Update to PixVerse (Legacy; EOL soon)

{
  "description": "Templates: search and browse PixVerse effects, compare previews, and create images or videos from a selected template and supplied media. Use for template discovery or template-based creation.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 260
    }
  ],
  "name": "pixverse-template",
  "skill_md_contents": "---\nname: pixverse-template\ndescription: \"Templates: search and browse PixVerse effects, compare previews, and create images or videos from a selected template and supplied media. Use for template discovery or template-based creation.\"\n---\n\n# Browse Templates And Create\n\n## Installed Command\n\nResolve the absolute plugin root from this `SKILL.md`: two directories above the\nskill directory, containing `.codex-plugin/plugin.json`. In each new shell set:\n\n```bash\nPVX=\"<absolute-plugin-root>/scripts/pvx\"\n```\n\nUse `\"${PVX}\"` from any working directory. This workflow requires managed CLI\n1.4.9 or later. For setup/login failures use `../pixverse-setup/SKILL.md`.\n\n## Find A Good Match\n\nFor discovery, use read-only commands through the installed wrapper:\n\n```bash\n\"${PVX}\" pixverse template search \"<keyword>\" --json\n\"${PVX}\" pixverse template categories --json\n\"${PVX}\" pixverse template list --category <id> --page 1 --limit 8 --json\n\"${PVX}\" pixverse template info <template-id> --json\n```\n\nSearch responses can contain hundreds of full records, translations and author\nmetadata. Save large JSON responses under `projects/<slug>/reference/` and extract\nonly the shortlist fields with Python; do not dump or load the entire catalog into\nconversation context. Category list pagination keeps browsing bounded. Ignore\nunrelated author/account metadata. Public search may work without login while\ndetail and generation still require a valid account; do not treat search success\nas proof that paid creation is authenticated.\n\nStart from the user's intent or supplied template ID. A specific ID goes directly\nto detail. A vague request can start with categories; a concrete theme starts with\nsearch. Preserve Unicode keywords. If results are poor, try a short synonym or\nan English keyword, then browse the relevant category. Search has no pagination\nor category flags; list supports pagination. An empty result is a catalog outcome,\nnot a reason to generate a replacement without the user's agreement.\n\nFor themed discovery (such as pets), inspect names and descriptions for relevance;\nkeyword results may contain unrelated entries. Try synonyms such as `pet`, `cat`\nand `dog` when useful, merge by template ID and keep only relevant candidates.\nFor popularity requests, sort the fetched candidates numerically by `hot_count`\ndescending. Missing or invalid counts are unknown, not zero; put them after known\ncounts. The CLI has no server-side popularity sort flag. A global popularity\nranking requires fetching every catalog page, deduplicating IDs and sorting the\ncomplete set. If fetching is incomplete, label the ranking as limited to retrieved\ncandidates. For themed popularity, filter for relevance before ranking. Describe\n`hot_count` as the API's reported popularity metric; do not claim it means views,\nuses, or a daily/weekly trend without an upstream definition. The `hot` marker and\ndefault result order do not establish a numerical popularity ranking.\n\nPresent a small shortlist, usually 3–5 useful options, with the real template ID,\nname, image/video output, required inputs and a one-line reason for the match.\n**Download the selected thumbnails before displaying them in chat.** Remote image\nembeds may fail even when a browser can open the URL. Use the saved catalog to\ncache only the chosen candidates, without another detail query or paid generation:\n\n```bash\n\"${PVX}\" template previews projects/<slug>/reference/catalog.json --ids <id1> <id2> <id3> --project <slug> --format markdown\n```\n\nThis reads search/list/detail JSON, downloads official thumbnails with the CDN's\nbrowser headers, validates the file format and caches by template ID and URL.\nIt returns a compact gallery using **absolute local image paths**. Present that\ngallery with translated labels if helpful; keep the same IDs and local paths.\nDo not embed the remote thumbnail URLs again. If one thumbnail fails, keep the\nother cards and give that candidate a text-only label rather than a broken image.\nSee the JSON result or `reference/template-previews/selection.json` for per-item\nstatus and errors. Cached files are reused on repeat browsing.