[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7EDRFFB7KBQKFRGPFBB16W4":3,"skill-same:01M7EDRFFB7KBQKFRGPFBB16W4":105},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":29,"commitSha":30,"skillDirectory":31,"forksCount":32,"repositoryHtmlUrl":33,"repositoryHomepageUrl":6,"ownerHtmlUrl":34,"ownerBlogUrl":6,"ownerDescription":35,"updatedAt":36,"install":37,"resources":50},"01M7EDRFFB7KBQKFRGPFBB16W4","vss-summarize-video",null,"Use when summarizing a recorded video through HITL-gated LVS, with an explicitly approved VLM fallback. Not for reports, archive search, or live RTSP captioning.","lza6\u002FArmor-piercing-system-prompts","lza6","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F128880206?v=4","01M7EB3K7WPX5T28EJ2NVHE6EG","Chinese-tingfeng","01M7EB3K97E1KQ16QWKJMFNGX8",11,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: vss-summarize-video\ndescription: Use when summarizing a recorded video through HITL-gated LVS, with an explicitly approved VLM fallback. Not for reports, archive search, or live RTSP captioning.\nlicense: Apache-2.0\nmetadata:\n  version: \"3.2.2\"\n  author: \"NVIDIA Video Search and Summarization team\"\n  github-url: \"https:\u002F\u002Fgithub.com\u002FNVIDIA-AI-Blueprints\u002Fvideo-search-and-summarization\"\n  tags: \"nvidia blueprint operational\"\n---\n\n# VSS Summarize Video\n\n## Instructions\n\n- Execute the five workflow stages below in order.\n- Run API commands yourself; do not tell the user to run them.\n- Use the required references at their named decision points.\n\n## Examples\n\nRunnable scenarios live under `evals\u002F`. The command implementations are in\n[`references\u002Fend-to-end-example.md`](references\u002Fend-to-end-example.md).\n\n## When to Use\n\nUse when summarizing a recorded video through HITL-gated LVS, to produce one\npolished narrative summary with timestamped events when LVS is available.\n\nDo not use this skill for:\n\n- Live RTSP captioning: use `vss-deploy-dense-captioning`.\n- Incident or alert-window reports: use `vss-generate-video-report` Mode B.\n- Archive search: use `vss-search-archive`.\n\n## Required References\n\nLoad these files only as directed:\n\n- [`references\u002Fend-to-end-example.md`](references\u002Fend-to-end-example.md): load\n  before executing the recorded-video workflow. It contains the exact\n  readiness, VIOS preparation, single-run summarize, and VLM fallback commands.\n- [`references\u002Fcli_usage.md`](references\u002Fcli_usage.md): load before Stage 4.\n  `vss summarize run` issues the summarize request and applies the operator's\n  configured memory policy; this reference has its flags, exit codes, output\n  shape, and read verbs.\n- [`references\u002Fvideo-summarization-api.md`](references\u002Fvideo-summarization-api.md):\n  load before constructing a live LVS operation **by hand** — a direct API\n  question, or the approved VLM fallback. Follow its **Runtime OpenAPI\n  Discovery** procedure on Docker. On Kubernetes, follow the K8s note there —\n  stock LVS Ingress does not publish LVS `\u002Fopenapi.json`. The ordered\n  workflow does not build a summarize payload; the CLI owns that.\n- [`references\u002Fhitl-prompts.md`](references\u002Fhitl-prompts.md): load when\n  collecting LVS scenario, events, and optional objects of interest.\n- [`references\u002Fvideo-summarization-debugging.md`](references\u002Fvideo-summarization-debugging.md):\n  load only when diagnosing a failed or empty response.\n- [`references\u002Fvideo-summarization-deployment.md`](references\u002Fvideo-summarization-deployment.md):\n  load only for deployment, configuration, logs, or service operations.\n- [`references\u002Fvideo-summarization-environment-variables.md`](references\u002Fvideo-summarization-environment-variables.md)\n  and `assets\u002Fvideo-summarization.env.example`: use when configuring the\n  service environment.\n- [`..\u002Fvss-build-vision-ai\u002Freferences\u002Fdeployment_resolution.md`](..\u002F..\u002Fvss-build-vision-ai\u002Freferences\u002Fdeployment_resolution.md):\n  Kubernetes `VSS_PUBLIC_URL` contract and the `\u002Flvs` mount.\n- [`references\u002Fdeploy-lvs-service.md`](references\u002Fdeploy-lvs-service.md): load\n  when asked about LVS's own container image, GPU\u002FCPU\u002Fstorage sizing, or\n  deployment contract as a peer service (heavier than\n  `video-summarization-deployment.md`, which covers operating an already\n  running deployment).\n- [`references\u002Fintegrate-lvs-service.md`](references\u002Fintegrate-lvs-service.md):\n  load when another agent or skill needs to integrate with LVS as a peer\n  service — required peers, integration interfaces, API schema, and network\n  requirements.\n\n## Core Invariants\n\n- Route by LVS readiness, never by video duration.\n- HTTP 200 from `\u002Fv1\u002Fready` selects LVS. Empty response bodies do not mean\n  unavailable.\n- Once LVS is selected, do not call a VLM `\u002Fv1\u002Fchat\u002Fcompletions` endpoint.\n- Issue exactly one `vss summarize run` per user summarize request. One run is\n  one `POST \u002Fv1\u002Fsummarize`. Never retry, hedge, broaden events, or run a second\n  backend automatically.\n- Endpoints come from the deployment `vss configure` recorded. Never pass an\n  endpoint, index, or model flag the caller did not name, and never replace a\n  failed run with hand-rolled curl against `\u002Fv1\u002Fsummarize`.\n- Save the complete command and its stdout. Diagnose failures from those files,\n  the run's own exit code, service logs, and non-mutating GET requests.\n- Render `video_summary` and every returned event verbatim. Do not paraphrase,\n  truncate descriptions, add fields, or fabricate `id`.\n- Direct VLM fallback requires explicit user approval unless the original\n  request pre-authorized it.\n\n## Prerequisites\n\n- VSS `lvs` profile reachable either on Docker (`$HOST_IP:38111`) or through\n  the public Ingress at `${VSS_PUBLIC_URL}\u002Flvs`.\n- `curl` and `jq` on the agent host.\n- Network reachability from the LVS service to the final VIOS clip URL (Docker:\n  from `vss-lvs`; Kubernetes: deploy must mint a URL the LVS pod can fetch).\n- A checkout containing `services\u002Fagent` and host `uv`, for `vss summarize run`.\n- One recorded deployment origin. Configure it once, before Stage 4:\n\n```bash\nVSS_REPO_ROOT=\"${VSS_REPO_ROOT:-$HOME\u002Fvideo-search-and-summarization}\"\ntest -f \"${VSS_REPO_ROOT}\u002Fservices\u002Fagent\u002Fpyproject.toml\" || {\n  echo \"VSS checkout not found at ${VSS_REPO_ROOT}; set VSS_REPO_ROOT explicitly\" >&2\n  exit 1\n}\nVSS=(uv run --project \"${VSS_REPO_ROOT}\u002Fservices\u002Fagent\" --no-dev --extra cli vss)\ncd \"${VSS_REPO_ROOT}\" && \"${VSS[@]}\" summarize run --help >\u002Fdev\u002Fnull || exit 1\n# Compose publishes the ingress on :7777; Kubernetes uses VSS_PUBLIC_URL.