A crowd in one chat

The chat's own model answers, then each other member in order, then the order runs
backwards asking each whether it disagrees, ending at the main model, which either
closes or sends them round again. Design and reasoning: LLeMbas.wiki/Crowd-chats.

THE SPEAKER SEAM, WHICH IS ALSO A BUG FIX

`chat_service.speaker_for` makes the *message* name the answering model and the
chat only the default. That closes a live half-wired feature -- `wake_chat` takes a
model override and `schedule/runner` passes one, and it reached the row and never
the request, so a schedule naming another model got the chat's model wearing the
other one's name.

The seam is wider than `build_request`: `{{model_name}}`, the authored prompt's
model layer, `vision` (where a wrong answer makes the endpoint reject the whole
request), the effort vocabulary (which raises inside the model's own chat template,
and whose refusal narrows every Model row sharing the id), `resolve_tools`,
`context_length` -> `_too_big`, and `ToolContext.model_id`. `resolve_endpoint` may
now only write back `chat.connection_id` when the speaker *is* the chat's model.

WHY N CHAINED REPLIES

`Generation` is one reply's state and `_follow` streams per message, so one
generation cannot stream into nine bubbles and `ensure` would not know which of the
nine it was after a restart. A subagent per speaker cannot work either: its answer
comes back as a tool result and tool results are never replayed, so speaker 3 could
not see speaker 2 -- which is the whole point. Chained, exactly one incomplete row
exists at a time, and `tests/test_crowd_chain.py` asserts that at every
observation.

The round lives on `Message.crowd_json`, not on the chat: the row is the authority,
and chat-level state would describe turns a rewind or a restart had removed.
`crowd.next_turn` is pure, so all eight refusals are tested with no endpoint.

THREE RULES, EACH A BUG WRITTEN THE OTHER WAY ROUND

- `if not _advance_crowd(g): _drain(g)` -- advancing must *suppress* draining, or a
  queued human turn puts a second incomplete row beside the next speaker's.
- `_advance_crowd` refuses unless the finishing row is the newest, or regenerating
  member 2 creates a second member 3 and two chains race down one turn.
- an error skips one speaker and two in a row end the round: the usual failure is a
  small member's window overflowing, and `_drain`'s stop-on-error would kill every
  crowd at whichever member is smallest.

Each other speaker's turn is relabelled as attributed user content, which is both
how a model can disagree with words it did not write and how the history keeps
alternating. The per-speaker instruction is payload-only -- as a row it could be
dropped from the request by a `created_at` tie, and every later speaker would answer
it. Compaction, titling and the notification are gated to once per turn; `_inject`
is off during a round; the way back gets no tools and a member is treated as
unattended.

Membership stores the model as text with no foreign key: "Test & refresh" deletes
and recreates Model rows, and a cascade would empty the crowd out of every chat.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

src/lembas/__init__.py

@@ -1,3 +1,3 @@
 """LLeMbas - a Middle-earth themed web UI for OpenAI-compatible LLM endpoints."""
 
-__version__ = "1.5.0"
+__version__ = "1.6.0"

src/lembas/api/admin_agents.py

@@ -66,6 +66,10 @@ async def agents_page(request: Request, db: Db, user: AdminUser, saved: bool = F
             # reply is allowed to set going on its own, and a nav entry for one
             # card would be worse than the near-miss.
             "subagents": settings_store.subagents(db),
+            # And a third group on the same page, for the same reason: a crowd is
+            # not an agent-chat feature either, but this is where somebody comes to
+            # find out what one turn is allowed to set going.
+            "crowd": settings_store.crowd(db),
             "saved": saved,
         },
     )
@@ -111,6 +115,35 @@ async def save_subagents(
     return RedirectResponse("/admin/agents?saved=1", status_code=status.HTTP_303_SEE_OTHER)
 
 
+@router.post("/crowd")
+async def save_crowd(
+    db: Db,
+    user: AdminUser,
+    enabled: bool = Form(False),
+    max_models: int = Form(4),
+    max_rounds: int = Form(2),
+    wall_seconds: int = Form(900),
+    collapse_agreement: bool = Form(False),
+) -> Response:
+    """Its own route, for the reason `save_subagents` gives above."""
+    settings_store.update(
+        db,
+        {
+            "enabled": enabled,
+            # Clamped here as well as on read. Every floor is one: a zero would be
+            # the feature switched off wearing the switch's clothes, and that is a
+            # thing to answer in one place.
+            "max_models": min(max(max_models, 1), 8),
+            "max_rounds": min(max(max_rounds, 1), 5),
+            "wall_seconds": min(max(wall_seconds, 60), 7200),
+            "collapse_agreement": collapse_agreement,
+        },
+        key=settings_store.CROWD,
+    )
+    log.info("crowd %s by %s", "enabled" if enabled else "disabled", user.email)
+    return RedirectResponse("/admin/agents?saved=1", status_code=status.HTTP_303_SEE_OTHER)
+
+
 @router.post("")
 async def save_agents(
     db: Db,

src/lembas/api/chats.py

@@ -1453,6 +1453,32 @@ def _queue_frames(
                 + "</div>"
             )
 
+    # The next speaker of a crowd round, on the same frame and by the same
+    # mechanism -- an incomplete assistant bubble carries `sse-connect`, so htmx
+    # opens the next stream itself and there is no new streaming machinery here at
+    # all.
+    #
+    # Its own branch and not the one above, deliberately. That one also re-renders
+    # "the last user turn at or before this bubble" to take Send now and Discard
+    # off it, and a crowd has no queued user turn: the swap would either re-render
+    # a node that was already correct or target one that is not in the document,
+    # where htmx silently does nothing. A branch that sometimes does nothing is a
+    # branch nobody can reason about.
+    if getattr(generation, "crowded", False):
+        following = list(
+            db.scalars(
+                select(Message)
+                .where(Message.chat_id == chat.id, Message.complete.is_(False))
+                .order_by(Message.created_at, Message.id)
+            )
+        )
+        for speaker_row in following:
+            out_of_band.append(
+                '<div hx-swap-oob="beforeend:#thread">'
+                + _render_bubble(db, chat, owner, speaker_row)
+                + "</div>"
+            )
+
     return "".join(moved), "".join(out_of_band)
 
 
@@ -2089,6 +2115,42 @@ async def update_chat(request: Request, db: Db, user: RequiredUser, chat_id: str
             else []
         )
 
+    if "crowd_model_ids" in form:
+        # The same shape as the bases above: one field always sent, so clearing
+        # every box clears the crowd. Checked against what this person can reach
+        # rather than against what exists, or the picker is advisory and a crafted
+        # request walks past it -- the reasoning the model branch carries.
+        from lembas.db.models import CrowdMember
+
+        settings = settings_store.crowd(db)
+        reachable = {
+            model.model_id for model in chat_service.available_models(db, user)
+        }
+        wanted: list[str] = []
+        for value in form.getlist("crowd_model_ids"):
+            value = str(value).strip()
+            # Never the chat's own model: it would answer twice in a row, which is
+            # nobody's idea of a second opinion.
+            if value and value in reachable and value != chat.model_id and value not in wanted:
+                wanted.append(value)
+        wanted = wanted[: int(settings["max_models"])]
+
+        chat.crowd = [
+            CrowdMember(
+                model_id=model_id,
+                connection_id=next(
+                    (
+                        model.connection_id
+                        for model in chat_service.available_models(db, user)
+                        if model.model_id == model_id
+                    ),
+                    None,
+                ),
+                position=index,
+            )
+            for index, model_id in enumerate(wanted)
+        ]
+
     submitted_params = {name: form[name] for name in _PARAM_RANGES if name in form}
     if submitted_params:
         if not allowed.get("chat.params"):
@@ -2228,7 +2290,14 @@ async def regenerate(
     message.content = ""
     message.error = ""
     message.complete = False
-    message.model_id = chat.model_id
+    # Whose reply this was stays whose reply it is, unless the chat's model has
+    # been changed since -- in which case regenerating is how somebody asks for
+    # the new one. Before 1.6.0 this always reset to the chat's model, which was
+    # merely a wrong label; now that the row *is* the model that answers, it would
+    # silently regenerate somebody else's turn as the chat's model.
+    if not (message.model_id or "").strip():
+        message.model_id = chat.model_id
+        message.connection_id = chat.connection_id
     _note_rewind(chat)
     db.commit()
     # restart, not ensure: this is the one caller that reuses a Message row, and

src/lembas/api/pages.py

@@ -82,6 +82,7 @@ def _chat_context(db: DBSession, user: User, chat: Chat | None) -> dict:
             else []
         ),
         "attached_base_ids": [base.id for base in chat.knowledge_bases] if chat else [],
+        **_crowd_context(db, user, chat, models),
         # What *this* model takes, not the three every model used to be assumed
         # to take. The vocabulary is per model -- gpt-oss has no `xhigh` and
         # Bonsai has no `high`, and sending the wrong one does not degrade, it
@@ -195,6 +196,42 @@ def _scope_context(db: DBSession, user: User, chat: Chat | None) -> dict:
     }
 
 
+def _crowd_context(db: DBSession, user: User, chat: Chat | None, models: list) -> dict:
+    """Who else could answer in this chat, and what that would cost.
+
+    Empty — and the panel then shows nothing rather than an empty control — when
+    the feature is off, when there is nobody else to add, or on the new-chat
+    screen, where there is no chat to attach anybody to yet.
+
+    The cost is spelled out because it is the thing somebody will not have thought
+    about: a turn is `speakers x rounds x 2 - 1` replies, and on one local endpoint
+    each change of speaker is also a model load.
+    """
+    from lembas.services import crowd as crowd_service
+
+    settings = settings_store.crowd(db)
+    if chat is None or not settings["enabled"]:
+        return {"crowd_available": [], "crowd_member_ids": [], "crowd_skipped": []}
+
+    others = [model for model in models if model.model_id != chat.model_id]
+    members = [
+        row.model_id
+        for row in sorted(chat.crowd, key=lambda row: (row.position, row.model_id))
+    ]
+    reachable = {model.model_id for model in others}
+    speakers = 1 + len([model_id for model_id in members if model_id in reachable])
+    rounds = int(settings["max_rounds"])
+    return {
+        "crowd_available": others,
+        "crowd_member_ids": [model_id for model_id in members if model_id in reachable],
+        "crowd_skipped": crowd_service.unreachable_members(db, chat, user),
+        # One round is out and back: everybody answers, everybody but the last is
+        # asked whether they disagree, and the main model closes.
+        "crowd_replies": max(1, speakers * 2 - 1),
+        "crowd_rounds": rounds,
+    }
+
+
 # What a gate is called in the menu. A gate covers several tools, so no single
 # tool's label is the right name for it.
 _GATE_LABELS = {

src/lembas/db/models/__init__.py

@@ -31,6 +31,7 @@ from lembas.db.models.chat import (
     ROLE_TOOL,
     ROLE_USER,
     Chat,
+    CrowdMember,
     Folder,
     Message,
 )
@@ -163,6 +164,7 @@ __all__ = [
     "Report",
     "Schedule",
     "Chat",
+    "CrowdMember",
     "Job",
     "Connection",
     "CustomTool",

src/lembas/db/models/chat.py

@@ -5,7 +5,15 @@ from __future__ import annotations
 from datetime import datetime
 from typing import TYPE_CHECKING, Any
 
