fix(gateway): complete worker registration cleanup

fix(gateway): 修复批量更新 API 格式时的认证通道状态冲突

.env.example

@@ -58,63 +58,33 @@ ADMIN_USERNAME=admin123456
 
 # ==================== 可选配置（有默认值） ====================
 
+# 可信反向代理 IP/CIDR，只有这些来源发送的 X-Real-IP / X-Forwarded-For 会被采用。
+# 默认仅信任本机回环代理：127.0.0.0/8,::1/128。
+# Docker/Nginx 位于独立容器时，请按实际容器网络设置，例如：172.16.0.0/12。
+# AETHER_TRUSTED_PROXY_CIDRS=127.0.0.0/8,::1/128,172.16.0.0/12
+
 # docker compose 下 app 启动前自动执行 pending migration/backfill（默认 true）
 # AETHER_GATEWAY_AUTO_PREPARE_DATABASE=true
 
-# 管理后台更新策略：
+# PostgreSQL 连接池配置（默认按 CPU 自动计算；正式高并发环境可显式预算）
-# - systemd/二进制部署使用 self：下载 GitHub Release 包，校验 SHA256 后切换 current 并重启。
+# AETHER_GATEWAY_DATA_POSTGRES_MIN_CONNECTIONS=12
-# - Docker Compose 使用 docker：后台只提示版本，实际更新请在 compose 目录执行 ./update.sh。
+# AETHER_GATEWAY_DATA_POSTGRES_MAX_CONNECTIONS=80
-# - 源码/本地构建使用 manual：手动拉取源码或下载 release。
+# AETHER_GATEWAY_MAX_IN_FLIGHT_REQUESTS=2048
-# Compose 默认把持久化文件放在 ./datas/{postgres,mysql,sqlite,redis}，日志放在 ./logs。
+# AETHER_GATEWAY_REQUEST_BODY_BUFFER_BUDGET_MB=256
-# 分布式/多节点部署不要使用 ./datas 作为共享数据目录；应使用外部共享 Postgres/MySQL 和 Redis。
+# AETHER_GATEWAY_REQUEST_BODY_READ_TIMEOUT_MS=120000
-# 多节点不要从管理后台一键更新单个节点，应使用镜像滚动更新、systemd 分批发布或外部编排。
+# 可选的 Payload 上限（MiB）；默认及 0 均表示不限制。
-# AETHER_BASE_DIR=/opt/aether
+# AETHER_MAX_REQUEST_BODY_MB=0
-# AETHER_UPDATE_STRATEGY=docker
+# AETHER_GATEWAY_SECURITY_CACHE_TTL_MS=1000
-# AETHER_DOCKER_UPDATE_COMMAND=./update.sh
+# AETHER_MAX_REDACTED_SYNC_RESPONSE_BODY_MB=0
-# AETHER_GATEWAY_DEPLOYMENT_TOPOLOGY=single-node
+# AETHER_MAX_INTERNAL_BUFFERED_BODY_MB=0
-# AETHER_GATEWAY_NODE_ROLE=all
+# AETHER_TUNNEL_NODE_STATUS_QUEUE_CAPACITY=1024
-# Docker Compose 默认强制把应用日志输出到 stdout/stderr，并由 Docker 轮转日志。
-# 如需文件日志，需要在 compose 里把 AETHER_LOG_DESTINATION 改成 file 或 both，
-# 并把容器用户可写目录挂载到 /opt/aether/logs。
-# AETHER_LOG_DESTINATION=stdout
-# AETHER_LOG_FORMAT=pretty
-# AETHER_LOG_DIR=/opt/aether/logs
-# 服务器访问 GitHub 需要代理时可配置；也兼容 UPDATE_PROXY_URL / HTTPS_PROXY / ALL_PROXY / HTTP_PROXY。
-# 如果 Aether 跑在 Docker 容器里，想走宿主机代理时请写 host.docker.internal，不要写 127.0.0.1。
-# AETHER_UPDATE_PROXY_URL=http://host.docker.internal:7890
-# 共享出口触发 GitHub API 限流时可配置只读 token；也兼容 GITHUB_TOKEN / GH_TOKEN。
-# AETHER_UPDATE_GITHUB_TOKEN=
-# 下载超时控制：总超时默认 600 秒；连续无响应/无数据默认 30 秒。
-# AETHER_UPDATE_DOWNLOAD_TIMEOUT_SECS=600
-# AETHER_UPDATE_DOWNLOAD_IDLE_TIMEOUT_SECS=30
-# 本地联调后台在线更新（配合 docker-compose.release-local.yml）：
-# 会用当前源码构建 release-layout 测试镜像，并伪装成较旧版本以触发升级入口。
-# AETHER_RELEASE_LOCAL_VERSION=v0.7.0
-# AETHER_RELEASE_LOCAL_PORT=18085
-# LOCAL_RELEASE_APP_IMAGE=aether-app:release-local
 
-# PostgreSQL 连接池配置（默认适合单实例/小型部署；高并发可按需调大）
+# PostgreSQL 容器调优：docker-compose.yml 已内置通用默认值，通常不用配置。
-# 推荐计算方式（单实例）：
+# 只有在 Postgres 独占大内存、或压测显示 DB 缓存/排序/维护任务成为瓶颈时再覆盖。
-#   MAX = CPU 核数 × 10（AI 网关偏 IO 等待，可激进些；纯 OLTP 用 × 4）
+# 内置默认：shared_buffers=1GB, effective_cache_size=3GB, shm_size=512mb,
-#   MIN = MAX × 0.2（保留常驻连接应对突发流量，避免冷启动握手开销）
+# work_mem=16MB, maintenance_work_mem=256MB。
-# 多实例部署时请按 实例数 × MAX 控制总和，PG 端 max_connections 至少为该总和 + 20 余量
+# POSTGRES_SHARED_BUFFERS=8GB
-# AETHER_GATEWAY_DATA_POSTGRES_MIN_CONNECTIONS=4
+# POSTGRES_EFFECTIVE_CACHE_SIZE=24GB
-# AETHER_GATEWAY_DATA_POSTGRES_MAX_CONNECTIONS=20
+# POSTGRES_SHM_SIZE=2gb
-# AETHER_GATEWAY_DATA_POSTGRES_STATEMENT_CACHE_CAPACITY=100
-# AETHER_GATEWAY_DATA_POSTGRES_ACQUIRE_TIMEOUT_MS=3000
-
-# PostgreSQL 性能调优（默认值适合 2核4GB 机器，按实际配置覆盖）
-# 参考：shared_buffers ≈ 可用内存 25%，effective_cache_size ≈ 可用内存 50-75%
-# POSTGRES_SHM_SIZE 控制 Docker 容器 /dev/shm；仪表盘统计等并行查询会使用它。
-# work_mem 是每个连接每个排序操作的内存，不要设太大（并发数 × work_mem 是实际占用）
-# | 系统内存 | shared_buffers | effective_cache_size | work_mem |
-# | 2GB      | 256MB          | 768MB                | 4MB      |
-# | 4GB      | 1GB            | 3GB                  | 16MB     |
-# | 8GB      | 2GB            | 6GB                  | 16MB     |
-# | 16GB     | 4GB            | 12GB                 | 32MB     |
-# | 32GB+    | 8GB            | 24GB                 | 32MB     |
-# POSTGRES_SHARED_BUFFERS=1GB
-# POSTGRES_EFFECTIVE_CACHE_SIZE=3GB
-# POSTGRES_SHM_SIZE=512mb
 # POSTGRES_WORK_MEM=16MB
-# POSTGRES_MAINTENANCE_WORK_MEM=256MB
+# POSTGRES_MAINTENANCE_WORK_MEM=1GB

.github/workflows/build-tunnel.yml

@@ -131,6 +131,7 @@ jobs:
             aether-tunnel-*.tar.gz
             aether-tunnel-*.zip
           if-no-files-found: error
+          retention-days: 1
 
   release:
     needs: build

.github/workflows/rust-ci.yml

@@ -25,6 +25,8 @@ concurrency:
 
 env:
   CARGO_INCREMENTAL: 0
+  CARGO_PROFILE_DEV_DEBUG: 0
+  CARGO_PROFILE_TEST_DEBUG: 0
   CARGO_TERM_COLOR: always
 
 jobs:
@@ -68,7 +70,7 @@ jobs:
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
-        run: cargo clippy -p aether-gateway --all-targets -- -D warnings
+        run: cargo clippy -p aether-gateway --lib --bins --examples -- -D warnings
 
       - name: Show sccache stats
         if: always()
@@ -136,7 +138,7 @@ jobs:
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
-        run: cargo clippy --workspace --exclude aether-gateway --exclude aether-data --all-targets -- -D warnings
+        run: cargo clippy --workspace --exclude aether-gateway --exclude aether-data --exclude aether-integration-tests --all-targets -- -D warnings
 
       - name: Show sccache stats
         if: always()
@@ -184,15 +186,27 @@ jobs:
       - name: Setup sccache
         uses: mozilla-actions/[email protected]
 
+      - name: Setup mold
+        uses: rui314/setup-mold@v1
+
       - name: Install nextest
         uses: taiki-e/install-action@nextest
 
-      - name: Test
+      - name: Test lib
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           RUST_MIN_STACK: "16777216"
-        run: cargo nextest run -p aether-gateway
+          RUSTFLAGS: "-C link-arg=-fuse-ld=mold"
+        run: cargo nextest run -p aether-gateway --lib
+
+      - name: Test bin
+        env:
+          RUSTC_WRAPPER: sccache
+          SCCACHE_GHA_ENABLED: "true"
+          RUST_MIN_STACK: "16777216"
+          RUSTFLAGS: "-C link-arg=-fuse-ld=mold"
+        run: cargo nextest run -p aether-gateway --bin aether-gateway
 
       - name: Show sccache stats
         if: always()
@@ -238,6 +252,45 @@ jobs:
           SCCACHE_GHA_ENABLED: "true"
         run: sccache --show-stats
 
+  check_data_features:
+    name: Check (Data Feature - ${{ matrix.feature }})
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        feature:
+          - postgres
+          - mysql
+          - sqlite
+          - all-drivers
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-toolchain@stable
+
+      - name: Rust cache
+        uses: Swatinem/rust-cache@v2
+        with:
+          shared-key: rust-ci-${{ runner.os }}
+          workspaces: . -> target
+
+      - name: Setup sccache
+        uses: mozilla-actions/[email protected]
+
+      - name: Check selected data driver
+        env:
+          RUSTC_WRAPPER: sccache
+          SCCACHE_GHA_ENABLED: "true"
+        run: cargo check -p aether-data --no-default-features --features ${{ matrix.feature }}
+
+      - name: Show sccache stats
+        if: always()
+        env:
+          RUSTC_WRAPPER: sccache
+          SCCACHE_GHA_ENABLED: "true"
+        run: sccache --show-stats
+
   test_rest:
     name: Test (Workspace Rest)
     runs-on: ubuntu-latest
@@ -266,7 +319,79 @@ jobs:
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
-        run: cargo nextest run --workspace --exclude aether-gateway --exclude aether-data
+        run: cargo nextest run --workspace --exclude aether-gateway --exclude aether-data --exclude aether-integration-tests
+
+      - name: Show sccache stats
+        if: always()
+        env:
+          RUSTC_WRAPPER: sccache
+          SCCACHE_GHA_ENABLED: "true"
+        run: sccache --show-stats
+
+  test_data_adapters:
+    name: Test (Data Adapter - ${{ matrix.package }})
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        package:
+          - aether-data-postgres
+          - aether-data-mysql
+          - aether-data-sqlite
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-toolchain@stable
+
+      - name: Rust cache
+        uses: Swatinem/rust-cache@v2
+        with:
+          shared-key: rust-ci-${{ runner.os }}
+          workspaces: . -> target
+
+      - name: Setup sccache
+        uses: mozilla-actions/[email protected]
+
+      - name: Install nextest
+        uses: taiki-e/install-action@nextest
+
+      - name: Test adapter
+        env:
+          RUSTC_WRAPPER: sccache
+          SCCACHE_GHA_ENABLED: "true"
+        run: cargo nextest run -p ${{ matrix.package }}
+
+      - name: Show sccache stats
+        if: always()
+        env:
+          RUSTC_WRAPPER: sccache
+          SCCACHE_GHA_ENABLED: "true"
+        run: sccache --show-stats
+
+  check_integration_scenarios:
+    name: Test (Integration Scenarios)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-toolchain@stable
+
+      - name: Rust cache
+        uses: Swatinem/rust-cache@v2
+        with:
+          shared-key: rust-ci-${{ runner.os }}
+          workspaces: . -> target
+
+      - name: Setup sccache
+        uses: mozilla-actions/[email protected]
+
+      - name: Test scenario binaries
+        env:
+          RUSTC_WRAPPER: sccache
+          SCCACHE_GHA_ENABLED: "true"
+        run: cargo test -p aether-integration-tests --bins
 
       - name: Show sccache stats
         if: always()
@@ -281,14 +406,20 @@ jobs:
     needs:
       - test_gateway
       - test_data
+      - check_data_features
       - test_rest
+      - test_data_adapters
+      - check_integration_scenarios
     if: ${{ always() }}
     steps:
       - name: Verify test jobs
         run: |
           if [ "${{ needs.test_gateway.result }}" != "success" ] || \
              [ "${{ needs.test_data.result }}" != "success" ] || \
-             [ "${{ needs.test_rest.result }}" != "success" ]; then
+             [ "${{ needs.check_data_features.result }}" != "success" ] || \
+             [ "${{ needs.test_rest.result }}" != "success" ] || \
+             [ "${{ needs.test_data_adapters.result }}" != "success" ] || \
+             [ "${{ needs.check_integration_scenarios.result }}" != "success" ]; then
             echo "Tests failed"
             exit 1
           fi
@@ -318,7 +449,7 @@ jobs:
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
-        run: cargo test -p aether-data sqlite --lib
+        run: cargo test -p aether-data --all-features sqlite --lib
 
       - name: Show sccache stats
         if: always()
@@ -362,26 +493,48 @@ jobs:
       - name: Setup sccache
         uses: mozilla-actions/[email protected]
 
+      - name: Add PostgreSQL server binaries to PATH
+        run: echo "$(pg_config --bindir)" >> "$GITHUB_PATH"
+
       - name: Run Postgres migration smoke test
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_POSTGRES_URL: postgres://aether:[email protected]:5432/aether_test
-        run: cargo test -p aether-data postgres_migrations_create_core_config_tables_when_url_is_set --lib -- --nocapture
+        run: cargo test -p aether-data --all-features postgres_migrations_create_core_config_tables_when_url_is_set --lib -- --nocapture
+
+      - name: Run Postgres provider metadata migration smoke test
+        env:
+          RUSTC_WRAPPER: sccache
+          SCCACHE_GHA_ENABLED: "true"
+          AETHER_TEST_POSTGRES_URL: postgres://aether:[email protected]:5432/aether_test
+        run: cargo test -p aether-data --all-features postgres_provider_upstream_metadata_migration_preserves_json_when_url_is_set --lib -- --nocapture
+
+      - name: Run Postgres API key lifecycle tests
+        env:
+          RUSTC_WRAPPER: sccache
+          SCCACHE_GHA_ENABLED: "true"
+          AETHER_REQUIRE_LOCAL_POSTGRES_TESTS: "true"
+        run: |
+          cargo test -p aether-data --all-features lifecycle::migrate::tests::postgres_request_candidates_preserve_deleted_api_key_identity --lib -- --exact --nocapture
+          cargo test -p aether-data --all-features lifecycle::migrate::tests::postgres_request_candidate_migration_decouples_legacy_api_key_foreign_key --lib -- --exact --nocapture
+          cargo test -p aether-data --all-features lifecycle::migrate::tests::postgres_stats_daily_api_key_migration_decouples_legacy_foreign_key --lib -- --exact --nocapture
+          cargo test -p aether-data --all-features lifecycle::migrate::tests::postgres_expired_api_key_cleanup_preserves_historical_identity --lib -- --exact --nocapture
+          cargo test -p aether-data --all-features lifecycle::migrate::tests::postgres_api_key_leaderboard_user_filter_preserves_aggregate_history --lib -- --exact --nocapture
 
       - name: Run Postgres core export smoke test
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_POSTGRES_URL: postgres://aether:[email protected]:5432/aether_test
-        run: cargo test -p aether-data postgres_core_export_reads_migrated_database_rows_when_url_is_set --lib -- --nocapture
+        run: cargo test -p aether-data --all-features postgres_core_export_reads_migrated_database_rows_when_url_is_set --lib -- --nocapture
 
       - name: Run SQLite-to-Postgres import smoke test
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_POSTGRES_URL: postgres://aether:[email protected]:5432/aether_test
-        run: cargo test -p aether-data sqlite_core_export_reads_migrated_database_rows --lib -- --nocapture
+        run: cargo test -p aether-data --all-features sqlite_core_export_reads_migrated_database_rows --lib -- --nocapture
 
       - name: Show sccache stats
         if: always()
@@ -431,56 +584,56 @@ jobs:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_MYSQL_URL: mysql://aether:[email protected]:3306/aether_test
-        run: cargo test -p aether-data mysql_migrations_create_core_config_tables_when_url_is_set --lib -- --nocapture
+        run: cargo test -p aether-data --all-features mysql_migrations_create_core_config_tables_when_url_is_set --lib -- --nocapture
 
       - name: Run MySQL usage write smoke test
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_MYSQL_URL: mysql://aether:[email protected]:3306/aether_test
-        run: cargo test -p aether-data mysql_usage_write_repository_upserts_when_url_is_set --lib -- --nocapture
+        run: cargo test -p aether-data-mysql mysql_usage_write_repository_upserts_and_flushes_counters_when_url_is_set --lib -- --nocapture
 
       - name: Run MySQL usage read smoke test
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_MYSQL_URL: mysql://aether:[email protected]:3306/aether_test
-        run: cargo test -p aether-data mysql_usage_read_repository_reads_usage_contract_views_when_url_is_set --lib -- --nocapture
+        run: cargo test -p aether-data-mysql mysql_usage_read_repository_reads_usage_contract_views_when_url_is_set --lib -- --nocapture
 
       - name: Run MySQL provider catalog smoke test
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_MYSQL_URL: mysql://aether:[email protected]:3306/aether_test
-        run: cargo test -p aether-data mysql_provider_catalog_repository_round_trips_when_url_is_set --lib -- --nocapture
+        run: cargo test -p aether-data-mysql mysql_provider_catalog_repository_round_trips_when_url_is_set --lib -- --nocapture
 
       - name: Run MySQL core export smoke test
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_MYSQL_URL: mysql://aether:[email protected]:3306/aether_test
-        run: cargo test -p aether-data mysql_core_export_reads_migrated_database_rows_when_url_is_set --lib -- --nocapture
+        run: cargo test -p aether-data --all-features mysql_core_export_reads_migrated_database_rows_when_url_is_set --lib -- --nocapture
 
       - name: Run MySQL wallet read smoke test
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_MYSQL_URL: mysql://aether:[email protected]:3306/aether_test
-        run: cargo test -p aether-data mysql_wallet_read_repository_reads_wallet_contract_views --lib -- --nocapture
+        run: cargo test -p aether-data-mysql mysql_wallet_read_repository_reads_wallet_contract_views --lib -- --nocapture
 
       - name: Run MySQL wallet daily usage aggregation smoke test
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_MYSQL_URL: mysql://aether:[email protected]:3306/aether_test
-        run: cargo test -p aether-data mysql_wallet_daily_usage_aggregation_uses_settlement_wallets_when_url_is_set --lib -- --nocapture
+        run: cargo test -p aether-data --all-features mysql_wallet_daily_usage_aggregation_uses_settlement_wallets_when_url_is_set --lib -- --nocapture
 
       - name: Run MySQL stats aggregation smoke test
         env:
           RUSTC_WRAPPER: sccache
           SCCACHE_GHA_ENABLED: "true"
           AETHER_TEST_MYSQL_URL: mysql://aether:[email protected]:3306/aether_test
-        run: cargo test -p aether-data mysql_stats_aggregation_runs_after_mysql_migrations_when_url_is_set --lib -- --nocapture
+        run: cargo test -p aether-data --all-features mysql_stats_aggregation_runs_after_mysql_migrations_when_url_is_set --lib -- --nocapture
 
       - name: Show sccache stats
         if: always()

Cargo.lock

@@ -69,6 +69,14 @@ dependencies = [
  "uuid",
 ]
 
+[[package]]
+name = "aether-admission-core"
+version = "0.1.0"
+dependencies = [
+ "serde",
+ "thiserror 2.0.18",
+]
+
 [[package]]
 name = "aether-ai-formats"
 version = "0.1.0"
@@ -95,6 +103,7 @@ dependencies = [
  "aether-pool-core",
  "aether-scheduler-core",
  "async-trait",
+ "base64 0.22.1",
  "http",
  "serde",
  "serde_json",
@@ -155,7 +164,9 @@ dependencies = [
  "aether-ai-formats",
  "aether-cache",
  "aether-data-contracts",
- "aether-data-query",
+ "aether-data-mysql",
+ "aether-data-postgres",
+ "aether-data-sqlite",
  "aether-wallet",
  "async-trait",
  "chrono",
@@ -183,7 +194,48 @@ dependencies = [
  "chrono",
  "serde",
  "serde_json",
+ "sha2",
  "thiserror 2.0.18",
+ "tokio",
+]
+
+[[package]]
+name = "aether-data-mysql"
+version = "0.1.0"
+dependencies = [
+ "aether-ai-formats",
+ "aether-data-contracts",
+ "aether-data-query",
+ "async-trait",
+ "chrono",
+ "chrono-tz",
+ "flate2",
+ "serde_json",
+ "sha2",
+ "sqlx",
+ "tokio",
+ "tracing",
+ "uuid",
+]
+
+[[package]]
+name = "aether-data-postgres"
+version = "0.1.0"
+dependencies = [
+ "aether-ai-formats",
+ "aether-data-contracts",
+ "aether-data-query",
+ "async-trait",
+ "chrono",
+ "chrono-tz",
+ "flate2",
+ "futures-util",
+ "serde_json",
+ "sha2",
+ "sqlx",
+ "tokio",
+ "tracing",
+ "uuid",
 ]
 
 [[package]]
@@ -203,6 +255,25 @@ dependencies = [
  "toml",
 ]
 
+[[package]]
+name = "aether-data-sqlite"
+version = "0.1.0"
+dependencies = [
+ "aether-ai-formats",
+ "aether-data-contracts",
+ "aether-data-query",
+ "async-trait",
+ "chrono",
+ "chrono-tz",
+ "flate2",
+ "serde_json",
+ "sha2",
+ "sqlx",
+ "tokio",
+ "tracing",
+ "uuid",
+]
+
 [[package]]
 name = "aether-dispatch-core"
 version = "0.1.0"
@@ -229,6 +300,11 @@ dependencies = [
  "aether-data",
  "aether-data-contracts",
  "aether-dispatch-core",
+ "aether-gateway-control",
+ "aether-gateway-execution",
+ "aether-gateway-frontdoor",
+ "aether-gateway-tunnel",
+ "aether-gateway-workers",
  "aether-http",
  "aether-model-fetch",
  "aether-oauth",
@@ -240,7 +316,7 @@ dependencies = [
  "aether-runtime-state",
  "aether-scheduler-core",
  "aether-task-runtime",
- "aether-testkit",
+ "aether-test-support",
  "aether-usage-runtime",
  "aether-video-tasks-core",
  "aether-wallet",
@@ -258,7 +334,11 @@ dependencies = [
  "futures-util",
  "hmac",
  "http",
+ "http-body-util",
+ "hyper",
+ "hyper-util",
  "ldap3",
+ "libc",
  "md-5",
  "object_store",
  "parking_lot",
@@ -270,9 +350,12 @@ dependencies = [
  "serde_json",
  "sha1",
  "sha2",
+ "socket2 0.6.3",
  "sqlx",
+ "sysinfo",
  "tar",
  "thiserror 2.0.18",
+ "tikv-jemalloc-sys",
  "tikv-jemallocator",
  "tokio",
  "tokio-util",
@@ -288,6 +371,63 @@ dependencies = [
  "zstd",
 ]
 
+[[package]]
+name = "aether-gateway-control"
+version = "0.1.0"
+dependencies = [
+ "http",
+]
+
+[[package]]
+name = "aether-gateway-execution"
+version = "0.1.0"
+dependencies = [
+ "aether-contracts",
+ "bytes",
+ "serde_json",
+]
+
+[[package]]
+name = "aether-gateway-frontdoor"
+version = "0.1.0"
+dependencies = [
+ "aether-ai-formats",
+ "axum",
+ "bytes",
+ "futures-util",
+ "http",
+ "serde_json",
+ "tokio",
+ "tower",
+ "tracing",
+ "tracing-subscriber",
+ "uuid",
+]
+
+[[package]]
+name = "aether-gateway-tunnel"
+version = "0.1.0"
+dependencies = [
+ "aether-admission-core",
+ "aether-contracts",
+ "base64 0.22.1",
+ "bytes",
+ "http",
+ "serde",
+ "serde_json",
+]
+
+[[package]]
+name = "aether-gateway-workers"
+version = "0.1.0"
+dependencies = [
+ "aether-runtime-state",
+ "aether-task-runtime",
+ "aether-test-support",
+ "tokio",
+ "tracing",
+]
+
 [[package]]
 name = "aether-http"
 version = "0.1.0"
@@ -296,6 +436,48 @@ dependencies = [
  "serde",
 ]
 
+[[package]]
+name = "aether-integration-tests"
+version = "0.1.0"
+dependencies = [
+ "aether-contracts",
+ "aether-data",
+ "aether-data-contracts",
+ "aether-gateway",
+ "aether-runtime-state",
+ "aether-testkit",
+ "async-stream",
+ "axum",
+ "futures-util",
+ "http",
+ "reqwest",
+ "serde",
+ "serde_json",
+ "sha2",
+ "sqlx",
+ "tokio",
+ "tokio-tungstenite 0.28.0",
+]
+
+[[package]]
+name = "aether-loadtools"
+version = "0.1.0"
+dependencies = [
+ "aether-http",
+ "aether-runtime",
+ "aether-runtime-state",
+ "aether-test-support",
+ "bytes",
+ "futures-util",
+ "http",
+ "libc",
+ "reqwest",
+ "serde",
+ "serde_json",
+ "sysinfo",
+ "tokio",
+]
+
 [[package]]
 name = "aether-model-fetch"
 version = "0.1.0"
@@ -341,12 +523,21 @@ dependencies = [
  "serde_json",
 ]
 
+[[package]]
+name = "aether-provider-core"
+version = "0.1.0"
+dependencies = [
+ "serde",
+ "thiserror 2.0.18",
+]
+
 [[package]]
 name = "aether-provider-pool"
 version = "0.1.0"
 dependencies = [
  "aether-data-contracts",
  "aether-pool-core",
+ "aether-provider-transport",
  "serde_json",
  "url",
  "uuid",
@@ -366,6 +557,9 @@ dependencies = [
  "async-trait",
  "axum",
  "base64 0.22.1",
+ "chrono",
+ "crypto_box",
+ "ed25519-dalek",
  "http",
  "regex",
  "reqwest",
@@ -440,11 +634,20 @@ dependencies = [
  "sha2",
 ]
 
+[[package]]
+name = "aether-task-core"
+version = "0.1.0"
+dependencies = [
+ "serde",
+ "thiserror 2.0.18",
+]
+
 [[package]]
 name = "aether-task-runtime"
 version = "0.1.0"
 dependencies = [
  "aether-runtime",
+ "aether-task-core",
  "serde",
  "serde_json",
  "tokio",
@@ -452,30 +655,25 @@ dependencies = [
  "tracing",
 ]
 
+[[package]]
+name = "aether-test-support"
+version = "0.1.0"
+dependencies = [
+ "tokio",
+]
+
 [[package]]
 name = "aether-testkit"
 version = "0.1.0"
 dependencies = [
- "aether-contracts",
  "aether-data",
- "aether-data-contracts",
  "aether-gateway",
- "aether-http",
+ "aether-loadtools",
  "aether-runtime",
  "aether-runtime-state",
- "async-stream",
  "axum",
- "bytes",
- "futures-util",
- "http",
- "libc",
- "reqwest",
- "serde",
- "serde_json",
  "sqlx",
- "sysinfo",
  "tokio",
- "tokio-tungstenite 0.28.0",
 ]
 
 [[package]]
@@ -484,6 +682,7 @@ version = "0.3.16"
 dependencies = [
  "aether-contracts",
  "aether-gateway",
+ "aether-gateway-tunnel",
  "aether-http",
  "aether-runtime",
  "aether-runtime-state",
@@ -522,6 +721,14 @@ dependencies = [
  "webpki-roots 0.26.11",
 ]
 
+[[package]]
+name = "aether-usage-core"
+version = "0.1.0"
+dependencies = [
+ "serde",
+ "thiserror 2.0.18",
+]
+
 [[package]]
 name = "aether-usage-runtime"
 version = "0.1.0"
@@ -533,6 +740,7 @@ dependencies = [
  "aether-runtime-state",
  "async-trait",
  "base64 0.22.1",
+ "futures-util",
  "serde",
  "serde_json",
  "tokio",
@@ -906,6 +1114,19 @@ dependencies = [
  "zeroize",
 ]
 
+[[package]]
+name = "bigdecimal"
+version = "0.4.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4d6867f1565b3aad85681f1015055b087fcfd840d6aeee6eee7f2da317603695"
+dependencies = [
+ "autocfg",
+ "libm",
+ "num-bigint",
+ "num-integer",
+ "num-traits",
+]
+
 [[package]]
 name = "bindgen"
 version = "0.72.1"
@@ -954,6 +1175,15 @@ dependencies = [
  "serde_core",
 ]
 
+[[package]]
+name = "blake2"
+version = "0.10.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "46502ad458c9a52b69d4d4d32775c788b7a1b85e8bc9d482d92250fc0e3f8efe"
+dependencies = [
+ "digest",
+]
+
 [[package]]
 name = "block-buffer"
 version = "0.10.4"
@@ -1135,6 +1365,7 @@ checksum = "773f3b9af64447d2ce9850330c473515014aa235e6a783b02db81ff39e4a3dad"
 dependencies = [
  "crypto-common",
  "inout",
+ "zeroize",
 ]
 
 [[package]]
@@ -1419,6 +1650,36 @@ dependencies = [
  "typenum",
 ]
 
+[[package]]
+name = "crypto_box"
+version = "0.9.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "16182b4f39a82ec8a6851155cc4c0cda3065bb1db33651726a29e1951de0f009"
+dependencies = [
+ "aead",
+ "blake2",
+ "crypto_secretbox",
+ "curve25519-dalek",
+ "salsa20",
+ "subtle",
+ "zeroize",
+]
+
+[[package]]
+name = "crypto_secretbox"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b9d6cf87adf719ddf43a805e92c6870a531aedda35ff640442cbaf8674e141e1"
+dependencies = [
+ "aead",
+ "cipher",
+ "generic-array",
+ "poly1305",
+ "salsa20",
+ "subtle",
+ "zeroize",
+]
+
 [[package]]
 name = "csscolorparser"
 version = "0.6.2"
@@ -1438,6 +1699,33 @@ dependencies = [
  "cipher",
 ]
 
+[[package]]
+name = "curve25519-dalek"
+version = "4.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "97fb8b7c4503de7d6ae7b42ab72a5a59857b4c937ec27a3d4539dba95b5ab2be"
+dependencies = [
+ "cfg-if",
+ "cpufeatures",
+ "curve25519-dalek-derive",
+ "digest",
+ "fiat-crypto",
+ "rustc_version",
+ "subtle",
+ "zeroize",
+]
+
+[[package]]
+name = "curve25519-dalek-derive"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f46882e17999c6cc590af592290432be3bce0428cb0d5f8b6715e4dc7b383eb3"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
 [[package]]
 name = "darling"
 version = "0.23.0"
@@ -1598,6 +1886,30 @@ version = "1.0.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "92773504d58c093f6de2459af4af33faa518c13451eb8f2b5698ed3d36e7c813"
 
+[[package]]
+name = "ed25519"
+version = "2.2.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "115531babc129696a58c64a4fef0a8bf9e9698629fb97e9e40767d235cfbcd53"
+dependencies = [
+ "pkcs8",
+ "signature",
+]
+
+[[package]]
+name = "ed25519-dalek"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "70e796c081cee67dc755e1a36a0a172b897fab85fc3f6bc48307991f64e4eca9"
+dependencies = [
+ "curve25519-dalek",
+ "ed25519",
+ "serde",
+ "sha2",
+ "subtle",
+ "zeroize",
+]
+
 [[package]]
 name = "either"
 version = "1.15.0"
@@ -1670,6 +1982,12 @@ version = "2.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9f1f227452a390804cdb637b74a86990f2a7d7ba4b7d5693aac9b4dd6defd8d6"
 
+[[package]]
+name = "fiat-crypto"
+version = "0.2.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "28dea519a9695b9977216879a3ebfddf92f1c08c05d984f8996aecd6ecdc811d"
+
 [[package]]
 name = "filedescriptor"
 version = "0.8.3"
@@ -1908,6 +2226,7 @@ checksum = "85649ca51fd72272d7821adaf274ad91c288277713d9c18820d8499a7ff69e9a"
 dependencies = [
  "typenum",
  "version_check",
+ "zeroize",
 ]
 
 [[package]]
@@ -2204,6 +2523,7 @@ dependencies = [
  "pin-project-lite",
  "socket2 0.6.3",
  "tokio",
+ "tower-layer",
  "tower-service",
  "tracing",
 ]
@@ -3162,6 +3482,17 @@ version = "0.2.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b4596b6d070b27117e987119b4dac604f3c58cfb0b191112e24771b2faeac1a6"
 
+[[package]]
+name = "poly1305"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8159bd90725d2df49889a078b54f4f79e87f1f8a8444194cdca81d38f5393abf"
+dependencies = [
+ "cpufeatures",
+ "opaque-debug",
+ "universal-hash",
+]
+
 [[package]]
 name = "polyval"
 version = "0.6.2"
@@ -3790,6 +4121,15 @@ version = "1.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9774ba4a74de5f7b1c1451ed6cd5285a32eddb5cccb8cc655a4e50009e06477f"
 
+[[package]]
+name = "salsa20"
+version = "0.10.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "97a22f5af31f73a954c10289c93e8a50cc23d971e80ee446f1f6f7137a088213"
+dependencies = [
+ "cipher",
+]
+
 [[package]]
 name = "schannel"
 version = "0.1.29"
@@ -4120,6 +4460,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ee6798b1838b6a0f69c007c133b8df5866302197e404e8b6ee8ed3e3a5e68dc6"
 dependencies = [
  "base64 0.22.1",
+ "bigdecimal",
  "bytes",
  "chrono",
  "crc",
@@ -4196,6 +4537,7 @@ checksum = "aa003f0038df784eb8fecbbac13affe3da23b45194bd57dba231c8f48199c526"
 dependencies = [
  "atoi",
  "base64 0.22.1",
+ "bigdecimal",
  "bitflags 2.11.0",
  "byteorder",
  "bytes",
@@ -4239,6 +4581,7 @@ checksum = "db58fcd5a53cf07c184b154801ff91347e4c30d17a3562a635ff028ad5deda46"
 dependencies = [
  "atoi",
  "base64 0.22.1",
+ "bigdecimal",
  "bitflags 2.11.0",
  "byteorder",
  "chrono",
@@ -4256,6 +4599,7 @@ dependencies = [
  "log",
  "md-5",
  "memchr",
+ "num-bigint",
  "once_cell",
  "rand 0.8.5",
  "serde",

Cargo.toml

@@ -1,34 +1,52 @@
 [workspace]
 members = [
     "apps/aether-tunnel",
-    "crates/aether-ai-formats",
+    "crates/aether-ai/formats",
     "crates/aether-admin",
-    "crates/aether-ai-serving",
+    "crates/aether-admission-core",
+    "crates/aether-ai/serving",
     "crates/aether-pool-core",
-    "crates/aether-provider-pool",
+    "crates/aether-provider/core",
+    "crates/aether-provider/pool",
     "crates/aether-routing-core",
-    "crates/aether-data-contracts",
+    "crates/aether-data/contracts",
-    "crates/aether-data-query",
+    "crates/aether-data/adapters/postgres",
-    "crates/aether-data-schema",
+    "crates/aether-data/adapters/mysql",
+    "crates/aether-data/adapters/sqlite",
+    "crates/aether-data/query",
+    "crates/aether-data/schema",
     "crates/aether-dispatch-core",
     "crates/aether-cache",
     "crates/aether-billing",
     "crates/aether-wallet",
     "crates/aether-crypto",
     "crates/aether-contracts",
-    "crates/aether-data",
+    "crates/aether-data/runtime",
     "crates/aether-model-fetch",
     "crates/aether-oauth",
-    "crates/aether-provider-transport",
+    "crates/aether-provider/transport",
     "crates/aether-scheduler-core",
-    "crates/aether-runtime-state",
+    "crates/aether-runtime/state",
-    "crates/aether-task-runtime",
+    "crates/aether-task/runtime",
-    "crates/aether-usage-runtime",
+    "crates/aether-task/core",
+    "crates/aether-gateway/frontdoor",
+    "crates/aether-gateway/control",
+    "crates/aether-gateway/execution",
+    "crates/aether-gateway/workers",
+    "crates/aether-gateway/tunnel",
+    "crates/aether-testing/loadtools",
+    "crates/aether-testing/integration",
+    "crates/aether-usage/core",
+    "crates/aether-testing/support",
+    "crates/aether-usage/runtime",
     "crates/aether-video-tasks-core",
     "apps/aether-gateway",
     "crates/aether-http",
-    "crates/aether-runtime",
+    "crates/aether-runtime/base",
-    "crates/aether-testkit",
+    "crates/aether-testing/testkit",
+]
+default-members = [
+    "apps/aether-gateway",
 ]
 resolver = "2"
 
@@ -39,33 +57,48 @@ repository = "https://github.com/fawney19/Aether.git"
 
 [workspace.dependencies]
 aether-admin = { path = "crates/aether-admin" }
-aether-ai-formats = { path = "crates/aether-ai-formats" }
+aether-admission-core = { path = "crates/aether-admission-core" }
-aether-ai-serving = { path = "crates/aether-ai-serving" }
+aether-ai-formats = { path = "crates/aether-ai/formats" }
+aether-ai-serving = { path = "crates/aether-ai/serving" }
 aether-pool-core = { path = "crates/aether-pool-core" }
-aether-provider-pool = { path = "crates/aether-provider-pool" }
+aether-provider-core = { path = "crates/aether-provider/core" }
+aether-provider-pool = { path = "crates/aether-provider/pool" }
 aether-routing-core = { path = "crates/aether-routing-core" }
-aether-data-contracts = { path = "crates/aether-data-contracts" }
+aether-data-contracts = { path = "crates/aether-data/contracts" }
-aether-data-query = { path = "crates/aether-data-query" }
+aether-data-postgres = { path = "crates/aether-data/adapters/postgres" }
-aether-data-schema = { path = "crates/aether-data-schema" }
+aether-data-mysql = { path = "crates/aether-data/adapters/mysql" }
+aether-data-sqlite = { path = "crates/aether-data/adapters/sqlite" }
+aether-data-query = { path = "crates/aether-data/query" }
+aether-data-schema = { path = "crates/aether-data/schema" }
 aether-dispatch-core = { path = "crates/aether-dispatch-core" }
 aether-cache = { path = "crates/aether-cache" }
 aether-billing = { path = "crates/aether-billing" }
 aether-wallet = { path = "crates/aether-wallet" }
 aether-crypto = { path = "crates/aether-crypto" }
 aether-contracts = { path = "crates/aether-contracts" }
-aether-data = { path = "crates/aether-data" }
+aether-data = { path = "crates/aether-data/runtime" }
 aether-model-fetch = { path = "crates/aether-model-fetch" }
 aether-oauth = { path = "crates/aether-oauth" }
-aether-provider-transport = { path = "crates/aether-provider-transport" }
+aether-provider-transport = { path = "crates/aether-provider/transport" }
 aether-scheduler-core = { path = "crates/aether-scheduler-core" }
-aether-runtime-state = { path = "crates/aether-runtime-state" }
+aether-runtime-state = { path = "crates/aether-runtime/state" }
-aether-task-runtime = { path = "crates/aether-task-runtime" }
+aether-task-runtime = { path = "crates/aether-task/runtime" }
-aether-usage-runtime = { path = "crates/aether-usage-runtime" }
+aether-task-core = { path = "crates/aether-task/core" }
+aether-gateway-frontdoor = { path = "crates/aether-gateway/frontdoor" }
+aether-gateway-control = { path = "crates/aether-gateway/control" }
+aether-gateway-execution = { path = "crates/aether-gateway/execution" }
+aether-gateway-workers = { path = "crates/aether-gateway/workers" }
+aether-gateway-tunnel = { path = "crates/aether-gateway/tunnel" }
+aether-loadtools = { path = "crates/aether-testing/loadtools" }
+aether-integration-tests = { path = "crates/aether-testing/integration" }
+aether-test-support = { path = "crates/aether-testing/support" }
+aether-usage-core = { path = "crates/aether-usage/core" }
+aether-usage-runtime = { path = "crates/aether-usage/runtime" }
 aether-video-tasks-core = { path = "crates/aether-video-tasks-core" }
 aether-gateway = { path = "apps/aether-gateway" }
 aether-http = { path = "crates/aether-http" }
-aether-runtime = { path = "crates/aether-runtime" }
+aether-runtime = { path = "crates/aether-runtime/base" }
-aether-testkit = { path = "crates/aether-testkit" }
+aether-testkit = { path = "crates/aether-testing/testkit" }
 aes = "0.8"
 aes-gcm = "0.10"
 async-stream = "0.3"
@@ -77,6 +110,8 @@ bytes = "1"
 cbc = "0.1"
 chrono = { version = "0.4", features = ["serde"] }
 chrono-tz = "0.10"
+crypto_box = { version = "0.9", features = ["seal"] }
+ed25519-dalek = { version = "2.2", features = ["pkcs8"] }
 flate2 = "1"
 futures-util = "0.3"
 hmac = "0.12"
@@ -92,8 +127,9 @@ serde = { version = "1", features = ["derive"] }
 serde_json = { version = "1", features = ["preserve_order"] }
 serde_path_to_error = "0.1"
 sha2 = "0.10"
+socket2 = "0.6"
 tar = "0.4"
-sqlx = { version = "0.8", default-features = false, features = ["postgres", "mysql", "sqlite", "runtime-tokio-rustls", "chrono"] }
+sqlx = { version = "0.8", default-features = false, features = ["runtime-tokio-rustls", "chrono"] }
 thiserror = "2"
 tokio = { version = "1", features = ["macros", "net", "rt-multi-thread", "signal", "sync", "time"] }
 tokio-util = { version = "0.7", features = ["codec", "io-util"] }

Dockerfile.app.local

@@ -62,7 +62,7 @@ COPY --from=gateway-planner /build/recipe.json ./recipe.json
 RUN --mount=type=cache,id=aether-cargo-registry,target=/usr/local/cargo/registry,sharing=locked \
     --mount=type=cache,id=aether-cargo-git,target=/usr/local/cargo/git,sharing=locked \
     --mount=type=cache,id=aether-cargo-target-local,target=/build/target,sharing=locked \
-    cargo chef cook --release --locked --package aether-gateway --bin aether-gateway --recipe-path recipe.json
+    cargo chef cook --release --locked --package aether-gateway --bin aether-gateway --features jemalloc --recipe-path recipe.json
 
 COPY Cargo.toml Cargo.lock ./
 COPY apps/ ./apps/
@@ -71,7 +71,7 @@ RUN --mount=type=cache,id=aether-cargo-registry,target=/usr/local/cargo/registry
     --mount=type=cache,id=aether-cargo-git,target=/usr/local/cargo/git,sharing=locked \
     --mount=type=cache,id=aether-cargo-target-local,target=/build/target,sharing=locked \
     set -eux; \
-    cargo build --release --locked -p aether-gateway --bin aether-gateway; \
+    cargo build --release --locked -p aether-gateway --bin aether-gateway --features jemalloc; \
     cp target/release/aether-gateway /tmp/aether-gateway
 
 # ==================== 最小运行时打包 ====================

Dockerfile.app.release-local

@@ -60,7 +60,7 @@ COPY --from=gateway-planner /build/recipe.json ./recipe.json
 RUN --mount=type=cache,id=aether-cargo-registry,target=/usr/local/cargo/registry,sharing=locked \
     --mount=type=cache,id=aether-cargo-git,target=/usr/local/cargo/git,sharing=locked \
     --mount=type=cache,id=aether-cargo-target-release-local,target=/build/target,sharing=locked \
-    cargo chef cook --release --locked --package aether-gateway --bin aether-gateway --recipe-path recipe.json
+    cargo chef cook --release --locked --package aether-gateway --bin aether-gateway --features jemalloc --recipe-path recipe.json
 
 COPY Cargo.toml Cargo.lock ./
 COPY apps/ ./apps/
@@ -68,7 +68,7 @@ COPY crates/ ./crates/
 RUN --mount=type=cache,id=aether-cargo-registry,target=/usr/local/cargo/registry,sharing=locked \
     --mount=type=cache,id=aether-cargo-git,target=/usr/local/cargo/git,sharing=locked \
     --mount=type=cache,id=aether-cargo-target-release-local,target=/build/target,sharing=locked \
-    cargo build --release --locked -p aether-gateway && \
+    cargo build --release --locked -p aether-gateway --features jemalloc && \
     cp target/release/aether-gateway /tmp/aether-gateway
 
 # ==================== 最小运行时打包 ====================

README.md

@@ -142,8 +142,17 @@ Aether Tunnel 是配套的正向代理节点，部署在海外 VPS 上，为墙
 
 - `APP_PORT`：`aether-gateway` 唯一监听端口，固定绑定 `0.0.0.0:${APP_PORT}`
 - `DATABASE_URL`：数据库连接串；SQLite 例如 `sqlite:///opt/aether/data/aether.db`，Postgres 例如 `postgresql://postgres:aether@postgres:5432/aether`
+- `AETHER_GATEWAY_DATA_POSTGRES_MIN_CONNECTIONS` / `AETHER_GATEWAY_DATA_POSTGRES_MAX_CONNECTIONS`：数据库连接池手动覆盖值；未配置时会自动推导，SQLite 固定 `1/1`，Postgres/MySQL 按 CPU 核心数计算并默认封顶 `100`
+- `AETHER_GATEWAY_MAX_IN_FLIGHT_REQUESTS`：单实例请求并发上限；未配置时按 CPU 自动推导（基础范围 `512-65536`），低文件描述符预算时会进一步下调
+- `AETHER_GATEWAY_REQUEST_BODY_BUFFER_BUDGET_MB`：单实例同时读取和解压请求体的加权内存预算，默认 `256MB`
+- `AETHER_GATEWAY_REQUEST_BODY_READ_TIMEOUT_MS`：请求体完整读取超时，默认 `120000ms`
+- `AETHER_MAX_REQUEST_BODY_MB`：可选的单请求解压后请求体上限；未配置或设为 `0` 时不限制
+- `AETHER_MAX_INTERNAL_BUFFERED_BODY_MB`：可选的 heartbeat、管理探测等内部整包响应体上限；未配置或设为 `0` 时不限制
+- `AETHER_TUNNEL_NODE_STATUS_QUEUE_CAPACITY`：隧道节点状态上报队列容量，默认 `1024`；满载时拒绝新事件，避免控制面故障导致无界内存增长
+- `AETHER_GATEWAY_SECURITY_CACHE_TTL_MS`：IP 黑白名单本地缓存时间，默认 `1000ms`，写操作会主动失效相关缓存
+- `AETHER_MAX_REDACTED_SYNC_RESPONSE_BODY_MB`：可选的 PII 恢复同步响应缓冲上限；未配置或设为 `0` 时不限制
 - `REDIS_URL`：Redis 连接串；仅 Postgres + Redis 的 Docker Compose 部署需要配置
-- `AETHER_RUNTIME_BACKEND=memory|redis`：运行时缓存/协调后端。SQLite 默认用 `memory`，不会连接 Redis
+- `AETHER_RUNTIME_BACKEND=memory|redis`：运行时缓存/协调后端。SQLite 默认用 `memory`，不会连接 Redis；多节点部署和需要跨 gateway 重启恢复 OpenAI Responses continuation history 的部署必须使用共享 Redis
 - `AETHER_GATEWAY_AUTO_PREPARE_DATABASE`：常规启动前自动执行挂起的 schema migration 和 backfill；仓库自带的 `docker-compose.yml` 默认开启
 - `JWT_SECRET_KEY` / `ENCRYPTION_KEY`：认证和敏感数据加密所需密钥
 - `API_KEY_PREFIX`：用户和管理员新建 API Key 时使用的前缀，默认 `sk`
@@ -168,4 +177,4 @@ Aether Tunnel 是配套的正向代理节点，部署在海外 VPS 上，为墙
 
 ## Star History
 
-[![Star History Chart](https://api.star-history.com/svg?repos=fawney19/Aether&type=Date)](https://star-history.com/#fawney19/Aether&Date)
+[![Star History Chart](https://api.star-history.com/svg?repos=fawney19/Aether&type=date&legend=top-left)](https://www.star-history.com/?repos=fawney19%2FAether&type=date&legend=top-left)

apps/aether-gateway/Cargo.toml

@@ -6,6 +6,16 @@ license.workspace = true
 repository.workspace = true
 description = "Rust ingress gateway for Aether phase 3a transparent proxy"
 