\nVSS_ORIGIN=\"${VSS_PUBLIC_URL:-http:\u002F\u002F${HOST_IP:-localhost}:7777}\"\n\"${VSS[@]}\" configure --base-url \"${VSS_ORIGIN%\u002F}\" || exit 1\n```\n\n`--extra cli` is mandatory: the base distribution holds the core libraries,\nwhile `nvidia-vss-cli` declares the `vss` executable. Configure against the\ningress origin, never `:38111` — that LVS container port exposes no\nElasticsearch, so a deployment recorded from it cannot persist.\n\nThe `vss-deploy-profile` skill can deploy the profile. A remote fallback VLM\nmust be able to fetch the clip URL; it generally cannot fetch localhost or\nprivate addresses.\n\n## Limitations\n\n- Direct VLM fallback cannot target LVS scenarios or events and is lower\n  quality.\n- Private VIOS URLs may be unreachable from remote VLM endpoints.\n- Each user request permits one `vss summarize run`, with no automatic retry.\n- Persistence needs a routed Elasticsearch. A deployment without one summarizes\n  and reports the result unpersisted rather than failing the job.\n- Both edges are configured to wait an hour, matching the CLI's own default, so\n  a long summarization is not cut short by a 504 that would be recorded as a\n  failed job. An Ingress the deployment overrides shorter still caps the wait.\n- Stock LVS Helm Ingress does not publish LVS `\u002Fmodels`, LVS `\u002Fopenapi.json`,\n  `\u002Frecommended_config`, or `\u002Fmetrics` — those remain Docker `:38111` only.\n\n## Endpoint resolution (Kubernetes vs Docker)\n\nResolve endpoints once before probing. Follow\n[`..\u002Fvss-build-vision-ai\u002Freferences\u002Fdeployment_resolution.md`](..\u002F..\u002Fvss-build-vision-ai\u002Freferences\u002Fdeployment_resolution.md).\n\n```bash\n# Prefer VSS_PUBLIC_URL; accept legacy VSS_ENDPOINT as the same public origin.\nif [ -z \"${VSS_PUBLIC_URL:-}\" ] && [ -n \"${VSS_ENDPOINT:-}\" ]; then\n  VSS_PUBLIC_URL=\"${VSS_ENDPOINT}\"\nfi\n\nif [ -n \"${VSS_PUBLIC_URL:-}\" ]; then\n  DEPLOYMENT_KIND=\"kubernetes\"\n  VSS_PUBLIC_URL=\"${VSS_PUBLIC_URL%\u002F}\"\n  # Force public origin — ignore leftover Docker LVS_BACKEND_URL \u002F VLM_* env.\n  # The \u002Flvs mount, not the origin — skill appends \u002Fv1\u002Fready and \u002Fv1\u002Fsummarize,\n  # and the gateway strips \u002Flvs before the backend sees them.\n  LVS_BACKEND_URL=\"${VSS_PUBLIC_URL}\u002Flvs\"\n  VIDEO_SUMMARIZATION_URL=\"${LVS_BACKEND_URL}\"\n  VSS_VIOS_URL=\"${VSS_PUBLIC_URL}\u002Fvst\"\n  VST_API_BASE=\"${VSS_VIOS_URL}\u002Fapi\u002Fv1\"\n  # RT-VLM is at its own mount; \u002Fv1\u002Fmodels and \u002Fv1\u002Fchat\u002Fcompletions hang off it.\n  VLM=\"${VSS_PUBLIC_URL}\u002Frtvi-vlm\"\nelse\n  DEPLOYMENT_KIND=\"docker\"\n  LVS_BACKEND_URL=\"${LVS_BACKEND_URL:-http:\u002F\u002F${HOST_IP:-localhost}:38111}\"\n  VIDEO_SUMMARIZATION_URL=\"${LVS_BACKEND_URL}\"\n  VSS_VIOS_URL=\"http:\u002F\u002F${HOST_IP:-localhost}:30888\u002Fvst\"\n  VST_API_BASE=\"${VSS_VIOS_URL}\u002Fapi\u002Fv1\"\n  VLM=\"${VLM_BASE_URL:-${RTVI_VLM_BASE_URL:-http:\u002F\u002F${HOST_IP:-localhost}:8018}}\"\n  VLM=\"${VLM%\u002Fv1}\"\nfi\n```\n\nOn Kubernetes, do not use `kubectl port-forward`, Service DNS, NodePorts,\n`docker exec`, or `docker inspect`. Do not append `\u002Fv1` to `LVS_BACKEND_URL`.\nIgnore Docker-derived `LVS_BACKEND_URL` \u002F `VLM_BASE_URL` \u002F `RTVI_VLM_BASE_URL`\nwhen `VSS_PUBLIC_URL` is set. Do not treat public `\u002Fopenapi.json` as the LVS\nschema (that path is Agent on stock Ingress).\n\n## Routing\n\n| Service | Base URL |\n|---|---|\n| LVS | `${VIDEO_SUMMARIZATION_URL}` (K8s: `${VSS_PUBLIC_URL}`; Docker: `http:\u002F\u002F${HOST_IP}:38111`) |\n| VLM \u002F RT-VLM | `${VLM}` then append `\u002Fv1\u002F...` (K8s: the `\u002Frtvi-vlm` mount; Docker: `:8018`) |\n| VIOS | `${VST_API_BASE}` |\n\nStrip a trailing `\u002Fv1` from the VLM base because this skill appends it. Do not\nscan ports or inspect configuration files to guess endpoints.\n\nProbe LVS `\u002Fv1\u002Fready` using the loop in the end-to-end reference. Readiness is\nthe HTTP status only: retry 503 warmup responses for about 30 seconds, and do\nnot inspect the body.\n\n| LVS result | Action |\n|---|---|\n| HTTP 200 | Use LVS for every video duration. |\n| Anything else | Ask to deploy LVS or ask before using VLM fallback. |\n\nIf LVS is unavailable, ask:\n\n> The VSS `lvs` profile isn't reachable\n> (`${VSS_PUBLIC_URL:-$HOST_IP:38111}`). Shall I deploy it now using\n> `\u002Fvss-deploy-profile -p lvs`? Reply `no` to stop here; I can use the\n> lower-quality VLM-only fallback only if you explicitly ask for it.\n\n- Deployment approved or pre-authorized: invoke `vss-deploy-profile`, re-probe,\n  and continue only after LVS returns 200.\n- Deployment declined: ask separately whether to use VLM fallback. Stop unless\n  the user approves it.\n- Fallback pre-authorized: use the fallback without another prompt.\n- Non-interactive run: the original task is the only approval source. If it\n  pre-authorizes neither deployment nor fallback, report blocked and stop.\n\n## Recorded Video Workflow\n\n### Stage 1: Select the Backend\n\nLoad the end-to-end and CLI references. Run the LVS readiness probe before\npreparing the clip. Also probe VLM `\u002Fv1\u002Fmodels` so an approved fallback can be\nvalidated, but do not infer against it while LVS is ready.\n\nThe summarization model needs no discovery: `vss configure` recorded the id LVS\nreports serving, and `vss summarize run` defaults to it on both Docker and\nKubernetes. Pass `--model` only when the caller named one, and read the recorded\nvalue from `vss configure show` when it has to be reported.\n\nDiscover a model id by hand only for an approved VLM fallback, which does not\nrun through the CLI:\n\n- **Docker:** honor `${VLM_NAME}` only if it matches an id from LVS `GET \u002Fmodels`;\n  otherwise use the sole advertised LVS id.\n- **Kubernetes:** prefer `${VLM_NAME}` when set; otherwise take the sole id from\n  `GET ${LVS_BACKEND_URL}\u002Fmodels` (LVS) or `GET ${VLM}\u002Fv1\u002Fmodels` (RT-VLM). If\n  multiple ids exist and no valid preference selects one, report them and stop.\n\nA non-200 LVS readiness result after warmup is the only unavailability signal.\nAn empty summary, empty events, missing optional fields, or empty readiness\nstdout must not trigger fallback.\n\n### Stage 2: Prepare the Video Through VIOS\n\nExecute VIOS API operations directly as part of this workflow; do not invoke a\nseparate skill. Follow **Prepare the video through VIOS** in the end-to-end\nreference (uses `${VST_API_BASE}`).\n\n1. List sensors and reuse the exact requested recording when present.