-from sqlalchemy import Boolean, DateTime, ForeignKey, Integer, String, Text
+from sqlalchemy import (
+    Boolean,
+    DateTime,
+    ForeignKey,
+    Integer,
+    String,
+    Text,
+    UniqueConstraint,
+)
 from sqlalchemy.orm import Mapped, mapped_column, relationship
 
 from lembas.db.base import Base, Timestamps, UUIDPrimaryKey
@@ -309,10 +317,60 @@ class Chat(UUIDPrimaryKey, Timestamps, Base):
         "KnowledgeBase", secondary="chat_knowledge_bases"
     )
 
+    # The other models answering in this chat, in the order they speak. Empty is
+    # every chat that has ever existed: one model, answering on its own.
+    crowd: Mapped[list[CrowdMember]] = relationship(
+        back_populates="chat",
+        cascade="all, delete-orphan",
+        order_by="CrowdMember.position",
+    )
+
     def __repr__(self) -> str:
         return f"<Chat {self.title!r}>"
 
 
+class CrowdMember(UUIDPrimaryKey, Timestamps, Base):
+    """One extra model answering in a chat, and where it sits in the order.
+
+    A row rather than an association table because it carries an order and has
+    nothing to associate *to*:
+
+    🚨 **the model is stored as text, with no foreign key to `models`.** "Test &
+    refresh" on the connection screen deletes every model the endpoint has
+    stopped listing and creates it again when it comes back, so a foreign key
+    with `ON DELETE CASCADE` -- which is what copying `chat_knowledge_bases`
+    would have given -- means one refresh taken while an endpoint happened to be
+    loading something else silently empties the crowd out of every chat, with no
+    row left to explain it. This is the reasoning `Chat.model_id`,
+    `ssh_profile_id` and `compacted_through_id` all carry, and the same trap that
+    lost the image reviewer its model in 1.4.x.
+
+    A member that no longer resolves is therefore skipped at send time and shown
+    struck through, rather than being deleted by something nobody asked.
+
+    `connection_id` is nullable and usually empty, meaning "resolve it from the
+    id"; it matters only where two connections offer the same model, since their
+    capabilities and effort lists are separate rows.
+    """
+
+    __tablename__ = "chat_crowd"
+    __table_args__ = (UniqueConstraint("chat_id", "model_id"),)
+
+    chat_id: Mapped[str] = mapped_column(
+        String(32), ForeignKey("chats.id", ondelete="CASCADE"), nullable=False, index=True
+    )
+    model_id: Mapped[str] = mapped_column(String(300), nullable=False)
+    connection_id: Mapped[str | None] = mapped_column(String(32), nullable=True)
+    # Where this member speaks. The chat's own model is always first and is not a
+    # row here, so these start at 1 in spirit and are only ever compared.
+    position: Mapped[int] = mapped_column(Integer, default=0, nullable=False)
+
+    chat: Mapped[Chat] = relationship(back_populates="crowd")
+
+    def __repr__(self) -> str:
+        return f"<CrowdMember {self.model_id} at {self.position}>"
+
+
 class Message(UUIDPrimaryKey, Timestamps, Base):
     __tablename__ = "messages"
 
@@ -340,14 +398,46 @@ class Message(UUIDPrimaryKey, Timestamps, Base):
     # Milliseconds spent producing the reasoning, for the "Thought for Xs" label.
     reasoning_ms: Mapped[int] = mapped_column(Integer, default=0, nullable=False)
 
+    # Which model wrote this, or is about to. Written on every assistant
+    # placeholder at creation and, from 1.6.0, **read back as the model that
+    # answers** -- `chat_service.speaker_for`. Before that it was a display
+    # snapshot only, and the two could disagree: `wake_chat` accepts a model
+    # override that reached this column and never reached the request, so a
+    # schedule naming another model got the chat's model wearing this label.
     model_id: Mapped[str] = mapped_column(String(300), default="")
 
+    # Which connection that model was reached through. Nullable and usually
+    # empty, meaning "resolve it from the model id as this application always
+    # has"; it matters only where the same id is offered by two connections,
+    # since `Model` is unique on the pair and their capabilities, context lengths
+    # and effort lists are separate rows.
+    #
+    # No foreign key, deliberately, and the same reasoning `Chat.model_id`
+    # carries: a transcript has to survive an administrator deleting a
+    # connection, and `migrations.py` compiles only the column type -- so a
+    # REFERENCES clause would exist on a fresh database and not on an upgraded
+    # one. Validated on read instead.
+    connection_id: Mapped[str | None] = mapped_column(String(32), nullable=True)
+
     # What the model did before answering: one entry per tool call, with its
     # arguments and results. Shown in the transcript so the sources behind an
     # answer stay visible, and deliberately NOT replayed as context on the next
     # turn -- see services/generation.py for why.
     tool_calls_json: Mapped[list[Any]] = mapped_column(JSONList, default=list)
 
+    # Where this message sits in a crowd round: the turn it belongs to, the
+    # round, the phase, and which speaker it is. NULL on every message that is
+    # not part of one, which is every message this application has ever written
+    # before 1.6.0.
+    #
+    # On the row and not on the chat, deliberately. "The row is the authority,
+    # not the registry" is the rule the reload story was won with, and round
+    # state on the chat reintroduces the split it was won against: a restart
+    # between speakers, or a rewind that deletes these rows, would leave
+    # chat-level state describing turns that no longer exist -- which is the
+    # problem `compacted_through_id` already documents.
+    crowd_json: Mapped[dict[str, Any] | None] = mapped_column(JSONDict, nullable=True)
+
     # Where each round's contribution ended, so `content`, `reasoning` and
     # `tool_calls_json` can be shown as the one sequence they actually were
     # rather than as three stacked zones. One entry per closed step, holding the

src/lembas/services/chat.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import logging
 import re
+from dataclasses import dataclass
 from datetime import UTC, datetime, timedelta
 from typing import Any
 
@@ -46,19 +47,71 @@ TITLE_MAX_TOKENS = 512
 TEMPORARY_LIFETIME = timedelta(hours=24)
 
 
-def resolve_endpoint(db: DBSession, chat: Chat) -> tuple[Endpoint, str]:
-    """Find the connection and model a chat should use.
+@dataclass(frozen=True)
+class Speaker:
+    """Which model is answering one reply, and through which connection.
+
+    The pair and not the id, because `Model` is unique on
+    `(connection_id, model_id)`: the same name can live behind two endpoints and
+    an id alone does not say which. `images/tool.py:_reviewer` already resolves a
+    model this way.
+
+    Frozen, and passed rather than re-derived, for the reason `Endpoint` is a
+    snapshot: a generation outlives the request that started it, and "who is
+    answering" must not be able to change underneath a reply that is already
+    streaming.
+    """
+
+    model_id: str
+    connection_id: str | None = None
+
+
+def speaker_for(db: DBSession, chat: Chat, message: Message | None = None) -> Speaker:
+    """Who is answering: the message being written into, or else the chat.
+
+    **The row names the model and the chat is only the default.** Until 1.6.0 the
+    answering model was `chat.model_id` and nothing else, while `Message.model_id`
+    was written on every placeholder and read only for display -- so the bubble's
+    avatar and the request could disagree, and did: `wake_chat` accepts a
+    `model_id` override and `schedule/runner` passes `schedule.model_id or
+    chat.model_id`, which reached the row and never reached the request. A
+    schedule naming another model got the chat's model wearing the other one's
+    name.
+
+    Reading it off the row is also what makes a reply survive a restart, because
+    `_follow` calls `ensure`, which starts a *new* generation against the same
+    row -- so anything the request depends on has to be durable, and the registry
+    is not. This is the rule the reload story was won with: the row is the
+    authority.
+    """
+    if message is not None and (message.model_id or "").strip():
+        return Speaker(message.model_id, getattr(message, "connection_id", None) or None)
+    return Speaker(chat.model_id, chat.connection_id)
+
+
+def resolve_endpoint(
+    db: DBSession, chat: Chat, speaker: Speaker | None = None
+) -> tuple[Endpoint, str]:
+    """Find the connection and model a reply should use.
 
     Chats store the model id as text rather than a foreign key so history
     survives an admin deleting a connection, which means the mapping back to a
     live connection has to be resolved at send time and can legitimately fail.
+
+    `speaker` defaults to the chat's own model, so every existing caller behaves
+    exactly as it did.
     """
-    if not chat.model_id:
+    speaker = speaker or speaker_for(db, chat)
+    if not speaker.model_id:
         raise LLMError("This chat has no model selected.")
+    # Whether resolving a fallback may be *written back* to the chat. It may only
+    # when the speaker is the chat's own model: a crowd member or a schedule's
+    # model finding its way to another connection must not repoint the chat.
+    speaks_for_chat = speaker.model_id == chat.model_id
 
     connection: Connection | None = None
-    if chat.connection_id:
-        connection = db.get(Connection, chat.connection_id)
+    if speaker.connection_id:
+        connection = db.get(Connection, speaker.connection_id)
 
     if connection is None or not connection.enabled:
         # The original connection is gone or disabled. Any enabled connection
@@ -67,7 +120,7 @@ def resolve_endpoint(db: DBSession, chat: Chat) -> tuple[Endpoint, str]:
             select(Model)
             .join(Connection)
             .where(
-                Model.model_id == chat.model_id,
+                Model.model_id == speaker.model_id,
                 Model.enabled.is_(True),
                 Connection.enabled.is_(True),
             )
@@ -76,13 +129,14 @@ def resolve_endpoint(db: DBSession, chat: Chat) -> tuple[Endpoint, str]:
         if model is None:
             raise LLMError(
                 f"No enabled connection currently offers the model "
-                f"'{chat.model_id}'. Pick another model for this chat."
+                f"'{speaker.model_id}'. Pick another model for this chat."
             )
         connection = model.connection
-        chat.connection_id = connection.id
-        db.commit()
+        if speaks_for_chat:
+            chat.connection_id = connection.id
+            db.commit()
 
-    return Endpoint.from_connection(connection), chat.model_id
+    return Endpoint.from_connection(connection), speaker.model_id
 
 
 def document_context(message: Message) -> str:
@@ -190,7 +244,9 @@ def folder_system_prompt(db: DBSession, chat: Chat) -> str:
     return ""
 
 
-def effective_system_prompt(db: DBSession, chat: Chat) -> str:
+def effective_system_prompt(
+    db: DBSession, chat: Chat, speaker: Speaker | None = None
+) -> str:
     """The system prompt a chat actually runs with.
 
     Four layers, most specific wins outright:
@@ -214,9 +270,9 @@ def effective_system_prompt(db: DBSession, chat: Chat) -> str:
     if inherited := folder_system_prompt(db, chat):
         return inherited
 
-    model = db.scalar(
-        select(Model).where(Model.model_id == chat.model_id).order_by(Model.position)
-    )
+    # The *answering* model's layer, which is not always the chat's: a crowd
+    # member speaking in somebody else's chat brings its own prompt with it.
+    model = model_row(db, speaker or Speaker(chat.model_id, chat.connection_id))
     if model is not None and (model.system_prompt or "").strip():
         return model.system_prompt.strip()
 
@@ -230,6 +286,7 @@ def build_messages(
     upto: Message | None = None,
     vision: bool = False,
     system_prompt: str | None = None,
+    speaker: Speaker | None = None,
 ) -> list[dict]:
     """Assemble the message list to send upstream.
 