+[features]
+default = []
+jemalloc = [
+    "dep:tikv-jemallocator",
+    "dep:tikv-jemalloc-sys",
+    "tikv-jemallocator/stats",
+    "tikv-jemalloc-sys/stats",
+]
+testkit = []
+
 [dependencies]
 aether-admin.workspace = true
 aether-ai-formats.workspace = true
@@ -14,10 +24,15 @@ aether-billing.workspace = true
 aether-cache.workspace = true
 aether-contracts.workspace = true
 aether-crypto.workspace = true
-aether-data.workspace = true
+aether-data = { workspace = true, features = ["all-drivers"] }
 aether-data-contracts.workspace = true
 aether-dispatch-core.workspace = true
+aether-gateway-frontdoor.workspace = true
+aether-gateway-control.workspace = true
+aether-gateway-execution.workspace = true
 aether-http.workspace = true
+aether-gateway-workers.workspace = true
+aether-gateway-tunnel.workspace = true
 aether-model-fetch.workspace = true
 aether-oauth.workspace = true
 aether-pool-core.workspace = true
@@ -46,7 +61,11 @@ flate2.workspace = true
 futures-util.workspace = true
 hmac.workspace = true
 http.workspace = true
+http-body-util = "0.1"
+hyper = { version = "1", features = ["client", "server", "http1", "http2"] }
+hyper-util = { version = "0.1", features = ["client-legacy", "client-pool", "server-auto", "service", "tokio"] }
 ldap3 = { version = "0.11", default-features = false, features = ["sync", "tls-rustls"] }
+libc = "0.2"
 md-5 = "0.10"
 object_store.workspace = true
 parking_lot = "0.12"
@@ -58,8 +77,10 @@ serde.workspace = true
 serde_json.workspace = true
 sha1 = "0.10"
 sha2 = { workspace = true, features = ["oid"] }
+socket2.workspace = true
 tar.workspace = true
-sqlx.workspace = true
+sqlx = { workspace = true, features = ["postgres", "mysql", "sqlite", "migrate"] }
+sysinfo = "0.32"
 thiserror.workspace = true
 tokio.workspace = true
 tokio-util.workspace = true
@@ -74,8 +95,9 @@ wreq-util.workspace = true
 zstd.workspace = true
 
 [target.'cfg(not(target_env = "msvc"))'.dependencies]
-tikv-jemallocator = "0.6"
+tikv-jemallocator = { version = "0.6", optional = true }
+tikv-jemalloc-sys = { version = "0.6", optional = true }
 
 [dev-dependencies]
-aether-testkit.workspace = true
+aether-test-support.workspace = true
 tracing-subscriber.workspace = true

apps/aether-gateway/src/ai_serving/api.rs

@@ -55,16 +55,20 @@ pub(crate) use aether_ai_formats::api::{
     ExecutionRuntimeAuthContext, LocalCoreSyncErrorKind, LocalOpenAiImageSpec,
     LocalSameFormatProviderFamily, LocalSameFormatProviderSpec, LocalStandardSourceFamily,
     LocalStandardSourceMode, LocalStandardSpec, OpenAIChatClientEmitter,
-    OpenAIResponsesClientEmitter, StreamingStandardTerminalObserver,
+    OpenAIResponsesClientEmitter, StreamingStandardTerminalObserver, CLAUDE_CHAT_STREAM_PLAN_KIND,
-    EXECUTION_RUNTIME_STREAM_DECISION_ACTION, EXECUTION_RUNTIME_SYNC_DECISION_ACTION,
+    CLAUDE_CLI_STREAM_PLAN_KIND, EXECUTION_RUNTIME_STREAM_DECISION_ACTION,
-    GEMINI_EMBEDDING_SYNC_PLAN_KIND, GEMINI_FILES_DOWNLOAD_PLAN_KIND,
+    EXECUTION_RUNTIME_SYNC_DECISION_ACTION, GEMINI_CHAT_STREAM_PLAN_KIND,
-    GEMINI_VIDEO_CANCEL_SYNC_PLAN_KIND, OPENAI_EMBEDDING_SYNC_PLAN_KIND,
+    GEMINI_CLI_STREAM_PLAN_KIND, GEMINI_EMBEDDING_SYNC_PLAN_KIND, GEMINI_FILES_DOWNLOAD_PLAN_KIND,
-    OPENAI_IMAGE_STREAM_PLAN_KIND, OPENAI_IMAGE_SYNC_FINALIZE_REPORT_KIND,
+    GEMINI_VIDEO_CANCEL_SYNC_PLAN_KIND, OPENAI_CHAT_STREAM_PLAN_KIND,
-    OPENAI_IMAGE_SYNC_PLAN_KIND, OPENAI_RERANK_SYNC_PLAN_KIND, OPENAI_VIDEO_CANCEL_SYNC_PLAN_KIND,
+    OPENAI_EMBEDDING_SYNC_PLAN_KIND, OPENAI_IMAGE_STREAM_PLAN_KIND,
+    OPENAI_IMAGE_SYNC_FINALIZE_REPORT_KIND, OPENAI_IMAGE_SYNC_PLAN_KIND,
+    OPENAI_RERANK_SYNC_PLAN_KIND, OPENAI_RESPONSES_COMPACT_STREAM_PLAN_KIND,
+    OPENAI_RESPONSES_STREAM_PLAN_KIND, OPENAI_VIDEO_CANCEL_SYNC_PLAN_KIND,
     OPENAI_VIDEO_CONTENT_PLAN_KIND, OPENAI_VIDEO_DELETE_SYNC_PLAN_KIND,
     OPENAI_VIDEO_REMIX_SYNC_PLAN_KIND,
 };
 pub(crate) use aether_ai_formats::protocol::stream::CanonicalUsage as StreamingCanonicalUsage;
+pub(crate) use aether_ai_formats::CODEX_RESPONSES_LITE_HEADER;
 
 pub(crate) fn parse_direct_request_body(
     parts: &http::request::Parts,
@@ -83,28 +87,36 @@ pub(crate) fn resolve_execution_runtime_stream_plan_kind(
     parts: &http::request::Parts,
     decision: &GatewayControlDecision,
 ) -> Option<&'static str> {
-    aether_ai_formats::api::resolve_execution_runtime_stream_plan_kind(
+    let plan_kind =
-        decision.route_class.as_deref(),
+        aether_ai_formats::api::resolve_execution_runtime_stream_plan_kind_with_client_surface(
-        decision.route_family.as_deref(),
+            decision.route_class.as_deref(),
-        decision.route_kind.as_deref(),
+            decision.route_family.as_deref(),
-        decision.request_auth_channel.as_deref(),
+            decision.route_kind.as_deref(),
-        &parts.method,
+            decision.client_surface,
-        parts.uri.path(),
+            decision.request_auth_channel.as_deref(),
-    )
+            &parts.method,
+            parts.uri.path(),
+        )?;
+    crate::ai_serving::plan_kind_matches_api_operation(plan_kind, true, decision.api_operation)
+        .then_some(plan_kind)
 }
 
 pub(crate) fn resolve_execution_runtime_sync_plan_kind(
     parts: &http::request::Parts,
     decision: &GatewayControlDecision,
 ) -> Option<&'static str> {
-    aether_ai_formats::api::resolve_execution_runtime_sync_plan_kind(
+    let plan_kind =
-        decision.route_class.as_deref(),
+        aether_ai_formats::api::resolve_execution_runtime_sync_plan_kind_with_client_surface(
-        decision.route_family.as_deref(),
+            decision.route_class.as_deref(),
-        decision.route_kind.as_deref(),
+            decision.route_family.as_deref(),
-        decision.request_auth_channel.as_deref(),
+            decision.route_kind.as_deref(),
-        &parts.method,
+            decision.client_surface,
-        parts.uri.path(),
+            decision.request_auth_channel.as_deref(),
-    )
+            &parts.method,
+            parts.uri.path(),
+        )?;
+    crate::ai_serving::plan_kind_matches_api_operation(plan_kind, false, decision.api_operation)
+        .then_some(plan_kind)
 }
 
 pub(crate) fn is_matching_stream_request(

apps/aether-gateway/src/ai_serving/finalize/internal/stream_rewrite.rs

@@ -2,6 +2,7 @@ use serde_json::Value;
 
 use crate::ai_serving::{
     maybe_build_ai_surface_stream_rewriter, AiSurfaceFinalizeError, AiSurfaceStreamRewriter,
+    ResponseHistoryRecord,
 };
 use crate::GatewayError;
 
@@ -24,6 +25,10 @@ impl LocalStreamRewriter<'_> {
     pub(crate) fn finish(&mut self) -> Result<Vec<u8>, GatewayError> {
         self.inner.finish().map_err(map_surface_error)
     }
+
+    pub(crate) fn take_response_history_record(&mut self) -> Option<ResponseHistoryRecord> {
+        self.inner.take_response_history_record()
+    }
 }
 
 fn map_surface_error(error: AiSurfaceFinalizeError) -> GatewayError {

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

@@ -13,6 +13,7 @@ fn same_format_claude_local_stream_rewriter_sanitizes_read_input_json_delta() {
     let report_context = json!({
         "provider_api_format": "claude:messages",
         "client_api_format": "claude:messages",
+        "anthropic_compatibility_profile": "claude_code_legacy",
         "needs_conversion": false,
     });
     let mut rewriter =

apps/aether-gateway/src/ai_serving/finalize/tests_sync.rs

@@ -25,12 +25,16 @@ fn test_decision() -> GatewayControlDecision {
         route_class: Some("ai_public".to_string()),
         route_family: Some("openai".to_string()),
         route_kind: Some("compact".to_string()),
+        client_surface: None,
+        api_operation: None,
+        gateway_credential_carrier: None,
         request_auth_channel: None,
         auth_endpoint_signature: Some("openai:responses:compact".to_string()),
         execution_runtime_candidate: true,
         auth_context: None,
         admin_principal: None,
         local_auth_rejection: None,
+        model_directive_policy: Default::default(),
     }
 }
 
@@ -1166,7 +1170,7 @@ fn local_finalize_handles_openai_responses_compact_cross_format_sync_response()
     assert_eq!(report.report_kind, "openai_responses_compact_sync_success");
     assert_eq!(
         report.client_body_json.expect("client body should exist")["object"],
-        "response"
+        "response.compaction"
     );
 }
 
@@ -1222,7 +1226,7 @@ fn local_finalize_handles_openai_responses_compact_cross_format_function_call_re
         .background_report
         .expect("compact tool-call should downgrade to success report");
     let client_body = report.client_body_json.expect("client body should exist");
-    assert_eq!(client_body["object"], "response");
+    assert_eq!(client_body["object"], "response.compaction");
     assert_eq!(client_body["output"][1]["type"], "function_call");
 }
 
@@ -1439,6 +1443,57 @@ fn local_finalize_handles_openai_responses_cross_format_stream_response_from_gem
     );
 }
 
+#[test]
+fn local_finalize_rejects_antigravity_usage_only_gemini_wrapper() {
+    let payload = GatewaySyncReportRequest {
+        trace_id: "trace-antigravity-empty-gemini-wrapper".to_string(),
+        report_kind: "gemini_chat_sync_finalize".to_string(),
+        report_context: Some(json!({
+            "client_api_format": "gemini:generate_content",
+            "provider_api_format": "gemini:generate_content",
+            "model": "gemini-3.5-flash",
+            "mapped_model": "gemini-3-flash-agent",
+            "needs_conversion": false,
+            "has_envelope": true,
+            "envelope_name": "antigravity:v1internal",
+            "upstream_is_stream": true,
+        })),
+        status_code: 200,
+        headers: BTreeMap::from([("content-type".to_string(), "application/json".to_string())]),
+        body_json: Some(json!({
+            "chunks": [{
+                "response": {
+                    "responseId": "resp-usage-only",
+                    "modelVersion": "gemini-3-flash-agent",
+                    "usageMetadata": {
+                        "promptTokenCount": 5528,
+                        "totalTokenCount": 5528
+                    }
+                },
+                "metadata": {},
+                "traceId": "trace-antigravity-empty-gemini-wrapper"
+            }],
+            "metadata": {
+                "stream": true,
+                "stored_chunks": 1,
+                "total_chunks": 1
+            }
+        })),
+        client_body_json: None,
+        body_base64: None,
+        telemetry: None,
+    };
+
+    let outcome = maybe_build_local_core_sync_finalize_response(
+        "trace-antigravity-empty-gemini-wrapper",
+        &test_decision(),
+        &payload,
+    )
+    .expect("local finalize should evaluate payload");
+
+    assert!(outcome.is_none());
+}
+
 #[test]
 fn local_finalize_handles_openai_responses_compact_openai_family_stream_response_even_when_conversion_flagged(
 ) {
@@ -1738,6 +1793,93 @@ fn local_finalize_handles_openai_chat_cross_format_sync_response_from_openai_res
     assert_eq!(client_body["usage"]["total_tokens"], 5);
 }
 
+#[test]
+fn local_finalize_aggregates_openai_responses_capture_envelope_before_chat_conversion() {
+    let payload = GatewaySyncReportRequest {
+        trace_id: "trace-openai-chat-capture-envelope-sync-123".to_string(),
+        report_kind: "openai_chat_sync_finalize".to_string(),
+        report_context: Some(json!({
+            "client_api_format": "openai:chat",
+            "provider_api_format": "openai:responses",
+            "model": "gpt-5.6-luna",
+            "mapped_model": "gpt-5.6-luna",
+            "needs_conversion": true,
+            "has_envelope": false,
+        })),
+        status_code: 200,
+        headers: BTreeMap::from([("content-type".to_string(), "application/json".to_string())]),
+        body_json: Some(json!({
+            "chunks": [
+                {
+                    "type": "response.output_text.delta",
+                    "response_id": "resp_capture_gateway_123",
+                    "output_index": 0,
+                    "content_index": 0,
+                    "delta": "Gateway "
+                },
+                {
+                    "type": "response.output_text.done",
+                    "response_id": "resp_capture_gateway_123",
+                    "output_index": 0,
+                    "content_index": 0,
+                    "text": "Gateway capture"
+                },
+                {
+                    "type": "response.completed",
+                    "response": {
+                        "id": "resp_capture_gateway_123",
+                        "object": "response",
+                        "status": "completed",
+                        "model": "gpt-5.6-luna",
+                        "output": [],
+                        "usage": {
+                            "input_tokens": 2,
+                            "output_tokens": 3,
+                            "total_tokens": 5
+                        }
+                    }
+                }
+            ],
+            "metadata": {}
+        })),
+        client_body_json: None,
+        body_base64: None,
+        telemetry: None,
+    };
+
+    let outcome = maybe_build_local_core_sync_finalize_response(
+        "trace-openai-chat-capture-envelope-sync-123",
+        &test_decision(),
+        &payload,
+    )
+    .expect("capture envelope finalize should succeed")
+    .expect("capture envelope finalize should match");
+
+    let report = outcome
+        .background_report
+        .expect("capture envelope conversion should produce a success report");
+    assert_eq!(report.report_kind, "openai_chat_sync_success");
+    let provider_body = report
+        .body_json
+        .expect("aggregated provider body should exist");
+    assert_eq!(provider_body["id"], "resp_capture_gateway_123");
+    assert_eq!(
+        provider_body["output"][0]["content"][0]["text"],
+        "Gateway capture"
+    );
+    assert!(provider_body.get("chunks").is_none());
+    let client_body = report
+        .client_body_json
+        .expect("converted client body should exist");
+    assert_eq!(
+        client_body["choices"][0]["message"]["content"],
+        "Gateway capture"
+    );
+    assert_eq!(client_body["usage"]["prompt_tokens"], 2);
+    assert_eq!(client_body["usage"]["completion_tokens"], 3);
+    assert_eq!(client_body["usage"]["total_tokens"], 5);
+}
+
 #[test]
 fn local_finalize_handles_claude_chat_cross_format_sync_response_from_openai_chat() {
     let payload = GatewaySyncReportRequest {
@@ -1784,12 +1926,16 @@ fn local_finalize_handles_claude_chat_cross_format_sync_response_from_openai_cha
             route_class: Some("ai_public".to_string()),
             route_family: Some("claude".to_string()),
             route_kind: Some("chat".to_string()),
+            client_surface: None,
+            api_operation: None,
+            gateway_credential_carrier: None,
             request_auth_channel: None,
             auth_endpoint_signature: Some("claude:messages".to_string()),
             execution_runtime_candidate: true,
             auth_context: None,
             admin_principal: None,
             local_auth_rejection: None,
+            model_directive_policy: Default::default(),
         },
         &payload,
     )
@@ -1851,12 +1997,16 @@ fn local_finalize_handles_gemini_cli_cross_format_sync_response_from_claude_cli(
             route_class: Some("ai_public".to_string()),
             route_family: Some("gemini".to_string()),
             route_kind: Some("cli".to_string()),
+            client_surface: None,
+            api_operation: None,
+            gateway_credential_carrier: None,
             request_auth_channel: None,
             auth_endpoint_signature: Some("gemini:generate_content".to_string()),
             execution_runtime_candidate: true,
             auth_context: None,
             admin_principal: None,
             local_auth_rejection: None,
+            model_directive_policy: Default::default(),
         },
         &payload,
     )

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

@@ -3,6 +3,7 @@ pub(crate) mod api;
 mod finalize;
 mod planner;
 mod pure;
+mod response_history;
 pub(crate) mod transport;
 
 use axum::body::Body;
@@ -13,7 +14,9 @@ use crate::{usage::GatewaySyncReportRequest, AppState, GatewayError};
 pub(crate) use self::adaptation::{
     maybe_build_provider_private_stream_normalizer, ProviderPrivateStreamNormalizer,
 };
-pub(crate) use self::api::gemini_generate_content_response_has_visible_output;
+pub(crate) use self::api::{
+    gemini_generate_content_response_has_visible_output, CODEX_RESPONSES_LITE_HEADER,
+};
 pub(crate) use self::finalize::common::LocalCoreSyncFinalizeOutcome;
 pub(crate) use self::finalize::internal::{
     maybe_bridge_standard_sync_json_to_stream, maybe_build_stream_response_rewriter,
@@ -48,18 +51,24 @@ pub(crate) use self::planner::{
     build_standard_family_stream_plan_and_reports, build_standard_family_sync_attempt_source,
     build_standard_family_sync_plan_and_reports, build_standard_stream_plan_from_decision,
     build_standard_sync_plan_from_decision, candidate_auth_channel_skip_reason,
-    extract_pool_sticky_session_token, maybe_build_stream_decision_payload,
+    codex_model_capabilities_for_transport, extract_pool_sticky_session_token,
-    maybe_build_stream_plan_payload, maybe_build_sync_decision_payload,
+    maybe_build_stream_decision_payload, maybe_build_stream_plan_payload,
-    maybe_build_sync_plan_payload, planner_is_matching_stream_request, provider_key_pool_score_id,
+    maybe_build_sync_decision_payload, maybe_build_sync_plan_payload,
-    provider_key_pool_score_scope, read_candidate_transport_snapshot,
+    planner_is_matching_stream_request, provider_key_pool_score_id, provider_key_pool_score_scope,
-    record_local_runtime_candidate_skip_reason,
+    read_candidate_transport_snapshot, record_local_runtime_candidate_skip_reason,
+    resolve_tunnel_scheduler_affinity_context, resolve_upstream_is_stream_for_provider,
     set_local_openai_chat_execution_exhausted_diagnostic,
-    set_local_openai_image_execution_exhausted_diagnostic, CandidateFailureDiagnostic,
+    set_local_openai_image_execution_exhausted_diagnostic, validate_final_openai_provider_request,
-    CandidateFailureDiagnosticKind, EligibleLocalExecutionCandidate, GatewayAuthApiKeySnapshot,
+    CandidateFailureDiagnostic, CandidateFailureDiagnosticKind, EligibleLocalExecutionCandidate,
-    GatewayProviderTransportSnapshot, LocalExecutionAttemptSource, LocalExecutionCandidateKind,
+    GatewayAuthApiKeySnapshot, GatewayProviderTransportSnapshot, LocalExecutionAttemptSource,
-    LocalResolvedOAuthRequestAuth, PlannerAppState, SkippedLocalExecutionCandidate,
+    LocalExecutionCandidateKind, LocalResolvedOAuthRequestAuth, PlannerAppState,
+    SkippedLocalExecutionCandidate,
 };
 pub(crate) use self::pure::*;
+pub(crate) use self::response_history::{
+    hydrate_openai_response_history, persist_converted_response_history,
+    persist_response_history_record,
+};
 pub(crate) use self::transport::{
     append_transport_diagnostics_to_value, build_request_trace_proxy_value,
     candidate_common_transport_skip_reason, candidate_transport_pair_skip_reason,
@@ -68,7 +77,7 @@ pub(crate) use self::transport::{
     request_pair_allowed_for_transport, request_pair_direct_auth,
     request_pair_transport_unsupported_reason, CandidateTransportPolicyFacts,
 };
-pub(crate) use crate::control::GatewayControlDecision;
+pub(crate) use crate::control::{GatewayControlDecision, GatewayCredentialCarrier};
 pub(crate) use crate::execution_runtime::{ConversionMode, ExecutionStrategy};
 pub(crate) use crate::headers::RequestOrigin;
 pub(crate) use aether_ai_serving::{
@@ -86,6 +95,7 @@ pub(crate) fn build_provider_transport_request_url(
     upstream_is_stream: bool,
     request_query: Option<&str>,
     kiro_api_region: Option<&str>,
+    api_operation: Option<ApiOperation>,
 ) -> Option<String> {
     self::transport::build_transport_request_url(
         transport,
@@ -95,6 +105,7 @@ pub(crate) fn build_provider_transport_request_url(
             upstream_is_stream,
             request_query,
             kiro_api_region,
+            api_operation,
         },
     )
 }
@@ -106,6 +117,7 @@ pub(crate) fn build_provider_transport_request_url_for_request_body(
     upstream_is_stream: bool,
     request_query: Option<&str>,
     kiro_api_region: Option<&str>,
+    api_operation: Option<ApiOperation>,
     provider_request_body: Option<&serde_json::Value>,
 ) -> Option<String> {
     self::transport::build_transport_request_url_for_request_body(
@@ -116,6 +128,7 @@ pub(crate) fn build_provider_transport_request_url_for_request_body(
             upstream_is_stream,
             request_query,
             kiro_api_region,
+            api_operation,
         },
         provider_request_body,
     )

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

@@ -0,0 +1,170 @@
+use std::collections::BTreeMap;
+use std::sync::Arc;
+
+use serde_json::Value;
+
+use crate::ai_serving::transport::antigravity::{
+    build_antigravity_safe_v1internal_request, build_antigravity_static_identity_headers,
+    classify_local_antigravity_request_support, AntigravityEnvelopeRequestType,
+    AntigravityRequestAuth, AntigravityRequestAuthUnsupportedReason,
+    AntigravityRequestEnvelopeSupport, AntigravityRequestSideSupport,
+    AntigravityRequestSideUnsupportedReason,
+};
+use crate::ai_serving::transport::{
+    build_standard_provider_request_headers, GatewayProviderTransportSnapshot,
+    StandardProviderRequestHeaders, StandardProviderRequestHeadersInput,
+};
+use crate::AppState;
+
+pub(crate) const ANTIGRAVITY_V1INTERNAL_ENVELOPE_NAME: &str = "antigravity:v1internal";
+
+pub(crate) enum AntigravityV1InternalRequestError {
+    TransportUnsupported,
+    EnvelopeUnsupported,
+    UpstreamUrlUnavailable,
+    HeaderRulesApplyFailed,
+}
+
+pub(crate) struct AntigravityV1InternalRequestInput<'a> {
+    pub(crate) state: &'a AppState,
+    pub(crate) parts: &'a http::request::Parts,
+    pub(crate) transport: &'a Arc<GatewayProviderTransportSnapshot>,
+    pub(crate) trace_id: &'a str,
+    pub(crate) mapped_model: &'a str,
+    pub(crate) provider_api_format: &'a str,
+    pub(crate) auth_header: &'a str,
+    pub(crate) auth_value: &'a str,
+    pub(crate) request_headers: &'a http::HeaderMap,
+    pub(crate) original_request_body: &'a Value,
+    pub(crate) gemini_request_body: &'a Value,
+    pub(crate) upstream_is_stream: bool,
+    pub(crate) same_format: bool,
+}
+
+pub(crate) struct AntigravityV1InternalRequest {
+    pub(crate) transport: Arc<GatewayProviderTransportSnapshot>,
+    pub(crate) body: Value,
+    pub(crate) headers: StandardProviderRequestHeaders,
+    pub(crate) upstream_url: String,
+}
+
+pub(crate) async fn build_antigravity_v1internal_provider_request(
+    input: AntigravityV1InternalRequestInput<'_>,
+) -> Result<AntigravityV1InternalRequest, AntigravityV1InternalRequestError> {
+    let payload = build_antigravity_v1internal_payload(
+        input.state,
+        input.transport,
+        input.trace_id,
+        input.mapped_model,
+        input.gemini_request_body,
+    )
+    .await?;
+
+    let upstream_url = crate::ai_serving::build_provider_transport_request_url_for_request_body(
+        &payload.transport,
+        input.provider_api_format,
+        Some(input.mapped_model),
+        input.upstream_is_stream,
+        input.parts.uri.query(),
+        None,
+        None,
+        Some(&payload.body),
+    )
+    .ok_or(AntigravityV1InternalRequestError::UpstreamUrlUnavailable)?;
+
+    let extra_headers: BTreeMap<String, String> =
+        build_antigravity_static_identity_headers(&payload.auth);
+    let mut headers =
+        build_standard_provider_request_headers(StandardProviderRequestHeadersInput {
+            transport: &payload.transport,
+            provider_api_format: input.provider_api_format,
+            same_format: input.same_format,
+            headers: input.request_headers,
+            auth_header: input.auth_header,
+            auth_value: input.auth_value,
+            extra_headers: &extra_headers,
+            header_rules: payload.transport.endpoint.header_rules.as_ref(),
+            provider_request_body: &payload.body,
+            original_request_body: input.original_request_body,
+            upstream_is_stream: input.upstream_is_stream,
+        })
+        .ok_or(AntigravityV1InternalRequestError::HeaderRulesApplyFailed)?;
+    headers
+        .headers
+        .insert("accept".to_string(), "text/event-stream".to_string());
+
+    Ok(AntigravityV1InternalRequest {
+        transport: payload.transport,
+        body: payload.body,
+        headers,
+        upstream_url,
+    })
+}
+
+struct AntigravityV1InternalPayload {
+    transport: Arc<GatewayProviderTransportSnapshot>,
+    auth: AntigravityRequestAuth,
+    body: Value,
+}
+
+async fn build_antigravity_v1internal_payload(
+    state: &AppState,
+    transport: &Arc<GatewayProviderTransportSnapshot>,
+    trace_id: &str,
+    mapped_model: &str,
+    gemini_request_body: &Value,
+) -> Result<AntigravityV1InternalPayload, AntigravityV1InternalRequestError> {
+    let mut resolved_transport = Arc::clone(transport);
+    let mut antigravity_support = classify_local_antigravity_request_support(
+        &resolved_transport,
+        gemini_request_body,
+        AntigravityEnvelopeRequestType::Agent,
+    );
+
+    if matches!(
+        antigravity_support,
+        AntigravityRequestSideSupport::Unsupported(
+            AntigravityRequestSideUnsupportedReason::UnsupportedAuth(
+                AntigravityRequestAuthUnsupportedReason::MissingProjectId
+            )
+        )
+    ) {
+        if let Some(hydrated) = state
+            .hydrate_antigravity_project_metadata_for_transport(&resolved_transport)
+            .await
+        {
+            resolved_transport = Arc::new(hydrated);
+            antigravity_support = classify_local_antigravity_request_support(
+                &resolved_transport,
+                gemini_request_body,
+                AntigravityEnvelopeRequestType::Agent,
+            );
+        }
+    }
+
+    let auth = match antigravity_support {
+        AntigravityRequestSideSupport::Supported(spec) => spec.auth,
+        AntigravityRequestSideSupport::Unsupported(_) => {
+            return Err(AntigravityV1InternalRequestError::TransportUnsupported);
+        }
+    };
+
+    let body = match build_antigravity_safe_v1internal_request(
+        &auth,
+        trace_id,
+        mapped_model,
+        gemini_request_body,
+        AntigravityEnvelopeRequestType::Agent,
+    ) {
+        AntigravityRequestEnvelopeSupport::Supported(envelope) => envelope,
+        AntigravityRequestEnvelopeSupport::Unsupported(_) => {
+            return Err(AntigravityV1InternalRequestError::EnvelopeUnsupported);
+        }
+    };
+
+    Ok(AntigravityV1InternalPayload {
+        transport: resolved_transport,
+        auth,
+        body,
+    })
+}

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

@@ -1,6 +1,8 @@
+use aether_routing_core::ResolvedRoutingPolicy;
 use aether_scheduler_core::{
-    build_scheduler_affinity_cache_key_for_api_key_id_with_client_session, ClientSessionAffinity,
+    build_scheduler_affinity_cache_key_for_api_key_id_with_client_session_and_scope,
-    SchedulerAffinityTarget, SchedulerMinimalCandidateSelectionCandidate,
+    ClientSessionAffinity, SchedulerAffinityScope, SchedulerAffinityTarget,
+    SchedulerMinimalCandidateSelectionCandidate,
 };
 
 use crate::ai_serving::{GatewayAuthApiKeySnapshot, PlannerAppState};
@@ -8,25 +10,38 @@ use crate::scheduler::affinity::SCHEDULER_AFFINITY_TTL;
 
 const PLANNER_SCHEDULER_AFFINITY_MAX_ENTRIES: usize = 10_000;
 
+pub(crate) fn has_explicit_session_affinity(
+    client_session_affinity: Option<&ClientSessionAffinity>,
+) -> bool {
+    client_session_affinity.is_some_and(ClientSessionAffinity::has_session_key)
+}
+
 pub(crate) fn read_cached_scheduler_affinity_target(
     state: PlannerAppState<'_>,
     auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
     client_session_affinity: Option<&ClientSessionAffinity>,
     client_api_format: &str,
     requested_model: Option<&str>,
+    routing_policy: Option<&ResolvedRoutingPolicy>,
 ) -> Option<SchedulerAffinityTarget> {
+    if !has_explicit_session_affinity(client_session_affinity) {
+        return None;
+    }
     let requested_model = requested_model
         .map(str::trim)
         .filter(|value| !value.is_empty())?;
     let api_key_id = auth_snapshot
         .map(|snapshot| snapshot.api_key_id.trim())
         .filter(|value| !value.is_empty())?;
-    let cache_key = build_scheduler_affinity_cache_key_for_api_key_id_with_client_session(
+    let affinity_scope = scheduler_affinity_scope_for_routing_policy(routing_policy);
-        api_key_id,
+    let cache_key =
-        client_api_format,
+        build_scheduler_affinity_cache_key_for_api_key_id_with_client_session_and_scope(
-        requested_model,
+            api_key_id,
-        client_session_affinity,
+            client_api_format,
-    )?;
+            requested_model,
+            client_session_affinity,
+            affinity_scope.as_ref(),
+        )?;
 
     state
         .app()
@@ -41,6 +56,9 @@ pub(crate) fn remember_scheduler_affinity_for_candidate(
     requested_model: &str,
     candidate: &SchedulerMinimalCandidateSelectionCandidate,
 ) {
+    if !has_explicit_session_affinity(client_session_affinity) {
+        return;
+    }
     remember_scheduler_affinity_for_candidate_at_epoch(
         state,
         auth_snapshot,
@@ -61,18 +79,71 @@ pub(crate) fn remember_scheduler_affinity_for_candidate_at_epoch(
     candidate: &SchedulerMinimalCandidateSelectionCandidate,
     expected_epoch: Option<u64>,
 ) {
+    remember_scheduler_affinity_for_candidate_with_scope_at_epoch(
+        state,
+        auth_snapshot,
+        client_session_affinity,
+        client_api_format,
+        requested_model,
+        candidate,
+        None,
+        expected_epoch,
+    );
+}
+
+#[allow(clippy::too_many_arguments)]
+pub(crate) fn remember_scheduler_affinity_for_candidate_with_routing_policy_at_epoch(
+    state: PlannerAppState<'_>,
+    auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
+    client_session_affinity: Option<&ClientSessionAffinity>,
+    client_api_format: &str,
+    requested_model: &str,
+    candidate: &SchedulerMinimalCandidateSelectionCandidate,
+    routing_policy: Option<&ResolvedRoutingPolicy>,
+    expected_epoch: Option<u64>,
+) {
+    let affinity_scope = scheduler_affinity_scope_for_routing_policy(routing_policy);
+    remember_scheduler_affinity_for_candidate_with_scope_at_epoch(
+        state,
+        auth_snapshot,
+        client_session_affinity,
+        client_api_format,
+        requested_model,
+        candidate,
+        affinity_scope.as_ref(),
+        expected_epoch,
+    );
+}
+
+#[allow(clippy::too_many_arguments)]
+fn remember_scheduler_affinity_for_candidate_with_scope_at_epoch(
+    state: PlannerAppState<'_>,
+    auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
+    client_session_affinity: Option<&ClientSessionAffinity>,
+    client_api_format: &str,
+    requested_model: &str,
+    candidate: &SchedulerMinimalCandidateSelectionCandidate,
+    affinity_scope: Option<&SchedulerAffinityScope>,
+    expected_epoch: Option<u64>,
+) {
+    if !has_explicit_session_affinity(client_session_affinity) {
+        return;
+    }
     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_with_client_session(
+    let Some(cache_key) =
-        api_key_id,
+        build_scheduler_affinity_cache_key_for_api_key_id_with_client_session_and_scope(
-        client_api_format,
+            api_key_id,
-        requested_model,
+            client_api_format,
-        client_session_affinity,
+            requested_model,
-    ) else {
+            client_session_affinity,
+            affinity_scope,
+        )
+    else {
         return;
     };
 
@@ -88,3 +159,15 @@ pub(crate) fn remember_scheduler_affinity_for_candidate_at_epoch(
         expected_epoch,
     );
 }
+
+fn scheduler_affinity_scope_for_routing_policy(
+    routing_policy: Option<&ResolvedRoutingPolicy>,
+) -> Option<SchedulerAffinityScope> {
+    let policy = routing_policy?;
+    let group_id = policy
+        .group_id
+        .as_deref()
+        .map(str::trim)
+        .filter(|group_id| !group_id.is_empty())?;
+    Some(SchedulerAffinityScope::new(group_id, policy.group_version))
+}

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

@@ -134,6 +134,7 @@ mod tests {
             global_model_id: "global-1".to_string(),
             global_model_name: "gpt-5.4".to_string(),
             selected_provider_model_name: "gpt-5.4".to_string(),
+            supports_streaming: true,
             mapping_matched_model: None,
         }
     }

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

@@ -169,6 +169,7 @@ mod tests {
             global_model_id: "global-model-1".to_string(),
             global_model_name: "gpt-test".to_string(),
             selected_provider_model_name: "gpt-test-upstream".to_string(),
+            supports_streaming: true,
             mapping_matched_model: None,
         }
     }

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

@@ -1,5 +1,3 @@
-use std::collections::BTreeMap;
-
 use aether_ai_serving::{
     ai_ranking_context, build_ai_rankable_candidate, run_ai_candidate_ranking,
     AiCandidateRankingPort, AiRankableCandidateParts, AiRankingContextConfig,
@@ -7,6 +5,7 @@ use aether_ai_serving::{
 };
 use aether_routing_core::{ResolvedRoutingPolicy, RoutingSchedulingMode, RoutingSetPriorityMode};
 use async_trait::async_trait;
+use tokio::sync::Mutex;
 use tracing::warn;
 
 use crate::ai_serving::{GatewayAuthApiKeySnapshot, PlannerAppState};
@@ -24,7 +23,7 @@ use aether_scheduler_core::{
 use super::candidate_affinity_cache::read_cached_scheduler_affinity_target;
 use super::candidate_resolution::{EligibleLocalExecutionCandidate, LocalExecutionCandidateKind};
 use super::candidate_transport_ranking_facts::{
-    resolve_cached_transport_ranking_facts, CandidateTransportRankingFacts,
+    resolve_cached_transport_ranking_facts, CandidateTransportRankingFactsCache,
 };
 
 struct GatewayLocalCandidateRankingPort<'a> {
@@ -35,6 +34,7 @@ struct GatewayLocalCandidateRankingPort<'a> {
     required_capabilities: Option<&'a serde_json::Value>,
     ordering_config: SchedulerOrderingConfig,
     routing_policy: Option<&'a ResolvedRoutingPolicy>,
+    transport_ranking_facts_cache: Mutex<CandidateTransportRankingFactsCache>,
 }
 
 #[async_trait]
@@ -66,6 +66,7 @@ impl AiCandidateRankingPort for GatewayLocalCandidateRankingPort<'_> {
             self.client_session_affinity,
             normalized_client_api_format,
             affinity_requested_model,
+            self.routing_policy,
         ))
     }
 
@@ -84,13 +85,17 @@ impl AiCandidateRankingPort for GatewayLocalCandidateRankingPort<'_> {
         normalized_client_api_format: &str,
         cached_affinity_match: bool,
     ) -> Result<SchedulerRankableCandidate, Self::Error> {
-        let ranking_facts = resolve_transport_ranking_facts_for_candidate(
+        let ranking_facts = {
-            self.state,
+            let mut cache = self.transport_ranking_facts_cache.lock().await;
-            &candidate.candidate,
+            resolve_cached_transport_ranking_facts(
-            candidate.transport.as_ref(),
+                self.state,
-            self.ordering_config,
+                &mut cache,
-        )
+                &candidate.candidate,
-        .await;
+                candidate.transport.as_ref(),
+                self.ordering_config,
+            )
+            .await
+        };
         let routing_overlaid_candidate =
             routing_overlaid_candidate(self.routing_policy, candidate.kind, &candidate.candidate);
         Ok(build_ai_rankable_candidate(AiRankableCandidateParts {
@@ -137,6 +142,7 @@ pub(crate) async fn rank_eligible_local_execution_candidates(
         required_capabilities,
         ordering_config,
         routing_policy,
+        transport_ranking_facts_cache: Mutex::new(CandidateTransportRankingFactsCache::default()),
     };
 
     match run_ai_candidate_ranking(&port, candidates, normalized_client_api_format).await {
@@ -145,23 +151,6 @@ pub(crate) async fn rank_eligible_local_execution_candidates(
     }
 }
 
-async fn resolve_transport_ranking_facts_for_candidate(
-    state: PlannerAppState<'_>,
-    candidate: &SchedulerMinimalCandidateSelectionCandidate,
-    transport: &crate::ai_serving::GatewayProviderTransportSnapshot,
-    ordering_config: SchedulerOrderingConfig,
-) -> CandidateTransportRankingFacts {
-    let mut ordering_cache = BTreeMap::new();
-    resolve_cached_transport_ranking_facts(
-        state,
-        &mut ordering_cache,
-        candidate,
-        transport,
-        ordering_config,
-    )
-    .await
-}
-
 fn cached_affinity_matches_local_execution_scope(
     eligible: &EligibleLocalExecutionCandidate,
     target: &SchedulerAffinityTarget,
@@ -233,17 +222,21 @@ fn routing_overlaid_candidate(
     let mut overlaid = candidate.clone();
     overlaid.provider_priority = policy
         .ranking_overlay
-        .provider_priority_or_unspecified(candidate.provider_id.as_str());
+        .provider_priority(candidate.provider_id.as_str(), candidate.provider_priority);
     let overlaid_key_priority = match kind {
         LocalExecutionCandidateKind::SingleKey => policy
             .ranking_overlay
-            .key_priority_or_unspecified(candidate.key_id.as_str()),
+            .key_priority_overrides
+            .get(candidate.key_id.as_str()),
         LocalExecutionCandidateKind::PoolGroup => policy
             .ranking_overlay
-            .pool_priority_or_unspecified(candidate.provider_id.as_str()),
+            .pool_priority_overrides
+            .get(candidate.provider_id.as_str()),
     };
-    overlaid.key_internal_priority = overlaid_key_priority;
+    if let Some(overlaid_key_priority) = overlaid_key_priority.copied() {
-    overlaid.key_global_priority_for_format = Some(overlaid_key_priority);
+        overlaid.key_internal_priority = overlaid_key_priority;
+        overlaid.key_global_priority_for_format = Some(overlaid_key_priority);
+    }
     overlaid
 }
 
@@ -284,7 +277,9 @@ mod tests {
     use serde_json::json;
 
     use super::super::candidate_affinity_cache::remember_scheduler_affinity_for_candidate;
-    use super::super::candidate_transport_ranking_facts::resolve_cached_candidate_transport_ranking_facts;
+    use super::super::candidate_transport_ranking_facts::{
+        resolve_cached_candidate_transport_ranking_facts, CandidateTransportRankingFactsCache,
+    };
     use super::{PlannerAppState, SchedulerMinimalCandidateSelectionCandidate};
     use crate::ai_serving::planner::candidate_resolution::{
         resolve_and_rank_local_execution_candidates,
@@ -306,7 +301,7 @@ mod tests {
         let ordering_config = super::read_scheduler_ordering_config_or_default(state).await;
         let mut candidates = candidates;
         let mut rankables = Vec::with_capacity(candidates.len());
-        let mut ordering_cache = BTreeMap::new();
+        let mut ordering_cache = CandidateTransportRankingFactsCache::default();
 
         for (original_index, candidate) in candidates.iter().enumerate() {
             let ranking_facts = resolve_cached_candidate_transport_ranking_facts(
@@ -358,12 +353,13 @@ mod tests {
             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(),
+            supports_streaming: true,
             mapping_matched_model: None,
         }
     }
 
     #[test]
-    fn routing_policy_priorities_do_not_fall_back_to_candidate_priorities() {
+    fn routing_policy_priorities_fall_back_to_candidate_priorities() {
         let mut candidate = sample_candidate("endpoint-1", "key-1");
         candidate.provider_priority = 7;
         candidate.key_internal_priority = 3;
@@ -389,18 +385,9 @@ mod tests {
             &candidate,
         );
 
-        assert_eq!(
+        assert_eq!(overlaid.provider_priority, 7);
-            overlaid.provider_priority,
+        assert_eq!(overlaid.key_internal_priority, 3);
-            aether_routing_core::ROUTING_PRIORITY_UNSPECIFIED
+        assert_eq!(overlaid.key_global_priority_for_format, Some(2));
-        );
-        assert_eq!(
-            overlaid.key_internal_priority,
-            aether_routing_core::ROUTING_PRIORITY_UNSPECIFIED
-        );
-        assert_eq!(
-            overlaid.key_global_priority_for_format,
-            Some(aether_routing_core::ROUTING_PRIORITY_UNSPECIFIED)
-        );
     }
 
     #[test]
@@ -634,6 +621,7 @@ mod tests {
             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(),
+            supports_streaming: true,
             mapping_matched_model: None,
         }
     }
@@ -1540,6 +1528,7 @@ mod tests {
             .expect("state should build")
             .with_data_state_for_tests(data_state);
         let auth_snapshot = sample_auth_snapshot();
+        let client_session_affinity = ClientSessionAffinity::from_session_key("session-1");
         let cached_candidate = sample_priority_candidate(
             "provider-cached",
             "endpoint-cached",
@@ -1551,7 +1540,7 @@ mod tests {
         remember_scheduler_affinity_for_candidate(
             PlannerAppState::new(&state),
             Some(&auth_snapshot),
-            None,
+            Some(&client_session_affinity),
             "openai:chat",
             "gpt-4.1",
             &cached_candidate,
@@ -1573,7 +1562,7 @@ mod tests {
             "openai:chat",
             "gpt-4.1",
             Some(&auth_snapshot),
-            None,
+            Some(&client_session_affinity),
             None,
             None,
             None,
@@ -1714,6 +1703,7 @@ mod tests {
             .expect("state should build")
             .with_data_state_for_tests(data_state);
         let auth_snapshot = sample_auth_snapshot();
+        let client_session_affinity = ClientSessionAffinity::from_session_key("session-1");
         let cached_cross_format = sample_priority_candidate(
             "provider-shared",
             "endpoint-openai",
@@ -1725,7 +1715,7 @@ mod tests {
         remember_scheduler_affinity_for_candidate(
             PlannerAppState::new(&state),
             Some(&auth_snapshot),
-            None,
+            Some(&client_session_affinity),
             "claude:messages",
             "gpt-4.1",
             &cached_cross_format,
@@ -1747,7 +1737,7 @@ mod tests {
             "claude:messages",
             "gpt-4.1",
             Some(&auth_snapshot),
-            None,
+            Some(&client_session_affinity),
             None,
             None,
             None,
@@ -1915,6 +1905,7 @@ mod tests {
             .expect("state should build")
             .with_data_state_for_tests(data_state);
         let auth_snapshot = sample_auth_snapshot();
+        let client_session_affinity = ClientSessionAffinity::from_session_key("session-1");
         let cached_candidate = sample_priority_candidate(
             "provider-pool",
             "endpoint-pool",
@@ -1926,7 +1917,7 @@ mod tests {
         remember_scheduler_affinity_for_candidate(
             PlannerAppState::new(&state),
             Some(&auth_snapshot),
-            None,
+            Some(&client_session_affinity),
             "openai:chat",
             "gpt-4.1",
             &cached_candidate,
@@ -1948,7 +1939,7 @@ mod tests {
             "openai:chat",
             Some("gpt-4.1"),
             Some(&auth_snapshot),
-            None,
+            Some(&client_session_affinity),
             None,
             None,
             None,
@@ -2008,6 +1999,7 @@ mod tests {
             .expect("state should build")
             .with_data_state_for_tests(data_state);
         let auth_snapshot = sample_auth_snapshot();
+        let client_session_affinity = ClientSessionAffinity::from_session_key("session-1");
         let cached_candidate = sample_priority_candidate(
             "provider-pool",
             "endpoint-pool",
@@ -2019,7 +2011,7 @@ mod tests {
         remember_scheduler_affinity_for_candidate(
             PlannerAppState::new(&state),
             Some(&auth_snapshot),
-            None,
+            Some(&client_session_affinity),
             "openai:chat",
             "gpt-4.1",
             &cached_candidate,
@@ -2041,7 +2033,7 @@ mod tests {
             "openai:chat",
             Some("gpt-4.1"),
             Some(&auth_snapshot),
-            None,
+            Some(&client_session_affinity),
             None,
             None,
             None,
@@ -2064,7 +2056,7 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn remembers_scheduler_affinity_for_candidate_using_requested_model_key() {
+    async fn ignores_scheduler_affinity_without_client_session_scope() {
         let state = AppState::new().expect("state should build");
         let auth_snapshot = sample_auth_snapshot();
         let candidate = sample_candidate("endpoint-1", "key-1");
@@ -2078,15 +2070,12 @@ mod tests {
             &candidate,
         );
 
-        let remembered = state
+        assert!(state
             .read_scheduler_affinity_target(
                 "scheduler_affinity:api-key-1:openai:chat:gpt-5",
                 SCHEDULER_AFFINITY_TTL,
             )
-            .expect("affinity target should be cached");
+            .is_none());
-        assert_eq!(remembered.provider_id, "provider-1");
-        assert_eq!(remembered.endpoint_id, "endpoint-1");
-        assert_eq!(remembered.key_id, "key-1");
     }
 
     #[tokio::test]

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

@@ -7,6 +7,7 @@ use aether_ai_serving::{
 use aether_routing_core::ResolvedRoutingPolicy;
 use async_trait::async_trait;
 use std::convert::Infallible;
+use std::time::Instant;
 use tracing::warn;
 
 use aether_scheduler_core::{
@@ -20,6 +21,7 @@ use crate::ai_serving::{
     PlannerAppState,
 };
 use crate::orchestration::LocalExecutionCandidateMetadata;
+use crate::stage_metrics::observe_gateway_stage_ms;
 
 use super::candidate_ranking::rank_eligible_local_execution_candidates;
 
@@ -68,7 +70,7 @@ struct GatewayLocalCandidateResolutionPort<'a> {
 #[async_trait]
 impl AiCandidateResolutionPort for GatewayLocalCandidateResolutionPort<'_> {
     type Candidate = SchedulerMinimalCandidateSelectionCandidate;
-    type Transport = GatewayProviderTransportSnapshot;
+    type Transport = Arc<GatewayProviderTransportSnapshot>;
     type Eligible = EligibleLocalExecutionCandidate;
     type Skipped = SkippedLocalExecutionCandidate;
     type Error = Infallible;
@@ -77,7 +79,12 @@ impl AiCandidateResolutionPort for GatewayLocalCandidateResolutionPort<'_> {
         &self,
         candidate: &Self::Candidate,
     ) -> Result<Option<Self::Transport>, Self::Error> {
-        Ok(read_candidate_transport_snapshot(self.state, candidate).await)
+        let started_at = Instant::now();
+        let transport = read_candidate_transport_snapshot_arc(self.state, candidate).await;
+        let elapsed_ms = started_at.elapsed().as_millis() as u64;
+        observe_gateway_stage_ms("candidate_transport_snapshot", elapsed_ms);
+        observe_gateway_stage_ms("candidate_resolution_transport_read", elapsed_ms);
+        Ok(transport)
     }
 
     fn build_missing_transport_skipped_candidate(
@@ -139,7 +146,7 @@ impl AiCandidateResolutionPort for GatewayLocalCandidateResolutionPort<'_> {
         SkippedLocalExecutionCandidate {
             candidate,
             skip_reason,
-            transport: Some(Arc::new(transport)),
+            transport: Some(transport),
             ranking: None,
             extra_data: None,
         }
@@ -159,7 +166,7 @@ impl AiCandidateResolutionPort for GatewayLocalCandidateResolutionPort<'_> {
         EligibleLocalExecutionCandidate {
             kind,
             candidate,
-            transport: Arc::new(transport),
+            transport,
             provider_api_format,
             orchestration: LocalExecutionCandidateMetadata::default(),
             ranking: None,
@@ -171,7 +178,8 @@ impl AiCandidateResolutionPort for GatewayLocalCandidateResolutionPort<'_> {
         candidates: Vec<Self::Eligible>,
         normalized_client_api_format: &str,
     ) -> Result<Vec<Self::Eligible>, Self::Error> {
-        Ok(rank_eligible_local_execution_candidates(
+        let started_at = Instant::now();
+        let ranked = rank_eligible_local_execution_candidates(
             self.state,
             candidates,
             normalized_client_api_format,
@@ -181,7 +189,12 @@ impl AiCandidateResolutionPort for GatewayLocalCandidateResolutionPort<'_> {
             self.required_capabilities,
             self.routing_policy,
         )
-        .await)
+        .await;
+        observe_gateway_stage_ms(
+            "candidate_resolution_rank",
+            started_at.elapsed().as_millis() as u64,
+        );
+        Ok(ranked)
     }
 
     async fn apply_pool_scheduler(
@@ -358,8 +371,13 @@ async fn resolve_and_rank_local_execution_candidates_with_pool_expansion(
         expand_pool_groups,
     };
 
