CrunTeam Crun Effect Templates
Use ../../../runtime/crun_cli.py. Treat the authenticated template-list endpoints as the source of truth for current
IDs, preview media, input requirements, supported settings, and credits. A model schema does not enumerate the valid
template IDs.
Read ../../crun-account-credits/SKILL.md before estimating and ../../crun-task-runner/SKILL.md before creating or
monitoring a task.
Discover templates
List one page for a known platform:
python <runtime>/crun_cli.py templates list --platform kling --page 1 --page-size 20
python <runtime>/crun_cli.py templates list --platform vidu --page 1 --page-size 20
python <runtime>/crun_cli.py templates list --platform bytedance --page 1 --page-size 20
Look up one exact ID with the unified flag; the CLI maps it to each platform's API parameter:
python <runtime>/crun_cli.py templates list --platform <platform> --template-id <id>
Use total to fetch additional pages when the requested effect is not on the first page. Never invent an ID, use a
stale remembered ID, or substitute a similarly named effect silently.
When the user has not named a platform and has not already requested a platform-specific search, search Kling first because it is the preferred default for effect quality. Evaluate its candidates against the user's actual intent, source-media requirements, preview, settings, and credits. If Kling has a suitable match, use or present that match and do not query another platform. If no relevant Kling template exists, or the available Kling templates do not satisfy the user's intent, stop and ask the user to choose whether to continue searching ByteDance or Vidu. Query only the platform the user selects; do not search both automatically.
When presenting a candidate list, show every candidate selected as relevant and display exactly these five fields for each candidate:
- Platform
- Template name
- Template input requirements
- Credit cost
- Preview link
Localize the candidate list to the user's language. Localize all five field labels, template names, input-requirement text, missing-preview text, and surrounding choice instructions. Prefer an official localized template-name field when the response provides one; otherwise translate the display name when safe while preserving brands and proper nouns. Keep platform names, numeric credit values, and preview URLs unchanged.
Do not display template IDs, categories, tags, resolutions, settings, cover links, or other metadata in the candidate
list. Preserve each template ID internally so the selected candidate can be queried and submitted later. Never omit
one of the five fields or replace content with an ellipsis. Condense the input requirements into clear localized prose,
removing examples and repetitive wording while retaining every constraint that can affect task success: required image
count and order, subject type or count, accepted aspect ratio or range, prompt requirements, and unsupported inputs.
Report the credit value exactly as returned. For ByteDance, include every Resolution and Credit pair from
ConsumeItems inside the single credit-cost field so no pricing option is lost. Prefer the video preview field
(video_url, video, or PreviewVideo); when no preview link exists, display a localized explicit value. If the
relevant candidates require
multiple messages, continue until every candidate and every required field has been shown. Template browsing or
choosing a preview is not authorization to spend credits.
Interpret each platform's response explicitly:
| Platform | ID and preview fields | Constraint and price fields |
|---|---|---|
| Kling | key, cover_url, video_url |
Read image-count and subject rules from info; use credit. |
| Vidu | template, cover, video |
Parse the JSON string in input_instruction; use credits, resolution, and aspect_ratio. |
| ByteDance | TemplateId, Cover, PreviewVideo |
Enforce InputNum and InputRequirement; use the selected resolution's Credit in ConsumeItems. |
Validate source media and map the task
Require the selected template's source media before task creation. Accept attached local files or directly usable URLs
from the current conversation. Upload each local file and use its returned file_url:
python <runtime>/crun_cli.py upload <local-file>
Never send local paths, Base64, or data URIs. Preserve the user's image order for multi-image effects. Read the selected template's returned instructions and enforce its exact image count and media constraints; do not infer them from another template.
Use this mapping, then inspect the live model schema before constructing input:
| Platform | Model | Template field | Media field | Platform-specific rules |
|---|---|---|---|---|
| Kling | kling/template |
template_id |
img_urls |
Follow the selected template's info instructions for image count. |
| Vidu | vidu/template |
template |
images |
Send optional prompt, aspect_ratio, area, beast, bgm, or story only when the live schema and selected template support them. |
| ByteDance | bytedance/template |
template_id |
img_urls |
Choose only a resolution published in ConsumeItems; report its matching Credit, but do not send ConsumeItems in task input. |
python <runtime>/crun_cli.py models describe --model <model>
Construct only fields allowed by input_schema. Re-fetch the exact template by ID immediately before estimation when
earlier discovery results may be stale.
If a live schema endpoint is temporarily unavailable, do not invent optional fields. Use only the minimal platform
mapping above, then let task estimate perform server-side validation before any creation. For Kling the minimal input
is exactly template_id plus img_urls.
Estimate, confirm, and run
Estimate the exact final input before every new task:
python <runtime>/crun_cli.py task estimate --model <model> --input-file <input.json>
Require affordable: true. If the effect or settings were discovered or selected during the conversation, show a
localized summary containing platform, effect name and ID, model, ordered source-media count, relevant settings,
estimated credits, and balance, then obtain explicit confirmation for that exact task. If the user's initial request
already supplied the exact platform, exact template ID, source media, settings, and an instruction to run it, follow the
shared task runner's user-specified-model confirmation rule.
After authorization, create exactly once, retain the task ID, and wait only on that task:
python <runtime>/crun_cli.py task create --model <model> --input-file <input.json>
python <runtime>/crun_cli.py task wait --task-id <task_id> --timeout-seconds 120
Follow crun-task-runner for timeouts, failures, recovery, downloads, and delivery. Never retry CreateTask
automatically. Return the localized effect name, platform, exact template ID, task ID, credits, and local media path;
include the preview or remote result URL only as a fallback.
Localize interaction
Write all choices, summaries, progress, errors, and delivery text in the user's language. Keep platform names, model IDs, template IDs, API field names, URLs, task IDs, and enum values unchanged.