Merge remote-tracking branch 'upstream/aether-rust-pioneer' into rust

.github/workflows/build-hub.yml

@@ -1,91 +0,0 @@
-name: Build aether-hub
-
-on:
-  push:
-    tags: ['hub-v*']
-  workflow_dispatch:
-
-permissions:
-  contents: write
-
-jobs:
-  build:
-    name: ${{ matrix.name }}
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - name: linux-amd64
-            target: x86_64-unknown-linux-gnu
-            use_cross: true
-          - name: linux-arm64
-            target: aarch64-unknown-linux-gnu
-            use_cross: true
-
-    steps:
-      - uses: actions/checkout@v5
-
-      - name: Install Rust toolchain
-        uses: dtolnay/rust-toolchain@stable
-        with:
-          targets: ${{ matrix.target }}
-
-      - name: Rust cache
-        uses: Swatinem/rust-cache@v2
-        with:
-          workspaces: aether-hub -> target
-          key: ${{ matrix.target }}
-
-      - name: Install cross
-        if: matrix.use_cross
-        uses: taiki-e/install-action@cross
-
-      - name: Build
-        working-directory: aether-hub
-        shell: bash
-        run: |
-          if [ "${{ matrix.use_cross }}" = "true" ]; then
-            cross build --release --target ${{ matrix.target }}
-          else
-            cargo build --release --target ${{ matrix.target }}
-          fi
-
-      - name: Package
-        shell: bash
-        run: |
-          cd aether-hub/target/${{ matrix.target }}/release
-          chmod +x aether-hub
-          tar czf ../../../../aether-hub-${{ matrix.name }}.tar.gz aether-hub
-
-      - name: Upload artifact
-        uses: actions/upload-artifact@v5
-        with:
-          name: aether-hub-${{ matrix.name }}
-          path: aether-hub-*.tar.gz
-          if-no-files-found: error
-
-  release:
-    needs: build
-    runs-on: ubuntu-latest
-    if: startsWith(github.ref, 'refs/tags/')
-    steps:
-      - name: Download all artifacts
-        uses: actions/download-artifact@v5
-        with:
-          merge-multiple: true
-          path: artifacts
-
-      - name: Generate checksums
-        working-directory: artifacts
-        run: sha256sum aether-hub-* > SHA256SUMS.txt
-
-      - name: Create GitHub Release
-        uses: softprops/action-gh-release@v2
-        with:
-          name: "${{ github.ref_name }}"
-          generate_release_notes: true
-          files: |
-            artifacts/aether-hub-*
-            artifacts/SHA256SUMS.txt
-          fail_on_unmatched_files: true

apps/aether-gateway/src/ai_pipeline/conversion/mod.rs

@@ -5,8 +5,9 @@ pub(crate) use crate::ai_pipeline::{
     core_error_default_client_api_format, is_core_error_finalize_kind, LocalCoreSyncErrorKind,
 };
 pub(crate) use crate::ai_pipeline::{
-    request_conversion_direct_auth, request_conversion_kind,
-    request_conversion_transport_supported, sync_chat_response_conversion_kind,
+    request_candidate_api_formats, request_conversion_direct_auth, request_conversion_kind,
+    request_conversion_requires_enable_flag, request_conversion_transport_supported,
+    request_pair_allowed_for_transport, sync_chat_response_conversion_kind,
     sync_cli_response_conversion_kind, RequestConversionKind, SyncChatResponseConversionKind,
     SyncCliResponseConversionKind,
 };
@@ -14,7 +15,7 @@ pub(crate) use crate::ai_pipeline::{
 #[cfg(test)]
 mod tests {
     use super::{
-        request_conversion_kind, sync_chat_response_conversion_kind,
+        request_candidate_api_formats, request_conversion_kind, sync_chat_response_conversion_kind,
         sync_cli_response_conversion_kind, RequestConversionKind, SyncChatResponseConversionKind,
         SyncCliResponseConversionKind,
     };
@@ -22,7 +23,15 @@ mod tests {
     #[test]
     fn request_conversion_registry_supports_bidirectional_standard_matrix() {
         assert_eq!(
-            request_conversion_kind("claude:chat", "openai:chat"),
+            request_conversion_kind("openai:chat", "openai:cli"),
+            Some(RequestConversionKind::ToOpenAIFamilyCli)
+        );
+        assert_eq!(
+            request_conversion_kind("openai:chat", "claude:cli"),
+            Some(RequestConversionKind::ToClaudeStandard)
+        );
+        assert_eq!(
+            request_conversion_kind("openai:cli", "openai:chat"),
             Some(RequestConversionKind::ToOpenAIChat)
         );
         assert_eq!(
@@ -30,12 +39,16 @@ mod tests {
             Some(RequestConversionKind::ToClaudeStandard)
         );
         assert_eq!(
-            request_conversion_kind("gemini:cli", "openai:compact"),
-            Some(RequestConversionKind::ToOpenAICompact)
+            request_conversion_kind("openai:compact", "gemini:cli"),
+            None
         );
         assert_eq!(
-            request_conversion_kind("openai:compact", "gemini:cli"),
-            Some(RequestConversionKind::ToGeminiStandard)
+            request_conversion_kind("gemini:cli", "openai:compact"),
+            None
+        );
+        assert_eq!(
+            request_conversion_kind("openai:chat", "openai:compact"),
+            None
         );
         assert_eq!(request_conversion_kind("claude:chat", "claude:chat"), None);
     }
@@ -59,12 +72,50 @@ mod tests {
             Some(SyncCliResponseConversionKind::ToGeminiCli)
         );
         assert_eq!(
-            sync_cli_response_conversion_kind("claude:cli", "openai:compact"),
+            sync_cli_response_conversion_kind("claude:chat", "openai:cli"),
             Some(SyncCliResponseConversionKind::ToOpenAIFamilyCli)
         );
+        assert_eq!(
+            sync_cli_response_conversion_kind("claude:cli", "openai:compact"),
+            None
+        );
+        assert_eq!(
+            sync_cli_response_conversion_kind("openai:compact", "claude:cli"),
+            None
+        );
         assert_eq!(
             sync_cli_response_conversion_kind("gemini:cli", "claude:cli"),
             Some(SyncCliResponseConversionKind::ToClaudeCli)
         );
     }
+
+    #[test]
+    fn request_candidate_registry_excludes_compact_as_cross_format_target() {
+        assert_eq!(
+            request_candidate_api_formats("openai:chat", false),
+            vec![
+                "openai:chat",
+                "openai:cli",
+                "claude:chat",
+                "claude:cli",
+                "gemini:chat",
+                "gemini:cli",
+            ]
+        );
+        assert_eq!(
+            request_candidate_api_formats("openai:cli", false),
+            vec![
+                "openai:chat",
+                "openai:cli",
+                "claude:chat",
+                "claude:cli",
+                "gemini:chat",
+                "gemini:cli",
+            ]
+        );
+        assert_eq!(
+            request_candidate_api_formats("openai:compact", false),
+            vec!["openai:compact"]
+        );
+    }
 }

apps/aether-gateway/src/ai_pipeline/finalize/tests_stream.rs

@@ -52,6 +52,130 @@ fn gemini_cli_v1internal_stream_rewriter_unwraps_response_object() {
     );
 }
 
+#[test]
+fn openai_chat_error_to_openai_cli_stream_rewriter_converts_to_response_failed() {
+    let report_context = json!({
+        "provider_api_format": "openai:chat",
+        "client_api_format": "openai:cli",
+        "needs_conversion": true,
+        "mapped_model": "gpt-5.4",
+    });
+    let mut rewriter =
+        maybe_build_local_stream_rewriter(Some(&report_context)).expect("rewriter should exist");
+    let output = rewriter
+        .push_chunk(
+            b"data: {\"error\":{\"message\":\"bad request\",\"type\":\"invalid_request_error\",\"code\":\"invalid_request\"}}\n\n",
+        )
+        .expect("rewrite should succeed");
+    let output_text = utf8(output);
+    assert!(output_text.contains("event: response.failed"));
+    assert!(output_text.contains("\"sequence_number\":1"));
+    assert!(output_text.contains("\"message\":\"bad request\""));
+    assert!(output_text.contains("\"type\":\"invalid_request_error\""));
+    assert!(!output_text.contains("data: [DONE]"));
+    assert!(rewriter.finish().expect("finish should succeed").is_empty());
+}
+
+#[test]
+fn claude_error_to_openai_cli_stream_rewriter_converts_to_response_failed() {
+    let report_context = json!({
+        "provider_api_format": "claude:chat",
+        "client_api_format": "openai:cli",
+        "needs_conversion": true,
+        "mapped_model": "claude-sonnet-4-5",
+    });
+    let mut rewriter =
+        maybe_build_local_stream_rewriter(Some(&report_context)).expect("rewriter should exist");
+    let output = rewriter
+        .push_chunk(
+            concat!(
+                "event: error\n",
+                "data: {\"type\":\"error\",\"error\":{\"type\":\"rate_limit_error\",\"message\":\"slow down\",\"code\":\"rate_limit\"}}\n\n"
+            )
+            .as_bytes(),
+        )
+        .expect("rewrite should succeed");
+    let output_text = utf8(output);
+    assert!(output_text.contains("event: response.failed"));
+    assert!(output_text.contains("\"sequence_number\":1"));
+    assert!(output_text.contains("\"message\":\"slow down\""));
+    assert!(output_text.contains("\"type\":\"rate_limit_error\""));
+    assert!(!output_text.contains("data: [DONE]"));
+    assert!(rewriter.finish().expect("finish should succeed").is_empty());
+}
+
+#[test]
+fn gemini_error_to_openai_cli_stream_rewriter_converts_to_response_failed() {
+    let report_context = json!({
+        "provider_api_format": "gemini:cli",
+        "client_api_format": "openai:cli",
+        "needs_conversion": true,
+        "mapped_model": "gemini-2.5-pro",
+    });
+    let mut rewriter =
+        maybe_build_local_stream_rewriter(Some(&report_context)).expect("rewriter should exist");
+    let output = rewriter
+        .push_chunk(
+            b"data: {\"error\":{\"code\":429,\"message\":\"quota exceeded\",\"status\":\"RESOURCE_EXHAUSTED\"}}\n\n",
+        )
+        .expect("rewrite should succeed");
+    let output_text = utf8(output);
+    assert!(output_text.contains("event: response.failed"));
+    assert!(output_text.contains("\"sequence_number\":1"));
+    assert!(output_text.contains("\"message\":\"quota exceeded\""));
+    assert!(output_text.contains("\"type\":\"rate_limit_error\""));
+    assert!(!output_text.contains("data: [DONE]"));
+    assert!(rewriter.finish().expect("finish should succeed").is_empty());
+}
+
+#[test]
+fn openai_chat_error_to_claude_chat_stream_rewriter_uses_error_event_line() {
+    let report_context = json!({
+        "provider_api_format": "openai:chat",
+        "client_api_format": "claude:chat",
+        "needs_conversion": true,
+        "mapped_model": "gpt-5.4",
+    });
+    let mut rewriter =
+        maybe_build_local_stream_rewriter(Some(&report_context)).expect("rewriter should exist");
+    let output = rewriter
+        .push_chunk(
+            b"data: {\"error\":{\"message\":\"bad request\",\"type\":\"invalid_request_error\",\"code\":\"invalid_request\"}}\n\n",
+        )
+        .expect("rewrite should succeed");
+    let output_text = utf8(output);
+    assert!(output_text.contains("event: error"));
+    assert!(output_text.contains("\"type\":\"error\""));
+    assert!(output_text.contains("\"message\":\"bad request\""));
+    assert!(output_text.contains("\"code\":\"invalid_request\""));
+    assert!(!output_text.contains("data: [DONE]"));
+    assert!(rewriter.finish().expect("finish should succeed").is_empty());
+}
+
+#[test]
+fn openai_chat_error_to_gemini_chat_stream_rewriter_keeps_data_only_error() {
+    let report_context = json!({
+        "provider_api_format": "openai:chat",
+        "client_api_format": "gemini:chat",
+        "needs_conversion": true,
+        "mapped_model": "gpt-5.4",
+    });
+    let mut rewriter =
+        maybe_build_local_stream_rewriter(Some(&report_context)).expect("rewriter should exist");
+    let output = rewriter
+        .push_chunk(
+            b"data: {\"error\":{\"message\":\"rate limited\",\"type\":\"rate_limit_error\",\"code\":\"rate_limit\"}}\n\n",
+        )
+        .expect("rewrite should succeed");
+    let output_text = utf8(output);
+    assert!(output_text.starts_with("data: {\"error\":"));
+    assert!(!output_text.contains("event: "));
+    assert!(output_text.contains("\"message\":\"rate limited\""));
+    assert!(output_text.contains("\"status\":\"RESOURCE_EXHAUSTED\""));
+    assert!(!output_text.contains("data: [DONE]"));
+    assert!(rewriter.finish().expect("finish should succeed").is_empty());
+}
+
 #[test]
 fn claude_to_openai_chat_stream_rewriter_converts_text_deltas() {
     let report_context = json!({
@@ -234,6 +358,22 @@ fn openai_cli_to_openai_chat_stream_rewriter_converts_text_deltas_immediately()
         .expect("rewrite should succeed");
     let completed_text = String::from_utf8(completed).expect("utf8 should decode");
     assert!(completed_text.contains("\"finish_reason\":\"stop\""));
+    assert!(
+        completed_text.contains("\"choices\":[]"),
+        "{completed_text}"
+    );
+    assert!(
+        completed_text.contains("\"prompt_tokens\":1"),
+        "{completed_text}"
+    );
+    assert!(
+        completed_text.contains("\"completion_tokens\":2"),
+        "{completed_text}"
+    );
+    assert!(
+        completed_text.contains("\"total_tokens\":3"),
+        "{completed_text}"
+    );
     assert!(completed_text.contains("data: [DONE]"));
     assert!(rewriter.finish().expect("finish should succeed").is_empty());
 }
@@ -262,6 +402,13 @@ fn openai_cli_to_openai_chat_stream_rewriter_converts_completed_event_without_bu
     assert!(output_text.contains("\"role\":\"assistant\""));
     assert!(output_text.contains("\"content\":\"Hello Codex\""));
     assert!(output_text.contains("\"finish_reason\":\"stop\""));
+    assert!(output_text.contains("\"choices\":[]"), "{output_text}");
+    assert!(output_text.contains("\"prompt_tokens\":1"), "{output_text}");
+    assert!(
+        output_text.contains("\"completion_tokens\":2"),
+        "{output_text}"
+    );
+    assert!(output_text.contains("\"total_tokens\":3"), "{output_text}");
     assert!(output_text.contains("data: [DONE]"));
     assert!(rewriter.finish().expect("finish should succeed").is_empty());
 }
@@ -514,6 +661,7 @@ fn openai_chat_to_claude_chat_stream_rewriter_converts_via_standard_matrix() {
     assert!(first_text.contains("event: content_block_start"));
     assert!(first_text.contains("event: content_block_delta"));
     assert!(first_text.contains("\"text\":\"Hello Claude\""));
+    assert!(first_text.contains("\"usage\":{\"input_tokens\":0,\"output_tokens\":0}"));
 
     let second = rewriter
         .push_chunk(
@@ -531,6 +679,44 @@ fn openai_chat_to_claude_chat_stream_rewriter_converts_via_standard_matrix() {
     assert!(rewriter.finish().expect("finish should succeed").is_empty());
 }
 
+#[test]
+fn openai_chat_to_claude_chat_stream_rewriter_injects_default_usage_when_finish_chunk_lacks_usage()
+{
+    let report_context = json!({
+        "provider_api_format": "openai:chat",
+        "client_api_format": "claude:chat",
+        "needs_conversion": true,
+        "mapped_model": "gpt-5.4",
+    });
+    let mut rewriter =
+        maybe_build_local_stream_rewriter(Some(&report_context)).expect("rewriter should exist");
+    let first = rewriter
+        .push_chunk(
+            "data: {\"id\":\"chatcmpl_usage_missing_123\",\"object\":\"chat.completion.chunk\",\"created\":1,\"model\":\"gpt-5.4\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"Hello\"},\"finish_reason\":null}]}\n\n"
+                .as_bytes(),
+        )
+        .expect("rewrite should succeed");
+    let first_text = utf8(first);
+    assert!(first_text.contains("event: message_start"));
+
+    let second = rewriter
+        .push_chunk(
+            concat!(
+                "data: {\"id\":\"chatcmpl_usage_missing_123\",\"object\":\"chat.completion.chunk\",\"model\":\"gpt-5.4\",\"choices\":[{\"index\":0,\"delta\":{},\"finish_reason\":\"stop\"}]}\n\n",
+                "data: [DONE]\n\n"
+            )
+            .as_bytes(),
+        )
+        .expect("rewrite should succeed");
+    let output_text = utf8(second);
+    assert!(output_text.is_empty(), "{output_text}");
+    let final_text = utf8(rewriter.finish().expect("finish should succeed"));
+    assert!(final_text.contains("event: message_delta"));
+    assert!(final_text.contains("\"stop_reason\":\"end_turn\""));
+    assert!(final_text.contains("\"usage\":{\"input_tokens\":0,\"output_tokens\":0}"));
+    assert!(final_text.contains("event: message_stop"));
+}
+
 #[test]
 fn openai_chat_to_gemini_cli_stream_rewriter_converts_via_standard_matrix() {
     let report_context = json!({

apps/aether-gateway/src/ai_pipeline/mod.rs

@@ -68,6 +68,8 @@ pub(crate) fn build_execution_runtime_auth_context(
     ExecutionRuntimeAuthContext {
         user_id: auth_context.user_id.clone(),
         api_key_id: auth_context.api_key_id.clone(),
+        username: auth_context.username.clone(),
+        api_key_name: auth_context.api_key_name.clone(),
         balance_remaining: auth_context.balance_remaining,
         access_allowed: auth_context.access_allowed,
     }

apps/aether-gateway/src/ai_pipeline/planner/candidate_affinity.rs

@@ -1,9 +1,18 @@
 use tracing::warn;
 
 use crate::ai_pipeline::transport::resolve_transport_proxy_snapshot;
-use crate::ai_pipeline::PlannerAppState;
-use crate::AppState;
-use aether_scheduler_core::SchedulerMinimalCandidateSelectionCandidate;
+use crate::ai_pipeline::{
+    GatewayAuthApiKeySnapshot, GatewayProviderTransportSnapshot, PlannerAppState,
+};
+use crate::scheduler::affinity::SCHEDULER_AFFINITY_TTL;
+use crate::scheduler::config::{read_scheduler_ordering_config, SchedulerOrderingConfig};
+use aether_scheduler_core::{
+    build_scheduler_affinity_cache_key_for_api_key_id, compare_candidates_by_priority_mode,
+    requested_capability_priority_for_candidate, SchedulerAffinityTarget,
+    SchedulerMinimalCandidateSelectionCandidate,
+};
+
+const PLANNER_SCHEDULER_AFFINITY_MAX_ENTRIES: usize = 10_000;
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
 enum TunnelOwnerAffinityBucket {
@@ -12,6 +21,12 @@ enum TunnelOwnerAffinityBucket {
     RemoteTunnel = 2,
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+struct CandidateExecutionOrdering {
+    tunnel_bucket: TunnelOwnerAffinityBucket,
+    keep_priority_on_conversion: bool,
+}
+
 pub(crate) async fn prefer_local_tunnel_owner_candidates(
     state: PlannerAppState<'_>,
     candidates: Vec<SchedulerMinimalCandidateSelectionCandidate>,
@@ -28,11 +43,127 @@ pub(crate) async fn prefer_local_tunnel_owner_candidates(
         .collect()
 }
 
+pub(crate) async fn rank_local_execution_candidates(
+    state: PlannerAppState<'_>,
+    candidates: Vec<SchedulerMinimalCandidateSelectionCandidate>,
+    client_api_format: &str,
+    required_capabilities: Option<&serde_json::Value>,
+) -> Vec<SchedulerMinimalCandidateSelectionCandidate> {
+    let normalized_client_api_format = client_api_format.trim().to_ascii_lowercase();
+    let ordering_config = read_scheduler_ordering_config_or_default(state).await;
+    let mut ranked = Vec::with_capacity(candidates.len());
+
+    for (original_index, candidate) in candidates.into_iter().enumerate() {
+        let ordering =
+            resolve_candidate_execution_ordering(state, &candidate, ordering_config).await;
+        let is_same_format = candidate
+            .endpoint_api_format
+            .trim()
+            .eq_ignore_ascii_case(normalized_client_api_format.as_str());
+        let demote_cross_format = !is_same_format && !ordering.keep_priority_on_conversion;
+        let capability_priority =
+            requested_capability_priority_for_candidate(required_capabilities, &candidate);
+        ranked.push((
+            capability_priority.0,
+            capability_priority.1,
+            ordering.tunnel_bucket,
+            demote_cross_format,
+            original_index,
+            candidate,
+        ));
+    }
+
+    ranked.sort_by(|left, right| {
+        left.0
+            .cmp(&right.0)
+            .then(left.1.cmp(&right.1))
+            .then(left.2.cmp(&right.2))
+            .then(left.3.cmp(&right.3))
+            .then_with(|| {
+                compare_candidates_by_priority_mode(
+                    &left.5,
+                    &right.5,
+                    ordering_config.priority_mode,
+                    None,
+                )
+            })
+            .then(left.4.cmp(&right.4))
+    });
+
+    ranked
+        .into_iter()
+        .map(|(_, _, _, _, _, candidate)| candidate)
+        .collect()
+}
+
+pub(crate) fn remember_scheduler_affinity_for_candidate(
+    state: PlannerAppState<'_>,
+    auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
+    client_api_format: &str,
+    requested_model: &str,
+    candidate: &SchedulerMinimalCandidateSelectionCandidate,
+) {
+    let Some(api_key_id) = auth_snapshot
+        .map(|snapshot| snapshot.api_key_id.trim())
+        .filter(|value| !value.is_empty())
+    else {
+        return;
+    };
+    let Some(cache_key) = build_scheduler_affinity_cache_key_for_api_key_id(
+        api_key_id,
+        client_api_format,
+        requested_model,
+    ) else {
+        return;
+    };
+
+    state.app().remember_scheduler_affinity_target(
+        &cache_key,
+        SchedulerAffinityTarget {
+            provider_id: candidate.provider_id.clone(),
+            endpoint_id: candidate.endpoint_id.clone(),
+            key_id: candidate.key_id.clone(),
+        },
+        SCHEDULER_AFFINITY_TTL,
+        PLANNER_SCHEDULER_AFFINITY_MAX_ENTRIES,
+    );
+}
+
 async fn resolve_candidate_tunnel_owner_affinity(
     state: PlannerAppState<'_>,
     candidate: &SchedulerMinimalCandidateSelectionCandidate,
 ) -> TunnelOwnerAffinityBucket {
-    let transport = match state
+    let Some(transport) = read_candidate_transport_snapshot(state, candidate).await else {
+        return TunnelOwnerAffinityBucket::Neutral;
+    };
+
+    resolve_tunnel_owner_affinity_from_transport(state, &transport).await
+}
+
+async fn resolve_candidate_execution_ordering(
+    state: PlannerAppState<'_>,
+    candidate: &SchedulerMinimalCandidateSelectionCandidate,
+    ordering_config: SchedulerOrderingConfig,
+) -> CandidateExecutionOrdering {
+    let Some(transport) = read_candidate_transport_snapshot(state, candidate).await else {
+        return CandidateExecutionOrdering {
+            tunnel_bucket: TunnelOwnerAffinityBucket::Neutral,
+            keep_priority_on_conversion: ordering_config.keep_priority_on_conversion,
+        };
+    };
+
+    CandidateExecutionOrdering {
+        tunnel_bucket: resolve_tunnel_owner_affinity_from_transport(state, &transport).await,
+        keep_priority_on_conversion: ordering_config.keep_priority_on_conversion
+            || transport.provider.keep_priority_on_conversion,
+    }
+}
+
+async fn read_candidate_transport_snapshot(
+    state: PlannerAppState<'_>,
+    candidate: &SchedulerMinimalCandidateSelectionCandidate,
+) -> Option<GatewayProviderTransportSnapshot> {
+    match state
         .read_provider_transport_snapshot(
             &candidate.provider_id,
             &candidate.endpoint_id,
@@ -40,8 +171,8 @@ async fn resolve_candidate_tunnel_owner_affinity(
         )
         .await
     {
-        Ok(Some(transport)) => transport,
-        Ok(None) => return TunnelOwnerAffinityBucket::Neutral,
+        Ok(Some(transport)) => Some(transport),
+        Ok(None) => None,
         Err(error) => {
             warn!(
                 event_name = "candidate_affinity_transport_load_failed",
@@ -50,12 +181,17 @@ async fn resolve_candidate_tunnel_owner_affinity(
                 endpoint_id = %candidate.endpoint_id,
                 key_id = %candidate.key_id,
                 error = ?error,
-                "failed to load provider transport while evaluating tunnel owner affinity"
+                "failed to load provider transport while evaluating execution ordering"
             );
-            return TunnelOwnerAffinityBucket::Neutral;
+            None
         }
-    };
+    }
+}
 
+async fn resolve_tunnel_owner_affinity_from_transport(
+    state: PlannerAppState<'_>,
+    transport: &GatewayProviderTransportSnapshot,
+) -> TunnelOwnerAffinityBucket {
     let Some(proxy) = resolve_transport_proxy_snapshot(&transport) else {
         return TunnelOwnerAffinityBucket::Neutral;
     };
@@ -99,6 +235,23 @@ async fn resolve_candidate_tunnel_owner_affinity(
     }
 }
 
+async fn read_scheduler_ordering_config_or_default(
+    state: PlannerAppState<'_>,
+) -> SchedulerOrderingConfig {
+    match read_scheduler_ordering_config(state.app()).await {
+        Ok(config) => config,
+        Err(error) => {
+            warn!(
+                event_name = "planner_scheduler_ordering_config_load_failed",
+                log_type = "event",
+                error = ?error,
+                "failed to load scheduler ordering config while ranking local execution candidates"
+            );
+            SchedulerOrderingConfig::default()
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use std::time::{SystemTime, UNIX_EPOCH};
@@ -110,11 +263,15 @@ mod tests {
     use serde_json::json;
 
     use super::{
-        prefer_local_tunnel_owner_candidates, AppState, PlannerAppState,
+        prefer_local_tunnel_owner_candidates, rank_local_execution_candidates,
+        remember_scheduler_affinity_for_candidate, PlannerAppState,
         SchedulerMinimalCandidateSelectionCandidate,
     };
+    use crate::data::auth::GatewayAuthApiKeySnapshot;
     use crate::data::GatewayDataState;
     use crate::tunnel::TunnelAttachmentRecord;
+    use crate::{scheduler::affinity::SCHEDULER_AFFINITY_TTL, AppState};
+    use aether_data::repository::auth::StoredAuthApiKeySnapshot;
 
     fn sample_candidate(
         endpoint_id: &str,
@@ -142,22 +299,55 @@ mod tests {
     }
 
     fn sample_provider() -> StoredProviderCatalogProvider {
+        sample_provider_with_options("provider-1", false, 0)
+    }
+
+    fn sample_provider_with_options(
+        id: &str,
+        keep_priority_on_conversion: bool,
+        provider_priority: i32,
+    ) -> StoredProviderCatalogProvider {
         StoredProviderCatalogProvider::new(
-            "provider-1".to_string(),
-            "provider-1".to_string(),
+            id.to_string(),
+            id.to_string(),
             Some("https://provider.example".to_string()),
             "custom".to_string(),
         )
         .expect("provider should build")
-        .with_transport_fields(true, false, false, None, None, None, None, None, None)
+        .with_transport_fields(
+            true,
+            keep_priority_on_conversion,
+            false,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+        )
+        .with_routing_fields(provider_priority)
     }
 
     fn sample_endpoint(id: &str) -> StoredProviderCatalogEndpoint {
+        sample_endpoint_for_provider("provider-1", id, "openai:chat")
+    }
+
+    fn sample_endpoint_for_provider(
+        provider_id: &str,
+        id: &str,
+        api_format: &str,
+    ) -> StoredProviderCatalogEndpoint {
         StoredProviderCatalogEndpoint::new(
             id.to_string(),
-            "provider-1".to_string(),
-            "openai:chat".to_string(),
-            Some("openai".to_string()),
+            provider_id.to_string(),
+            api_format.to_string(),
+            Some(
+                api_format
+                    .split(':')
+                    .next()
+                    .unwrap_or(api_format)
+                    .to_string(),
+            ),
             Some("chat".to_string()),
             true,
         )
@@ -176,9 +366,17 @@ mod tests {
     }
 
     fn sample_key(id: &str, node_id: &str) -> StoredProviderCatalogKey {
+        sample_key_for_provider("provider-1", id, node_id)
+    }
+
+    fn sample_key_for_provider(
+        provider_id: &str,
+        id: &str,
+        node_id: &str,
+    ) -> StoredProviderCatalogKey {
         StoredProviderCatalogKey::new(
             id.to_string(),
-            "provider-1".to_string(),
+            provider_id.to_string(),
             id.to_string(),
             "api_key".to_string(),
             None,
@@ -214,6 +412,35 @@ mod tests {
             .as_secs()
     }
 
