Two selects that never wrote anything, and a queue

The approval card in Auto mode and the missing /effort were one bug. Both
selects hung their hx-patch on an empty sibling form reached by form="…",
and htmx binds a trigger to the annotated element: change fires on the
select and bubbles to its ancestors, which a sibling is not. The live rows
read agent_mode=manual and params_json={} while the browser showed Auto and
Effort: high. policy.py was never involved.

The verb moves onto the control; the empty form stays as value scoping,
which is the half of the CLAUDE.md note that was right. conftest gains
control_named so a test asserts the element carrying the name carries the
verb, rather than asserting the markup that was there throughout.

The composer's highlight was a third instance of the same carelessness in
CSS: .tok-mention is written for the transcript and scoped to nothing, so
the mirror painted its token in accent-coloured monospace over the
textarea's own text. Scoped under .msg; the mirror restates transparency
and font rather than inheriting them, and bleeds by box-shadow.

/effort is now offered before the first prompt and _new_chat reads it.
/index re-walks the project directory on demand, file_write drops the
listing it just invalidated, and the index ladder falls through to SFTP on
a host that refuses exec instead of returning nothing.

A second message during a reply is queued rather than starting a second
concurrent generation: a real Message row with queued set, so it survives a
restart and can be withdrawn. _drain hands one on at the end of a reply,
_inject takes one in at a tool-round boundary so an agent can be steered
mid-task. Stop leaves the queue undelivered. The terminal's Auto toggle
becomes off/copy/send, and send posts straight to the chat without touching
the composer.

@ now offers notes, skills, this chat's attachments and a URL to fetch; a
knowledge base attaches as a reference rather than a copy. copy_document
carries provenance, which was the one attach path that dropped it.

Also fixes an unrelated live bug: the round loop compared against the
global MAX_ROUNDS of 3 while sizing itself from the agent budget of 40, so
agent replies stopped after three rounds and reported forty.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "lembas"
-version = "0.6.1"
+version = "0.6.2"
 description = "LLeMbas - a Middle-earth themed web UI for OpenAI-compatible LLM endpoints"
 readme = "README.md"
 requires-python = ">=3.11"

src/lembas/__init__.py

@@ -1,3 +1,3 @@
 """LLeMbas - a Middle-earth themed web UI for OpenAI-compatible LLM endpoints."""
 
-__version__ = "0.6.1"
+__version__ = "0.6.2"

src/lembas/api/chats.py

@@ -10,8 +10,8 @@ from collections.abc import AsyncIterator
 from datetime import UTC, datetime
 
 from fastapi import APIRouter, Depends, Form, HTTPException, Request, status
-from fastapi.responses import HTMLResponse, Response, StreamingResponse
+from fastapi.responses import HTMLResponse, JSONResponse, Response, StreamingResponse
-from sqlalchemy import select
+from sqlalchemy import func, select
 from sqlalchemy.orm import Session as DBSession
 
 from lembas.api.deps import Db, RequiredUser, require_permission
@@ -49,6 +49,12 @@ router = APIRouter(prefix="/api/chats", tags=["chats"])
 # Well under nginx's 60s default; see services/sse.py:KEEPALIVE.
 KEEPALIVE_AFTER = 15.0
 
+# How many prompts may wait behind a reply at once. The terminal panel's Auto
+# send is what this exists for: a `for` loop in a shell can produce commands
+# faster than any model answers them, and a bound with a sentence attached is
+# better than four hundred rows nobody meant to write.
+MAX_QUEUED = 10
+
 
 def _owned_chat(db: DBSession, chat_id: str, user_id: str) -> Chat:
     chat = db.get(Chat, chat_id)
@@ -70,6 +76,7 @@ def _new_chat(
     ssh_profile_id: str = "",
     project_dir: str = "",
     agent_mode: str = "",
+    reasoning_effort: str = "",
 ) -> Chat:
     """Create a chat row, resolving which model it should use.
 
@@ -81,7 +88,8 @@ def _new_chat(
     before the first word. Without it, reaching Plan mode meant starting a chat
     in Manual, sending something to make the chat exist, and only then being
     offered the control -- by which point the model had already answered under
-    the wrong rules.
+    the wrong rules. The reasoning effort is accepted for the same reason, and
+    wins over the model's default: an explicit choice beats an inherited one.
     """
     chosen = None
     if model_id:
@@ -126,6 +134,12 @@ def _new_chat(
     # Left alone entirely on a plain chat, where it means nothing.
     if profile is not None and agent_mode.strip() in agent_policy.MODES:
         chat.agent_mode = agent_mode.strip()
+    # After the model's defaults, so choosing one on the new-chat screen wins
+    # over the administrator's. Empty means "whatever the model said", not
+    # "none" -- clearing it is what the blank option on an existing chat does.
+    wanted_effort = reasoning_effort.strip().lower()
+    if wanted_effort in chat_service.EFFORTS:
+        chat.params_json = {**chat.params_json, "reasoning_effort": wanted_effort}
     db.add(chat)
     db.commit()
     return chat
@@ -144,6 +158,7 @@ async def start_chat(
     ssh_profile_id: str = Form(""),
     project_dir: str = Form(""),
     agent_mode: str = Form(""),
+    reasoning_effort: str = Form(""),
 ) -> Response:
     """Create a chat from its first message.
 
@@ -166,6 +181,7 @@ async def start_chat(
         ssh_profile_id=ssh_profile_id,
         project_dir=project_dir,
         agent_mode=agent_mode,
+        reasoning_effort=reasoning_effort,
     )
 
     user_message = chat_service.create_message(db, chat, ROLE_USER, content)
@@ -400,6 +416,108 @@ async def compact_chat(request: Request, db: Db, user: RequiredUser, chat_id: st
     )
 
 
+@router.post("/{chat_id}/index")
+async def reindex_chat(db: Db, user: RequiredUser, chat_id: str) -> Response:
+    """Walk the project directory again, now.
+
+    The listing is cached for five minutes and only ever built when a reply
+    starts, so a tree that has just changed under somebody's hands -- a checkout,
+    a build, anything done in the terminal panel rather than through
+    `file_write` -- stays wrong until the next reply after the TTL lapses. This
+    is the "look again" that was missing.
+
+    Read-only, and therefore outside `agent/policy.py` for the reason the
+    directory browser is: it is LLeMbas acting on a person's instruction, not a
+    model choosing to look, and a listing that asked permission would be
+    useless. The gate is ownership of the connection, checked here rather than
+    trusted from the chat.
+    """
+    from lembas.db.models import SshProfile
+    from lembas.services.agent import index as index_service
+    from lembas.services.agent import ssh as ssh_service
+
+    chat = _owned_chat(db, chat_id, user.id)
+    if chat.kind != KIND_AGENT or not chat.ssh_profile_id:
+        raise HTTPException(status.HTTP_409_CONFLICT, "This chat has no project directory.")
+    if not permissions.has(db, user, "tools.agent"):
+        raise HTTPException(status.HTTP_403_FORBIDDEN, "You may not use agent connections.")
+
+    profile = db.get(SshProfile, chat.ssh_profile_id)
+    if profile is None or profile.owner_id != user.id or not profile.enabled:
+        raise HTTPException(status.HTTP_404_NOT_FOUND, "That connection is not available.")
+    if not profile.host_key:
+        raise HTTPException(
+            status.HTTP_409_CONFLICT,
+            "This connection's host key has not been accepted yet.",
+        )
+
+    project_dir = chat.project_dir or profile.default_dir or ""
+    try:
+        found = await index_service.ensure(
+            ssh_service.SshExecutor(ssh_service.spec_from(profile), project_dir),
+            profile.id,
+            project_dir,
+            refresh=True,
+        )
+    except Exception as exc:  # noqa: BLE001 - surfaced to the reader, not swallowed
+        log.warning("could not index %s for chat %s: %s", project_dir, chat.id, exc)
+        raise HTTPException(
+            status.HTTP_502_BAD_GATEWAY, "Could not read the project directory."
+        ) from exc
+
+    listed = len(found.paths)
+    return JSONResponse(
+        {
+            "ok": True,
+            "files": found.total,
+            "listed": listed,
+            "truncated": found.truncated,
+            "message": (
+                f"{found.total} files under {project_dir or '~'}"
+                + (f", {listed} listed." if listed != found.total else ".")
+            ),
+        }
+    )
+
+
+@router.post("/{chat_id}/bases")
+async def attach_base(
+    request: Request, db: Db, user: RequiredUser, chat_id: str, base_id: str = Form("")
+) -> Response:
+    """Scope this chat to a knowledge base, from the `@` menu.
+
+    A base is a *reference*, not an attachment: `Chat.knowledge_bases` already
+    narrows `knowledge_search`, and the harness already names the attached bases
+    so the model can tell "there is nothing about this" from "I can only see
+    this folder". Copying a folder of documents into the window instead would
+    cost the context on every request forever to answer one question.
+
+    Additive, and idempotent -- choosing the same base twice is not an error.
+    Removing one is a checkbox in the chat's settings, where the whole set is
+    visible at once.
+    """
+    from lembas.db.models import KnowledgeBase
+    from lembas.services.library import documents as documents_service
+
+    chat = _owned_chat(db, chat_id, user.id)
+    if not permissions.has(db, user, "library.use"):
+        raise HTTPException(status.HTTP_403_FORBIDDEN, "You may not use the library.")
+
+    base = db.scalar(
+        documents_service.visible_bases(db, user).where(KnowledgeBase.id == base_id)
+    )
+    if base is None:
+        raise HTTPException(status.HTTP_404_NOT_FOUND, "That knowledge base is not available.")
+
+    if base.id not in {b.id for b in chat.knowledge_bases}:
+        chat.knowledge_bases = [*chat.knowledge_bases, base]
+        db.commit()
+
+    return templates.TemplateResponse(
+        request, "chat/_base_chip.html", {"request": request, "base": base}
+    )
+
+
 @router.post("/{chat_id}/keep")
 async def keep_chat(db: Db, user: RequiredUser, chat_id: str) -> Response:
     """Stop a temporary chat being temporary.
@@ -489,6 +607,22 @@ async def post_message(
     return _send(request, db, chat, user, content, file_ids=file_ids)
 
 
+def _reply_in_flight(db: DBSession, chat: Chat) -> bool:
+    """Whether this chat already has a reply being written.
+
+    The row is the authority, not the registry: a restart leaves an incomplete
+    assistant message behind with no `Generation` anywhere, and that row is what
+    starts the reply again on the next page load. The registry is consulted too,
+    for the sliver in which a generation is still running and its row has
+    already been written -- `_persist` sets `complete` before `_run` sets
+    `done`.
+    """
+    unfinished = db.scalar(
+        select(Message).where(Message.chat_id == chat.id, Message.complete.is_(False))
+    )
+    return unfinished is not None or generation_service.running_for(chat.id) is not None
+
+
 def _note_rewind(chat: Chat) -> None:
     """Record that an agent chat's transcript went back and the machine did not.
 
