diff --git a/.github/workflows/run-sweep.yml b/.github/workflows/run-sweep.yml
index b46c4c5901..33960c9331 100644
--- a/.github/workflows/run-sweep.yml
+++ b/.github/workflows/run-sweep.yml
@@ -249,7 +249,7 @@ jobs:
                       --model 'claude-fable-5'
                       --max-turns 8
                       --allowedTools "Read,Glob,Grep,Bash(git diff:*)"
-                      --json-schema '{"type":"object","properties":{"criteria":{"type":"array","items":{"type":"string","enum":["multi-node","agentic","eval-only","fp4","mtp","eagle","eagle3","sglang","vllm","dynamo-sglang","dynamo-vllm","glm5","glm5.1","kimik2.5","kimik3","dsv4","minimaxm3","qwen3.5","dsr1","checklist-complete","patchwork"]},"uniqueItems":true},"reason":{"type":"string"}},"required":["criteria","reason"]}'
+                      --json-schema '{"type":"object","properties":{"criteria":{"type":"array","items":{"type":"string","enum":["multi-node","agentic","eval-only","fp4","mtp","eagle","eagle3","sglang","vllm","sgl-router","vllm-router","dynamo-sglang","dynamo-vllm","glm5","glm5.1","kimik2.5","kimik3","dsv4","minimaxm3","qwen3.5","dsr1","checklist-complete","patchwork"]},"uniqueItems":true},"reason":{"type":"string"}},"required":["criteria","reason"]}'
                   prompt: |
                       Inspect this Actions run's change range.
 
@@ -264,7 +264,7 @@ jobs:
                       - eval-only: evaluation without throughput measurement
                       - fp4: FP4 precision
                       - mtp, eagle, eagle3: speculative decoding method
-                      - sglang, vllm, dynamo-vllm: runtime framework
+                      - sglang, vllm, sgl-router, vllm-router, dynamo-vllm: runtime framework
                       - model criterion: matching configured model family
                       - checklist-complete: PR checklist is satisfied
                       - patchwork: modified upstream engine or runtime source
diff --git a/benchmarks/multi_node/agentic_srt.sh b/benchmarks/multi_node/agentic_srt.sh
index 79a36da524..5e7c1a3ce8 100644
--- a/benchmarks/multi_node/agentic_srt.sh
+++ b/benchmarks/multi_node/agentic_srt.sh
@@ -32,7 +32,14 @@ install_agentic_deps
 wait_for_agentic_servers_idle() {
     local timeout_seconds="${AIPERF_DRAIN_TIMEOUT_SECONDS:-1800}"
     local poll_seconds="${AIPERF_DRAIN_POLL_SECONDS:-10}"
-    local frontend_metrics_url="http://localhost:${PORT}/metrics"
+    local frontend_metrics_url=""
+
+    # Dynamo exposes its own active-request gauge. Native static routers do
+    # not promise a metrics endpoint, so their drain signal comes entirely
+    # from the explicit logical worker endpoints supplied by srt-slurm.
+    case "$FRAMEWORK" in
+        dynamo-*) frontend_metrics_url="http://localhost:${PORT}/metrics" ;;
+    esac
 
     "$AIPERF_PYTHON" - \
         "$timeout_seconds" \
@@ -70,13 +77,19 @@ def metric_sum(metrics: str, name: str) -> float:
 
