harmonics-cli

harmonics-cli gives an agent or robot its own non-speech VOICE. It is the inverse of text-to-speech: instead of turning words into speech, it renders an agent's live meaning into short, pleasant sonic gestures — chimes, flutes, pulses, tonal motifs — that a listener recognizes by who is speaking and what they mean. It reproduces no words or phonemes; it maps meaning, not the sound of words (text-to-notes / text-to-audio).

A recognizable non-speech voice makes a headless agent or robot present and legible by ear: with no screen to read and no speech to synthesize or parse, you can tell who is working, who just succeeded, and who is stuck.

The design spine — intent → sound

Everything the domain renders resolves to five expressive axes:

Urgency is carried by tempo, attack, and repetition — never by dissonance or raw loudness — so the palette stays pleasant and non-fatiguing even at the urgent end, playing repeatedly next to a human without becoming an alarm.

Voice, not soundtrack

harmonics is a first-person utterance an agent emits as itself, live and in the moment, driven by its own axes. It is not a third-person spectator soundtrack. The contrast is league-of-agents' replay score, which narrates a match from the outside off an event log; harmonics is the inverse — the agent's own voice, not an ambient bed you tune out. And it is not TTS: no words, no phonemes, only meaning.

Status

The agent-first CLI, mesh identity, skill kit, and CI baseline are in place, and the voice domain has shipped. The pure text→notes core stays dependency-free and offline-testable — tests assert on note sequences, not on a speaker. The text-to-notes verbs:

• harmonics play --intent success --confidence high --urgency low — render explicit axes to a note sequence (dry-run by default; --play/--out produce sound or a file).
• harmonics say "done, tests all green" — infer the axes from a sentence, then render.
• harmonics demo — tour the whole voice in one command: dry-run by default, or --play / --html / --wav / --out / --json to hear it, write a browser-playable gallery, or stream the clips into your own code.

What you get today

• An agent-first CLI cited from teken (afi-cli) — the runtime package has no third-party dependencies.
• A mesh identity — culture.yaml (suffix + backend) and the matching resident prompt file (AGENTS.colleague.md, since this agent runs backend: colleague).
• The canonical guildmaster skill kit (11 skills) under .claude/skills/, vendored cite-don't-import. See docs/skill-sources.md.
• A build + deploy baseline — pytest, lint, the agent-first rubric gate, and PyPI Trusted Publishing wired into GitHub Actions.

Quickstart

uv sync
uv run pytest -n auto                 # run the test suite
uv run harmonics whoami      # identity from culture.yaml
uv run harmonics learn       # self-teaching prompt (add --json)
uv run teken cli doctor . --strict    # the agent-first rubric gate CI runs

CLI

Installed from PyPI as harmonics-cli (uv tool install harmonics-cli); the command you run is harmonics.

Every command supports --json. Results go to stdout, errors/diagnostics to stderr (never mixed). Exit codes: 0 success, 1 user error, 2 environment error, 3+ reserved.

Live playback (--play)

The core (text→notes plus offline WAV render) is dependency-free, so a plain install can already write audio — --wav/--out/--html need no extra. Live playback (--play) needs a real audio backend, kept out of the core deps and behind an opt-in extra:

uv tool install 'harmonics-cli[audio]'     # or: pip install 'harmonics-cli[audio]'
# already installed? re-run with --force, or inject into the existing tool env:
uv tool install --force --with sounddevice --with numpy harmonics-cli

The audio extra pulls in sounddevice (ships wheels bundling PortAudio) and numpy. A hand-installed simpleaudio also works. With no backend, --play fails with a friendly hint (exit 2) — it never crashes or silently no-ops — so fall back to --wav out.wav to write a file instead.

Contributing

See CLAUDE.md for the full conventions — the design-spine mapping, the offline-testable text→notes rule, dry-run-by-default for audio verbs, version-bump-every-PR, the cicd PR lane, and deploy setup.

License

Apache 2.0 — see LICENSE.