            "admin-user",
+            Some(json!({"chat_pii_redaction": {"enabled": false}})),
+        )
+        .await
+        .expect("admin feature settings should update");
+    user_repository
+        .update_user_feature_settings(
+            "target-user",
+            Some(json!({"chat_pii_redaction": {"enabled": true}})),
+        )
+        .await
+        .expect("target feature settings should update");
+
+    let data_state = GatewayDataState::with_auth_api_key_repository_for_tests(auth_repository)
+        .with_user_reader(user_repository);
+    let app_state = AppState::new()
+        .expect("gateway should build")
+        .with_data_state_for_tests(data_state);
+    let inspection_state = app_state.clone();
+    let gateway = build_router_with_state(app_state);
+    let (gateway_url, gateway_handle) = start_server(gateway).await;
+
+    let response = reqwest::Client::new()
+        .post(format!(
+            "{gateway_url}/api/admin/users/target-user/api-keys"
+        ))
+        .header(crate::constants::GATEWAY_HEADER, "rust-phase3b")
+        .header(TRUSTED_ADMIN_USER_ID_HEADER, "admin-user")
+        .header(TRUSTED_ADMIN_USER_ROLE_HEADER, "admin")
+        .header(TRUSTED_ADMIN_SESSION_ID_HEADER, "admin-session")
+        .json(&json!({"name": "target-key"}))
+        .send()
+        .await
+        .expect("request should succeed");
+    assert_eq!(response.status(), StatusCode::OK);
+    let payload: serde_json::Value = response.json().await.expect("response should parse");
+    let plaintext_key = payload["key"].as_str().expect("plaintext key should exist");
+    let api_key_id = payload["id"].as_str().expect("key id should exist");
+
+    let resolved = inspection_state
+        .read_cached_auth_api_key_snapshot_by_key_hash(
+            &hash_api_key(plaintext_key),
+            chrono::Utc::now().timestamp().max(0) as u64,
+        )
+        .await
+        .expect("created key snapshot should resolve")
+        .expect("created key snapshot should exist");
+    assert_eq!(resolved.user_id, "target-user");
+    assert_eq!(
+        resolved.effective_allowed_providers(),
+        Some(&["anthropic".to_string()][..])
+    );
+    assert_eq!(
+        resolved.effective_allowed_api_formats(),
+        Some(&["claude:messages".to_string()][..])
+    );
+    assert_eq!(
+        resolved.effective_allowed_models(),
+        Some(&["claude-sonnet-4-5".to_string()][..])
+    );
+    assert_eq!(resolved.user_rate_limit, Some(30));
+
+    inspection_state
+        .update_user_group(
+            &target_group.id,
+            UpsertUserGroupRecord {
+                name: "Target Gemini".to_string(),
+                description: None,
+                priority: 10,
+                allowed_providers: Some(vec!["google".to_string()]),
+                allowed_providers_mode: "specific".to_string(),
+                allowed_api_formats: Some(vec!["gemini:generate-content".to_string()]),
+                allowed_api_formats_mode: "specific".to_string(),
+                allowed_models: Some(vec!["gemini-2.5-pro".to_string()]),
+                allowed_models_mode: "specific".to_string(),
+                rate_limit: Some(15),
+                rate_limit_mode: "custom".to_string(),
+            },
+        )
+        .await
+        .expect("target group should update")
+        .expect("target group should still exist");
+    let updated = inspection_state
+        .read_cached_auth_api_key_snapshot_by_key_hash(
+            &hash_api_key(plaintext_key),
+            chrono::Utc::now().timestamp().max(0) as u64,
+        )
+        .await
+        .expect("created key snapshot should refresh")
+        .expect("created key snapshot should still exist");
+    assert_eq!(updated.user_id, "target-user");
+    assert_eq!(
+        updated.effective_allowed_providers(),
+        Some(&["google".to_string()][..])
+    );
+    assert_eq!(
+        updated.effective_allowed_api_formats(),
+        Some(&["gemini:generate-content".to_string()][..])
+    );
+    assert_eq!(
+        updated.effective_allowed_models(),
+        Some(&["gemini-2.5-pro".to_string()][..])
+    );
+    assert_eq!(updated.user_rate_limit, Some(15));
+
+    let target_features = inspection_state
+        .read_user_feature_settings("target-user")
+        .await
+        .expect("target feature settings should resolve");
+    let key_features = inspection_state
+        .read_auth_api_key_feature_settings("target-user", api_key_id, false)
+        .await
+        .expect("key feature settings should resolve");
+    assert_eq!(
+        target_features,
+        Some(json!({"chat_pii_redaction": {"enabled": true}}))
+    );
+    assert_eq!(
+        key_features, None,
+        "an omitted key override must inherit target settings"
+    );
+
+    gateway_handle.abort();
+}
+
 #[tokio::test]
 async fn gateway_returns_conflict_for_admin_create_user_api_key_when_writer_unavailable() {
     let upstream_hits = Arc::new(Mutex::new(0usize));

crates/aether-admin/src/observability/usage.rs

@@ -1231,6 +1231,9 @@ fn admin_usage_active_request_json(
     if let Some(reasoning_effort) = item.provider_reasoning_effort() {
         value["reasoning_effort"] = json!(reasoning_effort);
     }
+    if let Some(requested_reasoning_effort) = item.requested_reasoning_effort() {
+        value["requested_reasoning_effort"] = json!(requested_reasoning_effort);
+    }
     if let Some(service_tier) = item.provider_service_tier() {
         value["service_tier"] = json!(service_tier);
     }
@@ -1298,6 +1301,7 @@ pub fn admin_usage_record_json(
         "response_time_ms": item.response_time_ms,
         "first_byte_time_ms": item.first_byte_time_ms,
         "created_at": unix_secs_to_rfc3339(item.created_at_unix_ms),
+        "updated_at": unix_secs_to_rfc3339(item.updated_at_unix_secs),
         "input_price_per_1m": input_price_per_1m,
         "output_price_per_1m": output_price_per_1m,
         "cache_creation_price_per_1m": cache_creation_price_per_1m,
@@ -1354,6 +1358,12 @@ pub fn admin_usage_record_json(
     if let Some(reasoning_effort) = item.provider_reasoning_effort() {
         object.insert("reasoning_effort".to_string(), json!(reasoning_effort));
     }
+    if let Some(requested_reasoning_effort) = item.requested_reasoning_effort() {
+        object.insert(
+            "requested_reasoning_effort".to_string(),
+            json!(requested_reasoning_effort),
+        );
+    }
     if let Some(service_tier) = item.provider_service_tier() {
         object.insert("service_tier".to_string(), json!(service_tier));
     }
@@ -2702,10 +2712,13 @@ mod tests {
     }
 
     #[test]
-    fn admin_usage_record_includes_provider_reasoning_effort() {
+    fn admin_usage_record_includes_requested_and_provider_reasoning_efforts() {
         let item = StoredRequestUsageAudit {
+            request_body: Some(json!({
+                "reasoning": { "effort": "xhigh" }
+            })),
             provider_request_body: Some(json!({
-                "reasoning": { "effort": "xhigh" },
+                "reasoning": { "effort": "max" },
                 "service_tier": "priority"
             })),
             ..sample_usage("completed", Some(200), None)
@@ -2721,10 +2734,14 @@ mod tests {
         );
         let active = admin_usage_active_request_json(&item, None, None, None);
 
-        assert_eq!(record["reasoning_effort"], "xhigh");
-        assert_eq!(active["reasoning_effort"], "xhigh");
+        assert_eq!(record["requested_reasoning_effort"], "xhigh");
+        assert_eq!(active["requested_reasoning_effort"], "xhigh");
+        assert_eq!(record["reasoning_effort"], "max");
+        assert_eq!(active["reasoning_effort"], "max");
         assert_eq!(record["service_tier"], "priority");
         assert_eq!(active["service_tier"], "priority");
+        assert!(record["updated_at"].is_string());
+        assert_eq!(record["updated_at"], active["updated_at"]);
     }
 
     #[test]

crates/aether-admin/src/provider/quota.rs

@@ -630,6 +630,25 @@ fn codex_write_window(
     }
 }
 
+fn codex_window_has_active_limit(source: &serde_json::Map<String, serde_json::Value>) -> bool {
+    [
+        "window_minutes",
+        "limit_window_seconds",
+        "reset_after_seconds",
+        "reset_at",
+    ]
+    .iter()
+    .any(|key| {
+        source
+            .get(*key)
+            .and_then(coerce_json_u64)
+            .is_some_and(|value| value > 0)
+    }) || source
+        .get("used_percent")
+        .and_then(coerce_json_f64)
+        .is_some_and(|value| value > 0.0)
+}
+
 fn codex_find_spark_rate_limit(
     root: &serde_json::Map<String, serde_json::Value>,
 ) -> Option<&serde_json::Map<String, serde_json::Value>> {
@@ -719,7 +738,8 @@ pub fn parse_codex_wham_usage_response(
         .cloned()
         .unwrap_or_default();
 
-    let use_paid_windows = !secondary_window.is_empty() && plan_type.as_deref() != Some("free");
+    let use_paid_windows =
+        codex_window_has_active_limit(&secondary_window) && plan_type.as_deref() != Some("free");
     if use_paid_windows {
         codex_write_window(&mut result, &secondary_window, "primary");
         codex_write_window(&mut result, &primary_window, "secondary");
@@ -1173,7 +1193,8 @@ pub fn parse_codex_usage_headers(
 
     let primary_window = read_window("primary");
     let secondary_window = read_window("secondary");
-    let use_paid_windows = !secondary_window.is_empty() && plan_type.as_deref() != Some("free");
+    let use_paid_windows =
+        codex_window_has_active_limit(&secondary_window) && plan_type.as_deref() != Some("free");
     if use_paid_windows {
         codex_write_window(&mut result, &secondary_window, "primary");
         codex_write_window(&mut result, &primary_window, "secondary");
@@ -2091,8 +2112,8 @@ mod tests {
         codex_build_invalid_state, codex_runtime_invalid_reason,
         normalize_codex_reset_credit_consume_outcome, parse_antigravity_usage_response,
         parse_chatgpt_web_conversation_init_response, parse_codex_backend_me_response,
-        parse_codex_wham_reset_credits_detail_response, parse_codex_wham_usage_response,
-        parse_gemini_cli_retrieve_user_quota_response,
+        parse_codex_usage_headers, parse_codex_wham_reset_credits_detail_response,
+        parse_codex_wham_usage_response, parse_gemini_cli_retrieve_user_quota_response,
         parse_gemini_cli_v1internal_credits_response, parse_windsurf_model_configs_response,
         parse_windsurf_rate_limit_response, parse_windsurf_user_status_response,
         provider_auto_remove_quota_exhausted_keys, quota_refresh_success_invalid_state,
@@ -2101,6 +2122,7 @@ mod tests {
     };
     use aether_data_contracts::repository::provider_catalog::StoredProviderCatalogKey;
     use serde_json::json;
+    use std::collections::BTreeMap;
 
     #[test]
     fn provider_auto_remove_quota_exhausted_keys_defaults_to_false() {
@@ -2506,6 +2528,50 @@ mod tests {
         );
     }
 
+    #[test]
+    fn parses_codex_monthly_header_without_zero_secondary_placeholder() {
+        let headers = BTreeMap::from([
+            ("x-codex-plan-type".to_string(), "team".to_string()),
+            ("x-codex-primary-used-percent".to_string(), "14".to_string()),
+            (
+                "x-codex-primary-reset-after-seconds".to_string(),
+                "2627672".to_string(),
+            ),
+            (
+                "x-codex-primary-reset-at".to_string(),
+                "1786915122".to_string(),
+            ),
+            (
+                "x-codex-primary-window-minutes".to_string(),
+                "43800".to_string(),
+            ),
+            (
+                "x-codex-secondary-used-percent".to_string(),
+                "0".to_string(),
+            ),
+            (
+                "x-codex-secondary-reset-after-seconds".to_string(),
+                "0".to_string(),
+            ),
+            ("x-codex-secondary-reset-at".to_string(), "".to_string()),
+            (
+                "x-codex-secondary-window-minutes".to_string(),
+                "0".to_string(),
+            ),
+        ]);
+
+        let parsed = parse_codex_usage_headers(&headers, 1_784_287_450)
+            .expect("Codex usage headers should parse");
+
+        assert_eq!(parsed.get("primary_used_percent"), Some(&json!(14.0)));
+        assert_eq!(
+            parsed.get("primary_window_minutes"),
+            Some(&json!(43_800u64))
+        );
+        assert!(parsed.get("secondary_used_percent").is_none());
+        assert!(parsed.get("secondary_window_minutes").is_none());
+    }
+
     #[test]
     fn parses_codex_reset_credit_count_from_wham_usage() {
         let parsed = parse_codex_wham_usage_response(

crates/aether-ai/formats/src/api.rs

@@ -173,6 +173,7 @@ pub use crate::formats::{
             codex::{
                 apply_codex_openai_compact_terminal_headers,
                 apply_codex_openai_responses_chat_body_edits,
+                apply_codex_openai_responses_identity_headers,
                 apply_codex_openai_responses_lite_header_for_request_body_with_capabilities,
                 apply_codex_openai_responses_lite_header_with_capabilities,
                 apply_codex_openai_responses_special_body_edits,

crates/aether-ai/formats/src/formats/openai/responses/codex.rs

@@ -7,6 +7,8 @@ use serde_json::{json, Value};
 
 const CODEX_DEFAULT_REASONING_EFFORT: &str = "medium";
 const CODEX_REASONING_ENCRYPTED_CONTENT_INCLUDE: &str = "reasoning.encrypted_content";
+const CODEX_PROMPT_CACHE_IDENTITY_NAMESPACE: &str =
+    "https://github.com/fawney19/Aether/codex/prompt-cache-identity/v1/";
 pub const CODEX_RESPONSES_LITE_HEADER: &str = "x-openai-internal-codex-responses-lite";
 pub const CODEX_MODEL_CATALOG_METADATA_FIELD: &str = "codex_models";
 const CODEX_OPENAI_RESPONSES_UNSUPPORTED_BODY_FIELDS: &[&str] = &[
@@ -1466,6 +1468,148 @@ fn wrap_codex_responses_string_input_for_backend(
     );
 }
 
+fn non_empty_json_string(value: Option<&Value>) -> Option<&str> {
+    value
+        .and_then(Value::as_str)
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+}
+
+fn adapt_codex_prompt_cache_identity_for_backend(body_object: &mut serde_json::Map<String, Value>) {
+    // The standard OpenAI contract permits arbitrary prompt_cache_key strings, while the Codex
+    // backend's native session identity is UUID-shaped. Adapt only requests that carry a cache key
+    // but do not already carry a native Codex session identity.
+    let Some(prompt_cache_key) = body_object
+        .get("prompt_cache_key")
+        .and_then(Value::as_str)
+        .filter(|value| !value.trim().is_empty())
+        .map(ToOwned::to_owned)
+    else {
+        return;
+    };
+
+    match body_object.get("client_metadata") {
+        Some(Value::Object(metadata)) => {
+            match metadata.get("session_id") {
+                Some(Value::String(session_id)) if !session_id.trim().is_empty() => return,
+                None | Some(Value::Null) => {}
+                Some(_) => return,
+            }
+            match metadata.get("thread_id") {
+                None | Some(Value::Null) => {}
+                Some(Value::String(thread_id)) if !thread_id.trim().is_empty() => {}
+                Some(_) => return,
+            }
+        }
+        Some(Value::Null) | None => {}
+        Some(_) => return,
+    }
+
+    let cache_identity = uuid::Uuid::parse_str(&prompt_cache_key)
+        .unwrap_or_else(|_| {
+            uuid::Uuid::new_v5(
+                &uuid::Uuid::NAMESPACE_URL,
+                format!("{CODEX_PROMPT_CACHE_IDENTITY_NAMESPACE}{prompt_cache_key}").as_bytes(),
+            )
+        })
+        .to_string();
+
+    {
+        let metadata = body_object
+            .entry("client_metadata".to_string())
+            .or_insert_with(|| json!({}));
+        if metadata.is_null() {
+            *metadata = json!({});
+        }
+        let Some(metadata) = metadata.as_object_mut() else {
+            return;
+        };
+        metadata.insert(
+            "session_id".to_string(),
+            Value::String(cache_identity.clone()),
+        );
+        if metadata.get("thread_id").is_none_or(Value::is_null) {
+            metadata.insert(
+                "thread_id".to_string(),
+                Value::String(cache_identity.clone()),
+            );
+        }
+    }
+
+    body_object.insert(
+        "prompt_cache_key".to_string(),
+        Value::String(cache_identity),
+    );
+}
+
+fn valid_codex_identity_header(value: &str) -> Option<String> {
+    let value = value.trim();
+    if value.is_empty() || value.parse::<http::HeaderValue>().is_err() {
+        return None;
+    }
+    Some(value.to_string())
+}
+
+fn codex_prompt_cache_header_identity(provider_request_body: &Value) -> Option<(String, String)> {
+    let prompt_cache_key = non_empty_json_string(provider_request_body.get("prompt_cache_key"))?;
+
+    match provider_request_body.get("client_metadata") {
+        Some(Value::Object(metadata)) => match metadata.get("session_id") {
+            Some(Value::String(session_id)) if !session_id.trim().is_empty() => {
+                let session_id = valid_codex_identity_header(session_id)?;
+                let thread_id = match metadata.get("thread_id") {
+                    None | Some(Value::Null) => session_id.clone(),
+                    Some(Value::String(thread_id)) if !thread_id.trim().is_empty() => {
+                        valid_codex_identity_header(thread_id)?
+                    }
+                    Some(_) => return None,
+                };
+                return Some((session_id, thread_id));
+            }
+            None | Some(Value::Null) => {}
+            Some(_) => return None,
+        },
+        Some(Value::Null) | None => {}
+        Some(_) => return None,
+    }
+
+    let cache_identity = uuid::Uuid::parse_str(prompt_cache_key).ok()?.to_string();
+    Some((cache_identity.clone(), cache_identity))
+}
+
+fn insert_btree_header_if_missing(
+    headers: &mut BTreeMap<String, String>,
+    header_name: &str,
+    header_value: String,
+) {
+    if headers.iter().any(|(name, value)| {
+        name.trim().eq_ignore_ascii_case(header_name) && !value.trim().is_empty()
+    }) {
+        return;
+    }
+    remove_btree_header(headers, header_name);
+    headers.insert(header_name.to_string(), header_value);
+}
+
+pub fn apply_codex_openai_responses_identity_headers(
+    provider_request_headers: &mut BTreeMap<String, String>,
+    provider_request_body: &Value,
+    provider_type: &str,
+    provider_api_format: &str,
+) {
+    // Codex projects its body session identity into compatibility HTTP headers. Existing explicit
+    // headers remain authoritative; only missing values are completed here.
+    if !is_codex_openai_responses_request(provider_type, provider_api_format) {
+        return;
+    }
+    let Some((session_id, thread_id)) = codex_prompt_cache_header_identity(provider_request_body)
+    else {
+        return;
+    };
+    insert_btree_header_if_missing(provider_request_headers, "session-id", session_id);
+    insert_btree_header_if_missing(provider_request_headers, "thread-id", thread_id);
+}
+
 pub fn apply_codex_openai_responses_special_body_edits(
     provider_request_body: &mut Value,
     provider_type: &str,
@@ -1532,6 +1676,7 @@ pub fn apply_codex_openai_responses_special_body_edits_with_source_model_and_cap
             body_object.remove(*field);
         }
     }
+    adapt_codex_prompt_cache_identity_for_backend(body_object);
     if is_openai_responses_compact_request(provider_api_format) {
         body_object.remove("store");
     } else if !body_rules_handle_path(body_rules, "store") {

crates/aether-ai/formats/src/provider_compat/private_envelope.rs

@@ -2,6 +2,7 @@ use std::collections::BTreeMap;
 
 use serde_json::Value;
 
+use crate::formats::shared::stream_core::common::openai_stream_terminal_error_body;
 use crate::formats::shared::AiSurfaceFinalizeError;
 use crate::provider_compat::kiro_stream::KiroToClaudeCliStreamState;
 
@@ -719,6 +720,27 @@ fn extract_stream_error_event_body(body: &[u8]) -> Option<Value> {
         {
             return Some(normalize_provider_private_error_body(event));
         }
+        if let Some(mut error_body) = openai_stream_terminal_error_body(&event) {
+            let response_failed_without_provider_type = event
+                .get("type")
+                .and_then(Value::as_str)
+                .is_some_and(|event_type| event_type.eq_ignore_ascii_case("response.failed"))
+                && event
+                    .pointer("/response/error/type")
+                    .or_else(|| event.pointer("/error/type"))
+                    .and_then(Value::as_str)
+                    .map(str::trim)
+                    .is_none_or(str::is_empty);
+            if response_failed_without_provider_type {
+                if let Some(error) = error_body.get_mut("error").and_then(Value::as_object_mut) {
+                    error.insert(
+                        "type".to_string(),
+                        Value::String("stream_terminal_error".to_string()),
+                    );
+                }
+            }
+            return Some(error_body);
+        }
         current_event_type = None;
     }
     None
@@ -1179,6 +1201,53 @@ data: {"message":"bad"}
         assert!(stream_body_contains_error_event(body));
     }
 
+    #[test]
+    fn extracts_openai_response_failed_error_without_losing_provider_fields() {
+        let body = br#"event: response.failed
+data: {"type":"response.failed","response":{"status":"failed","error":{"type":"invalid_request","message":"cyber policy rejected the request","code":"cyber_policy_violation","param":"input","details":{"policy_category":"cybersecurity","appeal_allowed":true}}}}
+
+"#;
+
+        let error_body = extract_provider_private_stream_error_body(None, body)
+            .expect("response.failed should expose its provider error body");
+
+        assert_eq!(
+            error_body,
+            json!({
+                "error": {
+                    "type": "invalid_request",
+                    "message": "cyber policy rejected the request",
+                    "code": "cyber_policy_violation",
+                    "param": "input",
+                    "details": {
+                        "policy_category": "cybersecurity",
+                        "appeal_allowed": true
+                    }
+                }
+            })
+        );
+    }
+
+    #[test]
+    fn response_failed_without_provider_type_uses_stream_terminal_error_category() {
+        let body = br#"event: response.failed
+data: {"type":"response.failed","response":{"status":"failed","error":{"message":"cyber policy rejected the request","code":"cyber_policy"}}}
+
+"#;
+
+        let error_body = extract_provider_private_stream_error_body(None, body)
+            .expect("response.failed should expose its provider error body");
+
+        assert_eq!(
+            error_body.pointer("/error/type"),
+            Some(&json!("stream_terminal_error"))
+        );
+        assert_eq!(
+            error_body.pointer("/error/code"),
+            Some(&json!("cyber_policy"))
+        );
+    }
+
     #[test]
     fn windsurf_sync_error_message_is_not_normalized_as_success() {
         let report_context = json!({

crates/aether-billing/src/event_enrichment.rs

@@ -1,9 +1,7 @@
 use aether_data_contracts::repository::billing::StoredBillingModelContext;
 use aether_data_contracts::repository::usage::{
-    extract_provider_actual_service_tier_from_response,
-    extract_provider_cache_ttl_minutes_from_metadata, extract_provider_service_tier_from_body,
-    normalize_provider_service_tier, resolve_provider_cache_ttl_minutes,
-    PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY, PROVIDER_SERVICE_TIER_METADATA_KEY,
+    extract_provider_cache_ttl_minutes_from_metadata, resolve_provider_cache_ttl_minutes,
+    resolve_provider_service_tier_from_request_capture,
 };
 use aether_data_contracts::DataLayerError;
 use aether_usage_runtime::{UsageEvent, UsageEventType};
@@ -160,7 +158,9 @@ fn calculate_billing_computation(
             .clone()
             .or_else(|| event.data.api_format.clone()),
         requested_processing_tier: processing_tiers.requested,
-        actual_processing_tier: processing_tiers.actual,
+        // The response-reported tier remains usage audit data, but it is not authoritative for
+        // pricing. Settlement follows the final request that was sent upstream.
+        actual_processing_tier: None,
         request_count,
         input_tokens: event.data.input_tokens.unwrap_or_default() as i64,
         output_tokens: event.data.output_tokens.unwrap_or_default() as i64,
@@ -192,29 +192,18 @@ fn calculate_billing_computation(
 #[derive(Debug, Clone, PartialEq, Eq)]
 struct UsageEventProcessingTiers {
     requested: Option<String>,
-    actual: Option<String>,
 }
 
 fn usage_event_processing_tiers(
     data: &aether_usage_runtime::UsageEventData,
 ) -> UsageEventProcessingTiers {
-    let metadata = data.request_metadata.as_ref().and_then(Value::as_object);
-    let requested = extract_provider_service_tier_from_body(data.provider_request_body.as_ref())
-        .or_else(|| {
-            metadata
-                .and_then(|metadata| metadata.get(PROVIDER_SERVICE_TIER_METADATA_KEY))
-                .and_then(Value::as_str)
-                .and_then(normalize_provider_service_tier)
-        });
-    let actual = metadata
-        .and_then(|metadata| metadata.get(PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY))
-        .and_then(Value::as_str)
-        .and_then(normalize_provider_service_tier)
-        .or_else(|| {
-            extract_provider_actual_service_tier_from_response(data.response_body.as_ref())
-        });
+    let requested = resolve_provider_service_tier_from_request_capture(
+        data.provider_request_body.as_ref(),
+        data.provider_request_body_state,
+        data.request_metadata.as_ref(),
+    );
 
-    UsageEventProcessingTiers { requested, actual }
+    UsageEventProcessingTiers { requested }
 }
 
 fn usage_event_provider_cache_ttl_minutes(
@@ -406,6 +395,7 @@ fn build_settlement_snapshot(
 #[cfg(test)]
 mod tests {
     use aether_data_contracts::repository::billing::StoredBillingModelContext;
+    use aether_data_contracts::repository::usage::UsageBodyCaptureState;
     use aether_usage_runtime::{UsageEvent, UsageEventData, UsageEventType};
     use async_trait::async_trait;
     use serde_json::json;
@@ -444,7 +434,7 @@ mod tests {
     }
 
     #[test]
-    fn processing_tier_facts_keep_request_and_terminal_response_independent() {
+    fn processing_tier_fact_uses_final_provider_request_body() {
         let data = UsageEventData {
             provider_request_body: Some(json!({"service_tier": "Priority"})),
             response_body: Some(json!({"service_tier": "priority"})),
@@ -458,7 +448,6 @@ mod tests {
         let tiers = usage_event_processing_tiers(&data);
 
         assert_eq!(tiers.requested.as_deref(), Some("priority"));
-        assert_eq!(tiers.actual.as_deref(), Some("default"));
     }
 
     #[test]
@@ -474,7 +463,83 @@ mod tests {
         let tiers = usage_event_processing_tiers(&data);
 
         assert_eq!(tiers.requested.as_deref(), Some("fast"));
-        assert_eq!(tiers.actual.as_deref(), Some("fast"));
+    }
+
+    #[test]
+    fn processing_tier_does_not_fall_back_to_stale_metadata_when_body_is_present() {
+        let data = UsageEventData {
+            provider_request_body: Some(json!({"model": "gpt-5"})),
+            response_body: Some(json!({"service_tier": "priority"})),
+            request_metadata: Some(json!({
+                "provider_service_tier": "priority",
+                "provider_actual_service_tier": "priority"
+            })),
+            ..UsageEventData::default()
+        };
+
+        let tiers = usage_event_processing_tiers(&data);
+
+        assert_eq!(tiers.requested, None);
+    }
+
+    #[test]
+    fn processing_tier_uses_request_derived_metadata_when_body_capture_was_disabled() {
+        let data = UsageEventData {
+            provider_request_body: None,
+            provider_request_body_state: Some(UsageBodyCaptureState::Disabled),
+            response_body: Some(json!({"service_tier": "flex"})),
+            request_metadata: Some(json!({
+                "provider_service_tier": "priority",
+                "provider_actual_service_tier": "flex"
+            })),
+            ..UsageEventData::default()
+        };
+
+        let tiers = usage_event_processing_tiers(&data);
+
+        assert_eq!(tiers.requested.as_deref(), Some("priority"));
+    }
+
+    #[test]
+    fn processing_tier_does_not_use_metadata_or_response_when_final_request_body_is_missing() {
+        let data = UsageEventData {
+            provider_request_body: None,
+            provider_request_body_state: Some(UsageBodyCaptureState::None),
+            response_body: Some(json!({"service_tier": "priority"})),
+            request_metadata: Some(json!({
+                "provider_service_tier": "priority",
+                "provider_actual_service_tier": "priority"
+            })),
+            ..UsageEventData::default()
+        };
+
+        let tiers = usage_event_processing_tiers(&data);
+
+        assert_eq!(tiers.requested, None);
+    }
+
+    #[test]
+    fn processing_tier_uses_request_derived_metadata_when_body_capture_was_truncated() {
+        let data = UsageEventData {
+            provider_request_body: Some(json!({
+                "truncated": true,
+                "reason": "body_capture_limit_exceeded",
+                "max_bytes": 128,
+                "source_bytes": 4096,
+                "value_kind": "object"
+            })),
+            provider_request_body_state: Some(UsageBodyCaptureState::Truncated),
+            response_body: Some(json!({"service_tier": "default"})),
+            request_metadata: Some(json!({
+                "provider_service_tier": "priority",
+                "provider_actual_service_tier": "default"
+            })),
+            ..UsageEventData::default()
+        };
+
+        let tiers = usage_event_processing_tiers(&data);
+
+        assert_eq!(tiers.requested.as_deref(), Some("priority"));
     }
 
     #[tokio::test]
@@ -638,7 +703,7 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn settlement_uses_actual_processing_tier_catalog_and_source() {
+    async fn settlement_uses_requested_processing_tier_catalog_and_ignores_response_tier() {
         let lookup = TestLookup {
             name_context: Some(
                 StoredBillingModelContext::new(
@@ -706,13 +771,28 @@ mod tests {
             .and_then(|value| value.pointer("/settlement_snapshot/pricing_snapshot"))
             .expect("settlement pricing snapshot should exist");
         assert_eq!(pricing_snapshot["requested_processing_tier"], "priority");
-        assert_eq!(pricing_snapshot["actual_processing_tier"], "flex");
-        assert_eq!(pricing_snapshot["billing_processing_tier"], "flex");
-        assert_eq!(pricing_snapshot["tiered_pricing_source"], "global_default");
-        assert_eq!(pricing_snapshot["processing_tier_price_multiplier"], 0.5);
+        assert!(pricing_snapshot["actual_processing_tier"].is_null());
+        assert_eq!(pricing_snapshot["billing_processing_tier"], "priority");
+        assert_eq!(
+            pricing_snapshot["tiered_pricing_source"],
+            "provider_override"
+        );
+        assert_eq!(
+            pricing_snapshot["processing_tier_price_multiplier"],
+            Value::Null
+        );
         assert_eq!(
             pricing_snapshot["tiered_pricing"]["tiers"][0]["input_price_per_1m"],
-            2.5
+            10.0
+        );
+        // The response fact remains available for audit, but does not influence settlement.
+        assert_eq!(
+            event
+                .data
+                .request_metadata
+                .as_ref()
+                .and_then(|value| value.get("provider_actual_service_tier")),
+            Some(&json!("flex"))
         );
         assert_eq!(
             event
@@ -720,12 +800,12 @@ mod tests {
                 .request_metadata
                 .as_ref()
                 .and_then(|value| { value.pointer("/billing_dimensions/actual_processing_tier") }),
-            Some(&json!("flex"))
+            Some(&Value::Null)
         );
     }
 
     #[tokio::test]
-    async fn actual_processing_catalog_controls_image_price_with_independent_fixed_price() {
+    async fn requested_processing_catalog_controls_image_price_with_independent_fixed_price() {
         let lookup = TestLookup {
             name_context: Some(
                 StoredBillingModelContext::new(
@@ -783,27 +863,27 @@ mod tests {
             .await
             .expect("billing should succeed");
 
-        assert_eq!(event.data.total_cost_usd, Some(0.44));
-        assert_eq!(event.data.actual_total_cost_usd, Some(0.44));
+        assert_eq!(event.data.total_cost_usd, Some(0.84));
+        assert_eq!(event.data.actual_total_cost_usd, Some(0.84));
         let metadata = event.data.request_metadata.as_ref().expect("metadata");
         let pricing = metadata
             .pointer("/settlement_snapshot/pricing_snapshot")
             .expect("pricing snapshot");
-        assert_eq!(pricing["billing_processing_tier"], "flex");
-        assert_eq!(pricing["tiered_pricing_source"], "global_default");
+        assert_eq!(pricing["billing_processing_tier"], "priority");
+        assert_eq!(pricing["tiered_pricing_source"], "provider_override");
         assert_eq!(pricing["price_per_request_source"], "provider_override");
-        assert_eq!(pricing["pricing_source"], "mixed");
+        assert_eq!(pricing["pricing_source"], "provider_override");
         assert_eq!(
             metadata
                 .pointer("/billing_snapshot/resolved_variables/image_output_price_per_image")
                 .and_then(Value::as_f64),
-            Some(0.2)
+            Some(0.4)
         );
         assert_eq!(
             metadata
                 .pointer("/billing_snapshot/cost_breakdown/image_output_cost")
                 .and_then(Value::as_f64),
-            Some(0.4)
+            Some(0.8)
         );
         assert_eq!(
             metadata

crates/aether-billing/src/pricing.rs

@@ -1,5 +1,3 @@
-use std::collections::BTreeSet;
-
 use aether_data_contracts::repository::{
     billing::StoredBillingModelContext,
     global_models::{explicit_pricing_catalog_state, ExplicitPricingCatalogState},
@@ -108,17 +106,12 @@ impl BillingModelPricingSnapshot {
     ) -> BillingPricingResolution {
         let requested_processing_tier = normalize_processing_tier(requested_processing_tier);
         let actual_processing_tier = normalize_processing_tier(actual_processing_tier);
-        let billing_processing_tier = actual_processing_tier
+        // The provider-reported actual tier is retained in the resolution for audit only. The
+        // final upstream request is the sole authority for selecting a billing catalog.
+        let billing_processing_tier = requested_processing_tier
             .as_deref()
             .map(canonical_processing_tier)
-            .or_else(|| {
-                requested_processing_tier.as_deref().map_or(
-                    Some("standard".to_string()),
-                    |requested| {
-                        processing_tier_is_standard(requested).then(|| "standard".to_string())
-                    },
-                )
-            });
+            .or_else(|| Some("standard".to_string()));
 
         let (tiered_pricing, tiered_pricing_source, processing_tier_price_multiplier) =
             billing_processing_tier
@@ -151,17 +144,12 @@ impl BillingModelPricingSnapshot {
         self.validate_processing_tier_containers()?;
         let requested_processing_tier = normalize_processing_tier(requested_processing_tier);
         let actual_processing_tier = normalize_processing_tier(actual_processing_tier);
-        let billing_processing_tier = actual_processing_tier
+        // Keep this checked path aligned with `resolve_pricing`: response facts must never choose
+        // the catalog used for settlement.
+        let billing_processing_tier = requested_processing_tier
             .as_deref()
             .map(canonical_processing_tier)
-            .or_else(|| {
-                requested_processing_tier.as_deref().map_or(
-                    Some("standard".to_string()),
-                    |requested| {
-                        processing_tier_is_standard(requested).then(|| "standard".to_string())
-                    },
-                )
-            });
+            .or_else(|| Some("standard".to_string()));
 
         let (tiered_pricing, tiered_pricing_source, processing_tier_price_multiplier) =
             match billing_processing_tier.as_deref() {
@@ -208,34 +196,7 @@ impl BillingModelPricingSnapshot {
             return Ok(None);
         }
 
-        let mut billing_tiers = BTreeSet::from(["standard".to_string(), requested_billing_tier]);
-        for pricing in [
-            self.model_tiered_pricing.as_ref(),
-            self.default_tiered_pricing.as_ref(),
-        ]
-        .into_iter()
-        .flatten()
-        {
-            let Some(processing_tiers) = pricing.get("processing_tiers").and_then(Value::as_object)
-            else {
-                continue;
-            };
-            billing_tiers.extend(processing_tiers.keys().filter_map(|tier| {
-                normalize_processing_tier(Some(tier)).map(|tier| canonical_processing_tier(&tier))
-            }));
-        }
-
-        let mut candidates = Vec::new();
-        for billing_tier in billing_tiers {
-            let resolution = self.authorization_pricing_for_tier(
-                requested_processing_tier.clone(),
-                Some(billing_tier),
-            )?;
-            if resolution.bills_standard_processing_tier() || resolution.tiered_pricing.is_some() {
-                candidates.push(resolution);
-            }
-        }
-        Ok((!candidates.is_empty()).then_some(candidates))
+        Ok(Some(vec![requested_resolution]))
     }
 
     pub fn validate_authorization_pricing_configuration(
@@ -839,7 +800,7 @@ mod tests {
     }
 
     #[test]
-    fn explicit_nonstandard_request_requires_actual_tier() {
+    fn explicit_nonstandard_request_selects_requested_catalog_without_actual_tier() {
         let pricing = snapshot(
             None,
             Some(json!({
@@ -852,17 +813,27 @@ mod tests {
 
         let resolution = pricing.resolve_pricing(Some("Priority"), None);
 
-        assert!(resolution.requires_actual_processing_tier());
+        assert!(!resolution.requires_actual_processing_tier());
         assert_eq!(
             resolution.requested_processing_tier.as_deref(),
             Some("priority")
         );
-        assert_eq!(resolution.billing_processing_tier, None);
-        assert_eq!(resolution.tiered_pricing, None);
+        assert_eq!(
+            resolution.billing_processing_tier.as_deref(),
+            Some("priority")
+        );
+        assert_eq!(
+            resolution
+                .tiered_pricing
+                .as_ref()
+                .and_then(|value| value.pointer("/tiers/0/input_price_per_1m"))
+                .and_then(serde_json::Value::as_f64),
+            Some(6.0)
+        );
     }
 
     #[test]
-    fn actual_tier_selects_exact_catalog_and_source() {
+    fn response_actual_tier_is_audited_but_does_not_select_pricing_catalog() {
         let pricing = snapshot(
             Some(json!({
                 "processing_tiers": {
@@ -878,28 +849,29 @@ mod tests {
         );
 
         let flex = pricing.resolve_pricing(Some("priority"), Some("flex"));
-        assert_eq!(flex.billing_processing_tier.as_deref(), Some("flex"));
+        assert_eq!(flex.actual_processing_tier.as_deref(), Some("flex"));
+        assert_eq!(flex.billing_processing_tier.as_deref(), Some("priority"));
         assert_eq!(
             flex.tiered_pricing_source,
-            Some(BillingPricingSource::GlobalDefault)
+            Some(BillingPricingSource::ProviderOverride)
         );
         assert_eq!(
             flex.tiered_pricing
                 .as_ref()
                 .and_then(|value| value.pointer("/tiers/0/input_price_per_1m"))
                 .and_then(serde_json::Value::as_f64),
-            Some(1.5)
+            Some(9.0)
         );
 
         let standard = pricing.resolve_pricing(Some("priority"), Some("Default"));
         assert_eq!(standard.actual_processing_tier.as_deref(), Some("default"));
         assert_eq!(
             standard.billing_processing_tier.as_deref(),
-            Some("standard")
+            Some("priority")
         );
         assert_eq!(
             standard.tiered_pricing_source,
-            Some(BillingPricingSource::GlobalDefault)
+            Some(BillingPricingSource::ProviderOverride)
         );
     }
 
@@ -962,9 +934,7 @@ mod tests {
                 .and_then(serde_json::Value::as_f64),
             Some(6.0)
         );
-        assert!(candidates.iter().any(|resolution| {
-            resolution.billing_processing_tier.as_deref() == Some("standard")
-        }));
+        assert_eq!(candidates.len(), 1);
     }
 
     #[test]
@@ -1222,7 +1192,7 @@ mod tests {
     }
 
     #[test]
-    fn actual_claude_fast_uses_the_exact_provider_multiplier_overlay() {
+    fn requested_claude_fast_uses_the_exact_provider_multiplier_overlay() {
         let pricing = snapshot(
             Some(json!({
                 "tiers": [{"up_to": null, "input_price_per_1m": 2.0}],
@@ -1234,8 +1204,9 @@ mod tests {
             })),
         );
 
-        let resolved = pricing.resolve_pricing(Some("priority"), Some("fast"));
+        let resolved = pricing.resolve_pricing(Some("fast"), Some("priority"));
 
+        assert_eq!(resolved.actual_processing_tier.as_deref(), Some("priority"));
         assert_eq!(resolved.billing_processing_tier.as_deref(), Some("fast"));
         assert_eq!(
             resolved

crates/aether-billing/src/service.rs

@@ -1104,7 +1104,7 @@ mod tests {
     }
 
     #[test]
-    fn nonstandard_request_without_actual_tier_fails_closed() {
+    fn nonstandard_request_without_actual_tier_uses_requested_catalog() {
         let result = BillingService::new()
             .calculate(
                 &processing_pricing(),
@@ -1112,57 +1112,49 @@ mod tests {
             )
             .expect("billing should calculate");
 
-        assert_eq!(result.cost_result.status, BillingSnapshotStatus::NoRule);
-        assert_eq!(
-            result.cost_result.snapshot.missing_required,
-            vec!["actual_processing_tier"]
-        );
+        assert_eq!(result.cost_result.status, BillingSnapshotStatus::Complete);
         assert_eq!(
             result.cost_result.snapshot.resolved_dimensions["billing_processing_tier"],
-            json!(null)
+            json!("priority")
+        );
+        assert_eq!(
+            result.cost_result.snapshot.resolved_variables["input_price_per_1m"],
+            json!(10.0)
         );
     }
 
     #[test]
-    fn actual_tier_controls_standard_flex_and_priority_catalogs() {
-        let cases = [
-            (
-                "default",
-                100,
-                5.0,
-                6.25,
-                BillingPricingSource::GlobalDefault,
-            ),
-            ("flex", 100, 2.5, 3.125, BillingPricingSource::GlobalDefault),
-            (
-                "priority",
-                100,
-                10.0,
-                12.5,
-                BillingPricingSource::ProviderOverride,
-            ),
-        ];
+    fn response_actual_tier_does_not_override_requested_catalog() {
+        let cases = ["default", "flex", "priority"];
 
-        for (actual, input_tokens, input_price, cache_write_price, source) in cases {
+        for actual in cases {
             let result = BillingService::new()
                 .calculate(
                     &processing_pricing(),
-                    &processing_usage(Some("priority"), Some(actual), input_tokens),
+                    &processing_usage(Some("priority"), Some(actual), 100),
                 )
                 .expect("processing tier should resolve");
 
             assert_eq!(result.cost_result.status, BillingSnapshotStatus::Complete);
+            assert_eq!(
+                result.pricing_resolution.actual_processing_tier.as_deref(),
+                Some(actual)
+            );
+            assert_eq!(
+                result.pricing_resolution.billing_processing_tier.as_deref(),
+                Some("priority")
+            );
             assert_eq!(
                 result.cost_result.snapshot.resolved_variables["input_price_per_1m"],
-                json!(input_price)
+                json!(10.0)
             );
             assert_eq!(
                 result.cost_result.snapshot.resolved_variables["cache_creation_price_per_1m"],
-                json!(cache_write_price)
+                json!(12.5)
             );
             assert_eq!(
                 result.pricing_resolution.tiered_pricing_source,
-                Some(source)
+                Some(BillingPricingSource::ProviderOverride)
             );
         }
     }
@@ -1373,7 +1365,7 @@ mod tests {
     }
 
     #[test]
-    fn finite_processing_catalog_and_unknown_actual_tier_fail_closed() {
+    fn finite_processing_catalog_fails_only_on_requested_catalog_bounds() {
         let priority = BillingService::new()
             .calculate(
                 &processing_pricing(),
@@ -1386,16 +1378,22 @@ mod tests {
             vec!["input_context_tier"]
         );
 
-        let unknown = BillingService::new()
+        let conflicting_actual = BillingService::new()
             .calculate(
                 &processing_pricing(),
                 &processing_usage(Some("priority"), Some("expedited"), 100),
             )
             .expect("billing should calculate");
-        assert_eq!(unknown.cost_result.status, BillingSnapshotStatus::NoRule);
         assert_eq!(
-            unknown.cost_result.snapshot.missing_required,
-            vec!["processing_tier_catalog"]
+            conflicting_actual.cost_result.status,
+            BillingSnapshotStatus::Complete
+        );
+        assert_eq!(
+            conflicting_actual
+                .pricing_resolution
+                .billing_processing_tier
+                .as_deref(),
+            Some("priority")
         );
     }
 
@@ -1502,13 +1500,17 @@ mod tests {
         estimate.max_output_tokens = Some(0);
         estimate.cache_ttl_minutes = Some(30);
 
-        for requested_processing_tier in [None, Some("standard"), Some("flex")] {
+        for (requested_processing_tier, expected) in [
+            (None, 3.75),
+            (Some("standard"), 3.75),
+            (Some("flex"), 1.875),
+        ] {
             estimate.requested_processing_tier = requested_processing_tier.map(ToOwned::to_owned);
             assert_eq!(
                 service
                     .estimate_authorization_cost_upper_bound(&processing_pricing(), &estimate)
                     .expect("eligible processing catalogs should calculate"),
-                Some(3.75),
+                Some(expected),
                 "requested tier: {requested_processing_tier:?}"
             );
         }
@@ -1523,7 +1525,7 @@ mod tests {
     }
 
     #[test]
-    fn unknown_actual_tier_cannot_fall_back_to_fixed_request_price() {
+    fn unknown_actual_tier_does_not_override_requested_tier_or_fixed_price() {
         let pricing = BillingModelPricingSnapshot {
             default_price_per_request: Some(0.02),
             ..processing_pricing()
@@ -1535,16 +1537,20 @@ mod tests {
             )
             .expect("billing should calculate");
 
-        assert_eq!(result.cost_result.status, BillingSnapshotStatus::NoRule);
+        assert_eq!(result.cost_result.status, BillingSnapshotStatus::Complete);
         assert_eq!(
-            result.cost_result.snapshot.missing_required,
-            vec!["processing_tier_catalog"]
+            result.pricing_resolution.billing_processing_tier.as_deref(),
+            Some("priority")
+        );
+        assert_eq!(
+            result.pricing_resolution.actual_processing_tier.as_deref(),
+            Some("expedited")
         );
         assert_eq!(result.pricing_resolution.price_per_request, Some(0.02));
     }
 
     #[test]
-    fn authorization_estimate_bounds_requested_and_provider_actual_catalogs() {
+    fn authorization_estimate_bounds_only_the_requested_catalog() {
         let service = BillingService::new();
         let mut estimate = BillingAuthorizationEstimateInput::new("chat", 100_000);
         estimate.api_format = Some("openai:responses".to_string());
@@ -1563,8 +1569,8 @@ mod tests {
             .expect("flex estimate should be bounded");
 
         assert_eq!(priority, 61.25);
-        assert_eq!(flex, 61.25);
-        assert_eq!(priority, flex);
+        assert_eq!(flex, 15.3125);
+        assert!(priority > flex);
     }
 
     #[test]

crates/aether-data/adapters/mysql/src/usage.rs

@@ -135,22 +135,22 @@ INSERT INTO `usage` (
 ON DUPLICATE KEY UPDATE
   user_id = VALUES(user_id),
   api_key_id = VALUES(api_key_id),
-  provider_name = VALUES(provider_name),
-  model = VALUES(model),
-  target_model = VALUES(target_model),
-  provider_id = VALUES(provider_id),
-  provider_endpoint_id = VALUES(provider_endpoint_id),
-  provider_api_key_id = VALUES(provider_api_key_id),
-  request_type = VALUES(request_type),
-  api_format = VALUES(api_format),
-  api_family = VALUES(api_family),
-  endpoint_kind = VALUES(endpoint_kind),
-  endpoint_api_format = VALUES(endpoint_api_format),
-  provider_api_family = VALUES(provider_api_family),
-  provider_endpoint_kind = VALUES(provider_endpoint_kind),
-  has_format_conversion = VALUES(has_format_conversion),
-  is_stream = VALUES(is_stream),
-  upstream_is_stream = VALUES(upstream_is_stream),
+  provider_name = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN provider_name ELSE VALUES(provider_name) END,
+  model = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN model ELSE VALUES(model) END,
+  target_model = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN target_model ELSE VALUES(target_model) END,
+  provider_id = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN provider_id ELSE VALUES(provider_id) END,
+  provider_endpoint_id = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN provider_endpoint_id ELSE VALUES(provider_endpoint_id) END,
+  provider_api_key_id = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN provider_api_key_id ELSE VALUES(provider_api_key_id) END,
+  request_type = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN request_type ELSE VALUES(request_type) END,
+  api_format = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN api_format ELSE VALUES(api_format) END,
+  api_family = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN api_family ELSE VALUES(api_family) END,
+  endpoint_kind = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN endpoint_kind ELSE VALUES(endpoint_kind) END,
+  endpoint_api_format = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN endpoint_api_format ELSE VALUES(endpoint_api_format) END,
+  provider_api_family = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN provider_api_family ELSE VALUES(provider_api_family) END,
+  provider_endpoint_kind = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN provider_endpoint_kind ELSE VALUES(provider_endpoint_kind) END,
+  has_format_conversion = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN has_format_conversion ELSE VALUES(has_format_conversion) END,
+  is_stream = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN is_stream ELSE VALUES(is_stream) END,
+  upstream_is_stream = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN upstream_is_stream ELSE VALUES(upstream_is_stream) END,
   input_tokens = CASE
     WHEN status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming') THEN input_tokens
     ELSE VALUES(input_tokens)
@@ -231,15 +231,15 @@ ON DUPLICATE KEY UPDATE
     WHEN status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming') THEN billing_status
     ELSE VALUES(billing_status)
   END,
-  request_metadata = VALUES(request_metadata),
-  candidate_id = VALUES(candidate_id),
-  candidate_index = VALUES(candidate_index),
-  key_name = VALUES(key_name),
-  planner_kind = VALUES(planner_kind),
-  route_family = VALUES(route_family),
-  route_kind = VALUES(route_kind),
-  execution_path = VALUES(execution_path),
-  local_execution_runtime_miss_reason = VALUES(local_execution_runtime_miss_reason),
+  request_metadata = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN request_metadata ELSE VALUES(request_metadata) END,
+  candidate_id = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN candidate_id ELSE VALUES(candidate_id) END,
+  candidate_index = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN candidate_index ELSE VALUES(candidate_index) END,
+  key_name = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN key_name ELSE VALUES(key_name) END,
+  planner_kind = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN planner_kind ELSE VALUES(planner_kind) END,
+  route_family = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN route_family ELSE VALUES(route_family) END,
+  route_kind = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN route_kind ELSE VALUES(route_kind) END,
+  execution_path = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN execution_path ELSE VALUES(execution_path) END,
+  local_execution_runtime_miss_reason = CASE WHEN (status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming')) OR (status = 'streaming' AND VALUES(status) = 'pending') THEN local_execution_runtime_miss_reason ELSE VALUES(local_execution_runtime_miss_reason) END,
   finalized_at = CASE
     WHEN status IN ('completed', 'failed', 'cancelled') AND VALUES(status) IN ('pending', 'streaming') THEN finalized_at
     ELSE VALUES(finalized_at)

crates/aether-data/adapters/mysql/src/usage/tests.rs

@@ -65,6 +65,50 @@ fn mysql_usage_upsert_keeps_terminal_state_when_streaming_arrives_late() {
     ));
 }
 
+#[test]
+fn mysql_usage_upsert_guards_candidate_identity_metadata_and_routing_from_late_lifecycle() {
+    for field in [
+        "provider_name",
+        "model",
+        "target_model",
+        "provider_id",
+        "provider_endpoint_id",
+        "provider_api_key_id",
+        "request_type",
+        "api_format",
+        "api_family",
+        "endpoint_kind",
+        "endpoint_api_format",
+        "provider_api_family",
+        "provider_endpoint_kind",
+        "has_format_conversion",
+        "is_stream",
+        "upstream_is_stream",
+        "request_metadata",
+        "candidate_id",
+        "candidate_index",
+        "key_name",
+        "planner_kind",
+        "route_family",
+        "route_kind",
+        "execution_path",
+        "local_execution_runtime_miss_reason",
+    ] {
+        let assignment = format!("{field} = CASE WHEN (");
+        assert!(
+            super::UPSERT_USAGE_SQL.contains(&assignment),
+            "missing lifecycle guard for {field}"
+        );
+        let preserve = format!("THEN {field} ELSE VALUES({field}) END");
+        assert!(
+            super::UPSERT_USAGE_SQL.contains(&preserve),
+            "late lifecycle must preserve {field}"
+        );
+    }
+    assert!(super::UPSERT_USAGE_SQL
+        .contains("OR (status = 'streaming' AND VALUES(status) = 'pending')"));
+}
+
 #[tokio::test]
 async fn mysql_usage_write_repository_upserts_when_url_is_set() {
     let Some(database_url) = std::env::var("AETHER_TEST_MYSQL_URL")

crates/aether-data/adapters/postgres/src/usage/mod.rs

@@ -49,7 +49,9 @@ use aether_data_contracts::repository::usage::{
     StoredProviderUsageSummary, StoredRequestUsageAudit, StoredUsageDailySummary,
     UpsertUsageRecord, UsageAuditListQuery, UsageCounterFlushSummary, UsageCounterHealthSnapshot,
     UsageCounterPendingHealthSnapshot, UsageDailyHeatmapQuery, UsageReadRepository,
-    UsageWriteRepository,
+    UsageWriteRepository, PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY,
+    PROVIDER_REASONING_EFFORT_METADATA_KEY, PROVIDER_SERVICE_TIER_METADATA_KEY,
+    REQUESTED_REASONING_EFFORT_METADATA_KEY,
 };
 use aether_data_contracts::DataLayerError;
 
