Thread participants let you keep specialist agents attached to a conversation without changing the main thread owner.
Use them when you want the main agent to keep driving the thread, but you also want another agent to watch for a specific class of issue and jump in when it matters.
Hidden Delegation vs Participants
Use !agent ... for hidden delegation.
Examples:
!weles verify whether this claim is correct!rarog prepare a softer operator-facing summary
Hidden delegation does not register a visible participant on the thread. It stays on the internal path.
Use @agent ... for visible resident participants.
Examples:
@weles verify claims before answering@rarog jump in when implementation risks or regressions appear
Participant registration keeps the main thread owner unchanged. Repeating @agent ... updates the assignment instead of creating duplicates.
These commands all deactivate the participant without changing thread ownership:
@weles stop@weles leave@weles done@weles return
The participant stays in thread history as inactive state, but it stops observing new thread activity.
Active participants observe visible thread activity and decide whether they have something useful to add.
When a participant wants to contribute, the daemon places that contribution into the queued suggestion flow instead of immediately replacing the main agent's response.
Each suggestion includes:
- the participant agent name
- the suggested message text
- whether the suggestion is marked for force-send
- whether the suggestion is queued or failed
Failed suggestions stay visible so you can retry them.
Some participant suggestions are marked as force-send.
Force-send means:
- the suggestion is urgent enough to interrupt an active stream
- sending it will stop the current stream first
- the participant message is then posted as a visible assistant message authored by that participant
In the TUI and desktop UI, force-send suggestions are clearly marked before you send them.
Participants do not see everything.
The participant observer prompt excludes:
- hidden internal delegation content
- participant suggestion queue contents
- system-only coordination that should not feed back into participant observation
This prevents suggestion loops and keeps participant reasoning tied to the visible thread.
Current UI behavior:
- the TUI shows queued participant suggestions in the queued message flow
- the desktop chat panel shows active participants in the thread header
- the desktop chat panel shows queued participant suggestions inline with send and dismiss actions
- participant-authored assistant messages are labeled with the agent name