Changelog

The long_running tool config gains a wait_audio object that loops a bundled sound while the tool runs — hold music, or an ambient bed under the spoken filler — for the caller who expects to hear something other than an agent repeating itself.

wait_audio has one required field, preset, one of:

office · city · forest · crowded_room · keyboard_typing · hold_music

The clip is armed on the same dwell as the spoken filler (filler_delay_ms of the tool running with nobody speaking) and stops the moment the tool settles, so the real answer never queues behind it. It is independent of filler_message: set either, both, or neither — omit wait_audio and the wait stays silent, exactly as today. Presets share one roster and the same per-preset level-matching with the agent’s background-noise bed; the wait clip plays at a fixed level while the bed follows the agent’s configured background_noise_volume, so a given preset sounds identical on both only when that volume is left at its default.

This extends the long-running-tool controls first shipped on 2026-08-12; existing long_running configs are unaffected.