\n2. If absent and the exact local file is available, upload it through the VIOS\n   file API. For uploaded or sample media without a requested timestamp, use\n   `2025-01-01T00:00:00.000Z` so timeline resolution is deterministic.\n3. Poll the returned stream's timelines and obtain the complete minimum start\n   and maximum end time.\n4. Generate a fresh temporary MP4 URL for that full interval with audio\n   disabled. Pass that minted URL to `--url` **as returned** (after stripping a\n   doubled `http:\u002F\u002F` scheme if present). Do not rewrite it for browser Ingress\n   paths before the summarize run.\n5. If LVS was selected, verify one-byte reachability:\n   - **Docker:** `docker exec vss-lvs` Python range probe in the reference.\n   - **Kubernetes:** bounded Range GET of the minted URL from the agent host\n     (no `docker exec` \u002F `kubectl exec`). Deploy must mint a URL the LVS pod\n     can fetch.\n\nRequire the exact recording, full timeline, and fresh clip URL before\ncontinuing. When the source file is available, compare VIOS timeline duration\nwith source duration. An upload response or byte probe proves reachability, not\ncomplete media readiness.\n\nIf preparation fails, stop and report the missing prerequisite. Do not choose\nan arbitrary `\u002Ftmp` video, alternate recording, local HTTP server, NvStreamer,\nor RTSP source unless the user explicitly requested that source.\n\nDo not use the `vss-lvs` container's lightweight `curl` shim for reachability;\nit can write the entire video into tool output. Use the one-byte Python probe\non Docker.\n\n### Stage 3: Collect LVS Settings\n\nWhen LVS is selected, load the HITL reference and collect `scenario`, `events`,\nand optional `objects_of_interest` before the summarize run.\n\nWhen the caller explicitly says to run autonomously without prompting and asks\nfor defaults or supplies no settings, use these values verbatim:\n\n```text\nscenario=\"activity monitoring\"\nevents=[\"notable activity\"]\n```\n\nThis is the only HITL bypass. Do not infer defaults from filenames or sensor\nnames. Mention defaults in the final response and offer a separate rerun with\nspecific settings.\n\n### Stage 4: Submit Once Through the CLI\n\nLoad the CLI reference. `vss summarize run` issues the summarize request on both\nDocker and Kubernetes, and persists only when static memory policy enables it.\nDo not build a `\u002Fv1\u002Fsummarize` payload by hand, and do not fetch\n`\u002Fopenapi.json` to construct one — the CLI owns the request shape,\n`vss configure` owns the endpoint, and `vss configure memory` owns persistence.\n\nUse the invocation in the end-to-end reference. It passes the fresh VIOS URL\nfrom Stage 2, the exact HITL values from Stage 3, `--chunk-duration 10`, and\n`--seed 1`; repeat `--event` per event and add `--object-of-interest` only when\nthe caller provided objects. Pass no endpoint flag.\n\nDo not pass `--persist` or `--memory-index`; those per-request controls do not\nexist. The standard workflow also does not pass `--no-persist`, so the\noperator's configured persistence default applies. When persistence is enabled,\nthe record needs two values:\n\n- `--video-id`, required alongside `--url`. Use the recording's VIOS **sensor**\n  id — from `sensor\u002Flist`, or from the `sensorId` an upload returns — never the\n  stream id. It becomes the record's sensor, which is what `list --sensor-id`\n  and time-windowed recall key on. Without `--video-id` the run exits 2 before\n  summarizing rather than after.\n- `--creation-time`, the media's absolute start. LVS reports event times as\n  offsets into the clip unless this anchors them, and unified memory stores\n  instants — so without it the events cannot be written and the run degrades to\n  exit 6. For uploaded sample media use the same `2025-01-01T00:00:00.000Z`\n  Stage 2 uploaded with.\n\nDo not pass `--num-frames-per-chunk` in the standard workflow. RT-VLM owns frame\nsampling; unset fields are absent from the request, so the deployment's own\ndefault applies.\n\nThe final line of stdout is one JSON object naming the job. Read that line and\nthe exit code; the prose on stderr is a diagnostic, not the result. A call\nrefused before a job exists prints no marker at all and stderr is the whole\nresult — check stdout is non-empty before parsing it. Emptiness, not the exit\ncode, is what says whether a job was created.\n\n| exit | meaning | action |\n|---|---|---|\n| 0 | summarized; persistence followed configured policy | present the result and report `persisted` truthfully |\n| 2 | a flag the CLI refused, before anything was submitted | fix the call, then run once |\n| 2 | LVS rejected the request it was sent; the marker names a job closed as failed | report the failure with that `job_id` |\n| 3 | LVS unreachable or returned 5xx | report it with the marker's `job_id` |\n| 4 | deployment configuration is missing, or an explicit Markdown note lacks static sink configuration | no job, no marker — run the remediation command from stderr |\n| 6 | summary produced; Elasticsearch or Markdown cache write failed | present the summary; report ES and Markdown outcomes separately |\n| 7 | timed out | reconcile with `vss summarize get --job-id`; do not re-run |\n\nExits 6 and 7 both mean the summarization already happened. Never repeat the run\nto obtain a different view of it, and never repeat it for diagnosis — a second\nrun requires a separate user request. The exit 2 that carries a marker is the\nsame story earlier: the request reached LVS and came back refused, so the one\nsubmission this request had is spent and a corrected call belongs to a new\nrequest. Once a job exists, every outcome names its `job_id`; use it rather than\nre-running. A call refused before a job exists names\nnothing, which is why empty stdout is the test.\n\nThe completion marker's `persisted` boolean is the authoritative Elasticsearch\noutcome. The result body includes `persist` only when persistence was attempted\nand reports its index and event count; optional Markdown status is separate\nunder `memory_note`. `record` says what the `job_id` is worth to a later read:\n`closed`, `absent` when policy skipped persistence, or `stale` when a submitted\nrecord could not be closed. Do not read the record back to confirm it, and never\nread Elasticsearch directly — recalling memory is a separate skill's job. The\none read that belongs here is reconciling an exit 7, whose outcome is genuinely\nunknown until `vss summarize get --job-id \u003Cjob_id>` answers.\n\nIf `video_summary` and `events` are empty, inspect the same payload's\n`summary.usage.total_chunks_processed`. A positive integer confirms processing;\nzero or missing means processing was not confirmed. Do not claim \"no\ndetections.