[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7CSZDPGPAZCEP9SRQ43CR2W":3,"skill-same:01M7CSZDPGPAZCEP9SRQ43CR2W":71},{"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":28,"commitSha":29,"skillDirectory":30,"forksCount":31,"repositoryHtmlUrl":32,"repositoryHomepageUrl":33,"ownerHtmlUrl":34,"ownerBlogUrl":6,"ownerDescription":35,"updatedAt":36,"install":37,"resources":50},"01M7CSZDPGPAZCEP9SRQ43CR2W","vss-generate-video-report",null,"Use this skill when producing a VSS analysis report — Mode A per-clip VLM, Mode B incident-range via video-analytics. Not for real-time alerts or ad-hoc Q&A.","YYC-Cube\u002FYYC3-Knowledge-Base","YYC-Cube","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F235006543?v=4","01M79J2PH4SWD2X0825X6PN748","yyc3","01M79J2PHVK3F6ZBYT8K2SCZCW",1,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-Skills-技能包\u002FYYC3-Skills-技能包-skills\u002Fskills\u002Fvss-generate-video-report\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: vss-generate-video-report\ndescription: Use this skill when producing a VSS analysis report — Mode A per-clip VLM, Mode B incident-range via video-analytics. Not for real-time alerts or ad-hoc Q&A.\nlicense: Apache-2.0\nmetadata:\n  version: \"3.2.0\"\n  github-url: \"https:\u002F\u002Fgithub.com\u002FNVIDIA-AI-Blueprints\u002Fvideo-search-and-summarization\"\n  tags: \"nvidia blueprint operational\"\n---\n\n# Report\n\nGenerate a video analysis report by routing to one of two backends — **never via** `POST \u002Fgenerate` on the VSS agent.\n\n| Mode | Trigger | Backend |\n|---|---|---|\n| **A. Video clip** | \"report on `\u003Csensor>`\", \"report on this video\", \"analyze warehouse_01.mp4\" | `\u002Fvss-manage-video-io-storage` → clip URL → **VLM chat\u002Fcompletions** |\n| **B. Incident range** | \"report on incidents from `\u003Ct1>` to `\u003Ct2>`\", \"report on alerts today\", \"what incidents happened on `\u003Csensor>` last hour\" | `\u002Fvss-query-analytics` → incident list → narrative report |\n\nIf the request is ambiguous (e.g. \"report on `\u003Csensor>`\" with no time range and no incident wording), default to **Mode A**. Ask only if the user mentions both a sensor and a time range.\n\n---\n\n## When to Use\n\n- \"Generate a report for this video\" \u002F \"for `\u003Csensor-id>`\" — **Mode A**\n- \"Create an analysis report on the uploaded video\" — **Mode A**\n- \"Report on incidents from 12:31Z to 12:32Z\" — **Mode B**\n- \"Summarize alerts on `\u003Csensor>` between `\u003Ct1>` and `\u003Ct2>`\" — **Mode B**\n\n---\n\n## Deployment prerequisite\n\n**Mode A** needs the VSS **base** profile (VST + VLM NIM).\n**Mode B** needs the VSS **alerts** profile (VA-MCP + Elasticsearch).\n\nProbe:\n\n```bash\n# Mode A — VST + VLM reachability\ncurl -sf --max-time 5 \"http:\u002F\u002F${HOST_IP}:30888\u002Fvst\u002Fapi\u002Fv1\u002Fsensor\u002Fversion\" >\u002Fdev\u002Fnull\n\n# Mode B — VA-MCP\ncurl -sf --max-time 5 \"http:\u002F\u002F${HOST_IP}:9901\u002F\" >\u002Fdev\u002Fnull\n```\n\nIf the probe fails, hand off to `\u002Fvss-deploy-profile` with `-p base` (Mode A) or `-p alerts` (Mode B). **Always** confirm the deploy with the user first.\n\n---\n\n## Browser-playable clip URL (always do this before embedding any clip in the report)\n\nVST returns clip URLs using the agent-internal `${HOST_IP}:30888` host:port. Those work in-cluster (VLM frame pulls, agent backend) but the user's browser cannot reach them. The deploy layer already exports the browser-facing host:port as `$VSS_PUBLIC_HOST` \u002F `$VSS_PUBLIC_PORT` (and scheme as `$VSS_PUBLIC_HTTP_PROTOCOL`) in every profile `.env` — Brev or bare-metal — so the rewrite is a one-liner:\n\n```bash\nBROWSER_CLIP_URL=$(echo \"$RAW_URL\" | sed -E \"s|^https?:\u002F\u002F[^\u002F]+|${VSS_PUBLIC_HTTP_PROTOCOL}:\u002F\u002F${VSS_PUBLIC_HOST}:${VSS_PUBLIC_PORT}|\")\n```\n\nApply it to **every clip URL surfaced in the rendered report** (Mode A Step 4 Clip URL row; Mode B per-incident clip sub-bullet). Leave the VLM `video_url` content block in Mode A Step 3 on the original internal URL — the VLM is in-cluster.