← LeaderboardThe Skill Leaderboard

scenario-delete-restore

Use when deleting Scenario assets, models, or collections through MCP, or undoing a deletion: restoring images, videos, or LoRAs deleted by mistake, finding something a user says they deleted, listing what was deleted recently, or cleaning up a project before a bulk delete. Keywords: delete, undelete, restore, recover, undo, trash, bin, recently deleted, 30 days, deleted model, deleted asset, cleanup.

Published by scenario-labs ↗scenario-labs/skills

Try your first task

A source-grounded starter task has not been verified for this skill yet.

SKILL.md

name: scenario-delete-restoredescription: "Use when deleting Scenario assets, models, or collections through MCP, or undoing a deletion: restoring images, videos, or LoRAs deleted by mistake, finding something a user says they deleted, listing what was deleted recently, or cleaning up a project before a bulk delete. Keywords: delete, undelete, restore, recover, undo, trash, bin, recently deleted, 30 days, deleted model, deleted asset, cleanup."license: MIT

Scenario Delete and Restore

Overview

A deleted asset or model stays recoverable for 30 days, on the Max plan and above (Max, Team, legacy Premium, Enterprise). Below Max, the server refuses to list or restore deleted items, so a delete is final from the agent's side.

  • A restore brings the item back, not reliably its filing: tags and collection membership are documented as lost, yet some tags survive. Read the restored rows before re-tagging.
  • Collections have no restore tool. A deleted collection is rebuilt by hand or not at all.
  • Recovery is per project: the deleted listings cover the project in scope only.
  • The tools below marked read, write, or delete are catalog-only. teams_list, models_list, and asset_display are direct: called as is, scope at top level, never through an executor. scenario_tools_search returns a catalog tool's schema and lane; the matching executor runs it with {name, parameters}, scope ids inside parameters (see the scenario skill).
  • Deleting a whole project is irreversible and belongs to scenario-team-admin.

If a sibling skill named here is missing from your available skills, load it with skill_get (Scenario MCP) or ask the user to install it (https://github.com/scenario-labs/skills#install); unattended, proceed from tool schemas and flag the gap.

Quick reference

StepToolLaneNotes
Find deleted assetsassets_list with deleted: truereadMost recently deleted first, each with deletedAt
Restore assetsasset_restorewrite1 to 100 asset_ids; returns restored and skipped
Find deleted modelsmodels_list, filters: {"status": "deleted"}directYour own models only; never with privacy: "public"
Restore a modelmodel_restorewriteOne model_id per call
Delete assetsasset_deletedeleteReturns deleted; split sets past 100 ids
Delete a modelmodel_deletedeleteOne model_id per call
Delete a collectioncollection_deletedeleteNo restore tool exists
Re-fileasset_add_tags, collection_add_assetswritemodel_add_tags, collection_add_models for a model
Check the planteams_listdirectThe team row's plan decides whether a delete is undone

Restoring

  • Find. assets_list with deleted: true takes no filters: the server refuses the pair. Match the user's words client-side against deletedAt, kind, prompt, description, and modelId, paging with page_token (limit up to 100). The list is newest deletion first, so stop paging once deletedAt is older than the user's window. prompt comes in three forms:
    • An asset id (asset_...): the prompt is a live text asset. Read it on assets_get_bulk, whose record carries the full prompt text, and match that.
    • Text, cut at 120 characters. A deleted asset's own record is unreadable (asset_get returns 404), so a row in the window whose cut text shows no match is ambiguous, not a miss.
    • Absent, on an upload, which keeps no file name either: match its description, an automatic caption.
  • Models. status: "deleted" is a key inside filters on models_list. Combined with privacy: "public" the server refuses it: deleted models are always the team's own.
  • Confirm the set. Show the candidates (id, deletedAt, prompt or name) when the match is not exact. Unattended, restore only items the task's description matches unambiguously, and list the rest as unresolved.
  • Read skipped. Each entry carries asset_id and a reason:
    • not-found: a wrong id, or an asset of another project;
    • not-deleted: already live, nothing to do;
    • expired: past the retention window. Nothing recovers it: tell the user.
  • Models come back as they were. model_restore returns the model to its prior status. An interrupted training returns as training-canceled, so it needs a new training run. The server refuses a model that is not deleted or was deleted over 30 days ago: report which.
  • Plan refusal. A restore or deleted listing refused with a plan message is not retryable. Report that restoring needs Max or above; upgrading is the user's call (unattended: report and stop).
  • Re-file. Re-apply the tags and collections the user still needs. Before re-tagging, read the restored rows on assets_get_bulk and send only the missing tags: a tag the asset kept is refused as a duplicate (see the scenario skill). asset_add_tags takes one asset per call. collection_add_assets takes any number of ids and skips those already present, so it needs no read. A collection deleted too is recreated with collection_create first.

