liaison

idle

control url system prompt voice id (optional) first message resume message (SPEC §10) idle suspend, seconds (SPEC §10) wake on speech (docs/wake-on-speech.md)

Two parties, two places. wake is the SDK's and lives here; the idle window it undoes is config.idle.suspendAfterSec, which belongs to whoever mints the token. A short window without wake is a session the user has to tap back once a minute. And with wake on a suspended session is still capturing: the recording indicator stays lit on a session this page calls suspended, which is the honest cost of not needing a tap.

input level meter (onLevel, docs/sdk-web.md §7)

On by default — the meter below is what this row is for. But passing onLevel is the only switch there is, and on a wake-less session it is also what builds the audio graph, so unticking this and wake is the only way to see the default path from here: the raw capture track straight into the transceiver, no AudioContext, no worklet, nothing between the microphone and Opus. Mute is still live on that path — it disables the capture track instead of gating the worklet, which is what makes the button the proof. Read once at connect, like wake.

server api key (dev mint, SPEC §8)

Dev only. This is a server key: it mints tokens for any tenant, any config and any TTL. ?apiKey= puts it in your browser history and in every proxy log on the way here. In production your backend holds this key and mints the token; the browser only ever sees the token.

input level (measured before the mute gate)

input level — no capture graph

Mute is local and takes no frame on the wire: with the audio graph in the path the worklet gates its output to digital zeros, and without one the capture tracks are simply disabled. It survives suspend, resume, reconnect and a device switch.

microphone speaker

This browser has no setSinkId, so a page cannot choose the sink: output routing stays with the OS. The picker is disabled rather than silently doing nothing.

Device names are empty until a getUserMedia grant exists — the browser's rule, not ours — so before the first connect these read Microphone N / Speaker N and are re-enumerated once a session is up. Both pickers work before connecting, too: the choice is passed as inputDeviceId / outputDeviceId and stays sticky across every later capture.

running jobs

wake counters (kept across reloads — the overnight false-wake count)

raw frames (last 200)

set_note (client tool)