Devices: writing instruments and effects

Overdub's instruments and effects are devices: a small definition (id, name, params, look) plus either a Web Audio graph (build(c, kit), trusted built-ins like the ported pedals and amps) or a kernel: a few dozen lines of pure DSP that run inside an AudioWorklet. Kernels are what you and your agent write. They get no DOM, no clock and seeded randomness only, so they are deterministic and hot-reloadable; that rule is for determinism, not security (see Under the hood). A kernel written with define_device or imported from a device file is measured by the device check before it reaches a track. Kernels that arrive inside a song (a share link, a project file) run only once this browser trusts their code: until the person presses Play them they are held (an instrument silent, an effect bypassed, nothing compiled), and once allowed they are compiled as they are, without the check (ARCHITECTURE.md, "Who runs a song's code"). They add devices to the studio and never replace one it ships: a song whose device has a built-in's id (core.*, pedal.*, amp.*, cab.*, overdub.*) or a house shelf device's id brings it in as guest.<slug> (a share link) or you.<slug> (a song file), and a kernel's source can be up to 256 KB.

Where things live:

The loop: write, check, play, refine

1. Write a def (below). Start from one of the examples.
2. const report = await checkDevice(def) (or the agent's define_device tool, which runs it and registers the device when it passes). Fix every error; read the warnings.
3. Put it on a track. Change the kernel and the instance hot-reloads it: instance.reload(source) compiles the new code in the worklet and crossfades old to new over 20 ms (params kept, held notes released). A reload that fails to compile keeps the old code playing and reports the error with its line.
4. Measure, don't guess: render the song (engine.render) and measure() it.

The device check

checkDevice(def, { quick = false }) parses the kernel on the main thread without running it (a syntax error comes back with its line: compile: line 12: SyntaxError: Unexpected token ')'), then renders it in OfflineAudioContexts through the same worklet the studio uses (where a missing create() or a bad poly is refused, also as compile: ...), and reports:

ok is false on a compile error, NaN/Infinity, a peak over +6 dBTP at default settings, a runaway at an extreme setting (a raw peak over +24 dBFS), or a stuck note. Over +6 dBFS at an extreme is a warning (an EQ with every band at max is a boost the player chose, not a bug). Level, tail, cpu, latency and determinism problems are warnings too. Drum kits (cat: 'drums') are played the General MIDI drum phrase instead of the melodic one. Every message says what to change. quick: true skips the per-param extremes and shortens the renders (a few hundred ms instead of about a second).

How a device gets its face

Nobody writes UI for a device: app/src/ui/faces.js draws it from the def. Effects become a stompbox (more than six params make it a rack) in look's colour, finish, knobs and lettering, with an LED and a footswitch; instruments get a synth panel. Knobs follow each param's curve (log travel for Hz and times), show its unit, and use opts for switches. Missing look fields are picked from a hash of the id, so two devices never look alike. A device written by an agent wears its author's badge in the agent colour. A param with face: false stays off the face and lives in the device's window (Slide Rule's forty band controls do), and a device the studio ships may draw a small screen on its face: screen(ctx2d, { w, h, ink, dim }, values), a function on the def, so a song's device (JSON) never has one.

Its window. Open on a face (or a double-click on its name) opens the device big (ARCHITECTURE.md, "Device windows"): its nameplate in its look colours, every param as a big control, its presets, A/B and, for an instrument, a keyboard. The window lays the params out in sections: by a param's group (group: 'Filter'), else by key prefixes (flt_cut and flt_res make a Filter section; env1_a … env1_r an Envelope 1), else by role. A device the studio ships may name an editor of its own (editor: 'wavetable' loads ui/editors/wavetable.js); a device written in a song never runs page code of its own: whatever it says, it gets the generic window.

Slide Rule (core.eq8), the worked example. An eight-band EQ whose window (ui/editors/eq8.js) draws its bands as nodes over the live spectrum of what comes in and goes out (ctx.meter.tap). Its filter math is one module, devices/builtin/eq8-curve.js: the kernel embeds those functions as source (EQ_SOURCE, from Function.prototype.toString) and the window imports them, so the curve on screen is computed from the same coefficients the kernel runs. tools/eq-test.js holds the two together: an impulse through the kernel, its DFT at test frequencies, against the curve, for every band type (within 0.5 dB; they agree to a thousandth). The rest of it:

• Bands. b1_* … b8_*, each on, type (BELL, LOW SHELF, HIGH SHELF, LOW CUT 12/24/48, HIGH CUT 12/24/48, NOTCH, BAND PASS), freq (20 Hz to 20 kHz), gain (±24 dB, for bells and shelves) and q (a bell's, notch's or band pass's width; on a shelf or a cut the bump at its corner: 1 none, 2 a 6 dB bump, up to 4). Off by default, parked as 0 dB bells at 60, 150, 300, 700 Hz, 1.5, 3, 6 and 12 kHz; then out_gain and out_auto.
• Filters. The Audio EQ Cookbook's biquads in Direct Form I, double precision; cuts are Butterworth cascades of one, two or four sections. The host glides frequency, gain and Q each block and the kernel interpolates the coefficients sample by sample across it; a band switched on or off, or to another shape, crossfades over 12 ms (an S-curve). At its defaults it is the input, sample for sample (its golden scene's hash is the dry strum's).
• Auto gain. A first guess from the curve (a mix's long-term spectrum under K-weighting, instant), then a slow correction by what it hears: the input and the output K-weighted over about 3 s, the trim following their ratio over about 2 s, held in silence and while a band is soloed. After a few seconds of program the level is within about 0.1 LU of the dry signal on the house's test signals.
• Solo. A hidden param (solo, auto: false) the window holds on the live instance only, with an automation segment (inst.auto), never in the song: it plays one band's region of the input (a bell's band pass, what a cut takes away), and the window lets go of it on close.
• Cost. All eight bands with two 48 dB cuts and auto gain run at about 0.3% of real time.