 while time.monotonic() < deadline:
     try:
-        frontend_metrics = fetch_metrics(frontend_url)
-        frontend_active = metric_sum(frontend_metrics, "dynamo_frontend_active_requests")
+        frontend_active = 0.0
+        if frontend_url:
+            frontend_metrics = fetch_metrics(frontend_url)
+            frontend_active = metric_sum(
+                frontend_metrics, "dynamo_frontend_active_requests"
+            )
         worker_active = 0.0
         for worker_url in worker_urls:
             worker_metrics = fetch_metrics(worker_url)
             worker_active += metric_sum(worker_metrics, "vllm:num_requests_running")
             worker_active += metric_sum(worker_metrics, "vllm:num_requests_waiting")
+            worker_active += metric_sum(worker_metrics, "sglang:num_requests_running")
+            worker_active += metric_sum(worker_metrics, "sglang:num_requests_waiting")
         print(
             f"Agentic drain status: frontend_active={frontend_active:g} "
             f"worker_running_or_waiting={worker_active:g}",
diff --git a/benchmarks/multi_node/srt-slurm-recipes/configs/native-sgl-router-deps.sh b/benchmarks/multi_node/srt-slurm-recipes/configs/native-sgl-router-deps.sh
new file mode 100644
index 0000000000..a9cccf9c1c
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/configs/native-sgl-router-deps.sh
@@ -0,0 +1,5 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+python3 -m pip install --no-cache-dir --upgrade "sglang-router==0.3.2"
+python3 -c 'import sglang_router; print(sglang_router.__file__)'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/configs/native-sgl-router-pd-fix-deps.sh b/benchmarks/multi_node/srt-slurm-recipes/configs/native-sgl-router-pd-fix-deps.sh
new file mode 100644
index 0000000000..be44533a07
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/configs/native-sgl-router-pd-fix-deps.sh
@@ -0,0 +1,55 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Temporary exact-source pin for the upstream SGLang Model Gateway one-token
+# P/D response fix. The shared wheel cache and flock make the source build
+# happen once even though srt-slurm runs the recipe setup on every role.
+SGLANG_REPOSITORY="https://github.com/cquil11/sglang.git"
+SGLANG_REF="fix/pd-one-token-prefill-response"
+SGLANG_COMMIT="9072e0dc5aaf1962b4e2c6f1a51094356b5d3324"
+WHEEL_CACHE_ROOT="/router_wheels/sglang-router-${SGLANG_COMMIT}"
+LOCK_FILE="${WHEEL_CACHE_ROOT}.lock"
+
+mkdir -p "${WHEEL_CACHE_ROOT}"
+exec 9>"${LOCK_FILE}"
+flock -w 3600 9
+
+shopt -s nullglob
+router_wheels=("${WHEEL_CACHE_ROOT}"/sglang_router-*.whl)
+if (( ${#router_wheels[@]} == 0 )); then
+    build_dir=$(mktemp -d /tmp/sglang-router-build.XXXXXX)
+    trap 'rm -rf "${build_dir}"' EXIT
+
+    git clone --filter=blob:none --no-checkout --single-branch \
+        --branch "${SGLANG_REF}" "${SGLANG_REPOSITORY}" "${build_dir}"
+    git -C "${build_dir}" sparse-checkout init --cone
+    git -C "${build_dir}" sparse-checkout set sgl-model-gateway
+    git -C "${build_dir}" checkout --detach "${SGLANG_COMMIT}"
+    if [[ "$(git -C "${build_dir}" rev-parse HEAD)" != "${SGLANG_COMMIT}" ]]; then
+        echo "ERROR: SGLang checkout does not match ${SGLANG_COMMIT}" >&2
+        exit 1
+    fi
+
+    python3 -m pip install --break-system-packages --no-cache-dir --upgrade \
+        "maturin>=1.0,<2.0"
+    (
+        cd "${build_dir}/sgl-model-gateway/bindings/python"
+        CARGO_BUILD_JOBS="${CARGO_BUILD_JOBS:-16}" \
+            maturin build --release --out "${WHEEL_CACHE_ROOT}"
+    )
+    router_wheels=("${WHEEL_CACHE_ROOT}"/sglang_router-*.whl)
+fi
+
+if (( ${#router_wheels[@]} != 1 )); then
+    echo "ERROR: expected exactly one cached SGLang Router wheel, found ${#router_wheels[@]}" >&2
+    exit 1
+fi
+
+# Install the released package first to resolve its Python dependencies, then
+# replace only the native package with the wheel built from the exact fix SHA.
+python3 -m pip install --break-system-packages --no-cache-dir --upgrade \
+    "sglang-router==0.3.2"
+python3 -m pip install --break-system-packages --no-cache-dir \
+    --force-reinstall --no-deps "${router_wheels[0]}"
+python3 -c 'import sglang_router; print(sglang_router.__file__)'
+echo "Installed SGLang Router from ${SGLANG_COMMIT}"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/configs/native-vllm-router-deps.sh b/benchmarks/multi_node/srt-slurm-recipes/configs/native-vllm-router-deps.sh
new file mode 100644
index 0000000000..286ebc7fb8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/configs/native-vllm-router-deps.sh
@@ -0,0 +1,11 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+python3 -m pip install --no-cache-dir --upgrade "vllm-router==0.1.15"
+command -v vllm-router >/dev/null
+python3 - <<'PY'
+from importlib.metadata import version
+
+assert version("vllm-router") == "0.1.15"
+print(f"vllm-router {version('vllm-router')}")
+PY
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-b200-2xdep8-mtp-sgl-router-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-b200-2xdep8-mtp-sgl-router-agentic.yaml
new file mode 100644
index 0000000000..efe272db5f
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-b200-2xdep8-mtp-sgl-router-agentic.yaml
@@ -0,0 +1,96 @@
+name: "dsv4-b200-sgl-router-agg-2xdep8-mtp-agentic"
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:v0.5.16-cu130"
+  precision: "fp4"
+
+setup_script: native-sgl-router-deps.sh
+
+slurm:
+  time_limit: "4:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  agg_nodes: 2
+  agg_workers: 2
+  gpus_per_agg: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+
+frontend:
+  type: sglang
+  enable_multiple_frontends: false
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+
+backend:
+  type: sglang
+  aggregated_environment:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    PYTHONUNBUFFERED: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "high"
+    SGLANG_SIMULATE_ACC_LEN: "2.99"
+    SGLANG_SIMULATE_ACC_METHOD: "match-expected"
+    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+    SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
+    SGLANG_OPT_USE_JIT_NORM: "1"
+    SGLANG_OPT_USE_TOPK_V2: "True"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+  sglang_config:
+    aggregated:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      enable-metrics: true
+      enable-cache-report: true
+      trust-remote-code: true
+      tool-call-parser: "deepseekv4"
+      stream-interval: 10
+      watchdog-timeout: 1800
+      mem-fraction-static: 0.90
+      chunked-prefill-size: 8192
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      scheduler-recv-interval: 30
+      dp-size: 8
+      tp-size: 8
+      ep-size: 8
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      speculative-algorithm: "EAGLE"
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.2/agentic/disagg-h200-2p2d-pcp8-tp8-dp8-mtp-sgl-router.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.2/agentic/disagg-h200-2p2d-pcp8-tp8-dp8-mtp-sgl-router.yaml
new file mode 100644
index 0000000000..e602939b4f
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.2/agentic/disagg-h200-2p2d-pcp8-tp8-dp8-mtp-sgl-router.yaml
@@ -0,0 +1,152 @@
+name: "glm5.2-h200-sgl-router-disagg-2p2d-mtp-agentic"
+
+# Native SGL Router validation of the established GLM-5.2 H200 Mooncake
+# topology. This deliberately changes only the request plane: model, worker
+# layout, KV transport, MTP, and AgentX behavior match the Dynamo baseline.
+
+model:
+  path: "hf:zai-org/GLM-5.2-FP8"
+  container: "lmsysorg/sglang:v0.5.16-cu130"
+  precision: "fp8"
+
+identity:
+  model:
+    repo: "zai-org/GLM-5.2-FP8"
+    revision: "ba978f7d347eaf65d22f1a86833408afdb953541"
+  container:
+    image: "lmsysorg/sglang:v0.5.16-cu130"
+  frameworks:
+    sglang: "0.5.16"
+
+setup_script: native-sgl-router-pd-fix-deps.sh
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+resources:
+  gpu_type: h200
+  gpus_per_node: 8
+  prefill_nodes: 2
+  decode_nodes: 2
+  prefill_workers: 2
+  decode_workers: 2
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+
+frontend:
+  type: sglang
+  enable_multiple_frontends: false
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+
+backend:
+  type: sglang
+  prefill_environment:
+    HF_HOME: "/hf_hub_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    PYTHONUNBUFFERED: "1"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+    SGLANG_OPT_USE_TOPK_V2: "1"
+    SGLANG_TIMEOUT_KEEP_ALIVE: "900"
+  decode_environment:
+    HF_HOME: "/hf_hub_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    PYTHONUNBUFFERED: "1"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+    SGLANG_OPT_USE_TOPK_V2: "1"
+    SGLANG_TIMEOUT_KEEP_ALIVE: "900"
+    SGLANG_SIMULATE_ACC_LEN: "2.99"
+    SGLANG_SIMULATE_ACC_METHOD: "match-expected"
+    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
+
+  sglang_config:
+    prefill:
+      host: 0.0.0.0
+      model-path: /model/
+      served-model-name: zai-org/GLM-5.2-FP8
+      trust-remote-code: true
+      tool-call-parser: glm47
+      reasoning-parser: glm45
+      tp-size: 8
+      ep-size: 1
+      attn-cp-size: 8
+      enable-prefill-cp: true
+      cp-strategy: interleave
+      enable-dsa-cache-layer-split: true
+      disaggregation-transfer-backend: mooncake
+      disaggregation-mode: prefill
+      kv-cache-dtype: fp8_e4m3
+      context-length: 1048576
+      max-total-tokens: 1048576
+      chunked-prefill-size: 32768
+      mem-fraction-static: 0.85
+      max-running-requests: 32
+      speculative-algorithm: EAGLE
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      watchdog-timeout: 86400
+      stream-interval: 60
+      enable-metrics: true
+      enable-cache-report: true
+
+    decode:
+      host: 0.0.0.0
+      model-path: /model/
+      served-model-name: zai-org/GLM-5.2-FP8
+      trust-remote-code: true
+      tool-call-parser: glm47
+      reasoning-parser: glm45
+      tp-size: 8
+      dp-size: 8
+      ep-size: 1
+      enable-dp-attention: true
+      disaggregation-transfer-backend: mooncake
+      disaggregation-mode: decode
+      kv-cache-dtype: fp8_e4m3
+      dsa-decode-backend: flashmla_kv
+      context-length: 1048576
+      max-total-tokens: 1048576
+      mem-fraction-static: 0.85
+      max-running-requests: 200
+      page-size: 64
+      disable-radix-cache: true
+      speculative-algorithm: EAGLE
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      watchdog-timeout: 86400
+      stream-interval: 60
+      enable-metrics: true
+      enable-cache-report: true
+
+sbatch_directives:
+  mem: "0"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-b200-2xdep8-mtp-vllm-router-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-b200-2xdep8-mtp-vllm-router-agentic.yaml
new file mode 100644
index 0000000000..0bb0f9ba8e
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-b200-2xdep8-mtp-vllm-router-agentic.yaml
@@ -0,0 +1,100 @@
+name: "dsv4-b200-vllm-router-agg-2xdep8-mtp-agentic"
+
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec"
+  precision: "fp4"
+
+setup_script: native-vllm-router-deps.sh
+
+slurm:
+  time_limit: "4:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  agg_nodes: 2
+  agg_workers: 2
+  gpus_per_agg: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+
+frontend:
+  type: vllm-router
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  aggregated_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    # The released B200 DSV4 DEP path uses PYNCCL for local DP/EP collectives.
+    # NCCL symmetric-memory window registration fails during dummy profiling on
+    # this cluster, before the router can receive traffic.
+    VLLM_USE_NCCL_SYMM_MEM: "0"
+    TORCH_SYMMMEM: "NVSHMEM"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_TLS: "cuda_copy,rc"
+    NCCL_P2P_LEVEL: "NVL"
+  vllm_config:
+    aggregated:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      enable-prompt-tokens-details: true
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
+      max-model-len: 1048576
+      max-num-seqs: 128
+      max-num-batched-tokens: 8192
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      max-cudagraph-capture-size: 128
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3,"rejection_sample_method":"synthetic","synthetic_acceptance_length":2.99}'
+      gpu-memory-utilization: 0.90
+      stream-interval: 10
+      no-disable-hybrid-kv-cache-manager: true
+      tokenizer-mode: "deepseek_v4"
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-b200-1p1d-dep8-mtp-vllm-router-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-b200-1p1d-dep8-mtp-vllm-router-agentic.yaml
new file mode 100644
index 0000000000..1ef9ff90fb
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-b200-1p1d-dep8-mtp-vllm-router-agentic.yaml
@@ -0,0 +1,105 @@
+name: "dsv4-b200-vllm-router-disagg-1p1d-dep8-mtp-agentic"
+
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec"
+  precision: "fp4"
+
+setup_script: native-vllm-router-deps.sh
+
+slurm:
+  time_limit: "4:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 8
+  decode_nodes: 1
+  decode_workers: 1
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+
+frontend:
+  type: vllm-router
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  prefill_environment: &vllm_environment
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    # Match the known-green B200 DSV4 DEP communicator path. NCCL symmetric
+    # memory fails while registering its local DP/EP communication window.
+    VLLM_USE_NCCL_SYMM_MEM: "0"
+    TORCH_SYMMMEM: "NVSHMEM"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_TLS: "cuda_copy,rc"
+    NCCL_P2P_LEVEL: "NVL"
+  decode_environment: *vllm_environment
+  vllm_config:
+    prefill: &vllm_config
+      kv-transfer-config: '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      enable-prompt-tokens-details: true
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
+      max-model-len: 1048576
+      max-num-seqs: 128
+      max-num-batched-tokens: 8192
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      max-cudagraph-capture-size: 128
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3,"rejection_sample_method":"synthetic","synthetic_acceptance_length":2.99}'
+      gpu-memory-utilization: 0.90
+      stream-interval: 10
+      no-disable-hybrid-kv-cache-manager: true
+      tokenizer-mode: "deepseek_v4"
+    decode: *vllm_config
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/configs/ci-priority.yaml b/configs/ci-priority.yaml
index 34565e2810..11628c1b49 100644
--- a/configs/ci-priority.yaml
+++ b/configs/ci-priority.yaml
@@ -18,6 +18,9 @@ adjustments:
     eagle: 0.75
     eagle3: 0.75
   framework-prefix:
+    # Match the explicit router runtimes before their broader engine prefixes.
+    sgl-router: 0.5
+    vllm-router: 0.5
     sglang: 0.5
     vllm: 0.5
     dynamo-sglang: 0.5
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 393babf91b..0a8670bc8d 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -966,6 +966,97 @@ dsv4-fp4-b200-vllm-agentic-mtp:
       # DEP Mooncake + MTP (num_speculative_tokens=3)
       - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, spec-decoding: mtp, conc-list: [12, 20, 28, 36, 44, 52, 60, 68, 76], router: { name: vllm-router, version: "0.1.14" } }
 
