i/Aether
1
0
Fork 0
mirror of https://github.com/fawney19/Aether.git synced 2026-10-05 08:57:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

feat: expose upstream response model in usage records

Browse Source
This commit is contained in:
AAEE86
2026-09-21 09:51:33 +08:00
parent ba7c9f8b27
commit f960bbd2c8
24 changed files with 466 additions and 83 deletions
Download Patch File Download Diff File Expand all files Collapse all files
crates/aether-data/adapters/postgres/src/usage/queries/list_recent_usage_audits_prefix.sql
+3
View File
@@ -183,6 +183,7 @@ SELECT
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
OR NULLIF(BTRIM("usage".request_metadata->>'provider_response_model'), '') IS NOT NULL
OR ("usage".request_metadata->>'client_requested_stream') IN ('true', 'false')
OR ("usage".request_metadata->>'upstream_is_stream') IN ('true', 'false')
OR ("usage".request_metadata->>'websocket_mode') IN ('true', 'false')
@@ -208,6 +209,8 @@ SELECT
NULLIF(BTRIM("usage".request_metadata->>'provider_service_tier'), ''),
'provider_actual_service_tier',
NULLIF(BTRIM("usage".request_metadata->>'provider_actual_service_tier'), ''),
'provider_response_model',
NULLIF(BTRIM("usage".request_metadata->>'provider_response_model'), ''),
'client_requested_stream',
CASE
WHEN ("usage".request_metadata->>'client_requested_stream') IN ('true', 'false')
crates/aether-data/adapters/postgres/src/usage/queries/list_usage_audits_prefix.sql
+3
View File
@@ -183,6 +183,7 @@ SELECT
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
OR NULLIF(BTRIM("usage".request_metadata->>'provider_response_model'), '') IS NOT NULL
OR ("usage".request_metadata->>'client_requested_stream') IN ('true', 'false')
OR ("usage".request_metadata->>'upstream_is_stream') IN ('true', 'false')
OR ("usage".request_metadata->>'websocket_mode') IN ('true', 'false')
@@ -208,6 +209,8 @@ SELECT
NULLIF(BTRIM("usage".request_metadata->>'provider_service_tier'), ''),
'provider_actual_service_tier',
NULLIF(BTRIM("usage".request_metadata->>'provider_actual_service_tier'), ''),
'provider_response_model',
NULLIF(BTRIM("usage".request_metadata->>'provider_response_model'), ''),
'client_requested_stream',
CASE
WHEN ("usage".request_metadata->>'client_requested_stream') IN ('true', 'false')
crates/aether-data/adapters/postgres/src/usage/tests.rs
+2
View File
@@ -3611,6 +3611,8 @@ fn usage_sql_uses_json_null_placeholders_for_usage_payload_columns() {
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("request_metadata->>'provider_response_model'"));
assert!(sql.contains("'provider_response_model'"));
assert!(sql.contains("request_metadata->>'websocket_mode'"));
assert!(sql.contains("'websocket_mode'"));
assert!(sql.contains("AS client_family"));
crates/aether-data/contracts/src/repository/usage/metadata_policy.rs
+27 -3
View File
@@ -8,9 +8,10 @@ use serde_json::{Map, Value};
use crate::repository::candidates::sanitize_request_candidate_skip_reason;
use super::{
LIVE_SESSION_METADATA_KEY, PLAN_USAGE_RESERVATION_DEFERRED_METADATA_KEY,
PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY, PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY,
PROVIDER_REASONING_EFFORT_METADATA_KEY, PROVIDER_SERVICE_TIER_METADATA_KEY,
normalize_provider_response_model, LIVE_SESSION_METADATA_KEY,
PLAN_USAGE_RESERVATION_DEFERRED_METADATA_KEY, PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY,
PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY, PROVIDER_REASONING_EFFORT_METADATA_KEY,
PROVIDER_RESPONSE_MODEL_METADATA_KEY, PROVIDER_SERVICE_TIER_METADATA_KEY,
REALTIME_SESSION_METADATA_KEY, REQUESTED_REASONING_EFFORT_METADATA_KEY,
ROUTING_CANDIDATE_SKIP_REASON_METADATA_KEY, ROUTING_FAILURE_DIAGNOSTIC_METADATA_KEY,
USAGE_AVAILABLE_METADATA_KEY, USAGE_PRICING_AVAILABLE_METADATA_KEY,
@@ -94,6 +95,12 @@ pub fn sanitize_usage_request_metadata_object(source: &Map<String, Value>) -> Op
] {
insert_known_string(source, &mut target, key, sanitize_service_tier);
}