@@ -302,24 +359,196 @@ def build_messages(
             continue
         payload.append(message_payload(message, vision=vision))
 
+    if speaker is not None:
+        payload = _as_one_speaker_sees_it(db, payload, history, speaker, upto=upto)
+
     return payload
 
 
-def model_for(db: DBSession, chat: Chat) -> Model | None:
-    """The Model row a chat is using, or None if it has gone.
+def _as_one_speaker_sees_it(
+    db: DBSession,
+    payload: list[dict[str, Any]],
+    history: list[Message],
+    speaker: Speaker,
+    *,
+    upto: Message | None = None,
+) -> list[dict[str, Any]]:
+    """Rewrite a crowd transcript from one speaker's point of view.
+
+    Two problems, one pass.
+
+    **Another speaker's reply must not arrive as this one's own prior turn.** Sent
+    verbatim, every assistant message in the payload reads as something *this*
+    model said -- so it defends sentences it never wrote, and cannot disagree with
+    them, which is the whole point of the backward pass. Each other speaker's turn
+    is therefore relabelled as user content behind a fragment-driven "«Label»
+    said:".
+
+    **Consecutive assistant turns break strict-alternation chat templates**, which
+    this project already knows: `task.compact_ack` exists so a compacted history
+    still alternates, and several templates reject one that does not. Relabelling
+    fixes that by construction, and the adjacent user turns it creates are merged.
+
+    ⚠ The relabelled entry is built here rather than by calling `message_payload`
+    with a swapped role. That function attaches image parts when the role is
+    `user` and the model has vision, so a swapped assistant turn carrying a
+    generated image would silently become a multimodal list -- and an endpoint
+    that rejects one rejects every later turn with it.
+    """
+    from lembas.services import prompts as prompts_service
+
+    # Nothing to do for the ordinary case: one model, and every assistant turn in
+    # the payload is its own.
+    others = {
+        message.model_id
+        for message in history
+        if message.role == ROLE_ASSISTANT
+        and (message.model_id or "")
+        and message.model_id != speaker.model_id
+    }
+    if not others:
+        return payload
+
+    labels = {
+        model_id: (row.label if (row := model_row(db, Speaker(model_id))) else model_id)
+        for model_id in others
+    }
+    template = prompts_service.resolve(db, "crowd.said") or "{{crowd_speaker}} answered:"
+
+    # The payload and the history line up only over the message rows: the system
+    # turn and a compaction pair come first and belong to nobody. Walking from the
+    # end is what pairs them without counting.
+    rows = [
+        message
+        for message in history
+        if not (upto is not None and message.id == upto.id)
+    ]
+    rewritten: list[dict[str, Any]] = []
+    for index, entry in enumerate(payload):
+        row = None
+        offset = index - (len(payload) - len(rows))
+        if 0 <= offset < len(rows):
+            row = rows[offset]
+        if (
+            row is not None
+            and entry.get("role") == ROLE_ASSISTANT
+            and (row.model_id or "") in others
+        ):
+            lead = template.replace("{{crowd_speaker}}", labels[row.model_id])
+            body = entry.get("content")
+            rewritten.append(
+                {"role": ROLE_USER, "content": f"{lead}\n\n{body if isinstance(body, str) else ''}"}
+            )
+            continue
+        rewritten.append(entry)
+
+    return _merge_user_turns(rewritten)
+
+
+def _with_crowd_instruction(
+    db: DBSession, payload: list[dict[str, Any]], turn, *, again: bool
+) -> list[dict[str, Any]]:
+    """Append what this speaker has been asked to do, as the closing user turn.
+
+    🚨 **Payload only. No row is written for it.** Writing the instruction into the
+    transcript the way `wake_chat` writes a background job's turn was the first
+    design and is wrong three times over. `build_messages` orders history by
+    `created_at` alone and `break`s at the placeholder, so on a shared microsecond
+    the placeholder sorts first and the instruction is dropped from the request
+    entirely -- the hazard `thread_tail` already carries an explicit tiebreak for.
+    It would double the rows in a turn, all of them bubbles somebody has to scroll
+    past. And every later speaker would read the previous speaker's instruction as
+    an ordinary user turn and answer that too.
+
+    The compaction summary is inserted the same way and for the same reason: a
+    turn in the payload with nothing behind it (`build_messages`).
+    """
+    from lembas.services import crowd as crowd_service
+    from lembas.services import prompts as prompts_service
+
+    if turn.phase == crowd_service.PHASE_OUT:
+        key = "crowd.turn"
+    elif turn.phase == crowd_service.PHASE_BACK:
+        key = "crowd.disagree"
+    else:
+        # Two fragments, not one with a clause in it: inviting a choice the model
+        # cannot express is worse than not offering it, and a model without the
+        # tools capability has no `crowd_again` to call.
+        key = "crowd.close" if again else "crowd.close_final"
+
+    text = (prompts_service.resolve(db, key) or "").strip()
+    if not text:
+        # Cleared on purpose is the administrator switching this wording off, and
+        # an empty user turn is not a thing to send.
+        return payload
+    return _merge_user_turns([*payload, {"role": ROLE_USER, "content": text}])
+
+
+def _merge_user_turns(payload: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """Fold adjacent user turns into one, so the history still alternates.
+
+    Only where both are plain strings: a turn carrying content parts is a
+    multimodal message and joining one to a string would destroy it.
+    """
+    merged: list[dict[str, Any]] = []
+    for entry in payload:
+        last = merged[-1] if merged else None
+        if (
+            last is not None
+            and last.get("role") == ROLE_USER
+            and entry.get("role") == ROLE_USER
+            and isinstance(last.get("content"), str)
+            and isinstance(entry.get("content"), str)
+        ):
+            merged[-1] = {
+                **last,
+                "content": f"{last['content']}\n\n{entry['content']}",
+            }
+            continue
+        merged.append(entry)
+    return merged
+
+
+def model_row(db: DBSession, speaker: Speaker) -> Model | None:
+    """The Model row a speaker names, or None if it has gone.
 
     Looked up by id rather than held as a foreign key, for the same reason
     resolve_endpoint does: chats store the model as text so history survives an
-    administrator deleting a connection.
+    administrator deleting a connection. The connection narrows it when one is
+    named, because two connections may offer the same id and their capabilities,
+    context length and effort lists are separate rows.
     """
+    if not speaker.model_id:
+        return None
+    if speaker.connection_id:
+        exact = db.scalar(
+            select(Model).where(
+                Model.model_id == speaker.model_id,
+                Model.connection_id == speaker.connection_id,
+            )
+        )
+        if exact is not None:
+            return exact
     return db.scalar(
-        select(Model).where(Model.model_id == chat.model_id).order_by(Model.position)
+        select(Model).where(Model.model_id == speaker.model_id).order_by(Model.position)
     )
 
 
-def model_supports(db: DBSession, chat: Chat, capability: str) -> bool:
-    """Whether the chat's current model is marked as having a capability."""
-    model = model_for(db, chat)
+def model_for(db: DBSession, chat: Chat) -> Model | None:
+    """The Model row a chat is using. The display answer; see `model_row`."""
+    return model_row(db, Speaker(chat.model_id, chat.connection_id))
+
+
+def model_supports(
+    db: DBSession, chat: Chat, capability: str, speaker: Speaker | None = None
+) -> bool:
+    """Whether the answering model is marked as having a capability.
+
+    ⚠ Worth getting right per speaker rather than per chat: `vision` decides
+    whether image parts go into the body, and an endpoint sent an image by a
+    model that cannot take one rejects **the whole request**, not the image.
+    """
+    model = model_row(db, speaker) if speaker is not None else model_for(db, chat)
     return bool(model and (model.capabilities_json or {}).get(capability))
 
 
@@ -331,12 +560,21 @@ def build_request(
     tools: list[dict[str, Any]] | None = None,
     user=None,
     force_tool: str = "",
+    speaker: Speaker | None = None,
+    crowd_turn=None,
+    crowd_again: bool = False,
 ) -> dict[str, Any]:
     """The whole request body, tools and harness included.
 
     Composed here rather than in the generation loop so that "what gets sent"
     has one answer, and so the harness cannot be forgotten by a future caller
     that offers tools.
+
+    `speaker` is who is answering; it defaults to the chat's own model, so a
+    caller that does not care behaves exactly as it did. Everything that differs
+    per model is resolved from it and not from the chat: the model name sent, the
+    vision decision, the authored prompt's model layer, `{{model_name}}`, the
+    personality, and the reasoning-effort vocabulary.
     """
     from lembas.services import harness as harness_service
     from lembas.services import prompts as prompts_service
@@ -346,10 +584,15 @@ def build_request(
         for key, value in (chat.params_json or {}).items()
         if key in FORWARDED_PARAMS and value not in (None, "")
     }
+    speaker = speaker or speaker_for(db, chat, upto)
+    if crowd_turn is None and upto is not None:
+        from lembas.services import crowd as crowd_service
+
+        crowd_turn = crowd_service.state_of(upto)
     # Images are only sent to a model an administrator has marked as having
     # vision. Sending them to one that has not is not a graceful degradation:
     # most endpoints reject the whole request.
-    vision = model_supports(db, chat, "vision")
+    vision = model_supports(db, chat, "vision", speaker=speaker)
 
     if user is None:
         from lembas.db.models import User
@@ -360,18 +603,23 @@ def build_request(
     # behaviour. See services/harness.py for why these are joined rather than
     # being two competing layers.
     system = harness_service.join(
-        harness_service.compose(db, user, tools, chat),
-        effective_system_prompt(db, chat),
+        harness_service.compose(db, user, tools, chat, speaker=speaker),
+        effective_system_prompt(db, chat, speaker),
         lead=prompts_service.render(db, "seam.authored_lead", {}),
     )
 
     body: dict[str, Any] = {
-        "model": chat.model_id,
+        "model": speaker.model_id,
         "messages": build_messages(
-            db, chat, upto=upto, vision=vision, system_prompt=system
+            db, chat, upto=upto, vision=vision, system_prompt=system, speaker=speaker
         ),
         **params,
     }
+    if crowd_turn is not None:
+        body["messages"] = _with_crowd_instruction(
+            db, body["messages"], crowd_turn, again=crowd_again
+        )
+
     if tools:
         body["tools"] = tools
         # Making the model call one particular tool, for `/image` -- the whole
@@ -388,14 +636,22 @@ def build_request(
         ):
             body["tool_choice"] = {"type": "function", "function": {"name": force_tool}}
 
