Skip to content

wolfrine/velum-scene-lab

v0.1.2

Pick surprising sensual adult photography scenarios from the owner's Velum catalog and collect explicit image ratings for future creative experiments.

GitHub

Velum Scene Lab — private plugin

The source of this private plugin lives alongside PTI's shared MCP code. The plugin uses .app.json to depend on the owner's existing ChatGPT PTI Apps connection. It deliberately declares no mcp.json or local executable: this keeps it usable in ordinary ChatGPT web/mobile chats instead of marking an imported MCP package as Desktop only. The specialized tools themselves remain hosted on the existing authenticated PTI Cloud Run MCP server.

Workflow:

  1. velum_prompt_pick draws uniformly from the manifest's available prompt IDs, avoiding the last 15 selected IDs if alternatives exist. It writes one selections/{uuid} document for later feedback.
  2. The agent uses the chosen prompt with the currently available image generator and user-authorized adult reference images. It asks the owner to rate the actual generated image on a 1–5 scale.
  3. velum_prompt_rate writes an immutable ratings/{selectionId} record, linked to prompt ID/version and optional stable image reference. velum_prompt_feedback reads recent ratings for follow-up analysis.

Canonical Firestore data stays under users/{uid}/velumPromptData/workspace. Prompt text, feedback and selection history are private per the connected owner. Existing Drive media and browser-local Velum votes are unchanged. Do not upload original reference photos to Firestore.

The prompt catalog is populated separately through the owner-authenticated PTI Apps MCP. Updates to it do not require a plugin version bump. ChatGPT may cache PTI Apps tool metadata even though the server's pti_app_context('velum') advertises the newer dedicated Velum tool names. When velum_prompt_pick, velum_prompt_rate, or velum_prompt_feedback are absent from the callable catalog, the skill automatically uses the already-available generic owner-scoped firestore_get_document, firestore_batch_write, and firestore_query tools. Selection and rating records remain linked and immutable. Generic fallback timestamps are ISO strings rather than Firestore server timestamp values. No second MCP connection or manual tool refresh is required for this path.