@@ -516,11 +650,53 @@ def _send(
     Shared by the composer and by anything else that puts words into a
     conversation on somebody's behalf -- carrying out a plan, for one. One path
     rather than two, so a second way of sending cannot drift from the first.
+
+    If a reply is already being written, the turn is *queued* instead: written,
+    shown, and not sent. Starting a second reply here is what used to happen,
+    and it produced two generations answering the same chat from two different
+    prefixes of it, with Stop pointing at whichever bubble came first in the
+    document.
     """
-    user_message = chat_service.create_message(db, chat, ROLE_USER, content)
+    if queued := _reply_in_flight(db, chat):
+        waiting = db.scalar(
+            select(func.count())
+            .select_from(Message)
+            .where(Message.chat_id == chat.id, Message.queued.is_(True))
+        )
+        if waiting >= MAX_QUEUED:
+            raise HTTPException(
+                status.HTTP_409_CONFLICT,
+                f"There are already {MAX_QUEUED} messages waiting to be sent.",
+            )
+
+    user_message = chat_service.create_message(db, chat, ROLE_USER, content, queued=queued)
     if file_ids:
         files_service.claim(db, ids=file_ids, user_id=user.id, message_id=user_message.id)
         db.refresh(user_message)
+
+    if queued:
+        # One bubble and no assistant placeholder. The streaming shell is the
+        # only thing that starts a generation, so a placeholder here would be a
+        # second concurrent reply -- exactly what the queue exists to prevent.
+        if (live := generation_service.running_for(chat.id)) is not None:
+            # So a reply between two rounds of tool calls notices it, and so
+            # anybody following sees the status change.
+            live.touch()
+        return templates.TemplateResponse(
+            request,
+            "chat/_message.html",
+            {
+                "request": request,
+                "message": user_message,
+                "chat": chat,
+                "user": user,
+                "models_by_id": {
+                    m.model_id: m for m in chat_service.available_models(db, user)
+                },
+                **audio_service.template_flags(db, user),
+            },
+        )
+
     assistant_message = chat_service.create_message(
         db, chat, ROLE_ASSISTANT, "", complete_=False, model_id=chat.model_id
     )
@@ -680,10 +856,90 @@ async def _follow(chat_id: str, message_id: str) -> AsyncIterator[str]:
         )
         title_html = templates.get_template("chat/_title_oob.html").render({"chat": chat})
 
-    yield sse.event("done", final_html + title_html)
+        # What the queue did while this reply was running. There is no push
+        # channel that outlives one message's stream, and this is the last frame
+        # that reaches the browser -- so it carries the rest out of band, the
+        # way the chat title already does.
+        moved_html, queue_html = _queue_frames(db, chat, owner, generation)
+
+    yield sse.event("done", moved_html + final_html + queue_html + title_html)
     yield sse.event("close", "")
 
 
+def _render_bubble(db: DBSession, chat: Chat, owner: User | None, message: Message) -> str:
+    """One finished bubble, rendered the way the `done` frame renders its own."""
+    return templates.get_template("chat/_message.html").render(
+        {
+            "message": message,
+            "body_html": (
+                render_markdown(message.content) if message.role == ROLE_ASSISTANT else ""
+            ),
+            "chat": chat,
+            "user": owner,
+            "models_by_id": {m.model_id: m for m in chat_service.available_models(db, None)},
+            **audio_service.template_flags(db, owner),
+        }
+    )
+
+
+def _queue_frames(
+    db: DBSession, chat: Chat, owner: User | None, generation
+) -> tuple[str, str]:
+    """The bubbles the queue produced during this reply, as out-of-band HTML.
+
+    Two pieces, because they swap differently. Anything taken *into* this reply
+    mid-round now sorts before it, so it is rendered ahead of the finished
+    bubble in the same `outerHTML` swap and its stale node is deleted out of
+    band -- one frame, and the DOM ends up in the order the database is in.
+    Anything drained *after* the reply is a new pair appended to the thread.
+    """
+    moved: list[str] = []
+    out_of_band: list[str] = []
+
+    for injected_id in generation.injected_ids:
+        row = db.get(Message, injected_id)
+        if row is None:
+            continue
+        moved.append(_render_bubble(db, chat, owner, row))
+        # Removed where it was; it is about to reappear above the reply.
+        out_of_band.append(f'<article id="msg-{row.id}" hx-swap-oob="delete"></article>')
+
+    if generation.drained:
+        fresh = list(
+            db.scalars(
+                select(Message)
+                .where(Message.chat_id == chat.id, Message.complete.is_(False))
+                .order_by(Message.created_at)
+            )
+        )
+        for assistant in fresh:
+            # The user turn that was waiting has just lost its Send now and
+            # Discard, so it is re-rendered in place.
+            delivered = db.scalars(
+                select(Message)
+                .where(
+                    Message.chat_id == chat.id,
+                    Message.role == ROLE_USER,
+                    Message.created_at <= assistant.created_at,
+                )
+                .order_by(Message.created_at.desc())
+                .limit(1)
+            ).first()
+            if delivered is not None:
+                out_of_band.append(
+                    f'<div hx-swap-oob="outerHTML:#msg-{delivered.id}">'
+                    + _render_bubble(db, chat, owner, delivered)
+                    + "</div>"
+                )
+            out_of_band.append(
+                '<div hx-swap-oob="beforeend:#thread">'
+                + _render_bubble(db, chat, owner, assistant)
+                + "</div>"
+            )
+
+    return "".join(moved), "".join(out_of_band)
+
+
 def _metrics_html(generation) -> str:
     """The metric chips for a reply still being written.
 
@@ -799,6 +1055,16 @@ async def edit_message(
     if not content and not message.attachments:
         raise HTTPException(status.HTTP_400_BAD_REQUEST, "A message cannot be empty.")
 
+    # Editing rewinds and then starts a reply, unconditionally. Doing that while
+    # one is already being written is a second concurrent generation -- the
+    # thing the queue exists to prevent -- reachable here by a button that is on
+    # screen throughout. It was reachable before the queue too; nothing made it
+    # obvious.
+    if _reply_in_flight(db, chat):
+        raise HTTPException(
+            status.HTTP_409_CONFLICT, "Wait for the current reply to finish, or stop it."
+        )
+
     message.content = content
 
     # Attachments cascade with their message, so the files go too.
@@ -884,6 +1150,67 @@ async def stop_message(db: Db, user: RequiredUser, chat_id: str, message_id: str
     return Response(status_code=status.HTTP_204_NO_CONTENT)
 
 
+def _waiting_message(db: DBSession, chat: Chat, message_id: str) -> Message:
+    message = db.get(Message, message_id)
+    if message is None or message.chat_id != chat.id or not message.queued:
+        raise HTTPException(
+            status.HTTP_404_NOT_FOUND, "That message is not waiting to be sent."
+        )
+    return message
+
+
+@router.post("/{chat_id}/messages/{message_id}/discard")
+async def discard_queued(db: Db, user: RequiredUser, chat_id: str, message_id: str) -> Response:
+    """Withdraw a prompt that has not been sent.
+
+    Deleted outright rather than marked: it never reached a model, nothing in
+    the transcript refers to it, and a conversation full of tombstones for
+    things nobody said is worse than the row being gone. Attachments cascade.
+
+    An empty body rather than a 204, because htmx does not swap on a 204 and the
+    bubble has to disappear.
+    """
+    chat = _owned_chat(db, chat_id, user.id)
+    message = _waiting_message(db, chat, message_id)
+
+    db.delete(message)
+    db.commit()
+    return HTMLResponse("")
+
+
+@router.post("/{chat_id}/messages/{message_id}/send-now")
+async def send_queued_now(
+    request: Request, db: Db, user: RequiredUser, chat_id: str, message_id: str
+) -> Response:
+    """Deliver a waiting prompt at once.
+
+    Refused while a reply is being written rather than allowed to jump ahead of
+    it: that is what the queue *is*, and starting a second generation here is
+    the thing this whole mechanism exists to stop. Stop the reply first.
+
+    The whole thread comes back, which is the rewind and compaction idiom, and
+    is safe only because of the refusal above -- there is no live bubble to
+    destroy.
+    """
+    chat = _owned_chat(db, chat_id, user.id)
+    message = _waiting_message(db, chat, message_id)
+    if _reply_in_flight(db, chat):
+        raise HTTPException(
+            status.HTTP_409_CONFLICT, "Wait for the current reply to finish, or stop it."
+        )
+
+    message.queued = False
+    db.commit()
+    assistant = chat_service.create_message(
+        db, chat, ROLE_ASSISTANT, "", complete_=False, model_id=chat.model_id
+    )
+    generation_service.ensure(chat.id, assistant.id)
+
+    return templates.TemplateResponse(
+        request, "chat/_thread.html", {"request": request, **_thread_context(db, chat, user)}
+    )
+
+
 @router.post("/{chat_id}/interaction/{interaction_id}")
 async def answer_interaction(
     request: Request,

src/lembas/api/files.py

@@ -16,14 +16,17 @@ from fastapi import (
     status,
 )
 from fastapi.responses import FileResponse
+from sqlalchemy import select
 
 from lembas.api.deps import Db, RequiredUser, require_permission
-from lembas.db.models import Attachment, Document
+from lembas.db.models import Attachment, Document, KnowledgeBase, Note
 from lembas.security import permissions
 from lembas.services import files as files_service
 from lembas.services import settings_store
 from lembas.services.fetch import FetchError, fetch
 from lembas.services.library import documents as documents_service
+from lembas.services.library import notes as notes_service
+from lembas.services.library import skills as skills_service
 from lembas.web.templating import templates
 
 log = logging.getLogger(__name__)
@@ -144,6 +147,102 @@ async def attach_from_knowledge(
     )
 