\"\n\n### VLM Fallback for Stages 3-4\n\nUse the fallback command in the end-to-end reference only when LVS remained\nunavailable after warmup and the user explicitly approved fallback. Do not run\nLVS HITL, and never use fallback to repair or replace an LVS response.\n\nBefore the result, include:\n\n> **Note:** Input video `\u003Cname>` is `\u003CN>`s long. The video summarization\n> service is not deployed, so this summary was produced by the VLM alone with\n> a generic default prompt. Deploy the `lvs` profile for higher-quality\n> summaries with scenario\u002Fevents targeting.\n\nIf the VLM cannot fetch the VIOS URL, report that blocker instead of sending\nan inference request.\n\n### Stage 5: Present the Result\n\nStart with exactly one header:\n\n```text\nSummary of \u003Cvideo_name> (\u003Cduration>)\n```\n\nUse `Ns` below 60 seconds and `Mm Ss` otherwise.\n\nFor LVS, the CLI nests the service's own envelope under `summary`: parse the\nJSON string in `summary.choices[0].message.content` while preserving\n`summary.usage`. Render `video_summary` verbatim, followed by every event in\nservice order. Preserve every returned field and the full `description`; use a\nper-event list if a table would truncate text.\n\nClose with the job's identity: the `job_id`, the completion marker's\n`persisted` value, and any separate `memory_note` result. An absent `persist`\nobject with `persisted=false` means static policy chose stdout-only execution,\nnot a failure.\n\nFor VLM, render `choices[0].message.content` verbatim. For Cosmos output, omit\nthe `\u003Cthink>...\u003C\u002Fthink>` block and show the answer. Do not add emojis or\nre-voice either backend's content.\n\n## Troubleshooting\n\n| Symptom | Action |\n|---|---|\n| `\u002Fv1\u002Fready` remains 503 | Treat LVS as unavailable after the warmup loop. |\n| Readiness stdout is empty | Use the HTTP status; a 200 body may be empty. |\n| Summary and events are empty | Inspect saved `summary.usage.total_chunks_processed`; do not retry. |\n| `vss` not found | Keep `--extra cli` and verify `VSS_REPO_ROOT`; never install globally. |\n| Run exits 4 | Follow stderr: configure the deployment, or configure the Markdown sink requested explicitly. |\n| Run exits 6 | A post-operation memory write failed. Present the summary and separate ES\u002FMarkdown status; do not re-run. |\n| Run exits 7 | Timed out. `vss summarize get --job-id \u003Cid>`; do not re-run. |\n| VLM returns `\u003Cthink>` | Remove reasoning through `\u003C\u002Fthink>` when rendering. |\n| K8s `\u002Fopenapi.json` looks like Agent | Expected — do not use it as LVS schema. |\n| K8s `\u002Fmodels` 404 \u002F HTML | Probing the bare origin — use `${LVS_BACKEND_URL}\u002Fmodels` or `${VLM}\u002Fv1\u002Fmodels`. |\n\nUse the debugging reference for deeper diagnostics and the deployment\nreference for logs or configuration. The LVS image is a multi-arch manifest, so\n`LVS_TAG=3.3.0-rc2` is the x86\u002FJetson Thor default; use `3.3.0-rc2-sbsa` on SBSA\u002FDGX Spark\u002FGrace. RT-VLM likewise needs a host-matched tag (`3.3.0-26.08.2` on x86\u002FJetson Thor, `3.3.0-26.08.2-sbsa` on SBSA\u002FDGX Spark\u002FGrace).\n\n## Direct API and Service Operations\n\nFor direct API questions such as models, readiness, recommended configuration,\nmetrics, schemas, or 422 responses, use the API reference instead of the\nrecorded-video workflow. `\u002Flvs` is a Prefix mount, so everything LVS serves is\npublic under it on Kubernetes — `\u002Flvs\u002Fv1\u002Fready`, `\u002Flvs\u002Fv1\u002Fsummarize`,\n`\u002Flvs\u002Fmodels`, `\u002Flvs\u002Fmetrics` — where the previous Exact-path Ingress published\nonly readiness and summarize. For deployment, restart, teardown, backend\nselection, or service logs, prefer `vss-deploy-profile` and use the deployment\nreference.\n\n## Cross-reference\n\n- `vss-deploy-profile`: deploy the `lvs` profile.\n- `vss-manage-video-io-storage`: general VIOS administration outside this\n  ordered workflow.\n- `vss-search-archive`: search archived video.\n- `vss-query-analytics`: query stored incidents and events.\n- `vss-generate-video-report`: Mode A delegates here when LVS `\u002Fv1\u002Fready` is 200.\n\nbump:3\n","# VSS Summarize Video\n\n## Instructions\n\n- Execute the five workflow stages below in order.\n- Run API commands yourself; do not tell the user to run them.\n- Use the required references at their named decision points.\n\n## Examples\n\nRunnable scenarios live under `evals\u002F`. The command implementations are in\n[`references\u002Fend-to-end-example.md`](references\u002Fend-to-end-example.md).\n\n## When to Use\n\nUse when summarizing a recorded video through HITL-gated LVS, to produce one\npolished narrative summary with timestamped events when LVS is available.\n\nDo not use this skill for:\n\n- Live RTSP captioning: use `vss-deploy-dense-captioning`.\n- Incident or alert-window reports: use `vss-generate-video-report` Mode B.\n- Archive search: use `vss-search-archive`.\n\n## Required References\n\nLoad these files only as directed:\n\n- [`references\u002Fend-to-end-example.md`](references\u002Fend-to-end-example.md): load\n  before executing the recorded-video workflow. It contains the exact\n  readiness, VIOS preparation, single-run summarize, and VLM fallback commands.\n- [`references\u002Fcli_usage.md`](references\u002Fcli_usage.md): load before Stage 4.\n  `vss summarize run` issues the summarize request and applies the operator's\n  configured memory policy; this reference has its flags, exit codes, output\n  shape, and read verbs.\n- [`references\u002Fvideo-summarization-api.md`](references\u002Fvideo-summarization-api.md):\n  load before constructing a live LVS operation **by hand** — a direct API\n  question, or the approved VLM fallback. Follow its **Runtime OpenAPI\n  Discovery** procedure on Docker. On Kubernetes, follow the K8s note there —\n  stock LVS Ingress does not publish LVS `\u002Fopenapi.json`. The ordered\n  workflow does not build a summarize payload; the CLI owns that.\n- [`references\u002Fhitl-prompts.md`](references\u002Fhitl-prompts.md): load when\n  collecting LVS scenario, events, and optional objects of interest.\n- [`references\u002Fvideo-summarization-debugging.md`](references\u002Fvideo-summarization-debugging.md):\n  load only when diagnosing a failed or empty response.\n- [`references\u002Fvideo-summarization-deployment.md`](references\u002Fvideo-summarization-deployment.md):\n  load only for deployment, configuration, logs, or service operations.\n- [`references\u002Fvideo-summarization-environment-variables.md`](references\u002Fvideo-summarization-environment-variables.md)\n  and `assets\u002Fvideo-summarization.env.example`: use when configuring the\n  service environment.\n- [`..\u002Fvss-build-vision-ai\u002Freferences\u002Fdeployment_resolution.md`](..