@@ -7994,8 +7996,14 @@ ORDER BY "usage".user_id ASC
                     http_audit_capture_mode,
                     routing_snapshot,
                     settlement_pricing_snapshot,
-                    request_metadata_value,
-                    request_metadata_json,
+                    mut request_metadata_value,
+                    mut request_metadata_json,
+                    replace_client_request_body_facts,
+                    replace_provider_request_body_facts,
+                    clear_request_body,
+                    clear_provider_request_body,
+                    clear_response_body,
+                    clear_client_response_body,
                 } = prepared;
                 Box::pin(async move {
                     lock_usage_request_id_in_tx(tx, &usage.request_id).await?;
@@ -8018,6 +8026,37 @@ ORDER BY "usage".user_id ASC
 
                     let previous_usage =
                         find_usage_by_request_id_in_tx(tx, &usage.request_id).await?;
+                    let capture_update_allowed = usage_capture_update_allowed(
+                        previous_usage
+                            .as_ref()
+                            .map(|stored| (stored.status.as_str(), stored.billing_status.as_str())),
+                        usage.status.as_str(),
+                    );
+                    let replace_terminal_snapshots =
+                        matches!(usage.status.as_str(), "completed" | "failed" | "cancelled");
+                    if capture_update_allowed
+                        && request_metadata_value.is_none()
+                        && (replace_terminal_snapshots
+                            || replace_client_request_body_facts
+                            || replace_provider_request_body_facts)
+                    {
+                        let previous_metadata = previous_usage
+                            .as_ref()
+                            .and_then(|stored| stored.request_metadata.as_ref());
+                        request_metadata_value = Some(if replace_terminal_snapshots {
+                            retain_previous_request_audit_metadata(
+                                previous_metadata,
+                                !replace_client_request_body_facts,
+                            )
+                        } else {
+                            clear_previous_request_body_facts(
+                                previous_metadata,
+                                replace_client_request_body_facts,
+                                replace_provider_request_body_facts,
+                            )
+                        });
+                        request_metadata_json = json_bind_text(request_metadata_value.as_ref())?;
+                    }
                     let _row = sqlx::query(UPSERT_SQL)
                         .bind(Uuid::new_v4().to_string())
                         .bind(&usage.request_id)
@@ -8090,72 +8129,89 @@ ORDER BY "usage".user_id ASC
                                 usage.updated_at_unix_secs
                             ))
                         })?)
-                        .bind(request_body_storage.has_detached_blob())
-                        .bind(provider_request_body_storage.has_detached_blob())
-                        .bind(response_body_storage.has_detached_blob())
-                        .bind(client_response_body_storage.has_detached_blob())
+                        .bind(request_body_storage.has_detached_blob() || clear_request_body)
+                        .bind(
+                            provider_request_body_storage.has_detached_blob()
+                                || clear_provider_request_body,
+                        )
+                        .bind(response_body_storage.has_detached_blob() || clear_response_body)
+                        .bind(
+                            client_response_body_storage.has_detached_blob()
+                                || clear_client_response_body,
+                        )
+                        .bind(capture_update_allowed)
                         .fetch_one(&mut **tx)
                         .await
                         .map_postgres_err()?;
-                    sync_usage_body_blob_storage(
-                        &mut **tx,
-                        &usage.request_id,
-                        UsageBodyField::RequestBody,
-                        usage.request_body.as_ref(),
-                        &request_body_storage,
-                    )
-                    .await?;
-                    sync_usage_body_blob_storage(
-                        &mut **tx,
-                        &usage.request_id,
-                        UsageBodyField::ProviderRequestBody,
-                        usage.provider_request_body.as_ref(),
-                        &provider_request_body_storage,
-                    )
-                    .await?;
-                    sync_usage_body_blob_storage(
-                        &mut **tx,
-                        &usage.request_id,
-                        UsageBodyField::ResponseBody,
-                        usage.response_body.as_ref(),
-                        &response_body_storage,
-                    )
-                    .await?;
-                    sync_usage_body_blob_storage(
-                        &mut **tx,
-                        &usage.request_id,
-                        UsageBodyField::ClientResponseBody,
-                        usage.client_response_body.as_ref(),
-                        &client_response_body_storage,
-                    )
-                    .await?;
-                    let http_audit_headers = UsageHttpAuditHeaders {
-                        request_headers_json: request_headers_json.as_deref(),
-                        provider_request_headers_json: provider_request_headers_json.as_deref(),
-                        response_headers_json: response_headers_json.as_deref(),
-                        client_response_headers_json: client_response_headers_json.as_deref(),
-                    };
-                    sync_usage_http_audit_storage(
-                        &mut **tx,
-                        &usage.request_id,
-                        &http_audit_headers,
-                        &http_audit_refs,
-                        &http_audit_states,
-                        http_audit_capture_mode,
-                    )
-                    .await?;
-                    sync_usage_routing_snapshot_storage(
-                        &mut **tx,
-                        &usage.request_id,
-                        &routing_snapshot,
-                    )
-                    .await?;
-                    sync_usage_settlement_pricing_snapshot_storage(
-                        &mut **tx,
-                        &usage.request_id,
-                        &settlement_pricing_snapshot,
-                    )
-                    .await?;
+                    if capture_update_allowed {
+                        sync_usage_body_blob_storage(
+                            &mut **tx,
+                            &usage.request_id,
+                            UsageBodyField::RequestBody,
+                            usage.request_body.as_ref(),
+                            &request_body_storage,
+                            clear_request_body,
+                        )
+                        .await?;
+                        sync_usage_body_blob_storage(
+                            &mut **tx,
+                            &usage.request_id,
+                            UsageBodyField::ProviderRequestBody,
+                            usage.provider_request_body.as_ref(),
+                            &provider_request_body_storage,
+                            clear_provider_request_body,
+                        )
+                        .await?;
+                        sync_usage_body_blob_storage(
+                            &mut **tx,
+                            &usage.request_id,
+                            UsageBodyField::ResponseBody,
+                            usage.response_body.as_ref(),
+                            &response_body_storage,
+                            clear_response_body,
+                        )
+                        .await?;
+                        sync_usage_body_blob_storage(
+                            &mut **tx,
+                            &usage.request_id,
+                            UsageBodyField::ClientResponseBody,
+                            usage.client_response_body.as_ref(),
+                            &client_response_body_storage,
+                            clear_client_response_body,
+                        )
+                        .await?;
+                        let http_audit_headers = UsageHttpAuditHeaders {
+                            request_headers_json: request_headers_json.as_deref(),
+                            provider_request_headers_json: provider_request_headers_json.as_deref(),
+                            response_headers_json: response_headers_json.as_deref(),
+                            client_response_headers_json: client_response_headers_json.as_deref(),
+                        };
+                        sync_usage_http_audit_storage(
+                            &mut **tx,
+                            &usage.request_id,
+                            &http_audit_headers,
+                            &http_audit_refs,
+                            &http_audit_states,
+                            http_audit_capture_mode,
+                        )
+                        .await?;
+                    }
+                    if capture_update_allowed {
+                        sync_usage_routing_snapshot_storage(
+                            &mut **tx,
+                            &usage.request_id,
+                            &routing_snapshot,
+                            replace_terminal_snapshots,
+                        )
+                        .await?;
+                        sync_usage_settlement_pricing_snapshot_storage(
+                            &mut **tx,
+                            &usage.request_id,
+                            &settlement_pricing_snapshot,
+                            replace_terminal_snapshots,
+                        )
+                        .await?;
+                    }
 
                     let mut stored = find_usage_by_request_id_in_tx(tx, &usage.request_id)
                         .await?
@@ -8165,71 +8221,79 @@ ORDER BY "usage".user_id ASC
                                 usage.request_id
                             ))
                         })?;
-                    if request_body_storage.has_detached_blob() {
-                        stored.request_body = usage.request_body.clone();
+                    if capture_update_allowed {
+                        if request_body_storage.has_detached_blob() {
+                            stored.request_body = usage.request_body.clone();
+                        }
+                        stored.request_headers = usage.request_headers.clone();
+                        stored.provider_request_headers = usage.provider_request_headers.clone();
+                        if provider_request_body_storage.has_detached_blob() {
+                            stored.provider_request_body = usage.provider_request_body.clone();
+                        }
+                        stored.response_headers = usage.response_headers.clone();
+                        if response_body_storage.has_detached_blob() {
+                            stored.response_body = usage.response_body.clone();
+                        }
+                        stored.client_response_headers = usage.client_response_headers.clone();
+                        if client_response_body_storage.has_detached_blob() {
+                            stored.client_response_body = usage.client_response_body.clone();
+                        }
+                        stored.request_body_ref = if clear_request_body {
+                            None
+                        } else {
+                            resolved_write_usage_body_ref(
+                                usage.request_body_ref.as_deref(),
+                                &usage.request_id,
+                                UsageBodyField::RequestBody,
+                                request_body_storage.has_detached_blob(),
+                                http_audit_refs.request_body_ref.as_deref(),
+                            )
+                        };
+                        stored.provider_request_body_ref = if clear_provider_request_body {
+                            None
+                        } else {
+                            resolved_write_usage_body_ref(
+                                usage.provider_request_body_ref.as_deref(),
+                                &usage.request_id,
+                                UsageBodyField::ProviderRequestBody,
+                                provider_request_body_storage.has_detached_blob(),
+                                http_audit_refs.provider_request_body_ref.as_deref(),
+                            )
+                        };
+                        stored.response_body_ref = if clear_response_body {
+                            None
+                        } else {
+                            resolved_write_usage_body_ref(
+                                usage.response_body_ref.as_deref(),
+                                &usage.request_id,
+                                UsageBodyField::ResponseBody,
+                                response_body_storage.has_detached_blob(),
+                                http_audit_refs.response_body_ref.as_deref(),
+                            )
+                        };
+                        stored.client_response_body_ref = if clear_client_response_body {
+                            None
+                        } else {
+                            resolved_write_usage_body_ref(
+                                usage.client_response_body_ref.as_deref(),
+                                &usage.request_id,
+                                UsageBodyField::ClientResponseBody,
+                                client_response_body_storage.has_detached_blob(),
+                                http_audit_refs.client_response_body_ref.as_deref(),
+                            )
+                        };
+                        stored.request_body_state =
+                            usage.request_body_state.or(stored.request_body_state);
+                        stored.provider_request_body_state = usage
+                            .provider_request_body_state
+                            .or(stored.provider_request_body_state);
+                        stored.response_body_state =
+                            usage.response_body_state.or(stored.response_body_state);
+                        stored.client_response_body_state = usage
+                            .client_response_body_state
+                            .or(stored.client_response_body_state);
+                        stored.request_metadata = request_metadata_value;
                     }
-                    stored.request_headers = usage.request_headers.clone();
-                    stored.provider_request_headers = usage.provider_request_headers.clone();
-                    if provider_request_body_storage.has_detached_blob() {
-                        stored.provider_request_body = usage.provider_request_body.clone();
-                    }
-                    stored.response_headers = usage.response_headers.clone();
-                    if response_body_storage.has_detached_blob() {
-                        stored.response_body = usage.response_body.clone();
-                    }
-                    stored.client_response_headers = usage.client_response_headers.clone();
-                    if client_response_body_storage.has_detached_blob() {
-                        stored.client_response_body = usage.client_response_body.clone();
-                    }
-                    stored.request_body_ref = resolved_write_usage_body_ref(
-                        usage.request_body_ref.as_deref(),
-                        &usage.request_id,
-                        UsageBodyField::RequestBody,
-                        request_body_storage.has_detached_blob(),
-                        http_audit_refs.request_body_ref.as_deref(),
-                    );
-                    stored.provider_request_body_ref = resolved_write_usage_body_ref(
-                        usage.provider_request_body_ref.as_deref(),
-                        &usage.request_id,
-                        UsageBodyField::ProviderRequestBody,
-                        provider_request_body_storage.has_detached_blob(),
-                        http_audit_refs.provider_request_body_ref.as_deref(),
-                    );
-                    stored.response_body_ref = resolved_write_usage_body_ref(
-                        usage.response_body_ref.as_deref(),
-                        &usage.request_id,
-                        UsageBodyField::ResponseBody,
-                        response_body_storage.has_detached_blob(),
-                        http_audit_refs.response_body_ref.as_deref(),
-                    );
-                    stored.client_response_body_ref = resolved_write_usage_body_ref(
-                        usage.client_response_body_ref.as_deref(),
-                        &usage.request_id,
-                        UsageBodyField::ClientResponseBody,
-                        client_response_body_storage.has_detached_blob(),
-                        http_audit_refs.client_response_body_ref.as_deref(),
-                    );
-                    stored.request_body_state =
-                        usage.request_body_state.or(stored.request_body_state);
-                    stored.provider_request_body_state = usage
-                        .provider_request_body_state
-                        .or(stored.provider_request_body_state);
-                    stored.response_body_state =
-                        usage.response_body_state.or(stored.response_body_state);
-                    stored.client_response_body_state = usage
-                        .client_response_body_state
-                        .or(stored.client_response_body_state);
-                    stored.candidate_id = routing_snapshot.candidate_id.clone();
-                    stored.candidate_index = routing_snapshot.candidate_index;
-                    stored.key_name = routing_snapshot.key_name.clone();
-                    stored.planner_kind = routing_snapshot.planner_kind.clone();
-                    stored.route_family = routing_snapshot.route_family.clone();
-                    stored.route_kind = routing_snapshot.route_kind.clone();
-                    stored.execution_path = routing_snapshot.execution_path.clone();
-                    stored.local_execution_runtime_miss_reason =
-                        routing_snapshot.local_execution_runtime_miss_reason.clone();
-                    stored.output_price_per_1m = settlement_pricing_snapshot.output_price_per_1m;
-                    stored.request_metadata = request_metadata_value;
 
                     let before_api_key_contribution =
                         previous_usage.as_ref().and_then(api_key_usage_contribution);
@@ -10415,6 +10479,112 @@ struct PreparedUsageUpsert {
     settlement_pricing_snapshot: UsageSettlementPricingSnapshot,
     request_metadata_value: Option<Value>,
     request_metadata_json: Option<String>,
+    replace_client_request_body_facts: bool,
+    replace_provider_request_body_facts: bool,
+    clear_request_body: bool,
+    clear_provider_request_body: bool,
+    clear_response_body: bool,
+    clear_client_response_body: bool,
+}
+
+fn request_body_capture_replaces_derived_facts(
+    request_body: Option<&Value>,
+    request_body_state: Option<UsageBodyCaptureState>,
+) -> bool {
+    // A typed capture state belongs to the incoming request snapshot. Metadata derived before a
+    // body was externalized, truncated, disabled, or found unavailable is authoritative when
+    // present; its absence must clear facts from an older candidate instead of falling through to
+    // PostgreSQL's sparse-upsert COALESCE behavior.
+    if request_body_state.is_some() {
+        return true;
+    }
+
+    let Some(request_body) = request_body else {
+        return false;
+    };
+
+    !request_body.as_object().is_some_and(|body| {
+        body.get("truncated").and_then(Value::as_bool) == Some(true)
+            && body.get("reason").and_then(Value::as_str) == Some("body_capture_limit_exceeded")
+    })
+}
+
+fn clear_previous_request_body_facts(
+    previous_metadata: Option<&Value>,
+    clear_client_request_body_facts: bool,
+    clear_provider_request_body_facts: bool,
+) -> Value {
+    let mut metadata = previous_metadata
+        .and_then(Value::as_object)
+        .cloned()
+        .unwrap_or_default();
+    if clear_client_request_body_facts {
+        metadata.remove(REQUESTED_REASONING_EFFORT_METADATA_KEY);
+        metadata.remove("request_body_ref");
+    }
+    if clear_provider_request_body_facts {
+        metadata.remove(PROVIDER_REASONING_EFFORT_METADATA_KEY);
+        metadata.remove(PROVIDER_SERVICE_TIER_METADATA_KEY);
+        metadata.remove(PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY);
+        metadata.remove("provider_request_body_ref");
+    }
+    // Keep an explicit empty object as a tombstone. Binding SQL NULL here would make the upsert's
+    // COALESCE retain the previous candidate's request-derived facts.
+    Value::Object(metadata)
+}
+
+fn retain_previous_request_audit_metadata(
+    previous_metadata: Option<&Value>,
+    preserve_client_request_body_facts: bool,
+) -> Value {
+    let Some(previous_metadata) = previous_metadata.and_then(Value::as_object) else {
+        return Value::Object(Map::new());
+    };
+    let mut retained = Map::new();
+    for key in [
+        "trace_id",
+        "client_ip",
+        "user_agent",
+        "client_family",
+        "client_requested_stream",
+        "client_session_affinity",
+        "api_key_is_standalone",
+        "request_path",
+        "request_query_string",
+        "request_path_and_query",
+    ] {
+        if let Some(value) = previous_metadata.get(key) {
+            retained.insert(key.to_string(), value.clone());
+        }
+    }
+    if preserve_client_request_body_facts {
+        for key in [REQUESTED_REASONING_EFFORT_METADATA_KEY, "request_body_ref"] {
+            if let Some(value) = previous_metadata.get(key) {
+                retained.insert(key.to_string(), value.clone());
+            }
+        }
+    }
+    Value::Object(retained)
+}
+
+fn usage_capture_update_allowed(
+    previous_usage: Option<(&str, &str)>,
+    incoming_status: &str,
+) -> bool {
+    let Some((previous_status, previous_billing_status)) = previous_usage else {
+        return true;
+    };
+    if previous_billing_status != "pending" {
+        return false;
+    }
+
+    let previous_is_terminal = matches!(previous_status, "completed" | "failed" | "cancelled");
+    let incoming_is_non_terminal = matches!(incoming_status, "pending" | "streaming");
+    if previous_is_terminal && incoming_is_non_terminal {
+        return false;
+    }
+
+    !(previous_status == "streaming" && incoming_status == "pending")
 }
 
 fn prepare_usage_body_storage(value: Option<&Value>) -> Result<UsageBodyStorage, DataLayerError> {
@@ -10485,40 +10655,78 @@ fn json_bind_text(value: Option<&Value>) -> Result<Option<String>, DataLayerErro
 fn prepare_usage_upsert_context(
     usage: &UpsertUsageRecord,
 ) -> Result<PreparedUsageUpsert, DataLayerError> {
+    let replace_client_request_body_facts = request_body_capture_replaces_derived_facts(
+        usage.request_body.as_ref(),
+        usage.request_body_state,
+    );
+    let replace_provider_request_body_facts = request_body_capture_replaces_derived_facts(
+        usage.provider_request_body.as_ref(),
+        usage.provider_request_body_state,
+    );
+    let clear_request_body = usage.request_body_state == Some(UsageBodyCaptureState::None);
+    let clear_provider_request_body =
+        usage.provider_request_body_state == Some(UsageBodyCaptureState::None);
+    let clear_response_body = usage.response_body_state == Some(UsageBodyCaptureState::None);
+    let clear_client_response_body =
+        usage.client_response_body_state == Some(UsageBodyCaptureState::None);
+    // A typed `none` marker wins over residual values left on a reused event by an earlier
+    // candidate. Do not serialize those values or recreate their detached references.
+    let request_body = (!clear_request_body)
+        .then_some(usage.request_body.as_ref())
+        .flatten();
+    let provider_request_body = (!clear_provider_request_body)
+        .then_some(usage.provider_request_body.as_ref())
+        .flatten();
+    let response_body = (!clear_response_body)
+        .then_some(usage.response_body.as_ref())
+        .flatten();
+    let client_response_body = (!clear_client_response_body)
+        .then_some(usage.client_response_body.as_ref())
+        .flatten();
+    let request_body_ref = (!clear_request_body)
+        .then_some(usage.request_body_ref.as_deref())
+        .flatten();
+    let provider_request_body_ref = (!clear_provider_request_body)
+        .then_some(usage.provider_request_body_ref.as_deref())
+        .flatten();
+    let response_body_ref = (!clear_response_body)
+        .then_some(usage.response_body_ref.as_deref())
+        .flatten();
+    let client_response_body_ref = (!clear_client_response_body)
+        .then_some(usage.client_response_body_ref.as_deref())
+        .flatten();
     let request_headers_json = json_bind_text(usage.request_headers.as_ref())?;
-    let request_body_storage = prepare_usage_body_storage(usage.request_body.as_ref())?;
+    let request_body_storage = prepare_usage_body_storage(request_body)?;
     let provider_request_headers_json = json_bind_text(usage.provider_request_headers.as_ref())?;
-    let provider_request_body_storage =
-        prepare_usage_body_storage(usage.provider_request_body.as_ref())?;
+    let provider_request_body_storage = prepare_usage_body_storage(provider_request_body)?;
     let response_headers_json = json_bind_text(usage.response_headers.as_ref())?;
-    let response_body_storage = prepare_usage_body_storage(usage.response_body.as_ref())?;
+    let response_body_storage = prepare_usage_body_storage(response_body)?;
     let client_response_headers_json = json_bind_text(usage.client_response_headers.as_ref())?;
-    let client_response_body_storage =
-        prepare_usage_body_storage(usage.client_response_body.as_ref())?;
+    let client_response_body_storage = prepare_usage_body_storage(client_response_body)?;
     let http_audit_refs = UsageHttpAuditRefs {
         request_body_ref: resolved_write_usage_body_ref(
-            usage.request_body_ref.as_deref(),
+            request_body_ref,
             &usage.request_id,
             UsageBodyField::RequestBody,
             request_body_storage.has_detached_blob(),
             None,
         ),
         provider_request_body_ref: resolved_write_usage_body_ref(
-            usage.provider_request_body_ref.as_deref(),
+            provider_request_body_ref,
             &usage.request_id,
             UsageBodyField::ProviderRequestBody,
             provider_request_body_storage.has_detached_blob(),
             None,
         ),
         response_body_ref: resolved_write_usage_body_ref(
-            usage.response_body_ref.as_deref(),
+            response_body_ref,
             &usage.request_id,
             UsageBodyField::ResponseBody,
             response_body_storage.has_detached_blob(),
             None,
         ),
         client_response_body_ref: resolved_write_usage_body_ref(
-            usage.client_response_body_ref.as_deref(),
+            client_response_body_ref,
             &usage.request_id,
             UsageBodyField::ClientResponseBody,
             client_response_body_storage.has_detached_blob(),
@@ -10547,42 +10755,49 @@ fn prepare_usage_upsert_context(
             http_audit_refs.client_response_body_ref.as_deref(),
         ),
     };
-    let request_metadata_value = prepare_request_metadata_for_body_storage(
+    let mut request_metadata_value = prepare_request_metadata_for_body_storage(
         usage.request_metadata.clone(),
         [
             (
                 UsageBodyField::RequestBody,
                 &request_body_storage,
-                usage.request_body.as_ref(),
-                usage.request_body_ref.as_deref(),
+                request_body,
+                request_body_ref,
             ),
             (
                 UsageBodyField::ProviderRequestBody,
                 &provider_request_body_storage,
-                usage.provider_request_body.as_ref(),
-                usage.provider_request_body_ref.as_deref(),
+                provider_request_body,
+                provider_request_body_ref,
             ),
             (
                 UsageBodyField::ResponseBody,
                 &response_body_storage,
-                usage.response_body.as_ref(),
-                usage.response_body_ref.as_deref(),
+                response_body,
+                response_body_ref,
             ),
             (
                 UsageBodyField::ClientResponseBody,
                 &client_response_body_storage,
-                usage.client_response_body.as_ref(),
-                usage.client_response_body_ref.as_deref(),
+                client_response_body,
+                client_response_body_ref,
             ),
         ],
     );
+    if request_metadata_value.is_some() && (clear_request_body || clear_provider_request_body) {
+        request_metadata_value = Some(clear_previous_request_body_facts(
+            request_metadata_value.as_ref(),
+            clear_request_body,
+            clear_provider_request_body,
+        ));
+    }
     let http_audit_capture_mode = usage_http_audit_capture_mode(
         &http_audit_refs,
         [
-            usage.request_body.as_ref(),
-            usage.provider_request_body.as_ref(),
-            usage.response_body.as_ref(),
-            usage.client_response_body.as_ref(),
+            request_body,
+            provider_request_body,
+            response_body,
+            client_response_body,
         ],
     );
     let routing_snapshot =
@@ -10607,6 +10822,12 @@ fn prepare_usage_upsert_context(
         settlement_pricing_snapshot,
         request_metadata_value,
         request_metadata_json,
+        replace_client_request_body_facts,
+        replace_provider_request_body_facts,
+        clear_request_body,
+        clear_provider_request_body,
+        clear_response_body,
+        clear_client_response_body,
     })
 }
 
@@ -11442,11 +11663,20 @@ async fn sync_usage_body_blob_storage<'e, E>(
     field: UsageBodyField,
     value: Option<&Value>,
     storage: &UsageBodyStorage,
+    clear_existing: bool,
 ) -> Result<(), DataLayerError>
 where
     E: sqlx::Executor<'e, Database = Postgres>,
 {
     let body_ref = usage_body_ref(request_id, field);
+    if clear_existing {
+        sqlx::query(DELETE_USAGE_BODY_BLOB_SQL)
+            .bind(&body_ref)
+            .execute(executor)
+            .await
+            .map_postgres_err()?;
+        return Ok(());
+    }
     if let Some(payload_gzip) = storage.detached_blob_bytes.as_ref() {
         sqlx::query(UPSERT_USAGE_BODY_BLOB_SQL)
             .bind(&body_ref)
@@ -11523,11 +11753,12 @@ async fn sync_usage_routing_snapshot_storage<'e, E>(
     executor: E,
     request_id: &str,
     snapshot: &UsageRoutingSnapshot,
+    replace_existing: bool,
 ) -> Result<(), DataLayerError>
 where
     E: sqlx::Executor<'e, Database = Postgres>,
 {
-    if !snapshot.any_present() {
+    if !snapshot.any_present() && !replace_existing {
         return Ok(());
     }
 
@@ -11545,6 +11776,7 @@ where
         .bind(snapshot.selected_endpoint_id.as_deref())
         .bind(snapshot.selected_provider_api_key_id.as_deref())
         .bind(snapshot.has_format_conversion)
+        .bind(replace_existing)
         .execute(executor)
         .await
         .map_postgres_err()?;
@@ -11556,11 +11788,12 @@ async fn sync_usage_settlement_pricing_snapshot_storage<'e, E>(
     executor: E,
     request_id: &str,
     snapshot: &UsageSettlementPricingSnapshot,
+    replace_existing: bool,
 ) -> Result<(), DataLayerError>
 where
     E: sqlx::Executor<'e, Database = Postgres>,
 {
-    if !snapshot.any_present() {
+    if !snapshot.any_present() && !replace_existing {
         return Ok(());
     }
 
@@ -11594,6 +11827,7 @@ where
         .bind(snapshot.cache_creation_price_per_1m)
         .bind(snapshot.cache_read_price_per_1m)
         .bind(snapshot.price_per_request)
+        .bind(replace_existing)
         .execute(executor)
         .await
         .map_postgres_err()?;

crates/aether-data/adapters/postgres/src/usage/queries/list_recent_usage_audits_prefix.sql

@@ -179,6 +179,7 @@ SELECT
       OR NULLIF(BTRIM("usage".request_metadata->>'user_agent'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'request_path'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'request_path_and_query'), '') IS NOT NULL
+      OR NULLIF(BTRIM("usage".request_metadata->>'requested_reasoning_effort'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_reasoning_effort'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_service_tier'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_actual_service_tier'), '') IS NOT NULL
@@ -193,6 +194,8 @@ SELECT
         NULLIF(BTRIM("usage".request_metadata->>'request_path'), ''),
         'request_path_and_query',
         NULLIF(BTRIM("usage".request_metadata->>'request_path_and_query'), ''),
+        'requested_reasoning_effort',
+        NULLIF(BTRIM("usage".request_metadata->>'requested_reasoning_effort'), ''),
         'provider_reasoning_effort',
         NULLIF(BTRIM("usage".request_metadata->>'provider_reasoning_effort'), ''),
         'provider_service_tier',

crates/aether-data/adapters/postgres/src/usage/queries/list_usage_audits_prefix.sql

@@ -179,6 +179,7 @@ SELECT
       OR NULLIF(BTRIM("usage".request_metadata->>'user_agent'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'request_path'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'request_path_and_query'), '') IS NOT NULL
+      OR NULLIF(BTRIM("usage".request_metadata->>'requested_reasoning_effort'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_reasoning_effort'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_service_tier'), '') IS NOT NULL
       OR NULLIF(BTRIM("usage".request_metadata->>'provider_actual_service_tier'), '') IS NOT NULL
@@ -193,6 +194,8 @@ SELECT
         NULLIF(BTRIM("usage".request_metadata->>'request_path'), ''),
         'request_path_and_query',
         NULLIF(BTRIM("usage".request_metadata->>'request_path_and_query'), ''),
+        'requested_reasoning_effort',
+        NULLIF(BTRIM("usage".request_metadata->>'requested_reasoning_effort'), ''),
         'provider_reasoning_effort',
         NULLIF(BTRIM("usage".request_metadata->>'provider_reasoning_effort'), ''),
         'provider_service_tier',

crates/aether-data/adapters/postgres/src/usage/queries/upsert_sql.sql

@@ -157,22 +157,22 @@ DO UPDATE SET
   api_key_id = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.api_key_id, "usage".api_key_id) ELSE "usage".api_key_id END,
   username = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.username, "usage".username) ELSE "usage".username END,
   api_key_name = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.api_key_name, "usage".api_key_name) ELSE "usage".api_key_name END,
-  provider_name = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.provider_name, "usage".provider_name) ELSE "usage".provider_name END,
-  model = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.model, "usage".model) ELSE "usage".model END,
-  target_model = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.target_model, "usage".target_model) ELSE "usage".target_model END,
-  provider_id = CASE WHEN "usage".billing_status = 'pending' OR ("usage".provider_id IS NULL AND ("usage".provider_endpoint_id IS NULL OR "usage".provider_endpoint_id = EXCLUDED.provider_endpoint_id) AND ("usage".provider_api_key_id IS NULL OR "usage".provider_api_key_id = EXCLUDED.provider_api_key_id)) THEN COALESCE(EXCLUDED.provider_id, "usage".provider_id) ELSE "usage".provider_id END,
-  provider_endpoint_id = CASE WHEN "usage".billing_status = 'pending' OR ("usage".provider_endpoint_id IS NULL AND ("usage".provider_id IS NULL OR "usage".provider_id = EXCLUDED.provider_id) AND ("usage".provider_api_key_id IS NULL OR "usage".provider_api_key_id = EXCLUDED.provider_api_key_id)) THEN COALESCE(EXCLUDED.provider_endpoint_id, "usage".provider_endpoint_id) ELSE "usage".provider_endpoint_id END,
-  provider_api_key_id = CASE WHEN "usage".billing_status = 'pending' OR ("usage".provider_api_key_id IS NULL AND ("usage".provider_id IS NULL OR "usage".provider_id = EXCLUDED.provider_id) AND ("usage".provider_endpoint_id IS NULL OR "usage".provider_endpoint_id = EXCLUDED.provider_endpoint_id)) THEN COALESCE(EXCLUDED.provider_api_key_id, "usage".provider_api_key_id) ELSE "usage".provider_api_key_id END,
-  request_type = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.request_type, "usage".request_type) ELSE "usage".request_type END,
-  api_format = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.api_format, "usage".api_format) ELSE "usage".api_format END,
-  api_family = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.api_family, "usage".api_family) ELSE "usage".api_family END,
-  endpoint_kind = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.endpoint_kind, "usage".endpoint_kind) ELSE "usage".endpoint_kind END,
-  endpoint_api_format = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.endpoint_api_format, "usage".endpoint_api_format) ELSE "usage".endpoint_api_format END,
-  provider_api_family = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.provider_api_family, "usage".provider_api_family) ELSE "usage".provider_api_family END,
-  provider_endpoint_kind = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.provider_endpoint_kind, "usage".provider_endpoint_kind) ELSE "usage".provider_endpoint_kind END,
-  has_format_conversion = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.has_format_conversion, "usage".has_format_conversion) ELSE "usage".has_format_conversion END,
-  is_stream = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.is_stream, "usage".is_stream) ELSE "usage".is_stream END,
-  upstream_is_stream = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.upstream_is_stream, "usage".upstream_is_stream, "usage".is_stream, false) ELSE "usage".upstream_is_stream END,
+  provider_name = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.provider_name, "usage".provider_name) ELSE "usage".provider_name END,
+  model = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.model, "usage".model) ELSE "usage".model END,
+  target_model = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN CASE WHEN EXCLUDED.status IN ('completed', 'failed', 'cancelled') THEN EXCLUDED.target_model ELSE COALESCE(EXCLUDED.target_model, "usage".target_model) END ELSE "usage".target_model END,
+  provider_id = CASE WHEN ("usage".billing_status = 'pending' AND $61) OR ("usage".billing_status <> 'pending' AND "usage".provider_id IS NULL AND ("usage".provider_endpoint_id IS NULL OR "usage".provider_endpoint_id = EXCLUDED.provider_endpoint_id) AND ("usage".provider_api_key_id IS NULL OR "usage".provider_api_key_id = EXCLUDED.provider_api_key_id)) THEN COALESCE(EXCLUDED.provider_id, "usage".provider_id) ELSE "usage".provider_id END,
+  provider_endpoint_id = CASE WHEN ("usage".billing_status = 'pending' AND $61) OR ("usage".billing_status <> 'pending' AND "usage".provider_endpoint_id IS NULL AND ("usage".provider_id IS NULL OR "usage".provider_id = EXCLUDED.provider_id) AND ("usage".provider_api_key_id IS NULL OR "usage".provider_api_key_id = EXCLUDED.provider_api_key_id)) THEN COALESCE(EXCLUDED.provider_endpoint_id, "usage".provider_endpoint_id) ELSE "usage".provider_endpoint_id END,
+  provider_api_key_id = CASE WHEN ("usage".billing_status = 'pending' AND $61) OR ("usage".billing_status <> 'pending' AND "usage".provider_api_key_id IS NULL AND ("usage".provider_id IS NULL OR "usage".provider_id = EXCLUDED.provider_id) AND ("usage".provider_endpoint_id IS NULL OR "usage".provider_endpoint_id = EXCLUDED.provider_endpoint_id)) THEN COALESCE(EXCLUDED.provider_api_key_id, "usage".provider_api_key_id) ELSE "usage".provider_api_key_id END,
+  request_type = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.request_type, "usage".request_type) ELSE "usage".request_type END,
+  api_format = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.api_format, "usage".api_format) ELSE "usage".api_format END,
+  api_family = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.api_family, "usage".api_family) ELSE "usage".api_family END,
+  endpoint_kind = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.endpoint_kind, "usage".endpoint_kind) ELSE "usage".endpoint_kind END,
+  endpoint_api_format = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.endpoint_api_format, "usage".endpoint_api_format) ELSE "usage".endpoint_api_format END,
+  provider_api_family = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.provider_api_family, "usage".provider_api_family) ELSE "usage".provider_api_family END,
+  provider_endpoint_kind = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.provider_endpoint_kind, "usage".provider_endpoint_kind) ELSE "usage".provider_endpoint_kind END,
+  has_format_conversion = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.has_format_conversion, "usage".has_format_conversion) ELSE "usage".has_format_conversion END,
+  is_stream = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.is_stream, "usage".is_stream) ELSE "usage".is_stream END,
+  upstream_is_stream = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.upstream_is_stream, "usage".upstream_is_stream, "usage".is_stream, false) ELSE "usage".upstream_is_stream END,
   input_tokens = CASE WHEN "usage".billing_status = 'pending' AND EXCLUDED.status IN ('completed', 'failed', 'cancelled') THEN GREATEST("usage".input_tokens, EXCLUDED.input_tokens) ELSE "usage".input_tokens END,
   output_tokens = CASE WHEN "usage".billing_status = 'pending' AND EXCLUDED.status IN ('completed', 'failed', 'cancelled') THEN GREATEST("usage".output_tokens, EXCLUDED.output_tokens) ELSE "usage".output_tokens END,
   total_tokens = CASE WHEN "usage".billing_status = 'pending' AND EXCLUDED.status IN ('completed', 'failed', 'cancelled') THEN GREATEST("usage".total_tokens, EXCLUDED.total_tokens) ELSE "usage".total_tokens END,
@@ -223,42 +223,42 @@ DO UPDATE SET
   END ELSE "usage".status END,
   billing_status = CASE WHEN "usage".billing_status = 'pending' THEN EXCLUDED.billing_status ELSE "usage".billing_status END,
   request_headers = NULL,
-  request_body = CASE WHEN "usage".billing_status = 'pending' THEN CASE
+  request_body = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN CASE
     WHEN EXCLUDED.request_body_compressed IS NOT NULL OR $57 THEN NULL
     ELSE COALESCE(EXCLUDED.request_body, "usage".request_body)
   END ELSE "usage".request_body END,
-  request_body_compressed = CASE WHEN "usage".billing_status = 'pending' THEN CASE
+  request_body_compressed = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN CASE
     WHEN EXCLUDED.request_body IS NOT NULL OR $57 THEN NULL
     ELSE COALESCE(EXCLUDED.request_body_compressed, "usage".request_body_compressed)
   END ELSE "usage".request_body_compressed END,
   provider_request_headers = NULL,
-  provider_request_body = CASE WHEN "usage".billing_status = 'pending' THEN CASE
+  provider_request_body = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN CASE
     WHEN EXCLUDED.provider_request_body_compressed IS NOT NULL OR $58 THEN NULL
     ELSE COALESCE(EXCLUDED.provider_request_body, "usage".provider_request_body)
   END ELSE "usage".provider_request_body END,
-  provider_request_body_compressed = CASE WHEN "usage".billing_status = 'pending' THEN CASE
+  provider_request_body_compressed = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN CASE
     WHEN EXCLUDED.provider_request_body IS NOT NULL OR $58 THEN NULL
     ELSE COALESCE(EXCLUDED.provider_request_body_compressed, "usage".provider_request_body_compressed)
   END ELSE "usage".provider_request_body_compressed END,
   response_headers = NULL,
-  response_body = CASE WHEN "usage".billing_status = 'pending' THEN CASE
+  response_body = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN CASE
     WHEN EXCLUDED.response_body_compressed IS NOT NULL OR $59 THEN NULL
     ELSE COALESCE(EXCLUDED.response_body, "usage".response_body)
   END ELSE "usage".response_body END,
-  response_body_compressed = CASE WHEN "usage".billing_status = 'pending' THEN CASE
+  response_body_compressed = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN CASE
     WHEN EXCLUDED.response_body IS NOT NULL OR $59 THEN NULL
     ELSE COALESCE(EXCLUDED.response_body_compressed, "usage".response_body_compressed)
   END ELSE "usage".response_body_compressed END,
   client_response_headers = NULL,
-  client_response_body = CASE WHEN "usage".billing_status = 'pending' THEN CASE
+  client_response_body = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN CASE
     WHEN EXCLUDED.client_response_body_compressed IS NOT NULL OR $60 THEN NULL
     ELSE COALESCE(EXCLUDED.client_response_body, "usage".client_response_body)
   END ELSE "usage".client_response_body END,
-  client_response_body_compressed = CASE WHEN "usage".billing_status = 'pending' THEN CASE
+  client_response_body_compressed = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN CASE
     WHEN EXCLUDED.client_response_body IS NOT NULL OR $60 THEN NULL
     ELSE COALESCE(EXCLUDED.client_response_body_compressed, "usage".client_response_body_compressed)
   END ELSE "usage".client_response_body_compressed END,
-  request_metadata = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.request_metadata, "usage".request_metadata) ELSE "usage".request_metadata END,
+  request_metadata = CASE WHEN "usage".billing_status = 'pending' AND $61 THEN COALESCE(EXCLUDED.request_metadata, "usage".request_metadata) ELSE "usage".request_metadata END,
   finalized_at = CASE WHEN "usage".billing_status = 'pending' THEN COALESCE(EXCLUDED.finalized_at, "usage".finalized_at) ELSE "usage".finalized_at END,
   updated_at_unix_secs = CASE WHEN "usage".billing_status = 'pending' THEN
     GREATEST(

crates/aether-data/adapters/postgres/src/usage/queries/upsert_usage_http_audit_sql.sql

@@ -41,16 +41,25 @@ DO UPDATE SET
     EXCLUDED.client_response_headers,
     usage_http_audits.client_response_headers
   ),
-  request_body_ref = COALESCE(EXCLUDED.request_body_ref, usage_http_audits.request_body_ref),
-  provider_request_body_ref = COALESCE(
-    EXCLUDED.provider_request_body_ref,
-    usage_http_audits.provider_request_body_ref
-  ),
-  response_body_ref = COALESCE(EXCLUDED.response_body_ref, usage_http_audits.response_body_ref),
-  client_response_body_ref = COALESCE(
-    EXCLUDED.client_response_body_ref,
-    usage_http_audits.client_response_body_ref
-  ),
+  request_body_ref = CASE
+    WHEN EXCLUDED.request_body_state = 'none' THEN NULL
+    ELSE COALESCE(EXCLUDED.request_body_ref, usage_http_audits.request_body_ref)
+  END,
+  provider_request_body_ref = CASE
+    WHEN EXCLUDED.provider_request_body_state = 'none' THEN NULL
+    ELSE COALESCE(EXCLUDED.provider_request_body_ref, usage_http_audits.provider_request_body_ref)
+  END,
+  response_body_ref = CASE
+    WHEN EXCLUDED.response_body_state = 'none' THEN NULL
+    ELSE COALESCE(EXCLUDED.response_body_ref, usage_http_audits.response_body_ref)
+  END,
+  client_response_body_ref = CASE
+    WHEN EXCLUDED.client_response_body_state = 'none' THEN NULL
+    ELSE COALESCE(
+      EXCLUDED.client_response_body_ref,
+      usage_http_audits.client_response_body_ref
+    )
+  END,
   request_body_state = COALESCE(
     EXCLUDED.request_body_state,
     usage_http_audits.request_body_state
@@ -67,9 +76,19 @@ DO UPDATE SET
     EXCLUDED.client_response_body_state,
     usage_http_audits.client_response_body_state
   ),
-  body_capture_mode = COALESCE(
-    NULLIF(EXCLUDED.body_capture_mode, 'none'),
-    usage_http_audits.body_capture_mode,
-    'none'
-  ),
+  body_capture_mode = CASE
+    WHEN EXCLUDED.body_capture_mode = 'none'
+      AND (
+        EXCLUDED.request_body_state = 'none'
+        OR EXCLUDED.provider_request_body_state = 'none'
+        OR EXCLUDED.response_body_state = 'none'
+        OR EXCLUDED.client_response_body_state = 'none'
+      )
+      THEN 'none'
+    ELSE COALESCE(
+      NULLIF(EXCLUDED.body_capture_mode, 'none'),
+      usage_http_audits.body_capture_mode,
+      'none'
+    )
+  END,
   updated_at = NOW()

crates/aether-data/adapters/postgres/src/usage/queries/upsert_usage_routing_snapshot_sql.sql

@@ -29,34 +29,16 @@ INSERT INTO usage_routing_snapshots (
 )
 ON CONFLICT (request_id)
 DO UPDATE SET
-  candidate_id = COALESCE(EXCLUDED.candidate_id, usage_routing_snapshots.candidate_id),
-  candidate_index = COALESCE(
-    EXCLUDED.candidate_index,
-    usage_routing_snapshots.candidate_index
-  ),
-  key_name = COALESCE(EXCLUDED.key_name, usage_routing_snapshots.key_name),
-  planner_kind = COALESCE(EXCLUDED.planner_kind, usage_routing_snapshots.planner_kind),
-  route_family = COALESCE(EXCLUDED.route_family, usage_routing_snapshots.route_family),
-  route_kind = COALESCE(EXCLUDED.route_kind, usage_routing_snapshots.route_kind),
-  execution_path = COALESCE(EXCLUDED.execution_path, usage_routing_snapshots.execution_path),
-  local_execution_runtime_miss_reason = COALESCE(
-    EXCLUDED.local_execution_runtime_miss_reason,
-    usage_routing_snapshots.local_execution_runtime_miss_reason
-  ),
-  selected_provider_id = COALESCE(
-    EXCLUDED.selected_provider_id,
-    usage_routing_snapshots.selected_provider_id
-  ),
-  selected_endpoint_id = COALESCE(
-    EXCLUDED.selected_endpoint_id,
-    usage_routing_snapshots.selected_endpoint_id
-  ),
-  selected_provider_api_key_id = COALESCE(
-    EXCLUDED.selected_provider_api_key_id,
-    usage_routing_snapshots.selected_provider_api_key_id
-  ),
-  has_format_conversion = COALESCE(
-    EXCLUDED.has_format_conversion,
-    usage_routing_snapshots.has_format_conversion
-  ),
+  candidate_id = CASE WHEN $14 THEN EXCLUDED.candidate_id ELSE COALESCE(EXCLUDED.candidate_id, usage_routing_snapshots.candidate_id) END,
+  candidate_index = CASE WHEN $14 THEN EXCLUDED.candidate_index ELSE COALESCE(EXCLUDED.candidate_index, usage_routing_snapshots.candidate_index) END,
+  key_name = CASE WHEN $14 THEN EXCLUDED.key_name ELSE COALESCE(EXCLUDED.key_name, usage_routing_snapshots.key_name) END,
+  planner_kind = CASE WHEN $14 THEN EXCLUDED.planner_kind ELSE COALESCE(EXCLUDED.planner_kind, usage_routing_snapshots.planner_kind) END,
+  route_family = CASE WHEN $14 THEN EXCLUDED.route_family ELSE COALESCE(EXCLUDED.route_family, usage_routing_snapshots.route_family) END,
+  route_kind = CASE WHEN $14 THEN EXCLUDED.route_kind ELSE COALESCE(EXCLUDED.route_kind, usage_routing_snapshots.route_kind) END,
+  execution_path = CASE WHEN $14 THEN EXCLUDED.execution_path ELSE COALESCE(EXCLUDED.execution_path, usage_routing_snapshots.execution_path) END,
+  local_execution_runtime_miss_reason = CASE WHEN $14 THEN EXCLUDED.local_execution_runtime_miss_reason ELSE COALESCE(EXCLUDED.local_execution_runtime_miss_reason, usage_routing_snapshots.local_execution_runtime_miss_reason) END,
+  selected_provider_id = CASE WHEN $14 THEN EXCLUDED.selected_provider_id ELSE COALESCE(EXCLUDED.selected_provider_id, usage_routing_snapshots.selected_provider_id) END,
+  selected_endpoint_id = CASE WHEN $14 THEN EXCLUDED.selected_endpoint_id ELSE COALESCE(EXCLUDED.selected_endpoint_id, usage_routing_snapshots.selected_endpoint_id) END,
+  selected_provider_api_key_id = CASE WHEN $14 THEN EXCLUDED.selected_provider_api_key_id ELSE COALESCE(EXCLUDED.selected_provider_api_key_id, usage_routing_snapshots.selected_provider_api_key_id) END,
+  has_format_conversion = CASE WHEN $14 THEN EXCLUDED.has_format_conversion ELSE COALESCE(EXCLUDED.has_format_conversion, usage_routing_snapshots.has_format_conversion) END,
   updated_at = NOW()

crates/aether-data/adapters/postgres/src/usage/queries/upsert_usage_settlement_pricing_snapshot_sql.sql

@@ -61,112 +61,32 @@ INSERT INTO usage_settlement_snapshots (
 )
 ON CONFLICT (request_id)
 DO UPDATE SET
-  billing_snapshot_schema_version = COALESCE(
-    EXCLUDED.billing_snapshot_schema_version,
-    usage_settlement_snapshots.billing_snapshot_schema_version
-  ),
-  billing_snapshot_status = COALESCE(
-    EXCLUDED.billing_snapshot_status,
-    usage_settlement_snapshots.billing_snapshot_status
-  ),
-  settlement_snapshot_schema_version = COALESCE(
-    EXCLUDED.settlement_snapshot_schema_version,
-    usage_settlement_snapshots.settlement_snapshot_schema_version
-  ),
-  settlement_snapshot = COALESCE(
-    EXCLUDED.settlement_snapshot,
-    usage_settlement_snapshots.settlement_snapshot
-  ),
-  billing_dimensions = COALESCE(
-    EXCLUDED.billing_dimensions,
-    usage_settlement_snapshots.billing_dimensions
-  ),
-  billing_input_tokens = COALESCE(
-    EXCLUDED.billing_input_tokens,
-    usage_settlement_snapshots.billing_input_tokens
-  ),
-  billing_effective_input_tokens = COALESCE(
-    EXCLUDED.billing_effective_input_tokens,
-    usage_settlement_snapshots.billing_effective_input_tokens
-  ),
-  billing_output_tokens = COALESCE(
-    EXCLUDED.billing_output_tokens,
-    usage_settlement_snapshots.billing_output_tokens
-  ),
-  billing_cache_creation_tokens = COALESCE(
-    EXCLUDED.billing_cache_creation_tokens,
-    usage_settlement_snapshots.billing_cache_creation_tokens
-  ),
-  billing_cache_creation_5m_tokens = COALESCE(
-    EXCLUDED.billing_cache_creation_5m_tokens,
-    usage_settlement_snapshots.billing_cache_creation_5m_tokens
-  ),
-  billing_cache_creation_1h_tokens = COALESCE(
-    EXCLUDED.billing_cache_creation_1h_tokens,
-    usage_settlement_snapshots.billing_cache_creation_1h_tokens
-  ),
-  billing_cache_read_tokens = COALESCE(
-    EXCLUDED.billing_cache_read_tokens,
-    usage_settlement_snapshots.billing_cache_read_tokens
-  ),
-  billing_total_input_context = COALESCE(
-    EXCLUDED.billing_total_input_context,
-    usage_settlement_snapshots.billing_total_input_context
-  ),
-  billing_cache_creation_cost_usd = COALESCE(
-    EXCLUDED.billing_cache_creation_cost_usd,
-    usage_settlement_snapshots.billing_cache_creation_cost_usd
-  ),
-  billing_cache_read_cost_usd = COALESCE(
-    EXCLUDED.billing_cache_read_cost_usd,
-    usage_settlement_snapshots.billing_cache_read_cost_usd
-  ),
-  billing_total_cost_usd = COALESCE(
-    EXCLUDED.billing_total_cost_usd,
-    usage_settlement_snapshots.billing_total_cost_usd
-  ),
-  billing_actual_total_cost_usd = COALESCE(
-    EXCLUDED.billing_actual_total_cost_usd,
-    usage_settlement_snapshots.billing_actual_total_cost_usd
-  ),
-  billing_pricing_source = COALESCE(
-    EXCLUDED.billing_pricing_source,
-    usage_settlement_snapshots.billing_pricing_source
-  ),
-  billing_rule_id = COALESCE(
-    EXCLUDED.billing_rule_id,
-    usage_settlement_snapshots.billing_rule_id
-  ),
-  billing_rule_version = COALESCE(
-    EXCLUDED.billing_rule_version,
-    usage_settlement_snapshots.billing_rule_version
-  ),
-  rate_multiplier = COALESCE(
-    EXCLUDED.rate_multiplier,
-    usage_settlement_snapshots.rate_multiplier
-  ),
-  is_free_tier = COALESCE(
-    EXCLUDED.is_free_tier,
-    usage_settlement_snapshots.is_free_tier
-  ),
-  input_price_per_1m = COALESCE(
-    EXCLUDED.input_price_per_1m,
-    usage_settlement_snapshots.input_price_per_1m
-  ),
-  output_price_per_1m = COALESCE(
-    EXCLUDED.output_price_per_1m,
-    usage_settlement_snapshots.output_price_per_1m
-  ),
-  cache_creation_price_per_1m = COALESCE(
-    EXCLUDED.cache_creation_price_per_1m,
-    usage_settlement_snapshots.cache_creation_price_per_1m
-  ),
-  cache_read_price_per_1m = COALESCE(
-    EXCLUDED.cache_read_price_per_1m,
-    usage_settlement_snapshots.cache_read_price_per_1m
-  ),
-  price_per_request = COALESCE(
-    EXCLUDED.price_per_request,
-    usage_settlement_snapshots.price_per_request
-  ),
+  billing_status = CASE WHEN $30 THEN EXCLUDED.billing_status ELSE usage_settlement_snapshots.billing_status END,
+  billing_snapshot_schema_version = CASE WHEN $30 THEN EXCLUDED.billing_snapshot_schema_version ELSE COALESCE(EXCLUDED.billing_snapshot_schema_version, usage_settlement_snapshots.billing_snapshot_schema_version) END,
+  billing_snapshot_status = CASE WHEN $30 THEN EXCLUDED.billing_snapshot_status ELSE COALESCE(EXCLUDED.billing_snapshot_status, usage_settlement_snapshots.billing_snapshot_status) END,
+  settlement_snapshot_schema_version = CASE WHEN $30 THEN EXCLUDED.settlement_snapshot_schema_version ELSE COALESCE(EXCLUDED.settlement_snapshot_schema_version, usage_settlement_snapshots.settlement_snapshot_schema_version) END,
+  settlement_snapshot = CASE WHEN $30 THEN EXCLUDED.settlement_snapshot ELSE COALESCE(EXCLUDED.settlement_snapshot, usage_settlement_snapshots.settlement_snapshot) END,
+  billing_dimensions = CASE WHEN $30 THEN EXCLUDED.billing_dimensions ELSE COALESCE(EXCLUDED.billing_dimensions, usage_settlement_snapshots.billing_dimensions) END,
+  billing_input_tokens = CASE WHEN $30 THEN EXCLUDED.billing_input_tokens ELSE COALESCE(EXCLUDED.billing_input_tokens, usage_settlement_snapshots.billing_input_tokens) END,
+  billing_effective_input_tokens = CASE WHEN $30 THEN EXCLUDED.billing_effective_input_tokens ELSE COALESCE(EXCLUDED.billing_effective_input_tokens, usage_settlement_snapshots.billing_effective_input_tokens) END,
+  billing_output_tokens = CASE WHEN $30 THEN EXCLUDED.billing_output_tokens ELSE COALESCE(EXCLUDED.billing_output_tokens, usage_settlement_snapshots.billing_output_tokens) END,
+  billing_cache_creation_tokens = CASE WHEN $30 THEN EXCLUDED.billing_cache_creation_tokens ELSE COALESCE(EXCLUDED.billing_cache_creation_tokens, usage_settlement_snapshots.billing_cache_creation_tokens) END,
+  billing_cache_creation_5m_tokens = CASE WHEN $30 THEN EXCLUDED.billing_cache_creation_5m_tokens ELSE COALESCE(EXCLUDED.billing_cache_creation_5m_tokens, usage_settlement_snapshots.billing_cache_creation_5m_tokens) END,
+  billing_cache_creation_1h_tokens = CASE WHEN $30 THEN EXCLUDED.billing_cache_creation_1h_tokens ELSE COALESCE(EXCLUDED.billing_cache_creation_1h_tokens, usage_settlement_snapshots.billing_cache_creation_1h_tokens) END,
+  billing_cache_read_tokens = CASE WHEN $30 THEN EXCLUDED.billing_cache_read_tokens ELSE COALESCE(EXCLUDED.billing_cache_read_tokens, usage_settlement_snapshots.billing_cache_read_tokens) END,
+  billing_total_input_context = CASE WHEN $30 THEN EXCLUDED.billing_total_input_context ELSE COALESCE(EXCLUDED.billing_total_input_context, usage_settlement_snapshots.billing_total_input_context) END,
+  billing_cache_creation_cost_usd = CASE WHEN $30 THEN EXCLUDED.billing_cache_creation_cost_usd ELSE COALESCE(EXCLUDED.billing_cache_creation_cost_usd, usage_settlement_snapshots.billing_cache_creation_cost_usd) END,
+  billing_cache_read_cost_usd = CASE WHEN $30 THEN EXCLUDED.billing_cache_read_cost_usd ELSE COALESCE(EXCLUDED.billing_cache_read_cost_usd, usage_settlement_snapshots.billing_cache_read_cost_usd) END,
+  billing_total_cost_usd = CASE WHEN $30 THEN EXCLUDED.billing_total_cost_usd ELSE COALESCE(EXCLUDED.billing_total_cost_usd, usage_settlement_snapshots.billing_total_cost_usd) END,
+  billing_actual_total_cost_usd = CASE WHEN $30 THEN EXCLUDED.billing_actual_total_cost_usd ELSE COALESCE(EXCLUDED.billing_actual_total_cost_usd, usage_settlement_snapshots.billing_actual_total_cost_usd) END,
+  billing_pricing_source = CASE WHEN $30 THEN EXCLUDED.billing_pricing_source ELSE COALESCE(EXCLUDED.billing_pricing_source, usage_settlement_snapshots.billing_pricing_source) END,
+  billing_rule_id = CASE WHEN $30 THEN EXCLUDED.billing_rule_id ELSE COALESCE(EXCLUDED.billing_rule_id, usage_settlement_snapshots.billing_rule_id) END,
+  billing_rule_version = CASE WHEN $30 THEN EXCLUDED.billing_rule_version ELSE COALESCE(EXCLUDED.billing_rule_version, usage_settlement_snapshots.billing_rule_version) END,
+  rate_multiplier = CASE WHEN $30 THEN EXCLUDED.rate_multiplier ELSE COALESCE(EXCLUDED.rate_multiplier, usage_settlement_snapshots.rate_multiplier) END,
+  is_free_tier = CASE WHEN $30 THEN EXCLUDED.is_free_tier ELSE COALESCE(EXCLUDED.is_free_tier, usage_settlement_snapshots.is_free_tier) END,
+  input_price_per_1m = CASE WHEN $30 THEN EXCLUDED.input_price_per_1m ELSE COALESCE(EXCLUDED.input_price_per_1m, usage_settlement_snapshots.input_price_per_1m) END,
+  output_price_per_1m = CASE WHEN $30 THEN EXCLUDED.output_price_per_1m ELSE COALESCE(EXCLUDED.output_price_per_1m, usage_settlement_snapshots.output_price_per_1m) END,
+  cache_creation_price_per_1m = CASE WHEN $30 THEN EXCLUDED.cache_creation_price_per_1m ELSE COALESCE(EXCLUDED.cache_creation_price_per_1m, usage_settlement_snapshots.cache_creation_price_per_1m) END,
+  cache_read_price_per_1m = CASE WHEN $30 THEN EXCLUDED.cache_read_price_per_1m ELSE COALESCE(EXCLUDED.cache_read_price_per_1m, usage_settlement_snapshots.cache_read_price_per_1m) END,
+  price_per_request = CASE WHEN $30 THEN EXCLUDED.price_per_request ELSE COALESCE(EXCLUDED.price_per_request, usage_settlement_snapshots.price_per_request) END,
   updated_at = NOW()

crates/aether-data/adapters/postgres/src/usage/tests.rs

@@ -1,25 +1,497 @@
 use chrono::{TimeZone, Utc};
 use serde_json::json;
+use sqlx::Row;
 
 use super::{
     attach_compressed_body_refs, attach_usage_http_audit_body_refs,
     attach_usage_routing_snapshot_metadata, attach_usage_settlement_pricing_snapshot_metadata,
-    inflate_usage_json_value, prepare_request_metadata_for_body_storage,
-    prepare_usage_body_storage, resolved_read_usage_body_ref, resolved_write_usage_body_ref,
+    clear_previous_request_body_facts, inflate_usage_json_value,
+    prepare_request_metadata_for_body_storage, prepare_usage_body_storage,
+    prepare_usage_upsert_context, request_body_capture_replaces_derived_facts,
+    resolved_read_usage_body_ref, resolved_write_usage_body_ref,
     split_dashboard_daily_aggregate_range, split_dashboard_hourly_aggregate_range,
-    usage_body_capture_state_for_storage, usage_body_ref, usage_effective_input_tokens,
-    usage_http_audit_body_refs, usage_http_audit_capture_mode, usage_routing_snapshot_from_usage,
-    usage_settlement_pricing_snapshot_from_usage, usage_total_input_context, AggregateRangeSplit,
-    SqlxUsageReadRepository, UsageHttpAuditRefs, UsageRoutingSnapshot,
-    UsageSettlementPricingSnapshot, MAX_INLINE_USAGE_BODY_BYTES,
+    usage_body_capture_state_for_storage, usage_body_ref, usage_capture_update_allowed,
+    usage_effective_input_tokens, usage_http_audit_body_refs, usage_http_audit_capture_mode,
+    usage_routing_snapshot_from_usage, usage_settlement_pricing_snapshot_from_usage,
+    usage_total_input_context, AggregateRangeSplit, SqlxUsageReadRepository, UsageHttpAuditRefs,
+    UsageRoutingSnapshot, UsageSettlementPricingSnapshot, MAX_INLINE_USAGE_BODY_BYTES,
 };
 use crate::{PostgresPoolConfig, PostgresPoolFactory};
 use aether_data_contracts::repository::usage::{
-    UpsertUsageRecord, UsageBodyCaptureState, UsageBodyField, UsageCostSavingsSummaryQuery,
-    UsageDashboardDailyBreakdownQuery, UsageDashboardSummaryQuery, UsageProviderPerformanceQuery,
-    UsageTimeSeriesGranularity,
+    UpsertUsageRecord, UsageAuditListQuery, UsageBodyCaptureState, UsageBodyField,
+    UsageCostSavingsSummaryQuery, UsageDashboardDailyBreakdownQuery, UsageDashboardSummaryQuery,
+    UsageProviderPerformanceQuery, UsageTimeSeriesGranularity,
 };
 