Params that aren't knobs. A device whose state is more than a few knobs (a drawn shape, a step pattern) can still keep all of it in params, so it undoes, flips with A/B, sits in presets and travels in a song. Three fields help: hidden: true keeps a param off the face and out of the generic window (its editor draws it instead); auto: false keeps it out of the lanes (a param a pedal applies too early to follow one, or one of a shape's 195 points that would bury the Add-a-lane menu); quantum is the smallest step a window control makes (default: 1/200 of a short range, so a point on a 1/16 grid needs quantum: 1e-4). A device the studio ships may also give its def a describe(params) that returns one line of text: get_project prints that in place of the params' JSON (Scribble Strip prints its shapes).

Scribble Strip (core.shaper): a shaper you draw

Draw a shape over a beat or a bar and it moves the volume, a filter or the pan in time with the song: the pumping sidechain feel, trance gates, stutters, swells and auto-pan. app/src/devices/builtin/shaper.js; its window (editor: 'shaper') is app/src/ui/editors/shaper.js; the design note and what was measured is research/SHAPER.md.

• Three lanes: vol (a gain: the top of the shape is full level, the bottom the level pulled down by vol_depth), flt (a resonant low-pass: the top is flt_cut, the bottom closed by flt_depth, up to 8 octaves; flt_res) and pan (equal-power: at 100% pan_depth the top is hard right, 0.5 the middle). Each has <lane>_on, <lane>_depth (%) and <lane>_rate (1/32 … 2 BARS, triplets and dotted). Then smooth (ms: the edges' rise; 0.1 is a hard edge) and mix (%).
• The shapes are params: <lane>_n points (1..16), each <lane><i>_x (where in one pass, 0..1), _y (0..1), _c (the bend of the line that leaves it, -1..1: > 0 starts slow, as lanes bend) and _s (1: hold, then jump at the next point). Any order: the shape is them sorted by x, and it loops. They're hidden and auto: false.
• Timing: a pass is locked to the transport's beat (beat 0 starts one); stopped, it runs on at the tempo. A bar is four beats (a kernel isn't told the meter). The shape is read half the smoothing time ahead, so a smoothed edge is centred where it's drawn.
• At the defaults it is bypass, bit for bit (the volume lane flat at the top, the others off). Presets: Pump (quarter notes), Pump (eighths), Gate (sixteenths), Stutter, Swell over a bar, Auto-pan, Filter wobble (eighths), Half-time duck. tools/shaper-test.js holds it to all of this.

Gaffer Tape (core.multiband): three bands, up and down

The three-band compressor producers put on everything: in each band it lifts the quiet detail up (upward compression, below one threshold) and holds the loud parts down (downward compression, above another), and one DEPTH knob mixes it in. On a synth, a drum bus or a vocal it makes the sound loud, dense and finished. app/src/devices/builtin/multiband.js; its window (editor: 'multiband') is app/src/ui/editors/multiband.js; the design note, with sources and what was measured, is research/MULTIBAND.md.