\u002F..\u002Fvss-build-vision-ai\u002Freferences\u002Fdeployment_resolution.md):\n  Kubernetes `VSS_PUBLIC_URL` contract and the `\u002Flvs` mount.\n- [`references\u002Fdeploy-lvs-service.md`](references\u002Fdeploy-lvs-service.md): load\n  when asked about LVS's own container image, GPU\u002FCPU\u002Fstorage sizing, or\n  deployment contract as a peer service (heavier than\n  `video-summarization-deployment.md`, which covers operating an already\n  running deployment).\n- [`references\u002Fintegrate-lvs-service.md`](references\u002Fintegrate-lvs-service.md):\n  load when another agent or skill needs to integrate with LVS as a peer\n  service — required peers, integration interfaces, API schema, and network\n  requirements.\n\n## Core Invariants\n\n- Route by LVS readiness, never by video duration.\n- HTTP 200 from `\u002Fv1\u002Fready` selects LVS. Empty response bodies do not mean\n  unavailable.\n- Once LVS is selected, do not call a VLM `\u002Fv1\u002Fchat\u002Fcompletions` endpoint.\n- Issue exactly one `vss summarize run` per user summarize request. One run is\n  one `POST \u002Fv1\u002Fsummarize`. Never retry, hedge, broaden events, or run a second\n  backend automatically.\n- Endpoints come from the deployment `vss configure` recorded. Never pass an\n  endpoint, index, or model flag the caller did not name, and never replace a\n  failed run with hand-rolled curl against `\u002Fv1\u002Fsummarize`.\n- Save the complete command and its stdout. Diagnose failures from those files,\n  the run's own exit code, service logs, and non-mutating GET requests.\n- Render `video_summary` and every returned event verbatim. Do not paraphrase,\n  truncate descriptions, add fields, or fabricate `id`.\n- Direct VLM fallback requires explicit user approval unless the original\n  request pre-authorized it.\n\n## Prerequisites\n\n- VSS `lvs` profile reachable either on Docker (`$HOST_IP:38111`) or through\n  the public Ingress at `${VSS_PUBLIC_URL}\u002Flvs`.\n- `curl` and `jq` on the agent host.\n- Network reachability from the LVS service to the final VIOS clip URL (Docker:\n  from `vss-lvs`; Kubernetes: deploy must mint a URL the LVS pod can fetch).\n- A checkout containing `services\u002Fagent` and host `uv`, for `vss summarize run`.\n- One recorded deployment origin. Configure it once, before Stage 4:\n\n```bash\nVSS_REPO_ROOT=\"${VSS_REPO_ROOT:-$HOME\u002Fvideo-search-and-summarization}\"\ntest -f \"${VSS_REPO_ROOT}\u002Fservices\u002Fagent\u002Fpyproject.toml\" || {\n  echo \"VSS checkout not found at ${VSS_REPO_ROOT}; set VSS_REPO_ROOT explicitly\" >&2\n  exit 1\n}\nVSS=(uv run --project \"${VSS_REPO_ROOT}\u002Fservices\u002Fagent\" --no-dev --extra cli vss)\ncd \"${VSS_REPO_ROOT}\" && \"${VSS[@]}\" summarize run --help >\u002Fdev\u002Fnull || exit 1\n# Compose publishes the ingress on :7777; Kubernetes uses VSS_PUBLIC_URL.\nVSS_ORIGIN=\"${VSS_PUBLIC_URL:-http:\u002F\u002F${HOST_IP:-localhost}:7777}\"\n\"${VSS[@]}\" configure --base-url \"${VSS_ORIGIN%\u002F}\" || exit 1\n```\n\n`--extra cli` is mandatory: the base distribution holds the core libraries,\nwhile `nvidia-vss-cli` declares the `vss` executable. Configure against the\ningress origin, never `:38111` — that LVS container port exposes no\nElasticsearch, so a deployment recorded from it cannot persist.\n\nThe `vss-deploy-profile` skill can deploy the profile. A remote fallback VLM\nmust be able to fetch the clip URL; it generally cannot fetch localhost or\nprivate addresses.\n\n## Limitations\n\n- Direct VLM fallback cannot target LVS scenarios or events and is lower\n  quality.\n- Private VIOS URLs may be unreachable from remote VLM endpoints.\n- Each user request permits one `vss summarize run`, with no automatic retry.\n- Persistence needs a routed Elasticsearch. A deployment without one summarizes\n  and reports the result unpersisted rather than failing the job.\n- Both edges are configured to wait an hour, matching the CLI's own default, so\n  a long summarization is not cut short by a 504 that would be recorded as a\n  failed job. An Ingress the deployment overrides shorter still caps the wait.\n- Stock LVS Helm Ingress does not publish LVS `\u002Fmodels`, LVS `\u002Fopenapi.json`,\n  `\u002Frecommended_config`, or `\u002Fmetrics` — those remain Docker `:38111` only.\n\n## Endpoint resolution (Kubernetes vs Docker)\n\nResolve endpoints once before probing. Follow\n[`..\u002Fvss-build-vision-ai\u002Freferences\u002Fdeployment_resolution.md`](..\u002F..\u002Fvss-build-vision-ai\u002Freferences\u002Fdeployment_resolution.md).\n\n```bash\n# Prefer VSS_PUBLIC_URL; accept legacy VSS_ENDPOINT as the same public origin.\nif [ -z \"${VSS_PUBLIC_URL:-}\" ] && [ -n \"${VSS_ENDPOINT:-}\" ]; then\n  VSS_PUBLIC_URL=\"${VSS_ENDPOINT}\"\nfi\n\nif [ -n \"${VSS_PUBLIC_URL:-}\" ]; then\n  DEPLOYMENT_KIND=\"kubernetes\"\n  VSS_PUBLIC_URL=\"${VSS_PUBLIC_URL%\u002F}\"\n  # Force public origin — ignore leftover Docker LVS_BACKEND_URL \u002F VLM_* env.\n  # The \u002Flvs mount, not the origin — skill appends \u002Fv1\u002Fready and \u002Fv1\u002Fsummarize,\n  # and the gateway strips \u002Flvs before the backend sees them.\n  LVS_BACKEND_URL=\"${VSS_PUBLIC_URL}\u002Flvs\"\n  VIDEO_SUMMARIZATION_URL=\"${LVS_BACKEND_URL}\"\n  VSS_VIOS_URL=\"${VSS_PUBLIC_URL}\u002Fvst\"\n  VST_API_BASE=\"${VSS_VIOS_URL}\u002Fapi\u002Fv1\"\n  # RT-VLM is at its own mount; \u002Fv1\u002Fmodels and \u002Fv1\u002Fchat\u002Fcompletions hang off it.\n  VLM=\"${VSS_PUBLIC_URL}\u002Frtvi-vlm\"\nelse\n  DEPLOYMENT_KIND=\"docker\"\n  LVS_BACKEND_URL=\"${LVS_BACKEND_URL:-http:\u002F\u002F${HOST_IP:-localhost}:38111}\"\n  VIDEO_SUMMARIZATION_URL=\"${LVS_BACKEND_URL}\"\n  VSS_VIOS_URL=\"http:\u002F\u002F${HOST_IP:-localhost}:30888\u002Fvst\"\n  VST_API_BASE=\"${VSS_VIOS_URL}\u002Fapi\u002Fv1\"\n  VLM=\"${VLM_BASE_URL:-${RTVI_VLM_BASE_URL:-http:\u002F\u002F${HOST_IP:-localhost}:8018}}\"\n  VLM=\"${VLM%\u002Fv1}\"\nfi\n```\n\nOn Kubernetes, do not use `kubectl port-forward`, Service DNS, NodePorts,\n`docker exec`, or `docker inspect`. Do not append `\u002Fv1` to `LVS_BACKEND_URL`.\nIgnore Docker-derived `LVS_BACKEND_URL` \u002F `VLM_BASE_URL` \u002F `RTVI_VLM_BASE_URL`\nwhen `VSS_PUBLIC_URL` is set. Do not treat public `\u002Fopenapi.json` as the LVS\nschema (that path is Agent on stock Ingress).\n\n## Routing\n\n| Service | Base URL |\n|---|---|\n| LVS | `${VIDEO_SUMMARIZATION_URL}` (K8s: `${VSS_PUBLIC_URL}`; Docker: `http:\u002F\u002F${HOST_IP}:38111`) |\n| VLM \u002F RT-VLM | `${VLM}` then append `\u002Fv1\u002F...` (K8s: the `\u002Frtvi-vlm` mount; Docker: `:8018`) |\n| VIOS | `${VST_API_BASE}` |\n\nStrip a trailing `\u002Fv1` from the VLM base because this skill appends it. Do not\nscan ports or inspect configuration files to guess endpoints.