+    fn sample_priority_candidate(
+        provider_id: &str,
+        endpoint_id: &str,
+        key_id: &str,
+        endpoint_api_format: &str,
+        key_global_priority_for_format: Option<i32>,
+        provider_priority: i32,
+    ) -> SchedulerMinimalCandidateSelectionCandidate {
+        SchedulerMinimalCandidateSelectionCandidate {
+            provider_id: provider_id.to_string(),
+            provider_name: provider_id.to_string(),
+            provider_type: "custom".to_string(),
+            provider_priority,
+            endpoint_id: endpoint_id.to_string(),
+            endpoint_api_format: endpoint_api_format.to_string(),
+            key_id: key_id.to_string(),
+            key_name: key_id.to_string(),
+            key_auth_type: "api_key".to_string(),
+            key_internal_priority: 0,
+            key_global_priority_for_format,
+            key_capabilities: None,
+            model_id: format!("model-{provider_id}"),
+            global_model_id: "global-model-1".to_string(),
+            global_model_name: "gpt-4.1".to_string(),
+            selected_provider_model_name: "gpt-4.1".to_string(),
+            mapping_matched_model: None,
+        }
+    }
+
     #[tokio::test]
     async fn prefers_local_tunnel_owner_candidates_before_remote_tunnel_candidates() {
         let provider_catalog = InMemoryProviderCatalogReadRepository::seed(
@@ -300,4 +527,339 @@ mod tests {
         assert_eq!(reordered[0].endpoint_id, "endpoint-a");
         assert_eq!(reordered[1].endpoint_id, "endpoint-b");
     }
+
+    #[tokio::test]
+    async fn local_execution_ranking_demotes_cross_format_candidates_without_keep_priority() {
+        let provider_catalog = InMemoryProviderCatalogReadRepository::seed(
+            vec![
+                sample_provider_with_options("provider-same", false, 10),
+                sample_provider_with_options("provider-cross", false, 0),
+            ],
+            vec![
+                sample_endpoint_for_provider("provider-same", "endpoint-same", "openai:chat"),
+                sample_endpoint_for_provider("provider-cross", "endpoint-cross", "claude:chat"),
+            ],
+            vec![
+                sample_key_for_provider("provider-same", "key-same", ""),
+                sample_key_for_provider("provider-cross", "key-cross", ""),
+            ],
+        );
+        let data_state = GatewayDataState::with_provider_transport_reader_for_tests(
+            std::sync::Arc::new(provider_catalog),
+            "development-key",
+        );
+        let state = AppState::new()
+            .expect("state should build")
+            .with_data_state_for_tests(data_state);
+
+        let ranked = rank_local_execution_candidates(
+            PlannerAppState::new(&state),
+            vec![
+                sample_priority_candidate(
+                    "provider-cross",
+                    "endpoint-cross",
+                    "key-cross",
+                    "claude:chat",
+                    Some(0),
+                    0,
+                ),
+                sample_priority_candidate(
+                    "provider-same",
+                    "endpoint-same",
+                    "key-same",
+                    "openai:chat",
+                    Some(10),
+                    10,
+                ),
+            ],
+            "openai:chat",
+            None,
+        )
+        .await;
+
+        assert_eq!(ranked[0].endpoint_id, "endpoint-same");
+        assert_eq!(ranked[1].endpoint_id, "endpoint-cross");
+    }
+
+    #[tokio::test]
+    async fn local_execution_ranking_keeps_cross_format_priority_when_enabled() {
+        let provider_catalog = InMemoryProviderCatalogReadRepository::seed(
+            vec![
+                sample_provider_with_options("provider-same", false, 10),
+                sample_provider_with_options("provider-cross", true, 0),
+            ],
+            vec![
+                sample_endpoint_for_provider("provider-same", "endpoint-same", "openai:chat"),
+                sample_endpoint_for_provider("provider-cross", "endpoint-cross", "claude:chat"),
+            ],
+            vec![
+                sample_key_for_provider("provider-same", "key-same", ""),
+                sample_key_for_provider("provider-cross", "key-cross", ""),
+            ],
+        );
+        let data_state = GatewayDataState::with_provider_transport_reader_for_tests(
+            std::sync::Arc::new(provider_catalog),
+            "development-key",
+        );
+        let state = AppState::new()
+            .expect("state should build")
+            .with_data_state_for_tests(data_state);
+
+        let ranked = rank_local_execution_candidates(
+            PlannerAppState::new(&state),
+            vec![
+                sample_priority_candidate(
+                    "provider-cross",
+                    "endpoint-cross",
+                    "key-cross",
+                    "claude:chat",
+                    Some(0),
+                    0,
+                ),
+                sample_priority_candidate(
+                    "provider-same",
+                    "endpoint-same",
+                    "key-same",
+                    "openai:chat",
+                    Some(10),
+                    10,
+                ),
+            ],
+            "openai:chat",
+            None,
+        )
+        .await;
+
+        assert_eq!(ranked[0].endpoint_id, "endpoint-cross");
+        assert_eq!(ranked[1].endpoint_id, "endpoint-same");
+    }
+
+    #[tokio::test]
+    async fn local_execution_ranking_keeps_cross_format_priority_when_global_override_is_enabled() {
+        let provider_catalog = InMemoryProviderCatalogReadRepository::seed(
+            vec![
+                sample_provider_with_options("provider-same", false, 10),
+                sample_provider_with_options("provider-cross", false, 0),
+            ],
+            vec![
+                sample_endpoint_for_provider("provider-same", "endpoint-same", "openai:chat"),
+                sample_endpoint_for_provider("provider-cross", "endpoint-cross", "claude:chat"),
+            ],
+            vec![
+                sample_key_for_provider("provider-same", "key-same", ""),
+                sample_key_for_provider("provider-cross", "key-cross", ""),
+            ],
+        );
+        let data_state = GatewayDataState::with_provider_transport_reader_for_tests(
+            std::sync::Arc::new(provider_catalog),
+            "development-key",
+        )
+        .with_system_config_values_for_tests(vec![(
+            "keep_priority_on_conversion".to_string(),
+            json!(true),
+        )]);
+        let state = AppState::new()
+            .expect("state should build")
+            .with_data_state_for_tests(data_state);
+
+        let ranked = rank_local_execution_candidates(
+            PlannerAppState::new(&state),
+            vec![
+                sample_priority_candidate(
+                    "provider-cross",
+                    "endpoint-cross",
+                    "key-cross",
+                    "claude:chat",
+                    Some(0),
+                    0,
+                ),
+                sample_priority_candidate(
+                    "provider-same",
+                    "endpoint-same",
+                    "key-same",
+                    "openai:chat",
+                    Some(10),
+                    10,
+                ),
+            ],
+            "openai:chat",
+            None,
+        )
+        .await;
+
+        assert_eq!(ranked[0].endpoint_id, "endpoint-cross");
+        assert_eq!(ranked[1].endpoint_id, "endpoint-same");
+    }
+
+    #[tokio::test]
+    async fn local_execution_ranking_uses_provider_priority_mode_when_configured() {
+        let provider_catalog = InMemoryProviderCatalogReadRepository::seed(
+            vec![
+                sample_provider_with_options("provider-provider-first", false, 0),
+                sample_provider_with_options("provider-global-first", false, 10),
+            ],
+            vec![
+                sample_endpoint_for_provider(
+                    "provider-provider-first",
+                    "endpoint-provider-first",
+                    "openai:chat",
+                ),
+                sample_endpoint_for_provider(
+                    "provider-global-first",
+                    "endpoint-global-first",
+                    "openai:chat",
+                ),
+            ],
+            vec![
+                sample_key_for_provider("provider-provider-first", "key-provider-first", ""),
+                sample_key_for_provider("provider-global-first", "key-global-first", ""),
+            ],
+        );
+        let data_state = GatewayDataState::with_provider_transport_reader_for_tests(
+            std::sync::Arc::new(provider_catalog),
+            "development-key",
+        )
+        .with_system_config_values_for_tests(vec![(
+            "provider_priority_mode".to_string(),
+            json!("provider"),
+        )]);
+        let state = AppState::new()
+            .expect("state should build")
+            .with_data_state_for_tests(data_state);
+
+        let ranked = rank_local_execution_candidates(
+            PlannerAppState::new(&state),
+            vec![
+                sample_priority_candidate(
+                    "provider-global-first",
+                    "endpoint-global-first",
+                    "key-global-first",
+                    "openai:chat",
+                    Some(0),
+                    10,
+                ),
+                sample_priority_candidate(
+                    "provider-provider-first",
+                    "endpoint-provider-first",
+                    "key-provider-first",
+                    "openai:chat",
+                    Some(10),
+                    0,
+                ),
+            ],
+            "openai:chat",
+            None,
+        )
+        .await;
+
+        assert_eq!(ranked[0].endpoint_id, "endpoint-provider-first");
+        assert_eq!(ranked[1].endpoint_id, "endpoint-global-first");
+    }
+
+    #[tokio::test]
+    async fn local_execution_ranking_prefers_candidates_matching_requested_capabilities() {
+        let provider_catalog = InMemoryProviderCatalogReadRepository::seed(
+            vec![
+                sample_provider_with_options("provider-miss", false, 0),
+                sample_provider_with_options("provider-hit", false, 0),
+            ],
+            vec![
+                sample_endpoint_for_provider("provider-miss", "endpoint-miss", "openai:chat"),
+                sample_endpoint_for_provider("provider-hit", "endpoint-hit", "openai:chat"),
+            ],
+            vec![
+                sample_key_for_provider("provider-miss", "key-miss", ""),
+                sample_key_for_provider("provider-hit", "key-hit", ""),
+            ],
+        );
+        let data_state = GatewayDataState::with_provider_transport_reader_for_tests(
+            std::sync::Arc::new(provider_catalog),
+            "development-key",
+        );
+        let state = AppState::new()
+            .expect("state should build")
+            .with_data_state_for_tests(data_state);
+
+        let mut candidate_miss = sample_priority_candidate(
+            "provider-miss",
+            "endpoint-miss",
+            "key-miss",
+            "openai:chat",
+            Some(0),
+            0,
+        );
+        let mut candidate_hit = sample_priority_candidate(
+            "provider-hit",
+            "endpoint-hit",
+            "key-hit",
+            "openai:chat",
+            Some(0),
+            0,
+        );
+        candidate_miss.key_capabilities = Some(json!({"cache_1h": false}));
+        candidate_hit.key_capabilities = Some(json!({"cache_1h": true}));
+
+        let required_capabilities = json!({"cache_1h": true});
+        let ranked = rank_local_execution_candidates(
+            PlannerAppState::new(&state),
+            vec![candidate_miss, candidate_hit],
+            "openai:chat",
+            Some(&required_capabilities),
+        )
+        .await;
+
+        assert_eq!(ranked[0].endpoint_id, "endpoint-hit");
+        assert_eq!(ranked[1].endpoint_id, "endpoint-miss");
+    }
+
+    #[tokio::test]
+    async fn remembers_scheduler_affinity_for_candidate_using_requested_model_key() {
+        let state = AppState::new().expect("state should build");
+        let auth_snapshot = GatewayAuthApiKeySnapshot::from_stored(
+            StoredAuthApiKeySnapshot::new(
+                "user-1".to_string(),
+                "alice".to_string(),
+                Some("alice@example.com".to_string()),
+                "user".to_string(),
+                "local".to_string(),
+                true,
+                false,
+                None,
+                None,
+                None,
+                "api-key-1".to_string(),
+                Some("default".to_string()),
+                true,
+                false,
+                false,
+                Some(60),
+                Some(5),
+                Some(4_102_444_800),
+                None,
+                None,
+                None,
+            )
+            .expect("stored auth snapshot should build"),
+            current_unix_secs(),
+        );
+        let candidate = sample_candidate("endpoint-1", "key-1");
+
+        remember_scheduler_affinity_for_candidate(
+            PlannerAppState::new(&state),
+            Some(&auth_snapshot),
+            "openai:chat",
+            "gpt-5",
+            &candidate,
+        );
+
+        let remembered = state
+            .read_scheduler_affinity_target(
+                "scheduler_affinity:api-key-1:openai:chat:gpt-5",
+                SCHEDULER_AFFINITY_TTL,
+            )
+            .expect("affinity target should be cached");
+        assert_eq!(remembered.provider_id, "provider-1");
+        assert_eq!(remembered.endpoint_id, "endpoint-1");
+        assert_eq!(remembered.key_id, "key-1");
+    }
 }

apps/aether-gateway/src/ai_pipeline/planner/passthrough/provider/family/candidates.rs

@@ -2,7 +2,9 @@ use serde_json::json;
 use tracing::warn;
 use uuid::Uuid;
 
-use crate::ai_pipeline::planner::candidate_affinity::prefer_local_tunnel_owner_candidates;
+use crate::ai_pipeline::planner::candidate_affinity::{
+    rank_local_execution_candidates, remember_scheduler_affinity_for_candidate,
+};
 use crate::ai_pipeline::{
     resolve_local_decision_execution_runtime_auth_context, ConversionMode, ExecutionStrategy,
     GatewayControlDecision, PlannerAppState,
@@ -61,10 +63,20 @@ pub(crate) async fn resolve_local_same_format_provider_decision_input(
         }
     };
 
+    let required_capabilities = planner_state
+        .resolve_request_candidate_required_capabilities(
+            &auth_context.user_id,
+            &auth_context.api_key_id,
+            Some(requested_model.as_str()),
+            None,
+        )
+        .await;
+
     Some(LocalSameFormatProviderDecisionInput {
         auth_context,
         requested_model,
         auth_snapshot,
+        required_capabilities,
     })
 }
 
@@ -80,16 +92,34 @@ pub(crate) async fn materialize_local_same_format_provider_candidate_attempts(
             spec.api_format,
             &input.requested_model,
             spec.require_streaming,
+            input.required_capabilities.as_ref(),
             Some(&input.auth_snapshot),
             current_unix_secs(),
         )
         .await?;
-    let candidates = prefer_local_tunnel_owner_candidates(planner_state, candidates).await;
+    let candidates = rank_local_execution_candidates(
+        planner_state,
+        candidates,
+        spec.api_format,
+        input.required_capabilities.as_ref(),
+    )
+    .await;
 
-    let created_at_unix_secs = current_unix_secs();
+    let created_at_unix_ms = current_unix_secs();
     let mut attempts = Vec::with_capacity(candidates.len());
+    let mut affinity_remembered = false;
     for (candidate_index, candidate) in candidates.into_iter().enumerate() {
         let generated_candidate_id = Uuid::new_v4().to_string();
+        if !affinity_remembered {
+            remember_scheduler_affinity_for_candidate(
+                planner_state,
+                Some(&input.auth_snapshot),
+                spec.api_format,
+                &input.requested_model,
+                &candidate,
+            );
+            affinity_remembered = true;
+        }
         let extra_data = append_execution_contract_fields_to_value(
             json!({
                 "provider_api_format": spec.api_format,
@@ -116,8 +146,9 @@ pub(crate) async fn materialize_local_same_format_provider_candidate_attempts(
                 &candidate,
                 candidate_index as u32,
                 &generated_candidate_id,
+                input.required_capabilities.as_ref(),
                 Some(extra_data),
-                created_at_unix_secs,
+                created_at_unix_ms,
                 "gateway local same-format decision request candidate upsert failed",
             )
             .await;

apps/aether-gateway/src/ai_pipeline/planner/passthrough/provider/family/mod.rs

@@ -21,6 +21,7 @@ pub(crate) struct LocalSameFormatProviderDecisionInput {
     pub(crate) auth_context: ExecutionRuntimeAuthContext,
     pub(crate) requested_model: String,
     pub(crate) auth_snapshot: GatewayAuthApiKeySnapshot,
+    pub(crate) required_capabilities: Option<serde_json::Value>,
 }
 
 #[derive(Debug, Clone)]

apps/aether-gateway/src/ai_pipeline/planner/passthrough/provider/family/payload.rs

@@ -265,6 +265,8 @@ pub(crate) async fn maybe_build_local_same_format_provider_decision_payload_for_
         json!({
             "user_id": input.auth_context.user_id,
             "api_key_id": input.auth_context.api_key_id,
+            "username": input.auth_context.username,
+            "api_key_name": input.auth_context.api_key_name,
             "request_id": trace_id,
             "candidate_id": candidate_id,
             "candidate_index": candidate_index,
@@ -359,6 +361,7 @@ pub(super) async fn mark_skipped_local_same_format_provider_candidate(
             candidate,
             candidate_index,
             candidate_id,
+            input.required_capabilities.as_ref(),
             skip_reason,
             current_unix_secs(),
             "gateway local same-format decision failed to persist skipped candidate",

apps/aether-gateway/src/ai_pipeline/planner/specialized/files/decision.rs

@@ -281,6 +281,8 @@ pub(super) async fn maybe_build_local_gemini_files_decision_payload_for_candidat
         report_context: Some(json!({
             "user_id": input.auth_context.user_id,
             "api_key_id": input.auth_context.api_key_id,
+            "username": input.auth_context.username,
+            "api_key_name": input.auth_context.api_key_name,
             "request_id": trace_id,
             "candidate_id": candidate_id,
             "candidate_index": candidate_index,

apps/aether-gateway/src/ai_pipeline/planner/specialized/files/support.rs

@@ -4,7 +4,9 @@ use tracing::warn;
 use uuid::Uuid;
 
 use crate::ai_pipeline::contracts::ExecutionRuntimeAuthContext;
-use crate::ai_pipeline::planner::candidate_affinity::prefer_local_tunnel_owner_candidates;
+use crate::ai_pipeline::planner::candidate_affinity::{
+    rank_local_execution_candidates, remember_scheduler_affinity_for_candidate,
+};
 use crate::ai_pipeline::{
     resolve_local_decision_execution_runtime_auth_context, GatewayControlDecision,
 };
@@ -20,6 +22,7 @@ pub(super) const GEMINI_FILES_REQUIRED_CAPABILITY: &str = "gemini_files";
 pub(super) struct LocalGeminiFilesDecisionInput {
     pub(super) auth_context: ExecutionRuntimeAuthContext,
     pub(super) auth_snapshot: GatewayAuthApiKeySnapshot,
+    pub(super) required_capabilities: Option<serde_json::Value>,
 }
 
 #[derive(Debug, Clone)]
@@ -59,9 +62,20 @@ pub(super) async fn resolve_local_gemini_files_decision_input(
         }
     };
 
+    let explicit_required_capabilities = json!({ "gemini_files": true });
+    let required_capabilities = planner_state
+        .resolve_request_candidate_required_capabilities(
+            &auth_context.user_id,
+            &auth_context.api_key_id,
+            None,
+            Some(&explicit_required_capabilities),
+        )
+        .await;
+
     Some(LocalGeminiFilesDecisionInput {
         auth_context,
         auth_snapshot,
+        required_capabilities,
     })
 }
 
@@ -80,12 +94,29 @@ pub(super) async fn materialize_local_gemini_files_candidate_attempts(
             current_unix_secs(),
         )
         .await?;
-    let candidates = prefer_local_tunnel_owner_candidates(planner_state, candidates).await;
+    let candidates = rank_local_execution_candidates(
+        planner_state,
+        candidates,
+        GEMINI_FILES_CLIENT_API_FORMAT,
+        input.required_capabilities.as_ref(),
+    )
+    .await;
 
-    let created_at_unix_secs = current_unix_secs();
+    let created_at_unix_ms = current_unix_secs();
     let mut attempts = Vec::with_capacity(candidates.len());
+    let mut affinity_remembered = false;
     for (candidate_index, candidate) in candidates.into_iter().enumerate() {
         let generated_candidate_id = Uuid::new_v4().to_string();
+        if !affinity_remembered {
+            remember_scheduler_affinity_for_candidate(
+                planner_state,
+                Some(&input.auth_snapshot),
+                GEMINI_FILES_CLIENT_API_FORMAT,
+                &candidate.global_model_name,
+                &candidate,
+            );
+            affinity_remembered = true;
+        }
         let extra_data = json!({
             "provider_api_format": GEMINI_FILES_CLIENT_API_FORMAT,
             "client_api_format": GEMINI_FILES_CLIENT_API_FORMAT,
@@ -107,8 +138,9 @@ pub(super) async fn materialize_local_gemini_files_candidate_attempts(
                 &candidate,
                 candidate_index as u32,
                 &generated_candidate_id,
+                input.required_capabilities.as_ref(),
                 Some(extra_data),
-                created_at_unix_secs,
+                created_at_unix_ms,
                 "gateway local gemini files request candidate upsert failed",
             )
             .await;
@@ -140,6 +172,7 @@ pub(super) async fn mark_skipped_local_gemini_files_candidate(
             candidate,
             candidate_index,
             candidate_id,
+            input.required_capabilities.as_ref(),
             skip_reason,
             current_unix_secs(),
             "gateway local gemini files failed to persist skipped candidate",

apps/aether-gateway/src/ai_pipeline/planner/specialized/video/decision.rs

@@ -248,6 +248,8 @@ pub(super) async fn maybe_build_local_video_create_decision_payload_for_candidat
         report_context: Some(json!({
             "user_id": input.auth_context.user_id.clone(),
             "api_key_id": input.auth_context.api_key_id.clone(),
+            "username": input.auth_context.username.clone(),
+            "api_key_name": input.auth_context.api_key_name.clone(),
             "request_id": trace_id,
             "candidate_id": candidate_id,
             "candidate_index": candidate_index,

apps/aether-gateway/src/ai_pipeline/planner/specialized/video/support.rs

@@ -5,7 +5,9 @@ use uuid::Uuid;
 
 use super::{LocalVideoCreateFamily, LocalVideoCreateSpec};
 use crate::ai_pipeline::contracts::ExecutionRuntimeAuthContext;
-use crate::ai_pipeline::planner::candidate_affinity::prefer_local_tunnel_owner_candidates;
+use crate::ai_pipeline::planner::candidate_affinity::{
+    rank_local_execution_candidates, remember_scheduler_affinity_for_candidate,
+};
 use crate::ai_pipeline::{
     resolve_local_decision_execution_runtime_auth_context, GatewayControlDecision,
 };
@@ -18,6 +20,7 @@ pub(super) struct LocalVideoCreateDecisionInput {
     pub(super) auth_context: ExecutionRuntimeAuthContext,
     pub(super) requested_model: String,
     pub(super) auth_snapshot: GatewayAuthApiKeySnapshot,
+    pub(super) required_capabilities: Option<serde_json::Value>,
 }
 
 #[derive(Debug, Clone)]
@@ -71,10 +74,20 @@ pub(super) async fn resolve_local_video_create_decision_input(
         }
     };
 
+    let required_capabilities = planner_state
+        .resolve_request_candidate_required_capabilities(
+            &auth_context.user_id,
+            &auth_context.api_key_id,
+            Some(requested_model.as_str()),
+            None,
+        )
+        .await;
+
     Some(LocalVideoCreateDecisionInput {
         auth_context,
         requested_model,
         auth_snapshot,
+        required_capabilities,
     })
 }
 
@@ -91,6 +104,7 @@ pub(super) async fn list_local_video_create_candidate_attempts(
             api_format,
             &input.requested_model,
             false,
+            input.required_capabilities.as_ref(),
             Some(&input.auth_snapshot),
             current_unix_secs(),
         )
@@ -127,12 +141,29 @@ async fn materialize_local_video_create_candidate_attempts(
     candidates: Vec<SchedulerMinimalCandidateSelectionCandidate>,
     api_format: &str,
 ) -> Vec<LocalVideoCreateCandidateAttempt> {
-    let candidates = prefer_local_tunnel_owner_candidates(state, candidates).await;
-    let created_at_unix_secs = current_unix_secs();
+    let candidates = rank_local_execution_candidates(
+        state,
+        candidates,
+        api_format,
+        input.required_capabilities.as_ref(),
+    )
+    .await;
+    let created_at_unix_ms = current_unix_secs();
     let mut attempts = Vec::with_capacity(candidates.len());
+    let mut affinity_remembered = false;
 
     for (candidate_index, candidate) in candidates.into_iter().enumerate() {
         let generated_candidate_id = Uuid::new_v4().to_string();
+        if !affinity_remembered {
+            remember_scheduler_affinity_for_candidate(
+                state,
+                Some(&input.auth_snapshot),
+                api_format,
+                &input.requested_model,
+                &candidate,
+            );
+            affinity_remembered = true;
+        }
         let extra_data = json!({
             "provider_api_format": api_format,
             "client_api_format": api_format,
@@ -153,8 +184,9 @@ async fn materialize_local_video_create_candidate_attempts(
                 &candidate,
                 candidate_index as u32,
                 &generated_candidate_id,
+                input.required_capabilities.as_ref(),
                 Some(extra_data),
-                created_at_unix_secs,
+                created_at_unix_ms,
                 "gateway local video decision request candidate upsert failed",
             )
             .await;
@@ -186,6 +218,7 @@ pub(super) async fn mark_skipped_local_video_candidate(
             candidate,
             candidate_index,
             candidate_id,
+            input.required_capabilities.as_ref(),
             skip_reason,
             current_unix_secs(),
             "gateway local video decision failed to persist skipped candidate",

apps/aether-gateway/src/ai_pipeline/planner/standard/family/candidates.rs

@@ -5,7 +5,13 @@ use serde_json::json;
 use tracing::warn;
 use uuid::Uuid;
 
-use crate::ai_pipeline::planner::candidate_affinity::prefer_local_tunnel_owner_candidates;
+use crate::ai_pipeline::conversion::{
+    request_candidate_api_formats, request_conversion_kind,
+    request_conversion_requires_enable_flag, request_pair_allowed_for_transport,
+};
+use crate::ai_pipeline::planner::candidate_affinity::{
+    rank_local_execution_candidates, remember_scheduler_affinity_for_candidate,
+};
 use crate::ai_pipeline::{
     resolve_local_decision_execution_runtime_auth_context, ConversionMode, ExecutionStrategy,
     GatewayControlDecision,
@@ -16,7 +22,7 @@ use crate::{append_execution_contract_fields_to_value, AppState, GatewayError};
 
 use super::{
     LocalStandardCandidateAttempt, LocalStandardDecisionInput, LocalStandardSourceFamily,
-    LocalStandardSourceMode, LocalStandardSpec,
+    LocalStandardSpec,
 };
 
 pub(super) async fn resolve_local_standard_decision_input(
@@ -63,10 +69,20 @@ pub(super) async fn resolve_local_standard_decision_input(
         }
     };
 
+    let required_capabilities = planner_state
+        .resolve_request_candidate_required_capabilities(
+            &auth_context.user_id,
+            &auth_context.api_key_id,
+            Some(requested_model.as_str()),
+            None,
+        )
+        .await;
+
     Some(LocalStandardDecisionInput {
         auth_context,
         requested_model,
         auth_snapshot,
+        required_capabilities,
     })
 }
 
@@ -79,8 +95,10 @@ pub(super) async fn materialize_local_standard_candidate_attempts(
     let planner_state = PlannerAppState::new(state);
     let mut seen_candidates = BTreeSet::new();
     let mut candidates = Vec::new();
-    for candidate_api_format in candidate_api_formats_for_spec(spec) {
-        let auth_snapshot = if *candidate_api_format == spec.api_format {
+    for candidate_api_format in
+        request_candidate_api_formats(spec.api_format, spec.require_streaming)
+    {
+        let auth_snapshot = if candidate_api_format == spec.api_format {
             Some(&input.auth_snapshot)
         } else {
             None
@@ -90,6 +108,7 @@ pub(super) async fn materialize_local_standard_candidate_attempts(
                 candidate_api_format,
                 &input.requested_model,
                 spec.require_streaming,
+                input.required_capabilities.as_ref(),
                 auth_snapshot,
                 current_unix_secs(),
             )
@@ -118,13 +137,71 @@ pub(super) async fn materialize_local_standard_candidate_attempts(
             }
         }
     }
-    let candidates = prefer_local_tunnel_owner_candidates(planner_state, candidates).await;
+    let candidates = rank_local_execution_candidates(
+        planner_state,
+        candidates,
+        spec.api_format,
+        input.required_capabilities.as_ref(),
+    )
+    .await;
 
-    let created_at_unix_secs = current_unix_secs();
+    let created_at_unix_ms = current_unix_secs();
     let mut attempts = Vec::with_capacity(candidates.len());
+    let mut affinity_remembered = false;
     for (candidate_index, candidate) in candidates.into_iter().enumerate() {
         let candidate_id = Uuid::new_v4().to_string();
         let provider_api_format = candidate.endpoint_api_format.trim().to_ascii_lowercase();
+        if provider_api_format != spec.api_format {
+            if let Ok(Some(transport)) = planner_state
+                .read_provider_transport_snapshot(
+                    &candidate.provider_id,
+                    &candidate.endpoint_id,
+                    &candidate.key_id,
+                )
+                .await
+            {
+                if !request_pair_allowed_for_transport(
+                    &transport,
+                    spec.api_format,
+                    provider_api_format.as_str(),
+                ) {
+                    let skip_reason =
+                        if request_conversion_kind(spec.api_format, provider_api_format.as_str())
+                            .is_some()
+                            && request_conversion_requires_enable_flag(
+                                spec.api_format,
+                                provider_api_format.as_str(),
+                            )
+                            && !transport.provider.enable_format_conversion
+                        {
+                            "format_conversion_disabled"
+                        } else {
+                            "transport_unsupported"
+                        };
+                    super::payload::mark_skipped_local_standard_candidate(
+                        state,
+                        input,
+                        trace_id,
+                        &candidate,
+                        candidate_index as u32,
+                        &candidate_id,
+                        skip_reason,
+                    )
+                    .await;
+                    continue;
+                }
+            }
+        }
+        if !affinity_remembered {
+            remember_scheduler_affinity_for_candidate(
+                planner_state,
+                Some(&input.auth_snapshot),
+                spec.api_format,
+                &input.requested_model,
+                &candidate,
+            );
+            affinity_remembered = true;
+        }
         let execution_strategy = if provider_api_format == spec.api_format {
             ExecutionStrategy::LocalSameFormat
         } else {
@@ -166,8 +243,9 @@ pub(super) async fn materialize_local_standard_candidate_attempts(
                 &candidate,
                 candidate_index as u32,
                 &candidate_id,
+                input.required_capabilities.as_ref(),
                 Some(extra_data),
-                created_at_unix_secs,
+                created_at_unix_ms,
                 "gateway local standard decision request candidate upsert failed",
             )
             .await;
@@ -213,20 +291,6 @@ fn auth_snapshot_allows_cross_format_candidate(
     true
 }
 
-fn candidate_api_formats_for_spec(spec: LocalStandardSpec) -> &'static [&'static str] {
-    match spec.mode {
-        LocalStandardSourceMode::Chat | LocalStandardSourceMode::Cli => &[
-            "openai:chat",
-            "openai:cli",
-            "openai:compact",
-            "claude:chat",
-            "claude:cli",
-            "gemini:chat",
-            "gemini:cli",
-        ],
-    }
-}
-
 fn extract_gemini_model_from_path(path: &str) -> Option<String> {
     let marker = "/models/";
     let start = path.find(marker)? + marker.len();

apps/aether-gateway/src/ai_pipeline/planner/standard/family/mod.rs

@@ -18,6 +18,7 @@ pub(super) struct LocalStandardDecisionInput {
     pub(super) auth_context: ExecutionRuntimeAuthContext,
     pub(super) requested_model: String,
     pub(super) auth_snapshot: GatewayAuthApiKeySnapshot,
+    pub(super) required_capabilities: Option<serde_json::Value>,
 }
 
 #[derive(Debug, Clone)]

apps/aether-gateway/src/ai_pipeline/planner/standard/family/payload.rs

@@ -91,6 +91,33 @@ pub(super) async fn maybe_build_local_standard_decision_payload_for_candidate(
         }
     };
 
