pi-cavallo

Generate and edit video in pi using Alibaba Model Studio video models. Supports HappyHorse 1.1/1.0, Wan 3.0, and Wan2.7: text-to-video, image-to-video, reference-to-video, and video editing.

Notification contract

Notices use official Pi TUI components through a bundled shared module. They render without loading pi-fornace or modifying Pi. New notices use a package-owned entry type; historical Fornace notices remain readable. extensions/shared/fornace-messages.ts is generated from the source identified in SOURCE.json, never edited independently.

Models

Wan 3.0 (all-in-one, up to 30 seconds)

• wan3.0-video (T2V, I2V with first/last-frame stitching, R2V, editing; parses files and web links as references)
• wan3.0-video-prime (speed-optimized wan3.0-video)

HappyHorse 1.1 (native audio, 480P/720P/1080P)

• happyhorse-1.1-t2v (Text-to-Video, recommended)
• happyhorse-1.1-i2v (Image-to-Video, recommended)
• happyhorse-1.1-r2v (Reference-to-Video, recommended)

HappyHorse 1.0

• happyhorse-1.0-t2v (Text-to-Video)
• happyhorse-1.0-i2v (Image-to-Video)
• happyhorse-1.0-r2v (Reference-to-Video, up to 9 images)
• happyhorse-1.0-video-edit (Video editing via language + references)

Wan2.7

• wan2.7-t2v (Text-to-Video)
• wan2.7-i2v-2026-04-25 (Image-to-Video: first frame, last frame, video continuation)
• wan2.7-r2v (Reference-to-Video)
• wan2.7-videoedit (Video editing via instruction + reference images)

OpenAI-compatible providers (mantice)

Run /cavallo-setup to submit video generation through an OpenAI-compatible gateway instead of DashScope. The wizard probes the endpoint and saves the route under "cavallo" in settings.json:

"cavallo": { "baseUrl": "https://llm.fornace.net/v1", "apiKey": "sk-...", "model": "fornace-video" }

The gateway route supports text-to-video everywhere, plus image inputs when the configured model is a verified MiniMax H3 route (fornace-video, fornace-video-h3-max):

• imagePath (first frame) and lastImagePath map to content items with role=first_frame / role=last_frame.
• referenceImages (1-9) map to role=reference_image items.
• Frames and reference images are mutually exclusive; lastImagePath requires imagePath.
• Local images load as base64 data URIs; http(s):// and mm_file:// URLs pass through.
• Requested vs effective aspect ratio: with first/last frames the upstream always derives the ratio from the input image (adaptive); a requested aspectRatio is sent but ignored upstream. Reference-to-video may request a concrete ratio (for example 16:9); default adaptive.

DashScope stays the route for every other input (videoPath, firstClipPath, audioPath) and model.

Defaults and overrides

Omit model and the best default is picked automatically from the task type:

• /cavallo-models lists the catalog and current defaults.
• /cavallo-models t2v=wan3.0-video persists a default override (stored under cavallo.defaults in settings.json).
• Duration and resolution limits are validated per model (HappyHorse 3-15s; Wan 3.0 up to 30s).

Features

• Non-blocking Execution: Submits to DashScope and polls in the background, freeing up the chat.
• Smart Thumbnails: Automatically extracts and displays a thumbnail when the video finishes downloading using ffmpeg (if installed).
• Finder Integration: Clickable markdown links to quickly reveal the exported video in Finder without launching it (open -R).
• Headless Mode Support: Fully compatible with API usage or RPC mode.
• Granular Control: Supports resolution (H3 Max defaults to 768P), duration (5 seconds by default), aspectRatio (16:9 by default), promptExpansionMode (balanced by default), and seed. The fal safety checker is always enabled.
• Audio Support: Wan2.7 models accept an audioPath parameter for driving video with sound (Note: Audio requires a public HTTP/HTTPS URL per DashScope API).

Setup

The extension automatically pulls the Alibaba Cloud (API Key) from your internal Pi Models configuration (/models). Alternatively, set the environment variable:

export DASHSCOPE_API_KEY="your-key-here"

Install

npm install -g pi-cavallo

Then in pi:

/install pi-cavallo