scorebench

Agent-native workbench for scorekit. You talk to the agent; the agent writes the scene DSL and drives scorekit. An explicit raw-YAML editor remains available for manual work.

scorebench is a desktop app (Tauri 2 + Svelte 5) that hosts a minimal ReACT agent for composing and rendering game music with scorekit. It is the shell, scorekit is the compiler, the LLM is the composer.

you ──chat──► agent core (Rust, Responses | Chat Completions)
                 │  tool calls (subprocess, --json)
                 ▼
              scorekit  validate / lint / build / diff
                 │
                 ▼
              project dir ──► scene.yaml + out/*.ogg + meta.json
                 │
                 ▼
              WebAudio playback + spectrum (AnalyserNode, zero in-house DSP)

Positioning

Arranging expertise is a system property, not a user requirement. scorebench has two faces:

• Tool face — a music compiler. The deterministic lower half: scene YAML in, audio out, with a schema, exit codes, and machine-readable JSON. That half is scorekit.
• Human face — an emotion expresser. The probabilistic upper half: the user speaks an emotional, imagistic vocabulary ("lonelier", "like dusk"); the agent translates it into the technical vocabulary of tempo, mode, voicing, and instrumentation.

Technique is relocated, not eliminated. Emotion is carried by craft — a "sad" piece with broken voicing doesn't sound sad, it sounds wrong — so the craft moves from the user into the system: style packs carry arranging knowledge as data, the review panel audits the system's own craft, and the fidelity of emotional translation is bounded by what scorekit's primitives can express.

In one line: the user speaks the vocabulary of emotion, the system works in the vocabulary of technique, and the agent is the translator between the two.

Product shape

• One project per window. Opening scorebench means opening one project directory (scene YAML + rendered assets + agent memory). No multi-project tabs.
• Agent-first authoring. The agent is the primary scene writer. Experienced users can also use the raw-source editor with explicit Validate and Save actions; there is no autosave. Parameter panels remain read-only observations.
• Playback & spectrum in the webview. Decoding, FFT, and progress come from the browser's WebAudio API (AnalyserNode) — no Rust audio stack, no in-house DSP.
• ScoreKit 0.3 sound textures. Scenes can schedule looped ambience and one-shot effects through portable source keys; scorebench persists the selected texture profile, checks mappings, and passes it to ScoreKit without mixing audio itself.
• Project memory. The agent maintains a rolling project summary; when the conversation exceeds the configured context budget it compacts automatically.

Iron rules

1. Agent core stays minimal. Two hand-rolled OpenAI-compatible transports: responses | chat_completions (configurable base URL + key + model). Provider presets only fill connection defaults; there is no multi-provider abstraction, agent framework, or SDK.
2. scorebench never renders audio itself. All compilation/rendering/export goes through the scorekit CLI (--json). If scorekit can't do it, scorebench doesn't do it.
3. No structured editing UI. No piano roll, timeline, or form-based scene editor. The only manual in-app write path is the explicit raw-YAML editor; scorebench otherwise observes and plays.
4. Deterministic boundary respected. scorebench never post-processes rendered artifacts; what scorekit writes is what plays.

Status

Core milestones M0–M5 complete (walking skeleton → agent core → observation surfaces → project memory → spectrum modules → release engineering). See docs/roadmap.md.

Install (Users)

macOS (Homebrew, recommended)

brew tap talkincode/tap
brew trust --tap talkincode/tap
brew install --cask talkincode/tap/scorebench
scorekit doctor

talkincode/tap/scorebench declares talkincode/tap/scorekit as a dependency, so Homebrew installs ScoreKit for you.

If Homebrew says Xcode is too old even though a newer Xcode or Xcode beta is already installed, point xcode-select at the active developer directory before retrying:

sudo xcode-select -s /Applications/Xcode-beta.app/Contents/Developer
xcodebuild -version
brew install --cask talkincode/tap/scorebench

Linux and Windows

Install scorebench from Releases:

• Linux: use the .deb or .AppImage artifact.
• Windows: use the .msi or -setup.exe artifact.

Install ScoreKit 0.5.x separately (for example with Homebrew brew install talkincode/tap/scorekit, or from ScoreKit Releases), then confirm:

brew trust --tap talkincode/tap   # when installing ScoreKit via Homebrew tap
scorekit --version
scorekit doctor

If scorebench starts but cannot find ScoreKit, pin the executable's absolute path in Settings → scorekit binary (also the way to choose between several installed versions), or set SCOREBENCH_SCOREKIT and restart the app.

Development (Contributors)

npm install
npm run tauri dev    # requires Rust toolchain + scorekit on PATH

Working across the whole constellation (scorekit, samples, forge, ScoreData)? The committed multi-repo map is scorekit-workspace.json in the scorekit repo; run python3 scripts/workspace.py doctor there to check/bootstrap the layout.

Documentation

The English user guide covers the ScoreKit scene protocol, practical arrangement concepts, render backends, sound-source provenance, licensing, and troubleshooting. Read it on GitHub Pages or from docs-site/src.

Build the site locally with mdBook 0.5.3 or newer:

mdbook build docs-site
mdbook serve docs-site --open

The API endpoint, model, context budget, and API key are configured from the in-app Settings panel. Keys use the OS keychain; the explicit insecure fallback is stored only in the Tauri app-config directory with mode 0600.

Version and release

src-tauri/tauri.conf.json is the application-version source of truth:

npm run version:set -- 0.1.1
git commit -am "Release 0.1.1"
git tag v0.1.1
git push origin main v0.1.1

Version tags build macOS (aarch64 and x86_64), Windows, and Linux installers, generate SHA256SUMS, and assemble a draft GitHub release. Publishing that release updates the Homebrew cask in talkincode/homebrew-tap when HOMEBREW_TAP_TOKEN is configured. Apple signing/notarization activates when its repository secrets are present; local and non-macOS builds do not require those secrets. scorekit is discovered at runtime and is installed as a Homebrew cask dependency, but is not bundled.

License

The source code is licensed under the MIT License.

Project names, logos, the cat emblem, application icons, and other brand assets are excluded from that license. The "talkincode" name and the cat emblem are reserved brand assets; see TRADEMARKS.md.