agents-go docs

TurnOptions (set on AgentSessionOptions.Turn) tunes how the session decides a user turn ended, how interruptions work, and speculative generation. Every knob has a safe default — you only set what you need.

session := agents.MustNewAgentSession(agents.AgentSessionOptions{
    Turn: agents.TurnOptions{Mode: agents.TurnModeVAD},
})

Turn-detection mode

`Mode`	Value	Ends a turn on
`TurnModeAuto`	`""` (default)	VAD if configured, else STT, else manual.
`TurnModeVAD`	`"vad"`	VAD end-of-speech + endpointing delay.
`TurnModeSTT`	`"stt"`	STT final transcripts + endpointing delay.
`TurnModeManual`	`"manual"`	Nothing automatic — you call `GenerateReply`.

Detector (a TurnDetector) optionally supplies a model-assisted end-of-turn detector. It is opt-in and external: build one from the inference package (inference.NewHTTP(baseURL) → inference.NewEOUTurnDetector(client)) pointed at the Python EOU sidecar, and assign it here. Leave it nil to use plain VAD/STT endpointing with no sidecar. See Latency & interruption.

Endpointing

Field	Default	Notes
`MinEndpointingDelay`	500ms	Silence debounce before committing a turn.
`MaxEndpointingDelay`	3s	Upper bound when dynamic endpointing / an EOU model extends the delay.
`DynamicEndpointing`	false	Adapt the delay to the user's observed pauses.
`EndpointingAlpha`	0.9	EMA coefficient for dynamic endpointing (`0 < α < 1`).

Interruption

Field	Default	Notes
`AllowInterruptions`	true	Enable barge-in during playout.
`InterruptionMinDuration`	500ms	Minimum user speech before it interrupts.
`InterruptionMinWords`	0 (off)	Minimum recognised words before interrupting (STT only).
`DiscardAudioIfUninterruptible`	true	Exclude audio from STT while an uninterruptible reply plays (VAD still sees it).

False interruptions

Field	Default	Notes
`ResumeFalseInterruption`	true	Resume a paused reply if no real turn is committed. Requires an output that reports `CanPause`.
`FalseInterruptionTimeout`	2s	Silence before a paused reply is classed false and resumed. Non-positive disables.

Reply guards

Field	Default	Notes
`MaxSpeechDuration`	off	Cap one reply after its first audio frame.
`MaxSpeechRetries`	0	Replacement replies allowed after `MaxSpeechDuration` cuts one off.
`MaxUserTurnWords`	0 (off)	Cap final-transcript words before the agent must speak.
`MaxUserTurnDuration`	off	Wall-clock cap before the agent must speak.

Preemptive (speculative) generation

Speculative replies are generated from final STT transcripts and promoted only if the committed transcript and request snapshot are unchanged.

Field	Default	Notes
`PreemptiveGeneration`	true	Enable speculative replies.
`PreemptiveTTS`	false	Also start synthesis speculatively (buffered, never played until promotion).
`PreemptiveMaxSpeechDuration`	10s	Skip speculation past this much user speech.
`PreemptiveMaxRetries`	3	Speculative requests per user turn.

*bool fields (e.g. AllowInterruptions) distinguish "unset → default" from an explicit false. Use a pointer helper to set them.

TurnOptions