WonderAgentPlay

Docs

Operator notes for the Muse-Wonder playground. Prefer the live UI for interactive controls.

Playground controls

  • System prompt — editable; counts toward input cost; reset restores the default helper prompt.
  • Max output tokens — UI slider only; the server clamps to product max (default 8000, env-overridable).
  • Request JSON format — soft request + server parse check; not a native schema guarantee. Failures appear as jsonValidation.
  • Templates — Code fix, Unit tests, JSON format from business/muse-presets.
  • Clear session — wipes localStorage key muse-wonder-playground-v1.

Billing: reserve → settle

  1. Client POSTs to /api/muse/run.
  2. Server validates input and reserves a ceiling (estimated input + max output).
  3. Model runs; on known usage, settle actual input+output and refund unused reserve.
  4. On unknown usage or timeout, keep a conservative hold (pending_recon) — do not treat as free.
  5. GET /api/muse/run returns quota snapshot for the UI.

Response cache

Phase one keeps ENABLE_RESPONSE_CACHE=false. There is no on-by-default answer cache UI. Identical prompts may still consume credits on every run.

WonderAgentPlay is an independent third-party Muse-Wonder-1.0 playground powered via Cloudflare Workers AI. Not affiliated with, endorsed by, or operated by Meta. In-session browser context only — Meta native Active-Memory is not available.