+dsv4-fp4-b200-vllm-router-agentic-agg:
+  image: vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:b200-dgxc
+  precision: fp4
+  framework: vllm-router
+  router: { name: vllm-router, version: "0.1.15" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - spec-decoding: mtp
+        conc-list: [16]
+        prefill:
+          num-worker: 2
+          tp: 1
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SRT_SLURM_REPOSITORY=https://github.com/SemiAnalysisAI/srt-slurm.git"
+          - "SRT_SLURM_REF=ab98030f0993bdb1595acc25e883b59d76ad7791"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/agg-b200-2xdep8-mtp-vllm-router-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 1
+          ep: 8
+          dp-attn: true
+
+dsv4-fp4-b200-vllm-router-agentic-disagg:
+  image: vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:b200-dgxc
+  precision: fp4
+  framework: vllm-router
+  router: { name: vllm-router, version: "0.1.15" }
+  kv-p2p-transfer: nixl
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - spec-decoding: mtp
+        conc-list: [16]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SRT_SLURM_REPOSITORY=https://github.com/SemiAnalysisAI/srt-slurm.git"
+          - "SRT_SLURM_REF=ab98030f0993bdb1595acc25e883b59d76ad7791"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-b200-1p1d-dep8-mtp-vllm-router-agentic.yaml"
+        decode:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+
+dsv4-fp4-b200-sgl-router-agentic-agg:
+  image: lmsysorg/sglang:v0.5.16-cu130
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:b200-dgxc
+  precision: fp4
+  framework: sgl-router
+  router: { name: sgl-router, version: "0.3.2" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - spec-decoding: mtp
+        conc-list: [16]
+        prefill:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SRT_SLURM_REPOSITORY=https://github.com/SemiAnalysisAI/srt-slurm.git"
+          - "SRT_SLURM_REF=ab98030f0993bdb1595acc25e883b59d76ad7791"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/agg-b200-2xdep8-mtp-sgl-router-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 8
+          dp-attn: true
+
 dsv4-fp4-b200-trt:
   image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-c185066
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -7676,6 +7767,44 @@ glm5.2-fp8-h200-dynamo-sglang-agentic-mtp-2p2d:
           ep: 1
           dp-attn: true
 