• Bands: low, mid and high, split at xover_lo (120 Hz) and xover_hi (2.5 kHz); the upper split stays at least 1.5 times the lower. Each band has <band>_down_thresh and <band>_down_ratio (above the threshold it is pulled down), <band>_up_thresh and <band>_up_ratio (below it, lifted, by at most 30 dB, and nothing under about −70 dB, so silence stays silent; it never sits above the downward threshold), <band>_attack, <band>_release and <band>_gain. Then depth (%: how much of it you hear), in_gain and out_gain (±12 dB) and time (10% to 1000%: every attack and release, scaled together).
• Heard in bands, applied with shelves. A Linkwitz-Riley crossover (24 dB per octave) feeds each band's level detector; the sound itself is never split. The three gains go onto the whole sound: the mid band's on all of it, a low shelf at xover_lo for the low band's and a high shelf at xover_hi for the high band's (each two trapezoidal SVF shelves, 24 dB per octave at their steepest, halfway in dB at the split). When the gains agree the shelves are flat, so at depth 0 the output is the input, sample for sample, and the phase only turns where the bands differ.
• One source of truth. The crossover, each band's dynamics (mbStep: the detector, the static curve with 6 dB soft knees, the smoothing), the depth mix and the shelves are pure functions in devices/builtin/multiband-curve.js: the kernel embeds their source, as Slide Rule's does, and the window imports them to draw each band's curve and to run the same detectors on its live input for its readouts.
• Depth mixes each band's gain toward 0 dB before the shelves: at 0% every band is at 0 dB and the shelves are flat; at 100% it is all of it.
• Attack and the look-ahead. The sound reaches its gains 5 ms after the detectors heard it. Each band's downward side follows its mean square at the attack and release, and the same curve on the band's peak at the attack, let go within 10 ms: so a fast attack (under 5 ms) catches a hit's front, gain and makeup in place before it arrives, and a slow one lets the front through. The lift lets go within a millisecond as the band gets louder, so a hit after a quiet spot doesn't jump out, and comes back at the release.
• Levels. The defaults are the classic (each band lifted toward −26, −24 and −32 dB at 6:1 and back within 60, 50 and 40 ms; held at 3:1 above −4, −8 and −14 dB; made up by 2, 1.5 and 2 dB) at 40% depth: within a decibel of bypass on the house's test signals and drum kits. A safety ceiling, a look-ahead true-peak limiter (Red Line's method) on everything it puts out, holds it at −1 dBTP at any setting; it lets go of a short over within 10 ms and a long one over 150 ms. The look-ahead and the ceiling cost 328 samples (6.8 ms) of latency, declared. About 2% of real time.
• Presets: Full depth (the classic, all of it, a decibel up), Glue (bus), Drum smash, Vocal presence, Bass tighten, Subtle 30%. Full depth, Drum smash and Vocal presence come out 1 to 3 LU louder on drums, the crest factor down; Glue, Subtle and Bass tighten stay within a decibel or so of the input. tools/multiband-test.js holds it to all of this.
• The window says what it does: while the song plays, the loudness out against in (in LU, large, in the warning ink when it takes a decibel or more off), In and Out meters, and each band held or lifted, all worked out on the page from the window's own taps (ctx.meter.tap) and the shared functions.

Under the hood

• One processor for every kernel. ensureKernelWorklet(c) loads app/src/kernel/processor.js (the processor and the dsp stdlib) once per context, a file on the studio's own origin like every worklet module it loads: the page's policy refuses scripts from data: and blob: URLs. Each instance is an AudioWorkletNode given the kernel's source in processorOptions; the worklet compiles it with new Function in a scope where dsp, a frozen Math (whose random throws) and a small console are the only useful names. That guards against accidents; it is not a security boundary (a determined kernel can reach the worklet's globals), so treat kernels from strangers like code from strangers. What keeps the page safe is that kernels are evaluated only there: the worklet scope has no DOM, no localStorage and no network. The main thread only ever parses kernel source (compileKernel), never runs it.
• Faults are contained. A kernel that throws, or outputs NaN/Infinity, or runs past +60 dBFS is silenced and reports one { stage: 'process', message, line } error (instance.errors, instance.on('error', fn)) until it is reloaded.
• Notes are sample accurate. noteOn/noteOff/allOff(time) carry AudioContext times; the worklet converts them to frames and splits the block at each one. Offline, every context that hosts kernels gets one suspend(0) where the host waits for each kernel to acknowledge what was posted to it (instance.sync()), so notes scheduled before startRendering() land on their frames. (Don't schedule your own suspend at exactly 0 on such a context.)
• Params are posted whole (set(params), defaults merged), clamped to their range, and smoothed per block with a ~10 ms one-pole (geometric for curve: 'log'); switches and stepped params snap; set(params, { first: true }) jumps.
• Bypass (setOn(false)) crossfades wet to dry over ~10 ms in Web Audio gains; a trails device keeps its wet path open so the tail rings out. A bypassed non-trails kernel is put to sleep (skipped) once its tail has had time to die.
• Transport: the host posts { bpm, playing, beat } from the engine's clock every 50 ms and the worklet extrapolates per block, so t.beat is right at every block start. Offline it is the project tempo from beat 0, playing.
• Instance extras beyond the contract: errors, faulted, version, on('error' | 'log' | 'ready', fn), stats() ({ voices, maxVoices, steals, notes }), sync(), reload(source), node.

Expression: pitch bend, the mod wheel and the sustain pedal

A kernel reads the player's expression from t, beside the transport: t.bend (semitones), t.mod (0..1) and t.sustain (the pedal is down). A kernel that never reads them plays exactly as it did before they existed. Two things set them:

• The channel, as a MIDI keyboard sends it: instance.expr({ bend?, mod?, sustain? }, time), sample accurate like a note (the block is split there). input/midi.js sends the pitch bend wheel (its 14 bits times the bend range: 2 semitones, or what midi.setBendRange(st) or the controller's RPN 0 sets, 1..24), the mod wheel (CC 1, with CC 33's fine bits) and the pedal (CC 64) to the instrument you are playing; CC 121 resets them. The host keeps the sustain pedal for every instrument: a note-off that arrives while it is down is held (the voice is not released, and it can still be stolen first) until it lifts. (midi.js also holds the keys it captures while the pedal is down, so a kept take carries the pedal as note lengths.)
• The note, from the song, so a render repeats it: a note may carry bend (semitones) and mod (0..1), each a number for the whole note or [[beat, value], ...] from the note's start (linear between the points, the first value before the first point, the last after the last; up to 64 points, bend within ±48). The engine, the browser render and the canonical Node render send them with the note-on (noteExpr turns beats into seconds; a note chased from mid-song starts its curve that far in). Agents write them with notes.add and notes.replace.