\n\nProbe LVS `\u002Fv1\u002Fready` using the loop in the end-to-end reference. Readiness is\nthe HTTP status only: retry 503 warmup responses for about 30 seconds, and do\nnot inspect the body.\n\n| LVS result | Action |\n|---|---|\n| HTTP 200 | Use LVS for every video duration. |\n| Anything else | Ask to deploy LVS or ask before using VLM fallback. |\n\nIf LVS is unavailable, ask:\n\n> The VSS `lvs` profile isn't reachable\n> (`${VSS_PUBLIC_URL:-$HOST_IP:38111}`). Shall I deploy it now using\n> `\u002Fvss-deploy-profile -p lvs`? Reply `no` to stop here; I can use the\n> lower-quality VLM-only fallback only if you explicitly ask for it.\n\n- Deployment approved or pre-authorized: invoke `vss-deploy-profile`, re-probe,\n  and continue only after LVS returns 200.\n- Deployment declined: ask separately whether to use VLM fallback. Stop unless\n  the user approves it.\n- Fallback pre-authorized: use the fallback without another prompt.\n- Non-interactive run: the original task is the only approval source. If it\n  pre-authorizes neither deployment nor fallback, report blocked and stop.\n\n## Recorded Video Workflow\n\n### Stage 1: Select the Backend\n\nLoad the end-to-end and CLI references. Run the LVS readiness probe before\npreparing the clip. Also probe VLM `\u002Fv1\u002Fmodels` so an approved fallback can be\nvalidated, but do not infer against it while LVS is ready.\n\nThe summarization model needs no discovery: `vss configure` recorded the id LVS\nreports serving, and `vss summarize run` defaults to it on both Docker and\nKubernetes. Pass `--model` only when the caller named one, and read the recorded\nvalue from `vss configure show` when it has to be reported.\n\nDiscover a model id by hand only for an approved VLM fallback, which does not\nrun through the CLI:\n\n- **Docker:** honor `${VLM_NAME}` only if it matches an id from LVS `GET \u002Fmodels`;\n  otherwise use the sole advertised LVS id.\n- **Kubernetes:** prefer `${VLM_NAME}` when set; otherwise take the sole id from\n  `GET ${LVS_BACKEND_URL}\u002Fmodels` (LVS) or `GET ${VLM}\u002Fv1\u002Fmodels` (RT-VLM). If\n  multiple ids exist and no valid preference selects one, report them and stop.\n\nA non-200 LVS readiness result after warmup is the only unavailability signal.\nAn empty summary, empty events, missing optional fields, or empty readiness\nstdout must not trigger fallback.\n\n### Stage 2: Prepare the Video Through VIOS\n\nExecute VIOS API operations directly as part of this workflow; do not invoke a\nseparate skill. Follow **Prepare the video through VIOS** in the end-to-end\nreference (uses `${VST_API_BASE}`).\n\n1. List sensors and reuse the exact requested recording when present.\n2. If absent and the exact local file is available, upload it through the VIOS\n   file API. For uploaded or sample media without a requested timestamp, use\n   `2025-01-01T00:00:00.000Z` so timeline resolution is deterministic.\n3. Poll the returned stream's timelines and obtain the complete minimum start\n   and maximum end time.\n4. Generate a fresh temporary MP4 URL for that full interval with audio\n   disabled. Pass that minted URL to `--url` **as returned** (after stripping a\n   doubled `http:\u002F\u002F` scheme if present). Do not rewrite it for browser Ingress\n   paths before the summarize run.\n5. If LVS was selected, verify one-byte reachability:\n   - **Docker:** `docker exec vss-lvs` Python range probe in the reference.\n   - **Kubernetes:** bounded Range GET of the minted URL from the agent host\n     (no `docker exec` \u002F `kubectl exec`). Deploy must mint a URL the LVS pod\n     can fetch.\n\nRequire the exact recording, full timeline, and fresh clip URL before\ncontinuing. When the source file is available, compare VIOS timeline duration\nwith source duration. An upload response or byte probe proves reachability, not\ncomplete media readiness.\n\nIf preparation fails, stop and report the missing prerequisite. Do not choose\nan arbitrary `\u002Ftmp` video, alternate recording, local HTTP server, NvStreamer,\nor RTSP source unless the user explicitly requested that source.\n\nDo not use the `vss-lvs` container's lightweight `curl` shim for reachability;\nit can write the entire video into tool output. Use the one-byte Python probe\non Docker.\n\n### Stage 3: Collect LVS Settings\n\nWhen LVS is selected, load the HITL reference and collect `scenario`, `events`,\nand optional `objects_of_interest` before the summarize run.\n\nWhen the caller explicitly says to run autonomously without prompting and asks\nfor defaults or supplies no settings, use these values verbatim:\n\n```text\nscenario=\"activity monitoring\"\nevents=[\"notable activity\"]\n```\n\nThis is the only HITL bypass. Do not infer defaults from filenames or sensor\nnames. Mention defaults in the final response and offer a separate rerun with\nspecific settings.\n\n### Stage 4: Submit Once Through the CLI\n\nLoad the CLI reference. `vss summarize run` issues the summarize request on both\nDocker and Kubernetes, and persists only when static memory policy enables it.\nDo not build a `\u002Fv1\u002Fsummarize` payload by hand, and do not fetch\n`\u002Fopenapi.json` to construct one — the CLI owns the request shape,\n`vss configure` owns the endpoint, and `vss configure memory` owns persistence.\n\nUse the invocation in the end-to-end reference. It passes the fresh VIOS URL\nfrom Stage 2, the exact HITL values from Stage 3, `--chunk-duration 10`, and\n`--seed 1`; repeat `--event` per event and add `--object-of-interest` only when\nthe caller provided objects. Pass no endpoint flag.\n\nDo not pass `--persist` or `--memory-index`; those per-request controls do not\nexist. The standard workflow also does not pass `--no-persist`, so the\noperator's configured persistence default applies. When persistence is enabled,\nthe record needs two values:\n\n- `--video-id`, required alongside `--url`. Use the recording's VIOS **sensor**\n  id — from `sensor\u002Flist`, or from the `sensorId` an upload returns — never the\n  stream id. It becomes the record's sensor, which is what `list --sensor-id`\n  and time-windowed recall key on. Without `--video-id` the run exits 2 before\n  summarizing rather than after.\n- `--creation-time`, the media's absolute start. LVS reports event times as\n  offsets into the clip unless this anchors them, and unified memory stores\n  instants — so without it the events cannot be written and the run degrades to\n  exit 6. For uploaded sample media use the same `2025-01-01T00:00:00.000Z`\n  Stage 2 uploaded with.\n\nDo not pass `--num-frames-per-chunk` in the standard workflow. RT-VLM owns frame\nsampling; unset fields are absent from the request, so the deployment's own\ndefault applies.