+fn fast_clear_usage_record(
+    request_id: &str,
+    provider_name: &str,
+    now_unix_secs: u64,
+    terminal: bool,
+    terminal_state: UsageBodyCaptureState,
+    terminal_service_tier: Option<&str>,
+) -> UpsertUsageRecord {
+    UpsertUsageRecord {
+        request_id: request_id.to_string(),
+        user_id: None,
+        api_key_id: None,
+        username: None,
+        api_key_name: None,
+        provider_name: provider_name.to_string(),
+        model: "gpt-5".to_string(),
+        target_model: Some("gpt-5".to_string()),
+        provider_id: None,
+        provider_endpoint_id: None,
+        provider_api_key_id: None,
+        request_type: Some("chat".to_string()),
+        api_format: Some("openai:chat".to_string()),
+        api_family: Some("openai".to_string()),
+        endpoint_kind: Some("chat".to_string()),
+        endpoint_api_format: Some("openai:chat".to_string()),
+        provider_api_family: Some("openai".to_string()),
+        provider_endpoint_kind: Some("chat".to_string()),
+        has_format_conversion: Some(false),
+        is_stream: Some(false),
+        input_tokens: terminal.then_some(1),
+        output_tokens: terminal.then_some(1),
+        total_tokens: terminal.then_some(2),
+        cache_creation_input_tokens: None,
+        cache_creation_ephemeral_5m_input_tokens: None,
+        cache_creation_ephemeral_1h_input_tokens: None,
+        cache_read_input_tokens: None,
+        cache_creation_cost_usd: None,
+        cache_read_cost_usd: None,
+        output_price_per_1m: None,
+        total_cost_usd: None,
+        actual_total_cost_usd: None,
+        status_code: terminal.then_some(200),
+        error_message: None,
+        error_category: None,
+        response_time_ms: terminal.then_some(10),
+        first_byte_time_ms: None,
+        status: if terminal { "completed" } else { "pending" }.to_string(),
+        billing_status: "pending".to_string(),
+        request_headers: None,
+        request_body: None,
+        request_body_ref: None,
+        request_body_state: None,
+        provider_request_headers: None,
+        provider_request_body: (!terminal).then(|| {
+            json!({
+                "model": "gpt-5",
+                "service_tier": "priority"
+            })
+        }),
+        provider_request_body_ref: None,
+        provider_request_body_state: Some(if terminal {
+            terminal_state
+        } else {
+            UsageBodyCaptureState::Inline
+        }),
+        response_headers: None,
+        response_body: None,
+        response_body_ref: None,
+        response_body_state: None,
+        client_response_headers: None,
+        client_response_body: None,
+        client_response_body_ref: None,
+        client_response_body_state: None,
+        candidate_id: None,
+        candidate_index: None,
+        key_name: None,
+        planner_kind: None,
+        route_family: None,
+        route_kind: None,
+        execution_path: None,
+        local_execution_runtime_miss_reason: None,
+        request_metadata: if terminal {
+            terminal_service_tier.map(|tier| json!({"provider_service_tier": tier}))
+        } else {
+            Some(json!({"provider_service_tier": "priority"}))
+        },
+        finalized_at_unix_secs: terminal.then_some(now_unix_secs + 1),
+        created_at_unix_ms: Some(now_unix_secs),
+        updated_at_unix_secs: now_unix_secs + u64::from(terminal),
+    }
+}
+
+#[tokio::test]
+#[ignore = "requires AETHER_TEST_DATABASE_URL and PostgreSQL migrations"]
+async fn live_terminal_none_capture_clears_fast_from_detail_and_lightweight_lists() {
+    let database_url = std::env::var("AETHER_TEST_DATABASE_URL")
+        .expect("AETHER_TEST_DATABASE_URL must point at the test database");
+    let factory = PostgresPoolFactory::new(PostgresPoolConfig {
+        database_url,
+        min_connections: 1,
+        max_connections: 2,
+        acquire_timeout_ms: 10_000,
+        idle_timeout_ms: 30_000,
+        max_lifetime_ms: 60_000,
+        statement_cache_capacity: 64,
+        require_ssl: false,
+    })
+    .expect("factory should build");
+    let repository =
+        SqlxUsageReadRepository::new(factory.connect_lazy().expect("lazy pool should build"));
+    crate::run_migrations(repository.pool())
+        .await
+        .expect("test database migrations should succeed");
+
+    let suffix = uuid::Uuid::new_v4().simple().to_string();
+    let request_id = format!("req-fast-clear-{suffix}");
+    let provider_name = format!("fast-clear-{suffix}");
+    let now_unix_secs = Utc::now().timestamp().max(0) as u64;
+
+    let mut pending_record = fast_clear_usage_record(
+        &request_id,
+        &provider_name,
+        now_unix_secs,
+        false,
+        UsageBodyCaptureState::Inline,
+        Some("priority"),
+    );
+    pending_record.candidate_id = Some("candidate-a".to_string());
+    pending_record.candidate_index = Some(1);
+    pending_record.key_name = Some("key-a".to_string());
+    pending_record.planner_kind = Some("planner-a".to_string());
+    pending_record.route_family = Some("route-family-a".to_string());
+    pending_record.route_kind = Some("route-kind-a".to_string());
+    pending_record.execution_path = Some("path-a".to_string());
+    pending_record.local_execution_runtime_miss_reason = Some("miss-a".to_string());
+    pending_record.request_metadata = Some(json!({
+        "trace_id": "trace-a",
+        "provider_service_tier": "priority",
+        "provider_actual_service_tier": "priority",
+        "billing_snapshot": {
+            "schema_version": "2.0",
+            "status": "complete",
+            "resolved_variables": {
+                "input_price_per_1m": 30.0,
+                "output_price_per_1m": 150.0
+            }
+        },
+        "settlement_snapshot": {
+            "schema_version": "2.0",
+            "pricing_snapshot": {
+                "pricing_source": "processing_tier",
+                "service_tier": "priority"
+            },
+            "billing_plan_snapshot": {
+                "rule_id": "fast-rule",
+                "rule_version": "1"
+            }
+        },
+        "billing_dimensions": {"service_tier": "priority"},
+        "rate_multiplier": 2.0,
+        "input_price_per_1m": 30.0,
+        "output_price_per_1m": 150.0
+    }));
+    let pending = repository
+        .upsert(pending_record)
+        .await
+        .expect("pending usage should persist");
+    assert_eq!(pending.provider_service_tier().as_deref(), Some("priority"));
+
+    let mut terminal_record = fast_clear_usage_record(
+        &request_id,
+        &provider_name,
+        now_unix_secs,
+        true,
+        UsageBodyCaptureState::None,
+        None,
+    );
+    terminal_record.provider_id = Some("final-provider-id".to_string());
+    terminal_record.provider_endpoint_id = Some("final-endpoint-id".to_string());
+    terminal_record.provider_api_key_id = Some("final-key-id".to_string());
+    terminal_record.target_model = None;
+    let terminal = repository
+        .upsert(terminal_record)
+        .await
+        .expect("terminal usage should persist");
+    assert_eq!(terminal.provider_service_tier(), None);
+
+    let stored = repository
+        .find_by_request_id(&request_id)
+        .await
+        .expect("detail lookup should succeed")
+        .expect("usage should exist");
+    assert_eq!(
+        stored.provider_request_body_state,
+        Some(UsageBodyCaptureState::None)
+    );
+    assert_eq!(stored.provider_service_tier(), None);
+    assert_eq!(stored.provider_actual_service_tier(), None);
+    let stored_metadata = stored
+        .request_metadata
+        .as_ref()
+        .and_then(serde_json::Value::as_object)
+        .expect("terminal audit metadata should remain");
+    assert_eq!(
+        stored_metadata
+            .get("trace_id")
+            .and_then(serde_json::Value::as_str),
+        Some("trace-a")
+    );
+    for stale_key in [
+        "provider_actual_service_tier",
+        "billing_snapshot",
+        "settlement_snapshot",
+        "billing_dimensions",
+        "rate_multiplier",
+        "input_price_per_1m",
+        "output_price_per_1m",
+        "candidate_id",
+    ] {
+        assert!(
+            stored_metadata.get(stale_key).is_none(),
+            "terminal metadata retained stale key {stale_key}"
+        );
+    }
+    assert_eq!(stored.settlement_rate_multiplier(), None);
+    assert_eq!(stored.settlement_input_price_per_1m(), None);
+    assert_eq!(stored.settlement_output_price_per_1m(), None);
+
+    let settlement_row = sqlx::query(
+        "SELECT settlement_snapshot, billing_dimensions, CAST(rate_multiplier AS DOUBLE PRECISION) AS rate_multiplier, CAST(input_price_per_1m AS DOUBLE PRECISION) AS input_price_per_1m, CAST(output_price_per_1m AS DOUBLE PRECISION) AS output_price_per_1m FROM usage_settlement_snapshots WHERE request_id = $1",
+    )
+    .bind(&request_id)
+    .fetch_one(repository.pool())
+    .await
+    .expect("settlement snapshot row should be readable");
+    assert!(settlement_row
+        .try_get::<Option<serde_json::Value>, _>("settlement_snapshot")
+        .expect("settlement snapshot should decode")
+        .is_none());
+    assert!(settlement_row
+        .try_get::<Option<serde_json::Value>, _>("billing_dimensions")
+        .expect("billing dimensions should decode")
+        .is_none());
+    assert!(settlement_row
+        .try_get::<Option<f64>, _>("rate_multiplier")
+        .expect("rate multiplier should decode")
+        .is_none());
+    assert!(settlement_row
+        .try_get::<Option<f64>, _>("input_price_per_1m")
+        .expect("input price should decode")
+        .is_none());
+    assert!(settlement_row
+        .try_get::<Option<f64>, _>("output_price_per_1m")
+        .expect("output price should decode")
+        .is_none());
+
+    let physical_metadata = sqlx::query_scalar::<_, Option<serde_json::Value>>(
+        "SELECT request_metadata FROM \"usage\" WHERE request_id = $1",
+    )
+    .bind(&request_id)
+    .fetch_one(repository.pool())
+    .await
+    .expect("physical metadata should be readable")
+    .expect("clear tombstone should be stored");
+    assert!(physical_metadata.get("provider_service_tier").is_none());
+
+    let physical_body = sqlx::query(
+        "SELECT provider_request_body, provider_request_body_compressed FROM \"usage\" WHERE request_id = $1",
+    )
+    .bind(&request_id)
+    .fetch_one(repository.pool())
+    .await
+    .expect("physical body columns should be readable");
+    assert!(physical_body
+        .try_get::<Option<serde_json::Value>, _>("provider_request_body")
+        .expect("provider body column should decode")
+        .is_none());
+    assert!(physical_body
+        .try_get::<Option<Vec<u8>>, _>("provider_request_body_compressed")
+        .expect("compressed provider body column should decode")
+        .is_none());
+
+    let physical_http = sqlx::query(
+        "SELECT provider_request_body_ref, provider_request_body_state FROM usage_http_audits WHERE request_id = $1",
+    )
+    .bind(&request_id)
+    .fetch_one(repository.pool())
+    .await
+    .expect("HTTP audit row should be readable");
+    assert!(physical_http
+        .try_get::<Option<String>, _>("provider_request_body_ref")
+        .expect("provider body ref should decode")
+        .is_none());
+    assert_eq!(
+        physical_http
+            .try_get::<Option<String>, _>("provider_request_body_state")
+            .expect("provider body state should decode")
+            .as_deref(),
+        Some("none")
+    );
+    let physical_blob_count = sqlx::query_scalar::<_, i64>(
+        "SELECT COUNT(*) FROM usage_body_blobs WHERE request_id = $1 AND body_field = 'provider_request_body'",
+    )
+    .bind(&request_id)
+    .fetch_one(repository.pool())
+    .await
+    .expect("provider body blob count should be readable");
+    assert_eq!(physical_blob_count, 0);
+
+    // A late pending candidate must not resurrect the terminal capture, metadata, or detached
+    // blob after the final state has been written.
+    let mut late_record = fast_clear_usage_record(
+        &request_id,
+        "late-provider",
+        now_unix_secs + 2,
+        false,
+        UsageBodyCaptureState::Inline,
+        Some("priority"),
+    );
+    late_record.model = "late-model".to_string();
+    late_record.target_model = Some("late-target".to_string());
+    late_record.provider_id = Some("late-provider-id".to_string());
+    late_record.provider_endpoint_id = Some("late-endpoint-id".to_string());
+    late_record.provider_api_key_id = Some("late-key-id".to_string());
+    late_record.endpoint_api_format = Some("late:format".to_string());
+    late_record.candidate_id = Some("late-candidate".to_string());
+    let late = repository
+        .upsert(late_record)
+        .await
+        .expect("late pending usage should be accepted without regressing capture");
+    assert_eq!(late.provider_service_tier(), None);
+    assert_eq!(late.provider_name, provider_name);
+    assert_eq!(late.model, "gpt-5");
+    assert_eq!(late.target_model, None);
+    assert_eq!(late.provider_id.as_deref(), Some("final-provider-id"));
+    assert_eq!(
+        late.provider_endpoint_id.as_deref(),
+        Some("final-endpoint-id")
+    );
+    assert_eq!(late.provider_api_key_id.as_deref(), Some("final-key-id"));
+    assert_eq!(late.endpoint_api_format.as_deref(), Some("openai:chat"));
+    assert_eq!(late.candidate_id, None);
+    let late_blob_count = sqlx::query_scalar::<_, i64>(
+        "SELECT COUNT(*) FROM usage_body_blobs WHERE request_id = $1 AND body_field = 'provider_request_body'",
+    )
+    .bind(&request_id)
+    .fetch_one(repository.pool())
+    .await
+    .expect("late provider body blob count should be readable");
+    assert_eq!(late_blob_count, 0);
+
+    let list_item = repository
+        .list_usage_audits(&UsageAuditListQuery {
+            provider_name: Some(provider_name.clone()),
+            limit: Some(10),
+            newest_first: true,
+            ..UsageAuditListQuery::default()
+        })
+        .await
+        .expect("usage list should succeed")
+        .into_iter()
+        .find(|item| item.request_id == request_id)
+        .expect("usage list should contain the test row");
+    assert_eq!(list_item.provider_request_body_state, None);
+    assert_eq!(list_item.provider_service_tier(), None);
+    assert_eq!(list_item.target_model, None);
+    assert_eq!(list_item.candidate_id, None);
+    assert_eq!(list_item.candidate_index, None);
+    assert_eq!(list_item.key_name, None);
+    assert_eq!(list_item.planner_kind, None);
+    assert_eq!(list_item.route_family, None);
+    assert_eq!(list_item.route_kind, None);
+    assert_eq!(list_item.execution_path, None);
+    assert_eq!(list_item.local_execution_runtime_miss_reason, None);
+
+    let recent_item = repository
+        .list_recent_usage_audits(None, 20)
+        .await
+        .expect("recent usage list should succeed")
+        .into_iter()
+        .find(|item| item.request_id == request_id)
+        .expect("recent usage list should contain the test row");
+    assert_eq!(recent_item.provider_request_body_state, None);
+    assert_eq!(recent_item.provider_service_tier(), None);
+
+    sqlx::query("DELETE FROM \"usage\" WHERE request_id = $1")
+        .bind(&request_id)
+        .execute(repository.pool())
+        .await
+        .expect("test usage should be removed");
+
+    for (index, state, incoming_tier, expected_tier) in [
+        ("disabled", UsageBodyCaptureState::Disabled, None, None),
+        ("truncated", UsageBodyCaptureState::Truncated, None, None),
+        (
+            "unavailable",
+            UsageBodyCaptureState::Unavailable,
+            None,
+            None,
+        ),
+        (
+            "none-stale-metadata",
+            UsageBodyCaptureState::None,
+            Some("priority"),
+            None,
+        ),
+        (
+            "disabled-preserved",
+            UsageBodyCaptureState::Disabled,
+            Some("priority"),
+            Some("priority"),
+        ),
+        (
+            "truncated-preserved",
+            UsageBodyCaptureState::Truncated,
+            Some("priority"),
+            Some("priority"),
+        ),
+    ] {
+        let request_id = format!("req-fast-clear-{suffix}-{index}");
+        let provider_name = format!("fast-clear-{suffix}-{index}");
+        repository
+            .upsert(fast_clear_usage_record(
+                &request_id,
+                &provider_name,
+                now_unix_secs,
+                false,
+                UsageBodyCaptureState::Inline,
+                Some("priority"),
+            ))
+            .await
+            .expect("typed-state pending usage should persist");
+        let terminal = repository
+            .upsert(fast_clear_usage_record(
+                &request_id,
+                &provider_name,
+                now_unix_secs + 1,
+                true,
+                state,
+                incoming_tier,
+            ))
+            .await
+            .expect("typed-state terminal usage should persist");
+        assert_eq!(
+            terminal.provider_service_tier().as_deref(),
+            expected_tier,
+            "state={state:?} should use only same-source metadata"
+        );
+        let list_item = repository
+            .list_usage_audits(&UsageAuditListQuery {
+                provider_name: Some(provider_name.clone()),
+                limit: Some(10),
+                newest_first: true,
+                ..UsageAuditListQuery::default()
+            })
+            .await
+            .expect("typed-state list should succeed")
+            .into_iter()
+            .find(|item| item.request_id == request_id)
+            .expect("typed-state row should be listed");
+        assert_eq!(list_item.provider_service_tier().as_deref(), expected_tier);
+        sqlx::query("DELETE FROM \"usage\" WHERE request_id = $1")
+            .bind(&request_id)
+            .execute(repository.pool())
+            .await
+            .expect("typed-state test usage should be removed");
+    }
+}
+
 #[tokio::test]
 #[ignore = "requires AETHER_TEST_DATABASE_URL and a populated PostgreSQL database"]
 async fn live_dashboard_combined_summary_matches_separate_queries() {
@@ -1536,6 +2008,8 @@ fn usage_sql_uses_json_null_placeholders_for_usage_payload_columns() {
         assert!(sql.contains("request_metadata->>'client_ip'"));
         assert!(sql.contains("'user_agent'"));
         assert!(sql.contains("request_metadata->>'user_agent'"));
+        assert!(sql.contains("request_metadata->>'requested_reasoning_effort'"));
+        assert!(sql.contains("request_metadata->>'provider_reasoning_effort'"));
         assert!(sql.contains("request_metadata->>'provider_service_tier'"));
         assert!(sql.contains("request_metadata->>'provider_actual_service_tier'"));
         assert!(sql.contains("AS client_family"));
@@ -1656,7 +2130,7 @@ fn usage_sql_settlement_pricing_snapshot_billing_values_use_authoritative_incomi
         "billing_actual_total_cost_usd",
     ] {
         let assignment = format!(
-            "{field} = COALESCE(\n    EXCLUDED.{field},\n    usage_settlement_snapshots.{field}\n  )"
+            "{field} = CASE WHEN $30 THEN EXCLUDED.{field} ELSE COALESCE(EXCLUDED.{field}, usage_settlement_snapshots.{field}) END"
         );
         assert!(
             sql.contains(assignment.as_str()),
@@ -1672,9 +2146,9 @@ fn usage_sql_settlement_pricing_snapshot_billing_values_use_authoritative_incomi
 #[test]
 fn usage_sql_upsert_recovers_missing_provider_links_after_billing_finalizes() {
     for assignment in [
-        "provider_id = CASE WHEN \"usage\".billing_status = 'pending' OR (\"usage\".provider_id IS NULL AND (\"usage\".provider_endpoint_id IS NULL OR \"usage\".provider_endpoint_id = EXCLUDED.provider_endpoint_id) AND (\"usage\".provider_api_key_id IS NULL OR \"usage\".provider_api_key_id = EXCLUDED.provider_api_key_id)) THEN COALESCE(EXCLUDED.provider_id, \"usage\".provider_id) ELSE \"usage\".provider_id END",
-        "provider_endpoint_id = CASE WHEN \"usage\".billing_status = 'pending' OR (\"usage\".provider_endpoint_id IS NULL AND (\"usage\".provider_id IS NULL OR \"usage\".provider_id = EXCLUDED.provider_id) AND (\"usage\".provider_api_key_id IS NULL OR \"usage\".provider_api_key_id = EXCLUDED.provider_api_key_id)) THEN COALESCE(EXCLUDED.provider_endpoint_id, \"usage\".provider_endpoint_id) ELSE \"usage\".provider_endpoint_id END",
-        "provider_api_key_id = CASE WHEN \"usage\".billing_status = 'pending' OR (\"usage\".provider_api_key_id IS NULL AND (\"usage\".provider_id IS NULL OR \"usage\".provider_id = EXCLUDED.provider_id) AND (\"usage\".provider_endpoint_id IS NULL OR \"usage\".provider_endpoint_id = EXCLUDED.provider_endpoint_id)) THEN COALESCE(EXCLUDED.provider_api_key_id, \"usage\".provider_api_key_id) ELSE \"usage\".provider_api_key_id END",
+        "provider_id = CASE WHEN (\"usage\".billing_status = 'pending' AND $61) OR (\"usage\".billing_status <> 'pending' AND \"usage\".provider_id IS NULL AND (\"usage\".provider_endpoint_id IS NULL OR \"usage\".provider_endpoint_id = EXCLUDED.provider_endpoint_id) AND (\"usage\".provider_api_key_id IS NULL OR \"usage\".provider_api_key_id = EXCLUDED.provider_api_key_id)) THEN COALESCE(EXCLUDED.provider_id, \"usage\".provider_id) ELSE \"usage\".provider_id END",
+        "provider_endpoint_id = CASE WHEN (\"usage\".billing_status = 'pending' AND $61) OR (\"usage\".billing_status <> 'pending' AND \"usage\".provider_endpoint_id IS NULL AND (\"usage\".provider_id IS NULL OR \"usage\".provider_id = EXCLUDED.provider_id) AND (\"usage\".provider_api_key_id IS NULL OR \"usage\".provider_api_key_id = EXCLUDED.provider_api_key_id)) THEN COALESCE(EXCLUDED.provider_endpoint_id, \"usage\".provider_endpoint_id) ELSE \"usage\".provider_endpoint_id END",
+        "provider_api_key_id = CASE WHEN (\"usage\".billing_status = 'pending' AND $61) OR (\"usage\".billing_status <> 'pending' AND \"usage\".provider_api_key_id IS NULL AND (\"usage\".provider_id IS NULL OR \"usage\".provider_id = EXCLUDED.provider_id) AND (\"usage\".provider_endpoint_id IS NULL OR \"usage\".provider_endpoint_id = EXCLUDED.provider_endpoint_id)) THEN COALESCE(EXCLUDED.provider_api_key_id, \"usage\".provider_api_key_id) ELSE \"usage\".provider_api_key_id END",
     ] {
         assert!(
             super::UPSERT_SQL.contains(assignment),
@@ -1761,6 +2235,56 @@ fn usage_sql_detached_body_flags_clear_inline_and_compressed_columns() {
         .contains("WHEN EXCLUDED.client_response_body_compressed IS NOT NULL OR $60 THEN NULL"));
 }
 
+#[test]
+fn usage_sql_capture_guard_covers_bodies_metadata_and_http_ref_tombstones() {
+    for assignment in [
+        "request_body = CASE WHEN \"usage\".billing_status = 'pending' AND $61",
+        "provider_request_body = CASE WHEN \"usage\".billing_status = 'pending' AND $61",
+        "response_body = CASE WHEN \"usage\".billing_status = 'pending' AND $61",
+        "client_response_body = CASE WHEN \"usage\".billing_status = 'pending' AND $61",
+        "request_metadata = CASE WHEN \"usage\".billing_status = 'pending' AND $61",
+    ] {
+        assert!(
+            super::UPSERT_SQL.contains(assignment),
+            "missing guard: {assignment}"
+        );
+    }
+    for assignment in [
+        "WHEN EXCLUDED.request_body_state = 'none' THEN NULL",
+        "WHEN EXCLUDED.provider_request_body_state = 'none' THEN NULL",
+        "WHEN EXCLUDED.response_body_state = 'none' THEN NULL",
+        "WHEN EXCLUDED.client_response_body_state = 'none' THEN NULL",
+    ] {
+        assert!(
+            super::UPSERT_USAGE_HTTP_AUDIT_SQL.contains(assignment),
+            "missing HTTP ref tombstone: {assignment}"
+        );
+    }
+    let source = include_str!("mod.rs");
+    assert!(source.contains(".bind(capture_update_allowed)"));
+    assert!(source.contains("if capture_update_allowed {"));
+}
+
+#[test]
+fn usage_sql_terminal_snapshots_replace_sparse_routing_and_settlement_facts() {
+    assert!(super::UPSERT_SQL.contains(
+        "target_model = CASE WHEN \"usage\".billing_status = 'pending' AND $61 THEN CASE WHEN EXCLUDED.status IN ('completed', 'failed', 'cancelled') THEN EXCLUDED.target_model"
+    ));
+    let routing = super::UPSERT_USAGE_ROUTING_SNAPSHOT_SQL;
+    assert!(routing.contains("candidate_id = CASE WHEN $14 THEN EXCLUDED.candidate_id"));
+    assert!(
+        routing.contains("selected_provider_id = CASE WHEN $14 THEN EXCLUDED.selected_provider_id")
+    );
+    let settlement = super::UPSERT_USAGE_SETTLEMENT_PRICING_SNAPSHOT_SQL;
+    assert!(settlement
+        .contains("settlement_snapshot = CASE WHEN $30 THEN EXCLUDED.settlement_snapshot"));
+    assert!(
+        settlement.contains("billing_dimensions = CASE WHEN $30 THEN EXCLUDED.billing_dimensions")
+    );
+    assert!(settlement.contains("rate_multiplier = CASE WHEN $30 THEN EXCLUDED.rate_multiplier"));
+    assert!(include_str!("mod.rs").contains(".bind(replace_existing)"));
+}
+
 #[test]
 fn usage_sql_clears_stale_failure_fields_for_non_failed_status_updates() {
     assert!(super::UPSERT_SQL.contains(
@@ -1884,6 +2408,123 @@ fn usage_body_capture_state_for_storage_preserves_unavailable_states() {
     );
 }
 
+#[test]
+fn explicit_none_capture_replaces_stale_request_body_facts_even_with_a_residual_body() {
+    let stale_body = json!({
+        "reasoning_effort": "xhigh",
+        "service_tier": "priority"
+    });
+
+    assert!(request_body_capture_replaces_derived_facts(
+        Some(&stale_body),
+        Some(UsageBodyCaptureState::None),
+    ));
+    assert!(request_body_capture_replaces_derived_facts(
+        Some(&stale_body),
+        None,
+    ));
+    assert!(request_body_capture_replaces_derived_facts(
+        Some(&stale_body),
+        Some(UsageBodyCaptureState::Disabled),
+    ));
+    assert!(request_body_capture_replaces_derived_facts(
+        None,
+        Some(UsageBodyCaptureState::Truncated),
+    ));
+    assert!(!request_body_capture_replaces_derived_facts(None, None,));
+}
+
+#[test]
+fn explicit_none_capture_drops_residual_body_ref_and_incoming_fast_metadata_before_storage() {
+    let mut usage = fast_clear_usage_record(
+        "req-none-residual",
+        "none-residual",
+        100,
+        true,
+        UsageBodyCaptureState::None,
+        Some("priority"),
+    );
+    usage.provider_request_body = Some(json!({
+        "model": "gpt-5",
+        "service_tier": "priority"
+    }));
+    usage.provider_request_body_ref =
+        Some("usage://request/req-none-residual/provider_request_body".to_string());
+    usage.request_metadata = Some(json!({
+        "trace_id": "trace-1",
+        "provider_service_tier": "priority",
+        "provider_reasoning_effort": "high",
+        "provider_cache_ttl_minutes": 30,
+        "provider_request_body_ref": "usage://request/req-none-residual/provider_request_body"
+    }));
+
+    let prepared = prepare_usage_upsert_context(&usage).expect("usage should prepare");
+    assert!(prepared.clear_provider_request_body);
+    assert!(!prepared.provider_request_body_storage.has_detached_blob());
+    assert_eq!(prepared.http_audit_refs.provider_request_body_ref, None);
+    assert_eq!(
+        prepared.request_metadata_value,
+        Some(json!({"trace_id": "trace-1"}))
+    );
+}
+
+#[test]
+fn request_body_fact_clear_keeps_unrelated_metadata_and_emits_an_empty_tombstone() {
+    let previous = json!({
+        "trace_id": "trace-1",
+        "requested_reasoning_effort": "xhigh",
+        "provider_reasoning_effort": "max",
+        "provider_service_tier": "priority",
+        "provider_cache_ttl_minutes": 30,
+        "provider_actual_service_tier": "default"
+    });
+
+    assert_eq!(
+        clear_previous_request_body_facts(Some(&previous), true, true),
+        json!({
+            "trace_id": "trace-1",
+            "provider_actual_service_tier": "default"
+        })
+    );
+    assert_eq!(
+        clear_previous_request_body_facts(
+            Some(&json!({"provider_service_tier": "priority"})),
+            false,
+            true,
+        ),
+        json!({})
+    );
+}
+
+#[test]
+fn terminal_capture_rejects_late_non_terminal_updates() {
+    assert!(usage_capture_update_allowed(None, "pending"));
+    assert!(usage_capture_update_allowed(
+        Some(("pending", "pending")),
+        "streaming",
+    ));
+    assert!(usage_capture_update_allowed(
+        Some(("failed", "pending")),
+        "completed",
+    ));
+    assert!(!usage_capture_update_allowed(
+        Some(("completed", "pending")),
+        "pending",
+    ));
+    assert!(!usage_capture_update_allowed(
+        Some(("failed", "pending")),
+        "streaming",
+    ));
+    assert!(!usage_capture_update_allowed(
+        Some(("streaming", "pending")),
+        "pending",
+    ));
+    assert!(!usage_capture_update_allowed(
+        Some(("completed", "settled")),
+        "completed",
+    ));
+}
+
 #[test]
 fn prepare_request_metadata_for_body_storage_strips_body_ref_compatibility_keys() {
     let detached = prepare_usage_body_storage(Some(&json!({

crates/aether-data/adapters/sqlite/src/usage.rs

@@ -156,22 +156,22 @@ INSERT INTO "usage" (
 ON CONFLICT (request_id) DO UPDATE SET
   user_id = excluded.user_id,
   api_key_id = excluded.api_key_id,
-  provider_name = excluded.provider_name,
-  model = excluded.model,
-  target_model = excluded.target_model,
-  provider_id = excluded.provider_id,
-  provider_endpoint_id = excluded.provider_endpoint_id,
-  provider_api_key_id = excluded.provider_api_key_id,
-  request_type = excluded.request_type,
-  api_format = excluded.api_format,
-  api_family = excluded.api_family,
-  endpoint_kind = excluded.endpoint_kind,
-  endpoint_api_format = excluded.endpoint_api_format,
-  provider_api_family = excluded.provider_api_family,
-  provider_endpoint_kind = excluded.provider_endpoint_kind,
-  has_format_conversion = excluded.has_format_conversion,
-  is_stream = excluded.is_stream,
-  upstream_is_stream = excluded.upstream_is_stream,
+  provider_name = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".provider_name ELSE excluded.provider_name END,
+  model = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".model ELSE excluded.model END,
+  target_model = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".target_model ELSE excluded.target_model END,
+  provider_id = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".provider_id ELSE excluded.provider_id END,
+  provider_endpoint_id = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".provider_endpoint_id ELSE excluded.provider_endpoint_id END,
+  provider_api_key_id = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".provider_api_key_id ELSE excluded.provider_api_key_id END,
+  request_type = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".request_type ELSE excluded.request_type END,
+  api_format = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".api_format ELSE excluded.api_format END,
+  api_family = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".api_family ELSE excluded.api_family END,
+  endpoint_kind = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".endpoint_kind ELSE excluded.endpoint_kind END,
+  endpoint_api_format = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".endpoint_api_format ELSE excluded.endpoint_api_format END,
+  provider_api_family = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".provider_api_family ELSE excluded.provider_api_family END,
+  provider_endpoint_kind = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".provider_endpoint_kind ELSE excluded.provider_endpoint_kind END,
+  has_format_conversion = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".has_format_conversion ELSE excluded.has_format_conversion END,
+  is_stream = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".is_stream ELSE excluded.is_stream END,
+  upstream_is_stream = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".upstream_is_stream ELSE excluded.upstream_is_stream END,
   input_tokens = CASE
         WHEN "usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming') THEN "usage".input_tokens
         ELSE excluded.input_tokens
@@ -255,15 +255,15 @@ ON CONFLICT (request_id) DO UPDATE SET
         WHEN "usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming') THEN "usage".billing_status
         ELSE excluded.billing_status
     END,
-  request_metadata = excluded.request_metadata,
-    candidate_id = COALESCE(excluded.candidate_id, "usage".candidate_id),
-    candidate_index = COALESCE(excluded.candidate_index, "usage".candidate_index),
-    key_name = COALESCE(excluded.key_name, "usage".key_name),
-  planner_kind = excluded.planner_kind,
-  route_family = excluded.route_family,
-  route_kind = excluded.route_kind,
-  execution_path = excluded.execution_path,
-  local_execution_runtime_miss_reason = excluded.local_execution_runtime_miss_reason,
+  request_metadata = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".request_metadata ELSE excluded.request_metadata END,
+  candidate_id = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".candidate_id WHEN excluded.status IN ('completed', 'failed', 'cancelled') THEN excluded.candidate_id ELSE COALESCE(excluded.candidate_id, "usage".candidate_id) END,
+  candidate_index = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".candidate_index WHEN excluded.status IN ('completed', 'failed', 'cancelled') THEN excluded.candidate_index ELSE COALESCE(excluded.candidate_index, "usage".candidate_index) END,
+  key_name = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".key_name WHEN excluded.status IN ('completed', 'failed', 'cancelled') THEN excluded.key_name ELSE COALESCE(excluded.key_name, "usage".key_name) END,
+  planner_kind = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".planner_kind ELSE excluded.planner_kind END,
+  route_family = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".route_family ELSE excluded.route_family END,
+  route_kind = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".route_kind ELSE excluded.route_kind END,
+  execution_path = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".execution_path ELSE excluded.execution_path END,
+  local_execution_runtime_miss_reason = CASE WHEN ("usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming')) OR ("usage".status = 'streaming' AND excluded.status = 'pending') THEN "usage".local_execution_runtime_miss_reason ELSE excluded.local_execution_runtime_miss_reason END,
   finalized_at = CASE
         WHEN "usage".status IN ('completed', 'failed', 'cancelled') AND excluded.status IN ('pending', 'streaming') THEN "usage".finalized_at
         ELSE excluded.finalized_at

crates/aether-data/adapters/sqlite/src/usage/tests.rs

@@ -6,6 +6,49 @@ use aether_data_contracts::repository::usage::{
     UsageReadRepository, UsageTimeSeriesGranularity, UsageWriteRepository,
 };
 
+#[test]
+fn sqlite_usage_upsert_guards_candidate_identity_metadata_and_routing_from_late_lifecycle() {
+    for field in [
+        "provider_name",
+        "model",
+        "target_model",
+        "provider_id",
+        "provider_endpoint_id",
+        "provider_api_key_id",
+        "request_type",
+        "api_format",
+        "api_family",
+        "endpoint_kind",
+        "endpoint_api_format",
+        "provider_api_family",
+        "provider_endpoint_kind",
+        "has_format_conversion",
+        "is_stream",
+        "upstream_is_stream",
+        "request_metadata",
+        "candidate_id",
+        "candidate_index",
+        "key_name",
+        "planner_kind",
+        "route_family",
+        "route_kind",
+        "execution_path",
+        "local_execution_runtime_miss_reason",
+    ] {
+        let assignment = format!("{field} = CASE WHEN (");
+        assert!(
+            super::UPSERT_USAGE_SQL.contains(&assignment),
+            "missing lifecycle guard for {field}"
+        );
+        assert!(
+            super::UPSERT_USAGE_SQL.contains(&format!("THEN \"usage\".{field}")),
+            "late lifecycle must preserve {field}"
+        );
+    }
+    assert!(super::UPSERT_USAGE_SQL
+        .contains("OR (\"usage\".status = 'streaming' AND excluded.status = 'pending')"));
+}
+
 #[tokio::test]
 async fn sqlite_provider_performance_can_skip_timeline() {
     let pool = sqlx::sqlite::SqlitePoolOptions::new()
@@ -139,7 +182,7 @@ async fn sqlite_usage_write_repository_does_not_regress_void_usage() {
         .expect("sqlite migrations should run");
     seed_stats_targets(&pool).await;
 
-    let repository = SqliteUsageWriteRepository::new(pool);
+    let repository = SqliteUsageWriteRepository::new(pool.clone());
     repository
         .upsert(sample_usage("request-1", "failed", "void", 1_000))
         .await
@@ -205,9 +248,11 @@ async fn sqlite_usage_write_repository_does_not_regress_terminal_usage_from_late
         .expect("sqlite migrations should run");
     seed_stats_targets(&pool).await;
 
-    let repository = SqliteUsageWriteRepository::new(pool);
+    let repository = SqliteUsageWriteRepository::new(pool.clone());
+    let mut terminal = sample_usage("request-1", "completed", "pending", 1_000);
+    terminal.request_metadata = Some(serde_json::json!({"trace_id": "terminal-trace"}));
     repository
-        .upsert(sample_usage("request-1", "completed", "pending", 1_000))
+        .upsert(terminal)
         .await
         .expect("terminal usage should upsert");
 
@@ -222,6 +267,30 @@ async fn sqlite_usage_write_repository_does_not_regress_terminal_usage_from_late
     late_streaming.response_time_ms = Some(9_999);
     late_streaming.first_byte_time_ms = Some(9_999);
     late_streaming.finalized_at_unix_secs = None;
+    late_streaming.provider_name = "Late Provider".to_string();
+    late_streaming.model = "late-model".to_string();
+    late_streaming.target_model = Some("late-target".to_string());
+    late_streaming.request_type = Some("late-request".to_string());
+    late_streaming.api_format = Some("late:api".to_string());
+    late_streaming.api_family = Some("late-family".to_string());
+    late_streaming.endpoint_kind = Some("late-endpoint".to_string());
+    late_streaming.endpoint_api_format = Some("late:endpoint".to_string());
+    late_streaming.provider_api_family = Some("late-provider-family".to_string());
+    late_streaming.provider_endpoint_kind = Some("late-provider-endpoint".to_string());
+    late_streaming.has_format_conversion = Some(false);
+    late_streaming.is_stream = Some(true);
+    late_streaming.candidate_id = Some("late-candidate".to_string());
+    late_streaming.candidate_index = Some(99);
+    late_streaming.key_name = Some("late-key".to_string());
+    late_streaming.planner_kind = Some("late-planner".to_string());
+    late_streaming.route_family = Some("late-route-family".to_string());
+    late_streaming.route_kind = Some("late-route-kind".to_string());
+    late_streaming.execution_path = Some("late-path".to_string());
+    late_streaming.local_execution_runtime_miss_reason = Some("late-miss".to_string());
+    late_streaming.request_metadata = Some(serde_json::json!({
+        "provider_service_tier": "priority",
+        "upstream_is_stream": true
+    }));
 
     let current = repository
         .upsert(late_streaming)
@@ -238,6 +307,91 @@ async fn sqlite_usage_write_repository_does_not_regress_terminal_usage_from_late
     assert_eq!(current.first_byte_time_ms, Some(12));
     assert_eq!(current.finalized_at_unix_secs, Some(1_000));
     assert_eq!(current.updated_at_unix_secs, 1_000);
+    assert_eq!(current.provider_name, "Provider One");
+    assert_eq!(current.model, "model-1");
+    assert_eq!(current.target_model.as_deref(), Some("target-model"));
+    assert_eq!(current.request_type.as_deref(), Some("chat"));
+    assert_eq!(current.api_format.as_deref(), Some("openai"));
+    assert!(current.has_format_conversion);
+    assert!(!current.is_stream);
+    assert_eq!(current.candidate_id.as_deref(), Some("candidate-1"));
+    assert_eq!(current.candidate_index, Some(1));
+    assert_eq!(current.key_name.as_deref(), Some("key-one"));
+    assert_eq!(current.planner_kind.as_deref(), Some("default"));
+    assert_eq!(current.route_family.as_deref(), Some("chat"));
+    assert_eq!(current.route_kind.as_deref(), Some("completion"));
+    assert_eq!(current.execution_path.as_deref(), Some("remote"));
+    assert_eq!(current.provider_service_tier(), None);
+    assert_eq!(
+        current
+            .request_metadata
+            .as_ref()
+            .and_then(|value| value.get("trace_id"))
+            .and_then(serde_json::Value::as_str),
+        Some("terminal-trace")
+    );
+
+    let listed = SqliteUsageReadRepository::new(pool)
+        .list_usage_audits(&UsageAuditListQuery {
+            limit: Some(10),
+            newest_first: true,
+            ..UsageAuditListQuery::default()
+        })
+        .await
+        .expect("usage list should load")
+        .into_iter()
+        .find(|item| item.request_id == "request-1")
+        .expect("terminal usage should be listed");
+    assert_eq!(listed.provider_name, "Provider One");
+    assert_eq!(listed.model, "model-1");
+    assert_eq!(listed.candidate_id.as_deref(), Some("candidate-1"));
+    assert_eq!(listed.provider_service_tier(), None);
+}
+
+#[tokio::test]
+async fn sqlite_usage_write_repository_allows_authoritative_completed_recovery() {
+    let pool = sqlx::sqlite::SqlitePoolOptions::new()
+        .max_connections(1)
+        .connect("sqlite::memory:")
+        .await
+        .expect("sqlite pool should connect");
+    run_migrations(&pool)
+        .await
+        .expect("sqlite migrations should run");
+    seed_stats_targets(&pool).await;
+
+    let repository = SqliteUsageWriteRepository::new(pool);
+    repository
+        .upsert(sample_usage("request-recovery", "failed", "void", 1_000))
+        .await
+        .expect("void failure should upsert");
+
+    let mut recovery = sample_usage("request-recovery", "completed", "pending", 1_001);
+    recovery.provider_name = "Recovered Provider".to_string();
+    recovery.model = "recovered-model".to_string();
+    recovery.target_model = Some("recovered-target".to_string());
+    recovery.api_format = Some("recovered:api".to_string());
+    recovery.candidate_id = Some("recovered-candidate".to_string());
+    recovery.request_metadata = Some(serde_json::json!({"provider_service_tier": "priority"}));
+    let recovered = repository
+        .upsert(recovery)
+        .await
+        .expect("completed recovery should upsert");
+
+    assert_eq!(recovered.status, "completed");
+    assert_eq!(recovered.billing_status, "pending");
+    assert_eq!(recovered.provider_name, "Recovered Provider");
+    assert_eq!(recovered.model, "recovered-model");
+    assert_eq!(recovered.target_model.as_deref(), Some("recovered-target"));
+    assert_eq!(recovered.api_format.as_deref(), Some("recovered:api"));
+    assert_eq!(
+        recovered.candidate_id.as_deref(),
+        Some("recovered-candidate")
+    );
+    assert_eq!(
+        recovered.provider_service_tier().as_deref(),
+        Some("priority")
+    );
 }
 
 #[tokio::test]
@@ -279,6 +433,51 @@ async fn sqlite_usage_write_repository_preserves_streaming_response_start_from_l
     assert_eq!(current.first_byte_time_ms, Some(12));
 }
 
+#[tokio::test]
+async fn sqlite_usage_write_repository_keeps_streaming_capture_from_late_pending() {
+    let pool = sqlx::sqlite::SqlitePoolOptions::new()
+        .max_connections(1)
+        .connect("sqlite::memory:")
+        .await
+        .expect("sqlite pool should connect");
+    run_migrations(&pool)
+        .await
+        .expect("sqlite migrations should run");
+    seed_stats_targets(&pool).await;
+
+    let repository = SqliteUsageWriteRepository::new(pool);
+    let mut streaming = sample_usage("request-streaming-capture", "streaming", "pending", 1_000);
+    streaming.request_metadata = Some(serde_json::json!({"trace_id": "streaming-final"}));
+    repository
+        .upsert(streaming)
+        .await
+        .expect("streaming usage should upsert");
+
+    let mut late_pending = sample_usage("request-streaming-capture", "pending", "pending", 1_001);
+    late_pending.provider_name = "Late Provider".to_string();
+    late_pending.model = "late-model".to_string();
+    late_pending.candidate_id = Some("late-candidate".to_string());
+    late_pending.request_metadata = Some(serde_json::json!({"provider_service_tier": "priority"}));
+    let current = repository
+        .upsert(late_pending)
+        .await
+        .expect("late pending usage should not regress streaming capture");
+
+    assert_eq!(current.status, "streaming");
+    assert_eq!(current.provider_name, "Provider One");
+    assert_eq!(current.model, "model-1");
+    assert_eq!(current.candidate_id.as_deref(), Some("candidate-1"));
+    assert_eq!(current.provider_service_tier(), None);
+    assert_eq!(
+        current
+            .request_metadata
+            .as_ref()
+            .and_then(|value| value.get("trace_id"))
+            .and_then(serde_json::Value::as_str),
+        Some("streaming-final")
+    );
+}
+
 #[tokio::test]
 async fn sqlite_usage_write_repository_cleans_stale_pending_requests() {
     let pool = sqlx::sqlite::SqlitePoolOptions::new()

crates/aether-data/contracts/src/repository/usage/mod.rs

@@ -6,9 +6,10 @@ pub use types::{
     extract_provider_actual_service_tier_from_response,
     extract_provider_cache_ttl_minutes_from_metadata, extract_provider_reasoning_effort_from_body,
     extract_provider_service_tier_from_body, normalize_provider_service_tier, parse_usage_body_ref,
-    resolve_provider_cache_ttl_minutes, usage_body_ref, usage_request_metadata_client_family,
-    ApiKeyLastUsedDelta, ManagementTokenCounterDelta, PendingUsageCleanupSummary,
-    ProviderApiKeyWindowUsageRequest, ProxyNodeCounterDelta, StoredProviderApiKeyUsageSummary,
+    resolve_provider_cache_ttl_minutes, resolve_provider_service_tier_from_request_capture,
+    usage_body_ref, usage_request_metadata_client_family, ApiKeyLastUsedDelta,
+    ManagementTokenCounterDelta, PendingUsageCleanupSummary, ProviderApiKeyWindowUsageRequest,
+    ProxyNodeCounterDelta, StoredProviderApiKeyUsageSummary,
     StoredProviderApiKeyWindowUsageSummary, StoredProviderUsageSummary, StoredProviderUsageWindow,
     StoredRequestUsageAudit, StoredUsageAuditAggregation, StoredUsageAuditSummary,
     StoredUsageBreakdownSummaryRow, StoredUsageCacheAffinityHitSummary,
@@ -34,5 +35,5 @@ pub use types::{
     UsageReadRepository, UsageRepository, UsageSettledCostSummaryQuery, UsageTimeSeriesGranularity,
     UsageTimeSeriesQuery, UsageWriteRepository, PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY,
     PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY, PROVIDER_REASONING_EFFORT_METADATA_KEY,
-    PROVIDER_SERVICE_TIER_METADATA_KEY,
+    PROVIDER_SERVICE_TIER_METADATA_KEY, REQUESTED_REASONING_EFFORT_METADATA_KEY,
 };

crates/aether-data/contracts/src/repository/usage/types.rs

@@ -3,6 +3,7 @@ use chrono::{DateTime, Utc};
 use serde_json::Value;
 
 pub const PROVIDER_REASONING_EFFORT_METADATA_KEY: &str = "provider_reasoning_effort";
+pub const REQUESTED_REASONING_EFFORT_METADATA_KEY: &str = "requested_reasoning_effort";
 pub const PROVIDER_SERVICE_TIER_METADATA_KEY: &str = "provider_service_tier";
 pub const PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY: &str = "provider_actual_service_tier";
 pub const PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY: &str = "provider_cache_ttl_minutes";
@@ -95,6 +96,89 @@ pub fn normalize_provider_service_tier(value: &str) -> Option<String> {
     Some(normalized)
 }
 
+/// Resolves a provider processing tier exclusively from the final upstream request.
+///
+/// A complete captured body is authoritative, including when it contains no tier. The metadata
+/// fallback is reserved for bodies that were stripped, externalized, or truncated after the tier
+/// had already been derived from that same request.
+pub fn resolve_provider_service_tier_from_request_capture(
+    provider_request_body: Option<&Value>,
+    provider_request_body_state: Option<UsageBodyCaptureState>,
+    request_metadata: Option<&Value>,
+) -> Option<String> {
+    if request_body_capture_is_authoritative(provider_request_body, provider_request_body_state) {
+        return extract_provider_service_tier_from_body(provider_request_body);
+    }
+
+    if !matches!(
+        provider_request_body_state,
+        Some(
+            UsageBodyCaptureState::Inline
+                | UsageBodyCaptureState::Reference
+                | UsageBodyCaptureState::Truncated
+                | UsageBodyCaptureState::Disabled
+                | UsageBodyCaptureState::Unavailable
+        )
+    ) {
+        return None;
+    }
+
+    request_metadata
+        .and_then(Value::as_object)
+        .and_then(|metadata| metadata.get(PROVIDER_SERVICE_TIER_METADATA_KEY))
+        .and_then(Value::as_str)
+        .and_then(normalize_provider_service_tier)
+}
+
+fn request_body_capture_is_authoritative(
+    request_body: Option<&Value>,
+    request_body_state: Option<UsageBodyCaptureState>,
+) -> bool {
+    let Some(request_body) = request_body else {
+        return false;
+    };
+    if matches!(
+        request_body_state,
+        Some(
+            UsageBodyCaptureState::None
+                | UsageBodyCaptureState::Truncated
+                | UsageBodyCaptureState::Disabled
+                | UsageBodyCaptureState::Unavailable
+        )
+    ) {
+        return false;
+    }
+
+    !request_body.as_object().is_some_and(|body| {
+        body.get("truncated").and_then(Value::as_bool) == Some(true)
+            && body.get("reason").and_then(Value::as_str) == Some("body_capture_limit_exceeded")
+    })
+}
+
+fn resolve_reasoning_effort_from_request_capture(
+    request_body: Option<&Value>,
+    request_body_state: Option<UsageBodyCaptureState>,
+    request_metadata: Option<&Value>,
+    metadata_key: &str,
+) -> Option<String> {
+    if request_body_capture_is_authoritative(request_body, request_body_state) {
+        return extract_provider_reasoning_effort_from_body(request_body);
+    }
+
+    // An explicit `none` marker describes the final capture attempt and must win over any stale
+    // inline body/metadata left by an earlier candidate. `None` (the absence of a marker) remains
+    // the legacy list-query shape, where metadata is the only available representation.
+    if request_body_state == Some(UsageBodyCaptureState::None) {
+        return None;
+    }
+
+    request_metadata
+        .and_then(Value::as_object)
+        .and_then(|metadata| metadata.get(metadata_key))
+        .and_then(Value::as_str)
+        .and_then(normalize_provider_reasoning_effort)
+}
+
 pub fn resolve_provider_cache_ttl_minutes(
     provider_api_format: Option<&str>,
     provider_model: Option<&str>,
@@ -511,33 +595,40 @@ impl StoredRequestUsageAudit {
     }
 
     pub fn provider_reasoning_effort(&self) -> Option<String> {
-        if self
-            .provider_request_body
-            .as_ref()
-            .and_then(Value::as_object)
-            .is_some()
-        {
-            return extract_provider_reasoning_effort_from_body(
-                self.provider_request_body.as_ref(),
-            );
-        }
+        resolve_reasoning_effort_from_request_capture(
+            self.provider_request_body.as_ref(),
+            self.provider_request_body_state,
+            self.request_metadata.as_ref(),
+            PROVIDER_REASONING_EFFORT_METADATA_KEY,
+        )
+    }
 
-        self.request_metadata_string(PROVIDER_REASONING_EFFORT_METADATA_KEY)
-            .and_then(normalize_provider_reasoning_effort)
+    pub fn requested_reasoning_effort(&self) -> Option<String> {
+        resolve_reasoning_effort_from_request_capture(
+            self.request_body.as_ref(),
+            self.request_body_state,
+            self.request_metadata.as_ref(),
+            REQUESTED_REASONING_EFFORT_METADATA_KEY,
+        )
     }
 
     pub fn provider_service_tier(&self) -> Option<String> {
-        if self
-            .provider_request_body
-            .as_ref()
-            .and_then(Value::as_object)
-            .is_some()
-        {
-            return extract_provider_service_tier_from_body(self.provider_request_body.as_ref());
-        }
-
-        self.request_metadata_string(PROVIDER_SERVICE_TIER_METADATA_KEY)
-            .and_then(normalize_provider_service_tier)
+        resolve_provider_service_tier_from_request_capture(
+            self.provider_request_body.as_ref(),
+            self.provider_request_body_state,
+            self.request_metadata.as_ref(),
+        )
+        .or_else(|| {
+            // Lightweight list queries intentionally omit bodies and typed capture state. Their
+            // request metadata was normalized before persistence and is the only available copy
+            // of the final provider-request fact.
+            if self.provider_request_body.is_none() && self.provider_request_body_state.is_none() {
+                self.request_metadata_string(PROVIDER_SERVICE_TIER_METADATA_KEY)
+                    .and_then(normalize_provider_service_tier)
+            } else {
+                None
+            }
+        })
     }
 
     pub fn provider_actual_service_tier(&self) -> Option<String> {
@@ -2684,6 +2775,126 @@ mod tests {
         assert_eq!(usage.provider_service_tier(), None);
     }
 