Inside a voice's render, t.bend is the channel's bend plus the note's own and t.mod their sum (clamped to 0..1); in process they are the channel's. Evaluate them per block: they move at block rate (the host splits blocks at channel changes and evaluates a note's curve at the start of each piece). core.guitar (DI Box) bends its string with t.bend and adds finger vibrato from t.mod.

Presets

presets: [{ name, params, blurb? }] names sounds a newcomer picks by name. The registry checks and fills them in (normPresets): every param left out takes its default, values are clamped to their ranges and snapped to their steps, a switch may be given by its label ('NYLON') or its index, names are unique (any case), up to 24. So a preset's params is the whole sound: instrument.set { params } (or insert.set { patch: { params } }) applies it as it is, and presetParams(def, name) looks one up. list_devices lists their names and get_device their params for agents. A device file and a define_device call carry the field too (a bad preset is refused with the reason).

The library and device files (v0)

The house shelf is app/src/devices/library/: ten kernels Claude wrote the way any agent writes them, one plain request each (the def's request field), signed by: 'claude', registered at boot with source: 'library' (ids claude.<slug>; their faces wear Claude's badge). Instruments: Choir Loft (claude.choir-loft), Biscuit Tin (claude.biscuit-tin), Dust Sheet (claude.dust-sheet), Sub Basement (claude.sub-basement). Effects: Charity Shop (claude.charity-shop), Skylight (claude.skylight), Chopping Block (claude.chopping-block), Power Cut (claude.power-cut), Say Ahh (claude.say-ahh), Leading Edge (claude.leading-edge). A def may carry demo: { params }, the settings the library page plays it at. tools/library-test.js holds them to house levels (instruments -14 to -18 LUFS on the test phrase, effects within 1.5 LU of bypass, true peaks at or under -1 dBTP, no check warnings).

/app/library.html shows every built-in, showcase and library device as its face, with its request, its check summary (library/reports.js, written by WRITE=1 node tools/library-test.js; a summary only counts while its hash matches the kernel) and a ▶ that renders a demo in the page. "Use in a new song" opens /app/?new&device=<id>, which app/src/ui/devices-io.js turns into a track with that device and a few bars to hear it with.

Device files. .overdub-device.json = { format: 'overdub-device/0', exported, device: { id, name, kind, cat, blurb, nod?, request?, by, params, look, tail?, trails?, kernel } }. Export from a device's info card in the rack (⌥⌘E exports the selected one); import from the Song menu, ⇧⌘I, ⌘O, or a drop anywhere. An import runs checkDevice first and is refused with the report if it fails; one that passes is one device.define signed by you (the device keeps its author). A file can't take over a built-in id (core.verb comes in as you.verb), and the same file twice changes nothing. Until kernels run as WASM, a device file is code from whoever made it: import files from people you trust. Importing one trusts its code in this browser (you chose the file), as define_device trusts what your agent writes. A share link (#s=) or a song file carries the song's devices, kernels included; the ones this browser hasn't trusted are held, and the studio asks before they run.

Drum kits and the note map

Two built-in kits sit on the drums shelf (cat: 'drums'). Every kit plays the drum phrase in the device check, and a track with one gets the drum grid.

• Gobo Kit (core.drums): one hit, one voice. It has six characters: FIELD, MACHINE, DUST, 808, 909 and ACOUSTIC+.
• Studio A (core.drumroom): an acoustic kit in a big tracking room, miked like a recording. It has articulations, velocity that changes the sound, strokes that never repeat, and a mic mix you balance. Its design note is docs/research/STUDIO-A.md.

The note map. Both kits play General MIDI. Studio A plays these articulations GM has no note for:

core/music.js names them in DRUM_MAP, for grid rows: footsplash hatedge quarter half openedge flam drag roll snareedge ridechoke crashchoke crash2choke chinachoke splashchoke. It also names the GM notes that had no name: rimshot 40, lowfloor 41, himid 48, china 52, splash 55, crash2 57, rideedge 59, and tom4 43, the same note as floor. Studio A's own note map (below) labels every row.

A kit names its own notes. A drum kit's def can carry notes: what each MIDI note plays on it, in a row's words, and other, what any note it doesn't name plays.

notes: { 36: 'Kick', 38: 'Snare', 40: 'Rimshot', 31: 'Flam', 24: 'Hat 1/2 open', /* ... */ other: 'Side stick' }

• The Beat tab, the piano roll and the inspector name a track's drum rows from its kit's map (core/music.js kitNotes(def) and drumName(p, notes)). On Studio A, note 40 is "Rimshot"; on Gobo Kit it is "Snare (40)", since Gobo plays its snare there.
• A kit with no map gets General MIDI's names (GM_DRUMS). A note the kit doesn't name reads as its other and the number ("Side stick (60)").
• get_device gives agents the map, so an agent writing for a kit uses its articulations.
• Names are read as text: keys are MIDI notes 0-127 or other, values are strings of at most 40 characters.
• Gobo Kit's map sits in core/music.js (KIT_NOTES, by its id) until its def carries it.