+    if !crate::ai_pipeline::conversion::request_pair_allowed_for_transport(
+        &transport,
+        spec.api_format,
+        provider_api_format.as_str(),
+    ) {
+        let skip_reason = if crate::ai_pipeline::conversion::request_conversion_requires_enable_flag(
+            spec.api_format,
+            provider_api_format.as_str(),
+        ) && !transport.provider.enable_format_conversion
+        {
+            "format_conversion_disabled"
+        } else {
+            "transport_unsupported"
+        };
+        mark_skipped_local_standard_candidate(
+            state,
+            input,
+            trace_id,
+            &candidate,
+            candidate_index,
+            &candidate_id,
+            skip_reason,
+        )
+        .await;
+        return None;
+    }
+
     if !crate::ai_pipeline::conversion::request_conversion_transport_supported(
         &transport,
         conversion_kind,
@@ -301,6 +328,8 @@ pub(super) async fn maybe_build_local_standard_decision_payload_for_candidate(
             json!({
                 "user_id": input.auth_context.user_id,
                 "api_key_id": input.auth_context.api_key_id,
+                "username": input.auth_context.username,
+                "api_key_name": input.auth_context.api_key_name,
                 "request_id": trace_id,
                 "candidate_id": candidate_id,
                 "candidate_index": candidate_index,
@@ -349,6 +378,7 @@ pub(super) async fn mark_skipped_local_standard_candidate(
             candidate,
             candidate_index,
             candidate_id,
+            input.required_capabilities.as_ref(),
             skip_reason,
             current_unix_secs(),
             "gateway local standard decision failed to persist skipped candidate",

apps/aether-gateway/src/ai_pipeline/planner/standard/normalize/chat.rs

@@ -112,11 +112,6 @@ pub(crate) fn build_cross_format_openai_chat_upstream_url(
                 parts.uri.query(),
                 false,
             )),
-            RequestConversionKind::ToOpenAICompact => Some(build_openai_cli_url(
-                &transport.endpoint.base_url,
-                parts.uri.query(),
-                true,
-            )),
             _ => None,
         },
     }

apps/aether-gateway/src/ai_pipeline/planner/standard/normalize/cli.rs

@@ -10,8 +10,8 @@ use crate::ai_pipeline::transport::antigravity::{
 };
 use crate::ai_pipeline::transport::apply_local_body_rules;
 use crate::ai_pipeline::transport::url::{
-    build_claude_messages_url, build_gemini_content_url, build_openai_cli_url,
-    build_passthrough_path_url,
+    build_claude_messages_url, build_gemini_content_url, build_openai_chat_url,
+    build_openai_cli_url, build_passthrough_path_url,
 };
 use crate::ai_pipeline::{
     build_cross_format_openai_cli_request_body as pipeline_build_cross_format_openai_cli_request_body,
@@ -140,16 +140,15 @@ pub(crate) fn build_cross_format_openai_cli_upstream_url(
             build_passthrough_path_url(&transport.endpoint.base_url, path, parts.uri.query(), &[])
         }
         None => match conversion_kind {
+            RequestConversionKind::ToOpenAIChat => Some(build_openai_chat_url(
+                &transport.endpoint.base_url,
+                parts.uri.query(),
+            )),
             RequestConversionKind::ToOpenAIFamilyCli => Some(build_openai_cli_url(
                 &transport.endpoint.base_url,
                 parts.uri.query(),
                 false,
             )),
-            RequestConversionKind::ToOpenAICompact => Some(build_openai_cli_url(
-                &transport.endpoint.base_url,
-                parts.uri.query(),
-                true,
-            )),
             RequestConversionKind::ToClaudeStandard => Some(build_claude_messages_url(
                 &transport.endpoint.base_url,
                 parts.uri.query(),
@@ -160,7 +159,6 @@ pub(crate) fn build_cross_format_openai_cli_upstream_url(
                 upstream_is_stream,
                 parts.uri.query(),
             ),
-            _ => None,
         },
     }
 }

apps/aether-gateway/src/ai_pipeline/planner/standard/normalize/tests.rs

@@ -3,7 +3,7 @@ use serde_json::json;
 use super::build_cross_format_openai_cli_request_body;
 
 #[test]
-fn builds_openai_family_cross_format_request_body_from_compact_source() {
+fn builds_openai_chat_cross_format_request_body_from_openai_cli_source() {
     let body_json = json!({
         "model": "gpt-5",
         "input": "hello",
@@ -12,18 +12,18 @@ fn builds_openai_family_cross_format_request_body_from_compact_source() {
     let provider_request_body = build_cross_format_openai_cli_request_body(
         &body_json,
         "gpt-5-upstream",
-        "openai:compact",
         "openai:cli",
+        "openai:chat",
         false,
         "openai",
         None,
         None,
     )
-    .expect("compact to openai cli body should build");
+    .expect("openai cli to openai chat body should build");
 
     assert_eq!(provider_request_body["model"], "gpt-5-upstream");
-    assert_eq!(provider_request_body["input"][0]["type"], "message");
-    assert_eq!(provider_request_body["input"][0]["role"], "user");
+    assert_eq!(provider_request_body["messages"][0]["role"], "user");
+    assert_eq!(provider_request_body["messages"][0]["content"], "hello");
 }
 
 #[test]

apps/aether-gateway/src/ai_pipeline/planner/standard/openai/chat/decision.rs

@@ -96,7 +96,7 @@ pub(super) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
             )
             .await
         }
-        "claude:chat" | "gemini:chat" | "openai:cli" | "openai:compact" => {
+        "claude:chat" | "claude:cli" | "gemini:chat" | "gemini:cli" | "openai:cli" => {
             build_cross_format_local_openai_chat_decision_payload_for_candidate(
                 state,
                 parts,

apps/aether-gateway/src/ai_pipeline/planner/standard/openai/chat/decision/cross_format.rs

@@ -6,7 +6,9 @@ use tracing::warn;
 
 use crate::ai_pipeline::collect_control_headers;
 use crate::ai_pipeline::conversion::{
-    request_conversion_direct_auth, request_conversion_kind, request_conversion_transport_supported,
+    request_conversion_direct_auth, request_conversion_kind,
+    request_conversion_requires_enable_flag, request_conversion_transport_supported,
+    request_pair_allowed_for_transport,
 };
 use crate::ai_pipeline::planner::common::OPENAI_CHAT_STREAM_PLAN_KIND;
 use crate::ai_pipeline::planner::standard::{
@@ -50,6 +52,28 @@ pub(super) async fn build_cross_format_local_openai_chat_decision_payload_for_ca
     else {
         return None;
     };
+    if !request_pair_allowed_for_transport(&transport, "openai:chat", provider_api_format.as_str())
+    {
+        let skip_reason =
+            if request_conversion_requires_enable_flag("openai:chat", provider_api_format.as_str())
+                && !transport.provider.enable_format_conversion
+            {
+                "format_conversion_disabled"
+            } else {
+                "transport_unsupported"
+            };
+        mark_skipped_local_openai_chat_candidate(
+            state,
+            input,
+            trace_id,
+            candidate,
+            candidate_index,
+            candidate_id,
+            skip_reason,
+        )
+        .await;
+        return None;
+    }
     if !request_conversion_transport_supported(transport, conversion_kind) {
         mark_skipped_local_openai_chat_candidate(
             state,
@@ -63,7 +87,6 @@ pub(super) async fn build_cross_format_local_openai_chat_decision_payload_for_ca
         .await;
         return None;
     }
-
     let resolve_auth = request_conversion_direct_auth(transport, conversion_kind);
     let oauth_auth = if resolve_auth.is_none() {
         match planner_state
@@ -258,6 +281,8 @@ pub(super) async fn build_cross_format_local_openai_chat_decision_payload_for_ca
             json!({
                 "user_id": input.auth_context.user_id,
                 "api_key_id": input.auth_context.api_key_id,
+                "username": input.auth_context.username,
+                "api_key_name": input.auth_context.api_key_name,
                 "request_id": trace_id,
                 "candidate_id": candidate_id,
                 "candidate_index": candidate_index,

apps/aether-gateway/src/ai_pipeline/planner/standard/openai/chat/decision/same_format.rs

@@ -233,6 +233,8 @@ pub(super) async fn build_same_format_local_openai_chat_decision_payload_for_can
             json!({
                 "user_id": input.auth_context.user_id,
                 "api_key_id": input.auth_context.api_key_id,
+                "username": input.auth_context.username,
+                "api_key_name": input.auth_context.api_key_name,
                 "request_id": trace_id,
                 "candidate_id": candidate_id,
                 "candidate_index": candidate_index,

apps/aether-gateway/src/ai_pipeline/planner/standard/openai/chat/decision/support.rs

@@ -3,7 +3,13 @@ use serde_json::json;
 use uuid::Uuid;
 
 use crate::ai_pipeline::contracts::ExecutionRuntimeAuthContext;
-use crate::ai_pipeline::planner::candidate_affinity::prefer_local_tunnel_owner_candidates;
+use crate::ai_pipeline::conversion::{
+    request_conversion_kind, request_conversion_requires_enable_flag,
+    request_pair_allowed_for_transport,
+};
+use crate::ai_pipeline::planner::candidate_affinity::{
+    rank_local_execution_candidates, remember_scheduler_affinity_for_candidate,
+};
 use crate::ai_pipeline::GatewayAuthApiKeySnapshot;
 use crate::ai_pipeline::{ConversionMode, ExecutionStrategy, PlannerAppState};
 use crate::clock::current_unix_secs;
@@ -14,6 +20,7 @@ pub(crate) struct LocalOpenAiChatDecisionInput {
     pub(crate) auth_context: ExecutionRuntimeAuthContext,
     pub(crate) requested_model: String,
     pub(crate) auth_snapshot: GatewayAuthApiKeySnapshot,
+    pub(crate) required_capabilities: Option<serde_json::Value>,
 }
 
 #[derive(Debug, Clone)]
@@ -48,6 +55,7 @@ pub(crate) async fn mark_skipped_local_openai_chat_candidate(
             candidate,
             candidate_index,
             candidate_id,
+            input.required_capabilities.as_ref(),
             skip_reason,
             current_unix_secs(),
             "gateway local openai chat decision failed to persist skipped candidate",
@@ -62,13 +70,71 @@ pub(crate) async fn materialize_local_openai_chat_candidate_attempts(
     candidates: Vec<SchedulerMinimalCandidateSelectionCandidate>,
 ) -> Vec<LocalOpenAiChatCandidateAttempt> {
     let planner_state = PlannerAppState::new(state);
-    let candidates = prefer_local_tunnel_owner_candidates(planner_state, candidates).await;
-    let created_at_unix_secs = current_unix_secs();
+    let candidates = rank_local_execution_candidates(
+        planner_state,
+        candidates,
+        "openai:chat",
+        input.required_capabilities.as_ref(),
+    )
+    .await;
+    let created_at_unix_ms = current_unix_secs();
     let mut attempts = Vec::with_capacity(candidates.len());
+    let mut affinity_remembered = false;
 
     for (candidate_index, candidate) in candidates.into_iter().enumerate() {
         let generated_candidate_id = Uuid::new_v4().to_string();
         let provider_api_format = candidate.endpoint_api_format.trim().to_ascii_lowercase();
+        if provider_api_format != "openai:chat" {
+            if let Ok(Some(transport)) = planner_state
+                .read_provider_transport_snapshot(
+                    &candidate.provider_id,
+                    &candidate.endpoint_id,
+                    &candidate.key_id,
+                )
+                .await
+            {
+                if !request_pair_allowed_for_transport(
+                    &transport,
+                    "openai:chat",
+                    provider_api_format.as_str(),
+                ) {
+                    let skip_reason =
+                        if request_conversion_kind("openai:chat", provider_api_format.as_str())
+                            .is_some()
+                            && request_conversion_requires_enable_flag(
+                                "openai:chat",
+                                provider_api_format.as_str(),
+                            )
+                            && !transport.provider.enable_format_conversion
+                        {
+                            "format_conversion_disabled"
+                        } else {
+                            "transport_unsupported"
+                        };
+                    mark_skipped_local_openai_chat_candidate(
+                        state,
+                        input,
+                        trace_id,
+                        &candidate,
+                        candidate_index as u32,
+                        &generated_candidate_id,
+                        skip_reason,
+                    )
+                    .await;
+                    continue;
+                }
+            }
+        }
+        if !affinity_remembered {
+            remember_scheduler_affinity_for_candidate(
+                planner_state,
+                Some(&input.auth_snapshot),
+                "openai:chat",
+                &input.requested_model,
+                &candidate,
+            );
+            affinity_remembered = true;
+        }
         let (execution_strategy, conversion_mode) = if provider_api_format == "openai:chat" {
             (ExecutionStrategy::LocalSameFormat, ConversionMode::None)
         } else {
@@ -103,8 +169,9 @@ pub(crate) async fn materialize_local_openai_chat_candidate_attempts(
                 &candidate,
                 candidate_index as u32,
                 &generated_candidate_id,
+                input.required_capabilities.as_ref(),
                 Some(extra_data),
-                created_at_unix_secs,
+                created_at_unix_ms,
                 "gateway local openai chat decision request candidate upsert failed",
             )
             .await;

apps/aether-gateway/src/ai_pipeline/planner/standard/openai/chat/plans/candidates.rs

@@ -3,6 +3,7 @@ use std::collections::BTreeSet;
 use aether_scheduler_core::SchedulerMinimalCandidateSelectionCandidate;
 
 use super::super::{GatewayError, LocalOpenAiChatDecisionInput};
+use crate::ai_pipeline::conversion::request_candidate_api_formats;
 use crate::ai_pipeline::{GatewayAuthApiKeySnapshot, PlannerAppState};
 use crate::clock::current_unix_secs;
 use crate::AppState;
@@ -17,17 +18,7 @@ pub(crate) async fn list_local_openai_chat_candidates(
     let mut combined = Vec::new();
     let mut seen = BTreeSet::new();
 
-    let api_formats = if require_streaming {
-        vec!["openai:chat", "claude:chat", "gemini:chat", "openai:cli"]
-    } else {
-        vec![
-            "openai:chat",
-            "claude:chat",
-            "gemini:chat",
-            "openai:cli",
-            "openai:compact",
-        ]
-    };
+    let api_formats = request_candidate_api_formats("openai:chat", require_streaming);
 
     for api_format in api_formats {
         let auth_snapshot = if api_format == "openai:chat" {
@@ -40,6 +31,7 @@ pub(crate) async fn list_local_openai_chat_candidates(
                 api_format,
                 &input.requested_model,
                 require_streaming,
+                input.required_capabilities.as_ref(),
                 auth_snapshot,
                 now_unix_secs,
             )

apps/aether-gateway/src/ai_pipeline/planner/standard/openai/chat/plans/resolve.rs

@@ -108,9 +108,19 @@ pub(crate) async fn resolve_local_openai_chat_decision_input(
         }
     };
 
+    let required_capabilities = planner_state
+        .resolve_request_candidate_required_capabilities(
+            &auth_context.user_id,
+            &auth_context.api_key_id,
+            Some(requested_model.as_str()),
+            None,
+        )
+        .await;
+
     Some(LocalOpenAiChatDecisionInput {
         auth_context,
         requested_model,
         auth_snapshot,
+        required_capabilities,
     })
 }

apps/aether-gateway/src/ai_pipeline/planner/standard/openai/cli/decision/payload.rs

@@ -95,6 +95,8 @@ pub(crate) async fn maybe_build_local_openai_cli_decision_payload_for_candidate(
             json!({
                 "user_id": input.auth_context.user_id,
                 "api_key_id": input.auth_context.api_key_id,
+                "username": input.auth_context.username,
+                "api_key_name": input.auth_context.api_key_name,
                 "request_id": trace_id,
                 "candidate_id": candidate_id,
                 "candidate_index": candidate_index,

apps/aether-gateway/src/ai_pipeline/planner/standard/openai/cli/decision/request.rs

@@ -5,7 +5,9 @@ use serde_json::Value;
 use tracing::warn;
 
 use crate::ai_pipeline::conversion::{
-    request_conversion_direct_auth, request_conversion_kind, request_conversion_transport_supported,
+    request_conversion_direct_auth, request_conversion_kind,
+    request_conversion_requires_enable_flag, request_conversion_transport_supported,
+    request_pair_allowed_for_transport,
 };
 use crate::ai_pipeline::planner::common::force_upstream_streaming_for_provider;
 use crate::ai_pipeline::planner::standard::{
@@ -115,6 +117,36 @@ pub(crate) async fn resolve_local_openai_cli_candidate_payload_parts(
 
     let same_format = provider_api_format == spec.api_format.trim().to_ascii_lowercase();
     let conversion_kind = request_conversion_kind(spec.api_format, provider_api_format.as_str());
+    if !same_format
+        && !request_pair_allowed_for_transport(
+            &transport,
+            spec.api_format,
+            provider_api_format.as_str(),
+        )
+    {
+        let skip_reason = if conversion_kind.is_some()
+            && request_conversion_requires_enable_flag(
+                spec.api_format,
+                provider_api_format.as_str(),
+            )
+            && !transport.provider.enable_format_conversion
+        {
+            "format_conversion_disabled"
+        } else {
+            "transport_unsupported"
+        };
+        mark_skipped_local_openai_cli_candidate(
+            state,
+            input,
+            trace_id,
+            candidate,
+            candidate_index,
+            candidate_id,
+            skip_reason,
+        )
+        .await;
+        return None;
+    }
     let transport_supported = if same_format {
         supports_local_standard_transport_with_network(&transport, provider_api_format.as_str())
     } else {

apps/aether-gateway/src/ai_pipeline/planner/standard/openai/cli/decision/support.rs

@@ -6,8 +6,13 @@ use tracing::warn;
 use uuid::Uuid;
 
 use crate::ai_pipeline::contracts::ExecutionRuntimeAuthContext;
-use crate::ai_pipeline::conversion::request_conversion_kind;
-use crate::ai_pipeline::planner::candidate_affinity::prefer_local_tunnel_owner_candidates;
+use crate::ai_pipeline::conversion::{
+    request_candidate_api_formats, request_conversion_kind,
+    request_conversion_requires_enable_flag, request_pair_allowed_for_transport,
+};
+use crate::ai_pipeline::planner::candidate_affinity::{
+    rank_local_execution_candidates, remember_scheduler_affinity_for_candidate,
+};
 use crate::ai_pipeline::{
     resolve_local_decision_execution_runtime_auth_context, ConversionMode, ExecutionStrategy,
     GatewayControlDecision,
@@ -23,6 +28,7 @@ pub(crate) struct LocalOpenAiCliDecisionInput {
     pub(crate) auth_context: ExecutionRuntimeAuthContext,
     pub(crate) requested_model: String,
     pub(crate) auth_snapshot: GatewayAuthApiKeySnapshot,
+    pub(crate) required_capabilities: Option<serde_json::Value>,
 }
 
 #[derive(Debug, Clone)]
@@ -70,10 +76,20 @@ pub(crate) async fn resolve_local_openai_cli_decision_input(
         }
     };
 
+    let required_capabilities = planner_state
+        .resolve_request_candidate_required_capabilities(
+            &auth_context.user_id,
+            &auth_context.api_key_id,
+            Some(requested_model.as_str()),
+            None,
+        )
+        .await;
+
     Some(LocalOpenAiCliDecisionInput {
         auth_context,
         requested_model,
         auth_snapshot,
+        required_capabilities,
     })
 }
 
@@ -86,8 +102,10 @@ pub(crate) async fn materialize_local_openai_cli_candidate_attempts(
     let planner_state = PlannerAppState::new(state);
     let mut seen_candidates = BTreeSet::new();
     let mut candidates = Vec::new();
-    for candidate_api_format in candidate_api_formats_for_spec(spec) {
-        let auth_snapshot = if *candidate_api_format == spec.api_format {
+    for candidate_api_format in
+        request_candidate_api_formats(spec.api_format, spec.require_streaming)
+    {
+        let auth_snapshot = if candidate_api_format == spec.api_format {
             Some(&input.auth_snapshot)
         } else {
             None
@@ -97,6 +115,7 @@ pub(crate) async fn materialize_local_openai_cli_candidate_attempts(
                 candidate_api_format,
                 &input.requested_model,
                 spec.require_streaming,
+                input.required_capabilities.as_ref(),
                 auth_snapshot,
                 current_unix_secs(),
             )
@@ -125,13 +144,71 @@ pub(crate) async fn materialize_local_openai_cli_candidate_attempts(
             }
         }
     }
-    let candidates = prefer_local_tunnel_owner_candidates(planner_state, candidates).await;
+    let candidates = rank_local_execution_candidates(
+        planner_state,
+        candidates,
+        spec.api_format,
+        input.required_capabilities.as_ref(),
+    )
+    .await;
 
-    let created_at_unix_secs = current_unix_secs();
+    let created_at_unix_ms = current_unix_secs();
     let mut attempts = Vec::with_capacity(candidates.len());
+    let mut affinity_remembered = false;
     for (candidate_index, candidate) in candidates.into_iter().enumerate() {
         let generated_candidate_id = Uuid::new_v4().to_string();
         let provider_api_format = candidate.endpoint_api_format.trim().to_ascii_lowercase();
+        if provider_api_format != spec.api_format.trim().to_ascii_lowercase() {
+            if let Ok(Some(transport)) = planner_state
+                .read_provider_transport_snapshot(
+                    &candidate.provider_id,
+                    &candidate.endpoint_id,
+                    &candidate.key_id,
+                )
+                .await
+            {
+                if !request_pair_allowed_for_transport(
+                    &transport,
+                    spec.api_format,
+                    provider_api_format.as_str(),
+                ) {
+                    let skip_reason =
+                        if request_conversion_kind(spec.api_format, provider_api_format.as_str())
+                            .is_some()
+                            && request_conversion_requires_enable_flag(
+                                spec.api_format,
+                                provider_api_format.as_str(),
+                            )
+                            && !transport.provider.enable_format_conversion
+                        {
+                            "format_conversion_disabled"
+                        } else {
+                            "transport_unsupported"
+                        };
+                    mark_skipped_local_openai_cli_candidate(
+                        state,
+                        input,
+                        trace_id,
+                        &candidate,
+                        candidate_index as u32,
+                        &generated_candidate_id,
+                        skip_reason,
+                    )
+                    .await;
+                    continue;
+                }
+            }
+        }
+        if !affinity_remembered {
+            remember_scheduler_affinity_for_candidate(
+                planner_state,
+                Some(&input.auth_snapshot),
+                spec.api_format,
+                &input.requested_model,
+                &candidate,
+            );
+            affinity_remembered = true;
+        }
         let execution_strategy =
             if provider_api_format == spec.api_format.trim().to_ascii_lowercase() {
                 ExecutionStrategy::LocalSameFormat
@@ -170,8 +247,9 @@ pub(crate) async fn materialize_local_openai_cli_candidate_attempts(
                 &candidate,
                 candidate_index as u32,
                 &generated_candidate_id,
+                input.required_capabilities.as_ref(),
                 Some(extra_data),
-                created_at_unix_secs,
+                created_at_unix_ms,
                 "gateway local openai cli decision request candidate upsert failed",
             )
             .await;
@@ -217,14 +295,6 @@ fn auth_snapshot_allows_cross_format_openai_cli_candidate(
     true
 }
 
-fn candidate_api_formats_for_spec(spec: LocalOpenAiCliSpec) -> &'static [&'static str] {
-    match spec.api_format {
-        "openai:compact" => &["openai:compact", "openai:cli", "claude:cli", "gemini:cli"],
-        "openai:cli" => &["openai:cli", "claude:cli", "gemini:cli"],
-        _ => &[],
-    }
-}
-
 pub(crate) async fn mark_skipped_local_openai_cli_candidate(
     state: &AppState,
     input: &LocalOpenAiCliDecisionInput,
@@ -242,6 +312,7 @@ pub(crate) async fn mark_skipped_local_openai_cli_candidate(
             candidate,
             candidate_index,
             candidate_id,
+            input.required_capabilities.as_ref(),
             skip_reason,
             current_unix_secs(),
             "gateway local openai cli decision failed to persist skipped candidate",

apps/aether-gateway/src/ai_pipeline/planner/state/candidate_runtime.rs

@@ -4,6 +4,23 @@ use serde_json::Value;
 use super::PlannerAppState;
 
 impl<'a> PlannerAppState<'a> {
+    pub(crate) async fn resolve_request_candidate_required_capabilities(
+        self,
+        user_id: &str,
+        api_key_id: &str,
+        requested_model: Option<&str>,
+        explicit_required_capabilities: Option<&Value>,
+    ) -> Option<Value> {
+        crate::request_candidate_runtime::resolve_request_candidate_required_capabilities(
+            self.app(),
+            user_id,
+            api_key_id,
+            requested_model,
+            explicit_required_capabilities,
+        )
+        .await
+    }
+
     #[allow(clippy::too_many_arguments)]
     pub(crate) async fn persist_available_local_candidate(
         self,
@@ -13,8 +30,9 @@ impl<'a> PlannerAppState<'a> {
         candidate: &SchedulerMinimalCandidateSelectionCandidate,
         candidate_index: u32,
         candidate_id: &str,
+        required_capabilities: Option<&Value>,
         extra_data: Option<Value>,
-        created_at_unix_secs: u64,
+        created_at_unix_ms: u64,
         error_context: &'static str,
     ) -> String {
         crate::request_candidate_runtime::persist_available_local_candidate(
@@ -25,8 +43,9 @@ impl<'a> PlannerAppState<'a> {
             candidate,
             candidate_index,
             candidate_id,
+            required_capabilities,
             extra_data,
-            created_at_unix_secs,
+            created_at_unix_ms,
             error_context,
         )
         .await
@@ -41,8 +60,9 @@ impl<'a> PlannerAppState<'a> {
         candidate: &SchedulerMinimalCandidateSelectionCandidate,
         candidate_index: u32,
         candidate_id: &str,
+        required_capabilities: Option<&Value>,
         skip_reason: &str,
-        finished_at_unix_secs: u64,
+        finished_at_unix_ms: u64,
         error_context: &'static str,
     ) {
         crate::request_candidate_runtime::persist_skipped_local_candidate(
@@ -53,8 +73,9 @@ impl<'a> PlannerAppState<'a> {
             candidate,
             candidate_index,
             candidate_id,
+            required_capabilities,
             skip_reason,
-            finished_at_unix_secs,
+            finished_at_unix_ms,
             error_context,
         )
         .await

apps/aether-gateway/src/ai_pipeline/planner/state/scheduler.rs

@@ -9,6 +9,7 @@ impl<'a> PlannerAppState<'a> {
         api_format: &str,
         global_model_name: &str,
         require_streaming: bool,
+        required_capabilities: Option<&serde_json::Value>,
         auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
         now_unix_secs: u64,
     ) -> Result<Vec<SchedulerMinimalCandidateSelectionCandidate>, GatewayError> {
@@ -18,6 +19,7 @@ impl<'a> PlannerAppState<'a> {
             api_format,
             global_model_name,
             require_streaming,
+            required_capabilities,
             auth_snapshot,
             now_unix_secs,
         )

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

@@ -44,18 +44,20 @@ pub(crate) use aether_ai_pipeline::api::{
     provider_adaptation_descriptor_for_provider_type,
     provider_adaptation_requires_eventstream_accept,
     provider_adaptation_should_unwrap_stream_envelope,
-    provider_private_response_allows_sync_finalize, request_conversion_direct_auth,
-    request_conversion_kind, request_conversion_transport_supported, resolve_claude_stream_spec,
-    resolve_claude_sync_spec, resolve_execution_runtime_stream_plan_kind,
-    resolve_execution_runtime_sync_plan_kind, resolve_finalize_stream_rewrite_mode,
-    resolve_gemini_files_stream_spec, resolve_gemini_files_sync_spec, resolve_gemini_stream_spec,
-    resolve_gemini_sync_spec, resolve_local_same_format_stream_spec,
-    resolve_local_same_format_sync_spec, resolve_local_video_sync_spec,
-    resolve_openai_chat_max_tokens, resolve_openai_cli_stream_spec, resolve_openai_cli_sync_spec,
-    stream_body_contains_error_event, supports_stream_scheduler_decision_kind,
-    supports_sync_scheduler_decision_kind, sync_chat_response_conversion_kind,
-    sync_cli_response_conversion_kind, transform_provider_private_stream_line, value_as_u64,
-    CanonicalStreamFrame, ClaudeClientEmitter, ClaudeProviderState, ExecutionRuntimeAuthContext,
+    provider_private_response_allows_sync_finalize, request_candidate_api_formats,
+    request_conversion_direct_auth, request_conversion_kind,
+    request_conversion_requires_enable_flag, request_conversion_transport_supported,
+    request_pair_allowed_for_transport, resolve_claude_stream_spec, resolve_claude_sync_spec,
+    resolve_execution_runtime_stream_plan_kind, resolve_execution_runtime_sync_plan_kind,
+    resolve_finalize_stream_rewrite_mode, resolve_gemini_files_stream_spec,
+    resolve_gemini_files_sync_spec, resolve_gemini_stream_spec, resolve_gemini_sync_spec,
+    resolve_local_same_format_stream_spec, resolve_local_same_format_sync_spec,
+    resolve_local_video_sync_spec, resolve_openai_chat_max_tokens, resolve_openai_cli_stream_spec,
+    resolve_openai_cli_sync_spec, stream_body_contains_error_event,
+    supports_stream_scheduler_decision_kind, supports_sync_scheduler_decision_kind,
+    sync_chat_response_conversion_kind, sync_cli_response_conversion_kind,
+    transform_provider_private_stream_line, value_as_u64, CanonicalStreamFrame,
+    ClaudeClientEmitter, ClaudeProviderState, ExecutionRuntimeAuthContext,
     FinalizeStreamRewriteMode, GatewayControlPlanRequest, GatewayControlPlanResponse,
     GatewayControlSyncDecisionResponse, GeminiClientEmitter, GeminiProviderState,
     LocalCoreSyncErrorKind, LocalGeminiFilesSpec, LocalOpenAiCliSpec,

apps/aether-gateway/src/async_task/http/cancel.rs

@@ -253,7 +253,7 @@ async fn persist_cancelled_video_task(
             next_poll_at_unix_secs: None,
             poll_count: task.poll_count,
             max_poll_count: task.max_poll_count,
-            created_at_unix_secs: task.created_at_unix_secs,
+            created_at_unix_ms: task.created_at_unix_ms,
             submitted_at_unix_secs: task.submitted_at_unix_secs,
             completed_at_unix_secs: Some(now_unix_secs),
             updated_at_unix_secs: now_unix_secs,

apps/aether-gateway/src/async_task/runtime.rs

@@ -272,7 +272,7 @@ fn build_successful_poll_update(
     record.resolution = task.resolution.clone();
     record.aspect_ratio = task.aspect_ratio.clone();
     record.size = task.size.clone();
-    record.created_at_unix_secs = task.created_at_unix_secs;
+    record.created_at_unix_ms = task.created_at_unix_ms;
     record.submitted_at_unix_secs = task.submitted_at_unix_secs;
     record.updated_at_unix_secs = now_unix_secs;
     record.retry_count = task.retry_count;
@@ -379,7 +379,7 @@ fn stored_task_to_upsert(task: &StoredVideoTask) -> UpsertVideoTask {
         next_poll_at_unix_secs: task.next_poll_at_unix_secs,
         poll_count: task.poll_count,
         max_poll_count: task.max_poll_count.max(1),
-        created_at_unix_secs: task.created_at_unix_secs,
+        created_at_unix_ms: task.created_at_unix_ms,
         submitted_at_unix_secs: task.submitted_at_unix_secs,
         completed_at_unix_secs: task.completed_at_unix_secs,
         updated_at_unix_secs: task.updated_at_unix_secs,

apps/aether-gateway/src/audit/admin.rs

@@ -1,5 +1,5 @@
 use axum::body::Body;
-use axum::http::{self, Response};
+use axum::http::{self, Response, StatusCode};
 use tracing::{info, warn};
 
 use crate::control::GatewayControlDecision;
@@ -68,12 +68,8 @@ pub(crate) fn emit_admin_audit(
         )
     };
 