\n\n---\n\n## Mode A — Report on a recorded video clip\n\n**If the VSS `lvs` profile is deployed** — `curl -sf --max-time 5 \"http:\u002F\u002F${HOST_IP}:38111\u002Fv1\u002Fready\"` returns HTTP 200 — run `\u002Fvss-summarize-video` to produce the summary, then paste its output into the report template in Step 4 and skip Steps 1–3 (the VLM-direct path). Run Steps 1–3 only when `\u002Fv1\u002Fready` is non-200.\n\n### Step 1 — Resolve the clip URL\n\nHand off to `\u002Fvss-manage-video-io-storage` to:\n\n1. List sensors and confirm the named `\u003Csensor-id>` exists (upload first if not).\n2. Fetch `\u002Fstorage\u002F\u003CstreamId>\u002Ftimelines` for the recorded range when the user did not supply `startTime` \u002F `endTime`.\n3. Request a clip URL:\n\n   ```bash\n   curl -s \"http:\u002F\u002F${HOST_IP}:30888\u002Fvst\u002Fapi\u002Fv1\u002Fstorage\u002Ffile\u002F\u003CstreamId>\u002Furl?startTime=\u003CstartTime>&endTime=\u003CendTime>&container=mp4&disableAudio=true\" | jq -r .videoUrl\n   ```\n\n   That gives a direct `mp4` URL that the VLM can pull frames from. Bind it to `VIDEO_URL` (used in-cluster by the VLM in Step 3) **and** rewrite to `BROWSER_CLIP_URL` for the Step 4 report template using the one-liner from *Browser-playable clip URL* above — the user's browser cannot reach `$VIDEO_URL` directly.\n   Mode A requires the selected VLM endpoint to be able to fetch `VIDEO_URL`.\n   Local NIM\u002FRT-VLM deployments normally can; remote endpoints generally cannot\n   fetch `localhost`, private `HOST_IP`, or VST-internal URLs. If the live\n   `VLM_ENDPOINT` is remote, surface that reachability requirement instead of\n   making a chat request that will fail after `\u002Fv1\u002Fmodels` succeeds.\n\n### Step 2 — Resolve VLM endpoint and model\n\nThe deploy may serve the VLM through either of two stacks. Both expose an OpenAI-compatible `chat\u002Fcompletions` API — pick whichever is live:\n\n| Backend | Env vars | Typical host endpoint | Picked when |\n|---|---|---|---|\n| **NIM Cosmos** | `VLM_BASE_URL`, `VLM_NAME` | `${VLM_BASE_URL}\u002Fv1` (no trailing `\u002Fv1` on the env var; the agent appends it) | `VLM_MODE` ∈ {`local`, `local_shared`, `remote`} **and** `VLM_BASE_URL` is non-empty |\n| **RT-VLM Cosmos** | `RTVI_VLM_BASE_URL`, `RTVI_VLM_MODEL_TO_USE` (model identifier on the RT-VLM side, e.g. `cosmos-reason2`) | `${RTVI_VLM_BASE_URL}\u002Fv1` — alerts default `http:\u002F\u002F${HOST_IP}:8018\u002Fv1`, base default `http:\u002F\u002F${HOST_IP}:30082\u002Fv1` (`RTVI_VLM_ENDPOINT`) | `VLM_MODE=none` **or** `VLM_BASE_URL` empty; also the only path for `warehouse` |\n\nRead the live values off the running agent container — do not guess:\n\n```bash\ndocker exec vss-agent env | grep -E '^(VLM_BASE_URL|VLM_NAME|VLM_MODE|RTVI_VLM_BASE_URL|RTVI_VLM_ENDPOINT|RTVI_VLM_MODEL_TO_USE)='\n```\n\nSelection rule:\n\n```bash\nif [ -n \"${VLM_BASE_URL}\" ] && [ \"${VLM_MODE}\" != \"none\" ]; then\n  VLM_ENDPOINT=\"${VLM_BASE_URL%\u002F}\u002Fv1\"\n  VLM_MODEL=\"${VLM_NAME}\"\nelse\n  VLM_ENDPOINT=\"${RTVI_VLM_ENDPOINT:-${RTVI_VLM_BASE_URL%\u002F}\u002Fv1}\"\n  VLM_MODEL=\"${RTVI_VLM_MODEL_TO_USE}\"\nfi\n```\n\nProbe `\u002Fv1\u002Fmodels` before sending a chat request to confirm the chosen endpoint is alive and the model is loaded:\n\n```bash\ncurl -sf --max-time 5 \"${VLM_ENDPOINT}\u002Fmodels\" | jq -r '.data[].id'\n```\n\nIf the probe fails or the listed ids don't include `${VLM_MODEL}`, fall back to the other backend (or surface the error — never silently pick a model that isn't on the server).\n\n### Step 3 — Call the VLM directly\n\nUse the OpenAI-compatible `chat\u002Fcompletions` endpoint with a `video_url` content block — the same payload shape `video_understanding` builds in `src\u002Fvss_agents\u002Ftools\u002Fvideo_understanding.py` (`_build_vlm_messages`):\n\n```bash\nPROMPT='Describe in detail what happens in the video, with timestamps (start–end in seconds from clip start) for each segment or event. Cover scenes, objects, people, vehicles, and notable actions.'