What Studio A does with the GM notes:

• 40 is the rimshot, 59 the ride's edge, 53 its bell.
• 50 and 48 play rack tom 1, 47 rack tom 2, 45 floor tom 1, and 43 and 41 floor tom 2.
• A closed (42, 22) or pedal (44) note chokes an open hat.
• A hat note's mod (0..1) sets how open the hats are for that stroke.
• The roll (33) rolls for as long as the note is held, and its mod swells it.
• A choke is the hit, then a hand grabbing it.
• Notes the map doesn't name play the side stick, quietly.

The full table, with every piece, is in the design note.

Studio A's window (editor: 'drumroom', ui/editors/drumroom.js) draws the kit from above, from LAYOUT. That is the same layout drumroom.js builds the mics from, so the picture and the stereo image agree.

• Playing it. You play it by clicking where you'd hit (engine.liveNoteOn), and lower on a piece is louder.
• The strokes. Alt plays the other stroke (a choke, the side stick, the hat pedal), and a drag up on the hats opens them. A touch has no Alt, so a finger held on a cymbal chokes it.
• Where to hit. The toms cover most of the kick's shell, so the kick is also the floor before it, from the pedal across to floor tom 1, where its name sits. Every piece's name plays that piece.
• Lights. The pieces light as the song plays them, and fade with each piece's ring.
• Controls. Beside the kit sit the selected piece's tune, decay and level, with a key for each of its notes. Along the bottom, the mic mix is laid out as channel strips.

Studio A's params.

• The kit: kit (MAPLE BIRCH JAZZ ARENA DEAD), tune, decay, humanize, and velocity (the velocity curve).
• The mic mix: mix_close, mix_oh, mix_room, mix_crush (dB faders, -40 off), bleed, room_size, and view (DRUMMER or AUDIENCE).
• Each piece's <piece>_tune, <piece>_decay and <piece>_level for kick, snare, hat, tom1-tom4, ride, crash1, crash2, china and splash.
• snare_wires (0 is snares off) and perc_level.
• Every param carries a group (its piece, mics, kit or perc), so an editor can lay them out by piece.

A kernel technique it uses: probes. A voice renders into one stereo pair and doesn't know where in the block it starts. A kit whose mics need each piece on its own bus can't build its mix in its voices. Studio A's voices are probes:

• start queues the note.
• render counts the frames the host asks for and reads the note's t.mod, then returns true for that block only (or while a held roll lasts).
• In process, the frames counted give each note's offset in the block (n - frames).

The whole kit, its persistent piece models and its shared mic buses run there, sample accurate. Any kernel that needs per-voice buses (a mixer of sources, a sympathetic resonance between notes) can do the same.

A big instrument: Light Table (core.wavetable)

Light Table is the wavetable synth: two oscillators that sweep through tables of single-cycle frames, a sub, noise, a filter, three envelopes, four LFOs, an 8-slot mod matrix and FX. It is the largest built-in, with 114 params, and a worked example of three things a big kernel needs. The design note, with every param, table and number, is docs/research/LIGHT-TABLE.md.

• Data the page and the kernel share. A kernel sees only dsp. So when a face needs the same data as the sound, put the data in one self-contained function (no imports, nothing from the module's scope) and paste the function's source into the kernel: const LT = (${lightTables})();. The page imports the same function (app/src/devices/builtin/wavetables.js) to draw what plays. tools/wavetable-test.js checks that the kernel carries it verbatim.
• Big data, built lazily. The tables are built in the worklet. The frames nearest the playing position are built first, at once, and the rest four a block, so changing a table never holds the audio thread. Each instance holds 14.1 MB.
• Many params.
  • Keys are grouped by prefix: a_*, b_*, sub_*, noise_*, flt_*, env1_*, lfo1_*, m1_*, macro1, fx_* and voice_*. Each has a role and a desc.
  • Reads stay fast: once a block, the kernel copies the host's params into an object of fixed shape, so hundreds of reads a block cost little.
  • Nothing allocates once it runs.

To drive it, as an agent or by hand:

• Start from a preset. Each preset's blurb opens with its family, as in "Bass: a Reese, …".
• Pick a table. a_table is a switch over 14 tables. list_devices with detail: "params" lists them, and the design note says what each sounds like. a_pos moves through the table.
• Wire a mod slot. Slot n is mn_src (an index into SOURCES), mn_dst (an index into DESTS) and mn_amt, from −1 to 1.
  • The amount is in the destination knob's travel. At CUTOFF, 1 is 10 octaves, so 0.1 is an octave. PITCH is 24 semitones and FINE 1 semitone.
  • For example, { m1_src: 4, m1_dst: 1, m1_amt: 0.3 } sweeps osc A through 0.3 of its table with LFO1.
  • MACRO 1–4 are sources that do nothing until a slot uses them.
  • A slot aimed at M1 AMT scales slot 1, so MOD WHEEL → M1 AMT puts a vibrato's depth on the wheel.

For agents

Everything below is KERNEL_GUIDE from app/src/kernel/guide.js, word for word: the agent layer puts it in the agent's context. Both examples in it pass checkDevice with no warnings (tools/kernel-test.js checks that, and that the guide names every dsp function and no others).