-    let audit_status = if response.status().is_success() {
-        "completed"
-    } else {
-        "failed"
-    };
-    if response.status().is_success() {
+    let (audit_status, log_level) = classify_admin_audit_response(method, response.status());
+    if log_level == AdminAuditLogLevel::Info {
         info!(
             event_name,
             log_type = "audit",
@@ -116,6 +112,25 @@ pub(crate) fn emit_admin_audit(
     }
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum AdminAuditLogLevel {
+    Info,
+    Warn,
+}
+
+fn classify_admin_audit_response(
+    method: &http::Method,
+    status: StatusCode,
+) -> (&'static str, AdminAuditLogLevel) {
+    if status.is_success() {
+        return ("completed", AdminAuditLogLevel::Info);
+    }
+    if status == StatusCode::NOT_FOUND && is_admin_read_method(method) {
+        return ("not_found", AdminAuditLogLevel::Info);
+    }
+    ("failed", AdminAuditLogLevel::Warn)
+}
+
 fn default_target_type(route_family: &str) -> &str {
     route_family
         .strip_suffix("_manage")
@@ -129,3 +144,29 @@ fn is_admin_mutation_method(method: &http::Method) -> bool {
         http::Method::POST | http::Method::PUT | http::Method::PATCH | http::Method::DELETE
     )
 }
+
+fn is_admin_read_method(method: &http::Method) -> bool {
+    matches!(*method, http::Method::GET | http::Method::HEAD)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{classify_admin_audit_response, AdminAuditLogLevel};
+    use axum::http::{Method, StatusCode};
+
+    #[test]
+    fn classifies_read_not_found_as_info_not_found() {
+        assert_eq!(
+            classify_admin_audit_response(&Method::GET, StatusCode::NOT_FOUND),
+            ("not_found", AdminAuditLogLevel::Info)
+        );
+    }
+
+    #[test]
+    fn classifies_mutation_not_found_as_warn_failed() {
+        assert_eq!(
+            classify_admin_audit_response(&Method::DELETE, StatusCode::NOT_FOUND),
+            ("failed", AdminAuditLogLevel::Warn)
+        );
+    }
+}

apps/aether-gateway/src/cache/mod.rs

@@ -6,4 +6,6 @@ mod scheduler_affinity;
 pub(crate) use auth_api_key_last_used::AuthApiKeyLastUsedCache;
 pub(crate) use auth_context::AuthContextCache;
 pub(crate) use direct_plan_bypass::DirectPlanBypassCache;
-pub(crate) use scheduler_affinity::{SchedulerAffinityCache, SchedulerAffinityTarget};
+pub(crate) use scheduler_affinity::{
+    SchedulerAffinityCache, SchedulerAffinitySnapshotEntry, SchedulerAffinityTarget,
+};

apps/aether-gateway/src/cache/scheduler_affinity.rs

@@ -1,6 +1,6 @@
 use std::time::Duration;
 
-use aether_cache::ExpiringMap;
+use aether_cache::{ExpiringMap, ExpiringMapFreshEntry};
 pub(crate) use aether_scheduler_core::SchedulerAffinityTarget;
 
 #[derive(Debug, Default)]
@@ -8,6 +8,13 @@ pub(crate) struct SchedulerAffinityCache {
     entries: ExpiringMap<String, SchedulerAffinityTarget>,
 }
 
+#[derive(Debug, Clone)]
+pub(crate) struct SchedulerAffinitySnapshotEntry {
+    pub(crate) cache_key: String,
+    pub(crate) target: SchedulerAffinityTarget,
+    pub(crate) age: Duration,
+}
+
 impl SchedulerAffinityCache {
     pub(crate) fn get_fresh(
         &self,
@@ -31,4 +38,18 @@ impl SchedulerAffinityCache {
     pub(crate) fn remove(&self, cache_key: &str) -> Option<SchedulerAffinityTarget> {
         self.entries.remove(&cache_key.to_string())
     }
+
+    pub(crate) fn fresh_entries(&self, ttl: Duration) -> Vec<SchedulerAffinitySnapshotEntry> {
+        self.entries
+            .snapshot_fresh(ttl)
+            .into_iter()
+            .map(
+                |ExpiringMapFreshEntry { key, value, age }| SchedulerAffinitySnapshotEntry {
+                    cache_key: key,
+                    target: value,
+                    age,
+                },
+            )
+            .collect()
+    }
 }

apps/aether-gateway/src/clock.rs

@@ -6,3 +6,10 @@ pub(crate) fn current_unix_secs() -> u64 {
         .unwrap_or_default()
         .as_secs()
 }
+
+pub(crate) fn current_unix_ms() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .unwrap_or_default()
+        .as_millis() as u64
+}

apps/aether-gateway/src/control/auth/resolution.rs

@@ -27,6 +27,10 @@ const AUTH_CONTEXT_CACHE_MAX_ENTRIES: usize = 256;
 pub(crate) struct GatewayControlAuthContext {
     pub(crate) user_id: String,
     pub(crate) api_key_id: String,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(crate) username: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(crate) api_key_name: Option<String>,
     pub(crate) balance_remaining: Option<f64>,
     pub(crate) access_allowed: bool,
     #[serde(skip)]
@@ -489,6 +493,8 @@ pub(super) async fn resolve_data_backed_auth_context(
                 return Ok(Some(GatewayControlAuthContext {
                     user_id: String::new(),
                     api_key_id: String::new(),
+                    username: None,
+                    api_key_name: None,
                     balance_remaining: None,
                     access_allowed: false,
                     user_rate_limit: None,
@@ -539,6 +545,8 @@ async fn resolve_trusted_auth_context(
         return Ok(Some(GatewayControlAuthContext {
             user_id: trusted_headers.user_id,
             api_key_id: trusted_headers.api_key_id,
+            username: None,
+            api_key_name: None,
             balance_remaining: trusted_headers.balance_remaining,
             access_allowed: false,
             user_rate_limit: None,
@@ -620,6 +628,8 @@ fn build_data_backed_auth_context(
     };
 
     GatewayControlAuthContext {
+        username: Some(snapshot.username.clone()),
+        api_key_name: snapshot.api_key_name.clone(),
         user_id: snapshot.user_id,
         api_key_id: snapshot.api_key_id,
         balance_remaining: wallet_remaining.or(balance_remaining),

apps/aether-gateway/src/data/candidate_selection.rs

@@ -4,9 +4,10 @@ use aether_scheduler_core::{
     auth_constraints_allow_api_format, build_minimal_candidate_selection,
     collect_global_model_names_for_required_capability, normalize_api_format,
     resolve_requested_global_model_name, SchedulerAuthConstraints,
-    SchedulerMinimalCandidateSelectionCandidate,
+    SchedulerMinimalCandidateSelectionCandidate, SchedulerPriorityMode,
 };
 use async_trait::async_trait;
+use std::collections::BTreeSet;
 
 use super::auth::GatewayAuthApiKeySnapshot;
 
@@ -62,6 +63,91 @@ pub(crate) async fn read_minimal_candidate_selection(
     requested_model_name: &str,
     require_streaming: bool,
     auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
+) -> Result<Vec<SchedulerMinimalCandidateSelectionCandidate>, DataLayerError> {
+    read_minimal_candidate_selection_with_priority_mode(
+        state,
+        api_format,
+        requested_model_name,
+        require_streaming,
+        auth_snapshot,
+        SchedulerPriorityMode::Provider,
+    )
+    .await
+}
+
+pub(crate) async fn read_minimal_candidate_selection_with_priority_mode(
+    state: &(impl MinimalCandidateSelectionRowSource + Sync),
+    api_format: &str,
+    requested_model_name: &str,
+    require_streaming: bool,
+    auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
+    priority_mode: SchedulerPriorityMode,
+) -> Result<Vec<SchedulerMinimalCandidateSelectionCandidate>, DataLayerError> {
+    read_minimal_candidate_selection_with_priority_mode_and_affinity_key(
+        state,
+        api_format,
+        requested_model_name,
+        require_streaming,
+        auth_snapshot,
+        priority_mode,
+        auth_snapshot_affinity_key(auth_snapshot),
+    )
+    .await
+}
+
+pub(crate) async fn read_minimal_candidate_selection_with_priority_mode_and_required_capabilities(
+    state: &(impl MinimalCandidateSelectionRowSource + Sync),
+    api_format: &str,
+    requested_model_name: &str,
+    require_streaming: bool,
+    auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
+    priority_mode: SchedulerPriorityMode,
+    required_capabilities: Option<&serde_json::Value>,
+) -> Result<Vec<SchedulerMinimalCandidateSelectionCandidate>, DataLayerError> {
+    read_minimal_candidate_selection_with_priority_mode_and_affinity_key_and_required_capabilities(
+        state,
+        api_format,
+        requested_model_name,
+        require_streaming,
+        auth_snapshot,
+        priority_mode,
+        auth_snapshot_affinity_key(auth_snapshot),
+        required_capabilities,
+    )
+    .await
+}
+
+pub(crate) async fn read_minimal_candidate_selection_with_priority_mode_and_affinity_key(
+    state: &(impl MinimalCandidateSelectionRowSource + Sync),
+    api_format: &str,
+    requested_model_name: &str,
+    require_streaming: bool,
+    auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
+    priority_mode: SchedulerPriorityMode,
+    affinity_key: Option<&str>,
+) -> Result<Vec<SchedulerMinimalCandidateSelectionCandidate>, DataLayerError> {
+    read_minimal_candidate_selection_with_priority_mode_and_affinity_key_and_required_capabilities(
+        state,
+        api_format,
+        requested_model_name,
+        require_streaming,
+        auth_snapshot,
+        priority_mode,
+        affinity_key,
+        None,
+    )
+    .await
+}
+
+pub(crate) async fn read_minimal_candidate_selection_with_priority_mode_and_affinity_key_and_required_capabilities(
+    state: &(impl MinimalCandidateSelectionRowSource + Sync),
+    api_format: &str,
+    requested_model_name: &str,
+    require_streaming: bool,
+    auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
+    priority_mode: SchedulerPriorityMode,
+    affinity_key: Option<&str>,
+    required_capabilities: Option<&serde_json::Value>,
 ) -> Result<Vec<SchedulerMinimalCandidateSelectionCandidate>, DataLayerError> {
     let normalized_api_format = normalize_api_format(api_format);
     if normalized_api_format.is_empty() {
@@ -81,17 +167,16 @@ pub(crate) async fn read_minimal_candidate_selection(
         return Ok(Vec::new());
     };
     let auth_constraints = auth_snapshot.map(auth_snapshot_constraints);
-    let affinity_key = auth_snapshot
-        .map(|snapshot| snapshot.api_key_id.trim())
-        .filter(|value| !value.is_empty());
     build_minimal_candidate_selection(
         rows,
         &normalized_api_format,
         requested_model_name,
         resolved_global_model_name.as_str(),
         require_streaming,
+        required_capabilities,
         auth_constraints.as_ref(),
         affinity_key,
+        priority_mode,
     )
 }
 
@@ -128,6 +213,60 @@ pub(crate) async fn read_global_model_names_for_required_capability(
     ))
 }
 
+pub(crate) async fn read_global_model_names_for_api_format(
+    state: &(impl MinimalCandidateSelectionRowSource + Sync),
+    api_format: &str,
+    require_streaming: bool,
+    auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
+) -> Result<Vec<String>, DataLayerError> {
+    let normalized_api_format = normalize_api_format(api_format);
+    if normalized_api_format.is_empty() {
+        return Ok(Vec::new());
+    }
+
+    if !auth_constraints_allow_api_format(
+        auth_snapshot.map(auth_snapshot_constraints).as_ref(),
+        &normalized_api_format,
+    ) {
+        return Ok(Vec::new());
+    }
+
+    let rows = state
+        .read_minimal_candidate_selection_rows_for_api_format(&normalized_api_format)
+        .await?;
+    let auth_constraints = auth_snapshot.map(auth_snapshot_constraints);
+    let mut model_names = BTreeSet::new();
+
+    for row in rows {
+        if require_streaming && !row.supports_streaming() {
+            continue;
+        }
+        if !aether_scheduler_core::auth_constraints_allow_provider(
+            auth_constraints.as_ref(),
+            &row.provider_id,
+            &row.provider_name,
+        ) {
+            continue;
+        }
+        if !aether_scheduler_core::auth_constraints_allow_model(
+            auth_constraints.as_ref(),
+            &row.global_model_name,
+            &row.global_model_name,
+        ) {
+            continue;
+        }
+        model_names.insert(row.global_model_name);
+    }
+
+    Ok(model_names.into_iter().collect())
+}
+
+fn auth_snapshot_affinity_key(auth_snapshot: Option<&GatewayAuthApiKeySnapshot>) -> Option<&str> {
+    auth_snapshot
+        .map(|snapshot| snapshot.api_key_id.trim())
+        .filter(|value| !value.is_empty())
+}
+
 fn auth_snapshot_constraints(snapshot: &GatewayAuthApiKeySnapshot) -> SchedulerAuthConstraints {
     SchedulerAuthConstraints {
         allowed_providers: snapshot

apps/aether-gateway/src/data/candidates.rs

@@ -36,7 +36,7 @@ mod tests {
         request_id: &str,
         candidate_index: i32,
         status: RequestCandidateStatus,
-        started_at_unix_secs: Option<i64>,
+        started_at_unix_ms: Option<i64>,
         latency_ms: Option<i32>,
         status_code: Option<i32>,
     ) -> StoredRequestCandidate {
@@ -62,9 +62,9 @@ mod tests {
             Some(1),
             None,
             None,
-            100 + i64::from(candidate_index),
-            started_at_unix_secs,
-            started_at_unix_secs.map(|value| value + 1),
+            (100 + i64::from(candidate_index)) * 1_000,
+            started_at_unix_ms.map(|v| v * 1_000),
+            started_at_unix_ms.map(|value| (value + 1) * 1_000),
         )
         .expect("candidate should build")
     }

apps/aether-gateway/src/data/decision_trace.rs

@@ -101,9 +101,9 @@ mod tests {
             Some(1),
             None,
             Some(serde_json::json!({"cache_1h": true})),
-            100,
-            Some(101),
-            Some(102),
+            100_000,
+            Some(101_000),
+            Some(102_000),
         )
         .expect("candidate should build")
     }
@@ -174,11 +174,15 @@ mod tests {
                     provider_name: Some("OpenAI".to_string()),
                     provider_website: Some("https://openai.com".to_string()),
                     provider_type: Some("custom".to_string()),
+                    provider_priority: Some(0),
+                    provider_keep_priority_on_conversion: Some(false),
                     endpoint_api_format: Some("openai:chat".to_string()),
                     endpoint_api_family: Some("openai".to_string()),
                     endpoint_kind: Some("chat".to_string()),
                     provider_key_name: Some("prod-key".to_string()),
                     provider_key_auth_type: Some("api_key".to_string()),
+                    provider_key_internal_priority: Some(50),
+                    provider_key_global_priority_by_format: None,
                     provider_key_capabilities: Some(serde_json::json!({"cache_1h": true})),
                     provider_key_is_active: Some(true),
                 }],

apps/aether-gateway/src/data/tests.rs

@@ -28,7 +28,9 @@ use aether_data_contracts::repository::usage::StoredRequestUsageAudit;
 use aether_data_contracts::repository::video_tasks::{
     UpsertVideoTask, VideoTaskLookupKey, VideoTaskStatus, VideoTaskWriteRepository,
 };
-use aether_scheduler_core::{build_minimal_candidate_selection, SchedulerAuthConstraints};
+use aether_scheduler_core::{
+    build_minimal_candidate_selection, SchedulerAuthConstraints, SchedulerPriorityMode,
+};
 use async_trait::async_trait;
 use serde_json::json;
 
@@ -108,7 +110,7 @@ async fn data_state_find_uses_configured_read_repository() {
             next_poll_at_unix_secs: Some(100),
             poll_count: 0,
             max_poll_count: 360,
-            created_at_unix_secs: 100,
+            created_at_unix_ms: 100,
             submitted_at_unix_secs: Some(100),
             completed_at_unix_secs: None,
             updated_at_unix_secs: 100,
@@ -384,9 +386,9 @@ async fn data_state_reads_decision_trace_with_provider_catalog_metadata() {
             Some(1),
             None,
             Some(serde_json::json!({"cache_1h": true})),
-            100,
-            Some(101),
-            Some(102),
+            100_000,
+            Some(101_000),
+            Some(102_000),
         )
         .expect("candidate should build"),
     ]));
@@ -472,9 +474,9 @@ async fn data_state_reads_request_audit_bundle_from_multiple_readers() {
             Some(1),
             None,
             Some(serde_json::json!({"cache_1h": true})),
-            100,
-            Some(101),
-            Some(102),
+            100_000,
+            Some(101_000),
+            Some(102_000),
         )
         .expect("candidate should build"),
     ]));
@@ -668,8 +670,10 @@ async fn data_state_reads_minimal_candidate_selection_with_auth_filters() {
         "gpt-4.1",
         "gpt-4.1",
         false,
+        None,
         Some(&auth_constraints),
         Some(auth_snapshot.api_key_id.as_str()),
+        SchedulerPriorityMode::Provider,
     )
     .expect("selection should read");
 
@@ -722,7 +726,7 @@ async fn maps_openai_video_task_repository_row_into_read_response() {
             next_poll_at_unix_secs: Some(120),
             poll_count: 1,
             max_poll_count: 360,
-            created_at_unix_secs: 100,
+            created_at_unix_ms: 100,
             submitted_at_unix_secs: Some(100),
             completed_at_unix_secs: None,
             updated_at_unix_secs: 120,
@@ -781,7 +785,7 @@ async fn maps_gemini_video_task_repository_row_into_read_response() {
             next_poll_at_unix_secs: None,
             poll_count: 4,
             max_poll_count: 360,
-            created_at_unix_secs: 100,
+            created_at_unix_ms: 100,
             submitted_at_unix_secs: Some(100),
             completed_at_unix_secs: Some(120),
             updated_at_unix_secs: 120,
@@ -843,7 +847,7 @@ async fn data_state_write_uses_configured_shadow_result_writer() {
             match_status: ShadowResultMatchStatus::Pending,
             status_code: Some(200),
             error_message: None,
-            created_at_unix_secs: 100,
+            created_at_unix_ms: 100,
             updated_at_unix_secs: 100,
         })
         .await
@@ -906,7 +910,7 @@ async fn data_state_records_shadow_result_samples_and_merges_match_status() {
         .expect("second stored result should exist");
 
     assert_eq!(second.match_status, ShadowResultMatchStatus::Match);
-    assert_eq!(second.created_at_unix_secs, 100);
+    assert_eq!(second.created_at_unix_ms, 100);
     assert_eq!(second.updated_at_unix_secs, 200);
     assert_eq!(second.request_id.as_deref(), Some("req-1"));
 }
@@ -986,7 +990,7 @@ fn sample_request_candidate(
     request_id: &str,
     candidate_index: i32,
     status: RequestCandidateStatus,
-    started_at_unix_secs: Option<i64>,
+    started_at_unix_ms: Option<i64>,
     latency_ms: Option<i32>,
     status_code: Option<i32>,
 ) -> StoredRequestCandidate {
@@ -1013,8 +1017,8 @@ fn sample_request_candidate(
         None,
         None,
         100 + i64::from(candidate_index),
-        started_at_unix_secs,
-        started_at_unix_secs.map(|value| value + 1),
+        started_at_unix_ms,
+        started_at_unix_ms.map(|value| value + 1),
     )
     .expect("candidate should build")
 }

apps/aether-gateway/src/execution_runtime/fallback.rs

@@ -1,25 +1,83 @@
-use aether_contracts::ExecutionResult;
+use std::collections::BTreeSet;
 
-fn is_local_candidate_attempt(report_context: Option<&serde_json::Value>) -> bool {
+use aether_contracts::{ExecutionPlan, ExecutionResult};
+use regex::Regex;
+
+use crate::AppState;
+
+fn local_candidate_index(report_context: Option<&serde_json::Value>) -> Option<u64> {
     report_context
         .and_then(serde_json::Value::as_object)
         .and_then(|context| context.get("candidate_index"))
         .and_then(serde_json::Value::as_u64)
-        .is_some()
 }
 
 fn is_retryable_local_upstream_status(status_code: u16) -> bool {
     status_code == 429 || status_code >= 500
 }
 
-pub(crate) fn should_retry_next_local_candidate_sync(
-    plan_kind: &str,
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+struct LocalFailoverPolicy {
+    max_retries: Option<u64>,
+    stop_status_codes: BTreeSet<u16>,
+    continue_status_codes: BTreeSet<u16>,
+    success_failover_patterns: Vec<LocalFailoverRegexRule>,
+    error_stop_patterns: Vec<LocalFailoverRegexRule>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct LocalFailoverRegexRule {
+    pattern: String,
+    status_codes: BTreeSet<u16>,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum LocalFailoverDecision {
+    UseDefault,
+    RetryNextCandidate,
+    StopLocalFailover,
+}
+
+pub(crate) async fn should_retry_next_local_candidate_sync(
+    state: &AppState,
+    plan: &ExecutionPlan,
+    _plan_kind: &str,
     report_context: Option<&serde_json::Value>,
     result: &ExecutionResult,
+    response_text: Option<&str>,
 ) -> bool {
-    is_local_candidate_attempt(report_context)
-        && plan_kind == "openai_chat_sync"
-        && is_retryable_local_upstream_status(result.status_code)
+    matches!(
+        resolve_local_failover_decision(
+            state,
+            plan,
+            report_context,
+            result.status_code,
+            response_text,
+        )
+        .await,
+        LocalFailoverDecision::RetryNextCandidate
+    )
+}
+
+pub(crate) async fn should_stop_local_candidate_failover_sync(
+    state: &AppState,
+    plan: &ExecutionPlan,
+    _plan_kind: &str,
+    report_context: Option<&serde_json::Value>,
+    result: &ExecutionResult,
+    response_text: Option<&str>,
+) -> bool {
+    matches!(
+        resolve_local_failover_decision(
+            state,
+            plan,
+            report_context,
+            result.status_code,
+            response_text,
+        )
+        .await,
+        LocalFailoverDecision::StopLocalFailover
+    )
 }
 
 pub(crate) fn should_fallback_to_control_sync(
@@ -102,14 +160,245 @@ pub(crate) fn resolve_core_sync_error_finalize_report_kind(
     Some(report_kind.to_string())
 }
 
-pub(crate) fn should_retry_next_local_candidate_stream(
-    plan_kind: &str,
+pub(crate) async fn should_retry_next_local_candidate_stream(
+    state: &AppState,
+    plan: &ExecutionPlan,
+    _plan_kind: &str,
     report_context: Option<&serde_json::Value>,
     status_code: u16,
+    response_text: Option<&str>,
 ) -> bool {
-    is_local_candidate_attempt(report_context)
-        && plan_kind == "openai_chat_stream"
-        && is_retryable_local_upstream_status(status_code)
+    matches!(
+        resolve_local_failover_decision(state, plan, report_context, status_code, response_text)
+            .await,
+        LocalFailoverDecision::RetryNextCandidate
+    )
+}
+
+pub(crate) async fn should_stop_local_candidate_failover_stream(
+    state: &AppState,
+    plan: &ExecutionPlan,
+    _plan_kind: &str,
+    report_context: Option<&serde_json::Value>,
+    status_code: u16,
+    response_text: Option<&str>,
+) -> bool {
+    matches!(
+        resolve_local_failover_decision(state, plan, report_context, status_code, response_text)
+            .await,
+        LocalFailoverDecision::StopLocalFailover
+    )
+}
+
+pub(crate) fn local_failover_response_text(
+    body_json: Option<&serde_json::Value>,
+    body_bytes: &[u8],
+    fallback_text: Option<&str>,
+) -> Option<String> {
+    if let Some(body_json) = body_json {
+        return serde_json::to_string(body_json).ok();
+    }
+    if !body_bytes.is_empty() {
+        return Some(String::from_utf8_lossy(body_bytes).into_owned());
+    }
+    fallback_text
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+        .map(ToOwned::to_owned)
+}
+
+async fn resolve_local_failover_decision(
+    state: &AppState,
+    plan: &ExecutionPlan,
+    report_context: Option<&serde_json::Value>,
+    status_code: u16,
+    response_text: Option<&str>,
+) -> LocalFailoverDecision {
+    let Some(candidate_index) = local_candidate_index(report_context) else {
+        return LocalFailoverDecision::UseDefault;
+    };
+    let policy = resolve_local_failover_policy(state, plan).await;
+    let response_text = response_text
+        .map(str::trim)
+        .filter(|value| !value.is_empty());
+
+    if policy.stop_status_codes.contains(&status_code) {
+        return LocalFailoverDecision::StopLocalFailover;
+    }
+
+    if status_code >= 400
+        && response_text.is_some_and(|text| {
+            policy
+                .error_stop_patterns
+                .iter()
+                .any(|rule| local_failover_regex_rule_matches(rule, text, status_code))
+        })
+    {
+        return LocalFailoverDecision::StopLocalFailover;
+    }
+
+    if policy
+        .max_retries
+        .is_some_and(|max_retries| candidate_index >= max_retries)
+    {
+        return LocalFailoverDecision::UseDefault;
+    }
+
+    if status_code == 200
+        && response_text.is_some_and(|text| {
+            policy
+                .success_failover_patterns
+                .iter()
+                .any(|rule| local_failover_regex_rule_matches(rule, text, status_code))
+        })
+    {
+        return LocalFailoverDecision::RetryNextCandidate;
+    }
+
+    if policy.continue_status_codes.contains(&status_code) {
+        return LocalFailoverDecision::RetryNextCandidate;
+    }
+
+    if is_retryable_local_upstream_status(status_code) {
+        return LocalFailoverDecision::RetryNextCandidate;
+    }
+
+    LocalFailoverDecision::UseDefault
+}
+
+async fn resolve_local_failover_policy(
+    state: &AppState,
+    plan: &ExecutionPlan,
+) -> LocalFailoverPolicy {
+    let transport = match state
+        .read_provider_transport_snapshot(&plan.provider_id, &plan.endpoint_id, &plan.key_id)
+        .await
+    {
+        Ok(Some(transport)) => transport,
+        Ok(None) | Err(_) => return LocalFailoverPolicy::default(),
+    };
+
+    let rules = transport
+        .provider
+        .config
+        .as_ref()
+        .and_then(|config| config.get("failover_rules"))
+        .and_then(serde_json::Value::as_object);
+    let max_retries = rules
+        .and_then(|value| value.get("max_retries"))
+        .and_then(parse_u64_value)
+        .or_else(|| {
+            transport
+                .endpoint
+                .max_retries
+                .and_then(|value| u64::try_from(value).ok())
+        })
+        .or_else(|| {
+            transport
+                .provider
+                .max_retries
+                .and_then(|value| u64::try_from(value).ok())
+        });
+
+    LocalFailoverPolicy {
+        max_retries,
+        stop_status_codes: rules
+            .map(|value| {
+                parse_status_code_set(
+                    value,
+                    &[
+                        "stop_on_status_codes",
+                        "early_stop_status_codes",
+                        "non_retryable_status_codes",
+                        "stop_status_codes",
+                    ],
+                )
+            })
+            .unwrap_or_default(),
+        continue_status_codes: rules
+            .map(|value| {
+                parse_status_code_set(
+                    value,
+                    &[
+                        "continue_on_status_codes",
+                        "retryable_status_codes",
+                        "retry_on_status_codes",
+                        "continue_status_codes",
+                    ],
+                )
+            })
+            .unwrap_or_default(),
+        success_failover_patterns: rules
+            .map(|value| parse_regex_rules(value, "success_failover_patterns"))
+            .unwrap_or_default(),
+        error_stop_patterns: rules
+            .map(|value| parse_regex_rules(value, "error_stop_patterns"))
+            .unwrap_or_default(),
+    }
+}
+
+fn parse_regex_rules(
+    rules: &serde_json::Map<String, serde_json::Value>,
+    key: &str,
+) -> Vec<LocalFailoverRegexRule> {
+    rules
+        .get(key)
+        .and_then(serde_json::Value::as_array)
+        .into_iter()
+        .flat_map(|items| items.iter())
+        .filter_map(parse_regex_rule)
+        .collect()
+}
+
+fn parse_regex_rule(value: &serde_json::Value) -> Option<LocalFailoverRegexRule> {
+    let object = value.as_object()?;
+    let pattern = object
+        .get("pattern")
+        .and_then(serde_json::Value::as_str)
+        .map(str::trim)
+        .filter(|value| !value.is_empty())?;
+    Some(LocalFailoverRegexRule {
+        pattern: pattern.to_string(),
+        status_codes: object
+            .get("status_codes")
+            .and_then(serde_json::Value::as_array)
+            .into_iter()
+            .flat_map(|values| values.iter())
+            .filter_map(|value| parse_u64_value(value).and_then(|value| u16::try_from(value).ok()))
+            .collect(),
+    })
+}
+
+fn local_failover_regex_rule_matches(
+    rule: &LocalFailoverRegexRule,
+    response_text: &str,
+    status_code: u16,
+) -> bool {
+    if !rule.status_codes.is_empty() && !rule.status_codes.contains(&status_code) {
+        return false;
+    }
+
+    Regex::new(&rule.pattern)
+        .ok()
+        .is_some_and(|regex| regex.is_match(response_text))
+}
+
+fn parse_status_code_set(
+    rules: &serde_json::Map<String, serde_json::Value>,
+    keys: &[&str],
+) -> BTreeSet<u16> {
+    keys.iter()
+        .filter_map(|key| rules.get(*key))
+        .filter_map(serde_json::Value::as_array)
+        .flat_map(|values| values.iter())
+        .filter_map(|value| parse_u64_value(value).and_then(|value| u16::try_from(value).ok()))
+        .collect()
+}
+
+fn parse_u64_value(value: &serde_json::Value) -> Option<u64> {
+    value
+        .as_u64()
+        .or_else(|| value.as_i64().and_then(|value| u64::try_from(value).ok()))
 }
 
 pub(crate) fn should_fallback_to_control_stream(
@@ -172,14 +461,121 @@ pub(crate) fn resolve_core_stream_direct_finalize_report_kind(plan_kind: &str) -
 