+# Request-plane sibling of the established Dynamo deployment above. The H200
+# hosts provide the peer-memory support required by the proven Mooncake data
+# plane; this keeps native SGL Router validation independent of B200's missing
+# GPUDirect registration module.
+glm5.2-fp8-h200-sgl-router-agentic-mtp-2p2d:
+  image: lmsysorg/sglang:v0.5.16-cu130
+  model: zai-org/GLM-5.2-FP8
+  model-prefix: glm5.2
+  runner: cluster:h200-dgxc
+  precision: fp8
+  framework: sgl-router
+  router: { name: sgl-router, version: "0.3.2" }
+  kv-p2p-transfer: mooncake
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - spec-decoding: mtp
+        conc-list: [8]
+        kv-offloading: none
+        prefill:
+          num-worker: 2
+          tp: 1
+          pcp-size: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "SRT_SLURM_REPOSITORY=https://github.com/SemiAnalysisAI/srt-slurm.git"
+          - "SRT_SLURM_REF=ab98030f0993bdb1595acc25e883b59d76ad7791"
+          - "CONFIG_FILE=recipes/sglang/glm5.2/agentic/disagg-h200-2p2d-pcp8-tp8-dp8-mtp-sgl-router.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 1
+          dp-attn: true
+
 # GLM-5.2 B300 NVFP4 AgentX with EAGLE/MTP speculative decoding, following the
 # AgentX speculative-decoding policy in MODELS.md. SGLang EAGLE runs off
 # GLM-5.2's built-in nextn head (num-steps 3,
@@ -7877,7 +8006,7 @@ kimik3-fp4-b200-dynamo-vllm-agentic:
   runner: cluster:b200-dgxc
   precision: fp4
   # framework stays dynamo-vllm for launcher routing, but this variant serves
-  # DIRECTLY with vllm serve (srt-slurm PR #278 frontend.type: vllm + the
+  # DIRECTLY with vllm serve (srt-slurm PR #278, frontend.type: vllm, plus the
   # InferenceX multinode patch) — no dynamo frontend/worker/router involved.
   framework: dynamo-vllm
   multinode: true
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 471981f60c..fa1879f5cd 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5709,3 +5709,18 @@
     - "Enable SGLang metrics on every aggregate, prefill, and decode engine."
     - "Use supported header-based Dynamo session routing with the in-repo AIPerf build."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2520
+
+- config-keys:
+    - dsv4-fp4-b200-vllm-router-agentic-agg
+    - dsv4-fp4-b200-vllm-router-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Validate first-class vLLM Router 0.1.15 through srt-slurm on B200 without a Dynamo request plane."
+    - "Exercise vLLM Router with two-worker aggregate DEP8 and one-prefill/one-decode DEP8 DeepSeek V4 FP4 deployments on B200. Both validations use MTP and the AgentX trace workload."
+    - "Pin the test deployment to exact SemiAnalysisAI/srt-slurm PR head ab98030f0993bdb1595acc25e883b59d76ad7791, preserve each logical backend Prometheus endpoint for AIPerf, and require the vllm: metric family before accepting results."
+    - "Enable vLLM prompt-token details on both native Router deployments so AgentX captures response-level prompt-cache usage in addition to each logical backend's Prometheus metrics."
+    - "Verify the terminal Slurm state and exit code after log streaming so a failed benchmark cannot be reported as a successful GitHub job merely because its log tail ended cleanly."
+    - "Classify sgl-router and vllm-router as explicit runtime criteria in the CI scheduler, matching the router-specific framework keys before their broader engine prefixes."
+    - "Drain sequential native-router AgentX points from all explicit logical worker metrics without requiring a Dynamo frontend gauge; recognize both vllm:num_requests_* and sglang:num_requests_* worker families."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2543
diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
index 778eac1e51..ffad35ede4 100644
--- a/runners/launch_b200-dgxc.sh
+++ b/runners/launch_b200-dgxc.sh
@@ -4,6 +4,8 @@
 SLURM_PARTITION="${SLURM_PARTITION:-gpu-2}"
 SLURM_ACCOUNT="${SLURM_ACCOUNT:-benchmark}"
 