Deleting

Deletion stays a confirmed action.

  1. Read the plan off the team's teams_list row. On Max and above, tell the user deleted assets and models stay recoverable for 30 days, possibly without their tags and collections. Below it, say the delete is permanent. A collection is permanent on every plan. An unfamiliar plan value: assets_list with deleted: true is a free read gated to the same plans, so a plan refusal there settles it.
  2. List the exact items (ids plus a name or prompt) and get an explicit yes. A set found by tag comes from search with filters.tags (see scenario-asset-analysis), whose index can trail a tag change: confirm each hit's tags on assets_get_bulk before deleting it. When every match must go, also page assets_list and match tags locally, since a fresh tag can miss the index. Unattended, delete only the ids the task names, or the set its stated criterion matches exactly, and report every id deleted. Never widen a pattern ("old drafts") past what the task states.
  3. Record the filing before the call: each asset row from assets_list or assets_get_bulk, and each model row from models_list, carries tags and collectionIds. Keep them in your reply or a local file, since a restore may not bring them back. For a collection, record its name and member ids. collections_list gives only assetCount and modelCount, so the members are the assets_list rows, and the models_list rows when modelCount is above 0, whose collectionIds hold its id. Page both with page_token (limit up to 100) until the members found reach those counts.
  4. Delete through scenario_tool_execute_delete. asset_delete refuses more than 100 ids in one call (You can not delete more than 100 assets at once.): send batches of 100.

Worked example: undo yesterday's cleanup

"Bring back the dragon renders I deleted yesterday and put them back in Key Art."

  1. Scope per the scenario skill. The team's teams_list row carries plan: below Max, tell the user restoring is not on their plan and stop. Here it reads enterprise, so continue.
  2. scenario_tools_search with query="restore deleted assets": asset_restore (write) and assets_list (read), with schemas.
  3. scenario_tool_execute_read with name: "assets_list", parameters: {"team_id", "project_id", "deleted": true, "limit": 100}. Follow nextPaginationToken until deletedAt falls before yesterday.
  4. Keep the rows whose deletedAt falls on yesterday, in the user's time zone, and whose prompt makes a dragon the subject. A negated or incidental mention is ambiguous, and so is a cut prompt that shows no dragon. A cut prompt that names one counts only when the row's description shows a dragon too, since the omitted text could negate it. Fourteen mention a dragon: thirteen clearly depict one, their description agreeing, and one reads "a castle without a dragon". Two more are cut before naming a subject. Show all sixteen with deletedAt and confirm. Unattended: keep the thirteen, and list the castle and the two cut prompts as unresolved, neither restored nor filed.
  5. scenario_tool_execute_write with name: "asset_restore", parameters: {"team_id", "project_id", "asset_ids": [<13 ids>]}. restored holds 12; skipped holds one not-deleted, already restored by a teammate.
  6. scenario_tool_execute_read with name: "collections_list": Key Art still exists. Then scenario_tool_execute_write with name: "collection_add_assets", parameters: {"team_id", "project_id", "collection_id", "asset_ids": [<all 13 ids>]}, since it skips any already present.
  7. asset_display on one restored asset to confirm. Report 12 restored, one already live, all 13 back in Key Art, and three unresolved: the castle and the two cut prompts.

Common mistakes

  • Sending filters with deleted: true: refused. List unfiltered and match client-side.
  • Re-applying every recorded tag after a restore: some survive, and a duplicate is refused. Read the restored rows, then send only the missing tags.
  • Promising recovery below Max: the server refuses restores there, so the delete was final.
  • Running asset_restore or model_restore through scenario_tool_execute_delete: both are write-lane, and the delete executor rejects them.
  • Retrying a plan refusal, or an expired skip: neither clears.
  • Deleting a collection to tidy up and expecting to undo it: no restore tool exists.
  • Sending more than 100 ids to asset_delete or asset_restore in one call: split the set.
scenario-delete-restore by scenario-labs · The Skill Leaderboard