On this page

On this page

Messages and delivery

Typing indicators

Typing indicators are sent to the chat channel while a run is active. Use agents.defaults.typingMode to control when typing starts and typingIntervalSeconds to control how often it refreshes (keepalive cadence, default 6 seconds).

Defaults

When agents.defaults.typingMode is unset:

  • Direct chats: typing starts immediately once the model loop begins.
  • Group chats with a mention: typing starts immediately.
  • Group chats without a mention: typing starts when the admitted run has user-visible activity, such as harness execution activity or message text.
  • Message-tool-only replies: typing starts immediately, even in a group chat without a mention, so members can see the agent working when no automatic final reply is posted. This takes precedence over the group rules above; an explicit typingMode still wins over it.
  • Session events: background events suppress typing while their ordinary session turn runs.

Modes

Set agents.defaults.typingMode to one of:

  • never - no typing indicator, ever.
  • instant - start typing as soon as the model loop begins, even if the run later returns only the silent reply token.
  • thinking - start typing on the first reasoning delta, or on active harness execution after the turn is accepted.
  • message - start typing on the first user-visible reply activity, such as active harness execution or a non-silent text delta. Silent reply tokens such as NO_REPLY do not count as text activity.

Order of "how early it fires": never -> message/thinking -> instant.

Configuration

Set the agent-level default:

json5
{  agents: {    defaults: {      typingMode: "thinking",      typingIntervalSeconds: 6,    },  },}

Override the policy for one agent:

json5
{  agents: {    entries: {      support: {        typingMode: "message",      },    },  },}

Notes

  • message mode does not start from silent reply tokens, but active execution can still show typing before any assistant text is available.
  • thinking still reacts to streamed reasoning (reasoningLevel: "stream"), and can also start from active execution before reasoning deltas arrive.
  • Migrated heartbeat monitors use ordinary automation delivery. They no longer have a separate heartbeat typing policy or target.
  • Telegram also shows typing while a detached child that owes a visible completion is running, using the originating account, chat, and topic. Typing stops when the child ends or its requester is reset, deleted, or rebound to another route. Quiet children do not start typing.
  • agents.defaults.typingIntervalSeconds controls the refresh cadence for every agent, not the start time. Default: 6 seconds.
Was this useful?