+source "$(dirname "${BASH_SOURCE[0]}")/slurm_utils.sh"
+
 set -x
 
 # MODEL_PATH: Override with pre-downloaded paths on cluster-accessible storage.
@@ -135,14 +137,15 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     fi
 
     # Validate framework
-    if [[ $FRAMEWORK != "dynamo-sglang" && $FRAMEWORK != "dynamo-trt" && $FRAMEWORK != "dynamo-vllm" ]]; then
-        echo "Unsupported framework: $FRAMEWORK. Supported frameworks are: dynamo-trt, dynamo-sglang, dynamo-vllm"
+    if [[ $FRAMEWORK != "dynamo-sglang" && $FRAMEWORK != "dynamo-trt" && $FRAMEWORK != "dynamo-vllm" && $FRAMEWORK != "sgl-router" && $FRAMEWORK != "vllm-router" ]]; then
+        echo "Unsupported framework: $FRAMEWORK. Supported frameworks are: dynamo-trt, dynamo-sglang, dynamo-vllm, sgl-router, vllm-router"
         exit 1
     fi
 
-    # Multinode dsv4 currently only ships with the dynamo-vllm recipe
-    if [[ $MODEL_PREFIX == "dsv4" && $FRAMEWORK != "dynamo-vllm" ]]; then
-        echo "Unsupported framework for multinode dsv4: $FRAMEWORK (only dynamo-vllm)"
+    # Multinode dsv4 currently ships with Dynamo vLLM and the two native
+    # static-router integrations under test in this PR.
+    if [[ $MODEL_PREFIX == "dsv4" && $FRAMEWORK != "dynamo-vllm" && $FRAMEWORK != "sgl-router" && $FRAMEWORK != "vllm-router" ]]; then
+        echo "Unsupported framework for multinode dsv4: $FRAMEWORK"
         exit 1
     fi
 
@@ -155,10 +158,29 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         rm -rf "$SRT_REPO_DIR"
     fi
 