A device is a definition plus a kernel: the source of ONE JavaScript expression that evaluates to an object. It runs in an AudioWorklet and gets dsp (below) and nothing else: no DOM, fetch, Date or timers; Math.random throws. The host does stereo I/O, polyphony, sample-accurate notes, param smoothing, hot reload, bypass, and silencing faults.

The definition

{ id: '<author>.<slug>' (lowercase a-z 0-9 . _ -, forever: songs refer to it; e.g. 'claude.tape-echo'),
  name: 'Tape Echo' (shown everywhere; may change later, unlike the id), kind: 'effect' | 'instrument',
  cat: synth keys drums bass pluck sampler | dynamics eq filter pitch drive fuzz amp mod time ambient glitch utility other,
  blurb: '<= 60 chars: what it does for the player',
  params: [ParamSpec], look: { ... }, tail?: seconds it rings after the input/notes stop (default 0),
  drone?: true if it never falls silent on its own, trails?: true to let the tail ring out when bypassed,
  presets?: [{ name: 'Felt', params: { tone: 0.2 } }] (named sounds; a param left out keeps its default),
  kernel: '<source>' }

ParamSpec, continuous: { key, label, min, max, def, curve?: 'lin' | 'log' (log needs min > 0; use it for Hz and times), unit?: 'Hz' | 'dB' | 'ms' | 's' | '%' | 'st' | 'note' | 'x', role?, desc?, step?, group? } ParamSpec, switch: { key, label, opts: ['LP', 'BP', 'HP'], def: 0 } (the kernel sees the index 0, 1, 2). role (so agents and macros find the right knob): tone level drive mix time feedback rate depth size decay attack release pitch shape width gate sens. Keys are forever; never 'id', 'on' or 'uid'. 3-5 good params beat 10. group (optional): the section a param sits in, in the device's window ('Filter', 'Envelope'); key prefixes (flt_cut, flt_res) do the same.

look (the face is drawn from it; missing fields are picked from the id): color, ink, led (hex), shape box | wide | mini | round | wah | rack, finish flat | sparkle | brushed | hammer | stripe | check, knob black | chicken | cream | chrome | small, label script | block | plate | stencil. Instruments get a synth panel in color/ink/knob.

The kernel

Effect:

({ create({ sr, seed, dsp, params }) {
     // allocate here: filters, delay lines, buffers, lookup tables
     return { latency?: samples, process(L, R, n, p, t) { /* in place: L/R hold the input, write the output */ } };
} })

Instrument (the host owns voices: poly of them, default 8, max 64; it steals the oldest released voice, then the oldest, with a 5 ms fade):

({ poly: 8,
   create({ sr, seed, dsp, params }) {
     return {
       voice(i) { return {                     // called poly + 2 times up front; i = 0, 1, 2... (seed + i for variety)
         start(pitch, vel, p) {},               // MIDI pitch (60 = C4), vel 0..1; reset ALL per-note state here
         release(p) {},                         // note-off: begin the release
         render(L, R, n, p, t) { return alive }, // ADD into L[0..n-1] and R[0..n-1]; return false once silent
         stop?() {},                            // optional: the host cut this voice
       }; },
       process?(L, R, n, p, t) {},              // optional, after the voices are summed: shared filter, chorus, reverb
     };
} })

• n is the number of frames to do now. It varies (the host splits blocks at note events): loop to n, never L.length.
• p is the params object by key, smoothed by the host (continuous params glide ~10 ms, switches and stepped params snap, values are clamped to their range). Read it; never store or mutate it.
• t = { bpm, playing, beat, bend, mod, sustain }: the transport at the start of the block (beat advances while playing) and the player's expression. Sync time to t.bpm (seconds per beat = 60 / t.bpm) and LFOs with lfo.sync(beats, t). t.bend is the pitch bend in semitones (multiply your frequencies by 2^(t.bend / 12)), t.mod the mod wheel 0..1 (vibrato, brightness, a rotor: your choice), t.sustain whether the pedal is down (the host already holds note-offs while it is). In a voice's render they are that note's own (a note can carry its own bend and mod); in process, the channel's. A kernel that ignores them still plays.
• Mono sources arrive on both L and R. Output is always stereo.
• render must return false (e.g. return env.active()) when the voice has finished, or the note counts as stuck.

Rules (the device check enforces most of them)

1. No allocation in process/render/start/release: no new arrays, objects, closures or string building per block. Create everything in create() or voice(). (Float32Array via dsp.buffer(n) in create.)
2. No Math.random (it throws): use dsp.rng(seed) / dsp.noise(seed). Same seed, same sound, every render.
3. Effects keep their level: at default settings the output should measure within 3 LU of the input (the check says "level: +x LU"). Instruments: about -14 LUFS for the test phrase (-18 is fine for plucks and drums, whose peaks are high), true peaks under -1 dBTP. Over +6 dBTP at default settings fails the check; so does a runaway (over +24 dBFS) at any setting, while over +6 dBFS at an extreme setting is a warning.
4. Heavy nonlinearities (drive, fuzz, folding, clipping with gain) alias: run them through dsp.oversample2x() or oversample4x() and declare latency (the oversampler's .latency) if you time-align a dry path.
5. Every change must be smooth: p is smoothed, but glide anything you derive from it that jumps (a delay time: dsp.smooth) so knob moves never click.
6. Guard the ends of every range: each param is tested at min and max, all at min, all at max. Clamp before log/sqrt/division; keep feedback below 1.
7. Put pow/exp/tan in coefficient formulas per block, not per sample, where you can.

dsp (all factories allocate: call them in create() or voice())

Generators have next(); one-in/one-out processors have tick(x); stereo processors have tick(l, r) and leave the WET signal in .l and .r. Times in seconds unless named ms; frequencies in Hz; gains linear unless named dB.

