genTone: Musical chirps while Hermes streams

A Hermes Desktop Plugin (single ESM file, @hermes/plugin-sdk). While the agent streams text, it emits short robotic chirps whose pitch follows the generation speed, quantized to a three-octave scale ladder starting at C4 (261.63 Hz). Faster generation → higher notes, slower → lower, so the tempo of the model paints a melody in the scale. The melody also goes into little improvised "wanderings" every once in a while. When the model is thinking, the same speed-mapped notes continue at reduced volume (~70%), so the thinking phase sounds like the answer but quieter. Sound stops the moment the last stream ends.

Six scales are selectable from the status-bar chip, all rooted at C and defaulting to Blues:

• Blues (C minor blues) — C · E♭ · F · F♯ · G · B♭ — 18 degrees
• Pentatonic (C minor pentatonic) — C · E♭ · F · G · B♭ — 15 degrees
• Major (C major) — C · D · E · F · G · A · B — 21 degrees
• Phrygian Dominant (1–♭2–3–4–5–♭6–♭7) — C · D♭ · E · F · G · A♭ · B♭ — 21 degrees
• Diminished (half-whole diminished) — C · D♭ · E♭ · E · G♭ · G · A · B♭ — 24 degrees
• Whole Tone (all whole steps) — C · D · E · F♯ · G♯ · A♯ — 18 degrees (six per octave)

Every ladder spans three octaves from the C root; Blues/Pentatonic/Phrygian/ Diminished/Whole-Tone top out at B♭6 (≈1865 Hz) and Major at C7 (≈1976 Hz).

Install

1. Copy plugin.js to ~/.hermes/desktop-plugins/gentone/plugin.js (the folder name must match the plugin id: gentone; under a named profile it's ~/.hermes/profiles/<name>/desktop-plugins/gentone/).
2. The app loads it within a few seconds. If nothing appears, run ⌘K → Reload desktop plugins. On the latest Hermes builds the plugin may not register until you fully relaunch Hermes Desktop — quitting and starting the app again is the reliable way to force it to load.
3. If you hear nothing, click the ♪ chip once — that's the user gesture that unlocks the audio context (Electron autoplay policy).

The chip

The status-bar chip is the scale selector. It shows ♫ while streaming, ♪ idle, and ♪ off muted, followed by the current scale name. Picking any scale always enables the plugin. The Volume slider at the bottom of the picker sets the loudness: its minimum (0) mutes — the chip then reads ♪ off — and dragging above 0 re-enables at that volume. The slider rests at the original volume by default and tops out 20% louder than that.

How it works

• Signals — listens to gateway events via host.onEvent: message.start, message.delta, thinking.delta, reasoning.delta (all carry payload .text), message.complete, and error.
• Speed — chars streamed over a rolling 700 ms window (text and thinking tokens) → chars/sec, EMA-smoothed onto the selected scale's ladder via the dynamic mapping below.
• Dynamic speed → pitch mapping — the pitch bounds are no longer fixed: each tick records the smoothed cps into a rolling RANGE_WINDOW_MS (16 s) history, and every REMAP_INTERVAL_MS (8 s ≈ 2 phrases) the min/max of that history are fitted across the whole three-octave ladder. A steady tps band therefore still spans all three octaves instead of hovering on one note. MIN_SPAN_LOG2 keeps a very steady stream from becoming hypersensitive and RANGE_PAD keeps the top and bottom degrees reachable; a new stream resets to the initial MIN_CPS/MAX_CPS bounds.
• Melodic wander — each streaming note random-walks ±WANDER (2) scale degrees around the speed-mapped pitch, so steady generation wanders melodically instead of holding one note. Switching scales re-anchors the walk and clears any active jitter so a perturbation can't leak across scales.
• Smooth jitter — a constant-probability perturbation, independent of how monotone the melody is. With chance JITTER_PROB (0.2) per eligible note, a short "bump" starts: a pattern of 1–JITTER_MAX_LEN (4) notes whose peak deviation is 1–JITTER_MAX_PEAK (4) degrees. The bump rises toward the peak and falls after it (each step ±1, never overshooting the envelope), so the melody is bent continuously across a few notes instead of a single note jumping. The sign is random, so a jitter lifts or dips the line. A jitter never starts while one is active, and once it ends a JITTER_GAP (8) note quiet period passes before the next can start.
• Time-locked rhythm in phrases of 4 s — the scheduler is a setTimeout chain locked to the wall clock, not a fixed setInterval. Each phrase is exactly RHYTHM_PERIOD_MS (4 s) long, split into RHYTHM_BARS (8) bars of 500 ms, each subdivided into eighth-note slots (RHYTHM_SLOTS = 16, i.e. 250 ms). Every phrase rolls a fresh random rhythm: a note count drawn from RHYTHM_MIN_NOTES–RHYTHM_MAX_NOTES (5–9) is placed across the slots, with the downbeat (slot 0) always anchored and bar starts weighted RHYTHM_DOWNBEAT_WEIGHT (3×) more heavily than offbeats so the phrase feels grounded even as the count and placement vary. The grid is anchored to performance.now(), so the 4 s phrases stay steady with no drift and a new phrase always begins exactly on the boundary — even across silent gaps. Each note that fires still gets its pitch from the stream-speed mechanics above, so the rhythm varies but the melody follows generation.
• No stacking — each chirp's envelope (NOTE_LEN_MS = 70 ms) stays shorter than the shortest possible note spacing (one slot = 125 ms), so notes physically cannot overlap or queue. When the last stream ends the scheduler stops and the master gain fades in ~30 ms — no trailing notes.
• Pause "interlude" beat — while a stream is active but no tokens flow for STALE_MS (tool calls, waits), the same wall-clock-locked rhythm keeps playing as a static-frequency hat-like beat at BEAT_FREQ (1200 Hz) instead of silence — an interlude at the same tempo and placement complexity as the notes, using the same chirp synthesis with the note choice removed. The instant tokens resume the speed-mapped notes take over again; when the stream fully ends the beat stops with everything else (nothing plays after message.complete).
• Thinking at reduced volume — thinking/reasoning tokens (thinking.delta / reasoning.delta) feed the same speed window and play the same speed-mapped, wandering notes at ~70% of whatever the volume slider is set to (so thinking stays quieter at any volume). The latest token type decides the volume class, so the transition is instant. Gaps with no tokens at all (tool calls, waits) switch to the pause "interlude" beat once STALE_MS (900 ms) pass.
• Volume slider — the Volume slider at the bottom of the picker sets the streaming volume (userVolume, 0–VOLUME_MAX). Its default is the original VOLUME (0.07); VOLUME_MAX is VOLUME × 1.2 (0.084), so the ceiling is 20% louder than the default. The minimum (0) mutes and reads ♪ off in the chip — it replaces the old explicit Off row — and any value above 0 re-enables and resumes the audio. Thinking notes follow the slider via THINKING_RATIO (~70%).
• Any session — notes play while any session streams, focused or not (the tab being hidden also silences notes).
• Hot-reload safe — a global engine guard disposes the previous module's interval + listener on file save, so re-loading the plugin can't duplicate notes.