\n\n# Cosmos Reason 2 reasoning prompt suffix — matches video_understanding.py for is_cosmos_reason2 + reasoning=true.\n# Drop this suffix for non-cosmos-reason2 VLMs.\nPROMPT=\"${PROMPT}\n\nAnswer the question using the following format:\n\n\u003Cthink>\nYour reasoning.\n\u003C\u002Fthink>\n\nWrite your final answer immediately after the \u003C\u002Fthink> tag.\"\n\ncurl -s -X POST \"${VLM_ENDPOINT}\u002Fchat\u002Fcompletions\" \\\n  -H \"Content-Type: application\u002Fjson\" \\\n  -d @- \u003C\u003CEOF | jq -r '.choices[0].message.content'\n{\n  \"model\": \"${VLM_MODEL}\",\n  \"messages\": [\n    {\n      \"role\": \"user\",\n      \"content\": [\n        {\"type\": \"text\", \"text\": $(jq -Rs . \u003C\u003C\u003C \"${PROMPT}\")},\n        {\"type\": \"video_url\", \"video_url\": {\"url\": \"${VIDEO_URL}\"}}\n      ]\n    }\n  ],\n  \"max_tokens\": 1024,\n  \"temperature\": 0.0\n}\nEOF\n```\n\nIf the VLM returns a `\u003Cthink>…\u003C\u002Fthink>` block (Cosmos Reason reasoning mode), keep only the text after `\u003C\u002Fthink>` as the report body.\n\n### Step 4 — Fill the Video Analysis Report template\n\n```markdown\n# Video Analysis Report\n\n## Basic Information\n\n| Field | Value |\n|-------|-------|\n| **Report Identifier** | vss_report_\u003CYYYYMMDD_HHMMSS> |\n| **Date of Analysis** | \u003CYYYY-MM-DD> |\n| **Time of Analysis** | \u003CHH:MM:SS> |\n| **Video Source** | \u003Csensor_id or filename> |\n| **Clip Range** | \u003CstartTime> – \u003CendTime> |\n| **Clip URL** | `\u003CBROWSER_CLIP_URL>` (apply the `$VSS_PUBLIC_HOST:$VSS_PUBLIC_PORT` rewrite — NEVER paste the raw `HOST_IP:30888` URL here) |\n| **VLM** | \u003CVLM_MODEL (NIM or RT-VLM)> |\n| **Analysis Request** | \u003Cuser's request> |\n\n## Analysis Results\n\n\u003CVLM output: timestamped caption \u002F summary>\n```\n\nReturn the rendered markdown to the user.\n\n---\n\n## Mode B — Report on incidents in a time range\n\n### Step 1 — Resolve the time range and (optionally) sensor\n\n- `start_time` \u002F `end_time` must be ISO 8601 UTC (`YYYY-MM-DDTHH:MM:SS.sssZ`). Resolve relative phrases (\"last hour\", \"today\") against the current host clock.\n- If the user names a sensor, capture it as `source` + `source_type=sensor`. Otherwise leave both unset for an all-sensors query.\n\n### Step 2 — Fetch incidents via `\u002Fvss-query-analytics`\n\nHand off to `\u002Fvss-query-analytics` (initialize → `tools\u002Fcall`) with:\n\n```json\n{\n  \"name\": \"video_analytics__get_incidents\",\n  \"arguments\": {\n    \"source\": \"\u003Csensor-id-or-omit>\",\n    \"source_type\": \"sensor\",\n    \"start_time\": \"\u003CISO>\",\n    \"end_time\": \"\u003CISO>\",\n    \"max_count\": 100,\n    \"includes\": [\"objectIds\", \"info\"]\n  }\n}\n```\n\nFor each incident keep: `id`, `sensorId`, `timestamp`, `end`, `category`, `place.name`, `info.verdict`, `info.reasoning`, `objectIds`, and the clip URL (commonly `info.clip_url`, `clip_url`, or whichever clip-pointer field the response carries). **Apply the `$VSS_PUBLIC_HOST:$VSS_PUBLIC_PORT` rewrite (see *Browser-playable clip URL* above) to every clip URL before pasting it into the report** — the raw value is a `HOST_IP:30888` URL the user's browser cannot reach.\n\n### Step 3 — Fill the Incident Range Report template\n\nGroup by sensor (or by category if no sensor scope), tally verdicts, list each incident as a bullet with timestamp \u002F category \u002F verdict \u002F reasoning.\n\n```markdown\n# Incident Range Report\n\n## Basic Information\n\n| Field | Value |\n|-------|-------|\n| **Report Identifier** | vss_report_\u003CYYYYMMDD_HHMMSS> |\n| **Range** | \u003Cstart_time> – \u003Cend_time> |\n| **Scope** | \u003Csensor_id> | all sensors |\n| **Total Incidents** | \u003CN> |\n| **Confirmed \u002F Rejected \u002F Unverified** | \u003Cc> \u002F \u003Cr> \u002F \u003Cu> |\n\n## Incidents\n\n### \u003Csensor_id_or_category>\n\n- **\u003Ctimestamp>** — \u003Ccategory> — verdict: **\u003Cconfirmed|rejected|unverified>**\n  - \u003Cinfo.reasoning (1–2 lines)>\n  - clip: `\u003Crewritten URL>` (omit row when the incident carries no clip URL — never paste a raw `HOST_IP:30888` URL)\n  - objects: \u003CobjectIds joined>\n- …\n\n## Summary\n\n\u003C2–4 sentences synthesizing what dominates the range — top categories, sensors with the most confirmed incidents, any clusters in time.