 
+def _chip(request: Request, attachment: Attachment) -> Response:
+    return templates.TemplateResponse(
+        request, "chat/_attachment_chip.html", {"request": request, "attachment": attachment}
+    )
+
+
+def _not_available(request: Request, what: str) -> Response:
+    return templates.TemplateResponse(
+        request,
+        "chat/_attachment_error.html",
+        {"request": request, "filename": what, "error": f"That {what} is not available."},
+    )
+
+
+@router.post("/from-note", dependencies=[Depends(require_permission("files.upload"))])
+async def attach_from_note(
+    request: Request, db: Db, user: RequiredUser, note_id: str = Form(""), chat_id: str = Form("")
+) -> Response:
+    """Attach a note the model wrote earlier.
+
+    A copy, like every other attach path: a note is edited far more often than a
+    document, and a transcript that changes underneath itself because somebody
+    tidied a note later is the thing all of this is arranged to prevent.
+    """
+    note = notes_service.get(db, note_id, user)
+    if note is None:
+        return _not_available(request, "note")
+
+    return _chip(
+        request,
+        files_service.store_text(
+            db,
+            user_id=user.id,
+            chat_id=chat_id or None,
+            filename=f"{note.title or 'note'}.txt",
+            text=note.body,
+            source_path=note.title or "",
+            source_label="Note",
+        ),
+    )
+
+
+@router.post("/from-skill", dependencies=[Depends(require_permission("files.upload"))])
+async def attach_from_skill(
+    request: Request, db: Db, user: RequiredUser, skill_id: str = Form(""), chat_id: str = Form("")
+) -> Response:
+    """Hand a skill over directly, rather than hoping the model fetches it.
+
+    The index of enabled skills is already in the harness and `skill_get` pulls
+    a body on demand -- but only if the model decides to. `@` is the reader
+    saying "use this one", which is a different act and deserves a way to say it.
+    """
+    skill = skills_service.get(db, skill_id, user)
+    if skill is None:
+        return _not_available(request, "skill")
+
+    return _chip(
+        request,
+        files_service.store_text(
+            db,
+            user_id=user.id,
+            chat_id=chat_id or None,
+            filename=f"{skill.name}.md",
+            text=skill.body,
+            source_path=skill.name,
+            source_label="Skill",
+        ),
+    )
+
+
+@router.post("/from-attachment", dependencies=[Depends(require_permission("files.upload"))])
+async def attach_from_attachment(
+    request: Request,
+    db: Db,
+    user: RequiredUser,
+    attachment_id: str = Form(""),
+    chat_id: str = Form(""),
+) -> Response:
+    """Point at something already in this conversation, without uploading again.
+
+    Copied rather than referenced, like everything else here -- an attachment
+    belongs to the message it was sent with, and two messages sharing one row
+    would make deleting either of them a question rather than an answer.
+    """
+    original = db.get(Attachment, attachment_id)
+    if original is None or original.user_id != user.id:
+        return _not_available(request, "attachment")
+
+    return _chip(
+        request,
+        files_service.copy_attachment(
+            db, user_id=user.id, chat_id=chat_id or None, attachment=original
+        ),
+    )
+
+
 @router.get("/knowledge-picker", dependencies=[Depends(require_permission("files.upload"))])
 async def knowledge_picker(
     request: Request, db: Db, user: RequiredUser, q: str = "", chat_id: str = ""
@@ -210,9 +309,14 @@ async def mention_picker(
                 ][:20]
 
     documents: list = []
+    notes: list = []
+    skills: list = []
+    bases: list = []
     if permissions.has(db, user, "library.use"):
         if needle:
             documents = documents_service.search(db, user, q, limit=10)
+            notes = notes_service.search(db, user, q, limit=5)
+            skills = skills_service.search(db, user, q, limit=5)
         else:
             documents = list(
                 db.scalars(
@@ -221,6 +325,48 @@ async def mention_picker(
                     .limit(10)
                 )
             )
+            notes = list(
+                db.scalars(
+                    notes_service.visible(db, user).order_by(Note.updated_at.desc()).limit(5)
+                )
+            )
+            skills = list(db.scalars(skills_service.visible(db, user).limit(5)))
+
+        # A whole base is a *reference*, not a copy: attaching one scopes the
+        # chat to it and the model searches inside it. Dumping the contents of
+        # a folder of contracts into the window would be the wrong shape
+        # entirely, and `Chat.knowledge_bases` already means exactly this.
+        # Only in an existing chat, because there is nothing to attach it to
+        # before one exists -- the same reason project files are absent there.
+        if chat_id:
+            bases = [
+                base
+                for base in db.scalars(
+                    documents_service.visible_bases(db, user).order_by(KnowledgeBase.name)
+                )
+                if not needle or needle in base.name.lower()
+            ][:5]
+
+    # A URL typed after `@` is a page to read, not a name to look up. The
+    # fetcher, its SSRF guard and its HTML-to-text already live behind
+    # `/api/files/link`; this only offers it.
+    website = q.strip() if q.strip().lower().startswith(("http://", "https://")) else ""
+
+    attachments: list = []
+    if chat_id and needle:
+        attachments = list(
+            db.scalars(
+                select(Attachment)
+                .where(
+                    Attachment.user_id == user.id,
+                    Attachment.chat_id == chat_id,
+                    Attachment.message_id.is_not(None),
+                )
+                .order_by(Attachment.created_at.desc())
+                .limit(20)
+            )
+        )
+        attachments = [a for a in attachments if needle in a.filename.lower()][:5]
 
     return templates.TemplateResponse(
         request,
@@ -230,6 +376,11 @@ async def mention_picker(
             "user": user,
             "files": files,
             "documents": documents,
+            "notes": notes,
+            "skills": skills,
+            "bases": bases,
+            "attachments": attachments,
+            "website": website,
             "q": q,
             "chat_id": chat_id,
             "profile_id": profile_id,

src/lembas/db/models/chat.py

@@ -223,6 +223,13 @@ class Message(UUIDPrimaryKey, Timestamps, Base):
     # True when the reader pressed Stop. Distinct from `error`: the text that
     # did arrive is kept and is perfectly usable, it is just cut short.
     stopped: Mapped[bool] = mapped_column(Boolean, default=False, nullable=False)
+    # Typed while a reply was still being written, and not yet handed to a
+    # model. A row rather than something held in the browser: it survives a
+    # restart, it is in the transcript the moment it is typed, and it can be
+    # withdrawn before it is ever sent. `build_messages` skips it; delivery --
+    # `generation._drain` at the end of a reply, or `_inject` between two rounds
+    # of tool calls -- is the only thing that clears it.
+    queued: Mapped[bool] = mapped_column(Boolean, default=False, nullable=False)
 
     chat: Mapped[Chat] = relationship(back_populates="messages")
     attachments: Mapped[list[Attachment]] = relationship(  # noqa: F821

src/lembas/services/agent/index.py

@@ -223,12 +223,21 @@ async def build(executor: Executor, project_dir: str) -> ProjectIndex:
     """Walk the directory, by whichever means works first."""
     started = time.monotonic()
     try:
+        found = None
         for attempt in (_from_git, _from_find):
-            found = await attempt(executor, project_dir)
+            try:
+                found = await attempt(executor, project_dir)
+            except ExecError as exc:
+                # A rung that cannot run at all is a rung that did not answer,
+                # not the end of the ladder. A host that refuses exec entirely
+                # -- an SFTP-only account, a forced command -- is the exact case
+                # the SFTP rung below exists for, and letting this out skipped
+                # straight past it to an empty listing.
+                log.debug("indexing %s: %s did not run: %s", project_dir, attempt.__name__,
+                          exc.message)
+                found = None
             if found is not None:
                 break
-        else:
-            found = None
         if found is None:
             found = await _from_sftp(executor, project_dir)
     except ExecError as exc:
@@ -491,6 +500,17 @@ def forget(profile_id: str) -> int:
     return len(doomed)
 
 
+def forget_dir(profile_id: str, project_dir: str) -> None:
+    """Drop one tree's listing, because something just changed it.
+
+    The TTL exists for drift nobody can see coming. A write through `file_write`
+    is not that: it is this process changing the tree it has just described, and
+    leaving five minutes of a listing that is known to be wrong is worse than
+    having none -- a model reading it concludes the file it created is missing.
+    """
+    _CACHE.pop((profile_id, project_dir), None)
+
+
 def clear() -> None:
     _CACHE.clear()
 
@@ -503,5 +523,6 @@ __all__ = [
     "clear",
     "ensure",
     "forget",
+    "forget_dir",
     "render",
 ]

src/lembas/services/agent/session.py

@@ -36,6 +36,10 @@ class AgentContext:
     chat_id: str
     label: str
     project_dir: str
+    # The connection's id, carried so a runner can drop the project listing it
+    # has just invalidated. `index` is keyed on the connection and the
+    # directory, not on the chat -- two chats on one tree share a listing.
+    profile_id: str = ""
     mode: str = policy.MODE_MANUAL
     allow: tuple[str, ...] = ()
     deny: tuple[str, ...] = ()
@@ -112,6 +116,7 @@ def resolve(db: DBSession, chat: Chat, user: User | None) -> AgentContext | None
         chat_id=chat.id,
         label=profile.label,
         project_dir=chat.project_dir or profile.default_dir or "",
+        profile_id=profile.id,
         mode=chat.agent_mode if chat.agent_mode in policy.MODES else policy.MODE_MANUAL,
         allow=tuple(values.get("allow_default") or ()),
         deny=tuple(values.get("deny_default") or ()),

src/lembas/services/agent/tools.py

@@ -21,7 +21,7 @@ import json
 import logging
 from typing import Any
 
-from lembas.services.agent import policy
+from lembas.services.agent import index, policy
 from lembas.services.agent.base import ExecError, ExecRequest
 from lembas.services.agent.session import AgentContext
 from lembas.services.tools import (
@@ -198,6 +198,13 @@ async def _run_write(context: ToolContext, args: dict[str, Any]) -> ToolOutcome:
             exc.message, _event("file_write", agent, path, status="error", error=exc.message)
         )
 
