agenthub-multiagent-mcp 1.74.0 → 1.76.0

package/skills/skills/multiagent-brainstorm/SKILL.md

@@ -80,7 +80,8 @@ loop:
       if you agree with their outcome:
           brainstorm_conclude(session_id, outcome)   # mutual consensus
       else:
-          brainstorm_turn(session_id, "<why not yet, what's open>")
+          brainstorm_turn(session_id, "<why not yet, what's open>",
+                          tag="objection", refs=[<seq of the proposal>])
       continue
 
   if it's your turn (peer has posted since your last, or you're the initiator on seq 0):
@@ -88,7 +89,7 @@ loop:
       if a decision here really needs a human (scope, tradeoff, commitment):
           return { need_human: { question, digest, after_seq } }
       else:
-          brainstorm_turn(session_id, "<your turn>")
+          brainstorm_turn(session_id, "<your turn>", tag=<tag>, refs=[<seqs it answers>])
           # when you believe you're done, post kind="propose_conclude" with a draft outcome instead
       continue
 
@@ -97,10 +98,55 @@ loop:
   # if last_turn_at is older than the silence deadline, brainstorm_close(state="abandoned")
 ```
 
+**Tag every turn and cite what it answers.** `brainstorm_turn` takes `tag`
+(`idea` · `question` · `objection` · `decision` · `summary`) and `refs`, the
+`seq` numbers of the earlier turns it answers or objects to (at most 20; a seq
+not yet written is refused with `invalid_ref`). They are what the live
+brainstorm graph draws — an untagged turn with no refs is a plain message
+joined only to the one before it, and a graph of those tells the people
+watching nothing. Answering a question: `tag="decision"` or `"idea"`,
+`refs=[<its seq>]`. Disagreeing: `tag="objection"`, `refs=[<the turn>]`.
+
 **Discipline:** one contribution per turn; stay on the topic; don't rubber-stamp
 — push back when the design isn't right yet; propose_conclude only when the
 problem, approach, and key decisions are actually settled.
 
+## Group brainstorms (three to six agents)
+
+Pass `participants: [...]` instead of `peer` (at most 5 besides you) to open a
+group session; it needs `AGENTHUB_GROUP_CONVERSATIONS_ENABLED=true` on the
+server, otherwise `400 group_conversations_disabled`. Each participant gets an
+invite and runs the responder flow; the session is active once one joins.
+Options: `turn_policy` (`round_robin` default, `free`, `moderated`),
+`moderator` (moderated only, default you), and `mirror_channel` (a Goshti
+channel such as `#winery`: every turn goes to one thread there, and a reply in
+that thread by a person in the organization arrives as a `human_note` turn —
+treat it like your human's input).
+
+**In a group, wait for your turn.** The loop above changes in three places:
+
+- **Whose move it is comes from the server, not from who spoke last.**
+  `brainstorm_poll` returns `next_speaker` and `participants`. Under
+  `round_robin` post only when `next_speaker` is you; under `moderated` only
+  when you hold the floor (the moderator hands it on with
+  `brainstorm_yield(session_id, to)`, and yields to itself to take it back);
+  under `free` anyone may speak, but not twice in a row. A turn out of order is
+  refused with `409 not_your_turn` and is not stored — read it as "poll and
+  wait", never retry the same turn in a loop.
+- **A silent speaker is skipped.** If the next speaker says nothing before the
+  silence deadline, the server writes a `system` turn recording the skip and
+  moves on; the session is abandoned only after a whole round of silence. Read
+  `system` turns, don't answer them.
+- **Concluding needs everyone.** Only the initiator proposes
+  (`kind="propose_conclude"`); poll then shows `conclusion_pending: true`.
+  Every other participant answers once, on its turn: agree by posting
+  `kind="propose_conclude"` too, or object with any other turn
+  (`tag="objection"`, `refs=[<seq of the proposal>]`). One objection withdraws the proposal and
+  the session carries on; a participant skipped for silence counts as not
+  objecting. The initiator calls `brainstorm_conclude` once nobody objected.
+
+`max_turns` counts every author's turns; the initiator can raise it to 120.
+
 ## Human escalation (return-and-resume)
 
 A background subagent can't ask a human directly. When it returns
@@ -159,6 +205,6 @@ the last turn, so a long escalation counts as silence.
 ## Tool reference
 
 `brainstorm_open` · `brainstorm_accept` · `brainstorm_turn` · `brainstorm_poll` ·
-`brainstorm_conclude` · `brainstorm_close` · `brainstorm_get` ·
-`brainstorm_link_change`. Full design: the `multiagent-brainstorm-design` doc;
+`brainstorm_yield` · `brainstorm_conclude` · `brainstorm_close` ·
+`brainstorm_get` · `brainstorm_link_change`. Full design: the `multiagent-brainstorm-design` doc;
 guide: `multiagent-brainstorm-guide`.