+    #[test]
+    fn requested_and_provider_reasoning_efforts_remain_independent() {
+        let mut usage = sample_usage();
+        usage.request_body = Some(json!({
+            "reasoning": { "effort": "XHigh" }
+        }));
+        usage.provider_request_body = Some(json!({
+            "reasoning_effort": "max"
+        }));
+        usage.request_metadata = Some(json!({
+            "requested_reasoning_effort": "low",
+            "provider_reasoning_effort": "medium"
+        }));
+
+        assert_eq!(usage.requested_reasoning_effort().as_deref(), Some("xhigh"));
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("max"));
+
+        usage.request_body = Some(json!({ "model": "gpt-5" }));
+        assert_eq!(usage.requested_reasoning_effort(), None);
+
+        usage.request_body = None;
+        assert_eq!(usage.requested_reasoning_effort().as_deref(), Some("low"));
+    }
+
+    #[test]
+    fn explicit_none_capture_ignores_stale_request_bodies_and_metadata() {
+        let mut usage = sample_usage();
+        usage.request_body = Some(json!({
+            "reasoning": { "effort": "xhigh" }
+        }));
+        usage.request_body_state = Some(UsageBodyCaptureState::None);
+        usage.provider_request_body = Some(json!({
+            "reasoning_effort": "max",
+            "service_tier": "priority"
+        }));
+        usage.provider_request_body_state = Some(UsageBodyCaptureState::None);
+        usage.request_metadata = Some(json!({
+            "requested_reasoning_effort": "high",
+            "provider_reasoning_effort": "medium",
+            "provider_service_tier": "priority"
+        }));
+
+        assert_eq!(usage.requested_reasoning_effort(), None);
+        assert_eq!(usage.provider_reasoning_effort(), None);
+        assert_eq!(usage.provider_service_tier(), None);
+    }
+
+    #[test]
+    fn reasoning_mapping_uses_derived_metadata_after_request_bodies_are_truncated() {
+        let truncated = json!({
+            "truncated": true,
+            "reason": "body_capture_limit_exceeded"
+        });
+        let mut usage = sample_usage();
+        usage.request_body = Some(truncated.clone());
+        usage.request_body_state = Some(UsageBodyCaptureState::Truncated);
+        usage.provider_request_body = Some(truncated);
+        usage.provider_request_body_state = Some(UsageBodyCaptureState::Truncated);
+        usage.request_metadata = Some(json!({
+            "requested_reasoning_effort": "xhigh",
+            "provider_reasoning_effort": "max"
+        }));
+
+        assert_eq!(usage.requested_reasoning_effort().as_deref(), Some("xhigh"));
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("max"));
+    }
+
+    #[test]
+    fn provider_service_tier_does_not_infer_fast_from_response_or_stale_metadata() {
+        let mut usage = sample_usage();
+        usage.provider_request_body = Some(json!({
+            "model": "gpt-5"
+        }));
+        usage.request_metadata = Some(json!({
+            "provider_service_tier": "priority",
+            "provider_actual_service_tier": "priority"
+        }));
+        usage.response_body = Some(json!({
+            "service_tier": "priority"
+        }));
+
+        assert_eq!(usage.provider_service_tier(), None);
+    }
+
+    #[test]
+    fn provider_service_tier_uses_derived_metadata_after_final_body_is_stripped() {
+        let mut usage = sample_usage();
+        usage.provider_request_body = None;
+        usage.provider_request_body_state = Some(UsageBodyCaptureState::Disabled);
+        usage.request_metadata = Some(json!({
+            "provider_service_tier": "priority"
+        }));
+        usage.response_body = Some(json!({
+            "service_tier": "default"
+        }));
+
+        assert_eq!(usage.provider_service_tier().as_deref(), Some("priority"));
+    }
+
+    #[test]
+    fn provider_service_tier_uses_request_derived_metadata_after_body_truncation() {
+        let mut usage = sample_usage();
+        usage.provider_request_body = Some(json!({
+            "truncated": true,
+            "reason": "body_capture_limit_exceeded",
+            "max_bytes": 128,
+            "source_bytes": 4096,
+            "value_kind": "object"
+        }));
+        usage.provider_request_body_state = Some(UsageBodyCaptureState::Truncated);
+        usage.request_metadata = Some(json!({
+            "provider_service_tier": "priority"
+        }));
+        usage.response_body = Some(json!({
+            "service_tier": "default"
+        }));
+
+        assert_eq!(usage.provider_service_tier().as_deref(), Some("priority"));
+    }
+
     #[test]
     fn provider_cache_ttl_uses_final_openai_contract_then_preserved_metadata() {
         let mut usage = sample_usage();

crates/aether-data/runtime/src/repository/usage/memory.rs

@@ -14,14 +14,16 @@ use aether_data_contracts::repository::usage::{
     StoredUsageProviderPerformanceTimelineRow, StoredUsageSettledCostSummary,
     StoredUsageTimeSeriesBucket, StoredUsageUserTotals, UsageAuditAggregationGroupBy,
     UsageAuditAggregationQuery, UsageAuditKeywordSearchQuery, UsageAuditSummaryQuery,
-    UsageBodyField, UsageBreakdownGroupBy, UsageBreakdownSummaryQuery,
+    UsageBodyCaptureState, UsageBodyField, UsageBreakdownGroupBy, UsageBreakdownSummaryQuery,
     UsageCacheAffinityHitSummaryQuery, UsageCacheAffinityIntervalGroupBy,
     UsageCacheAffinityIntervalQuery, UsageCacheHitSummaryQuery, UsageCostSavingsSummaryQuery,
     UsageDashboardDailyBreakdownQuery, UsageDashboardProviderCountsQuery,
     UsageDashboardSummaryQuery, UsageErrorDistributionQuery, UsageLeaderboardGroupBy,
     UsageLeaderboardQuery, UsageMonitoringErrorCountQuery, UsageMonitoringErrorListQuery,
     UsagePerformancePercentilesQuery, UsageProviderPerformanceQuery, UsageSettledCostSummaryQuery,
-    UsageTimeSeriesGranularity, UsageTimeSeriesQuery,
+    UsageTimeSeriesGranularity, UsageTimeSeriesQuery, PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY,
+    PROVIDER_REASONING_EFFORT_METADATA_KEY, PROVIDER_SERVICE_TIER_METADATA_KEY,
+    REQUESTED_REASONING_EFFORT_METADATA_KEY,
 };
 use async_trait::async_trait;
 use chrono::Utc;
@@ -2703,10 +2705,14 @@ fn hydrate_client_family(item: &mut StoredRequestUsageAudit) {
 fn persisted_usage_body_ref(
     incoming_ref: Option<&str>,
     incoming_body: Option<&Value>,
+    incoming_state: Option<UsageBodyCaptureState>,
     _metadata: Option<&Value>,
     existing: Option<&StoredRequestUsageAudit>,
     field: UsageBodyField,
 ) -> Option<String> {
+    if incoming_state == Some(UsageBodyCaptureState::None) {
+        return None;
+    }
     if incoming_body.is_some() {
         return None;
     }
@@ -2724,6 +2730,78 @@ fn persisted_usage_body_ref(
         })
 }
 
+fn request_body_capture_replaces_derived_facts(
+    request_body: Option<&Value>,
+    request_body_state: Option<UsageBodyCaptureState>,
+) -> bool {
+    if request_body_state.is_some() {
+        return true;
+    }
+    let Some(request_body) = request_body else {
+        return false;
+    };
+    !request_body.as_object().is_some_and(|body| {
+        body.get("truncated").and_then(Value::as_bool) == Some(true)
+            && body.get("reason").and_then(Value::as_str) == Some("body_capture_limit_exceeded")
+    })
+}
+
+fn clear_request_body_facts(
+    metadata: Option<&Value>,
+    clear_client_request_body_facts: bool,
+    clear_provider_request_body_facts: bool,
+) -> Value {
+    let mut metadata = metadata
+        .and_then(Value::as_object)
+        .cloned()
+        .unwrap_or_default();
+    if clear_client_request_body_facts {
+        metadata.remove(REQUESTED_REASONING_EFFORT_METADATA_KEY);
+        metadata.remove("request_body_ref");
+    }
+    if clear_provider_request_body_facts {
+        metadata.remove(PROVIDER_REASONING_EFFORT_METADATA_KEY);
+        metadata.remove(PROVIDER_SERVICE_TIER_METADATA_KEY);
+        metadata.remove(PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY);
+        metadata.remove("provider_request_body_ref");
+    }
+    Value::Object(metadata)
+}
+
+fn retain_previous_request_audit_metadata(
+    metadata: Option<&Value>,
+    preserve_client_request_body_facts: bool,
+) -> Value {
+    let Some(metadata) = metadata.and_then(Value::as_object) else {
+        return Value::Object(serde_json::Map::new());
+    };
+    let mut retained = serde_json::Map::new();
+    for key in [
+        "trace_id",
+        "client_ip",
+        "user_agent",
+        "client_family",
+        "client_requested_stream",
+        "client_session_affinity",
+        "api_key_is_standalone",
+        "request_path",
+        "request_query_string",
+        "request_path_and_query",
+    ] {
+        if let Some(value) = metadata.get(key) {
+            retained.insert(key.to_string(), value.clone());
+        }
+    }
+    if preserve_client_request_body_facts {
+        for key in [REQUESTED_REASONING_EFFORT_METADATA_KEY, "request_body_ref"] {
+            if let Some(value) = metadata.get(key) {
+                retained.insert(key.to_string(), value.clone());
+            }
+        }
+    }
+    Value::Object(retained)
+}
+
 fn merge_usage_status_code(
     existing: Option<&StoredRequestUsageAudit>,
     incoming_status: &str,
@@ -2789,14 +2867,57 @@ impl UsageWriteRepository for InMemoryUsageReadRepository {
             return Ok(existing.expect("existing usage should be present").clone());
         }
 
-        let request_metadata = usage.request_metadata.clone().or_else(|| {
-            existing
-                .as_ref()
-                .and_then(|existing| existing.request_metadata.clone())
+        let replace_client_request_body_facts = request_body_capture_replaces_derived_facts(
+            usage.request_body.as_ref(),
+            usage.request_body_state,
+        );
+        let replace_provider_request_body_facts = request_body_capture_replaces_derived_facts(
+            usage.provider_request_body.as_ref(),
+            usage.provider_request_body_state,
+        );
+        let clear_request_body = usage.request_body_state == Some(UsageBodyCaptureState::None);
+        let clear_provider_request_body =
+            usage.provider_request_body_state == Some(UsageBodyCaptureState::None);
+        let clear_response_body = usage.response_body_state == Some(UsageBodyCaptureState::None);
+        let clear_client_response_body =
+            usage.client_response_body_state == Some(UsageBodyCaptureState::None);
+        let replace_routing_snapshot = usage_status_is_finalized(&usage.status);
+        let mut incoming_request_metadata = usage.request_metadata.clone();
+        if incoming_request_metadata.is_some()
+            && (clear_request_body || clear_provider_request_body)
+        {
+            incoming_request_metadata = Some(clear_request_body_facts(
+                incoming_request_metadata.as_ref(),
+                clear_request_body,
+                clear_provider_request_body,
+            ));
+        }
+        let request_metadata = incoming_request_metadata.or_else(|| {
+            if replace_routing_snapshot {
+                Some(retain_previous_request_audit_metadata(
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.request_metadata.as_ref()),
+                    !replace_client_request_body_facts,
+                ))
+            } else if replace_client_request_body_facts || replace_provider_request_body_facts {
+                Some(clear_request_body_facts(
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.request_metadata.as_ref()),
+                    replace_client_request_body_facts,
+                    replace_provider_request_body_facts,
+                ))
+            } else {
+                existing
+                    .as_ref()
+                    .and_then(|existing| existing.request_metadata.clone())
+            }
         });
         let request_body_ref = persisted_usage_body_ref(
             usage.request_body_ref.as_deref(),
             usage.request_body.as_ref(),
+            usage.request_body_state,
             request_metadata.as_ref(),
             existing.as_ref(),
             UsageBodyField::RequestBody,
@@ -2804,6 +2925,7 @@ impl UsageWriteRepository for InMemoryUsageReadRepository {
         let provider_request_body_ref = persisted_usage_body_ref(
             usage.provider_request_body_ref.as_deref(),
             usage.provider_request_body.as_ref(),
+            usage.provider_request_body_state,
             request_metadata.as_ref(),
             existing.as_ref(),
             UsageBodyField::ProviderRequestBody,
@@ -2811,6 +2933,7 @@ impl UsageWriteRepository for InMemoryUsageReadRepository {
         let response_body_ref = persisted_usage_body_ref(
             usage.response_body_ref.as_deref(),
             usage.response_body.as_ref(),
+            usage.response_body_state,
             request_metadata.as_ref(),
             existing.as_ref(),
             UsageBodyField::ResponseBody,
@@ -2818,10 +2941,34 @@ impl UsageWriteRepository for InMemoryUsageReadRepository {
         let client_response_body_ref = persisted_usage_body_ref(
             usage.client_response_body_ref.as_deref(),
             usage.client_response_body.as_ref(),
+            usage.client_response_body_state,
             request_metadata.as_ref(),
             existing.as_ref(),
             UsageBodyField::ClientResponseBody,
         );
+        if clear_request_body
+            || clear_provider_request_body
+            || clear_response_body
+            || clear_client_response_body
+        {
+            let mut detached_bodies = self.detached_bodies.write().expect("usage repository lock");
+            for (clear, field) in [
+                (clear_request_body, UsageBodyField::RequestBody),
+                (
+                    clear_provider_request_body,
+                    UsageBodyField::ProviderRequestBody,
+                ),
+                (clear_response_body, UsageBodyField::ResponseBody),
+                (
+                    clear_client_response_body,
+                    UsageBodyField::ClientResponseBody,
+                ),
+            ] {
+                if clear {
+                    detached_bodies.remove(&usage_body_ref(&usage.request_id, field));
+                }
+            }
+        }
         let stored = StoredRequestUsageAudit {
             id: existing
                 .as_ref()
@@ -2935,11 +3082,15 @@ impl UsageWriteRepository for InMemoryUsageReadRepository {
                     .as_ref()
                     .and_then(|existing| existing.request_headers.clone())
             }),
-            request_body: usage.request_body.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.request_body.clone())
-            }),
+            request_body: if clear_request_body {
+                None
+            } else {
+                usage.request_body.or_else(|| {
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.request_body.clone())
+                })
+            },
             request_body_ref,
             request_body_state: usage.request_body_state.or_else(|| {
                 existing
@@ -2951,11 +3102,15 @@ impl UsageWriteRepository for InMemoryUsageReadRepository {
                     .as_ref()
                     .and_then(|existing| existing.provider_request_headers.clone())
             }),
-            provider_request_body: usage.provider_request_body.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.provider_request_body.clone())
-            }),
+            provider_request_body: if clear_provider_request_body {
+                None
+            } else {
+                usage.provider_request_body.or_else(|| {
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.provider_request_body.clone())
+                })
+            },
             provider_request_body_ref,
             provider_request_body_state: usage.provider_request_body_state.or_else(|| {
                 existing
@@ -2967,11 +3122,15 @@ impl UsageWriteRepository for InMemoryUsageReadRepository {
                     .as_ref()
                     .and_then(|existing| existing.response_headers.clone())
             }),
-            response_body: usage.response_body.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.response_body.clone())
-            }),
+            response_body: if clear_response_body {
+                None
+            } else {
+                usage.response_body.or_else(|| {
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.response_body.clone())
+                })
+            },
             response_body_ref,
             response_body_state: usage.response_body_state.or_else(|| {
                 existing
@@ -2983,61 +3142,95 @@ impl UsageWriteRepository for InMemoryUsageReadRepository {
                     .as_ref()
                     .and_then(|existing| existing.client_response_headers.clone())
             }),
-            client_response_body: usage.client_response_body.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.client_response_body.clone())
-            }),
+            client_response_body: if clear_client_response_body {
+                None
+            } else {
+                usage.client_response_body.or_else(|| {
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.client_response_body.clone())
+                })
+            },
             client_response_body_ref,
             client_response_body_state: usage.client_response_body_state.or_else(|| {
                 existing
                     .as_ref()
                     .and_then(|existing| existing.client_response_body_state)
             }),
-            candidate_id: usage.candidate_id.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.routing_candidate_id().map(ToOwned::to_owned))
-            }),
-            candidate_index: usage.candidate_index.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.routing_candidate_index())
-            }),
-            key_name: usage.key_name.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.routing_key_name().map(ToOwned::to_owned))
-            }),
-            planner_kind: usage.planner_kind.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.routing_planner_kind().map(ToOwned::to_owned))
-            }),
-            route_family: usage.route_family.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.routing_route_family().map(ToOwned::to_owned))
-            }),
-            route_kind: usage.route_kind.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.routing_route_kind().map(ToOwned::to_owned))
-            }),
-            execution_path: usage.execution_path.or_else(|| {
-                existing
-                    .as_ref()
-                    .and_then(|existing| existing.routing_execution_path().map(ToOwned::to_owned))
-            }),
-            local_execution_runtime_miss_reason: usage.local_execution_runtime_miss_reason.or_else(
-                || {
+            candidate_id: if replace_routing_snapshot {
+                usage.candidate_id
+            } else {
+                usage.candidate_id.or_else(|| {
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.routing_candidate_id().map(ToOwned::to_owned))
+                })
+            },
+            candidate_index: if replace_routing_snapshot {
+                usage.candidate_index
+            } else {
+                usage.candidate_index.or_else(|| {
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.routing_candidate_index())
+                })
+            },
+            key_name: if replace_routing_snapshot {
+                usage.key_name
+            } else {
+                usage.key_name.or_else(|| {
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.routing_key_name().map(ToOwned::to_owned))
+                })
+            },
+            planner_kind: if replace_routing_snapshot {
+                usage.planner_kind
+            } else {
+                usage.planner_kind.or_else(|| {
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.routing_planner_kind().map(ToOwned::to_owned))
+                })
+            },
+            route_family: if replace_routing_snapshot {
+                usage.route_family
+            } else {
+                usage.route_family.or_else(|| {
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.routing_route_family().map(ToOwned::to_owned))
+                })
+            },
+            route_kind: if replace_routing_snapshot {
+                usage.route_kind
+            } else {
+                usage.route_kind.or_else(|| {
+                    existing
+                        .as_ref()
+                        .and_then(|existing| existing.routing_route_kind().map(ToOwned::to_owned))
+                })
+            },
+            execution_path: if replace_routing_snapshot {
+                usage.execution_path
+            } else {
+                usage.execution_path.or_else(|| {
+                    existing.as_ref().and_then(|existing| {
+                        existing.routing_execution_path().map(ToOwned::to_owned)
+                    })
+                })
+            },
+            local_execution_runtime_miss_reason: if replace_routing_snapshot {
+                usage.local_execution_runtime_miss_reason
+            } else {
+                usage.local_execution_runtime_miss_reason.or_else(|| {
                     existing.as_ref().and_then(|existing| {
                         existing
                             .routing_local_execution_runtime_miss_reason()
                             .map(ToOwned::to_owned)
                     })
-                },
-            ),
+                })
+            },
             client_family: usage_request_metadata_client_family(request_metadata.as_ref())
                 .map(ToOwned::to_owned)
                 .or_else(|| {

crates/aether-data/runtime/src/repository/usage/memory/tests.rs

@@ -15,7 +15,7 @@ use crate::repository::usage::{
 };
 use aether_data_contracts::repository::usage::{
     usage_body_ref, ProviderApiKeyWindowUsageRequest, UsageAuditAggregationGroupBy,
-    UsageAuditAggregationQuery, UsageBodyField, UsageDashboardSummaryQuery,
+    UsageAuditAggregationQuery, UsageBodyCaptureState, UsageBodyField, UsageDashboardSummaryQuery,
     UsageLeaderboardGroupBy, UsageLeaderboardQuery, UsageProviderPerformanceQuery,
     UsageTimeSeriesGranularity,
 };
@@ -135,6 +135,127 @@ fn sample_upsert_usage_record(request_id: &str) -> UpsertUsageRecord {
     }
 }
 
+#[tokio::test]
+async fn upsert_uses_typed_provider_capture_as_the_fast_fact_snapshot() {
+    for (name, state, incoming_tier, expected_tier) in [
+        (
+            "disabled-clear",
+            UsageBodyCaptureState::Disabled,
+            None,
+            None,
+        ),
+        (
+            "truncated-clear",
+            UsageBodyCaptureState::Truncated,
+            None,
+            None,
+        ),
+        (
+            "unavailable-clear",
+            UsageBodyCaptureState::Unavailable,
+            None,
+            None,
+        ),
+        (
+            "disabled-preserve",
+            UsageBodyCaptureState::Disabled,
+            Some("priority"),
+            Some("priority"),
+        ),
+        (
+            "truncated-preserve",
+            UsageBodyCaptureState::Truncated,
+            Some("priority"),
+            Some("priority"),
+        ),
+        (
+            "none-clears-residual",
+            UsageBodyCaptureState::None,
+            Some("priority"),
+            None,
+        ),
+    ] {
+        let request_id = format!("req-memory-fast-{name}");
+        let repository = InMemoryUsageReadRepository::default();
+        let mut pending = sample_upsert_usage_record(&request_id);
+        pending.provider_request_body = Some(json!({
+            "model": "gpt-5",
+            "service_tier": "priority"
+        }));
+        pending.provider_request_body_state = Some(UsageBodyCaptureState::Inline);
+        pending.request_metadata = Some(json!({"provider_service_tier": "priority"}));
+        pending.target_model = Some("candidate-a-target".to_string());
+        pending.candidate_id = Some("candidate-a".to_string());
+        pending.candidate_index = Some(1);
+        pending.key_name = Some("key-a".to_string());
+        pending.planner_kind = Some("planner-a".to_string());
+        pending.route_family = Some("route-family-a".to_string());
+        pending.route_kind = Some("route-kind-a".to_string());
+        pending.execution_path = Some("path-a".to_string());
+        let pending = repository
+            .upsert(pending)
+            .await
+            .expect("pending usage should upsert");
+        assert_eq!(pending.provider_service_tier().as_deref(), Some("priority"));
+
+        let mut terminal = sample_upsert_usage_record(&request_id);
+        terminal.status = "completed".to_string();
+        terminal.provider_request_body_state = Some(state);
+        terminal.target_model = None;
+        terminal.request_metadata =
+            incoming_tier.map(|tier| json!({"provider_service_tier": tier}));
+        terminal.updated_at_unix_secs += 1;
+        terminal.finalized_at_unix_secs = Some(terminal.updated_at_unix_secs);
+        if state == UsageBodyCaptureState::None {
+            terminal.provider_request_body = Some(json!({
+                "model": "stale-model",
+                "service_tier": "priority"
+            }));
+            terminal.provider_request_body_ref = Some(usage_body_ref(
+                &request_id,
+                UsageBodyField::ProviderRequestBody,
+            ));
+        }
+        let terminal = repository
+            .upsert(terminal)
+            .await
+            .expect("terminal usage should upsert");
+        assert_eq!(
+            terminal.provider_service_tier().as_deref(),
+            expected_tier,
+            "state={state:?} must use only incoming facts"
+        );
+        assert_eq!(terminal.target_model, None);
+        assert_eq!(terminal.candidate_id, None);
+        assert_eq!(terminal.candidate_index, None);
+        assert_eq!(terminal.key_name, None);
+        assert_eq!(terminal.planner_kind, None);
+        assert_eq!(terminal.route_family, None);
+        assert_eq!(terminal.route_kind, None);
+        assert_eq!(terminal.execution_path, None);
+        if state == UsageBodyCaptureState::None {
+            assert_eq!(terminal.provider_request_body, None);
+            assert_eq!(terminal.provider_request_body_ref, None);
+
+            let mut late = sample_upsert_usage_record(&request_id);
+            late.provider_request_body = Some(json!({
+                "model": "late-model",
+                "service_tier": "priority"
+            }));
+            late.provider_request_body_state = Some(UsageBodyCaptureState::Inline);
+            late.request_metadata = Some(json!({"provider_service_tier": "priority"}));
+            late.updated_at_unix_secs += 2;
+            let late = repository
+                .upsert(late)
+                .await
+                .expect("late pending usage should not regress terminal capture");
+            assert_eq!(late.provider_service_tier(), None);
+            assert_eq!(late.provider_request_body, None);
+            assert_eq!(late.provider_request_body_ref, None);
+        }
+    }
+}
+
 #[tokio::test]
 async fn finds_usage_by_request_id() {
     let repository = InMemoryUsageReadRepository::seed(vec![

crates/aether-usage/runtime/src/record.rs

@@ -2,7 +2,10 @@ use aether_data_contracts::repository::usage::UpsertUsageRecord;
 use aether_data_contracts::DataLayerError;
 
 use crate::request_metadata::{
-    attach_provider_request_body_metadata, sanitize_usage_request_metadata,
+    attach_client_request_body_metadata, attach_provider_request_body_metadata,
+    clear_client_request_body_metadata, clear_provider_request_body_metadata,
+    request_body_derived_facts_action, sanitize_usage_request_metadata,
+    RequestBodyDerivedFactsAction,
 };
 use crate::{UsageEvent, UsageEventType};
 
@@ -38,15 +41,42 @@ pub fn build_upsert_usage_record_from_event(
         }
     };
     let mut data = event.data.clone();
-    data.request_metadata = attach_provider_request_body_metadata(
-        data.request_metadata,
-        data.endpoint_api_format
-            .as_deref()
-            .or(data.api_format.as_deref()),
-        data.target_model.as_deref().or(Some(data.model.as_str())),
-        Some(data.model.as_str()),
+    // Request-derived facts are captured before body capture policy is applied. Do not let a
+    // truncation/disabled placeholder clear those facts while converting the queued event into a
+    // database record. Inline (or ref-loaded) bodies remain authoritative and may clear stale
+    // metadata when the final upstream request no longer contains a value.
+    match request_body_derived_facts_action(data.request_body.as_ref(), data.request_body_state) {
+        RequestBodyDerivedFactsAction::Refresh => {
+            data.request_metadata = attach_client_request_body_metadata(
+                data.request_metadata,
+                data.request_body.as_ref(),
+            );
+        }
+        RequestBodyDerivedFactsAction::Clear => {
+            data.request_metadata = clear_client_request_body_metadata(data.request_metadata);
+        }
+        RequestBodyDerivedFactsAction::Preserve => {}
+    }
+    match request_body_derived_facts_action(
         data.provider_request_body.as_ref(),
-    );
+        data.provider_request_body_state,
+    ) {
+        RequestBodyDerivedFactsAction::Refresh => {
+            data.request_metadata = attach_provider_request_body_metadata(
+                data.request_metadata,
+                data.endpoint_api_format
+                    .as_deref()
+                    .or(data.api_format.as_deref()),
+                data.target_model.as_deref().or(Some(data.model.as_str())),
+                Some(data.model.as_str()),
+                data.provider_request_body.as_ref(),
+            );
+        }
+        RequestBodyDerivedFactsAction::Clear => {
+            data.request_metadata = clear_provider_request_body_metadata(data.request_metadata);
+        }
+        RequestBodyDerivedFactsAction::Preserve => {}
+    }
     let now_unix_secs = event.timestamp_ms / 1_000;
 
     Ok(UpsertUsageRecord {
@@ -165,6 +195,8 @@ fn empty_to_none(value: Option<String>) -> Option<String> {
 
 #[cfg(test)]
 mod tests {
+    use aether_data_contracts::repository::usage::UsageBodyCaptureState;
+
     use crate::{UsageEvent, UsageEventData, UsageEventType};
 
     use super::build_upsert_usage_record_from_event;
@@ -186,6 +218,9 @@ mod tests {
                 output_tokens: Some(20),
                 total_tokens: Some(30),
                 status_code: Some(200),
+                request_body: Some(serde_json::json!({
+                    "reasoning": { "effort": "xhigh" }
+                })),
                 provider_request_body: Some(serde_json::json!({
                     "reasoning": { "effort": "max" },
                     "service_tier": "priority"
@@ -202,6 +237,14 @@ mod tests {
         assert_eq!(record.status, "completed");
         assert_eq!(record.billing_status, "pending");
         assert_eq!(record.total_tokens, Some(30));
+        assert_eq!(
+            record
+                .request_metadata
+                .as_ref()
+                .and_then(|value| value.get("requested_reasoning_effort"))
+                .and_then(serde_json::Value::as_str),
+            Some("xhigh")
+        );
         assert_eq!(
             record
                 .request_metadata
@@ -229,6 +272,139 @@ mod tests {
         assert_eq!(record.finalized_at_unix_secs, Some(1_700_000_000));
     }
 
+    #[test]
+    fn truncated_request_placeholders_preserve_pre_capture_request_facts() {
+        let truncated = serde_json::json!({
+            "truncated": true,
+            "reason": "body_capture_limit_exceeded",
+            "max_bytes": 128,
+            "source_bytes": 2048,
+            "value_kind": "object"
+        });
+        let record = build_upsert_usage_record_from_event(&UsageEvent {
+            event_type: UsageEventType::Completed,
+            request_id: "req-truncated-request-facts".to_string(),
+            timestamp_ms: 1_700_000_000_000,
+            data: UsageEventData {
+                provider_name: "OpenAI".to_string(),
+                model: "gpt-5".to_string(),
+                api_format: Some("openai:responses".to_string()),
+                endpoint_api_format: Some("openai:responses".to_string()),
+                request_body: Some(truncated.clone()),
+                request_body_state: Some(UsageBodyCaptureState::Truncated),
+                provider_request_body: Some(truncated),
+                provider_request_body_state: Some(UsageBodyCaptureState::Truncated),
+                request_metadata: Some(serde_json::json!({
+                    "requested_reasoning_effort": "xhigh",
+                    "provider_reasoning_effort": "max",
+                    "provider_service_tier": "priority"
+                })),
+                ..UsageEventData::default()
+            },
+        })
+        .expect("record should build");
+
+        let metadata = record
+            .request_metadata
+            .as_ref()
+            .expect("derived request facts should remain");
+        assert_eq!(metadata["requested_reasoning_effort"], "xhigh");
+        assert_eq!(metadata["provider_reasoning_effort"], "max");
+        assert_eq!(metadata["provider_service_tier"], "priority");
+    }
+
+    #[test]
+    fn disabled_reference_and_unavailable_bodies_preserve_pre_capture_request_facts() {
+        for state in [
+            UsageBodyCaptureState::Disabled,
+            UsageBodyCaptureState::Reference,
+            UsageBodyCaptureState::Unavailable,
+        ] {
+            let record = build_upsert_usage_record_from_event(&UsageEvent {
+                event_type: UsageEventType::Completed,
+                request_id: format!("req-preserved-{state:?}"),
+                timestamp_ms: 1_700_000_000_000,
+                data: UsageEventData {
+                    provider_name: "OpenAI".to_string(),
+                    model: "gpt-5".to_string(),
+                    request_body_state: Some(state),
+                    provider_request_body_state: Some(state),
+                    request_metadata: Some(serde_json::json!({
+                        "requested_reasoning_effort": "xhigh",
+                        "provider_reasoning_effort": "max",
+                        "provider_service_tier": "priority"
+                    })),
+                    ..UsageEventData::default()
+                },
+            })
+            .expect("record should build");
+
+            let metadata = record
+                .request_metadata
+                .as_ref()
+                .expect("derived request facts should remain");
+            assert_eq!(metadata["requested_reasoning_effort"], "xhigh");
+            assert_eq!(metadata["provider_reasoning_effort"], "max");
+            assert_eq!(metadata["provider_service_tier"], "priority");
+        }
+    }
+
+    #[test]
+    fn missing_or_complete_factless_final_bodies_clear_stale_request_facts() {
+        let cases = [
+            ("typed-missing", Some(UsageBodyCaptureState::None), None),
+            (
+                "typed-missing-with-stale-body",
+                Some(UsageBodyCaptureState::None),
+                Some(serde_json::json!({
+                    "reasoning_effort": "xhigh",
+                    "service_tier": "priority"
+                })),
+            ),
+            ("untyped-missing", None, None),
+            (
+                "inline-without-facts",
+                Some(UsageBodyCaptureState::Inline),
+                Some(serde_json::json!({"model": "gpt-5"})),
+            ),
+        ];
+
+        for (name, state, body) in cases {
+            let record = build_upsert_usage_record_from_event(&UsageEvent {
+                event_type: UsageEventType::Completed,
+                request_id: format!("req-clear-{name}"),
+                timestamp_ms: 1_700_000_000_000,
+                data: UsageEventData {
+                    provider_name: "OpenAI".to_string(),
+                    model: "gpt-5".to_string(),
+                    request_body: body.clone(),
+                    request_body_state: state,
+                    provider_request_body: body,
+                    provider_request_body_state: state,
+                    request_metadata: Some(serde_json::json!({
+                        "trace_id": "trace-1",
+                        "requested_reasoning_effort": "xhigh",
+                        "provider_reasoning_effort": "max",
+                        "provider_service_tier": "priority",
+                        "provider_actual_service_tier": "priority"
+                    })),
+                    ..UsageEventData::default()
+                },
+            })
+            .expect("record should build");
+
+            let metadata = record
+                .request_metadata
+                .as_ref()
+                .expect("audit facts remain");
+            assert_eq!(metadata["trace_id"], "trace-1");
+            assert_eq!(metadata["provider_actual_service_tier"], "priority");
+            assert!(metadata.get("requested_reasoning_effort").is_none());
+            assert!(metadata.get("provider_reasoning_effort").is_none());
+            assert!(metadata.get("provider_service_tier").is_none());
+        }
+    }
+
     #[test]
     fn cancelled_terminal_record_is_void_for_billing() {
         let record = build_upsert_usage_record_from_event(&UsageEvent {

crates/aether-usage/runtime/src/request_metadata.rs

@@ -6,9 +6,10 @@ use aether_contracts::ExecutionPlan;
 use aether_data_contracts::repository::usage::{
     extract_provider_actual_service_tier_from_response,
     extract_provider_reasoning_effort_from_body, extract_provider_service_tier_from_body,
-    normalize_provider_service_tier, resolve_provider_cache_ttl_minutes,
+    normalize_provider_service_tier, resolve_provider_cache_ttl_minutes, UsageBodyCaptureState,
     PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY, PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY,
     PROVIDER_REASONING_EFFORT_METADATA_KEY, PROVIDER_SERVICE_TIER_METADATA_KEY,
+    REQUESTED_REASONING_EFFORT_METADATA_KEY,
 };
 use serde_json::{json, Map, Value};
 
@@ -17,6 +18,53 @@ const MAX_USAGE_REQUEST_METADATA_NODES: usize = 4_000;
 const MAX_USAGE_REQUEST_METADATA_BYTES: usize = 16 * 1024;
 const MAX_USAGE_REQUEST_METADATA_STRING_BYTES: usize = 1_024;
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum RequestBodyDerivedFactsAction {
+    Refresh,
+    Preserve,
+    Clear,
+}
+
+pub(crate) fn request_body_derived_facts_action(
+    request_body: Option<&Value>,
+    state: Option<UsageBodyCaptureState>,
+) -> RequestBodyDerivedFactsAction {
+    // A typed `none` marker is produced by the final capture attempt. It must take precedence over
+    // a body value that may have survived from an earlier candidate. A missing marker (`None`)
+    // remains compatible with legacy events, where a present body is still authoritative.
+    if state == Some(UsageBodyCaptureState::None) {
+        return RequestBodyDerivedFactsAction::Clear;
+    }
+
+    if let Some(request_body) = request_body {
+        if matches!(
+            state,
+            Some(
+                UsageBodyCaptureState::Truncated
+                    | UsageBodyCaptureState::Disabled
+                    | UsageBodyCaptureState::Unavailable
+            )
+        ) || request_body.as_object().is_some_and(|body| {
+            body.get("truncated").and_then(Value::as_bool) == Some(true)
+                && body.get("reason").and_then(Value::as_str) == Some("body_capture_limit_exceeded")
+        }) {
+            return RequestBodyDerivedFactsAction::Preserve;
+        }
+        return RequestBodyDerivedFactsAction::Refresh;
+    }
+
+    match state {
+        Some(
+            UsageBodyCaptureState::Inline
+            | UsageBodyCaptureState::Reference
+            | UsageBodyCaptureState::Truncated
+            | UsageBodyCaptureState::Disabled
+            | UsageBodyCaptureState::Unavailable,
+        ) => RequestBodyDerivedFactsAction::Preserve,
+        Some(UsageBodyCaptureState::None) | None => RequestBodyDerivedFactsAction::Clear,
+    }
+}
+
 pub(crate) fn build_usage_request_metadata_seed(
     _plan: &ExecutionPlan,
     context: Option<&Map<String, Value>>,
@@ -76,6 +124,32 @@ pub(crate) fn sanitize_usage_request_metadata_ref(value: Option<&Value>) -> Opti
     (!filtered.is_empty()).then_some(Value::Object(filtered))
 }
 
+pub(crate) fn attach_client_request_body_metadata(
+    metadata: Option<Value>,
+    request_body: Option<&Value>,
+) -> Option<Value> {
+    let request_body_is_object = request_body.and_then(Value::as_object).is_some();
+    let reasoning_effort = extract_provider_reasoning_effort_from_body(request_body);
+    if !request_body_is_object && reasoning_effort.is_none() {
+        return metadata;
+    }
+
+    let mut object = match metadata {
+        Some(Value::Object(object)) => object,
+        _ => Map::new(),
+    };
+    if request_body_is_object {
+        object.remove(REQUESTED_REASONING_EFFORT_METADATA_KEY);
+    }
+    if let Some(reasoning_effort) = reasoning_effort {
+        object.insert(
+            REQUESTED_REASONING_EFFORT_METADATA_KEY.to_string(),
+            Value::String(reasoning_effort),
+        );
+    }
+    (!object.is_empty()).then_some(Value::Object(object))
+}
+
 pub(crate) fn attach_provider_request_body_metadata(
     metadata: Option<Value>,
     provider_api_format: Option<&str>,
@@ -129,6 +203,31 @@ pub(crate) fn attach_provider_request_body_metadata(
     (!object.is_empty()).then_some(Value::Object(object))
 }
 
+pub(crate) fn clear_client_request_body_metadata(metadata: Option<Value>) -> Option<Value> {
+    clear_request_metadata_fields(metadata, &[REQUESTED_REASONING_EFFORT_METADATA_KEY])
+}
+
+pub(crate) fn clear_provider_request_body_metadata(metadata: Option<Value>) -> Option<Value> {
+    clear_request_metadata_fields(
+        metadata,
+        &[
+            PROVIDER_REASONING_EFFORT_METADATA_KEY,
+            PROVIDER_SERVICE_TIER_METADATA_KEY,
+            PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY,
+        ],
+    )
+}
+
+fn clear_request_metadata_fields(metadata: Option<Value>, keys: &[&str]) -> Option<Value> {
+    let Some(Value::Object(mut object)) = metadata else {
+        return None;
+    };
+    for key in keys {
+        object.remove(*key);
+    }
+    (!object.is_empty()).then_some(Value::Object(object))
+}
+
 pub(crate) fn attach_provider_response_body_metadata(
     metadata: Option<Value>,
     provider_response_body: Option<&Value>,
@@ -148,6 +247,47 @@ pub(crate) fn attach_provider_response_body_metadata(
     attach_provider_actual_service_tier_metadata(metadata, actual_service_tier.as_deref())
 }
 
+/// Refreshes the response-derived tier for a terminal snapshot. Complete response objects are
+/// authoritative even when they contain no tier (which clears a stale candidate value). Capture
+/// placeholders/absent bodies are not authoritative, so a terminal summary already present in
+/// metadata is preserved for those cases.
+pub(crate) fn refresh_provider_response_body_metadata(
+    metadata: Option<Value>,
+    provider_response_body: Option<&Value>,
+) -> Option<Value> {
+    let is_capture_placeholder = provider_response_body
+        .and_then(Value::as_object)
+        .is_some_and(|body| {
+            body.get("truncated").and_then(Value::as_bool) == Some(true)
+                && body.get("reason").and_then(Value::as_str) == Some("body_capture_limit_exceeded")
+        });
+    let body_is_complete_object =
+        provider_response_body.and_then(Value::as_object).is_some() && !is_capture_placeholder;
+    let actual_service_tier =
+        extract_provider_actual_service_tier_from_response(provider_response_body)
+            .and_then(|value| normalize_provider_service_tier(&value));
+    let Some(actual_service_tier) = actual_service_tier else {
+        if !body_is_complete_object {
+            return metadata;
+        }
+        let Some(Value::Object(mut object)) = metadata else {
+            return None;
+        };
+        object.remove(PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY);
+        return (!object.is_empty()).then_some(Value::Object(object));
+    };
+
+    let mut object = match metadata {
+        Some(Value::Object(object)) => object,
+        _ => Map::new(),
+    };
+    object.insert(
+        PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY.to_string(),
+        Value::String(actual_service_tier),
+    );
+    (!object.is_empty()).then_some(Value::Object(object))
+}
+
 pub(crate) fn attach_provider_actual_service_tier_metadata(
     metadata: Option<Value>,
     actual_service_tier: Option<&str>,
@@ -179,6 +319,7 @@ fn copy_allowed_metadata_fields(source: &Map<String, Value>, target: &mut Map<St
     copy_non_empty_string(source, target, "request_path");
     copy_non_empty_string(source, target, "request_query_string");
     copy_non_empty_string(source, target, "request_path_and_query");
+    copy_non_empty_string(source, target, REQUESTED_REASONING_EFFORT_METADATA_KEY);
     copy_non_empty_string(source, target, PROVIDER_REASONING_EFFORT_METADATA_KEY);
     copy_non_empty_string(source, target, PROVIDER_SERVICE_TIER_METADATA_KEY);
     copy_non_empty_string(source, target, PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY);
@@ -226,6 +367,7 @@ fn move_allowed_metadata_fields(mut source: Map<String, Value>, target: &mut Map
     remove_non_empty_string(&mut source, target, "request_path");
     remove_non_empty_string(&mut source, target, "request_query_string");
     remove_non_empty_string(&mut source, target, "request_path_and_query");
+    remove_non_empty_string(&mut source, target, REQUESTED_REASONING_EFFORT_METADATA_KEY);
     remove_non_empty_string(&mut source, target, PROVIDER_REASONING_EFFORT_METADATA_KEY);
     remove_non_empty_string(&mut source, target, PROVIDER_SERVICE_TIER_METADATA_KEY);
     remove_non_empty_string(
@@ -517,10 +659,16 @@ mod tests {
     use serde_json::{json, Value};
     use std::collections::BTreeMap;
 
+    use crate::{
+        apply_usage_body_capture_policy_to_event, UsageBodyCapturePolicy, UsageEvent,
+        UsageEventData, UsageEventType, UsageRequestRecordLevel,
+    };
+
     use super::{
-        attach_provider_actual_service_tier_metadata, attach_provider_request_body_metadata,
-        attach_provider_response_body_metadata, build_usage_request_metadata_seed,
-        merge_usage_request_metadata, merge_usage_request_metadata_owned,
+        attach_client_request_body_metadata, attach_provider_actual_service_tier_metadata,
+        attach_provider_request_body_metadata, attach_provider_response_body_metadata,
+        build_usage_request_metadata_seed, merge_usage_request_metadata,
+        merge_usage_request_metadata_owned, refresh_provider_response_body_metadata,
         sanitize_usage_request_metadata, sanitize_usage_request_metadata_ref,
         MAX_USAGE_REQUEST_METADATA_BYTES, MAX_USAGE_REQUEST_METADATA_DEPTH,
         MAX_USAGE_REQUEST_METADATA_NODES,
@@ -830,6 +978,27 @@ mod tests {
         assert_eq!(metadata, None);
     }
 
+    #[test]
+    fn client_request_body_metadata_preserves_requested_reasoning_mapping_source() {
+        let updated = attach_client_request_body_metadata(
+            Some(json!({
+                "trace_id": "trace-1",
+                "requested_reasoning_effort": "low"
+            })),
+            Some(&json!({
+                "reasoning": { "effort": "XHigh" }
+            })),
+        )
+        .expect("metadata should remain");
+
+        assert_eq!(updated["requested_reasoning_effort"], "xhigh");
+
+        let cleared =
+            attach_client_request_body_metadata(Some(updated), Some(&json!({ "model": "gpt-5" })))
+                .expect("trace metadata should remain");
+        assert!(cleared.get("requested_reasoning_effort").is_none());
+    }
+
     #[test]
     fn provider_request_body_metadata_uses_final_provider_body_as_source_of_truth() {
         let metadata = Some(json!({
@@ -880,6 +1049,75 @@ mod tests {
         );
     }
 
+    #[test]
+    fn final_provider_request_tier_survives_basic_body_capture_as_derived_metadata() {
+        let request_body = json!({
+            "model": "gpt-5",
+            "reasoning": { "effort": "xhigh" }
+        });
+        let provider_request_body = json!({
+            "model": "gpt-5",
+            "reasoning": { "effort": "max" },
+            "service_tier": "priority"
+        });
+        let request_metadata = attach_client_request_body_metadata(None, Some(&request_body));
+        let request_metadata = attach_provider_request_body_metadata(
+            request_metadata,
+            Some("openai:responses"),
+            Some("gpt-5"),
+            Some("gpt-5"),
+            Some(&provider_request_body),
+        );
+        let mut event = UsageEvent::new(
+            UsageEventType::Completed,
+            "req-final-provider-tier",
+            UsageEventData {
+                provider_name: "OpenAI".to_string(),
+                model: "gpt-5".to_string(),
+                request_body: Some(request_body),
+                provider_request_body: Some(provider_request_body),
+                request_metadata,
+                ..UsageEventData::default()
+            },
+        );
+
+        apply_usage_body_capture_policy_to_event(
+            UsageBodyCapturePolicy {
+                record_level: UsageRequestRecordLevel::Basic,
+                max_request_body_bytes: Some(1024),
+                max_response_body_bytes: Some(1024),
+            },
+            &mut event,
+        );
+
+        assert_eq!(event.data.request_body, None);
+        assert_eq!(event.data.provider_request_body, None);
+        assert_eq!(
+            event
+                .data
+                .request_metadata
+                .as_ref()
+                .and_then(|metadata| metadata.get("requested_reasoning_effort")),
+            Some(&json!("xhigh"))
+        );
+        assert_eq!(
+            event
+                .data
+                .request_metadata
+                .as_ref()
+                .and_then(|metadata| metadata.get("provider_reasoning_effort")),
+            Some(&json!("max"))
+        );
+        assert_eq!(
+            event
+                .data
+                .request_metadata
+                .as_ref()
+                .and_then(|metadata| metadata.get("provider_service_tier")),
+            Some(&json!("priority"))
+        );
+    }
+
     #[test]
     fn provider_response_metadata_preserves_terminal_actual_service_tier() {
         let metadata = attach_provider_response_body_metadata(
@@ -902,6 +1140,39 @@ mod tests {
         );
     }
 
+    #[test]
+    fn terminal_response_refresh_replaces_stale_actual_tier() {
+        let metadata = refresh_provider_response_body_metadata(
+            Some(json!({"provider_actual_service_tier": "priority"})),
+            Some(&json!({"service_tier": "Default"})),
+        )
+        .expect("actual tier should remain");
+        assert_eq!(metadata["provider_actual_service_tier"], "default");
+
+        let metadata = refresh_provider_response_body_metadata(
+            Some(json!({
+                "trace_id": "trace-1",
+                "provider_actual_service_tier": "priority"
+            })),
+            Some(&json!({"id": "response-without-tier"})),
+        )
+        .expect("un-tiered complete response should remain auditable");
+        assert_eq!(metadata, json!({"trace_id": "trace-1"}));
+    }
+
+    #[test]
+    fn terminal_response_refresh_preserves_summary_over_capture_placeholder() {
+        let metadata = refresh_provider_response_body_metadata(
+            Some(json!({"provider_actual_service_tier": "priority"})),
+            Some(&json!({
+                "truncated": true,
+                "reason": "body_capture_limit_exceeded"
+            })),
+        )
+        .expect("summary should survive placeholder capture");
+        assert_eq!(metadata["provider_actual_service_tier"], "priority");
+    }
+
     #[test]
     fn terminal_summary_tier_uses_the_same_normalized_metadata_field() {
         let metadata = attach_provider_actual_service_tier_metadata(

crates/aether-usage/runtime/src/runtime.rs

@@ -13,7 +13,12 @@ use tokio::sync::mpsc;
 use tracing::{info, warn};
 
 use crate::executor::spawn_on_usage_background_runtime;
-use crate::request_metadata::attach_provider_response_body_metadata;
+use crate::request_metadata::{
+    attach_client_request_body_metadata, attach_provider_request_body_metadata,
+    attach_provider_response_body_metadata, clear_client_request_body_metadata,
+    clear_provider_request_body_metadata, request_body_derived_facts_action,
+    RequestBodyDerivedFactsAction,
+};
 use crate::worker::{
     build_usage_queue_worker_with_record_gate, UsageWorkerControl, UsageWorkerObservation,
 };
@@ -1264,6 +1269,7 @@ impl UsageRuntime {
     where
         T: UsageRuntimeAccess,
     {
+        preserve_request_facts(event);
         preserve_provider_response_facts(event);
         match self.cached_body_capture_policy(data).await {
             Ok(policy) => apply_usage_body_capture_policy_to_event(policy, event),
@@ -1734,6 +1740,44 @@ impl UsageRuntime {
     }
 }
 
+fn preserve_request_facts(event: &mut UsageEvent) {
+    let data = &mut event.data;
+    match request_body_derived_facts_action(data.request_body.as_ref(), data.request_body_state) {
+        RequestBodyDerivedFactsAction::Refresh => {
+            data.request_metadata = attach_client_request_body_metadata(
+                data.request_metadata.take(),
+                data.request_body.as_ref(),
+            );
+        }
+        RequestBodyDerivedFactsAction::Clear => {
+            data.request_metadata =
+                clear_client_request_body_metadata(data.request_metadata.take());
+        }
+        RequestBodyDerivedFactsAction::Preserve => {}
+    }
+    match request_body_derived_facts_action(
+        data.provider_request_body.as_ref(),
+        data.provider_request_body_state,
+    ) {
+        RequestBodyDerivedFactsAction::Refresh => {
+            data.request_metadata = attach_provider_request_body_metadata(
+                data.request_metadata.take(),
+                data.endpoint_api_format
+                    .as_deref()
+                    .or(data.api_format.as_deref()),
+                data.target_model.as_deref().or(Some(data.model.as_str())),
+                Some(data.model.as_str()),
+                data.provider_request_body.as_ref(),
+            );
+        }
+        RequestBodyDerivedFactsAction::Clear => {
+            data.request_metadata =
+                clear_provider_request_body_metadata(data.request_metadata.take());
+        }
+        RequestBodyDerivedFactsAction::Preserve => {}
+    }
+}
+
 fn preserve_provider_response_facts(event: &mut UsageEvent) {
     let metadata = event.data.request_metadata.take();
     event.data.request_metadata =
@@ -2989,7 +3033,9 @@ mod tests {
     use aether_data_contracts::repository::settlement::{
         StoredUsageSettlement, UsageSettlementInput,
     };
-    use aether_data_contracts::repository::usage::{StoredRequestUsageAudit, UpsertUsageRecord};
+    use aether_data_contracts::repository::usage::{
+        StoredRequestUsageAudit, UpsertUsageRecord, UsageBodyCaptureState,
+    };
     use aether_data_contracts::DataLayerError;
     use aether_runtime_state::{MemoryRuntimeStateConfig, RuntimeQueueStore, RuntimeState};
     use async_trait::async_trait;
@@ -2998,9 +3044,10 @@ mod tests {
     use tokio::time::{sleep, timeout, Duration};
 
     use super::{
-        preserve_provider_response_facts, LifecycleEventCoalescer, UsageBillingEventEnricher,
-        UsageBodyCapturePolicy, UsageEnqueueRetryDispatcher, UsageRequestRecordLevel,
-        UsageRuntimeAccess, UsageWorkerObservation, UsageWorkerSupervisorState,
+        preserve_provider_response_facts, preserve_request_facts, LifecycleEventCoalescer,
+        UsageBillingEventEnricher, UsageBodyCapturePolicy, UsageEnqueueRetryDispatcher,
+        UsageRequestRecordLevel, UsageRuntimeAccess, UsageWorkerObservation,
+        UsageWorkerSupervisorState,
     };
     use crate::worker::ManualProxyNodeCounter;
     use crate::{
@@ -5816,6 +5863,76 @@ mod tests {
         );
     }
 
+    #[test]
+    fn preserve_request_facts_ignores_post_capture_truncation_placeholders() {
+        let truncated = json!({
+            "truncated": true,
+            "reason": "body_capture_limit_exceeded"
+        });
+        let mut event = UsageEvent::new(
+            UsageEventType::Completed,
+            "req-truncated-preserve",
+            UsageEventData {
+                provider_name: "OpenAI".to_string(),
+                model: "gpt-5".to_string(),
+                request_body: Some(truncated.clone()),
+                request_body_state: Some(UsageBodyCaptureState::Truncated),
+                provider_request_body: Some(truncated),
+                provider_request_body_state: Some(UsageBodyCaptureState::Truncated),
+                request_metadata: Some(json!({
+                    "requested_reasoning_effort": "xhigh",
+                    "provider_reasoning_effort": "max",
+                    "provider_service_tier": "priority"
+                })),
+                ..UsageEventData::default()
+            },
+        );
+
+        preserve_request_facts(&mut event);
+
+        let metadata = event
+            .data
+            .request_metadata
+            .as_ref()
+            .expect("metadata remains");
+        assert_eq!(metadata["requested_reasoning_effort"], "xhigh");
+        assert_eq!(metadata["provider_reasoning_effort"], "max");
+        assert_eq!(metadata["provider_service_tier"], "priority");
+    }
+
+    #[test]
+    fn preserve_request_facts_clears_stale_facts_when_final_bodies_are_missing() {
+        let mut event = UsageEvent::new(
+            UsageEventType::Completed,
+            "req-missing-final-body",
+            UsageEventData {
+                provider_name: "OpenAI".to_string(),
+                model: "gpt-5".to_string(),
+                request_body_state: Some(UsageBodyCaptureState::None),
+                provider_request_body_state: Some(UsageBodyCaptureState::None),
+                request_metadata: Some(json!({
+                    "requested_reasoning_effort": "xhigh",
+                    "provider_reasoning_effort": "max",
+                    "provider_service_tier": "priority",
+                    "provider_actual_service_tier": "priority"
+                })),
+                ..UsageEventData::default()
+            },
+        );
+
+        preserve_request_facts(&mut event);
+
+        let metadata = event
+            .data
+            .request_metadata
+            .as_ref()
+            .expect("response audit fact remains");
+        assert!(metadata.get("requested_reasoning_effort").is_none());
+        assert!(metadata.get("provider_reasoning_effort").is_none());
+        assert!(metadata.get("provider_service_tier").is_none());
+        assert_eq!(metadata["provider_actual_service_tier"], "priority");
+    }
+
     #[test]
     fn basic_request_record_level_strips_body_capture_but_preserves_derived_fields() {
         let mut event = UsageEvent::new(
@@ -5826,9 +5943,16 @@ mod tests {
                 model: "gpt-5".to_string(),
                 total_tokens: Some(42),
                 error_message: Some("upstream failed".to_string()),
-                request_body: Some(json!({"messages":[{"role":"user","content":"hello"}]})),
+                request_body: Some(json!({
+                    "messages":[{"role":"user","content":"hello"}],
+                    "reasoning": {"effort": "xhigh"}
+                })),
                 request_body_ref: Some("usage://request/req-basic-1/request_body".to_string()),
-                provider_request_body: Some(json!({"model":"gpt-5"})),
+                provider_request_body: Some(json!({
+                    "model":"gpt-5",
+                    "reasoning": {"effort": "max"},
+                    "service_tier": "priority"
+                })),
                 provider_request_body_ref: Some(
                     "usage://request/req-basic-1/provider_request_body".to_string(),
                 ),
@@ -5841,11 +5965,16 @@ mod tests {
                 client_response_body_ref: Some(
                     "usage://request/req-basic-1/client_response_body".to_string(),
                 ),
-                request_metadata: Some(json!({"provider_service_tier": "priority"})),
+                request_metadata: Some(json!({
+                    "requested_reasoning_effort": "low",
+                    "provider_reasoning_effort": "medium",
+                    "provider_service_tier": "standard"
+                })),
                 ..UsageEventData::default()
             },
         );
 
+        preserve_request_facts(&mut event);
         preserve_provider_response_facts(&mut event);
         apply_usage_body_capture_policy_to_event(
             UsageBodyCapturePolicy {
@@ -5865,6 +5994,33 @@ mod tests {
         assert!(event.data.response_body_ref.is_none());
         assert!(event.data.client_response_body.is_none());
         assert!(event.data.client_response_body_ref.is_none());
+        assert_eq!(
+            event
+                .data
+                .request_metadata
+                .as_ref()
+                .and_then(|metadata| metadata.get("requested_reasoning_effort"))
+                .and_then(serde_json::Value::as_str),
+            Some("xhigh")
+        );
+        assert_eq!(
+            event
+                .data
+                .request_metadata
+                .as_ref()
+                .and_then(|metadata| metadata.get("provider_reasoning_effort"))
+                .and_then(serde_json::Value::as_str),
+            Some("max")
+        );
+        assert_eq!(
+            event
+                .data
+                .request_metadata
+                .as_ref()
+                .and_then(|metadata| metadata.get("provider_service_tier"))
+                .and_then(serde_json::Value::as_str),
+            Some("priority")
+        );
         assert_eq!(
             event
                 .data

crates/aether-usage/runtime/src/write.rs

@@ -13,10 +13,11 @@ use crate::body_capture::{
     RuntimeBodyCaptureMetadataInput,
 };
 use crate::request_metadata::{
-    attach_provider_actual_service_tier_metadata, attach_provider_request_body_metadata,
-    attach_provider_response_body_metadata, build_usage_request_metadata_seed,
+    attach_client_request_body_metadata, attach_provider_actual_service_tier_metadata,
+    attach_provider_request_body_metadata, build_usage_request_metadata_seed,
     merge_usage_request_metadata, merge_usage_request_metadata_owned,
-    sanitize_usage_request_metadata, sanitize_usage_request_metadata_ref,
+    refresh_provider_response_body_metadata, sanitize_usage_request_metadata,
+    sanitize_usage_request_metadata_ref,
 };
 use crate::{
     map_usage_from_response, stream_capture_terminal_state, GatewayStreamReportRequest,
@@ -696,6 +697,8 @@ fn build_terminal_usage_event_from_seed_impl(
     } else {
         merge_usage_request_metadata(request_metadata, audit_payload)
     };
+    let request_metadata =
+        attach_client_request_body_metadata(request_metadata, request_body.as_ref());
     let request_metadata = attach_provider_request_body_metadata(
         request_metadata,
         Some(provider_contract.as_str()),
@@ -1021,7 +1024,7 @@ pub fn build_sync_terminal_usage_seed(
         status_code,
         provider_response_full.as_ref(),
     );
-    let request_metadata = attach_provider_response_body_metadata(
+    let request_metadata = refresh_provider_response_body_metadata(
         context_seed.request_metadata,
         provider_response_full.as_ref(),
     );
@@ -1199,12 +1202,16 @@ pub fn build_stream_terminal_usage_seed(
         missing_observed_finish,
         terminal_error_message.is_some(),
     );
-    let request_metadata = attach_provider_actual_service_tier_metadata(
+    let request_metadata = refresh_provider_response_body_metadata(
         context_seed.request_metadata,
+        provider_response_full.as_ref(),
+    );
+    // The parser's terminal summary is authoritative when a response body is truncated or the
+    // body and summary disagree; attach it after the body refresh so it wins.
+    let request_metadata = attach_provider_actual_service_tier_metadata(
+        request_metadata,
         provider_actual_service_tier.as_deref(),
     );
-    let request_metadata =
-        attach_provider_response_body_metadata(request_metadata, provider_response_full.as_ref());
 
     TerminalUsageSeed {
         terminal_state,
@@ -2127,7 +2134,10 @@ fn build_runtime_request_metadata_seed_from_parts(
         metadata.insert("body_size".to_string(), body_size);
     }
 
-    (!metadata.is_empty()).then_some(Value::Object(metadata))
+    attach_client_request_body_metadata(
+        (!metadata.is_empty()).then_some(Value::Object(metadata)),
+        context_value_ref(context, "original_request_body"),
+    )
 }
 
 fn build_runtime_body_size_metadata(

frontend/src/api/__tests__/models-dev-pricing.spec.ts

@@ -239,6 +239,20 @@ describe('resolveModelsDevTieredPricing', () => {
     })
   })
 
