-
1.6.0 - a crowd in one chat Stable
Homer released this
2026-09-26 13:38:52 +00:00 | 1 commits to main since this releaseLLeMbas 1.6.0 - a crowd in one chat
A chat can have more than one model in it. Switch it on under Admin -> Agents and
each chat's settings panel offers the others. The chat's own model answers first,
then each of the others in turn; then the order runs BACKWARDS, each one asked
whether it disagrees with anything said; and it ends back at the first model,
which either writes the final answer or sends them round again.Every contribution is its own bubble, with its own avatar, its own metrics and a
chip saying which speaker it is and which pass it belongs to.WHAT IT COSTS, SAID WHERE YOU TURN IT ON
One turn is models x rounds x 2 - 1 replies, so four models over two rounds is
fifteen. On a single local endpoint every change of speaker also loads a different
model. Your own warning is in the defaults - larger crowds of smaller models, and
sometimes of bigger ones, start going round in circles - so the round limit is
two, and it is a limit ordinary work will reach rather than a runaway backstop.THE DETAILS THAT MAKE IT WORK RATHER THAN LOOK LIKE IT WORKS
Each model sees the others' answers quoted and attributed, never as its own words.
Sent verbatim they would read as something it had said itself, so it would defend
sentences it never wrote and could not disagree with them - which is the entire
point of the way back. It is also what keeps the conversation alternating, which
several chat templates require.A member you can no longer reach is skipped and said so, not silently dropped. A
member whose endpoint fails is skipped, and two failures in a row end the round -
the commonest failure in a crowd is a small model's context window overflowing on
a transcript several models have written into, and ending there would kill every
crowd at whichever member is smallest.Stop ends the round, not just the model writing at the time. A message typed
during a round waits for the round instead of interleaving with it. One reply
counts as one reply against your concurrency limit. The unread dot and the
notification fire once for the turn rather than once per speaker.On the way back a model gets no tools: the question is whether it disagrees with
what is already in front of it. A member cannot ask you a question, cannot
delegate, and cannot rewrite its own personality - it is answering a
machine-composed instruction with several models' words quoted into it, which is
not a position to do any of those from.Every sentence a crowd sends is a fragment you can edit under Admin -> Prompts.
ALSO FIXED
A schedule that named its own model was ignored. The model was written on the
reply and never sent, so the bubble showed the model you chose while the answer
came from the chat's model. The same fix is what makes a crowd possible: the reply
itself now says which model is answering, rather than the conversation deciding
for all of them. Regenerating somebody's turn in a crowd keeps that model instead
of silently switching to the chat's.Downloads