+    # The tree just changed, and this process is what changed it. The listing's
+    # TTL is for drift nobody can see coming; leaving five more minutes of a
+    # listing known to be wrong makes a model conclude the file it has just
+    # written does not exist.
+    if agent.profile_id:
+        index.forget_dir(agent.profile_id, agent.project_dir)
+
     return ToolOutcome(
         f"Wrote {written} bytes to {path}.",
         _event("file_write", agent, path, status="ok", text=f"{written} bytes"),

src/lembas/services/chat.py

@@ -226,6 +226,11 @@ def build_messages(
             message
         ) <= compaction_service.moment(cutoff):
             continue
+        # Typed while the previous reply was still being written, and not yet
+        # handed to a model. It is in the transcript and it is not in the
+        # request; delivery is what moves it from one to the other.
+        if message.queued:
+            continue
         # Skip turns that failed or produced nothing -- but a message carrying
         # only an attachment has no text and must still be sent.
         if message.error:
@@ -447,6 +452,7 @@ def create_message(
     *,
     complete_: bool = True,
     model_id: str = "",
+    queued: bool = False,
 ) -> Message:
     message = Message(
         chat_id=chat.id,
@@ -454,6 +460,7 @@ def create_message(
         content=content,
         complete=complete_,
         model_id=model_id,
+        queued=queued,
     )
     db.add(message)
     db.commit()

src/lembas/services/compaction.py

@@ -98,8 +98,12 @@ def split(
         return [], list(messages)
     boundary = moment(cutoff)
     return (
-        [m for m in messages if moment(m) <= boundary],
+        # A prompt still waiting to be sent stays on the live side whatever its
-        [m for m in messages if moment(m) > boundary],
+        # timestamp says. Folding one into the "earlier messages" details would
+        # hide the only place its Send now and Discard exist, and it has not
+        # been part of any request to summarise.
+        [m for m in messages if not m.queued and moment(m) <= boundary],
+        [m for m in messages if m.queued or moment(m) > boundary],
     )
 
 
@@ -133,6 +137,9 @@ def transcript(db: DBSession, chat: Chat, *, upto: Message) -> str:
         Message.chat_id == chat.id,
         Message.created_at <= upto.created_at,
         Message.error == "",
+        # Not yet sent to anything. Summarising it would fold words the model
+        # has never seen into the record, and then deliver them again later.
+        Message.queued.is_(False),
     )
     if previous is not None:
         query = query.where(Message.created_at > previous.created_at)

src/lembas/services/files.py

@@ -375,6 +375,43 @@ def store_text(
     return attachment
 
 
+def copy_attachment(
+    db: DBSession, *, user_id: str, chat_id: str | None, attachment: Attachment
+) -> Attachment:
+    """Duplicate something already sent, so it can ride along with a new message.
+
+    A copy and not a second reference to one row: an attachment belongs to the
+    message it was sent with, and sharing one between two would make deleting
+    either of them a question rather than an answer.
+    """
+    stored_name = ""
+    source = attachments_dir() / attachment.stored_name if attachment.stored_name else None
+    if source is not None and source.exists():
+        stored_name = f"{secrets.token_hex(16)}{Path(source.name).suffix}"
+        (attachments_dir() / stored_name).write_bytes(source.read_bytes())
+
+    copy = Attachment(
+        user_id=user_id,
+        chat_id=chat_id,
+        filename=attachment.filename,
+        stored_name=stored_name,
+        media_type=attachment.media_type,
+        size_bytes=attachment.size_bytes,
+        kind=attachment.kind,
+        width=attachment.width,
+        height=attachment.height,
+        extracted_text=attachment.extracted_text,
+        pages=attachment.pages,
+        truncated=attachment.truncated,
+        extraction_error=attachment.extraction_error,
+        source_path=attachment.source_path,
+        source_label=attachment.source_label,
+    )
+    db.add(copy)
+    db.commit()
+    return copy
+
+
 def copy_document(
     db: DBSession, *, user_id: str, chat_id: str | None, document
 ) -> Attachment:
@@ -407,6 +444,12 @@ def copy_document(
         pages=document.pages,
         truncated=document.truncated,
         extraction_error=document.extraction_error,
+        # Where it came from, for the same reason a project file carries it: a
+        # model handed four documents cannot tell which is which, and cannot
+        # name one back when asked to work on it. This was the one attach path
+        # that dropped provenance.
+        source_path=(document.title or "")[:1000],
+        source_label=(document.base.name if document.base else "Knowledge")[:200],
     )
     db.add(attachment)
     db.commit()

src/lembas/services/generation.py

@@ -129,6 +129,13 @@ class Generation:
     # the reply and is written onto the message, so the Execute button sends
     # exactly what was proposed rather than something parsed back out of prose.
     plan: dict | None = None
+    # The queue, seen from the reply's side. `drained` says this reply's ending
+    # handed the next waiting prompt to a fresh one; `injected_ids` names the
+    # prompts taken into *this* reply between two rounds of tool calls. Both are
+    # read only by `_follow`, which turns them into bubbles on the `done` frame
+    # -- the one frame that reaches a browser after a reply is over.
+    drained: bool = False
+    injected_ids: list[str] = field(default_factory=list)
 
     def touch(self) -> None:
         self.version += 1
@@ -187,6 +194,22 @@ def answer(
     return False
 
 
+def running_for(chat_id: str) -> Generation | None:
+    """The reply being written in this chat, if there is one.
+
+    A linear scan for the reason `answer` gives above: one entry per reply in
+    flight, consulted at human speed. `_prune` first, because a finished
+    generation lingers `KEEP_FINISHED` so that late followers still get the
+    final frames -- and without the sweep those five minutes would look like a
+    chat that is permanently busy, and queue everything typed into it.
+    """
+    _prune()
+    for generation in _RUNNING.values():
+        if generation.chat_id == chat_id and not generation.done:
+            return generation
+    return None
+
+
 _VERDICTS = (interaction.ALLOW, interaction.ALLOW_ALWAYS, interaction.DENY)
 
 
@@ -328,6 +351,12 @@ async def _run(generation: Generation) -> None:
 
             model = chat_service.model_for(db, chat)
             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")
 
         generation.prompt_estimate = tokens.estimate_request(payload)
 
@@ -406,10 +435,17 @@ async def _run(generation: Generation) -> None:
             if generation.stopped or not calls:
                 break
 
-            if round_number == tools_service.MAX_ROUNDS:
+            if round_number == budget:
                 # Out of rounds with the model still asking for tools. Recorded
                 # rather than silently dropped: an answer that stops here needs
                 # to be explicable.
+                #
+                # `budget`, not `MAX_ROUNDS`. The loop is sized by the budget on
+                # the line above and the message below has always reported it,
+                # but the comparison was against the global 3 -- so an agent
+                # chat allowed forty steps stopped after three and said it had
+                # taken forty. Two numbers, one of them wrong, in code whose
+                # whole job is to say what happened.
                 generation.tool_events.append(
                     {
                         "name": calls[0]["name"],
@@ -454,6 +490,22 @@ async def _run(generation: Generation) -> None:
                     generation.plan = outcome.event["plan"]
             generation.touch()
 
+            # Something typed while this reply was working. Taken in here, at a
+            # round boundary, rather than made to wait for the whole reply: an
+            # agent that has just finished one loop and is about to start
+            # another is exactly when "actually, do it the other way" is worth
+            # having.
+            #
+            # Only while there is a round left to answer in. Injecting into the
+            # last one would deliver the prompt into a reply that then runs out
+            # of budget without addressing it -- and it is marked delivered, so
+            # nothing would ever send it again. Below that line it waits for
+            # `_drain`, which always gives it a reply of its own.
+            if round_number + 1 < budget and (
+                added := _inject(generation, generation.chat_id, vision)
+            ):
+                messages.append(added)
+
             payload = {**payload, "messages": messages}
 
             # A plan ends the turn. One more request so the model can say what
@@ -525,6 +577,11 @@ async def _run(generation: Generation) -> None:
         # the row. The other order left a window in which the finished frame
         # showed the previous turn's stored values.
         _persist(generation, title, time.monotonic() - started)
+        # After the row is authoritative and before `done`, for the same reason
+        # `_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)
         generation.done = True
         generation.finished_at = datetime.now(UTC)
         generation.touch()
@@ -1006,6 +1063,117 @@ def _question_from(payload: dict) -> str:
     return ""
 