>\n```\n\nIf `get_incidents` returns zero results, return a one-line report stating the range and scope produced no incidents — do not invent content and do not fall back to Mode A.\n\n---\n\n## Cross-Reference\n\n- **`\u002Fvss-manage-video-io-storage`** — sensor list, timelines, and clip URL for Mode A Step 1.\n- **`\u002Fvss-query-analytics`** — incident retrieval (and verdict \u002F reasoning enrichment) for Mode B Step 2.\n- **`\u002Fvss-ask-video`** — ad-hoc VLM Q&A on a single clip (not a structured report).\n- **`\u002Fvss-summarize-video`** — used by Mode A to produce the summary body when the `lvs` profile is deployed; the report template (Step 4) is still filled here.\n\n","# Report\n\nGenerate a video analysis report by routing to one of two backends — **never via** `POST \u002Fgenerate` on the VSS agent.\n\n| Mode | Trigger | Backend |\n|---|---|---|\n| **A. Video clip** | \"report on `\u003Csensor>`\", \"report on this video\", \"analyze warehouse_01.mp4\" | `\u002Fvss-manage-video-io-storage` → clip URL → **VLM chat\u002Fcompletions** |\n| **B. Incident range** | \"report on incidents from `\u003Ct1>` to `\u003Ct2>`\", \"report on alerts today\", \"what incidents happened on `\u003Csensor>` last hour\" | `\u002Fvss-query-analytics` → incident list → narrative report |\n\nIf the request is ambiguous (e.g. \"report on `\u003Csensor>`\" with no time range and no incident wording), default to **Mode A**. Ask only if the user mentions both a sensor and a time range.\n\n---\n\n## When to Use\n\n- \"Generate a report for this video\" \u002F \"for `\u003Csensor-id>`\" — **Mode A**\n- \"Create an analysis report on the uploaded video\" — **Mode A**\n- \"Report on incidents from 12:31Z to 12:32Z\" — **Mode B**\n- \"Summarize alerts on `\u003Csensor>` between `\u003Ct1>` and `\u003Ct2>`\" — **Mode B**\n\n---\n\n## Deployment prerequisite\n\n**Mode A** needs the VSS **base** profile (VST + VLM NIM).\n**Mode B** needs the VSS **alerts** profile (VA-MCP + Elasticsearch).\n\nProbe:\n\n```bash\n# Mode A — VST + VLM reachability\ncurl -sf --max-time 5 \"http:\u002F\u002F${HOST_IP}:30888\u002Fvst\u002Fapi\u002Fv1\u002Fsensor\u002Fversion\" >\u002Fdev\u002Fnull\n\n# Mode B — VA-MCP\ncurl -sf --max-time 5 \"http:\u002F\u002F${HOST_IP}:9901\u002F\" >\u002Fdev\u002Fnull\n```\n\nIf the probe fails, hand off to `\u002Fvss-deploy-profile` with `-p base` (Mode A) or `-p alerts` (Mode B). **Always** confirm the deploy with the user first.\n\n---\n\n## Browser-playable clip URL (always do this before embedding any clip in the report)\n\nVST returns clip URLs using the agent-internal `${HOST_IP}:30888` host:port. Those work in-cluster (VLM frame pulls, agent backend) but the user's browser cannot reach them. The deploy layer already exports the browser-facing host:port as `$VSS_PUBLIC_HOST` \u002F `$VSS_PUBLIC_PORT` (and scheme as `$VSS_PUBLIC_HTTP_PROTOCOL`) in every profile `.env` — Brev or bare-metal — so the rewrite is a one-liner:\n\n```bash\nBROWSER_CLIP_URL=$(echo \"$RAW_URL\" | sed -E \"s|^https?:\u002F\u002F[^\u002F]+|${VSS_PUBLIC_HTTP_PROTOCOL}:\u002F\u002F${VSS_PUBLIC_HOST}:${VSS_PUBLIC_PORT}|\")\n```\n\nApply it to **every clip URL surfaced in the rendered report** (Mode A Step 4 Clip URL row; Mode B per-incident clip sub-bullet). Leave the VLM `video_url` content block in Mode A Step 3 on the original internal URL — the VLM is in-cluster.\n\n---\n\n## Mode A — Report on a recorded video clip\n\n**If the VSS `lvs` profile is deployed** — `curl -sf --max-time 5 \"http:\u002F\u002F${HOST_IP}:38111\u002Fv1\u002Fready\"` returns HTTP 200 — run `\u002Fvss-summarize-video` to produce the summary, then paste its output into the report template in Step 4 and skip Steps 1–3 (the VLM-direct path). Run Steps 1–3 only when `\u002Fv1\u002Fready` is non-200.