 #[cfg(test)]
 mod tests {
+    use std::collections::BTreeSet;
+
     use aether_contracts::ExecutionResult;
+    use aether_data::repository::provider_catalog::InMemoryProviderCatalogReadRepository;
+    use aether_data_contracts::repository::provider_catalog::{
+        StoredProviderCatalogEndpoint, StoredProviderCatalogKey, StoredProviderCatalogProvider,
+    };
 
     use super::{
         resolve_core_stream_error_finalize_report_kind,
-        resolve_core_sync_error_finalize_report_kind, should_fallback_to_control_stream,
-        should_fallback_to_control_sync, should_retry_next_local_candidate_stream,
-        should_retry_next_local_candidate_sync,
+        resolve_core_sync_error_finalize_report_kind, resolve_local_failover_policy,
+        should_fallback_to_control_stream, should_fallback_to_control_sync,
+        should_retry_next_local_candidate_stream, should_retry_next_local_candidate_sync,
+        should_stop_local_candidate_failover_stream, should_stop_local_candidate_failover_sync,
+        LocalFailoverPolicy, LocalFailoverRegexRule,
     };
+    use crate::data::GatewayDataState;
+    use crate::AppState;
+
+    fn sample_plan() -> aether_contracts::ExecutionPlan {
+        aether_contracts::ExecutionPlan {
+            request_id: "req-1".to_string(),
+            candidate_id: Some("cand-1".to_string()),
+            provider_name: Some("provider-1".to_string()),
+            provider_id: "provider-1".to_string(),
+            endpoint_id: "endpoint-1".to_string(),
+            key_id: "key-1".to_string(),
+            method: "POST".to_string(),
+            url: "https://example.com/v1/chat/completions".to_string(),
+            headers: Default::default(),
+            content_type: Some("application/json".to_string()),
+            content_encoding: None,
+            body: aether_contracts::RequestBody::from_json(serde_json::json!({"model":"gpt-5"})),
+            stream: false,
+            client_api_format: "openai:chat".to_string(),
+            provider_api_format: "openai:chat".to_string(),
+            model_name: Some("gpt-5".to_string()),
+            proxy: None,
+            tls_profile: None,
+            timeouts: None,
+        }
+    }
+
+    fn sample_provider(config: Option<serde_json::Value>) -> StoredProviderCatalogProvider {
+        StoredProviderCatalogProvider::new(
+            "provider-1".to_string(),
+            "provider-1".to_string(),
+            Some("https://provider.example".to_string()),
+            "custom".to_string(),
+        )
+        .expect("provider should build")
+        .with_transport_fields(true, false, false, None, Some(3), None, None, None, config)
+    }
+
+    fn sample_endpoint() -> StoredProviderCatalogEndpoint {
+        StoredProviderCatalogEndpoint::new(
+            "endpoint-1".to_string(),
+            "provider-1".to_string(),
+            "openai:chat".to_string(),
+            Some("openai".to_string()),
+            Some("chat".to_string()),
+            true,
+        )
+        .expect("endpoint should build")
+        .with_transport_fields(
+            "https://api.provider.example".to_string(),
+            None,
+            None,
+            Some(2),
+            None,
+            None,
+            None,
+            None,
+        )
+        .expect("endpoint transport should build")
+    }
+
+    fn sample_key() -> StoredProviderCatalogKey {
+        StoredProviderCatalogKey::new(
+            "key-1".to_string(),
+            "provider-1".to_string(),
+            "key-1".to_string(),
+            "api_key".to_string(),
+            None,
+            true,
+        )
+        .expect("key should build")
+        .with_transport_fields(
+            Some(serde_json::json!(["openai:chat"])),
+            "plain-upstream-key".to_string(),
+            None,
+            None,
+            Some(serde_json::json!({"openai:chat": 1})),
+            None,
+            None,
+            None,
+            None,
+        )
+        .expect("key transport should build")
+    }
+
+    fn build_state_with_provider_config(config: Option<serde_json::Value>) -> AppState {
+        let provider_catalog = InMemoryProviderCatalogReadRepository::seed(
+            vec![sample_provider(config)],
+            vec![sample_endpoint()],
+            vec![sample_key()],
+        );
+        let data_state = GatewayDataState::with_provider_transport_reader_for_tests(
+            std::sync::Arc::new(provider_catalog),
+            "development-key",
+        );
+        AppState::new()
+            .expect("state should build")
+            .with_data_state_for_tests(data_state)
+    }
 
     #[test]
     fn sync_failover_marks_chat_errors() {
@@ -220,8 +616,8 @@ mod tests {
         );
     }
 
-    #[test]
-    fn sync_retry_next_candidate_is_local_openai_chat_only() {
+    #[tokio::test]
+    async fn sync_retry_next_candidate_requires_local_candidate_context() {
         let result = ExecutionResult {
             request_id: "req-1".to_string(),
             candidate_id: None,
@@ -235,26 +631,57 @@ mod tests {
             "candidate_index": 0,
             "retry_index": 0,
         });
+        let state = build_state_with_provider_config(None);
+        let plan = sample_plan();
 
-        assert!(should_retry_next_local_candidate_sync(
-            "openai_chat_sync",
-            Some(&local_report_context),
-            &result,
-        ));
-        assert!(!should_retry_next_local_candidate_sync(
-            "openai_chat_sync",
-            None,
-            &result,
-        ));
-        assert!(!should_retry_next_local_candidate_sync(
-            "claude_chat_sync",
-            None,
-            &result,
-        ));
+        assert!(
+            should_retry_next_local_candidate_sync(
+                &state,
+                &plan,
+                "openai_chat_sync",
+                Some(&local_report_context),
+                &result,
+                None,
+            )
+            .await
+        );
+        assert!(
+            should_retry_next_local_candidate_sync(
+                &state,
+                &plan,
+                "claude_cli_sync",
+                Some(&local_report_context),
+                &result,
+                None,
+            )
+            .await
+        );
+        assert!(
+            !should_retry_next_local_candidate_sync(
+                &state,
+                &plan,
+                "openai_chat_sync",
+                None,
+                &result,
+                None,
+            )
+            .await
+        );
+        assert!(
+            !should_retry_next_local_candidate_sync(
+                &state,
+                &plan,
+                "claude_chat_sync",
+                None,
+                &result,
+                None,
+            )
+            .await
+        );
     }
 
-    #[test]
-    fn sync_retry_next_candidate_treats_rate_limit_as_retryable() {
+    #[tokio::test]
+    async fn sync_retry_next_candidate_treats_rate_limit_as_retryable() {
         let result = ExecutionResult {
             request_id: "req-1".to_string(),
             candidate_id: None,
@@ -268,49 +695,304 @@ mod tests {
             "candidate_index": 0,
             "retry_index": 0,
         });
+        let state = build_state_with_provider_config(None);
+        let plan = sample_plan();
 
-        assert!(should_retry_next_local_candidate_sync(
-            "openai_chat_sync",
-            Some(&local_report_context),
-            &result,
-        ));
+        assert!(
+            should_retry_next_local_candidate_sync(
+                &state,
+                &plan,
+                "openai_chat_sync",
+                Some(&local_report_context),
+                &result,
+                None,
+            )
+            .await
+        );
     }
 
-    #[test]
-    fn stream_retry_next_candidate_is_local_openai_chat_only() {
+    #[tokio::test]
+    async fn stream_retry_next_candidate_requires_local_candidate_context() {
         let local_report_context = serde_json::json!({
             "candidate_index": 0,
             "retry_index": 0,
         });
+        let state = build_state_with_provider_config(None);
+        let plan = sample_plan();
 
-        assert!(should_retry_next_local_candidate_stream(
-            "openai_chat_stream",
-            Some(&local_report_context),
-            502,
-        ));
-        assert!(!should_retry_next_local_candidate_stream(
-            "openai_chat_stream",
-            None,
-            502,
-        ));
-        assert!(!should_retry_next_local_candidate_stream(
-            "claude_chat_stream",
-            Some(&local_report_context),
-            502,
-        ));
+        assert!(
+            should_retry_next_local_candidate_stream(
+                &state,
+                &plan,
+                "openai_chat_stream",
+                Some(&local_report_context),
+                502,
+                None,
+            )
+            .await
+        );
+        assert!(
+            should_retry_next_local_candidate_stream(
+                &state,
+                &plan,
+                "gemini_cli_stream",
+                Some(&local_report_context),
+                502,
+                None,
+            )
+            .await
+        );
+        assert!(
+            !should_retry_next_local_candidate_stream(
+                &state,
+                &plan,
+                "openai_chat_stream",
+                None,
+                502,
+                None,
+            )
+            .await
+        );
+        assert!(
+            !should_retry_next_local_candidate_stream(
+                &state,
+                &plan,
+                "claude_chat_stream",
+                None,
+                502,
+                None,
+            )
+            .await
+        );
     }
 
-    #[test]
-    fn stream_retry_next_candidate_treats_rate_limit_as_retryable() {
+    #[tokio::test]
+    async fn stream_retry_next_candidate_treats_rate_limit_as_retryable() {
         let local_report_context = serde_json::json!({
             "candidate_index": 0,
             "retry_index": 0,
         });
+        let state = build_state_with_provider_config(None);
+        let plan = sample_plan();
 
-        assert!(should_retry_next_local_candidate_stream(
-            "openai_chat_stream",
-            Some(&local_report_context),
-            429,
-        ));
+        assert!(
+            should_retry_next_local_candidate_stream(
+                &state,
+                &plan,
+                "openai_chat_stream",
+                Some(&local_report_context),
+                429,
+                None,
+            )
+            .await
+        );
+    }
+
+    #[test]
+    fn resolve_local_failover_policy_reads_provider_rules() {
+        let state = build_state_with_provider_config(Some(serde_json::json!({
+            "failover_rules": {
+                "max_retries": 1,
+                "stop_on_status_codes": [503],
+                "continue_on_status_codes": [409, 429]
+            }
+        })));
+        let plan = sample_plan();
+        let runtime = tokio::runtime::Runtime::new().expect("runtime should build");
+
+        let policy = runtime.block_on(resolve_local_failover_policy(&state, &plan));
+        assert_eq!(
+            policy,
+            LocalFailoverPolicy {
+                max_retries: Some(1),
+                stop_status_codes: [503].into_iter().collect(),
+                continue_status_codes: [409, 429].into_iter().collect(),
+                success_failover_patterns: Vec::new(),
+                error_stop_patterns: Vec::new(),
+            }
+        );
+    }
+
+    #[tokio::test]
+    async fn local_failover_policy_can_stop_retryable_statuses_and_continue_non_retryable_statuses()
+    {
+        let state = build_state_with_provider_config(Some(serde_json::json!({
+            "failover_rules": {
+                "max_retries": 2,
+                "stop_on_status_codes": [503],
+                "continue_on_status_codes": [409]
+            }
+        })));
+        let plan = sample_plan();
+        let first_candidate = serde_json::json!({
+            "candidate_index": 0,
+            "retry_index": 0,
+        });
+        let third_candidate = serde_json::json!({
+            "candidate_index": 2,
+            "retry_index": 0,
+        });
+
+        assert!(
+            !should_retry_next_local_candidate_stream(
+                &state,
+                &plan,
+                "openai_chat_stream",
+                Some(&first_candidate),
+                503,
+                None,
+            )
+            .await
+        );
+        assert!(
+            should_stop_local_candidate_failover_stream(
+                &state,
+                &plan,
+                "openai_chat_stream",
+                Some(&first_candidate),
+                503,
+                None,
+            )
+            .await
+        );
+        assert!(
+            should_retry_next_local_candidate_stream(
+                &state,
+                &plan,
+                "openai_chat_stream",
+                Some(&first_candidate),
+                409,
+                None,
+            )
+            .await
+        );
+        assert!(
+            !should_retry_next_local_candidate_stream(
+                &state,
+                &plan,
+                "openai_chat_stream",
+                Some(&third_candidate),
+                429,
+                None,
+            )
+            .await
+        );
+    }
+
+    #[test]
+    fn resolve_local_failover_policy_reads_regex_rules() {
+        let state = build_state_with_provider_config(Some(serde_json::json!({
+            "failover_rules": {
+                "success_failover_patterns": [
+                    {"pattern": "relay:.*格式错误"}
+                ],
+                "error_stop_patterns": [
+                    {"pattern": "content_policy_violation", "status_codes": [400, 403]}
+                ]
+            }
+        })));
+        let plan = sample_plan();
+        let runtime = tokio::runtime::Runtime::new().expect("runtime should build");
+
+        let policy = runtime.block_on(resolve_local_failover_policy(&state, &plan));
+        assert_eq!(
+            policy.success_failover_patterns,
+            vec![LocalFailoverRegexRule {
+                pattern: "relay:.*格式错误".to_string(),
+                status_codes: BTreeSet::new(),
+            }]
+        );
+        assert_eq!(
+            policy.error_stop_patterns,
+            vec![LocalFailoverRegexRule {
+                pattern: "content_policy_violation".to_string(),
+                status_codes: [400, 403].into_iter().collect(),
+            }]
+        );
+    }
+
+    #[tokio::test]
+    async fn success_failover_pattern_can_retry_sync_candidate() {
+        let result = ExecutionResult {
+            request_id: "req-1".to_string(),
+            candidate_id: None,
+            status_code: 200,
+            headers: Default::default(),
+            body: None,
+            telemetry: None,
+            error: None,
+        };
+        let local_report_context = serde_json::json!({
+            "candidate_index": 0,
+            "retry_index": 0,
+        });
+        let state = build_state_with_provider_config(Some(serde_json::json!({
+            "failover_rules": {
+                "success_failover_patterns": [
+                    {"pattern": "relay:.*格式错误"}
+                ]
+            }
+        })));
+        let plan = sample_plan();
+
+        assert!(
+            should_retry_next_local_candidate_sync(
+                &state,
+                &plan,
+                "openai_chat_sync",
+                Some(&local_report_context),
+                &result,
+                Some("{\"error\":\"relay: 返回格式错误\"}"),
+            )
+            .await
+        );
+    }
+
+    #[tokio::test]
+    async fn error_stop_pattern_can_stop_sync_failover() {
+        let result = ExecutionResult {
+            request_id: "req-1".to_string(),
+            candidate_id: None,
+            status_code: 400,
+            headers: Default::default(),
+            body: None,
+            telemetry: None,
+            error: None,
+        };
+        let local_report_context = serde_json::json!({
+            "candidate_index": 0,
+            "retry_index": 0,
+        });
+        let state = build_state_with_provider_config(Some(serde_json::json!({
+            "failover_rules": {
+                "error_stop_patterns": [
+                    {"pattern": "content_policy_violation", "status_codes": [400]}
+                ]
+            }
+        })));
+        let plan = sample_plan();
+
+        assert!(
+            should_stop_local_candidate_failover_sync(
+                &state,
+                &plan,
+                "openai_chat_sync",
+                Some(&local_report_context),
+                &result,
+                Some("{\"error\":\"content_policy_violation\"}"),
+            )
+            .await
+        );
+        assert!(
+            !should_retry_next_local_candidate_sync(
+                &state,
+                &plan,
+                "openai_chat_sync",
+                Some(&local_report_context),
+                &result,
+                Some("{\"error\":\"content_policy_violation\"}"),
+            )
+            .await
+        );
     }
 }

apps/aether-gateway/src/execution_runtime/mod.rs

@@ -19,11 +19,12 @@ pub(crate) use self::constants::{
     MAX_ERROR_BODY_BYTES, MAX_STREAM_PREFETCH_BYTES, MAX_STREAM_PREFETCH_FRAMES,
 };
 pub(crate) use self::fallback::{
-    resolve_core_stream_direct_finalize_report_kind,
+    local_failover_response_text, resolve_core_stream_direct_finalize_report_kind,
     resolve_core_stream_error_finalize_report_kind, resolve_core_sync_error_finalize_report_kind,
     should_fallback_to_control_stream, should_fallback_to_control_sync,
     should_finalize_sync_response, should_retry_next_local_candidate_stream,
-    should_retry_next_local_candidate_sync,
+    should_retry_next_local_candidate_sync, should_stop_local_candidate_failover_stream,
+    should_stop_local_candidate_failover_sync,
 };
 pub use server::{
     build_execution_runtime_router, build_execution_runtime_router_with_request_concurrency_limit,

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

@@ -1,6 +1,7 @@
+use std::collections::VecDeque;
 use std::io::Error as IoError;
 
-use aether_contracts::{ExecutionPlan, ExecutionTelemetry, StreamFramePayload};
+use aether_contracts::{ExecutionPlan, ExecutionTelemetry, StreamFrame, StreamFramePayload};
 use aether_data_contracts::repository::candidates::RequestCandidateStatus;
 use aether_scheduler_core::SchedulerRequestCandidateStatusUpdate;
 use async_stream::stream;
@@ -32,7 +33,7 @@ use crate::ai_pipeline_api::{
 use crate::api::response::{
     attach_control_metadata_headers, build_client_response, build_client_response_from_parts,
 };
-use crate::clock::current_unix_secs as current_request_candidate_unix_secs;
+use crate::clock::current_unix_ms as current_request_candidate_unix_ms;
 use crate::constants::{CONTROL_CANDIDATE_ID_HEADER, CONTROL_REQUEST_ID_HEADER};
 use crate::control::GatewayControlDecision;
 use crate::execution_runtime::build_direct_execution_frame_stream;
@@ -45,9 +46,9 @@ use crate::execution_runtime::transport::{
     DirectSyncExecutionRuntime, DirectUpstreamStreamExecution,
 };
 use crate::execution_runtime::{
-    resolve_core_stream_direct_finalize_report_kind,
+    local_failover_response_text, resolve_core_stream_direct_finalize_report_kind,
     resolve_core_stream_error_finalize_report_kind, should_fallback_to_control_stream,
-    should_retry_next_local_candidate_stream,
+    should_retry_next_local_candidate_stream, should_stop_local_candidate_failover_stream,
 };
 use crate::execution_runtime::{MAX_STREAM_PREFETCH_BYTES, MAX_STREAM_PREFETCH_FRAMES};
 use crate::log_ids::short_request_id;
@@ -165,7 +166,7 @@ pub(crate) async fn execute_execution_runtime_stream(
         };
 
         if response.status() != http::StatusCode::OK {
-            let terminal_unix_secs = current_request_candidate_unix_secs();
+            let terminal_unix_secs = current_request_candidate_unix_ms();
             record_local_request_candidate_status(
                 state,
                 &plan,
@@ -179,8 +180,8 @@ pub(crate) async fn execute_execution_runtime_stream(
                         response.status()
                     )),
                     latency_ms: None,
-                    started_at_unix_secs: Some(terminal_unix_secs),
-                    finished_at_unix_secs: Some(terminal_unix_secs),
+                    started_at_unix_ms: Some(terminal_unix_secs),
+                    finished_at_unix_ms: Some(terminal_unix_secs),
                 },
             )
             .await;
@@ -209,6 +210,60 @@ pub(crate) async fn execute_execution_runtime_stream(
     }
 }
 
+fn decode_stream_data_chunk(
+    chunk_b64: Option<&str>,
+    text: Option<&str>,
+) -> Result<Vec<u8>, GatewayError> {
+    if let Some(chunk_b64) = chunk_b64 {
+        return base64::engine::general_purpose::STANDARD
+            .decode(chunk_b64)
+            .map_err(|err| GatewayError::Internal(err.to_string()));
+    }
+    Ok(text.unwrap_or_default().as_bytes().to_vec())
+}
+
+async fn next_stream_frame<R>(
+    buffered_frames: &mut VecDeque<StreamFrame>,
+    lines: &mut FramedRead<R, LinesCodec>,
+) -> Result<Option<StreamFrame>, GatewayError>
+where
+    R: tokio::io::AsyncRead + Unpin,
+{
+    if let Some(frame) = buffered_frames.pop_front() {
+        return Ok(Some(frame));
+    }
+    read_next_frame(lines).await
+}
+
+async fn probe_local_stream_success_failover_text<R>(
+    buffered_frames: &mut VecDeque<StreamFrame>,
+    lines: &mut FramedRead<R, LinesCodec>,
+) -> Result<Option<String>, GatewayError>
+where
+    R: tokio::io::AsyncRead + Unpin,
+{
+    while let Some(frame) = read_next_frame(lines).await? {
+        let probe_text = match &frame.payload {
+            StreamFramePayload::Data { chunk_b64, text } => {
+                match decode_stream_data_chunk(chunk_b64.as_deref(), text.as_deref()) {
+                    Ok(chunk) if !chunk.is_empty() => {
+                        Some(String::from_utf8_lossy(&chunk).into_owned())
+                    }
+                    Ok(_) | Err(_) => None,
+                }
+            }
+            StreamFramePayload::Error { .. } | StreamFramePayload::Eof { .. } => None,
+            StreamFramePayload::Headers { .. } | StreamFramePayload::Telemetry { .. } => None,
+        };
+        buffered_frames.push_back(frame);
+        if probe_text.is_some() {
+            return Ok(probe_text);
+        }
+    }
+
+    Ok(None)
+}
+
 async fn execute_stream_from_frame_stream(
     state: &AppState,
     plan: ExecutionPlan,
@@ -237,69 +292,137 @@ async fn execute_stream_from_frame_stream(
             "execution runtime stream must start with headers frame".to_string(),
         ));
     };
+    let mut buffered_frames = VecDeque::new();
 
-    if should_retry_next_local_candidate_stream(plan_kind, report_context.as_ref(), status_code) {
-        let terminal_unix_secs = current_request_candidate_unix_secs();
-        record_local_request_candidate_status(
+    if status_code == 200 {
+        let success_probe_text =
+            probe_local_stream_success_failover_text(&mut buffered_frames, &mut lines).await?;
+        if should_retry_next_local_candidate_stream(
             state,
             &plan,
+            plan_kind,
             report_context.as_ref(),
-            SchedulerRequestCandidateStatusUpdate {
-                status: RequestCandidateStatus::Failed,
-                status_code: Some(status_code),
-                error_type: Some("retryable_upstream_status".to_string()),
-                error_message: Some(format!(
-                    "execution runtime stream returned retryable status {status_code}"
-                )),
-                latency_ms: None,
-                started_at_unix_secs: Some(terminal_unix_secs),
-                finished_at_unix_secs: Some(terminal_unix_secs),
-            },
-        )
-        .await;
-        warn!(
-            event_name = "local_stream_candidate_retry_scheduled",
-            log_type = "event",
-            trace_id = %trace_id,
-            request_id = %request_id_for_log,
             status_code,
-            "gateway local stream decision retrying next candidate after retryable execution runtime status"
-        );
-        return Ok(None);
+            success_probe_text.as_deref(),
+        )
+        .await
+        {
+            let terminal_unix_secs = current_request_candidate_unix_ms();
+            record_local_request_candidate_status(
+                state,
+                &plan,
+                report_context.as_ref(),
+                SchedulerRequestCandidateStatusUpdate {
+                    status: RequestCandidateStatus::Failed,
+                    status_code: Some(status_code),
+                    error_type: Some("success_failover_pattern".to_string()),
+                    error_message: Some(
+                        "execution runtime stream matched provider success failover rule"
+                            .to_string(),
+                    ),
+                    latency_ms: None,
+                    started_at_unix_ms: Some(terminal_unix_secs),
+                    finished_at_unix_ms: Some(terminal_unix_secs),
+                },
+            )
+            .await;
+            warn!(
+                event_name = "local_stream_candidate_retry_scheduled",
+                log_type = "event",
+                trace_id = %trace_id,
+                request_id = %request_id_for_log,
+                status_code,
+                "gateway local stream decision retrying next candidate after success failover rule match"
+            );
+            return Ok(None);
+        }
     }
 
     let stream_error_finalize_kind =
         resolve_core_stream_error_finalize_report_kind(plan_kind, status_code);
 
-    if should_fallback_to_control_stream(
-        plan_kind,
-        status_code,
-        stream_error_finalize_kind.is_some(),
-    ) {
-        let terminal_unix_secs = current_request_candidate_unix_secs();
-        record_local_request_candidate_status(
-            state,
-            &plan,
-            report_context.as_ref(),
-            SchedulerRequestCandidateStatusUpdate {
-                status: RequestCandidateStatus::Failed,
-                status_code: Some(status_code),
-                error_type: Some("control_fallback".to_string()),
-                error_message: Some(format!(
-                    "stream decision fell back to control after status {status_code}"
-                )),
-                latency_ms: None,
-                started_at_unix_secs: Some(terminal_unix_secs),
-                finished_at_unix_secs: Some(terminal_unix_secs),
-            },
-        )
-        .await;
-        return Ok(None);
-    }
-
     if status_code >= 400 {
         let error_body = collect_error_body(&mut lines).await?;
         let (body_json, body_base64) = decode_stream_error_body(&headers, &error_body);
+        let error_response_text =
+            local_failover_response_text(body_json.as_ref(), &error_body, None);
+        let stop_local_failover = should_stop_local_candidate_failover_stream(
+            state,
+            &plan,
+            plan_kind,
+            report_context.as_ref(),
+            status_code,
+            error_response_text.as_deref(),
+        )
+        .await;
+        if !stop_local_failover
+            && should_retry_next_local_candidate_stream(
+                state,
+                &plan,
+                plan_kind,
+                report_context.as_ref(),
+                status_code,
+                error_response_text.as_deref(),
+            )
+            .await
+        {
+            let terminal_unix_secs = current_request_candidate_unix_ms();
+            record_local_request_candidate_status(
+                state,
+                &plan,
+                report_context.as_ref(),
+                SchedulerRequestCandidateStatusUpdate {
+                    status: RequestCandidateStatus::Failed,
+                    status_code: Some(status_code),
+                    error_type: Some("retryable_upstream_status".to_string()),
+                    error_message: Some(format!(
+                        "execution runtime stream returned retryable status {status_code}"
+                    )),
+                    latency_ms: None,
+                    started_at_unix_ms: Some(terminal_unix_secs),
+                    finished_at_unix_ms: Some(terminal_unix_secs),
+                },
+            )
+            .await;
+            warn!(
+                event_name = "local_stream_candidate_retry_scheduled",
+                log_type = "event",
+                trace_id = %trace_id,
+                request_id = %request_id_for_log,
+                status_code,
+                "gateway local stream decision retrying next candidate after retryable execution runtime status"
+            );
+            return Ok(None);
+        }
+
+        if !stop_local_failover
+            && should_fallback_to_control_stream(
+                plan_kind,
+                status_code,
+                stream_error_finalize_kind.is_some(),
+            )
+        {
+            let terminal_unix_secs = current_request_candidate_unix_ms();
+            record_local_request_candidate_status(
+                state,
+                &plan,
+                report_context.as_ref(),
+                SchedulerRequestCandidateStatusUpdate {
+                    status: RequestCandidateStatus::Failed,
+                    status_code: Some(status_code),
+                    error_type: Some("control_fallback".to_string()),
+                    error_message: Some(format!(
+                        "stream decision fell back to control after status {status_code}"
+                    )),
+                    latency_ms: None,
+                    started_at_unix_ms: Some(terminal_unix_secs),
+                    finished_at_unix_ms: Some(terminal_unix_secs),
+                },
+            )
+            .await;
+            return Ok(None);
+        }
+
         let usage_report_kind = stream_error_finalize_kind
             .clone()
             .or_else(|| report_kind.clone())
@@ -324,7 +447,7 @@ async fn execute_stream_from_frame_stream(
                 &usage_payload,
             )
             .await;
-        let terminal_unix_secs = current_request_candidate_unix_secs();
+        let terminal_unix_secs = current_request_candidate_unix_ms();
         record_local_request_candidate_status(
             state,
             &plan,
@@ -337,8 +460,8 @@ async fn execute_stream_from_frame_stream(
                     "execution runtime stream returned error status {status_code}"
                 )),
                 latency_ms: None,
-                started_at_unix_secs: Some(terminal_unix_secs),
-                finished_at_unix_secs: Some(terminal_unix_secs),
+                started_at_unix_ms: Some(terminal_unix_secs),
+                finished_at_unix_ms: Some(terminal_unix_secs),
             },
         )
         .await;
@@ -399,7 +522,7 @@ async fn execute_stream_from_frame_stream(
         while prefetched_chunks.len() < MAX_STREAM_PREFETCH_FRAMES
             && prefetched_inspection_body.len() < MAX_STREAM_PREFETCH_BYTES
         {
-            let Some(frame) = (match read_next_frame(&mut lines).await {
+            let Some(frame) = (match next_stream_frame(&mut buffered_frames, &mut lines).await {
                 Ok(frame) => frame,
                 Err(err) => {
                     let failure = build_stream_failure_report(
@@ -429,14 +552,14 @@ async fn execute_stream_from_frame_stream(
             };
             match frame.payload {
                 StreamFramePayload::Data { chunk_b64, text } => {
-                    let chunk = if let Some(chunk_b64) = chunk_b64 {
-                        match base64::engine::general_purpose::STANDARD.decode(chunk_b64) {
-                            Ok(decoded) => decoded,
+                    let chunk =
+                        match decode_stream_data_chunk(chunk_b64.as_deref(), text.as_deref()) {
+                            Ok(chunk) => chunk,
                             Err(err) => {
                                 let failure = build_stream_failure_report(
                                     "execution_runtime_stream_chunk_decode_error",
                                     format!(
-                                        "failed to decode execution runtime stream chunk: {err}"
+                                        "failed to decode execution runtime stream chunk: {err:?}"
                                     ),
                                     502,
                                 );
@@ -456,12 +579,7 @@ async fn execute_stream_from_frame_stream(
                                 )
                                 .await;
                             }
-                        }
-                    } else if let Some(text) = text {
-                        text.into_bytes()
-                    } else {
-                        Vec::new()
-                    };
+                        };
 
                     if chunk.is_empty() {
                         continue;
@@ -621,7 +739,7 @@ async fn execute_stream_from_frame_stream(
         }
     }
 