 
+def _next_waiting(db, chat_id: str) -> Message | None:
+    """The oldest prompt in this chat that has not been sent."""
+    return db.scalars(
+        select(Message)
+        .where(
+            Message.chat_id == chat_id,
+            Message.role == ROLE_USER,
+            Message.queued.is_(True),
+        )
+        .order_by(Message.created_at)
+        .limit(1)
+    ).first()
+
+
+def _drain(generation: Generation) -> None:
+    """Hand the next waiting prompt to a reply of its own, if there is one.
+
+    Exactly one, not all of them. Draining the lot would put two consecutive
+    user turns into the next request, which several local chat templates refuse
+    outright -- `build_messages` already goes to some trouble over that around
+    the compaction lead. "One after another" is also what was asked for: the
+    second waiting prompt is drained by the reply the first one starts, and so
+    on down the chain.
+
+    Three refusals, and none of them is a special case:
+
+    - **Superseded.** The same test `_persist` makes, for the same reason: a
+      regeneration cancels its predecessor and the predecessor's `finally:`
+      still runs. Without this, regenerating would drain the queue *and* leave
+      a third generation running.
+    - **Stopped.** Stop means stop, and the queue stays visible and
+      undelivered with Send now beside it. This is also what makes shutdown
+      safe -- cancellation sets `stopped`, so a restart never fires off a reply
+      with nobody watching.
+    - **Errored.** The endpoint has just failed. Feeding the next prompt into it
+      produces a second failure and spends somebody's words to do it.
+    """
+    owner = _RUNNING.get(generation.message_id)
+    if owner is not None and owner is not generation:
+        return
+    if generation.stopped or generation.error:
+        return
+
+    try:
+        with session_scope() as db:
+            chat = db.get(Chat, generation.chat_id)
+            if chat is None:
+                return
+            waiting = _next_waiting(db, chat.id)
+            if waiting is None:
+                return
+
+            waiting.queued = False
+            assistant = chat_service.create_message(
+                db, chat, ROLE_ASSISTANT, "", complete_=False, model_id=chat.model_id
+            )
+            chat_id, assistant_id = chat.id, assistant.id
+    except Exception:  # noqa: BLE001 - the reply is over either way
+        log.exception("could not drain the queue for chat %s", generation.chat_id)
+        return
+
+    # Outside the session: this starts a task, and a task is not something to
+    # hold a database session open across.
+    ensure(chat_id, assistant_id)
+    generation.drained = True
+
+
+def _inject(generation: Generation, chat_id: str, vision: bool) -> dict | None:
+    """Take the oldest waiting prompt into this reply, between two rounds.
+
+    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
+    way round would ask the same question twice and let an agent act on it
+    twice, which is not.
+
+    Sent verbatim, in the user role, with no framing. Everything else this
+    codebase injects is quoted and attributed because it came out of a file, a
+    page or a machine; this one genuinely *is* the person at the keyboard,
+    authenticated by the session cookie and stored as a `Message` whose role
+    says so. Wrapping it would teach a model that a user turn can be a
+    quotation, which is the exact distinction the other two rely on. What the
+    model needs -- that this can happen at all -- is one sentence in the
+    harness, where authored wording lives.
+    """
+    try:
+        with session_scope() as db:
+            waiting = _next_waiting(db, chat_id)
+            if waiting is None:
+                return None
+
+            waiting.queued = False
+            entry = chat_service.message_payload(waiting, vision=vision)
+            # The reply that answers it must sort *before* it, or the next
+            # turn's transcript reads "answer, then the question it answered"
+            # and a small model dutifully answers again. Moving the placeholder
+            # rather than the prompt keeps several interjections in the order
+            # they were typed.
+            placeholder = db.get(Message, generation.message_id)
+            if placeholder is not None:
+                placeholder.created_at = datetime.now(UTC)
+            generation.injected_ids.append(waiting.id)
+    except Exception:  # noqa: BLE001 - a lost interjection is not a failed reply
+        log.exception("could not take a queued prompt into chat %s", chat_id)
+        return None
+
+    generation.status = "Taking in what you just added…"
+    generation.touch()
+    return entry
+
+
 def _persist(generation: Generation, title: str, elapsed: float) -> None:
     """Write the finished reply, name the chat, and set the unread flag.
 

src/lembas/services/prompts.py

@@ -587,6 +587,22 @@ BUILTIN: tuple[Fragment, ...] = (
             "within that budget: two careful searches beat six that run out halfway."
         ),
     ),
+    Fragment(
+        key="core.interjection",
+        label="Being interrupted",
+        group=GROUP_CORE,
+        order=115,
+        when_tools=True,
+        hint="A message typed while you are working is handed to you between two "
+        "rounds of tool calls. Without this a model reads it as a fresh "
+        "conversation and starts the whole task again.",
+        default=(
+            "A new message from the person you are working for can arrive between "
+            "rounds of tool calls, while you are still working. Take it into account "
+            "from that point on. You do not need to start again or to re-explain what "
+            "you have already done — carry on, adjusted."
+        ),
+    ),
     Fragment(
         key="core.no_replay",
         label="Results are not kept",

src/lembas/web/static/css/chat.css

@@ -576,6 +576,24 @@
 .msg:focus-within .msg__actions { opacity: 1; }
 .msg__actions .is-copied { color: var(--success); }
 
+/* --- A turn that is waiting to be sent ------------------------------------
+   Its actions do not fade in on hover like the others: they are the only way
+   to withdraw something that has not happened yet, and a control you have to
+   find by hovering is one somebody will not find. */
+.msg--queued { opacity: 0.75; }
+.msg--queued .msg__body {
+  border-inline-start: 2px dashed var(--border-strong);
+  padding-inline-start: var(--sp-2);
+}
+.msg__actions--queued { opacity: 1; align-items: center; }
+.msg__note {
+  display: inline-flex;
+  align-items: center;
+  gap: var(--sp-1);
+  color: var(--ink-muted);
+  font-size: var(--text-xs);
+}
+
 /* --- Rendered Markdown ---------------------------------------------------- */
 .msg__body > :first-child { margin-top: 0; }
 .msg__body > :last-child { margin-bottom: 0; }
@@ -761,16 +779,32 @@
 .composer__mirror .tok-mention,
 .composer__mirror .tok-command {
   border-radius: var(--radius-sm);
-  /* Bled sideways so the rectangle does not sit hard against the next word,
+  /* Restated rather than inherited. `color: transparent` on the mirror is an
-     and the negative margin keeps the text metrics identical. */
+     inherited value, and a colour the span declares itself beats it -- which is
-  padding: 0 2px;
+     exactly what the transcript's rule below used to do from across the file,
-  margin: 0 -2px;
+     painting the token in accent-coloured mono at 0.95em on top of the
+     textarea's own text. Doubled, and shifted from there on, because the
+     metrics differ. The font must be restated for the same reason. */
+  color: transparent;
+  font: inherit;
+  /* Bled sideways by a shadow, not by padding: a rectangle that spreads cannot
+     move a glyph, and the negative margin that used to do this was the only
+     thing in the mirror that could. */
+  padding: 0;
+  margin: 0;
+}
+.composer__mirror .tok-mention {
+  background: var(--accent-soft);
+  box-shadow: 0 0 0 2px var(--accent-soft);
+}
+.composer__mirror .tok-command {
+  background: var(--leaf-soft);
+  box-shadow: 0 0 0 2px var(--leaf-soft);
 }
-.composer__mirror .tok-mention { background: var(--accent-soft); }
-.composer__mirror .tok-command { background: var(--leaf-soft); }
 
-/* The same two in a sent message, where they are text rather than a backdrop. */
+/* The same two in a sent message, where they are text rather than a backdrop --
-.tok-mention {
+   and scoped to it, because unscoped they also matched the mirror's spans. */
+.msg .tok-mention {
   border-radius: var(--radius-sm);
   padding: 0 2px;
   background: var(--accent-soft);

src/lembas/web/static/js/commands.js

@@ -75,7 +75,11 @@
       name: "effort",
       summary: "How hard a reasoning model should think",
       argument: "low | medium | high",
-      when: function () { return !!chat() && !!el("[data-effort]"); },
+      /* Offered wherever there is a model, not only where the control is.
+         `available()` filters `find()` and `run()` as well as the menu, so a
+         command hidden here is not merely unlisted -- typing it in full stops
+         being a command and gets sent as a message. Better to answer. */
+      when: function () { return !!el('[name="model_id"]'); },
       run: function (rest) { setEffort(rest); }
     },
     {
@@ -115,6 +119,12 @@
           });
       }
     },
+    {
+      name: "index",
+      summary: "Read the project directory again",
+      when: function () { return !!chat() && isAgent(); },
+      run: function () { reindex(); }
+    },
     {
       name: "terminal",
       summary: "Show or hide the terminal",
@@ -163,6 +173,32 @@
     return function () { window.location = url; };
   }
 
+  /* --- Reading the project directory again --------------------------------
+     The listing is cached for five minutes and only ever built when a reply
+     starts, so anything done in the terminal panel -- a checkout, a build --
+     is invisible to it until then. This is the "look again now". */
+  var indexing = false;
+
+  function reindex() {
+    if (indexing) return note("Already reading the project directory.");
+    indexing = true;
+    note("Reading the project directory…");
+    post("/api/chats/" + chat() + "/index")
+      .then(function (response) {
+        return response.json().then(function (body) {
+          return { ok: response.ok, body: body };
+        });
+      })
+      .then(function (result) {
+        if (!result.ok) {
+          return note(result.body.detail || "Could not read the project directory.", "error");
+        }
+        note(result.body.message);
+      })
+      .catch(function () { note("Could not read the project directory.", "error"); })
+      .finally(function () { indexing = false; });
+  }
+
   /* --- Reasoning effort ---------------------------------------------------
      The command drives the same select the composer shows, so there is one
      piece of state and the control updates itself when the command is used. */
@@ -171,7 +207,11 @@
   function setEffort(rest) {
     var select = el("[data-effort]");
     if (!select) {
-      return note("This model is not marked as a reasoning model.", "error");
+      return note(
+        "This model is not marked as a reasoning model, so effort would do " +
+          "nothing. An administrator can mark it on the model's page.",
+        "error"
+      );
     }
     var wanted = (rest || "").trim().toLowerCase();
     if (!wanted) {

src/lembas/web/static/js/composer.js

@@ -273,6 +273,24 @@
     } else if (option.dataset.mentionKnowledge) {
       body.append("document_id", option.dataset.mentionKnowledge);
       attach("/api/files/from-knowledge", body);
+    } else if (option.dataset.mentionNote) {
+      body.append("note_id", option.dataset.mentionNote);
+      attach("/api/files/from-note", body);
+    } else if (option.dataset.mentionSkill) {
+      body.append("skill_id", option.dataset.mentionSkill);
+      attach("/api/files/from-skill", body);
+    } else if (option.dataset.mentionAttachment) {
+      body.append("attachment_id", option.dataset.mentionAttachment);
+      attach("/api/files/from-attachment", body);
+    } else if (option.dataset.mentionUrl) {
+      body.append("url", option.dataset.mentionUrl);
+      attach("/api/files/link", body);
+    } else if (option.dataset.mentionBase) {
+      /* Not an attachment: nothing is copied, and what changes is what this
+         chat is allowed to search. It goes on the chat, so the route is the
+         chat's own. */
+      body.append("base_id", option.dataset.mentionBase);
+      attach("/api/chats/" + where.chatId + "/bases", body);
     }
   }
 