-    # The model's own vocabulary, looked up here rather than passed in: every
-    # caller of `build_request` would otherwise have to remember, which is the
-    # trap `audio_service.template_flags` fell into.
-    chat_model = model_for(db, chat)
+    # The *answering* model's own vocabulary, looked up here rather than passed
+    # in: every caller of `build_request` would otherwise have to remember, which
+    # is the trap `audio_service.template_flags` fell into.
+    #
+    # ⚠ Per speaker and not per chat, and this one is not cosmetic: the
+    # vocabularies genuinely differ -- gpt-oss takes low/medium/high, a Bonsai
+    # takes low/medium/xhigh and *raises inside its chat template* on high -- so
+    # a chat's effort handed to another model fails the whole reply rather than
+    # being ignored. `_learn_refused_effort` then narrows every Model row sharing
+    # that id, so getting this wrong would also corrupt other models' lists as a
+    # side effect.
+    speaking_model = model_row(db, speaker)
     apply_effort(
         body,
         (chat.params_json or {}).get("reasoning_effort"),
-        efforts_for(chat_model) if chat_model is not None else None,
+        efforts_for(speaking_model) if speaking_model is not None else None,
     )
     return body
 

src/lembas/services/crowd.py

@@ -0,0 +1,383 @@
+"""Several models answering one turn, in order, then again in reverse.
+
+The shape the owner asked for: the chat's own model answers, then each other
+member in order; then the order runs **backwards**, each member asked whether it
+disagrees with anything; and it ends at the main model, which decides whether to
+go round again or stop.
+
+## Why N chained replies and not one clever one
+
+One `Generation` per speaker, one `Message` per speaker, chained where `_drain`
+already chains a queued turn. That is not the cheapest shape, it is the only one
+in which every existing invariant keeps holding for the reason it already holds:
+
+* `Generation` is **one reply's** state and `_follow` streams **per message**,
+  keyed on `generation.message_id`. One generation cannot stream into nine
+  bubbles without a second streaming protocol, and `ensure(chat_id, message_id)`
+  would have no answer to "which of the nine am I" after a restart.
+* Exactly one incomplete assistant row exists at any moment, so
+  `_reply_in_flight` needs no teaching and the composer queues for the whole
+  round.
+* Each speaker gets its own `steps_json`, `usage_json` and `model_id`, so the
+  avatar, the metrics chip and the regenerate button are per speaker with no new
+  rendering.
+
+A subagent per speaker was rejected outright: a helper is handed a *serialisation*
+of the conversation, its answer comes back as a tool result, and tool results are
+never replayed -- so speaker 3 could not see speaker 2, which is the entire point
+of a crowd. That feature already exists and is called `ask_friend`.
+
+## Where the round lives
+
+On the **message row**, in `Message.crowd_json`, and not on the chat. "The row is
+the authority, not the registry" is the rule the reload story was won with, and
+round state on the chat reintroduces exactly the split it was won against: a
+restart between speakers, or a rewind that deletes the rows, would leave
+chat-level state describing turns that no longer exist -- which is the problem
+`Chat.compacted_through_id` already documents.
+
+`Message.parent_id` is **not** used for grouping. It is reserved for conversation
+branching and says so in its own comment.
+
+## The scheduler is a pure function
+
+`next_turn` takes numbers and returns numbers. Every refusal -- out of rounds, out
+of time, nobody to ask, not the newest message -- is therefore testable without an
+endpoint, which matters because the refusals are the interesting half.
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, replace
+from datetime import UTC, datetime
+from typing import Any
+
+from sqlalchemy import select
+from sqlalchemy.orm import Session as DBSession
+
+from lembas.db.models import Chat, Message
+
+log = logging.getLogger(__name__)
+
+# The forward pass: everybody answers in order.
+PHASE_OUT = "out"
+# The way back: each member is asked whether it disagrees, in reverse order,
+# stopping one short of the main model.
+PHASE_BACK = "back"
+# The main model's last word, where it decides whether to go round again.
+PHASE_CLOSE = "close"
+
+PHASES = (PHASE_OUT, PHASE_BACK, PHASE_CLOSE)
+
+# Why a round ended, when it ended for a reason rather than by finishing.
+STOPPED_ROUNDS = "rounds"
+STOPPED_TIME = "time"
+STOPPED_ERRORS = "errors"
+
+# How many speaker errors in a row end the round. One is skipped: the commonest
+# failure in a crowd is not a dead endpoint but a small member's context window
+# overflowing on a transcript several models have been writing into, and killing
+# the round at whichever member is smallest is the wrong answer. Two in a row is
+# an endpoint that has actually gone, which is what `_drain`'s refusal protects
+# against and is worth keeping.
+MAX_CONSECUTIVE_ERRORS = 2
+
+
+@dataclass(frozen=True)
+class Turn:
+    """Where one crowd round has got to, as it is stored on a message."""
+
+    turn: str
+    round: int
+    phase: str
+    index: int
+    of: int
+    started_at: str
+    errors: int = 0
+    stopped: str = ""
+
+    def as_json(self) -> dict[str, Any]:
+        return {
+            "turn": self.turn,
+            "round": self.round,
+            "phase": self.phase,
+            "index": self.index,
+            "of": self.of,
+            "started_at": self.started_at,
+            "errors": self.errors,
+            "stopped": self.stopped,
+        }
+
+    @property
+    def is_main(self) -> bool:
+        return self.index == 0
+
+
+def state_of(message: Message | None) -> Turn | None:
+    """The round state on a message, or None if it is not part of one."""
+    raw = getattr(message, "crowd_json", None) or None
+    if not raw or not isinstance(raw, dict):
+        return None
+    try:
+        return Turn(
+            turn=str(raw.get("turn") or ""),
+            round=int(raw.get("round") or 1),
+            phase=str(raw.get("phase") or PHASE_OUT),
+            index=int(raw.get("index") or 0),
+            of=int(raw.get("of") or 1),
+            started_at=str(raw.get("started_at") or ""),
+            errors=int(raw.get("errors") or 0),
+            stopped=str(raw.get("stopped") or ""),
+        )
+    except (TypeError, ValueError):  # pragma: no cover - a hand-edited row
+        return None
+
+
+def now_stamp() -> str:
+    return datetime.now(UTC).isoformat()
+
+
+def elapsed(started_at: str) -> float:
+    """Seconds since a round began, or 0.0 if the stamp is unreadable.
+
+    Unreadable reads as "no time has passed" rather than as "out of time": a
+    round abandoned because of a bad timestamp would be a feature failing for a
+    reason nobody could see.
+    """
+    try:
+        began = datetime.fromisoformat(started_at)
+    except (TypeError, ValueError):
+        return 0.0
+    if began.tzinfo is None:
+        began = began.replace(tzinfo=UTC)
+    return max(0.0, (datetime.now(UTC) - began).total_seconds())
+
+
+def next_turn(
+    *,
+    speakers: int,
+    state: Turn | None,
+    turn_id: str,
+    again: bool = False,
+    errored: bool = False,
+    max_rounds: int = 2,
+    wall_seconds: int = 900,
+) -> Turn | None:
+    """Who speaks next, or None when the round is over.
+
+    Pure: numbers in, numbers out, no session and no clock beyond the stamp it is
+    handed. `speakers` counts the main model as one of them.
+
+    `state=None` means the reply that has just finished was the ordinary first
+    one, started by the composer as it always is -- so this is where a round
+    begins rather than continues.
+    """
+    if speakers < 2:
+        return None
+
+    if state is None:
+        return Turn(
+            turn=turn_id,
+            round=1,
+            phase=PHASE_OUT,
+            index=1,
+            of=speakers,
+            started_at=now_stamp(),
+        )
+
+    # Errors are counted consecutively, so one member timing out is skipped and
+    # an endpoint that has gone ends the round.
+    errors = state.errors + 1 if errored else 0
+    if errors >= MAX_CONSECUTIVE_ERRORS:
+        return replace(state, stopped=STOPPED_ERRORS)
+
+    if wall_seconds and elapsed(state.started_at) >= wall_seconds:
+        return replace(state, errors=errors, stopped=STOPPED_TIME)
+
+    carry = {
+        "turn": state.turn,
+        "of": speakers,
+        "started_at": state.started_at,
+        "errors": errors,
+    }
+
+    if state.phase == PHASE_OUT:
+        if state.index + 1 <= speakers - 1:
+            return Turn(round=state.round, phase=PHASE_OUT, index=state.index + 1, **carry)
+        # The forward pass is done. The way back starts one short of the speaker
+        # that has just finished -- asking it whether it disagrees with itself is
+        # a round spent on nothing.
+        if speakers - 2 >= 1:
+            return Turn(round=state.round, phase=PHASE_BACK, index=speakers - 2, **carry)
+        return Turn(round=state.round, phase=PHASE_CLOSE, index=0, **carry)
+
+    if state.phase == PHASE_BACK:
+        if state.index - 1 >= 1:
+            return Turn(round=state.round, phase=PHASE_BACK, index=state.index - 1, **carry)
+        return Turn(round=state.round, phase=PHASE_CLOSE, index=0, **carry)
+
+    # The main model has had its last word. Another round only if it asked for
+    # one *and* there is one left.
+    if not again:
+        return None
+    if state.round + 1 > max_rounds:
+        return replace(state, errors=errors, stopped=STOPPED_ROUNDS)
+    return Turn(round=state.round + 1, phase=PHASE_OUT, index=1, **carry)
+
+
+# --- Resolving the membership --------------------------------------------------
+def member_speakers(db: DBSession, chat: Chat, user=None) -> list:
+    """Every member that can actually be reached, in order, main model first.
+
+    Filtered through `permissions.models_visible_to` by way of
+    `chat_service.roster_models`, so a member whose access has been revoked, whose
+    model has been disabled, or whose row has gone is skipped rather than
+    attempted -- and the skip is visible in the transcript rather than silent.
+
+    Deduplicated against the main model: adding the chat's own model to the crowd
+    would have it answer twice in a row, which is not what anybody meant by it.
+    """
+    from lembas.services import chat as chat_service
+
+    reachable = {
+        model.model_id: model for model in chat_service.roster_models(db, user, exclude="")
+    }
+    speakers = [chat_service.Speaker(chat.model_id, chat.connection_id)]
+    seen = {chat.model_id}
+    for member in sorted(chat.crowd, key=lambda row: (row.position, row.model_id)):
+        if member.model_id in seen or member.model_id not in reachable:
+            continue
+        seen.add(member.model_id)
+        speakers.append(chat_service.Speaker(member.model_id, member.connection_id))
+    return speakers
+
+
+def unreachable_members(db: DBSession, chat: Chat, user=None) -> list[str]:
+    """Members that will be skipped, so a screen can say so rather than lie."""
+    from lembas.services import chat as chat_service
+
+    reachable = {
+        model.model_id for model in chat_service.roster_models(db, user, exclude="")
+    }
+    return [
+        member.model_id
+        for member in chat.crowd
+        if member.model_id not in reachable or member.model_id == chat.model_id
+    ]
+
+
+def is_newest(db: DBSession, message: Message) -> bool:
+    """Whether this is the last message in its chat.
+
+    The guard that stops a regenerate from forking the round. `restart` re-runs
+    `_run`, whose `finally` advances the crowd again -- and speakers further down
+    already exist, so without this, regenerating member 2 creates a second member
+    3 and two chains race down one turn. `_drain` never needed it, because a
+    queued row only ever exists *forward* of the reply.
+    """
+    latest = db.scalars(
+        select(Message)
+        .where(Message.chat_id == message.chat_id)
+        .order_by(Message.created_at.desc(), Message.id.desc())
+        .limit(1)
+    ).first()
+    return latest is not None and latest.id == message.id
+
+
+# --- Asking for another round ---------------------------------------------------
+async def _run_crowd_again(context, args: dict[str, Any]):
+    """Record that the main model wants the crowd to go round again.
+
+    Written onto the running `Generation` rather than onto the row, because it is
+    a fact about *this* reply and dies with it -- and onto a field rather than
+    parsed back out of the prose, for the reason `plan_json` exists: a sentinel
+    phrase in an answer is a decision nobody can see and a wording nobody can
+    change.
+
+    Offered only on the main model's closing turn and only while a round is left,
+    so a call arriving anywhere else is a call that was never on the table.
+    """
+    from lembas.services import generation as generation_service
+    from lembas.services.tools import ToolOutcome
+
+    reason = str(args.get("focus") or "").strip()
+    running = generation_service.running_for(context.chat_id) if context.chat_id else None
+    if running is None:
+        return ToolOutcome(
+            "There is no round to continue.",
+            {"name": "crowd_again", "status": "error", "error": "no round"},
+        )
+
+    running.crowd_again = True
+    return ToolOutcome(
+        "The others will answer again."
+        + (f" You have asked them to focus on: {reason}" if reason else "")
+        + " Finish your answer now: what you write is what the person reads for "
+        "this round.",
+        {
+            "name": "crowd_again",
+            "status": "ok",
+            "query": reason[:160],
+            "detail": "another round",
+        },
+    )
+
+
+def tool_defs() -> list:
+    """The one tool, offered only to the closing speaker of a crowd round."""
+    from lembas.services.tools import FAMILY_CROWD, RISK_READ, ToolDef
+
+    return [
+        ToolDef(
+            name="crowd_again",
+            family=FAMILY_CROWD,
+            description=(
+                "Send the other models round again, because the disagreement is "
+                "real and another pass would settle it. Say what they should focus "
+                "on. Use it sparingly: every round costs the person another wait, "
+                "and a crowd asked to go round because the discussion was "
+                "interesting will keep finding things to discuss. If the answers "
+                "have converged, or the disagreement is a matter of taste, or "
+                "nobody has said anything new on the way back, do not call this -- "
+                "write the answer instead."
+            ),
+            parameters={
+                "type": "object",
+                "properties": {
+                    "focus": {
+                        "type": "string",
+                        "description": (
+                            "What the next round should settle, in one sentence."
+                        ),
+                    }
+                },
+                "required": [],
+            },
+            run=_run_crowd_again,
+            # It changes nothing in the world; what it costs is more replies, and
+            # that is bounded by `crowd.max_rounds` rather than by an approval.
+            risk=RISK_READ,
+        ),
+    ]
+
+
+__all__ = [
+    "MAX_CONSECUTIVE_ERRORS",
+    "PHASES",
+    "PHASE_BACK",
+    "PHASE_CLOSE",
+    "PHASE_OUT",
+    "STOPPED_ERRORS",
+    "STOPPED_ROUNDS",
+    "STOPPED_TIME",
+    "Turn",
+    "elapsed",
+    "is_newest",
+    "member_speakers",
+    "next_turn",
+    "now_stamp",
+    "state_of",
+    "tool_defs",
+    "unreachable_members",
+]

