AGENT SKILL

RTVI-CV Detection & Tracking 2D

暂无简介

License 许可
MIT
所属仓库
rhCat/skillChip
最近更新
2026-07-31
Stars
★ 0
Forks
0

01 前言 Front Matter

name
RTVI-CV Detection & Tracking 2D
perks
[ "deploy", "discover-streams", "render-box", "collect-metrics", "synthesize-docker-run", "check-container-gpu", "clean-cache" ]
skill
vss-deploy-detection-tracking-2d

02 资源文件 78

03 SKILL.md 正文

vss-deploy-detection-tracking-2d — RTVI-CV Detection & Tracking 2D

Deploy, debug, and operate the RTVI-CV (Real Time Video Intelligence CV) 2D detection / tracking microservice and drive its REST API.

Perks

perk tool(s) nature
deploy load_defaults, fetch_resources, apply_in_container, start_app_in_container, add_streams destructive (runs Docker, pulls NGC images, launches a GPU perception app, adds live streams)
discover-streams discover_streams read-only — deterministic, layout-agnostic enumeration of local .mp4/.mkv videos into exactly STREAM_COUNT unique (id, url) stream descriptors
render-box render_box pure formatter — render a fixed-width light-box step receipt from body rows
collect-metrics collect_metrics read-only — sample /api/v1/metrics (+ nvidia-smi) N times and print averaged GPU/CPU/RAM + per-stream FPS; degrades gracefully offline
synthesize-docker-run synthesize_docker_run read-only — reconstruct the full docker run … command for an existing container from docker inspect, secrets redacted
check-container-gpu check_container_gpu read-only — probe CUDA/NVML inside a running container (nvidia-smi -L via docker exec) to catch the stale-GPU-handle state
clean-cache clean_engine_cache tidy — relocate non-*.engine/*.plan files out of the TRT engine cache into .quarantine/; idempotent, never deletes

The deploy perk runs the full DEPLOY workflow as an ordered sequence: resolve per-use-case defaults (load_defaults, the read-only entry step) → fetch + extract NGC model/video assets (fetch_resources) → apply in-container pipeline configuration (apply_in_container) → launch the perception app and wait for readiness (start_app_in_container) → add streams over REST (add_streams). Because it pulls images, mutates a container, and launches a GPU service, it is declared destructive: true and the executor gates it accordingly.

The remaining perks are the skill's independent, standalone operations — each invocable on its own, each emitting one structured-JSON audit line and writing its artifact under record_store. discover-streams (filesystem scan), render-box (stdin→box), and clean-cache (cache tidy, with DRY_RUN=1) run fully offline; collect-metrics, synthesize-docker-run, and check-container-gpu target a running instance / container and degrade gracefully (placeholder artifact, exit 0) when the REST endpoint, Docker, or GPU is unavailable.

How to use it

Copy ledger.json -> task-ledger.json, set the vars (USECASE, CONTAINER_NAME, STREAM_COUNT) + record_store, then validate -> compose -> compile -> oversight -> executor.

Localized from NVIDIA/skills vss-deploy-detection-tracking-2d (Apache-2.0).

04 同仓库更多 Skills