Math: sr, TAU, PI, clamp(x, lo, hi), lerp(a, b, t), mtof(midi), ftom(hz), dB(db) -> gain, toDb(gain) -> dB, sstep(t) smoothstep, tanh(x) (fast rational; exactly +-1 beyond +-3), softclip(x) (cubic; +-1 beyond +-1.5), hardclip(x, lim = 1), fold(x) (triangle wavefolder: identity in -1..1, folds beyond), crush(x, bits).

Random: rng(seed) -> r; r() in [0, 1), r.bi() in [-1, 1), r.gauss(). noise(seed, 'white' | 'pink' | 'brown').next() (white: uniform +-1; pink and brown: RMS about 0.3).

Oscillators: osc(shape = 'saw'), shapes 'sine' 'saw' 'square' 'pulse' 'tri' (band-limited) -> .freq(hz) (chainable) .next() (-1..1) .reset(phase = 0) .wave(shape) .phase .pw (pulse width 0.02..0.98). blep(t, dt) / blamp(t, dt): the polyBLEP / polyBLAMP residuals (phase t, increment dt) for band-limiting your own shapes. lfo(shape = 'sine', hz = 1, seed?), shapes 'sine' 'tri' 'saw' 'square' 'sh' (sample and hold) 'drift' (smooth random) -> .next() (-1..1) .uni() (0..1) .rate(hz) .sync(beats, t) (call once per block: one cycle per beats, locked to t.beat while playing) .phase .offset (0..1, phase offset used by sync).

Filters: svf() (zero-delay state variable; stable at any setting; retune every sample if you like) -> .set(fc, q = 0.707, gainDb = 0), then ONE of .lp(x) .bp(x) .hp(x) .notch(x) .peak(x) (resonant peak) .allpass(x) .bell(x) (EQ bell by gainDb) .lowshelf(x) .highshelf(x) per sample (each call advances the filter); or .tick(x) then read .low .band .high. q: 0.5 soft, 0.707 flat, 2-10 resonant, 20+ ringing. One filter per channel. onepole(fc?) -> .set(fc) .lp(x) .hp(x). dcblock() -> .tick(x). biquad() -> .set(type, fc, q = 0.707, gainDb = 0) with type 'lp' 'hp' 'bp' 'notch' 'allpass' 'peak' 'lowshelf' 'highshelf', then .tick(x). (set costs trig: per block.)

Delays: delay(maxSamples) -> .read(d) (linear, d >= 1 samples) .cubic(d) (Hermite, d >= 2, for modulated delays) .write(x) .ms(ms) -> samples .clear(). Read BEFORE you write each sample: read(d) is the input from d samples ago. allpass(len, g = 0.5) -> .tick(x). comb(len, fb = 0.8, damp = 0.2) -> .tick(x) .set(fb, damp).

Envelopes: adsr(a = 0.005, d = 0.1, s = 0.7, r = 0.2) -> .gate(on) .hit() (attack, then release on its own) .next() (0..1) .active() .set(a, d, s, r) (cheap when unchanged: fine per block) .reset(). Exponential segments; a retrigger attacks from the current level (no click). ar(a = 0.002, r = 0.3): the same, holding at 1 while gated. follower(attackMs = 5, releaseMs = 100) -> .tick(x) (the level of x). smooth(ms = 10, init = 0) -> .tick(target) .reset(v) .value. slew(riseMs = 10, fallMs = riseMs, init = 0) -> .tick(target).

Oversampling: oversample2x() / oversample4x() -> .process(x, fn) runs fn (a one-sample function you make ONCE in create, reading variables you update per block) at 2x / 4x and returns one band-limited sample. .latency = 23 / 27.5 samples. Use one per channel.

Physical models: karplus(hz = 220, decay = 3, bright = 0.5, seed?) (plucked string) -> .pluck(vel = 1, hz?) .next() .freq(hz) .set(decay, bright) (from the next pluck) .mute(t60 = 0.08) (damp it: a release) .active(). modal(freqs, decays, gains?) (resonator bank: drums, bells, bars; arrays; decays are T60 s; gains default 1/n) -> .strike(vel = 1) .tick(x) (excite with a signal) .next() .tune(ratio) .damp(k) (k < 1 chokes) .active().

Space: fdn(size = 0.6, decay = 2, damp = 0.4, seed?) (8-line modulated reverb; size 0..1 room to hall, decay = T60 s, damp 0 bright .. 1 dark) -> .tick(l, r) then .l .r (wet) .set(size, decay, damp). chorus(depth = 0.5, rate = 0.8) (two voices in quadrature) -> .tick(l, r) then .l .r (wet) .set(depth, rate). buffer(n) -> Float32Array(n).