+  it('uses the standard catalog when an imported tier has a zero default ratio', () => {
+    expect(resolveModelsDevTieredPricing('openai', 'gpt-5.6-sol', {
+      input: 5,
+      output: 30,
+    }, {
+      fast: {
+        cost: { input: 0, output: 0 },
+        provider: { body: { service_tier: 'priority' } },
+      },
+    })?.processing_tiers).toEqual({
+      priority: { price_multiplier: 1 },
+    })
+  })
+
   it('prefers Anthropic speed=fast when the mode body also carries a standard service tier', () => {
     expect(resolveModelsDevTieredPricing('anthropic', 'claude-opus-fast', {
       input: 5,

frontend/src/api/__tests__/provider-keys.spec.ts

@@ -0,0 +1,38 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+
+const { getMock } = vi.hoisted(() => ({ getMock: vi.fn() }))
+
+vi.mock('@/api/client', () => ({
+  default: {
+    get: getMock,
+  },
+}))
+
+import { getProviderKeysPage } from '@/api/endpoints/keys'
+
+describe('getProviderKeysPage', () => {
+  beforeEach(() => {
+    getMock.mockReset()
+  })
+
+  it('normalizes a legacy array response for the provider drawer', async () => {
+    getMock.mockResolvedValue({
+      data: [{ id: 'key-1' }, { id: 'key-2' }],
+    })
+
+    const result = await getProviderKeysPage('provider-demo', { page: 1, page_size: 1 })
+
+    expect(result).toMatchObject({ total: 2, page: 1, page_size: 1 })
+    expect(result.keys).toEqual([{ id: 'key-1' }])
+  })
+
+  it('normalizes a malformed object without exposing a non-array keys field', async () => {
+    getMock.mockResolvedValue({
+      data: { total: null, page: null, page_size: null, keys: {} },
+    })
+
+    const result = await getProviderKeysPage('provider-demo', { page: 2, page_size: 3 })
+
+    expect(result).toEqual({ total: 0, page: 2, page_size: 3, keys: [] })
+  })
+})

frontend/src/api/__tests__/providers-summary.spec.ts

@@ -0,0 +1,101 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+
+const { getMock } = vi.hoisted(() => ({ getMock: vi.fn() }))
+
+vi.mock('@/api/client', () => ({
+  default: {
+    get: getMock,
+  },
+}))
+
+import { getProviderMappingPreview, getProvidersSummary } from '@/api/endpoints/providers'
+
+const provider = {
+  id: 'provider-1',
+  name: 'Provider 1',
+  provider_type: 'openai',
+  is_active: true,
+  endpoints: [],
+}
+
+describe('getProvidersSummary', () => {
+  beforeEach(() => {
+    getMock.mockReset()
+  })
+
+  it('normalizes the paginated summary response', async () => {
+    getMock.mockResolvedValue({
+      data: { total: 1, page: 1, page_size: 20, items: [provider] },
+    })
+
+    const result = await getProvidersSummary({ page: 1, page_size: 20, search: 'paged' })
+
+    expect(result.total).toBe(1)
+    expect(result.items).toHaveLength(1)
+    expect(result.items[0]?.kiro_simulated_cache_enabled).toBe(false)
+  })
+
+  it('supports the legacy array response without reading an undefined items field', async () => {
+    getMock.mockResolvedValue({ data: [provider] })
+
+    const result = await getProvidersSummary({ page: 2, page_size: 20, search: 'legacy' })
+
+    expect(result).toMatchObject({ total: 1, page: 2, page_size: 20 })
+    expect(result.items).toHaveLength(1)
+  })
+})
+
+describe('getProviderMappingPreview', () => {
+  beforeEach(() => {
+    getMock.mockReset()
+  })
+
+  it('normalizes a non-contract response instead of exposing missing arrays to the UI', async () => {
+    getMock.mockResolvedValue({
+      data: { message: '演示模式：该接口暂未模拟', demo_mode: true },
+    })
+
+    const result = await getProviderMappingPreview('provider-demo')
+
+    expect(result).toEqual({
+      provider_id: 'provider-demo',
+      provider_name: '',
+      keys: [],
+      total_keys: 0,
+      total_matches: 0,
+      truncated: false,
+      truncated_keys: 0,
+      truncated_models: 0,
+    })
+  })
+
+  it('normalizes missing nested mapping arrays', async () => {
+    getMock.mockResolvedValue({
+      data: {
+        provider_id: 'provider-nested',
+        provider_name: 'Nested Provider',
+        keys: [{
+          key_id: 'key-1',
+          key_name: 'Primary',
+          masked_key: 'sk-***',
+          is_active: true,
+          allowed_models: null,
+          matching_global_models: [{
+            global_model_id: 'model-1',
+            global_model_name: 'gpt-5',
+            display_name: 'GPT-5',
+            is_active: true,
+            matched_models: null,
+          }],
+        }],
+      },
+    })
+
+    const result = await getProviderMappingPreview('provider-nested')
+
+    expect(result.keys[0]?.allowed_models).toEqual([])
+    expect(result.keys[0]?.matching_global_models[0]?.matched_models).toEqual([])
+    expect(result.total_keys).toBe(1)
+    expect(result.total_matches).toBe(1)
+  })
+})

frontend/src/api/__tests__/users.spec.ts

@@ -1,13 +1,15 @@
 import { beforeEach, describe, expect, it, vi } from 'vitest'
 
-const { getMock, cachedRequestMock } = vi.hoisted(() => ({
+const { getMock, postMock, cachedRequestMock } = vi.hoisted(() => ({
   getMock: vi.fn(),
+  postMock: vi.fn(),
   cachedRequestMock: vi.fn(async (_key: string, fn: () => Promise<unknown>) => fn()),
 }))
 
 vi.mock('@/api/client', () => ({
   default: {
     get: getMock,
+    post: postMock,
   },
 }))
 
@@ -20,6 +22,7 @@ import { usersApi } from '@/api/users'
 describe('usersApi admin list query', () => {
   beforeEach(() => {
     getMock.mockReset()
+    postMock.mockReset()
     cachedRequestMock.mockClear()
     getMock.mockResolvedValue({
       data: {
@@ -49,4 +52,55 @@ describe('usersApi admin list query', () => {
       },
     })
   })
+
+  it('keeps user management renderable when the group response has no items array', async () => {
+    getMock.mockResolvedValueOnce({
+      data: {
+        message: '演示模式：该接口暂未模拟',
+        demo_mode: true,
+      },
+    })
+
+    await expect(usersApi.listUserGroups()).resolves.toEqual({
+      message: '演示模式：该接口暂未模拟',
+      demo_mode: true,
+      items: [],
+    })
+  })
+
+  it('creates a managed key through the selected target user route', async () => {
+    postMock.mockResolvedValueOnce({
+      data: {
+        id: 'target-key',
+        key: 'sk-target',
+      },
+    })
+    const payload = {
+      name: 'target key',
+      feature_settings: {
+        chat_pii_redaction: { enabled: true },
+      },
+    }
+
+    await usersApi.createApiKey('target-user', payload)
+
+    expect(postMock).toHaveBeenCalledWith(
+      '/api/admin/users/target-user/api-keys',
+      payload,
+    )
+  })
+
+  it('reads managed keys from the production api_keys envelope', async () => {
+    getMock.mockResolvedValueOnce({
+      data: {
+        api_keys: [{ id: 'target-key', created_at: '2026-07-17T00:00:00Z' }],
+        total: 1,
+      },
+    })
+
+    await expect(usersApi.getUserApiKeys('target-user')).resolves.toEqual([
+      { id: 'target-key', created_at: '2026-07-17T00:00:00Z' },
+    ])
+    expect(getMock).toHaveBeenCalledWith('/api/admin/users/target-user/api-keys')
+  })
 })

frontend/src/api/dashboard.ts

@@ -209,6 +209,7 @@ export interface RequestDetail {
   has_format_conversion?: boolean | null
   model: string
   target_model?: string | null  // 映射后的目标模型名
+  requested_reasoning_effort?: string | null
   reasoning_effort?: string | null
   service_tier?: string | null
   actual_service_tier?: string | null
@@ -266,6 +267,7 @@ export interface RequestDetail {
   response_time_ms: number
   first_byte_time_ms?: number | null
   created_at: string
+  updated_at?: string | null
   request_headers?: Record<string, unknown>
   request_body?: Record<string, unknown>
   provider_request_headers?: Record<string, unknown>

frontend/src/api/endpoints/keys.ts

@@ -121,17 +121,50 @@ export interface ProviderKeysPageQuery {
   page_size?: number
 }
 
+type ProviderKeysPagePayload = ProviderKeysPageResponse | EndpointAPIKey[]
+
+function normalizeProviderKeysPage(
+  value: ProviderKeysPagePayload,
+  page: number,
+  pageSize: number,
+): ProviderKeysPageResponse {
+  if (Array.isArray(value)) {
+    const start = value.length > pageSize ? (page - 1) * pageSize : 0
+    const keys = value.slice(start, start + pageSize)
+    return {
+      total: value.length,
+      page,
+      page_size: pageSize,
+      keys,
+    }
+  }
+
+  const keys = Array.isArray(value.keys) ? value.keys : []
+  return {
+    total: typeof value.total === 'number' && Number.isFinite(value.total)
+      ? value.total
+      : keys.length,
+    page: typeof value.page === 'number' && Number.isFinite(value.page)
+      ? value.page
+      : page,
+    page_size: typeof value.page_size === 'number' && Number.isFinite(value.page_size)
+      ? value.page_size
+      : pageSize,
+    keys,
+  }
+}
+
 export async function getProviderKeysPage(
   providerId: string,
   params: ProviderKeysPageQuery = {},
 ): Promise<ProviderKeysPageResponse> {
   const page = params.page ?? 1
   const pageSize = params.page_size ?? 20
-  const response = await client.get<ProviderKeysPageResponse>(
+  const response = await client.get<ProviderKeysPagePayload>(
     `/api/admin/endpoints/providers/${providerId}/keys`,
     { params: { page, page_size: pageSize } },
   )
-  return response.data
+  return normalizeProviderKeysPage(response.data, page, pageSize)
 }
 
 export async function getProviderKeys(providerId: string): Promise<EndpointAPIKey[]> {

frontend/src/api/endpoints/providers.ts

@@ -42,6 +42,8 @@ export interface ProviderSummaryPageResponse {
   items: ProviderWithEndpointsSummary[]
 }
 
+type ProviderSummaryResponse = ProviderSummaryPageResponse | ProviderWithEndpointsSummary[]
+
 function normalizeProviderSummary(
   provider: ProviderWithEndpointsSummary,
 ): ProviderWithEndpointsSummary {
@@ -62,16 +64,26 @@ export async function getProvidersSummary(
   return cachedRequest(
     cacheKey,
     async () => {
-      const response = await client.get<ProviderSummaryPageResponse>(
+      const response = await client.get<ProviderSummaryResponse>(
         '/api/admin/providers/summary',
         {
           params,
           timeout: options.timeout,
         },
       )
+      const data = response.data
+      if (Array.isArray(data)) {
+        return {
+          total: data.length,
+          page: params.page ?? 1,
+          page_size: params.page_size ?? data.length,
+          items: data.map(normalizeProviderSummary),
+        }
+      }
+
       return {
-        ...response.data,
-        items: response.data.items.map(normalizeProviderSummary),
+        ...data,
+        items: (data.items ?? []).map(normalizeProviderSummary),
       }
     },
     cacheTtlMs,
@@ -371,6 +383,84 @@ export interface ProviderMappingPreviewResponse {
   truncated_models: number
 }
 
+function mappingPreviewRecord(value: unknown): Record<string, unknown> {
+  return value !== null && typeof value === 'object' && !Array.isArray(value)
+    ? value as Record<string, unknown>
+    : {}
+}
+
+function mappingPreviewString(value: unknown, fallback = ''): string {
+  return typeof value === 'string' ? value : fallback
+}
+
+function mappingPreviewCount(value: unknown, fallback: number): number {
+  return typeof value === 'number' && Number.isFinite(value) && value >= 0
+    ? value
+    : fallback
+}
+
+function normalizeProviderMappingPreview(
+  value: unknown,
+  providerId: string,
+): ProviderMappingPreviewResponse {
+  const source = mappingPreviewRecord(value)
+  const rawKeys = Array.isArray(source.keys) ? source.keys : []
+  const keys = rawKeys.map((rawKey) => {
+    const key = mappingPreviewRecord(rawKey)
+    const rawGlobalModels = Array.isArray(key.matching_global_models)
+      ? key.matching_global_models
+      : []
+
+    return {
+      key_id: mappingPreviewString(key.key_id),
+      key_name: mappingPreviewString(key.key_name),
+      masked_key: mappingPreviewString(key.masked_key, '***'),
+      is_active: key.is_active === true,
+      allowed_models: Array.isArray(key.allowed_models)
+        ? key.allowed_models.filter((item): item is string => typeof item === 'string')
+        : [],
+      matching_global_models: rawGlobalModels.map((rawGlobalModel) => {
+        const globalModel = mappingPreviewRecord(rawGlobalModel)
+        const rawMatchedModels = Array.isArray(globalModel.matched_models)
+          ? globalModel.matched_models
+          : []
+
+        return {
+          global_model_id: mappingPreviewString(globalModel.global_model_id),
+          global_model_name: mappingPreviewString(globalModel.global_model_name),
+          display_name: mappingPreviewString(
+            globalModel.display_name,
+            mappingPreviewString(globalModel.global_model_name),
+          ),
+          is_active: globalModel.is_active === true,
+          matched_models: rawMatchedModels.map((rawMatchedModel) => {
+            const matchedModel = mappingPreviewRecord(rawMatchedModel)
+            return {
+              allowed_model: mappingPreviewString(matchedModel.allowed_model),
+              mapping_pattern: mappingPreviewString(matchedModel.mapping_pattern),
+            }
+          }),
+        }
+      }),
+    }
+  })
+  const inferredMatches = keys.reduce(
+    (total, key) => total + key.matching_global_models.length,
+    0,
+  )
+
+  return {
+    provider_id: mappingPreviewString(source.provider_id, providerId),
+    provider_name: mappingPreviewString(source.provider_name),
+    keys,
+    total_keys: mappingPreviewCount(source.total_keys, keys.length),
+    total_matches: mappingPreviewCount(source.total_matches, inferredMatches),
+    truncated: source.truncated === true,
+    truncated_keys: mappingPreviewCount(source.truncated_keys, 0),
+    truncated_models: mappingPreviewCount(source.truncated_models, 0),
+  }
+}
+
 /**
  * 获取 Provider 映射预览
  */
@@ -379,6 +469,6 @@ export async function getProviderMappingPreview(
 ): Promise<ProviderMappingPreviewResponse> {
   return dedupedRequest(`providers:mapping-preview:${providerId}`, async () => {
     const response = await client.get<ProviderMappingPreviewResponse>(`/api/admin/providers/${providerId}/mapping-preview`)
-    return response.data
+    return normalizeProviderMappingPreview(response.data, providerId)
   })
 }

frontend/src/api/me.ts

@@ -55,6 +55,7 @@ export interface UsageRecordDetail {
   provider?: string // 仅管理员可见
   model: string
   request_type?: string | null
+  requested_reasoning_effort?: string | null
   reasoning_effort?: string | null
   service_tier?: string | null
   actual_service_tier?: string | null
@@ -370,6 +371,7 @@ export const meApi = {
       has_fallback?: boolean | null
       target_model?: string | null
       request_type?: string | null
+      requested_reasoning_effort?: string | null
       reasoning_effort?: string | null
       service_tier?: string | null
       actual_service_tier?: string | null

frontend/src/api/models-dev-pricing.ts

@@ -28,6 +28,7 @@ const TOKEN_PRICE_FIELDS = [
   'cache_read_price_per_1m',
 ] as const
 const PROCESSING_MODE_FALLBACK_KEYS = new Set(['fast', 'priority', 'flex', 'batch'])
+const DEFAULT_PROCESSING_TIER_MULTIPLIER = 1
 
 function isRecord(value: unknown): value is Record<string, unknown> {
   return typeof value === 'object' && value !== null && !Array.isArray(value)
@@ -128,7 +129,9 @@ function uniformPriceMultiplier(
       if (Math.abs(processingPrice - standardPrice * candidate) > 1e-9) return null
     }
   }
-  return candidate
+  // A zero ratio from an imported experimental mode is a missing/default price marker,
+  // not a free processing tier. Keep the tier on the Standard catalog so it remains billable.
+  return candidate === 0 ? DEFAULT_PROCESSING_TIER_MULTIPLIER : candidate
 }
 
 export function resolveModelsDevTieredPricing(

frontend/src/api/usage.ts

@@ -15,6 +15,7 @@ export interface UsageRecord {
   provider_name?: string
   model: string
   request_type?: string | null
+  requested_reasoning_effort?: string | null
   reasoning_effort?: string | null
   service_tier?: string | null
   actual_service_tier?: string | null
@@ -570,6 +571,7 @@ export const usageApi = {
       has_fallback?: boolean | null
       target_model?: string | null
       request_type?: string | null
+      requested_reasoning_effort?: string | null
       reasoning_effort?: string | null
       service_tier?: string | null
       actual_service_tier?: string | null

frontend/src/api/users.ts

@@ -374,7 +374,10 @@ export const usersApi = {
 
   async listUserGroups(): Promise<ListUserGroupsResponse> {
     const response = await apiClient.get<ListUserGroupsResponse>('/api/admin/user-groups')
-    return response.data
+    return {
+      ...response.data,
+      items: Array.isArray(response.data?.items) ? response.data.items : [],
+    }
   },
 
   async createUserGroup(payload: UpsertUserGroupRequest): Promise<UserGroup> {
@@ -417,8 +420,9 @@ export const usersApi = {
   },
 
   async getUserApiKeys(userId: string): Promise<ApiKey[]> {
-    const response = await apiClient.get<{ api_keys: ApiKey[] }>(`/api/admin/users/${userId}/api-keys`)
-    return response.data.api_keys
+    const response = await apiClient.get<{ api_keys?: ApiKey[] } | ApiKey[]>(`/api/admin/users/${userId}/api-keys`)
+    if (Array.isArray(response.data)) return response.data
+    return Array.isArray(response.data?.api_keys) ? response.data.api_keys : []
   },
 
   async getUserSessions(userId: string): Promise<SessionRecord[]> {

frontend/src/components/ui/__tests__/badge.spec.ts

@@ -0,0 +1,31 @@
+import { afterEach, describe, expect, it } from 'vitest'
+import { createApp, h, type App } from 'vue'
+
+import Badge from '../badge.vue'
+
+const mountedApps: Array<{ app: App, root: HTMLElement }> = []
+
+afterEach(() => {
+  for (const { app, root } of mountedApps.splice(0)) {
+    app.unmount()
+    root.remove()
+  }
+})
+
+describe('Badge', () => {
+  it('renders the transparent outline variant without the card background', () => {
+    const root = document.createElement('div')
+    document.body.appendChild(root)
+    const app = createApp({
+      render: () => h(Badge, { variant: 'outline-transparent' }, () => 'Fast'),
+    })
+
+    app.mount(root)
+    mountedApps.push({ app, root })
+
+    const badge = root.firstElementChild
+    expect(badge?.classList.contains('border-border')).toBe(true)
+    expect(badge?.classList.contains('bg-transparent')).toBe(true)
+    expect(badge?.classList.contains('bg-card/50')).toBe(false)
+  })
+})

frontend/src/components/ui/badge.vue

@@ -20,6 +20,7 @@ const badgeVariants = cva(
         destructive:
           'border-transparent bg-destructive text-destructive-foreground hover:bg-destructive/80',
         outline: 'text-foreground border-border bg-card/50',
+        'outline-transparent': 'text-foreground border-border bg-transparent',
         success:
           'border-transparent bg-primary text-primary-foreground hover:bg-primary/80',
         warning:
@@ -35,7 +36,7 @@ const badgeVariants = cva(
 )
 
 interface Props {
-  variant?: 'default' | 'secondary' | 'destructive' | 'outline' | 'success' | 'warning' | 'dark'
+  variant?: 'default' | 'secondary' | 'destructive' | 'outline' | 'outline-transparent' | 'success' | 'warning' | 'dark'
   class?: string
 }
 

frontend/src/features/pool/components/PoolAccountBatchDialog.vue

@@ -135,10 +135,10 @@
                     :title="row.statusBadgeTitle"
                   >{{ row.statusBadgeLabel }}</Badge>
                   <Badge
-                    v-if="row.key.oauth_plan_type"
+                    v-if="row.planLabel"
                     variant="outline"
                     class="text-[10px] px-1 py-0 h-4 shrink-0"
-                  >{{ row.key.oauth_plan_type }}</Badge>
+                  >{{ row.planLabel }}</Badge>
                   <Badge
                     v-if="row.oauthOrgBadge"
                     variant="secondary"
@@ -499,6 +499,7 @@ import { exportKey, refreshProviderQuota } from '@/api/endpoints/keys'
 import { refreshProviderOAuth } from '@/api/endpoints/provider_oauth'
 import { useProxyNodesStore } from '@/stores/proxy-nodes'
 import { getOAuthOrgBadge } from '@/utils/oauthIdentity'
+import { formatOAuthPlanType } from '@/utils/oauthPlanType'
 import {
   canExportOAuthCredential,
   canRefreshOAuthCredential,
@@ -552,6 +553,7 @@ type BatchActionOption = {
 
 type PageKeyRow = {
   key: PoolKeyDetail
+  planLabel: string
   authTypeLabel: string
   statusBadgeLabel: string | null
   statusBadgeTitle: string
@@ -650,6 +652,7 @@ const pageKeyRows = computed<PageKeyRow[]>(() => pageKeys.value.map((key) => {
 
   return {
     key,
+    planLabel: formatOAuthPlanType(key.oauth_plan_type),
     authTypeLabel: normalizeAuthTypeLabel(key),
     statusBadgeLabel,
     statusBadgeTitle: statusBadgeLabel ? getStatusBadgeTitle(key) : '',

frontend/src/features/pool/components/PoolKeyQuotaPanel.vue

@@ -109,7 +109,10 @@ const QuotaProgressRows = defineComponent({
         : 'flex flex-col gap-1 min-w-[140px] max-w-[208px]',
     }, [
       h('div', { class: 'flex items-center justify-between text-[10px] leading-none' }, [
-        h('span', { class: 'text-muted-foreground font-medium shrink-0' }, item.label),
+        h('span', {
+          'data-testid': 'pool-quota-period-label',
+          class: 'text-muted-foreground font-medium shrink-0',
+        }, item.label),
         item.resetText
           ? h('span', {
             'data-testid': 'pool-quota-reset-text',

frontend/src/features/pool/components/PoolKeyStatsPanel.vue

@@ -1,41 +1,40 @@
 <template>
   <div
-    v-if="cycle"
+    v-if="cycle && cycleMetricRows.length > 0"
     :class="cycleContainerClass"
-    :data-testid="variant === 'desktop' ? 'pool-stats-cycle-groups' : undefined"
+    :data-testid="variant === 'desktop' ? 'pool-stats-cycle-text' : 'pool-mobile-stats-cycle-text'"
   >
     <div
-      :class="cycleGridClass"
-      :data-testid="variant === 'desktop' ? 'pool-stats-cycle-grid' : 'pool-mobile-stats-cycle-grid'"
+      v-for="row in cycleMetricRows"
+      :key="`${row.key}-${variant}-cycle-row`"
+      class="flex items-baseline justify-between gap-3"
+      :title="`${row.label} ${row.valueText}`"
     >
-      <span aria-hidden="true" />
+      <span class="shrink-0 text-muted-foreground">
+        {{ row.label }}
+      </span>
       <span
-        :class="cycleGroupLabelClass"
-        :data-testid="variant === 'desktop' ? 'pool-stats-cycle-group-5h' : 'pool-mobile-stats-cycle-group-5h'"
-      >5H</span>
-      <span class="text-center text-muted-foreground/50">|</span>
-      <span
-        :class="cycleGroupLabelClass"
-        :data-testid="variant === 'desktop' ? 'pool-stats-cycle-group-weekly' : 'pool-mobile-stats-cycle-group-weekly'"
-      >{{ legacyT('周') }}</span>
-
-      <template
-        v-for="row in cycleRows"
-        :key="`${row.key}-${variant}-cycle-row`"
+        class="grid w-[112px] shrink-0 grid-cols-[minmax(0,1fr)_auto_minmax(0,1fr)] items-baseline gap-x-1 font-medium text-foreground"
+        :data-testid="variant === 'desktop' ? `pool-stats-cycle-${row.key}` : undefined"
       >
-        <span class="text-muted-foreground truncate">{{ row.label }}</span>
+        <span class="min-w-0 truncate text-right">{{ row.hasComparison ? row.smallValue : '-' }}</span>
         <span
-          :class="[cycleValueClass, row.fiveH.missing ? 'text-muted-foreground/80' : '']"
-          :data-testid="variant === 'desktop' ? `pool-stats-5h-${row.key}` : undefined"
-          :title="row.fiveH.value"
-        >{{ row.fiveH.value }}</span>
-        <span class="text-center text-muted-foreground/50">|</span>
-        <span
-          :class="[cycleValueClass, row.weekly.missing ? 'text-muted-foreground/80' : '']"
-          :data-testid="variant === 'desktop' ? `pool-stats-weekly-${row.key}` : undefined"
-          :title="row.weekly.value"
-        >{{ row.weekly.value }}</span>
-      </template>
+          class="w-1.5 text-center text-muted-foreground/60"
+          data-cycle-stat-part="divider"
+          aria-hidden="true"
+        >/</span>
+        <span class="min-w-0 truncate text-left">{{ row.largeValue }}</span>
+      </span>
+    </div>
+  </div>
+
+  <div
+    v-else-if="cycle"
+    :class="cycleContainerClass"
+    :data-testid="variant === 'desktop' ? 'pool-stats-cycle-empty' : 'pool-mobile-stats-cycle-empty'"
+  >
+    <div class="flex min-h-16 items-center justify-center text-muted-foreground">
+      —
     </div>
   </div>
 
@@ -68,47 +67,70 @@
 
 <script setup lang="ts">
 import { computed } from 'vue'
-import { useI18n } from '@/i18n'
-import type { PoolStatsMetric } from '@/features/pool/utils/poolStatsDisplay'
-
-export interface PoolKeyCycleStatsRow {
-  key: PoolStatsMetric['key']
-  label: string
-  fiveH: PoolStatsMetric
-  weekly: PoolStatsMetric
-}
+import type {
+  PoolCodexCycleStatsGroup,
+  PoolStatsMetric,
+  PoolStatsMetricKey,
+} from '@/features/pool/utils/poolStatsDisplay'
 
 const props = withDefaults(defineProps<{
   cycle: boolean
-  cycleRows: PoolKeyCycleStatsRow[]
+  cycleGroups: PoolCodexCycleStatsGroup[]
   accountMetrics: PoolStatsMetric[]
   variant?: 'desktop' | 'mobile'
 }>(), {
   variant: 'desktop',
 })
 
-const { legacyT } = useI18n()
+const CYCLE_METRIC_KEYS: PoolStatsMetricKey[] = ['request_count', 'total_tokens', 'total_cost_usd']
+const CYCLE_METRIC_LABELS: Record<PoolStatsMetricKey, string> = {
+  request_count: '请求',
+  total_tokens: 'Token',
+  total_cost_usd: '费用',
+}
 
-const cycleContainerClass = computed(() => props.variant === 'desktop'
-  ? 'mx-auto w-[188px] text-[10px] leading-4'
-  : ''
-)
+function missingMetric(key: PoolStatsMetricKey): PoolStatsMetric {
+  return {
+    key,
+    label: CYCLE_METRIC_LABELS[key],
+    value: '-',
+    missing: true,
+    numericValue: null,
+  }
+}
 
-const cycleGridClass = computed(() => [
-  'grid min-h-16 w-[188px] grid-cols-[38px_64px_10px_64px] items-center gap-x-1',
-  props.variant === 'mobile' ? 'text-left' : '',
+function metricForGroup(
+  group: PoolCodexCycleStatsGroup | undefined,
+  key: PoolStatsMetricKey,
+): PoolStatsMetric {
+  return group?.metrics.find(metric => metric.key === key) ?? missingMetric(key)
+}
+
+const cycleMetricRows = computed(() => {
+  const smallGroup = props.cycleGroups.length > 1 ? props.cycleGroups[0] : undefined
+  const largeGroup = props.cycleGroups.at(-1)
+  if (!largeGroup) return []
+
+  return CYCLE_METRIC_KEYS.map((key) => {
+    const smallMetric = metricForGroup(smallGroup, key)
+    const largeMetric = metricForGroup(largeGroup, key)
+    const hasComparison = Boolean(smallGroup)
+    return {
+      key,
+      label: CYCLE_METRIC_LABELS[key],
+      hasComparison,
+      smallValue: smallMetric.value,
+      largeValue: largeMetric.value,
+      valueText: hasComparison ? `${smallMetric.value}/${largeMetric.value}` : largeMetric.value,
+    }
+  })
+})
+
+const cycleContainerClass = computed(() => [
+  'w-full space-y-1 text-[11px] leading-4 tabular-nums',
+  props.variant === 'desktop' ? 'mx-auto max-w-[168px]' : 'py-0.5',
 ].filter(Boolean).join(' '))
 
-const cycleGroupLabelClass = computed(() => props.variant === 'desktop'
-  ? 'text-center text-[9px] font-semibold text-muted-foreground/80'
-  : 'text-center text-[10px] font-semibold text-foreground'
-)
-
-const cycleValueClass = computed(() => [
-  'min-w-0 truncate text-center text-foreground/90',
-  props.variant === 'desktop' ? 'tabular-nums' : 'font-medium tabular-nums',
-].join(' '))
-
 const accountContainerClass = computed(() => props.variant === 'desktop'
   ? 'grid min-h-16 w-[188px] grid-rows-4 gap-0 mx-auto text-[10px] leading-4'
   : ''

frontend/src/features/pool/components/__tests__/PoolKeyDisplayPanels.spec.ts

@@ -11,20 +11,41 @@ describe('pool key display panels', () => {
     document.body.appendChild(root)
     const app = createApp(PoolKeyStatsPanel, {
       cycle: true,
-      cycleRows: [{
-        key: 'request_count',
-        label: '请求',
-        fiveH: { key: 'request_count', label: '请求', value: '12', missing: false },
-        weekly: { key: 'request_count', label: '请求', value: '88', missing: false },
-      }],
+      cycleGroups: [
+        {
+          code: '5h',
+          label: '5H',
+          metrics: [{ key: 'request_count', label: '请求', value: '12', missing: false, numericValue: 12 }],
+        },
+        {
+          code: 'weekly',
+          label: '周',
+          metrics: [{ key: 'request_count', label: '请求', value: '88', missing: false, numericValue: 88 }],
+        },
+      ],
       accountMetrics: [],
     })
     app.use(createI18n())
     app.mount(root)
 
-    expect(root.querySelector('[data-testid="pool-stats-cycle-grid"]')).toBeTruthy()
-    expect(root.querySelector('[data-testid="pool-stats-5h-request_count"]')?.textContent).toBe('12')
-    expect(root.querySelector('[data-testid="pool-stats-weekly-request_count"]')?.textContent).toBe('88')
+    const stats = root.querySelector('[data-testid="pool-stats-cycle-text"]')
+    const requestValue = root.querySelector('[data-testid="pool-stats-cycle-request_count"]')
+    expect(stats).toBeTruthy()
+    expect(stats?.className).toContain('w-full')
+    expect(stats?.className).toContain('max-w-[168px]')
+    expect(requestValue?.textContent?.trim()).toBe('12/88')
+    expect(requestValue?.previousElementSibling?.textContent?.trim()).toBe('请求')
+    expect(requestValue?.parentElement?.className).toContain('justify-between')
+    expect(requestValue?.className).toContain('grid-cols-[minmax(0,1fr)_auto_minmax(0,1fr)]')
+    expect(requestValue?.className).toContain('w-[112px]')
+    expect(requestValue?.children[0]?.className).toContain('text-right')
+    expect(requestValue?.children[1]?.textContent).toBe('/')
+    expect(requestValue?.children[2]?.className).toContain('text-left')
+    expect(root.querySelectorAll('[data-cycle-stat-part="divider"]')).toHaveLength(3)
+    expect(root.querySelector('[data-testid="pool-stats-cycle-small-overlay"]')).toBeNull()
+    expect(root.querySelector('[data-testid="pool-stats-cycle-large-base"]')).toBeNull()
+    expect(root.textContent).not.toContain('5H')
+    expect(root.textContent).not.toContain('周')
 
     app.unmount()
     root.remove()
@@ -68,4 +89,40 @@ describe('pool key display panels', () => {
     app.unmount()
     root.remove()
   })
+
+  it('renders single-cycle stats as plain text', () => {
+    const root = document.createElement('div')
+    document.body.appendChild(root)
+    const app = createApp(PoolKeyStatsPanel, {
+      cycle: true,
+      cycleGroups: [{
+        code: 'monthly',
+        label: '月',
+        metrics: [
+          { key: 'request_count', label: '请求', value: '31', missing: false, numericValue: 31 },
+          { key: 'total_tokens', label: 'Token', value: '38.8K', missing: false, numericValue: 38_800 },
+          { key: 'total_cost_usd', label: '费用', value: '$0.077', missing: false, numericValue: 0.077 },
+        ],
+      }],
+      accountMetrics: [],
+    })
+    app.use(createI18n())
+    app.mount(root)
+
+    const requestValue = root.querySelector('[data-testid="pool-stats-cycle-request_count"]')
+    expect(requestValue?.textContent?.trim()).toBe('-/31')
+    expect(requestValue?.previousElementSibling?.textContent?.trim()).toBe('请求')
+    expect(requestValue?.className).toContain('grid-cols-[minmax(0,1fr)_auto_minmax(0,1fr)]')
+    expect(requestValue?.children[0]?.textContent).toBe('-')
+    expect(requestValue?.children[1]?.textContent).toBe('/')
+    expect(requestValue?.children[1]?.className).toContain('w-1.5')
+    expect(requestValue?.children[2]?.textContent).toBe('31')
+    expect(requestValue?.children[2]?.className).toContain('text-left')
+    expect(root.querySelector('[data-testid="pool-stats-cycle-single-marker"]')).toBeNull()
+    expect(root.querySelector('[data-testid="pool-stats-cycle-bar-request_count"]')).toBeNull()
+    expect(root.textContent).not.toContain('月')
+
+    app.unmount()
+    root.remove()
+  })
 })

frontend/src/features/pool/utils/__tests__/poolStatsDisplay.spec.ts

@@ -19,6 +19,7 @@ function createCodexKey(overrides: Partial<PoolStatsKeyInput> = {}): PoolStatsKe
         windows: [
           {
             code: '5h',
+            window_minutes: 300,
             usage: {
               request_count: 5,
               total_tokens: 2500,
@@ -27,10 +28,11 @@ function createCodexKey(overrides: Partial<PoolStatsKeyInput> = {}): PoolStatsKe
           },
           {
             code: 'weekly',
+            window_minutes: 10_080,
             usage: {
-              request_count: 0,
-              total_tokens: 0,
-              total_cost_usd: '0.00000000',
+              request_count: 8,
+              total_tokens: 5000,
+              total_cost_usd: '0.012',
             },
           },
         ],
@@ -54,9 +56,9 @@ describe('poolStatsDisplay', () => {
       total_cost_usd: '$0.0045',
     })
     expect(metricValues(display.groups[1].metrics)).toEqual({
-      request_count: '0',
-      total_tokens: '0',
-      total_cost_usd: '0',
+      request_count: '8',
+      total_tokens: '5K',
+      total_cost_usd: '$0.012',
     })
   })
 
@@ -65,7 +67,7 @@ describe('poolStatsDisplay', () => {
       createCodexKey({
         status_snapshot: {
           quota: {
-            windows: [{ code: '5h', usage: null }],
+            windows: [{ code: '5h', window_minutes: 300, usage: null }],
           },
         },
       }),
@@ -81,10 +83,48 @@ describe('poolStatsDisplay', () => {
       total_tokens: '—',
       total_cost_usd: '—',
     })
-    expect(metricValues(display.groups[1].metrics)).toEqual({
-      request_count: '—',
-      total_tokens: '—',
-      total_cost_usd: '—',
+    expect(display.groups).toHaveLength(1)
+  })
+
+  it('builds monthly stats from the actual quota window and ignores zero placeholders', () => {
+    const display = buildPoolStatsDisplay(
+      createCodexKey({
+        status_snapshot: {
+          quota: {
+            windows: [
+              {
+                code: 'monthly',
+                label: '月',
+                window_minutes: 43_800,
+                usage: {
+                  request_count: 12,
+                  total_tokens: 3456,
+                  total_cost_usd: '0.125',
+                },
+              },
+              {
+                code: 'weekly',
+                label: '周',
+                window_minutes: 0,
+                usage: {
+                  request_count: 99,
+                },
+              },
+            ],
+          },
+        },
+      }),
+      'codex',
+      'current_cycle',
+    )
+
+    expect(display.kind).toBe('codex_cycle')
+    if (display.kind !== 'codex_cycle') throw new Error('expected codex cycle display')
+    expect(display.groups.map(group => group.label)).toEqual(['月'])
+    expect(metricValues(display.groups[0].metrics)).toEqual({
+      request_count: '12',
+      total_tokens: '3.5K',
+      total_cost_usd: '$0.125',
     })
   })
 

frontend/src/features/pool/utils/poolStatsDisplay.ts

@@ -1,10 +1,11 @@
 import type { QuotaWindowUsageSnapshot } from '@/api/endpoints/types/statusSnapshot'
 import type { PoolManagementStatsMode } from '@/features/pool/utils/poolManagementState'
+import { getCodexQuotaWindowPresentation } from '@/utils/codexQuotaWindow'
 import { formatCompactNumber } from '@/utils/format'
 
 export type PoolStatsMetricKey = 'request_count' | 'total_tokens' | 'total_cost_usd'
 export type PoolStatsDisplayKind = 'account_total' | 'codex_cycle'
-export type PoolCodexCycleWindowCode = '5h' | 'weekly'
+export type PoolCodexCycleWindowCode = string
 
 export interface PoolStatsKeyInput {
   request_count?: number | null
@@ -14,6 +15,9 @@ export interface PoolStatsKeyInput {
     quota?: {
       windows?: Array<{
         code?: string | null
+        label?: string | null
+        scope?: string | null
+        window_minutes?: number | null
         usage?: QuotaWindowUsageSnapshot | null
       } | null> | null
     } | null
@@ -25,6 +29,7 @@ export interface PoolStatsMetric {
   label: string
   value: string
   missing: boolean
+  numericValue?: number | null
 }
 
 export interface PoolAccountTotalStatsDisplay {
@@ -46,10 +51,6 @@ export interface PoolCodexCycleStatsDisplay {
 export type PoolStatsDisplay = PoolAccountTotalStatsDisplay | PoolCodexCycleStatsDisplay
 
 const MISSING_STAT_VALUE = '—'
-const CODEX_CYCLE_WINDOWS: Array<{ code: PoolCodexCycleWindowCode, label: string }> = [
-  { code: '5h', label: '5H' },
-  { code: 'weekly', label: '周' },
-]
 
 export function isCodexProviderType(providerType: string | null | undefined): boolean {
   return String(providerType || '').trim().toLowerCase() === 'codex'
@@ -103,12 +104,14 @@ function createMetric(
   key: PoolStatsMetricKey,
   label: string,
   value: string | null,
+  numericValue?: number | null,
 ): PoolStatsMetric {
   return {
     key,
     label,
     value: value ?? MISSING_STAT_VALUE,
     missing: value == null,
+    numericValue: numericValue ?? null,
   }
 }
 
@@ -116,17 +119,40 @@ function normalizeWindowCode(value: unknown): string {
   return String(value || '').trim().toLowerCase()
 }
 
-function getQuotaWindowUsage(
-  key: PoolStatsKeyInput,
-  code: PoolCodexCycleWindowCode,
-): QuotaWindowUsageSnapshot | null {
+function getCodexCycleStatsGroups(key: PoolStatsKeyInput): PoolCodexCycleStatsGroup[] {
   const windows = key.status_snapshot?.quota?.windows
-  if (!Array.isArray(windows)) return null
+  if (!Array.isArray(windows)) return []
 
-  const window = windows.find(item => normalizeWindowCode(item?.code) === code)
-  return window?.usage ?? null
+  const seenCodes = new Set<string>()
+  return windows
+    .map((window) => {
+      if (!window) return null
+      const code = normalizeWindowCode(window.code)
+      const scope = String(window.scope || 'account').trim().toLowerCase()
+      if (!code || scope !== 'account' || code.startsWith('spark_') || seenCodes.has(code)) {
+        return null
+      }
+      const presentation = getCodexQuotaWindowPresentation({
+        code,
+        label: window.label,
+        scope,
+        window_minutes: window.window_minutes,
+      })
+      if (!presentation) return null
+      seenCodes.add(code)
+      return {
+        code,
+        label: presentation.label,
+        sortOrder: presentation.sortOrder,
+        metrics: buildCycleMetrics(window.usage ?? null),
+      }
+    })
+    .filter((group): group is PoolCodexCycleStatsGroup & { sortOrder: number } => group != null)
+    .sort((left, right) => left.sortOrder - right.sortOrder)
+    .map(({ sortOrder: _sortOrder, ...group }) => group)
 }
 
+
 function buildAccountTotalMetrics(key: PoolStatsKeyInput): PoolStatsMetric[] {
   return [
     createMetric('request_count', '请求', formatPoolStatInteger(key.request_count)),
@@ -136,10 +162,28 @@ function buildAccountTotalMetrics(key: PoolStatsKeyInput): PoolStatsMetric[] {
 }
 
 function buildCycleMetrics(usage: QuotaWindowUsageSnapshot | null): PoolStatsMetric[] {
+  const requestCount = usage?.request_count == null ? null : Number(usage.request_count)
+  const totalTokens = usage?.total_tokens == null ? null : Number(usage.total_tokens)
+  const totalCostUsd = usage?.total_cost_usd == null ? null : Number(usage.total_cost_usd)
   return [
-    createMetric('request_count', '请求', formatCycleInteger(usage?.request_count)),
-    createMetric('total_tokens', 'Token', formatCycleTokenCount(usage?.total_tokens)),
-    createMetric('total_cost_usd', '费用', formatCycleUsd(usage?.total_cost_usd)),
+    createMetric(
+      'request_count',
+      '请求',
+      formatCycleInteger(usage?.request_count),
+      Number.isFinite(requestCount) ? Math.max(requestCount ?? 0, 0) : null,
+    ),
+    createMetric(
+      'total_tokens',
+      'Token',
+      formatCycleTokenCount(usage?.total_tokens),
+      Number.isFinite(totalTokens) ? Math.max(totalTokens ?? 0, 0) : null,
+    ),
+    createMetric(
+      'total_cost_usd',
+      '费用',
+      formatCycleUsd(usage?.total_cost_usd),
+      Number.isFinite(totalCostUsd) ? Math.max(totalCostUsd ?? 0, 0) : null,
+    ),
   ]
 }
 
@@ -157,10 +201,7 @@ export function buildCodexCycleStatsDisplay(
 ): PoolCodexCycleStatsDisplay {
   return {
     kind: 'codex_cycle',
-    groups: CODEX_CYCLE_WINDOWS.map(window => ({
-      ...window,
-      metrics: buildCycleMetrics(getQuotaWindowUsage(key, window.code)),
-    })),
+    groups: getCodexCycleStatsGroups(key),
   }
 }
 

frontend/src/features/providers/components/ProviderDetailDrawer.vue

@@ -782,7 +782,7 @@
                 :models="providerModels"
                 :endpoints="endpoints"
                 :provider-keys="providerKeys"
-                :loading="loadingProviderModels || loadingProviderKeys"
+                :loading="loadingProviderModels"
                 @edit-model="handleEditModel"
                 @batch-assign="handleBatchAssign"
                 @refresh="loadEndpoints"
@@ -798,7 +798,7 @@
                 :provider-keys="providerKeys"
                 :models="providerModels"
                 :mapping-preview="providerMappingPreview"
-                :loading="loadingProviderEndpoints || loadingProviderKeys || loadingProviderModels || loadingProviderMappingPreview"
+                :loading="loadingProviderMappingPreview"
                 @refresh="handleModelMappingChanged"
               />
             </div>
@@ -1303,8 +1303,6 @@ watch(
         loading.value = false
       }
       void loadSystemFormatConversionConfig()
-      // mapping-preview 较慢，不阻塞首屏渲染
-      void loadMappingPreview()
       if (!hasInitialProvider) {
         await loadProvider()
       }
@@ -1313,7 +1311,13 @@ watch(
       if (newOpen && !oldOpen) {
         startCountdownTimer()
       }
-      void endpointsPromise.then(() => autoRefreshQuotaInBackground())
+      // 优先完成端点、密钥和模型的首屏数据，再请求计算量较大的映射预览。
+      // 同时校验抽屉状态，避免关闭或切换 Provider 后启动无用请求。
+      void endpointsPromise.then(() => {
+        if (!props.open || props.providerId !== newId) return
+        void loadMappingPreview()
+        void autoRefreshQuotaInBackground()
+      })
     } else if (!newOpen && oldOpen) {
       // 使在途请求失效，避免关闭后旧响应回写
       providerLoadRequestId += 1

frontend/src/features/providers/components/__tests__/ModelMappingTab.contract.spec.ts

@@ -0,0 +1,38 @@
+import { describe, expect, it } from 'vitest'
+import { createSSRApp, h } from 'vue'
+import { renderToString } from '@vue/server-renderer'
+
+import type { ProviderWithEndpointsSummary } from '@/api/endpoints'
+import ModelMappingTab from '../provider-tabs/ModelMappingTab.vue'
+
+const provider = {
+  id: 'provider-demo',
+  name: 'Demo Provider',
+  provider_type: 'custom',
+  is_active: true,
+  active_keys: 0,
+  api_formats: [],
+} as ProviderWithEndpointsSummary
+
+describe('ModelMappingTab response contracts', () => {
+  it('keeps the module visible when a legacy or malformed preview reaches the component', async () => {
+    const app = createSSRApp({
+      render: () => h(ModelMappingTab, {
+        provider,
+        models: [],
+        endpoints: [],
+        providerKeys: [],
+        mappingPreview: {
+          message: '演示模式：该接口暂未模拟',
+          demo_mode: true,
+        },
+        loading: false,
+      }),
+    })
+
+    const html = await renderToString(app)
+
+    expect(html).toContain('模型映射')
+    expect(html).toContain('暂无模型映射')
+  })
+})

frontend/src/features/providers/components/__tests__/ProviderDetailDrawer.loading.spec.ts

@@ -0,0 +1,32 @@
+import { readFileSync } from 'node:fs'
+import { resolve } from 'node:path'
+import { describe, expect, it } from 'vitest'
+
+const source = readFileSync(
+  resolve(process.cwd(), 'src/features/providers/components/ProviderDetailDrawer.vue'),
+  'utf8',
+)
+
+describe('ProviderDetailDrawer loading priorities', () => {
+  it('loads mapping preview after first-screen provider data', () => {
+    const openWatcher = source
+      .split('// 合并监听 providerId 和 open')[1]
+      ?.split('} else if (!newOpen && oldOpen)')[0]
+
+    expect(openWatcher).toBeTruthy()
+    expect(openWatcher).toContain('const endpointsPromise = loadEndpoints()')
+    expect(openWatcher).toContain('endpointsPromise.then(() => {')
+    expect(openWatcher).toContain('if (!props.open || props.providerId !== newId) return')
+    expect(openWatcher).toContain('void loadMappingPreview()')
+
+    const beforeEndpoints = openWatcher?.split('const endpointsPromise = loadEndpoints()')[0]
+    expect(beforeEndpoints).not.toContain('loadMappingPreview()')
+  })
+
+  it('keeps model and mapping loading states independent', () => {
+    expect(source).toContain(':loading="loadingProviderModels"')
+    expect(source).toContain(':loading="loadingProviderMappingPreview"')
+    expect(source).not.toContain(':loading="loadingProviderModels || loadingProviderKeys"')
+    expect(source).not.toContain(':loading="loadingProviderEndpoints || loadingProviderKeys || loadingProviderModels || loadingProviderMappingPreview"')
+  })
+})

frontend/src/features/providers/components/provider-tabs/ModelMappingTab.vue

@@ -599,13 +599,20 @@ const regexMappings = computed<CombinedMapping[]>(() => {
   const result: CombinedMapping[] = []
   const modelMap = new Map<string, CombinedMapping>()
 
-  for (const keyInfo of aliasMappingPreview.value.keys) {
-    for (const gm of keyInfo.matching_global_models) {
+  const previewKeys = Array.isArray(aliasMappingPreview.value.keys)
+    ? aliasMappingPreview.value.keys
+    : []
+  for (const keyInfo of previewKeys) {
+    const matchingGlobalModels = Array.isArray(keyInfo.matching_global_models)
+      ? keyInfo.matching_global_models
+      : []
+    for (const gm of matchingGlobalModels) {
+      const matchedModels = Array.isArray(gm.matched_models) ? gm.matched_models : []
       if (!modelMap.has(gm.global_model_id)) {
         modelMap.set(gm.global_model_id, {
           key: `regex-${gm.global_model_id}`,
           type: 'regex',
-          targetModelName: gm.display_name,
+          targetModelName: gm.display_name || gm.global_model_name || gm.global_model_id,
           targetModelId: gm.global_model_id,
           globalModelName: gm.global_model_name,
           mappings: [],
@@ -618,7 +625,7 @@ const regexMappings = computed<CombinedMapping[]>(() => {
       if (!mapping) continue
 
       // 添加 Key 信息
-      const keyMatches: MappingItem[] = gm.matched_models.map(m => ({
+      const keyMatches: MappingItem[] = matchedModels.map(m => ({
         name: m.allowed_model,
         pattern: m.mapping_pattern
       }))
@@ -631,7 +638,7 @@ const regexMappings = computed<CombinedMapping[]>(() => {
       })
 
       // 收集所有映射（去重）
-      for (const match of gm.matched_models) {
+      for (const match of matchedModels) {
         if (!mapping.mappings.some(m => m.name === match.allowed_model)) {
           mapping.mappings.push({
             name: match.allowed_model,

frontend/src/features/usage/components/ElapsedTimeText.vue

@@ -22,6 +22,9 @@ const props = withDefaults(defineProps<{
 const now = ref(Date.now())
 const precision = computed(() => Math.max(0, props.precision))
 const isActive = computed(() => props.status === 'pending' || props.status === 'streaming')
+// Usage timestamps have second precision while durations have millisecond precision.
+// Switching anchors can therefore introduce a sub-second phase shift at first byte.
+const ACTIVE_CLOCK_TIMESTAMP_PRECISION_MS = 1000
 
 let rafId: number | null = null
 
@@ -72,19 +75,29 @@ const displayText = computed(() => {
     return `${(responseTimeMs / 1000).toFixed(precision.value)}s`
   }
 
+  const createdAtMs = parseCreatedAtMs(props.createdAt)
+  const createdAtElapsedMs = Number.isNaN(createdAtMs)
+    ? null
+    : Math.max(0, now.value - createdAtMs)
+
   const responseTimeMs = finiteNonNegativeMs(props.responseTimeMs)
   const updatedAtMs = parseCreatedAtMs(props.responseTimeUpdatedAt)
   if (responseTimeMs != null && !Number.isNaN(updatedAtMs)) {
     const elapsedSinceUpdateMs = Math.max(0, now.value - updatedAtMs)
-    return `${((responseTimeMs + elapsedSinceUpdateMs) / 1000).toFixed(precision.value)}s`
+    const responseElapsedMs = responseTimeMs + elapsedSinceUpdateMs
+
+    // When both clocks differ only by timestamp truncation, keep the original
+    // created-at clock so the first-byte snapshot cannot make total time pause
+    // or move backwards. A larger difference is a real calibration signal
+    // (for example an audit row created before execution) and remains authoritative.
+    if (createdAtElapsedMs != null &&
+      Math.abs(responseElapsedMs - createdAtElapsedMs) <= ACTIVE_CLOCK_TIMESTAMP_PRECISION_MS) {
+      return `${(createdAtElapsedMs / 1000).toFixed(precision.value)}s`
+    }
+    return `${(responseElapsedMs / 1000).toFixed(precision.value)}s`
   }
 
-  if (!props.createdAt) return '-'
-
-  const createdAtMs = parseCreatedAtMs(props.createdAt)
-  if (Number.isNaN(createdAtMs)) return '-'
-
-  const elapsedMs = Math.max(0, now.value - createdAtMs)
-  return `${(elapsedMs / 1000).toFixed(precision.value)}s`
+  if (createdAtElapsedMs == null) return '-'
+  return `${(createdAtElapsedMs / 1000).toFixed(precision.value)}s`
 })
 </script>

frontend/src/features/usage/components/HorizontalRequestTimeline.vue

@@ -462,7 +462,7 @@
                   </span>
                 </div>
 
-                <!-- 错误信息：真实上游响应合并在此处展示 -->
+                <!-- 错误信息：将实际上游响应头和响应体作为同一个对象展示 -->
                 <div
                   v-if="currentAttempt.status === 'failed' && currentAttemptRequestError"
                   class="error-block"
@@ -485,12 +485,24 @@
                   </div>
                   <div
                     v-if="currentAttemptRequestError.upstreamResponse"
-                    class="error-json"
+                    class="error-json error-upstream-response-json"
                   >
                     <JsonContentPanel
                       :data="currentAttemptRequestError.upstreamResponse"
                       :is-dark="isDark"
-                      empty-message="无上游响应信息"
+                      title="上游响应"
+                      empty-message="无上游响应"
+                    />
+                  </div>
+                  <div
+                    v-if="currentAttemptRequestError.diagnostic"
+                    class="error-json error-diagnostic-json"
+                  >
+                    <JsonContentPanel
+                      :data="currentAttemptRequestError.diagnostic"
+                      :is-dark="isDark"
+                      title="失败诊断"
+                      empty-message="无失败诊断信息"
                     />
                   </div>
                 </div>
@@ -1227,7 +1239,7 @@ const normalizeUpstreamResponseDisplay = (value: unknown): Record<string, unknow
   const raw = extractObject(value)
   if (!raw) return null
   const statusCode = readNumberField(raw, 'status_code') ?? readNumberField(raw, 'statusCode')
-  const headers = raw.headers
+  const headers = raw.headers ?? raw.header
   const body = raw.body
   const bodyRef = readStringField(raw, 'body_ref') ?? readStringField(raw, 'bodyRef')
   const bodyState = readStringField(raw, 'body_state') ?? readStringField(raw, 'bodyState')
@@ -1627,6 +1639,7 @@ const currentAttemptRequestError = computed<{
   message: string
   statusCode?: number
   upstreamResponse: Record<string, unknown> | null
+  diagnostic: Record<string, unknown> | null
 } | null>(() => {
   const attempt = currentAttempt.value
   if (!attempt || attempt.status !== 'failed') return null
@@ -1667,10 +1680,20 @@ const currentAttemptRequestError = computed<{
         rawMessage,
       )
     : null
-  const upstreamResponseWithDiagnostic = diagnostic
-    ? { ...(upstreamResponseDisplay ?? {}), diagnostic }
-    : upstreamResponseDisplay
-  if (!message && statusCode == null && !upstreamResponseWithDiagnostic) return null
+  const upstreamResponseData: Record<string, unknown> = {}
+  const responseHeader = upstreamResponseDisplay?.headers
+  const responseBody = upstreamResponseDisplay?.body
+  if (hasRenderableValue(responseHeader)) upstreamResponseData.header = responseHeader
+  if (hasRenderableValue(responseBody)) upstreamResponseData.body = responseBody
+  const response = Object.keys(upstreamResponseData).length > 0
+    ? upstreamResponseData
+    : null
+  if (
+    !message
+    && statusCode == null
+    && !response
+    && !diagnostic
+  ) return null
   const showMessage = shouldShowAttemptMessageWithUpstreamResponse(
     rawMessage || fallbackType,
     upstreamResponseDisplay,
@@ -1679,7 +1702,8 @@ const currentAttemptRequestError = computed<{
   return {
     message: showMessage ? (message || '未知错误') : '',
     statusCode,
-    upstreamResponse: upstreamResponseWithDiagnostic,
+    upstreamResponse: response,
+    diagnostic,
   }
 })
 

frontend/src/features/usage/components/RequestDetailDrawer.vue

@@ -23,23 +23,21 @@
                 <h3 class="text-lg font-semibold">
                   请求详情
                 </h3>
-                <div class="flex min-w-0 max-w-[10rem] items-center gap-1 text-sm font-mono text-muted-foreground bg-muted px-2 py-0.5 rounded sm:max-w-none">
-                  <span class="truncate">{{ detail?.model || '-' }}</span>
-                  <template v-if="detail?.target_model && detail.target_model !== detail.model">
-                    <svg
-                      xmlns="http://www.w3.org/2000/svg"
-                      viewBox="0 0 20 20"
-                      fill="currentColor"
-                      class="w-3 h-3 flex-shrink-0"
-                    >
-                      <path
-                        fill-rule="evenodd"
-                        d="M3 10a.75.75 0 01.75-.75h10.638L10.23 5.29a.75.75 0 111.04-1.08l5.5 5.25a.75.75 0 010 1.08l-5.5 5.25a.75.75 0 11-1.04-1.08l4.158-3.96H3.75A.75.75 0 013 10z"
-                        clip-rule="evenodd"
-                      />
-                    </svg>
-                    <span class="truncate">{{ detail.target_model }}</span>
-                  </template>
+                <UsageModelDisplay
+                  v-if="headerModelRecord"
+                  :record="headerModelRecord"
+                  :cyber="detailCyberPolicyError"
+                  context="detail"
+                  data-request-detail-model-display
+                  class="min-w-0 max-w-[18rem] text-sm font-mono text-muted-foreground sm:max-w-none"
+                  model-row-class="rounded bg-muted px-2 py-0.5"
+                />
+                <div
+                  v-else
+                  data-request-detail-model-display
+                  class="rounded bg-muted px-2 py-0.5 text-sm font-mono text-muted-foreground"
+                >
+                  -
                 </div>
                 <Badge
                   v-if="detail?.status_code === 200"
@@ -159,47 +157,6 @@
               v-else-if="detail"
               class="space-y-4"
             >
-              <!-- 执行失败原因：优先展示本地调度/运行时失败摘要 -->
-              <Card
-                v-if="failureNotice"
-                class="border-red-200 bg-red-50/80 shadow-sm dark:border-red-900/60 dark:bg-red-950/30"
-              >
-                <div class="p-3 sm:p-4 flex gap-3">
-                  <div class="mt-0.5 flex h-8 w-8 shrink-0 items-center justify-center rounded-full bg-red-100 text-red-600 dark:bg-red-900/50 dark:text-red-300">
-                    <AlertTriangle class="h-4 w-4" />
-                  </div>
-                  <div class="min-w-0 flex-1 space-y-2">
-                    <div class="flex flex-wrap items-center gap-2">
-                      <h4 class="text-sm font-semibold text-red-950 dark:text-red-100">
-                        {{ failureNotice.title }}
-                      </h4>
-                      <Badge
-                        v-if="failureNotice.isSchedulingFailure"
-                        variant="outline"
-                        class="border-red-300 bg-white/60 text-[10px] text-red-700 dark:border-red-800 dark:bg-red-950/40 dark:text-red-200"
-                      >
-                        调度阶段
-                      </Badge>
-                    </div>
-                    <p class="text-sm leading-6 text-red-900 dark:text-red-100">
-                      {{ failureNotice.message }}
-                    </p>
-                    <div
-                      v-if="failureNotice.meta.length > 0"
-                      class="flex flex-wrap gap-1.5"
-                    >
-                      <span
-                        v-for="item in failureNotice.meta"
-                        :key="item"
-                        class="rounded-full border border-red-200 bg-white/70 px-2 py-0.5 text-[11px] font-mono text-red-700 dark:border-red-900 dark:bg-red-950/50 dark:text-red-200"
-                      >
-                        {{ item }}
-                      </span>
-                    </div>
-                  </div>
-                </div>
-              </Card>
-
               <!-- 费用与性能概览 -->
               <Card>
                 <div class="p-3 sm:p-4">
@@ -253,8 +210,6 @@
                       v-if="hasServiceTierFacts || processingTierPriceMultiplier !== null"
                       class="mt-3"
                       :requested="serviceTierFacts.requested"
-                      :actual="serviceTierFacts.actual"
-                      :billing="serviceTierFacts.billing"
                       :price-multiplier="processingTierPriceMultiplier"
                     />
                   </div>
@@ -886,6 +841,7 @@ import TabsContent from '@/components/ui/tabs-content.vue'
 import { AlertTriangle, Check, Columns2, RefreshCw, X, Monitor, Server, MessageSquareText, Code2, Terminal, Play } from 'lucide-vue-next'
 import { dashboardApi, type RequestDetail } from '@/api/dashboard'
 import type { ImageProgress, RequestTrace } from '@/api/requestTrace'
+import type { UsageRecord } from '../types'
 import { formatApiFormat } from '@/api/endpoints/types/api-format'
 import {
   formatByteSize,
@@ -906,7 +862,11 @@ import {
   resolveDisplayRequestStatus,
   resolveUsageStreamLabelSegments,
 } from '../utils/status'
-import { resolveRequestFailureNotice } from '../utils/errorNotice'
+import { isCyberPolicyError } from '../utils/cyberError'
+import {
+  mergeUsageRecordErrorMessage,
+  parseUsageTimestampMs,
+} from '../utils/recordSync'
 import {
   formatPricePerMillion,
   resolveProcessingTierPriceMultiplier,
@@ -922,7 +882,11 @@ import ConversationView from './RequestDetailDrawer/ConversationView.vue'
 import HorizontalRequestTimeline from './HorizontalRequestTimeline.vue'
 import ReplayDialog from './ReplayDialog.vue'
 import ServiceTierFacts from './ServiceTierFacts.vue'
-import { hasServiceTierFact, resolveServiceTierFacts } from '../utils/service-tier'
+import UsageModelDisplay from './UsageModelDisplay.vue'
+import {
+  hasServiceTierFact,
+  resolveServiceTierFacts,
+} from '../utils/service-tier'
 