+    let started_at = Instant::now();
     match run_ai_candidate_resolution(&port, candidates, request).await {
         Ok(mut outcome) => {
+            observe_gateway_stage_ms(
+                "candidate_resolution_core",
+                started_at.elapsed().as_millis() as u64,
+            );
             for candidate in &mut outcome.eligible_candidates {
                 candidate.orchestration.scheduler_affinity_epoch = Some(scheduler_affinity_epoch);
             }
@@ -506,8 +524,17 @@ pub(crate) async fn read_candidate_transport_snapshot(
     state: PlannerAppState<'_>,
     candidate: &SchedulerMinimalCandidateSelectionCandidate,
 ) -> Option<GatewayProviderTransportSnapshot> {
+    read_candidate_transport_snapshot_arc(state, candidate)
+        .await
+        .map(|transport| (*transport).clone())
+}
+
+pub(crate) async fn read_candidate_transport_snapshot_arc(
+    state: PlannerAppState<'_>,
+    candidate: &SchedulerMinimalCandidateSelectionCandidate,
+) -> Option<Arc<GatewayProviderTransportSnapshot>> {
     match state
-        .read_provider_transport_snapshot(
+        .read_provider_transport_snapshot_arc(
             &candidate.provider_id,
             &candidate.endpoint_id,
             &candidate.key_id,
@@ -616,6 +643,7 @@ mod tests {
             global_model_id: "global-model-1".to_string(),
             global_model_name: "claude-sonnet".to_string(),
             selected_provider_model_name: "claude-sonnet".to_string(),
+            supports_streaming: true,
             mapping_matched_model: None,
         }
     }

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

@@ -1,8 +1,10 @@
 use std::collections::BTreeMap;
 
+use aether_contracts::ProxySnapshot;
 use aether_scheduler_core::{
     SchedulerMinimalCandidateSelectionCandidate, SchedulerTunnelAffinityBucket,
 };
+use serde_json::Value;
 use tracing::warn;
 
 use crate::ai_serving::{GatewayProviderTransportSnapshot, PlannerAppState};
@@ -10,7 +12,9 @@ use crate::scheduler::config::SchedulerOrderingConfig;
 
 use super::candidate_resolution::read_candidate_transport_snapshot;
 
-pub(super) type CandidateTransportIdentity<'a> = (&'a str, &'a str, &'a str);
+const TUNNEL_OWNER_INSTANCE_ID_EXTRA_KEY: &str = "tunnel_owner_instance_id";
+
+pub(super) type CandidateTransportIdentity = (String, String, String);
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub(super) struct CandidateTransportRankingFacts {
@@ -18,38 +22,51 @@ pub(super) struct CandidateTransportRankingFacts {
     pub(super) keep_priority_on_conversion: bool,
 }
 
-pub(super) async fn resolve_cached_candidate_transport_ranking_facts<'a>(
+#[derive(Debug, Default)]
-    state: PlannerAppState<'_>,
+pub(super) struct CandidateTransportRankingFactsCache {
-    cache: &mut BTreeMap<CandidateTransportIdentity<'a>, CandidateTransportRankingFacts>,
+    candidate_facts: BTreeMap<CandidateTransportIdentity, CandidateTransportRankingFacts>,
-    candidate: &'a SchedulerMinimalCandidateSelectionCandidate,
+    configured_proxy_snapshots: BTreeMap<String, Option<ProxySnapshot>>,
-    ordering_config: SchedulerOrderingConfig,
+    system_proxy_snapshot: Option<Option<ProxySnapshot>>,
-) -> CandidateTransportRankingFacts {
+    tunnel_buckets_by_node_id: BTreeMap<String, SchedulerTunnelAffinityBucket>,
-    let identity = candidate_transport_identity(candidate);
-    if let Some(facts) = cache.get(&identity).copied() {
-        return facts;
-    }
-
-    let facts = resolve_candidate_transport_ranking_facts(state, candidate, ordering_config).await;
-    cache.insert(identity, facts);
-    facts
 }
 
-pub(super) async fn resolve_cached_transport_ranking_facts<'a>(
+pub(super) async fn resolve_cached_candidate_transport_ranking_facts(
     state: PlannerAppState<'_>,
-    cache: &mut BTreeMap<CandidateTransportIdentity<'a>, CandidateTransportRankingFacts>,
+    cache: &mut CandidateTransportRankingFactsCache,
-    candidate: &'a SchedulerMinimalCandidateSelectionCandidate,
+    candidate: &SchedulerMinimalCandidateSelectionCandidate,
-    transport: &GatewayProviderTransportSnapshot,
     ordering_config: SchedulerOrderingConfig,
 ) -> CandidateTransportRankingFacts {
     let identity = candidate_transport_identity(candidate);
-    if let Some(facts) = cache.get(&identity).copied() {
+    if let Some(facts) = cache.candidate_facts.get(&identity).copied() {
         return facts;
     }
 
     let facts =
-        resolve_candidate_transport_ranking_facts_from_transport(state, transport, ordering_config)
+        resolve_candidate_transport_ranking_facts(state, cache, candidate, ordering_config).await;
-            .await;
+    cache.candidate_facts.insert(identity, facts);
-    cache.insert(identity, facts);
+    facts
+}
+
+pub(super) async fn resolve_cached_transport_ranking_facts(
+    state: PlannerAppState<'_>,
+    cache: &mut CandidateTransportRankingFactsCache,
+    candidate: &SchedulerMinimalCandidateSelectionCandidate,
+    transport: &GatewayProviderTransportSnapshot,
+    ordering_config: SchedulerOrderingConfig,
+) -> CandidateTransportRankingFacts {
+    let identity = candidate_transport_identity(candidate);
+    if let Some(facts) = cache.candidate_facts.get(&identity).copied() {
+        return facts;
+    }
+
+    let facts = resolve_candidate_transport_ranking_facts_from_transport(
+        state,
+        cache,
+        transport,
+        ordering_config,
+    )
+    .await;
+    cache.candidate_facts.insert(identity, facts);
     facts
 }
 
@@ -58,13 +75,15 @@ pub(super) async fn candidate_keeps_priority_on_conversion(
     candidate: &SchedulerMinimalCandidateSelectionCandidate,
     ordering_config: SchedulerOrderingConfig,
 ) -> bool {
-    resolve_candidate_transport_ranking_facts(state, candidate, ordering_config)
+    let mut cache = CandidateTransportRankingFactsCache::default();
+    resolve_candidate_transport_ranking_facts(state, &mut cache, candidate, ordering_config)
         .await
         .keep_priority_on_conversion
 }
 
 async fn resolve_candidate_transport_ranking_facts(
     state: PlannerAppState<'_>,
+    cache: &mut CandidateTransportRankingFactsCache,
     candidate: &SchedulerMinimalCandidateSelectionCandidate,
     ordering_config: SchedulerOrderingConfig,
 ) -> CandidateTransportRankingFacts {
@@ -75,17 +94,23 @@ async fn resolve_candidate_transport_ranking_facts(
         };
     };
 
-    resolve_candidate_transport_ranking_facts_from_transport(state, &transport, ordering_config)
+    resolve_candidate_transport_ranking_facts_from_transport(
-        .await
+        state,
+        cache,
+        &transport,
+        ordering_config,
+    )
+    .await
 }
 
 async fn resolve_candidate_transport_ranking_facts_from_transport(
     state: PlannerAppState<'_>,
+    cache: &mut CandidateTransportRankingFactsCache,
     transport: &GatewayProviderTransportSnapshot,
     ordering_config: SchedulerOrderingConfig,
 ) -> CandidateTransportRankingFacts {
     CandidateTransportRankingFacts {
-        tunnel_bucket: resolve_tunnel_owner_affinity_from_transport(state, transport).await,
+        tunnel_bucket: resolve_tunnel_owner_affinity_from_transport(state, cache, transport).await,
         keep_priority_on_conversion: ordering_config.keep_priority_on_conversion
             || transport.provider.keep_priority_on_conversion,
     }
@@ -93,12 +118,11 @@ async fn resolve_candidate_transport_ranking_facts_from_transport(
 
 async fn resolve_tunnel_owner_affinity_from_transport(
     state: PlannerAppState<'_>,
+    cache: &mut CandidateTransportRankingFactsCache,
     transport: &GatewayProviderTransportSnapshot,
 ) -> SchedulerTunnelAffinityBucket {
-    let Some(proxy) = state
+    let Some(proxy) =
-        .app()
+        resolve_transport_proxy_snapshot_with_tunnel_affinity_cached(state, cache, transport).await
-        .resolve_transport_proxy_snapshot_with_tunnel_affinity(transport)
-        .await
     else {
         return SchedulerTunnelAffinityBucket::Neutral;
     };
@@ -114,10 +138,74 @@ async fn resolve_tunnel_owner_affinity_from_transport(
         return SchedulerTunnelAffinityBucket::Neutral;
     };
 
+    if let Some(bucket) = cache.tunnel_buckets_by_node_id.get(node_id).copied() {
+        return bucket;
+    }
+
+    let bucket = resolve_tunnel_owner_affinity_from_proxy(state, &proxy, node_id).await;
+    cache
+        .tunnel_buckets_by_node_id
+        .insert(node_id.to_string(), bucket);
+    bucket
+}
+
+async fn resolve_transport_proxy_snapshot_with_tunnel_affinity_cached(
+    state: PlannerAppState<'_>,
+    cache: &mut CandidateTransportRankingFactsCache,
+    transport: &GatewayProviderTransportSnapshot,
+) -> Option<ProxySnapshot> {
+    for raw in [
+        transport.key.proxy.as_ref(),
+        transport.endpoint.proxy.as_ref(),
+        transport.provider.proxy.as_ref(),
+    ]
+    .into_iter()
+    .flatten()
+    {
+        let cache_key = proxy_config_cache_key(raw);
+        if let Some(snapshot) = cache.configured_proxy_snapshots.get(&cache_key) {
+            if snapshot.is_some() {
+                return snapshot.clone();
+            }
+            continue;
+        }
+        let snapshot = state
+            .app()
+            .resolve_configured_proxy_snapshot_with_tunnel_affinity(Some(raw))
+            .await;
+        cache
+            .configured_proxy_snapshots
+            .insert(cache_key, snapshot.clone());
+        if snapshot.is_some() {
+            return snapshot;
+        }
+    }
+
+    if let Some(snapshot) = cache.system_proxy_snapshot.as_ref() {
+        return snapshot.clone();
+    }
+    let snapshot = state.app().resolve_system_proxy_snapshot().await;
+    cache.system_proxy_snapshot = Some(snapshot.clone());
+    snapshot
+}
+
+async fn resolve_tunnel_owner_affinity_from_proxy(
+    state: PlannerAppState<'_>,
+    proxy: &ProxySnapshot,
+    node_id: &str,
+) -> SchedulerTunnelAffinityBucket {
     if state.app().tunnel.has_local_proxy(node_id) {
         return SchedulerTunnelAffinityBucket::LocalTunnel;
     }
 
+    if let Some(owner_instance_id) = proxy_tunnel_owner_instance_id(proxy) {
+        return if owner_instance_id == state.app().tunnel.local_instance_id() {
+            SchedulerTunnelAffinityBucket::LocalTunnel
+        } else {
+            SchedulerTunnelAffinityBucket::RemoteTunnel
+        };
+    }
+
     match state
         .app()
         .tunnel
@@ -144,10 +232,25 @@ async fn resolve_tunnel_owner_affinity_from_transport(
 
 fn candidate_transport_identity(
     candidate: &SchedulerMinimalCandidateSelectionCandidate,
-) -> CandidateTransportIdentity<'_> {
+) -> CandidateTransportIdentity {
     (
-        candidate.provider_id.as_str(),
+        candidate.provider_id.clone(),
-        candidate.endpoint_id.as_str(),
+        candidate.endpoint_id.clone(),
-        candidate.key_id.as_str(),
+        candidate.key_id.clone(),
     )
 }
+
+fn proxy_config_cache_key(raw: &Value) -> String {
+    serde_json::to_string(raw).unwrap_or_else(|_| raw.to_string())
+}
+
+fn proxy_tunnel_owner_instance_id(proxy: &ProxySnapshot) -> Option<&str> {
+    proxy
+        .extra
+        .as_ref()
+        .and_then(Value::as_object)
+        .and_then(|extra| extra.get(TUNNEL_OWNER_INSTANCE_ID_EXTRA_KEY))
+        .and_then(Value::as_str)
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+}

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

@@ -1,15 +1,16 @@
 use axum::body::Bytes;
 
+use crate::ai_serving::is_json_request;
 use crate::ai_serving::{
     endpoint_config_forces_upstream_stream_policy as endpoint_config_forces_upstream_stream_policy_impl,
     enforce_request_body_stream_field as enforce_request_body_stream_field_impl,
     force_upstream_streaming_for_provider as force_upstream_streaming_for_provider_impl,
-    is_json_request, parse_direct_request_body as parse_direct_request_body_impl,
+    parse_direct_request_body as parse_direct_request_body_impl,
-    resolve_upstream_is_stream_from_endpoint_config as resolve_upstream_is_stream_from_endpoint_config_impl,
+    resolve_format_upstream_is_stream_for_provider as resolve_upstream_is_stream_for_provider_impl,
 };
 pub(crate) use crate::ai_serving::{
     CLAUDE_CHAT_STREAM_PLAN_KIND, CLAUDE_CHAT_SYNC_PLAN_KIND, CLAUDE_CLI_STREAM_PLAN_KIND,
-    CLAUDE_CLI_SYNC_PLAN_KIND, EXECUTION_RUNTIME_STREAM_ACTION,
+    CLAUDE_CLI_SYNC_PLAN_KIND, CLAUDE_COUNT_TOKENS_SYNC_PLAN_KIND, EXECUTION_RUNTIME_STREAM_ACTION,
     EXECUTION_RUNTIME_STREAM_DECISION_ACTION, EXECUTION_RUNTIME_SYNC_ACTION,
     EXECUTION_RUNTIME_SYNC_DECISION_ACTION, GEMINI_CHAT_STREAM_PLAN_KIND,
     GEMINI_CHAT_SYNC_PLAN_KIND, GEMINI_CLI_STREAM_PLAN_KIND, GEMINI_CLI_SYNC_PLAN_KIND,
@@ -20,9 +21,10 @@ pub(crate) use crate::ai_serving::{
     OPENAI_EMBEDDING_SYNC_PLAN_KIND, OPENAI_IMAGE_STREAM_PLAN_KIND, OPENAI_IMAGE_SYNC_PLAN_KIND,
     OPENAI_RERANK_SYNC_PLAN_KIND, OPENAI_RESPONSES_COMPACT_STREAM_PLAN_KIND,
     OPENAI_RESPONSES_COMPACT_SYNC_PLAN_KIND, OPENAI_RESPONSES_STREAM_PLAN_KIND,
-    OPENAI_RESPONSES_SYNC_PLAN_KIND, OPENAI_VIDEO_CANCEL_SYNC_PLAN_KIND,
+    OPENAI_RESPONSES_SYNC_PLAN_KIND, OPENAI_SEARCH_SYNC_PLAN_KIND,
-    OPENAI_VIDEO_CONTENT_PLAN_KIND, OPENAI_VIDEO_CREATE_SYNC_PLAN_KIND,
+    OPENAI_VIDEO_CANCEL_SYNC_PLAN_KIND, OPENAI_VIDEO_CONTENT_PLAN_KIND,
-    OPENAI_VIDEO_DELETE_SYNC_PLAN_KIND, OPENAI_VIDEO_REMIX_SYNC_PLAN_KIND,
+    OPENAI_VIDEO_CREATE_SYNC_PLAN_KIND, OPENAI_VIDEO_DELETE_SYNC_PLAN_KIND,
+    OPENAI_VIDEO_REMIX_SYNC_PLAN_KIND,
 };
 
 pub(crate) use aether_ai_serving::AiRequestedModelFamily as RequestedModelFamily;
@@ -54,10 +56,10 @@ pub(crate) fn resolve_upstream_is_stream_for_provider(
     client_is_stream: bool,
     hard_requires_streaming: bool,
 ) -> bool {
-    let hard_requires_streaming = hard_requires_streaming
+    resolve_upstream_is_stream_for_provider_impl(
-        || force_upstream_streaming_for_provider(provider_type, provider_api_format);
-    resolve_upstream_is_stream_from_endpoint_config_impl(
         endpoint_config,
+        provider_type,
+        provider_api_format,
         client_is_stream,
         hard_requires_streaming,
     )
@@ -178,6 +180,27 @@ mod tests {
             true,
             false,
         ));
+        assert!(!resolve_upstream_is_stream_for_provider(
+            Some(&json!({"upstream_stream_policy": "force_stream"})),
+            "codex",
+            "openai:image",
+            true,
+            true,
+        ));
+        assert!(!resolve_upstream_is_stream_for_provider(
+            Some(&json!({"upstream_stream_policy": "force_stream"})),
+            "codex",
+            "openai:responses:compact",
+            true,
+            true,
+        ));
+        assert!(!resolve_upstream_is_stream_for_provider(
+            Some(&json!({"upstream_stream_policy": "force_stream"})),
+            "custom",
+            "openai:responses:compact",
+            true,
+            true,
+        ));
     }
 
     #[test]

apps/aether-gateway/src/ai_serving/planner/decision/control_plan.rs

@@ -1,14 +1,15 @@
 use crate::ai_serving::planner::common::{
     CLAUDE_CHAT_STREAM_PLAN_KIND, CLAUDE_CHAT_SYNC_PLAN_KIND, CLAUDE_CLI_STREAM_PLAN_KIND,
-    CLAUDE_CLI_SYNC_PLAN_KIND, GEMINI_CHAT_STREAM_PLAN_KIND, GEMINI_CHAT_SYNC_PLAN_KIND,
+    CLAUDE_CLI_SYNC_PLAN_KIND, CLAUDE_COUNT_TOKENS_SYNC_PLAN_KIND, GEMINI_CHAT_STREAM_PLAN_KIND,
-    GEMINI_CLI_STREAM_PLAN_KIND, GEMINI_CLI_SYNC_PLAN_KIND, GEMINI_EMBEDDING_SYNC_PLAN_KIND,
+    GEMINI_CHAT_SYNC_PLAN_KIND, GEMINI_CLI_STREAM_PLAN_KIND, GEMINI_CLI_SYNC_PLAN_KIND,
-    GEMINI_FILES_DELETE_PLAN_KIND, GEMINI_FILES_DOWNLOAD_PLAN_KIND, GEMINI_FILES_GET_PLAN_KIND,
+    GEMINI_EMBEDDING_SYNC_PLAN_KIND, GEMINI_FILES_DELETE_PLAN_KIND,
-    GEMINI_FILES_LIST_PLAN_KIND, GEMINI_VIDEO_CANCEL_SYNC_PLAN_KIND,
+    GEMINI_FILES_DOWNLOAD_PLAN_KIND, GEMINI_FILES_GET_PLAN_KIND, GEMINI_FILES_LIST_PLAN_KIND,
-    GEMINI_VIDEO_CREATE_SYNC_PLAN_KIND, OPENAI_CHAT_STREAM_PLAN_KIND, OPENAI_CHAT_SYNC_PLAN_KIND,
+    GEMINI_VIDEO_CANCEL_SYNC_PLAN_KIND, GEMINI_VIDEO_CREATE_SYNC_PLAN_KIND,
-    OPENAI_EMBEDDING_SYNC_PLAN_KIND, OPENAI_IMAGE_STREAM_PLAN_KIND, OPENAI_IMAGE_SYNC_PLAN_KIND,
+    OPENAI_CHAT_STREAM_PLAN_KIND, OPENAI_CHAT_SYNC_PLAN_KIND, OPENAI_EMBEDDING_SYNC_PLAN_KIND,
-    OPENAI_RERANK_SYNC_PLAN_KIND, OPENAI_RESPONSES_COMPACT_STREAM_PLAN_KIND,
+    OPENAI_IMAGE_STREAM_PLAN_KIND, OPENAI_IMAGE_SYNC_PLAN_KIND, OPENAI_RERANK_SYNC_PLAN_KIND,
-    OPENAI_RESPONSES_COMPACT_SYNC_PLAN_KIND, OPENAI_RESPONSES_STREAM_PLAN_KIND,
+    OPENAI_RESPONSES_COMPACT_STREAM_PLAN_KIND, OPENAI_RESPONSES_COMPACT_SYNC_PLAN_KIND,
-    OPENAI_RESPONSES_SYNC_PLAN_KIND, OPENAI_VIDEO_CANCEL_SYNC_PLAN_KIND,
+    OPENAI_RESPONSES_STREAM_PLAN_KIND, OPENAI_RESPONSES_SYNC_PLAN_KIND,
+    OPENAI_SEARCH_SYNC_PLAN_KIND, OPENAI_VIDEO_CANCEL_SYNC_PLAN_KIND,
     OPENAI_VIDEO_CONTENT_PLAN_KIND, OPENAI_VIDEO_CREATE_SYNC_PLAN_KIND,
     OPENAI_VIDEO_DELETE_SYNC_PLAN_KIND, OPENAI_VIDEO_REMIX_SYNC_PLAN_KIND,
 };
@@ -100,12 +101,15 @@ fn build_sync_plan_payload_from_decision(
         OPENAI_RESPONSES_SYNC_PLAN_KIND => {
             build_openai_responses_sync_plan_from_decision(parts, body_json, payload, false)?
         }
-        OPENAI_IMAGE_SYNC_PLAN_KIND => build_passthrough_sync_plan_from_decision(parts, payload)?,
+        OPENAI_IMAGE_SYNC_PLAN_KIND | OPENAI_SEARCH_SYNC_PLAN_KIND => {
+            build_passthrough_sync_plan_from_decision(parts, payload)?
+        }
         OPENAI_RESPONSES_COMPACT_SYNC_PLAN_KIND => {
             build_openai_responses_sync_plan_from_decision(parts, body_json, payload, true)?
         }
         CLAUDE_CHAT_SYNC_PLAN_KIND
         | CLAUDE_CLI_SYNC_PLAN_KIND
+        | CLAUDE_COUNT_TOKENS_SYNC_PLAN_KIND
         | OPENAI_EMBEDDING_SYNC_PLAN_KIND
         | OPENAI_RERANK_SYNC_PLAN_KIND => {
             build_standard_sync_plan_from_decision(parts, body_json, payload)?

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

@@ -59,6 +59,7 @@ pub(crate) async fn build_gemini_cli_v1internal_provider_request(
         input.upstream_is_stream,
         input.parts.uri.query(),
         None,
+        None,
         Some(&payload.body),
     )
     .ok_or(GeminiCliV1InternalRequestError::UpstreamUrlUnavailable)?;

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

@@ -1,6 +1,7 @@
 use crate::ai_serving::{AiExecutionDecision, AiExecutionPlanPayload, GatewayControlDecision};
 use crate::{AppState, GatewayError};
 
+mod antigravity;
 mod candidate_affinity_cache;
 mod candidate_materialization;
 mod candidate_metadata;
@@ -33,6 +34,7 @@ pub(crate) use self::candidate_resolution::{
     candidate_auth_channel_skip_reason, read_candidate_transport_snapshot,
     EligibleLocalExecutionCandidate, LocalExecutionCandidateKind, SkippedLocalExecutionCandidate,
 };
+pub(crate) use self::common::resolve_upstream_is_stream_for_provider;
 pub(crate) use self::passthrough::{
     build_local_same_format_stream_attempt_source, build_local_same_format_stream_plan_and_reports,
     build_local_same_format_sync_attempt_source, build_local_same_format_sync_plan_and_reports,
@@ -47,7 +49,7 @@ pub(crate) use self::plan_builders::{
 pub(crate) use self::pool_scores::{
     build_provider_key_pool_score_upsert, provider_key_pool_score_id, provider_key_pool_score_scope,
 };
-pub(crate) use self::request_gzip::resolve_transport_request_gzip_policy;
+pub(crate) use self::request_gzip::resolve_transport_request_encoding_policy;
 pub(crate) use self::route::is_matching_stream_request as planner_is_matching_stream_request;
 pub(crate) use self::runtime_miss::{
     apply_local_runtime_candidate_terminal_reason, record_local_runtime_candidate_skip_reason,
@@ -78,7 +80,8 @@ pub(crate) use self::standard::{
     build_local_stream_plan_and_reports as build_standard_family_stream_plan_and_reports,
     build_local_sync_attempt_source as build_standard_family_sync_attempt_source,
     build_local_sync_plan_and_reports as build_standard_family_sync_plan_and_reports,
-    set_local_openai_chat_execution_exhausted_diagnostic,
+    codex_model_capabilities_for_transport, set_local_openai_chat_execution_exhausted_diagnostic,
+    validate_final_openai_provider_request,
 };
 pub(crate) use self::state::{
     GatewayAuthApiKeySnapshot, GatewayProviderTransportSnapshot, LocalResolvedOAuthRequestAuth,
@@ -90,6 +93,71 @@ pub(crate) use aether_ai_serving::{
     CandidateFailureDiagnostic, CandidateFailureDiagnosticKind,
 };
 
+pub(crate) struct ResolvedTunnelSchedulerAffinityContext {
+    pub(crate) requested_model: String,
+    pub(crate) client_session_affinity: Option<aether_scheduler_core::ClientSessionAffinity>,
+    pub(crate) policy_context: Option<crate::scheduler::affinity::SchedulerAffinityPolicyContext>,
+    pub(crate) routing_overlay: Option<aether_routing_core::RankingOverlay>,
+}
+
+pub(crate) async fn resolve_tunnel_scheduler_affinity_context(
+    state: &AppState,
+    parts: &http::request::Parts,
+    decision: &GatewayControlDecision,
+    requested_model: String,
+    body_json: &serde_json::Value,
+    client_api_format: &str,
+) -> Result<Option<ResolvedTunnelSchedulerAffinityContext>, GatewayError> {
+    let Some(auth_context) = decision.auth_context.as_ref() else {
+        return Ok(None);
+    };
+    let execution_auth_context =
+        crate::ai_serving::build_execution_runtime_auth_context(auth_context);
+    let Some(auth_snapshot) = state
+        .read_cached_auth_api_key_snapshot(
+            &execution_auth_context.user_id,
+            &execution_auth_context.api_key_id,
+            crate::clock::current_unix_secs(),
+        )
+        .await?
+    else {
+        return Ok(None);
+    };
+    let resolved_auth_input = decision_input::ResolvedLocalDecisionAuthInput {
+        auth_context: execution_auth_context,
+        auth_snapshot,
+        required_capabilities: None,
+        model_directive_policy: decision.model_directive_policy.clone(),
+    };
+    let mut input = decision_input::build_local_requested_model_decision_input(
+        resolved_auth_input,
+        requested_model,
+    );
+    decision_input::attach_routing_policy_to_local_requested_model_input(
+        state,
+        parts,
+        &mut input,
+        body_json,
+        client_api_format,
+    )
+    .await?;
+    let policy_context = input
+        .routing_policy
+        .as_ref()
+        .map(crate::scheduler::affinity::SchedulerAffinityPolicyContext::from_routing_policy);
+    let routing_overlay = input
+        .routing_policy
+        .as_ref()
+        .map(|policy| policy.ranking_overlay.clone());
+
+    Ok(Some(ResolvedTunnelSchedulerAffinityContext {
+        requested_model: input.requested_model,
+        client_session_affinity: input.client_session_affinity,
+        policy_context,
+        routing_overlay,
+    }))
+}
+
 pub(crate) async fn maybe_build_sync_decision_payload(
     state: &AppState,
     parts: &http::request::Parts,

apps/aether-gateway/src/ai_serving/planner/passthrough/plan_builders.rs

@@ -71,6 +71,7 @@ pub(crate) fn build_passthrough_stream_plan_from_decision(
         .content_type
         .take()
         .or_else(|| provider_request_headers.get("content-type").cloned());
+    let stream = payload.upstream_is_stream;
     let plan = build_ai_execution_plan_from_decision(
         &mut payload,
         AiExecutionPlanFromDecisionParts {
@@ -84,7 +85,7 @@ pub(crate) fn build_passthrough_stream_plan_from_decision(
                 body_bytes_b64: None,
                 body_ref: None,
             },
-            stream: true,
+            stream,
         },
     );
 

apps/aether-gateway/src/ai_serving/planner/passthrough/provider/family/build.rs

@@ -66,7 +66,7 @@ pub(crate) async fn maybe_build_sync_local_same_format_provider_decision_payload
         candidate_count,
     );
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) =
             maybe_build_local_same_format_provider_decision_payload_for_candidate(
                 state, parts, trace_id, body_json, &input, attempt, spec,
@@ -134,7 +134,7 @@ pub(crate) async fn maybe_build_stream_local_same_format_provider_decision_paylo
         candidate_count,
     );
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) =
             maybe_build_local_same_format_provider_decision_payload_for_candidate(
                 state, parts, trace_id, body_json, &input, attempt, spec,

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

@@ -24,7 +24,7 @@ use crate::ai_serving::{
     ai_local_execution_contract_for_formats, extract_pool_sticky_session_token,
     resolve_local_decision_execution_runtime_auth_context, GatewayControlDecision, PlannerAppState,
 };
-use crate::client_session_affinity::client_session_affinity_from_parts;
+use crate::client_session_affinity::client_session_affinity_from_api_request;
 use crate::clock::current_unix_secs;
 use crate::{AppState, GatewayError};
 
@@ -60,7 +60,9 @@ pub(crate) async fn resolve_local_same_format_provider_decision_input(
         state,
         auth_context,
         Some(requested_model.as_str()),
+        decision.auth_endpoint_signature.as_deref(),
         None,
+        &decision.model_directive_policy,
     )
     .await
     {
@@ -79,7 +81,13 @@ pub(crate) async fn resolve_local_same_format_provider_decision_input(
 
     let mut input = build_local_requested_model_decision_input(resolved_input, requested_model);
     input.request_auth_channel = decision.request_auth_channel.clone();
-    input.client_session_affinity = client_session_affinity_from_parts(parts, Some(body_json));
+    input.client_surface = decision.client_surface;
+    input.gateway_credential_carrier = decision.gateway_credential_carrier;
+    input.client_session_affinity = client_session_affinity_from_api_request(
+        spec_metadata.api_format,
+        &parts.headers,
+        Some(body_json),
+    );
     if let Err(err) = attach_routing_policy_to_local_requested_model_input(
         state,
         parts,
@@ -115,15 +123,23 @@ pub(crate) async fn materialize_local_same_format_provider_candidate_attempts(
         input.required_capabilities.as_ref(),
         LocalCandidatePersistencePolicyKind::SameFormatProviderDecision,
     );
+    let model_directive_resolution = input
+        .model_directive_policy
+        .resolve_reasoning(spec_metadata.api_format, Some(&input.requested_model));
+    let routing_model = model_directive_resolution
+        .base_model()
+        .unwrap_or(&input.requested_model);
     let (candidates, preselection_skipped) = planner_state
-        .list_selectable_candidates_with_skip_reasons(
+        .list_selectable_candidates_with_skip_reasons_for_request_operation(
             spec_metadata.api_format,
-            &input.requested_model,
+            routing_model,
             spec_metadata.require_streaming,
             input.required_capabilities.as_ref(),
             Some(&input.auth_snapshot),
             input.client_session_affinity.as_ref(),
             current_unix_secs(),
+            false,
+            spec.operation.map(|operation| operation.as_str()),
         )
         .await?;
     let outcome = materialize_local_execution_candidates_with_serving(
@@ -212,15 +228,23 @@ pub(crate) async fn build_local_same_format_provider_candidate_attempt_source<'a
         input.required_capabilities.as_ref(),
         LocalCandidatePersistencePolicyKind::SameFormatProviderDecision,
     );
+    let model_directive_resolution = input
+        .model_directive_policy
+        .resolve_reasoning(spec_metadata.api_format, Some(&input.requested_model));
+    let routing_model = model_directive_resolution
+        .base_model()
+        .unwrap_or(&input.requested_model);
     let (candidates, preselection_skipped) = planner_state
-        .list_selectable_candidates_with_skip_reasons(
+        .list_selectable_candidates_with_skip_reasons_for_request_operation(
             spec_metadata.api_format,
-            &input.requested_model,
+            routing_model,
             spec_metadata.require_streaming,
             input.required_capabilities.as_ref(),
             Some(&input.auth_snapshot),
             input.client_session_affinity.as_ref(),
             current_unix_secs(),
+            false,
+            spec.operation.map(|operation| operation.as_str()),
         )
         .await?;
 

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

@@ -1,5 +1,8 @@
 use serde_json::json;
 
+use aether_ai_serving::{AdaptationMode, AiRequestGzipPolicy, OriginalRequestPayload};
+use aether_contracts::{ExecutionResponseBodyMode, EXECUTION_RESPONSE_BODY_MODE_HEADER};
+
 use crate::ai_serving::ai_local_execution_contract_for_formats;
 use crate::ai_serving::build_request_trace_proxy_value;
 use crate::ai_serving::planner::candidate_materialization::{
@@ -17,7 +20,7 @@ use crate::ai_serving::planner::report_context::{
 use crate::ai_serving::planner::spec_metadata::local_same_format_provider_spec_metadata;
 use crate::ai_serving::planner::CandidateFailureDiagnostic;
 use crate::ai_serving::planner::{
-    build_ai_execution_decision_response, resolve_transport_request_gzip_policy,
+    build_ai_execution_decision_response, resolve_transport_request_encoding_policy,
     AiExecutionDecisionResponseParts,
 };
 use crate::ai_serving::transport::{
@@ -61,6 +64,8 @@ pub(crate) async fn maybe_build_local_same_format_provider_decision_payload_for_
     else {
         return Ok(None);
     };
+    let request_redacted = resolved.request_redacted;
+    let compatibility_edits_empty = resolved.compatibility_edits.is_empty();
     let original_request_body_json = if resolved.request_redacted {
         Some(&resolved.provider_request_body)
     } else {
@@ -82,6 +87,51 @@ pub(crate) async fn maybe_build_local_same_format_provider_decision_payload_for_
         .clone()
         .or_else(|| resolve_transport_profile(&resolved.transport));
     let mut extra_fields = serde_json::Map::new();
+    extra_fields.insert(
+        "provider_type".to_string(),
+        json!(resolved.transport.provider.provider_type.as_str()),
+    );
+    if let Some(operation) = spec.operation {
+        extra_fields.insert("api_operation".to_string(), json!(operation.as_str()));
+    }
+    if let Some(client_surface) = input.client_surface {
+        extra_fields.insert("client_surface".to_string(), json!(client_surface.as_str()));
+    }
+    if let Some(carrier) = input.gateway_credential_carrier {
+        extra_fields.insert(
+            "gateway_credential_carrier".to_string(),
+            json!(carrier.as_str()),
+        );
+    }
+    extra_fields.insert(
+        "upstream_credential_mode".to_string(),
+        json!(resolved.transport.key.auth_type.trim().to_ascii_lowercase()),
+    );
+    let mut adaptation_mode = if resolved.compatibility_edits.is_empty() {
+        AdaptationMode::NativeTransparent
+    } else {
+        AdaptationMode::SameFormatCompat
+    };
+    if crate::ai_serving::normalize_api_format_alias(&resolved.provider_api_format)
+        == "claude:messages"
+    {
+        let compatibility_profile =
+            crate::ai_serving::transport::resolve_anthropic_compatibility_profile(
+                &resolved.transport,
+                &resolved.provider_api_format,
+            );
+        extra_fields.insert(
+            "anthropic_compatibility_profile".to_string(),
+            json!(compatibility_profile.as_str()),
+        );
+        if compatibility_profile.uses_claude_code_compatibility() {
+            adaptation_mode = AdaptationMode::SameFormatCompat;
+        }
+    }
+    extra_fields.insert(
+        "adaptation_mode".to_string(),
+        json!(adaptation_mode.as_str()),
+    );
     if let Some(proxy_value) =
         build_request_trace_proxy_value(Some(&resolved.transport), proxy.as_ref())
     {
@@ -149,6 +199,7 @@ pub(crate) async fn maybe_build_local_same_format_provider_decision_payload_for_
                 original_request_body_json,
                 original_request_body_base64: None,
                 client_session_affinity: input.client_session_affinity.as_ref(),
+                routing_policy: input.routing_policy.as_ref(),
                 scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
                 client_requested_stream: body_json
                     .get("stream")
@@ -184,7 +235,7 @@ pub(crate) async fn maybe_build_local_same_format_provider_decision_payload_for_
         compatibility_edits: _,
         request_redacted: _,
     } = resolved;
-    let request_gzip = resolve_transport_request_gzip_policy(&transport);
+    let request_encoding = resolve_transport_request_encoding_policy(&transport);
 
     let mut decision = build_ai_execution_decision_response(AiExecutionDecisionResponseParts {
         decision_is_stream: spec_metadata.require_streaming,
@@ -194,6 +245,7 @@ pub(crate) async fn maybe_build_local_same_format_provider_decision_payload_for_
         request_id: trace_id.to_string(),
         candidate_id: candidate_id.to_string(),
         provider_name: transport.provider.name.clone(),
+        provider_type: transport.provider.provider_type.clone(),
         provider_id: candidate.provider_id.clone(),
         endpoint_id: candidate.endpoint_id.clone(),
         key_id: candidate.key_id.clone(),
@@ -211,8 +263,8 @@ pub(crate) async fn maybe_build_local_same_format_provider_decision_payload_for_
         provider_request_body: Some(provider_request_body),
         provider_request_body_base64: None,
         content_type: Some("application/json".to_string()),
-        content_encoding: None,
+        content_encoding: request_encoding.content_encoding,
-        request_gzip,
+        request_gzip: request_encoding.request_gzip,
         proxy,
         transport_profile,
         timeouts: resolve_transport_execution_timeouts(&transport),
@@ -221,10 +273,84 @@ pub(crate) async fn maybe_build_local_same_format_provider_decision_payload_for_
         report_context: Some(report_context),
         auth_context: input.auth_context.clone(),
     });
-    apply_provider_request_routing_policy_to_decision(input, &mut decision)?;
+    apply_provider_request_routing_policy_to_decision(
+        input,
+        &mut decision,
+        Some(transport.as_ref()),
+    )?;
+    enforce_provider_api_operation_invariants(
+        spec.operation,
+        decision.provider_request_body.as_mut(),
+        &mut decision.provider_request_headers,
+    );
+    decision.provider_request_body_base64 = original_request_body_base64(
+        parts,
+        decision.provider_request_body.as_ref(),
+        adaptation_mode,
+        request_redacted,
+        compatibility_edits_empty,
+        decision.content_encoding.as_deref(),
+        decision.request_gzip.as_ref(),
+    );
+    decision
+        .provider_request_headers
+        .retain(|name, _| !name.eq_ignore_ascii_case(EXECUTION_RESPONSE_BODY_MODE_HEADER));
+    if !spec_metadata.require_streaming && decision.provider_request_body_base64.is_some() {
+        decision.provider_request_headers.insert(
+            EXECUTION_RESPONSE_BODY_MODE_HEADER.to_string(),
+            ExecutionResponseBodyMode::PreserveBytes
+                .as_str()
+                .to_string(),
+        );
+    }
     Ok(Some(decision))
 }
 
+fn enforce_provider_api_operation_invariants(
+    operation: Option<crate::ai_serving::ApiOperation>,
+    provider_request_body: Option<&mut serde_json::Value>,
+    provider_request_headers: &mut std::collections::BTreeMap<String, String>,
+) {
+    if operation != Some(crate::ai_serving::ApiOperation::ClaudeCountTokens) {
+        return;
+    }
+
+    if let Some(provider_request_body) = provider_request_body {
+        crate::ai_serving::transport::enforce_same_format_provider_api_operation_body_policy(
+            provider_request_body,
+            operation,
+        );
+    }
+    for header_name in ["accept", "content-type"] {
+        provider_request_headers.retain(|name, _| !name.eq_ignore_ascii_case(header_name));
+        provider_request_headers.insert(header_name.to_string(), "application/json".to_string());
+    }
+}
+
+fn original_request_body_base64(
+    parts: &http::request::Parts,
+    provider_request_body: Option<&serde_json::Value>,
+    adaptation_mode: AdaptationMode,
+    request_redacted: bool,
+    compatibility_edits_empty: bool,
+    content_encoding: Option<&str>,
+    request_gzip: Option<&AiRequestGzipPolicy>,
+) -> Option<String> {
+    if adaptation_mode != AdaptationMode::NativeTransparent
+        || request_redacted
+        || !compatibility_edits_empty
+        || content_encoding.is_some_and(|value| !value.trim().is_empty())
+        || request_gzip.is_some_and(|policy| policy.enabled != Some(false))
+    {
+        return None;
+    }
+
+    parts
+        .extensions
+        .get::<OriginalRequestPayload>()?
+        .body_bytes_base64_if_unchanged(provider_request_body?)
+}
+
 pub(super) async fn mark_skipped_local_same_format_provider_candidate(
     state: &AppState,
     input: &LocalSameFormatProviderDecisionInput,
@@ -308,3 +434,177 @@ pub(super) async fn mark_skipped_local_same_format_provider_candidate_with_failu
     )
     .await;
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeMap;
+
+    use base64::Engine as _;
+
+    use super::{
+        enforce_provider_api_operation_invariants, original_request_body_base64, AdaptationMode,
+        AiRequestGzipPolicy, OriginalRequestPayload,
+    };
+    use crate::ai_serving::ApiOperation;
+
+    fn request_parts_with_original_payload(
+        body_json: serde_json::Value,
+        body_bytes: &[u8],
+    ) -> http::request::Parts {
+        let (mut parts, ()) = http::Request::new(()).into_parts();
+        parts
+            .extensions
+            .insert(OriginalRequestPayload::from_parsed_json(
+                body_json, body_bytes,
+            ));
+        parts
+    }
+
+    #[test]
+    fn count_tokens_invariants_win_after_provider_routing_mutations() {
+        let mut body = serde_json::json!({
+            "model": "claude-sonnet-4",
+            "messages": [],
+            "stream": true
+        });
+        let mut headers = BTreeMap::from([
+            ("Accept".to_string(), "text/event-stream".to_string()),
+            ("Content-Type".to_string(), "text/plain".to_string()),
+            ("x-provider-route".to_string(), "kept".to_string()),
+        ]);
+
+        enforce_provider_api_operation_invariants(
+            Some(ApiOperation::ClaudeCountTokens),
+            Some(&mut body),
+            &mut headers,
+        );
+
+        assert!(body.get("stream").is_none());
+        assert_eq!(
+            headers.get("accept").map(String::as_str),
+            Some("application/json")
+        );
+        assert_eq!(
+            headers.get("content-type").map(String::as_str),
+            Some("application/json")
+        );
+        assert_eq!(
+            headers.get("x-provider-route").map(String::as_str),
+            Some("kept")
+        );
+        assert_eq!(
+            headers
+                .keys()
+                .filter(|name| name.eq_ignore_ascii_case("accept"))
+                .count(),
+            1
+        );
+        assert_eq!(
+            headers
+                .keys()
+                .filter(|name| name.eq_ignore_ascii_case("content-type"))
+                .count(),
+            1
+        );
+    }
+
+    #[test]
+    fn unchanged_same_format_body_preserves_original_json_bytes() {
+        let raw = br#"{ "unknown": {"enabled":true}, "messages": [], "model": "claude-sonnet-4" }"#;
+        let body_json: serde_json::Value = serde_json::from_slice(raw).expect("body should parse");
+        let parts = request_parts_with_original_payload(body_json.clone(), raw);
+
+        let encoded = original_request_body_base64(
+            &parts,
+            Some(&body_json),
+            AdaptationMode::NativeTransparent,
+            false,
+            true,
+            None,
+            None,
+        )
+        .expect("unchanged request should retain exact bytes");
+
+        assert_eq!(
+            base64::engine::general_purpose::STANDARD
+                .decode(encoded)
+                .expect("body should decode"),
+            raw
+        );
+    }
+
+    #[test]
+    fn request_edits_or_encoding_disable_original_json_bytes() {
+        let raw = br#"{"model":"claude-sonnet-4","messages":[]}"#;
+        let body_json: serde_json::Value = serde_json::from_slice(raw).expect("body should parse");
+        let parts = request_parts_with_original_payload(body_json.clone(), raw);
+        let changed_body = serde_json::json!({
+            "model": "claude-sonnet-4-5",
+            "messages": []
+        });
+
+        assert!(original_request_body_base64(
+            &parts,
+            Some(&changed_body),
+            AdaptationMode::NativeTransparent,
+            false,
+            true,
+            None,
+            None,
+        )
+        .is_none());
+        assert!(original_request_body_base64(
+            &parts,
+            Some(&body_json),
+            AdaptationMode::NativeTransparent,
+            true,
+            true,
+            None,
+            None,
+        )
+        .is_none());
+        assert!(original_request_body_base64(
+            &parts,
+            Some(&body_json),
+            AdaptationMode::NativeTransparent,
+            false,
+            false,
+            None,
+            None,
+        )
+        .is_none());
+        assert!(original_request_body_base64(
+            &parts,
+            Some(&body_json),
+            AdaptationMode::SameFormatCompat,
+            false,
+            true,
+            None,
+            None,
+        )
+        .is_none());
+        assert!(original_request_body_base64(
+            &parts,
+            Some(&body_json),
+            AdaptationMode::NativeTransparent,
+            false,
+            true,
+            Some("gzip"),
+            None,
+        )
+        .is_none());
+        assert!(original_request_body_base64(
+            &parts,
+            Some(&body_json),
+            AdaptationMode::NativeTransparent,
+            false,
+            true,
+            None,
+            Some(&AiRequestGzipPolicy {
+                enabled: Some(true),
+                min_bytes: Some(1),
+            }),
+        )
+        .is_none());
+    }
+}

apps/aether-gateway/src/ai_serving/planner/passthrough/provider/family/request.rs

@@ -13,7 +13,8 @@ use crate::ai_serving::planner::redaction::{
 use crate::ai_serving::transport::antigravity::{
     build_antigravity_safe_v1internal_request, build_antigravity_static_identity_headers,
     classify_local_antigravity_request_support, AntigravityEnvelopeRequestType,
-    AntigravityRequestEnvelopeSupport, AntigravityRequestSideSupport,
+    AntigravityRequestAuthUnsupportedReason, AntigravityRequestEnvelopeSupport,
+    AntigravityRequestSideSupport, AntigravityRequestSideUnsupportedReason,
 };
 use crate::ai_serving::transport::{
     build_gemini_cli_v1internal_request, build_grok_browser_headers, build_grok_upstream_url,
@@ -39,7 +40,10 @@ use super::{
     LocalSameFormatProviderCandidateAttempt, LocalSameFormatProviderDecisionInput,
     LocalSameFormatProviderSpec,
 };
-use crate::ai_serving::planner::standard::same_format_provider_request_body_failure_extra_data;
+use crate::ai_serving::planner::standard::{
+    codex_model_capabilities_for_transport, openai_provider_request_contract_failure_extra_data,
+    same_format_provider_request_body_failure_extra_data,
+};
 
 pub(crate) fn resolve_same_format_provider_transport_unsupported_reason_for_trace(
     transport: &GatewayProviderTransportSnapshot,
@@ -50,6 +54,7 @@ pub(crate) fn resolve_same_format_provider_transport_unsupported_reason_for_trac
             "openai:chat" => "openai:chat",
             "openai:responses" => "openai:responses",
             "openai:responses:compact" => "openai:responses:compact",
+            "openai:search" => "openai:search",
             "openai:embedding" => "openai:embedding",
             "openai:rerank" => "openai:rerank",
             "claude:messages" => "claude:messages",
@@ -71,9 +76,10 @@ pub(crate) fn resolve_same_format_provider_transport_unsupported_reason_for_trac
             decision_kind: "trace_candidate_metadata",
             report_kind: Some("trace_candidate_metadata"),
         },
+        None,
     );
     if !behavior.is_antigravity
-        && !behavior.is_claude_code
+        && !behavior.is_claude_code_transport
         && !behavior.is_gemini_cli
         && !behavior.is_vertex
         && !behavior.is_kiro
@@ -123,6 +129,23 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
     spec: LocalSameFormatProviderSpec,
 ) -> Result<Option<LocalSameFormatProviderCandidatePayloadParts>, GatewayError> {
     let candidate = &attempt.eligible.candidate;
+    if let Some(skip_reason) = same_format_provider_operation_skip_reason(
+        &attempt.eligible.transport,
+        attempt.eligible.provider_api_format.as_str(),
+        spec.operation,
+    ) {
+        mark_skipped_local_same_format_provider_candidate(
+            state,
+            input,
+            trace_id,
+            candidate,
+            attempt.candidate_index,
+            &attempt.candidate_id,
+            skip_reason,
+        )
+        .await;
+        return Ok(None);
+    }
     let Some(prepared) = prepare_local_same_format_provider_candidate(
         state,
         trace_id,
@@ -136,13 +159,26 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
     else {
         return Ok(None);
     };
-    let enable_model_directives =
+    let model_directive_resolution = input
-        crate::system_features::reasoning_model_directive_enabled_for_api_format_and_model(
+        .model_directive_policy
-            state,
+        .resolve_reasoning(spec.api_format, Some(&input.requested_model));
-            spec.api_format,
+    let model_directive_mapping =
-            Some(&input.requested_model),
+        match model_directive_resolution.mapping_patch_for_mapped_model(&prepared.mapped_model) {
-        )
+            Ok(mapping) => mapping,
-        .await;
+            Err(skip_reason) => {
+                mark_skipped_local_same_format_provider_candidate(
+                    state,
+                    input,
+                    trace_id,
+                    candidate,
+                    attempt.candidate_index,
+                    &attempt.candidate_id,
+                    skip_reason,
+                )
+                .await;
+                return Ok(None);
+            }
+        };
     let effective_headers = input.effective_headers(&parts.headers);
     let redaction = resolve_provider_chat_pii_redaction(
         state,
@@ -168,7 +204,7 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
             prepared.force_body_stream_field,
             prepared.kiro_auth.as_ref(),
             prepared.is_claude_code,
-            enable_model_directives,
+            false,
         )
     else {
         mark_skipped_local_same_format_provider_candidate_with_extra_data(
@@ -195,18 +231,11 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
     };
     let mut base_provider_request_body = base_provider_request.body;
     let mut compatibility_edits = base_provider_request.compatibility_edits;
-    if let Some(mapping) =
+    if let Some(mapping) = model_directive_mapping.as_ref() {
-        crate::system_features::reasoning_model_directive_mapping_for_api_format_and_model(
-            state,
-            spec.api_format,
-            Some(&input.requested_model),
-        )
-        .await
-    {
         let before_mapping = base_provider_request_body.clone();
         crate::ai_serving::apply_model_directive_mapping_patch(
             &mut base_provider_request_body,
-            &mapping,
+            mapping,
         );
         if before_mapping != base_provider_request_body {
             compatibility_edits.push(SameFormatProviderCompatibilityEdit {
@@ -229,12 +258,81 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
         }
     }
 