src/lembas/web/static/js/terminal.js

@@ -35,7 +35,11 @@
   /* Whether this shell tells us where commands begin and end -- "live",
      "loading" or "none". Everything the three buttons do keys off it. */
   var integration = "loading";
-  var autoSend = false;
+  /* "off" | "copy" | "send". Three states rather than a boolean, because the
+     old one did the wrong one of them: it appended into the composer, on top of
+     whatever was being typed there. A select rather than a cycling button --
+     a button cannot say which of three states it is in. */
+  var autoMode = "off";
   var lastCommand = null;
 
   function say(text, isError) {
@@ -169,8 +173,12 @@
          and the buttons fetch what they need when they are pressed. */
       if (integration !== "live") { integration = "live"; applyIntegration(); }
       showLast(payload.command);
-      if (autoSend) {
+      if (autoMode !== "off") {
-        capture(true).then(function (text) { intoComposer(text, true); });
+        capture(true).then(function (text) {
+          if (!text) return;
+          if (autoMode === "copy") return intoComposer(text, true);
+          sendStraightToChat(text);
+        });
       }
       return;
     }
@@ -287,10 +295,10 @@
     var usable = integration === "live";
     auto.disabled = !usable;
     auto.title = usable
-      ? "Attach every command you run to your next message"
+      ? "What to do with each command you run"
       : "This shell did not load LLeMbas's command markers, so there is no way " +
         "to tell where one command's output ends.";
-    if (!usable && autoSend) setAuto(false);
+    if (!usable && autoMode !== "off") setAuto("off");
   }
 
   function showLast(command) {
@@ -301,16 +309,32 @@
     slot.textContent = lastCommand ? lastCommand.summary : "";
   }
 
-  function setAuto(on) {
+  var AUTO_SAID = {
-    autoSend = !!on;
+    off: "Commands are no longer attached automatically.",
-    var button = panel.querySelector("[data-terminal-auto]");
+    copy: "Every command you run will be put into the message box.",
-    if (button) {
+    send: "Every command you run will be sent as a message on its own."
-      button.setAttribute("aria-pressed", autoSend ? "true" : "false");
+  };
-      button.classList.toggle("is-active", autoSend);
+
-    }
+  function setAuto(mode) {
-    say(autoSend
+    autoMode = AUTO_SAID[mode] ? mode : "off";
-      ? "Every command you run will be attached to your next message."
+    var select = panel && panel.querySelector("[data-terminal-auto]");
-      : "Commands are no longer attached automatically.");
+    if (select && select.value !== autoMode) select.value = autoMode;
+    say(AUTO_SAID[autoMode]);
+  }
+
+  /* Sent, not typed. The composer is left entirely alone -- somebody may be
+     half-way through a sentence in it, and overwriting that is the complaint
+     this replaces. The thread receives whatever the server decides the message
+     is: a streaming pair, or a single queued bubble if a reply is already being
+     written. Nothing here needs to know which. */
+  function sendStraightToChat(text) {
+    var url = panel.dataset.url.replace(/\/terminal\/ws$/, "/messages");
+    if (!window.htmx) return;
+    window.htmx.ajax("POST", url, {
+      target: "#thread",
+      swap: "beforeend",
+      values: { content: text }
+    });
   }
 
   /* A selection always wins, in every state. People rely on it, and it is the
@@ -427,10 +451,11 @@
         event.preventDefault();
         return copyToClipboard();
       }
-      if (event.target.closest("[data-terminal-auto]")) {
+    });
-        event.preventDefault();
+
-        return setAuto(!autoSend);
+    panel.addEventListener("change", function (event) {
-      }
+      var select = event.target.closest("[data-terminal-auto]");
+      if (select) setAuto(select.value);
     });
 
     /* xterm holds colours as values, not as variables, so a theme change has

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

@@ -0,0 +1,19 @@
+{% from "_macros.html" import icon %}
+{#
+  A knowledge base attached from the `@` menu.
+
+  Deliberately not an attachment chip: nothing was copied and there is no
+  `file_ids` input to submit. The base is already on the chat, and what it does
+  is narrow what `knowledge_search` may see -- so this says so, and says it in
+  the present tense, because it is already in force before the message is sent.
+
+  No remove button. Removing one is a checkbox in the chat's settings, where
+  the whole set is visible at once rather than only whichever was added last.
+#}
+<div class="attach-chip attach-chip--base" id="base-chip-{{ base.id }}">
+  <span class="attach-chip__icon">{{ icon("archive", "icon--sm") }}</span>
+  <span class="attach-chip__body">
+    <span class="attach-chip__name" title="{{ base.name }}">{{ base.name }}</span>
+    <span class="attach-chip__meta">This chat now searches only the bases it is attached to</span>
+  </span>
+</div>

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

@@ -215,10 +215,16 @@
            they were taking a slot in a row that has work to do.
 
            Its own form: nesting one inside the composer's form is invalid HTML
-           and the browser drops the inner one. #}
+           and the browser drops the inner one.
+
+           The verb is on the select, not on that form. htmx binds a trigger to
+           the annotated element itself, and `change` fires here and bubbles
+           through this element's *ancestors* -- which a sibling form is not.
+           `form=` scopes the values, and only the values. #}
         <div class="composer__context">
           <select class="select select--sm" name="agent_mode" aria-label="Approval mode"
-                  form="agent-mode-form">
+                  form="agent-mode-form"
+                  hx-patch="/api/chats/{{ chat.id }}" hx-swap="none">
             {% for value, label, hint in agent_modes %}
             <option value="{{ value }}" title="{{ hint }}"
                     {{ 'selected' if value == chat.agent_mode }}>{{ label }}</option>
@@ -234,17 +240,27 @@
           Only on a model an administrator has marked as **reasoning**: that
           flag has existed since the beginning with no reader at all, and
           offering the control everywhere would be offering a setting that does
-          nothing almost everywhere. Its own form, for the reason the mode has
+          nothing almost everywhere.
-          one -- a form cannot nest inside another.
+
+          On an existing chat it writes on change, and needs the same treatment
+          the mode select gets: the verb on the control, `form=` for the values.
+          Before there is a chat there is nothing to PATCH, so it is an ordinary
+          field of the composer's own form and `_new_chat` reads it -- which is
+          what makes the setting choosable before the first prompt rather than
+          after it.
         #}
-        {% if chat and current_model and current_model.capabilities_json.get("reasoning") %}
+        {% if current_model and current_model.capabilities_json.get("reasoning") %}
         <select class="select select--sm" name="reasoning_effort" data-effort
                 aria-label="Reasoning effort" title="How hard this model should think"
-                form="chat-params-form">
+                {% if chat %}
+                form="chat-params-form"
+                hx-patch="/api/chats/{{ chat.id }}" hx-swap="none"
+                {% endif %}>
+          {% set chosen = chat.params_json.get('reasoning_effort') if chat
+                          else (current_model.params_json or {}).get('reasoning_effort') %}
           <option value="">Effort: default</option>
           {% for value in efforts %}
-          <option value="{{ value }}"
+          <option value="{{ value }}" {{ 'selected' if chosen == value }}>
-                  {{ 'selected' if chat.params_json.get('reasoning_effort') == value }}>
             Effort: {{ value }}
           </option>
           {% endfor %}
@@ -286,17 +302,22 @@
       </div>
     </form>
 
-    {# Outside the composer's form, and referenced by the mode select's `form`
+    {# Outside the composer's form, and referenced by the two selects' `form`
-       attribute above. hx-patch and not hx-post: there is no POST for a chat,
+       attributes above. These carry no htmx of their own: they exist so that
-       only PATCH, and htmx shows nothing when a request 405s -- which is how
+       `Nt(e)` -- htmx's "which form do the values come from", which reads
-       this control spent its whole life doing nothing. #}
+       `e.form` before falling back to `closest("form")` -- resolves to a form
+       holding exactly one control. Without them the PATCH would carry the
+       composer's `content`, `model_id` and `project_dir`, and `update_chat`
+       answers `project_dir` with a 409.
+
+       hx-patch and not hx-post: there is no POST for a chat, only PATCH, and
+       htmx shows nothing when a request 405s -- which is how these controls
+       spent the first half of their lives doing nothing. #}
     {% if chat and chat.kind == "agent" %}
-    <form id="agent-mode-form" hx-patch="/api/chats/{{ chat.id }}" hx-swap="none"
+    <form id="agent-mode-form"></form>
-          hx-trigger="change"></form>
     {% endif %}
     {% if chat %}
-    <form id="chat-params-form" hx-patch="/api/chats/{{ chat.id }}" hx-swap="none"
+    <form id="chat-params-form"></form>
-          hx-trigger="change"></form>
     {% endif %}
 
     <p class="composer__hint">

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

@@ -11,7 +11,8 @@
   all of it is escaped by autoescaping and none of it is marked safe.
 #}
 <div id="mention-results">
-  {% if not files and not documents %}
+  {% if not files and not documents and not notes and not skills
+        and not bases and not attachments and not website %}
   <p class="muted text-sm" style="padding: var(--sp-3)">
     {% if q %}
       Nothing matches “{{ q }}”.
@@ -21,6 +22,22 @@
   </p>
   {% else %}
 
+  {% if website %}
+  <p class="picker__group">A page to read</p>
+  <ul class="picker__list">
+    <li>
+      <button class="picker__option" type="button"
+              data-mention-url="{{ website }}" data-mention-token="{{ website }}">
+        {{ icon("link", "icon--sm") }}
+        <span class="picker__option-body">
+          <span class="picker__option-name">Fetch this page</span>
+          <span class="picker__option-note mono">{{ website }}</span>
+        </span>
+      </button>
+    </li>
+  </ul>
+  {% endif %}
+
   {% if files %}
   <p class="picker__group">In the project</p>
   <ul class="picker__list">
@@ -61,5 +78,82 @@
   </ul>
   {% endif %}
 