insert_known_string(
source,
&mut target,
PROVIDER_RESPONSE_MODEL_METADATA_KEY,
normalize_provider_response_model,
);
insert_bounded_u64(
source,
&mut target,
@@ -1279,6 +1286,23 @@ mod tests {
}
}
#[test]
fn persistence_projection_keeps_bounded_response_model_only_as_a_string() {
let metadata = sanitize_usage_request_metadata(Some(json!({
"provider_response_model": " GPT-5.1 "
})))
.expect("response model should remain");
assert_eq!(metadata["provider_response_model"], "GPT-5.1");
assert!(sanitize_usage_request_metadata(Some(json!({
"provider_response_model": 42
})))
.is_none());
assert!(sanitize_usage_request_metadata(Some(json!({
"provider_response_model": "x".repeat(257)
})))
.is_none());
}
#[test]
fn persistence_projection_keeps_only_bounded_settlement_facts() {
let metadata = sanitize_usage_request_metadata(Some(json!({
crates/aether-data/contracts/src/repository/usage/mod.rs
+10 -9
View File
@@ -15,11 +15,12 @@ pub use policy::*;
pub use types::{
canonical_usage_body_ref_for, 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,
extract_provider_response_model_from_bodies, extract_provider_service_tier_from_body,
normalize_provider_response_model, normalize_provider_service_tier, parse_usage_body_ref,
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,
usage_body_capture_is_authoritative, usage_body_ref, usage_request_metadata_client_family,
ApiKeyLastUsedDelta, ManagementTokenCounterDelta, PendingUsageCleanupSummary,
ProviderApiKeyWindowUsageRequest, ProxyNodeCounterDelta, StoredProviderApiKeyUsageSummary,
StoredProviderApiKeyWindowUsageSummary, StoredProviderUsageSummary, StoredProviderUsageWindow,
StoredRequestUsageAudit, StoredUsageAuditAggregation, StoredUsageAuditSummary,
StoredUsageBodyPayload, StoredUsageBreakdownSummaryRow, StoredUsageCacheAffinityHitSummary,
@@ -46,9 +47,9 @@ pub use types::{
UsageTimeSeriesQuery, UsageWriteRepository, LIVE_SESSION_METADATA_KEY,
PLAN_USAGE_RESERVATION_DEFERRED_METADATA_KEY, PROVIDER_ACTUAL_SERVICE_TIER_METADATA_KEY,
PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY, PROVIDER_REASONING_EFFORT_METADATA_KEY,
PROVIDER_SERVICE_TIER_METADATA_KEY, REALTIME_SESSION_METADATA_KEY,
REQUESTED_REASONING_EFFORT_METADATA_KEY, ROUTING_CANDIDATE_SKIP_REASON_METADATA_KEY,
ROUTING_FAILURE_DIAGNOSTIC_METADATA_KEY, USAGE_AVAILABLE_METADATA_KEY,
USAGE_PRICING_AVAILABLE_METADATA_KEY, WEBSOCKET_MODE_METADATA_KEY,
WEBSOCKET_TRANSPORT_METADATA_KEY,
PROVIDER_RESPONSE_MODEL_METADATA_KEY, PROVIDER_SERVICE_TIER_METADATA_KEY,
REALTIME_SESSION_METADATA_KEY, REQUESTED_REASONING_EFFORT_METADATA_KEY,
ROUTING_CANDIDATE_SKIP_REASON_METADATA_KEY, ROUTING_FAILURE_DIAGNOSTIC_METADATA_KEY,
USAGE_AVAILABLE_METADATA_KEY, USAGE_PRICING_AVAILABLE_METADATA_KEY,
WEBSOCKET_MODE_METADATA_KEY, WEBSOCKET_TRANSPORT_METADATA_KEY,
};
crates/aether-data/contracts/src/repository/usage/types.rs
+99 -4
View File
@@ -6,6 +6,7 @@ pub const PROVIDER_REASONING_EFFORT_METADATA_KEY: &str = "provider_reasoning_eff
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_RESPONSE_MODEL_METADATA_KEY: &str = "provider_response_model";
pub const PROVIDER_CACHE_TTL_MINUTES_METADATA_KEY: &str = "provider_cache_ttl_minutes";
pub const ROUTING_CANDIDATE_SKIP_REASON_METADATA_KEY: &str = "routing_candidate_skip_reason";
pub const ROUTING_FAILURE_DIAGNOSTIC_METADATA_KEY: &str = "routing_failure_diagnostic";
@@ -119,6 +120,42 @@ pub fn normalize_provider_service_tier(value: &str) -> Option<String> {
Some(value.to_ascii_lowercase())
}
/// 清洗响应体顶层 `model`，保留大小写，只去除首尾空白。
pub fn normalize_provider_response_model(value: &str) -> Option<String> {
let value = value.trim();
if value.is_empty() || value.len() > 256 {
return None;
}
Some(value.to_string())
}
/// 只有请求体和响应体都可作为完整事实时，才计算响应模型，避免用截断内容猜测。
pub fn extract_provider_response_model_from_bodies(
request_body: Option<&Value>,
request_body_state: Option<UsageBodyCaptureState>,
response_body: Option<&Value>,
response_body_state: Option<UsageBodyCaptureState>,
) -> Option<String> {