+    let source_model = body_json
+        .get("model")
+        .and_then(Value::as_str)
+        .unwrap_or(input.requested_model.as_str());
+    let codex_model_capabilities = codex_model_capabilities_for_transport(
+        &transport,
+        prepared.provider_api_format.as_str(),
+        prepared.mapped_model.as_str(),
+        source_model,
+    );
+    if let Err(violation) =
+        crate::ai_serving::finalize_openai_provider_request_with_codex_model_capabilities(
+            &mut base_provider_request_body,
+            crate::ai_serving::OpenAiProviderRequestFinalization {
+                source_api_format: spec.api_format,
+                provider_api_format: prepared.provider_api_format.as_str(),
+                provider_type: transport.provider.provider_type.as_str(),
+                provider_model: prepared.mapped_model.as_str(),
+                source_model,
+                body_rules: transport.endpoint.body_rules.as_ref(),
+                upstream_is_stream: prepared.upstream_is_stream,
+                require_body_stream_field: request_requires_body_stream_field(
+                    body_json,
+                    prepared.force_body_stream_field,
+                ),
+            },
+            codex_model_capabilities.as_ref(),
+        )
+    {
+        mark_skipped_local_same_format_provider_candidate_with_extra_data(
+            state,
+            input,
+            trace_id,
+            candidate,
+            attempt.candidate_index,
+            &attempt.candidate_id,
+            "provider_request_body_build_failed",
+            Some(openai_provider_request_contract_failure_extra_data(
+                &violation,
+                spec.api_format,
+                prepared.provider_api_format.as_str(),
+                "same_format_provider_request_finalization",
+            )),
+        )
+        .await;
+        return Ok(None);
+    }
+
     let antigravity_auth = if prepared.is_antigravity {
-        match classify_local_antigravity_request_support(
+        let mut antigravity_support = classify_local_antigravity_request_support(
             &transport,
             &base_provider_request_body,
             AntigravityEnvelopeRequestType::Agent,
+        );
+        if matches!(
+            antigravity_support,
+            AntigravityRequestSideSupport::Unsupported(
+                AntigravityRequestSideUnsupportedReason::UnsupportedAuth(
+                    AntigravityRequestAuthUnsupportedReason::MissingProjectId
+                )
+            )
         ) {
+            if let Some(hydrated) = state
+                .hydrate_antigravity_project_metadata_for_transport(&transport)
+                .await
+            {
+                transport = Arc::new(hydrated);
+                antigravity_support = classify_local_antigravity_request_support(
+                    &transport,
+                    &base_provider_request_body,
+                    AntigravityEnvelopeRequestType::Agent,
+                );
+            }
+        }
+        match antigravity_support {
             AntigravityRequestSideSupport::Supported(spec) => Some(spec.auth),
             AntigravityRequestSideSupport::Unsupported(_) => {
                 mark_skipped_local_same_format_provider_candidate(
@@ -291,7 +389,7 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
     } else {
         None
     };
-    let provider_request_body = if let Some(antigravity_auth) = antigravity_auth.as_ref() {
+    let mut provider_request_body = if let Some(antigravity_auth) = antigravity_auth.as_ref() {
         match build_antigravity_safe_v1internal_request(
             antigravity_auth,
             trace_id,
@@ -351,6 +449,16 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
     } else {
         base_provider_request_body
     };
+    if crate::ai_serving::transport::enforce_same_format_provider_api_operation_body_policy(
+        &mut provider_request_body,
+        spec.operation,
+    ) {
+        compatibility_edits.push(SameFormatProviderCompatibilityEdit {
+            field: "stream".to_string(),
+            action: SameFormatProviderCompatibilityEditAction::RuntimeRewrite,
+            detail: "removed stream field for non-streaming API operation".to_string(),
+        });
+    }
 
     let is_grok = prepared
         .transport
@@ -417,10 +525,10 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
             original_request_body: body_json,
             header_rules: transport.endpoint.header_rules.as_ref(),
             behavior: prepared.behavior,
+            api_operation: spec.operation,
             auth_header: prepared.auth_header.as_deref(),
             auth_value: prepared.auth_value.as_deref(),
             extra_headers: &extra_headers,
-            key_fingerprint: transport.key.fingerprint.as_ref(),
             kiro_auth_config: prepared.kiro_auth.as_ref().map(|auth| &auth.auth_config),
             kiro_machine_id: prepared
                 .kiro_auth
@@ -445,6 +553,28 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
         .await;
         return Ok(None);
     };
+    crate::ai_serving::apply_codex_openai_special_headers(
+        &mut provider_request_headers,
+        &provider_request_body,
+        effective_headers,
+        transport.provider.provider_type.as_str(),
+        prepared.provider_api_format.as_str(),
+        Some(trace_id),
+        transport.key.decrypted_auth_config.as_deref(),
+    );
+    let provider_model = provider_request_body
+        .get("model")
+        .and_then(Value::as_str)
+        .unwrap_or(prepared.mapped_model.as_str());
+    crate::ai_serving::apply_codex_openai_responses_lite_header_for_request_body_with_capabilities(
+        &mut provider_request_headers,
+        Some(&provider_request_body),
+        transport.provider.provider_type.as_str(),
+        prepared.provider_api_format.as_str(),
+        provider_model,
+        source_model,
+        codex_model_capabilities.as_ref(),
+    );
     request_identity_response_encoding_when_redacted(
         &mut provider_request_headers,
         redaction.redacted,
@@ -469,3 +599,98 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
         request_redacted: redaction.redacted,
     }))
 }
+
+fn same_format_provider_operation_skip_reason(
+    transport: &GatewayProviderTransportSnapshot,
+    provider_api_format: &str,
+    operation: Option<crate::ai_serving::ApiOperation>,
+) -> Option<&'static str> {
+    (!crate::ai_serving::transport::transport_supports_api_operation(
+        transport,
+        provider_api_format,
+        operation,
+    ))
+    .then_some("transport_operation_unsupported")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::same_format_provider_operation_skip_reason;
+    use crate::ai_serving::transport::snapshot::{
+        GatewayProviderTransportEndpoint, GatewayProviderTransportKey,
+        GatewayProviderTransportProvider,
+    };
+    use crate::ai_serving::{ApiOperation, GatewayProviderTransportSnapshot};
+
+    fn private_adapter_transport(provider_type: &str) -> GatewayProviderTransportSnapshot {
+        GatewayProviderTransportSnapshot {
+            provider: GatewayProviderTransportProvider {
+                id: "provider-1".to_string(),
+                name: provider_type.to_string(),
+                provider_type: provider_type.to_string(),
+                website: None,
+                is_active: true,
+                keep_priority_on_conversion: false,
+                enable_format_conversion: true,
+                concurrent_limit: None,
+                max_retries: None,
+                proxy: None,
+                request_timeout_secs: None,
+                stream_first_byte_timeout_secs: None,
+                config: None,
+            },
+            endpoint: GatewayProviderTransportEndpoint {
+                id: "endpoint-1".to_string(),
+                provider_id: "provider-1".to_string(),
+                api_format: "claude:messages".to_string(),
+                api_family: Some("claude".to_string()),
+                endpoint_kind: Some("chat".to_string()),
+                is_active: true,
+                base_url: "https://private.example".to_string(),
+                header_rules: None,
+                body_rules: None,
+                max_retries: None,
+                custom_path: None,
+                config: None,
+                format_acceptance_config: None,
+                proxy: None,
+            },
+            key: GatewayProviderTransportKey {
+                id: "key-1".to_string(),
+                provider_id: "provider-1".to_string(),
+                name: "key".to_string(),
+                auth_type: "oauth".to_string(),
+                is_active: true,
+                api_formats: None,
+                auth_type_by_format: None,
+                allow_auth_channel_mismatch_formats: None,
+                allowed_models: None,
+                capabilities: None,
+                rate_multipliers: None,
+                global_priority_by_format: None,
+                expires_at_unix_secs: None,
+                proxy: None,
+                fingerprint: None,
+                upstream_metadata: None,
+                decrypted_api_key: String::new(),
+                decrypted_auth_config: None,
+            },
+        }
+    }
+
+    #[test]
+    fn private_adapter_count_tokens_is_rejected_by_pre_auth_operation_gate() {
+        for provider_type in ["kiro", "grok"] {
+            let transport = private_adapter_transport(provider_type);
+            assert_eq!(
+                same_format_provider_operation_skip_reason(
+                    &transport,
+                    "claude:messages",
+                    Some(ApiOperation::ClaudeCountTokens),
+                ),
+                Some("transport_operation_unsupported"),
+                "provider_type={provider_type}"
+            );
+        }
+    }
+}

apps/aether-gateway/src/ai_serving/planner/passthrough/provider/family/request/policy.rs

@@ -1,6 +1,6 @@
 use crate::ai_serving::planner::spec_metadata::LocalExecutionSurfaceSpecMetadata;
 use crate::ai_serving::transport::{
-    classify_same_format_provider_request_behavior as classify_same_format_provider_request_behavior_impl,
+    classify_same_format_provider_request_behavior_for_operation as classify_same_format_provider_request_behavior_impl,
     resolve_same_format_provider_direct_auth as resolve_same_format_provider_direct_auth_impl,
     same_format_provider_transport_supported as same_format_provider_transport_supported_impl,
     same_format_provider_transport_unsupported_reason as same_format_provider_transport_unsupported_reason_impl,
@@ -15,6 +15,7 @@ pub(super) fn classify_same_format_provider_request_behavior(
     transport: &GatewayProviderTransportSnapshot,
     provider_api_format: &str,
     spec_metadata: LocalExecutionSurfaceSpecMetadata,
+    api_operation: Option<crate::ai_serving::ApiOperation>,
 ) -> SameFormatProviderRequestBehavior {
     classify_same_format_provider_request_behavior_impl(
         transport,
@@ -25,6 +26,7 @@ pub(super) fn classify_same_format_provider_request_behavior(
                 .report_kind
                 .expect("same-format provider specs should declare report kind"),
         },
+        api_operation,
     )
 }
 

apps/aether-gateway/src/ai_serving/planner/passthrough/provider/family/request/prepare.rs

@@ -59,6 +59,7 @@ pub(super) async fn prepare_local_same_format_provider_candidate(
         &transport,
         provider_api_format,
         spec_metadata,
+        spec.operation,
     );
 
     if !same_format_provider_transport_supported(

apps/aether-gateway/src/ai_serving/planner/passthrough/provider/plans.rs

@@ -190,7 +190,7 @@ pub(crate) async fn build_local_stream_attempt_source<'a>(
 #[async_trait]
 impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalSameFormatProviderSyncAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiSyncAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_sync_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -213,6 +213,21 @@ impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalSameFormatProviderSyncA
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 #[async_trait]
@@ -220,7 +235,7 @@ impl LocalExecutionAttemptSource<AiStreamAttempt>
     for LocalSameFormatProviderStreamAttemptSource<'_>
 {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiStreamAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_stream_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -243,6 +258,21 @@ impl LocalExecutionAttemptSource<AiStreamAttempt>
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 impl LocalSameFormatProviderSyncAttemptSource<'_> {
@@ -372,7 +402,7 @@ pub(crate) async fn build_local_sync_plan_and_reports(
     }
 
     let mut plans = Vec::new();
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_same_format_provider_decision_payload_for_candidate(
             state, parts, trace_id, body_json, &input, attempt, spec,
         )
@@ -458,7 +488,7 @@ pub(crate) async fn build_local_stream_plan_and_reports(
     }
 
     let mut plans = Vec::new();
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_same_format_provider_decision_payload_for_candidate(
             state, parts, trace_id, body_json, &input, attempt, spec,
         )

apps/aether-gateway/src/ai_serving/planner/passthrough/provider/request/url.rs

@@ -24,6 +24,7 @@ pub(crate) fn build_same_format_upstream_url(
             upstream_is_stream,
             request_query: parts.uri.query(),
             kiro_api_region: kiro_auth.map(|auth| auth.auth_config.effective_api_region()),
+            api_operation: spec.operation,
             provider_request_body,
         },
     )

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

@@ -127,7 +127,7 @@ fn provider_key_score_input(
         oauth_invalid_reason: key.oauth_invalid_reason.clone(),
         success_count: key.success_count.unwrap_or(0).into(),
         error_count: key.error_count.unwrap_or(0).into(),
-        total_response_time_ms: key.total_response_time_ms.unwrap_or(0).into(),
+        total_response_time_ms: key.total_response_time_ms.unwrap_or(0),
         total_tokens: key.total_tokens,
         total_cost_usd: key.total_cost_usd,
         last_used_at: key.last_used_at_unix_secs,

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

@@ -1,5 +1,5 @@
 use std::borrow::Cow;
-use std::time::{SystemTime, UNIX_EPOCH};
+use std::time::{Instant, SystemTime, UNIX_EPOCH};
 
 use serde_json::Value;
 use tracing::warn;
@@ -7,9 +7,11 @@ use tracing::warn;
 use crate::ai_serving::ExecutionRuntimeAuthContext;
 use crate::privacy::{
     build_redaction_session_config, read_chat_pii_redaction_runtime_config,
-    try_mask_chat_pii_request_json_with_cache_options, ChatPiiRedactionRequestFormat,
+    try_mask_chat_pii_request_value_with_cache_options, CachedRequestRedaction,
-    MaskChatRequestOptions, RedactionMaskError, RedactionSessionSlot, RedisRedactionMappingCache,
+    ChatPiiRedactionRequestFormat, MaskChatRequestOptions, RedactionMaskError,
+    RedactionSessionSlot, RedisRedactionMappingCache,
 };
+use crate::stage_metrics::observe_gateway_stage_ms;
 use crate::{AppState, GatewayError};
 
 pub(crate) struct ProviderRequestRedaction<'a> {
@@ -73,6 +75,20 @@ pub(crate) async fn resolve_provider_chat_pii_redaction<'a>(
     let Some(slot) = parts.extensions.get::<RedactionSessionSlot>() else {
         return Ok(ProviderRequestRedaction::disabled(body_json));
     };
+    let request_cache_key = request_redaction_cache_key(format, body_json);
+    if let Some(cached) = slot.cached_request_redaction(&request_cache_key) {
+        crate::stage_metrics::record_chat_pii_redaction_request_cache_hit();
+        observe_gateway_stage_ms("chat_pii_redaction_request_cache_hit", 0);
+        return Ok(provider_redaction_from_cached(
+            slot,
+            candidate_id,
+            body_json,
+            cached,
+        ));
+    }
+    crate::stage_metrics::record_chat_pii_redaction_request_cache_miss();
+
+    let runtime_config_started_at = Instant::now();
     let runtime_config = read_chat_pii_redaction_runtime_config(state)
         .await
         .map_err(|err| {
@@ -82,11 +98,22 @@ pub(crate) async fn resolve_provider_chat_pii_redaction<'a>(
             );
             GatewayError::Internal("chat pii redaction setup failed".to_string())
         })?;
+    observe_gateway_stage_ms(
+        "chat_pii_redaction_runtime_config",
+        runtime_config_started_at.elapsed().as_millis() as u64,
+    );
     if !runtime_config.enabled {
+        slot.put_cached_request_redaction(request_cache_key, CachedRequestRedaction::unredacted());
         return Ok(ProviderRequestRedaction::disabled(body_json));
     }
+    let feature_settings_started_at = Instant::now();
     let feature_settings = resolve_chat_pii_redaction_feature_settings(state, auth_context).await?;
+    observe_gateway_stage_ms(
+        "chat_pii_redaction_feature_settings",
+        feature_settings_started_at.elapsed().as_millis() as u64,
+    );
     if !feature_settings.effective_enabled() {
+        slot.put_cached_request_redaction(request_cache_key, CachedRequestRedaction::unredacted());
         return Ok(ProviderRequestRedaction::disabled(body_json));
     }
     let Some(hmac_key) = state.encryption_key().map(str::as_bytes).map(Vec::from) else {
@@ -95,20 +122,14 @@ pub(crate) async fn resolve_provider_chat_pii_redaction<'a>(
             "chat pii redaction setup failed".to_string(),
         ));
     };
-    let body_bytes = serde_json::to_vec(body_json).map_err(|err| {
-        warn!(
-            error = ?err,
-            "gateway failed to serialize provider chat pii redaction body"
-        );
-        GatewayError::Internal("chat pii redaction setup failed".to_string())
-    })?;
     let now_unix_secs = SystemTime::now()
         .duration_since(UNIX_EPOCH)
         .unwrap_or_default()
         .as_secs();
     let cache = RedisRedactionMappingCache::new(state.runtime_state.as_ref());
-    let masked = try_mask_chat_pii_request_json_with_cache_options(
+    let mask_started_at = Instant::now();
-        &body_bytes,
+    let masked = try_mask_chat_pii_request_value_with_cache_options(
+        body_json,
         format,
         build_redaction_session_config(hmac_key, &runtime_config, now_unix_secs),
         MaskChatRequestOptions::runtime(),
@@ -116,19 +137,27 @@ pub(crate) async fn resolve_provider_chat_pii_redaction<'a>(
     )
     .await
     .map_err(redaction_mask_error_to_gateway_error)?;
+    observe_gateway_stage_ms(
+        "chat_pii_redaction_mask_body",
+        mask_started_at.elapsed().as_millis() as u64,
+    );
     if !masked.redacted {
+        slot.put_cached_request_redaction(request_cache_key, CachedRequestRedaction::unredacted());
         return Ok(ProviderRequestRedaction {
             body_json: Cow::Borrowed(body_json),
             redacted: false,
         });
     }
-    let masked_body_json = serde_json::from_slice::<Value>(&masked.body).map_err(|err| {
+    let Some(masked_body_json) = masked.body_json else {
-        warn!(
+        warn!("gateway pii redaction reported redacted without masked body");
-            error = ?err,
+        return Err(GatewayError::Internal(
-            "gateway failed to decode redacted provider chat pii body"
+            "chat pii redaction setup failed".to_string(),
-        );
+        ));
-        GatewayError::Internal("chat pii redaction setup failed".to_string())
+    };
-    })?;
+    slot.put_cached_request_redaction(
+        request_cache_key,
+        CachedRequestRedaction::redacted(masked_body_json.clone(), masked.session.clone()),
+    );
     slot.put_for_candidate(candidate_id, masked.session);
     Ok(ProviderRequestRedaction {
         body_json: Cow::Owned(masked_body_json),
@@ -136,31 +165,46 @@ pub(crate) async fn resolve_provider_chat_pii_redaction<'a>(
     })
 }
 
+fn request_redaction_cache_key(format: ChatPiiRedactionRequestFormat, body_json: &Value) -> String {
+    format!("{format:?}:{:p}", body_json)
+}
+
+fn provider_redaction_from_cached<'a>(
+    slot: &RedactionSessionSlot,
+    candidate_id: &str,
+    body_json: &'a Value,
+    cached: CachedRequestRedaction,
+) -> ProviderRequestRedaction<'a> {
+    if !cached.redacted {
+        return ProviderRequestRedaction::disabled(body_json);
+    }
+    let Some(masked_body_json) = cached.body_json else {
+        return ProviderRequestRedaction::disabled(body_json);
+    };
+    if let Some(session) = cached.session {
+        slot.put_for_candidate(candidate_id, session);
+    }
+    ProviderRequestRedaction {
+        body_json: Cow::Owned(masked_body_json),
+        redacted: true,
+    }
+}
+
 async fn resolve_chat_pii_redaction_feature_settings(
     state: &AppState,
     auth_context: &ExecutionRuntimeAuthContext,
 ) -> Result<ChatPiiRedactionFeatureSettings, GatewayError> {
-    let user_settings = state
+    let user_settings_fut = state.read_user_feature_settings(&auth_context.user_id);
-        .read_user_feature_settings(&auth_context.user_id)
+    let key_settings_fut = state.read_auth_api_key_feature_settings(
-        .await
+        &auth_context.user_id,
+        &auth_context.api_key_id,
+        auth_context.api_key_is_standalone,
+    );
+    let (user_settings, key_settings) = tokio::try_join!(user_settings_fut, key_settings_fut)
         .map_err(|err| {
             warn!(
                 error = ?err,
-                "gateway failed to read user chat pii redaction feature settings"
+                "gateway failed to read chat pii redaction feature settings"
-            );
-            GatewayError::Internal("chat pii redaction setup failed".to_string())
-        })?;
-    let key_settings = state
-        .read_auth_api_key_feature_settings(
-            &auth_context.user_id,
-            &auth_context.api_key_id,
-            auth_context.api_key_is_standalone,
-        )
-        .await
-        .map_err(|err| {
-            warn!(
-                error = ?err,
-                "gateway failed to read api key chat pii redaction feature settings"
             );
             GatewayError::Internal("chat pii redaction setup failed".to_string())
         })?;
@@ -172,12 +216,7 @@ async fn resolve_chat_pii_redaction_feature_settings(
 }
 
 fn redaction_mask_error_to_gateway_error(error: RedactionMaskError) -> GatewayError {
-    match error {
+    match error {}
-        RedactionMaskError::Limit(limit) => GatewayError::Client {
-            status: limit.client_status(),
-            message: limit.safe_message().to_string(),
-        },
-    }
 }
 
 #[cfg(test)]

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

@@ -6,6 +6,7 @@ use aether_ai_serving::{
     provider_stream_event_api_format_for_provider_type as ai_provider_stream_event_api_format_for_provider_type,
     AiExecutionReportContextParts, AiRequestOrigin,
 };
+use aether_routing_core::ResolvedRoutingPolicy;
 use aether_runtime_state::RuntimeLockLease;
 use aether_scheduler_core::{ClientSessionAffinity, SchedulerRankingOutcome};
 use serde_json::{Map, Value};
@@ -20,8 +21,9 @@ use crate::client_session_affinity::{
 };
 use crate::orchestration::{
     insert_pool_key_lease_report_context_fields, ExecutionAttemptIdentity,
-    SCHEDULER_AFFINITY_EPOCH_REPORT_FIELD,
+    ROUTING_POOL_POLICY_OVERRIDE_REPORT_FIELD, SCHEDULER_AFFINITY_EPOCH_REPORT_FIELD,
 };
+use crate::scheduler::affinity::insert_scheduler_affinity_policy_report_context_field;
 
 pub(crate) struct LocalExecutionReportContextParts<'a> {
     pub(crate) auth_context: &'a ExecutionRuntimeAuthContext,
@@ -55,6 +57,7 @@ pub(crate) struct LocalExecutionReportContextParts<'a> {
     pub(crate) original_request_body_json: Option<&'a Value>,
     pub(crate) original_request_body_base64: Option<&'a str>,
     pub(crate) client_session_affinity: Option<&'a ClientSessionAffinity>,
+    pub(crate) routing_policy: Option<&'a ResolvedRoutingPolicy>,
     pub(crate) scheduler_affinity_epoch: Option<u64>,
     pub(crate) client_requested_stream: bool,
     pub(crate) upstream_is_stream: bool,
@@ -105,6 +108,16 @@ pub(crate) fn build_local_execution_report_context(
         merge_incoming_tls_fingerprint(&mut extra_fields, incoming_tls);
     }
     insert_pool_key_lease_report_context_fields(&mut extra_fields, parts.pool_key_lease);
+    insert_scheduler_affinity_policy_report_context_field(&mut extra_fields, parts.routing_policy);
+    if let Some(override_policy) = parts
+        .routing_policy
+        .and_then(|policy| policy.pool_policy_overrides.get(parts.provider_id))
+        .filter(|override_policy| !override_policy.scheduling_presets.is_empty())
+    {
+        if let Ok(value) = serde_json::to_value(override_policy) {
+            extra_fields.insert(ROUTING_POOL_POLICY_OVERRIDE_REPORT_FIELD.to_string(), value);
+        }
+    }
     if let Some(epoch) = parts.scheduler_affinity_epoch {
         extra_fields.insert(
             SCHEDULER_AFFINITY_EPOCH_REPORT_FIELD.to_string(),
@@ -315,6 +328,7 @@ mod tests {
                 original_request_body_json: Some(&json!({"model": "gpt-5"})),
                 original_request_body_base64: None,
                 client_session_affinity: Some(&client_session_affinity),
+                routing_policy: None,
                 scheduler_affinity_epoch: None,
                 client_requested_stream: false,
                 upstream_is_stream: false,
@@ -397,6 +411,7 @@ mod tests {
                 })),
                 original_request_body_base64: None,
                 client_session_affinity: None,
+                routing_policy: None,
                 scheduler_affinity_epoch: None,
                 client_requested_stream: false,
                 upstream_is_stream: true,
@@ -463,6 +478,7 @@ mod tests {
                 original_request_body_json: Some(&json!({"model": "gpt-5"})),
                 original_request_body_base64: None,
                 client_session_affinity: None,
+                routing_policy: None,
                 scheduler_affinity_epoch: None,
                 client_requested_stream: false,
                 upstream_is_stream: false,

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

@@ -1,23 +1,50 @@
 use aether_ai_serving::AiRequestGzipPolicy;
 use serde_json::Value;
 
-use crate::ai_serving::is_openai_responses_family_format;
+use crate::ai_serving::{normalize_api_format_alias, parse_codex_auth_identity};
 
 use super::state::GatewayProviderTransportSnapshot;
 
-const DEFAULT_CODEX_REQUEST_GZIP_MIN_BYTES: usize = 64 * 1024;
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
-
+pub(crate) struct TransportRequestEncodingPolicy {
-pub(crate) fn resolve_transport_request_gzip_policy(
+    pub content_encoding: Option<String>,
-    transport: &GatewayProviderTransportSnapshot,
+    pub request_gzip: Option<AiRequestGzipPolicy>,
-) -> Option<AiRequestGzipPolicy> {
-    transport_request_gzip_policy_from_config(transport.endpoint.config.as_ref())
-        .or_else(|| transport_request_gzip_policy_from_config(transport.provider.config.as_ref()))
-        .or_else(|| default_transport_request_gzip_policy(transport))
 }
 
-fn default_transport_request_gzip_policy(
+pub(crate) fn resolve_transport_request_encoding_policy(
     transport: &GatewayProviderTransportSnapshot,
-) -> Option<AiRequestGzipPolicy> {
+) -> TransportRequestEncodingPolicy {
+    if transport
+        .provider
+        .provider_type
+        .trim()
+        .eq_ignore_ascii_case("codex")
+        && normalize_api_format_alias(transport.endpoint.api_format.as_str())
+            == "openai:responses:compact"
+    {
+        return TransportRequestEncodingPolicy::default();
+    }
+
+    let request_gzip = transport_request_gzip_policy_from_config(
+        transport.endpoint.config.as_ref(),
+    )
+    .or_else(|| transport_request_gzip_policy_from_config(transport.provider.config.as_ref()));
+    if request_gzip.is_some() {
+        return TransportRequestEncodingPolicy {
+            content_encoding: None,
+            request_gzip,
+        };
+    }
+
+    TransportRequestEncodingPolicy {
+        content_encoding: default_transport_request_content_encoding(transport),
+        request_gzip: None,
+    }
+}
+
+fn default_transport_request_content_encoding(
+    transport: &GatewayProviderTransportSnapshot,
+) -> Option<String> {
     if !transport
         .provider
         .provider_type
@@ -26,19 +53,24 @@ fn default_transport_request_gzip_policy(
     {
         return None;
     }
-    if !is_codex_request_gzip_endpoint_api_format(transport.endpoint.api_format.as_str()) {
+    if !is_codex_request_compression_api_format(transport.endpoint.api_format.as_str()) {
+        return None;
+    }
+    let auth_type =
+        crate::ai_serving::transport::auth::resolve_local_auth_type_for_transport_format(transport);
+    let uses_codex_backend = auth_type == "oauth"
+        || (auth_type == "bearer"
+            && parse_codex_auth_identity(transport.key.decrypted_auth_config.as_deref())
+                .uses_codex_backend);
+    if !uses_codex_backend {
         return None;
     }
 
-    Some(AiRequestGzipPolicy {
+    Some("zstd".to_string())
-        enabled: Some(true),
-        min_bytes: Some(DEFAULT_CODEX_REQUEST_GZIP_MIN_BYTES),
-    })
 }
 
-fn is_codex_request_gzip_endpoint_api_format(api_format: &str) -> bool {
+fn is_codex_request_compression_api_format(api_format: &str) -> bool {
-    is_openai_responses_family_format(api_format)
+    normalize_api_format_alias(api_format) == "openai:responses"
-        || api_format.trim().eq_ignore_ascii_case("openai:image")
 }
 
 fn transport_request_gzip_policy_from_config(
@@ -216,6 +248,16 @@ mod tests {
         }
     }
 
+    fn resolved_gzip_policy(
+        transport: &GatewayProviderTransportSnapshot,
+    ) -> Option<AiRequestGzipPolicy> {
+        resolve_transport_request_encoding_policy(transport).request_gzip
+    }
+
+    fn resolved_content_encoding(transport: &GatewayProviderTransportSnapshot) -> Option<String> {
+        resolve_transport_request_encoding_policy(transport).content_encoding
+    }
+
     #[test]
     fn endpoint_request_gzip_policy_overrides_provider_policy() {
         let transport = sample_transport(
@@ -226,7 +268,7 @@ mod tests {
         );
 
         assert_eq!(
-            resolve_transport_request_gzip_policy(&transport),
+            resolved_gzip_policy(&transport),
             Some(AiRequestGzipPolicy {
                 enabled: Some(true),
                 min_bytes: Some(1024),
@@ -244,7 +286,7 @@ mod tests {
         );
 
         assert_eq!(
-            resolve_transport_request_gzip_policy(&transport),
+            resolved_gzip_policy(&transport),
             Some(AiRequestGzipPolicy {
                 enabled: Some(false),
                 min_bytes: None,
@@ -265,7 +307,7 @@ mod tests {
         );
 
         assert_eq!(
-            resolve_transport_request_gzip_policy(&transport),
+            resolved_gzip_policy(&transport),
             Some(AiRequestGzipPolicy {
                 enabled: Some(true),
                 min_bytes: Some(4096),
@@ -283,7 +325,7 @@ mod tests {
         );
 
         assert_eq!(
-            resolve_transport_request_gzip_policy(&transport),
+            resolved_gzip_policy(&transport),
             Some(AiRequestGzipPolicy {
                 enabled: Some(true),
                 min_bytes: Some(1),
@@ -292,35 +334,73 @@ mod tests {
     }
 
     #[test]
-    fn codex_responses_endpoint_gets_default_request_gzip_policy() {
+    fn codex_responses_endpoint_uses_zstd_without_a_size_threshold() {
-        let transport = sample_transport("codex", "openai:responses", None, None);
+        let mut transport = sample_transport("codex", "openai:responses", None, None);
+        transport.key.auth_type = "oauth".to_string();
 
         assert_eq!(
-            resolve_transport_request_gzip_policy(&transport),
+            resolved_content_encoding(&transport).as_deref(),
-            Some(AiRequestGzipPolicy {
+            Some("zstd")
-                enabled: Some(true),
-                min_bytes: Some(DEFAULT_CODEX_REQUEST_GZIP_MIN_BYTES),
-            })
         );
+        assert_eq!(resolved_gzip_policy(&transport), None);
     }
 
     #[test]
-    fn codex_image_endpoint_gets_default_request_gzip_policy() {
+    fn codex_responses_api_key_auth_does_not_enable_default_compression() {
-        let transport = sample_transport("codex", "openai:image", None, None);
+        let transport = sample_transport("codex", "openai:responses", None, None);
+
+        assert_eq!(resolved_content_encoding(&transport), None);
+        assert_eq!(resolved_gzip_policy(&transport), None);
+    }
+
+    #[test]
+    fn codex_responses_bearer_auth_uses_identity_metadata_for_backend_compression() {
+        let mut transport = sample_transport("codex", "openai:responses", None, None);
+        transport.key.auth_type = "bearer".to_string();
+        transport.key.decrypted_auth_config =
+            Some(r#"{"provider_type":"codex","account_id":"account-1"}"#.to_string());
 
         assert_eq!(
-            resolve_transport_request_gzip_policy(&transport),
+            resolved_content_encoding(&transport).as_deref(),
-            Some(AiRequestGzipPolicy {
+            Some("zstd")
-                enabled: Some(true),
-                min_bytes: Some(DEFAULT_CODEX_REQUEST_GZIP_MIN_BYTES),
-            })
         );
+        assert_eq!(resolved_gzip_policy(&transport), None);
+    }
+
+    #[test]
+    fn codex_image_endpoint_does_not_get_responses_request_gzip_policy() {
+        let transport = sample_transport("codex", "openai:image", None, None);
+
+        assert_eq!(resolved_content_encoding(&transport), None);
+        assert_eq!(resolved_gzip_policy(&transport), None);
+    }
+
+    #[test]
+    fn codex_compact_endpoint_does_not_get_default_request_gzip_policy() {
+        let transport = sample_transport("codex", "openai:responses:compact", None, None);
+
+        assert_eq!(resolved_content_encoding(&transport), None);
+        assert_eq!(resolved_gzip_policy(&transport), None);
+    }
+
+    #[test]
+    fn codex_compact_endpoint_rejects_an_explicit_request_gzip_policy() {
+        let transport = sample_transport(
+            "codex",
+            "openai:responses:compact",
+            None,
+            Some(json!({"request_gzip": {"enabled": true, "min_bytes": 2048}})),
+        );
+
+        assert_eq!(resolved_gzip_policy(&transport), None);
+        assert_eq!(resolved_content_encoding(&transport), None);
     }
 
     #[test]
     fn non_codex_endpoint_does_not_get_default_request_gzip_policy() {
         let transport = sample_transport("openai", "openai:responses", None, None);
 
-        assert_eq!(resolve_transport_request_gzip_policy(&transport), None);
+        assert_eq!(resolved_content_encoding(&transport), None);
+        assert_eq!(resolved_gzip_policy(&transport), None);
     }
 }

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

@@ -1,8 +1,8 @@
 use crate::ai_serving::GatewayControlDecision;
 use crate::ai_serving::{
     is_matching_stream_http_request as is_matching_stream_http_request_impl,
-    resolve_execution_runtime_stream_plan_kind as resolve_execution_runtime_stream_plan_kind_impl,
+    resolve_execution_runtime_stream_plan_kind_with_client_surface as resolve_execution_runtime_stream_plan_kind_impl,
-    resolve_execution_runtime_sync_plan_kind as resolve_execution_runtime_sync_plan_kind_impl,
+    resolve_execution_runtime_sync_plan_kind_with_client_surface as resolve_execution_runtime_sync_plan_kind_impl,
     supports_stream_execution_decision_kind as supports_stream_execution_decision_kind_impl,
     supports_sync_execution_decision_kind as supports_sync_execution_decision_kind_impl,
 };
@@ -11,28 +11,34 @@ pub(crate) fn resolve_execution_runtime_stream_plan_kind(
     parts: &http::request::Parts,
     decision: &GatewayControlDecision,
 ) -> Option<&'static str> {
-    resolve_execution_runtime_stream_plan_kind_impl(
+    let plan_kind = resolve_execution_runtime_stream_plan_kind_impl(
         decision.route_class.as_deref(),
         decision.route_family.as_deref(),
         decision.route_kind.as_deref(),
+        decision.client_surface,
         decision.request_auth_channel.as_deref(),
         &parts.method,
         parts.uri.path(),
-    )
+    )?;
+    crate::ai_serving::plan_kind_matches_api_operation(plan_kind, true, decision.api_operation)
+        .then_some(plan_kind)
 }
 
 pub(crate) fn resolve_execution_runtime_sync_plan_kind(
     parts: &http::request::Parts,
     decision: &GatewayControlDecision,
 ) -> Option<&'static str> {
-    resolve_execution_runtime_sync_plan_kind_impl(
+    let plan_kind = resolve_execution_runtime_sync_plan_kind_impl(
         decision.route_class.as_deref(),
         decision.route_family.as_deref(),
         decision.route_kind.as_deref(),
+        decision.client_surface,
         decision.request_auth_channel.as_deref(),
         &parts.method,
         parts.uri.path(),
-    )
+    )?;
+    crate::ai_serving::plan_kind_matches_api_operation(plan_kind, false, decision.api_operation)
+        .then_some(plan_kind)
 }
 
 pub(crate) fn is_matching_stream_request(
@@ -62,7 +68,7 @@ mod tests {
         resolve_execution_runtime_sync_plan_kind, supports_stream_execution_decision_kind,
         supports_sync_execution_decision_kind,
     };
-    use crate::ai_serving::GatewayControlDecision;
+    use crate::ai_serving::{ApiOperation, ClientSurface, GatewayControlDecision};
 
     fn sample_decision(route_family: &str, route_kind: &str) -> GatewayControlDecision {
         GatewayControlDecision {
@@ -71,12 +77,16 @@ mod tests {
             route_class: Some("ai_public".to_string()),
             route_family: Some(route_family.to_string()),
             route_kind: Some(route_kind.to_string()),
+            client_surface: None,
+            api_operation: None,
+            gateway_credential_carrier: None,
             request_auth_channel: None,
             auth_context: None,
             admin_principal: None,
             auth_endpoint_signature: None,
             execution_runtime_candidate: true,
             local_auth_rejection: None,
+            model_directive_policy: Default::default(),
         }
     }
 
@@ -120,7 +130,9 @@ mod tests {
         let (claude_parts, _) = claude_request.into_parts();
 
         let claude_api_key = sample_decision_with_auth_channel("claude", "messages", "api_key");
-        let claude_bearer = sample_decision_with_auth_channel("claude", "messages", "bearer_like");
+        let mut claude_bearer =
+            sample_decision_with_auth_channel("claude", "messages", "bearer_like");
+        claude_bearer.client_surface = Some(ClientSurface::ClaudeCode);
         assert_eq!(
             resolve_execution_runtime_sync_plan_kind(&claude_parts, &claude_api_key),
             Some("claude_chat_sync")
@@ -130,6 +142,13 @@ mod tests {
             Some("claude_cli_stream")
         );
 
+        let claude_sdk_bearer =
+            sample_decision_with_auth_channel("claude", "messages", "bearer_like");
+        assert_eq!(
+            resolve_execution_runtime_sync_plan_kind(&claude_parts, &claude_sdk_bearer),
+            Some("claude_chat_sync")
+        );
+
         let gemini_request = Request::builder()
             .method(Method::POST)
             .uri("/v1beta/models/gemini-2.5-pro:generateContent")
@@ -151,6 +170,36 @@ mod tests {
         );
     }
 
+    #[test]
+    fn resolves_claude_count_tokens_as_native_sync_operation() {
+        let request = Request::builder()
+            .method(Method::POST)
+            .uri("/v1/messages/count_tokens")
+            .body(())
+            .expect("request should build");
+        let (parts, _) = request.into_parts();
+        let mut decision = sample_decision("claude", "count_tokens");
+        decision.api_operation = Some(ApiOperation::ClaudeCountTokens);
+
+        assert_eq!(
+            resolve_execution_runtime_sync_plan_kind(&parts, &decision),
+            Some("claude_count_tokens_sync")
+        );
+        assert!(supports_sync_execution_decision_kind(
+            "claude_count_tokens_sync"
+        ));
+
+        decision.api_operation = Some(ApiOperation::ClaudeMessagesCreate);
+        assert_eq!(
+            resolve_execution_runtime_sync_plan_kind(&parts, &decision),
+            None
+        );
+        assert_eq!(
+            resolve_execution_runtime_stream_plan_kind(&parts, &decision),
+            None
+        );
+    }
+
     #[test]
     fn stream_matching_uses_surface_route_logic() {
         let request = Request::builder()

apps/aether-gateway/src/ai_serving/planner/specialized/files.rs

@@ -175,7 +175,7 @@ pub(crate) async fn build_local_gemini_files_stream_attempt_source_for_kind<'a>(
 #[async_trait]
 impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalGeminiFilesSyncAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiSyncAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_sync_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -193,12 +193,27 @@ impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalGeminiFilesSyncAttemptS
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 #[async_trait]
 impl LocalExecutionAttemptSource<AiStreamAttempt> for LocalGeminiFilesStreamAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiStreamAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_stream_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -216,6 +231,21 @@ impl LocalExecutionAttemptSource<AiStreamAttempt> for LocalGeminiFilesStreamAtte
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 impl LocalGeminiFilesSyncAttemptSource<'_> {
@@ -323,7 +353,7 @@ pub(crate) async fn maybe_build_sync_local_gemini_files_decision_payload(
     let (mut source, _) =
         build_local_gemini_files_candidate_attempt_source(state, trace_id, &input).await?;
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) = maybe_build_local_gemini_files_decision_payload_for_candidate(
             state,
             parts,
@@ -365,7 +395,7 @@ pub(crate) async fn maybe_build_stream_local_gemini_files_decision_payload(
         build_local_gemini_files_candidate_attempt_source(state, trace_id, &input).await?;
 
     let empty_body_json = serde_json::Value::Null;
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) = maybe_build_local_gemini_files_decision_payload_for_candidate(
             state,
             parts,
@@ -414,7 +444,7 @@ async fn build_local_sync_plan_and_reports(
         build_local_gemini_files_candidate_attempt_source(state, trace_id, &input).await?;
 
     let mut plans = Vec::new();
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_gemini_files_decision_payload_for_candidate(
             state,
             parts,
@@ -467,7 +497,7 @@ async fn build_local_stream_plan_and_reports(
 
     let mut plans = Vec::new();
     let empty_body_json = serde_json::Value::Null;
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_gemini_files_decision_payload_for_candidate(
             state,
             parts,

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

@@ -7,14 +7,14 @@ use crate::ai_serving::planner::report_context::{
 };
 use crate::ai_serving::planner::spec_metadata::local_gemini_files_spec_metadata;
 use crate::ai_serving::planner::{
-    build_ai_execution_decision_response, resolve_transport_request_gzip_policy,
+    build_ai_execution_decision_response, resolve_transport_request_encoding_policy,
     AiExecutionDecisionResponseParts,
 };
 use crate::ai_serving::transport::{
     resolve_transport_execution_timeouts, resolve_transport_profile,
 };
 use crate::ai_serving::{ai_local_execution_contract_for_formats, PlannerAppState};
-use crate::{AiExecutionDecision, AppState, GatewayError};
+use crate::{append_local_failover_policy_to_value, AiExecutionDecision, AppState, GatewayError};
 
 use super::request::resolve_local_gemini_files_candidate_payload_parts;
 use super::support::{
@@ -107,6 +107,7 @@ pub(super) async fn maybe_build_local_gemini_files_decision_payload_for_candidat
         original_request_body_json: Some(body_json),
         original_request_body_base64: resolved.provider_request_body_base64.as_deref(),
         client_session_affinity: input.client_session_affinity.as_ref(),
+        routing_policy: input.routing_policy.as_ref(),
         scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
         client_requested_stream: spec_metadata.require_streaming,
         upstream_is_stream: spec_metadata.require_streaming,
@@ -114,6 +115,7 @@ pub(super) async fn maybe_build_local_gemini_files_decision_payload_for_candidat
         needs_conversion: false,
         extra_fields,
     });
+    let report_context = append_local_failover_policy_to_value(report_context, &transport);
     let super::request::LocalGeminiFilesCandidatePayloadParts {
         transport: _,
         auth_header,
@@ -124,7 +126,7 @@ pub(super) async fn maybe_build_local_gemini_files_decision_payload_for_candidat
         upstream_url,
         file_name: _,
     } = resolved;
-    let request_gzip = resolve_transport_request_gzip_policy(&transport);
+    let request_encoding = resolve_transport_request_encoding_policy(&transport);
 
     let mut decision = build_ai_execution_decision_response(AiExecutionDecisionResponseParts {
         decision_is_stream: spec_metadata.require_streaming,
@@ -134,6 +136,7 @@ pub(super) async fn maybe_build_local_gemini_files_decision_payload_for_candidat
         request_id: trace_id.to_string(),
         candidate_id: candidate_id.clone(),
         provider_name: transport.provider.name.clone(),
+        provider_type: transport.provider.provider_type.clone(),
         provider_id: candidate.provider_id.clone(),
         endpoint_id: candidate.endpoint_id.clone(),
         key_id: candidate.key_id.clone(),
@@ -156,8 +159,8 @@ pub(super) async fn maybe_build_local_gemini_files_decision_payload_for_candidat
             .map(str::trim)
             .filter(|value| !value.is_empty())
             .map(ToOwned::to_owned),
-        content_encoding: None,
+        content_encoding: request_encoding.content_encoding,
-        request_gzip,
+        request_gzip: request_encoding.request_gzip,
         proxy,
         transport_profile,
         timeouts: resolve_transport_execution_timeouts(&transport),
@@ -166,6 +169,10 @@ pub(super) async fn maybe_build_local_gemini_files_decision_payload_for_candidat
         report_context: Some(report_context),
         auth_context: input.auth_context.clone(),
     });
-    apply_provider_request_routing_policy_to_decision(input, &mut decision)?;
+    apply_provider_request_routing_policy_to_decision(
+        input,
+        &mut decision,
+        Some(transport.as_ref()),
+    )?;
     Ok(Some(decision))
 }