+  {% if notes %}
+  <p class="picker__group">Notes</p>
+  <ul class="picker__list">
+    {% for note in notes %}
+    <li>
+      <button class="picker__option" type="button"
+              data-mention-note="{{ note.id }}" data-mention-token="{{ note.title }}">
+        {{ icon("file-text", "icon--sm") }}
+        <span class="picker__option-body">
+          <span class="picker__option-name">{{ note.title }}</span>
+          <span class="picker__option-note">
+            {{ note.body | truncate(70) }}{% if note.owner_id != user.id %} · shared{% endif %}
+          </span>
+        </span>
+      </button>
+    </li>
+    {% endfor %}
+  </ul>
+  {% endif %}
+
+  {% if skills %}
+  <p class="picker__group">Skills</p>
+  <ul class="picker__list">
+    {% for skill in skills %}
+    <li>
+      <button class="picker__option" type="button"
+              data-mention-skill="{{ skill.id }}" data-mention-token="{{ skill.name }}">
+        {{ icon("sparkle", "icon--sm") }}
+        <span class="picker__option-body">
+          <span class="picker__option-name mono">{{ skill.name }}</span>
+          <span class="picker__option-note">{{ skill.description }}</span>
+        </span>
+      </button>
+    </li>
+    {% endfor %}
+  </ul>
+  {% endif %}
+
+  {% if bases %}
+  {# A base is a reference and not a copy: choosing one narrows what this chat
+     may search rather than putting anything into the message. #}
+  <p class="picker__group">Search only these</p>
+  <ul class="picker__list">
+    {% for base in bases %}
+    <li>
+      <button class="picker__option" type="button"
+              data-mention-base="{{ base.id }}" data-mention-token="{{ base.name }}">
+        {{ icon("archive", "icon--sm") }}
+        <span class="picker__option-body">
+          <span class="picker__option-name">{{ base.name }}</span>
+          <span class="picker__option-note">Scope this chat to this knowledge base</span>
+        </span>
+      </button>
+    </li>
+    {% endfor %}
+  </ul>
+  {% endif %}
+
+  {% if attachments %}
+  <p class="picker__group">Already in this chat</p>
+  <ul class="picker__list">
+    {% for attachment in attachments %}
+    <li>
+      <button class="picker__option" type="button"
+              data-mention-attachment="{{ attachment.id }}"
+              data-mention-token="{{ attachment.filename }}">
+        {{ icon("attach", "icon--sm") }}
+        <span class="picker__option-body">
+          <span class="picker__option-name">{{ attachment.filename }}</span>
+          <span class="picker__option-note">{{ attachment.human_size }}</span>
+        </span>
+      </button>
+    </li>
+    {% endfor %}
+  </ul>
+  {% endif %}
+
   {% endif %}
 </div>

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

@@ -13,8 +13,16 @@
   escaped plain text for everyone else.
 #}
 {% set streaming = (message.role == "assistant" and not message.complete) %}
+{#
+  Typed while the previous reply was still being written, and not yet handed to
+  a model. It is in the transcript and it is not in the request. It must never
+  carry `sse-connect` -- a queued turn with a streaming shell on it would be the
+  second concurrent reply the queue exists to prevent.
+#}
+{% set queued = (message.role == "user" and message.queued) %}
 
-<article class="msg msg--{{ message.role }}" id="msg-{{ message.id }}"
+<article class="msg msg--{{ message.role }}{{ ' msg--queued' if queued }}"
+         id="msg-{{ message.id }}"
   {% if streaming %}
          hx-ext="sse"
          sse-connect="/api/chats/{{ chat.id }}/messages/{{ message.id }}/stream"
@@ -210,7 +218,24 @@
     </div>
     {% endif %}
 
-    {% if not streaming %}
+    {% if queued %}
+    {# Nothing has been sent to any model. Both buttons sit on the bubble they
+       act on rather than in a toast somewhere, and Edit is deliberately absent:
+       editing rewinds and then starts a reply, which on a row you can press
+       while another reply is streaming is a second concurrent generation behind
+       a pencil. Discard and retype is the honest affordance. #}
+    <footer class="msg__actions msg__actions--queued">
+      <span class="msg__note">{{ icon("clock", "icon--sm") }} Waiting to be sent</span>
+      <button class="btn btn--sm" type="button"
+              hx-post="/api/chats/{{ chat.id }}/messages/{{ message.id }}/send-now"
+              hx-target="#thread" hx-swap="innerHTML">Send now</button>
+      <button class="btn btn--sm" type="button"
+              hx-post="/api/chats/{{ chat.id }}/messages/{{ message.id }}/discard"
+              hx-target="#msg-{{ message.id }}" hx-swap="outerHTML"
+              data-confirm-button="Discard this message?">Discard</button>
+    </footer>
+    <div hidden id="msg-body-{{ message.id }}">{{ message.content }}</div>
+    {% elif not streaming %}
     <footer class="msg__actions">
       <button class="btn btn--icon btn--sm" type="button"
               data-copy="msg-body-{{ message.id }}" aria-label="Copy message">

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

@@ -50,12 +50,21 @@
             aria-label="Send to chat">
       {{ icon("arrow-up", "icon--sm") }}
     </button>
-    <button class="btn btn--icon btn--sm" type="button" data-terminal-auto
+    {#
-            aria-pressed="false"
+      Three states, not two. A cycling icon button cannot say which of three it
-            title="Attach every command you run to your next message"
+      is in, and this one decides whether things are sent to a model without
-            aria-label="Send every command automatically">
+      being asked again -- so it says so in words.
-      {{ icon("sparkle", "icon--sm") }}
+
-    </button>
+      Not remembered between page loads on purpose: a switch that forwards every
+      command you run to a model is not something to inherit from last week.
+    #}
+    <select class="select select--sm" data-terminal-auto
+            aria-label="What to do when a command finishes"
+            title="What to do with each command you run">
+      <option value="off" selected>Auto: off</option>
+      <option value="copy">Auto: copy</option>
+      <option value="send">Auto: send</option>
+    </select>
     <button class="btn btn--icon btn--sm" type="button" data-toggle="#terminal"
             aria-label="Close terminal">
       {{ icon("x", "icon--sm") }}

tests/conftest.py

@@ -209,6 +209,32 @@ def mock_http():
     httpx.AsyncClient = original
 
 
+def control_named(html: str, name: str) -> dict[str, str]:
+    """The attributes of the one element carrying `name="…"`.
+
+    Exists so a test can ask "does the control that carries the name also carry
+    the verb?". Two selects in the composer once delegated their `hx-patch` to
+    an empty sibling form through the `form=` attribute, which scopes values but
+    routes no events -- htmx binds a trigger to the annotated element, and
+    `change` reaches ancestors, never siblings. Both controls were decorative
+    for a whole release, and the tests passed the entire time because they
+    asserted the markup that was there rather than the property that mattered.
+    """
+    from html.parser import HTMLParser
+
+    found: list[dict[str, str]] = []
+
+    class Finder(HTMLParser):
+        def handle_starttag(self, tag: str, attrs: list[tuple[str, str | None]]) -> None:
+            got = {key: (value or "") for key, value in attrs}
+            if got.get("name") == name:
+                found.append(got)
+
+    Finder().feed(html)
+    assert len(found) == 1, f"expected one element named {name!r}, found {len(found)}"
+    return found[0]
+
+
 @pytest.fixture
 def user_id(db: Session, registered: dict[str, str]) -> str:
     """The registered user's id.

tests/test_agent_browse.py

@@ -254,3 +254,78 @@ def test_somebody_elses_connection_is_not_browsable(
     # 404 and not 403: whether that connection exists is not this endpoint's to
     # reveal to somebody who does not own it.
     assert client.get(f"/api/agents/{profile.id}/browse").status_code == 404
+
+
+# --- Reading the project directory again -------------------------------------
+def _agent_chat(db, profile):
+    from lembas.db.models import KIND_AGENT, Chat, Connection, Model
+
+    connection = Connection(name="c", base_url="http://127.0.0.1:1", api_key_encrypted="")
+    db.add(connection)
+    db.commit()
+    db.add(Model(connection_id=connection.id, model_id="m", capabilities_json={"tools": True}))
+    db.commit()
+
+    chat = Chat(
+        user_id=profile.owner_id,
+        model_id="m",
+        connection_id=connection.id,
+        kind=KIND_AGENT,
+        ssh_profile_id=profile.id,
+        project_dir=profile.default_dir,
+    )
+    db.add(chat)
+    db.commit()
+    return chat
+
+
+def test_reindexing_walks_the_tree_again(client: TestClient, db, registered, served_tree):
+    """The listing is built only when a reply starts and then held for five
+    minutes, so anything done in the terminal panel is invisible to it until
+    then. This is the way to say "look again"."""
+    from lembas.services.agent import index as index_service
+
+    profile = _profile(db, served_tree)
+    chat = _agent_chat(db, profile)
+
+    response = client.post(f"/api/chats/{chat.id}/index")
+    assert response.status_code == 200, response.text
+    body = response.json()
+
+    assert body["ok"] is True
+    assert body["files"] >= 2  # README.md and src/
+    assert index_service.cached(profile.id, served_tree["root"]) is not None
+
+
+def test_reindexing_a_plain_chat_says_there_is_nothing_to_read(
+    client: TestClient, db, registered, served_tree
+):
+    from lembas.db.models import Chat, Connection, Model
+
+    connection = Connection(name="c", base_url="http://127.0.0.1:1", api_key_encrypted="")
+    db.add(connection)
+    db.commit()
+    db.add(Model(connection_id=connection.id, model_id="m"))
+    db.commit()
+    chat = Chat(user_id=_profile(db, served_tree).owner_id, model_id="m",
+                connection_id=connection.id)
+    db.add(chat)
+    db.commit()
+
+    assert client.post(f"/api/chats/{chat.id}/index").status_code == 409
+
+
+def test_reindexing_somebody_elses_chat_is_not_possible(
+    client: TestClient, db, registered, served_tree
+):
+    profile = _profile(db, served_tree)
+    chat = _agent_chat(db, profile)
+
+    client.post("/auth/logout", follow_redirects=False)
+    client.post(
+        "/auth/register",
+        data={"name": "Sam", "email": "sam@shire.test", "password": "potatoes-po-ta-toes"},
+        follow_redirects=False,
+    )
+
+    assert client.post(f"/api/chats/{chat.id}/index").status_code == 404

tests/test_agent_index.py

@@ -180,6 +180,43 @@ async def test_forgetting_a_connection_drops_its_listings():
     assert index_service.cached("profile-2", "/work") is not None
 
