Changelog

​

Changed

• A refused command fails its own call. The failure carries a stable code, such as queue_full or invalid_image. In JavaScript the call resolves undefined and the error event reports operation: "sendCommand" with the code. In Python, send_command() raises a ReactorError. The model still broadcasts command_error beside the failure, and a future major release removes that message. The schema lists every code.
• A reference upload that the model cannot decode gives a fixed reason. enqueue fails with invalid_image or invalid_audio and a standard message, such as “Reference 1: The reference image cannot be decoded safely.” The message no longer repeats the decoder’s own text. Use the code to tell one refusal from another.

​

New

• Prompt-only clips. Reference images are now optional. Omit reference_image and reference_images, or send an empty list, and the clip comes from the prompt alone. A list still holds up to nine images.
• Audio on a continued clip. A clip that carries audio no longer needs an image of its own. A continue_from_clip_id whose clip the session still holds stands in for one, because that clip’s own soundtrack carries across the boundary.

​

Changed

• When a clip carries audio and no image, an unknown or dropped continue_from_clip_id is refused with command_error. Without audio, such an ID still falls back to an independent clip.

​

New

• Reference audio. enqueue now takes audio references that condition the generated soundtrack: one reference_audio, or an ordered reference_audios list of up to three. Each clip is 2–15 seconds, mono or stereo, in WAV, MP3, AAC/M4A, OGG/Opus, FLAC, or WebM. Audio is optional and must accompany at least one reference image, and a clip carries at most twelve references in total.
• Two roles for an audio reference. Name each clip in the prompt as Audio 1, Audio 2, and so on, then say whether it is a character’s voice or a track to reuse. A label alone does not tell the model what the audio is for. The prompt guide has both patterns.
• Two clip fields. Every clip message and queue entry now reports has_reference_audio and reference_audio_count beside the existing image fields.

​

Changed

• A clip queued with continue_from_clip_id uses one of its three audio references for the previous clip’s soundtrack, so it accepts at most two of your own. Its budget of nine reference images is unchanged.