File: C:/Users/fred/.codex/.tmp/plugins/plugins/nvidia/skills/dynamo-interconnect-check/evals/evals.json
{
"skill": "dynamo-interconnect-check",
"cases": [
{
"id": "confirm-nixl-before-trusting-disagg",
"question": "I just brought up a disaggregated recipe. Confirm the NIXL/UCX transport is actually working before I trust the throughput numbers.",
"expected_skill": "dynamo-interconnect-check",
"expected_script": "scripts/check_interconnect.py",
"ground_truth": "Run read-only interconnect checks (transport env vars on the recipe, then runtime NIXL/UCX/NCCL reachability over RDMA/NVLink) and report whether KV transport is on the fast path.",
"expected_behavior": [
"Check transport env vars on the recipe",
"Run read-only runtime interconnect checks",
"Report whether NIXL/UCX can reach peer workers over RDMA/NVLink",
"Never mutate the cluster or print secrets"
]
},
{
"id": "agg-ok-disagg-slow-suspect-fabric",
"question": "Aggregated serving works fine but disaggregated is slow and sometimes hangs — I suspect the fabric, not the model.",
"expected_skill": "dynamo-interconnect-check",
"expected_script": "scripts/check_interconnect.py",
"ground_truth": "Validate the disagg transport path; a deployment can pass an endpoint smoke test while KV transfer silently falls back to a slow/broken path.",
"expected_behavior": [
"Check transport configuration and runtime reachability",
"Determine if KV transfer is on a fallback path",
"Report fabric findings"
]
},
{
"id": "rdma-nvlink-readiness-multinode",
"question": "Check RDMA/NVLink readiness for my multi-node disagg deployment.",
"expected_skill": "dynamo-interconnect-check",
"expected_script": "scripts/check_interconnect.py",
"ground_truth": "Run the interconnect readiness checks for a multi-node disagg deployment and report transport health.",
"expected_behavior": [
"Run interconnect readiness checks for multi-node disagg",
"Report RDMA/NVLink transport health"
]
},
{
"id": "neg-deploy-disagg-recipe",
"question": "Deploy a disaggregated recipe for DeepSeek on my cluster.",
"expected_skill": "dynamo-recipe-runner",
"ground_truth": "Bringing a recipe up belongs to dynamo-recipe-runner; interconnect-check runs afterward.",
"expected_behavior": ["dynamo-interconnect-check stays silent; dynamo-recipe-runner handles it"]
},
{
"id": "neg-disagg-pods-crashing",
"question": "My disagg prefill pods are crashing and unschedulable.",
"expected_skill": "dynamo-troubleshoot",
"ground_truth": "Already-crashing/unschedulable pods are failure diagnosis, which belongs to dynamo-troubleshoot.",
"expected_behavior": ["dynamo-interconnect-check stays silent; dynamo-troubleshoot handles it"]
},
{
"id": "neg-switch-router-mode",
"question": "Change my router to least-loaded mode.",
"expected_skill": "dynamo-router-starter",
"ground_truth": "Router mode work belongs to dynamo-router-starter.",
"expected_behavior": ["dynamo-interconnect-check stays silent; dynamo-router-starter handles it"]
}
]
}