Conversation timing

A Conversation holds a dialogue. A Task processes supplied evidence without waiting for participant input. Use Conversation → Task → Conversation to conduct an interview, generate its assessment, and discuss feedback.

Completion and time

Minimum and maximum time are independent of model completion. Times begin when the Conversation starts, including speech, silence, thinking pauses, and tool work.

Control	Default	Behavior
Minimum seconds	0	Prevents model-declared completion before this time.
Maximum seconds	Unset	Starts a bounded ending even when the participant is silent.
Allow model completion	Enabled	Allows completion when the authored objective and optional completion condition are satisfied.
Wrap up seconds	30	Warns the model before the maximum so it can close naturally.
Grace seconds	30	Allows the current turn to finish after the maximum; remaining speech stops when grace expires.

Minimum cannot exceed maximum. Set the execution timeout beyond maximum plus grace. The workflow’s overall maximum remains authoritative. Explicit user exit, inactivity, and failures can end a node before its minimum. For a 45-minute interview that may finish after 10 minutes when its objective is satisfied:

completion:
  minimum_seconds: 600
  maximum_seconds: 2700
  model_completion: true
  condition: Cover the supplied interview topics and their follow-up questions.
  wrap_up_seconds: 30
  grace_seconds: 30

To run until the scheduled time, set model_completion: false. For model-led completion with no minimum, set minimum_seconds: 0. Question-count instructions are prompt guidance rather than enforced question-bank progress counts.

Inactivity

Check-ins begin after assistant playback finishes while the Conversation is waiting for the participant. Speech, model generation, playback, and relevant tool work suppress them.

inactivity:
  enabled: true
  after_seconds: 10
  retry_seconds: 30
  max_check_ins: 2
  thinking_seconds: 60
  instructions: Gently offer clarification or more thinking time without supplying an answer.

With these defaults, the assistant checks in after 10 seconds, waits 30 seconds after playback before checking again, then waits another 30 seconds before returning inactive. Assistant speech does not reset the unanswered-check-in count. Participant speech cancels pending reminders; a committed response resets the count. “Let me think” defers the next check-in for 60 seconds. Explicit stop/pause requests suppress reminders until the participant resumes. Overall time limits still apply. Set enabled: false to disable check-ins for a node.

Route each ending

Outcome	Meaning
`complete`	The model finished and returned schema-valid output after closing speech was delivered.
`time_limit`	The configured maximum ended the conversation, with bounded turn grace.
`inactive`	The unanswered-check-in budget was exhausted.
`user_ended`	The participant explicitly requested the end of this conversation.
`error` / `timeout`	Processing failed or an execution limit was exceeded.

Only complete guarantees structured output. Other endings retain the transcript. Route partial interviews into a grading Task that marks unasked or unanswered topics as unassessed. Route feedback endings to End. Collection steps should use a failure or recovery route when required fields are missing.

Get started

Console

Workflows

Session API

API reference

Help

Conversation timing

Completion and time

Inactivity

Route each ending

​Completion and time

​Inactivity

​Route each ending

Completion and time

Inactivity

Route each ending