Analysis hooks

A voice agent can start ordinary workflows while it talks, and one more when it hangs up. This is where the rest of Assemblix reaches the conversation: extract the caller's details, write them to a CRM, score the call, raise an alert.

Two rules hold for both hooks:

• The conversation never waits. A hook runs in the background; the agent keeps talking.
• The result does not come back. In this version hooks observe, they do not steer. A workflow that fails is logged and dropped — it cannot end a call.

The workflows are normal workflows. Nothing about them is voice-specific, and you can run and debug them by hand like any other.

Per-turn workflow

Runs once for every finished thing the caller says.

{
  "message": "I'd like to book an appointment for Tuesday",
  "voice": {
    "session_id": "…",
    "turn_index": 3,
    "agent_reply": "Of course — what day suits you?"
  }
}

message is the caller's utterance. agent_reply is what the agent said just before, because a single utterance out of context is rarely enough to act on.

Reach these from a node with input.message, input.voice.turn_index, and so on.

Final workflow

Runs once, after the call ends.

{
  "message": "user: hello\nassistant: hi, how can I help?\n…",
  "voice": {
    "session_id": "…",
    "transcript": [{ "role": "user", "text": "hello" }],
    "duration_sec": 74.2,
    "end_reason": "user_hangup"
  }
}

message is the whole conversation flattened into text — usually what you want to hand an agent node. voice.transcript is the same thing structured, for anything that needs to walk the turns. end_reason is one of user_hangup, timeout or error.

Reading the results

Every hook run is attached to the call that started it. Open a call from the agent's Calls tab and its analysis runs are listed beside the transcript, each linking into the execution viewer.