Tunables (top of plugin.js)

Constant	Default	Meaning
ROOT_FREQ	261.63	Hz of scale degree 0 (C4)
MINOR_BLUES	[0,3,5,…,34]	semitone offsets, 3-octave minor blues (18 degrees)
MINOR_PENTA	[0,3,5,…,34]	semitone offsets, 3-octave minor pentatonic (15 degrees)
MAJOR	[0,2,4,…,35]	semitone offsets, 3-octave major (21 degrees)
PHRYGIAN_DOM	[0,1,4,…,34]	semitone offsets, 3-octave phrygian dominant (21 degrees)
NOTE_LEN_MS	70	chirp length — keep ≤ 125 ms (one rhythm slot)
BEAT_FREQ	1200	Hz of the static-frequency pause "interlude" beat (hat-like)
RHYTHM_PERIOD_MS	4000	fixed phrase length — every phrase is exactly this long
RHYTHM_BARS / RHYTHM_SLOTS	8 / 32	bars per phrase, and placement resolution (sixteenth notes)
RHYTHM_MIN_NOTES / RHYTHM_MAX_NOTES	10 / 18	random note count range per phrase
RHYTHM_DOWNBEAT_WEIGHT	3	bar starts weigh this much more than offbeats when placing notes
WINDOW_MS	700	speed rolling window
MIN_CPS / MAX_CPS	6 / 100	initial mapping bounds until samples arrive (see dynamic range below)
RANGE_WINDOW_MS	16000	rolling cps history used to estimate the speed range
REMAP_INTERVAL_MS	8000	how often the dynamic pitch bounds are recomputed (~2 phrases)
MIN_SPAN_LOG2	0.5	min span the dynamic mapping covers (sensitivity floor)
RANGE_PAD	0.10	bounds padded outward by this fraction of the span
STALE_MS	900	token gap before notes switch to the pause "interlude" beat
EMA	0.3	pitch smoothing (1 = none, 0 = frozen)
WANDER	2	±scale degrees of pitch variation around the mapped note
JITTER_PROB	0.2	per-note chance a smooth jitter starts (constant throughout, not monotony-based)
JITTER_MAX_LEN / JITTER_MAX_PEAK	4 / 4	max notes in a jitter bump, and max peak deviation in scale degrees
JITTER_GAP	8	quiet notes after a jitter ends before the next can start
VOLUME	0.07	default streaming volume (0–1) — where the picker slider rests with no input
VOLUME_MAX	0.084	slider ceiling = VOLUME × 1.2 (20% louder than the default)
THINKING_VOLUME	0.05	default thinking-note volume (~70% of VOLUME)
THINKING_RATIO	0.05/0.07	thinking notes scale by this (~70%) of whatever the slider sets
SHIMMER	0.007	detune of the second oscillator (robot color)

Notes

• Built against the new desktop plugin SDK (@hermes/plugin-sdk disk format, hot-reload, no build step) — needs a recent desktop build.
• Requires Web Audio API in the renderer (standard Electron/Chromium — no extra deps, no imports beyond @hermes/plugin-sdk, react, react/jsx-runtime).
• Verified: JS syntax (node --check), and all four scale ladders parse from the source as monotonically ascending from C4 (261.63 Hz) to ≈1865–1976 Hz. Not yet runtime-tested on a real desktop app instance — that's the copy-and-try step on your side. If the chip never appears or JS errors, check hermes logs gui -f and the app's error toast.