\n\n### Step 1 — Resolve the clip URL\n\nHand off to `\u002Fvss-manage-video-io-storage` to:\n\n1. List sensors and confirm the named `\u003Csensor-id>` exists (upload first if not).\n2. Fetch `\u002Fstorage\u002F\u003CstreamId>\u002Ftimelines` for the recorded range when the user did not supply `startTime` \u002F `endTime`.\n3. Request a clip URL:\n\n   ```bash\n   curl -s \"http:\u002F\u002F${HOST_IP}:30888\u002Fvst\u002Fapi\u002Fv1\u002Fstorage\u002Ffile\u002F\u003CstreamId>\u002Furl?startTime=\u003CstartTime>&endTime=\u003CendTime>&container=mp4&disableAudio=true\" | jq -r .videoUrl\n   ```\n\n   That gives a direct `mp4` URL that the VLM can pull frames from. Bind it to `VIDEO_URL` (used in-cluster by the VLM in Step 3) **and** rewrite to `BROWSER_CLIP_URL` for the Step 4 report template using the one-liner from *Browser-playable clip URL* above — the user's browser cannot reach `$VIDEO_URL` directly.\n   Mode A requires the selected VLM endpoint to be able to fetch `VIDEO_URL`.\n   Local NIM\u002FRT-VLM deployments normally can; remote endpoints generally cannot\n   fetch `localhost`, private `HOST_IP`, or VST-internal URLs. If the live\n   `VLM_ENDPOINT` is remote, surface that reachability requirement instead of\n   making a chat request that will fail after `\u002Fv1\u002Fmodels` succeeds.\n\n### Step 2 — Resolve VLM endpoint and model\n\nThe deploy may serve the VLM through either of two stacks. Both expose an OpenAI-compatible `chat\u002Fcompletions` API — pick whichever is live:\n\n| Backend | Env vars | Typical host endpoint | Picked when |\n|---|---|---|---|\n| **NIM Cosmos** | `VLM_BASE_URL`, `VLM_NAME` | `${VLM_BASE_URL}\u002Fv1` (no trailing `\u002Fv1` on the env var; the agent appends it) | `VLM_MODE` ∈ {`local`, `local_shared`, `remote`} **and** `VLM_BASE_URL` is non-empty |\n| **RT-VLM Cosmos** | `RTVI_VLM_BASE_URL`, `RTVI_VLM_MODEL_TO_USE` (model identifier on the RT-VLM side, e.g. `cosmos-reason2`) | `${RTVI_VLM_BASE_URL}\u002Fv1` — alerts default `http:\u002F\u002F${HOST_IP}:8018\u002Fv1`, base default `http:\u002F\u002F${HOST_IP}:30082\u002Fv1` (`RTVI_VLM_ENDPOINT`) | `VLM_MODE=none` **or** `VLM_BASE_URL` empty; also the only path for `warehouse` |\n\nRead the live values off the running agent container — do not guess:\n\n```bash\ndocker exec vss-agent env | grep -E '^(VLM_BASE_URL|VLM_NAME|VLM_MODE|RTVI_VLM_BASE_URL|RTVI_VLM_ENDPOINT|RTVI_VLM_MODEL_TO_USE)='\n```\n\nSelection rule:\n\n```bash\nif [ -n \"${VLM_BASE_URL}\" ] && [ \"${VLM_MODE}\" != \"none\" ]; then\n  VLM_ENDPOINT=\"${VLM_BASE_URL%\u002F}\u002Fv1\"\n  VLM_MODEL=\"${VLM_NAME}\"\nelse\n  VLM_ENDPOINT=\"${RTVI_VLM_ENDPOINT:-${RTVI_VLM_BASE_URL%\u002F}\u002Fv1}\"\n  VLM_MODEL=\"${RTVI_VLM_MODEL_TO_USE}\"\nfi\n```\n\nProbe `\u002Fv1\u002Fmodels` before sending a chat request to confirm the chosen endpoint is alive and the model is loaded:\n\n```bash\ncurl -sf --max-time 5 \"${VLM_ENDPOINT}\u002Fmodels\" | jq -r '.data[].id'\n```\n\nIf the probe fails or the listed ids don't include `${VLM_MODEL}`, fall back to the other backend (or surface the error — never silently pick a model that isn't on the server).\n\n### Step 3 — Call the VLM directly\n\nUse the OpenAI-compatible `chat\u002Fcompletions` endpoint with a `video_url` content block — the same payload shape `video_understanding` builds in `src\u002Fvss_agents\u002Ftools\u002Fvideo_understanding.py` (`_build_vlm_messages`):\n\n```bash\nPROMPT='Describe in detail what happens in the video, with timestamps (start–end in seconds from clip start) for each segment or event. Cover scenes, objects, people, vehicles, and notable actions.'\n\n# Cosmos Reason 2 reasoning prompt suffix — matches video_understanding.py for is_cosmos_reason2 + reasoning=true.\n# Drop this suffix for non-cosmos-reason2 VLMs.\nPROMPT=\"${PROMPT}\n\nAnswer the question using the following format:\n\n\u003Cthink>\nYour reasoning.\n\u003C\u002Fthink>\n\nWrite your final answer immediately after the \u003C\u002Fthink> tag.