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

@@ -53,7 +53,9 @@ pub(super) async fn resolve_local_gemini_files_decision_input(
         state,
         auth_context,
         None,
+        decision.auth_endpoint_signature.as_deref(),
         Some(&explicit_required_capabilities),
+        &decision.model_directive_policy,
     )
     .await
     {

apps/aether-gateway/src/ai_serving/planner/specialized/image.rs

@@ -253,7 +253,7 @@ pub(crate) async fn build_local_image_stream_attempt_source_for_kind<'a>(
 #[async_trait]
 impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalOpenAiImageSyncAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiSyncAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_sync_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -271,12 +271,27 @@ impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalOpenAiImageSyncAttemptS
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 #[async_trait]
 impl LocalExecutionAttemptSource<AiStreamAttempt> for LocalOpenAiImageStreamAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiStreamAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_stream_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -294,6 +309,21 @@ impl LocalExecutionAttemptSource<AiStreamAttempt> for LocalOpenAiImageStreamAtte
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 impl LocalOpenAiImageSyncAttemptSource<'_> {
@@ -421,7 +451,7 @@ pub(crate) async fn maybe_build_sync_local_image_decision_payload(
         return Ok(None);
     };
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) = maybe_build_local_openai_image_decision_payload_for_candidate(
             state,
             parts,
@@ -482,7 +512,7 @@ pub(crate) async fn maybe_build_stream_local_image_decision_payload(
         return Ok(None);
     };
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) = maybe_build_local_openai_image_decision_payload_for_candidate(
             state,
             parts,
@@ -540,7 +570,7 @@ async fn build_local_sync_plan_and_reports(
     };
 
     let mut plans = Vec::new();
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_openai_image_decision_payload_for_candidate(
             state,
             parts,
@@ -617,7 +647,7 @@ async fn build_local_stream_plan_and_reports(
     };
 
     let mut plans = Vec::new();
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_openai_image_decision_payload_for_candidate(
             state,
             parts,

apps/aether-gateway/src/ai_serving/planner/specialized/image/decision.rs

@@ -5,7 +5,7 @@ use crate::ai_serving::planner::report_context::{
 };
 use crate::ai_serving::planner::spec_metadata::local_openai_image_spec_metadata;
 use crate::ai_serving::planner::{
-    build_ai_execution_decision_response, resolve_transport_request_gzip_policy,
+    build_ai_execution_decision_response, resolve_transport_request_encoding_policy,
     AiExecutionDecisionResponseParts,
 };
 use crate::ai_serving::transport::{
@@ -13,7 +13,8 @@ use crate::ai_serving::transport::{
 };
 use crate::ai_serving::{ai_local_execution_contract_for_formats, PlannerAppState};
 use crate::{
-    append_execution_contract_fields_to_value, AiExecutionDecision, AppState, GatewayError,
+    append_execution_contract_fields_to_value, append_local_failover_policy_to_value,
+    AiExecutionDecision, AppState, GatewayError,
 };
 
 use super::request::resolve_local_openai_image_candidate_payload_parts;
@@ -84,11 +85,7 @@ pub(super) async fn maybe_build_local_openai_image_decision_payload_for_candidat
             serde_json::Value::Bool(true),
         );
     }
-    let upstream_is_stream = resolved
+    let upstream_is_stream = resolved.upstream_is_stream;
-        .provider_request_body
-        .get("stream")
-        .and_then(serde_json::Value::as_bool)
-        .unwrap_or(spec_metadata.require_streaming);
     let effective_headers = input.effective_headers(&parts.headers);
     let report_context = append_execution_contract_fields_to_value(
         build_local_execution_report_context(LocalExecutionReportContextParts {
@@ -123,6 +120,7 @@ pub(super) async fn maybe_build_local_openai_image_decision_payload_for_candidat
             original_request_body_json: Some(body_json),
             original_request_body_base64: body_base64,
             client_session_affinity: input.client_session_affinity.as_ref(),
+            routing_policy: input.routing_policy.as_ref(),
             scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
             client_requested_stream: spec_metadata.require_streaming,
             upstream_is_stream,
@@ -135,7 +133,8 @@ pub(super) async fn maybe_build_local_openai_image_decision_payload_for_candidat
         spec_metadata.api_format,
         provider_api_format.as_str(),
     );
-    let request_gzip = resolve_transport_request_gzip_policy(&transport);
+    let report_context = append_local_failover_policy_to_value(report_context, &transport);
+    let request_encoding = resolve_transport_request_encoding_policy(&transport);
 
     let mut decision = build_ai_execution_decision_response(AiExecutionDecisionResponseParts {
         decision_is_stream: spec_metadata.require_streaming,
@@ -145,6 +144,7 @@ pub(super) async fn maybe_build_local_openai_image_decision_payload_for_candidat
         request_id: trace_id.to_string(),
         candidate_id: candidate_id.clone(),
         provider_name: transport.provider.name.clone(),
+        provider_type: transport.provider.provider_type.clone(),
         provider_id: candidate.provider_id.clone(),
         endpoint_id: candidate.endpoint_id.clone(),
         key_id: candidate.key_id.clone(),
@@ -162,8 +162,8 @@ pub(super) async fn maybe_build_local_openai_image_decision_payload_for_candidat
         provider_request_body: Some(resolved.provider_request_body),
         provider_request_body_base64: None,
         content_type: Some("application/json".to_string()),
-        content_encoding: None,
+        content_encoding: request_encoding.content_encoding,
-        request_gzip,
+        request_gzip: request_encoding.request_gzip,
         proxy,
         transport_profile,
         timeouts: resolve_transport_execution_timeouts(&transport),
@@ -172,6 +172,10 @@ pub(super) async fn maybe_build_local_openai_image_decision_payload_for_candidat
         report_context: Some(report_context),
         auth_context: input.auth_context.clone(),
     });
-    apply_provider_request_routing_policy_to_decision(input, &mut decision)?;
+    apply_provider_request_routing_policy_to_decision(
+        input,
+        &mut decision,
+        Some(transport.as_ref()),
+    )?;
     Ok(Some(decision))
 }

apps/aether-gateway/src/ai_serving/planner/specialized/image/request.rs

@@ -16,8 +16,8 @@ use crate::ai_serving::transport::{
     ProviderOpenAiImageHeadersInput, StandardProviderRequestHeadersInput, GROK_CHAT_PATH,
 };
 use crate::ai_serving::{
-    apply_codex_openai_responses_special_body_edits, apply_codex_openai_responses_special_headers,
+    apply_codex_openai_special_headers, build_chatgpt_web_image_request_body,
-    build_chatgpt_web_image_request_body,
+    build_codex_openai_image_api_provider_request_body,
     build_gemini_image_request_body_from_openai_image_request,
     build_openai_image_api_provider_request_body, build_openai_image_provider_request_body,
     default_model_for_openai_image_operation, normalize_openai_image_request,
@@ -48,6 +48,7 @@ pub(super) struct LocalOpenAiImageCandidatePayloadParts {
     pub(super) upstream_url: String,
     pub(super) input_summary: Value,
     pub(super) transport_profile: Option<ResolvedTransportProfile>,
+    pub(super) upstream_is_stream: bool,
 }
 
 pub(super) async fn resolve_local_openai_image_candidate_payload_parts(
@@ -130,7 +131,10 @@ pub(super) async fn resolve_local_openai_image_candidate_payload_parts(
         parts,
         body_json,
         body_base64,
-        openai_image_normalize_options_for_provider(&transport.provider.provider_type),
+        openai_image_normalize_options_for_provider(
+            &transport.provider.provider_type,
+            Some(prepared_candidate.mapped_model.as_str()),
+        ),
     );
     let Some(normalized_request) = normalized_request else {
         mark_skipped_local_openai_image_candidate_with_failure_diagnostic(
@@ -174,29 +178,56 @@ pub(super) async fn resolve_local_openai_image_candidate_payload_parts(
     } else {
         build_openai_image_upstream_url(transport, Some(parts.uri.path()), parts.uri.query())
     };
-    let mut provider_request_body = if is_chatgpt_web {
+    let upstream_is_stream =
-        match build_chatgpt_web_image_request_body(parts, body_json, body_base64) {
+        crate::ai_serving::planner::common::resolve_upstream_is_stream_for_provider(
-            Ok(body) => body,
+            transport.endpoint.config.as_ref(),
-            Err(err) => err.to_error_json(),
+            transport.provider.provider_type.as_str(),
-        }
+            spec_metadata.api_format,
-    } else if is_codex || is_grok {
+            spec_metadata.require_streaming && candidate.supports_streaming,
-        build_openai_image_provider_request_body(&normalized_request)
+            false,
+        );
+    let provider_request_body = if is_chatgpt_web {
+        Some(
+            match build_chatgpt_web_image_request_body(parts, body_json, body_base64) {
+                Ok(body) => body,
+                Err(err) => err.to_error_json(),
+            },
+        )
+    } else if is_codex {
+        build_codex_openai_image_api_provider_request_body(
+            &normalized_request,
+            Some(prepared_candidate.mapped_model.as_str()),
+            upstream_is_stream,
+        )
+    } else if is_grok {
+        Some(build_openai_image_provider_request_body(
+            &normalized_request,
+        ))
     } else {
         build_openai_image_api_provider_request_body(
             &normalized_request,
             Some(prepared_candidate.mapped_model.as_str()),
+            upstream_is_stream,
         )
     };
-    if !is_chatgpt_web {
+    let Some(provider_request_body) = provider_request_body else {
-        apply_codex_openai_responses_special_body_edits(
+        mark_skipped_local_openai_image_candidate_with_failure_diagnostic(
-            &mut provider_request_body,
+            state,
-            transport.provider.provider_type.as_str(),
+            input,
-            spec_metadata.api_format,
+            trace_id,
-            transport.endpoint.body_rules.as_ref(),
+            candidate,
-            Some(candidate.key_id.as_str()),
+            attempt.candidate_index,
-        );
+            &attempt.candidate_id,
-    }
+            "provider_request_body_missing",
-
+            CandidateFailureDiagnostic::provider_request_body_missing(
+                spec_metadata.api_format,
+                spec_metadata.api_format,
+                "codex_openai_images_request_contract",
+            ),
+        )
+        .await;
+        return None;
+    };
     let Some(mut provider_request_headers) = (if is_grok {
         build_grok_browser_headers(GrokHeaderInput {
             transport,
@@ -210,13 +241,16 @@ pub(super) async fn resolve_local_openai_image_candidate_payload_parts(
         })
     } else {
         build_openai_image_headers(ProviderOpenAiImageHeadersInput {
+            transport,
             headers: effective_headers,
             auth_header: &auth_header,
             auth_value: &auth_value,
-            accept: if is_codex || is_chatgpt_web {
+            accept: if is_codex {
-                "text/event-stream"
+                None
+            } else if upstream_is_stream {
+                Some("text/event-stream")
             } else {
-                "application/json"
+                Some("application/json")
             },
             header_rules: transport.endpoint.header_rules.as_ref(),
             provider_request_body: &provider_request_body,
@@ -244,7 +278,7 @@ pub(super) async fn resolve_local_openai_image_candidate_payload_parts(
         provider_request_headers.insert("x-aether-chatgpt-web-image".to_string(), "1".to_string());
     } else if is_grok {
     } else {
-        apply_codex_openai_responses_special_headers(
+        apply_codex_openai_special_headers(
             &mut provider_request_headers,
             &provider_request_body,
             effective_headers,
@@ -286,6 +320,7 @@ pub(super) async fn resolve_local_openai_image_candidate_payload_parts(
         upstream_url,
         input_summary,
         transport_profile,
+        upstream_is_stream,
     })
 }
 
@@ -402,7 +437,14 @@ async fn resolve_local_openai_image_to_gemini_candidate_payload_parts(
                 return None;
             }
         };
-    let upstream_is_stream = spec_metadata.require_streaming;
+    let upstream_is_stream =
+        crate::ai_serving::planner::common::resolve_upstream_is_stream_for_provider(
+            transport.endpoint.config.as_ref(),
+            transport.provider.provider_type.as_str(),
+            provider_api_format,
+            spec_metadata.require_streaming && candidate.supports_streaming,
+            false,
+        );
     let Some(upstream_url) = crate::ai_serving::planner::standard::build_standard_upstream_url(
         parts,
         transport,
@@ -473,6 +515,7 @@ async fn resolve_local_openai_image_to_gemini_candidate_payload_parts(
         upstream_url,
         input_summary: converted.summary_json,
         transport_profile: None,
+        upstream_is_stream,
     })
 }
 

apps/aether-gateway/src/ai_serving/planner/specialized/image/support.rs

@@ -58,7 +58,9 @@ pub(super) async fn resolve_local_openai_image_decision_input(
         state,
         auth_context,
         Some(requested_model.as_str()),
+        decision.auth_endpoint_signature.as_deref(),
         None,
+        &decision.model_directive_policy,
     )
     .await
     {
@@ -124,6 +126,7 @@ pub(super) async fn list_local_openai_image_candidate_attempts(
                 matches_client_format.then_some(&input.auth_snapshot),
                 input.client_session_affinity.as_ref(),
                 current_unix_secs(),
+                false,
             )
             .await
         {
@@ -144,8 +147,8 @@ pub(super) async fn list_local_openai_image_candidate_attempts(
                 auth_snapshot_allows_cross_format_candidate(
                     &input.auth_snapshot,
                     &input.requested_model,
+                    None,
                     candidate,
-                    false,
                 )
             });
         }
@@ -197,6 +200,7 @@ pub(super) async fn build_local_openai_image_candidate_attempt_source<'a>(
                 matches_client_format.then_some(&input.auth_snapshot),
                 input.client_session_affinity.as_ref(),
                 current_unix_secs(),
+                false,
             )
             .await
         {
@@ -206,16 +210,16 @@ pub(super) async fn build_local_openai_image_candidate_attempt_source<'a>(
                         auth_snapshot_allows_cross_format_candidate(
                             &input.auth_snapshot,
                             &input.requested_model,
+                            None,
                             candidate,
-                            false,
                         )
                     });
                     format_skipped.retain(|candidate| {
                         auth_snapshot_allows_cross_format_candidate(
                             &input.auth_snapshot,
                             &input.requested_model,
+                            None,
                             &candidate.candidate,
-                            false,
                         )
                     });
                 }

apps/aether-gateway/src/ai_serving/planner/specialized/video.rs

@@ -105,7 +105,7 @@ pub(crate) async fn build_local_video_sync_attempt_source_for_kind<'a>(
 #[async_trait]
 impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalVideoCreateSyncAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiSyncAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_sync_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -123,6 +123,21 @@ impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalVideoCreateSyncAttemptS
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 impl LocalVideoCreateSyncAttemptSource<'_> {
@@ -195,7 +210,7 @@ pub(crate) async fn maybe_build_sync_local_video_decision_payload(
         return Ok(None);
     };
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) = maybe_build_local_video_create_decision_payload_for_candidate(
             state, parts, body_json, trace_id, &input, attempt, spec,
         )
@@ -240,7 +255,7 @@ async fn build_local_sync_plan_and_reports(
     };
 
     let mut plans = Vec::new();
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_video_create_decision_payload_for_candidate(
             state, parts, body_json, trace_id, &input, attempt, spec,
         )

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

@@ -5,14 +5,14 @@ use crate::ai_serving::planner::report_context::{
 };
 use crate::ai_serving::planner::spec_metadata::local_video_create_spec_metadata;
 use crate::ai_serving::planner::{
-    build_ai_execution_decision_response, resolve_transport_request_gzip_policy,
+    build_ai_execution_decision_response, resolve_transport_request_encoding_policy,
     AiExecutionDecisionResponseParts,
 };
 use crate::ai_serving::transport::{
     resolve_transport_execution_timeouts, resolve_transport_profile,
 };
 use crate::ai_serving::{ai_local_execution_contract_for_formats, PlannerAppState};
-use crate::{AiExecutionDecision, AppState, GatewayError};
+use crate::{append_local_failover_policy_to_value, AiExecutionDecision, AppState, GatewayError};
 
 use super::request::resolve_local_video_create_candidate_payload_parts;
 use super::support::{LocalVideoCreateCandidateAttempt, LocalVideoCreateDecisionInput};
@@ -88,6 +88,7 @@ pub(super) async fn maybe_build_local_video_create_decision_payload_for_candidat
         original_request_body_json: Some(body_json),
         original_request_body_base64: None,
         client_session_affinity: input.client_session_affinity.as_ref(),
+        routing_policy: input.routing_policy.as_ref(),
         scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
         client_requested_stream: false,
         upstream_is_stream: false,
@@ -95,6 +96,7 @@ pub(super) async fn maybe_build_local_video_create_decision_payload_for_candidat
         needs_conversion: false,
         extra_fields,
     });
+    let report_context = append_local_failover_policy_to_value(report_context, &transport);
     let super::request::LocalVideoCreateCandidatePayloadParts {
         transport: _,
         auth_header,
@@ -104,7 +106,7 @@ pub(super) async fn maybe_build_local_video_create_decision_payload_for_candidat
         provider_request_body,
         upstream_url,
     } = resolved;
-    let request_gzip = resolve_transport_request_gzip_policy(&transport);
+    let request_encoding = resolve_transport_request_encoding_policy(&transport);
 
     let mut decision = build_ai_execution_decision_response(AiExecutionDecisionResponseParts {
         decision_is_stream: false,
@@ -114,6 +116,7 @@ pub(super) async fn maybe_build_local_video_create_decision_payload_for_candidat
         request_id: trace_id.to_string(),
         candidate_id: candidate_id.clone(),
         provider_name: transport.provider.name.clone(),
+        provider_type: transport.provider.provider_type.clone(),
         provider_id: candidate.provider_id.clone(),
         endpoint_id: candidate.endpoint_id.clone(),
         key_id: candidate.key_id.clone(),
@@ -137,8 +140,8 @@ pub(super) async fn maybe_build_local_video_create_decision_payload_for_candidat
             .map(str::trim)
             .filter(|value| !value.is_empty())
             .map(ToOwned::to_owned),
-        content_encoding: None,
+        content_encoding: request_encoding.content_encoding,
-        request_gzip,
+        request_gzip: request_encoding.request_gzip,
         proxy,
         transport_profile,
         timeouts: resolve_transport_execution_timeouts(&transport),
@@ -147,6 +150,10 @@ pub(super) async fn maybe_build_local_video_create_decision_payload_for_candidat
         report_context: Some(report_context),
         auth_context: input.auth_context.clone(),
     });
-    apply_provider_request_routing_policy_to_decision(input, &mut decision)?;
+    apply_provider_request_routing_policy_to_decision(
+        input,
+        &mut decision,
+        Some(transport.as_ref()),
+    )?;
     Ok(Some(decision))
 }

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

@@ -62,7 +62,9 @@ pub(super) async fn resolve_local_video_create_decision_input(
         state,
         auth_context,
         Some(requested_model.as_str()),
+        decision.auth_endpoint_signature.as_deref(),
         None,
+        &decision.model_directive_policy,
     )
     .await
     {
@@ -130,6 +132,7 @@ pub(super) async fn list_local_video_create_candidate_attempts(
             Some(&input.auth_snapshot),
             input.client_session_affinity.as_ref(),
             current_unix_secs(),
+            false,
         )
         .await
     {
@@ -186,6 +189,7 @@ pub(super) async fn build_local_video_create_candidate_attempt_source<'a>(
             Some(&input.auth_snapshot),
             input.client_session_affinity.as_ref(),
             current_unix_secs(),
+            false,
         )
         .await
     {

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

@@ -3,5 +3,43 @@
 mod tests;
 
 pub(crate) use crate::ai_serving::{
-    apply_codex_openai_responses_special_body_edits, apply_codex_openai_responses_special_headers,
+    apply_codex_openai_responses_identity_headers, apply_codex_openai_responses_special_body_edits,
+    apply_codex_openai_special_headers,
 };
+
+pub(crate) fn codex_model_capabilities_for_transport(
+    transport: &crate::ai_serving::GatewayProviderTransportSnapshot,
+    provider_api_format: &str,
+    provider_model: &str,
+    source_model: &str,
+) -> Option<crate::ai_serving::CodexResponsesModelCapabilities> {
+    codex_model_capabilities(
+        &transport.provider.provider_type,
+        provider_api_format,
+        provider_model,
+        source_model,
+        transport.key.upstream_metadata.as_ref(),
+    )
+}
+
+fn codex_model_capabilities(
+    provider_type: &str,
+    provider_api_format: &str,
+    provider_model: &str,
+    source_model: &str,
+    upstream_metadata: Option<&serde_json::Value>,
+) -> Option<crate::ai_serving::CodexResponsesModelCapabilities> {
+    let uses_codex_model_catalog =
+        crate::ai_serving::is_openai_responses_family_format(provider_api_format)
+            || crate::ai_serving::api_format_alias_matches(provider_api_format, "openai:search");
+    if !provider_type.trim().eq_ignore_ascii_case("codex") || !uses_codex_model_catalog {
+        return None;
+    }
+    Some(
+        crate::ai_serving::resolve_codex_responses_model_capabilities(
+            provider_model,
+            source_model,
+            upstream_metadata,
+        ),
+    )
+}

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

@@ -1,16 +1,58 @@
 use std::collections::BTreeMap;
 
 use super::{
-    apply_codex_openai_responses_special_body_edits, apply_codex_openai_responses_special_headers,
+    apply_codex_openai_responses_identity_headers, apply_codex_openai_responses_special_body_edits,
+    apply_codex_openai_special_headers, codex_model_capabilities,
+};
+use crate::ai_serving::planner::standard::{
+    build_cross_format_openai_responses_request_body, build_local_openai_responses_request_body,
 };
-use crate::ai_serving::planner::standard::build_local_openai_responses_request_body;
 use http::{HeaderMap, HeaderValue};
 use serde_json::json;
 
+#[test]
+fn search_uses_live_codex_model_catalog_capabilities() {
+    let metadata = crate::ai_serving::build_codex_model_catalog_metadata(&[json!({
+        "slug": "gpt-search-custom",
+        "default_reasoning_level": "low",
+        "supported_reasoning_levels": [
+            {"effort": "low"},
+            {"effort": "max"}
+        ],
+        "supports_parallel_tool_calls": true
+    })]);
+
+    let capabilities = codex_model_capabilities(
+        "codex",
+        "openai:search",
+        "gpt-search-custom",
+        "gpt-search-custom",
+        Some(&metadata),
+    )
+    .expect("Search should resolve capabilities from the Codex model catalog");
+
+    assert_eq!(
+        capabilities.default_reasoning_effort.as_deref(),
+        Some("low")
+    );
+    assert_eq!(
+        capabilities.supported_reasoning_efforts,
+        vec!["low".to_string(), "max".to_string()]
+    );
+    assert!(codex_model_capabilities(
+        "codex",
+        "openai:chat",
+        "gpt-search-custom",
+        "gpt-search-custom",
+        Some(&metadata),
+    )
+    .is_none());
+}
+
 #[test]
 fn applies_codex_defaults_when_body_rules_do_not_handle_fields() {
     let mut body = json!({
-        "model": "gpt-5",
+        "model": "gpt-5.4",
         "max_output_tokens": 128,
         "temperature": 0.3,
         "top_p": 0.9,
@@ -31,10 +73,11 @@ fn applies_codex_defaults_when_body_rules_do_not_handle_fields() {
     assert!(body.get("top_p").is_none());
     assert!(body.get("metadata").is_none());
     assert_eq!(body["store"], false);
-    assert_eq!(body["instructions"], "");
+    assert!(body.get("instructions").is_none());
     assert_eq!(body["include"], json!(["reasoning.encrypted_content"]));
     assert_eq!(body["parallel_tool_calls"], true);
-    assert!(body.get("reasoning").is_none());
+    assert_eq!(body["reasoning"]["effort"], "medium");
+    assert!(body["reasoning"].get("summary").is_none());
 }
 
 #[test]
@@ -80,7 +123,7 @@ fn strips_store_for_compact_even_when_body_rules_handle_it() {
         {"action":"set","path":"top_p","value":0.5}
     ]);
     let mut body = json!({
-        "model": "gpt-5",
+        "model": "gpt-5.4",
         "max_output_tokens": 128,
         "metadata": {"client": "desktop", "mode": "custom"},
         "store": true,
@@ -99,12 +142,29 @@ fn strips_store_for_compact_even_when_body_rules_handle_it() {
     assert!(body.get("max_output_tokens").is_none());
     assert!(body.get("store").is_none());
     assert_eq!(body["instructions"], "Keep custom");
-    assert_eq!(body["metadata"]["mode"], "custom");
+    assert!(body.get("metadata").is_none());
-    assert_eq!(body["top_p"], 0.5);
+    assert!(body.get("top_p").is_none());
+    assert_eq!(body["parallel_tool_calls"], true);
+    assert!(body.as_object().is_some_and(|object| {
+        object.keys().all(|field| {
+            matches!(
+                field.as_str(),
+                "model"
+                    | "input"
+                    | "instructions"
+                    | "tools"
+                    | "parallel_tool_calls"
+                    | "reasoning"
+                    | "service_tier"
+                    | "prompt_cache_key"
+                    | "text"
+            )
+        })
+    }));
 }
 
 #[test]
-fn injects_stable_prompt_cache_key_for_codex_requests() {
+fn does_not_synthesize_prompt_cache_key_from_api_key_identity() {
     let mut body = json!({
         "model": "gpt-5",
         "input": "hello",
@@ -118,83 +178,478 @@ fn injects_stable_prompt_cache_key_for_codex_requests() {
         Some("key-123"),
     );
 
+    assert!(body.get("prompt_cache_key").is_none());
+}
+
+#[test]
+fn adapts_generic_prompt_cache_key_to_codex_native_identity() {
+    let mut body = json!({
+        "model": "gpt-5",
+        "input": "hello",
+        "prompt_cache_key": "ltm-pc-v2-5557e02f5c9b447a97673ba330dbe77a",
+    });
+
+    apply_codex_openai_responses_special_body_edits(
+        &mut body,
+        "codex",
+        "openai:responses",
+        None,
+        Some("key-123"),
+    );
+
+    let expected_identity = "d9c5d122-7c1c-5fb1-ba9d-656062eda44e";
+    assert_eq!(body["prompt_cache_key"], expected_identity);
+    assert_eq!(body["client_metadata"]["session_id"], expected_identity);
+    assert_eq!(body["client_metadata"]["thread_id"], expected_identity);
+}
+
+#[test]
+fn preserves_native_codex_cache_identity_and_metadata() {
+    let mut body = json!({
+        "model": "gpt-5",
+        "input": "hello",
+        "prompt_cache_key": "guardian:parent-thread",
+        "client_metadata": {
+            "session_id": "native-session",
+            "thread_id": "native-thread",
+            "turn_id": "native-turn"
+        }
+    });
+    let expected = body.clone();
+
+    apply_codex_openai_responses_special_body_edits(
+        &mut body,
+        "codex",
+        "openai:responses",
+        None,
+        Some("key-123"),
+    );
+
+    assert_eq!(body["prompt_cache_key"], expected["prompt_cache_key"]);
+    assert_eq!(body["client_metadata"], expected["client_metadata"]);
+}
+
+#[test]
+fn preserves_uuid_prompt_cache_key_while_completing_codex_identity() {
+    let identity = "172c39e6-c0a0-5a70-8b63-e0f8e0d185a3";
+    let mut body = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": identity
+    });
+
+    apply_codex_openai_responses_special_body_edits(
+        &mut body,
+        "codex",
+        "openai:responses",
+        None,
+        None,
+    );
+
+    assert_eq!(body["prompt_cache_key"], identity);
+    assert_eq!(body["client_metadata"]["session_id"], identity);
+    assert_eq!(body["client_metadata"]["thread_id"], identity);
+}
+
+#[test]
+fn keeps_codex_prompt_cache_domains_distinct() {
+    let mut first = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "tenant-a"
+    });
+    let mut second = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "tenant-b"
+    });
+
+    for body in [&mut first, &mut second] {
+        apply_codex_openai_responses_special_body_edits(
+            body,
+            "codex",
+            "openai:responses",
+            None,
+            None,
+        );
+    }
+
+    assert_ne!(first["prompt_cache_key"], second["prompt_cache_key"]);
     assert_eq!(
-        body["prompt_cache_key"],
+        first["prompt_cache_key"],
-        "53363264-dbb0-5f9d-b9c7-3e92c45c5bdf"
+        first["client_metadata"]["session_id"]
+    );
+    assert_eq!(
+        second["prompt_cache_key"],
+        second["client_metadata"]["session_id"]
     );
 }
 
 #[test]
-fn keeps_existing_prompt_cache_key_for_codex_requests() {
+fn completes_partial_and_null_codex_client_metadata() {
-    let mut body = json!({
+    let mut partial = json!({
-        "model": "gpt-5",
+        "model": "gpt-5.6-luna",
         "input": "hello",
-        "prompt_cache_key": "existing-key",
+        "prompt_cache_key": "generic-affinity",
+        "client_metadata": {
+            "thread_id": "native-thread",
+            "caller": "sdk"
+        }
+    });
+    let mut null_metadata = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity",
+        "client_metadata": null
+    });
+    let mut null_session = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity",
+        "client_metadata": {
+            "session_id": null,
+            "thread_id": null,
+            "caller": "sdk"
+        }
     });
 
-    apply_codex_openai_responses_special_body_edits(
+    for body in [&mut partial, &mut null_metadata, &mut null_session] {
-        &mut body,
+        apply_codex_openai_responses_special_body_edits(
-        "codex",
+            body,
-        "openai:responses",
+            "codex",
-        None,
+            "openai:responses",
-        Some("key-123"),
+            None,
-    );
+            None,
+        );
+    }
 
-    assert_eq!(body["prompt_cache_key"], "existing-key");
+    assert_eq!(partial["client_metadata"]["thread_id"], "native-thread");
+    assert_eq!(partial["client_metadata"]["caller"], "sdk");
+    assert_eq!(
+        partial["client_metadata"]["session_id"],
+        partial["prompt_cache_key"]
+    );
+    assert_eq!(
+        null_metadata["client_metadata"]["session_id"],
+        null_metadata["prompt_cache_key"]
+    );
+    assert_eq!(
+        null_metadata["client_metadata"]["thread_id"],
+        null_metadata["prompt_cache_key"]
+    );
+    assert_eq!(
+        null_session["client_metadata"]["session_id"],
+        null_session["prompt_cache_key"]
+    );
+    assert_eq!(
+        null_session["client_metadata"]["thread_id"],
+        null_session["prompt_cache_key"]
+    );
+    assert_eq!(null_session["client_metadata"]["caller"], "sdk");
 }
 
 #[test]
-fn injects_chatgpt_account_id_and_session_headers_for_codex_requests() {
+fn leaves_malformed_codex_client_metadata_unchanged() {
+    let mut body = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity",
+        "client_metadata": "invalid"
+    });
+    let mut malformed_fields = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity",
+        "client_metadata": {
+            "session_id": 42,
+            "thread_id": ""
+        }
+    });
+    let expected_malformed_metadata = malformed_fields["client_metadata"].clone();
+
+    for candidate in [&mut body, &mut malformed_fields] {
+        apply_codex_openai_responses_special_body_edits(
+            candidate,
+            "codex",
+            "openai:responses",
+            None,
+            None,
+        );
+    }
+
+    assert_eq!(body["prompt_cache_key"], "generic-affinity");
+    assert_eq!(body["client_metadata"], "invalid");
+    assert_eq!(malformed_fields["prompt_cache_key"], "generic-affinity");
+    assert_eq!(
+        malformed_fields["client_metadata"],
+        expected_malformed_metadata
+    );
+}
+
+#[test]
+fn limits_prompt_cache_identity_adaptation_to_codex_responses_family() {
+    let original = json!({
+        "model": "gpt-5.6-luna",
+        "input": "hello",
+        "prompt_cache_key": "generic-affinity"
+    });
+    let mut standard_openai = original.clone();
+    let mut codex_compact = original.clone();
+
+    apply_codex_openai_responses_special_body_edits(
+        &mut standard_openai,
+        "openai",
+        "openai:responses",
+        None,
+        None,
+    );
+    apply_codex_openai_responses_special_body_edits(
+        &mut codex_compact,
+        "codex",
+        "openai:responses:compact",
+        None,
+        None,
+    );
+
+    assert_eq!(standard_openai, original);
+    assert_ne!(codex_compact["prompt_cache_key"], "generic-affinity");
+    assert!(codex_compact.get("client_metadata").is_none());
+}
+
+#[test]
+fn chat_to_codex_responses_adapts_prompt_cache_identity_end_to_end() {
+    let body = json!({
+        "model": "gpt-5.6-luna",
+        "messages": [{"role": "user", "content": "hello"}],
+        "prompt_cache_key": "ltm-pc-v2-5557e02f5c9b447a97673ba330dbe77a"
+    });
+
+    let provider_request_body = build_cross_format_openai_responses_request_body(
+        &body,
+        "gpt-5.6-luna",
+        "openai:chat",
+        "openai:responses",
+        true,
+        false,
+        "codex",
+        None,
+        None,
+        &HeaderMap::new(),
+        false,
+    )
+    .expect("chat to Codex Responses request should build");
+
+    let expected_identity = "d9c5d122-7c1c-5fb1-ba9d-656062eda44e";
+    assert_eq!(provider_request_body["prompt_cache_key"], expected_identity);
+    assert_eq!(
+        provider_request_body["client_metadata"]["session_id"],
+        expected_identity
+    );
+    assert_eq!(
+        provider_request_body["client_metadata"]["thread_id"],
+        expected_identity
+    );
+
+    let mut provider_request_headers = BTreeMap::new();
+    apply_codex_openai_special_headers(
+        &mut provider_request_headers,
+        &provider_request_body,
+        &HeaderMap::new(),
+        "codex",
+        "openai:responses",
+        Some("trace-codex-cache-identity"),
+        None,
+    );
+    apply_codex_openai_responses_identity_headers(
+        &mut provider_request_headers,
+        &provider_request_body,
+        "codex",
+        "openai:responses",
+    );
+    assert_eq!(
+        provider_request_headers
+            .get("session-id")
+            .map(String::as_str),
+        Some(expected_identity)
+    );
+    assert_eq!(
+        provider_request_headers
+            .get("thread-id")
+            .map(String::as_str),
+        Some(expected_identity)
+    );
+}
+
+#[test]
+fn projects_uuid_prompt_cache_identity_into_missing_session_headers() {
     let mut headers = BTreeMap::new();
     let body = json!({
         "model": "gpt-5",
         "prompt_cache_key": "172c39e6-c0a0-5a70-8b63-e0f8e0d185a3",
     });
 
-    apply_codex_openai_responses_special_headers(
+    apply_codex_openai_special_headers(
         &mut headers,
         &body,
         &HeaderMap::new(),
         "codex",
         "openai:responses",
         Some("trace-codex-123"),
-        Some(r#"{"account_id":"acc-123"}"#),
+        Some(r#"{"account_id":"acc-123","is_fedramp":true}"#),
     );
+    apply_codex_openai_responses_identity_headers(&mut headers, &body, "codex", "openai:responses");
+    assert_eq!(
+        headers.get("chatgpt-account-id"),
+        Some(&"acc-123".to_string())
+    );
+    assert_eq!(headers.get("x-client-request-id"), None);
+    assert_eq!(
+        headers.get("user-agent"),
+        Some(&"codex_cli_rs/0.144.1".to_string())
+    );
+    assert_eq!(headers.get("originator"), Some(&"codex_cli_rs".to_string()));
+    assert!(!headers.contains_key("version"));
+    assert_eq!(headers.get("x-openai-fedramp"), Some(&"true".to_string()));
+    assert_eq!(
+        headers.get("session-id").map(String::as_str),
+        Some("172c39e6-c0a0-5a70-8b63-e0f8e0d185a3")
+    );
+    assert_eq!(
+        headers.get("thread-id").map(String::as_str),
+        Some("172c39e6-c0a0-5a70-8b63-e0f8e0d185a3")
+    );
+}
 
+#[test]
+fn projects_native_codex_metadata_for_non_uuid_cache_overrides() {
+    let mut headers = BTreeMap::new();
+    let body = json!({
+        "model": "gpt-5.6-luna",
+        "prompt_cache_key": "guardian:parent-thread",
+        "client_metadata": {
+            "session_id": "019f687b-8e92-7842-9631-d5bf0dba0a3b",
+            "thread_id": "019f6d20-1111-7222-8333-444455556666"
+        }
+    });
+
+    apply_codex_openai_special_headers(
+        &mut headers,
+        &body,
+        &HeaderMap::new(),
+        "codex",
+        "openai:responses",
+        None,
+        None,
+    );
+    apply_codex_openai_responses_identity_headers(&mut headers, &body, "codex", "openai:responses");
+
+    assert_eq!(
+        headers.get("session-id").map(String::as_str),
+        Some("019f687b-8e92-7842-9631-d5bf0dba0a3b")
+    );
+    assert_eq!(
+        headers.get("thread-id").map(String::as_str),
+        Some("019f6d20-1111-7222-8333-444455556666")
+    );
+}
+
+#[test]
+fn leaves_non_native_cache_keys_out_of_identity_headers() {
+    let mut headers = BTreeMap::new();
+    let body = json!({
+        "model": "gpt-5.6-luna",
+        "prompt_cache_key": "generic-cache-key"
+    });
+
+    apply_codex_openai_special_headers(
+        &mut headers,
+        &body,
+        &HeaderMap::new(),
+        "codex",
+        "openai:responses",
+        None,
+        None,
+    );
+    apply_codex_openai_responses_identity_headers(&mut headers, &body, "codex", "openai:responses");
+
+    assert!(!headers.contains_key("session-id"));
+    assert!(!headers.contains_key("thread-id"));
+}
+
+#[test]
+fn leaves_malformed_native_metadata_out_of_identity_headers() {
+    let mut headers = BTreeMap::new();
+    let body = json!({
+        "model": "gpt-5.6-luna",
+        "prompt_cache_key": "172c39e6-c0a0-5a70-8b63-e0f8e0d185a3",
+        "client_metadata": {
+            "session_id": 42,
+            "thread_id": ""
+        }
+    });
+
+    apply_codex_openai_responses_identity_headers(&mut headers, &body, "codex", "openai:responses");
+
+    assert!(!headers.contains_key("session-id"));
+    assert!(!headers.contains_key("thread-id"));
+}
+
+#[test]
+fn injects_only_codex_client_headers_for_images_requests() {
+    let mut headers = BTreeMap::new();
+    apply_codex_openai_special_headers(
+        &mut headers,
+        &json!({
+            "model": "gpt-image-2",
+            "prompt": "draw a city"
+        }),
+        &HeaderMap::new(),
+        "codex",
+        "openai:image",
+        Some("trace-codex-image-123"),
+        Some(r#"{"account_id":"acc-123","is_fedramp":true}"#),
+    );
     assert_eq!(
         headers.get("chatgpt-account-id"),
         Some(&"acc-123".to_string())
     );
-    assert_eq!(
-        headers.get("x-client-request-id"),
-        Some(&"trace-codex-123".to_string())
-    );
     assert_eq!(
         headers.get("user-agent"),
-        Some(
+        Some(&"codex_cli_rs/0.144.1".to_string())
-            &"codex-tui/0.122.0 (Mac OS 15.2.0; arm64) vscode/2.6.11 (codex-tui; 0.122.0)"
-                .to_string()
-        )
-    );
-    assert_eq!(headers.get("originator"), Some(&"codex-tui".to_string()));
-    assert_eq!(
-        headers.get("session_id"),
-        Some(&"ab5ecce4f0d110fe".to_string())
-    );
-    assert_eq!(
-        headers.get("conversation_id"),
-        Some(&"ab5ecce4f0d110fe".to_string())
     );
+    assert_eq!(headers.get("originator"), Some(&"codex_cli_rs".to_string()));
+    assert!(!headers.contains_key("version"));
+    assert_eq!(headers.get("x-openai-fedramp"), Some(&"true".to_string()));
+    for name in ["x-client-request-id", "session-id", "thread-id"] {
+        assert!(
+            !headers.contains_key(name),
+            "unexpected Images header: {name}"
+        );
+    }
 }
 
 #[test]
-fn respects_existing_codex_request_and_session_headers() {
+fn preserves_client_context_headers_and_enforces_codex_provider_identity() {
     let mut headers = BTreeMap::new();
     headers.insert(
         "x-client-request-id".to_string(),
         "kept-by-rule-request".to_string(),
     );
-    headers.insert("session_id".to_string(), "kept-by-rule".to_string());
+    headers.insert("session-id".to_string(), "kept-by-rule-session".to_string());
+    headers.insert("thread-id".to_string(), "kept-by-rule-thread".to_string());
+    headers.insert(
+        "chatgpt-account-id".to_string(),
+        "configured-spoof".to_string(),
+    );
+    headers.insert(
+        "x-openai-fedramp".to_string(),
+        "configured-false".to_string(),
+    );
+    headers.insert(
+        "User-Agent".to_string(),
+        "AsyncOpenAI/Python 2.44.0".to_string(),
+    );
+    headers.insert("ORIGINATOR".to_string(), "sdk-client".to_string());
     let body = json!({
         "model": "gpt-5",
         "prompt_cache_key": "172c39e6-c0a0-5a70-8b63-e0f8e0d185a3",
@@ -205,12 +660,12 @@ fn respects_existing_codex_request_and_session_headers() {
         HeaderValue::from_static("user-specified-request"),
     );
     original_headers.insert(
-        "session_id",
+        "session-id",
         HeaderValue::from_static("user-specified-session"),
     );
     original_headers.insert(
-        "conversation_id",
+        "thread-id",
-        HeaderValue::from_static("user-specified-conversation"),
+        HeaderValue::from_static("user-specified-thread"),
     );
     original_headers.insert(
         "user-agent",
@@ -220,64 +675,105 @@ fn respects_existing_codex_request_and_session_headers() {
         "originator",
         HeaderValue::from_static("user-specified-originator"),
     );
+    original_headers.insert("version", HeaderValue::from_static("user-version"));
+    original_headers.insert("x-openai-fedramp", HeaderValue::from_static("user-fedramp"));
+    original_headers.insert(
+        "chatgpt-account-id",
+        HeaderValue::from_static("user-account"),
+    );
 
-    apply_codex_openai_responses_special_headers(
+    apply_codex_openai_special_headers(
         &mut headers,
         &body,
         &original_headers,
         "codex",
         "openai:responses",
         Some("trace-codex-123"),
-        Some(r#"{"account_id":"acc-123"}"#),
+        Some(r#"{"account_id":"acc-123","is_fedramp":true}"#),
     );
+    apply_codex_openai_responses_identity_headers(&mut headers, &body, "codex", "openai:responses");
 
     assert_eq!(
         headers.get("x-client-request-id"),
         Some(&"kept-by-rule-request".to_string())
     );
-    assert!(!headers.contains_key("user-agent"));
+    assert_eq!(
-    assert!(!headers.contains_key("originator"));
+        headers.get("user-agent"),
-    assert_eq!(headers.get("session_id"), Some(&"kept-by-rule".to_string()));
+        Some(&"codex_cli_rs/0.144.1".to_string())
-    assert!(!headers.contains_key("conversation_id"));
+    );
+    assert_eq!(headers.get("originator"), Some(&"codex_cli_rs".to_string()));
+    assert_eq!(
+        headers
+            .keys()
+            .filter(|name| name.eq_ignore_ascii_case("user-agent"))
+            .count(),
+        1
+    );
+    assert_eq!(
+        headers
+            .keys()
+            .filter(|name| name.eq_ignore_ascii_case("originator"))
+            .count(),
+        1
+    );
+    assert!(!headers.contains_key("version"));
+    assert_eq!(
+        headers.get("chatgpt-account-id"),
+        Some(&"acc-123".to_string())
+    );
+    assert_eq!(headers.get("x-openai-fedramp"), Some(&"true".to_string()));
+    assert_eq!(
+        headers.get("session-id"),
+        Some(&"kept-by-rule-session".to_string())
+    );
+    assert_eq!(
+        headers.get("thread-id"),
+        Some(&"kept-by-rule-thread".to_string())
+    );
 }
 
 #[test]
-fn skips_conversation_id_for_compact_codex_requests() {
+fn compact_projects_uuid_prompt_cache_identity_into_session_headers() {
     let mut headers = BTreeMap::new();
     let body = json!({
         "model": "gpt-5",
         "prompt_cache_key": "172c39e6-c0a0-5a70-8b63-e0f8e0d185a3",
     });
 
-    apply_codex_openai_responses_special_headers(
+    apply_codex_openai_special_headers(
         &mut headers,
         &body,
         &HeaderMap::new(),
         "codex",
         "openai:responses:compact",
         Some("trace-codex-compact-123"),
-        Some(r#"{"account_id":"acc-123"}"#),
+        Some(r#"{"account_id":"acc-123","is_fedramp":true}"#),
+    );
+    apply_codex_openai_responses_identity_headers(
+        &mut headers,
+        &body,
+        "codex",
+        "openai:responses:compact",
     );
 
     assert_eq!(
         headers.get("chatgpt-account-id"),
         Some(&"acc-123".to_string())
     );
-    assert_eq!(
+    assert_eq!(headers.get("x-client-request-id"), None);
-        headers.get("x-client-request-id"),
-        Some(&"trace-codex-compact-123".to_string())
-    );
     assert_eq!(
         headers.get("user-agent"),
-        Some(
+        Some(&"codex_cli_rs/0.144.1".to_string())
-            &"codex-tui/0.122.0 (Mac OS 15.2.0; arm64) vscode/2.6.11 (codex-tui; 0.122.0)"
-                .to_string()
-        )
     );
-    assert_eq!(headers.get("originator"), Some(&"codex-tui".to_string()));
+    assert_eq!(headers.get("originator"), Some(&"codex_cli_rs".to_string()));
+    assert!(!headers.contains_key("version"));
+    assert_eq!(headers.get("x-openai-fedramp"), Some(&"true".to_string()));
     assert_eq!(
-        headers.get("session_id"),
+        headers.get("session-id").map(String::as_str),
-        Some(&"ab5ecce4f0d110fe".to_string())
+        Some("172c39e6-c0a0-5a70-8b63-e0f8e0d185a3")
+    );
+    assert_eq!(
+        headers.get("thread-id").map(String::as_str),
+        Some("172c39e6-c0a0-5a70-8b63-e0f8e0d185a3")
     );
-    assert!(!headers.contains_key("conversation_id"));
 }

apps/aether-gateway/src/ai_serving/planner/standard/family/build.rs

@@ -178,7 +178,7 @@ pub(crate) async fn build_local_stream_attempt_source<'a>(
 #[async_trait]
 impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalStandardSyncAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiSyncAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_sync_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -201,12 +201,27 @@ impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalStandardSyncAttemptSour
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 #[async_trait]
 impl LocalExecutionAttemptSource<AiStreamAttempt> for LocalStandardStreamAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiStreamAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_stream_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -229,6 +244,21 @@ impl LocalExecutionAttemptSource<AiStreamAttempt> for LocalStandardStreamAttempt
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 impl LocalStandardSyncAttemptSource<'_> {
@@ -340,7 +370,7 @@ pub(crate) async fn maybe_build_sync_via_standard_family_payload(
             .await?;
     apply_local_runtime_candidate_evaluation_progress(state, trace_id, candidate_count);
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) = maybe_build_local_standard_decision_payload_for_candidate(
             state, parts, trace_id, body_json, &input, attempt, spec,
         )
@@ -390,7 +420,7 @@ pub(crate) async fn maybe_build_stream_via_standard_family_payload(
             .await?;
     apply_local_runtime_candidate_evaluation_progress(state, trace_id, candidate_count);
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) = maybe_build_local_standard_decision_payload_for_candidate(
             state, parts, trace_id, body_json, &input, attempt, spec,
         )
@@ -449,7 +479,7 @@ pub(crate) async fn build_local_sync_plan_and_reports(
         return Ok(Vec::new());
     }
     let mut plans = Vec::new();
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_standard_decision_payload_for_candidate(
             state, parts, trace_id, body_json, &input, attempt, spec,
         )
@@ -524,7 +554,7 @@ pub(crate) async fn build_local_stream_plan_and_reports(
         return Ok(Vec::new());
     }
     let mut plans = Vec::new();
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_standard_decision_payload_for_candidate(
             state, parts, trace_id, body_json, &input, attempt, spec,
         )

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

@@ -60,7 +60,9 @@ pub(super) async fn resolve_local_standard_decision_input(
         state,
         auth_context,
         Some(requested_model.as_str()),
+        decision.auth_endpoint_signature.as_deref(),
         None,
+        &decision.model_directive_policy,
     )
     .await
     {
@@ -119,8 +121,10 @@ pub(super) async fn materialize_local_standard_candidate_attempts(
     );
     let preselection = preselect_local_execution_candidates_with_serving(
         planner_state,
+        &input.model_directive_policy,
         spec_metadata.api_format,
         &input.requested_model,
+        None,
         false,
         input.required_capabilities.as_ref(),
         &input.auth_snapshot,
@@ -243,9 +247,11 @@ pub(super) async fn build_local_standard_candidate_attempt_source<'a>(
     let (source, candidate_count) =
         build_lazy_requested_model_execution_candidate_attempt_source_with_serving(
             planner_state,
+            &input.model_directive_policy,
             trace_id,
             spec_metadata.api_format,
             &input.requested_model,
+            None,
             spec_metadata.require_streaming,
             &input.auth_snapshot,
             input.client_session_affinity.as_ref(),
@@ -338,8 +344,10 @@ async fn maybe_append_gemini_image_openai_image_preselection(
 
     let image_preselection = preselect_local_execution_candidates_for_api_formats_with_serving(
         planner_state,
+        &input.model_directive_policy,
         spec_metadata.api_format,
         &input.requested_model,
+        None,
         spec_metadata.require_streaming,
         input.required_capabilities.as_ref(),
         &input.auth_snapshot,

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

@@ -15,7 +15,7 @@ use crate::ai_serving::planner::report_context::{
 use crate::ai_serving::planner::spec_metadata::local_standard_spec_metadata;
 use crate::ai_serving::planner::CandidateFailureDiagnostic;
 use crate::ai_serving::planner::{
-    build_ai_execution_decision_response, resolve_transport_request_gzip_policy,
+    build_ai_execution_decision_response, resolve_transport_request_encoding_policy,
     AiExecutionDecisionResponseParts,
 };
 use crate::ai_serving::transport::{
@@ -140,6 +140,7 @@ pub(super) async fn maybe_build_local_standard_decision_payload_for_candidate(
                 original_request_body_json,
                 original_request_body_base64: None,
                 client_session_affinity: input.client_session_affinity.as_ref(),
+                routing_policy: input.routing_policy.as_ref(),
                 scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
                 client_requested_stream: body_json
                     .get("stream")
@@ -176,7 +177,7 @@ pub(super) async fn maybe_build_local_standard_decision_payload_for_candidate(
         transport_profile: _,
         request_redacted: _,
     } = resolved;
-    let request_gzip = resolve_transport_request_gzip_policy(&transport);
+    let request_encoding = resolve_transport_request_encoding_policy(&transport);
 
     let mut decision = build_ai_execution_decision_response(AiExecutionDecisionResponseParts {
         decision_is_stream: spec_metadata.require_streaming,
@@ -186,6 +187,7 @@ pub(super) async fn maybe_build_local_standard_decision_payload_for_candidate(
         request_id: trace_id.to_string(),
         candidate_id: candidate_id.to_string(),
         provider_name: candidate.provider_name.clone(),
+        provider_type: transport.provider.provider_type.clone(),
         provider_id: candidate.provider_id.clone(),
         endpoint_id: candidate.endpoint_id.clone(),
         key_id: candidate.key_id.clone(),
@@ -203,8 +205,8 @@ pub(super) async fn maybe_build_local_standard_decision_payload_for_candidate(
         provider_request_body: Some(provider_request_body),
         provider_request_body_base64: None,
         content_type: Some("application/json".to_string()),
-        content_encoding: None,
+        content_encoding: request_encoding.content_encoding,
-        request_gzip,
+        request_gzip: request_encoding.request_gzip,
         proxy,
         transport_profile,
         timeouts,
@@ -213,7 +215,11 @@ pub(super) async fn maybe_build_local_standard_decision_payload_for_candidate(
         report_context: Some(report_context),
         auth_context: input.auth_context.clone(),
     });
-    apply_provider_request_routing_policy_to_decision(input, &mut decision)?;
+    apply_provider_request_routing_policy_to_decision(
+        input,
+        &mut decision,
+        Some(transport.as_ref()),
+    )?;
     Ok(Some(decision))
 }
 
@@ -368,10 +374,13 @@ mod tests {
             auth_snapshot: sample_auth_snapshot(),
             required_capabilities: None,
             request_auth_channel: None,
+            client_surface: None,
+            gateway_credential_carrier: None,
             client_session_affinity: None,
             routing_policy: None,
             routing_trace_seed: None,
             routing_context: None,
+            model_directive_policy: Default::default(),
         }
     }
 
@@ -475,6 +484,7 @@ mod tests {
             } else {
                 "gpt-4o-upstream".to_string()
             },
+            supports_streaming: true,
             mapping_matched_model: None,
         }
     }

apps/aether-gateway/src/ai_serving/planner/standard/family/request.rs

@@ -21,8 +21,9 @@ use crate::ai_serving::planner::redaction::{
 };
 use crate::ai_serving::planner::spec_metadata::local_standard_spec_metadata;
 use crate::ai_serving::planner::standard::{
-    apply_codex_openai_responses_special_headers, apply_deepseek_tool_call_thinking_compat,
+    apply_codex_openai_special_headers, apply_deepseek_tool_call_thinking_compat,
-    is_deepseek_provider, request_body_build_failure_extra_data,
+    codex_model_capabilities_for_transport, is_deepseek_provider,
+    openai_provider_request_contract_failure_extra_data, request_body_build_failure_extra_data,
     request_conversion_failure_extra_data,
 };
 use crate::ai_serving::transport::kiro::{
@@ -44,7 +45,9 @@ use crate::ai_serving::transport::{
 };
 use crate::ai_serving::{
     build_openai_image_request_body_from_gemini_image_request, gemini_request_is_image_generation,
+    project_codex_openai_image_api_request_body, project_openai_image_api_request_body,
     CandidateFailureDiagnostic, GatewayProviderTransportSnapshot, LocalResolvedOAuthRequestAuth,
+    OpenAiImageOperation,
 };
 use crate::{AppState, GatewayError};
 