+    if [[ -n "${SRT_SLURM_REPOSITORY:-}" || -n "${SRT_SLURM_REF:-}" ]]; then
+        if [[ -z "${SRT_SLURM_REPOSITORY:-}" || -z "${SRT_SLURM_REF:-}" ]]; then
+            echo "SRT_SLURM_REPOSITORY and SRT_SLURM_REF must be set together" >&2
+            exit 1
+        fi
+        git clone "$SRT_SLURM_REPOSITORY" "$SRT_REPO_DIR" || exit 1
+        cd "$SRT_REPO_DIR" || exit 1
+        git checkout --detach "$SRT_SLURM_REF" || exit 1
+        if [[ "$(git rev-parse HEAD)" != "$SRT_SLURM_REF" ]]; then
+            echo "srt-slurm checkout does not match requested exact commit: $SRT_SLURM_REF" >&2
+            exit 1
+        fi
+        mkdir -p recipes/vllm/deepseek-v4 recipes/sglang/deepseek-v4 configs || exit 1
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" \
+            recipes/vllm/deepseek-v4 || exit 1
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4" \
+            recipes/sglang/deepseek-v4 || exit 1
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs" \
+            configs || exit 1
     # TODO(CJQ): make first class upon srt-slurm upstream refactor
-    if [[ "$IS_AGENTIC" == "1" && $MODEL_PREFIX == "kimik3" ]]; then
+    elif [[ "$IS_AGENTIC" == "1" && $MODEL_PREFIX == "kimik3" ]]; then
         # Direct-vLLM agentic experiment (Variant D): srt-slurm PR #278
-        # (kylliang/direct-aggregate-vllm) adds frontend.type: vllm — `vllm
+        # (kylliang/direct-aggregate-vllm, frontend.type: vllm) adds direct `vllm
         # serve` owns the OpenAI port itself, no Dynamo layer. The fork branch
         # carries PR #278 plus the multi-node extension (vLLM-native
         # --master-addr/--nnodes/--node-rank serve + headless non-leader
@@ -322,6 +344,8 @@ containers:
   dynamo-trtllm: "${SQUASH_FILE}"
   dynamo-sglang: "${SQUASH_FILE}"
   dynamo-vllm: "${SQUASH_FILE}"
+  sgl-router: "${SQUASH_FILE}"
+  vllm-router: "${SQUASH_FILE}"
   sglang-v0.5.11-cu130: "${SQUASH_FILE}"
   "${IMAGE}": "${SQUASH_FILE}"
   nginx-sqsh: "${NGINX_SQUASH_FILE}"
@@ -407,6 +431,12 @@ EOF
 
     wait $POLL_PID
 
+    # Native router validation is all-or-nothing. Other B200 workflows retain
+    # the launcher's historical result-collection behavior.
+    if [[ "$FRAMEWORK" == "sgl-router" || "$FRAMEWORK" == "vllm-router" ]]; then
+        wait_for_slurm_job_success "$JOB_ID" || exit 1
+    fi
+
     set -x
 
     echo "Job $JOB_ID completed!"
diff --git a/runners/launch_h200-dgxc-slurm.sh b/runners/launch_h200-dgxc-slurm.sh
index 8991e04182..4f353562cc 100755
--- a/runners/launch_h200-dgxc-slurm.sh
+++ b/runners/launch_h200-dgxc-slurm.sh
@@ -21,7 +21,7 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     # MODEL_PATH: Override with pre-downloaded paths on H200 runner
     # The yaml files specify HuggingFace model IDs for portability, but we use
     # local paths to avoid repeated downloading on the shared H200 cluster.
-    if [[ $FRAMEWORK == "dynamo-sglang" ]]; then
+    if [[ $FRAMEWORK == "dynamo-sglang" || $FRAMEWORK == "sgl-router" ]]; then
         if [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp8" ]]; then
             export MODEL_PATH="/models/DeepSeek-R1-0528"
             export SRT_SLURM_MODEL_PREFIX="dsr1-fp8"
@@ -32,7 +32,7 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
             fi
             export SRT_SLURM_MODEL_PREFIX="glm5.2-fp8"
         else
-            echo "Unsupported model prefix/precision for dynamo-sglang: $MODEL_PREFIX/$PRECISION"
+            echo "Unsupported model prefix/precision for SGLang: $MODEL_PREFIX/$PRECISION"
             exit 1
         fi
     elif [[ $FRAMEWORK == "dynamo-trt" ]]; then
@@ -44,16 +44,16 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
             echo "Unsupported model prefix/precision for dynamo-trt: $MODEL_PREFIX/$PRECISION"
             exit 1
         fi
-    elif [[ $FRAMEWORK == "vllm" ]]; then
+    elif [[ $FRAMEWORK == "vllm" || $FRAMEWORK == "vllm-router" ]]; then
         if [[ $MODEL_PREFIX == "kimik3" && $PRECISION == "fp4" ]]; then
             export MODEL_PATH="/models/gharunners/hf-hub-cache/Kimi-K3"
             export SRT_SLURM_MODEL_PREFIX="kimik3"
         else
-            echo "Unsupported model prefix/precision for vllm: $MODEL_PREFIX/$PRECISION"
+            echo "Unsupported model prefix/precision for vLLM: $MODEL_PREFIX/$PRECISION"
             exit 1
         fi
     else
-        echo "Unsupported framework: $FRAMEWORK. Supported frameworks are: dynamo-trt, dynamo-sglang, vllm"
+        echo "Unsupported framework: $FRAMEWORK. Supported frameworks are: dynamo-trt, dynamo-sglang, vllm, sgl-router, vllm-router"
         exit 1
     fi
 
@@ -64,7 +64,21 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         rm -rf "$SRT_REPO_DIR"
     fi
 