\"\n\ncurl -s -X POST \"${VLM_ENDPOINT}\u002Fchat\u002Fcompletions\" \\\n  -H \"Content-Type: application\u002Fjson\" \\\n  -d @- \u003C\u003CEOF | jq -r '.choices[0].message.content'\n{\n  \"model\": \"${VLM_MODEL}\",\n  \"messages\": [\n    {\n      \"role\": \"user\",\n      \"content\": [\n        {\"type\": \"text\", \"text\": $(jq -Rs . \u003C\u003C\u003C \"${PROMPT}\")},\n        {\"type\": \"video_url\", \"video_url\": {\"url\": \"${VIDEO_URL}\"}}\n      ]\n    }\n  ],\n  \"max_tokens\": 1024,\n  \"temperature\": 0.0\n}\nEOF\n```\n\nIf the VLM returns a `\u003Cthink>…\u003C\u002Fthink>` block (Cosmos Reason reasoning mode), keep only the text after `\u003C\u002Fthink>` as the report body.\n\n### Step 4 — Fill the Video Analysis Report template\n\n```markdown\n# Video Analysis Report\n\n## Basic Information\n\n| Field | Value |\n|-------|-------|\n| **Report Identifier** | vss_report_\u003CYYYYMMDD_HHMMSS> |\n| **Date of Analysis** | \u003CYYYY-MM-DD> |\n| **Time of Analysis** | \u003CHH:MM:SS> |\n| **Video Source** | \u003Csensor_id or filename> |\n| **Clip Range** | \u003CstartTime> – \u003CendTime> |\n| **Clip URL** | `\u003CBROWSER_CLIP_URL>` (apply the `$VSS_PUBLIC_HOST:$VSS_PUBLIC_PORT` rewrite — NEVER paste the raw `HOST_IP:30888` URL here) |\n| **VLM** | \u003CVLM_MODEL (NIM or RT-VLM)> |\n| **Analysis Request** | \u003Cuser's request> |\n\n## Analysis Results\n\n\u003CVLM output: timestamped caption \u002F summary>\n```\n\nReturn the rendered markdown to the user.\n\n---\n\n## Mode B — Report on incidents in a time range\n\n### Step 1 — Resolve the time range and (optionally) sensor\n\n- `start_time` \u002F `end_time` must be ISO 8601 UTC (`YYYY-MM-DDTHH:MM:SS.sssZ`). Resolve relative phrases (\"last hour\", \"today\") against the current host clock.\n- If the user names a sensor, capture it as `source` + `source_type=sensor`. Otherwise leave both unset for an all-sensors query.\n\n### Step 2 — Fetch incidents via `\u002Fvss-query-analytics`\n\nHand off to `\u002Fvss-query-analytics` (initialize → `tools\u002Fcall`) with:\n\n```json\n{\n  \"name\": \"video_analytics__get_incidents\",\n  \"arguments\": {\n    \"source\": \"\u003Csensor-id-or-omit>\",\n    \"source_type\": \"sensor\",\n    \"start_time\": \"\u003CISO>\",\n    \"end_time\": \"\u003CISO>\",\n    \"max_count\": 100,\n    \"includes\": [\"objectIds\", \"info\"]\n  }\n}\n```\n\nFor each incident keep: `id`, `sensorId`, `timestamp`, `end`, `category`, `place.name`, `info.verdict`, `info.reasoning`, `objectIds`, and the clip URL (commonly `info.clip_url`, `clip_url`, or whichever clip-pointer field the response carries). **Apply the `$VSS_PUBLIC_HOST:$VSS_PUBLIC_PORT` rewrite (see *Browser-playable clip URL* above) to every clip URL before pasting it into the report** — the raw value is a `HOST_IP:30888` URL the user's browser cannot reach.\n\n### Step 3 — Fill the Incident Range Report template\n\nGroup by sensor (or by category if no sensor scope), tally verdicts, list each incident as a bullet with timestamp \u002F category \u002F verdict \u002F reasoning.\n\n```markdown\n# Incident Range Report\n\n## Basic Information\n\n| Field | Value |\n|-------|-------|\n| **Report Identifier** | vss_report_\u003CYYYYMMDD_HHMMSS> |\n| **Range** | \u003Cstart_time> – \u003Cend_time> |\n| **Scope** | \u003Csensor_id> | all sensors |\n| **Total Incidents** | \u003CN> |\n| **Confirmed \u002F Rejected \u002F Unverified** | \u003Cc> \u002F \u003Cr> \u002F \u003Cu> |\n\n## Incidents\n\n### \u003Csensor_id_or_category>\n\n- **\u003Ctimestamp>** — \u003Ccategory> — verdict: **\u003Cconfirmed|rejected|unverified>**\n  - \u003Cinfo.reasoning (1–2 lines)>\n  - clip: `\u003Crewritten URL>` (omit row when the incident carries no clip URL — never paste a raw `HOST_IP:30888` URL)\n  - objects: \u003CobjectIds joined>\n- …\n\n## Summary\n\n\u003C2–4 sentences synthesizing what dominates the range — top categories, sensors with the most confirmed incidents, any clusters in time.>\n```\n\nIf `get_incidents` returns zero results, return a one-line report stating the range and scope produced no incidents — do not invent content and do not fall back to Mode A.