kbot Styled Rhythm Generator

An Ableton Live Extension that generates styled drum patterns into a MIDI clip — a native-TS re-build of Side Brain's Styled Rhythm Generator (overview video).

Two context-menu commands on any MIDI clip:

• kbot: Generate Rhythm — the styled mixer panel (algorithmic).
• kbot: Pattern Drop (AI) — describe a groove in words; a local Ollama model writes the grid. Falls back to the built-in generator if no model answers.

The pattern is written to the clip as one undo step.

Features

• 13 styles — Hip Hop, Trap, Afrobeats, House, Techno, Drum & Bass, Reggaeton, Funk, Disco, Pop, Rock, UK Garage, Breakbeat.
• Mixer-style density — a slider per drum element (8 elements). Center = the groove as written; 0 = muted; right = busier.
• Per-element controls (L S P on each element row):
  • Lock — hold a voice through Re-shuffle while every other voice re-rolls.
  • No-swing — exclude that voice from swing (keep it dead-straight).
  • No-pocket — exclude that voice from the pocket push/lay-back.
• Live preview grid — every element row shows a 16-step firing grid that mirrors the host generator step-for-step. Re-shuffle re-rolls in place and redraws live (locked rows visibly hold); Generate writes to the clip.
• Feel controls — Swing (delays off-beat 16ths), Pocket (push/lay-back), Humanize (velocity + micro-timing spread, plus Live's own velocityDeviation).
• Fills — 1 Beat, ½ Bar, or Full Bar snare/tom roll on the last bar.
• Deterministic — the same seed + config reproduces a pattern exactly (per-element mulberry32 RNG; locking freezes one voice's seed).
• Real kit mapping — reads the track's Drum Rack and targets each pad by name (then by GM note); falls back to General MIDI when no rack is present. Elements with no matching pad are flagged with * in the panel.

Pattern Drop (AI) — local models

src/ai.ts builds a strict-JSON prompt and posts it to a local Ollama (http://localhost:11434), forcing format: "json". The default model is chosen by an eval, not a guess.

npm run eval                 # score the installed models on the real prompt
npm run eval qwen3:8b phi4:14b   # only these models

The harness (eval/run.ts) runs the extension's actual buildMessages prompt across local models, parses each response with the extension's own parsePattern, and scores JSON validity · exactly-16-step rows · hat presence · backbeat/four-on-the-floor placement · latency. Results land in eval/results.json; the winner becomes DEFAULT_MODEL in src/ai.ts and tops the panel's model dropdown.

Develop / run / test

npm install
npm run build     # tsc type-check + esbuild bundle (inlines interface.html)
npm test          # vitest — pins the groove math (36 tests)
npm run start     # load into Ableton Live 12 (uses EXTENSION_HOST_PATH in .env)
npm run package   # produce a distributable .ablx

.env must point EXTENSION_HOST_PATH at your Live install's ExtensionHost/ExtensionHostNodeModule.node.

Architecture

File	Role
src/grooves.ts	Genre model: 13 styles × 8 elements × 16-step prob/velocity + feel defaults. Tune patterns here.
src/generate.ts	Pure generate(config, noteMap) → NoteDescription[] — density, swing, pocket, humanize, fills, per-element lock seeds, deterministic RNG.
src/ai.ts	Pattern Drop: prompt builder, model-output sanitizer (parsePattern), grid→notes. Host-agnostic so it stays testable.
src/interface.html	Mixer panel with per-element lock/exclude toggles + live firing-grid preview. Posts a config JSON back to the extension.
src/interface-ai.html	Pattern Drop panel — prompt + model picker.
src/extension.ts	Both context-menu commands; resolves the Drum Rack mapping, shows the panel, writes notes in a transaction.
eval/run.ts	Local-model eval harness for Pattern Drop.
*.test.ts	Vitest suites for the pure generators.

Parity with the original (and SDK limits)

Closed: all 13 styles, mixer density, fills, custom bar length, swing, pocket, per-element swing/pocket exclude, re-shuffle with per-element lock.

Blocked by the Extensions SDK beta (no API exists yet):

• Built-in Drum-Rack kit loader — the SDK exposes no preset/browser-load call (insertDevice only inserts an empty chain). We read & map an existing rack instead.
• Persistent (modeless) panel — the SDK offers only showModalDialog, so Re-shuffle previews visually in the panel rather than auditioning into the clip live. Generate commits the result.
• Auto-BPM — intentionally omitted: a step generator is beat-relative, so tempo doesn't affect the grid.

Notes / follow-ups

• Generation replaces the clip's notes (not merge).
• Playback length is governed by the clip's own loop length; Bars only sets how many bars of pattern are written.
• The preview grid shows the base groove (one bar, no fill); the fill is applied at Generate.