-    if [[ $IS_AGENTIC == "1" && $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5.2" ]]; then
+    if [[ -n "${SRT_SLURM_REPOSITORY:-}" || -n "${SRT_SLURM_REF:-}" ]]; then
+        if [[ -z "${SRT_SLURM_REPOSITORY:-}" || -z "${SRT_SLURM_REF:-}" ]]; then
+            echo "SRT_SLURM_REPOSITORY and SRT_SLURM_REF must be set together" >&2
+            exit 1
+        fi
+        git clone "$SRT_SLURM_REPOSITORY" "$SRT_REPO_DIR"
+        cd "$SRT_REPO_DIR"
+        git checkout --detach "$SRT_SLURM_REF"
+        if [[ "$(git rev-parse HEAD)" != "$SRT_SLURM_REF" ]]; then
+            echo "srt-slurm checkout does not match requested exact commit: $SRT_SLURM_REF" >&2
+            exit 1
+        fi
+        mkdir -p configs
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs" configs
+    elif [[ $IS_AGENTIC == "1" && $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5.2" ]]; then
         # v1.0.44 includes the AgentX custom benchmark integration and passes
         # every logical SGLang worker's Prometheus URL to AIPerf.
         git clone --branch v1.0.44 --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
@@ -103,7 +117,7 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     # Map container images to local squash files based on framework
     NGINX_SQUASH_FILE="/data/containers/nginx+1.27.4.sqsh"
 
-    if [[ $FRAMEWORK == "dynamo-sglang" ]]; then
+    if [[ $FRAMEWORK == "dynamo-sglang" || $FRAMEWORK == "sgl-router" ]]; then
         # SGLang container mapping
         if [[ $MODEL_PREFIX == "glm5.2" ]]; then
             SQUASH_FILE="/data/gharunners/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
@@ -115,7 +129,7 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         # TRT-LLM container mapping - convert IMAGE to srt-slurm format (nvcr.io/ -> nvcr.io#)
         CONTAINER_KEY=$(echo "$IMAGE" | sed 's|nvcr.io/|nvcr.io#|')
         SQUASH_FILE="/data/containers/$(echo "$IMAGE" | sed 's|nvcr.io/||' | sed 's/[\/:@#]/+/g').sqsh"
-    elif [[ $FRAMEWORK == "vllm" ]]; then
+    elif [[ $FRAMEWORK == "vllm" || $FRAMEWORK == "vllm-router" ]]; then
         CONTAINER_KEY="$IMAGE"
         SQUASH_FILE="/data/gharunners/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
     fi
@@ -150,10 +164,13 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     if [[ "$IS_AGENTIC" == "1" ]]; then
         AIPERF_MMAP_CACHE_HOST_PATH="/home/sa-shared/gharunners/ai-perf-cache"
         HF_HUB_CACHE_HOST_PATH="/models/gharunners/hf-hub-cache"
+        ROUTER_WHEEL_CACHE_HOST_PATH="/home/sa-shared/gharunners/router-wheels"
         mkdir -p "$AIPERF_MMAP_CACHE_HOST_PATH"
+        mkdir -p "$ROUTER_WHEEL_CACHE_HOST_PATH"
         DEFAULT_MOUNTS_BLOCK="default_mounts:
   ${AIPERF_MMAP_CACHE_HOST_PATH}: /aiperf_mmap_cache
-  ${HF_HUB_CACHE_HOST_PATH}: /hf_hub_cache"
+  ${HF_HUB_CACHE_HOST_PATH}: /hf_hub_cache
+  ${ROUTER_WHEEL_CACHE_HOST_PATH}: /router_wheels"
     fi
     echo "Creating srtslurm.yaml configuration..."
     cat > srtslurm.yaml <<EOF
@@ -176,6 +193,8 @@ containers:
   dynamo-trtllm: "${SQUASH_FILE}"
   dynamo-sglang: "${SQUASH_FILE}"
   dynamo-vllm: "${SQUASH_FILE}"
+  sgl-router: "${SQUASH_FILE}"
+  vllm-router: "${SQUASH_FILE}"
   nginx-sqsh: "${NGINX_SQUASH_FILE}"
   latest: "${SQUASH_FILE}"
   "${CONTAINER_KEY}": "${SQUASH_FILE}"
@@ -232,6 +251,12 @@ EOF
     trap 'rc=$?; bundle_server_logs "$LOGS_DIR" "$GITHUB_WORKSPACE/multinode_server_logs.tar.gz"; scancel "$JOB_ID" 2>/dev/null || true; exit "$rc"' EXIT INT TERM HUP
 
     stream_slurm_job_log "$JOB_ID" "$LOG_FILE" || exit 1
+    # Native router validation is all-or-nothing. Keep the shared log helper's
+    # legacy behavior for unrelated launchers that intentionally retain
+    # partial results from timed-out allocations.
+    if [[ "$FRAMEWORK" == "sgl-router" || "$FRAMEWORK" == "vllm-router" ]]; then
+        wait_for_slurm_job_success "$JOB_ID" || exit 1
+    fi
 
     set -x
 
diff --git a/runners/slurm_utils.sh b/runners/slurm_utils.sh
index 60ce818eb5..3a0f876f8f 100644
--- a/runners/slurm_utils.sh
+++ b/runners/slurm_utils.sh
@@ -18,6 +18,49 @@ slurm_job_is_active() {
     squeue -j "$job_id" --noheader 2>/dev/null | grep -q "$job_id"
 }
 