\n\n---\n\n## Cross-Reference\n\n- **`\u002Fvss-manage-video-io-storage`** — sensor list, timelines, and clip URL for Mode A Step 1.\n- **`\u002Fvss-query-analytics`** — incident retrieval (and verdict \u002F reasoning enrichment) for Mode B Step 2.\n- **`\u002Fvss-ask-video`** — ad-hoc VLM Q&A on a single clip (not a structured report).\n- **`\u002Fvss-summarize-video`** — used by Mode A to produce the summary body when the `lvs` profile is deployed; the report template (Step 4) is still filled here.\n\n",{"name":5,"license":15,"metadata":24,"description":7},{"tags":25,"version":26,"github-url":27},"nvidia blueprint operational","3.2.0","https:\u002F\u002Fgithub.com\u002FNVIDIA-AI-Blueprints\u002Fvideo-search-and-summarization","main","87afe6ce96220accf50a345dea46784c8c9f210e","YYC3-Skills-技能包\u002FYYC3-Skills-技能包-skills\u002Fskills\u002Fvss-generate-video-report",0,"https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base","https:\u002F\u002Fkb.yyc3.top\u002F","https:\u002F\u002Fgithub.com\u002FYYC-Cube","万象归元于云枢 | 深栈智启新纪元","2026-11-10 10:37:32",{"source":8,"skillName":5,"commands":38,"prompt":48,"zipAvailable":49},[39,42,45],{"tool":40,"command":41},"npx","npx skills add YYC-Cube\u002FYYC3-Knowledge-Base --skill vss-generate-video-report",{"tool":43,"command":44},"bunx","bunx skills add YYC-Cube\u002FYYC3-Knowledge-Base --skill vss-generate-video-report",{"tool":46,"command":47},"pnpm","pnpm dlx skills add YYC-Cube\u002FYYC3-Knowledge-Base --skill vss-generate-video-report","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-Skills-技能包\u002FYYC3-Skills-技能包-skills\u002Fskills\u002Fvss-generate-video-report\u002FSKILL.md",true,[51,57,61,65],{"path":52,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":55,"sourceHtmlUrl":56},"BENCHMARK.md","REFERENCE","参考文档",2811,"https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-Skills-技能包\u002FYYC3-Skills-技能包-skills\u002Fskills\u002Fvss-generate-video-report\u002FBENCHMARK.md",{"path":58,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":59,"sourceHtmlUrl":60},"evals\u002Fbase_profile_report.json",8611,"https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-Skills-技能包\u002FYYC3-Skills-技能包-skills\u002Fskills\u002Fvss-generate-video-report\u002Fevals\u002Fbase_profile_report.json",{"path":62,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":63,"sourceHtmlUrl":64},"skill-card.md",2810,"https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-Skills-技能包\u002FYYC3-Skills-技能包-skills\u002Fskills\u002Fvss-generate-video-report\u002Fskill-card.md",{"path":66,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":69,"sourceHtmlUrl":70},"skill.oms.sig","OTHER","其他",4621,"https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-Skills-技能包\u002FYYC3-Skills-技能包-skills\u002Fskills\u002Fvss-generate-video-report\u002Fskill.oms.sig",[72,80,87,94,101,108],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":76,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":19},"01M7CTDWHERCXPQCP0K03A86AS","vercel-sandbox","Run agent-browser + Chrome inside Vercel Sandbox microVMs for browser automation from any Vercel-deployed app. Use when the user needs browser automation in a Vercel app (Next.js, SvelteKit, Nuxt, Remix, Astro, etc.), wants to run headless Chrome without binary size limits, needs persistent browser sessions across commands, or wants ephemeral isolated browser environments. Triggers include \"Vercel Sandbox browser\", \"microVM Chrome\", \"agent-browser in sandbox\", \"browser automation on Vercel\", or any task requiring Chrome in a Vercel Sandbox.","MIT","https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-WebSpec-Web标准\u002FYYC3-WebSpec-Web标准-standards\u002Fbrowser-agent\u002Fskills\u002Fvercel-sandbox\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":76,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":19},"01M7CTD8D2BNJY54S65M47HEJA","slack","Interact with Slack workspaces using browser automation. Use when the user needs to check unread channels, navigate Slack, send