Skip to content

the ultimate vocal synthesizer - sing, speak, beatbox, and more

Rust
3.0 KB
Raw Blame

Contributing

PRs are welcome, and most of them here are written with coding agents. This page is for you and your agent; AGENTS.md holds the rules for the code itself.

Start from the vision

README → Vision & scope decides what gets merged. Work outside it is declined however well it's built: a new mode that isn't the voice, an export pipeline, tooling for one person's workflow, a sequencer. If you're unsure whether an idea fits, open an issue or a draft PR and ask before you build much.

Solve one problem per PR. Split unrelated fixes, and leave drive-by refactors for their own PR.

Use a strong model, and review the work

  • Use the most capable model you have, with reasoning effort to match the work's complexity and risk.
  • Before you open the PR, have a fresh agent review the diff and fix what it finds (or say in the PR why you didn't). A reviewer from a different model family tends to catch more; we recommend one but don't require it.
  • Follow AGENTS.md: the engine's contract, the golden tests, and the checks to run for the files you touched.

Show the change works

Put the evidence in the PR, not in git.

  • Sound: before/after audio of the same phrase, or before/after spectrograms (bun tools/spectrogram.ts), and say what you hear change. Re-bless the goldens. Voice tuning follows docs/hill-climb/.
  • UI: before/after screenshots, including at phone width; a short recording when motion or timing matters.
  • Bug fix: the reproduction, kept as a test that fails before the fix and passes after.
  • Language: a native speaker's check, and English unchanged (docs/language-front-ends.md).
  • Say what you couldn't check.

Write the PR top-down

Someone who reads only the first screen should know what changed, why, how you know it works, and what could break. Details go below. Use the sections in the PR template, and keep prose short.

  • Title: the area and what changed, in searchable words, for example Play: grabbing the face before any audio starts it.
  • Summary: two to four sentences. Add a small visual if it makes the point faster: a diff of a call tree or file tree, a spectrogram, a screenshot.
  • Evidence: before and after, as above.
  • Risk: is this easy to undo (a two-way door) or hard to undo (one-way: a wire format, saved songs, a public API)? What else could it break: the goldens, the plugin, phones, other languages?
  • Details: design choices, what you ruled out, the checks you ran with their numbers.
  • Attribution, last line: the model and harness that wrote the code, and the reviewer's model if you used one, for example Written by Claude Opus 5.5 in Claude Code; reviewed by GPT-5 in Codex. Write No agent if you wrote it yourself.

Nothing copyrighted

No copyrighted audio, lyrics or scores in git, and nothing derived from them. Contributions are under the MIT license.