src/lembas/services/generation.py

@@ -33,6 +33,7 @@ from lembas.security import permissions
 from lembas.services import canvas as canvas_service
 from lembas.services import chat as chat_service
 from lembas.services import compaction as compaction_service
+from lembas.services import crowd as crowd_service
 from lembas.services import interaction, settings_store, tokens, tool_labels
 from lembas.services import metrics as metrics_service
 from lembas.services import prompts as prompts_service
@@ -222,6 +223,15 @@ class Generation:
     # -- the one frame that reaches a browser after a reply is over.
     drained: bool = False
     injected_ids: list[str] = field(default_factory=list)
+    # A crowd round, seen from one speaker's side. `crowded` says this reply's
+    # ending handed the turn to the next speaker -- read by `_follow`, exactly as
+    # `drained` is, to put the next bubble on the `done` frame. `crowd_again` is
+    # the main model having called `crowd_again` on its closing turn: a field
+    # rather than a parse of the prose, for the reason `plan_json` exists, and
+    # on the generation rather than the row because it is a fact about this reply
+    # and dies with it.
+    crowded: bool = False
+    crowd_again: bool = False
     # Images this reply produced, waiting to be bound to its message row. The
     # runner writes the file and the `Attachment`; only `_persist` may say which
     # turn it belongs to, which is the same division of labour `canvas` above
@@ -599,7 +609,16 @@ async def _run(generation: Generation) -> None:
         # assembly path. Here rather than in post_message because that route's
         # whole contract is to return immediately, and a three-second
         # summarisation in front of it would break exactly that.
-        await _maybe_compact(generation)
+        # Once per turn, on the reply that opens it. Three reasons, and the
+        # first is the one that bites: `should_compact` reads `context_limit` off
+        # the *last complete* assistant turn's usage, which mid-crowd is the
+        # previous **speaker** -- so an 8k member at position three tells a 128k
+        # member at position four to compact. `last_complete`'s own promise that
+        # the cut lands on a reply and therefore leaves a history starting on a
+        # user turn is also false mid-round. And compacting during a round would
+        # ask the way back whether it disagrees with a summary of itself.
+        if _opens_the_turn_id(generation):
+            await _maybe_compact(generation)
 
         # Before the session opens, for the same reason compaction is: the
         # listing is an SSH round trip, and holding a database session across
@@ -614,7 +633,14 @@ async def _run(generation: Generation) -> None:
                 generation.error = "That chat no longer exists."
                 return
 
-            endpoint, model_id = chat_service.resolve_endpoint(db, chat)
+            # Who is answering, from the row being written into rather than
+            # from the chat. The row is durable and this generation is not: a
+            # restart turns `_follow` into `ensure`, which starts a brand new
+            # `_run` against the same message, and everything the request depends
+            # on has to survive that. It is also the only thing that can make the
+            # bubble's avatar and the model actually asked agree.
+            speaker = chat_service.speaker_for(db, chat, message)
+            endpoint, model_id = chat_service.resolve_endpoint(db, chat, speaker)
             owner = db.get(User, chat.user_id)
 
             # Before the request is built, not while it streams. Every other
@@ -631,13 +657,42 @@ async def _run(generation: Generation) -> None:
             # Resolved once, so that what the loop is allowed to *run* is the
             # same set the endpoint was *offered* -- not whatever happens to
             # exist by the time a call comes back.
-            toolset = tools_service.resolve_tools(db, chat, owner)
+            # Where this speaker sits in a crowd round, if it is in one. Read
+            # once, here, and used for three decisions: which tools it may have,
+            # which instruction closes its request, and whether it may ask for
+            # another round.
+            crowd_state = crowd_service.state_of(message)
+            crowd_settings = settings_store.crowd(db)
+            may_ask_again = bool(
+                crowd_state is not None
+                and crowd_state.phase == crowd_service.PHASE_CLOSE
+                and crowd_state.round < int(crowd_settings["max_rounds"])
+            )
+            toolset = tools_service.resolve_tools(
+                db, chat, owner, speaker, crowd_turn=crowd_state, crowd_again=may_ask_again
+            )
             offered = toolset.schemas
             payload = chat_service.build_request(
-                db, chat, upto=message, tools=offered, user=owner, force_tool=generation.force_tool
+                db,
+                chat,
+                upto=message,
+                tools=offered,
+                user=owner,
+                force_tool=generation.force_tool,
+                speaker=speaker,
+                crowd_turn=crowd_state,
+                # Asked of the resolved set rather than of the settings: a model
+                # without the tools capability gets no tools at all, so inviting it
+                # to call `crowd_again` would be offering a choice it cannot
+                # express -- and `crowd.close_final` is the wording for that.
+                crowd_again="crowd_again" in toolset.by_name,
             )
             question = _question_from(payload)
-            needs_title = not chat.title_generated
+            # Once per turn. A crowd member titling the chat would name it after
+            # `_question_from`'s last user turn, which under the crowd relabelling
+            # is another model's quoted answer -- so the chat gets called after a
+            # quotation. The main model's first reply is the one that titles.
+            needs_title = not chat.title_generated and _opens_the_turn(message)
             # An agent chat is titled from its opening words and never costs a
             # model call for it. That prompt is a good title already -- somebody
             # starting one states an objective, not a topic -- while an ordinary
@@ -654,16 +709,22 @@ async def _run(generation: Generation) -> None:
             # Read here, with the rest, because titling happens after this
             # session has closed and must not open another one.
             title_prompt = prompts_service.resolve(db, "task.title")
-            tool_context = tools_service.context_for(db, owner, chat, tools=toolset)
+            tool_context = tools_service.context_for(
+                db, owner, chat, tools=toolset, speaker=speaker
+            )
 
-            model = chat_service.model_for(db, chat)
+            # The answering model's window, not the chat's. `_too_big` is the one
+            # budget that stops a reply dead rather than asking it to wrap up, so
+            # judging a small model's request against a large model's ceiling is
+            # how a reply fails with no explanation in it.
+            model = chat_service.model_row(db, speaker)
             generation.context_limit = model.context_length if model is not None else 0
             # Kept for `_inject`, which builds a user turn after this session
             # has closed. A turn taken in mid-reply has to be shaped exactly as
             # the same words typed a moment later would have been -- images to a
             # vision model, a plain string to anything else, or the endpoint
             # rejects the whole request.
-            vision = chat_service.model_supports(db, chat, "vision")
+            vision = chat_service.model_supports(db, chat, "vision", speaker=speaker)
             # Resolved while the session is open, like everything else here.
             # Empty for an admin and for a user in no group, which is every
             # instance that has not set one -- see permissions.limits_for.
@@ -1082,7 +1143,18 @@ async def _run(generation: Generation) -> None:
         # `_persist` is: `_follow` breaks the instant it sees that flag, and the
         # frame it then sends is the one that has to carry the next turn's
         # bubbles. There is no push channel that outlives a single reply.
-        _drain(generation)
+        #
+        # 🚨 Advancing a crowd round *suppresses* the drain, and the order of this
+        # sentence is the whole of it. Written the other way round -- advance, then
+        # drain -- a queued human turn typed during a round would create a second
+        # incomplete assistant row beside the next speaker's, which is two
+        # generations in one chat: the state `_reply_in_flight`, `_too_many_replies`,
+        # `wake.lock_for` and the superseded guards in `_persist`/`_drain` all exist
+        # to make unreachable, and whose symptom is a Stop button pointing at
+        # whichever bubble comes first in the document. The queue waits for the
+        # round; that is what a queue is for.
+        if not _advance_crowd(generation):
+            _drain(generation)
         generation.done = True
         generation.finished_at = datetime.now(UTC)
         generation.touch()
@@ -2112,9 +2184,166 @@ def _drain(generation: Generation) -> None:
     generation.drained = True
 