if !usage_body_capture_is_authoritative(request_body, request_body_state)
|| !usage_body_capture_is_authoritative(response_body, response_body_state)
{
return None;
}
let request_model = request_body
.and_then(Value::as_object)
.and_then(|body| body.get("model"))
.and_then(Value::as_str)
.and_then(normalize_provider_response_model)?;
let response_model = response_body
.and_then(Value::as_object)
.and_then(|body| body.get("model"))
.and_then(Value::as_str)
.and_then(normalize_provider_response_model)?;
(request_model != response_model).then_some(response_model)
}
/// 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
@@ -129,7 +166,7 @@ pub fn resolve_provider_service_tier_from_request_capture(
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) {
if usage_body_capture_is_authoritative(provider_request_body, provider_request_body_state) {
return extract_provider_service_tier_from_body(provider_request_body);
}
@@ -153,7 +190,7 @@ pub fn resolve_provider_service_tier_from_request_capture(
.and_then(normalize_provider_service_tier)
}
fn request_body_capture_is_authoritative(
pub fn usage_body_capture_is_authoritative(
request_body: Option<&Value>,
request_body_state: Option<UsageBodyCaptureState>,
) -> bool {
@@ -184,7 +221,7 @@ fn resolve_reasoning_effort_from_request_capture(
request_metadata: Option<&Value>,
metadata_key: &str,
) -> Option<String> {
if request_body_capture_is_authoritative(request_body, request_body_state) {
if usage_body_capture_is_authoritative(request_body, request_body_state) {
return extract_provider_reasoning_effort_from_body(request_body);
}
@@ -701,6 +738,11 @@ impl StoredRequestUsageAudit {
})
}
pub fn provider_response_model(&self) -> Option<String> {
self.request_metadata_string(PROVIDER_RESPONSE_MODEL_METADATA_KEY)
.and_then(normalize_provider_response_model)
}
pub fn provider_cache_ttl_minutes(&self) -> Option<i64> {
resolve_provider_cache_ttl_minutes(
self.endpoint_api_format
@@ -2646,7 +2688,8 @@ fn parse_timestamp(value: i64, field_name: &str) -> Result<u64, crate::DataLayer
mod tests {
use super::{
canonical_usage_body_ref_for, extract_provider_actual_service_tier_from_response,
extract_provider_service_tier_from_body, normalize_provider_reasoning_effort,
extract_provider_response_model_from_bodies, extract_provider_service_tier_from_body,
normalize_provider_reasoning_effort, normalize_provider_response_model,
normalize_provider_service_tier, resolve_provider_cache_ttl_minutes, usage_body_ref,
StoredRequestUsageAudit, UpsertUsageRecord, UsageBodyCaptureState, UsageBodyCaptureStorage,
UsageBodyField, UsageProviderPerformanceQuery, REALTIME_SESSION_METADATA_KEY,
@@ -2672,6 +2715,58 @@ mod tests {
}
}
#[test]
fn response_model_requires_authoritative_different_top_level_models() {
let request = json!({"model": " gpt-5 "});
let response = json!({"model": " gpt-5.1 "});
assert_eq!(
extract_provider_response_model_from_bodies(
Some(&request),
Some(UsageBodyCaptureState::Inline),
Some(&response),
Some(UsageBodyCaptureState::Inline),
),
Some("gpt-5.1".to_string())
);
assert_eq!(
extract_provider_response_model_from_bodies(
Some(&request),
Some(UsageBodyCaptureState::Inline),
Some(&json!({"model": "gpt-5"})),
Some(UsageBodyCaptureState::Inline),
),
None
);
assert_eq!(
extract_provider_response_model_from_bodies(
Some(&request),
Some(UsageBodyCaptureState::Truncated),
Some(&response),
Some(UsageBodyCaptureState::Inline),
),
None
);
}
#[test]
fn response_model_keeps_case_and_rejects_invalid_values() {
assert_eq!(
normalize_provider_response_model(" GPT-5.1 "),
Some("GPT-5.1".to_string())
);
assert_eq!(normalize_provider_response_model(" "), None);
assert_eq!(normalize_provider_response_model(&"x".repeat(257)), None);
assert_eq!(
extract_provider_response_model_from_bodies(
Some(&json!({"model": "gpt-5"})),
None,
Some(&json!({"model": 42})),
None,
),
None
);
}
#[test]
fn provider_fact_normalization_preserves_non_ascii_case_and_byte_count() {
for normalize in [