Voice agent detector pack

The voice pack is a first-party detector pack you activate per org. Built-in detectors keep running regardless; the voice pack adds to them, it doesn't replace anything.

Pack detectors run server-side only (in the detector worker), and, like every newly added detector, start in shadow mode: they evaluate and are visible in the dashboard's shadow view, but don't fire alerts until an operator promotes them.

1. Activate the pack

Activation is binary per org — the whole pack is on or off, no per-detector toggles.

SDK:

from dunetrace import Dunetrace

dt = Dunetrace(api_key="dt_live_...", api_url="https://your-dunetrace-host")
dt.enable_pack("voice")
print(dt.enabled_packs())   # ["voice"]
# dt.disable_pack("voice")  # turn it back off

API:

curl -X POST https://your-dunetrace-host/v1/orgs/packs/voice \
  -H "Authorization: Bearer dt_live_..."

Dashboard: Packs in the sidebar → Activate on the voice card. The org is always taken from your API key, never from the URL.

2. Emit voice events

The pack reads four event types off the run. Emit them from your voice pipeline via the dt.run() context manager, alongside the usual llm_called / tool_called hooks. All existing hooks are unchanged; these are purely additive.

voice_activity_detected's type must be one of speech_start, speech_end, silence, barge_in. turn_taking's action must be one of agent_speaking, user_speaking, both_speaking, neither. An invalid value raises ValueError at the call site.

with dt.run("voice-agent", user_input=transcript, model="gpt-4o-realtime") as run:
    run.voice_activity_detected("speech_start")
    run.transcription_received(transcript, confidence=0.94, latency_ms=110)
    run.turn_taking("user_speaking", to_user=True)

    run.llm_called("gpt-4o-realtime", prompt_tokens=200)
    reply = generate_reply(transcript)
    run.llm_responded(completion_tokens=40, latency_ms=850, finish_reason="stop", output=reply)

    run.turn_taking("agent_speaking", from_agent=True)
    run.tts_generated(reply, latency_ms=90, truncated=False)

3. The nine detectors

Every detector is silent on a run with no voice events, so activating the pack on a non-voice agent produces zero signals. Thresholds are class attributes, tunable the same way as any built-in.

VOICE_TTS_TRUNCATION and VOICE_BARGE_IN_FAILURE deliberately don't overlap: a truncation immediately after a barge_in is the correct behavior (the agent yielded), so it's excluded from the truncation count. VOICE_VAD_FALSE_TRIGGER only counts events carrying a positive duration_ms; a 0 or absent duration means "unknown", not "zero-length".

4. Voice policy actions

Three policy actions target voice agents. They follow the same contract as switch_model: Dunetrace sets intent, your agent's own loop enforces it. A policy fires, Dunetrace sets a structured attribute on the run, and your code reads it between turns. Dunetrace never touches your audio pipeline.

These are ordinary policy actions, driven by the existing trigger set (llm_latency_ms, error_count, step_count, signal, …), not gated on pack activation. Example: recover on a slow model response.

dt.add_policy(
    name="voice-latency-recovery",
    condition={"trigger": "llm_latency_ms", "operator": "gt", "value": 5000},
    action={"type": "inject_recovery_prompt",
            "params": {"prompt": "Sorry for the delay — one moment."}},
)

with dt.run("voice-agent", model="gpt-4o-realtime") as run:
    run.llm_called("gpt-4o-realtime")
    reply = generate_reply(transcript)      # slow
    run.llm_responded(latency_ms=6200, finish_reason="stop", output=reply)

    if run.recovery_prompt:                 # policy fired
        speak(run.recovery_prompt)
    if run.escalate_to_human:
        transfer_to_agent(reason=run.escalation_reason)

inject_recovery_prompt content is checked for prompt-injection patterns server-side, the same as inject_prompt, since it's spoken text with the same risk surface.

Scope

• Pack detectors run only in the detector worker (server-side). The SDK's in-process runtime path uses the fixed built-in battery and never evaluates pack detectors.
• No new RunState fields — the detectors read the four event types straight out of the run's event list.
• The voice policy actions are independent of the voice pack: a text agent that never reads run.stop_tts simply ignores it, and the actions work whether or not the pack is activated.