 
+def _advance_crowd(generation: Generation) -> bool:
+    """Start the next speaker of a crowd round. True if one was started.
+
+    The imperative shell around `crowd.next_turn`, which is pure -- so everything
+    interesting about this (the eight ways a round declines to continue) is tested
+    without an endpoint, and what is left here is reading rows and writing one.
+
+    Three refusals of its own, and each is a bug if it is left out:
+
+    * **Superseded.** The same guard `_persist` and `_drain` carry: this reply is
+      no longer the one registered for its message.
+    * **Stopped.** A person pressing Stop ends the round, not just the speaker
+      writing at the time. `_drain` refuses after a stop for the same reason and
+      it is the same reason here -- somebody asked for it to end.
+    * **Not the newest message.** `regenerate` calls `restart`, whose `finally`
+      runs this again -- and the speakers after it already exist. Without this,
+      regenerating member 2 creates a second member 3 and two chains race down one
+      turn. `_drain` never needed the guard because a queued row only ever exists
+      *forward* of the reply.
+
+    An **error** does not end the round: `crowd.next_turn` counts consecutive
+    failures and abandons after two, because the commonest failure in a crowd is a
+    small member's context window overflowing rather than a dead endpoint, and
+    ending the round there would kill every crowd at whichever member is smallest.
+    """
+    owner = _RUNNING.get(generation.message_id)
+    if owner is not None and owner is not generation:
+        return False
+    if generation.stopped:
+        return False
+
+    try:
+        with session_scope() as db:
+            chat = db.get(Chat, generation.chat_id)
+            message = db.get(Message, generation.message_id)
+            if chat is None or message is None:
+                return False
+
+            settings = settings_store.crowd(db)
+            if not settings["enabled"] or not chat.crowd:
+                return False
+            if not crowd_service.is_newest(db, message):
+                return False
+
+            owner_user = db.get(User, chat.user_id)
+            speakers = crowd_service.member_speakers(db, chat, owner_user)
+            speakers = speakers[: int(settings["max_models"]) + 1]
+
+            state = crowd_service.state_of(message)
+            # The turn a round belongs to: the user message this all answers.
+            turn_id = state.turn if state is not None else _turn_anchor(db, message)
+            following = crowd_service.next_turn(
+                speakers=len(speakers),
+                state=state,
+                turn_id=turn_id,
+                again=generation.crowd_again,
+                errored=bool(generation.error),
+                max_rounds=int(settings["max_rounds"]),
+                wall_seconds=int(settings["wall_seconds"]),
+            )
+            if following is None:
+                return False
+            if following.stopped:
+                # Recorded on the row that ended it, so the transcript can say
+                # why a round stopped rather than simply stopping. Nothing else
+                # needs writing: there is no next speaker.
+                message.crowd_json = following.as_json()
+                db.commit()
+                return False
+
+            speaker = speakers[following.index]
+            placeholder = chat_service.create_message(
+                db,
+                chat,
+                ROLE_ASSISTANT,
+                "",
+                complete_=False,
+                model_id=speaker.model_id,
+            )
+            placeholder.connection_id = speaker.connection_id
+            placeholder.crowd_json = following.as_json()
+            db.commit()
+            chat_id, next_id = chat.id, placeholder.id
+    except Exception:  # noqa: BLE001 - the reply is over either way
+        log.exception("could not advance the crowd in chat %s", generation.chat_id)
+        return False
+
+    # Outside the session, like `_drain`: this starts a task.
+    ensure(chat_id, next_id)
+    generation.crowded = True
+    return True
+
+
+def _opens_the_turn(message: Message) -> bool:
+    """Whether this reply is the first one answering a question.
+
+    True for every ordinary reply, and for a crowd only for the main model's
+    opening turn -- which is the one with no crowd state on it at all, because a
+    round begins when that reply *finishes*.
+    """
+    return crowd_service.state_of(message) is None
+
+
+def _opens_the_turn_id(generation: Generation) -> bool:
+    """`_opens_the_turn` before the session is open, by message id.
+
+    `_maybe_compact` runs before `_run` reads anything, so this opens its own
+    session -- one primary-key lookup, and only on a chat that has a crowd.
+    """
+    try:
+        with session_scope() as db:
+            message = db.get(Message, generation.message_id)
+            return message is None or _opens_the_turn(message)
+    except Exception:  # noqa: BLE001 - compaction is best-effort anyway
+        return True
+
+
+def _ends_the_turn(message: Message) -> bool:
+    """Whether this reply is the last one the person is waiting for.
+
+    True for every ordinary reply, and for a crowd only on the main model's
+    closing turn. What is gated on it is everything that should happen once per
+    question rather than once per speaker: the unread dot, the web push, and the
+    chat's title.
+    """
+    state = crowd_service.state_of(message)
+    if state is None:
+        return True
+    return state.phase == crowd_service.PHASE_CLOSE
+
+
+def _turn_anchor(db, message: Message) -> str:
+    """The user turn a round answers, for a round that is only now beginning.
+
+    The last user message at or before this reply. Only read once per round -- it
+    is carried on every later turn's state -- and it exists so a rewind can tell
+    which rows belonged to which question.
+    """
+    row = db.scalars(
+        select(Message)
+        .where(
+            Message.chat_id == message.chat_id,
+            Message.role == ROLE_USER,
+            Message.created_at <= message.created_at,
+        )
+        .order_by(Message.created_at.desc(), Message.id.desc())
+        .limit(1)
+    ).first()
+    return row.id if row is not None else ""
+
+
 def _inject(generation: Generation, chat_id: str, vision: bool) -> dict | None:
     """Take the oldest waiting prompt into this reply, between two rounds.
 
+    ⚠ Never during a crowd round. This restamps the placeholder's `created_at` so
+    the reply sorts after the prompt it answers, which mid-round reorders the
+    speakers underneath themselves -- and the round's own bookkeeping counts an
+    anchor that has moved. The turn stays queued and arrives after the round as a
+    clean new question with a round of its own, which is what `_drain` is for.
+
     Marked delivered and committed *before* the request goes out, so this is
     at-most-once. A crash in between loses the turn, which is recoverable --
     the words are still in the transcript with Send now beside them. The other
@@ -2132,6 +2361,9 @@ def _inject(generation: Generation, chat_id: str, vision: bool) -> dict | None:
     """
     try:
         with session_scope() as db:
+            message = db.get(Message, generation.message_id)
+            if crowd_service.state_of(message) is not None:
+                return None
             waiting = _next_waiting(db, chat_id)
             if waiting is None:
                 return None
@@ -2265,7 +2497,13 @@ def _persist(generation: Generation, title: str, elapsed: float) -> None:
             # clears this when it is next opened. Not for a temporary chat:
             # there is no sidebar row for the dot, and the toast would name a
             # chat nobody can navigate to.
-            if generation.followers == 0 and not chat.temporary:
+            # 🚨 Once per *turn*, not once per speaker. `announce_later` has no
+            # dedupe of its own -- its docstring says so, because every site that
+            # calls it runs once per arrival -- so a five-model crowd with nobody
+            # watching would be nine web pushes and nine sidebar toasts for one
+            # question. The closing speaker is the arrival; everybody before it is
+            # the middle of one.
+            if generation.followers == 0 and not chat.temporary and _ends_the_turn(message):
                 chat.unread = True
                 chat.unread_notified = False
                 # And out to any browser that asked to be told, which is the

src/lembas/services/harness.py

@@ -166,6 +166,7 @@ def context_variables(
     user: User | None,
     tools: list[dict[str, Any]] | None,
     chat=None,
+    speaker=None,
 ) -> dict[str, str]:
     """What every ``{{name}}`` in a fragment resolves to for this request.
 
@@ -294,8 +295,12 @@ def context_variables(
         from lembas.services import chat as chat_service
         from lembas.services.subagent import ROLE_FRIEND
 
-        model = chat_service.model_for(db, chat)
-        values["model_name"] = model.label if model is not None else chat.model_id
+        # The *answering* model, not the chat's: telling a crowd member it is the
+        # main model is a lie it will then reason from, and its personality is
+        # keyed on whichever model is speaking.
+        speaking = speaker or chat_service.speaker_for(db, chat)
+        model = chat_service.model_row(db, speaking)
+        values["model_name"] = model.label if model is not None else speaking.model_id
         # Naming the bases a chat is scoped to matters: without it the model
         # cannot tell "there is nothing about this" from "I am only allowed to
         # see the contracts folder", and phrases a miss as the former.
@@ -333,7 +338,7 @@ def context_variables(
         # reason the roster and the tool are one checkbox rather than two.
         if "friend" in families:
             values["model_roster"] = chat_service.roster_block(
-                db, user, exclude=chat.model_id
+                db, user, exclude=speaking.model_id
             )
 
         if "persona" in families:
@@ -341,7 +346,7 @@ def context_variables(
             # administrator's default until the model has written one with them;
             # and this model's impression of them, which has no default and never
             # could.
-            key = chat.model_id
+            key = speaking.model_id
             values["persona"] = personas_service.block(db, key, user)
             values["person_view"] = personas_service.view_block(db, key, user)
 
@@ -512,12 +517,13 @@ def compose(
     user: User | None,
     tools: list[dict[str, Any]] | None,
     chat=None,
+    speaker=None,
 ) -> str:
     """The operational preamble for this request, or "" when there is nothing to say."""
     offered = tools or []
     return compose_from(
         db,
-        variables=context_variables(db, user, offered, chat),
+        variables=context_variables(db, user, offered, chat, speaker),
         families=_families(db, offered),
         has_tools=bool(offered),
     )

src/lembas/services/prompts.py

@@ -182,6 +182,14 @@ VARIABLES: tuple[Variable, ...] = (
         "Per model and per person, so two models may hold different views and "
         "nobody sees anybody else's. The person can read and delete it.",
     ),
+    Variable(
+        "crowd_speaker",
+        "The model being quoted",
+        "Inside the crowd fragments only: the name of the model whose words "
+        "follow, or whose turn it is. Blank everywhere else, because it is a "
+        "property of one quotation rather than of a request — which is why the "
+        "legend cannot show you a value for it.",
+    ),
     Variable(
         "timezone",
         "Timezone",
@@ -1995,6 +2003,106 @@ BUILTIN: tuple[Fragment, ...] = (
             "{{transcript}}"
         ),
     ),
+    Fragment(
+        key="crowd.said",
+        label="Quoting another model in a crowd",
+        group=GROUP_TASKS,
+        order=450,
+        variables=("crowd_speaker",),
+        hint="What another speaker's answer is labelled as when it reaches this "
+        "one. It matters more than it looks: sent unlabelled, every earlier reply "
+        "arrives as something *this* model said, so it defends sentences it never "
+        "wrote and cannot disagree with them — which is the whole point of the "
+        "way back. Relabelling is also what keeps the history alternating, which "
+        "several chat templates require.",
+        default="{{crowd_speaker}} answered:",
+    ),
+    Fragment(
+        key="crowd.turn",
+        label="A crowd member's turn on the way out",
+        group=GROUP_TASKS,
+        order=451,
+        hint="Added as the last turn when a member speaks on the forward pass. "
+        "The failure to word against is a member that repeats what has already "
+        "been said in different words, which is what makes a crowd feel like an "
+        "echo rather than a second opinion.",
+        default=(
+            "You are one of several models answering this. The answers above are "
+            "quoted with the name of whoever wrote them; yours comes next.\n"
+            "\n"
+            "Add what is missing, correct what is wrong, and say what you would "
+            "have done differently. Do not restate what has already been said to "
+            "show that you agree with it — if you have nothing to add, say so in "
+            "one line and stop. Be brief: somebody is reading all of these."
+        ),
+    ),
+    Fragment(
+        key="crowd.disagree",
+        label="A crowd member's turn on the way back",
+        group=GROUP_TASKS,
+        order=452,
+        hint="Added as the last turn on the backward pass, which is where the "
+        "value of a crowd actually is: everybody has now been heard, and this is "
+        "the chance to object. Worded to ask for disagreement rather than for a "
+        "summary, because a model asked to review will produce a review whether "
+        "it has one or not.",
+        default=(
+            "Everybody has now answered. Read the whole exchange again.\n"
+            "\n"
+            "Do you disagree with anything said above — a claim that is wrong, a "
+            "risk nobody named, an answer to the wrong question? Say so plainly, "
+            "and say which part you mean. **If you have no disagreement, reply "
+            "with one short sentence saying so and nothing else.** Do not "
+            "summarise, do not praise the other answers, and do not repeat your "
+            "own."
+        ),
+    ),
+    Fragment(
+        key="crowd.close",
+        label="The main model's last word, with another round available",
+        group=GROUP_TASKS,
+        order=453,
+        hint="The main model's closing turn when it can still ask for another "
+        "round. Its own fragment rather than a sentence inside the one below, "
+        "because inviting a choice a model cannot express is worse than not "
+        "offering it: on a model without the tools capability there is no "
+        "crowd_again to call, and that is the case the next fragment covers.",
+        default=(
+            "You opened this and you are closing it. The others have answered and "
+            "have had the chance to disagree.\n"
+            "\n"
+            "Write the answer the person actually asked for. Take what the others "
+            "got right, say where you disagree with them and why, and name "
+            "anything still unresolved rather than papering over it. Attribute "
+            "what you took from whom.\n"
+            "\n"
+            "If the disagreement is real and another round would settle it, call "
+            "crowd_again and say what you want them to address. Do not call it "
+            "because the discussion was interesting — every round costs the person "
+            "another wait."
+        ),
+    ),
+    Fragment(
+        key="crowd.close_final",
+        label="The main model's last word, with no round left",
+        group=GROUP_TASKS,
+        order=454,
+        hint="The same turn when another round is not on offer — the round limit "
+        "is reached, or this model has no tools and so cannot ask. It says the "
+        "answer has to be final rather than inviting a choice that would be "
+        "ignored, which is the difference between a feature and a feature that "
+        "looks like one.",
+        default=(
+            "You opened this and you are closing it, and this is the last turn: "
+            "there will be no further round.\n"
+            "\n"
+            "Write the answer the person actually asked for. Take what the others "
+            "got right, say where you disagree with them and why, and attribute "
+            "what you took from whom. Where the disagreement is unresolved, say so "
+            "and say what would settle it — that is more useful than a confident "
+            "answer papered over the top of it."
+        ),
+    ),
     Fragment(
         key="task.compact_lead",
         label="How a summary is introduced",

src/lembas/services/settings_store.py

@@ -32,6 +32,7 @@ AGENTS = "agents"
 IMAGES = "images"
 SCHEDULES = "schedules"
 SUBAGENTS = "subagents"
+CROWD = "crowd"
 BRANDING = "branding"
 EXTRACTION = "extraction"
 