 // 对话解析器
 import {
@@ -937,6 +901,7 @@ type RequestStateStatus = 'pending' | 'streaming' | 'completed' | 'failed' | 'ca
 const props = defineProps<{
   isOpen: boolean
   requestId: string | null
+  summaryRecord?: UsageRecord | null
 }>()
 
 const emit = defineEmits<{
@@ -966,11 +931,13 @@ const emit = defineEmits<{
     endpointApiFormat?: string | null
     hasFormatConversion?: boolean | null
     targetModel?: string | null
+    requestedReasoningEffort?: string | null
     reasoningEffort?: string | null
     serviceTier?: string | null
     actualServiceTier?: string | null
     imageProgress?: ImageProgress | null
     errorMessage?: string | null
+    updatedAt?: string | null
   }]
 }>()
 
@@ -1098,6 +1065,114 @@ function resolveRequestStateStatusFromDetail(nextDetail: Pick<RequestDetail, 'st
   return resolveRequestStateStatus(nextDetail.status, nextDetail.status_code, nextDetail.error_message)
 }
 
+type HeaderModelTextField =
+  | 'model'
+  | 'target_model'
+  | 'model_version'
+  | 'requested_reasoning_effort'
+  | 'reasoning_effort'
+  | 'service_tier'
+  | 'actual_service_tier'
+
+const FINAL_PROVIDER_HEADER_FIELDS = new Set<HeaderModelTextField>([
+  'target_model',
+  'reasoning_effort',
+  'service_tier',
+  'actual_service_tier',
+])
+
+let modelSnapshotRevision = 0
+const summaryModelRevision = ref(0)
+const detailModelRevision = ref(0)
+
+function usageSnapshotUpdatedAtMs(
+  source: UsageRecord | RequestDetail | null | undefined,
+): number | null {
+  const value = source?.updated_at
+  if (typeof value !== 'string' || !value.trim()) return null
+  return parseUsageTimestampMs(value)
+}
+
+function summaryNullIsNewerForProviderField(
+  field: HeaderModelTextField,
+  nextDetail: RequestDetail | null | undefined,
+): boolean {
+  if (!FINAL_PROVIDER_HEADER_FIELDS.has(field) || !props.summaryRecord) return false
+
+  const summaryUpdatedAt = usageSnapshotUpdatedAtMs(props.summaryRecord)
+  const detailUpdatedAt = usageSnapshotUpdatedAtMs(nextDetail)
+  if (summaryUpdatedAt != null && detailUpdatedAt != null && summaryUpdatedAt !== detailUpdatedAt) {
+    return summaryUpdatedAt > detailUpdatedAt
+  }
+
+  if (summaryModelRevision.value > detailModelRevision.value) return true
+
+  // A terminal list row is a complete final-provider snapshot. When no
+  // comparable timestamps exist, its explicit null must beat a cached detail
+  // from an earlier candidate. Non-terminal rows may still be filled by a
+  // detail request that completed after the lightweight list response.
+  return ['completed', 'failed', 'cancelled'].includes(props.summaryRecord.status ?? '')
+}
+
+function readHeaderModelTextField(
+  source: UsageRecord | RequestDetail | null | undefined,
+  field: HeaderModelTextField,
+): { resolved: boolean, value: string | null } {
+  if (!source || !Object.prototype.hasOwnProperty.call(source, field)) {
+    return { resolved: false, value: null }
+  }
+
+  const value = (source as unknown as Record<string, unknown>)[field]
+  if (value === null) return { resolved: true, value: null }
+  if (typeof value !== 'string') return { resolved: false, value: null }
+
+  const normalized = value.trim()
+  return { resolved: true, value: normalized || null }
+}
+
+function resolveHeaderModelTextField(
+  field: HeaderModelTextField,
+  nextDetail: RequestDetail | null | undefined,
+): string | null | undefined {
+  // Prefer a populated list/active fact so sparse detail cannot make the header
+  // flicker. A summary null is often only a lightweight-contract placeholder,
+  // though, so a later populated detail is still useful. Final-provider stale
+  // facts are cleared when full list/active snapshots merge into the summary.
+  const summaryValue = readHeaderModelTextField(props.summaryRecord, field)
+  if (summaryValue.value) return summaryValue.value
+
+  const detailValue = readHeaderModelTextField(nextDetail, field)
+  if (detailValue.value) {
+    if (
+      summaryValue.resolved
+      && summaryValue.value === null
+      && summaryNullIsNewerForProviderField(field, nextDetail)
+    ) return null
+    return detailValue.value
+  }
+
+  return summaryValue.resolved || detailValue.resolved ? null : undefined
+}
+
+watch(
+  () => [
+    props.requestId,
+    props.summaryRecord?.status,
+    props.summaryRecord?.updated_at,
+    props.summaryRecord?.model,
+    props.summaryRecord?.target_model,
+    props.summaryRecord?.model_version,
+    props.summaryRecord?.requested_reasoning_effort,
+    props.summaryRecord?.reasoning_effort,
+    props.summaryRecord?.service_tier,
+    props.summaryRecord?.actual_service_tier,
+  ],
+  () => {
+    summaryModelRevision.value = ++modelSnapshotRevision
+  },
+  { immediate: true },
+)
+
 function detailTotalCost(nextDetail: RequestDetail): number | null {
   const structuredCost = typeof nextDetail.cost === 'object' ? nextDetail.cost?.total : null
   const totalCost = toNumber(nextDetail.total_cost)
@@ -1124,6 +1199,15 @@ function emitDetailRequestState(nextDetail: RequestDetail) {
   const id = props.requestId
   if (!id) return
 
+  const targetModel = resolveHeaderModelTextField('target_model', nextDetail)
+  const requestedReasoningEffort = resolveHeaderModelTextField(
+    'requested_reasoning_effort',
+    nextDetail,
+  )
+  const reasoningEffort = resolveHeaderModelTextField('reasoning_effort', nextDetail)
+  const serviceTier = resolveHeaderModelTextField('service_tier', nextDetail)
+  const actualServiceTier = resolveHeaderModelTextField('actual_service_tier', nextDetail)
+
   emit('requestState', {
     id,
     requestId: nextDetail.request_id || nextDetail.id || null,
@@ -1148,11 +1232,13 @@ function emitDetailRequestState(nextDetail: RequestDetail) {
     apiFormat: nextDetail.api_format ?? null,
     endpointApiFormat: nextDetail.endpoint_api_format ?? null,
     hasFormatConversion: nextDetail.has_format_conversion ?? null,
-    targetModel: nextDetail.target_model ?? null,
-    reasoningEffort: nextDetail.reasoning_effort ?? null,
-    serviceTier: nextDetail.service_tier ?? null,
-    actualServiceTier: nextDetail.actual_service_tier ?? null,
+    ...(targetModel ? { targetModel } : {}),
+    ...(requestedReasoningEffort ? { requestedReasoningEffort } : {}),
+    ...(reasoningEffort ? { reasoningEffort } : {}),
+    ...(serviceTier ? { serviceTier } : {}),
+    ...(actualServiceTier ? { actualServiceTier } : {}),
     errorMessage: nextDetail.error_message ?? undefined,
+    updatedAt: nextDetail.updated_at ?? undefined,
   })
 }
 
@@ -1358,10 +1444,106 @@ const metadataPanelData = computed<Record<string, unknown> | null>(() => {
     : null
 })
 
-const failureNotice = computed(() => resolveRequestFailureNotice(detail.value))
+const detailForCurrentRequest = computed(() => (
+  detailMatchesRequestId(detail.value, props.requestId) ? detail.value : null
+))
 
-const serviceTierFacts = computed(() => resolveServiceTierFacts(detail.value))
+type AuthoritativeErrorSource = 'summary' | 'detail' | null
+
+function isTerminalRequestState(status: RequestStateStatus | undefined): boolean {
+  return status === 'completed' || status === 'failed' || status === 'cancelled'
+}
+
+function isSuccessfulTerminalRequestState(status: RequestStateStatus | undefined): boolean {
+  return status === 'completed' || status === 'cancelled'
+}
+
+const authoritativeErrorSource = computed<AuthoritativeErrorSource>(() => {
+  const summary = props.summaryRecord
+  const currentDetail = detailForCurrentRequest.value
+  if (!summary || !currentDetail) return null
+
+  const summaryStatus = resolveRequestStateStatus(
+    summary.status,
+    summary.status_code,
+    summary.error_message,
+  )
+  const detailStatus = resolveRequestStateStatusFromDetail(currentDetail)
+  const summaryUpdatedAtMs = usageSnapshotUpdatedAtMs(summary)
+  const detailUpdatedAtMs = usageSnapshotUpdatedAtMs(currentDetail)
+
+  if (summaryUpdatedAtMs != null && detailUpdatedAtMs != null &&
+    summaryUpdatedAtMs !== detailUpdatedAtMs) {
+    if (detailUpdatedAtMs > summaryUpdatedAtMs && isTerminalRequestState(detailStatus)) {
+      return 'detail'
+    }
+    if (summaryUpdatedAtMs > detailUpdatedAtMs && isTerminalRequestState(summaryStatus)) {
+      return 'summary'
+    }
+  }
+
+  // Without a comparable timestamp, a successful/cancelled terminal snapshot
+  // still has to clear a failure from the other source. Generic failed detail
+  // remains non-authoritative so opening the drawer cannot flash away a Cyber
+  // refusal already resolved by the list.
+  const detailSucceeded = isSuccessfulTerminalRequestState(detailStatus)
+  const summarySucceeded = isSuccessfulTerminalRequestState(summaryStatus)
+  if (detailSucceeded && !summarySucceeded) return 'detail'
+  if (summarySucceeded && !detailSucceeded) return 'summary'
+  if (detailSucceeded && summarySucceeded) {
+    return detailModelRevision.value >= summaryModelRevision.value ? 'detail' : 'summary'
+  }
+
+  return null
+})
+
+const headerModelRecord = computed(() => {
+  const summary = props.summaryRecord
+  const currentDetail = detailForCurrentRequest.value
+  if (!summary && !currentDetail) return null
+
+  const authoritativeSource = authoritativeErrorSource.value
+  const errorMessage = authoritativeSource === 'summary'
+    ? mergeUsageRecordErrorMessage(undefined, summary?.error_message, { authoritative: true })
+    : mergeUsageRecordErrorMessage(
+      summary?.error_message,
+      currentDetail?.error_message,
+      { authoritative: authoritativeSource === 'detail' },
+    )
+
+  return {
+    model: resolveHeaderModelTextField('model', currentDetail) ?? '-',
+    target_model: resolveHeaderModelTextField('target_model', currentDetail),
+    model_version: resolveHeaderModelTextField('model_version', currentDetail),
+    requested_reasoning_effort: resolveHeaderModelTextField(
+      'requested_reasoning_effort',
+      currentDetail,
+    ),
+    reasoning_effort: resolveHeaderModelTextField('reasoning_effort', currentDetail),
+    service_tier: resolveHeaderModelTextField('service_tier', currentDetail),
+    error_message: errorMessage,
+  }
+})
+const serviceTierFacts = computed(() => resolveServiceTierFacts(headerModelRecord.value))
 const hasServiceTierFacts = computed(() => hasServiceTierFact(serviceTierFacts.value))
+const detailCyberPolicyError = computed(() => {
+  const summaryError = props.summaryRecord?.error_message
+  const currentDetail = detailForCurrentRequest.value
+  const detailErrors = [
+    currentDetail?.error_message,
+    currentDetail?.upstream_error,
+    currentDetail?.failure_summary,
+    currentDetail?.response_body,
+  ]
+
+  if (authoritativeErrorSource.value === 'summary') {
+    return isCyberPolicyError(summaryError)
+  }
+  if (authoritativeErrorSource.value === 'detail') {
+    return isCyberPolicyError(detailErrors)
+  }
+  return isCyberPolicyError([summaryError, ...detailErrors])
+})
 const processingTierPriceMultiplier = computed(() => (
   resolveProcessingTierPriceMultiplier(detail.value)
 ))
@@ -2257,15 +2439,12 @@ const visibleTabs = computed(() => {
   })
 })
 
-watch(() => props.requestId, async (newId) => {
-  if (newId && props.isOpen) {
-    await loadDetail(newId)
-  }
-})
-
-watch(() => props.isOpen, async (isOpen) => {
-  if (isOpen && props.requestId) {
-    await loadDetail(props.requestId)
+watch([() => props.isOpen, () => props.requestId], async ([isOpen, requestId]) => {
+  if (isOpen && requestId) {
+    if (!detailMatchesRequestId(detail.value, requestId)) {
+      detail.value = null
+    }
+    await loadDetail(requestId)
   } else if (!isOpen) {
     stopAutoRefresh()
     showTimeline.value = false
@@ -2276,6 +2455,14 @@ watch(() => props.isOpen, async (isOpen) => {
   }
 })
 
+function detailMatchesRequestId(
+  candidate: RequestDetail | null | undefined,
+  requestId: string | null | undefined,
+): boolean {
+  if (!candidate || !requestId) return false
+  return candidate.id === requestId || candidate.request_id === requestId
+}
+
 async function ensureBodyContentLoaded() {
   if (!props.requestId || !detail.value) return
 
@@ -2334,6 +2521,9 @@ async function loadDetail(id: string, silent = false) {
   const requestId = ++loadDetailRequestId
   loadDetailInFlight = true
   if (!silent) {
+    if (!detailMatchesRequestId(detail.value, id)) {
+      detail.value = null
+    }
     loading.value = true
     historicalPricing.value = null
     timelineLoaded.value = false
@@ -2371,6 +2561,7 @@ async function loadDetail(id: string, silent = false) {
       error_flow: response.error_flow,
       scheduling_failure: response.scheduling_failure,
     }
+    detailModelRevision.value = ++modelSnapshotRevision
     detail.value = nextDetail
     bodiesLoadedForRequestId.value = sameRequest ? bodiesLoadedForRequestId.value : null
     emitDetailRequestState(nextDetail)

frontend/src/features/usage/components/ServiceTierFacts.vue

@@ -1,12 +1,12 @@
 <template>
   <dl
     class="grid grid-cols-1 gap-x-4 gap-y-1.5 text-xs"
-    :class="hasPriceMultiplier ? 'sm:grid-cols-4' : 'sm:grid-cols-3'"
+    :class="hasPriceMultiplier ? 'sm:grid-cols-3' : 'sm:grid-cols-2'"
     data-testid="service-tier-facts"
   >
     <div class="flex min-w-0 items-baseline justify-between gap-3 sm:block">
       <dt class="text-muted-foreground">
-        请求层级
+        上游请求层级
       </dt>
       <dd
         class="truncate font-mono font-medium text-foreground sm:mt-0.5"
@@ -15,26 +15,15 @@
         {{ formatServiceTierFact(requested) || '-' }}
       </dd>
     </div>
-    <div class="flex min-w-0 items-baseline justify-between gap-3 sm:block">
-      <dt class="text-muted-foreground">
-        实际层级
-      </dt>
-      <dd
-        class="truncate font-mono font-medium text-foreground sm:mt-0.5"
-        :title="formatServiceTierFact(actual) || '-'"
-      >
-        {{ formatServiceTierFact(actual) || '-' }}
-      </dd>
-    </div>
     <div class="flex min-w-0 items-baseline justify-between gap-3 sm:block">
       <dt class="text-muted-foreground">
         计费层级
       </dt>
       <dd
         class="truncate font-mono font-medium text-foreground sm:mt-0.5"
-        :title="formatServiceTierFact(billing) || '-'"
+        :title="formatServiceTierFact(requested) || '-'"
       >
-        {{ formatServiceTierFact(billing) || '-' }}
+        {{ formatServiceTierFact(requested) || '-' }}
       </dd>
     </div>
     <div
@@ -58,8 +47,6 @@ import { formatServiceTierFact } from '../utils/service-tier'
 
 const props = defineProps<{
   requested: string | null
-  actual: string | null
-  billing: string | null
   priceMultiplier?: number | null
 }>()
 
@@ -70,7 +57,7 @@ const hasPriceMultiplier = computed(() => (
 ))
 
 const multiplierTierLabel = computed(() => (
-  formatServiceTierFact(props.billing ?? props.actual ?? props.requested) ?? '处理层级'
+  formatServiceTierFact(props.requested) ?? '处理层级'
 ))
 
 const formattedPriceMultiplier = computed(() => (

frontend/src/features/usage/components/UsageModelDisplay.vue

@@ -0,0 +1,184 @@
+<template>
+  <div
+    class="flex min-w-0 max-w-full flex-col gap-0.5"
+    :class="shouldStackBadges && stackFullWidth ? 'w-full items-start' : 'items-start'"
+    :data-usage-model-layout="shouldStackBadges ? 'stacked' : 'inline'"
+    :data-request-detail-model-layout="context === 'detail'
+      ? (shouldStackBadges ? 'stacked' : 'inline')
+      : undefined"
+  >
+    <div
+      class="flex min-w-0 max-w-full items-center gap-1"
+      :class="modelRowClass"
+    >
+      <span
+        class="min-w-0 truncate"
+        :class="modelClass"
+        data-usage-model-source
+      >{{ record.model }}</span>
+      <template v-if="actualModel">
+        <span class="shrink-0 text-muted-foreground/70">-&gt;</span>
+        <span
+          class="min-w-0 truncate"
+          :class="modelClass"
+          data-usage-model-target
+        >{{ actualModel }}</span>
+      </template>
+      <template v-if="!shouldStackBadges">
+        <Badge
+          v-for="badge in modelBadges"
+          :key="badge.key"
+          :data-usage-model-badge="badge.key"
+          :data-request-detail-model-badge="context === 'detail' ? badge.key : undefined"
+          :variant="badge.variant"
+          class="h-4 shrink-0 whitespace-nowrap rounded-full px-1.5 text-[10px] leading-4"
+          :class="badge.className"
+          :title="badge.title"
+          :aria-label="badge.ariaLabel"
+        >
+          {{ badge.label }}
+        </Badge>
+      </template>
+    </div>
+
+    <div
+      v-if="shouldStackBadges && modelBadges.length > 0"
+      class="flex min-w-0 max-w-full flex-wrap items-center gap-1"
+      data-usage-model-badges-row
+      :data-request-detail-model-badges-row="context === 'detail' ? '' : undefined"
+    >
+      <Badge
+        v-for="badge in modelBadges"
+        :key="badge.key"
+        :data-usage-model-badge="badge.key"
+        :data-request-detail-model-badge="context === 'detail' ? badge.key : undefined"
+        :variant="badge.variant"
+        class="h-4 shrink-0 whitespace-nowrap rounded-full px-1.5 text-[10px] leading-4"
+        :class="badge.className"
+        :title="badge.title"
+        :aria-label="badge.ariaLabel"
+      >
+        {{ badge.label }}
+      </Badge>
+    </div>
+  </div>
+</template>
+
+<script setup lang="ts">
+import { computed } from 'vue'
+
+import { Badge } from '@/components/ui'
+import { isCyberPolicyError } from '../utils/cyberError'
+import { formatServiceTierFact } from '../utils/service-tier'
+
+type ModelBadgeKey = 'compact' | 'reasoning' | 'fast' | 'cyber'
+
+interface ModelBadgePresentation {
+  key: ModelBadgeKey
+  label: string
+  variant: 'outline' | 'outline-transparent'
+  className: string
+  title: string
+  ariaLabel: string
+}
+
+interface UsageModelDisplayRecord {
+  model: string
+  target_model?: string | null
+  model_version?: string | null
+  request_type?: string | null
+  requested_reasoning_effort?: string | null
+  reasoning_effort?: string | null
+  service_tier?: string | null
+  error_message?: string | null
+}
+
+const props = withDefaults(defineProps<{
+  record: UsageModelDisplayRecord
+  modelClass?: string
+  modelRowClass?: string
+  context?: 'usage' | 'detail'
+  cyber?: boolean | null
+  stackFullWidth?: boolean
+}>(), {
+  modelClass: '',
+  modelRowClass: '',
+  context: 'usage',
+  cyber: null,
+  stackFullWidth: false,
+})
+
+const actualModel = computed(() => {
+  const targetModel = normalizeText(props.record.target_model)
+  if (targetModel && targetModel !== props.record.model) return targetModel
+
+  const modelVersion = normalizeText(props.record.model_version)
+  if (modelVersion && modelVersion !== props.record.model) return modelVersion
+  return null
+})
+
+const reasoningLabel = computed(() => {
+  const requested = normalizeText(props.record.requested_reasoning_effort)
+  const actual = normalizeText(props.record.reasoning_effort)
+  if (requested && actual && requested.toLowerCase() !== actual.toLowerCase()) {
+    return `${requested} -> ${actual}`
+  }
+  return actual ?? requested
+})
+
+const modelBadges = computed<ModelBadgePresentation[]>(() => {
+  const badges: ModelBadgePresentation[] = []
+  if (normalizeText(props.record.request_type)?.toLowerCase() === 'compact') {
+    badges.push({
+      key: 'compact',
+      label: '会话压缩',
+      variant: 'outline',
+      className: 'border-sky-500/30 bg-sky-500/5 text-sky-700 dark:text-sky-300',
+      title: '会话压缩',
+      ariaLabel: '会话压缩',
+    })
+  }
+  if (reasoningLabel.value) {
+    badges.push({
+      key: 'reasoning',
+      label: reasoningLabel.value,
+      variant: 'outline',
+      className: 'border-primary/30 bg-primary/5 text-primary',
+      title: `Reasoning: ${reasoningLabel.value}`,
+      ariaLabel: `Reasoning: ${reasoningLabel.value}`,
+    })
+  }
+
+  if (formatServiceTierFact(props.record.service_tier) === 'Fast') {
+    badges.push({
+      key: 'fast',
+      label: 'Fast',
+      variant: 'outline-transparent',
+      className: 'text-amber-700 dark:text-amber-300',
+      title: '上游请求档位：Fast\n计费档位：Fast',
+      ariaLabel: '上游请求档位：Fast，计费档位：Fast',
+    })
+  }
+
+  if (props.cyber ?? isCyberPolicyError(props.record.error_message)) {
+    badges.push({
+      key: 'cyber',
+      label: 'Cyber',
+      variant: 'outline',
+      className: 'border-primary/30 bg-primary/5 text-rose-600 dark:text-rose-300',
+      title: '上游 Cyber Policy 拒绝',
+      ariaLabel: '上游 Cyber Policy 拒绝',
+    })
+  }
+  return badges
+})
+
+const shouldStackBadges = computed(() => (
+  actualModel.value !== null || modelBadges.value.length >= 3
+))
+
+function normalizeText(value: string | null | undefined): string | null {
+  const normalized = value?.trim()
+  return normalized || null
+}
+</script>

frontend/src/features/usage/components/UsageRecordsTable.vue

@@ -242,34 +242,12 @@
         <!-- 第一行：模型 + 费用 -->
         <div class="flex items-start justify-between gap-2">
           <div class="min-w-0 flex-1">
-            <div class="flex min-w-0 items-center gap-1.5">
-              <span class="min-w-0 truncate text-[15px] font-semibold leading-5">{{ record.model }}</span>
-              <Badge
-                v-if="getRequestTypeLabel(record)"
-                variant="outline"
-                class="h-4 rounded-full border-sky-500/30 bg-sky-500/5 px-1.5 text-[10px] leading-4 text-sky-700 dark:text-sky-300 flex-shrink-0"
-                title="会话压缩"
-              >
-                {{ getRequestTypeLabel(record) }}
-              </Badge>
-              <Badge
-                v-if="getReasoningEffort(record)"
-                variant="outline"
-                class="h-4 rounded-full border-primary/30 bg-primary/5 px-1.5 text-[10px] leading-4 text-primary flex-shrink-0"
-                :title="getReasoningEffortTitle(record)"
-              >
-                {{ getReasoningEffort(record) }}
-              </Badge>
-              <Badge
-                v-if="getServiceTierBadge(record)"
-                variant="outline"
-                class="h-4 whitespace-nowrap rounded-full px-1.5 text-[10px] leading-4 flex-shrink-0"
-                :class="getServiceTierBadge(record)?.className"
-                :title="getServiceTierBadge(record)?.title"
-                :aria-label="getServiceTierBadge(record)?.ariaLabel"
-              >
-                {{ getServiceTierBadge(record)?.label }}
-              </Badge>
+            <div class="flex min-w-0 flex-wrap items-center gap-1.5">
+              <UsageModelDisplay
+                :record="record"
+                model-class="text-[15px] font-semibold leading-5"
+                stack-full-width
+              />
               <!-- 状态 Badge -->
               <Badge
                 v-if="isUsageRecordFailed(record)"
@@ -320,10 +298,6 @@
                 {{ getStreamModeLabel(record) }}
               </Badge>
             </div>
-            <span
-              v-if="getActualModel(record)"
-              class="text-[11px] text-muted-foreground truncate block"
-            >-> {{ getActualModel(record) }}</span>
           </div>
           <div class="flex flex-col items-end flex-shrink-0">
             <span class="text-sm text-primary font-semibold leading-5">{{ formatCurrency(record.cost || 0) }}</span>
@@ -745,85 +719,10 @@
             :class="[isAdmin ? 'w-[14%]' : 'w-[22%]']"
             :title="getModelTooltip(record)"
           >
-            <div
-              v-if="getActualModel(record)"
-              class="flex flex-col text-xs gap-0.5"
-            >
-              <div class="flex min-w-0 items-center gap-1">
-                <span class="truncate">{{ record.model }}</span>
-                <svg
-                  xmlns="http://www.w3.org/2000/svg"
-                  viewBox="0 0 20 20"
-                  fill="currentColor"
-                  class="w-3 h-3 text-muted-foreground flex-shrink-0"
-                >
-                  <path
-                    fill-rule="evenodd"
-                    d="M3 10a.75.75 0 01.75-.75h10.638L10.23 5.29a.75.75 0 111.04-1.08l5.5 5.25a.75.75 0 010 1.08l-5.5 5.25a.75.75 0 11-1.04-1.08l4.158-3.96H3.75A.75.75 0 013 10z"
-                    clip-rule="evenodd"
-                  />
-                </svg>
-                <Badge
-                  v-if="getRequestTypeLabel(record)"
-                  variant="outline"
-                  class="h-4 rounded-full border-sky-500/30 bg-sky-500/5 px-1.5 text-[10px] leading-4 text-sky-700 dark:text-sky-300 flex-shrink-0"
-                  title="会话压缩"
-                >
-                  {{ getRequestTypeLabel(record) }}
-                </Badge>
-                <Badge
-                  v-if="getReasoningEffort(record)"
-                  variant="outline"
-                  class="h-4 rounded-full border-primary/30 bg-primary/5 px-1.5 text-[10px] leading-4 text-primary flex-shrink-0"
-                  :title="getReasoningEffortTitle(record)"
-                >
-                  {{ getReasoningEffort(record) }}
-                </Badge>
-                <Badge
-                  v-if="getServiceTierBadge(record)"
-                  variant="outline"
-                  class="h-4 whitespace-nowrap rounded-full px-1.5 text-[10px] leading-4 flex-shrink-0"
-                  :class="getServiceTierBadge(record)?.className"
-                  :title="getServiceTierBadge(record)?.title"
-                  :aria-label="getServiceTierBadge(record)?.ariaLabel"
-                >
-                  {{ getServiceTierBadge(record)?.label }}
-                </Badge>
-              </div>
-              <span class="text-muted-foreground truncate">{{ getActualModel(record) }}</span>
-            </div>
-            <span
-              v-else
-              class="flex min-w-0 items-center gap-1"
-            >
-              <span class="truncate">{{ record.model }}</span>
-              <Badge
-                v-if="getRequestTypeLabel(record)"
-                variant="outline"
-                class="h-4 rounded-full border-sky-500/30 bg-sky-500/5 px-1.5 text-[10px] leading-4 text-sky-700 dark:text-sky-300 flex-shrink-0"
-                title="会话压缩"
-              >
-                {{ getRequestTypeLabel(record) }}
-              </Badge>
-              <Badge
-                v-if="getReasoningEffort(record)"
-                variant="outline"
-                class="h-4 rounded-full border-primary/30 bg-primary/5 px-1.5 text-[10px] leading-4 text-primary flex-shrink-0"
-                :title="getReasoningEffortTitle(record)"
-              >
-                {{ getReasoningEffort(record) }}
-              </Badge>
-              <Badge
-                v-if="getServiceTierBadge(record)"
-                variant="outline"
-                class="h-4 whitespace-nowrap rounded-full px-1.5 text-[10px] leading-4 flex-shrink-0"
-                :class="getServiceTierBadge(record)?.className"
-                :title="getServiceTierBadge(record)?.title"
-                :aria-label="getServiceTierBadge(record)?.ariaLabel"
-              >
-                {{ getServiceTierBadge(record)?.label }}
-              </Badge>
-            </span>
+            <UsageModelDisplay
+              :record="record"
+              class="text-xs"
+            />
           </TableCell>
           <TableCell
             v-if="isAdmin && isColumnVisible('provider')"
@@ -1130,11 +1029,13 @@ import { useDarkMode } from '@/composables/useDarkMode'
 import { API_FORMAT_ORDER, formatApiFormat } from '@/api/endpoints/types/api-format'
 import { formatClientFamily } from '@/features/usage/utils/clientFamily'
 import { formatServiceTierFact } from '../utils/service-tier'
+import { isCyberPolicyError } from '../utils/cyberError'
 import type { DateRangeParams, UsageRecord } from '../types'
 import { MultiSelect, TimeRangePicker } from '@/components/common'
 import type { MultiSelectOption } from '@/components/common/MultiSelect.vue'
 import ElapsedTimeText from './ElapsedTimeText.vue'
 import ServerUserSelector from './ServerUserSelector.vue'
+import UsageModelDisplay from './UsageModelDisplay.vue'
 
 export interface UserOption {
   id: string
@@ -1632,24 +1533,20 @@ function getActualModel(record: UsageRecord): string | null {
 }
 
 function getReasoningEffort(record: UsageRecord): string | null {
-  const effort = record.reasoning_effort?.trim()
-  return effort || null
+  const requested = record.requested_reasoning_effort?.trim()
+  const actual = record.reasoning_effort?.trim()
+  if (requested && actual && requested.toLowerCase() !== actual.toLowerCase()) {
+    return `${requested} -> ${actual}`
+  }
+  return actual || requested || null
 }
 
-function getRequestTypeLabel(record: UsageRecord): string | null {
-  return record.request_type?.trim().toLowerCase() === 'compact' ? '会话压缩' : null
+function hasCyberPolicyError(record: UsageRecord): boolean {
+  return isCyberPolicyError(record.error_message)
 }
 
-function getReasoningEffortTitle(record: UsageRecord): string {
-  const effort = getReasoningEffort(record)
-  return effort ? `Reasoning: ${effort}` : ''
-}
-
-type ServiceTierBadgeState = 'confirmed' | 'downgraded' | 'upgraded' | 'pending' | 'unconfirmed'
-
 interface ServiceTierBadgePresentation {
   label: string
-  state: ServiceTierBadgeState
   className: string
   title: string
   ariaLabel: string
@@ -1670,44 +1567,19 @@ function canonicalServiceTier(value: string | null): string | null {
   return value
 }
 
-function serviceTierBadgeClass(state: ServiceTierBadgeState): string {
-  switch (state) {
-    case 'confirmed':
-      return 'border-emerald-500/40 bg-emerald-500/10 text-emerald-700 dark:text-emerald-300'
-    case 'downgraded':
-      return 'border-amber-500/50 bg-amber-500/10 text-amber-700 dark:text-amber-300'
-    case 'upgraded':
-      return 'border-sky-500/40 bg-sky-500/10 text-sky-700 dark:text-sky-300'
-    case 'pending':
-      return 'border-dashed border-muted-foreground/30 bg-muted/30 text-muted-foreground'
-    case 'unconfirmed':
-      return 'border-dashed border-amber-500/40 bg-amber-500/5 text-amber-700 dark:text-amber-300'
-  }
-}
-
 function buildServiceTierBadgePresentation(
-  label: string,
-  state: ServiceTierBadgeState,
   requestedRaw: string | null,
-  actualRaw: string | null,
-  billingTier: string | null,
 ): ServiceTierBadgePresentation {
   const titleLines: string[] = []
   const requestedLabel = formatServiceTierFact(requestedRaw)
-  const actualLabel = formatServiceTierFact(actualRaw)
-  const billingLabel = formatServiceTierFact(billingTier)
-  if (requestedLabel) titleLines.push(`请求档位：${requestedLabel}`)
-  if (actualLabel) titleLines.push(`实际档位：${actualLabel}`)
-  if (billingLabel) {
-    titleLines.push(`计费档位：${billingLabel}`)
-  } else {
-    titleLines.push(`计费档位：${state === 'pending' ? '待上游确认' : '未确认'}`)
-  }
+  if (requestedLabel) titleLines.push(`上游请求档位：${requestedLabel}`)
+  // Billing is resolved from the same final provider request tier. Keep it
+  // explicit in the tooltip without consulting a response-side tier.
+  if (requestedLabel) titleLines.push(`计费档位：${requestedLabel}`)
   const title = titleLines.join('\n')
   return {
-    label,
-    state,
-    className: serviceTierBadgeClass(state),
+    label: 'Fast',
+    className: '!bg-transparent text-amber-700 dark:text-amber-300',
     title,
     ariaLabel: titleLines.join('，'),
   }
@@ -1715,54 +1587,10 @@ function buildServiceTierBadgePresentation(
 
 function getServiceTierBadge(record: UsageRecord): ServiceTierBadgePresentation | null {
   const requestedRaw = normalizeServiceTier(record.service_tier)
-  const actualRaw = normalizeServiceTier(record.actual_service_tier)
   const requested = canonicalServiceTier(requestedRaw)
-  const actual = canonicalServiceTier(actualRaw)
   const requestedFast = requested === 'priority'
-  const actualFast = actual === 'priority'
-
-  if (actual) {
-    if (requestedFast && !actualFast) {
-      return buildServiceTierBadgePresentation(
-        `Fast → ${actual}`,
-        'downgraded',
-        requestedRaw,
-        actualRaw,
-        actual,
-      )
-    }
-    if (!requestedFast && actualFast) {
-      const requestedLabel = requested ?? 'standard'
-      return buildServiceTierBadgePresentation(
-        requested ? `${requestedLabel} → Fast` : 'Fast',
-        requested ? 'upgraded' : 'confirmed',
-        requestedRaw,
-        actualRaw,
-        actual,
-      )
-    }
-    if (actualFast) {
-      return buildServiceTierBadgePresentation(
-        'Fast',
-        'confirmed',
-        requestedRaw,
-        actualRaw,
-        actual,
-      )
-    }
-    return null
-  }
-
   if (!requestedFast) return null
-  const displayStatus = getDisplayStatus(record)
-  const isActive = displayStatus === 'pending' || displayStatus === 'streaming'
-  return buildServiceTierBadgePresentation(
-    isActive ? 'Fast · 待确认' : 'Fast · 未确认',
-    isActive ? 'pending' : 'unconfirmed',
-    requestedRaw,
-    null,
-    null,
-  )
+  return buildServiceTierBadgePresentation(requestedRaw)
 }
 
 function getServiceTierTitle(record: UsageRecord): string {
@@ -1770,10 +1598,9 @@ function getServiceTierTitle(record: UsageRecord): string {
   if (badge) return badge.title
 
   const requested = formatServiceTierFact(record.service_tier)
-  const actual = formatServiceTierFact(record.actual_service_tier)
   return [
-    requested ? `请求档位：${requested}` : null,
-    actual ? `实际档位：${actual}` : null,
+    requested ? `上游请求档位：${requested}` : null,
+    requested ? `计费档位：${requested}` : null,
   ].filter((line): line is string => Boolean(line)).join('\n')
 }
 
@@ -1781,10 +1608,10 @@ function getServiceTierTitle(record: UsageRecord): string {
 function getModelTooltip(record: UsageRecord): string {
   const actualModel = getActualModel(record)
   const reasoningEffort = getReasoningEffort(record)
-  const requestType = getRequestTypeLabel(record)
   const serviceTierTitle = getServiceTierTitle(record)
   const tierSuffix = serviceTierTitle ? `\n${serviceTierTitle}` : ''
-  const suffix = `${requestType ? `\n操作: ${requestType}` : ''}${reasoningEffort ? `\nReasoning: ${reasoningEffort}` : ''}${tierSuffix}`
+  const cyberSuffix = hasCyberPolicyError(record) ? '\nCyber Policy: blocked' : ''
+  const suffix = `${reasoningEffort ? `\nReasoning: ${reasoningEffort}` : ''}${tierSuffix}${cyberSuffix}`
   if (actualModel) {
     return `${record.model} -> ${actualModel}${suffix}`
   }

frontend/src/features/usage/components/__tests__/ElapsedTimeText.spec.ts

@@ -1,5 +1,5 @@
 import { afterEach, describe, expect, it, vi } from 'vitest'
-import { createApp, h, nextTick, type App } from 'vue'
+import { createApp, h, nextTick, reactive, type App } from 'vue'
 import ElapsedTimeText from '../ElapsedTimeText.vue'
 
 const mountedApps: Array<{ app: App, root: HTMLElement }> = []
@@ -15,6 +15,18 @@ function mountElapsedTimeText(props: Record<string, unknown>) {
   return root
 }
 
+function mountReactiveElapsedTimeText(initialProps: Record<string, unknown>) {
+  const props = reactive(initialProps)
+  const root = document.createElement('div')
+  document.body.appendChild(root)
+  const app = createApp({
+    render: () => h(ElapsedTimeText, { ...props }),
+  })
+  app.mount(root)
+  mountedApps.push({ app, root })
+  return { props, root }
+}
+
 afterEach(() => {
   vi.useRealTimers()
   for (const { app, root } of mountedApps.splice(0)) {
@@ -53,4 +65,30 @@ describe('ElapsedTimeText', () => {
 
     expect(root.textContent).toBe('4.00s')
   })
+
+  it('does not pause or move total time backwards when the first-byte clock arrives', async () => {
+    vi.useFakeTimers()
+    vi.setSystemTime(new Date('2026-07-17T12:00:06.250Z'))
+
+    const { props, root } = mountReactiveElapsedTimeText({
+      status: 'pending',
+      createdAt: '2026-07-17T12:00:00Z',
+      responseTimeUpdatedAt: null,
+      responseTimeMs: null,
+    })
+    await nextTick()
+    expect(root.textContent).toBe('6.25s')
+
+    // The first-byte snapshot implies 5.85s at the same instant because its
+    // timestamp is truncated to seconds. The visible clock must stay continuous.
+    props.status = 'streaming'
+    props.responseTimeUpdatedAt = '2026-07-17T12:00:06Z'
+    props.responseTimeMs = 5600
+    await nextTick()
+    expect(root.textContent).toBe('6.25s')
+
+    vi.advanceTimersByTime(500)
+    await nextTick()
+    expect(Number.parseFloat(root.textContent ?? '')).toBeGreaterThanOrEqual(6.74)
+  })
 })

frontend/src/features/usage/components/__tests__/HorizontalRequestTimeline.spec.ts

@@ -58,9 +58,13 @@ vi.mock('../JsonContentPanel.vue', async () => {
           type: null,
           default: null,
         },
+        title: {
+          type: String,
+          default: 'JSON',
+        },
       },
       setup(props) {
-        return () => h('pre', JSON.stringify(props.data))
+        return () => h('pre', { 'data-title': props.title }, JSON.stringify(props.data))
       },
     }),
   }
@@ -517,7 +521,8 @@ describe('HorizontalRequestTimeline', () => {
     expect(requestPathCode?.textContent).toContain('/v1/images/generations')
   })
 
-  it('shows upstream response JSON inside the error block on trace nodes', async () => {
+  it('shows upstream response headers and body in one error envelope', async () => {
+    const upstreamErrorMessage = 'This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber'
     const trace = buildTrace([
       buildCandidate({
         id: 'cand-upstream-response',
@@ -527,21 +532,33 @@ describe('HorizontalRequestTimeline', () => {
         key_name: 'Upstream Key',
         candidate_index: 0,
         status: 'failed',
-        error_message: 'execution runtime stream returned non-success status 302',
+        error_message: 'execution runtime stream returned non-success status 400',
         extra_data: {
           upstream_response: {
-            status_code: 302,
-            headers: { location: '/' },
+            status_code: 400,
+            headers: {
+              'content-type': 'application/json',
+              'x-request-id': 'req_usage-cyber-risk-demo',
+            },
+            body: {
+              error: {
+                type: 'invalid_request',
+                message: upstreamErrorMessage,
+                code: 400,
+              },
+            },
+            body_ref: 'usage://request/req-1/response_body',
+            body_state: 'reference',
           },
           error_flow: {
             source: 'upstream_response',
-            status_code: 302,
+            status_code: 400,
             classification: 'use_default',
             decision: 'use_default',
             propagation: 'none',
             retryable: false,
             safe_to_expose: false,
-            message: 'execution runtime stream returned non-success status 302',
+            message: 'execution runtime stream returned non-success status 400',
           },
           client_response: {
             status_code: 502,
@@ -555,15 +572,31 @@ describe('HorizontalRequestTimeline', () => {
     await nextTick()
 
     expect(root.textContent).toContain('错误信息')
-    expect(root.textContent).toContain('HTTP 302')
-    expect(root.textContent).not.toContain('上游返回非成功状态 302')
-    expect(root.querySelector('.error-block .error-json')?.textContent).toContain('"status_code":302')
-    expect(root.querySelector('.error-block .error-json')?.textContent).toContain('"headers"')
+    expect(root.textContent).toContain('HTTP 400')
+    expect(root.textContent).not.toContain('上游返回非成功状态 400')
+    const upstreamResponse = root.querySelector<HTMLElement>('.error-upstream-response-json pre')
+    expect(upstreamResponse?.dataset.title).toBe('上游响应')
+    expect(JSON.parse(upstreamResponse?.textContent ?? '{}')).toEqual({
+      header: {
+        'content-type': 'application/json',
+        'x-request-id': 'req_usage-cyber-risk-demo',
+      },
+      body: {
+        error: {
+          type: 'invalid_request',
+          message: upstreamErrorMessage,
+          code: 400,
+        },
+      },
+    })
+    expect(upstreamResponse?.textContent).not.toContain('"status_code"')
+    expect(upstreamResponse?.textContent).not.toContain('"headers"')
+    expect(upstreamResponse?.textContent).not.toContain('"body_ref"')
+    expect(upstreamResponse?.textContent).not.toContain('"body_state"')
     expect(root.textContent).not.toContain('上游真实响应')
-    expect(root.textContent).not.toContain('execution runtime stream returned non-success status 302')
+    expect(root.textContent).not.toContain('execution runtime stream returned non-success status 400')
     expect(root.textContent).not.toContain('真实请求错误')
     expect(root.textContent).not.toContain('返回客户端响应')
-    expect(root.textContent).not.toContain('上游响应')
     expect(root.textContent).not.toContain('默认处理')
     expect(root.textContent).not.toContain('none')
     expect(root.textContent).not.toContain('不再重试')
@@ -601,12 +634,12 @@ describe('HorizontalRequestTimeline', () => {
     expect(root.textContent).toContain('流式格式转换失败')
     expect(root.textContent).toContain('上游返回了当前不支持的 stream event')
     expect(root.textContent).toContain('字段 $.type = "response.future.delta"')
-    const errorJsonText = root.querySelector('.error-block .error-json')?.textContent ?? ''
-    expect(errorJsonText).toContain('"body_state":"disabled"')
-    expect(errorJsonText).toContain('"diagnostic"')
-    expect(errorJsonText).toContain('"breakpoint":"$.type"')
-    expect(errorJsonText).toContain('"analysis_hint"')
-    expect(errorJsonText).toContain('"raw"')
+    const diagnosticText = root.querySelector('.error-diagnostic-json')?.textContent ?? ''
+    expect(diagnosticText).toContain('"breakpoint":"$.type"')
+    expect(diagnosticText).toContain('"analysis_hint"')
+    expect(diagnosticText).toContain('"raw"')
+    expect(diagnosticText).toContain('"body_state":"disabled"')
+    expect(root.querySelector('.error-upstream-response-json')?.textContent).toContain('"header":{"content-type":"application/json"}')
   })
 
   it('formats request conversion diagnostics with field paths on skipped trace nodes', async () => {
@@ -664,9 +697,8 @@ describe('HorizontalRequestTimeline', () => {
     expect(root.textContent).toContain('流式格式转换失败')
     expect(root.textContent).toContain('finish reason')
     expect(root.textContent).toContain('字段 $.finish_reason = "future_reason"')
-    const errorJsonText = root.querySelector('.error-block .error-json')?.textContent ?? ''
-    expect(errorJsonText).toContain('"diagnostic"')
-    expect(errorJsonText).toContain('"breakpoint":"$.finish_reason"')
+    const diagnosticText = root.querySelector('.error-diagnostic-json')?.textContent ?? ''
+    expect(diagnosticText).toContain('"breakpoint":"$.finish_reason"')
   })
 
   it('uses conversion messages from error_flow as the diagnostic breakpoint source', async () => {
@@ -701,10 +733,10 @@ describe('HorizontalRequestTimeline', () => {
     expect(root.textContent).toContain('OpenAI Chat → OpenAI Responses')
     expect(root.textContent).toContain('字段 $.n 会丢失信息')
     expect(root.textContent).not.toContain('上游返回非成功状态 500')
-    const errorJsonText = root.querySelector('.error-block .error-json')?.textContent ?? ''
-    expect(errorJsonText).toContain('"body_state":"disabled"')
-    expect(errorJsonText).toContain('"breakpoint":"$.n"')
-    expect(errorJsonText).toContain('断点在请求/响应格式转换器')
+    const diagnosticText = root.querySelector('.error-diagnostic-json')?.textContent ?? ''
+    expect(diagnosticText).toContain('"body_state":"disabled"')
+    expect(diagnosticText).toContain('"breakpoint":"$.n"')
+    expect(diagnosticText).toContain('断点在请求/响应格式转换器')
   })
 
   it('shows failed diagnostic messages even when the only response panel data is diagnostic metadata', async () => {
@@ -735,9 +767,8 @@ describe('HorizontalRequestTimeline', () => {
     expect(root.textContent).toContain('错误信息')
     expect(root.textContent).toContain('格式转换失败')
     expect(root.textContent).toContain('OpenAI Responses 不支持字段 $.temperature')
-    const errorJsonText = root.querySelector('.error-block .error-json')?.textContent ?? ''
-    expect(errorJsonText).toContain('"diagnostic"')
-    expect(errorJsonText).toContain('"breakpoint":"$.temperature"')
+    const diagnosticText = root.querySelector('.error-diagnostic-json')?.textContent ?? ''
+    expect(diagnosticText).toContain('"breakpoint":"$.temperature"')
   })
 
   it('keeps the failure message when upstream response only records an empty body state', async () => {

frontend/src/features/usage/components/__tests__/RequestDetailDrawer.pricing.spec.ts