@@ -313,7 +316,13 @@ pub(crate) async fn resolve_local_standard_candidate_payload_parts(
     {
         return Ok(
             resolve_local_gemini_image_to_openai_image_candidate_payload_parts(
-                state, parts, trace_id, body_json, input, attempt,
+                state,
+                parts,
+                trace_id,
+                body_json,
+                input,
+                attempt,
+                spec_metadata.require_streaming,
             )
             .await,
         );
@@ -555,13 +564,35 @@ pub(crate) async fn resolve_local_standard_candidate_payload_parts(
     );
     let force_body_stream_field =
         endpoint_config_forces_body_stream_field(transport.endpoint.config.as_ref());
-    let enable_model_directives =
+    let model_directive_resolution = input
-        crate::system_features::reasoning_model_directive_enabled_for_api_format_and_model(
+        .model_directive_policy
-            state,
+        .resolve_reasoning(provider_api_format, Some(&input.requested_model));
-            provider_api_format,
+    let model_directive_mapping = match model_directive_resolution
-            Some(&input.requested_model),
+        .mapping_patch_for_mapped_model(&prepared_candidate.mapped_model)
-        )
+    {
-        .await;
+        Ok(mapping) => mapping,
+        Err(skip_reason) => {
+            mark_skipped_local_standard_candidate(
+                state,
+                input,
+                trace_id,
+                candidate,
+                attempt.candidate_index,
+                &attempt.candidate_id,
+                skip_reason,
+            )
+            .await;
+            return Ok(None);
+        }
+    };
+    crate::ai_serving::hydrate_openai_response_history(
+        state.runtime_state(),
+        body_json,
+        spec_metadata.api_format,
+        provider_api_format,
+        input.auth_context.api_key_id.as_str(),
+    )
+    .await?;
     let redaction = resolve_provider_chat_pii_redaction(
         state,
         parts,
@@ -588,7 +619,7 @@ pub(crate) async fn resolve_local_standard_candidate_payload_parts(
             },
             Some(input.auth_context.api_key_id.as_str()),
             Some(effective_headers),
-            enable_model_directives,
+            false,
         ) {
             Some(body) => body,
             None => {
@@ -655,18 +686,8 @@ pub(crate) async fn resolve_local_standard_candidate_payload_parts(
         provider_api_format,
         Some(body_json),
     );
-    if let Some(mapping) =
+    if let Some(mapping) = model_directive_mapping.as_ref() {
-        crate::system_features::reasoning_model_directive_mapping_for_api_format_and_model(
+        crate::ai_serving::apply_model_directive_mapping_patch(&mut provider_request_body, mapping);
-            state,
-            provider_api_format,
-            Some(&input.requested_model),
-        )
-        .await
-    {
-        crate::ai_serving::apply_model_directive_mapping_patch(
-            &mut provider_request_body,
-            &mapping,
-        );
         // Directive mapping is a deep-merge patch and may overwrite/add `stream`;
         // re-enforce stream-field policy afterward.
         enforce_provider_body_stream_policy(
@@ -712,6 +733,61 @@ pub(crate) async fn resolve_local_standard_candidate_payload_parts(
         );
     }
 
+    let normalized_provider_api_format =
+        crate::ai_serving::normalize_api_format_alias(provider_api_format);
+    if matches!(
+        normalized_provider_api_format.as_str(),
+        "openai:chat" | "openai:responses" | "openai:responses:compact"
+    ) {
+        let source_model = body_json
+            .get("model")
+            .and_then(Value::as_str)
+            .unwrap_or(input.requested_model.as_str());
+        let codex_model_capabilities = codex_model_capabilities_for_transport(
+            transport,
+            provider_api_format,
+            prepared_candidate.mapped_model.as_str(),
+            source_model,
+        );
+        if let Err(violation) =
+            crate::ai_serving::finalize_openai_provider_request_with_codex_model_capabilities(
+                &mut provider_request_body,
+                crate::ai_serving::OpenAiProviderRequestFinalization {
+                    source_api_format: spec_metadata.api_format,
+                    provider_api_format,
+                    provider_type: transport.provider.provider_type.as_str(),
+                    provider_model: prepared_candidate.mapped_model.as_str(),
+                    source_model,
+                    body_rules: transport.endpoint.body_rules.as_ref(),
+                    upstream_is_stream,
+                    require_body_stream_field: request_requires_body_stream_field(
+                        body_json,
+                        force_body_stream_field,
+                    ),
+                },
+                codex_model_capabilities.as_ref(),
+            )
+        {
+            mark_skipped_local_standard_candidate_with_extra_data(
+                state,
+                input,
+                trace_id,
+                candidate,
+                attempt.candidate_index,
+                &attempt.candidate_id,
+                "provider_request_body_build_failed",
+                Some(openai_provider_request_contract_failure_extra_data(
+                    &violation,
+                    spec_metadata.api_format,
+                    provider_api_format,
+                    "standard_family_request_finalization",
+                )),
+            )
+            .await;
+            return Ok(None);
+        }
+    }
+
     if let Some(kiro_auth) = kiro_auth.as_ref() {
         return Ok(build_kiro_cross_format_payload_parts(
             state,
@@ -752,8 +828,6 @@ pub(crate) async fn resolve_local_standard_candidate_payload_parts(
         .await);
     }
 
-    let normalized_provider_api_format =
-        crate::ai_serving::normalize_api_format_alias(provider_api_format);
     if normalized_provider_api_format == "gemini:generate_content"
         && is_gemini_cli_provider_transport(transport)
     {
@@ -838,7 +912,7 @@ pub(crate) async fn resolve_local_standard_candidate_payload_parts(
         return Ok(None);
     };
     let mut provider_request_headers = resolved_headers.headers;
-    apply_codex_openai_responses_special_headers(
+    apply_codex_openai_special_headers(
         &mut provider_request_headers,
         &provider_request_body,
         effective_headers,
@@ -988,7 +1062,7 @@ async fn build_gemini_cli_cross_format_payload_parts(
         };
 
     let mut provider_request_headers = resolved.headers.headers;
-    apply_codex_openai_responses_special_headers(
+    apply_codex_openai_special_headers(
         &mut provider_request_headers,
         &resolved.body,
         effective_headers,
@@ -1146,6 +1220,7 @@ async fn resolve_local_gemini_image_to_openai_image_candidate_payload_parts(
     body_json: &serde_json::Value,
     input: &LocalStandardDecisionInput,
     attempt: &LocalStandardCandidateAttempt,
+    client_requires_streaming: bool,
 ) -> Option<LocalStandardCandidatePayloadParts> {
     let client_api_format = "gemini:generate_content";
     let provider_api_format = "openai:image";
@@ -1221,18 +1296,60 @@ async fn resolve_local_gemini_image_to_openai_image_candidate_payload_parts(
         return None;
     };
 
-    let upstream_is_stream = true;
+    let upstream_is_stream = resolve_upstream_is_stream_for_provider(
-    let upstream_url =
+        transport.endpoint.config.as_ref(),
-        build_openai_image_upstream_url(transport, Some("/v1/images/generations"), None);
+        transport.provider.provider_type.as_str(),
+        provider_api_format,
+        client_requires_streaming && candidate.supports_streaming,
+        false,
+    );
+    let is_codex = transport
+        .provider
+        .provider_type
+        .trim()
+        .eq_ignore_ascii_case("codex");
+    let mut provider_request_body = converted.body_json;
+    if upstream_is_stream {
+        provider_request_body
+            .as_object_mut()?
+            .insert("stream".to_string(), Value::Bool(true));
+    }
+    provider_request_body = project_openai_image_api_request_body(
+        &provider_request_body,
+        &prepared_candidate.mapped_model,
+        converted.operation,
+        crate::image_capabilities::openai_image_provider_max_generation_count_for_model(
+            transport.provider.provider_type.as_str(),
+            Some(prepared_candidate.mapped_model.as_str()),
+        ),
+    )?;
+    if is_codex {
+        provider_request_body = project_codex_openai_image_api_request_body(
+            &provider_request_body,
+            converted.operation,
+        )?;
+    }
+    let request_path = match converted.operation {
+        OpenAiImageOperation::Generate => "/v1/images/generations",
+        OpenAiImageOperation::Edit => "/v1/images/edits",
+    };
+    let upstream_url = build_openai_image_upstream_url(transport, Some(request_path), None);
     let effective_headers = input.effective_headers(&parts.headers);
     let Some(mut provider_request_headers) =
         build_openai_image_headers(ProviderOpenAiImageHeadersInput {
+            transport,
             headers: effective_headers,
             auth_header: &prepared_candidate.auth_header,
             auth_value: &prepared_candidate.auth_value,
-            accept: "text/event-stream",
+            accept: if is_codex {
+                None
+            } else if upstream_is_stream {
+                Some("text/event-stream")
+            } else {
+                Some("application/json")
+            },
             header_rules: transport.endpoint.header_rules.as_ref(),
-            provider_request_body: &converted.body_json,
+            provider_request_body: &provider_request_body,
             original_request_body: body_json,
         })
     else {
@@ -1253,9 +1370,9 @@ async fn resolve_local_gemini_image_to_openai_image_candidate_payload_parts(
         .await;
         return None;
     };
-    apply_codex_openai_responses_special_headers(
+    apply_codex_openai_special_headers(
         &mut provider_request_headers,
-        &converted.body_json,
+        &provider_request_body,
         effective_headers,
         transport.provider.provider_type.as_str(),
         provider_api_format,
@@ -1268,7 +1385,7 @@ async fn resolve_local_gemini_image_to_openai_image_candidate_payload_parts(
         auth_value: prepared_candidate.auth_value,
         mapped_model: converted.mapped_model,
         provider_api_format: provider_api_format.to_string(),
-        provider_request_body: converted.body_json,
+        provider_request_body,
         provider_request_headers,
         upstream_url,
         upstream_is_stream,

apps/aether-gateway/src/ai_serving/planner/standard/gemini/plan_builders.rs

@@ -1,12 +1,10 @@
 use std::collections::BTreeMap;
 
-use aether_contracts::RequestBody;
-
 use super::{
     augment_sync_report_context, build_ai_execution_plan_from_decision,
-    generic_decision_missing_exact_provider_request, take_ai_decision_plan_core,
+    generic_decision_missing_exact_provider_request, resolve_ai_passthrough_sync_request_body,
-    take_ai_upstream_auth_pair, take_non_empty_string, AiExecutionPlanFromDecisionParts,
+    take_ai_decision_plan_core, take_ai_upstream_auth_pair, take_non_empty_string,
-    AiStreamAttempt, AiSyncAttempt,
+    AiExecutionPlanFromDecisionParts, AiStreamAttempt, AiSyncAttempt,
 };
 use crate::ai_serving::transport::{
     build_standard_plan_fallback_headers, StandardPlanFallbackAcceptPolicy,
@@ -61,6 +59,10 @@ pub(crate) fn build_gemini_sync_plan_from_decision(
         &provider_request_headers,
         &provider_request_body_value,
     )?;
+    let request_body = resolve_ai_passthrough_sync_request_body(
+        Some(provider_request_body_value),
+        payload.provider_request_body_base64.take(),
+    );
     let stream = payload.upstream_is_stream;
     let plan = build_ai_execution_plan_from_decision(
         &mut payload,
@@ -70,7 +72,7 @@ pub(crate) fn build_gemini_sync_plan_from_decision(
             url,
             headers: std::mem::take(&mut provider_request_headers),
             content_type,
-            body: RequestBody::from_json(provider_request_body_value),
+            body: request_body,
             stream,
         },
     );
@@ -129,6 +131,10 @@ pub(crate) fn build_gemini_stream_plan_from_decision(
         &provider_request_headers,
         &provider_request_body_value,
     )?;
+    let request_body = resolve_ai_passthrough_sync_request_body(
+        Some(provider_request_body_value),
+        payload.provider_request_body_base64.take(),
+    );
     let plan = build_ai_execution_plan_from_decision(
         &mut payload,
         AiExecutionPlanFromDecisionParts {
@@ -137,7 +143,7 @@ pub(crate) fn build_gemini_stream_plan_from_decision(
             url,
             headers: std::mem::take(&mut provider_request_headers),
             content_type,
-            body: RequestBody::from_json(provider_request_body_value),
+            body: request_body,
             stream: true,
         },
     );

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

@@ -15,7 +15,8 @@ mod normalize;
 mod openai;
 
 pub(crate) use self::codex::{
-    apply_codex_openai_responses_special_body_edits, apply_codex_openai_responses_special_headers,
+    apply_codex_openai_responses_special_body_edits, apply_codex_openai_special_headers,
+    codex_model_capabilities_for_transport,
 };
 pub(crate) use self::deepseek::{apply_deepseek_tool_call_thinking_compat, is_deepseek_provider};
 pub(crate) use self::family::{
@@ -25,9 +26,11 @@ pub(crate) use self::family::{
 pub(crate) use self::normalize::{
     build_cross_format_openai_chat_request_body, build_cross_format_openai_chat_upstream_url,
     build_cross_format_openai_responses_request_body,
+    build_cross_format_openai_responses_request_body_with_codex_model_capabilities,
     build_cross_format_openai_responses_upstream_url, build_local_openai_chat_request_body,
     build_local_openai_chat_upstream_url, build_local_openai_responses_request_body,
-    build_local_openai_responses_upstream_url,
+    build_local_openai_responses_request_body_with_codex_model_capabilities,
+    build_local_openai_responses_upstream_url, validate_final_openai_provider_request,
 };
 pub(crate) use self::openai::{
     build_local_openai_chat_stream_attempt_source_for_kind,
@@ -62,8 +65,8 @@ pub(crate) use crate::ai_serving::{
     normalize_openai_responses_request_to_openai_chat_request, parse_openai_tool_result_content,
 };
 pub(crate) use aether_ai_serving::{
-    request_body_build_failure_extra_data, request_conversion_failure_extra_data,
+    openai_provider_request_contract_failure_extra_data, request_body_build_failure_extra_data,
-    same_format_provider_request_body_failure_extra_data,
+    request_conversion_failure_extra_data, same_format_provider_request_body_failure_extra_data,
 };
 
 pub(crate) fn build_standard_upstream_url(
@@ -81,6 +84,7 @@ pub(crate) fn build_standard_upstream_url(
         upstream_is_stream,
         parts.uri.query(),
         None,
+        None,
         provider_request_body,
     )
 }
@@ -297,7 +301,7 @@ mod tests {
         let converted = build_standard_request_body(
             &request,
             "claude:messages",
-            "gpt-5",
+            "gpt-5.4",
             "codex",
             "openai:responses",
             "/v1/messages",
@@ -309,7 +313,7 @@ mod tests {
 
         assert!(converted.get("metadata").is_none());
         assert_eq!(converted["store"], false);
-        assert_eq!(converted["instructions"], "");
+        assert!(converted.get("instructions").is_none());
         assert_eq!(converted["include"], json!(["reasoning.encrypted_content"]));
         assert_eq!(converted["parallel_tool_calls"], true);
         assert_eq!(converted["reasoning"]["effort"], "medium");

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

@@ -12,9 +12,38 @@ pub(crate) use self::chat::{
 };
 pub(crate) use self::responses::{
     build_cross_format_openai_responses_request_body,
+    build_cross_format_openai_responses_request_body_with_codex_model_capabilities,
     build_cross_format_openai_responses_upstream_url, build_local_openai_responses_request_body,
+    build_local_openai_responses_request_body_with_codex_model_capabilities,
     build_local_openai_responses_upstream_url,
 };
 pub(super) use crate::ai_serving::planner::common::{
     enforce_provider_body_stream_policy, request_requires_body_stream_field,
 };
+
+pub(crate) fn validate_final_openai_provider_request(
+    provider_api_format: &str,
+    mapped_model: &str,
+    source_request_body: &serde_json::Value,
+    provider_request_body: &serde_json::Value,
+) -> Option<()> {
+    let provider_model = provider_request_body
+        .get("model")
+        .and_then(serde_json::Value::as_str)
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+        .unwrap_or(mapped_model);
+    let source_model = source_request_body
+        .get("model")
+        .and_then(serde_json::Value::as_str)
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+        .unwrap_or(mapped_model);
+    crate::ai_serving::validate_openai_provider_request_contract(
+        provider_api_format,
+        provider_model,
+        source_model,
+        provider_request_body,
+    )
+    .ok()
+}

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

@@ -9,7 +9,10 @@ use crate::ai_serving::{
     GatewayProviderTransportSnapshot,
 };
 
-use super::{enforce_provider_body_stream_policy, request_requires_body_stream_field};
+use super::{
+    enforce_provider_body_stream_policy, request_requires_body_stream_field,
+    validate_final_openai_provider_request,
+};
 
 pub(crate) fn build_local_openai_chat_request_body(
     body_json: &Value,
@@ -39,6 +42,12 @@ pub(crate) fn build_local_openai_chat_request_body(
         upstream_is_stream,
         request_requires_body_stream_field(body_json, force_body_stream_field),
     );
+    validate_final_openai_provider_request(
+        "openai:chat",
+        mapped_model,
+        body_json,
+        &provider_request_body,
+    )?;
     Some(provider_request_body)
 }
 
@@ -92,6 +101,12 @@ pub(crate) fn build_cross_format_openai_chat_request_body(
         upstream_is_stream,
         request_requires_body_stream_field(body_json, force_body_stream_field),
     );
+    validate_final_openai_provider_request(
+        provider_api_format,
+        mapped_model,
+        body_json,
+        &provider_request_body,
+    )?;
     Some(provider_request_body)
 }
 

apps/aether-gateway/src/ai_serving/planner/standard/normalize/responses.rs

@@ -2,14 +2,16 @@ use serde_json::Value;
 
 use crate::ai_serving::transport::apply_standard_provider_request_body_rules_with_request_headers;
 use crate::ai_serving::{
-    apply_codex_openai_responses_special_body_edits,
     apply_openai_responses_compact_special_body_edits,
-    build_cross_format_openai_responses_request_body_with_model_directives as surface_build_cross_format_openai_responses_request_body,
+    build_cross_format_openai_responses_request_body_with_model_directives_and_history_scope as surface_build_cross_format_openai_responses_request_body,
     build_local_openai_responses_request_body_with_model_directives as surface_build_local_openai_responses_request_body,
     GatewayProviderTransportSnapshot,
 };
 
-use super::{enforce_provider_body_stream_policy, request_requires_body_stream_field};
+use super::{
+    enforce_provider_body_stream_policy, request_requires_body_stream_field,
+    validate_final_openai_provider_request,
+};
 
 pub(crate) fn build_local_openai_responses_request_body(
     body_json: &Value,
@@ -19,9 +21,35 @@ pub(crate) fn build_local_openai_responses_request_body(
     provider_type: &str,
     provider_api_format: &str,
     body_rules: Option<&Value>,
-    user_api_key_id: Option<&str>,
+    _user_api_key_id: Option<&str>,
     request_headers: &http::HeaderMap,
     enable_model_directives: bool,
+) -> Option<Value> {
+    build_local_openai_responses_request_body_with_codex_model_capabilities(
+        body_json,
+        mapped_model,
+        require_streaming,
+        force_body_stream_field,
+        provider_type,
+        provider_api_format,
+        body_rules,
+        request_headers,
+        None,
+        enable_model_directives,
+    )
+}
+
+pub(crate) fn build_local_openai_responses_request_body_with_codex_model_capabilities(
+    body_json: &Value,
+    mapped_model: &str,
+    require_streaming: bool,
+    force_body_stream_field: bool,
+    provider_type: &str,
+    provider_api_format: &str,
+    body_rules: Option<&Value>,
+    request_headers: &http::HeaderMap,
+    model_capabilities: Option<&crate::ai_serving::CodexResponsesModelCapabilities>,
+    enable_model_directives: bool,
 ) -> Option<Value> {
     let provider_request_body = surface_build_local_openai_responses_request_body(
         body_json,
@@ -36,23 +64,39 @@ pub(crate) fn build_local_openai_responses_request_body(
             body_json,
             request_headers,
         )?;
-    apply_codex_openai_responses_special_body_edits(
+    let source_model = body_json
+        .get("model")
+        .and_then(Value::as_str)
+        .unwrap_or(mapped_model);
+    crate::ai_serving::apply_codex_openai_responses_special_body_edits_with_source_model_and_capabilities(
         &mut provider_request_body,
         provider_type,
         provider_api_format,
+        mapped_model,
+        source_model,
+        model_capabilities,
         body_rules,
-        user_api_key_id,
     );
     apply_openai_responses_compact_special_body_edits(
         &mut provider_request_body,
         provider_api_format,
     );
+    crate::ai_serving::strip_incompatible_openai_responses_reasoning_items(
+        &mut provider_request_body,
+        provider_api_format,
+    );
     enforce_provider_body_stream_policy(
         &mut provider_request_body,
         provider_api_format,
         require_streaming,
         request_requires_body_stream_field(body_json, force_body_stream_field),
     );
+    validate_final_openai_provider_request(
+        provider_api_format,
+        mapped_model,
+        body_json,
+        &provider_request_body,
+    )?;
     Some(provider_request_body)
 }
 
@@ -68,6 +112,36 @@ pub(crate) fn build_cross_format_openai_responses_request_body(
     user_api_key_id: Option<&str>,
     request_headers: &http::HeaderMap,
     enable_model_directives: bool,
+) -> Option<Value> {
+    build_cross_format_openai_responses_request_body_with_codex_model_capabilities(
+        body_json,
+        mapped_model,
+        client_api_format,
+        provider_api_format,
+        upstream_is_stream,
+        force_body_stream_field,
+        provider_type,
+        body_rules,
+        request_headers,
+        user_api_key_id,
+        None,
+        enable_model_directives,
+    )
+}
+
+pub(crate) fn build_cross_format_openai_responses_request_body_with_codex_model_capabilities(
+    body_json: &Value,
+    mapped_model: &str,
+    client_api_format: &str,
+    provider_api_format: &str,
+    upstream_is_stream: bool,
+    force_body_stream_field: bool,
+    provider_type: &str,
+    body_rules: Option<&Value>,
+    request_headers: &http::HeaderMap,
+    history_scope: Option<&str>,
+    model_capabilities: Option<&crate::ai_serving::CodexResponsesModelCapabilities>,
+    enable_model_directives: bool,
 ) -> Option<Value> {
     let provider_request_body = surface_build_cross_format_openai_responses_request_body(
         body_json,
@@ -76,6 +150,7 @@ pub(crate) fn build_cross_format_openai_responses_request_body(
         provider_api_format,
         upstream_is_stream,
         enable_model_directives,
+        history_scope,
     )?;
     let mut provider_request_body =
         apply_standard_provider_request_body_rules_with_request_headers(
@@ -84,23 +159,39 @@ pub(crate) fn build_cross_format_openai_responses_request_body(
             body_json,
             request_headers,
         )?;
-    apply_codex_openai_responses_special_body_edits(
+    let source_model = body_json
+        .get("model")
+        .and_then(Value::as_str)
+        .unwrap_or(mapped_model);
+    crate::ai_serving::apply_codex_openai_responses_special_body_edits_with_source_model_and_capabilities(
         &mut provider_request_body,
         provider_type,
         provider_api_format,
+        mapped_model,
+        source_model,
+        model_capabilities,
         body_rules,
-        user_api_key_id,
     );
     apply_openai_responses_compact_special_body_edits(
         &mut provider_request_body,
         provider_api_format,
     );
+    crate::ai_serving::strip_incompatible_openai_responses_reasoning_items(
+        &mut provider_request_body,
+        provider_api_format,
+    );
     enforce_provider_body_stream_policy(
         &mut provider_request_body,
         provider_api_format,
         upstream_is_stream,
         request_requires_body_stream_field(body_json, force_body_stream_field),
     );
+    validate_final_openai_provider_request(
+        provider_api_format,
+        mapped_model,
+        body_json,
+        &provider_request_body,
+    )?;
     Some(provider_request_body)
 }
 

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

@@ -6,8 +6,8 @@ use http::Request;
 use serde_json::{json, Value};
 
 use super::{
-    build_cross_format_openai_responses_request_body, build_local_openai_responses_request_body,
+    build_cross_format_openai_responses_request_body, build_local_openai_chat_request_body,
-    build_local_openai_responses_upstream_url,
+    build_local_openai_responses_request_body, build_local_openai_responses_upstream_url,
 };
 
 fn object_keys(value: &Value) -> Vec<&str> {
@@ -146,12 +146,47 @@ fn local_openai_responses_wrapper_preserves_body_order_after_edits() {
             "reasoning",
             "tool_choice",
             "parallel_tool_calls",
-            "instructions",
-            "prompt_cache_key",
         ]
     );
     assert_eq!(provider_request_body["parallel_tool_calls"], json!(true));
-    assert_eq!(provider_request_body["instructions"], json!(""));
+    assert!(provider_request_body.get("instructions").is_none());
+}
+
+#[test]
+fn local_openai_responses_wrapper_strips_foreign_reasoning_item_ids() {
+    let body_json = json!({
+        "model": "gpt-5.4",
+        "input": [
+            {"type": "reasoning", "id": "rs_provider_123", "summary": []},
+            {
+                "type": "reasoning",
+                "id": "item_72d3bd8d367d01977ace23f1",
+                "summary": []
+            },
+            {"type": "message", "role": "user", "content": "continue"}
+        ]
+    });
+
+    let provider_request_body = build_local_openai_responses_request_body(
+        &body_json,
+        "gpt-5.4",
+        false,
+        false,
+        "codex",
+        "openai:responses",
+        None,
+        None,
+        &http::HeaderMap::new(),
+        false,
+    )
+    .expect("local OpenAI Responses body should build");
+
+    let input = provider_request_body["input"]
+        .as_array()
+        .expect("input array");
+    assert_eq!(input.len(), 2);
+    assert_eq!(input[0]["id"], "rs_provider_123");
+    assert_eq!(input[1]["type"], "message");
 }
 
 #[test]
@@ -181,18 +216,49 @@ fn local_openai_responses_compact_wrapper_strips_store_for_same_format_requests(
 }
 
 #[test]
-fn local_openai_responses_compact_wrapper_strips_include_for_codex_requests() {
+fn local_codex_compact_wrapper_applies_the_complete_request_projection() {
     let body_json = json!({
-        "model": "gpt-5.4",
+        "model": "gpt-5.6-sol",
-        "input": [],
+        "input": [{
+            "type": "message",
+            "role": "user",
+            "content": [{"type": "input_text", "text": "hello"}]
+        }],
+        "instructions": "Work carefully",
+        "client_metadata": {"origin": "codex"},
         "include": ["reasoning.encrypted_content"],
         "store": true,
-        "stream": true
+        "stream": true,
+        "stream_options": {"reasoning_summary_delivery": "sequential_cutoff"},
+        "tool_choice": "auto",
+        "parallel_tool_calls": true,
+        "reasoning": {"effort": "max", "summary": "auto", "context": "all_turns"},
+        "text": {"verbosity": "medium"},
+        "tools": [{
+            "type": "function",
+            "name": "lookup",
+            "parameters": {"type": "object", "properties": {}}
+        }],
+        "service_tier": "priority",
+        "prompt_cache_key": "thread-compact"
     });
 
-    let provider_request_body = build_local_openai_responses_request_body(
+    let regular = build_local_openai_responses_request_body(
         &body_json,
-        "gpt-5.4",
+        "gpt-5.6-sol",
+        true,
+        false,
+        "codex",
+        "openai:responses",
+        None,
+        Some("key-123"),
+        &http::HeaderMap::new(),
+        false,
+    )
+    .expect("local Codex Responses body should build");
+    let compact = build_local_openai_responses_request_body(
+        &body_json,
+        "gpt-5.6-sol",
         false,
         false,
         "codex",
@@ -202,22 +268,44 @@ fn local_openai_responses_compact_wrapper_strips_include_for_codex_requests() {
         &http::HeaderMap::new(),
         false,
     )
-    .expect("local codex compact body should build");
+    .expect("local Codex Compact body should build");
 
-    assert!(provider_request_body.get("include").is_none());
+    for field in [
-    assert!(provider_request_body.get("store").is_none());
+        "client_metadata",
-    assert!(provider_request_body.get("stream").is_none());
+        "include",
-    assert_eq!(provider_request_body["instructions"], "");
+        "store",
-    assert_eq!(
+        "stream",
-        provider_request_body["prompt_cache_key"],
+        "stream_options",
-        "3d2e2842-74cb-55dd-803a-b8940b3500c2"
+        "tool_choice",
-    );
+    ] {
+        assert!(
+            regular.get(field).is_some(),
+            "Responses should contain {field}"
+        );
+        assert!(compact.get(field).is_none(), "Compact should omit {field}");
+    }
+    for field in [
+        "model",
+        "input",
+        "instructions",
+        "parallel_tool_calls",
+        "reasoning",
+        "text",
+        "tools",
+        "service_tier",
+        "prompt_cache_key",
+    ] {
+        assert_eq!(
+            compact[field], regular[field],
+            "Compact should preserve {field}"
+        );
+    }
 }
 
 #[test]
 fn local_openai_responses_wrapper_applies_model_directive_before_body_rules() {
     let body_json = json!({
-        "model": "gpt-5.4-max",
+        "model": "gpt-5.6-sol-max",
         "input": "hello",
         "reasoning": {"effort": "low", "summary": "auto"}
     });
@@ -227,7 +315,7 @@ fn local_openai_responses_wrapper_applies_model_directive_before_body_rules() {
 
     let provider_request_body = build_local_openai_responses_request_body(
         &body_json,
-        "gpt-5.4",
+        "gpt-5.6-sol",
         false,
         false,
         "openai",
@@ -239,11 +327,137 @@ fn local_openai_responses_wrapper_applies_model_directive_before_body_rules() {
     )
     .expect("local openai responses body should build");
 
-    assert_eq!(provider_request_body["reasoning"]["effort"], "xhigh");
+    assert_eq!(provider_request_body["reasoning"]["effort"], "max");
     assert_eq!(provider_request_body["reasoning"]["summary"], "auto");
     assert_eq!(provider_request_body["metadata"]["override_seen"], true);
 }
 
+#[test]
+fn final_openai_provider_contract_uses_the_mapped_model_for_reasoning() {
+    let alias = json!({
+        "model": "deployment-alias",
+        "input": "hello",
+        "reasoning": {"effort": "max"}
+    });
+    assert!(build_local_openai_responses_request_body(
+        &alias,
+        "gpt-5.6-sol",
+        false,
+        false,
+        "openai",
+        "openai:responses",
+        None,
+        None,
+        &http::HeaderMap::new(),
+        false,
+    )
+    .is_some());
+    assert!(build_local_openai_responses_request_body(
+        &alias,
+        "gpt-5.4",
+        false,
+        false,
+        "openai",
+        "openai:responses",
+        None,
+        None,
+        &http::HeaderMap::new(),
+        false,
+    )
+    .is_none());
+
+    let minimal = json!({
+        "model": "deployment-alias",
+        "messages": [{"role": "user", "content": "hello"}],
+        "reasoning_effort": "minimal"
+    });
+    assert!(build_local_openai_chat_request_body(
+        &minimal,
+        "gpt-5.6-terra",
+        false,
+        false,
+        None,
+        &http::HeaderMap::new(),
+        false,
+    )
+    .is_none());
+
+    let opaque_mapping = json!({
+        "model": "gpt-5.6-sol-max",
+        "input": "hello",
+        "reasoning": {"effort": "max", "mode": "pro"},
+        "prompt_cache_options": {"mode": "explicit", "ttl": "30m"}
+    });
+    assert!(build_local_openai_responses_request_body(
+        &opaque_mapping,
+        "azure-production",
+        false,
+        false,
+        "openai",
+        "openai:responses",
+        None,
+        None,
+        &http::HeaderMap::new(),
+        false,
+    )
+    .is_some());
+    assert!(build_local_openai_responses_request_body(
+        &opaque_mapping,
+        "gpt-5.4",
+        false,
+        false,
+        "openai",
+        "openai:responses",
+        None,
+        None,
+        &http::HeaderMap::new(),
+        false,
+    )
+    .is_none());
+}
+
+#[test]
+fn final_openai_provider_contract_validates_body_rule_output() {
+    let body = json!({
+        "model": "gpt-5.6-sol",
+        "input": "hello",
+        "reasoning": {"effort": "max"}
+    });
+    let model_override = json!([
+        {"action":"set","path":"model","value":"gpt-5.4"}
+    ]);
+    assert!(build_local_openai_responses_request_body(
+        &body,
+        "gpt-5.6-sol",
+        false,
+        false,
+        "openai",
+        "openai:responses",
+        Some(&model_override),
+        None,
+        &http::HeaderMap::new(),
+        false,
+    )
+    .is_none());
+
+    let cache_override = json!([
+        {"action":"set","path":"prompt_cache_options.ttl","value":"1h"}
+    ]);
+    assert!(build_local_openai_responses_request_body(
+        &json!({"model":"gpt-5.6-sol","input":"hello"}),
+        "gpt-5.6-sol",
+        false,
+        false,
+        "openai",
+        "openai:responses",
+        Some(&cache_override),
+        None,
+        &http::HeaderMap::new(),
+        false,
+    )
+    .is_none());
+}
+
 #[test]
 fn local_openai_responses_upstream_url_preserves_codex_base_path() {
     let request = Request::builder()
@@ -371,7 +585,7 @@ fn applies_codex_defaults_unless_body_rules_handle_the_field() {
 }
 
 #[test]
-fn injects_codex_prompt_cache_key_for_openai_responses_cross_format_requests() {
+fn omits_codex_prompt_cache_key_for_openai_responses_cross_format_requests() {
     let body_json = json!({
         "model": "claude-sonnet-4-5",
         "messages": [{
@@ -395,14 +609,11 @@ fn injects_codex_prompt_cache_key_for_openai_responses_cross_format_requests() {
     )
     .expect("claude cli to codex request should build");
 
-    assert_eq!(
+    assert!(provider_request_body.get("prompt_cache_key").is_none());
-        provider_request_body["prompt_cache_key"],
-        "4ee6ea6e-3ac6-5a18-8cb8-1f8b956419e5"
-    );
 }
 
 #[test]
-fn injects_codex_prompt_cache_key_for_openai_chat_cross_format_requests() {
+fn omits_codex_prompt_cache_key_for_openai_chat_cross_format_requests() {
     let body_json = json!({
         "model": "gpt-5",
         "messages": [{
@@ -425,8 +636,5 @@ fn injects_codex_prompt_cache_key_for_openai_chat_cross_format_requests() {
     )
     .expect("openai chat to codex request should build");
 
-    assert_eq!(
+    assert!(provider_request_body.get("prompt_cache_key").is_none());
-        provider_request_body["prompt_cache_key"],
-        "4ee6ea6e-3ac6-5a18-8cb8-1f8b956419e5"
-    );
 }

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

@@ -6,6 +6,7 @@ mod request;
 mod support;
 
 pub(super) use self::payload::maybe_build_local_openai_chat_decision_payload_for_candidate;
+pub(super) use self::request::LocalOpenAiChatRequestPreparation;
 pub(super) use self::support::{
     build_lazy_local_openai_chat_candidate_attempt_source,
     build_local_openai_chat_candidate_attempt_source,

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

@@ -6,18 +6,21 @@ use crate::ai_serving::planner::report_context::{
     insert_provider_stream_event_api_format, LocalExecutionReportContextParts,
 };
 use crate::ai_serving::planner::{
-    build_ai_execution_decision_response, resolve_transport_request_gzip_policy,
+    build_ai_execution_decision_response, resolve_transport_request_encoding_policy,
     AiExecutionDecisionResponseParts,
 };
 use crate::ai_serving::transport::{
     resolve_transport_execution_timeouts, resolve_transport_profile,
 };
+use crate::stage_metrics::observe_gateway_stage_ms;
 use crate::{
     append_execution_contract_fields_to_value, append_local_failover_policy_to_value,
     AiExecutionDecision, AppState, GatewayError,
 };
 
-use super::request::resolve_local_openai_chat_candidate_payload_parts;
+use super::request::{
+    resolve_local_openai_chat_candidate_payload_parts, LocalOpenAiChatRequestPreparation,
+};
 use super::support::{LocalOpenAiChatCandidateAttempt, LocalOpenAiChatDecisionInput};
 
 #[allow(clippy::too_many_arguments)]
@@ -27,6 +30,7 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
     trace_id: &str,
     body_json: &serde_json::Value,
     input: &LocalOpenAiChatDecisionInput,
+    preparation: Option<&mut LocalOpenAiChatRequestPreparation>,
     attempt: LocalOpenAiChatCandidateAttempt,
     decision_kind: &str,
     report_kind: &str,
@@ -40,12 +44,15 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
         candidate_id,
         ..
     } = attempt;
+    let upstream_is_stream = upstream_is_stream && eligible.candidate.supports_streaming;
+    let payload_started_at = std::time::Instant::now();
     let Some(resolved) = resolve_local_openai_chat_candidate_payload_parts(
         state,
         parts,
         trace_id,
         body_json,
         input,
+        preparation,
         &eligible,
         candidate_index,
         &candidate_id,
@@ -55,9 +62,25 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
     )
     .await?
     else {
+        observe_gateway_stage_ms(
+            "stream_candidate_payload_parts",
+            payload_started_at.elapsed().as_millis() as u64,
+        );
         return Ok(None);
     };
+    observe_gateway_stage_ms(
+        "stream_candidate_payload_parts",
+        payload_started_at.elapsed().as_millis() as u64,
+    );
     let candidate = &eligible.candidate;
+    let upstream_is_stream =
+        crate::ai_serving::planner::common::resolve_upstream_is_stream_for_provider(
+            resolved.transport.endpoint.config.as_ref(),
+            resolved.transport.provider.provider_type.as_str(),
+            resolved.provider_api_format.as_str(),
+            upstream_is_stream,
+            false,
+        );
 
     let prompt_cache_key = resolved
         .provider_request_body
@@ -66,9 +89,14 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
         .map(str::trim)
         .filter(|value| !value.is_empty())
         .map(ToOwned::to_owned);
+    let proxy_started_at = std::time::Instant::now();
     let proxy = state
         .resolve_transport_proxy_snapshot_with_tunnel_affinity(&resolved.transport)
         .await;
+    observe_gateway_stage_ms(
+        "stream_candidate_proxy",
+        proxy_started_at.elapsed().as_millis() as u64,
+    );
     let transport_profile = resolved
         .transport_profile
         .clone()
@@ -130,6 +158,7 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
         Some(body_json)
     };
     let effective_headers = input.effective_headers(&parts.headers);
+    let report_context_started_at = std::time::Instant::now();
     let report_context = append_local_failover_policy_to_value(
         append_execution_contract_fields_to_value(
             build_local_execution_report_context(LocalExecutionReportContextParts {
@@ -164,6 +193,7 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
                 original_request_body_json,
                 original_request_body_base64: None,
                 client_session_affinity: input.client_session_affinity.as_ref(),
+                routing_policy: input.routing_policy.as_ref(),
                 scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
                 client_requested_stream: body_json
                     .get("stream")
@@ -184,8 +214,13 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
         ),
         &transport,
     );
-    let request_gzip = resolve_transport_request_gzip_policy(&transport);
+    observe_gateway_stage_ms(
+        "stream_candidate_report_context",
+        report_context_started_at.elapsed().as_millis() as u64,
+    );
+    let request_encoding = resolve_transport_request_encoding_policy(&transport);
 
+    let decision_started_at = std::time::Instant::now();
     let mut decision = build_ai_execution_decision_response(AiExecutionDecisionResponseParts {
         decision_is_stream,
         decision_kind: decision_kind.to_string(),
@@ -194,6 +229,7 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
         request_id: trace_id.to_string(),
         candidate_id: candidate_id.clone(),
         provider_name: transport.provider.name.clone(),
+        provider_type: transport.provider.provider_type.clone(),
         provider_id: candidate.provider_id.clone(),
         endpoint_id: candidate.endpoint_id.clone(),
         key_id: candidate.key_id.clone(),
@@ -211,8 +247,8 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
         provider_request_body: Some(provider_request_body),
         provider_request_body_base64: None,
         content_type: Some("application/json".to_string()),
-        content_encoding: None,
+        content_encoding: request_encoding.content_encoding,
-        request_gzip,
+        request_gzip: request_encoding.request_gzip,
         proxy,
         transport_profile,
         timeouts,
@@ -221,6 +257,14 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
         report_context: Some(report_context),
         auth_context: input.auth_context.clone(),
     });
-    apply_provider_request_routing_policy_to_decision(input, &mut decision)?;
+    apply_provider_request_routing_policy_to_decision(
+        input,
+        &mut decision,
+        Some(transport.as_ref()),
+    )?;
+    observe_gateway_stage_ms(
+        "stream_candidate_decision_build",
+        decision_started_at.elapsed().as_millis() as u64,
+    );
     Ok(Some(decision))
 }

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

@@ -293,9 +293,11 @@ pub(crate) async fn build_lazy_local_openai_chat_candidate_attempt_source<'a>(
     );
     build_lazy_requested_model_execution_candidate_attempt_source_with_serving(
         planner_state,
+        &input.model_directive_policy,
         trace_id,
         "openai:chat",
         &input.requested_model,
+        None,
         require_streaming,
         &input.auth_snapshot,
         input.client_session_affinity.as_ref(),

apps/aether-gateway/src/ai_serving/planner/standard/openai/chat/mod.rs

@@ -13,6 +13,7 @@ use self::decision::{
     build_lazy_local_openai_chat_candidate_attempt_source,
     maybe_build_local_openai_chat_decision_payload_for_candidate, LocalOpenAiChatCandidateAttempt,
     LocalOpenAiChatCandidateAttemptSource, LocalOpenAiChatDecisionInput,
+    LocalOpenAiChatRequestPreparation,
 };
 use self::plans::{
     build_local_openai_chat_stream_attempt_source, build_local_openai_chat_stream_plan_and_reports,
@@ -147,7 +148,7 @@ pub(crate) async fn maybe_build_sync_local_decision_payload(
     )
     .await;
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let upstream_is_stream = self::plans::openai_chat_upstream_is_stream_for_candidate(
             &attempt.eligible.transport,
             attempt.eligible.provider_api_format.as_str(),
@@ -159,6 +160,7 @@ pub(crate) async fn maybe_build_sync_local_decision_payload(
             trace_id,
             body_json,
             &input,
+            None,
             attempt,
             OPENAI_CHAT_SYNC_PLAN_KIND,
             "openai_chat_sync_success",
@@ -199,7 +201,7 @@ pub(crate) async fn maybe_build_stream_local_decision_payload(
     )
     .await;
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let upstream_is_stream = self::plans::openai_chat_upstream_is_stream_for_candidate(
             &attempt.eligible.transport,
             attempt.eligible.provider_api_format.as_str(),
@@ -211,6 +213,7 @@ pub(crate) async fn maybe_build_stream_local_decision_payload(
             trace_id,
             body_json,
             &input,
+            None,
             attempt,
             OPENAI_CHAT_STREAM_PLAN_KIND,
             "openai_chat_stream_success",

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

@@ -31,11 +31,12 @@ pub(super) fn openai_chat_upstream_is_stream_for_candidate(
         crate::ai_serving::transport::kiro::is_kiro_claude_messages_transport(
             transport,
             provider_api_format,
-        ) || openai_chat_gemini_cli_client_stream_requires_upstream_streaming(
+        ) || openai_chat_antigravity_requires_upstream_streaming(transport, provider_api_format)
-            transport,
+            || openai_chat_gemini_cli_client_stream_requires_upstream_streaming(
-            provider_api_format,
+                transport,
-            client_is_stream,
+                provider_api_format,
-        );
+                client_is_stream,
+            );
     resolve_upstream_is_stream_for_provider(
         transport.endpoint.config.as_ref(),
         transport.provider.provider_type.as_str(),
@@ -45,6 +46,15 @@ pub(super) fn openai_chat_upstream_is_stream_for_candidate(
     )
 }
 
+fn openai_chat_antigravity_requires_upstream_streaming(
+    transport: &GatewayProviderTransportSnapshot,
+    provider_api_format: &str,
+) -> bool {
+    crate::ai_serving::transport::antigravity::is_antigravity_provider_transport(transport)
+        && crate::ai_serving::normalize_api_format_alias(provider_api_format)
+            == "gemini:generate_content"
+}
+
 fn openai_chat_gemini_cli_client_stream_requires_upstream_streaming(
     transport: &GatewayProviderTransportSnapshot,
     provider_api_format: &str,
@@ -201,4 +211,24 @@ mod tests {
             false,
         ));
     }
+
+    #[test]
+    fn openai_chat_policy_resolver_preserves_antigravity_streaming_envelope() {
+        let antigravity = sample_transport(
+            "antigravity",
+            "gemini:generate_content",
+            Some(json!({"upstream_stream_policy": "force_non_stream"})),
+        );
+
+        assert!(openai_chat_upstream_is_stream_for_candidate(
+            &antigravity,
+            "gemini:generate_content",
+            false,
+        ));
+        assert!(openai_chat_upstream_is_stream_for_candidate(
+            &antigravity,
+            "gemini:generate_content",
+            true,
+        ));
+    }
 }

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

@@ -21,8 +21,10 @@ pub(crate) async fn list_local_openai_chat_candidates(
 > {
     let outcome = preselect_local_execution_candidates_with_serving(
         PlannerAppState::new(state),
+        &input.model_directive_policy,
         "openai:chat",
         &input.requested_model,
+        None,
         require_streaming,
         input.required_capabilities.as_ref(),
         &input.auth_snapshot,

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

@@ -9,6 +9,7 @@ use crate::ai_serving::planner::decision_input::{
 };
 use crate::ai_serving::resolve_local_decision_execution_runtime_auth_context;
 use crate::client_session_affinity::client_session_affinity_from_parts;
+use crate::stage_metrics::observe_gateway_stage_ms;
 use crate::{AppState, GatewayError};
 
 pub(crate) async fn resolve_local_openai_chat_decision_input(
@@ -59,11 +60,14 @@ pub(crate) async fn resolve_local_openai_chat_decision_input(
         return Ok(None);
     };
 
+    let auth_started_at = std::time::Instant::now();
     let resolved_input = match resolve_local_authenticated_decision_input(
         state,
         auth_context.clone(),
         Some(requested_model.as_str()),
+        decision.auth_endpoint_signature.as_deref(),
         None,
+        &decision.model_directive_policy,
     )
     .await
     {
@@ -106,10 +110,20 @@ pub(crate) async fn resolve_local_openai_chat_decision_input(
             return Err(err);
         }
     };
+    observe_gateway_stage_ms(
+        "openai_chat_decision_input_auth",
+        auth_started_at.elapsed().as_millis() as u64,
+    );
 
     let mut input = build_local_requested_model_decision_input(resolved_input, requested_model);
     input.request_auth_channel = decision.request_auth_channel.clone();
+    let affinity_started_at = std::time::Instant::now();
     input.client_session_affinity = client_session_affinity_from_parts(parts, Some(body_json));
+    observe_gateway_stage_ms(
+        "openai_chat_decision_input_affinity",
+        affinity_started_at.elapsed().as_millis() as u64,
+    );
+    let routing_started_at = std::time::Instant::now();
     if let Err(err) = attach_routing_policy_to_local_requested_model_input(
         state,
         parts,
@@ -126,5 +140,9 @@ pub(crate) async fn resolve_local_openai_chat_decision_input(
         );
         return Err(err);
     }
+    observe_gateway_stage_ms(
+        "openai_chat_decision_input_routing",
+        routing_started_at.elapsed().as_millis() as u64,
+    );
     Ok(Some(input))
 }

apps/aether-gateway/src/ai_serving/planner/standard/openai/chat/plans/stream.rs

@@ -1,11 +1,12 @@
 use async_trait::async_trait;
+use std::collections::VecDeque;
 use tracing::warn;
 
 use super::super::{
     build_lazy_local_openai_chat_candidate_attempt_source,
     maybe_build_local_openai_chat_decision_payload_for_candidate, AppState, GatewayControlDecision,
     GatewayError, LocalOpenAiChatCandidateAttempt, LocalOpenAiChatCandidateAttemptSource,
-    LocalOpenAiChatDecisionInput,
+    LocalOpenAiChatDecisionInput, LocalOpenAiChatRequestPreparation,
 };
 use super::diagnostic::{
     set_local_openai_chat_candidate_evaluation_diagnostic, set_local_openai_chat_miss_diagnostic,
@@ -18,6 +19,23 @@ use crate::ai_serving::planner::plan_builders::{
     build_openai_chat_stream_plan_from_decision, AiStreamAttempt,
 };
 use crate::ai_serving::planner::runtime_miss::apply_local_runtime_candidate_terminal_reason;
+use crate::ai_serving::planner::standard::build_local_openai_chat_upstream_url;
+use crate::ai_serving::transport::{
+    is_windsurf_provider_transport, local_openai_chat_transport_unsupported_reason,
+};
+use crate::clock::request_distribution_seed;
+use crate::stage_metrics::{
+    observe_gateway_stage_ms, record_openai_chat_stream_payload_build_prefetch_avoided,
+    record_openai_chat_stream_payload_build_selected,
+    record_openai_chat_stream_raw_candidates_scanned,
+    record_openai_chat_stream_target_select_selected_rank,
+};
+use crate::upstream_admission::upstream_target_key_from_url;
+
+const OPENAI_CHAT_STREAM_TARGET_SELECT_WINDOW_ENV: &str =
+    "AETHER_GATEWAY_OPENAI_CHAT_STREAM_TARGET_SELECT_WINDOW";
+const DEFAULT_OPENAI_CHAT_STREAM_TARGET_SELECT_WINDOW: usize = 2;
+const MAX_OPENAI_CHAT_STREAM_TARGET_SELECT_WINDOW: usize = 8;
 
 pub(crate) struct LocalOpenAiChatStreamAttemptSource<'a> {
     state: &'a AppState,
@@ -26,6 +44,8 @@ pub(crate) struct LocalOpenAiChatStreamAttemptSource<'a> {
     body_json: serde_json::Value,
     input: LocalOpenAiChatDecisionInput,
     candidates: LocalOpenAiChatCandidateAttemptSource<'a>,
+    prefetched_attempts: VecDeque<LocalOpenAiChatCandidateAttempt>,
+    request_preparation: LocalOpenAiChatRequestPreparation,
 }
 
 pub(crate) async fn build_local_openai_chat_stream_attempt_source<'a>(
@@ -40,6 +60,7 @@ pub(crate) async fn build_local_openai_chat_stream_attempt_source<'a>(
         return Ok(None);
     }
 
+    let attempt_source_started_at = std::time::Instant::now();
     let Some(input) = resolve_local_openai_chat_decision_input(
         state, parts, trace_id, decision, body_json, plan_kind, true,
     )
@@ -76,6 +97,10 @@ pub(crate) async fn build_local_openai_chat_stream_attempt_source<'a>(
         Some(input.requested_model.as_str()),
         candidate_count,
     );
+    observe_gateway_stage_ms(
+        "openai_chat_attempt_source_build",
+        attempt_source_started_at.elapsed().as_millis() as u64,
+    );
 