@@ -343,6 +344,41 @@ def _schedules_defaults() -> dict[str, Any]:
     }
 
 
+def _crowd_defaults() -> dict[str, Any]:
+    """Several models answering one turn, in order, then again in reverse.
+
+    Off until an administrator turns it on, and the reason is arithmetic: one
+    turn costs **models x rounds x 2 - 1** replies, so four models over two
+    rounds is fifteen. On a single local endpoint every change of speaker is also
+    a model load, because llama-swap holds one at a time.
+
+    The owner's own warning, recorded because it is the failure this feature
+    actually has: *larger crowds of smaller models -- and sometimes of bigger
+    ones -- start cycling, or never stop.* So the numbers below are a ceiling
+    reached by ordinary work, not a runaway backstop, which is the opposite of
+    how `subagents.max_rounds` is set and is deliberate: a round of a crowd is a
+    visible, expensive thing somebody is waiting through.
+    """
+    return {
+        "enabled": False,
+        # Besides the chat's own model. Four speakers is already eight replies a
+        # turn at one round each.
+        "max_models": 4,
+        # One round is out-and-back: everyone answers, then everyone is asked
+        # whether they disagree, ending at the main model. Two is one chance to
+        # change its mind after hearing the objections, which is the whole point;
+        # three is where cycling starts.
+        "max_rounds": 2,
+        # The whole turn, across every speaker, so a member whose endpoint has
+        # stalled cannot hold a round open all afternoon.
+        "wall_seconds": 900,
+        # Whether a short "I agree" on the way back is collapsed in the
+        # transcript. On by default: N-1 bubbles saying nothing is what makes
+        # somebody switch the feature off, and the disagreements are the point.
+        "collapse_agreement": True,
+    }
+
+
 def _subagents_defaults() -> dict[str, Any]:
     """Delegating a piece of a reply to a second, unattended model.
 
@@ -389,6 +425,7 @@ _DEFAULTS: dict[str, Any] = {
     IMAGES: _images_defaults,
     SCHEDULES: _schedules_defaults,
     SUBAGENTS: _subagents_defaults,
+    CROWD: _crowd_defaults,
     # Whose instance this is. The defaults live in `services/branding.py`
     # beside the code that reads them, because every one of them is paired with
     # a label and a hint for the admin page and splitting the three across two
@@ -668,6 +705,22 @@ def subagents(db: DBSession) -> dict[str, Any]:
     return values
 
 
+def crowd(db: DBSession) -> dict[str, Any]:
+    """Crowd settings, clamped on read for the reason `agents` gives.
+
+    Every bound has a floor of one: a `max_models` of zero is the feature
+    switched off wearing the switch's clothes, and that is a thing to answer in
+    one place rather than two.
+    """
+    values = get_group(db, CROWD)
+    values["max_models"] = min(max(int(values.get("max_models") or 1), 1), 8)
+    values["max_rounds"] = min(max(int(values.get("max_rounds") or 1), 1), 5)
+    values["wall_seconds"] = min(max(int(values.get("wall_seconds") or 1), 60), 7200)
+    values["enabled"] = bool(values.get("enabled"))
+    values["collapse_agreement"] = bool(values.get("collapse_agreement"))
+    return values
+
+
 def images_ready(db: DBSession) -> bool:
     """Whether image generation can actually happen.
 

src/lembas/services/tool_labels.py

@@ -76,6 +76,8 @@ LABELS: dict[str, str] = {
     "subagent_run": "Helper",
     # A question put to one of the other models here.
     "ask_friend": "Asked another model",
+    # The main model sending a crowd round again.
+    "crowd_again": "Another round",
     # What a model keeps about itself and about the person it is talking to.
     "persona_write": "Personality rewritten",
     "impression_write": "Impression updated",
@@ -121,6 +123,7 @@ ICONS: dict[str, str] = {
     "schedule_cancel": "stop-circle",
     "subagent_run": "sparkle",
     "ask_friend": "users",
+    "crowd_again": "refresh",
     "persona_write": "user",
     "impression_write": "user",
     "memory_add": "star",
@@ -169,6 +172,7 @@ ACTIONS: dict[str, str] = {
     "schedule_cancel": "Stop a schedule",
     "subagent_run": "Send a helper",
     "ask_friend": "Ask another model",
+    "crowd_again": "Send the crowd round again",
     "persona_write": "Rewrite its own personality",
     "impression_write": "Update what it makes of you",
     "memory_add": "Remember something",
@@ -216,6 +220,9 @@ DETAIL_KEYS: dict[str, str] = {
     # question carrying a wrong assumption comes back as a confident answer
     # about the wrong thing -- the same reason `subagent_run` names the task.
     "ask_friend": "question",
+    # What the next round is for. The only field it has, and the one thing worth
+    # correcting before several models spend a reply each on it.
+    "crowd_again": "focus",
     # The whole text, because for these two the text *is* the thing being agreed
     # to: there is no shorter field that says what the model would become.
     "persona_write": "content",

src/lembas/services/tools.py

@@ -162,6 +162,15 @@ FAMILY_FRIEND = "friend"
 # a model up: either it may form and keep opinions of this kind or it may not.
 FAMILY_PERSONA = "persona"
 
+# Sending a crowd round again. Its own family so `harness._families` can map the
+# name back to one, and deliberately **not in `FAMILIES`**: that tuple is the list
+# of things an administrator switches on, and this is mechanism. Being in it would
+# mint a `tool_crowd` capability checkbox and demand a `tools.crowd` permission
+# that does not exist -- which, because `_family_allowed` falls through to
+# `allowed.get(...)`, would mean the tool could never be offered at all. Its real
+# gate is `resolve_tools(crowd_again=…)`: one turn of one round.
+FAMILY_CROWD = "crowd"
+
 # The built-in families, in the order they are offered.
 FAMILIES = (
     FAMILY_SEARCH,
@@ -1635,6 +1644,13 @@ def _family_allowed(
         # separate switch would be a second door to the cost with nothing
         # naming it. `Helpers` on /admin/agents is where both are bounded.
         return bool(allowed.get("tools.friend") and subagents)
+    if gate == FAMILY_CROWD:
+        # Always allowed, because whether it is *offered* is decided before this:
+        # `resolve_tools` puts it in the book only on the main model's closing turn
+        # with a round still left. A permission here would be a second switch for
+        # one already-enabled feature, and an absent one would silently make the
+        # crowd a single round for ever.
+        return True
     if gate in (
         FAMILY_CUSTOM,
         FAMILY_MCP,
@@ -1721,6 +1737,13 @@ def _friend_defs() -> list[ToolDef]:
     return subagent_service.friend_tool_defs()
 
 
+def _crowd_defs() -> list[ToolDef]:
+    """The go-round-again tool. Imported inside the call for the reason above."""
+    from lembas.services import crowd as crowd_service
+
+    return crowd_service.tool_defs()
+
+
 def _image_defs(db: DBSession, values: dict | None = None) -> list[ToolDef]:
     """The image tool, whose schema carries this instance's own choices.
 
@@ -1776,6 +1799,7 @@ def registry(db: DBSession) -> dict[str, ToolDef]:
             *_schedule_defs(),
             *_subagent_defs(),
             *_friend_defs(),
+            *_crowd_defs(),
         ]
     )
 
@@ -1786,13 +1810,31 @@ def families(db: DBSession) -> tuple[str, ...]:
     return (*FAMILIES, *rows)
 
 
-def resolve_tools(db: DBSession, chat: Chat, user: User | None) -> ToolSet:
-    """Every tool this chat may call right now, with its runner attached."""
+def resolve_tools(
+    db: DBSession,
+    chat: Chat,
+    user: User | None,
+    speaker=None,
+    *,
+    crowd_turn=None,
+    crowd_again: bool = False,
+) -> ToolSet:
+    """Every tool this chat may call right now, with its runner attached.
+
+    The capabilities are the **answering** model's. `tools` being off is the first
+    gate and returns nothing at all, so handing a crowd member the main model's
+    switches would offer a tool list to an endpoint that rejects the request for
+    carrying one.
+    """
     from lembas.security import permissions
     from lembas.services import chat as chat_service
 
     capabilities = {}
-    model = chat_service.model_for(db, chat)
+    model = (
+        chat_service.model_row(db, speaker)
+        if speaker is not None
+        else chat_service.model_for(db, chat)
+    )
     if model is not None:
         capabilities = model.capabilities_json or {}
 
@@ -1819,6 +1861,12 @@ def resolve_tools(db: DBSession, chat: Chat, user: User | None) -> ToolSet:
             *(_schedule_defs() if schedules_on else []),
             *(_subagent_defs() if subagents_on else []),
             *(_friend_defs() if subagents_on else []),
+            # Only on the closing turn, and only with a round left. Not gated on a
+            # capability or a permission: a tool that exists on exactly one turn of
+            # one feature is mechanism, and an administrator switching it off would
+            # be switching off the main model's ability to use the feature it
+            # already enabled.
+            *(_crowd_defs() if crowd_again else []),
         ]
     )
 
@@ -1832,6 +1880,26 @@ def resolve_tools(db: DBSession, chat: Chat, user: User | None) -> ToolSet:
     off = scoped_off(chat)
     empty_library = not skills_service.count_enabled(db, user, exclude=scoped_skills_off(chat))
 
