{
  "skills": ["vss-deploy-detection-tracking-3d"],
  "resources": {
    "platforms": {
      "RTXPRO6000BW": {
        "gpu_count": 1,
        "modes": ["standalone"]
      }
    }
  },
  "env": "A GPU host matching `{{platform}}` with Docker, NVIDIA Container Toolkit, NGC credentials at `~/.ngc/config` (the skill bootstraps `~/.ngc/config` from `NGC_CLI_API_KEY` if needed), and ≥50 GB free disk for ~15 container images plus the extracted `vss-warehouse-app-data` (`videos/warehouse-4cams-20mx20m-synthetic/`, `models/mv3dt/BodyPose3DNet/`). `VSS_DATA_DIR` must point at the extracted app-data directory (NOT at the repo). Free TCP ports: 30888 (VST UI), 9092 (Kafka), 1883 (mosquitto), 5000 (auto-calibration UI), 8010 (auto-calibration MS), 31000 (nvstreamer), 9200 (Elasticsearch — extended profile only). `BP_PROFILE=bp_wh_kafka`; `STREAM_TYPE=kafka`. Default profile size is **extended** (`MINIMAL_PROFILE=\"\"`) so the VST video wall can render bounding-box overlays — the skill explains the trade-off in `verify-and-view.md`. **Cases run in declared order with state preserved between them**: case 1 deploys MV3DT on the sample → case 2 verifies the running stack → case 3 tears down. The framework must NOT reset Docker / container state between cases. Before case 1 runs, the host must have no MV3DT containers running (`vss-rtvi-cv-mv3dt`, `vss-rtvi-cv-bev-fusion`, `mosquitto`, `kafka`, `redis`, `vss-vios-sensor`) and the named volumes `mdx_mdx-kafka` and `mdx_vios_pg_data` must be absent. This eval exercises the warehouse-blueprint compose tree at `deploy/docker/compose.yml` with `--env-file industry-profiles/warehouse-operations/.env`, gated on the `bp_wh_kafka_mv3dt` compose profile — it does not use `/vss-deploy-profile` or `scripts/dev-profile.sh`.",
  "expects": [
    {
      "query": "Deploy rtvi-cv-3d on the sample dataset.",
      "checks": [
        "The agent loads the `vss-deploy-detection-tracking-3d` skill (not `vss-deploy-profile`'s warehouse reference). Trajectory shows it reading `skills/vss-deploy-detection-tracking-3d/SKILL.md` and at least one of `references/deploy-rtvi-cv-3d-stack.md` or `references/configure-cameras.md`.",
        "The agent acknowledges the profile-size dimension (Q0). Pass condition: at least one of `extended`, `minimal`, `MINIMAL_PROFILE`, or `bbox overlay` appears in the agent's text response, AskUserQuestion options, or trajectory metadata. The agent either defaults to extended or asks the user to choose — silent picking of minimal without surfacing the trade-off fails this check.",
        "The agent does NOT fabricate `MODE` / `BP_PROFILE` values outside the supported set. `MODE` must end up as `mv3dt` (not `3d`, `mv-3d`, `multi-view`, etc.); `BP_PROFILE` must be one of `bp_wh_kafka` or `bp_wh_redis` for an MV3DT deployment — NOT `bp_wh_auto_calib`, which gates only the AMC microservice + VST/nvstreamer supporting services and leaves the core `vss-rtvi-cv-mv3dt` / `vss-rtvi-cv-bev-fusion` containers undeployed. Fabrications or `bp_wh_auto_calib` fail this check.",
        "After deploy completes, the core MV3DT containers are running — verified by `docker ps --format '{{.Names}}' | grep -qx vss-rtvi-cv-mv3dt && docker ps --format '{{.Names}}' | grep -qx vss-rtvi-cv-bev-fusion && docker ps --format '{{.Names}}' | grep -qx mosquitto && docker ps --format '{{.Names}}' | grep -qxE '^(kafka|redis)$'`.",
        "BEV Fusion reports healthy via the compose health check — verified by `docker inspect --format '{{.State.Health.Status}}' vss-rtvi-cv-bev-fusion 2>/dev/null | grep -qx healthy`. The agent must NOT verify this via `docker exec ... test -f /tmp/fusion_ready` (the `test` binary is not on PATH in the shipped image).",
        "Under the extended profile, the overlay-supporting services are also up — verified by `docker ps --format '{{.Names}}' | grep -qx elasticsearch && docker ps --format '{{.Names}}' | grep -qx vss-video-analytics-api-mv3dt`. If the agent deliberately chose minimal, this check passes by default (judge: confirm by inspecting agent's `MINIMAL_PROFILE` value in trajectory).",
        "NGC credentials at `~/.ngc/config` are preserved — verified by `test -f ~/.ngc/config`.",
        "The final response does not contain plaintext API tokens matching the pattern `(Bearer |sk-|glpat-|nvapi-)[A-Za-z0-9+/=_-]{10,}`."
      ]
    },
    {
      "query": "Verify the deployment is healthy and data is flowing.",
      "checks": [
        "The agent uses `docker inspect --format '{{.State.Health.Status}}' vss-rtvi-cv-bev-fusion` (or equivalent) to check fusion health — NOT `docker exec ... test -f /tmp/fusion_ready` which fails on the shipped image.",
        "The agent uses `kafka-get-offsets` (or `redis-cli XLEN` when `STREAM_TYPE=redis`) to check broker offsets — NOT `kafka-run-class kafka.tools.GetOffsetShell` which raises `ClassNotFoundException` on `confluentinc/cp-kafka:8.2.0`.",
        "Broker offsets are actually growing — `mdx-raw` and `mdx-bev` topics both show offset > 0 by the time the agent finishes. Judge: run `docker exec kafka kafka-get-offsets --bootstrap-server localhost:9092 --topic mdx-bev` and confirm a numeric offset is reported.",
        "The agent surfaces the VST video wall URL at `http://<HOST_IP>:30888/vst` (or via HAProxy if extended deploys it) and explains that overlays render only when the extended profile / Elasticsearch is up.",
        "The final response does not contain plaintext API tokens matching the pattern `(Bearer |sk-|glpat-|nvapi-)[A-Za-z0-9+/=_-]{10,}`."
      ]
    },
    {
      "query": "Tear down the MV3DT deployment.",
      "checks": [
        "After the agent completes the teardown, none of the MV3DT containers are running — verified by `docker ps --filter status=running --format '{{.Names}}' | grep -E 'vss-rtvi-cv-mv3dt|vss-rtvi-cv-bev-fusion|mosquitto|vss-vios-sensor|vss-configurator-mv3dt' | wc -l | grep -qE '^0$'`.",
        "The MV3DT named volumes are removed — verified by `docker volume ls -q | grep -qE '^mdx_mdx-kafka$|^mdx_vios_pg_data$' && exit 1 || exit 0`.",
        "The trajectory shows the agent running `docker compose ... down` (or equivalent) against the warehouse-operations compose file (any of `docker compose down`, `compose.yml`, or `industry-profiles/warehouse-operations/.env` appears in the trajectory).",
        "The `.env` file is preserved — verified by `test -f deploy/docker/industry-profiles/warehouse-operations/.env`. The agent must not `git checkout` or otherwise wipe the user's env edits during teardown.",
        "NGC credentials at `~/.ngc/config` are preserved — verified by `test -f ~/.ngc/config`.",
        "The final response does not contain plaintext API tokens matching the pattern `(Bearer |sk-|glpat-|nvapi-)[A-Za-z0-9+/=_-]{10,}`."
      ]
    }
  ]
}