-    let candidate_started_unix_secs = current_request_candidate_unix_secs();
+    let candidate_started_unix_secs = current_request_candidate_unix_ms();
     state
         .usage_runtime
         .record_pending(state.data.as_ref(), &plan, report_context.as_ref())
@@ -649,8 +767,8 @@ async fn execute_stream_from_frame_stream(
             latency_ms: prefetched_telemetry
                 .as_ref()
                 .and_then(|telemetry| telemetry.elapsed_ms),
-            started_at_unix_secs: Some(candidate_started_unix_secs),
-            finished_at_unix_secs: None,
+            started_at_unix_ms: Some(candidate_started_unix_secs),
+            finished_at_unix_ms: None,
         },
     )
     .await;
@@ -672,6 +790,7 @@ async fn execute_stream_from_frame_stream(
     let request_id_for_report = request_id.to_string();
     let request_id_for_report_log = short_request_id(request_id);
     let candidate_id_for_report = candidate_id.map(ToOwned::to_owned);
+    let mut buffered_frames = buffered_frames;
     tokio::spawn(async move {
         let mut provider_buffered_body = provider_prefetched_body_for_report;
         let mut buffered_body = prefetched_body_for_report;
@@ -682,7 +801,7 @@ async fn execute_stream_from_frame_stream(
 
         if !reached_eof {
             loop {
-                let next_frame = match read_next_frame(&mut lines).await {
+                let next_frame = match next_stream_frame(&mut buffered_frames, &mut lines).await {
                     Ok(frame) => frame,
                     Err(err) => {
                         warn!(
@@ -707,9 +826,9 @@ async fn execute_stream_from_frame_stream(
                 };
                 match frame.payload {
                     StreamFramePayload::Data { chunk_b64, text } => {
-                        let chunk = if let Some(chunk_b64) = chunk_b64 {
-                            match base64::engine::general_purpose::STANDARD.decode(chunk_b64) {
-                                Ok(decoded) => decoded,
+                        let chunk =
+                            match decode_stream_data_chunk(chunk_b64.as_deref(), text.as_deref()) {
+                                Ok(chunk) => chunk,
                                 Err(err) => {
                                     warn!(
                                         event_name = "stream_execution_chunk_decode_failed",
@@ -717,22 +836,19 @@ async fn execute_stream_from_frame_stream(
                                         trace_id = %trace_id_owned,
                                         request_id = %request_id_for_report_log,
                                         candidate_id = ?candidate_id_for_report.as_deref(),
-                                        error = %err,
+                                        error = ?err,
                                         "gateway failed to decode execution runtime chunk"
                                     );
                                     terminal_failure = Some(build_stream_failure_report(
                                         "execution_runtime_stream_chunk_decode_error",
-                                        format!("failed to decode execution runtime stream chunk: {err}"),
+                                        format!(
+                                        "failed to decode execution runtime stream chunk: {err:?}"
+                                    ),
                                         502,
                                     ));
                                     break;
                                 }
-                            }
-                        } else if let Some(text) = text {
-                            text.into_bytes()
-                        } else {
-                            Vec::new()
-                        };
+                            };
 
                         if chunk.is_empty() {
                             continue;
@@ -998,8 +1114,8 @@ async fn execute_stream_from_frame_stream(
                     error_type: Some("downstream_disconnect".to_string()),
                     error_message: Some("client disconnected before stream completion".to_string()),
                     latency_ms: telemetry.as_ref().and_then(|value| value.elapsed_ms),
-                    started_at_unix_secs: Some(candidate_started_unix_secs_for_report),
-                    finished_at_unix_secs: Some(current_request_candidate_unix_secs()),
+                    started_at_unix_ms: Some(candidate_started_unix_secs_for_report),
+                    finished_at_unix_ms: Some(current_request_candidate_unix_ms()),
                 },
             )
             .await;
@@ -1055,8 +1171,8 @@ async fn execute_stream_from_frame_stream(
                 error_type: None,
                 error_message: None,
                 latency_ms: telemetry.as_ref().and_then(|value| value.elapsed_ms),
-                started_at_unix_secs: Some(candidate_started_unix_secs_for_report),
-                finished_at_unix_secs: Some(current_request_candidate_unix_secs()),
+                started_at_unix_ms: Some(candidate_started_unix_secs_for_report),
+                finished_at_unix_ms: Some(current_request_candidate_unix_ms()),
             },
         )
         .await;

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

@@ -8,7 +8,7 @@ use serde_json::{Map, Value};
 use tracing::warn;
 
 use crate::api::response::attach_control_metadata_headers;
-use crate::clock::current_unix_secs as current_request_candidate_unix_secs;
+use crate::clock::current_unix_ms as current_request_candidate_unix_ms;
 use crate::control::GatewayControlDecision;
 use crate::execution_runtime::submission::{
     resolve_core_error_background_report_kind, submit_local_core_error_or_sync_finalize,
@@ -117,13 +117,13 @@ async fn record_stream_sync_failure(
     report_context: Option<&Value>,
     payload: &GatewaySyncReportRequest,
     failure: &StreamFailureReport,
-    started_at_unix_secs: Option<u64>,
+    started_at_unix_ms: Option<u64>,
 ) {
     state
         .usage_runtime
         .record_sync_terminal(state.data.as_ref(), plan, report_context, payload)
         .await;
-    let terminal_unix_secs = current_request_candidate_unix_secs();
+    let terminal_unix_secs = current_request_candidate_unix_ms();
     record_report_request_candidate_status(
         state,
         report_context,
@@ -136,8 +136,8 @@ async fn record_stream_sync_failure(
                 .telemetry
                 .as_ref()
                 .and_then(|telemetry| telemetry.elapsed_ms),
-            started_at_unix_secs: started_at_unix_secs.or(Some(terminal_unix_secs)),
-            finished_at_unix_secs: Some(terminal_unix_secs),
+            started_at_unix_ms: started_at_unix_ms.or(Some(terminal_unix_secs)),
+            finished_at_unix_ms: Some(terminal_unix_secs),
         },
     )
     .await;
@@ -195,7 +195,7 @@ pub(super) async fn submit_midstream_stream_failure(
     headers: &std::collections::BTreeMap<String, String>,
     telemetry: Option<ExecutionTelemetry>,
     buffered_body: &[u8],
-    started_at_unix_secs: u64,
+    started_at_unix_ms: u64,
     failure: StreamFailureReport,
 ) {
     let Some(report_kind) =
@@ -219,7 +219,7 @@ pub(super) async fn submit_midstream_stream_failure(
         report_context,
         &payload,
         &failure,
-        Some(started_at_unix_secs),
+        Some(started_at_unix_ms),
     )
     .await;
     if let Err(err) = submit_sync_report(state, trace_id, payload).await {

apps/aether-gateway/src/execution_runtime/submission.rs

@@ -297,29 +297,39 @@ fn resolve_local_sync_error_status_code(status_code: u16, body_json: &serde_json
         return status_code;
     }
 
-    let Some(error_object) = body_json.get("error").and_then(|value| value.as_object()) else {
-        return 400;
-    };
+    let body_object = body_json.as_object();
+    let error_object = body_object
+        .and_then(|object| object.get("error"))
+        .and_then(|value| value.as_object());
 
-    for key in ["code", "status"] {
-        let Some(value) = error_object.get(key) else {
-            continue;
-        };
-        if let Some(number) = value.as_u64() {
+    let raw_code = first_non_empty_error_text(error_object, body_object, &["code"]);
+    let raw_status = first_non_empty_error_text(error_object, body_object, &["status"]);
+    for numeric_hint in [raw_code.as_deref(), raw_status.as_deref()]
+        .into_iter()
+        .flatten()
+    {
+        if let Ok(number) = numeric_hint.parse::<u16>() {
             if (400..600).contains(&number) {
-                return number as u16;
-            }
-        }
-        if let Some(text) = value.as_str() {
-            if let Ok(number) = text.parse::<u16>() {
-                if (400..600).contains(&number) {
-                    return number;
-                }
+                return number;
             }
         }
     }
 
-    400
+    let raw_type = first_non_empty_error_text(error_object, body_object, &["type", "__type"]);
+    let message = first_non_empty_error_text(
+        error_object,
+        body_object,
+        &["message", "detail", "reason", "status", "type", "__type"],
+    )
+    .unwrap_or_else(|| "HTTP 400".to_string());
+    let kind = classify_local_sync_error_kind(
+        status_code,
+        raw_type.as_deref(),
+        raw_status.as_deref(),
+        raw_code.as_deref(),
+        message.as_str(),
+    );
+    default_status_code_for_local_sync_error_kind(kind)
 }
 
 fn extract_local_sync_error_details(
@@ -436,6 +446,20 @@ fn classify_local_sync_error_kind(
     LocalCoreSyncErrorKind::InvalidRequest
 }
 
+fn default_status_code_for_local_sync_error_kind(kind: LocalCoreSyncErrorKind) -> u16 {
+    match kind {
+        LocalCoreSyncErrorKind::InvalidRequest | LocalCoreSyncErrorKind::ContextLengthExceeded => {
+            400
+        }
+        LocalCoreSyncErrorKind::Authentication => 401,
+        LocalCoreSyncErrorKind::PermissionDenied => 403,
+        LocalCoreSyncErrorKind::NotFound => 404,
+        LocalCoreSyncErrorKind::RateLimit => 429,
+        LocalCoreSyncErrorKind::Overloaded => 503,
+        LocalCoreSyncErrorKind::ServerError => 500,
+    }
+}
+
 pub(crate) fn strip_utf8_bom_and_ws(mut body: &[u8]) -> &[u8] {
     loop {
         while let Some(first) = body.first() {
@@ -534,3 +558,128 @@ pub(crate) async fn submit_local_core_error_or_sync_finalize(
 
     Ok(response)
 }
+
+#[cfg(test)]
+mod tests {
+    use axum::body::to_bytes;
+    use serde_json::json;
+
+    use super::maybe_build_local_core_error_response;
+    use crate::control::GatewayControlDecision;
+    use crate::usage::GatewaySyncReportRequest;
+
+    fn test_decision() -> GatewayControlDecision {
+        GatewayControlDecision::synthetic(
+            "/v1/chat/completions",
+            Some("ai_public".to_string()),
+            Some("openai".to_string()),
+            Some("chat".to_string()),
+            Some("openai:chat".to_string()),
+        )
+        .with_execution_runtime_candidate(true)
+    }
+
+    fn core_finalize_payload(
+        report_kind: &str,
+        client_api_format: &str,
+        provider_api_format: &str,
+        status_code: u16,
+        body_json: serde_json::Value,
+    ) -> GatewaySyncReportRequest {
+        GatewaySyncReportRequest {
+            trace_id: "trace-core-error-status-123".to_string(),
+            report_kind: report_kind.to_string(),
+            report_context: Some(json!({
+                "client_api_format": client_api_format,
+                "provider_api_format": provider_api_format,
+            })),
+            status_code,
+            headers: Default::default(),
+            body_json: Some(body_json),
+            client_body_json: None,
+            body_base64: None,
+            telemetry: None,
+        }
+    }
+
+    #[tokio::test]
+    async fn maybe_build_local_core_error_response_infers_status_from_semantic_error_type() {
+        let payload = core_finalize_payload(
+            "openai_chat_sync_finalize",
+            "openai:chat",
+            "claude:chat",
+            200,
+            json!({
+                "type": "error",
+                "error": {
+                    "type": "rate_limit_error",
+                    "message": "slow down"
+                }
+            }),
+        );
+
+        let response = maybe_build_local_core_error_response(
+            "trace-sync-status-type",
+            &test_decision(),
+            &payload,
+        )
+        .expect("response build should not error")
+        .expect("response should exist");
+
+        assert_eq!(response.status(), http::StatusCode::TOO_MANY_REQUESTS);
+        assert_eq!(
+            serde_json::from_slice::<serde_json::Value>(
+                &to_bytes(response.into_body(), usize::MAX)
+                    .await
+                    .expect("body should read"),
+            )
+            .expect("body should decode"),
+            json!({
+                "error": {
+                    "message": "slow down",
+                    "type": "rate_limit_error"
+                }
+            })
+        );
+    }
+
+    #[tokio::test]
+    async fn maybe_build_local_core_error_response_infers_status_from_gemini_status_text() {
+        let payload = core_finalize_payload(
+            "gemini_chat_sync_finalize",
+            "gemini:chat",
+            "gemini:chat",
+            200,
+            json!({
+                "error": {
+                    "message": "quota reached",
+                    "status": "RESOURCE_EXHAUSTED"
+                }
+            }),
+        );
+
+        let response = maybe_build_local_core_error_response(
+            "trace-sync-status-gemini",
+            &test_decision(),
+            &payload,
+        )
+        .expect("response build should not error")
+        .expect("response should exist");
+
+        assert_eq!(response.status(), http::StatusCode::TOO_MANY_REQUESTS);
+        assert_eq!(
+            serde_json::from_slice::<serde_json::Value>(
+                &to_bytes(response.into_body(), usize::MAX)
+                    .await
+                    .expect("body should read"),
+            )
+            .expect("body should decode"),
+            json!({
+                "error": {
+                    "message": "quota reached",
+                    "status": "RESOURCE_EXHAUSTED"
+                }
+            })
+        );
+    }
+}

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

@@ -15,7 +15,7 @@ use crate::ai_pipeline_api::{
 use crate::api::response::{
     attach_control_metadata_headers, build_client_response, build_client_response_from_parts,
 };
-use crate::clock::current_unix_secs as current_request_candidate_unix_secs;
+use crate::clock::current_unix_ms as current_request_candidate_unix_ms;
 use crate::constants::{CONTROL_CANDIDATE_ID_HEADER, CONTROL_REQUEST_ID_HEADER};
 use crate::control::GatewayControlDecision;
 #[cfg(test)]
@@ -23,8 +23,9 @@ use crate::execution_runtime::remote_compat::post_sync_plan_to_remote_execution_
 use crate::execution_runtime::submission::submit_local_core_error_or_sync_finalize;
 use crate::execution_runtime::transport::DirectSyncExecutionRuntime;
 use crate::execution_runtime::{
-    resolve_core_sync_error_finalize_report_kind, should_fallback_to_control_sync,
-    should_finalize_sync_response, should_retry_next_local_candidate_sync,
+    local_failover_response_text, resolve_core_sync_error_finalize_report_kind,
+    should_fallback_to_control_sync, should_finalize_sync_response,
+    should_retry_next_local_candidate_sync, should_stop_local_candidate_failover_sync,
 };
 use crate::log_ids::short_request_id;
 use crate::request_candidate_runtime::{
@@ -85,6 +86,7 @@ pub(crate) async fn execute_execution_runtime_sync(
     let plan_request_id = plan.request_id.as_str();
     let plan_request_id_for_log = short_request_id(plan_request_id);
     let plan_candidate_id = plan.candidate_id.as_deref();
+    let candidate_started_unix_secs = current_request_candidate_unix_ms();
     #[cfg(not(test))]
     let result = {
         match DirectSyncExecutionRuntime::new()
@@ -140,6 +142,7 @@ pub(crate) async fn execute_execution_runtime_sync(
                 plan_request_id,
                 plan_candidate_id,
                 report_context.as_ref(),
+                candidate_started_unix_secs,
             )
             .await?;
             match remote_outcome {
@@ -159,8 +162,34 @@ pub(crate) async fn execute_execution_runtime_sync(
         .telemetry
         .as_ref()
         .and_then(|telemetry| telemetry.elapsed_ms);
-    if should_retry_next_local_candidate_sync(plan_kind, report_context.as_ref(), &result) {
-        let terminal_unix_secs = current_request_candidate_unix_secs();
+    let mut headers = result.headers.clone();
+    let (body_bytes, body_json, body_base64) = decode_execution_result_body(&result, &mut headers)?;
+    let local_failover_response_text = local_failover_response_text(
+        body_json.as_ref(),
+        &body_bytes,
+        result.error.as_ref().map(|error| error.message.as_str()),
+    );
+    let stop_local_failover = should_stop_local_candidate_failover_sync(
+        state,
+        &plan,
+        plan_kind,
+        report_context.as_ref(),
+        &result,
+        local_failover_response_text.as_deref(),
+    )
+    .await;
+    if should_retry_next_local_candidate_sync(
+        state,
+        &plan,
+        plan_kind,
+        report_context.as_ref(),
+        &result,
+        local_failover_response_text.as_deref(),
+    )
+    .await
+        && !stop_local_failover
+    {
+        let terminal_unix_secs = current_request_candidate_unix_ms();
         record_local_request_candidate_status(
             state,
             &plan,
@@ -171,8 +200,8 @@ pub(crate) async fn execute_execution_runtime_sync(
                 error_type: result_error_type.clone(),
                 error_message: result_error_message.clone(),
                 latency_ms: result_latency_ms,
-                started_at_unix_secs: Some(terminal_unix_secs),
-                finished_at_unix_secs: Some(terminal_unix_secs),
+                started_at_unix_ms: Some(candidate_started_unix_secs),
+                finished_at_unix_ms: Some(terminal_unix_secs),
             },
         )
         .await;
@@ -191,8 +220,6 @@ pub(crate) async fn execute_execution_runtime_sync(
         .or(Some(plan_request_id));
     let request_id_for_log = short_request_id(request_id.unwrap_or("-"));
     let candidate_id = result.candidate_id.as_deref().or(plan_candidate_id);
-    let mut headers = result.headers.clone();
-    let (body_bytes, body_json, body_base64) = decode_execution_result_body(&result, &mut headers)?;
     let has_body_bytes = body_base64.is_some();
     let explicit_finalize = should_finalize_sync_response(report_kind.as_deref());
     let mapped_error_finalize_kind =
@@ -220,15 +247,17 @@ pub(crate) async fn execute_execution_runtime_sync(
         mapped_error_finalize_kind.clone()
     };
 
-    if should_fallback_to_control_sync(
-        plan_kind,
-        &result,
-        body_json.as_ref(),
-        has_body_bytes,
-        explicit_finalize || implicit_finalize.is_some(),
-        mapped_error_finalize_kind.is_some(),
-    ) {
-        let terminal_unix_secs = current_request_candidate_unix_secs();
+    if !stop_local_failover
+        && should_fallback_to_control_sync(
+            plan_kind,
+            &result,
+            body_json.as_ref(),
+            has_body_bytes,
+            explicit_finalize || implicit_finalize.is_some(),
+            mapped_error_finalize_kind.is_some(),
+        )
+    {
+        let terminal_unix_secs = current_request_candidate_unix_ms();
         record_local_request_candidate_status(
             state,
             &plan,
@@ -239,8 +268,8 @@ pub(crate) async fn execute_execution_runtime_sync(
                 error_type: result_error_type.clone(),
                 error_message: result_error_message.clone(),
                 latency_ms: result_latency_ms,
-                started_at_unix_secs: Some(terminal_unix_secs),
-                finished_at_unix_secs: Some(terminal_unix_secs),
+                started_at_unix_ms: Some(candidate_started_unix_secs),
+                finished_at_unix_ms: Some(terminal_unix_secs),
             },
         )
         .await;
@@ -251,7 +280,7 @@ pub(crate) async fn execute_execution_runtime_sync(
         .usage_runtime
         .record_pending(state.data.as_ref(), &plan, report_context.as_ref())
         .await;
-    let terminal_unix_secs = current_request_candidate_unix_secs();
+    let terminal_unix_secs = current_request_candidate_unix_ms();
     record_local_request_candidate_status(
         state,
         &plan,
@@ -266,8 +295,8 @@ pub(crate) async fn execute_execution_runtime_sync(
             error_type: result_error_type.clone(),
             error_message: result_error_message.clone(),
             latency_ms: result_latency_ms,
-            started_at_unix_secs: Some(terminal_unix_secs),
-            finished_at_unix_secs: Some(terminal_unix_secs),
+            started_at_unix_ms: Some(candidate_started_unix_secs),
+            finished_at_unix_ms: Some(terminal_unix_secs),
         },
     )
     .await;
@@ -579,6 +608,7 @@ async fn execute_sync_via_remote_execution_runtime(
     plan_request_id: &str,
     plan_candidate_id: Option<&str>,
     report_context: Option<&serde_json::Value>,
+    candidate_started_unix_secs: u64,
 ) -> Result<RemoteSyncFallbackOutcome, GatewayError> {
     let response = match post_sync_plan_to_remote_execution_runtime(
         state,
@@ -604,7 +634,7 @@ async fn execute_sync_via_remote_execution_runtime(
     };
 
     if response.status() != http::StatusCode::OK {
-        let terminal_unix_secs = current_request_candidate_unix_secs();
+        let terminal_unix_secs = current_request_candidate_unix_ms();
         record_local_request_candidate_status(
             state,
             plan,
@@ -618,8 +648,8 @@ async fn execute_sync_via_remote_execution_runtime(
                     response.status()
                 )),
                 latency_ms: None,
-                started_at_unix_secs: Some(terminal_unix_secs),
-                finished_at_unix_secs: Some(terminal_unix_secs),
+                started_at_unix_ms: Some(candidate_started_unix_secs),
+                finished_at_unix_ms: Some(terminal_unix_secs),
             },
         )
         .await;

apps/aether-gateway/src/execution_runtime/tests.rs

@@ -228,6 +228,160 @@ fn build_best_effort_local_core_error_body_converts_claude_cli_error_to_openai_c
     );
 }
 
+#[test]
+fn build_best_effort_local_core_error_body_converts_sync_errors_across_standard_families() {
+    let cases = vec![
+        (
+            "claude chat -> openai chat",
+            core_finalize_payload(
+                "openai_chat_sync_finalize",
+                "openai:chat",
+                "claude:chat",
+                429,
+                json!({
+                    "type": "error",
+                    "error": {
+                        "type": "rate_limit_error",
+                        "message": "slow down"
+                    }
+                }),
+            ),
+            json!({
+                "error": {
+                    "message": "slow down",
+                    "type": "rate_limit_error"
+                }
+            }),
+        ),
+        (
+            "openai chat -> claude chat",
+            core_finalize_payload(
+                "claude_chat_sync_finalize",
+                "claude:chat",
+                "openai:chat",
+                404,
+                json!({
+                    "error": {
+                        "message": "missing model",
+                        "type": "not_found_error",
+                        "code": "model_missing"
+                    }
+                }),
+            ),
+            json!({
+                "type": "error",
+                "error": {
+                    "message": "missing model",
+                    "type": "not_found_error",
+                    "code": "model_missing"
+                }
+            }),
+        ),
+        (
+            "openai chat -> gemini chat",
+            core_finalize_payload(
+                "gemini_chat_sync_finalize",
+                "gemini:chat",
+                "openai:chat",
+                401,
+                json!({
+                    "error": {
+                        "message": "bad auth",
+                        "type": "authentication_error"
+                    }
+                }),
+            ),
+            json!({
+                "error": {
+                    "code": 401,
+                    "message": "bad auth",
+                    "status": "UNAUTHENTICATED"
+                }
+            }),
+        ),
+        (
+            "gemini cli -> openai cli",
+            core_finalize_payload(
+                "openai_cli_sync_finalize",
+                "openai:cli",
+                "gemini:cli",
+                429,
+                json!({
+                    "error": {
+                        "message": "quota hit",
+                        "status": "RESOURCE_EXHAUSTED",
+                        "code": 429
+                    }
+                }),
+            ),
+            json!({
+                "error": {
+                    "message": "quota hit",
+                    "type": "rate_limit_error",
+                    "code": "429"
+                }
+            }),
+        ),
+        (
+            "gemini cli -> claude cli",
+            core_finalize_payload(
+                "claude_cli_sync_finalize",
+                "claude:cli",
+                "gemini:cli",
+                503,
+                json!({
+                    "error": {
+                        "message": "backend busy",
+                        "status": "UNAVAILABLE"
+                    }
+                }),
+            ),
+            json!({
+                "type": "error",
+                "error": {
+                    "message": "backend busy",
+                    "type": "api_error",
+                    "code": "UNAVAILABLE"
+                }
+            }),
+        ),
+        (
+            "claude cli -> gemini cli",
+            core_finalize_payload(
+                "gemini_cli_sync_finalize",
+                "gemini:cli",
+                "claude:cli",
+                404,
+                json!({
+                    "type": "error",
+                    "error": {
+                        "type": "not_found_error",
+                        "message": "resource missing"
+                    }
+                }),
+            ),
+            json!({
+                "error": {
+                    "code": 404,
+                    "message": "resource missing",
+                    "status": "NOT_FOUND"
+                }
+            }),
+        ),
+    ];
+
+    for (label, payload, expected) in cases {
+        let converted = build_best_effort_local_core_error_body(
+            &payload,
+            payload.body_json.as_ref().expect("body_json should exist"),
+        )
+        .expect("conversion should not error")
+        .expect("conversion should produce a client error body");
+
+        assert_eq!(converted, expected, "unexpected conversion for {label}");
+    }
+}
+
 #[test]
 fn resolve_local_core_error_response_body_json_parses_body_base64_json_for_cross_format_cli_error()
 {

apps/aether-gateway/src/executor/candidate_loop.rs

@@ -125,8 +125,8 @@ where
                 error_type: None,
                 error_message: None,
                 latency_ms: None,
-                started_at_unix_secs: None,
-                finished_at_unix_secs: None,
+                started_at_unix_ms: None,
+                finished_at_unix_ms: None,
             },
         )
         .await;
@@ -153,8 +153,8 @@ pub(crate) async fn mark_unused_local_candidate_items<T, FPlan, FContext>(
                 error_type: None,
                 error_message: None,
                 latency_ms: None,
-                started_at_unix_secs: None,
-                finished_at_unix_secs: None,
+                started_at_unix_ms: None,
+                finished_at_unix_ms: None,
             },
         )
         .await;

apps/aether-gateway/src/handlers/admin/billing/collectors/support.rs

@@ -53,7 +53,7 @@ pub(super) fn build_admin_billing_collector_payload_from_record(
         "default_value": record.default_value,
         "priority": record.priority,
         "is_enabled": record.is_enabled,
-        "created_at": unix_secs_to_rfc3339(record.created_at_unix_secs),
+        "created_at": unix_secs_to_rfc3339(record.created_at_unix_ms),
         "updated_at": unix_secs_to_rfc3339(record.updated_at_unix_secs),
     })
 }

apps/aether-gateway/src/handlers/admin/billing/payments/shared.rs

@@ -213,7 +213,7 @@ pub(super) fn build_admin_payment_order_payload(
         "gateway_order_id": record.gateway_order_id,
         "gateway_response": record.gateway_response,
         "status": admin_payment_effective_status(&record.status, record.expires_at_unix_secs),
-        "created_at": unix_secs_to_rfc3339(record.created_at_unix_secs),
+        "created_at": unix_secs_to_rfc3339(record.created_at_unix_ms),
         "paid_at": record.paid_at_unix_secs.and_then(unix_secs_to_rfc3339),
         "credited_at": record.credited_at_unix_secs.and_then(unix_secs_to_rfc3339),
         "expires_at": record.expires_at_unix_secs.and_then(unix_secs_to_rfc3339),
@@ -236,7 +236,7 @@ pub(super) fn build_admin_payment_callback_payload(
         "payload": row.try_get::<Option<serde_json::Value>, _>("payload").map_err(|err| GatewayError::Internal(err.to_string()))?,
         "error_message": row.try_get::<Option<String>, _>("error_message").map_err(|err| GatewayError::Internal(err.to_string()))?,
         "created_at": row
-            .try_get::<Option<i64>, _>("created_at_unix_secs")
+            .try_get::<Option<i64>, _>("created_at_unix_ms")
             .map_err(|err| GatewayError::Internal(err.to_string()))?
             .and_then(|value| u64::try_from(value).ok())
             .and_then(unix_secs_to_rfc3339),
@@ -263,7 +263,7 @@ pub(super) fn build_admin_payment_callback_payload_from_record(
         "status": record.status,
         "payload": record.payload,
         "error_message": record.error_message,
-        "created_at": unix_secs_to_rfc3339(record.created_at_unix_secs),
+        "created_at": unix_secs_to_rfc3339(record.created_at_unix_ms),
         "processed_at": record.processed_at_unix_secs.and_then(unix_secs_to_rfc3339),
     })
 }

apps/aether-gateway/src/handlers/admin/billing/rules.rs

@@ -53,7 +53,7 @@ fn build_admin_billing_rule_payload_from_record(
         "variables": record.variables,
         "dimension_mappings": record.dimension_mappings,
         "is_enabled": record.is_enabled,
-        "created_at": unix_secs_to_rfc3339(record.created_at_unix_secs),
+        "created_at": unix_secs_to_rfc3339(record.created_at_unix_ms),
         "updated_at": unix_secs_to_rfc3339(record.updated_at_unix_secs),
     })
 }

apps/aether-gateway/src/handlers/admin/billing/wallets/mutations/adjust.rs

@@ -97,7 +97,7 @@ pub(in super::super) async fn build_admin_wallet_adjust_response(
         transaction.link_id.as_deref(),
         transaction.operator_id.as_deref(),
         transaction.description.as_deref(),
-        unix_secs_to_rfc3339(transaction.created_at_unix_secs),
+        unix_secs_to_rfc3339(transaction.created_at_unix_ms),
     );
     let response = Json(json!({
         "wallet": wallet_payload,

apps/aether-gateway/src/handlers/admin/billing/wallets/mutations/fail_refund.rs

@@ -84,7 +84,7 @@ pub(in super::super) async fn build_admin_wallet_fail_refund_response(
                             transaction.link_id.as_deref(),
                             transaction.operator_id.as_deref(),
                             transaction.description.as_deref(),
-                            unix_secs_to_rfc3339(transaction.created_at_unix_secs),
+                            unix_secs_to_rfc3339(transaction.created_at_unix_ms),
                         )
                     })
                     .unwrap_or(serde_json::Value::Null),

apps/aether-gateway/src/handlers/admin/billing/wallets/mutations/process_refund.rs