+    # What a crowd speaker may do, which is narrower than what the chat may.
+    if crowd_turn is not None:
+        from lembas.services import crowd as crowd_service
+
+        if crowd_turn.phase == crowd_service.PHASE_BACK:
+            # The way back is "do you disagree with any of this", which needs
+            # nothing looked up: everything it is about is already in front of it.
+            # An empty toolset also guarantees the turn ends in words, which is the
+            # shape `_wrap_up` relies on.
+            return ToolSet()
+        if not crowd_turn.is_main:
+            # A member answers a machine-composed instruction with several models'
+            # words quoted into it, and nobody is waiting on *it* in particular.
+            # So: it cannot stop the round for an approval or a question -- one
+            # card would park every remaining speaker for `approval_timeout` -- it
+            # cannot fan out, and it cannot rewrite a personality under wording it
+            # did not choose. The same set `unattended` withdraws, for the same
+            # reasons, applied for a different one.
+            off = off | {FAMILY_ASK, FAMILY_SUBAGENT, FAMILY_FRIEND, FAMILY_PERSONA}
+
     # A scheduled task runs with nobody present, so `ask_user` cannot work here:
     # it pauses the reply and waits for a POST that will never come, until
     # `approval_timeout` expires -- a run that silently does nothing for fifteen
@@ -2017,10 +2085,21 @@ def context_for(
     chat: Chat | None = None,
     *,
     tools: ToolSet | None = None,
+    speaker=None,
 ) -> ToolContext:
-    """The snapshot a running tool needs, taken while the session is open."""
+    """The snapshot a running tool needs, taken while the session is open.
+
+    `speaker` is the model answering, and it decides which model a tool acts *as*:
+    which personality `persona_write` rewrites, and whose endpoint the image
+    reviewer and the Preserve-VRAM unload reach for. It defaults to the chat's own
+    model.
+    """
+    from lembas.services import chat as chat_service
     from lembas.services.agent import session as agent_session
 
+    if chat is not None and speaker is None:
+        speaker = chat_service.speaker_for(db, chat)
+
     return ToolContext(
         agent=agent_session.resolve(db, chat, user) if chat is not None else None,
         owner_id=user.id if user else "",
@@ -2029,8 +2108,8 @@ def context_for(
         image_config=settings_store.images(db),
         image_workflow_id=(chat.image_workflow_id or "") if chat is not None else "",
         image_checkpoint=(chat.image_checkpoint or "") if chat is not None else "",
-        model_id=(chat.model_id or "") if chat is not None else "",
-        connection_id=(chat.connection_id or "") if chat is not None else "",
+        model_id=(speaker.model_id or "") if speaker is not None else "",
+        connection_id=(speaker.connection_id or "") if speaker is not None else "",
         base_ids=[base.id for base in chat.knowledge_bases] if chat is not None else [],
         skills_off=scoped_skills_off(chat),
         tools=tools.by_name if tools is not None else None,

src/lembas/web/static/css/app.css

@@ -417,6 +417,12 @@ input.visually-hidden[type="checkbox"] {
   color: var(--ink-muted);
 }
 .badge--leaf { background: var(--leaf-soft); color: var(--leaf); }
+/* A crowd's backward pass: everybody has answered and each is being asked
+   whether it disagrees. Quieter than an answer, because most of these are one
+   line saying "no" -- and deliberately *not* hidden, because the one that says
+   yes is the whole reason the feature exists. */
+.msg--crowd-back .msg__body { color: var(--ink-muted); }
+.msg--crowd-back .msg__author { font-weight: 500; }
 .badge--success { background: var(--success-soft); color: var(--success); }
 .badge--danger { background: var(--danger-soft); color: var(--danger); }
 .badge--warning { background: var(--warning-soft); color: var(--warning); }

src/lembas/web/templates/admin/agents.html

@@ -445,6 +445,93 @@
   button was pressed, which is what keeps each group's save handler writing one
   key.
 #}
+{# A third settings group on this page, saved by its own form -- the reason the
+   Helpers card gives. A crowd is not an agent-chat feature either, but this is the
+   page somebody opens to find out what one turn may set going. #}
+<form method="post" action="/admin/agents/crowd" class="form-grid">
+  <section class="card">
+    <h2 class="card__title">A crowd</h2>
+    <p class="field__hint">
+      A chat can have more than one model in it. The chat's own model answers, then
+      each of the others in turn; then the order runs <strong>backwards</strong>,
+      each one asked whether it disagrees with anything; and it ends back at the
+      first, which either closes or sends them round again.
+    </p>
+
+    <div class="alert">
+      {{ icon("warning", "icon--sm") }}
+      <span>
+        One turn costs <strong>models × rounds × 2 − 1</strong> replies — four
+        models over two rounds is fifteen — and on a single local endpoint every
+        change of speaker also loads a different model. Larger crowds of smaller
+        models, and sometimes of bigger ones, start going round in circles: that is
+        what the round limit is for, and it is a limit ordinary work will reach
+        rather than a runaway backstop.
+      </span>
+    </div>
+
+    <div class="field">
+      <label class="checkbox">
+        <input type="checkbox" name="enabled" value="true"
+               {{ 'checked' if crowd.enabled }}>
+        <span>Let a chat have a crowd</span>
+      </label>
+      <p class="field__hint">
+        Off by default. With it on, each chat's settings panel offers the other
+        models; a chat with none ticked behaves exactly as it always has.
+      </p>
+    </div>
+
+    <div class="field">
+      <label class="field__label" for="crowd_max_models">Most models besides the chat's own</label>
+      <input class="input" id="crowd_max_models" name="max_models"
+             type="number" min="1" max="8" step="1" value="{{ crowd.max_models }}">
+      <p class="field__hint">
+        Four is already eight replies a turn at one round each. More voices past
+        that tend to repeat each other rather than add anything.
+      </p>
+    </div>
+
+    <div class="field">
+      <label class="field__label" for="crowd_max_rounds">Most rounds</label>
+      <input class="input" id="crowd_max_rounds" name="max_rounds"
+             type="number" min="1" max="5" step="1" value="{{ crowd.max_rounds }}">
+      <p class="field__hint">
+        A round is out and back. Two gives the first model one chance to change its
+        mind after hearing the objections, which is the point of the whole thing;
+        three is where going in circles starts.
+      </p>
+    </div>
+
+    <div class="field">
+      <label class="field__label" for="crowd_wall_seconds">Longest a turn may take</label>
+      <input class="input" id="crowd_wall_seconds" name="wall_seconds"
+             type="number" min="60" max="7200" step="30" value="{{ crowd.wall_seconds }}">
+      <p class="field__hint">
+        Across every speaker, not each. A member whose endpoint has stalled cannot
+        then hold the round open all afternoon.
+      </p>
+    </div>
+
+    <div class="field">
+      <label class="checkbox">
+        <input type="checkbox" name="collapse_agreement" value="true"
+               {{ 'checked' if crowd.collapse_agreement }}>
+        <span>Fold away a short "I agree" on the way back</span>
+      </label>
+      <p class="field__hint">
+        The disagreements are what a crowd is for; a column of bubbles saying
+        nothing is what makes somebody switch it off. The text is still there
+        behind a disclosure.
+      </p>
+    </div>
+
+    <div class="btn-row">
+      <button class="btn btn--primary" type="submit">Save</button>
+    </div>
+  </section>
+</form>
+
 <form method="post" action="/admin/agents/subagents" class="form-grid">
   <section class="card">
     <h2 class="card__title">Helpers</h2>

src/lembas/web/templates/chat/_message.html

@@ -29,8 +29,17 @@
   `msg--machine` only overrides what should differ.
 #}
 {% set machine = (message.role == "user" and message.machine) %}
+{#
+  Where this bubble sits in a crowd round, if it is in one. Nine bubbles for one
+  question need orienting, and a `<details>` wrapper round the round is the wrong
+  way to do it: bubbles arrive with `beforeend:#thread`, which appends *after* any
+  container, so the live and reloaded renderings would disagree and a reload would
+  rearrange nine bubbles under the reader. A chip on each one is the same markup
+  either way.
+#}
+{% set crowd = message.crowd_json or None %}
 
-<article class="msg msg--{{ message.role }}{{ ' msg--machine' if machine }}{{ ' msg--queued' if queued }}"
+<article class="msg msg--{{ message.role }}{{ ' msg--machine' if machine }}{{ ' msg--queued' if queued }}{{ ' msg--crowd-back' if crowd and crowd.get('phase') == 'back' }}"
          id="msg-{{ message.id }}"
   {% if streaming %}
          hx-ext="sse"
@@ -74,6 +83,33 @@
       {# Only worth showing when it adds something the author line does not. #}
       <span class="msg__model" title="{{ message.model_id }}">{{ message.model_id }}</span>
       {% endif %}
+      {% if crowd %}
+      {# Which speaker, which pass. The count is of speakers rather than of
+         replies: a round produces more bubbles than it has models in it. #}
+      <span class="badge">
+        {% if crowd.get("phase") == "out" %}
+          {{ crowd.get("index", 0) + 1 }} of {{ crowd.get("of", 1) }}
+        {% elif crowd.get("phase") == "back" %}
+          on the way back
+        {% else %}
+          closing
+        {% endif %}
+        {% if crowd.get("round", 1) > 1 %} · round {{ crowd.get("round") }}{% endif %}
+      </span>
+      {% if crowd.get("stopped") %}
+      {# Why a round ended, where it ended. Without this a crowd that ran out of
+         rounds or time simply stops, which reads as the feature failing. #}
+      <span class="badge badge--warning" title="The round ended here">
+        {% if crowd.get("stopped") == "rounds" %}
+          no rounds left
+        {% elif crowd.get("stopped") == "time" %}
+          out of time
+        {% else %}
+          two endpoints failed
+        {% endif %}
+      </span>
+      {% endif %}
+      {% endif %}
     </header>
 
     {% if message.attachments %}

src/lembas/web/templates/chat/index.html

@@ -243,6 +243,52 @@
         </div>
         {% endif %}
 
+        {% if crowd_available %}
+        {# Who else answers. Nothing ticked is every chat that has ever existed:
+           one model, answering on its own. #}
+        <div class="field">
+          <label class="field__label">Crowd</label>
+          <form hx-patch="/api/chats/{{ chat.id }}" hx-swap="none" hx-trigger="change">
+            {# Always submitted, for the reason the bases above are. #}
+            <input type="hidden" name="crowd_model_ids" value="">
+            <div class="checkbox-row">
+              {% for model in crowd_available %}
+              <label class="checkbox">
+                <input type="checkbox" name="crowd_model_ids" value="{{ model.model_id }}"
+                       {{ 'checked' if model.model_id in crowd_member_ids }}>
+                <span>{{ model.label }}</span>
+              </label>
+              {% endfor %}
+            </div>
+          </form>
+          <p class="field__hint">
+            {% if crowd_member_ids %}
+              {{ crowd_member_ids|length + 1 }} models answer each turn: this one
+              first, then the others, then back through them asking whether they
+              disagree, ending here.
+              <strong>That is {{ crowd_replies }} replies a turn</strong>, and up to
+              {{ crowd_rounds }} rounds of it.
+            {% else %}
+              Tick a model to have it answer after this one, then be asked whether
+              it disagrees. Useful for a second opinion; expensive, because each
+              one is a whole reply, and slow on one local endpoint because every
+              change of speaker loads a different model.
+            {% endif %}
+          </p>
+          {# Outside the two branches above, deliberately. A member whose model has
+             gone is filtered out of `crowd_member_ids`, so if it was the only one
+             this would fall into the "tick a model" branch and never mention the
+             row that is still there -- which is the one thing somebody needs to
+             know to tidy it up. #}
+          {% if crowd_skipped %}
+          <p class="field__hint">
+            Skipped, because you cannot reach {{ "them" if crowd_skipped|length > 1 else "it" }}
+            any more: <s>{{ crowd_skipped|join(", ") }}</s>. Untick to clear.
+          </p>
+          {% endif %}
+        </div>
+        {% endif %}
+
         {% if can.get("chat.params") %}
         <div class="grid grid--3">
           <div class="field">