     Ok(Some((
         LocalOpenAiChatStreamAttemptSource {
@@ -85,6 +110,8 @@ pub(crate) async fn build_local_openai_chat_stream_attempt_source<'a>(
             body_json: effective_body_json,
             input,
             candidates,
+            prefetched_attempts: VecDeque::new(),
+            request_preparation: LocalOpenAiChatRequestPreparation,
         },
         candidate_count,
     )))
@@ -93,18 +120,13 @@ pub(crate) async fn build_local_openai_chat_stream_attempt_source<'a>(
 #[async_trait]
 impl LocalExecutionAttemptSource<AiStreamAttempt> for LocalOpenAiChatStreamAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiStreamAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        let select_started_at = std::time::Instant::now();
-            match self.build_stream_attempt(attempt).await? {
+        let selected = self.next_execution_attempt_with_target_select().await?;
-                Some(attempt) => return Ok(Some(attempt)),
+        observe_gateway_stage_ms(
-                None => continue,
+            "openai_chat_stream_target_select",
-            }
+            select_started_at.elapsed().as_millis() as u64,
-        }
-        apply_local_runtime_candidate_terminal_reason(
-            self.state,
-            self.trace_id,
-            "no_local_stream_plans",
         );
-        Ok(None)
+        Ok(selected)
     }
 
     async fn drain_execution_attempts(&mut self) -> Result<Vec<AiStreamAttempt>, GatewayError> {
@@ -116,11 +138,178 @@ impl LocalExecutionAttemptSource<AiStreamAttempt> for LocalOpenAiChatStreamAttem
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.prefetched_attempts
+            .retain(|attempt| attempt.eligible.candidate.key_id != key_id);
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.prefetched_attempts
+            .retain(|attempt| attempt.eligible.candidate.endpoint_id != endpoint_id);
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.prefetched_attempts
+            .retain(|attempt| attempt.eligible.candidate.provider_id != provider_id);
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 impl LocalOpenAiChatStreamAttemptSource<'_> {
-    async fn build_stream_attempt(
+    async fn next_execution_attempt_with_target_select(
+        &mut self,
+    ) -> Result<Option<AiStreamAttempt>, GatewayError> {
+        loop {
+            let Some(attempt) = self.next_raw_attempt_with_target_select().await? else {
+                apply_local_runtime_candidate_terminal_reason(
+                    self.state,
+                    self.trace_id,
+                    "no_local_stream_plans",
+                );
+                return Ok(None);
+            };
+
+            let plan_started_at = std::time::Instant::now();
+            record_openai_chat_stream_payload_build_selected();
+            match self.build_stream_attempt(attempt).await? {
+                Some(attempt) => {
+                    observe_gateway_stage_ms(
+                        "stream_candidate_plan_build",
+                        plan_started_at.elapsed().as_millis() as u64,
+                    );
+                    return Ok(Some(attempt));
+                }
+                None => {
+                    observe_gateway_stage_ms(
+                        "stream_candidate_plan_build",
+                        plan_started_at.elapsed().as_millis() as u64,
+                    );
+                    continue;
+                }
+            }
+        }
+    }
+
+    async fn next_raw_attempt_with_target_select(
+        &mut self,
+    ) -> Result<Option<LocalOpenAiChatCandidateAttempt>, GatewayError> {
+        let select_window = openai_chat_stream_target_select_window();
+        if select_window <= 1 {
+            return self.next_raw_attempt_linear().await;
+        }
+        let mut attempts = Vec::with_capacity(select_window);
+        for _ in 0..select_window {
+            match self.next_raw_attempt_linear().await? {
+                Some(attempt) => attempts.push(attempt),
+                None => break,
+            }
+        }
+        if attempts.is_empty() {
+            return Ok(None);
+        }
+        record_openai_chat_stream_raw_candidates_scanned(attempts.len());
+        let seed = request_distribution_seed();
+        let target_keys = attempts
+            .iter()
+            .map(|attempt| self.lightweight_target_key_for_attempt(attempt))
+            .collect::<Vec<_>>();
+        for target_key in target_keys.iter().flatten() {
+            self.state
+                .upstream_target_admission
+                .record_raw_seen_for_target_key(target_key);
+        }
+        let selected_index = if target_keys.iter().all(Option::is_some) {
+            let choices = attempts
+                .iter()
+                .zip(target_keys.iter())
+                .map(|(attempt, target_key)| {
+                    let target_key = target_key.as_deref().unwrap_or("-");
+                    let snapshot = self
+                        .state
+                        .upstream_target_admission
+                        .snapshot_for_target_key(target_key);
+                    TargetSelectChoice {
+                        target_key,
+                        identity: target_select_candidate_identity(attempt),
+                        in_flight: snapshot
+                            .as_ref()
+                            .map(|snapshot| snapshot.in_flight)
+                            .unwrap_or(0),
+                        selection_pressure_total: snapshot
+                            .as_ref()
+                            .map(|snapshot| snapshot.selection_pressure_total)
+                            .unwrap_or(0),
+                    }
+                })
+                .collect::<Vec<_>>();
+            select_target_index(seed, &choices)
+        } else {
+            0
+        };
+        record_openai_chat_stream_target_select_selected_rank(selected_index);
+        record_openai_chat_stream_payload_build_prefetch_avoided(attempts.len().saturating_sub(1));
+        if let Some(Some(target_key)) = target_keys.get(selected_index) {
+            self.state
+                .upstream_target_admission
+                .record_preselect_for_target_key(target_key);
+        }
+        let selected = attempts.remove(selected_index);
+        self.prefetched_attempts.extend(attempts);
+        Ok(Some(selected))
+    }
+
+    async fn next_raw_attempt_linear(
+        &mut self,
+    ) -> Result<Option<LocalOpenAiChatCandidateAttempt>, GatewayError> {
+        if let Some(attempt) = self.prefetched_attempts.pop_front() {
+            return Ok(Some(attempt));
+        }
+        let source_started_at = std::time::Instant::now();
+        let attempt = self.candidates.next_attempt().await?;
+        observe_gateway_stage_ms(
+            "stream_candidate_source_next",
+            source_started_at.elapsed().as_millis() as u64,
+        );
+        Ok(attempt)
+    }
+
+    fn lightweight_target_key_for_attempt(
         &self,
+        attempt: &LocalOpenAiChatCandidateAttempt,
+    ) -> Option<String> {
+        let provider_api_format = attempt.eligible.provider_api_format.trim();
+        if !provider_api_format.eq_ignore_ascii_case("openai:chat") {
+            return None;
+        }
+        let transport = &attempt.eligible.transport;
+        if transport
+            .provider
+            .provider_type
+            .trim()
+            .eq_ignore_ascii_case("grok")
+            || is_windsurf_provider_transport(transport)
+            || local_openai_chat_transport_unsupported_reason(transport).is_some()
+        {
+            return None;
+        }
+        if transport.provider.proxy.is_some()
+            || transport.endpoint.proxy.is_some()
+            || transport.key.proxy.is_some()
+        {
+            return None;
+        }
+        let upstream_url = build_local_openai_chat_upstream_url(self.parts, transport)?;
+        upstream_target_key_from_url(upstream_url.as_str(), None)
+    }
+
+    async fn build_stream_attempt(
+        &mut self,
         attempt: LocalOpenAiChatCandidateAttempt,
     ) -> Result<Option<AiStreamAttempt>, GatewayError> {
         let upstream_is_stream = openai_chat_upstream_is_stream_for_candidate(
@@ -134,6 +323,7 @@ impl LocalOpenAiChatStreamAttemptSource<'_> {
             self.trace_id,
             &self.body_json,
             &self.input,
+            Some(&mut self.request_preparation),
             attempt,
             OPENAI_CHAT_STREAM_PLAN_KIND,
             "openai_chat_stream_success",
@@ -158,6 +348,97 @@ impl LocalOpenAiChatStreamAttemptSource<'_> {
     }
 }
 
+fn openai_chat_stream_target_select_window() -> usize {
+    std::env::var(OPENAI_CHAT_STREAM_TARGET_SELECT_WINDOW_ENV)
+        .ok()
+        .and_then(|value| value.trim().parse::<usize>().ok())
+        .filter(|value| *value > 0)
+        .unwrap_or(DEFAULT_OPENAI_CHAT_STREAM_TARGET_SELECT_WINDOW)
+        .clamp(1, MAX_OPENAI_CHAT_STREAM_TARGET_SELECT_WINDOW)
+}
+
+#[derive(Clone, Copy)]
+struct TargetSelectCandidateIdentity<'a> {
+    provider_id: &'a str,
+    endpoint_id: &'a str,
+    key_id: &'a str,
+    candidate_id: &'a str,
+}
+
+#[derive(Clone, Copy)]
+struct TargetSelectChoice<'a> {
+    target_key: &'a str,
+    identity: TargetSelectCandidateIdentity<'a>,
+    in_flight: usize,
+    selection_pressure_total: u64,
+}
+
+fn select_target_index(seed: u64, choices: &[TargetSelectChoice<'_>]) -> usize {
+    choices
+        .iter()
+        .enumerate()
+        .min_by_key(|(index, choice)| {
+            target_select_score(
+                seed,
+                choice.target_key,
+                &choice.identity,
+                *index,
+                choice.in_flight,
+                choice.selection_pressure_total,
+            )
+        })
+        .map(|(index, _)| index)
+        .unwrap_or(0)
+}
+
+fn target_select_candidate_identity(
+    attempt: &LocalOpenAiChatCandidateAttempt,
+) -> TargetSelectCandidateIdentity<'_> {
+    TargetSelectCandidateIdentity {
+        provider_id: &attempt.eligible.candidate.provider_id,
+        endpoint_id: &attempt.eligible.candidate.endpoint_id,
+        key_id: &attempt.eligible.candidate.key_id,
+        candidate_id: &attempt.candidate_id,
+    }
+}
+
+fn target_select_tie_break(
+    seed: u64,
+    target_key: &str,
+    identity: &TargetSelectCandidateIdentity<'_>,
+    index: usize,
+) -> u64 {
+    let mut hash = seed ^ ((index as u64).wrapping_mul(0x9E37_79B9_7F4A_7C15));
+    hash = hash_string(hash, target_key);
+    hash = hash_string(hash, identity.provider_id);
+    hash = hash_string(hash, identity.endpoint_id);
+    hash = hash_string(hash, identity.key_id);
+    hash_string(hash, identity.candidate_id)
+}
+
+fn hash_string(mut hash: u64, value: &str) -> u64 {
+    for byte in value.as_bytes() {
+        hash ^= u64::from(*byte);
+        hash = hash.wrapping_mul(0x100_0000_01B3);
+    }
+    hash
+}
+
+fn target_select_score(
+    seed: u64,
+    target_key: &str,
+    identity: &TargetSelectCandidateIdentity<'_>,
+    index: usize,
+    in_flight: usize,
+    selected_total: u64,
+) -> (usize, u64, u64) {
+    (
+        in_flight,
+        selected_total,
+        target_select_tie_break(seed, target_key, identity, index),
+    )
+}
+
 pub(crate) async fn build_local_openai_chat_stream_plan_and_reports(
     state: &AppState,
     parts: &http::request::Parts,
@@ -207,3 +488,82 @@ pub(crate) async fn build_local_openai_chat_stream_plan_and_reports(
 
     Ok(plans)
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn identity<'a>(
+        endpoint_id: &'a str,
+        candidate_id: &'a str,
+    ) -> TargetSelectCandidateIdentity<'a> {
+        TargetSelectCandidateIdentity {
+            provider_id: "provider",
+            endpoint_id,
+            key_id: "key",
+            candidate_id,
+        }
+    }
+
+    #[test]
+    fn target_select_score_prefers_lower_in_flight() {
+        let busy = identity("endpoint-a", "candidate-a");
+        let idle = identity("endpoint-b", "candidate-b");
+
+        assert!(
+            target_select_score(7, "http://127.0.0.1:18182|proxy=-", &idle, 1, 0, 10)
+                < target_select_score(7, "http://127.0.0.1:18181|proxy=-", &busy, 0, 5, 0)
+        );
+    }
+
+    #[test]
+    fn target_select_tie_break_distinguishes_equivalent_targets() {
+        let left = identity("endpoint-a", "candidate-a");
+        let right = identity("endpoint-b", "candidate-b");
+
+        assert_ne!(
+            target_select_tie_break(11, "http://127.0.0.1:18181|proxy=-", &left, 0),
+            target_select_tie_break(11, "http://127.0.0.1:18182|proxy=-", &right, 1)
+        );
+    }
+
+    #[test]
+    fn select_target_index_prefers_lower_in_flight_target() {
+        let choices = [
+            TargetSelectChoice {
+                target_key: "http://127.0.0.1:18181|proxy=-",
+                identity: identity("endpoint-a", "candidate-a"),
+                in_flight: 8,
+                selection_pressure_total: 0,
+            },
+            TargetSelectChoice {
+                target_key: "http://127.0.0.1:18182|proxy=-",
+                identity: identity("endpoint-b", "candidate-b"),
+                in_flight: 1,
+                selection_pressure_total: 100,
+            },
+        ];
+
+        assert_eq!(select_target_index(17, &choices), 1);
+    }
+
+    #[test]
+    fn select_target_index_uses_selection_pressure_before_tie_break() {
+        let choices = [
+            TargetSelectChoice {
+                target_key: "http://127.0.0.1:18181|proxy=-",
+                identity: identity("endpoint-a", "candidate-a"),
+                in_flight: 0,
+                selection_pressure_total: 20,
+            },
+            TargetSelectChoice {
+                target_key: "http://127.0.0.1:18182|proxy=-",
+                identity: identity("endpoint-b", "candidate-b"),
+                in_flight: 0,
+                selection_pressure_total: 1,
+            },
+        ];
+
+        assert_eq!(select_target_index(19, &choices), 1);
+    }
+}

apps/aether-gateway/src/ai_serving/planner/standard/openai/chat/plans/sync.rs

@@ -93,7 +93,7 @@ pub(crate) async fn build_local_openai_chat_sync_attempt_source<'a>(
 #[async_trait]
 impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalOpenAiChatSyncAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiSyncAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_sync_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -116,6 +116,21 @@ impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalOpenAiChatSyncAttemptSo
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 impl LocalOpenAiChatSyncAttemptSource<'_> {
@@ -134,6 +149,7 @@ impl LocalOpenAiChatSyncAttemptSource<'_> {
             self.trace_id,
             &self.body_json,
             &self.input,
+            None,
             attempt,
             OPENAI_CHAT_SYNC_PLAN_KIND,
             "openai_chat_sync_success",

apps/aether-gateway/src/ai_serving/planner/standard/openai/plan_builders/stream.rs

@@ -129,7 +129,7 @@ pub(crate) fn build_openai_chat_stream_plan_from_decision(
             headers: std::mem::take(&mut provider_request_headers),
             content_type,
             body: RequestBody::from_json(provider_request_body_value),
-            stream: true,
+            stream: effective_upstream_is_stream,
         },
     );
 
@@ -229,7 +229,7 @@ pub(crate) fn build_openai_responses_stream_plan_from_decision(
             headers: std::mem::take(&mut provider_request_headers),
             content_type,
             body: RequestBody::from_json(provider_request_body_value),
-            stream: true,
+            stream: effective_upstream_is_stream,
         },
     );
 
@@ -291,6 +291,7 @@ mod tests {
             request_id: Some("req_123".to_string()),
             candidate_id: Some("cand_123".to_string()),
             provider_name: Some("Codex".to_string()),
+            provider_type: Some("codex".to_string()),
             provider_id: Some("prov_123".to_string()),
             endpoint_id: Some("ep_123".to_string()),
             key_id: Some("key_123".to_string()),
@@ -385,6 +386,46 @@ mod tests {
         );
     }
 
+    #[test]
+    fn build_compact_stream_plan_preserves_non_stream_upstream_mode() {
+        let parts = http::Request::builder()
+            .uri("http://localhost/v1/responses/compact")
+            .body(())
+            .expect("request should build")
+            .into_parts()
+            .0;
+        let mut payload = sample_responses_payload();
+        payload.decision_kind = Some("openai_responses_compact_stream".to_string());
+        payload.upstream_url = Some("https://example.com/v1/responses/compact".to_string());
+        payload.provider_api_format = Some("openai:responses:compact".to_string());
+        payload.client_api_format = Some("openai:responses:compact".to_string());
+        payload.upstream_is_stream = false;
+        payload.provider_request_body = Some(json!({
+            "model": "gpt-5.6-sol",
+            "input": [],
+            "instructions": "You are Codex.",
+            "tools": [],
+            "parallel_tool_calls": true,
+            "reasoning": {"effort": "high"},
+            "prompt_cache_key": "cache-key",
+            "text": {"verbosity": "low"}
+        }));
+
+        let built =
+            build_openai_responses_stream_plan_from_decision(&parts, &json!({}), payload, true)
+                .expect("plan build should succeed")
+                .expect("plan should be produced");
+
+        assert!(!built.plan.stream);
+        assert!(built
+            .plan
+            .body
+            .json_body
+            .as_ref()
+            .is_some_and(|body| body.get("stream").is_none()));
+        assert!(!built.plan.headers.contains_key("accept"));
+    }
+
     #[test]
     fn build_openai_chat_stream_plan_fallback_preserves_complete_same_format_headers() {
         let parts = http::Request::builder()
@@ -404,6 +445,7 @@ mod tests {
             request_id: Some("req_stream_456".to_string()),
             candidate_id: Some("cand_stream_456".to_string()),
             provider_name: Some("OpenAI".to_string()),
+            provider_type: Some("openai".to_string()),
             provider_id: Some("prov_stream_456".to_string()),
             endpoint_id: Some("ep_stream_456".to_string()),
             key_id: Some("key_stream_456".to_string()),
@@ -462,7 +504,7 @@ mod tests {
     }
 
     #[test]
-    fn build_openai_chat_stream_plan_keeps_downstream_stream_for_force_non_stream_upstream() {
+    fn build_openai_chat_stream_plan_preserves_force_non_stream_upstream_mode() {
         fn force_non_stream_payload(provider_request_body: Option<Value>) -> AiExecutionDecision {
             AiExecutionDecision {
                 action: "stream".to_string(),
@@ -472,6 +514,7 @@ mod tests {
                 request_id: Some("req_force_non_stream".to_string()),
                 candidate_id: Some("cand_force_non_stream".to_string()),
                 provider_name: Some("OpenAI".to_string()),
+                provider_type: Some("openai".to_string()),
                 provider_id: Some("prov_force_non_stream".to_string()),
                 endpoint_id: Some("ep_force_non_stream".to_string()),
                 key_id: Some("key_force_non_stream".to_string()),
@@ -523,7 +566,7 @@ mod tests {
         .expect("plan build should succeed")
         .expect("plan should be produced");
 
-        assert!(built.plan.stream);
+        assert!(!built.plan.stream);
         assert_eq!(
             built
                 .plan
@@ -548,7 +591,7 @@ mod tests {
         .expect("fallback plan build should succeed")
         .expect("fallback plan should be produced");
 
-        assert!(built.plan.stream);
+        assert!(!built.plan.stream);
         assert_eq!(
             built
                 .plan
@@ -579,6 +622,7 @@ mod tests {
             request_id: Some("req_stream_789".to_string()),
             candidate_id: Some("cand_stream_789".to_string()),
             provider_name: Some("Claude".to_string()),
+            provider_type: Some("anthropic".to_string()),
             provider_id: Some("prov_stream_789".to_string()),
             endpoint_id: Some("ep_stream_789".to_string()),
             key_id: Some("key_stream_789".to_string()),

apps/aether-gateway/src/ai_serving/planner/standard/openai/plan_builders/sync.rs

@@ -257,6 +257,7 @@ mod tests {
             request_id: Some("req_123".to_string()),
             candidate_id: Some("cand_123".to_string()),
             provider_name: Some("Codex".to_string()),
+            provider_type: Some("codex".to_string()),
             provider_id: Some("prov_123".to_string()),
             endpoint_id: Some("ep_123".to_string()),
             key_id: Some("key_123".to_string()),
@@ -369,6 +370,7 @@ mod tests {
             request_id: Some("req_456".to_string()),
             candidate_id: Some("cand_456".to_string()),
             provider_name: Some("OpenAI".to_string()),
+            provider_type: Some("openai".to_string()),
             provider_id: Some("prov_456".to_string()),
             endpoint_id: Some("ep_456".to_string()),
             key_id: Some("key_456".to_string()),
@@ -440,6 +442,7 @@ mod tests {
             request_id: Some("req_789".to_string()),
             candidate_id: Some("cand_789".to_string()),
             provider_name: Some("Claude".to_string()),
+            provider_type: Some("anthropic".to_string()),
             provider_id: Some("prov_789".to_string()),
             endpoint_id: Some("ep_789".to_string()),
             key_id: Some("key_789".to_string()),

apps/aether-gateway/src/ai_serving/planner/standard/openai/responses/decision/payload.rs

@@ -9,7 +9,7 @@ use crate::ai_serving::planner::report_context::{
 };
 use crate::ai_serving::planner::spec_metadata::local_openai_responses_spec_metadata;
 use crate::ai_serving::planner::{
-    build_ai_execution_decision_response, resolve_transport_request_gzip_policy,
+    build_ai_execution_decision_response, resolve_transport_request_encoding_policy,
     AiExecutionDecisionResponseParts,
 };
 use crate::ai_serving::transport::{
@@ -142,6 +142,7 @@ pub(crate) async fn maybe_build_local_openai_responses_decision_payload_for_cand
                 original_request_body_json,
                 original_request_body_base64: None,
                 client_session_affinity: input.client_session_affinity.as_ref(),
+                routing_policy: input.routing_policy.as_ref(),
                 scheduler_affinity_epoch: eligible.orchestration.scheduler_affinity_epoch,
                 client_requested_stream: body_json
                     .get("stream")
@@ -204,7 +205,7 @@ pub(crate) async fn maybe_build_local_openai_responses_decision_payload_for_cand
         image_request_summary: _,
         request_redacted: _,
     } = resolved;
-    let request_gzip = resolve_transport_request_gzip_policy(&transport);
+    let request_encoding = resolve_transport_request_encoding_policy(&transport);
 
     let mut decision = build_ai_execution_decision_response(AiExecutionDecisionResponseParts {
         decision_is_stream: spec_metadata.require_streaming,
@@ -214,6 +215,7 @@ pub(crate) async fn maybe_build_local_openai_responses_decision_payload_for_cand
         request_id: trace_id.to_string(),
         candidate_id: candidate_id.clone(),
         provider_name: transport.provider.name.clone(),
+        provider_type: transport.provider.provider_type.clone(),
         provider_id: candidate.provider_id.clone(),
         endpoint_id: candidate.endpoint_id.clone(),
         key_id: candidate.key_id.clone(),
@@ -231,8 +233,8 @@ pub(crate) async fn maybe_build_local_openai_responses_decision_payload_for_cand
         provider_request_body: Some(provider_request_body),
         provider_request_body_base64: None,
         content_type: Some("application/json".to_string()),
-        content_encoding: None,
+        content_encoding: request_encoding.content_encoding,
-        request_gzip,
+        request_gzip: request_encoding.request_gzip,
         proxy,
         transport_profile,
         timeouts,
@@ -241,6 +243,10 @@ pub(crate) async fn maybe_build_local_openai_responses_decision_payload_for_cand
         report_context: Some(report_context),
         auth_context: input.auth_context.clone(),
     });
-    apply_provider_request_routing_policy_to_decision(input, &mut decision)?;
+    apply_provider_request_routing_policy_to_decision(
+        input,
+        &mut decision,
+        Some(transport.as_ref()),
+    )?;
     Ok(Some(decision))
 }

apps/aether-gateway/src/ai_serving/planner/standard/openai/responses/decision/support.rs

@@ -31,8 +31,8 @@ use crate::ai_serving::planner::spec_metadata::local_openai_responses_spec_metad
 use crate::ai_serving::planner::CandidateFailureDiagnostic;
 use crate::ai_serving::{
     ai_local_execution_contract_for_formats, extract_pool_sticky_session_token,
-    resolve_local_decision_execution_runtime_auth_context, ExecutionRuntimeAuthContext,
+    openai_responses_request_operation, resolve_local_decision_execution_runtime_auth_context,
-    GatewayControlDecision, PlannerAppState,
+    ExecutionRuntimeAuthContext, GatewayControlDecision, PlannerAppState,
 };
 use crate::client_session_affinity::client_session_affinity_from_parts;
 use crate::{AppState, GatewayError};
@@ -91,7 +91,9 @@ pub(crate) async fn resolve_local_openai_responses_decision_input(
         state,
         auth_context.clone(),
         Some(requested_model.as_str()),
+        decision.auth_endpoint_signature.as_deref(),
         None,
+        &decision.model_directive_policy,
     )
     .await
     {
@@ -161,6 +163,7 @@ pub(crate) async fn materialize_local_openai_responses_candidate_attempts(
     spec: LocalOpenAiResponsesSpec,
 ) -> Result<(Vec<LocalOpenAiResponsesCandidateAttempt>, usize), GatewayError> {
     let spec_metadata = local_openai_responses_spec_metadata(spec);
+    let request_operation = openai_responses_request_operation(spec_metadata.api_format, body_json);
     let planner_state = PlannerAppState::new(state);
     let sticky_session_token = extract_pool_sticky_session_token(body_json);
     let auth_context: &ExecutionRuntimeAuthContext = &input.auth_context;
@@ -171,8 +174,10 @@ pub(crate) async fn materialize_local_openai_responses_candidate_attempts(
     );
     let preselection = preselect_local_execution_candidates_with_serving(
         planner_state,
+        &input.model_directive_policy,
         spec_metadata.api_format,
         &input.requested_model,
+        request_operation,
         spec_metadata.require_streaming,
         input.required_capabilities.as_ref(),
         &input.auth_snapshot,
@@ -259,6 +264,7 @@ pub(crate) async fn build_local_openai_responses_candidate_attempt_source<'a>(
     spec: LocalOpenAiResponsesSpec,
 ) -> Result<(LocalOpenAiResponsesCandidateAttemptSource<'a>, usize), GatewayError> {
     let spec_metadata = local_openai_responses_spec_metadata(spec);
+    let request_operation = openai_responses_request_operation(spec_metadata.api_format, body_json);
     let planner_state = PlannerAppState::new(state);
     let sticky_session_token = extract_pool_sticky_session_token(body_json);
     let auth_context: &ExecutionRuntimeAuthContext = &input.auth_context;
@@ -280,9 +286,11 @@ pub(crate) async fn build_local_openai_responses_candidate_attempt_source<'a>(
     Ok(
         build_lazy_requested_model_execution_candidate_attempt_source_with_serving(
             planner_state,
+            &input.model_directive_policy,
             trace_id,
             spec_metadata.api_format,
             &input.requested_model,
+            request_operation,
             spec_metadata.require_streaming,
             &input.auth_snapshot,
             input.client_session_affinity.as_ref(),
@@ -365,8 +373,10 @@ pub(crate) async fn build_local_openai_responses_image_candidate_attempt_source<
     );
     let preselection = preselect_local_execution_candidates_for_api_formats_with_serving(
         planner_state,
+        &input.model_directive_policy,
         spec_metadata.api_format,
         &input.requested_model,
+        None,
         false,
         input.required_capabilities.as_ref(),
         &input.auth_snapshot,

apps/aether-gateway/src/ai_serving/planner/standard/openai/responses/mod.rs

@@ -114,7 +114,7 @@ pub(crate) async fn maybe_build_sync_local_openai_responses_decision_payload(
     )
     .await?;
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) = maybe_build_local_openai_responses_decision_payload_for_candidate(
             state, parts, trace_id, body_json, &input, attempt, spec,
         )
@@ -153,7 +153,7 @@ pub(crate) async fn maybe_build_stream_local_openai_responses_decision_payload(
     )
     .await?;
 
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         if let Some(payload) = maybe_build_local_openai_responses_decision_payload_for_candidate(
             state, parts, trace_id, body_json, &input, attempt, spec,
         )

apps/aether-gateway/src/ai_serving/planner/standard/openai/responses/plans.rs

@@ -162,7 +162,7 @@ pub(super) async fn build_local_stream_attempt_source<'a>(
 #[async_trait]
 impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalOpenAiResponsesSyncAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiSyncAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_sync_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -185,12 +185,27 @@ impl LocalExecutionAttemptSource<AiSyncAttempt> for LocalOpenAiResponsesSyncAtte
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 #[async_trait]
 impl LocalExecutionAttemptSource<AiStreamAttempt> for LocalOpenAiResponsesStreamAttemptSource<'_> {
     async fn next_execution_attempt(&mut self) -> Result<Option<AiStreamAttempt>, GatewayError> {
-        while let Some(attempt) = self.candidates.next_attempt().await {
+        while let Some(attempt) = self.candidates.next_attempt().await? {
             match self.build_stream_attempt(attempt).await? {
                 Some(attempt) => return Ok(Some(attempt)),
                 None => continue,
@@ -213,6 +228,21 @@ impl LocalExecutionAttemptSource<AiStreamAttempt> for LocalOpenAiResponsesStream
         }
         Ok(drained)
     }
+
+    async fn skip_credential(&mut self, key_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_credential(key_id);
+        Ok(())
+    }
+
+    async fn skip_endpoint(&mut self, endpoint_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_endpoint(endpoint_id);
+        Ok(())
+    }
+
+    async fn skip_provider(&mut self, provider_id: &str) -> Result<(), GatewayError> {
+        self.candidates.skip_provider(provider_id);
+        Ok(())
+    }
 }
 
 impl LocalOpenAiResponsesSyncAttemptSource<'_> {
@@ -331,7 +361,7 @@ pub(super) async fn build_local_sync_plan_and_reports(
     }
 
     let mut plans = Vec::new();
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_openai_responses_decision_payload_for_candidate(
             state, parts, trace_id, body_json, &input, attempt, spec,
         )
@@ -403,7 +433,7 @@ pub(super) async fn build_local_stream_plan_and_reports(
     }
 
     let mut plans = Vec::new();
-    while let Some(attempt) = source.next_attempt().await {
+    while let Some(attempt) = source.next_attempt().await? {
         let Some(payload) = maybe_build_local_openai_responses_decision_payload_for_candidate(
             state, parts, trace_id, body_json, &input, attempt, spec,
         )

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

@@ -1,9 +1,8 @@
 use std::collections::BTreeMap;
 
-use aether_contracts::RequestBody;
-
 use super::{
-    augment_sync_report_context, build_ai_execution_plan_from_decision, take_ai_decision_plan_core,
+    augment_sync_report_context, build_ai_execution_plan_from_decision,
+    resolve_ai_passthrough_sync_request_body, take_ai_decision_plan_core,
     take_ai_upstream_auth_pair, take_non_empty_string, AiExecutionPlanFromDecisionParts,
     AiStreamAttempt, AiSyncAttempt,
 };
@@ -63,6 +62,10 @@ pub(crate) fn build_standard_sync_plan_from_decision(
         &provider_request_headers,
         &provider_request_body_value,
     )?;
+    let request_body = resolve_ai_passthrough_sync_request_body(
+        Some(provider_request_body_value),
+        payload.provider_request_body_base64.take(),
+    );
     let stream = payload.upstream_is_stream;
     let plan = build_ai_execution_plan_from_decision(
         &mut payload,
@@ -72,7 +75,7 @@ pub(crate) fn build_standard_sync_plan_from_decision(
             url,
             headers: std::mem::take(&mut provider_request_headers),
             content_type,
-            body: RequestBody::from_json(provider_request_body_value),
+            body: request_body,
             stream,
         },
     );
@@ -146,6 +149,11 @@ pub(crate) fn build_standard_stream_plan_from_decision(
         &provider_request_headers,
         &provider_request_body_value,
     )?;
+    let request_body = resolve_ai_passthrough_sync_request_body(
+        Some(provider_request_body_value),
+        payload.provider_request_body_base64.take(),
+    );
+    let stream = payload.upstream_is_stream;
     let plan = build_ai_execution_plan_from_decision(
         &mut payload,
         AiExecutionPlanFromDecisionParts {
@@ -154,8 +162,8 @@ pub(crate) fn build_standard_stream_plan_from_decision(
             url,
             headers: std::mem::take(&mut provider_request_headers),
             content_type,
-            body: RequestBody::from_json(provider_request_body_value),
+            body: request_body,
-            stream: true,
+            stream,
         },
     );
 
@@ -165,3 +173,88 @@ pub(crate) fn build_standard_stream_plan_from_decision(
         report_context,
     }))
 }
+
+#[cfg(test)]
+mod tests {
+    use aether_contracts::{ExecutionResponseBodyMode, EXECUTION_RESPONSE_BODY_MODE_HEADER};
+    use serde_json::json;
+
+    use super::{
+        build_standard_stream_plan_from_decision, build_standard_sync_plan_from_decision,
+        AiExecutionDecision,
+    };
+
+    fn decision_with_raw_body(upstream_is_stream: bool) -> AiExecutionDecision {
+        serde_json::from_value(json!({
+            "action": if upstream_is_stream { "stream" } else { "sync" },
+            "request_id": "req-raw",
+            "provider_id": "provider-raw",
+            "endpoint_id": "endpoint-raw",
+            "key_id": "key-raw",
+            "upstream_url": "https://api.anthropic.test/v1/messages",
+            "provider_api_format": "claude:messages",
+            "client_api_format": "claude:messages",
+            "provider_request_headers": {
+                "content-type": "application/json",
+                (EXECUTION_RESPONSE_BODY_MODE_HEADER): ExecutionResponseBodyMode::PreserveBytes.as_str()
+            },
+            "provider_request_body": {
+                "model": "claude-sonnet-4",
+                "messages": []
+            },
+            "provider_request_body_base64": "eyAibW9kZWwiOiAiY2xhdWRlLXNvbm5ldC00IiwgIm1lc3NhZ2VzIjogW10gfQ==",
+            "content_type": "application/json",
+            "upstream_is_stream": upstream_is_stream
+        }))
+        .expect("decision should deserialize")
+    }
+
+    fn request_parts() -> http::request::Parts {
+        http::Request::builder()
+            .uri("http://localhost/v1/messages")
+            .body(())
+            .expect("request should build")
+            .into_parts()
+            .0
+    }
+
+    #[test]
+    fn standard_sync_plan_prefers_exact_request_body_bytes() {
+        let built = build_standard_sync_plan_from_decision(
+            &request_parts(),
+            &json!({}),
+            decision_with_raw_body(false),
+        )
+        .expect("plan should build")
+        .expect("plan should exist");
+
+        assert!(built.plan.body.json_body.is_none());
+        assert_eq!(
+            built.plan.body.body_bytes_b64.as_deref(),
+            Some("eyAibW9kZWwiOiAiY2xhdWRlLXNvbm5ldC00IiwgIm1lc3NhZ2VzIjogW10gfQ==")
+        );
+        assert_eq!(
+            built
+                .plan
+                .headers
+                .get(EXECUTION_RESPONSE_BODY_MODE_HEADER)
+                .map(String::as_str),
+            Some(ExecutionResponseBodyMode::PreserveBytes.as_str())
+        );
+    }
+
+    #[test]
+    fn standard_stream_plan_prefers_exact_request_body_bytes() {
+        let built = build_standard_stream_plan_from_decision(
+            &request_parts(),
+            &json!({}),
+            decision_with_raw_body(true),
+            false,
+        )
+        .expect("plan should build")
+        .expect("plan should exist");
+
+        assert!(built.plan.body.json_body.is_none());
+        assert!(built.plan.body.body_bytes_b64.is_some());
+    }
+}

apps/aether-gateway/src/ai_serving/planner/state/auth.rs

@@ -10,9 +10,7 @@ impl<'a> PlannerAppState<'a> {
         now_unix_secs: u64,
     ) -> Result<Option<GatewayAuthApiKeySnapshot>, GatewayError> {
         self.app()
-            .data
+            .read_cached_auth_api_key_snapshot(user_id, api_key_id, now_unix_secs)
-            .read_auth_api_key_snapshot(user_id, api_key_id, now_unix_secs)
             .await
-            .map_err(|err| GatewayError::Internal(err.to_string()))
     }
 }

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

@@ -10,16 +10,15 @@ impl<'a> PlannerAppState<'a> {
         api_key_id: &str,
         requested_model: Option<&str>,
         explicit_required_capabilities: Option<&Value>,
+        model_directive_base_model: Option<&str>,
     ) -> Option<Value> {
-        let enable_model_directives =
-            crate::system_features::reasoning_model_directive_enabled(self.app()).await;
         crate::request_candidate_runtime::resolve_request_candidate_required_capabilities(
             self.app(),
             user_id,
             api_key_id,
             requested_model,
             explicit_required_capabilities,
-            enable_model_directives,
+            model_directive_base_model,
         )
         .await
     }

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

@@ -20,14 +20,8 @@ impl<'a> PlannerAppState<'a> {
         auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
         client_session_affinity: Option<&ClientSessionAffinity>,
         now_unix_secs: u64,
+        enable_model_directives: bool,
     ) -> Result<Vec<SchedulerMinimalCandidateSelectionCandidate>, GatewayError> {
-        let enable_model_directives =
-            crate::system_features::reasoning_model_directive_enabled_for_api_format_and_model(
-                self.app(),
-                api_format,
-                Some(global_model_name),
-            )
-            .await;
         crate::scheduler::candidate::list_selectable_candidates(
             self.app().data.as_ref(),
             self.app(),
@@ -52,6 +46,39 @@ impl<'a> PlannerAppState<'a> {
         auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
         client_session_affinity: Option<&ClientSessionAffinity>,
         now_unix_secs: u64,
+        enable_model_directives: bool,
+    ) -> Result<
+        (
+            Vec<SchedulerMinimalCandidateSelectionCandidate>,
+            Vec<SchedulerSkippedCandidate>,
+        ),
+        GatewayError,
+    > {
+        self.list_selectable_candidates_with_skip_reasons_for_request_operation(
+            api_format,
+            global_model_name,
+            require_streaming,
+            required_capabilities,
+            auth_snapshot,
+            client_session_affinity,
+            now_unix_secs,
+            enable_model_directives,
+            None,
+        )
+        .await
+    }
+
+    pub(crate) async fn list_selectable_candidates_with_skip_reasons_for_request_operation(
+        self,
+        api_format: &str,
+        global_model_name: &str,
+        require_streaming: bool,
+        required_capabilities: Option<&serde_json::Value>,
+        auth_snapshot: Option<&GatewayAuthApiKeySnapshot>,
+        client_session_affinity: Option<&ClientSessionAffinity>,
+        now_unix_secs: u64,
+        enable_model_directives: bool,
+        request_operation: Option<&str>,
     ) -> Result<
         (
             Vec<SchedulerMinimalCandidateSelectionCandidate>,
@@ -63,16 +90,8 @@ impl<'a> PlannerAppState<'a> {
         let wait_interval = Duration::from_millis(API_KEY_CONCURRENCY_WAIT_POLL_INTERVAL_MS.max(1));
         let wait_deadline = Instant::now() + wait_timeout;
         let mut attempt_now_unix_secs = now_unix_secs;
-        let enable_model_directives =
-            crate::system_features::reasoning_model_directive_enabled_for_api_format_and_model(
-                self.app(),
-                api_format,
-                Some(global_model_name),
-            )
-            .await;
-
         loop {
-            let result = crate::scheduler::candidate::list_selectable_candidates_with_skip_reasons(
+            let result = crate::scheduler::candidate::list_selectable_candidates_with_skip_reasons_for_request_operation(
                 self.app().data.as_ref(),
                 self.app(),
                 api_format,
@@ -83,6 +102,7 @@ impl<'a> PlannerAppState<'a> {
                 client_session_affinity,
                 attempt_now_unix_secs,
                 enable_model_directives,
+                request_operation,
             )
             .await?;
 

apps/aether-gateway/src/ai_serving/planner/state/transport.rs

@@ -3,8 +3,20 @@ pub(crate) use crate::ai_serving::transport::{
     GatewayProviderTransportSnapshot, LocalResolvedOAuthRequestAuth,
 };
 use crate::GatewayError;
+use std::sync::Arc;
 
 impl<'a> PlannerAppState<'a> {
+    pub(crate) async fn read_provider_transport_snapshot_arc(
+        self,
+        provider_id: &str,
+        endpoint_id: &str,
+        key_id: &str,
+    ) -> Result<Option<Arc<GatewayProviderTransportSnapshot>>, GatewayError> {
+        self.app()
+            .read_provider_transport_snapshot_arc(provider_id, endpoint_id, key_id)
+            .await
+    }
+
     pub(crate) async fn read_provider_transport_snapshot(
         self,
         provider_id: &str,

