Release video

A GitHub Action that turns a release's notes into a short oxframe video and attaches the MP4 to the release.

Connect an LLM agent for a video worth watching. With one (Claude Code, Codex, the Anthropic API or any agent CLI), the agent reads the notes, comes up with a visual treatment for this release (a diagram of what was fixed, a before and after, a counter) and writes the oxframe scene; oxframe check reviews it until it passes. See Connect an agent. Without one, a template renders the notes: the version opens it, each section gets its heading in its color (added green, fixed blue, changed yellow, breaking red), and its changes are written on word by word over a calm sea.

on:
  release:
    types: [published]
jobs:
  video:
    runs-on: ubuntu-latest
    permissions:
      contents: write
    steps:
      - uses: oxygenframe/[email protected]
        with:
          # From `claude setup-token` (a Claude subscription); or see below.
          claude-code-oauth-token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

oxframe check runs first. Changes share a card only while they fit above the sea; a card that still cannot fit its text is split into two cards and checked again, not clipped. Each change holds long enough to read. The end card says "made by Oxframe" (with the oxframe logo) unless credit is false. A [theme] table in the repository's oxframe.toml is copied onto the generated project.

The frame is 16:9. portrait: true renders 1080×1920 instead.

A release created with GITHUB_TOKEN does not start other workflows, so on: release never runs for it. Render it in the job that creates it, or any time later, with tag; the notes are read from the release:

      - run: gh release create "$GITHUB_REF_NAME" --notes-file notes.md
      - uses: oxygenframe/[email protected]
        with:
          tag: ${{ github.ref_name }}

notes renders text of your own instead of the release's notes.

Connect an agent

The agent gets the release notes, the generated project's agent docs (AGENTS.md and the oxframe skill, for the SDK version it builds with) and the template scene as an example, with the request: come up with a visual treatment of this changelog that can be built in oxframe, then build it. oxframe check reviews each version, and its report goes back to the agent until the scene passes (at most five tries). If none passes, or the agent fails, the template renders the notes as usual. The treatment is printed in the job log. Pick one of:

Claude Code with a Claude subscription. Run claude setup-token once on your machine and store the token it prints (sk-ant-oat…) as a secret. The action installs Claude Code and asks with claude -p, no tools.

          claude-code-oauth-token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

The Anthropic API. An API key from the Claude Console:

          anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}
          # model: claude-opus-5-5

Any other agent CLI (Codex, Gemini CLI, …). Install and sign it in in an earlier step; agent-command runs in bash, gets the prompt on stdin and answers on stdout, or writes the answer to $REPLY_FILE:

      - run: npm install -g @openai/codex
      - uses: oxygenframe/[email protected]
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        with:
          agent-command: codex exec --skip-git-repo-check --ephemeral -s read-only -o "$REPLY_FILE"

A video is one to five agent requests. Use a strong model: the scene is a whole Rust program against the oxframe API.

The generated project builds the oxframe SDK. The build is cached between runs (per CLI version), so only the first video after a new oxframe release compiles it.

License

Licensed under either of MIT or Apache-2.0 at your option.