@@ -69,7 +69,7 @@ pub(in super::super) async fn build_admin_wallet_process_refund_response(
                     transaction.link_id.as_deref(),
                     transaction.operator_id.as_deref(),
                     transaction.description.as_deref(),
-                    unix_secs_to_rfc3339(transaction.created_at_unix_secs),
+                    unix_secs_to_rfc3339(transaction.created_at_unix_ms),
                 ),
             }))
             .into_response();

apps/aether-gateway/src/handlers/admin/billing/wallets/mutations/recharge.rs

@@ -87,7 +87,7 @@ pub(in super::super) async fn build_admin_wallet_recharge_response(
             payment_order.amount_usd,
             payment_order.payment_method,
             payment_order.status,
-            unix_secs_to_rfc3339(payment_order.created_at_unix_secs),
+            unix_secs_to_rfc3339(payment_order.created_at_unix_ms),
             payment_order
                 .credited_at_unix_secs
                 .and_then(unix_secs_to_rfc3339),

apps/aether-gateway/src/handlers/admin/billing/wallets/reads/ledger.rs

@@ -69,7 +69,7 @@ pub(in super::super) async fn build_admin_wallet_ledger_response(
                 "operator_name": entry.operator_name,
                 "operator_email": entry.operator_email,
                 "description": entry.description,
-                "created_at": entry.created_at_unix_secs.and_then(unix_secs_to_rfc3339),
+                "created_at": entry.created_at_unix_ms.and_then(unix_secs_to_rfc3339),
             })
         })
         .collect::<Vec<_>>();

apps/aether-gateway/src/handlers/admin/billing/wallets/reads/list.rs

@@ -59,7 +59,7 @@ pub(in super::super) async fn build_admin_wallet_list_response(
                 "total_consumed": wallet.total_consumed,
                 "total_refunded": wallet.total_refunded,
                 "total_adjusted": wallet.total_adjusted,
-                "created_at": wallet.created_at_unix_secs.and_then(unix_secs_to_rfc3339),
+                "created_at": wallet.created_at_unix_ms.and_then(unix_secs_to_rfc3339),
                 "updated_at": wallet.updated_at_unix_secs.and_then(unix_secs_to_rfc3339),
             })
         })

apps/aether-gateway/src/handlers/admin/billing/wallets/reads/refund_requests.rs

@@ -79,7 +79,7 @@ pub(in super::super) async fn build_admin_wallet_refund_requests_response(
             "requested_by": refund.requested_by,
             "approved_by": refund.approved_by,
             "processed_by": refund.processed_by,
-            "created_at": refund.created_at_unix_secs.and_then(unix_secs_to_rfc3339),
+            "created_at": refund.created_at_unix_ms.and_then(unix_secs_to_rfc3339),
             "updated_at": refund.updated_at_unix_secs.and_then(unix_secs_to_rfc3339),
             "processed_at": refund.processed_at_unix_secs.and_then(unix_secs_to_rfc3339),
             "completed_at": refund.completed_at_unix_secs.and_then(unix_secs_to_rfc3339),

apps/aether-gateway/src/handlers/admin/billing/wallets/reads/transactions.rs

@@ -80,7 +80,7 @@ pub(in super::super) async fn build_admin_wallet_transactions_response(
             "operator_name": operator_name,
             "operator_email": operator_email,
             "description": transaction.description,
-            "created_at": transaction.created_at_unix_secs.and_then(unix_secs_to_rfc3339),
+            "created_at": transaction.created_at_unix_ms.and_then(unix_secs_to_rfc3339),
         }));
     }
 

apps/aether-gateway/src/handlers/admin/billing/wallets/shared/payloads.rs

@@ -188,7 +188,7 @@ pub(in super::super) fn build_admin_wallet_refund_payload(
         "requested_by": refund.requested_by.clone(),
         "approved_by": refund.approved_by.clone(),
         "processed_by": refund.processed_by.clone(),
-        "created_at": unix_secs_to_rfc3339(refund.created_at_unix_secs),
+        "created_at": unix_secs_to_rfc3339(refund.created_at_unix_ms),
         "updated_at": unix_secs_to_rfc3339(refund.updated_at_unix_secs),
         "processed_at": refund.processed_at_unix_secs.and_then(unix_secs_to_rfc3339),
         "completed_at": refund.completed_at_unix_secs.and_then(unix_secs_to_rfc3339),

apps/aether-gateway/src/handlers/admin/endpoint/health_builders/status.rs

@@ -3,6 +3,7 @@ use crate::handlers::admin::shared::unix_secs_to_rfc3339;
 use crate::handlers::public::{
     api_format_display_name, build_public_health_timeline, provider_key_api_formats,
 };
+use crate::handlers::shared::unix_ms_to_rfc3339;
 use aether_data_contracts::repository::candidates::PublicHealthTimelineBucket;
 use aether_scheduler_core::{is_provider_key_circuit_open, provider_key_health_score};
 use serde_json::json;
@@ -124,28 +125,24 @@ pub(crate) async fn build_admin_endpoint_health_status_payload(
                 total_count: 0,
                 success_count: 0,
                 failed_count: 0,
-                min_created_at_unix_secs: None,
-                max_created_at_unix_secs: None,
+                min_created_at_unix_ms: None,
+                max_created_at_unix_ms: None,
             });
         bucket.total_count += row.total_count;
         bucket.success_count += row.success_count;
         bucket.failed_count += row.failed_count;
-        bucket.min_created_at_unix_secs = match (
-            bucket.min_created_at_unix_secs,
-            row.min_created_at_unix_secs,
-        ) {
-            (Some(left), Some(right)) => Some(left.min(right)),
-            (None, Some(right)) => Some(right),
-            (left, None) => left,
-        };
-        bucket.max_created_at_unix_secs = match (
-            bucket.max_created_at_unix_secs,
-            row.max_created_at_unix_secs,
-        ) {
-            (Some(left), Some(right)) => Some(left.max(right)),
-            (None, Some(right)) => Some(right),
-            (left, None) => left,
-        };
+        bucket.min_created_at_unix_ms =
+            match (bucket.min_created_at_unix_ms, row.min_created_at_unix_ms) {
+                (Some(left), Some(right)) => Some(left.min(right)),
+                (None, Some(right)) => Some(right),
+                (left, None) => left,
+            };
+        bucket.max_created_at_unix_ms =
+            match (bucket.max_created_at_unix_ms, row.max_created_at_unix_ms) {
+                (Some(left), Some(right)) => Some(left.max(right)),
+                (None, Some(right)) => Some(right),
+                (left, None) => left,
+            };
     }
 
     let mut payload = endpoint_ids_by_format
@@ -182,8 +179,8 @@ pub(crate) async fn build_admin_endpoint_health_status_payload(
                 "display_name": api_format_display_name(&api_format),
                 "health_score": health_score,
                 "timeline": timeline,
-                "time_range_start": time_range_start.and_then(unix_secs_to_rfc3339),
-                "time_range_end": time_range_end.or(Some(now_unix_secs)).and_then(unix_secs_to_rfc3339),
+                "time_range_start": time_range_start.and_then(unix_ms_to_rfc3339),
+                "time_range_end": time_range_end.map(|ms| unix_ms_to_rfc3339(ms)).unwrap_or_else(|| unix_secs_to_rfc3339(now_unix_secs)),
                 "total_endpoints": endpoint_ids.len(),
                 "total_keys": total_keys,
                 "active_keys": active_keys_by_format.get(&api_format).map(BTreeSet::len).unwrap_or(0),

apps/aether-gateway/src/handlers/admin/features/gemini_files/read_routes.rs

@@ -140,7 +140,7 @@ fn build_admin_gemini_file_mapping_payload(
         "username": username,
         "display_name": mapping.display_name,
         "mime_type": mapping.mime_type,
-        "created_at": unix_secs_to_rfc3339(mapping.created_at_unix_secs),
+        "created_at": unix_secs_to_rfc3339(mapping.created_at_unix_ms),
         "expires_at": unix_secs_to_rfc3339(mapping.expires_at_unix_secs),
         "is_expired": mapping.expires_at_unix_secs <= now_unix_secs,
     })

apps/aether-gateway/src/handlers/admin/features/video_tasks/builders.rs

@@ -94,7 +94,7 @@ pub(super) fn build_admin_video_task_list_item(
         "error_message": task.error_message,
         "poll_count": task.poll_count,
         "max_poll_count": task.max_poll_count,
-        "created_at": admin_video_task_timestamp(Some(task.created_at_unix_secs)),
+        "created_at": admin_video_task_timestamp(Some(task.created_at_unix_ms)),
         "completed_at": admin_video_task_timestamp(task.completed_at_unix_secs),
         "submitted_at": admin_video_task_timestamp(task.submitted_at_unix_secs),
     })

apps/aether-gateway/src/handlers/admin/features/video_tasks/routes.rs

@@ -272,7 +272,7 @@ pub(super) async fn maybe_build_local_admin_video_tasks_response(
             payload.insert("max_poll_count".to_string(), json!(task.max_poll_count));
             payload.insert(
                 "created_at".to_string(),
-                json!(admin_video_task_timestamp(Some(task.created_at_unix_secs))),
+                json!(admin_video_task_timestamp(Some(task.created_at_unix_ms))),
             );
             payload.insert(
                 "updated_at".to_string(),

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

@@ -49,6 +49,7 @@ pub(super) async fn build_admin_monitoring_cache_stats_response(
             "affinity_stats": {
                 "storage_type": snapshot.storage_type,
                 "total_affinities": snapshot.total_affinities,
+                "active_affinities": snapshot.total_affinities,
                 "cache_hits": snapshot.cache_hits,
                 "cache_misses": snapshot.cache_misses,
                 "cache_hit_rate": snapshot.cache_hit_rate,

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

@@ -1,6 +1,8 @@
 use super::cache_types::AdminMonitoringCacheAffinityRecord;
+use crate::cache::SchedulerAffinityTarget;
 use crate::handlers::admin::request::AdminAppState;
 use crate::GatewayError;
+use std::time::Duration;
 
 fn parse_admin_monitoring_cache_affinity_key(raw_key: &str) -> Option<(String, String, String)> {
     let parts = raw_key.split(':').collect::<Vec<_>>();
@@ -25,6 +27,48 @@ fn parse_admin_monitoring_cache_affinity_key(raw_key: &str) -> Option<(String, S
     Some((affinity_key.to_string(), api_format, model_name))
 }
 
+fn parse_admin_monitoring_scheduler_affinity_key(
+    raw_key: &str,
+) -> Option<(String, String, String)> {
+    let parts = raw_key.split(':').collect::<Vec<_>>();
+    let start = parts
+        .iter()
+        .position(|segment| *segment == "scheduler_affinity")?;
+    let affinity_key = parts.get(start + 1)?.trim();
+    if affinity_key.is_empty() {
+        return None;
+    }
+
+    let remaining = parts.get(start + 2..)?;
+    if remaining.len() < 2 {
+        return None;
+    }
+
+    let (api_format, model_name_parts) = if remaining.len() == 2 {
+        (remaining[0].trim().to_string(), &remaining[1..])
+    } else {
+        (
+            format!("{}:{}", remaining[0].trim(), remaining[1].trim()),
+            &remaining[2..],
+        )
+    };
+    if api_format.trim().is_empty() {
+        return None;
+    }
+
+    let model_name = model_name_parts
+        .iter()
+        .map(|segment| segment.trim())
+        .filter(|segment| !segment.is_empty())
+        .collect::<Vec<_>>()
+        .join(":");
+    if model_name.is_empty() {
+        return None;
+    }
+
+    Some((affinity_key.to_string(), api_format, model_name))
+}
+
 pub(super) fn admin_monitoring_scheduler_affinity_cache_key(
     record: &AdminMonitoringCacheAffinityRecord,
 ) -> Option<String> {
@@ -92,6 +136,93 @@ pub(super) fn admin_monitoring_cache_affinity_record(
     })
 }
 
+pub(super) fn admin_monitoring_scheduler_affinity_record(
+    cache_key: &str,
+    target: &SchedulerAffinityTarget,
+    age: Duration,
+    ttl: Duration,
+    now_unix_secs: u64,
+) -> Option<AdminMonitoringCacheAffinityRecord> {
+    let (affinity_key, api_format, model_name) =
+        parse_admin_monitoring_scheduler_affinity_key(cache_key)?;
+    let age_secs = age.as_secs();
+    let created_at = now_unix_secs.saturating_sub(age_secs);
+    let expire_at = created_at.saturating_add(ttl.as_secs());
+
+    Some(AdminMonitoringCacheAffinityRecord {
+        raw_key: cache_key.to_string(),
+        affinity_key,
+        api_format,
+        model_name,
+        provider_id: Some(target.provider_id.clone()),
+        endpoint_id: Some(target.endpoint_id.clone()),
+        key_id: Some(target.key_id.clone()),
+        created_at: Some(serde_json::json!(created_at)),
+        expire_at: Some(serde_json::json!(expire_at)),
+        request_count: 0,
+    })
+}
+
+pub(super) fn admin_monitoring_scheduler_affinity_record_from_raw(
+    raw_key: &str,
+    raw_value: &str,
+) -> Option<AdminMonitoringCacheAffinityRecord> {
+    let payload = serde_json::from_str::<serde_json::Value>(raw_value).ok()?;
+    let object = payload.as_object()?;
+    let (affinity_key, parsed_api_format, parsed_model_name) =
+        parse_admin_monitoring_scheduler_affinity_key(raw_key)?;
+    let api_format = object
+        .get("api_format")
+        .and_then(serde_json::Value::as_str)
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+        .unwrap_or(parsed_api_format.as_str())
+        .to_string();
+    let model_name = object
+        .get("model_name")
+        .and_then(serde_json::Value::as_str)
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+        .unwrap_or(parsed_model_name.as_str())
+        .to_string();
+    let request_count = object
+        .get("request_count")
+        .and_then(|value| {
+            value
+                .as_u64()
+                .or_else(|| value.as_i64().and_then(|number| u64::try_from(number).ok()))
+        })
+        .unwrap_or(0);
+
+    Some(AdminMonitoringCacheAffinityRecord {
+        raw_key: raw_key.to_string(),
+        affinity_key,
+        api_format,
+        model_name,
+        provider_id: object
+            .get("provider_id")
+            .and_then(serde_json::Value::as_str)
+            .map(ToOwned::to_owned),
+        endpoint_id: object
+            .get("endpoint_id")
+            .and_then(serde_json::Value::as_str)
+            .map(ToOwned::to_owned),
+        key_id: object
+            .get("key_id")
+            .and_then(serde_json::Value::as_str)
+            .map(ToOwned::to_owned),
+        created_at: object.get("created_at").cloned(),
+        expire_at: object.get("expire_at").cloned(),
+        request_count,
+    })
+}
+
+pub(super) fn admin_monitoring_cache_affinity_record_identity(
+    record: &AdminMonitoringCacheAffinityRecord,
+) -> String {
+    admin_monitoring_scheduler_affinity_cache_key(record).unwrap_or_else(|| record.raw_key.clone())
+}
+
 pub(super) fn clear_admin_monitoring_scheduler_affinity_entries(
     state: &AdminAppState<'_>,
     records: &[AdminMonitoringCacheAffinityRecord],

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

@@ -8,6 +8,7 @@ use super::super::cache_route_helpers::{
     admin_monitoring_cache_affinity_unavailable_response,
 };
 use super::super::cache_store::{
+    admin_monitoring_has_runtime_scheduler_affinity_entries,
     list_admin_monitoring_cache_affinity_records_by_affinity_keys,
     load_admin_monitoring_cache_affinity_entries_for_tests,
 };
@@ -33,6 +34,7 @@ pub(in super::super) async fn build_admin_monitoring_cache_affinity_delete_respo
 
     if state.redis_kv_runner().is_none()
         && load_admin_monitoring_cache_affinity_entries_for_tests(state).is_empty()
+        && !admin_monitoring_has_runtime_scheduler_affinity_entries(state)
     {
         return Ok(admin_monitoring_cache_affinity_unavailable_response());
     }

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

@@ -11,6 +11,7 @@ use super::super::cache_route_helpers::{
     admin_monitoring_cache_users_user_identifier_from_path,
 };
 use super::super::cache_store::{
+    admin_monitoring_has_runtime_scheduler_affinity_entries,
     list_admin_monitoring_cache_affinity_records_by_affinity_keys,
     load_admin_monitoring_cache_affinity_entries_for_tests,
 };
@@ -37,6 +38,7 @@ pub(in super::super) async fn build_admin_monitoring_cache_users_delete_response
 
     if state.redis_kv_runner().is_none()
         && load_admin_monitoring_cache_affinity_entries_for_tests(state).is_empty()
+        && !admin_monitoring_has_runtime_scheduler_affinity_entries(state)
     {
         return Ok(admin_monitoring_cache_affinity_unavailable_response());
     }

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

@@ -1,41 +1,25 @@
-use super::cache_affinity::admin_monitoring_cache_affinity_record;
+use super::cache_affinity::{
+    admin_monitoring_cache_affinity_record, admin_monitoring_cache_affinity_record_identity,
+    admin_monitoring_scheduler_affinity_record,
+    admin_monitoring_scheduler_affinity_record_from_raw,
+};
 use super::cache_types::{AdminMonitoringCacheAffinityRecord, AdminMonitoringCacheSnapshot};
 use crate::handlers::admin::observability::stats::round_to;
 use crate::handlers::admin::request::AdminAppState;
+use crate::scheduler::affinity::SCHEDULER_AFFINITY_TTL;
 use crate::GatewayError;
 use aether_data_contracts::repository::usage::UsageAuditListQuery;
 
 async fn count_admin_monitoring_cache_affinity_entries(state: &AdminAppState<'_>) -> usize {
-    let Some(runner) = state.redis_kv_runner() else {
-        return 0;
-    };
-    let mut connection = match runner.client().get_multiplexed_async_connection().await {
-        Ok(value) => value,
-        Err(_) => return 0,
-    };
-    let pattern = runner.keyspace().key("cache_affinity:*");
-    let mut cursor = 0u64;
-    let mut total = 0usize;
-    loop {
-        let (next_cursor, keys) = match redis::cmd("SCAN")
-            .arg(cursor)
-            .arg("MATCH")
-            .arg(&pattern)
-            .arg("COUNT")
-            .arg(200)
-            .query_async::<(u64, Vec<String>)>(&mut connection)
-            .await
-        {
-            Ok(value) => value,
-            Err(_) => return total,
-        };
-        total += keys.len();
-        if next_cursor == 0 {
-            break;
-        }
-        cursor = next_cursor;
-    }
-    total
+    list_admin_monitoring_cache_affinity_records(state)
+        .await
+        .map(|items| items.len())
+        .unwrap_or_else(|_| {
+            state
+                .as_ref()
+                .list_scheduler_affinity_entries(SCHEDULER_AFFINITY_TTL)
+                .len()
+        })
 }
 
 async fn scan_admin_monitoring_namespaced_keys(
@@ -191,12 +175,27 @@ pub(super) async fn list_admin_monitoring_cache_affinity_records_by_affinity_key
     list_admin_monitoring_cache_affinity_records_matching(state, Some(affinity_keys)).await
 }
 
+pub(super) fn admin_monitoring_has_runtime_scheduler_affinity_entries(
+    state: &AdminAppState<'_>,
+) -> bool {
+    !state
+        .as_ref()
+        .list_scheduler_affinity_entries(SCHEDULER_AFFINITY_TTL)
+        .is_empty()
+}
+
 async fn list_admin_monitoring_cache_affinity_records_matching(
     state: &AdminAppState<'_>,
     affinity_keys: Option<&std::collections::BTreeSet<String>>,
 ) -> Result<Vec<AdminMonitoringCacheAffinityRecord>, GatewayError> {
     let mut records = Vec::new();
-    let mut seen_raw_keys = std::collections::BTreeSet::new();
+    let mut seen_record_ids = std::collections::BTreeSet::new();
+
+    let mut push_record = |record: AdminMonitoringCacheAffinityRecord| {
+        if seen_record_ids.insert(admin_monitoring_cache_affinity_record_identity(&record)) {
+            records.push(record);
+        }
+    };
 
     if let Some(runner) = state.redis_kv_runner() {
         let mut connection = runner
@@ -209,14 +208,24 @@ async fn list_admin_monitoring_cache_affinity_records_matching(
         let patterns = affinity_keys
             .map(|keys| {
                 keys.iter()
-                    .map(|affinity_key| {
-                        runner
-                            .keyspace()
-                            .key(&format!("cache_affinity:{affinity_key}:*"))
+                    .flat_map(|affinity_key| {
+                        [
+                            runner
+                                .keyspace()
+                                .key(&format!("cache_affinity:{affinity_key}:*")),
+                            runner
+                                .keyspace()
+                                .key(&format!("scheduler_affinity:{affinity_key}:*")),
+                        ]
                     })
                     .collect::<Vec<_>>()
             })
-            .unwrap_or_else(|| vec![runner.keyspace().key("cache_affinity:*")]);
+            .unwrap_or_else(|| {
+                vec![
+                    runner.keyspace().key("cache_affinity:*"),
+                    runner.keyspace().key("scheduler_affinity:*"),
+                ]
+            });
 
         for pattern in patterns {
             let mut cursor = 0u64;
@@ -246,16 +255,18 @@ async fn list_admin_monitoring_cache_affinity_records_matching(
                         let Some(raw_value) = raw_value else {
                             continue;
                         };
-                        let Some(record) = admin_monitoring_cache_affinity_record(&key, &raw_value)
-                        else {
+                        let record = if key.contains("scheduler_affinity:") {
+                            admin_monitoring_scheduler_affinity_record_from_raw(&key, &raw_value)
+                        } else {
+                            admin_monitoring_cache_affinity_record(&key, &raw_value)
+                        };
+                        let Some(record) = record else {
                             continue;
                         };
                         if affinity_keys.is_some_and(|keys| !keys.contains(&record.affinity_key)) {
                             continue;
                         }
-                        if seen_raw_keys.insert(record.raw_key.clone()) {
-                            records.push(record);
-                        }
+                        push_record(record);
                     }
                 }
                 if next_cursor == 0 {
@@ -264,7 +275,6 @@ async fn list_admin_monitoring_cache_affinity_records_matching(
                 cursor = next_cursor;
             }
         }
-        return Ok(records);
     }
 
     for (key, raw_value) in load_admin_monitoring_cache_affinity_entries_for_tests(state) {
@@ -274,9 +284,27 @@ async fn list_admin_monitoring_cache_affinity_records_matching(
         if affinity_keys.is_some_and(|keys| !keys.contains(&record.affinity_key)) {
             continue;
         }
-        if seen_raw_keys.insert(record.raw_key.clone()) {
-            records.push(record);
+        push_record(record);
+    }
+
+    let now_unix_secs = chrono::Utc::now().timestamp().max(0) as u64;
+    for entry in state
+        .as_ref()
+        .list_scheduler_affinity_entries(SCHEDULER_AFFINITY_TTL)
+    {
+        let Some(record) = admin_monitoring_scheduler_affinity_record(
+            &entry.cache_key,
+            &entry.target,
+            entry.age,
+            SCHEDULER_AFFINITY_TTL,
+            now_unix_secs,
+        ) else {
+            continue;
+        };
+        if affinity_keys.is_some_and(|keys| !keys.contains(&record.affinity_key)) {
+            continue;
         }
+        push_record(record);
     }
 
     Ok(records)

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

@@ -161,7 +161,7 @@ pub(super) async fn build_admin_monitoring_resilience_snapshot(
         .filter(admin_monitoring_usage_is_error)
         .collect::<Vec<_>>();
     recent_usage_errors
-        .sort_by(|left, right| right.created_at_unix_secs.cmp(&left.created_at_unix_secs));
+        .sort_by(|left, right| right.created_at_unix_ms.cmp(&left.created_at_unix_ms));
 
     let total_errors = recent_usage_errors.len();
     let mut error_breakdown = BTreeMap::<String, usize>::new();
@@ -200,7 +200,7 @@ pub(super) async fn build_admin_monitoring_resilience_snapshot(
                 "error_id": item.id,
                 "error_type": error_type,
                 "operation": operation,
-                "timestamp": unix_secs_to_rfc3339(item.created_at_unix_secs),
+                "timestamp": unix_secs_to_rfc3339(item.created_at_unix_ms),
                 "context": {
                     "request_id": item.request_id,
                     "provider_id": item.provider_id,

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

@@ -31,7 +31,7 @@ pub(super) fn sample_usage(
     total_cost_usd: f64,
     status: &str,
     status_code: Option<i32>,
-    created_at_unix_secs: i64,
+    created_at_unix_ms: i64,
 ) -> StoredRequestUsageAudit {
     let is_error = status_code.is_some_and(|value| value >= 400)
         || status.trim().eq_ignore_ascii_case("failed")
@@ -70,9 +70,9 @@ pub(super) fn sample_usage(
         Some(30),
         status.to_string(),
         "billed".to_string(),
-        created_at_unix_secs,
-        created_at_unix_secs,
-        Some(created_at_unix_secs),
+        created_at_unix_ms,
+        created_at_unix_ms,
+        Some(created_at_unix_ms),
     )
     .expect("usage should build")
 }
@@ -82,7 +82,7 @@ pub(super) fn sample_candidate(
     request_id: &str,
     candidate_index: i32,
     status: RequestCandidateStatus,
-    started_at_unix_secs: Option<i64>,
+    started_at_unix_ms: Option<i64>,
     latency_ms: Option<i32>,
     status_code: Option<i32>,
 ) -> StoredRequestCandidate {
@@ -108,9 +108,9 @@ pub(super) fn sample_candidate(
         Some(1),
         None,
         Some(json!({"cache_1h": true})),
-        100 + i64::from(candidate_index),
-        started_at_unix_secs,
-        started_at_unix_secs.map(|value| value + 1),
+        (100 + i64::from(candidate_index)) * 1_000,
+        started_at_unix_ms.map(|v| v * 1_000),
+        started_at_unix_ms.map(|value| (value + 1) * 1_000),
     )
     .expect("candidate should build")
 }

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

@@ -258,6 +258,56 @@ async fn admin_monitoring_resilience_status_returns_local_payload() {
     assert!(payload["timestamp"].as_str().is_some());
 }
 
+#[tokio::test]
+async fn admin_monitoring_cache_stats_count_runtime_scheduler_affinities() {
+    let state = AppState::new().expect("state should build");
+    let affinity_cache_key =
+        aether_scheduler_core::build_scheduler_affinity_cache_key_for_api_key_id(
+            "user-key-1",
+            "openai:chat",
+            "model-alpha",
+        )
+        .expect("scheduler affinity cache key should build");
+    state.scheduler_affinity_cache.insert(
+        affinity_cache_key,
+        crate::cache::SchedulerAffinityTarget {
+            provider_id: "provider-1".to_string(),
+            endpoint_id: "endpoint-1".to_string(),
+            key_id: "provider-key-1".to_string(),
+        },
+        crate::scheduler::affinity::SCHEDULER_AFFINITY_TTL,
+        128,
+    );
+
+    let response = local_monitoring_response(
+        &state,
+        &request_context(http::Method::GET, "/api/admin/monitoring/cache/stats"),
+    )
+    .await
+    .expect("handler should not error")
+    .expect("monitoring 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");
+    assert_eq!(payload["status"], json!("ok"));
+    assert_eq!(payload["data"]["total_affinities"], json!(1));
+    assert_eq!(
+        payload["data"]["affinity_stats"]["total_affinities"],
+        json!(1)
+    );
+    assert_eq!(
+        payload["data"]["affinity_stats"]["active_affinities"],
+        json!(1)
+    );
+    assert_eq!(
+        payload["data"]["affinity_stats"]["storage_type"],
+        json!("memory")
+    );
+}
+
 #[tokio::test]
 async fn admin_monitoring_cache_stats_returns_local_payload() {
     let now = chrono::Utc::now().timestamp();

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

@@ -209,6 +209,136 @@ async fn admin_monitoring_cache_affinities_and_affinity_return_local_payload_fro
     assert_eq!(detail_payload["total_endpoints"], json!(1));
 }
 
+#[tokio::test]
+async fn admin_monitoring_cache_affinities_and_delete_use_runtime_scheduler_affinity_cache() {
+    let user_repository = Arc::new(
+        InMemoryUserReadRepository::seed_auth_users(vec![sample_monitoring_auth_user("user-1")])
+            .with_export_users(vec![sample_monitoring_export_user("user-1")]),
+    );
+    let auth_repository = Arc::new(
+        InMemoryAuthApiKeySnapshotRepository::default().with_export_records(vec![
+            sample_monitoring_export_api_key("user-1", "user-key-1"),
+        ]),
+    );
+    let provider_catalog = Arc::new(InMemoryProviderCatalogReadRepository::seed(
+        vec![sample_provider()],
+        vec![sample_monitoring_catalog_endpoint()],
+        vec![sample_monitoring_catalog_key()],
+    ));
+    let state = AppState::new()
+        .expect("state should build")
+        .with_data_state_for_tests(
+            crate::data::GatewayDataState::with_provider_catalog_reader_for_tests(provider_catalog)
+                .with_user_reader(user_repository)
+                .with_auth_api_key_reader(auth_repository),
+        );
+    let affinity_cache_key =
+        aether_scheduler_core::build_scheduler_affinity_cache_key_for_api_key_id(
+            "user-key-1",
+            "openai:chat",
+            "model-alpha",
+        )
+        .expect("scheduler affinity cache key should build");
+    state.scheduler_affinity_cache.insert(
+        affinity_cache_key.clone(),
+        crate::cache::SchedulerAffinityTarget {
+            provider_id: "provider-1".to_string(),
+            endpoint_id: "endpoint-1".to_string(),
+            key_id: "provider-key-1".to_string(),
+        },
+        crate::scheduler::affinity::SCHEDULER_AFFINITY_TTL,
+        128,
+    );
+
+    let list_context = request_context(
+        http::Method::GET,
+        "/api/admin/monitoring/cache/affinities?keyword=alice&limit=20&offset=0",
+    );
+    let list_response = local_monitoring_response(&state, &list_context)
+        .await
+        .expect("handler should not error")
+        .expect("route should be handled locally");
+    assert_eq!(list_response.status(), http::StatusCode::OK);
+    let list_body = to_bytes(list_response.into_body(), usize::MAX)
+        .await
+        .expect("body should read");
+    let list_payload: serde_json::Value =
+        serde_json::from_slice(&list_body).expect("json body should parse");
+    assert_eq!(list_payload["status"], json!("ok"));
+    assert_eq!(list_payload["data"]["meta"]["total"], json!(1));
+    assert_eq!(list_payload["data"]["matched_user_id"], json!("user-1"));
+    assert_eq!(
+        list_payload["data"]["items"][0]["affinity_key"],
+        json!("user-key-1")
+    );
+    assert_eq!(
+        list_payload["data"]["items"][0]["api_format"],
+        json!("openai:chat")
+    );
+    assert_eq!(
+        list_payload["data"]["items"][0]["provider_name"],
+        json!("OpenAI")
+    );
+    assert_eq!(
+        list_payload["data"]["items"][0]["endpoint_url"],
+        json!("https://api.openai.example/v1")
+    );
+    assert_eq!(list_payload["data"]["items"][0]["request_count"], json!(0));
+    assert!(list_payload["data"]["items"][0]["expire_at"]
+        .as_u64()
+        .is_some_and(|value| value > 0));
+
+    let detail_context = request_context(
+        http::Method::GET,
+        "/api/admin/monitoring/cache/affinity/alice",
+    );
+    let detail_response = local_monitoring_response(&state, &detail_context)
+        .await
+        .expect("handler should not error")
+        .expect("route should be handled locally");
+    assert_eq!(detail_response.status(), http::StatusCode::OK);
+    let detail_body = to_bytes(detail_response.into_body(), usize::MAX)
+        .await
+        .expect("body should read");
+    let detail_payload: serde_json::Value =
+        serde_json::from_slice(&detail_body).expect("json body should parse");
+    assert_eq!(detail_payload["status"], json!("ok"));
+    assert_eq!(
+        detail_payload["affinities"][0]["api_format"],
+        json!("openai:chat")
+    );
+    assert_eq!(detail_payload["total_endpoints"], json!(1));
+
+    let delete_response = local_monitoring_response(
+        &state,
+        &request_context(
+            http::Method::DELETE,
+            "/api/admin/monitoring/cache/affinity/user-key-1/endpoint-1/model-alpha/openai:chat",
+        ),
+    )
+    .await
+    .expect("handler should not error")
+    .expect("route should be handled locally");
+    assert_eq!(delete_response.status(), http::StatusCode::OK);
+    let delete_body = to_bytes(delete_response.into_body(), usize::MAX)
+        .await
+        .expect("body should read");
+    let delete_payload: serde_json::Value =
+        serde_json::from_slice(&delete_body).expect("json body should parse");
+    assert_eq!(
+        delete_payload["message"],
+        json!("已清除缓存亲和性: Alice Key")
+    );
+    assert_eq!(delete_payload["affinity_key"], json!("user-key-1"));
+    assert_eq!(
+        state.read_scheduler_affinity_target(
+            &affinity_cache_key,
+            crate::scheduler::affinity::SCHEDULER_AFFINITY_TTL,
+        ),
+        None
+    );
+}
+
 #[tokio::test]
 async fn admin_monitoring_cache_users_delete_returns_local_payload_from_test_store() {
     let user_repository = Arc::new(