\n\nVideo templates automatically use their official GIF/WebP motion thumbnail in the\nlocal gallery. If it is missing or fails, the helper downloads the MP4 example and\nembeds that local video instead; playback controls depend on the host. Image templates\nkeep a static thumbnail even when a motion example exists. With `--videos`, also cache\nthe selected full example clips for local playback. Do not claim that MP4 previews\nautoplay, or download every catalog video while browsing. If both motion sources fail,\nkeep the available static cover and explain that the motion preview is unavailable.\nExamples are template demonstrations, not newly generated work and not a promise\nthat a new result will match exactly. Catalog/preview fetching spends no generation\ncredits and does not require paid preflight. Template selection still precedes any\npaid creation. Do not invent preview URLs or describe unseen animation.\n\nShow more candidates on request. Keep upstream names intact; explain the choices\nin the user's language. A missing preview does not authorize generating a substitute.\n\nDiscovery alone never authorizes generation. Let the user select a template,\nunless they already supplied an ID or authorized you to choose and create.\n\n## Prepare The Selected Template\n\nRead detail again before generation. Explain what the selected effect does and\ncollect only the missing inputs. Keep the user's subject/product media intact.\n\n- Ordinary templates: `template_type == 2` means image output. `effect_type`\n  declares the image count; zero is text-only. Video input is allowed only where\n  the detail supports it. Image-template prompts are not editable through the\n  ordinary image-template route; do not promise that a supplied prompt changes it.\n- Published Canvas templates: `template_source` 21 or 22 uses ordered\n  `input_fields`. Describe each slot by its name and content type, preserve image\n  order and optional slots, and respect size/duration constraints. CLI 1.4.9\n  supports at most one text and one video slot. Unsupported schemas stop before\n  paid work; offer another template. The CLI rejects ambiguous optional bindings.\n- Templates own their model and supported output settings. Do not add ordinary\n  Sunburst/Seedance defaults, model overrides or prompt enhancement to a fixed\n  template. Published Canvas templates reject quality, duration, aspect ratio,\n  count, seed, off-peak and idempotency overrides. Their result may contain several\n  images/videos; do not promise one file or a specific resolution without evidence.\n\nTemplate services may return a different resolution from a requested quality tier.\nWhen verifying output, report the measured dimensions; a successful submission is\nnot proof that the quality flag was honored.\n\nUsing a published Canvas template does not require binding or opening a Canvas\nproject. If the user explicitly wants to edit a Canvas project or graph instead,\nuse `../pixverse-canvas/SKILL.md`.\n\n## Generate And Deliver\n\nRead `../../skills-shared/cli-workflow.md` and\n`../../skills-internal/pixverse-agent-gateway/SKILL.md` before paid work.\nApply `../../skills-shared/quality-policy.md` for membership/account checks;\nthe selected template's fixed settings take precedence over generic model defaults.\nFree/Basic still stops for upgrade or a separately agreed fallback route. Fallback\nconsent cannot change a template's model or bypass its entitlement. Never silently\nsubstitute another effect. Browsing and selection need no paid confirmation.\n\nUse a manual queue, not `route queue`, to preserve template settings. Inspect\n`\"${PVX}\" pixverse capabilities create template --json` before composing an\nunfamiliar command. For example, after collecting the actual required image:\n\n```bash\n\"${PVX}\" project init <slug>\n\"${PVX}\" queue write projects/<slug>/queue.json --project <slug> --id effect --preflight --format markdown -- pixverse create template --template-id <id> --image <input>\n\"${PVX}\" queue run projects/<slug>/queue.json\n```\n\nThe creation fragment varies with detail: omit `--image` for text-only templates,\nadd the ordered images for multiple slots, and use `--video` / `--prompt` only\nwhen accepted. Preflight reads detail and the account. State the submission count,\nfixed settings and variable output count; follow the effective `skip` / `require`\nconfirmation policy, adding `--confirmed` only after required approval.\n\nShow the first usable local preview promptly. Deliver every returned `outputs`\nentry, including its media type, ID and local path; the top-level path is only the\nfirst output. Preserve partial successes and report missing/failed results clearly.\nUse the recorded credit evidence without guessing a total for an incomplete receipt.\nA download failure retries the free download, never generation. Resume/reconcile\nexisting receipts after interruption. Published templates have no server idempotency\nkey: an unknown submission must not be automatically resubmitted. Ordinary generation\nhas no automatic QA stage.\n"
}