messages, extract data, find information, search conversations, or automate any Slack task. Triggers include \"check my Slack\", \"what channels have unreads\", \"send a message to\", \"search Slack for\", \"extract from Slack\", \"find who said\", or any task requiring programmatic Slack interaction.","https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-WebSpec-Web标准\u002FYYC3-WebSpec-Web标准-standards\u002Fbrowser-agent\u002Fskills\u002Fslack\u002FSKILL.md",[],[],{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":76,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":19},"01M7CTCSSB8ZP3Q58C3YZDZK1W","electron","Automate Electron desktop apps (VS Code, Slack, Discord, Figma, Notion, Spotify, etc.) using agent-browser via Chrome DevTools Protocol. Use when the user needs to interact with an Electron app, automate a desktop app, connect to a running app, control a native app, or test an Electron application. Triggers include \"automate Slack app\", \"control VS Code\", \"interact with Discord app\", \"test this Electron app\", \"connect to desktop app\", or any task requiring automation of a native Electron application.","https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-WebSpec-Web标准\u002FYYC3-WebSpec-Web标准-standards\u002Fbrowser-agent\u002Fskills\u002Felectron\u002FSKILL.md",[],[],{"skillNo":95,"name":96,"nameZh":6,"description":97,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":76,"licenseFileUrl":6,"sourceHtmlUrl":98,"categories":99,"keywords":100,"publishedAt":19},"01M7CTCBJY76688V4DCZGWFHGQ","dogfood","Systematically explore and test a web application to find bugs, UX issues, and other problems. Use when asked to \"dogfood\", \"QA\", \"exploratory test\", \"find issues\", \"bug hunt\", \"test this app\u002Fsite\u002Fplatform\", or review the quality of a web application. Produces a structured report with full reproduction evidence -- step-by-step screenshots, repro videos, and detailed repro steps for every issue -- so findings can be handed directly to the responsible teams.","https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-WebSpec-Web标准\u002FYYC3-WebSpec-Web标准-standards\u002Fbrowser-agent\u002Fskills\u002Fdogfood\u002FSKILL.md",[],[],{"skillNo":102,"name":103,"nameZh":6,"description":104,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":76,"licenseFileUrl":6,"sourceHtmlUrl":105,"categories":106,"keywords":107,"publishedAt":19},"01M7CTC4H0FFGVEAQR9PA4GA1A","agent-browser","Browser automation CLI for AI agents. Use when the user needs to interact with websites, including navigating pages, filling forms, clicking buttons, taking screenshots, extracting data, testing web apps, or automating any browser task. Triggers include requests to \"open a website\", \"fill out a form\", \"click a button\", \"take a screenshot\", \"scrape data from a page\", \"test this web app\", \"login to a site\", \"automate browser actions\", or any task requiring programmatic web interaction.","https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-WebSpec-Web标准\u002FYYC3-WebSpec-Web标准-standards\u002Fbrowser-agent\u002Fskills\u002Fagent-browser\u002FSKILL.md",[],[],{"skillNo":109,"name":110,"nameZh":6,"description":111,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":76,"licenseFileUrl":6,"sourceHtmlUrl":112,"categories":113,"keywords":114,"publishedAt":19},"01M7CTBKVFYBG78FDSABSDJTTQ","shadcn","Manages shadcn components and projects — adding, searching, fixing, debugging, styling, and composing UI, including chat interfaces. Provides project context, component docs, and usage examples. Applies when working with shadcn\u002Fui, component registries, presets, --preset codes, or any project with a components.json file. Also triggers for \"shadcn init\", \"create an app with --preset\", \"switch to --preset\", and for building a design system or applying a DESIGN.md or brand spec to shadcn\u002Fui.","https:\u002F\u002Fgithub.com\u002FYYC-Cube\u002FYYC3-Knowledge-Base\u002Fblob\u002Fmain\u002FYYC3-UI-界面组件\u002FYYC3-UI-界面组件-ui\u002Fskills\u002Fshadcn\u002FSKILL.md",[],[]]