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

@@ -13,7 +13,7 @@ use axum::{
     body::Body,
     response::{IntoResponse, Response},
 };
-use tracing::warn;
+use tracing::debug;
 
 pub(super) async fn build_admin_monitoring_trace_request_response(
     state: &AdminAppState<'_>,
@@ -38,7 +38,7 @@ pub(super) async fn build_admin_monitoring_trace_request_response(
         .await
         .map_err(|err| GatewayError::Internal(err.to_string()))?
     else {
-        warn!(
+        debug!(
             event_name = "admin_monitoring_request_trace_not_found",
             log_type = "admin_monitoring",
             request_id = %short_request_id(request_id.as_str()),

apps/aether-gateway/src/handlers/admin/observability/usage/analytics/cache_affinity.rs

@@ -20,8 +20,8 @@ pub(in super::super) async fn list_recent_completed_usage_for_cache_affinity(
         .await?;
     items.retain(|item| item.status == "completed");
     items.sort_by(|left, right| {
-        left.created_at_unix_secs
-            .cmp(&right.created_at_unix_secs)
+        left.created_at_unix_ms
+            .cmp(&right.created_at_unix_ms)
             .then_with(|| left.id.cmp(&right.id))
     });
     Ok(items)

apps/aether-gateway/src/handlers/admin/observability/usage/analytics_routes/cache_affinity_interval_timeline.rs

@@ -50,14 +50,14 @@ pub(super) async fn build_admin_usage_cache_affinity_interval_timeline_response(
             }
         }
 
-        let mut previous_created_at_unix_secs = None;
+        let mut previous_created_at_unix_ms = None;
         for item in items {
-            if let Some(previous) = previous_created_at_unix_secs {
+            if let Some(previous) = previous_created_at_unix_ms {
                 let interval_minutes =
-                    item.created_at_unix_secs.saturating_sub(previous) as f64 / 60.0;
+                    item.created_at_unix_ms.saturating_sub(previous) as f64 / 60.0;
                 if interval_minutes <= 120.0 {
                     let mut point = json!({
-                        "x": unix_secs_to_rfc3339(item.created_at_unix_secs),
+                        "x": unix_secs_to_rfc3339(item.created_at_unix_ms),
                         "y": ((interval_minutes * 100.0).round()) / 100.0,
                     });
                     if !item.model.trim().is_empty() {
@@ -78,7 +78,7 @@ pub(super) async fn build_admin_usage_cache_affinity_interval_timeline_response(
                         .push(point);
                 }
             }
-            previous_created_at_unix_secs = Some(item.created_at_unix_secs);
+            previous_created_at_unix_ms = Some(item.created_at_unix_ms);
         }
     }
 

apps/aether-gateway/src/handlers/admin/observability/usage/summary_routes.rs

@@ -78,8 +78,8 @@ pub(super) async fn maybe_build_local_admin_usage_summary_response(
                 .collect();
             items.sort_by(|left, right| {
                 right
-                    .created_at_unix_secs
-                    .cmp(&left.created_at_unix_secs)
+                    .created_at_unix_ms
+                    .cmp(&left.created_at_unix_ms)
                     .then_with(|| left.id.cmp(&right.id))
             });
             if requested_ids.is_none() && items.len() > 50 {
@@ -147,8 +147,8 @@ pub(super) async fn maybe_build_local_admin_usage_summary_response(
             });
             usage.sort_by(|left, right| {
                 right
-                    .created_at_unix_secs
-                    .cmp(&left.created_at_unix_secs)
+                    .created_at_unix_ms
+                    .cmp(&left.created_at_unix_ms)
                     .then_with(|| left.id.cmp(&right.id))
             });
             let total = usage.len();

apps/aether-gateway/src/handlers/admin/provider/endpoints_admin/reads.rs

@@ -28,9 +28,9 @@ pub(crate) async fn build_admin_provider_endpoints_payload(
         .unwrap_or_default();
     endpoints.sort_by(|left, right| {
         right
-            .created_at_unix_secs
+            .created_at_unix_ms
             .unwrap_or_default()
-            .cmp(&left.created_at_unix_secs.unwrap_or_default())
+            .cmp(&left.created_at_unix_ms.unwrap_or_default())
             .then_with(|| left.id.cmp(&right.id))
     });
     let keys = state

apps/aether-gateway/src/handlers/admin/provider/oauth/dispatch/device/authorize.rs

@@ -165,7 +165,7 @@ pub(super) async fn handle_admin_provider_oauth_device_authorize(
             .map(str::trim)
             .filter(|value| !value.is_empty())
             .map(ToOwned::to_owned),
-        created_at_unix_secs: now_unix_secs,
+        created_at_unix_ms: now_unix_secs,
         key_id: None,
         email: None,
         replaced: false,

apps/aether-gateway/src/handlers/admin/provider/oauth/provisioning.rs

@@ -128,7 +128,7 @@ pub(crate) async fn create_provider_oauth_catalog_key(
     record.total_response_time_ms = Some(0);
     record.health_by_format = Some(json!({}));
     record.circuit_breaker_by_format = Some(json!({}));
-    record.created_at_unix_secs = Some(now_unix_secs);
+    record.created_at_unix_ms = Some(now_unix_secs);
     record.updated_at_unix_secs = Some(now_unix_secs);
     state.create_provider_catalog_key(&record).await
 }

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

@@ -803,7 +803,7 @@ pub(super) fn build_admin_pool_key_payload(
     );
     payload.insert(
         "created_at".to_string(),
-        json!(key.created_at_unix_secs.and_then(unix_secs_to_rfc3339)),
+        json!(key.created_at_unix_ms.and_then(unix_secs_to_rfc3339)),
     );
     payload.insert(
         "last_used_at".to_string(),

apps/aether-gateway/src/handlers/admin/provider/summary/aggregates.rs

@@ -177,7 +177,7 @@ pub(crate) async fn build_admin_providers_summary_payload(
             .is_active
             .cmp(&left.is_active)
             .then_with(|| left.provider_priority.cmp(&right.provider_priority))
-            .then_with(|| left.created_at_unix_secs.cmp(&right.created_at_unix_secs))
+            .then_with(|| left.created_at_unix_ms.cmp(&right.created_at_unix_ms))
     });
 
     let total = providers.len();

apps/aether-gateway/src/handlers/admin/provider/summary/health.rs

@@ -1,6 +1,7 @@
 use crate::handlers::admin::request::AdminAppState;
 use crate::handlers::admin::shared::unix_secs_to_rfc3339;
-use crate::handlers::public::{request_candidate_event_unix_secs, request_candidate_status_label};
+use crate::handlers::public::{request_candidate_event_unix_ms, request_candidate_status_label};
+use crate::handlers::shared::unix_ms_to_rfc3339;
 use aether_data_contracts::repository::candidates::{
     RequestCandidateStatus, StoredRequestCandidate,
 };
@@ -82,14 +83,14 @@ pub(crate) async fn build_admin_provider_health_monitor_payload(
     for candidates in attempts_by_endpoint.values_mut() {
         candidates.sort_by(|left, right| {
             right
-                .created_at_unix_secs
-                .cmp(&left.created_at_unix_secs)
+                .created_at_unix_ms
+                .cmp(&left.created_at_unix_ms)
                 .then_with(|| right.id.cmp(&left.id))
         });
         candidates.truncate(per_endpoint_limit);
         candidates.sort_by(|left, right| {
-            request_candidate_event_unix_secs(left)
-                .cmp(&request_candidate_event_unix_secs(right))
+            request_candidate_event_unix_ms(left)
+                .cmp(&request_candidate_event_unix_ms(right))
                 .then_with(|| left.id.cmp(&right.id))
         });
     }
@@ -118,12 +119,12 @@ pub(crate) async fn build_admin_provider_health_monitor_payload(
             };
             let last_event_at = candidates
                 .last()
-                .and_then(|candidate| unix_secs_to_rfc3339(request_candidate_event_unix_secs(candidate)));
+                .and_then(|candidate| unix_ms_to_rfc3339(request_candidate_event_unix_ms(candidate)));
             let events = candidates
                 .into_iter()
                 .filter_map(|candidate| {
                     Some(json!({
-                        "timestamp": unix_secs_to_rfc3339(request_candidate_event_unix_secs(&candidate))?,
+                        "timestamp": unix_ms_to_rfc3339(request_candidate_event_unix_ms(&candidate))?,
                         "status": request_candidate_status_label(candidate.status),
                         "status_code": candidate.status_code,
                         "latency_ms": candidate.latency_ms,

apps/aether-gateway/src/handlers/admin/provider/summary/list.rs

@@ -88,7 +88,7 @@ pub(crate) async fn build_admin_providers_payload(
                     "api_key": has_any_key_by_provider.contains(&provider_id).then_some("***"),
                     "priority": provider.provider_priority,
                     "is_active": provider.is_active,
-                    "created_at": provider.created_at_unix_secs.and_then(unix_secs_to_rfc3339),
+                    "created_at": provider.created_at_unix_ms.and_then(unix_secs_to_rfc3339),
                     "updated_at": provider.updated_at_unix_secs.and_then(unix_secs_to_rfc3339),
                 })
             })

apps/aether-gateway/src/handlers/admin/provider/summary/value.rs

@@ -1,6 +1,6 @@
 use crate::handlers::admin::shared::unix_secs_to_rfc3339;
 use crate::handlers::public::{
-    provider_key_api_formats, request_candidate_event_unix_secs, request_candidate_status_label,
+    provider_key_api_formats, request_candidate_event_unix_ms, request_candidate_status_label,
 };
 use aether_data_contracts::repository::candidates::{
     RequestCandidateStatus, StoredRequestCandidate,
@@ -172,7 +172,7 @@ pub(crate) fn build_admin_provider_summary_value(
         "endpoint_health_details": endpoint_health_details,
         "ops_configured": ops_configured,
         "ops_architecture_id": ops_architecture_id,
-        "created_at": endpoint_timestamp_or_now(provider.created_at_unix_secs, now_unix_secs),
+        "created_at": endpoint_timestamp_or_now(provider.created_at_unix_ms, now_unix_secs),
         "updated_at": endpoint_timestamp_or_now(provider.updated_at_unix_secs, now_unix_secs),
     })
 }

apps/aether-gateway/src/handlers/admin/provider/write/keys/create.rs

@@ -176,7 +176,7 @@ pub(crate) async fn build_admin_create_provider_key_record(
         normalize_string_list(payload.model_exclude_patterns).map(|value| json!(value));
     key.health_by_format = Some(json!({}));
     key.circuit_breaker_by_format = Some(json!({}));
-    key.created_at_unix_secs = Some(now_unix_secs);
+    key.created_at_unix_ms = Some(now_unix_secs);
     key.updated_at_unix_secs = Some(now_unix_secs);
     Ok(key)
 }

apps/aether-gateway/src/handlers/admin/provider/write/keys/payload.rs

@@ -24,9 +24,9 @@ pub(crate) async fn build_admin_provider_keys_payload(
         left.internal_priority
             .cmp(&right.internal_priority)
             .then_with(|| {
-                left.created_at_unix_secs
+                left.created_at_unix_ms
                     .unwrap_or_default()
-                    .cmp(&right.created_at_unix_secs.unwrap_or_default())
+                    .cmp(&right.created_at_unix_ms.unwrap_or_default())
             })
             .then_with(|| left.id.cmp(&right.id))
     });

apps/aether-gateway/src/handlers/admin/system/shared/export/providers.rs

@@ -82,9 +82,9 @@ pub(crate) async fn build_admin_system_export_providers_payload(
                 left.internal_priority
                     .cmp(&right.internal_priority)
                     .then(
-                        left.created_at_unix_secs
+                        left.created_at_unix_ms
                             .unwrap_or(0)
-                            .cmp(&right.created_at_unix_secs.unwrap_or(0)),
+                            .cmp(&right.created_at_unix_ms.unwrap_or(0)),
                     )
                     .then(left.id.cmp(&right.id))
             });

apps/aether-gateway/src/handlers/internal/gateway_helpers.rs

@@ -463,7 +463,7 @@ pub(crate) fn build_management_token_payload(
         "last_used_ip": token.last_used_ip,
         "usage_count": token.usage_count,
         "is_active": token.is_active,
-        "created_at": token.created_at_unix_secs.and_then(unix_secs_to_rfc3339),
+        "created_at": token.created_at_unix_ms.and_then(unix_secs_to_rfc3339),
         "updated_at": token.updated_at_unix_secs.and_then(unix_secs_to_rfc3339),
     });
     if let Some(user) = user {

apps/aether-gateway/src/handlers/public/catalog_helpers.rs

@@ -1,5 +1,7 @@
 use crate::api::ai::public_api_format_local_path;
-use crate::handlers::shared::{query_param_optional_bool, query_param_value, unix_secs_to_rfc3339};
+use crate::handlers::shared::{
+    query_param_optional_bool, query_param_value, unix_ms_to_rfc3339, unix_secs_to_rfc3339,
+};
 use crate::AppState;
 use aether_data_contracts::repository::candidates::{
     PublicHealthTimelineBucket, RequestCandidateStatus, StoredRequestCandidate,
@@ -25,11 +27,11 @@ pub(crate) fn request_candidate_status_label(status: RequestCandidateStatus) ->
     }
 }
 
-pub(crate) fn request_candidate_event_unix_secs(candidate: &StoredRequestCandidate) -> u64 {
+pub(crate) fn request_candidate_event_unix_ms(candidate: &StoredRequestCandidate) -> u64 {
     candidate
-        .finished_at_unix_secs
-        .or(candidate.started_at_unix_secs)
-        .unwrap_or(candidate.created_at_unix_secs)
+        .finished_at_unix_ms
+        .or(candidate.started_at_unix_ms)
+        .unwrap_or(candidate.created_at_unix_ms)
 }
 
 pub(crate) fn normalize_admin_base_url(base_url: &str) -> Result<String, String> {
@@ -410,28 +412,24 @@ pub(crate) async fn build_api_format_health_monitor_payload(
                 total_count: 0,
                 success_count: 0,
                 failed_count: 0,
-                min_created_at_unix_secs: None,
-                max_created_at_unix_secs: None,
+                min_created_at_unix_ms: None,
+                max_created_at_unix_ms: None,
             });
         bucket.total_count += row.total_count;
         bucket.success_count += row.success_count;
         bucket.failed_count += row.failed_count;
-        bucket.min_created_at_unix_secs = match (
-            bucket.min_created_at_unix_secs,
-            row.min_created_at_unix_secs,
-        ) {
-            (Some(left), Some(right)) => Some(left.min(right)),
-            (None, Some(right)) => Some(right),
-            (left, None) => left,
-        };
-        bucket.max_created_at_unix_secs = match (
-            bucket.max_created_at_unix_secs,
-            row.max_created_at_unix_secs,
-        ) {
-            (Some(left), Some(right)) => Some(left.max(right)),
-            (None, Some(right)) => Some(right),
-            (left, None) => left,
-        };
+        bucket.min_created_at_unix_ms =
+            match (bucket.min_created_at_unix_ms, row.min_created_at_unix_ms) {
+                (Some(left), Some(right)) => Some(left.min(right)),
+                (None, Some(right)) => Some(right),
+                (left, None) => left,
+            };
+        bucket.max_created_at_unix_ms =
+            match (bucket.max_created_at_unix_ms, row.max_created_at_unix_ms) {
+                (Some(left), Some(right)) => Some(left.max(right)),
+                (None, Some(right)) => Some(right),
+                (left, None) => left,
+            };
     }
 
     let mut formats = Vec::new();
@@ -456,19 +454,19 @@ pub(crate) async fn build_api_format_health_monitor_payload(
         };
         let last_event_at = attempts.first().and_then(|candidate| {
             candidate
-                .finished_at_unix_secs
-                .or(candidate.started_at_unix_secs)
-                .or(Some(candidate.created_at_unix_secs))
+                .finished_at_unix_ms
+                .or(candidate.started_at_unix_ms)
+                .or(Some(candidate.created_at_unix_ms))
         });
         let events = attempts
             .into_iter()
             .filter_map(|candidate| {
                 let timestamp = candidate
-                    .finished_at_unix_secs
-                    .or(candidate.started_at_unix_secs)
-                    .unwrap_or(candidate.created_at_unix_secs);
+                    .finished_at_unix_ms
+                    .or(candidate.started_at_unix_ms)
+                    .unwrap_or(candidate.created_at_unix_ms);
                 Some(json!({
-                    "timestamp": unix_secs_to_rfc3339(timestamp)?,
+                    "timestamp": unix_ms_to_rfc3339(timestamp)?,
                     "status": request_candidate_status_label(candidate.status),
                     "status_code": candidate.status_code,
                     "latency_ms": candidate.latency_ms,
@@ -490,11 +488,11 @@ pub(crate) async fn build_api_format_health_monitor_payload(
             "failed_count": failed_count,
             "skipped_count": skipped_count,
             "success_rate": success_rate,
-            "last_event_at": last_event_at.and_then(unix_secs_to_rfc3339),
+            "last_event_at": last_event_at.and_then(unix_ms_to_rfc3339),
             "events": events,
             "timeline": timeline,
-            "time_range_start": time_range_start.and_then(unix_secs_to_rfc3339),
-            "time_range_end": time_range_end.or(Some(now_unix_secs)).and_then(unix_secs_to_rfc3339),
+            "time_range_start": time_range_start.and_then(unix_ms_to_rfc3339),
+            "time_range_end": time_range_end.map(|ms| unix_ms_to_rfc3339(ms)).unwrap_or_else(|| unix_secs_to_rfc3339(now_unix_secs)),
         });
         if options.include_api_path {
             format_payload["api_path"] = json!(public_api_format_local_path(&api_format));
@@ -536,12 +534,12 @@ pub(crate) fn build_public_health_timeline(
             continue;
         }
 
-        earliest_time = match (earliest_time, bucket.min_created_at_unix_secs) {
+        earliest_time = match (earliest_time, bucket.min_created_at_unix_ms) {
             (Some(left), Some(right)) => Some(left.min(right)),
             (None, Some(right)) => Some(right),
             (left, None) => left,
         };
-        latest_time = match (latest_time, bucket.max_created_at_unix_secs) {
+        latest_time = match (latest_time, bucket.max_created_at_unix_ms) {
             (Some(left), Some(right)) => Some(left.max(right)),
             (None, Some(right)) => Some(right),
             (left, None) => left,
@@ -592,3 +590,50 @@ pub(crate) fn api_format_display_name(api_format: &str) -> String {
     };
     format!("{family_label} {kind_label}")
 }
+
+#[cfg(test)]
+mod tests {
+    use super::request_candidate_event_unix_ms;
+    use crate::handlers::shared::unix_ms_to_rfc3339;
+    use aether_data_contracts::repository::candidates::{
+        RequestCandidateStatus, StoredRequestCandidate,
+    };
+
+    #[test]
+    fn request_candidate_event_timestamp_uses_millisecond_precision() {
+        let candidate = StoredRequestCandidate::new(
+            "cand-1".to_string(),
+            "req-1".to_string(),
+            None,
+            None,
+            None,
+            None,
+            0,
+            0,
+            Some("provider-1".to_string()),
+            Some("endpoint-1".to_string()),
+            Some("key-1".to_string()),
+            RequestCandidateStatus::Success,
+            None,
+            false,
+            Some(200),
+            None,
+            None,
+            Some(42),
+            Some(1),
+            None,
+            None,
+            1_700_000_000_000,
+            Some(1_700_000_000_111),
+            Some(1_700_000_000_123),
+        )
+        .expect("candidate should build");
+
+        let event_unix_ms = request_candidate_event_unix_ms(&candidate);
+        assert_eq!(event_unix_ms, 1_700_000_000_123);
+        assert_eq!(
+            unix_ms_to_rfc3339(event_unix_ms).as_deref(),
+            Some("2023-11-14T22:13:20.123Z")
+        );
+    }
+}

apps/aether-gateway/src/handlers/public/mod.rs

@@ -10,7 +10,7 @@ pub(crate) use self::catalog_helpers::{
     admin_requested_force_stream, api_format_display_name, build_api_format_health_monitor_payload,
     build_public_catalog_models_payload, build_public_catalog_search_models_payload,
     build_public_health_timeline, build_public_providers_payload, normalize_admin_base_url,
-    provider_key_api_formats, request_candidate_event_unix_secs, request_candidate_status_label,
+    provider_key_api_formats, request_candidate_event_unix_ms, request_candidate_status_label,
     ApiFormatHealthMonitorOptions,
 };
 pub(crate) use self::system_modules_helpers::{

apps/aether-gateway/src/handlers/public/support/announcements/shared.rs

@@ -84,7 +84,7 @@ pub(super) fn build_public_announcement_payload(
         },
         "start_time": format_optional_unix_datetime(announcement.start_time_unix_secs),
         "end_time": format_optional_unix_datetime(announcement.end_time_unix_secs),
-        "created_at": format_required_unix_datetime(announcement.created_at_unix_secs),
+        "created_at": format_required_unix_datetime(announcement.created_at_unix_ms),
         "updated_at": format_required_unix_datetime(announcement.updated_at_unix_secs),
     })
 }

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

@@ -361,7 +361,7 @@ fn dashboard_usage_local_date(
     item: &StoredRequestUsageAudit,
     tz_offset_minutes: i32,
 ) -> Option<chrono::NaiveDate> {
-    let timestamp = i64::try_from(item.created_at_unix_secs).ok()?;
+    let timestamp = i64::try_from(item.created_at_unix_ms).ok()?;
     let datetime = chrono::DateTime::<chrono::Utc>::from_timestamp(timestamp, 0)?;
     Some((datetime + chrono::Duration::minutes(i64::from(tz_offset_minutes))).date_naive())
 }
@@ -1008,7 +1008,7 @@ pub(super) async fn handle_dashboard_recent_requests_get(
                 "user": username,
                 "model": dashboard_non_empty_value(&item.model, "N/A"),
                 "tokens": item.total_tokens,
-                "time": dashboard_format_time_hhmm(item.created_at_unix_secs),
+                "time": dashboard_format_time_hhmm(item.created_at_unix_ms),
                 "is_stream": item.is_stream,
             })
         })

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

@@ -95,7 +95,7 @@ fn users_me_usage_total_input_context(item: &StoredRequestUsageAudit) -> u64 {
 
 fn users_me_usage_effective_unix_secs(item: &StoredRequestUsageAudit) -> u64 {
     item.finalized_at_unix_secs
-        .unwrap_or(item.created_at_unix_secs)
+        .unwrap_or(item.created_at_unix_ms)
 }
 
 fn users_me_usage_cache_hit_rate(total_input_context: u64, cache_read_tokens: u64) -> f64 {
@@ -159,7 +159,7 @@ fn build_users_me_usage_record_payload(
         "first_byte_time_ms": item.first_byte_time_ms,
         "is_stream": item.is_stream,
         "status": item.status,
-        "created_at": unix_secs_to_rfc3339(item.created_at_unix_secs),
+        "created_at": unix_secs_to_rfc3339(item.created_at_unix_ms),
         "cache_creation_input_tokens": item.cache_creation_input_tokens,
         "cache_read_input_tokens": item.cache_read_input_tokens,
         "status_code": item.status_code,
@@ -649,8 +649,8 @@ pub(super) async fn handle_users_me_usage_get(
         .collect::<Vec<_>>();
     records.sort_by(|left, right| {
         right
-            .created_at_unix_secs
-            .cmp(&left.created_at_unix_secs)
+            .created_at_unix_ms
+            .cmp(&left.created_at_unix_ms)
             .then_with(|| left.id.cmp(&right.id))
     });
     let total_record_count = records.len();
@@ -735,8 +735,8 @@ pub(super) async fn handle_users_me_usage_active_get(
         .collect::<Vec<_>>();
     items.sort_by(|left, right| {
         right
-            .created_at_unix_secs
-            .cmp(&left.created_at_unix_secs)
+            .created_at_unix_ms
+            .cmp(&left.created_at_unix_ms)
             .then_with(|| left.id.cmp(&right.id))
     });
     if ids.is_none() && items.len() > 50 {
@@ -798,20 +798,19 @@ pub(super) async fn handle_users_me_usage_interval_timeline_get(
     };
     items.retain(|item| item.status == "completed");
     items.sort_by(|left, right| {
-        left.created_at_unix_secs
-            .cmp(&right.created_at_unix_secs)
+        left.created_at_unix_ms
+            .cmp(&right.created_at_unix_ms)
             .then_with(|| left.id.cmp(&right.id))
     });
 
     let mut points = Vec::new();
-    let mut previous_created_at_unix_secs = None;
+    let mut previous_created_at_unix_ms = None;
     for item in items {
-        if let Some(previous) = previous_created_at_unix_secs {
-            let interval_minutes =
-                (item.created_at_unix_secs.saturating_sub(previous) as f64) / 60.0;
+        if let Some(previous) = previous_created_at_unix_ms {
+            let interval_minutes = (item.created_at_unix_ms.saturating_sub(previous) as f64) / 60.0;
             if interval_minutes <= 120.0 {
                 points.push(json!({
-                    "x": unix_secs_to_rfc3339(item.created_at_unix_secs),
+                    "x": unix_secs_to_rfc3339(item.created_at_unix_ms),
                     "y": round_to(interval_minutes, 2),
                     "model": item.model,
                 }));
@@ -820,7 +819,7 @@ pub(super) async fn handle_users_me_usage_interval_timeline_get(
                 }
             }
         }
-        previous_created_at_unix_secs = Some(item.created_at_unix_secs);
+        previous_created_at_unix_ms = Some(item.created_at_unix_ms);
     }
 
     Json(json!({

apps/aether-gateway/src/handlers/public/support/wallet/reads.rs

@@ -153,7 +153,7 @@ pub(super) fn wallet_transaction_payload_from_record(
         "link_id": record.link_id.clone(),
         "operator_id": record.operator_id.clone(),
         "description": record.description.clone(),
-        "created_at": record.created_at_unix_secs.and_then(unix_secs_to_rfc3339),
+        "created_at": record.created_at_unix_ms.and_then(unix_secs_to_rfc3339),
     })
 }
 

apps/aether-gateway/src/handlers/public/support/wallet/recharge.rs

@@ -207,7 +207,7 @@ pub(crate) fn wallet_payment_order_payload_from_row(
     row: &sqlx::postgres::PgRow,
 ) -> Result<serde_json::Value, GatewayError> {
     let created_at = row
-        .try_get::<Option<i64>, _>("created_at_unix_secs")
+        .try_get::<Option<i64>, _>("created_at_unix_ms")
         .map_err(|err| GatewayError::Internal(err.to_string()))?
         .and_then(|value| u64::try_from(value).ok())
         .and_then(unix_secs_to_rfc3339);
@@ -280,7 +280,7 @@ fn wallet_payment_order_payload_from_record(
         record.gateway_order_id.clone(),
         record.gateway_response.clone(),
         record.status.clone(),
-        Some(unix_secs_to_rfc3339(record.created_at_unix_secs)).flatten(),
+        Some(unix_secs_to_rfc3339(record.created_at_unix_ms)).flatten(),
         record.paid_at_unix_secs.and_then(unix_secs_to_rfc3339),
         record.credited_at_unix_secs.and_then(unix_secs_to_rfc3339),
         record.expires_at_unix_secs.and_then(unix_secs_to_rfc3339),

apps/aether-gateway/src/handlers/public/support/wallet/refunds.rs

@@ -98,7 +98,7 @@ fn wallet_refund_payload_from_row(
     row: &sqlx::postgres::PgRow,
 ) -> Result<serde_json::Value, GatewayError> {
     let created_at = row
-        .try_get::<Option<i64>, _>("created_at_unix_secs")
+        .try_get::<Option<i64>, _>("created_at_unix_ms")
         .map_err(|err| GatewayError::Internal(err.to_string()))?
         .and_then(|value| u64::try_from(value).ok())
         .and_then(unix_secs_to_rfc3339);
@@ -157,7 +157,7 @@ fn wallet_refund_payload_from_record(
         "payout_method": record.payout_method,
         "payout_reference": record.payout_reference,
         "payout_proof": record.payout_proof,
-        "created_at": unix_secs_to_rfc3339(record.created_at_unix_secs),
+        "created_at": unix_secs_to_rfc3339(record.created_at_unix_ms),
         "updated_at": unix_secs_to_rfc3339(record.updated_at_unix_secs),
         "processed_at": record.processed_at_unix_secs.and_then(unix_secs_to_rfc3339),
         "completed_at": record.completed_at_unix_secs.and_then(unix_secs_to_rfc3339),
@@ -247,7 +247,7 @@ pub(super) async fn handle_wallet_refunds_list(
                 "payout_method": record.payout_method,
                 "payout_reference": record.payout_reference,
                 "payout_proof": record.payout_proof,
-                "created_at": unix_secs_to_rfc3339(record.created_at_unix_secs),
+                "created_at": unix_secs_to_rfc3339(record.created_at_unix_ms),
                 "updated_at": unix_secs_to_rfc3339(record.updated_at_unix_secs),
                 "processed_at": record.processed_at_unix_secs.and_then(unix_secs_to_rfc3339),
                 "completed_at": record.completed_at_unix_secs.and_then(unix_secs_to_rfc3339),

apps/aether-gateway/src/handlers/public/system_modules_helpers/keys_grouped.rs

@@ -156,7 +156,7 @@ pub(crate) async fn build_admin_keys_grouped_by_format_payload(
                     .and_then(serde_json::Value::as_bool)
                     .unwrap_or(false),
                 "last_used_at": key.last_used_at_unix_secs.and_then(unix_secs_to_rfc3339),
-                "created_at": unix_secs_to_rfc3339(key.created_at_unix_secs.unwrap_or(now_unix_secs)),
+                "created_at": unix_secs_to_rfc3339(key.created_at_unix_ms.unwrap_or(now_unix_secs)),
                 "updated_at": unix_secs_to_rfc3339(key.updated_at_unix_secs.unwrap_or(now_unix_secs)),
             }));
         }

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

@@ -598,7 +598,7 @@ pub(crate) fn build_admin_provider_key_response(
     payload.insert(
         "created_at".to_string(),
         json!(unix_secs_to_rfc3339(
-            key.created_at_unix_secs.unwrap_or(now_unix_secs)
+            key.created_at_unix_ms.unwrap_or(now_unix_secs)
         )),
     );
     payload.insert(