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

@@ -3,14 +3,21 @@ pub(crate) use aether_ai_formats::api::{
     aggregate_openai_chat_stream_sync_response, aggregate_openai_responses_stream_sync_response,
     aggregate_standard_chat_stream_sync_response, aggregate_standard_cli_stream_sync_response,
     api_format_alias_matches, api_format_storage_aliases,
-    apply_codex_openai_responses_chat_body_edits, apply_codex_openai_responses_special_body_edits,
+    apply_codex_openai_compact_terminal_headers, apply_codex_openai_responses_chat_body_edits,
-    apply_codex_openai_responses_special_headers, apply_model_directive_mapping_patch,
+    apply_codex_openai_responses_identity_headers,
+    apply_codex_openai_responses_lite_header_for_request_body_with_capabilities,
+    apply_codex_openai_responses_lite_header_with_capabilities,
+    apply_codex_openai_responses_special_body_edits,
+    apply_codex_openai_responses_special_body_edits_with_source_model_and_capabilities,
+    apply_codex_openai_special_headers, apply_model_directive_mapping_patch,
     apply_model_directive_overrides_from_model, apply_model_directive_overrides_from_request,
     apply_openai_responses_compact_special_body_edits, build_chatgpt_web_image_request_body,
+    build_codex_model_catalog_metadata, build_codex_openai_image_api_provider_request_body,
     build_core_error_body_for_client_format, build_cross_format_openai_chat_request_body,
     build_cross_format_openai_chat_request_body_with_model_directives,
     build_cross_format_openai_responses_request_body,
     build_cross_format_openai_responses_request_body_with_model_directives,
+    build_cross_format_openai_responses_request_body_with_model_directives_and_history_scope,
     build_gemini_image_request_body_from_openai_image_request,
     build_gemini_image_response_from_openai_image_response,
     build_gemini_image_response_from_openai_responses_image_response, build_generated_tool_call_id,
@@ -41,17 +48,21 @@ pub(crate) use aether_ai_formats::api::{
     convert_standard_chat_response, convert_standard_cli_response, copy_request_number_field,
     copy_request_number_field_as, core_error_background_report_kind,
     core_error_default_client_api_format, core_success_background_report_kind,
+    default_model_directive_mapping_patch, default_model_directive_suffixes,
     default_model_for_openai_image_operation, encode_done_sse, encode_json_sse,
     encode_kiro_sse_events, endpoint_config_forces_upstream_stream_policy,
     enforce_request_body_stream_field, estimate_kiro_tokens, extract_openai_text_content,
+    finalize_openai_provider_request,
+    finalize_openai_provider_request_with_codex_model_capabilities,
     find_kiro_real_thinking_end_tag, find_kiro_real_thinking_end_tag_at_buffer_end,
-    find_kiro_real_thinking_start_tag, force_upstream_streaming_for_provider,
+    find_kiro_real_thinking_start_tag, forbid_upstream_streaming_for_provider,
-    gemini_request_is_image_generation, implicit_sync_finalize_report_kind,
+    force_upstream_streaming_for_provider, gemini_request_is_image_generation,
-    is_core_error_finalize_kind, is_matching_stream_http_request, is_matching_stream_request,
+    hydrate_response_history, implicit_sync_finalize_report_kind, is_core_error_finalize_kind,
-    is_openai_image_stream_request, is_openai_responses_family_format, is_openai_responses_format,
+    is_matching_stream_http_request, is_matching_stream_request, is_openai_image_stream_request,
-    kiro_crc32, map_claude_stop_reason, map_openai_reasoning_effort_to_claude_output,
+    is_openai_responses_compact_format, is_openai_responses_family_format,
-    map_openai_reasoning_effort_to_gemini_budget, maybe_bridge_standard_sync_json_to_stream,
+    is_openai_responses_format, kiro_crc32, map_claude_stop_reason,
-    maybe_build_ai_surface_stream_rewriter,
+    map_openai_reasoning_effort_to_claude_output, map_openai_reasoning_effort_to_gemini_budget,
+    maybe_bridge_standard_sync_json_to_stream, maybe_build_ai_surface_stream_rewriter,
     maybe_build_openai_chat_cross_format_sync_product_from_normalized_payload,
     maybe_build_openai_image_sync_finalize_product,
     maybe_build_openai_responses_cross_format_sync_product_from_normalized_payload,
@@ -60,23 +71,31 @@ pub(crate) use aether_ai_formats::api::{
     maybe_build_standard_cross_format_sync_product_from_normalized_payload,
     maybe_build_standard_same_format_sync_body_from_normalized_payload,
     maybe_build_standard_sync_finalize_product_from_normalized_payload, model_directive_base_model,
-    normalize_api_format_alias, normalize_claude_request_to_openai_chat_request,
+    model_directive_builtin_suffix_supported_for_source_model,
-    normalize_gemini_request_to_openai_chat_request, normalize_openai_image_request,
+    model_directive_suffix_has_builtin_mapping, normalize_api_format_alias,
-    normalize_openai_image_request_with_options,
+    normalize_claude_request_to_openai_chat_request,
+    normalize_gemini_request_to_openai_chat_request, normalize_openai_image_quality,
+    normalize_openai_image_request, normalize_openai_image_request_with_options,
     normalize_openai_responses_request_to_openai_chat_request,
     normalize_provider_private_report_context, normalize_provider_private_response_value,
     normalize_standard_request_to_openai_chat_request, openai_image_operation_from_path,
-    parse_direct_request_body, parse_openai_stop_sequences, parse_openai_tool_result_content,
+    parse_codex_auth_identity, parse_direct_request_body, parse_model_directive,
-    prepare_local_success_response_parts, prepare_local_success_response_parts_owned,
+    parse_model_directive_with_suffixes, parse_openai_stop_sequences,
-    provider_adaptation_allows_sync_finalize_envelope, provider_adaptation_anchor_api_format,
+    parse_openai_tool_result_content, prepare_local_success_response_parts,
-    provider_adaptation_descriptor_for_envelope, provider_adaptation_descriptor_for_provider_type,
+    prepare_local_success_response_parts_owned, project_codex_openai_image_api_request_body,
+    project_openai_image_api_request_body, provider_adaptation_allows_sync_finalize_envelope,
+    provider_adaptation_anchor_api_format, provider_adaptation_descriptor_for_envelope,
+    provider_adaptation_descriptor_for_provider_type,
     provider_adaptation_requires_eventstream_accept,
     provider_adaptation_should_unwrap_stream_envelope,
-    provider_private_response_allows_sync_finalize, request_candidate_api_format_preference,
+    provider_private_response_allows_sync_finalize, record_converted_response_history,
-    request_candidate_api_formats, request_conversion_kind,
+    request_candidate_api_format_preference, request_candidate_api_formats,
-    request_conversion_requires_enable_flag, request_path_implies_stream_request,
+    request_conversion_kind, request_conversion_requires_enable_flag,
-    resolve_claude_stream_spec, resolve_claude_sync_spec,
+    request_path_implies_stream_request, resolve_claude_stream_spec, resolve_claude_sync_spec,
-    resolve_execution_runtime_stream_plan_kind, resolve_execution_runtime_sync_plan_kind,
+    resolve_codex_responses_model_capabilities, resolve_execution_runtime_stream_plan_kind,
+    resolve_execution_runtime_stream_plan_kind_with_client_surface,
+    resolve_execution_runtime_sync_plan_kind,
+    resolve_execution_runtime_sync_plan_kind_with_client_surface,
     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_image_stream_spec, resolve_local_image_sync_spec,
@@ -85,24 +104,28 @@ pub(crate) use aether_ai_formats::api::{
     resolve_openai_embedding_sync_spec, resolve_openai_responses_stream_spec,
     resolve_openai_responses_sync_spec, resolve_requested_gemini_image_model_for_request,
     resolve_requested_openai_image_model_for_request,
-    resolve_upstream_is_stream_from_endpoint_config, sanitize_request_path,
+    resolve_upstream_is_stream_for_provider as resolve_format_upstream_is_stream_for_provider,
-    sanitize_request_path_and_query, sanitize_request_query_string,
+    resolve_upstream_is_stream_from_endpoint_config, response_history_is_loaded,
-    stream_body_contains_error_event, supports_stream_execution_decision_kind,
+    response_history_storage_key, sanitize_request_path, sanitize_request_path_and_query,
-    supports_sync_execution_decision_kind, sync_chat_response_conversion_kind,
+    sanitize_request_query_string, stream_body_contains_error_event,
-    sync_cli_response_conversion_kind, transform_provider_private_stream_line, value_as_u64,
+    supports_stream_execution_decision_kind, supports_sync_execution_decision_kind,
-    AiControlPlanRequest, AiSurfaceFinalizeError, AiSurfaceStreamRewriter, CanonicalStreamFrame,
+    sync_chat_response_conversion_kind, sync_cli_response_conversion_kind,
-    ChatGptWebImageRequestError, ClaudeClientEmitter, ClaudeProviderState,
+    transform_provider_private_stream_line, validate_openai_provider_request_contract,
-    ExecutionRuntimeAuthContext, FinalizeStreamRewriteMode, FormatContext, GeminiClientEmitter,
+    value_as_u64, AiControlPlanRequest, AiSurfaceFinalizeError, AiSurfaceStreamRewriter,
-    GeminiImageRequestForOpenAi, GeminiProviderState, KiroToClaudeCliStreamState,
+    CanonicalStreamFrame, ChatGptWebImageRequestError, ClaudeClientEmitter, ClaudeProviderState,
-    LocalCoreSyncErrorKind, LocalGeminiFilesSpec, LocalOpenAiImageSpec, LocalOpenAiResponsesSpec,
+    CodexResponsesModelCapabilities, ExecutionRuntimeAuthContext, FinalizeStreamRewriteMode,
-    LocalSameFormatProviderFamily, LocalSameFormatProviderSpec, LocalStandardSourceFamily,
+    FormatContext, GeminiClientEmitter, GeminiImageRequestForOpenAi, GeminiProviderState,
-    LocalStandardSourceMode, LocalStandardSpec, LocalSyncReportParts, LocalVideoCreateFamily,
+    KiroToClaudeCliStreamState, LocalCoreSyncErrorKind, LocalGeminiFilesSpec, LocalOpenAiImageSpec,
-    LocalVideoCreateSpec, NormalizedOpenAiImageRequest, OpenAIChatClientEmitter,
+    LocalOpenAiResponsesSpec, LocalSameFormatProviderFamily, LocalSameFormatProviderSpec,
-    OpenAIChatProviderState, OpenAIResponsesClientEmitter, OpenAIResponsesProviderState,
+    LocalStandardSourceFamily, LocalStandardSourceMode, LocalStandardSpec, LocalSyncReportParts,
-    OpenAiImageNormalizeOptions, OpenAiImageOperation, OpenAiImageRequestForGemini,
+    LocalVideoCreateFamily, LocalVideoCreateSpec, NormalizedOpenAiImageRequest,
-    OpenAiImageResponseFormat, OpenAiImageStreamState, OpenAiImageSyncFinalizeProduct,
+    OpenAIChatClientEmitter, OpenAIChatProviderState, OpenAIResponsesClientEmitter,
+    OpenAIResponsesProviderState, OpenAiImageNormalizeOptions, OpenAiImageOperation,
+    OpenAiImageRequestForGemini, OpenAiImageResponseFormat, OpenAiImageStreamState,
+    OpenAiImageSyncFinalizeProduct, OpenAiProviderRequestFinalization,
     ProviderAdaptationDescriptor, ProviderAdaptationSurface, ProviderPrivateStreamNormalizer,
-    RequestConversionKind, StandardCrossFormatSyncProduct, StandardSyncFinalizeNormalizedProduct,
+    ReasoningEffort, RequestConversionKind, ResponseHistoryRecord, ServiceTier,
+    StandardCrossFormatSyncProduct, StandardSyncFinalizeNormalizedProduct,
     StreamingStandardFormatMatrix, SyncChatResponseConversionKind, SyncCliResponseConversionKind,
     SyncToStreamBridgeOutcome, ANTIGRAVITY_V1INTERNAL_ENVELOPE_NAME, CLAUDE_CHAT_STREAM_PLAN_KIND,
     CLAUDE_CHAT_STREAM_SUCCESS_REPORT_KIND, CLAUDE_CHAT_SYNC_ERROR_REPORT_KIND,
@@ -110,14 +133,15 @@ pub(crate) use aether_ai_formats::api::{
     CLAUDE_CHAT_SYNC_SUCCESS_REPORT_KIND, CLAUDE_CLI_STREAM_PLAN_KIND,
     CLAUDE_CLI_STREAM_SUCCESS_REPORT_KIND, CLAUDE_CLI_SYNC_ERROR_REPORT_KIND,
     CLAUDE_CLI_SYNC_FINALIZE_REPORT_KIND, CLAUDE_CLI_SYNC_PLAN_KIND,
-    CLAUDE_CLI_SYNC_SUCCESS_REPORT_KIND, CODEX_OPENAI_IMAGE_DEFAULT_MODEL,
+    CLAUDE_CLI_SYNC_SUCCESS_REPORT_KIND, CLAUDE_COUNT_TOKENS_SYNC_PLAN_KIND,
-    CODEX_OPENAI_IMAGE_DEFAULT_OUTPUT_FORMAT, CODEX_OPENAI_IMAGE_DEFAULT_VARIATION_MODEL,
+    CODEX_OPENAI_IMAGE_DEFAULT_MODEL, CODEX_OPENAI_IMAGE_DEFAULT_OUTPUT_FORMAT,
-    CODEX_OPENAI_IMAGE_DEFAULT_VARIATION_PROMPT, CODEX_OPENAI_IMAGE_INTERNAL_MODEL,
+    CODEX_OPENAI_IMAGE_DEFAULT_VARIATION_MODEL, CODEX_OPENAI_IMAGE_DEFAULT_VARIATION_PROMPT,
-    EXECUTION_RUNTIME_STREAM_ACTION, EXECUTION_RUNTIME_STREAM_DECISION_ACTION,
+    CODEX_OPENAI_IMAGE_INTERNAL_MODEL, EXECUTION_RUNTIME_STREAM_ACTION,
-    EXECUTION_RUNTIME_SYNC_ACTION, EXECUTION_RUNTIME_SYNC_DECISION_ACTION,
+    EXECUTION_RUNTIME_STREAM_DECISION_ACTION, EXECUTION_RUNTIME_SYNC_ACTION,
-    GEMINI_CHAT_STREAM_PLAN_KIND, GEMINI_CHAT_STREAM_SUCCESS_REPORT_KIND,
+    EXECUTION_RUNTIME_SYNC_DECISION_ACTION, GEMINI_CHAT_STREAM_PLAN_KIND,
-    GEMINI_CHAT_SYNC_ERROR_REPORT_KIND, GEMINI_CHAT_SYNC_FINALIZE_REPORT_KIND,
+    GEMINI_CHAT_STREAM_SUCCESS_REPORT_KIND, GEMINI_CHAT_SYNC_ERROR_REPORT_KIND,
-    GEMINI_CHAT_SYNC_PLAN_KIND, GEMINI_CHAT_SYNC_SUCCESS_REPORT_KIND, GEMINI_CLI_STREAM_PLAN_KIND,
+    GEMINI_CHAT_SYNC_FINALIZE_REPORT_KIND, GEMINI_CHAT_SYNC_PLAN_KIND,
+    GEMINI_CHAT_SYNC_SUCCESS_REPORT_KIND, GEMINI_CLI_STREAM_PLAN_KIND,
     GEMINI_CLI_STREAM_SUCCESS_REPORT_KIND, GEMINI_CLI_SYNC_ERROR_REPORT_KIND,
     GEMINI_CLI_SYNC_FINALIZE_REPORT_KIND, GEMINI_CLI_SYNC_PLAN_KIND,
     GEMINI_CLI_SYNC_SUCCESS_REPORT_KIND, GEMINI_CLI_V1INTERNAL_ENVELOPE_NAME,
@@ -125,22 +149,53 @@ pub(crate) use aether_ai_formats::api::{
     GEMINI_FILES_DELETE_PLAN_KIND, GEMINI_FILES_DOWNLOAD_PLAN_KIND, GEMINI_FILES_GET_PLAN_KIND,
     GEMINI_FILES_LIST_PLAN_KIND, GEMINI_FILES_UPLOAD_PLAN_KIND, GEMINI_VIDEO_CANCEL_SYNC_PLAN_KIND,
     GEMINI_VIDEO_CREATE_SYNC_FINALIZE_REPORT_KIND, GEMINI_VIDEO_CREATE_SYNC_PLAN_KIND,
-    KIRO_ENVELOPE_NAME, KIRO_MAX_THINKING_BUFFER, OPENAI_CHAT_STREAM_PLAN_KIND,
+    KIRO_ENVELOPE_NAME, KIRO_MAX_THINKING_BUFFER, MODEL_DIRECTIVE_API_FORMATS,
-    OPENAI_CHAT_STREAM_SUCCESS_REPORT_KIND, OPENAI_CHAT_SYNC_ERROR_REPORT_KIND,
+    OPENAI_CHAT_STREAM_PLAN_KIND, OPENAI_CHAT_STREAM_SUCCESS_REPORT_KIND,
-    OPENAI_CHAT_SYNC_FINALIZE_REPORT_KIND, OPENAI_CHAT_SYNC_PLAN_KIND,
+    OPENAI_CHAT_SYNC_ERROR_REPORT_KIND, OPENAI_CHAT_SYNC_FINALIZE_REPORT_KIND,
-    OPENAI_CHAT_SYNC_SUCCESS_REPORT_KIND, OPENAI_EMBEDDING_SYNC_PLAN_KIND,
+    OPENAI_CHAT_SYNC_PLAN_KIND, OPENAI_CHAT_SYNC_SUCCESS_REPORT_KIND,
-    OPENAI_IMAGE_STREAM_PLAN_KIND, OPENAI_IMAGE_STREAM_SUCCESS_REPORT_KIND,
+    OPENAI_EMBEDDING_SYNC_PLAN_KIND, OPENAI_IMAGE_STREAM_PLAN_KIND,
-    OPENAI_IMAGE_SYNC_FINALIZE_REPORT_KIND, OPENAI_IMAGE_SYNC_PLAN_KIND,
+    OPENAI_IMAGE_STREAM_SUCCESS_REPORT_KIND, OPENAI_IMAGE_SYNC_FINALIZE_REPORT_KIND,
-    OPENAI_IMAGE_SYNC_SUCCESS_REPORT_KIND, OPENAI_RERANK_SYNC_PLAN_KIND,
+    OPENAI_IMAGE_SYNC_PLAN_KIND, OPENAI_IMAGE_SYNC_SUCCESS_REPORT_KIND,
-    OPENAI_RESPONSES_COMPACT_STREAM_PLAN_KIND, OPENAI_RESPONSES_COMPACT_STREAM_SUCCESS_REPORT_KIND,
+    OPENAI_RERANK_SYNC_PLAN_KIND, OPENAI_RESPONSES_COMPACT_STREAM_PLAN_KIND,
+    OPENAI_RESPONSES_COMPACT_STREAM_SUCCESS_REPORT_KIND,
     OPENAI_RESPONSES_COMPACT_SYNC_ERROR_REPORT_KIND,
     OPENAI_RESPONSES_COMPACT_SYNC_FINALIZE_REPORT_KIND, OPENAI_RESPONSES_COMPACT_SYNC_PLAN_KIND,
     OPENAI_RESPONSES_COMPACT_SYNC_SUCCESS_REPORT_KIND, OPENAI_RESPONSES_STREAM_PLAN_KIND,
     OPENAI_RESPONSES_STREAM_SUCCESS_REPORT_KIND, OPENAI_RESPONSES_SYNC_ERROR_REPORT_KIND,
     OPENAI_RESPONSES_SYNC_FINALIZE_REPORT_KIND, OPENAI_RESPONSES_SYNC_PLAN_KIND,
-    OPENAI_RESPONSES_SYNC_SUCCESS_REPORT_KIND, OPENAI_VIDEO_CANCEL_SYNC_PLAN_KIND,
+    OPENAI_RESPONSES_SYNC_SUCCESS_REPORT_KIND, OPENAI_SEARCH_SYNC_PLAN_KIND,
+    OPENAI_SEARCH_SYNC_SUCCESS_REPORT_KIND, OPENAI_VIDEO_CANCEL_SYNC_PLAN_KIND,
     OPENAI_VIDEO_CONTENT_PLAN_KIND, OPENAI_VIDEO_CREATE_SYNC_FINALIZE_REPORT_KIND,
     OPENAI_VIDEO_CREATE_SYNC_PLAN_KIND, OPENAI_VIDEO_DELETE_SYNC_PLAN_KIND,
     OPENAI_VIDEO_REMIX_SYNC_PLAN_KIND,
 };
-pub(crate) use aether_ai_formats::{is_embedding_api_format, is_rerank_api_format};
+pub(crate) use aether_ai_formats::{
+    api_format_defaults_to_client_error_failover, api_format_defaults_to_non_stream,
+    api_format_permission_covers, intersect_api_format_allowed_lists, is_embedding_api_format,
+    is_rerank_api_format, openai_responses_request_operation,
+    openai_responses_synthetic_reasoning_item_id,
+    strip_incompatible_openai_responses_reasoning_items, ApiOperation, ClientSurface,
+};
+
+pub(crate) fn plan_kind_matches_api_operation(
+    plan_kind: &str,
+    require_streaming: bool,
+    expected_operation: Option<ApiOperation>,
+) -> bool {
+    let Some(expected_operation) = expected_operation else {
+        return true;
+    };
+    if expected_operation == ApiOperation::OpenAiResponsesCompact {
+        return if require_streaming {
+            plan_kind == OPENAI_RESPONSES_COMPACT_STREAM_PLAN_KIND
+        } else {
+            plan_kind == OPENAI_RESPONSES_COMPACT_SYNC_PLAN_KIND
+        };
+    }
+    let resolved_operation = if require_streaming {
+        resolve_local_same_format_stream_spec(plan_kind).and_then(|spec| spec.operation)
+    } else {
+        resolve_local_same_format_sync_spec(plan_kind).and_then(|spec| spec.operation)
+    };
+    resolved_operation == Some(expected_operation)
+}

apps/aether-gateway/src/ai_serving/response_history.rs

@@ -0,0 +1,96 @@
+use crate::ai_serving::{
+    hydrate_response_history, normalize_api_format_alias, record_converted_response_history,
+    response_history_is_loaded, response_history_storage_key, ResponseHistoryRecord,
+};
+use aether_runtime_state::RuntimeState;
+use serde_json::Value;
+use tracing::warn;
+
+use crate::GatewayError;
+
+pub(crate) async fn hydrate_openai_response_history(
+    runtime_state: &RuntimeState,
+    request: &Value,
+    client_api_format: &str,
+    provider_api_format: &str,
+    history_scope: &str,
+) -> Result<(), GatewayError> {
+    if normalize_api_format_alias(client_api_format) != "openai:responses"
+        || normalize_api_format_alias(provider_api_format) != "openai:chat"
+    {
+        return Ok(());
+    }
+    let Some(previous_response_id) = request
+        .get("previous_response_id")
+        .and_then(Value::as_str)
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+    else {
+        return Ok(());
+    };
+    if response_history_is_loaded(previous_response_id, Some(history_scope)) {
+        return Ok(());
+    }
+
+    let storage_key = response_history_storage_key(previous_response_id, Some(history_scope));
+    let payload = runtime_state.kv_get(&storage_key).await.map_err(|error| {
+        warn!(
+            event_name = "openai_response_history_read_failed",
+            log_type = "ops",
+            backend = runtime_state.backend_kind().as_str(),
+            error = ?error,
+            "gateway failed to read shared OpenAI response history"
+        );
+        GatewayError::Internal("OpenAI response history lookup failed".to_string())
+    })?;
+    let Some(payload) = payload else {
+        return Ok(());
+    };
+    if let Err(error) =
+        hydrate_response_history(previous_response_id, Some(history_scope), &payload)
+    {
+        let _ = runtime_state.kv_delete(&storage_key).await;
+        warn!(
+            event_name = "openai_response_history_invalid",
+            log_type = "ops",
+            backend = runtime_state.backend_kind().as_str(),
+            error = %error,
+            "gateway rejected invalid shared OpenAI response history"
+        );
+        return Err(GatewayError::Internal(
+            "OpenAI response history validation failed".to_string(),
+        ));
+    }
+    Ok(())
+}
+
+pub(crate) async fn persist_response_history_record(
+    runtime_state: &RuntimeState,
+    record: ResponseHistoryRecord,
+) {
+    if let Err(error) = runtime_state
+        .kv_set(&record.storage_key, record.payload, Some(record.ttl))
+        .await
+    {
+        warn!(
+            event_name = "openai_response_history_write_failed",
+            log_type = "ops",
+            backend = runtime_state.backend_kind().as_str(),
+            error = ?error,
+            "gateway failed to persist shared OpenAI response history"
+        );
+    }
+}
+
+pub(crate) async fn persist_converted_response_history(
+    runtime_state: &RuntimeState,
+    report_context: &Value,
+    response: Option<&Value>,
+) {
+    let Some(response) = response else {
+        return;
+    };
+    if let Some(record) = record_converted_response_history(report_context, response) {
+        persist_response_history_record(runtime_state, record).await;
+    }
+}

apps/aether-gateway/src/ai_serving/transport.rs

@@ -59,8 +59,8 @@ pub(crate) mod windsurf {
 }
 
 pub(crate) use aether_provider_transport::{
-    append_transport_diagnostics_to_value, apply_local_body_rules,
+    append_transport_diagnostics_to_value, apply_local_auth_config_header_overrides,
-    apply_local_body_rules_with_request_headers, apply_local_header_rules,
+    apply_local_body_rules, apply_local_body_rules_with_request_headers, apply_local_header_rules,
     apply_local_header_rules_with_request_headers, apply_standard_provider_request_body_rules,
     apply_standard_provider_request_body_rules_with_request_headers,
     apply_transport_request_body_semantics, body_rules_are_locally_supported,
@@ -82,10 +82,11 @@ pub(crate) use aether_provider_transport::{
     build_windsurf_cascade_headers, build_windsurf_cascade_request_body,
     build_windsurf_cascade_upstream_url, candidate_common_transport_skip_reason,
     candidate_transport_pair_skip_reason, classify_same_format_provider_request_behavior,
-    ensure_upstream_auth_header, gemini_files_transport_unsupported_reason,
+    classify_same_format_provider_request_behavior_for_operation,
-    header_rules_are_locally_supported, header_rules_have_enabled_rules,
+    enforce_same_format_provider_api_operation_body_policy, ensure_upstream_auth_header,
-    is_gemini_cli_provider_transport, is_windsurf_provider_transport,
+    gemini_files_transport_unsupported_reason, header_rules_are_locally_supported,
-    local_gemini_transport_unsupported_reason_with_network,
+    header_rules_have_enabled_rules, is_gemini_cli_provider_transport,
+    is_windsurf_provider_transport, local_gemini_transport_unsupported_reason_with_network,
     local_openai_chat_transport_unsupported_reason,
     local_standard_transport_unsupported_reason_with_network,
     local_windsurf_request_transport_unsupported_reason_with_network,
@@ -93,22 +94,22 @@ pub(crate) use aether_provider_transport::{
     request_conversion_enabled_for_transport, request_conversion_transport_supported,
     request_conversion_transport_unsupported_reason, request_pair_allowed_for_transport,
     request_pair_direct_auth, request_pair_transport_unsupported_reason,
-    resolve_gemini_cli_project_id, resolve_gemini_files_auth, resolve_grok_session_auth,
+    resolve_anthropic_compatibility_profile, resolve_gemini_cli_project_id,
-    resolve_local_gemini_cli_request_auth, resolve_openai_image_auth,
+    resolve_gemini_files_auth, resolve_grok_session_auth, resolve_local_gemini_cli_request_auth,
-    resolve_same_format_provider_direct_auth, resolve_transport_execution_timeouts,
+    resolve_openai_image_auth, resolve_same_format_provider_direct_auth,
-    resolve_transport_profile, resolve_transport_proxy_snapshot,
+    resolve_transport_execution_timeouts, resolve_transport_profile,
-    resolve_transport_proxy_snapshot_with_tunnel_affinity, resolve_video_create_auth,
+    resolve_transport_proxy_snapshot, resolve_transport_proxy_snapshot_with_tunnel_affinity,
-    same_format_provider_transport_supported, same_format_provider_transport_unsupported_reason,
+    resolve_video_create_auth, same_format_provider_transport_supported,
-    should_skip_upstream_passthrough_header, should_try_same_format_provider_oauth_auth,
+    same_format_provider_transport_unsupported_reason, should_skip_upstream_passthrough_header,
-    supports_local_gemini_transport_with_network,
+    should_try_same_format_provider_oauth_auth, supports_local_gemini_transport_with_network,
     supports_local_generic_oauth_request_auth_resolution,
     supports_local_oauth_request_auth_resolution, transport_proxy_is_locally_supported,
-    video_create_transport_unsupported_reason, CandidateTransportPolicyFacts,
+    transport_supports_api_operation, video_create_transport_unsupported_reason,
-    GatewayProviderTransportSnapshot, GeminiCliRequestAuth, GeminiCliRequestAuthSupport,
+    AnthropicCompatibilityProfile, CandidateTransportPolicyFacts, GatewayProviderTransportSnapshot,
-    GeminiCliRequestAuthUnsupportedReason, GeminiCliRequestEnvelopeSupport,
+    GeminiCliRequestAuth, GeminiCliRequestAuthSupport, GeminiCliRequestAuthUnsupportedReason,
-    GeminiFilesHeadersInput, GeminiFilesRequestBodyError, GeminiFilesRequestBodyParts,
+    GeminiCliRequestEnvelopeSupport, GeminiFilesHeadersInput, GeminiFilesRequestBodyError,
-    GrokHeaderInput, LocalResolvedOAuthRequestAuth, ProviderOpenAiImageHeadersInput,
+    GeminiFilesRequestBodyParts, GrokHeaderInput, LocalResolvedOAuthRequestAuth,
-    ProviderVideoCreateFamily, ProviderVideoCreateHeadersInput,
+    ProviderOpenAiImageHeadersInput, ProviderVideoCreateFamily, ProviderVideoCreateHeadersInput,
     SameFormatProviderCompatibilityEdit, SameFormatProviderCompatibilityEditAction,
     SameFormatProviderFamily, SameFormatProviderHeadersInput, SameFormatProviderRequestBehavior,
     SameFormatProviderRequestBehaviorParams, SameFormatProviderRequestBodyInput,

apps/aether-gateway/src/allocator_metrics.rs

@@ -0,0 +1,218 @@
+use aether_runtime::{MetricKind, MetricSample};
+
+pub(crate) fn gateway_allocator_metric_samples() -> Vec<MetricSample> {
+    match allocator_snapshot() {
+        Some(snapshot) => snapshot.to_metric_samples(),
+        None => unavailable_metric_samples(),
+    }
+}
+
+#[derive(Debug, Clone, Copy, Default)]
+struct AllocatorSnapshot {
+    allocated_bytes: u64,
+    active_bytes: u64,
+    resident_bytes: u64,
+    mapped_bytes: u64,
+    retained_bytes: u64,
+    metadata_bytes: u64,
+}
+
+impl AllocatorSnapshot {
+    fn to_metric_samples(self) -> Vec<MetricSample> {
+        vec![
+            gauge(
+                "gateway_allocator_observability_available",
+                "Whether gateway allocator heap metrics were available for this scrape.",
+                1,
+            ),
+            gauge(
+                "gateway_allocator_allocated_bytes",
+                "Bytes currently allocated by the gateway allocator.",
+                self.allocated_bytes,
+            ),
+            gauge(
+                "gateway_allocator_active_bytes",
+                "Bytes in active pages managed by the gateway allocator.",
+                self.active_bytes,
+            ),
+            gauge(
+                "gateway_allocator_resident_bytes",
+                "Bytes resident in physical memory for the gateway allocator.",
+                self.resident_bytes,
+            ),
+            gauge(
+                "gateway_allocator_mapped_bytes",
+                "Bytes mapped by the gateway allocator.",
+                self.mapped_bytes,
+            ),
+            gauge(
+                "gateway_allocator_retained_bytes",
+                "Bytes retained by the gateway allocator for future use.",
+                self.retained_bytes,
+            ),
+            gauge(
+                "gateway_allocator_metadata_bytes",
+                "Bytes used for allocator metadata.",
+                self.metadata_bytes,
+            ),
+            gauge(
+                "gateway_allocator_active_to_allocated_basis_points",
+                "Active allocator bytes divided by allocated bytes in basis points.",
+                ratio_basis_points(self.active_bytes, self.allocated_bytes),
+            ),
+            gauge(
+                "gateway_allocator_resident_to_allocated_basis_points",
+                "Resident allocator bytes divided by allocated bytes in basis points.",
+                ratio_basis_points(self.resident_bytes, self.allocated_bytes),
+            ),
+        ]
+    }
+}
+
+fn unavailable_metric_samples() -> Vec<MetricSample> {
+    vec![
+        gauge(
+            "gateway_allocator_observability_available",
+            "Whether gateway allocator heap metrics were available for this scrape.",
+            0,
+        ),
+        gauge(
+            "gateway_allocator_allocated_bytes",
+            "Bytes currently allocated by the gateway allocator.",
+            0,
+        ),
+        gauge(
+            "gateway_allocator_active_bytes",
+            "Bytes in active pages managed by the gateway allocator.",
+            0,
+        ),
+        gauge(
+            "gateway_allocator_resident_bytes",
+            "Bytes resident in physical memory for the gateway allocator.",
+            0,
+        ),
+        gauge(
+            "gateway_allocator_mapped_bytes",
+            "Bytes mapped by the gateway allocator.",
+            0,
+        ),
+        gauge(
+            "gateway_allocator_retained_bytes",
+            "Bytes retained by the gateway allocator for future use.",
+            0,
+        ),
+        gauge(
+            "gateway_allocator_metadata_bytes",
+            "Bytes used for allocator metadata.",
+            0,
+        ),
+        gauge(
+            "gateway_allocator_active_to_allocated_basis_points",
+            "Active allocator bytes divided by allocated bytes in basis points.",
+            0,
+        ),
+        gauge(
+            "gateway_allocator_resident_to_allocated_basis_points",
+            "Resident allocator bytes divided by allocated bytes in basis points.",
+            0,
+        ),
+    ]
+}
+
+#[cfg(all(feature = "jemalloc", not(target_env = "msvc")))]
+fn allocator_snapshot() -> Option<AllocatorSnapshot> {
+    refresh_jemalloc_epoch()?;
+    Some(AllocatorSnapshot {
+        allocated_bytes: read_jemalloc_stat("stats.allocated\0")?,
+        active_bytes: read_jemalloc_stat("stats.active\0")?,
+        resident_bytes: read_jemalloc_stat("stats.resident\0")?,
+        mapped_bytes: read_jemalloc_stat("stats.mapped\0")?,
+        retained_bytes: read_jemalloc_stat("stats.retained\0")?,
+        metadata_bytes: read_jemalloc_stat("stats.metadata\0")?,
+    })
+}
+
+#[cfg(not(all(feature = "jemalloc", not(target_env = "msvc"))))]
+fn allocator_snapshot() -> Option<AllocatorSnapshot> {
+    None
+}
+
+#[cfg(all(feature = "jemalloc", not(target_env = "msvc")))]
+fn refresh_jemalloc_epoch() -> Option<()> {
+    let mut epoch = 1_u64;
+    let result = unsafe {
+        tikv_jemalloc_sys::mallctl(
+            c"epoch".as_ptr(),
+            std::ptr::null_mut(),
+            std::ptr::null_mut(),
+            (&mut epoch as *mut u64).cast(),
+            std::mem::size_of::<u64>(),
+        )
+    };
+    if result == 0 {
+        Some(())
+    } else {
+        None
+    }
+}
+
+#[cfg(all(feature = "jemalloc", not(target_env = "msvc")))]
+fn read_jemalloc_stat(name: &str) -> Option<u64> {
+    let mut value = 0_usize;
+    let mut size = std::mem::size_of::<usize>();
+    let result = unsafe {
+        tikv_jemalloc_sys::mallctl(
+            name.as_ptr().cast(),
+            (&mut value as *mut usize).cast(),
+            &mut size,
+            std::ptr::null_mut(),
+            0,
+        )
+    };
+    if result == 0 {
+        Some(u64_from_usize(value))
+    } else {
+        None
+    }
+}
+
+fn gauge(name: &'static str, help: &'static str, value: u64) -> MetricSample {
+    MetricSample::new(name, help, MetricKind::Gauge, value)
+}
+
+fn ratio_basis_points(numerator: u64, denominator: u64) -> u64 {
+    if denominator == 0 {
+        return 0;
+    }
+    numerator.saturating_mul(10_000) / denominator
+}
+
+fn u64_from_usize(value: usize) -> u64 {
+    u64::try_from(value).unwrap_or(u64::MAX)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{gateway_allocator_metric_samples, ratio_basis_points};
+
+    #[test]
+    fn renders_allocator_metric_samples() {
+        let samples = gateway_allocator_metric_samples();
+
+        assert!(samples
+            .iter()
+            .any(|sample| sample.name == "gateway_allocator_observability_available"));
+        assert!(samples
+            .iter()
+            .any(|sample| sample.name == "gateway_allocator_allocated_bytes"));
+        assert!(samples
+            .iter()
+            .any(|sample| sample.name == "gateway_allocator_active_to_allocated_basis_points"));
+    }
+
+    #[test]
+    fn computes_ratio_basis_points() {
+        assert_eq!(ratio_basis_points(150, 100), 15_000);
+        assert_eq!(ratio_basis_points(1, 0), 0);
+    }
+}

apps/aether-gateway/src/api/ai/gemini.rs

@@ -1,6 +1,7 @@
 pub(crate) fn normalized_signature(api_format: &str) -> Option<&'static str> {
     match crate::ai_serving::normalize_api_format_alias(api_format).as_str() {
         "gemini:generate_content" => Some("gemini:generate_content"),
+        "gemini:interactions" => Some("gemini:interactions"),
         "gemini:embedding" => Some("gemini:embedding"),
         "gemini:video" => Some("gemini:video"),
         "gemini:files" => Some("gemini:files"),
@@ -11,6 +12,7 @@ pub(crate) fn normalized_signature(api_format: &str) -> Option<&'static str> {
 pub(crate) fn local_path(api_format: &str) -> Option<&'static str> {
     match crate::ai_serving::normalize_api_format_alias(api_format).as_str() {
         "gemini" | "gemini:generate_content" => Some("/v1beta/models/{model}:{action}"),
+        "gemini:interactions" => Some("/v1/interactions"),
         "gemini:embedding" => Some("/v1beta/models/{model}:{action}"),
         "gemini:video" => Some("/v1beta/models/{model}:predictLongRunning"),
         "gemini:files" => Some("/v1beta/files"),

apps/aether-gateway/src/api/ai/openai.rs

@@ -5,6 +5,7 @@ pub(crate) fn normalized_signature(api_format: &str) -> Option<&'static str> {
         "openai:rerank" => Some("openai:rerank"),
         "openai:responses" => Some("openai:responses"),
         "openai:responses:compact" => Some("openai:responses:compact"),
+        "openai:search" => Some("openai:search"),
         "openai:image" => Some("openai:image"),
         "openai:video" => Some("openai:video"),
         _ => None,
@@ -18,6 +19,7 @@ pub(crate) fn local_path(api_format: &str) -> Option<&'static str> {
         "openai:rerank" => Some("/v1/rerank"),
         "openai:responses" => Some("/v1/responses"),
         "openai:responses:compact" => Some("/v1/responses/compact"),
+        "openai:search" => Some("/v1/alpha/search"),
         "openai:image" => Some("/v1/images/generations"),
         "openai:video" => Some("/v1/videos"),
         _ => None,

apps/aether-gateway/src/api/ai/registry.rs

@@ -1,8 +1,13 @@
+use axum::body::Body;
+use axum::extract::Request;
+use axum::http::{header, HeaderValue, Response, StatusCode};
 use axum::routing::{any, post};
 use axum::Router;
 
 use super::{aliyun, claude, doubao, gemini, jina, openai};
-use crate::{handlers::proxy::proxy_request, state::AppState};
+use crate::api::response::build_local_http_error_response_with_request_path;
+use crate::headers::extract_or_generate_trace_id;
+use crate::{handlers::proxy::proxy_request, state::AppState, GatewayError};
 
 // Router registration patterns live here so AI public ingress has a single mount registry.
 // They intentionally stay separate from manifest-facing route inventories in constants.rs,
@@ -11,22 +16,27 @@ const AI_POST_ROUTE_PATTERNS: &[&str] = &[
     "/v1/chat/completions",
     "/v1/embeddings",
     "/v1/rerank",
-    "/v1/messages",
-    "/v1/messages/count_tokens",
     "/v1/responses",
     "/v1/responses/compact",
+    "/v1/alpha/search",
     "/v1/images/generations",
     "/v1/images/edits",
+    "/v1/interactions",
+    "/v1beta/interactions",
     "/v1internal:loadCodeAssist",
     "/v1internal:fetchAvailableModels",
+    "/v1internal:retrieveUserQuotaSummary",
     "/v1internal:fetchUserInfo",
     "/v1internal:fetchAdminControls",
     "/v1internal:setUserSettings",
     "/v1internal:listExperiments",
     "/v1internal:recordCodeAssistMetrics",
+    "/v1internal:writeTrajectoryAcls",
     "/v1internal:streamGenerateContent",
 ];
 
+const CLAUDE_POST_ROUTE_PATTERNS: &[&str] = &["/v1/messages", "/v1/messages/count_tokens"];
+
 const AI_ANY_ROUTE_PATTERNS: &[&str] = &[
     "/v1/models/{*gemini_path}",
     "/v1beta/models/{*gemini_path}",
@@ -43,12 +53,33 @@ pub(crate) fn mount_ai_routes(mut router: Router<AppState>) -> Router<AppState>
     for path in AI_POST_ROUTE_PATTERNS {
         router = router.route(path, post(proxy_request));
     }
+    for path in CLAUDE_POST_ROUTE_PATTERNS {
+        router = router.route(
+            path,
+            post(proxy_request).fallback(claude_method_not_allowed),
+        );
+    }
     for path in AI_ANY_ROUTE_PATTERNS {
         router = router.route(path, any(proxy_request));
     }
     router
 }
 
+async fn claude_method_not_allowed(request: Request) -> Result<Response<Body>, GatewayError> {
+    let trace_id = extract_or_generate_trace_id(request.headers());
+    let mut response = build_local_http_error_response_with_request_path(
+        &trace_id,
+        None,
+        Some(request.uri().path()),
+        StatusCode::METHOD_NOT_ALLOWED,
+        "Method not allowed",
+    )?;
+    response
+        .headers_mut()
+        .insert(header::ALLOW, HeaderValue::from_static("POST"));
+    Ok(response)
+}
+
 pub(crate) fn public_api_format_local_path(api_format: &str) -> &'static str {
     let normalized = api_format.trim().to_ascii_lowercase();
     openai::local_path(&normalized)
@@ -95,6 +126,12 @@ mod tests {
     fn supports_data_api_endpoint_signatures_and_public_paths() {
         for (api_format, family, kind, path) in [
             ("openai:embedding", "openai", "embedding", "/v1/embeddings"),
+            (
+                "gemini:interactions",
+                "gemini",
+                "interactions",
+                "/v1/interactions",
+            ),
             (
                 "gemini:embedding",
                 "gemini",
@@ -110,6 +147,7 @@ mod tests {
                 "/api/v1/services/embeddings/multimodal-embedding/multimodal-embedding",
             ),
             ("openai:rerank", "openai", "rerank", "/v1/rerank"),
+            ("openai:search", "openai", "search", "/v1/alpha/search"),
             ("jina:rerank", "jina", "rerank", "/v1/rerank"),
         ] {
             assert_eq!(

apps/aether-gateway/src/api/backend/public.rs

@@ -21,6 +21,7 @@ pub(crate) fn mount_public_support_routes(router: Router<AppState>) -> Router<Ap
         .route("/api/public/global-models", get(proxy_request))
         .route("/api/public/health/api-formats", get(proxy_request))
         .route("/api/public/health/models", get(proxy_request))
+        .route("/api/public/health/related", get(proxy_request))
         .route("/api/modules/auth-status", get(proxy_request))
         .route("/api/capabilities", get(proxy_request))
         .route("/api/capabilities/user-configurable", get(proxy_request))

apps/aether-gateway/src/api/response.rs

@@ -6,6 +6,7 @@ use axum::http::Response;
 use axum::http::StatusCode;
 use serde_json::json;
 
+use crate::ai_serving::{build_core_error_body_for_client_format, LocalCoreSyncErrorKind};
 use crate::constants::*;
 use crate::control::GatewayControlDecision;
 use crate::control::GatewayLocalAuthRejection;
@@ -191,7 +192,7 @@ pub(crate) fn build_local_balance_denied_response(
         Some(remaining) => format!("余额不足（剩余: ${remaining:.2}）"),
         None => "余额不足".to_string(),
     };
-    let payload = json!({
+    let fallback_payload = json!({
         "error": {
             "type": "balance_exceeded",
             "message": message,
@@ -201,6 +202,13 @@ pub(crate) fn build_local_balance_denied_response(
             }
         }
     });
+    let payload = build_local_error_payload(
+        control_decision,
+        None,
+        &message,
+        LocalCoreSyncErrorKind::RateLimit,
+        fallback_payload,
+    );
     let body =
         serde_json::to_vec(&payload).map_err(|err| GatewayError::Internal(err.to_string()))?;
     let headers = BTreeMap::from([("content-type".to_string(), "application/json".to_string())]);
@@ -218,12 +226,20 @@ pub(crate) fn build_local_user_rpm_limited_response(
     control_decision: Option<&GatewayControlDecision>,
     rejection: &FrontdoorUserRpmRejection,
 ) -> Result<Response<Body>, GatewayError> {
-    let payload = json!({
+    let message = "请求过于频繁，请稍后重试";
+    let fallback_payload = json!({
         "error": {
             "type": "rate_limit_exceeded",
-            "message": "请求过于频繁，请稍后重试",
+            "message": message,
         }
     });
+    let payload = build_local_error_payload(
+        control_decision,
+        None,
+        message,
+        LocalCoreSyncErrorKind::RateLimit,
+        fallback_payload,
+    );
     let body =
         serde_json::to_vec(&payload).map_err(|err| GatewayError::Internal(err.to_string()))?;
     let headers = BTreeMap::from([
@@ -248,12 +264,35 @@ pub(crate) fn build_local_http_error_response(
     status_code: StatusCode,
     message: &str,
 ) -> Result<Response<Body>, GatewayError> {
-    let payload = json!({
+    build_local_http_error_response_with_request_path(
+        trace_id,
+        control_decision,
+        None,
+        status_code,
+        message,
+    )
+}
+
+pub(crate) fn build_local_http_error_response_with_request_path(
+    trace_id: &str,
+    control_decision: Option<&GatewayControlDecision>,
+    request_path: Option<&str>,
+    status_code: StatusCode,
+    message: &str,
+) -> Result<Response<Body>, GatewayError> {
+    let fallback_payload = json!({
         "error": {
             "type": "http_error",
             "message": message,
         }
     });
+    let payload = build_local_error_payload(
+        control_decision,
+        request_path,
+        message,
+        local_error_kind_for_status(status_code),
+        fallback_payload,
+    );
     let body =
         serde_json::to_vec(&payload).map_err(|err| GatewayError::Internal(err.to_string()))?;
     let headers = BTreeMap::from([("content-type".to_string(), "application/json".to_string())]);
@@ -329,19 +368,28 @@ pub(crate) fn build_local_auth_rejection_response(
 pub(crate) fn build_local_overloaded_response(
     trace_id: &str,
     control_decision: Option<&GatewayControlDecision>,
+    request_path: Option<&str>,
     gate: &str,
     limit: usize,
 ) -> Result<Response<Body>, GatewayError> {
-    let payload = json!({
+    let message = "服务繁忙，请稍后重试";
+    let fallback_payload = json!({
         "error": {
             "type": "overloaded",
-            "message": "服务繁忙，请稍后重试",
+            "message": message,
             "details": {
                 "gate": gate,
                 "limit": limit,
             }
         }
     });
+    let payload = build_local_error_payload(
+        control_decision,
+        request_path,
+        message,
+        LocalCoreSyncErrorKind::Overloaded,
+        fallback_payload,
+    );
     let body =
         serde_json::to_vec(&payload).map_err(|err| GatewayError::Internal(err.to_string()))?;
     let headers = BTreeMap::from([("content-type".to_string(), "application/json".to_string())]);
@@ -354,10 +402,65 @@ pub(crate) fn build_local_overloaded_response(
     )
 }
 
+fn build_local_error_payload(
+    control_decision: Option<&GatewayControlDecision>,
+    request_path: Option<&str>,
+    message: &str,
+    kind: LocalCoreSyncErrorKind,
+    fallback_payload: serde_json::Value,
+) -> serde_json::Value {
+    if !local_error_uses_claude_format(control_decision, request_path) {
+        return fallback_payload;
+    }
+
+    build_core_error_body_for_client_format("claude:messages", message, None, kind)
+        .unwrap_or(fallback_payload)
+}
+
+fn local_error_uses_claude_format(
+    control_decision: Option<&GatewayControlDecision>,
+    request_path: Option<&str>,
+) -> bool {
+    control_decision.is_some_and(|decision| {
+        decision.route_family.as_deref() == Some("claude")
+            || decision
+                .auth_endpoint_signature
+                .as_deref()
+                .is_some_and(|format| {
+                    crate::ai_serving::normalize_api_format_alias(format)
+                        .eq_ignore_ascii_case("claude:messages")
+                })
+    }) || request_path.is_some_and(|path| {
+        matches!(
+            path.trim_end_matches('/'),
+            "/v1/messages" | "/v1/messages/count_tokens"
+        )
+    })
+}
+
+fn local_error_kind_for_status(status: StatusCode) -> LocalCoreSyncErrorKind {
+    match status.as_u16() {
+        400 | 405 | 422 => LocalCoreSyncErrorKind::InvalidRequest,
+        401 => LocalCoreSyncErrorKind::Authentication,
+        403 => LocalCoreSyncErrorKind::PermissionDenied,
+        404 => LocalCoreSyncErrorKind::NotFound,
+        413 => LocalCoreSyncErrorKind::RequestTooLarge,
+        429 => LocalCoreSyncErrorKind::RateLimit,
+        503 | 529 => LocalCoreSyncErrorKind::Overloaded,
+        _ => LocalCoreSyncErrorKind::ServerError,
+    }
+}
+
 #[cfg(test)]
 mod tests {
-    use super::build_client_response_from_parts;
+    use super::{
-    use axum::body::Body;
+        build_client_response_from_parts, build_local_auth_rejection_response,
+        build_local_http_error_response_with_request_path, build_local_overloaded_response,
+        build_local_user_rpm_limited_response,
+    };
+    use crate::control::{GatewayControlDecision, GatewayLocalAuthRejection};
+    use crate::rate_limit::FrontdoorUserRpmRejection;
+    use axum::body::{to_bytes, Body};
     use std::collections::BTreeMap;
 
     #[test]
@@ -386,4 +489,96 @@ mod tests {
             Some("no")
         );
     }
+
+    fn claude_decision() -> GatewayControlDecision {
+        GatewayControlDecision::synthetic(
+            "/v1/messages",
+            Some("ai_public".to_string()),
+            Some("claude".to_string()),
+            Some("messages".to_string()),
+            Some("claude:messages".to_string()),
+        )
+    }
+
+    async fn response_json(response: http::Response<Body>) -> serde_json::Value {
+        let body = to_bytes(response.into_body(), usize::MAX)
+            .await
+            .expect("response body should read");
+        serde_json::from_slice(&body).expect("response body should be JSON")
+    }
+
+    #[tokio::test]
+    async fn claude_local_errors_use_anthropic_envelopes() {
+        let decision = claude_decision();
+        let invalid_key = build_local_auth_rejection_response(
+            "trace-auth",
+            Some(&decision),
+            &GatewayLocalAuthRejection::InvalidApiKey,
+        )
+        .expect("invalid-key response should build");
+        let invalid_key = response_json(invalid_key).await;
+        assert_eq!(invalid_key["type"], "error");
+        assert_eq!(invalid_key["error"]["type"], "authentication_error");
+
+        let rpm = build_local_user_rpm_limited_response(
+            "trace-rpm",
+            Some(&decision),
+            &FrontdoorUserRpmRejection {
+                scope: "api_key",
+                limit: 1,
+                retry_after: 60,
+            },
+        )
+        .expect("RPM response should build");
+        let rpm = response_json(rpm).await;
+        assert_eq!(rpm["type"], "error");
+        assert_eq!(rpm["error"]["type"], "rate_limit_error");
+
+        let overloaded = build_local_overloaded_response(
+            "trace-overload",
+            None,
+            Some("/v1/messages/count_tokens"),
+            "requests",
+            10,
+        )
+        .expect("overload response should build");
+        let overloaded = response_json(overloaded).await;
+        assert_eq!(overloaded["type"], "error");
+        assert_eq!(overloaded["error"]["type"], "overloaded_error");
+    }
+
+    #[tokio::test]
+    async fn claude_path_shapes_pre_control_http_errors_and_413() {
+        for path in ["/v1/messages", "/v1/messages/count_tokens"] {
+            let forbidden = build_local_http_error_response_with_request_path(
+                "trace-pre-control",
+                None,
+                Some(path),
+                http::StatusCode::FORBIDDEN,
+                "blocked",
+            )
+            .expect("forbidden response should build");
+            let forbidden = response_json(forbidden).await;
+            assert_eq!(forbidden["type"], "error", "path: {path}");
+            assert_eq!(
+                forbidden["error"]["type"], "permission_error",
+                "path: {path}"
+            );
+
+            let too_large = build_local_http_error_response_with_request_path(
+                "trace-too-large",
+                None,
+                Some(path),
+                http::StatusCode::PAYLOAD_TOO_LARGE,
+                "too large",
+            )
+            .expect("payload-too-large response should build");
+            let too_large = response_json(too_large).await;
+            assert_eq!(too_large["type"], "error", "path: {path}");
+            assert_eq!(
+                too_large["error"]["type"], "request_too_large",
+                "path: {path}"
+            );
+        }
+    }
 }

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

@@ -144,34 +144,38 @@ pub(crate) fn spawn_video_task_poller(state: AppState) -> Option<JoinHandle<()>>
         return None;
     }
 
-    Some(tokio::spawn(async move {
+    Some(crate::task_runtime::spawn_singleton_worker(
-        let mut interval = tokio::time::interval(config.interval);
+        state,
-        interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
+        crate::task_runtime::TASK_KEY_VIDEO_TASK_POLLER,
-        interval.tick().await;
+        move |state| async move {
-        let mut deferred_since = None;
+            let mut interval = tokio::time::interval(config.interval);
-        loop {
+            interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
             interval.tick().await;
-            if state
+            let mut deferred_since = None;
-                .data
+            loop {
-                .should_defer_maintenance_for_database_pool_pressure(&mut deferred_since)
+                interval.tick().await;
-            {
+                if state
-                debug!(
+                    .data
-                    event_name = "video_task_poller_deferred",
+                    .should_defer_maintenance_for_database_pool_pressure(&mut deferred_since)
-                    log_type = "event",
+                {
-                    "gateway video task poller deferred because database pool has no idle reserve"
+                    debug!(
-                );
+                        event_name = "video_task_poller_deferred",
-                continue;
+                        log_type = "event",
+                        "gateway video task poller deferred because database pool has no idle reserve"
+                    );
+                    continue;
+                }
+                if let Err(err) = poll_video_tasks_once(&state, config.batch_size).await {
+                    warn!(
+                        event_name = "video_task_poller_tick_failed",
+                        log_type = "event",
+                        error = ?err,
+                        "gateway video task poller tick failed"
+                    );
+                }
             }
-            if let Err(err) = poll_video_tasks_once(&state, config.batch_size).await {
+        },
-                warn!(
+    ))
-                    event_name = "video_task_poller_tick_failed",
-                    log_type = "event",
-                    error = ?err,
-                    "gateway video task poller tick failed"
-                );
-            }
-        }
-    }))
 }
 
 async fn fetch_video_task_refresh_attempt(

apps/aether-gateway/src/backup/config.rs

@@ -11,6 +11,7 @@ pub(crate) struct S3BackupConfig {
     pub(crate) scope: BackupScope,
     pub(crate) endpoint: String,
     pub(crate) region: String,
+    pub(crate) user_agent: String,
     pub(crate) bucket: String,
     pub(crate) prefix: String,
     pub(crate) access_key_id: String,
@@ -104,6 +105,8 @@ impl S3BackupConfig {
             endpoint,
             region: optional_string(entries, "backup_s3_region")?
                 .unwrap_or_else(|| "auto".to_string()),
+            user_agent: optional_string(entries, "backup_s3_user_agent")?
+                .unwrap_or_else(|| "rclone/v1.68.0".to_string()),
             bucket,
             prefix: optional_string(entries, "backup_s3_prefix")?
                 .unwrap_or_else(|| "aether/backups/".to_string()),
@@ -179,6 +182,7 @@ fn config_label(key: &str) -> &str {
     match key {
         "backup_s3_endpoint" => "Endpoint（S3 地址）",
         "backup_s3_region" => "Region（S3 区域）",
+        "backup_s3_user_agent" => "User-Agent（请求头标识）",
         "backup_s3_bucket" => "Bucket（存储桶）",
         "backup_s3_prefix" => "Prefix（备份前缀）",
         "backup_s3_access_key_id" => "Access Key ID（访问密钥 ID）",
@@ -384,6 +388,7 @@ mod tests {
 
         assert_eq!(config.scope, BackupScope::Data);
         assert_eq!(config.region, "auto");
+        assert_eq!(config.user_agent, "rclone/v1.68.0");
         assert_eq!(config.prefix, "aether/backups/");
         assert_eq!(config.path_style, true);
         assert_eq!(config.compression, "zstd");

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

@@ -132,6 +132,7 @@ mod tests {
             scope,
             endpoint: "https://example.com".to_string(),
             region: "auto".to_string(),
+            user_agent: "rclone/v1.68.0".to_string(),
             bucket: "aether-backups".to_string(),
             prefix: "prod/".to_string(),
             access_key_id: "test-access-key".to_string(),

apps/aether-gateway/src/backup/store.rs

@@ -6,7 +6,8 @@ use bytes::Bytes;
 use futures_util::TryStreamExt;
 use object_store::aws::AmazonS3Builder;
 use object_store::path::Path;
-use object_store::ObjectStore;
+use object_store::{ClientOptions, ObjectStore};
+use reqwest::header::HeaderValue;
 use tokio::sync::RwLock;
 
 use super::config::S3BackupConfig;
@@ -84,7 +85,19 @@ pub(crate) struct ObjectStoreS3BackupStore {
 
 impl ObjectStoreS3BackupStore {
     pub(crate) fn from_config(config: &S3BackupConfig) -> Result<Self, BackupStoreError> {
+        // 部分 S3 兼容网关（如中国科技云 s3.cstcloud.cn）按 User-Agent 放行请求，
+        // object_store 默认 UA 会被拒，因此允许自定义 User-Agent。
+        let client_options = if config.user_agent.trim().is_empty() {
+            ClientOptions::new()
+        } else {
+            ClientOptions::new().with_user_agent(
+                HeaderValue::from_str(config.user_agent.trim()).map_err(|error| {
+                    BackupStoreError::new(format!("S3 备份 User-Agent 配置无效: {error}"))
+                })?,
+            )
+        };
         let store = AmazonS3Builder::new()
+            .with_client_options(client_options)
             .with_endpoint(config.endpoint.clone())
             .with_region(config.region.clone())
             .with_bucket_name(config.bucket.clone())

apps/aether-gateway/src/backup/task.rs

@@ -31,6 +31,7 @@ const S3_BACKUP_CONFIG_KEYS: &[&str] = &[
     "backup_s3_scope",
     "backup_s3_endpoint",
     "backup_s3_region",
+    "backup_s3_user_agent",
     "backup_s3_bucket",
     "backup_s3_prefix",
     "backup_s3_access_key_id",