Limits and timeouts

Backend

Limit	Value	Configurable	Effect when reached
Text of one turn	64 KiB (65 536 bytes, UTF-8)	no	Refused with too_large.
Audio of one turn	60 seconds	no	Refused with too_large.
Time to produce an answer	30 seconds, from the start of the agent choice to the end of the answer, tool calls included	no	Turn fails with the standard message.
Tool calls per turn	maxToolCalls	per agent	Turn fails with the standard message.
Turns per conversation at a time	1	no	Further turns on that conversation are refused with occupied. Turns on other conversations run in parallel.
Incoming audio streams per session	1	no	session.error with audio_busy.
Model request retries	0	no	A model error fails the turn.
Second generation after an answer without words	1	no	The agent is asked once more to write its answer, without tools. If that is empty too, the turn fails with the standard message.
Turns the models receive	AMBROGIO_HISTORY_TURNS (20), the new turn included	yes	Earlier turns stay stored and visible in the clients, but the orchestrator and the agents do not see them.
Session ping interval	30 seconds	no	A session that does not answer by the next ping is closed.
Shutdown after SIGTERM or SIGINT	10 seconds	no	The backend logs server_shutdown_timeout and exits with code 1. Give the container at least 15 seconds to stop.
Conversations per list page	AMBROGIO_LIST_PAGE_SIZE (20), at most AMBROGIO_MAX_LIST_PAGE_SIZE (100)	yes	See Backend configuration.
Speech synthesis threads	4, or fewer if the CPU has fewer	no
Speech recognition threads	2, or fewer if the CPU has fewer	no
Silence added at the start of a voice request	1 second	no	Keeps the first word, which the bundled model loses when speech starts at once.
Silence added at the end of a voice request	0.8 seconds	no	Lets the recognizer finish the last words before the transcription is returned.

Speech recognition runs while the request is open: each audio frame goes to the recognizer as it arrives, and at input.end only the last part remains to decode. The recognizer runs in one worker thread shared by all voice requests, which processes frames in the order they arrive. If the CPU cannot keep up with the audio, frames wait and the text arrives later; no audio is lost. Real-time recognition needs an x86_64 CPU with AVX2; see Before you install. If the worker stops because of a fault, the transcriptions in progress fail, their turns are refused with transcription_failed (“I didn’t understand the request.”), and the next voice request starts a new worker.

Turns are never queued and never retried automatically.

Satellite

Setting	Value	Configurable
Wake word threshold	0.25	AMBROGIO_WAKE_WORD_THRESHOLD
Length of the wake word	64 characters	no
Speech needed around a wake word	some speech in the 1.5 seconds that end with the detection; otherwise the detection is discarded	no
Pause between wake-ups	2 seconds	no
Audio kept from before you start speaking	1.5 seconds	no
Speech detection window	32 ms (512 samples at 16 kHz)	no
Speech probability that counts as speech	above 0.26	no
Speech probability that counts as silence	0.11 or lower	no
Speech needed to start a request	128 ms	no
Silence that ends a request	640 ms	no
Continuous speech before the request is closed at the next short pause	55 seconds	no
Wait for the backend to confirm a new session	5 seconds; then the satellite closes the connection and treats the backend as unavailable	no
Wait for an outcome after you stop speaking	60 seconds	no
Wait for a follow-up reply	15 seconds, restarted after each answer	AMBROGIO_REPLY_TIMEOUT_MS
Level of other local audio during an interaction	set per installation, 100% otherwise	AMBROGIO_ATTENUATION_LEVEL
Time allowed for each amixer call	2 seconds	no
Shutdown after SIGTERM or SIGINT	5 seconds	no

The satellite does not listen while it is speaking: you cannot interrupt an answer.

Terminal client

Limit	Value
Minimum terminal size	70 × 14 characters
Largest message	64 KiB in UTF-8, checked before sending