+wait_for_slurm_job_success() {
+    local job_id="$1"
+    local attempts="${SLURM_ACCOUNTING_ATTEMPTS:-24}"
+    local interval="${SLURM_ACCOUNTING_INTERVAL_SECONDS:-5}"
+    local attempt record state exit_code
+
+    if ! command -v sacct >/dev/null 2>&1; then
+        echo "ERROR: sacct is required to verify Slurm job $job_id" >&2
+        return 1
+    fi
+
+    for ((attempt = 1; attempt <= attempts; attempt++)); do
+        record=$(sacct -X -n -P -j "$job_id" -o JobIDRaw,State,ExitCode 2>/dev/null \
+            | awk -F'|' -v job_id="$job_id" '$1 == job_id { print $2 "|" $3; exit }')
+        if [[ -n "$record" ]]; then
+            state="${record%%|*}"
+            exit_code="${record#*|}"
+            # Slurm may suffix terminal states with "+" when the display is
+            # truncated; compare the canonical state name.
+            state="${state%% *}"
+            state="${state%%+}"
+
+            if [[ "$state" == "COMPLETED" && "$exit_code" == "0:0" ]]; then
+                echo "Slurm job $job_id completed successfully ($state, exit $exit_code)"
+                return 0
+            fi
+
+            case "$state" in
+                BOOT_FAIL|CANCELLED|DEADLINE|FAILED|NODE_FAIL|OUT_OF_MEMORY|PREEMPTED|REVOKED|SPECIAL_EXIT|TIMEOUT)
+                    echo "ERROR: Slurm job $job_id ended in $state with exit $exit_code" >&2
+                    return 1
+                    ;;
+            esac
+        fi
+
+        sleep "$interval"
+    done
+
+    echo "ERROR: Slurm accounting did not report a successful terminal state for job $job_id" >&2
+    [[ -z "${record:-}" ]] || echo "Last accounting record: $record" >&2
+    return 1
+}
+
 stream_slurm_job_log() {
     local job_id="$1"
     local log_file="$2"
diff --git a/runners/test_slurm_utils.py b/runners/test_slurm_utils.py
index c027cc9b2f..a76cd262f8 100644
--- a/runners/test_slurm_utils.py
+++ b/runners/test_slurm_utils.py
@@ -1,6 +1,8 @@
 import subprocess
 from pathlib import Path
 
+import pytest
+
 
 REPO_ROOT = Path(__file__).resolve().parents[1]
 SLURM_UTILS = REPO_ROOT / "runners" / "slurm_utils.sh"
@@ -15,6 +17,42 @@ def run_bash(command: str, *args: Path | str) -> subprocess.CompletedProcess[str
     )
 
 
+def write_executable(path: Path, body: str) -> None:
+    path.write_text(f"#!/usr/bin/env bash\n{body}\n")
+    path.chmod(0o755)
+
+
+@pytest.mark.parametrize(
+    ("accounting_record", "expected_returncode", "expected_message"),
+    [
+        ("4242|COMPLETED|0:0", 0, "completed successfully"),
+        ("4242|FAILED|1:0", 1, "ended in FAILED with exit 1:0"),
+    ],
+)
+def test_wait_for_slurm_job_success_reports_accounting_status(
+    tmp_path: Path,
+    accounting_record: str,
+    expected_returncode: int,
+    expected_message: str,
+) -> None:
+    mock_bin = tmp_path / "bin"
+    mock_bin.mkdir()
+    write_executable(mock_bin / "sacct", 'printf \'%s\\n\' "$MOCK_SACCT_RECORD"')
+
+    result = run_bash(
+        'export PATH="$3:$PATH" MOCK_SACCT_RECORD="$4" '
+        "SLURM_ACCOUNTING_ATTEMPTS=1 SLURM_ACCOUNTING_INTERVAL_SECONDS=0; "
+        'source "$1"; wait_for_slurm_job_success 4242',
+        SLURM_UTILS,
+        tmp_path,
+        mock_bin,
+        accounting_record,
+    )
+
+    assert result.returncode == expected_returncode
+    assert expected_message in result.stdout + result.stderr
+
+
 def test_copy_agentic_results_stages_only_matching_points(tmp_path: Path) -> None:
     source = tmp_path / "source"
     workspace = tmp_path / "workspace"
diff --git a/utils/agentic/aggregation/backends/sglang.py b/utils/agentic/aggregation/backends/sglang.py
index acbbc12b02..0e7f8d6e18 100644
--- a/utils/agentic/aggregation/backends/sglang.py
+++ b/utils/agentic/aggregation/backends/sglang.py
@@ -27,7 +27,7 @@ class SglangBackend(ServerMetricsBackend):
     def matches(self, metrics: dict[str, dict[str, Any]], framework: str) -> bool:
         metric_names = set(metrics)
         return any(name.startswith("sglang:") for name in metric_names) or (
-            not metrics and framework.lower() == "sglang"
+            not metrics and framework.lower() in {"sglang", "sgl-router"}
         )
 
     def populate(
diff --git a/utils/agentic/aggregation/backends/vllm.py b/utils/agentic/aggregation/backends/vllm.py
index 18ed30ece8..a39735c3ed 100644
--- a/utils/agentic/aggregation/backends/vllm.py
+++ b/utils/agentic/aggregation/backends/vllm.py
@@ -27,7 +27,7 @@ class VllmBackend(ServerMetricsBackend):
     def matches(self, metrics: dict[str, dict[str, Any]], framework: str) -> bool:
         metric_names = set(metrics)
         return any(name.startswith("vllm:") for name in metric_names) or (
-            not metrics and framework.lower() == "vllm"
+            not metrics and framework.lower() in {"vllm", "vllm-router"}
         )
 
     def populate(