Agent Notify Server

One-shot alerts

Endpoint	Behavior
GET /alert_beep	Play alert_beep once.
GET /alert_done	Play alert_done once.
GET /alert_await	Play alert_await_user_input once.

Looping alerts replaces existing loop

Endpoint	Behavior
GET /loop_beep	Loop alert_beep. Replaces any prior loop.
GET /loop_done	Loop alert_done. Replaces any prior loop.
GET /loop_await	Loop alert_await_user_input. Replaces any prior loop.

Between iterations a single voice waits loop_alert_timeout_millis before replaying. Stages 1–3 can override via loop_alert_timeout_millis_{1,2,3} — each falls back to the previous stage if unset, so the loop progressively speeds up as it escalates. loop_alert_jitter_millis[_1|_2|_3] apply +/- rand(0..=jitter) per sleep.

Escalation: after the loop starts, a second voice joins at escalate_wait_1 seconds, a third at escalate_wait_2, and a fourth at escalate_wait_3. Each new voice is drawn from extra_alert_<state>_sounds; when those run out the pool cycles, doubling already-playing voices. All voices in a session share the current stage's gap, so existing voices speed up too when the supervisor advances. Voices are independent threads — they drift relative to one another.

Stop

Endpoint	Behavior
GET /stop	Stop all playback (loops and queued one-shots).

State read-only

Endpoint	Behavior
GET /state	JSON snapshot of audio engine + config. Does not change playback.

The audio object shows the current loop (loop_name = beep / done / await), how many concurrent voices are active, the current escalation stage, the gap and jitter being applied right now, and how long the loop has been running. The config object mirrors the loaded YAML so you can verify it parsed correctly.

Mixing rules

• A one-shot mixes with whatever is looping — both play simultaneously.
• Requesting a new loop replaces the prior loop.
• /stop halts everything.