\n\nThe final line of stdout is one JSON object naming the job. Read that line and\nthe exit code; the prose on stderr is a diagnostic, not the result. A call\nrefused before a job exists prints no marker at all and stderr is the whole\nresult — check stdout is non-empty before parsing it. Emptiness, not the exit\ncode, is what says whether a job was created.\n\n| exit | meaning | action |\n|---|---|---|\n| 0 | summarized; persistence followed configured policy | present the result and report `persisted` truthfully |\n| 2 | a flag the CLI refused, before anything was submitted | fix the call, then run once |\n| 2 | LVS rejected the request it was sent; the marker names a job closed as failed | report the failure with that `job_id` |\n| 3 | LVS unreachable or returned 5xx | report it with the marker's `job_id` |\n| 4 | deployment configuration is missing, or an explicit Markdown note lacks static sink configuration | no job, no marker — run the remediation command from stderr |\n| 6 | summary produced; Elasticsearch or Markdown cache write failed | present the summary; report ES and Markdown outcomes separately |\n| 7 | timed out | reconcile with `vss summarize get --job-id`; do not re-run |\n\nExits 6 and 7 both mean the summarization already happened. Never repeat the run\nto obtain a different view of it, and never repeat it for diagnosis — a second\nrun requires a separate user request. The exit 2 that carries a marker is the\nsame story earlier: the request reached LVS and came back refused, so the one\nsubmission this request had is spent and a corrected call belongs to a new\nrequest. Once a job exists, every outcome names its `job_id`; use it rather than\nre-running. A call refused before a job exists names\nnothing, which is why empty stdout is the test.\n\nThe completion marker's `persisted` boolean is the authoritative Elasticsearch\noutcome. The result body includes `persist` only when persistence was attempted\nand reports its index and event count; optional Markdown status is separate\nunder `memory_note`. `record` says what the `job_id` is worth to a later read:\n`closed`, `absent` when policy skipped persistence, or `stale` when a submitted\nrecord could not be closed. Do not read the record back to confirm it, and never\nread Elasticsearch directly — recalling memory is a separate skill's job. The\none read that belongs here is reconciling an exit 7, whose outcome is genuinely\nunknown until `vss summarize get --job-id \u003Cjob_id>` answers.\n\nIf `video_summary` and `events` are empty, inspect the same payload's\n`summary.usage.total_chunks_processed`. A positive integer confirms processing;\nzero or missing means processing was not confirmed. Do not claim \"no\ndetections.\"\n\n### VLM Fallback for Stages 3-4\n\nUse the fallback command in the end-to-end reference only when LVS remained\nunavailable after warmup and the user explicitly approved fallback. Do not run\nLVS HITL, and never use fallback to repair or replace an LVS response.\n\nBefore the result, include:\n\n> **Note:** Input video `\u003Cname>` is `\u003CN>`s long. The video summarization\n> service is not deployed, so this summary was produced by the VLM alone with\n> a generic default prompt. Deploy the `lvs` profile for higher-quality\n> summaries with scenario\u002Fevents targeting.\n\nIf the VLM cannot fetch the VIOS URL, report that blocker instead of sending\nan inference request.\n\n### Stage 5: Present the Result\n\nStart with exactly one header:\n\n```text\nSummary of \u003Cvideo_name> (\u003Cduration>)\n```\n\nUse `Ns` below 60 seconds and `Mm Ss` otherwise.\n\nFor LVS, the CLI nests the service's own envelope under `summary`: parse the\nJSON string in `summary.choices[0].message.content` while preserving\n`summary.usage`. Render `video_summary` verbatim, followed by every event in\nservice order. Preserve every returned field and the full `description`; use a\nper-event list if a table would truncate text.\n\nClose with the job's identity: the `job_id`, the completion marker's\n`persisted` value, and any separate `memory_note` result. An absent `persist`\nobject with `persisted=false` means static policy chose stdout-only execution,\nnot a failure.\n\nFor VLM, render `choices[0].message.content` verbatim. For Cosmos output, omit\nthe `\u003Cthink>...\u003C\u002Fthink>` block and show the answer. Do not add emojis or\nre-voice either backend's content.\n\n## Troubleshooting\n\n| Symptom | Action |\n|---|---|\n| `\u002Fv1\u002Fready` remains 503 | Treat LVS as unavailable after the warmup loop. |\n| Readiness stdout is empty | Use the HTTP status; a 200 body may be empty. |\n| Summary and events are empty | Inspect saved `summary.usage.total_chunks_processed`; do not retry. |\n| `vss` not found | Keep `--extra cli` and verify `VSS_REPO_ROOT`; never install globally. |\n| Run exits 4 | Follow stderr: configure the deployment, or configure the Markdown sink requested explicitly. |\n| Run exits 6 | A post-operation memory write failed. Present the summary and separate ES\u002FMarkdown status; do not re-run. |\n| Run exits 7 | Timed out. `vss summarize get --job-id \u003Cid>`; do not re-run. |\n| VLM returns `\u003Cthink>` | Remove reasoning through `\u003C\u002Fthink>` when rendering. |\n| K8s `\u002Fopenapi.json` looks like Agent | Expected — do not use it as LVS schema. |\n| K8s `\u002Fmodels` 404 \u002F HTML | Probing the bare origin — use `${LVS_BACKEND_URL}\u002Fmodels` or `${VLM}\u002Fv1\u002Fmodels`. |\n\nUse the debugging reference for deeper diagnostics and the deployment\nreference for logs or configuration. The LVS image is a multi-arch manifest, so\n`LVS_TAG=3.3.0-rc2` is the x86\u002FJetson Thor default; use `3.3.0-rc2-sbsa` on SBSA\u002FDGX Spark\u002FGrace. RT-VLM likewise needs a host-matched tag (`3.3.0-26.08.2` on x86\u002FJetson Thor, `3.3.0-26.08.2-sbsa` on SBSA\u002FDGX Spark\u002FGrace).\n\n## Direct API and Service Operations\n\nFor direct API questions such as models, readiness, recommended configuration,\nmetrics, schemas, or 422 responses, use the API reference instead of the\nrecorded-video workflow. `\u002Flvs` is a Prefix mount, so everything LVS serves is\npublic under it on Kubernetes — `\u002Flvs\u002Fv1\u002Fready`, `\u002Flvs\u002Fv1\u002Fsummarize`,\n`\u002Flvs\u002Fmodels`, `\u002Flvs\u002Fmetrics` — where the previous Exact-path Ingress published\nonly readiness and summarize. For deployment, restart, teardown, backend\nselection, or service logs, prefer `vss-deploy-profile` and use the deployment\nreference.\n\n## Cross-reference\n\n- `vss-deploy-profile`: deploy the `lvs` profile.\n- `vss-manage-video-io-storage`: general VIOS administration outside this\n  ordered workflow.