Problem#
GET /v1/openapi/lovable?activity=video copies the OpenAI Sora request schema, so it is wrong for the Lovable gateway:
seconds is VideoSeconds ("4" | "8" | "12"), described as "allowed values: 4, 8, 12". The gateway runs Veo, which takes 4, 6 or 8 seconds. A 6-second request is valid but fails the schema; a 12-second request passes the schema but fails at the gateway.
model says "allowed values: sora-2, sora-2-pro. Defaults to sora-2", but its enum is the Google models (google/veo-3.1, -fast, -lite, google/gemini-omni-1.1-flash).
size lists the Sora sizes (720x1280, 1024x1792, ...).
Ask#
Give Lovable its own video schema, per model if the limits differ: seconds from Veo (4, 6, 8) and Omni, the gateway's sizes, and descriptions that match.
Acceptance#
seconds on the Veo models allows 6 and does not allow 12.
- No description mentions Sora.
TanStack AI type-checks @tanstack/ai-lovable video requests against this document (TanStack/ai#2104, TanStack/ai#2105). The adapter casts seconds today because of this.