@@ -97,4 +97,388 @@ describe('RequestDetailDrawer settlement pricing', () => {
     })
     expect(document.body.textContent).not.toContain('输出 $0/M')
   })
+
+  it('shows mapping, reasoning, Fast, and Cyber together in the model header', async () => {
+    apiMocks.getRequestDetail.mockResolvedValue({
+      ...buildEmbeddingDetail(),
+      id: 'usage-cyber-risk-demo',
+      request_id: 'req_usage-cyber-risk-demo',
+      model: 'gpt-5',
+      target_model: 'gpt-5.1',
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: 'max',
+      request_body: {
+        model: 'gpt-5',
+        reasoning: { effort: 'xhigh' },
+      },
+      service_tier: 'priority',
+      // A response-side tier must not be used for the Fast badge or billing.
+      actual_service_tier: 'default',
+      provider_request_body: {
+        model: 'gpt-5.1',
+        reasoning: { effort: 'max' },
+        service_tier: 'priority',
+      },
+      status: 'failed',
+      status_code: 400,
+      error_message: 'This content was flagged for possible cybersecurity risk. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber',
+      response_body: {
+        error: {
+          type: 'invalid_request',
+          message: 'This content was flagged for possible cybersecurity risk.',
+          code: 400,
+        },
+      },
+    })
+
+    let isOpen!: Ref<boolean>
+    const Host = defineComponent({
+      setup() {
+        isOpen = ref(false)
+        return () => h(RequestDetailDrawer, {
+          isOpen: isOpen.value,
+          requestId: 'usage-cyber-risk-demo',
+        })
+      },
+    })
+
+    const root = document.createElement('div')
+    document.body.appendChild(root)
+    const app = createApp(Host)
+    app.mount(root)
+    mountedApps.push({ app, root })
+
+    isOpen.value = true
+    await nextTick()
+
+    await vi.waitFor(() => {
+      expect(document.body.querySelector('[data-request-detail-model-display]')?.textContent)
+        .toContain('gpt-5')
+      expect(document.body.querySelector('[data-request-detail-model-display]')?.textContent)
+        .toContain('gpt-5.1')
+      expect(document.body.querySelector('[data-request-detail-model-badge="reasoning"]')?.textContent)
+        .toContain('xhigh -> max')
+      expect(document.body.querySelector('[data-request-detail-model-badge="fast"]')?.textContent)
+        .toContain('Fast')
+      expect(document.body.querySelector('[data-request-detail-model-badge="fast"]')?.textContent?.trim())
+        .toBe('Fast')
+      expect(document.body.querySelector('[data-request-detail-model-badge="cyber"]')?.textContent)
+        .toContain('Cyber')
+      const modelLayout = document.body.querySelector(
+        '[data-request-detail-model-layout="stacked"]',
+      )
+      expect(modelLayout?.firstElementChild?.textContent).toContain('gpt-5')
+      expect(modelLayout?.firstElementChild?.textContent).toContain('->')
+      expect(modelLayout?.firstElementChild?.textContent).toContain('gpt-5.1')
+      expect(modelLayout?.firstElementChild?.querySelector('[data-request-detail-model-badge]'))
+        .toBeNull()
+      const modelBadgesRow = modelLayout?.querySelector(
+        '[data-request-detail-model-badges-row]',
+      )
+      expect(modelBadgesRow?.textContent).toContain('xhigh -> max')
+      expect(modelBadgesRow?.textContent).toContain('Fast')
+      expect(modelBadgesRow?.textContent).toContain('Cyber')
+      const serviceTierFacts = document.body.querySelector('[data-testid="service-tier-facts"]')
+      expect([...serviceTierFacts?.querySelectorAll('dt') ?? []].map(node => node.textContent?.trim()))
+        .toEqual(['上游请求层级', '计费层级'])
+      expect([...serviceTierFacts?.querySelectorAll('dd') ?? []].map(node => node.textContent?.trim()))
+        .toEqual(['Fast', 'Fast'])
+    })
+  })
+
+  it('keeps the Cyber badge stable from the selected row while lightweight detail loads', async () => {
+    let resolveDetail!: (value: RequestDetail) => void
+    apiMocks.getRequestDetail.mockReturnValue(new Promise<RequestDetail>((resolve) => {
+      resolveDetail = resolve
+    }))
+
+    let isOpen!: Ref<boolean>
+    const requestId = ref('usage-cyber-summary')
+    const summaryRecord = ref<Record<string, unknown>>({
+      id: 'usage-cyber-summary',
+      model: 'gpt-5',
+      target_model: 'gpt-5.1',
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: 'max',
+      service_tier: 'priority',
+      error_message: 'This content was flagged for possible cybersecurity risk. https://chatgpt.com/cyber',
+    })
+    const Host = defineComponent({
+      setup() {
+        isOpen = ref(false)
+        return () => h(RequestDetailDrawer, {
+          isOpen: isOpen.value,
+          requestId: requestId.value,
+          summaryRecord: summaryRecord.value as never,
+        })
+      },
+    })
+
+    const root = document.createElement('div')
+    document.body.appendChild(root)
+    const app = createApp(Host)
+    app.mount(root)
+    mountedApps.push({ app, root })
+
+    isOpen.value = true
+    await nextTick()
+
+    expect(apiMocks.getRequestDetail).toHaveBeenCalledTimes(1)
+    expect(document.body.querySelector('[data-request-detail-model-badge="cyber"]')?.textContent)
+      .toContain('Cyber')
+    expect(document.body.querySelector('[data-request-detail-model-badge="fast"]')?.textContent)
+      .toContain('Fast')
+
+    resolveDetail({
+      ...buildEmbeddingDetail(),
+      id: 'usage-cyber-summary',
+      request_id: 'usage-cyber-summary',
+      model: 'gpt-5',
+      // Lightweight detail can legitimately omit these final-provider facts.
+      target_model: null,
+      requested_reasoning_effort: null,
+      reasoning_effort: null,
+      service_tier: null,
+      status: 'failed',
+      status_code: 400,
+      error_message: 'execution runtime stream returned non-success status 400',
+      response_body: null,
+    })
+    await vi.waitFor(() => {
+      expect(document.body.querySelector('[data-request-detail-model-badge="cyber"]')?.textContent)
+        .toContain('Cyber')
+      expect(document.body.querySelector('[data-usage-model-target]')?.textContent)
+        .toContain('gpt-5.1')
+      expect(document.body.querySelector('[data-request-detail-model-badge="reasoning"]')?.textContent)
+        .toContain('xhigh -> max')
+      expect(document.body.querySelector('[data-request-detail-model-badge="fast"]')?.textContent)
+        .toContain('Fast')
+      const tierValues = [
+        ...document.body.querySelectorAll('[data-testid="service-tier-facts"] dd'),
+      ].map(node => node.textContent?.trim())
+      expect(tierValues).toEqual(['Fast', 'Fast'])
+    })
+  })
+
+  it('clears a stale summary Cyber badge when newer detail completed successfully', async () => {
+    apiMocks.getRequestDetail.mockResolvedValue({
+      ...buildEmbeddingDetail(),
+      id: 'usage-cyber-recovered',
+      request_id: 'usage-cyber-recovered',
+      model: 'gpt-5',
+      status: 'completed',
+      status_code: 200,
+      error_message: undefined,
+      updated_at: '2026-07-17T00:00:02Z',
+    })
+
+    let isOpen!: Ref<boolean>
+    const Host = defineComponent({
+      setup() {
+        isOpen = ref(false)
+        return () => h(RequestDetailDrawer, {
+          isOpen: isOpen.value,
+          requestId: 'usage-cyber-recovered',
+          summaryRecord: {
+            id: 'usage-cyber-recovered',
+            model: 'gpt-5',
+            input_tokens: 0,
+            output_tokens: 0,
+            total_tokens: 0,
+            cost: 0,
+            is_stream: false,
+            status: 'failed',
+            status_code: 400,
+            error_message: 'This content was flagged for possible cybersecurity risk. https://chatgpt.com/cyber',
+            created_at: '2026-07-17T00:00:00Z',
+            updated_at: '2026-07-17T00:00:01Z',
+          },
+        })
+      },
+    })
+
+    const root = document.createElement('div')
+    document.body.appendChild(root)
+    const app = createApp(Host)
+    app.mount(root)
+    mountedApps.push({ app, root })
+
+    isOpen.value = true
+    await nextTick()
+
+    await vi.waitFor(() => {
+      expect(apiMocks.getRequestDetail).toHaveBeenCalledTimes(1)
+      expect(document.body.querySelector('[data-request-detail-model-badge="cyber"]'))
+        .toBeNull()
+    })
+  })
+
+  it('uses populated detail fallbacks without overriding a populated summary tier', async () => {
+    apiMocks.getRequestDetail.mockResolvedValue({
+      ...buildEmbeddingDetail(),
+      id: 'usage-standard-summary',
+      request_id: 'usage-standard-summary',
+      model: 'gpt-5',
+      target_model: 'gpt-5.1',
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: 'max',
+      // The non-empty summary tier remains authoritative over this stale fact.
+      service_tier: 'priority',
+    })
+
+    let isOpen!: Ref<boolean>
+    const summaryRecord = ref<Record<string, unknown>>({
+      id: 'usage-standard-summary',
+      model: 'gpt-5',
+      target_model: null,
+      model_version: null,
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: null,
+      service_tier: 'default',
+    })
+    const Host = defineComponent({
+      setup() {
+        isOpen = ref(false)
+        return () => h(RequestDetailDrawer, {
+          isOpen: isOpen.value,
+          requestId: 'usage-standard-summary',
+          summaryRecord: summaryRecord.value as never,
+        })
+      },
+    })
+
+    const root = document.createElement('div')
+    document.body.appendChild(root)
+    const app = createApp(Host)
+    app.mount(root)
+    mountedApps.push({ app, root })
+
+    isOpen.value = true
+    await nextTick()
+
+    await vi.waitFor(() => {
+      expect(apiMocks.getRequestDetail).toHaveBeenCalledTimes(1)
+      expect(document.body.querySelector('[data-usage-model-target]')?.textContent?.trim())
+        .toBe('gpt-5.1')
+      expect(document.body.querySelector('[data-request-detail-model-badge="reasoning"]')?.textContent?.trim())
+        .toBe('xhigh -> max')
+      expect(document.body.querySelector('[data-request-detail-model-badge="fast"]')).toBeNull()
+      const tierValues = [
+        ...document.body.querySelectorAll('[data-testid="service-tier-facts"] dd'),
+      ].map(node => node.textContent?.trim())
+      expect(tierValues).toEqual(['default', 'default'])
+    })
+  })
+
+  it('uses detail model_version when the lightweight summary has null model facts', async () => {
+    apiMocks.getRequestDetail.mockResolvedValue({
+      ...buildEmbeddingDetail(),
+      id: 'usage-version-summary',
+      request_id: 'usage-version-summary',
+      model: 'gpt-5',
+      target_model: null,
+      model_version: 'gpt-5.1-2026-07-17',
+    })
+
+    let isOpen!: Ref<boolean>
+    const Host = defineComponent({
+      setup() {
+        isOpen = ref(false)
+        return () => h(RequestDetailDrawer, {
+          isOpen: isOpen.value,
+          requestId: 'usage-version-summary',
+          summaryRecord: {
+            id: 'usage-version-summary',
+            model: 'gpt-5',
+            target_model: null,
+            model_version: null,
+            input_tokens: 0,
+            output_tokens: 0,
+            total_tokens: 0,
+            cost: 0,
+            is_stream: false,
+            created_at: '2026-07-17T00:00:00Z',
+          },
+        })
+      },
+    })
+
+    const root = document.createElement('div')
+    document.body.appendChild(root)
+    const app = createApp(Host)
+    app.mount(root)
+    mountedApps.push({ app, root })
+
+    isOpen.value = true
+    await nextTick()
+
+    await vi.waitFor(() => {
+      expect(document.body.querySelector('[data-usage-model-target]')?.textContent?.trim())
+        .toBe('gpt-5.1-2026-07-17')
+    })
+  })
+
+  it('lets a newer final-provider summary clear facts cached from an earlier candidate', async () => {
+    apiMocks.getRequestDetail.mockResolvedValue({
+      ...buildEmbeddingDetail(),
+      id: 'usage-final-candidate',
+      request_id: 'usage-final-candidate',
+      model: 'gpt-5',
+      target_model: 'gpt-5.1',
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: 'max',
+      service_tier: 'priority',
+      status: 'streaming',
+      updated_at: '2026-07-17T00:00:01Z',
+    })
+
+    let isOpen!: Ref<boolean>
+    const summaryRecord = ref<Record<string, unknown>>({
+      id: 'usage-final-candidate',
+      model: 'gpt-5',
+      target_model: 'gpt-5.1',
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: 'max',
+      service_tier: 'priority',
+      status: 'streaming',
+      updated_at: '2026-07-17T00:00:01Z',
+    })
+    const Host = defineComponent({
+      setup() {
+        isOpen = ref(false)
+        return () => h(RequestDetailDrawer, {
+          isOpen: isOpen.value,
+          requestId: 'usage-final-candidate',
+          summaryRecord: summaryRecord.value as never,
+        })
+      },
+    })
+
+    const root = document.createElement('div')
+    document.body.appendChild(root)
+    const app = createApp(Host)
+    app.mount(root)
+    mountedApps.push({ app, root })
+
+    isOpen.value = true
+    await vi.waitFor(() => {
+      expect(document.body.querySelector('[data-request-detail-model-badge="fast"]')).not.toBeNull()
+    })
+
+    summaryRecord.value = {
+      ...summaryRecord.value,
+      target_model: null,
+      reasoning_effort: null,
+      service_tier: null,
+      updated_at: '2026-07-17T00:00:02Z',
+    }
+    await nextTick()
+
+    expect(document.body.querySelector('[data-usage-model-target]')).toBeNull()
+    expect(document.body.querySelector('[data-request-detail-model-badge="reasoning"]')?.textContent?.trim())
+      .toBe('xhigh')
+    expect(document.body.querySelector('[data-request-detail-model-badge="fast"]')).toBeNull()
+    expect([...document.body.querySelectorAll('[data-testid="service-tier-facts"] dd')]
+      .some(node => node.textContent?.trim() === 'Fast')).toBe(false)
+  })
 })

frontend/src/features/usage/components/__tests__/ServiceTierFacts.spec.ts

@@ -13,14 +13,12 @@ afterEach(() => {
 })
 
 describe('ServiceTierFacts', () => {
-  it('renders all three facts and marks a missing actual tier explicitly', () => {
+  it('renders request and billing facts from the same request tier', () => {
     const root = document.createElement('div')
     document.body.appendChild(root)
     const app = createApp({
       render: () => h(ServiceTierFacts, {
         requested: 'priority',
-        actual: null,
-        billing: 'flex',
       }),
     })
     app.mount(root)
@@ -28,30 +26,25 @@ describe('ServiceTierFacts', () => {
 
     expect(root.querySelector('[data-testid="service-tier-facts"]')).not.toBeNull()
     expect([...root.querySelectorAll('dt')].map(node => node.textContent?.trim())).toEqual([
-      '请求层级',
-      '实际层级',
+      '上游请求层级',
       '计费层级',
     ])
     expect([...root.querySelectorAll('dd')].map(node => node.textContent?.trim())).toEqual([
       'Fast',
-      '-',
-      'flex',
+      'Fast',
     ])
     expect([...root.querySelectorAll('dd')].map(node => node.getAttribute('title'))).toEqual([
       'Fast',
-      '-',
-      'flex',
+      'Fast',
     ])
   })
 
-  it('uses the same Fast label for raw priority and fast facts', () => {
+  it('uses the Fast label for a raw fast request tier', () => {
     const root = document.createElement('div')
     document.body.appendChild(root)
     const app = createApp({
       render: () => h(ServiceTierFacts, {
-        requested: 'priority',
-        actual: 'fast',
-        billing: 'priority',
+        requested: 'fast',
       }),
     })
     app.mount(root)
@@ -60,12 +53,10 @@ describe('ServiceTierFacts', () => {
     expect([...root.querySelectorAll('dd')].map(node => node.textContent?.trim())).toEqual([
       'Fast',
       'Fast',
-      'Fast',
     ])
     expect([...root.querySelectorAll('dd')].map(node => node.getAttribute('title'))).toEqual([
       'Fast',
       'Fast',
-      'Fast',
     ])
   })
 
@@ -75,8 +66,6 @@ describe('ServiceTierFacts', () => {
     const app = createApp({
       render: () => h(ServiceTierFacts, {
         requested: 'priority',
-        actual: 'fast',
-        billing: 'fast',
         priceMultiplier: 2.5,
       }),
     })
@@ -94,8 +83,6 @@ describe('ServiceTierFacts', () => {
     const app = createApp({
       render: () => h(ServiceTierFacts, {
         requested: 'priority',
-        actual: null,
-        billing: null,
         priceMultiplier: null,
       }),
     })

frontend/src/features/usage/components/__tests__/UsageRecordsTable.spec.ts

@@ -290,19 +290,31 @@ describe('UsageRecordsTable', () => {
   })
 
   it('shows reasoning effort next to the model name', () => {
-    const root = mountUsageRecordsTable([buildRecord({ reasoning_effort: 'xhigh' })])
+    const root = mountUsageRecordsTable([buildRecord({
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: 'xhigh',
+      service_tier: 'priority',
+    })])
 
     expect(root.textContent).toContain('gpt-5')
     expect(root.textContent).toContain('xhigh')
+    const inlineLayout = root.querySelector('[data-usage-model-layout="inline"]')
+    expect(inlineLayout).not.toBeNull()
+    expect(inlineLayout?.querySelector('[data-usage-model-badge="reasoning"]')?.textContent?.trim())
+      .toBe('xhigh')
+    expect(inlineLayout?.querySelector('[data-usage-model-badge="fast"]')?.textContent?.trim())
+      .toBe('Fast')
   })
 
   it('shows request reasoning effort while the record is pending', () => {
     const root = mountUsageRecordsTable([buildRecord({
       status: 'pending',
-      reasoning_effort: 'max',
+      requested_reasoning_effort: 'max',
+      reasoning_effort: null,
     })])
 
-    expect(root.textContent).toContain('max')
+    expect(root.querySelector('[data-usage-model-badge="reasoning"]')?.textContent?.trim())
+      .toBe('max')
   })
 
   it('marks Responses compaction while the record is pending', () => {
@@ -311,65 +323,145 @@ describe('UsageRecordsTable', () => {
       request_type: 'compact',
     })])
 
-    expect(root.textContent).toContain('会话压缩')
+    expect(root.querySelector('[data-usage-model-badge="compact"]')?.textContent?.trim())
+      .toBe('会话压缩')
   })
 
-  it.each([
-    ['priority', 'priority'],
-    ['fast', 'fast'],
-    ['priority', 'fast'],
-    ['fast', 'priority'],
-  ])('shows confirmed Fast for requested %s and actual %s', (requested, actual) => {
-    const root = mountUsageRecordsTable([buildRecord({
-      service_tier: requested,
-      actual_service_tier: actual,
-    })])
-
-    const badge = expectServiceTierBadge(root, 'Fast')
-    expect(badge.getAttribute('title')).toBe([
-      '请求档位：Fast',
-      '实际档位：Fast',
-      '计费档位：Fast',
-    ].join('\n'))
-    expect(badge.getAttribute('aria-label')).toBe(
-      '请求档位：Fast，实际档位：Fast，计费档位：Fast',
-    )
-  })
-
-  it('shows fast to standard when the provider downgrades a priority request', () => {
+  it('shows mapping, reasoning, Fast, and Cyber together in the model area', () => {
     const root = mountUsageRecordsTable([buildRecord({
+      model: 'gpt-5',
+      target_model: 'gpt-5.1',
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: 'max',
       service_tier: 'priority',
+      // A conflicting response-side value must not affect the Fast badge.
       actual_service_tier: 'default',
+      status: 'failed',
+      status_code: 400,
+      error_message: 'This content was flagged for possible cybersecurity risk. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber',
     })])
 
-    const badge = expectServiceTierBadge(root, 'Fast → standard')
-    expect(badge.getAttribute('title')).toBe([
-      '请求档位：Fast',
-      '实际档位：default',
-      '计费档位：standard',
-    ].join('\n'))
+    expect(root.textContent).toContain('gpt-5')
+    expect(root.textContent).toContain('gpt-5.1')
+    expect(root.textContent).toContain('xhigh -> max')
+    expect(root.textContent).toContain('Fast')
+    const reasoningBadge = root.querySelector<HTMLElement>('[data-usage-model-badge="reasoning"]')
+    const fastBadge = root.querySelector<HTMLElement>('[data-usage-model-badge="fast"]')
+    const cyberBadges = root.querySelectorAll<HTMLElement>('[data-usage-model-badge="cyber"]')
+    const cyberBadge = cyberBadges[0]
+    for (const badge of [reasoningBadge, fastBadge, cyberBadge]) {
+      expect(badge).not.toBeNull()
+      expect(badge?.classList.contains('h-4')).toBe(true)
+      expect(badge?.classList.contains('rounded-full')).toBe(true)
+      expect(badge?.classList.contains('px-1.5')).toBe(true)
+      expect(badge?.classList.contains('text-[10px]')).toBe(true)
+      expect(badge?.classList.contains('leading-4')).toBe(true)
+    }
+    expect(reasoningBadge?.classList.contains('border-primary/30')).toBe(true)
+    expect(reasoningBadge?.classList.contains('bg-primary/5')).toBe(true)
+    expect(reasoningBadge?.classList.contains('text-primary')).toBe(true)
+    expect(fastBadge?.getAttribute('variant')).toBe('outline-transparent')
+    expect(fastBadge?.classList.contains('border-amber-400/50')).toBe(false)
+    expect(fastBadge?.classList.contains('!bg-transparent')).toBe(false)
+    expect(fastBadge?.classList.contains('bg-amber-400/10')).toBe(false)
+    expect(fastBadge?.classList.contains('text-amber-700')).toBe(true)
+    expect(cyberBadge?.classList.contains('border-primary/30')).toBe(true)
+    expect(cyberBadge?.classList.contains('bg-primary/5')).toBe(true)
+    expect(cyberBadge?.classList.contains('text-rose-600')).toBe(true)
+    expect(cyberBadges.length).toBeGreaterThan(0)
+    expect([...cyberBadges].every(badge => badge.textContent?.trim() === 'Cyber')).toBe(true)
+    expect([...cyberBadges].every(badge => badge.title === '上游 Cyber Policy 拒绝')).toBe(true)
+
+    const stackedLayout = root.querySelector('[data-usage-model-layout="stacked"]')
+    expect(stackedLayout).not.toBeNull()
+    const modelRow = stackedLayout?.firstElementChild
+    expect(modelRow?.textContent).toContain('gpt-5')
+    expect(modelRow?.textContent).toContain('->')
+    expect(modelRow?.textContent).toContain('gpt-5.1')
+    expect(modelRow?.querySelector('[data-usage-model-badge]')).toBeNull()
+    const badgesRow = stackedLayout?.querySelector('[data-usage-model-badges-row]')
+    expect(badgesRow?.textContent).toContain('xhigh -> max')
+    expect(badgesRow?.textContent).toContain('Fast')
+    expect(badgesRow?.textContent).toContain('Cyber')
   })
 
-  it('shows fast to flex when the provider moves a priority request to flex', () => {
+  it('stacks three model badges even without a model mapping', () => {
     const root = mountUsageRecordsTable([buildRecord({
+      model: 'gpt-5',
+      target_model: null,
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: 'xhigh',
       service_tier: 'priority',
-      actual_service_tier: 'flex',
+      status: 'failed',
+      status_code: 400,
+      error_message: 'This content was flagged for possible cybersecurity risk. https://chatgpt.com/cyber',
     })])
 
-    expectServiceTierBadge(root, 'Fast → flex')
+    const stackedLayout = root.querySelector('[data-usage-model-layout="stacked"]')
+    expect(stackedLayout?.firstElementChild?.textContent?.trim()).toBe('gpt-5')
+    expect(stackedLayout?.querySelector('[data-usage-model-badges-row]')?.textContent)
+      .toContain('xhigh')
+    expect(stackedLayout?.querySelector('[data-usage-model-badges-row]')?.textContent)
+      .toContain('Fast')
+    expect(stackedLayout?.querySelector('[data-usage-model-badges-row]')?.textContent)
+      .toContain('Cyber')
   })
 
-  it('shows standard to fast when the provider upgrades a default request', () => {
+  it.each(['priority', 'fast', ' Priority ', 'FAST'])(
+    'shows Fast from the final provider request tier %s',
+    (requested) => {
+      const root = mountUsageRecordsTable([buildRecord({
+        service_tier: requested,
+        actual_service_tier: 'default',
+      })])
+
+      const badge = expectServiceTierBadge(root, 'Fast')
+      expect(badge.getAttribute('title')).toBe([
+        '上游请求档位：Fast',
+        '计费档位：Fast',
+      ].join('\n'))
+      expect(badge.getAttribute('aria-label')).toBe(
+        '上游请求档位：Fast，计费档位：Fast',
+      )
+      expect(badge.textContent).not.toContain('→')
+      expect(badge.textContent).not.toContain('待确认')
+      expect(badge.textContent).not.toContain('未确认')
+    },
+  )
+
+  it.each(['default', 'flex', null])(
+    'ignores the response-side tier %s when the request tier is Fast',
+    (actualServiceTier) => {
+      const root = mountUsageRecordsTable([buildRecord({
+        service_tier: 'priority',
+        actual_service_tier: actualServiceTier,
+      })])
+
+      expectServiceTierBadge(root, 'Fast')
+      expect(root.textContent).not.toContain('Fast →')
+    },
+  )
+
+  it('does not infer Fast from a response-side priority tier', () => {
     const root = mountUsageRecordsTable([buildRecord({
       service_tier: 'default',
       actual_service_tier: 'priority',
     })])
 
-    expectServiceTierBadge(root, 'standard → Fast')
+    expect(root.querySelector('[data-usage-model-badge="fast"]')).toBeNull()
+  })
+
+  it('does not infer Fast when only the response has a tier', () => {
+    const root = mountUsageRecordsTable([buildRecord({
+      service_tier: null,
+      actual_service_tier: 'priority',
+    })])
+
+    expect(root.querySelector('[data-usage-model-badge="fast"]')).toBeNull()
   })
 
   it.each(['pending', 'streaming'] as const)(
-    'shows fast as pending confirmation while a priority request is %s',
+    'keeps Fast stable while a priority request is %s',
     (status) => {
       const root = mountUsageRecordsTable([buildRecord({
         service_tier: 'priority',
@@ -377,18 +469,18 @@ describe('UsageRecordsTable', () => {
         status,
       })])
 
-      expectServiceTierBadge(root, 'Fast · 待确认')
+      expectServiceTierBadge(root, 'Fast')
     },
   )
 
-  it('shows fast as unconfirmed when a completed priority request has no actual tier', () => {
+  it('keeps Fast stable for a completed request without a response tier', () => {
     const root = mountUsageRecordsTable([buildRecord({
       service_tier: 'priority',
       actual_service_tier: null,
       status: 'completed',
     })])
 
-    expectServiceTierBadge(root, 'Fast · 未确认')
+    expectServiceTierBadge(root, 'Fast')
   })
 
   it('offers embedding API formats in the usage record filter', () => {

frontend/src/features/usage/composables/__tests__/useUsageData.spec.ts

@@ -162,7 +162,146 @@ describe('useUsageData', () => {
     })
   })
 
-  it('preserves detail-filled usage metrics when a later list refresh is still empty', async () => {
+  it('clears a failed candidate Cyber snapshot when the final candidate completes', async () => {
+    const isAdminPage = ref(true)
+    const { loadRecords, currentRecords } = useUsageData({ isAdminPage })
+    const dateRange = { preset: 'today', tz_offset_minutes: 0 }
+    const cyberMessage = 'This content was flagged for possible cybersecurity risk. https://chatgpt.com/cyber'
+
+    getAllUsageRecordsMock.mockResolvedValueOnce({
+      records: [buildUsageRecord({
+        status: 'failed',
+        status_code: 400,
+        error_message: cyberMessage,
+        updated_at: '2026-07-17T00:00:01Z',
+      })],
+      total: 1,
+      limit: 20,
+      offset: 0,
+    })
+    await loadRecords({ page: 1, pageSize: 20 }, undefined, dateRange)
+
+    getAllUsageRecordsMock.mockResolvedValueOnce({
+      records: [buildUsageRecord({
+        status: 'completed',
+        status_code: 200,
+        error_message: undefined,
+        updated_at: '2026-07-17T00:00:02Z',
+      })],
+      total: 1,
+      limit: 20,
+      offset: 0,
+    })
+    await loadRecords({ page: 1, pageSize: 20 }, undefined, dateRange)
+
+    expect(currentRecords.value[0]).toMatchObject({
+      status: 'completed',
+      status_code: 200,
+      updated_at: '2026-07-17T00:00:02Z',
+    })
+    expect(currentRecords.value[0]?.error_message).toBeUndefined()
+  })
+
+  it('rejects an older same-rank terminal snapshot as a unit', async () => {
+    const isAdminPage = ref(true)
+    const { loadRecords, currentRecords } = useUsageData({ isAdminPage })
+    const dateRange = { preset: 'today', tz_offset_minutes: 0 }
+    const cyberMessage = 'This content was flagged for possible cybersecurity risk. https://chatgpt.com/cyber'
+
+    getAllUsageRecordsMock.mockResolvedValueOnce({
+      records: [buildUsageRecord({
+        status: 'completed',
+        status_code: 200,
+        error_message: undefined,
+        updated_at: '2026-07-17T00:00:02Z',
+      })],
+      total: 1,
+      limit: 20,
+      offset: 0,
+    })
+    await loadRecords({ page: 1, pageSize: 20 }, undefined, dateRange)
+
+    getAllUsageRecordsMock.mockResolvedValueOnce({
+      records: [buildUsageRecord({
+        status: 'failed',
+        status_code: 400,
+        error_message: cyberMessage,
+        updated_at: '2026-07-17T00:00:01Z',
+      })],
+      total: 1,
+      limit: 20,
+      offset: 0,
+    })
+    await loadRecords({ page: 1, pageSize: 20 }, undefined, dateRange)
+
+    expect(currentRecords.value[0]).toMatchObject({
+      status: 'completed',
+      status_code: 200,
+      updated_at: '2026-07-17T00:00:02Z',
+    })
+    expect(currentRecords.value[0]?.error_message).toBeUndefined()
+  })
+
+  it('keeps live response duration and its anchor atomic across stale refreshes', async () => {
+    const isAdminPage = ref(true)
+    const { loadRecords, currentRecords } = useUsageData({ isAdminPage })
+    const dateRange = { preset: 'today', tz_offset_minutes: 0 }
+
+    getAllUsageRecordsMock.mockResolvedValueOnce({
+      records: [buildUsageRecord({
+        status: 'streaming',
+        response_time_ms: 5500,
+        response_time_updated_at: '2026-07-17T12:00:06Z',
+        first_byte_time_ms: 2000,
+      })],
+      total: 1,
+      limit: 20,
+      offset: 0,
+    })
+    await loadRecords({ page: 1, pageSize: 20 }, undefined, dateRange)
+
+    getAllUsageRecordsMock.mockResolvedValueOnce({
+      records: [buildUsageRecord({
+        status: 'streaming',
+        response_time_ms: 5000,
+        response_time_updated_at: '2026-07-17T12:00:07Z',
+        first_byte_time_ms: 1500,
+      })],
+      total: 1,
+      limit: 20,
+      offset: 0,
+    })
+    await loadRecords({ page: 1, pageSize: 20 }, undefined, dateRange)
+
+    expect(currentRecords.value[0]).toMatchObject({
+      status: 'streaming',
+      response_time_ms: 5500,
+      response_time_updated_at: '2026-07-17T12:00:06Z',
+      first_byte_time_ms: 2000,
+    })
+
+    getAllUsageRecordsMock.mockResolvedValueOnce({
+      records: [buildUsageRecord({
+        status: 'completed',
+        response_time_ms: 5200,
+        response_time_updated_at: '2026-07-17T12:00:07Z',
+        first_byte_time_ms: 1800,
+      })],
+      total: 1,
+      limit: 20,
+      offset: 0,
+    })
+    await loadRecords({ page: 1, pageSize: 20 }, undefined, dateRange)
+
+    expect(currentRecords.value[0]).toMatchObject({
+      status: 'completed',
+      response_time_ms: 5200,
+      response_time_updated_at: '2026-07-17T12:00:07Z',
+      first_byte_time_ms: 2000,
+    })
+  })
+
+  it('preserves client/detail metrics but clears stale final-provider facts from the next list snapshot', async () => {
     const isAdminPage = ref(true)
     const { loadRecords, currentRecords } = useUsageData({ isAdminPage })
     const dateRange = { preset: 'today', tz_offset_minutes: 0 }
@@ -217,9 +356,10 @@ describe('useUsageData', () => {
       has_retry: true,
       target_model: 'gpt-5.5',
       request_type: 'compact',
-      reasoning_effort: 'xhigh',
-      service_tier: 'auto',
-      actual_service_tier: 'priority',
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: 'max',
+      service_tier: 'priority',
+      actual_service_tier: 'default',
     })
 
     getAllUsageRecordsMock.mockResolvedValueOnce({
@@ -245,11 +385,12 @@ describe('useUsageData', () => {
         endpoint_api_format: undefined,
         has_format_conversion: undefined,
         has_retry: false,
-        target_model: null,
+        target_model: undefined,
         request_type: null,
-        reasoning_effort: null,
-        service_tier: null,
-        actual_service_tier: null,
+        requested_reasoning_effort: null,
+        reasoning_effort: undefined,
+        service_tier: undefined,
+        actual_service_tier: undefined,
       })],
       total: 1,
       limit: 20,
@@ -280,11 +421,12 @@ describe('useUsageData', () => {
       endpoint_api_format: 'openai:responses',
       has_format_conversion: false,
       has_retry: true,
-      target_model: 'gpt-5.5',
+      target_model: null,
       request_type: 'compact',
-      reasoning_effort: 'xhigh',
-      service_tier: 'auto',
-      actual_service_tier: 'priority',
+      requested_reasoning_effort: 'xhigh',
+      reasoning_effort: null,
+      service_tier: null,
+      actual_service_tier: null,
     })
   })
 

frontend/src/features/usage/composables/useUsageData.ts

@@ -14,6 +14,12 @@ import { createDefaultStats } from '../types'
 import { log } from '@/utils/logger'
 import { getErrorStatus } from '@/types/api-error'
 import { isUsageProviderVisible, normalizeUsageProviderStats } from '../utils/providerStats'
+import {
+  mergeUsageRecordErrorMessage,
+  mergeUsageRecordFirstByteTimeMs,
+  mergeUsageRecordResponseTiming,
+  parseUsageTimestampMs,
+} from '../utils/recordSync'
 
 export interface UseUsageDataOptions {
   isAdminPage: Ref<boolean>
@@ -464,25 +470,6 @@ export function useUsageData(options: UseUsageDataOptions) {
     }
   }
 
-  function mergePositiveDurationMs(
-    existingValue: number | null | undefined,
-    nextValue: number | null | undefined
-  ): number | null | undefined {
-    const existingIsPositive = typeof existingValue === 'number' && Number.isFinite(existingValue) && existingValue > 0
-    const nextIsPositive = typeof nextValue === 'number' && Number.isFinite(nextValue) && nextValue > 0
-
-    if (existingIsPositive && nextIsPositive) {
-      return Math.max(existingValue, nextValue)
-    }
-    if (existingIsPositive) {
-      return existingValue
-    }
-    if (nextIsPositive) {
-      return nextValue
-    }
-    return existingValue ?? nextValue
-  }
-
   function mergeSparseRecordMetric(
     existingValue: number | null | undefined,
     nextValue: number | null | undefined
@@ -539,10 +526,18 @@ export function useUsageData(options: UseUsageDataOptions) {
       const hasNextStatus = typeof record.status === 'string' && record.status.length > 0
       const currentRank = hasExistingStatus ? (statusPriority[existing.status] ?? -1) : -1
       const nextRank = hasNextStatus ? (statusPriority[record.status] ?? -1) : -1
-      const statusProgressed = hasNextStatus && (
+      const existingUpdatedAtMs = parseUsageTimestampMs(existing.updated_at)
+      const nextUpdatedAtMs = parseUsageTimestampMs(record.updated_at)
+      const nextStatusSnapshotIsStale = existingUpdatedAtMs != null &&
+        nextUpdatedAtMs != null &&
+        nextUpdatedAtMs < existingUpdatedAtMs
+      const sameRankTerminalTransition = currentRank === 2 && nextRank === 2
+      const statusProgressed = hasNextStatus && !nextStatusSnapshotIsStale && (
         !hasExistingStatus ||
         nextRank > currentRank ||
-        (nextRank === currentRank && existing.status === record.status)
+        (nextRank === currentRank && (
+          existing.status === record.status || sameRankTerminalTransition
+        ))
       )
       const mergedStatus = statusProgressed ? record.status : existing.status
 
@@ -586,12 +581,27 @@ export function useUsageData(options: UseUsageDataOptions) {
           ? existing.client_requested_stream
           : undefined
       const clientIsStream = mergeBooleanTrueWins(existingClientIsStream, recordClientIsStream) ?? clientRequestedStream
+      const nextTimingIsAuthoritative = statusProgressed &&
+        (record.status === 'completed' || record.status === 'failed' || record.status === 'cancelled')
+      const responseTiming = mergeUsageRecordResponseTiming(
+        {
+          response_time_ms: existing.response_time_ms,
+          response_time_updated_at: existing.response_time_updated_at,
+        },
+        {
+          response_time_ms: record.response_time_ms,
+          response_time_updated_at: record.response_time_updated_at,
+        },
+        { preferNext: nextTimingIsAuthoritative },
+      )
 
       return {
         ...record,
         // 保留详情抽屉/活跃轮询已经拿到的完整指标，避免列表刷新用 0 或空值回退。
         status: mergedStatus,
-        provider: protectProvider ? existing.provider : (record.provider || existing.provider),
+        provider: statusProgressed
+          ? (protectProvider ? existing.provider : (record.provider || existing.provider))
+          : existing.provider,
         input_tokens: mergeSparseRecordMetric(existing.input_tokens, record.input_tokens) ?? record.input_tokens,
         effective_input_tokens: mergeSparseRecordMetric(existing.effective_input_tokens, record.effective_input_tokens) ?? record.effective_input_tokens,
         output_tokens: mergeSparseRecordMetric(existing.output_tokens, record.output_tokens) ?? record.output_tokens,
@@ -610,13 +620,31 @@ export function useUsageData(options: UseUsageDataOptions) {
         cache_read_input_tokens: mergeSparseRecordMetric(existing.cache_read_input_tokens, record.cache_read_input_tokens) ?? record.cache_read_input_tokens,
         cost: mergeSparseRecordMetric(existing.cost, record.cost) ?? record.cost,
         actual_cost: mergeSparseRecordMetric(existing.actual_cost, record.actual_cost) ?? record.actual_cost,
-        response_time_ms: mergePositiveDurationMs(existing.response_time_ms, record.response_time_ms),
-        first_byte_time_ms: mergePositiveDurationMs(existing.first_byte_time_ms, record.first_byte_time_ms),
-        updated_at: record.updated_at ?? existing.updated_at,
-        response_time_updated_at: record.response_time_updated_at ?? existing.response_time_updated_at,
-        status_code: record.status_code ?? existing.status_code,
-        error_message: record.error_message ?? existing.error_message,
-        image_progress: record.image_progress ?? existing.image_progress,
+        response_time_ms: responseTiming.response_time_ms,
+        first_byte_time_ms: mergeUsageRecordFirstByteTimeMs(
+          existing.first_byte_time_ms,
+          record.first_byte_time_ms,
+        ),
+        updated_at: statusProgressed
+          ? (record.updated_at ?? existing.updated_at)
+          : existing.updated_at,
+        response_time_updated_at: responseTiming.response_time_updated_at,
+        // Status, code and error are one lifecycle snapshot. An accepted full
+        // list snapshot may clear an earlier candidate's 400/Cyber failure;
+        // a rejected stale status snapshot must not mutate either field.
+        status_code: statusProgressed
+          ? (record.status_code ?? undefined)
+          : existing.status_code,
+        error_message: statusProgressed
+          ? mergeUsageRecordErrorMessage(
+            existing.error_message,
+            record.error_message,
+            { authoritative: true },
+          )
+          : existing.error_message,
+        image_progress: statusProgressed
+          ? (record.image_progress ?? existing.image_progress)
+          : existing.image_progress,
         is_stream: upstreamIsStream,
         upstream_is_stream: upstreamIsStream,
         client_requested_stream: clientRequestedStream,
@@ -627,13 +655,50 @@ export function useUsageData(options: UseUsageDataOptions) {
         has_fallback: existing.has_fallback === true || record.has_fallback === true,
         has_retry: existing.has_retry === true || record.has_retry === true,
         api_key_name: record.api_key_name || existing.api_key_name,
-        provider_key_name: record.provider_key_name || existing.provider_key_name,
-        rate_multiplier: record.rate_multiplier ?? existing.rate_multiplier,
-        target_model: record.target_model ?? existing.target_model,
-        request_type: record.request_type ?? existing.request_type,
-        reasoning_effort: record.reasoning_effort ?? existing.reasoning_effort,
-        service_tier: record.service_tier ?? existing.service_tier,
-        actual_service_tier: record.actual_service_tier ?? existing.actual_service_tier
+        provider_key_name: statusProgressed
+          ? (record.provider_key_name || existing.provider_key_name)
+          : existing.provider_key_name,
+        rate_multiplier: statusProgressed
+          ? (record.rate_multiplier ?? existing.rate_multiplier)
+          : existing.rate_multiplier,
+        // Full list snapshots describe the final provider candidate. Missing/null means the
+        // final request did not map the model and must clear an earlier candidate's arrow.
+        target_model: statusProgressed
+          ? (typeof record.target_model === 'string' && record.target_model.trim()
+              ? record.target_model
+              : null)
+          : existing.target_model,
+        // Request type is client-request identity, not a provider-candidate fact. Preserve a
+        // known compact operation when a later sparse snapshot omits it.
+        request_type:
+          typeof record.request_type === 'string' && record.request_type.trim()
+            ? record.request_type
+            : existing.request_type,
+        requested_reasoning_effort:
+          typeof record.requested_reasoning_effort === 'string'
+            && record.requested_reasoning_effort.trim()
+            ? record.requested_reasoning_effort
+            : existing.requested_reasoning_effort,
+        // Provider reasoning belongs to the final candidate just like service_tier; do not
+        // retain a previous candidate's `max` when the final request has no reasoning field.
+        reasoning_effort: statusProgressed
+          ? (typeof record.reasoning_effort === 'string' && record.reasoning_effort.trim()
+              ? record.reasoning_effort
+              : null)
+          : existing.reasoning_effort,
+        // The list response is the authoritative snapshot of the final provider request. Do not
+        // carry a tier forward when this response has no tier; doing so can leave a stale Fast
+        // badge after the final upstream request falls back to Standard.
+        service_tier: statusProgressed
+          ? (typeof record.service_tier === 'string' && record.service_tier.trim()
+              ? record.service_tier
+              : null)
+          : existing.service_tier,
+        actual_service_tier: statusProgressed
+          ? (typeof record.actual_service_tier === 'string' && record.actual_service_tier.trim()
+              ? record.actual_service_tier
+              : null)
+          : existing.actual_service_tier
       }
     })
   }

frontend/src/features/usage/types.ts

@@ -97,9 +97,10 @@ export interface UsageRecord {
   target_model?: string | null  // 映射后的目标模型名（若无映射则为空）
   model_version?: string | null  // Provider 返回的实际模型版本（列表轻量字段）
   request_type?: string | null  // 由请求语义识别出的操作类型
+  requested_reasoning_effort?: string | null  // 用户请求侧 reasoning 级别，用于展示转换关系
   reasoning_effort?: string | null  // 从发送给 Provider 的请求体提取的 reasoning 级别
   service_tier?: string | null  // 从发送给 Provider 的请求体提取的服务层级
-  actual_service_tier?: string | null  // Provider 响应确认的实际服务层级
+  actual_service_tier?: string | null  // 响应侧审计事实，不用于 Fast 展示或计费
   api_format?: string
   endpoint_api_format?: string  // 端点原生格式
   has_format_conversion?: boolean  // 是否发生了格式转换

frontend/src/features/usage/utils/__tests__/cyberError.spec.ts

@@ -0,0 +1,45 @@
+import { describe, expect, it } from 'vitest'
+
+import { isCyberPolicyError } from '../cyberError'
+
+describe('isCyberPolicyError', () => {
+  it('recognizes the provider cybersecurity refusal message', () => {
+    expect(isCyberPolicyError({
+      error: {
+        type: 'invalid_request',
+        message: 'This content was flagged for possible cybersecurity risk. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber',
+        code: 400,
+      },
+    })).toBe(true)
+  })
+
+  it('recognizes an explicit cyber_policy code', () => {
+    expect(isCyberPolicyError({ error: { code: 'CYBER_POLICY' } })).toBe(true)
+  })
+
+  it('recognizes explicit Cyber Policy types and reasons', () => {
+    expect(isCyberPolicyError({ error: { type: 'cyber_policy' } })).toBe(true)
+    expect(isCyberPolicyError({ error: { type: 'CYBER' } })).toBe(true)
+    expect(isCyberPolicyError({ error: { reason: 'cyber-policy' } })).toBe(true)
+    expect(isCyberPolicyError({ error: { category: 'cyber_policy_violation' } })).toBe(true)
+    expect(isCyberPolicyError({ error: { type: 'cybersecurity-risk' } })).toBe(true)
+  })
+
+  it('recognizes structured Cyber classifiers inside a serialized error', () => {
+    expect(isCyberPolicyError('{"error":{"type":"cyber"}}')).toBe(true)
+  })
+
+  it('does not classify ordinary invalid requests as Cyber Policy failures', () => {
+    expect(isCyberPolicyError({
+      error: {
+        type: 'invalid_request',
+        message: 'The request payload is malformed',
+        code: 400,
+      },
+    })).toBe(false)
+  })
+
+  it('does not classify a generic use of the word cyber', () => {
+    expect(isCyberPolicyError('The cyber security report was generated successfully')).toBe(false)
+  })
+})

frontend/src/features/usage/utils/__tests__/recordSync.spec.ts

@@ -2,7 +2,10 @@ import { describe, expect, it } from 'vitest'
 
 import type { UsageRecord } from '../../types'
 import {
+  mergeUsageRecordErrorMessage,
   mergeUsageRecordFirstByteTimeMs,
+  mergeUsageRecordLifecycleSnapshot,
+  mergeUsageRecordResponseTiming,
   syncUsageRecordStreamResolution,
 } from '../recordSync'
 
@@ -75,3 +78,140 @@ describe('mergeUsageRecordFirstByteTimeMs', () => {
     expect(mergeUsageRecordFirstByteTimeMs(-1, null)).toBeUndefined()
   })
 })
+
+describe('mergeUsageRecordResponseTiming', () => {
+  it('keeps duration and update timestamp as one monotonic active snapshot', () => {
+    const existing = {
+      response_time_ms: 5500,
+      response_time_updated_at: '2026-07-17T12:00:06Z',
+    }
+    const stale = {
+      response_time_ms: 5000,
+      response_time_updated_at: '2026-07-17T12:00:07Z',
+    }
+
+    expect(mergeUsageRecordResponseTiming(existing, stale)).toBe(existing)
+  })
+
+  it('accepts a live snapshot whose projected elapsed time has advanced', () => {
+    const existing = {
+      response_time_ms: 5500,
+      response_time_updated_at: '2026-07-17T12:00:06Z',
+    }
+    const advanced = {
+      response_time_ms: 7000,
+      response_time_updated_at: '2026-07-17T12:00:07Z',
+    }
+
+    expect(mergeUsageRecordResponseTiming(existing, advanced)).toBe(advanced)
+  })
+
+  it('does not combine an unanchored detail estimate with an existing anchor', () => {
+    const existing = {
+      response_time_ms: 5500,
+      response_time_updated_at: '2026-07-17T12:00:06Z',
+    }
+    const detailEstimate = {
+      response_time_ms: 6000,
+      response_time_updated_at: null,
+    }
+
+    expect(mergeUsageRecordResponseTiming(existing, detailEstimate)).toBe(existing)
+  })
+
+  it('lets a terminal snapshot replace the active estimate', () => {
+    const terminal = {
+      response_time_ms: 5200,
+      response_time_updated_at: null,
+    }
+
+    expect(mergeUsageRecordResponseTiming({
+      response_time_ms: 5500,
+      response_time_updated_at: '2026-07-17T12:00:06Z',
+    }, terminal, { preferNext: true })).toBe(terminal)
+  })
+})
+
+describe('mergeUsageRecordErrorMessage', () => {
+  const cyberMessage = 'This content was flagged for possible cybersecurity risk. Join the Trusted Access for Cyber program: https://chatgpt.com/cyber'
+
+  it('keeps an authoritative Cyber Policy message when trace reports a generic error', () => {
+    expect(mergeUsageRecordErrorMessage(
+      cyberMessage,
+      'execution runtime stream ended with a terminal error',
+    )).toBe(cyberMessage)
+  })
+
+  it('keeps an existing error when trace omits its error message', () => {
+    expect(mergeUsageRecordErrorMessage(cyberMessage, undefined)).toBe(cyberMessage)
+    expect(mergeUsageRecordErrorMessage(cyberMessage, null)).toBe(cyberMessage)
+    expect(mergeUsageRecordErrorMessage(cyberMessage, '   ')).toBe(cyberMessage)
+  })
+
+  it('accepts a Cyber Policy message discovered by trace', () => {
+    expect(mergeUsageRecordErrorMessage('Request failed', cyberMessage)).toBe(cyberMessage)
+  })
+
+  it('updates ordinary errors when the next snapshot has a more specific message', () => {
+    expect(mergeUsageRecordErrorMessage('Request failed', 'rate limit exceeded'))
+      .toBe('rate limit exceeded')
+    expect(mergeUsageRecordErrorMessage(undefined, 'Request failed')).toBe('Request failed')
+  })
+
+  it('lets an authoritative final-candidate snapshot replace or clear Cyber', () => {
+    expect(mergeUsageRecordErrorMessage(
+      cyberMessage,
+      'rate limit exceeded',
+      { authoritative: true },
+    )).toBe('rate limit exceeded')
+    expect(mergeUsageRecordErrorMessage(
+      cyberMessage,
+      null,
+      { authoritative: true },
+    )).toBeUndefined()
+  })
+})
+
+describe('mergeUsageRecordLifecycleSnapshot', () => {
+  const cyberMessage = 'This content was flagged for possible cybersecurity risk. https://chatgpt.com/cyber'
+
+  it('rejects an older failed detail without changing status, code, or error', () => {
+    expect(mergeUsageRecordLifecycleSnapshot({
+      status: 'completed',
+      status_code: 200,
+      error_message: undefined,
+      updated_at: '2026-07-17T00:00:02Z',
+    }, {
+      status: 'failed',
+      statusCode: 400,
+      errorMessage: cyberMessage,
+      updatedAt: '2026-07-17T00:00:01Z',
+    })).toEqual({
+      status: 'completed',
+      status_code: 200,
+      error_message: undefined,
+      updated_at: '2026-07-17T00:00:02Z',
+      accepted: false,
+    })
+  })
+
+  it('accepts a newer completed detail and clears an earlier Cyber failure', () => {
+    expect(mergeUsageRecordLifecycleSnapshot({
+      status: 'failed',
+      status_code: 400,
+      error_message: cyberMessage,
+      updated_at: '2026-07-17T00:00:01Z',
+    }, {
+      status: 'completed',
+      statusCode: 200,
+      errorMessage: null,
+      updatedAt: '2026-07-17T00:00:02Z',
+    })).toEqual({
+      status: 'completed',
+      status_code: 200,
+      error_message: undefined,
+      updated_at: '2026-07-17T00:00:02Z',
+      accepted: true,
+    })
+  })
+})

frontend/src/features/usage/utils/__tests__/service-tier.spec.ts

@@ -8,8 +8,8 @@ import {
 } from '../service-tier'
 
 describe('service tier facts', () => {
-  it('keeps requested, actual and billing tiers independent', () => {
-    const facts = resolveServiceTierFacts({
+  it('uses the final provider request tier for display and billing', () => {
+    const source = {
       service_tier: 'priority',
       actual_service_tier: 'default',
       settlement: {
@@ -21,17 +21,28 @@ describe('service tier facts', () => {
           },
         },
       },
-    })
+    }
+    const facts = resolveServiceTierFacts(source)
 
-    expect(facts).toEqual({ requested: 'priority', actual: 'default', billing: 'standard' })
+    expect(facts).toEqual({ requested: 'priority' })
     expect(hasServiceTierFact(facts)).toBe(true)
   })
 
-  it('does not infer billing from requested or actual tiers', () => {
-    expect(resolveServiceTierFacts({
-      service_tier: 'priority',
+  it('does not infer a tier from the provider response or settlement snapshot', () => {
+    const source = {
       actual_service_tier: 'flex',
-    })).toEqual({ requested: 'priority', actual: 'flex', billing: null })
+      settlement: {
+        settlement_snapshot: {
+          pricing_snapshot: {
+            billing_processing_tier: 'priority',
+          },
+        },
+      },
+    }
+
+    const facts = resolveServiceTierFacts(source)
+    expect(facts).toEqual({ requested: null })
+    expect(hasServiceTierFact(facts)).toBe(false)
   })
 
   it('normalizes only non-empty string facts', () => {

frontend/src/features/usage/utils/cyberError.ts

@@ -0,0 +1,70 @@
+const CYBER_POLICY_TEXT_MARKERS = [
+  'possible cybersecurity risk',
+  'trusted access for cyber',
+  'chatgpt.com/cyber',
+]
+
+const CYBER_POLICY_CLASSIFIER_FIELDS = [
+  'code',
+  'type',
+  'category',
+  'reason',
+] as const
+
+const CYBER_ERROR_OBJECT_FIELDS = [
+  'error',
+  'errors',
+  'message',
+  'error_message',
+  'detail',
+  'body',
+  'response_body',
+  'upstream_error',
+  'failure_summary',
+] as const
+
+function normalizeCyberClassifier(value: string): string {
+  return value.trim().toLowerCase().replace(/[\s-]+/g, '_')
+}
+
+function isCyberPolicyClassifier(value: unknown): boolean {
+  if (typeof value !== 'string') return false
+  const normalized = normalizeCyberClassifier(value)
+  if (normalized === 'cyber' || normalized === 'cyber_policy') return true
+
+  // Providers have used nearby classifier spellings while keeping the same
+  // structured error contract. Keep this deliberately narrower than a generic
+  // substring check so ordinary cybersecurity content is not badged.
+  return /^(?:cyber|cybersecurity)_(?:policy|safety|risk)(?:_(?:violation|error|refusal|blocked))?$/.test(normalized)
+}
+
+function isCyberPolicyText(value: string): boolean {
+  const normalized = value.trim().toLowerCase()
+  if (!normalized) return false
+  return CYBER_POLICY_TEXT_MARKERS.some(marker => normalized.includes(marker))
+    || /["'](?:code|type|category|reason)["']\s*:\s*["'](?:cyber|cyber[-_ ]policy|cyber[-_ ]safety|cybersecurity[-_ ](?:policy|risk))["']/i.test(normalized)
+}
+
+function detectCyberPolicyError(value: unknown, seen: WeakSet<object>): boolean {
+  if (typeof value === 'string') return isCyberPolicyText(value)
+  if (value === null || typeof value !== 'object') return false
+  if (seen.has(value)) return false
+  seen.add(value)
+
+  if (Array.isArray(value)) return value.some(item => detectCyberPolicyError(item, seen))
+
+  const record = value as Record<string, unknown>
+  if (CYBER_POLICY_CLASSIFIER_FIELDS.some(field => isCyberPolicyClassifier(record[field]))) {
+    return true
+  }
+
+  return CYBER_ERROR_OBJECT_FIELDS.some(field => detectCyberPolicyError(record[field], seen))
+}
+
+/**
+ * Detects the provider's Cyber Policy refusal without treating generic HTTP 400,
+ * invalid_request, or ordinary uses of the word "cyber" as policy failures.
+ */
+export function isCyberPolicyError(value: unknown): boolean {
+  return detectCyberPolicyError(value, new WeakSet<object>())
+}

frontend/src/features/usage/utils/recordSync.ts

@@ -1,10 +1,63 @@
-import type { UsageRecord } from '../types'
+import type { RequestStatus, UsageRecord } from '../types'
+import { isCyberPolicyError } from './cyberError'
 
 export type UsageRecordStreamResolution = Pick<
   UsageRecord,
   'id' | 'is_stream' | 'upstream_is_stream' | 'client_requested_stream' | 'client_is_stream'
 >
 
+export type UsageRecordResponseTiming = Pick<
+  UsageRecord,
+  'response_time_ms' | 'response_time_updated_at'
+>
+
+function finiteNonNegativeDurationMs(value: number | null | undefined): number | null {
+  return typeof value === 'number' && Number.isFinite(value) && value >= 0 ? value : null
+}
+
+export function parseUsageTimestampMs(value: string | null | undefined): number | null {
+  if (!value) return null
+  const normalized = /(?:Z|[+-]\d{2}:\d{2})$/i.test(value) ? value : `${value}Z`
+  const timestampMs = new Date(normalized).getTime()
+  return Number.isFinite(timestampMs) ? timestampMs : null
+}
+
+/**
+ * Merge a live response duration together with the timestamp that anchors it.
+ *
+ * These fields form one clock snapshot: active elapsed time is projected as
+ * `response_time_ms + (now - response_time_updated_at)`. Merging the larger
+ * duration with a newer timestamp can therefore manufacture a shorter clock
+ * that never existed. Keep the pair atomic and, while active, retain whichever
+ * snapshot projects the larger elapsed value.
+ */
+export function mergeUsageRecordResponseTiming(
+  existing: UsageRecordResponseTiming,
+  next: UsageRecordResponseTiming,
+  options: { preferNext?: boolean } = {},
+): UsageRecordResponseTiming {
+  const existingDurationMs = finiteNonNegativeDurationMs(existing.response_time_ms)
+  const nextDurationMs = finiteNonNegativeDurationMs(next.response_time_ms)
+
+  if (nextDurationMs == null) return existing
+  if (options.preferNext || existingDurationMs == null) return next
+
+  const existingUpdatedAtMs = parseUsageTimestampMs(existing.response_time_updated_at)
+  const nextUpdatedAtMs = parseUsageTimestampMs(next.response_time_updated_at)
+
+  if (existingUpdatedAtMs != null && nextUpdatedAtMs != null) {
+    const existingStartedAtMs = existingUpdatedAtMs - existingDurationMs
+    const nextStartedAtMs = nextUpdatedAtMs - nextDurationMs
+    return nextStartedAtMs <= existingStartedAtMs ? next : existing
+  }
+
+  // An anchored snapshot is safer than an unanchored duration for a live clock.
+  if (existingUpdatedAtMs != null) return existing
+  if (nextUpdatedAtMs != null) return next
+
+  return nextDurationMs >= existingDurationMs ? next : existing
+}
+
 export function mergeUsageRecordFirstByteTimeMs(
   existingValue: number | null | undefined,
   nextValue: number | null | undefined
@@ -29,6 +82,109 @@ export function mergeUsageRecordFirstByteTimeMs(
   return existingValue == null ? existingValue : undefined
 }
 