\n- `vss-search-archive`: search archived video.\n- `vss-query-analytics`: query stored incidents and events.\n- `vss-generate-video-report`: Mode A delegates here when LVS `\u002Fv1\u002Fready` is 200.\n\nbump:3\n",{"name":5,"license":15,"metadata":24,"description":7},{"tags":25,"author":26,"version":27,"github-url":28},"nvidia blueprint operational","NVIDIA Video Search and Summarization team","3.2.2","https:\u002F\u002Fgithub.com\u002FNVIDIA-AI-Blueprints\u002Fvideo-search-and-summarization","main","d69096e239f2a89aabc2cd078c47664a9d3b1e11","03-逆向工程\u002Fvss-summarize-video",3,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts","https:\u002F\u002Fgithub.com\u002Flza6","个人vx：Tf00798\r\n个人✈：@TFGY999","2026-10-08 14:29:22",{"source":8,"skillName":5,"commands":38,"prompt":48,"zipAvailable":49},[39,42,45],{"tool":40,"command":41},"npx","npx skills add lza6\u002FArmor-piercing-system-prompts --skill vss-summarize-video",{"tool":43,"command":44},"bunx","bunx skills add lza6\u002FArmor-piercing-system-prompts --skill vss-summarize-video",{"tool":46,"command":47},"pnpm","pnpm dlx skills add lza6\u002FArmor-piercing-system-prompts --skill vss-summarize-video","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002FSKILL.md",true,[51,57,61,65,69,73,77,81,85,89,93,97,101],{"path":52,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":55,"sourceHtmlUrl":56},"evals\u002Fevals.json","REFERENCE","参考文档",1534,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Fevals\u002Fevals.json",{"path":58,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":59,"sourceHtmlUrl":60},"evals\u002Flvs_api_ops.json",4093,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Fevals\u002Flvs_api_ops.json",{"path":62,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":63,"sourceHtmlUrl":64},"evals\u002Flvs_profile_summarize.json",10495,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Fevals\u002Flvs_profile_summarize.json",{"path":66,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":67,"sourceHtmlUrl":68},"references\u002Fcli_usage.md",9297,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Freferences\u002Fcli_usage.md",{"path":70,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":71,"sourceHtmlUrl":72},"references\u002Fdeploy-lvs-service.md",8431,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Freferences\u002Fdeploy-lvs-service.md",{"path":74,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":75,"sourceHtmlUrl":76},"references\u002Fend-to-end-example.md",13050,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Freferences\u002Fend-to-end-example.md",{"path":78,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":79,"sourceHtmlUrl":80},"references\u002Fhitl-prompts.md",4910,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Freferences\u002Fhitl-prompts.md",{"path":82,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":83,"sourceHtmlUrl":84},"references\u002Fintegrate-lvs-service.md",10416,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Freferences\u002Fintegrate-lvs-service.md",{"path":86,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":87,"sourceHtmlUrl":88},"references\u002Fvideo-summarization-api.md",11596,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Freferences\u002Fvideo-summarization-api.md",{"path":90,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":91,"sourceHtmlUrl":92},"references\u002Fvideo-summarization-debugging.md",3683,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Freferences\u002Fvideo-summarization-debugging.md",{"path":94,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":95,"sourceHtmlUrl":96},"references\u002Fvideo-summarization-deployment.md",12209,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Freferences\u002Fvideo-summarization-deployment.md",{"path":98,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":99,"sourceHtmlUrl":100},"references\u002Fvideo-summarization-environment-variables.md",9613,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Freferences\u002Fvideo-summarization-environment-variables.md",{"path":102,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":103,"sourceHtmlUrl":104},"skill-card.md",3878,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F03-逆向工程\u002Fvss-summarize-video\u002Fskill-card.md",[106,115,120,127,134,141],{"skillNo":107,"name":108,"nameZh":6,"description":109,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":110,"licenseFileUrl":6,"sourceHtmlUrl":111,"categories":112,"keywords":113,"publishedAt":114},"01M7FFG485KY83TD6HSC4YPCMC","obsidian-clipper-template-creator","Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.","MIT","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator\u002FSKILL.md",[],[],"2026-11-16 10:05:34",{"skillNo":116,"name":108,"nameZh":6,"description":109,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":110,"licenseFileUrl":6,"sourceHtmlUrl":117,"categories":118,"keywords":119,"publishedAt":114},"01M7FFG0T6AZ0STEJPSB3CX84K","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator-antigravity-awesome-skills-main\u002FSKILL.md",[],[],{"skillNo":121,"name":122,"nameZh":6,"description":123,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":110,"licenseFileUrl":6,"sourceHtmlUrl":124,"categories":125,"keywords":126,"publishedAt":114},"01M7FFFY0QEM14SKWPDYGMXM2P","python-api","Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP\u002FmAP semantics, and source-style import diagnostics.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Fpython-api\u002FSKILL.md",[],[],{"skillNo":128,"name":129,"nameZh":6,"description":130,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":110,"licenseFileUrl":6,"sourceHtmlUrl":131,"categories":132,"keywords":133,"publishedAt":114},"01M7FFFVZQZ2GJ0C0TGP7JHD20","file-evaluation","Folder-based PASCAL VOC AP\u002FmAP evaluation from ground-truth and detection text files.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Ffile-evaluation\u002FSKILL.md",[],[],{"skillNo":135,"name":136,"nameZh":6,"description":137,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":110,"licenseFileUrl":6,"sourceHtmlUrl":138,"categories":139,"keywords":140,"publishedAt":114},"01M7FFFSFNYSE77SJ6PE1CK2DA","object-detection-metrics","Guides legacy Object-Detection-Metrics PASCAL VOC AP\u002FmAP evaluation from detection text folders and direct Python API classes.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002FSKILL.md",[],[],{"skillNo":142,"name":143,"nameZh":6,"description":144,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":110,"licenseFileUrl":6,"sourceHtmlUrl":145,"categories":146,"keywords":147,"publishedAt":114},"01M7FFFP0GCK1RVXBCPQ8APBMT","nx-workspace-patterns","Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fnx-workspace-patterns\u002FSKILL.md",[],[]]