What the check reports (define_device returns it)

{ ok, errors, warnings, level: { lufs, deltaLU }, truePeak, nan, tail: { seconds, decays }, cpu: { pct }, latency: { samples }, deterministic, extremes: { cases, failed }, voices?: { poly, maxVoices, steals }, stuck? }. ok is false on a compile error (with the line), NaN/Infinity, a peak over +6 dBTP at defaults, a runaway at an extreme setting, or a stuck note. Effects are rendered with a DI guitar strum and a drum loop; instruments play chords, a melody, a fast run, low to high notes and soft to hard velocities. Fix every error; act on warnings unless you mean them.

Example: an effect

{id: "claude.tape-echo", name: "Tape Echo", kind: "effect", cat: "time", by: "claude", blurb: "Tempo-synced echoes that darken as they repeat", params: [{key: "division", label: "TIME", opts: ["1/16", "1/8", "1/8.", "1/4", "1/2"], def: 2}, {key: "feedback", label: "REPEATS", min: 0, max: 95, def: 40, unit: "%", role: "feedback"}, {key: "tone", label: "TONE", min: 500, max: 12000, def: 3500, curve: "log", unit: "Hz", role: "tone"}, {key: "mix", label: "MIX", min: 0, max: 100, def: 30, unit: "%", role: "mix"}], look: {color: "#8a5a2b", ink: "#fff3e0", shape: "box", finish: "hammer", knob: "cream", label: "script", led: "#ffb347"}, tail: 6, trails: true, kernel: `
({
  create({ sr, seed, dsp }) {
    const BEATS = [0.25, 0.5, 0.75, 1, 2];
    const max = sr * 4;
    const dl = dsp.delay(max), dr = dsp.delay(max);
    const tl = dsp.onepole(), tr = dsp.onepole();
    const time = dsp.smooth(80, sr / 4);          // the delay time glides like tape (in samples)
    const wow = dsp.lfo('sine', 0.6);
    return {
      process(L, R, n, p, t) {
        const target = Math.min(max - 8, BEATS[p.division] * 60 / t.bpm * sr);
        tl.set(p.tone); tr.set(p.tone);
        const fb = p.feedback / 100, mix = p.mix / 100;
        for (let i = 0; i < n; i++) {
          const d = time.tick(target) + 6 * wow.next();
          const yl = dl.cubic(d), yr = dr.cubic(d);
          dl.write(L[i] + dsp.tanh(tl.lp(yl) * fb));
          dr.write(R[i] + dsp.tanh(tr.lp(yr) * fb));
          L[i] += yl * mix;
          R[i] += yr * mix;
        }
      },
    };
  },
})` }

Example: an instrument

{id: "claude.glass-harp", name: "Glass Harp", kind: "instrument", cat: "pluck", by: "claude", blurb: "Plucked strings with a soft glassy shimmer", params: [{key: "decay", label: "DECAY", min: 0.3, max: 8, def: 3, curve: "log", unit: "s", role: "decay"}, {key: "bright", label: "BRIGHT", min: 0, max: 1, def: 0.6, role: "tone"}, {key: "shimmer", label: "SHIMMER", min: 0, max: 100, def: 30, unit: "%", role: "mix"}], look: {color: "#3b4f7a", ink: "#eef3ff", knob: "chrome", led: "#9fd8ff"}, tail: 8, kernel: `
({
  poly: 8,
  create({ sr, seed, dsp }) {
    const ch = dsp.chorus(0.6, 0.35);
    return {
      voice(i) {
        const s = dsp.karplus(220, 3, 0.6, seed + i);
        const glass = dsp.osc('sine'), env = dsp.adsr(0.002, 1.2, 0, 0.4);
        let gl = 0, gr = 0, v = 0;
        return {
          start(pitch, vel, p) {
            s.set(p.decay, p.bright).pluck(vel, dsp.mtof(pitch));
            glass.freq(dsp.mtof(pitch + 12)).reset(0);
            env.gate(true);
            const pan = dsp.clamp((pitch - 60) / 48, -0.5, 0.5);   // low notes left, high notes right
            gl = Math.cos((pan + 0.5) * Math.PI / 2); gr = Math.sin((pan + 0.5) * Math.PI / 2);
            v = vel;
          },
          release(p) { s.mute(0.15); env.gate(false); },
          render(L, R, n, p) {
            const sh = 0.25 * p.shimmer / 100 * v;
            for (let i = 0; i < n; i++) {
              const y = 0.62 * s.next() + sh * glass.next() * env.next();
              L[i] += y * gl; R[i] += y * gr;
            }
            return s.active() || env.active();
          },
        };
      },
      process(L, R, n, p) {
        for (let i = 0; i < n; i++) {
          ch.tick(L[i], R[i]);
          L[i] = dsp.softclip(L[i] + 0.3 * ch.l);     // gentle ceiling for big chords
          R[i] = dsp.softclip(R[i] + 0.3 * ch.r);
        }
      },
    };
  },
})` }