Problem#
GET /v1/openapi/byteplus is live, but three things stop a generated client from covering the BytePlus adapter.
- The video status call has the wrong method. The document lists
POST /contents/generations/tasks/{id}. Ark polls a task with GET /contents/generations/tasks/{id}. Cancel (DELETE /contents/generations/tasks/{id}) is missing.
- Durations are only in prose.
duration is a plain integer. The per-model limits are in the description only: "2-12s on 1.0, 4-12s on 1.5-pro, 4-15s on 2.0, 4-30s on 2.5", and -1 lets the model choose. A generator can't read that.
- Audio is not in the main document. TTS (
/tts/create) and speech recognition (/auc/bigmodel/recognize/flash) are only in ?activity=audio, because they use the voice host (voice.ap-southeast-1.bytepluses.com), not Ark.
Ask#
- Make the status operation
GET, and add DELETE for cancel.
- Publish per-model video schemas (as fal does, with
x-modelschemas-model), each with its own duration bounds (minimum / maximum, plus -1 where allowed).
- Put the audio paths in the main document with a path-level
servers override for the voice host.
Acceptance#
paths["/contents/generations/tasks/{id}"] has get and delete, and no post.
?activity=video gives each Seedance model its own duration schema with machine-readable bounds.
GET /v1/openapi/byteplus includes /tts/create and /auc/bigmodel/recognize/flash with a servers entry on each.
TanStack AI generates its provider clients and video duration tables from these documents (TanStack/ai#2104, TanStack/ai#2105). BYTEPLUS_VIDEO_DURATIONS in @tanstack/ai-byteplus is hand-written from live probes until this lands.