+export function mergeUsageRecordErrorMessage(
+  existingValue: string | null | undefined,
+  nextValue: string | null | undefined,
+  options: { authoritative?: boolean } = {},
+): string | undefined {
+  const existing = typeof existingValue === 'string' && existingValue.trim()
+    ? existingValue
+    : undefined
+  const next = typeof nextValue === 'string' && nextValue.trim()
+    ? nextValue
+    : undefined
+
+  // Complete list/active snapshots describe the current final candidate. They
+  // must be able to replace *and clear* an error left by an earlier candidate.
+  if (options.authoritative) return next
+
+  if (!next) return existing
+
+  // Detail/trace snapshots may carry a generic runtime message. Do not let that
+  // downgrade a provider Cyber Policy refusal already resolved by the usage list.
+  if (isCyberPolicyError(existing) && !isCyberPolicyError(next)) return existing
+
+  return next
+}
+
+export type UsageRecordLifecycleSnapshot = Pick<
+  UsageRecord,
+  'status' | 'status_code' | 'error_message' | 'updated_at'
+>
+
+export type UsageRecordLifecycleUpdate = {
+  status?: RequestStatus
+  statusCode?: number | null
+  errorMessage?: string | null
+  updatedAt?: string | null
+}
+
+/**
+ * Merge the sparse lifecycle state emitted by the detail drawer.
+ *
+ * Status, status code and error belong to one snapshot. If a detail response
+ * is older (or its status would regress), none of those fields may leak into
+ * the newer row. A completed/cancelled or explicitly newer terminal snapshot
+ * is authoritative for errors; a same-snapshot generic failure still keeps a
+ * provider Cyber refusal already known by the list.
+ */
+export function mergeUsageRecordLifecycleSnapshot(
+  existing: UsageRecordLifecycleSnapshot,
+  update: UsageRecordLifecycleUpdate,
+): UsageRecordLifecycleSnapshot & { accepted: boolean } {
+  const statusPriority: Record<RequestStatus, number> = {
+    pending: 0,
+    streaming: 1,
+    completed: 2,
+    failed: 2,
+    cancelled: 2,
+  }
+  const existingUpdatedAtMs = parseUsageTimestampMs(existing.updated_at)
+  const nextUpdatedAtMs = parseUsageTimestampMs(update.updatedAt)
+  const nextSnapshotIsOlder = existingUpdatedAtMs != null &&
+    nextUpdatedAtMs != null &&
+    nextUpdatedAtMs < existingUpdatedAtMs
+  const currentRank = existing.status ? statusPriority[existing.status] : -1
+  const nextRank = update.status ? statusPriority[update.status] : -1
+  const statusAccepted = update.status != null &&
+    !nextSnapshotIsOlder &&
+    nextRank >= currentRank
+  const accepted = !nextSnapshotIsOlder && (update.status == null || statusAccepted)
+
+  if (!accepted) {
+    return { ...existing, accepted: false }
+  }
+
+  const terminalSnapshotIsStrictlyNewer = statusAccepted &&
+    (update.status === 'completed' || update.status === 'failed' || update.status === 'cancelled') &&
+    existingUpdatedAtMs != null &&
+    nextUpdatedAtMs != null &&
+    nextUpdatedAtMs > existingUpdatedAtMs
+  const errorIsAuthoritative = statusAccepted && (
+    update.status === 'completed' ||
+    update.status === 'cancelled' ||
+    terminalSnapshotIsStrictlyNewer
+  )
+  const hasStatusCode = Object.prototype.hasOwnProperty.call(update, 'statusCode')
+  const hasErrorMessage = Object.prototype.hasOwnProperty.call(update, 'errorMessage')
+
+  return {
+    status: statusAccepted ? update.status : existing.status,
+    status_code: hasStatusCode ? (update.statusCode ?? undefined) : existing.status_code,
+    error_message: hasErrorMessage
+      ? mergeUsageRecordErrorMessage(
+        existing.error_message,
+        update.errorMessage,
+        { authoritative: errorIsAuthoritative },
+      )
+      : existing.error_message,
+    updated_at: typeof update.updatedAt === 'string'
+      ? update.updatedAt
+      : existing.updated_at,
+    accepted: true,
+  }
+}
+
 export function syncUsageRecordStreamResolution(
   records: UsageRecord[],
   resolved: UsageRecordStreamResolution

frontend/src/features/usage/utils/service-tier.ts

@@ -1,30 +1,26 @@
 export interface ServiceTierFacts {
   requested: string | null
-  actual: string | null
-  billing: string | null
 }
 
 export interface ServiceTierFactSource {
   service_tier?: unknown
-  actual_service_tier?: unknown
-  settlement?: unknown
 }
 
 export function resolveServiceTierFacts(
   source: ServiceTierFactSource | null | undefined,
 ): ServiceTierFacts {
-  const settlement = asRecord(source?.settlement)
-  const settlementSnapshot = asRecord(settlement?.settlement_snapshot)
-  const pricingSnapshot = asRecord(settlementSnapshot?.pricing_snapshot)
+  // The processing tier is an input-side fact: it must come from the final
+  // request body sent to the provider. Response-advertised tiers and old
+  // settlement snapshots can describe a different/legacy value, so they are
+  // deliberately not consulted here. The billing display uses this same
+  // authoritative request tier.
   return {
     requested: normalizeServiceTierFact(source?.service_tier),
-    actual: normalizeServiceTierFact(source?.actual_service_tier),
-    billing: normalizeServiceTierFact(pricingSnapshot?.billing_processing_tier),
   }
 }
 
 export function hasServiceTierFact(facts: ServiceTierFacts): boolean {
-  return facts.requested !== null || facts.actual !== null || facts.billing !== null
+  return facts.requested !== null
 }
 
 export function normalizeServiceTierFact(value: unknown): string | null {
@@ -47,9 +43,3 @@ export function formatServiceTierFact(value: unknown): string | null {
     ? 'Fast'
     : normalized
 }
-
-function asRecord(value: unknown): Record<string, unknown> | null {
-  return value !== null && typeof value === 'object' && !Array.isArray(value)
-    ? value as Record<string, unknown>
-    : null
-}

frontend/src/features/users/__tests__/apiKeyFeatureSettings.spec.ts

@@ -0,0 +1,69 @@
+import { describe, expect, it } from 'vitest'
+
+import {
+  buildApiKeyRedactionFeatureSettingsPatch,
+  resolveApiKeyRedactionFormState,
+} from '../apiKeyFeatureSettings'
+
+const redaction = {
+  enabled: true,
+  inject_model_instruction: false,
+}
+
+describe('managed API key feature setting inheritance', () => {
+  it('uses the target user value for an inherited key form', () => {
+    expect(resolveApiKeyRedactionFormState(null, {
+      chat_pii_redaction: redaction,
+    })).toEqual({
+      mode: 'inherit',
+      ...redaction,
+    })
+  })
+
+  it('omits feature_settings when a created key keeps inheritance', () => {
+    expect(buildApiKeyRedactionFeatureSettingsPatch({
+      isEditing: false,
+      currentFeatureSettings: undefined,
+      mode: 'inherit',
+      value: redaction,
+    })).toEqual({})
+  })
+
+  it('writes an override only when custom mode is selected', () => {
+    expect(buildApiKeyRedactionFeatureSettingsPatch({
+      isEditing: false,
+      currentFeatureSettings: undefined,
+      mode: 'custom',
+      value: redaction,
+    })).toEqual({
+      feature_settings: {
+        chat_pii_redaction: redaction,
+      },
+    })
+  })
+
+  it('does not create an override when only another field of an inherited key changes', () => {
+    expect(buildApiKeyRedactionFeatureSettingsPatch({
+      isEditing: true,
+      currentFeatureSettings: null,
+      mode: 'inherit',
+      value: redaction,
+    })).toEqual({})
+  })
+
+  it('removes only the existing redaction override when inheritance is restored', () => {
+    expect(buildApiKeyRedactionFeatureSettingsPatch({
+      isEditing: true,
+      currentFeatureSettings: {
+        chat_pii_redaction: { enabled: false },
+        notification_push_service: { enabled: true },
+      },
+      mode: 'inherit',
+      value: redaction,
+    })).toEqual({
+      feature_settings: {
+        notification_push_service: { enabled: true },
+      },
+    })
+  })
+})

frontend/src/features/users/apiKeyFeatureSettings.ts

@@ -0,0 +1,64 @@
+import type { FeatureSettings } from '@/api/users'
+import {
+  hasChatPiiRedactionFeatureSettings,
+  mergeChatPiiRedactionFeatureSettings,
+  readChatPiiRedactionFeatureSettings,
+  removeChatPiiRedactionFeatureSettings,
+  type ChatPiiRedactionFeatureSettings,
+} from '@/utils/featureSettings'
+
+export type ApiKeyRedactionMode = 'inherit' | 'custom'
+
+export interface ApiKeyRedactionFormState extends ChatPiiRedactionFeatureSettings {
+  mode: ApiKeyRedactionMode
+}
+
+export function resolveApiKeyRedactionFormState(
+  apiKeyFeatureSettings: FeatureSettings | null | undefined,
+  inheritedUserFeatureSettings: FeatureSettings | null | undefined,
+): ApiKeyRedactionFormState {
+  const hasCustomRedaction = hasChatPiiRedactionFeatureSettings(apiKeyFeatureSettings)
+  const value = readChatPiiRedactionFeatureSettings(
+    hasCustomRedaction ? apiKeyFeatureSettings : inheritedUserFeatureSettings,
+  )
+  return {
+    mode: hasCustomRedaction ? 'custom' : 'inherit',
+    ...value,
+  }
+}
+
+/**
+ * Builds only the feature-settings portion of an API-key mutation.
+ *
+ * An omitted field preserves inheritance. `null` (or an object with the
+ * redaction key removed) is emitted only when an existing custom override is
+ * explicitly switched back to inheritance.
+ */
+export function buildApiKeyRedactionFeatureSettingsPatch(options: {
+  isEditing: boolean
+  currentFeatureSettings: FeatureSettings | null | undefined
+  mode: ApiKeyRedactionMode
+  value: ChatPiiRedactionFeatureSettings
+}): { feature_settings?: FeatureSettings | null } {
+  if (options.mode === 'custom') {
+    return {
+      feature_settings: mergeChatPiiRedactionFeatureSettings(
+        options.isEditing ? options.currentFeatureSettings : null,
+        options.value,
+      ),
+    }
+  }
+
+  if (
+    options.isEditing
+    && hasChatPiiRedactionFeatureSettings(options.currentFeatureSettings)
+  ) {
+    return {
+      feature_settings: removeChatPiiRedactionFeatureSettings(
+        options.currentFeatureSettings,
+      ),
+    }
+  }
+
+  return {}
+}

frontend/src/features/users/components/UserApiKeyFormDialog.vue

@@ -104,15 +104,47 @@
 
       <div class="space-y-3 rounded-lg border border-border bg-muted/30 p-3">
         <div class="flex items-center justify-between gap-3">
+          <div>
+            <Label class="text-sm font-medium">
+              {{ legacyT('敏感信息保护') }}
+            </Label>
+            <p class="mt-1 text-xs text-muted-foreground">
+              {{ legacyT(form.chat_pii_redaction_mode === 'inherit' ? '跟随目标用户设置' : '仅覆盖此 API Key') }}
+            </p>
+          </div>
+          <div class="flex items-center gap-2">
+            <Button
+              size="sm"
+              :variant="form.chat_pii_redaction_mode === 'inherit' ? 'default' : 'outline'"
+              @click="updateField('chat_pii_redaction_mode', 'inherit')"
+            >
+              {{ legacyT('跟随用户') }}
+            </Button>
+            <Button
+              size="sm"
+              :variant="form.chat_pii_redaction_mode === 'custom' ? 'default' : 'outline'"
+              @click="updateField('chat_pii_redaction_mode', 'custom')"
+            >
+              {{ legacyT('单独配置') }}
+            </Button>
+          </div>
+        </div>
+        <div
+          v-if="form.chat_pii_redaction_mode === 'custom'"
+          class="flex items-center justify-between gap-3 border-t border-border/50 pt-3"
+        >
           <Label class="text-sm font-medium">
-            {{ legacyT('敏感信息保护') }}
+            {{ legacyT('启用保护') }}
           </Label>
           <Switch
             :model-value="form.chat_pii_redaction_enabled"
             @update:model-value="updateField('chat_pii_redaction_enabled', $event)"
           />
         </div>
-        <div class="flex items-center justify-between gap-3">
+        <div
+          v-if="form.chat_pii_redaction_mode === 'custom' && form.chat_pii_redaction_enabled"
+          class="flex items-center justify-between gap-3 border-t border-border/50 pt-3"
+        >
           <Label class="text-sm font-medium">
             {{ legacyT('占位符说明') }}
           </Label>
@@ -156,6 +188,7 @@ export interface UserApiKeyFormState {
   rate_limit?: number
   concurrent_limit?: number
   ip_rules_text: string
+  chat_pii_redaction_mode: 'inherit' | 'custom'
   chat_pii_redaction_enabled: boolean
   chat_pii_redaction_placeholder_notice: boolean
 }

frontend/src/mocks/__tests__/pool-quota-contract.spec.ts

@@ -0,0 +1,79 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+
+vi.mock('@/config/demo', () => ({
+  isDemoMode: () => true,
+  DEMO_ACCOUNTS: {
+    admin: { email: 'admin@demo.aether.io', password: 'demo123' },
+    user: { email: 'user@demo.aether.io', password: 'demo123' },
+  },
+}))
+
+import type { QuotaWindowSnapshot } from '@/api/endpoints/types'
+import { getCodexQuotaWindowPresentation } from '@/utils/codexQuotaWindow'
+import { handleMockRequest, setMockUserToken } from '../handler'
+
+interface MockPoolKey {
+  key_id: string
+  oauth_plan_type: string
+  status_snapshot: {
+    quota: {
+      windows: QuotaWindowSnapshot[]
+    }
+  }
+}
+
+describe('pool quota demo contracts', () => {
+  beforeEach(() => {
+    setMockUserToken('demo-access-token-admin')
+  })
+
+  it('exposes a dedicated Codex pool in the overview and provider summary', async () => {
+    const overviewResponse = await handleMockRequest({
+      method: 'GET',
+      url: '/api/admin/pool/overview',
+    })
+    const overview = overviewResponse?.data as {
+      items: Array<{ provider_id: string; provider_type: string; total_keys: number }>
+    }
+    const provider = overview.items[0]
+
+    expect(provider).toMatchObject({
+      provider_id: 'provider-codex-pool-demo',
+      provider_type: 'codex',
+      total_keys: 4,
+    })
+
+    const summaryResponse = await handleMockRequest({
+      method: 'GET',
+      url: `/api/admin/providers/${provider.provider_id}/summary`,
+    })
+    expect(summaryResponse?.data).toMatchObject({
+      id: provider.provider_id,
+      provider_type: 'codex',
+      name: 'Codex 周期额度演示',
+    })
+  })
+
+  it('covers dual, weekly-only, monthly-only, and 5H-only quota windows', async () => {
+    const response = await handleMockRequest({
+      method: 'GET',
+      url: '/api/admin/pool/provider-codex-pool-demo/keys',
+      params: { page: 1, page_size: 50, status: 'all' },
+    })
+    const page = response?.data as { total: number; keys: MockPoolKey[] }
+    const keys = new Map(page.keys.map(key => [key.key_id, key]))
+    const labelsFor = (keyId: string) => keys.get(keyId)?.status_snapshot.quota.windows
+      .map(getCodexQuotaWindowPresentation)
+      .filter((item): item is NonNullable<typeof item> => item != null)
+      .sort((left, right) => left.sortOrder - right.sortOrder)
+      .map(item => item.label)
+
+    expect(page.total).toBe(4)
+    expect(labelsFor('codex-pool-plus-dual')).toEqual(['5H', '周'])
+    expect(labelsFor('codex-pool-team-weekly')).toEqual(['周'])
+    expect(labelsFor('codex-pool-business-monthly')).toEqual(['月'])
+    expect(labelsFor('codex-pool-free-five-hour')).toEqual(['5H'])
+    expect(keys.get('codex-pool-business-monthly')?.oauth_plan_type)
+      .toBe('self_serve_business_usage_based')
+  })
+})

frontend/src/mocks/__tests__/provider-detail-contract.spec.ts

@@ -0,0 +1,72 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+
+vi.mock('@/config/demo', () => ({
+  isDemoMode: () => true,
+  DEMO_ACCOUNTS: {
+    admin: { email: 'admin@demo.aether.io', password: 'demo123' },
+    user: { email: 'user@demo.aether.io', password: 'demo123' },
+  },
+}))
+
+import { handleMockRequest, setMockUserToken } from '../handler'
+
+describe('provider detail demo contracts', () => {
+  beforeEach(() => {
+    setMockUserToken('demo-access-token-admin')
+  })
+
+  it('returns the paginated key contract used by the provider drawer', async () => {
+    const firstPageResponse = await handleMockRequest({
+      method: 'GET',
+      url: '/api/admin/endpoints/providers/provider-004/keys',
+      params: { page: 1, page_size: 1 },
+    })
+    const secondPageResponse = await handleMockRequest({
+      method: 'GET',
+      url: '/api/admin/endpoints/providers/provider-004/keys',
+      params: { page: 2, page_size: 1 },
+    })
+    const firstPage = firstPageResponse?.data as {
+      total: number
+      page: number
+      page_size: number
+      keys: Array<{ id: string }>
+    }
+    const secondPage = secondPageResponse?.data as typeof firstPage
+
+    expect(firstPage).toMatchObject({ total: 2, page: 1, page_size: 1 })
+    expect(firstPage.keys).toHaveLength(1)
+    expect(secondPage).toMatchObject({ total: 2, page: 2, page_size: 1 })
+    expect(secondPage.keys).toHaveLength(1)
+    expect(secondPage.keys[0]?.id).not.toBe(firstPage.keys[0]?.id)
+  })
+
+  it('preserves the legacy array contract for skip/limit callers', async () => {
+    const response = await handleMockRequest({
+      method: 'GET',
+      url: '/api/admin/endpoints/providers/provider-004/keys',
+      params: { skip: 1, limit: 1 },
+    })
+
+    expect(Array.isArray(response?.data)).toBe(true)
+    expect(response?.data).toHaveLength(1)
+  })
+
+  it('returns a complete mapping-preview envelope', async () => {
+    const response = await handleMockRequest({
+      method: 'GET',
+      url: '/api/admin/providers/provider-004/mapping-preview',
+    })
+
+    expect(response?.data).toEqual({
+      provider_id: 'provider-004',
+      provider_name: 'IKunCode',
+      keys: [],
+      total_keys: 0,
+      total_matches: 0,
+      truncated: false,
+      truncated_keys: 0,
+      truncated_models: 0,
+    })
+  })
+})

frontend/src/mocks/__tests__/usage-detail-contract.spec.ts

@@ -0,0 +1,52 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+
+vi.mock('@/config/demo', () => ({
+  isDemoMode: () => true,
+  DEMO_ACCOUNTS: {
+    admin: { email: 'admin@demo.aether.io', password: 'demo123' },
+    user: { email: 'user@demo.aether.io', password: 'demo123' },
+  },
+}))
+
+import { handleMockRequest, setMockUserToken } from '../handler'
+
+describe('usage detail demo contracts', () => {
+  beforeEach(() => {
+    setMockUserToken('demo-access-token-admin')
+  })
+
+  it('keeps body availability while omitting bodies from lightweight detail', async () => {
+    const response = await handleMockRequest({
+      method: 'GET',
+      url: '/api/admin/usage/usage-cyber-risk-demo',
+      params: { include_bodies: false },
+    })
+
+    expect(response?.data).toMatchObject({
+      has_request_body: true,
+      has_provider_request_body: true,
+      has_response_body: true,
+      request_body: null,
+      provider_request_body: null,
+      response_body: null,
+      client_response_body: null,
+      body_load_errors: null,
+    })
+  })
+
+  it('returns the exact Cyber error body when bodies are requested', async () => {
+    const response = await handleMockRequest({
+      method: 'GET',
+      url: '/api/admin/usage/usage-cyber-risk-demo',
+      params: { include_bodies: true },
+    })
+
+    expect(response?.data?.response_body).toEqual({
+      error: {
+        type: 'invalid_request',
+        message: 'This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber',
+        code: 400,
+      },
+    })
+  })
+})

frontend/src/mocks/__tests__/user-management-contract.spec.ts

@@ -0,0 +1,62 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+
+vi.mock('@/config/demo', () => ({
+  isDemoMode: () => true,
+  DEMO_ACCOUNTS: {
+    admin: { email: 'admin@demo.aether.io', password: 'demo123' },
+    user: { email: 'user@demo.aether.io', password: 'demo123' },
+  },
+}))
+
+import { handleMockRequest, setMockUserToken } from '../handler'
+
+describe('user management demo contracts', () => {
+  beforeEach(() => {
+    setMockUserToken('demo-access-token-admin')
+  })
+
+  it('returns a list-shaped user group response', async () => {
+    const response = await handleMockRequest({
+      method: 'GET',
+      url: '/api/admin/user-groups',
+    })
+
+    expect(response?.data).toEqual({
+      items: [],
+      default_group_id: null,
+    })
+  })
+
+  it('creates and lists managed keys only for the selected target user', async () => {
+    const aliceId = 'demo-user-uuid-0003'
+    const bobId = 'demo-user-uuid-0004'
+    const created = await handleMockRequest({
+      method: 'POST',
+      url: `/api/admin/users/${aliceId}/api-keys`,
+      data: JSON.stringify({ name: 'Alice inherited key' }),
+    })
+
+    expect(created?.data).toMatchObject({
+      name: 'Alice inherited key',
+      feature_settings: null,
+      is_standalone: false,
+    })
+    expect(created?.data?.key).toMatch(/^sk-ae-demo-/)
+
+    const aliceKeys = await handleMockRequest({
+      method: 'GET',
+      url: `/api/admin/users/${aliceId}/api-keys`,
+    })
+    const bobKeys = await handleMockRequest({
+      method: 'GET',
+      url: `/api/admin/users/${bobId}/api-keys`,
+    })
+
+    expect(aliceKeys?.data).toMatchObject({ total: 1 })
+    expect(aliceKeys?.data?.api_keys).toHaveLength(1)
+    expect(aliceKeys?.data?.api_keys[0]).toMatchObject({ name: 'Alice inherited key' })
+    expect(aliceKeys?.data?.api_keys[0]).not.toHaveProperty('key')
+    expect(aliceKeys?.data?.api_keys[0]).not.toHaveProperty('fullKey')
+    expect(bobKeys?.data).toEqual({ api_keys: [], total: 0 })
+  })
+})

frontend/src/mocks/handler.ts

@@ -705,6 +705,77 @@ function getActivityHeatmap() {
   return cachedHeatmap
 }
 
+const MOCK_CYBER_POLICY_USAGE_ID = 'usage-cyber-risk-demo'
+const MOCK_CYBER_POLICY_ERROR_MESSAGE = 'This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber'
+const MOCK_CYBER_POLICY_ERROR_BODY = {
+  error: {
+    type: 'invalid_request',
+    message: MOCK_CYBER_POLICY_ERROR_MESSAGE,
+    code: 400
+  }
+}
+
+interface MockManagedUserApiKey {
+  id: string
+  fullKey: string
+  key_display: string
+  name: string
+  created_at: string
+  last_used_at?: string
+  is_active: boolean
+  is_locked: boolean
+  is_standalone: false
+  feature_settings?: Record<string, unknown> | null
+  rate_limit?: number | null
+  concurrent_limit?: number | null
+  ip_rules?: string[] | null
+  total_requests: number
+  total_cost_usd: number
+  force_capabilities?: Record<string, unknown> | null
+}
+
+const mockManagedUserApiKeysByUserId = new Map<string, MockManagedUserApiKey[]>([
+  [MOCK_NORMAL_USER.id ?? '', MOCK_USER_API_KEYS.map((key, index) => ({
+    ...key,
+    fullKey: `sk-ae-demo-user-${index + 1}`,
+    is_locked: false,
+    is_standalone: false as const,
+  }))],
+])
+let mockManagedUserApiKeySequence = 0
+
+function mockManagedUserApiKeys(userId: string): MockManagedUserApiKey[] {
+  if (!MOCK_ALL_USERS.some(user => user.id === userId)) {
+    throw { response: createMockResponse({ detail: '用户不存在' }, 404) }
+  }
+  let keys = mockManagedUserApiKeysByUserId.get(userId)
+  if (!keys) {
+    keys = []
+    mockManagedUserApiKeysByUserId.set(userId, keys)
+  }
+  return keys
+}
+
+function publicMockManagedUserApiKey(key: MockManagedUserApiKey) {
+  const { fullKey: _fullKey, ...publicKey } = key
+  void _fullKey
+  return publicKey
+}
+
+function mockRequestObject(config: AxiosRequestConfig): Record<string, unknown> {
+  if (typeof config.data === 'string') {
+    try {
+      const parsed = JSON.parse(config.data)
+      return parsed && typeof parsed === 'object' && !Array.isArray(parsed) ? parsed : {}
+    } catch {
+      return {}
+    }
+  }
+  return config.data && typeof config.data === 'object' && !Array.isArray(config.data)
+    ? config.data as Record<string, unknown>
+    : {}
+}
+
 // 生成更真实的使用记录
 function generateMockUsageRecords(count: number = 100) {
   const records = []
@@ -798,6 +869,47 @@ function generateMockUsageRecords(count: number = 100) {
     })
   }
 
+  // 固定在首屏的失败记录，用于预览候选链路中的实际上游错误响应。
+  records.unshift({
+    id: MOCK_CYBER_POLICY_USAGE_ID,
+    user_id: 'demo-admin-uuid-0001',
+    username: 'Demo Admin',
+    user_email: 'admin@demo.aether.ai',
+    api_key: {
+      id: 'key-demo-cyber-risk',
+      name: 'OpenAI Cyber Risk Demo',
+      display: 'sk-ae...demo'
+    },
+    provider: 'openai',
+    api_key_name: 'openai-cyber-risk-demo',
+    rate_multiplier: 1.0,
+    model: 'gpt-5',
+    target_model: 'gpt-5.1',
+    requested_reasoning_effort: 'xhigh',
+    reasoning_effort: 'max',
+    service_tier: 'priority',
+    // Deliberately conflicts with the final provider request. UI and billing
+    // must use the request-side `service_tier`, never this response fact.
+    actual_service_tier: 'default',
+    api_format: 'openai:responses',
+    input_tokens: 0,
+    output_tokens: 0,
+    cache_creation_input_tokens: 0,
+    cache_read_input_tokens: 0,
+    total_tokens: 0,
+    cost: 0,
+    actual_cost: 0,
+    response_time_ms: 428,
+    is_stream: true,
+    status_code: 400,
+    error_message: MOCK_CYBER_POLICY_ERROR_MESSAGE,
+    status: 'failed',
+    created_at: new Date(now).toISOString(),
+    updated_at: new Date(now).toISOString(),
+    has_fallback: false,
+    model_version: undefined
+  })
+
   return records
 }
 
@@ -979,6 +1091,204 @@ const MOCK_CAPABILITIES = [
   { name: 'context_1m', display_name: '1M上下文', description: '支持1M上下文窗口', match_mode: 'compatible', short_name: '1M' }
 ]
 
+const MOCK_CODEX_POOL_PROVIDER_ID = 'provider-codex-pool-demo'
+const MOCK_CODEX_POOL_PROVIDER = {
+  id: MOCK_CODEX_POOL_PROVIDER_ID,
+  name: 'Codex 周期额度演示',
+  provider_type: 'codex',
+  description: '展示 5H、周、月及组合额度窗口',
+  website: 'https://openai.com/codex',
+  provider_priority: 0,
+  billing_type: 'free_tier',
+  monthly_used_usd: 0,
+  is_active: true,
+  total_endpoints: 1,
+  active_endpoints: 1,
+  total_keys: 4,
+  active_keys: 4,
+  total_models: 3,
+  active_models: 3,
+  avg_health_score: 0.97,
+  unhealthy_endpoints: 0,
+  api_formats: ['openai:responses'],
+  endpoint_health_details: [
+    { api_format: 'openai:responses', health_score: 0.97, is_active: true, active_keys: 4 }
+  ],
+  pool_advanced: {
+    enabled: true,
+    probing_enabled: true,
+  },
+  claude_code_advanced: null,
+  proxy: null,
+  created_at: '2026-07-01T00:00:00Z',
+  updated_at: new Date().toISOString(),
+}
+
+function createMockCodexQuotaWindow(
+  code: string,
+  label: string,
+  windowMinutes: number,
+  remainingRatio: number,
+  resetSeconds: number,
+  observedAt: number,
+  requestCount: number,
+) {
+  return {
+    code,
+    label,
+    scope: 'account',
+    unit: 'percent',
+    used_ratio: 1 - remainingRatio,
+    remaining_ratio: remainingRatio,
+    reset_at: resetSeconds > 0 ? observedAt + resetSeconds : null,
+    reset_seconds: resetSeconds,
+    window_minutes: windowMinutes,
+    usage: {
+      request_count: requestCount,
+      total_tokens: requestCount * 1250,
+      total_cost_usd: (requestCount * 0.0025).toFixed(8),
+    },
+  }
+}
+
+function createMockCodexPoolKeys() {
+  const nowSeconds = Math.floor(Date.now() / 1000)
+  const common = {
+    provider_type: 'codex',
+    is_active: true,
+    auth_type: 'oauth',
+    credential_kind: 'oauth_session',
+    runtime_auth_kind: 'bearer',
+    oauth_managed: true,
+    oauth_header_auth: true,
+    can_refresh_oauth: true,
+    can_export_oauth: true,
+    can_edit_oauth: true,
+    oauth_expires_at: nowSeconds + 14 * 24 * 3600,
+    api_formats: ['openai:responses'],
+    rate_multipliers: null,
+    internal_priority: 50,
+    rpm_limit: null,
+    cache_ttl_minutes: 5,
+    max_probe_interval_minutes: 32,
+    health_score: 0.97,
+    circuit_breaker_open: false,
+    proxy: null,
+    cooldown_reason: null,
+    cooldown_ttl_seconds: null,
+    cost_window_usage: 0,
+    cost_limit: null,
+    sticky_sessions: 0,
+    lru_score: null,
+    created_at: '2026-07-01T00:00:00Z',
+    imported_at: '2026-07-01T00:00:00Z',
+    last_used_at: new Date(nowSeconds * 1000 - 10 * 60 * 1000).toISOString(),
+    scheduling_status: 'available',
+    scheduling_reason: 'available',
+    scheduling_label: '可调度',
+    scheduling_reasons: [],
+  }
+
+  const buildKey = (
+    keyId: string,
+    keyName: string,
+    planType: string,
+    accountQuota: string,
+    windows: ReturnType<typeof createMockCodexQuotaWindow>[],
+    requestCount: number,
+  ) => ({
+    ...common,
+    key_id: keyId,
+    key_name: keyName,
+    oauth_plan_type: planType,
+    oauth_account_id: `acct-${keyId}`,
+    oauth_account_name: keyName,
+    quota_updated_at: nowSeconds - 10 * 60,
+    account_quota: accountQuota,
+    request_count: requestCount,
+    total_tokens: requestCount * 2400,
+    total_cost_usd: (requestCount * 0.004).toFixed(8),
+    status_snapshot: {
+      oauth: {
+        code: 'valid',
+        label: '有效',
+        expires_at: nowSeconds + 14 * 24 * 3600,
+        requires_reauth: false,
+        expiring_soon: false,
+      },
+      account: {
+        code: 'ok',
+        label: null,
+        reason: null,
+        blocked: false,
+        source: null,
+        recoverable: false,
+      },
+      quota: {
+        version: 2,
+        provider_type: 'codex',
+        code: 'ok',
+        label: null,
+        reason: null,
+        freshness: 'fresh',
+        source: 'response_headers',
+        observed_at: nowSeconds,
+        updated_at: nowSeconds,
+        exhausted: false,
+        usage_ratio: windows.reduce((max, window) => Math.max(max, window.used_ratio), 0),
+        plan_type: planType,
+        credits: { has_credits: false, unlimited: false },
+        windows,
+      },
+    },
+  })
+
+  return [
+    buildKey(
+      'codex-pool-plus-dual',
+      'Plus · 5H + 周',
+      'plus',
+      '5H剩余 62.0% | 周剩余 84.0%',
+      [
+        createMockCodexQuotaWindow('5h', '5H', 300, 0.62, 3 * 3600, nowSeconds, 18),
+        createMockCodexQuotaWindow('weekly', '周', 10_080, 0.84, 5 * 24 * 3600, nowSeconds, 42),
+      ],
+      128,
+    ),
+    buildKey(
+      'codex-pool-team-weekly',
+      'Team · 仅周',
+      'team',
+      '周剩余 71.0%',
+      [
+        createMockCodexQuotaWindow('weekly', '周', 10_080, 0.71, 4 * 24 * 3600, nowSeconds, 31),
+      ],
+      96,
+    ),
+    buildKey(
+      'codex-pool-business-monthly',
+      'Codex · 仅月（含空占位）',
+      'self_serve_business_usage_based',
+      '月剩余 86.0%',
+      [
+        createMockCodexQuotaWindow('monthly', '月', 43_800, 0.86, 2_627_672, nowSeconds, 54),
+        createMockCodexQuotaWindow('weekly', '周', 0, 1, 0, nowSeconds, 0),
+      ],
+      214,
+    ),
+    buildKey(
+      'codex-pool-free-five-hour',
+      'Free · 仅5H',
+      'free',
+      '5H剩余 93.0%',
+      [
+        createMockCodexQuotaWindow('5h', '5H', 300, 0.93, 4 * 3600, nowSeconds, 7),
+      ],
+      37,
+    ),
+  ]
+}
+
 /**
  * Mock API 路由处理器
  */
@@ -1330,6 +1640,15 @@ const mockHandlers: Record<string, (config: AxiosRequestConfig) => Promise<Axios
     return createMockResponse(MOCK_ALL_USERS)
   },
 
+  'GET /api/admin/user-groups': async () => {
+    await delay()
+    requireAdmin()
+    return createMockResponse({
+      items: [],
+      default_group_id: null,
+    })
+  },
+
   'POST /api/admin/users': async (config) => {
     await delay()
     requireAdmin()
@@ -1378,7 +1697,12 @@ const mockHandlers: Record<string, (config: AxiosRequestConfig) => Promise<Axios
   'GET /api/admin/providers/summary': async () => {
     await delay()
     requireAdmin()
-    return createMockResponse(MOCK_PROVIDERS)
+    return createMockResponse({
+      total: MOCK_PROVIDERS.length,
+      page: 1,
+      page_size: MOCK_PROVIDERS.length,
+      items: MOCK_PROVIDERS,
+    })
   },
 
   'GET /api/admin/providers': async () => {
@@ -1387,6 +1711,33 @@ const mockHandlers: Record<string, (config: AxiosRequestConfig) => Promise<Axios
     return createMockResponse(MOCK_PROVIDERS)
   },
 
+  'GET /api/admin/pool/overview': async () => {
+    await delay()
+    requireAdmin()
+    return createMockResponse({
+      items: [{
+        provider_id: MOCK_CODEX_POOL_PROVIDER_ID,
+        provider_name: MOCK_CODEX_POOL_PROVIDER.name,
+        provider_type: 'codex',
+        total_keys: 4,
+        active_keys: 4,
+        cooldown_count: 0,
+        pool_enabled: true,
+        provider_hot_count: 2,
+        provider_desired_hot: 3,
+        provider_in_flight: 1,
+        provider_ema_in_flight: 0.8,
+        provider_burst_pending: false,
+      }]
+    })
+  },
+
+  'GET /api/admin/pool/scheduling-presets': async () => {
+    await delay()
+    requireAdmin()
+    return createMockResponse([])
+  },
+
   'POST /api/admin/providers': async (config) => {
     await delay()
     requireAdmin()
@@ -2404,6 +2755,9 @@ registerDynamicRoute('PUT', '/api/admin/modules/status/:moduleName/enabled', asy
 registerDynamicRoute('GET', '/api/admin/providers/:providerId/summary', async (_config, params) => {
   await delay()
   requireAdmin()
+  if (params.providerId === MOCK_CODEX_POOL_PROVIDER_ID) {
+    return createMockResponse(MOCK_CODEX_POOL_PROVIDER)
+  }
   const provider = MOCK_PROVIDERS.find(p => p.id === params.providerId)
   if (!provider) {
     throw { response: createMockResponse({ detail: '提供商不存在' }, 404) }
@@ -2411,6 +2765,76 @@ registerDynamicRoute('GET', '/api/admin/providers/:providerId/summary', async (_
   return createMockResponse(provider)
 })
 
+registerDynamicRoute('GET', '/api/admin/pool/:providerId/keys', async (config, params) => {
+  await delay()
+  requireAdmin()
+  if (params.providerId !== MOCK_CODEX_POOL_PROVIDER_ID) {
+    return createMockResponse({ total: 0, page: 1, page_size: 50, keys: [] })
+  }
+
+  const query = (config.params || {}) as Record<string, unknown>
+  const search = String(query.search || '').trim().toLowerCase()
+  const status = String(query.status || 'all').trim().toLowerCase()
+  const sortBy = String(query.sort_by || 'imported_at').trim()
+  const sortOrder = String(query.sort_order || 'desc').trim().toLowerCase()
+  let keys = createMockCodexPoolKeys()
+
+  if (search) {
+    keys = keys.filter(key => [
+      key.key_name,
+      key.oauth_plan_type,
+      key.oauth_account_id,
+      key.account_quota,
+    ].some(value => String(value || '').toLowerCase().includes(search)))
+  }
+  if (status === 'enabled') {
+    keys = keys.filter(key => key.is_active)
+  } else if (status === 'disabled') {
+    keys = keys.filter(key => !key.is_active)
+  } else if (status !== 'all') {
+    keys = keys.filter(key => key.scheduling_status === status || key.scheduling_reason === status)
+  }
+
+  keys.sort((left, right) => {
+    const leftValue = String((left as Record<string, unknown>)[sortBy] ?? left.imported_at ?? '')
+    const rightValue = String((right as Record<string, unknown>)[sortBy] ?? right.imported_at ?? '')
+    const comparison = leftValue.localeCompare(rightValue)
+    return sortOrder === 'asc' ? comparison : -comparison
+  })
+
+  const rawPage = Number(query.page)
+  const rawPageSize = Number(query.page_size)
+  const page = Number.isFinite(rawPage) && rawPage >= 1 ? Math.floor(rawPage) : 1
+  const pageSize = Number.isFinite(rawPageSize) && rawPageSize >= 1 ? Math.floor(rawPageSize) : 50
+  const start = (page - 1) * pageSize
+  return createMockResponse({
+    total: keys.length,
+    page,
+    page_size: pageSize,
+    keys: keys.slice(start, start + pageSize),
+  })
+})
+
+// Provider 模型映射预览
+registerDynamicRoute('GET', '/api/admin/providers/:providerId/mapping-preview', async (_config, params) => {
+  await delay()
+  requireAdmin()
+  const provider = MOCK_PROVIDERS.find(p => p.id === params.providerId)
+  if (!provider) {
+    throw { response: createMockResponse({ detail: '提供商不存在' }, 404) }
+  }
+  return createMockResponse({
+    provider_id: provider.id,
+    provider_name: provider.name,
+    keys: [],
+    total_keys: 0,
+    total_matches: 0,
+    truncated: false,
+    truncated_keys: 0,
+    truncated_models: 0,
+  })
+})
+
 // Provider 更新
 registerDynamicRoute('PATCH', '/api/admin/providers/:providerId', async (config, params) => {
   await delay()
@@ -2486,13 +2910,37 @@ registerDynamicRoute('DELETE', '/api/admin/endpoints/:endpointId', async (_confi
 })
 
 // Provider Keys 列表
-registerDynamicRoute('GET', '/api/admin/endpoints/providers/:providerId/keys', async (_config, params) => {
+registerDynamicRoute('GET', '/api/admin/endpoints/providers/:providerId/keys', async (config, params) => {
   await delay()
   requireAdmin()
   if (!PROVIDER_KEYS_CACHE[params.providerId]) {
     PROVIDER_KEYS_CACHE[params.providerId] = generateMockKeysForProvider(params.providerId, 2)
   }
-  return createMockResponse(PROVIDER_KEYS_CACHE[params.providerId])
+  const keys = PROVIDER_KEYS_CACHE[params.providerId]
+  const query = config.params || {}
+
+  // 当前详情抽屉使用 page/page_size 分页；其他调用仍使用 skip/limit 并期望裸数组。
+  if (query.page !== undefined || query.page_size !== undefined) {
+    const rawPage = Number(query.page)
+    const rawPageSize = Number(query.page_size)
+    const page = Number.isFinite(rawPage) && rawPage >= 1 ? Math.floor(rawPage) : 1
+    const pageSize = Number.isFinite(rawPageSize) && rawPageSize >= 1
+      ? Math.floor(rawPageSize)
+      : 20
+    const start = (page - 1) * pageSize
+    return createMockResponse({
+      total: keys.length,
+      page,
+      page_size: pageSize,
+      keys: keys.slice(start, start + pageSize),
+    })
+  }
+
+  const rawSkip = Number(query.skip)
+  const rawLimit = Number(query.limit)
+  const skip = Number.isFinite(rawSkip) && rawSkip >= 0 ? Math.floor(rawSkip) : 0
+  const limit = Number.isFinite(rawLimit) && rawLimit >= 1 ? Math.floor(rawLimit) : keys.length
+  return createMockResponse(keys.slice(skip, skip + limit))
 })
 
 // 为 Provider 创建 Key
@@ -2545,6 +2993,25 @@ registerDynamicRoute('POST', '/api/admin/endpoints/providers/:providerId/keys',
 registerDynamicRoute('POST', '/api/admin/endpoints/providers/:providerId/refresh-quota', async (config, params) => {
   await delay()
   requireAdmin()
+  if (params.providerId === MOCK_CODEX_POOL_PROVIDER_ID) {
+    const body = JSON.parse(config.data || '{}')
+    const requestedKeyIds = Array.isArray(body.key_ids)
+      ? body.key_ids.map((id: unknown) => String(id).trim()).filter(Boolean)
+      : createMockCodexPoolKeys().map(key => key.key_id)
+    const keyNames = new Map(createMockCodexPoolKeys().map(key => [key.key_id, key.key_name]))
+    const results = requestedKeyIds.map((keyId: string) => ({
+      key_id: keyId,
+      key_name: keyNames.get(keyId) || keyId,
+      status: 'success',
+      metadata: { updated_at: new Date().toISOString() },
+    }))
+    return createMockResponse({
+      success: results.length,
+      failed: 0,
+      total: results.length,
+      results,
+    })
+  }
   if (!PROVIDER_KEYS_CACHE[params.providerId]) {
     PROVIDER_KEYS_CACHE[params.providerId] = generateMockKeysForProvider(params.providerId, 2)
   }
@@ -2684,6 +3151,20 @@ registerDynamicRoute('POST', '/api/admin/endpoints/keys/:keyId/clear-oauth-inval
   return createMockResponse({ message: 'OAuth invalid cleared (demo)', key_id: params.keyId })
 })
 
+registerDynamicRoute('POST', '/api/admin/endpoints/keys/:keyId/reset-cycle-stats', async (_config, params) => {
+  await delay()
+  requireAdmin()
+  const key = createMockCodexPoolKeys().find(item => item.key_id === params.keyId)
+  const windows = key?.status_snapshot.quota.windows.filter(window => (
+    window.window_minutes > 0 && !window.code.startsWith('spark_')
+  )).length ?? 0
+  return createMockResponse({
+    message: '已重置周期统计（演示模式）',
+    reset_at: Math.floor(Date.now() / 1000),
+    windows,
+  })
+})
+
 
 // Keys grouped by format
 mockHandlers['GET /api/admin/endpoints/keys/grouped-by-format'] = async () => {
@@ -3206,10 +3687,122 @@ registerDynamicRoute('DELETE', '/api/admin/users/:userId', async (_config, param
 })
 
 // 用户 API Keys
-registerDynamicRoute('GET', '/api/admin/users/:userId/api-keys', async (_config, _params) => {
+registerDynamicRoute('GET', '/api/admin/users/:userId/api-keys', async (_config, params) => {
   await delay()
   requireAdmin()
-  return createMockResponse(MOCK_USER_API_KEYS)
+  const apiKeys = mockManagedUserApiKeys(params.userId).map(publicMockManagedUserApiKey)
+  return createMockResponse({
+    api_keys: apiKeys,
+    total: apiKeys.length,
+  })
+})
+
+registerDynamicRoute('POST', '/api/admin/users/:userId/api-keys', async (config, params) => {
+  await delay()
+  requireAdmin()
+  const keys = mockManagedUserApiKeys(params.userId)
+  const body = mockRequestObject(config)
+  const sequence = ++mockManagedUserApiKeySequence
+  const fullKey = `sk-ae-demo-${params.userId.slice(0, 8)}-${sequence}`
+  const key: MockManagedUserApiKey = {
+    id: `managed-key-${params.userId}-${sequence}`,
+    fullKey,
+    key_display: `${fullKey.slice(0, 10)}...${fullKey.slice(-4)}`,
+    name: typeof body.name === 'string' && body.name.trim()
+      ? body.name.trim()
+      : `Key-${sequence}`,
+    created_at: new Date().toISOString(),
+    is_active: true,
+    is_locked: false,
+    is_standalone: false,
+    feature_settings: body.feature_settings && typeof body.feature_settings === 'object'
+      ? body.feature_settings as Record<string, unknown>
+      : null,
+    rate_limit: typeof body.rate_limit === 'number' ? body.rate_limit : 0,
+    concurrent_limit: typeof body.concurrent_limit === 'number'
+      ? body.concurrent_limit
+      : null,
+    ip_rules: Array.isArray(body.ip_rules)
+      ? body.ip_rules.filter((value): value is string => typeof value === 'string')
+      : null,
+    total_requests: 0,
+    total_cost_usd: 0,
+    force_capabilities: null,
+  }
+  keys.unshift(key)
+  return createMockResponse({
+    ...publicMockManagedUserApiKey(key),
+    key: fullKey,
+    message: 'API Key创建成功，请妥善保存完整密钥',
+  })
+})
+
+registerDynamicRoute('PUT', '/api/admin/users/:userId/api-keys/:keyId', async (config, params) => {
+  await delay()
+  requireAdmin()
+  const keys = mockManagedUserApiKeys(params.userId)
+  const index = keys.findIndex(key => key.id === params.keyId)
+  if (index < 0) {
+    throw { response: createMockResponse({ detail: 'API Key 不存在' }, 404) }
+  }
+  const body = mockRequestObject(config)
+  const existing = keys[index]
+  const updated: MockManagedUserApiKey = {
+    ...existing,
+    ...(typeof body.name === 'string' ? { name: body.name.trim() } : {}),
+    ...(typeof body.rate_limit === 'number' ? { rate_limit: body.rate_limit } : {}),
+    ...(typeof body.concurrent_limit === 'number' || body.concurrent_limit === null
+      ? { concurrent_limit: body.concurrent_limit }
+      : {}),
+    ...(Array.isArray(body.ip_rules) || body.ip_rules === null
+      ? { ip_rules: body.ip_rules as string[] | null }
+      : {}),
+    ...('feature_settings' in body
+      ? { feature_settings: body.feature_settings as Record<string, unknown> | null }
+      : {}),
+  }
+  keys[index] = updated
+  return createMockResponse({
+    ...publicMockManagedUserApiKey(updated),
+    message: 'API Key更新成功',
+  })
+})
+
+registerDynamicRoute('DELETE', '/api/admin/users/:userId/api-keys/:keyId', async (_config, params) => {
+  await delay()
+  requireAdmin()
+  const keys = mockManagedUserApiKeys(params.userId)
+  const index = keys.findIndex(key => key.id === params.keyId)
+  if (index < 0) {
+    throw { response: createMockResponse({ detail: 'API Key 不存在' }, 404) }
+  }
+  keys.splice(index, 1)
+  return createMockResponse({ message: 'API Key删除成功' })
+})
+
+registerDynamicRoute('PATCH', '/api/admin/users/:userId/api-keys/:keyId/lock', async (_config, params) => {
+  await delay()
+  requireAdmin()
+  const key = mockManagedUserApiKeys(params.userId).find(key => key.id === params.keyId)
+  if (!key) {
+    throw { response: createMockResponse({ detail: 'API Key 不存在' }, 404) }
+  }
+  key.is_locked = !key.is_locked
+  return createMockResponse({
+    id: key.id,
+    is_locked: key.is_locked,
+    message: key.is_locked ? 'API Key已锁定' : 'API Key已解锁',
+  })
+})
+
+registerDynamicRoute('GET', '/api/admin/users/:userId/api-keys/:keyId/full-key', async (_config, params) => {
+  await delay()
+  requireAdmin()
+  const key = mockManagedUserApiKeys(params.userId).find(key => key.id === params.keyId)
+  if (!key) {
+    throw { response: createMockResponse({ detail: 'API Key 不存在' }, 404) }
+  }
+  return createMockResponse({ key: key.fullKey })
 })
 
 // 管理员 - 用户会话列表
@@ -3295,10 +3888,13 @@ registerDynamicRoute('DELETE', '/api/users/me/api-keys/:keyId', async (_config,
 })
 
 // 使用记录详情 - /api/admin/usage/:requestId
-registerDynamicRoute('GET', '/api/admin/usage/:requestId', async (_config, params) => {
+registerDynamicRoute('GET', '/api/admin/usage/:requestId', async (config, params) => {
   await delay()
   requireAdmin()
 
+  const includeBodies = config.params?.include_bodies !== false
+    && config.params?.include_bodies !== 'false'
+
   const records = getUsageRecords()
   const record = records.find(r => r.id === params.requestId)
 
@@ -3318,6 +3914,9 @@ registerDynamicRoute('GET', '/api/admin/usage/:requestId', async (_config, param
   // 生成模拟的请求/响应数据
   const mockRequestBody = {
     model: record.model,
+    ...(record.requested_reasoning_effort
+      ? { reasoning: { effort: record.requested_reasoning_effort } }
+      : {}),
     max_tokens: 4096,
     messages: [
       {
@@ -3328,7 +3927,27 @@ registerDynamicRoute('GET', '/api/admin/usage/:requestId', async (_config, param
     stream: record.is_stream
   }
 
-  const mockResponseBody = record.status === 'failed' ? {
+  const mockProviderRequestBody = record.id === MOCK_CYBER_POLICY_USAGE_ID
+    ? {
+        model: record.target_model || record.model,
+        reasoning: { effort: record.reasoning_effort },
+        service_tier: record.service_tier,
+        input: [
+          {
+            role: 'user',
+            content: 'Help me with an authorized cybersecurity research task.'
+          }
+        ],
+        stream: record.is_stream
+      }
+    : {
+        ...mockRequestBody,
+        model: record.target_model || record.model
+      }
+
+  const mockResponseBody = record.id === MOCK_CYBER_POLICY_USAGE_ID
+    ? MOCK_CYBER_POLICY_ERROR_BODY
+    : record.status === 'failed' ? {
     error: {
       type: 'api_error',
       message: record.error_message || 'An error occurred'
@@ -3376,6 +3995,10 @@ registerDynamicRoute('GET', '/api/admin/usage/:requestId', async (_config, param
     api_format: record.api_format,
     model: record.model,
     target_model: record.target_model,
+    requested_reasoning_effort: record.requested_reasoning_effort,
+    reasoning_effort: record.reasoning_effort,
+    service_tier: record.service_tier,
+    actual_service_tier: record.actual_service_tier,
     tokens: {
       input: record.input_tokens,
       output: record.output_tokens,
@@ -3406,6 +4029,7 @@ registerDynamicRoute('GET', '/api/admin/usage/:requestId', async (_config, param
     error_message: record.error_message,
     response_time_ms: record.response_time_ms,
     created_at: record.created_at,
+    updated_at: record.updated_at ?? record.created_at,
     request_headers: {
       'Content-Type': 'application/json',
       'Authorization': 'Bearer sk-aether-***',
@@ -3414,7 +4038,12 @@ registerDynamicRoute('GET', '/api/admin/usage/:requestId', async (_config, param
       'Accept': 'application/json',
       'X-Request-ID': `req_${record.id}`
     },
-    request_body: mockRequestBody,
+    has_request_body: true,
+    has_provider_request_body: true,
+    has_response_body: true,
+    has_client_response_body: false,
+    request_body: includeBodies ? mockRequestBody : null,
+    provider_request_body: includeBodies ? mockProviderRequestBody : null,
     provider_request_headers: {
       'Content-Type': 'application/json',
       'Authorization': `Bearer sk-${record.provider}-***`,
@@ -3428,7 +4057,9 @@ registerDynamicRoute('GET', '/api/admin/usage/:requestId', async (_config, param
       'X-RateLimit-Remaining': '999',
       'X-RateLimit-Reset': new Date(Date.now() + 60000).toISOString()
     },
-    response_body: mockResponseBody,
+    response_body: includeBodies ? mockResponseBody : null,
+    client_response_body: null,
+    body_load_errors: null,
     metadata: {
       client_ip: '192.168.1.100',
       user_agent: 'Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7)',
@@ -3576,7 +4207,8 @@ registerDynamicRoute('GET', '/api/admin/monitoring/trace/:requestId', async (_co
     })
   } else if (record.status === 'failed') {
     // 失败请求：多个候选都失败
-    const attemptCount = 2 + Math.floor(Math.random() * 2)
+    const isCyberPolicyDemo = record.id === MOCK_CYBER_POLICY_USAGE_ID
+    const attemptCount = isCyberPolicyDemo ? 1 : 2 + Math.floor(Math.random() * 2)
 
     for (let i = 0; i < attemptCount; i++) {
       const attemptStarted = new Date(now.getTime() + i * 200)
@@ -3611,7 +4243,20 @@ registerDynamicRoute('GET', '/api/admin/monitoring/trace/:requestId', async (_co
           ranking_mode: 'FixedOrder',
           priority_mode: 'Provider',
           ranking_index: i,
-          priority_slot: i + 1
+          priority_slot: i + 1,
+          ...(isCyberPolicyDemo ? {
+            upstream_response: {
+              source: 'upstream_response',
+              status_code: 400,
+              headers: {
+                'content-type': 'application/json',
+                'x-request-id': `req_${MOCK_CYBER_POLICY_USAGE_ID}`
+              },
+              body: MOCK_CYBER_POLICY_ERROR_BODY,
+              body_ref: `usage://request/req_${MOCK_CYBER_POLICY_USAGE_ID}/response_body`,
+              body_state: 'reference'
+            }
+          } : {})
         },
         latency_ms: attemptLatency,
         created_at: attemptStarted.toISOString(),

frontend/src/utils/__tests__/codexQuotaWindow.spec.ts

@@ -0,0 +1,40 @@
+import { describe, expect, it } from 'vitest'
+
+import { getCodexQuotaWindowPresentation } from '../codexQuotaWindow'
+
+describe('getCodexQuotaWindowPresentation', () => {
+  it.each([
+    [300, '5H'],
+    [10_080, '周'],
+    [43_200, '月'],
+    [43_800, '月'],
+    [44_640, '月'],
+  ])('labels a %i-minute window as %s', (windowMinutes, expectedLabel) => {
+    expect(getCodexQuotaWindowPresentation({
+      code: 'primary',
+      window_minutes: windowMinutes,
+    })?.label).toBe(expectedLabel)
+  })
+
+  it('supports simultaneous 5H and weekly windows', () => {
+    const windows = [
+      getCodexQuotaWindowPresentation({ code: 'secondary', window_minutes: 10_080 }),
+      getCodexQuotaWindowPresentation({ code: 'primary', window_minutes: 300 }),
+    ].filter((item): item is NonNullable<typeof item> => item != null)
+
+    expect(windows.sort((a, b) => a.sortOrder - b.sortOrder).map(item => item.label)).toEqual(['5H', '周'])
+  })
+
+  it('drops zero-minute placeholder windows', () => {
+    expect(getCodexQuotaWindowPresentation({
+      code: 'weekly',
+      label: '周',
+      window_minutes: 0,
+    })).toBeNull()
+  })
+
+  it('keeps legacy labels when old snapshots have no window duration', () => {
+    expect(getCodexQuotaWindowPresentation({ code: '5h' })?.label).toBe('5H')
+    expect(getCodexQuotaWindowPresentation({ code: 'weekly' })?.label).toBe('周')
+  })
+})

frontend/src/utils/__tests__/oauthPlanType.spec.ts

@@ -0,0 +1,15 @@
+import { describe, expect, it } from 'vitest'
+
+import { formatOAuthPlanType } from '../oauthPlanType'
+
+describe('formatOAuthPlanType', () => {
+  it('uses the compact Codex label for the usage-based business plan', () => {
+    expect(formatOAuthPlanType('self_serve_business_usage_based')).toBe('Codex')
+    expect(formatOAuthPlanType(' SELF_SERVE_BUSINESS_USAGE_BASED ')).toBe('Codex')
+  })
+
+  it('keeps existing known plan labels intact', () => {
+    expect(formatOAuthPlanType('plus')).toBe('Plus')
+    expect(formatOAuthPlanType('team')).toBe('Team')
+  })
+})

frontend/src/utils/__tests__/providerKeyQuota.spec.ts

@@ -43,6 +43,34 @@ describe('providerKeyQuota', () => {
     }, 'codex')).toBe('周剩余 90.0% | 5H剩余 80.0% | Spark5H剩余 60.0% | Spark周剩余 95.0%')
   })
 
+  it('uses actual Codex window durations and ignores zero placeholders', () => {
+    expect(getQuotaDisplayText({
+      status_snapshot: {
+        oauth: { code: 'valid' },
+        account: { code: 'ok', blocked: false },
+        quota: {
+          provider_type: 'codex',
+          code: 'ok',
+          exhausted: false,
+          windows: [
+            {
+              code: 'weekly',
+              label: '周',
+              window_minutes: 0,
+              remaining_ratio: 1,
+            },
+            {
+              code: '5h',
+              label: '5H',
+              window_minutes: 43_800,
+              remaining_ratio: 0.86,
+            },
+          ],
+        },
+      },
+    }, 'codex')).toBe('月剩余 86.0%')
+  })
+
   it('formats Grok account quota from structured quota windows', () => {
     expect(getQuotaDisplayText({
       status_snapshot: {

frontend/src/utils/codexQuotaWindow.ts

@@ -0,0 +1,62 @@
+import type { QuotaWindowSnapshot } from '@/api/endpoints/types'
+
+const MINUTES_PER_HOUR = 60
+const MINUTES_PER_DAY = 24 * MINUTES_PER_HOUR
+const MINUTES_PER_WEEK = 7 * MINUTES_PER_DAY
+const MIN_MONTH_MINUTES = 28 * MINUTES_PER_DAY
+const MAX_MONTH_MINUTES = 31 * MINUTES_PER_DAY
+
+export interface CodexQuotaWindowPresentation {
+  label: string
+  sortOrder: number
+}
+
+function formatCodexQuotaPeriod(windowMinutes: number): string {
+  if (windowMinutes === 5 * MINUTES_PER_HOUR) return '5H'
+  if (windowMinutes === MINUTES_PER_WEEK) return '周'
+  if (windowMinutes >= MIN_MONTH_MINUTES && windowMinutes <= MAX_MONTH_MINUTES) return '月'
+
+  if (windowMinutes % MINUTES_PER_WEEK === 0) {
+    return `${windowMinutes / MINUTES_PER_WEEK}周`
+  }
+  if (windowMinutes % MINUTES_PER_DAY === 0) {
+    return `${windowMinutes / MINUTES_PER_DAY}天`
+  }
+  if (windowMinutes % MINUTES_PER_HOUR === 0) {
+    return `${windowMinutes / MINUTES_PER_HOUR}H`
+  }
+  return `${windowMinutes}分钟`
+}
+
+function getLegacyCodexQuotaPeriod(code: string, label: string): string | null {
+  if (code === '5h') return '5H'
+  if (code === 'weekly') return '周'
+  if (code === 'monthly') return '月'
+  return label || null
+}
+
+export function getCodexQuotaWindowPresentation(
+  window: QuotaWindowSnapshot,
+): CodexQuotaWindowPresentation | null {
+  const code = String(window.code || '').trim().toLowerCase()
+  const isSpark = code.startsWith('spark_')
+  const baseCode = isSpark ? code.slice('spark_'.length) : code
+  const rawLabel = String(window.label || '').trim().replace(/^Spark\s*/i, '')
+  const hasExplicitWindowMinutes = window.window_minutes != null
+  const windowMinutes = Number(window.window_minutes)
+
+  if (hasExplicitWindowMinutes && (!Number.isFinite(windowMinutes) || windowMinutes <= 0)) {
+    return null
+  }
+
+  const period = hasExplicitWindowMinutes
+    ? formatCodexQuotaPeriod(windowMinutes)
+    : getLegacyCodexQuotaPeriod(baseCode, rawLabel)
+  if (!period) return null
+
+  const fallbackOrder = baseCode === '5h' ? 300 : baseCode === 'weekly' ? 10_080 : 1_000_000
+  return {
+    label: isSpark ? `Spark${period}` : period,
+    sortOrder: (isSpark ? 10_000_000 : 0) + (hasExplicitWindowMinutes ? windowMinutes : fallbackOrder),
+  }
+}