 
+async def test_a_host_that_refuses_to_run_commands_still_gets_a_listing():
+    """SFTP is the rung for exactly this, and the ladder used to skip it.
+
+    An `ExecError` from git or find -- an SFTP-only account, a forced command,
+    a shell that is `/bin/false` -- escaped the loop and was caught outside it,
+    which returned an empty index without ever trying the one method that would
+    have worked.
+    """
+
+    class _NoExec(_Fake):
+        async def run(self, request):
+            raise ExecError("This account may not run commands.")
+
+    executor = _NoExec(tree={"/work": [RemoteEntry("README.md", False, 5)]})
+
+    found = await index_service.build(executor, "/work")
+
+    assert found.source == "sftp"
+    assert "README.md" in found.paths
+
+
+async def test_forgetting_one_tree_leaves_the_others():
+    """What a write invalidates is the directory it wrote into, not the machine.
+
+    Two chats on one box in different trees share nothing but the connection,
+    and dropping both would make every write cost somebody else a walk.
+    """
+    executor = _Fake(answers={"git ls-files": _ok("a.py\n")})
+    await index_service.ensure(executor, "profile-1", "/work")
+    await index_service.ensure(executor, "profile-1", "/other")
+
+    index_service.forget_dir("profile-1", "/work")
+
+    assert index_service.cached("profile-1", "/work") is None
+    assert index_service.cached("profile-1", "/other") is not None
+
+
 def test_reading_the_cache_never_does_work():
     """`harness` calls this synchronously while assembling the system message,
     so it must never be the thing that opens a connection."""

tests/test_agent_mode.py

@@ -18,6 +18,8 @@ from lembas.db.models import KIND_AGENT, Chat, Connection, Model, SshProfile, Us
 from lembas.services import settings_store
 from lembas.services.agent import policy
 
+from .conftest import control_named
+
 
 def _agent_chat(db, *, mode: str = policy.MODE_MANUAL) -> Chat:
     """An agent chat pointed at a profile that is never actually connected to.
@@ -93,22 +95,28 @@ def test_the_mode_form_uses_a_method_the_route_serves(client: TestClient, db, re
     assert chat.agent_mode == policy.MODE_MANUAL
 
 
-def test_the_rendered_form_patches(client: TestClient, db, registered):
+def test_the_mode_select_carries_its_own_verb(client: TestClient, db, registered):
-    """And that the template actually carries it, since that is where it broke.
+    """The request must hang off the element the event fires on.
 
-    Pinned to the mode form by its id rather than to any `hx-patch` on the
+    This is the invariant, and the previous version of this test did not check
-    page: the model picker and the system-prompt box patch the same URL, so a
+    it. `hx-patch` lived on an empty sibling `<form>` that the select pointed at
-    looser assertion would have passed throughout the entire life of the bug.
+    with `form="…"`, which was enough to make the markup look right and enough
+    to make every assertion here pass -- while htmx bound the `change` listener
+    to the form, and `change` fires on the select and bubbles to its *ancestors*
+    only. The mode never once reached the database.
+
+    So: assert on the control, by name, whichever element that turns out to be.
     """
     chat = _agent_chat(db)
     body = client.get(f"/chat/{chat.id}").text
 
-    assert 'id="agent-mode-form"' in body
+    select = control_named(body, "agent_mode")
-    form = body[body.index('id="agent-mode-form"') :][:200]
+    assert select["hx-patch"] == f"/api/chats/{chat.id}"
-    assert f'hx-patch="/api/chats/{chat.id}"' in form
+    assert "hx-post" not in select
-    assert "hx-post" not in form
+    # And the empty form still scopes the values, so the PATCH carries this
-    # And that the select outside it is actually submitted by it.
+    # field alone rather than the whole composer -- `project_dir` in a PATCH is
-    assert 'form="agent-mode-form"' in body
+    # a 409.
+    assert select["form"] == "agent-mode-form"
 
 
 def test_the_mode_is_offered_beside_the_composer_not_in_the_topbar(

tests/test_agent_tools.py

@@ -9,6 +9,7 @@ from __future__ import annotations
 
 import asyncio
 import json as _json
+import time
 
 import pytest
 
@@ -205,6 +206,31 @@ async def test_files_are_written_and_read_back(db, user_id, machine, tmp_path):
     assert "note.txt" in listed.content
 
 
+async def test_writing_a_file_drops_the_project_listing(db, user_id, machine):
+    """Otherwise the model is shown a five-minute-old tree that it knows is
+    wrong, and concludes the file it has just created does not exist.
+
+    The TTL is for drift nobody can see coming. This is not that: it is this
+    process changing the tree it has just described.
+    """
+    from lembas.services.agent import index as index_service
+
+    chat, profile = _setup(db, user_id, machine, mode=policy.MODE_AUTO)
+    user = db.get(User, user_id)
+    resolved = tools_service.resolve_tools(db, chat, user)
+    context = tools_service.context_for(db, user, chat, tools=resolved)
+
+    index_service._CACHE[(profile.id, machine["dir"])] = index_service.ProjectIndex(
+        paths=("stale.txt",), total=1, source="git", built_at=time.monotonic()
+    )
+
+    await tools_service.run_tool(
+        context, "file_write", '{"path": "fresh.txt", "content": "hi"}'
+    )
+
+    assert index_service.cached(profile.id, machine["dir"]) is None
+
+
 # --- The runner backstop ----------------------------------------------------------
 async def test_a_runner_refuses_what_the_mode_forbids(db, user_id, machine):
     """`_authorise` is the real gate and runs first. This is the belt to that
@@ -486,6 +512,140 @@ async def test_an_agent_chat_is_told_its_real_round_budget(db, user_id, machine)
     assert values["max_rounds"] == "25"
 
 
+async def test_an_agent_chat_gets_the_rounds_it_was_promised(db, user_id, machine, monkeypatch):
+    """And is then allowed to use them, which is the half that was missing.
+
+    The budget sizes the loop and names itself in the out-of-rounds message, and
+    the harness above tells the model the same number. But the comparison that
+    ends the loop read the global `MAX_ROUNDS` of three. So an agent chat
+    allowed forty rounds stopped after three and reported that it had taken
+    forty: two wrong answers to "why did it stop", with no way to tell them
+    apart from the outside.
+    """
+    settings_store.update(db, {"max_steps": 5}, key=settings_store.AGENTS)
+    chat, _profile = _setup(db, user_id, machine, mode=policy.MODE_AUTO)
+    assistant = Message(chat_id=chat.id, role=ROLE_ASSISTANT, content="", complete=False)
+    db.add(assistant)
+    db.commit()
+
+    payloads: list[dict] = []
+
+    async def stream_chat(_endpoint, payload):
+        payloads.append(payload)
+        yield {
+            "choices": [
+                {"delta": {"tool_calls": [{"index": 0, "id": "c1", "function": {
+                    "name": "file_list", "arguments": '{"path": "."}'}}]}}
+            ]
+        }
+
+    monkeypatch.setattr(generation_service, "stream_chat", stream_chat)
+
+    async def _no_title(*_args, **_kwargs):
+        return ""
+
+    monkeypatch.setattr("lembas.services.chat.generate_title", _no_title)
+
+    generation = generation_service.Generation(chat_id=chat.id, message_id=assistant.id)
+    await generation_service._run(generation)
+
+    # Five rounds that may call tools, then the one that gives up.
+    assert len(payloads) == 6
+    assert "after 5 rounds" in generation.tool_events[-1]["error"]
+
+
+# --- Interjecting while it works --------------------------------------------------
+async def test_a_queued_prompt_is_taken_in_between_rounds(db, user_id, machine, monkeypatch):
+    """The point of queueing in an agent chat: steering work already under way.
+
+    An agent that has just finished one loop and is about to start another is
+    exactly when "actually, do it the other way" is worth having, and making it
+    wait for the whole reply would mean it arrives after the thing it was meant
+    to change.
+    """
+    from lembas.services import chat as chat_service
+
+    chat, _profile = _setup(db, user_id, machine, mode=policy.MODE_AUTO)
+    message_id = _pending_reply(db, chat)
+    chat_service.create_message(
+        db, chat, "user", "actually, check the other directory first", queued=True
+    )
+
+    payloads: list[dict] = []
+    monkeypatch.setattr(
+        generation_service,
+        "stream_chat",
+        _stub_stream(
+            [[_chunk("file_list", '{"path": "."}')], [_text("Done.")]],
+            payloads,
+        ),
+    )
+
+    generation = generation_service.Generation(chat_id=chat.id, message_id=message_id)
+    await generation_service._run(generation)
+
+    # Verbatim, in the user role, with nothing wrapped around it: this genuinely
+    # is the person at the keyboard, and quoting it would teach the model that a
+    # user turn can be a quotation -- the distinction `execute_plan` relies on.
+    assert payloads[1]["messages"][-1] == {
+        "role": "user",
+        "content": "actually, check the other directory first",
+    }
+
+
+async def test_an_interjection_is_delivered_only_once(db, user_id, machine, monkeypatch):
+    """Marked delivered before the request goes out, so a crash loses it rather
+    than asking the same thing twice and letting an agent act on it twice."""
+    from lembas.services import chat as chat_service
+
+    chat, _profile = _setup(db, user_id, machine, mode=policy.MODE_AUTO)
+    message_id = _pending_reply(db, chat)
+    waiting = chat_service.create_message(db, chat, "user", "one more thing", queued=True)
+
+    monkeypatch.setattr(
+        generation_service,
+        "stream_chat",
+        _stub_stream(
+            [
+                [_chunk("file_list", '{"path": "."}')],
+                [_chunk("file_list", '{"path": "src"}')],
+                [_text("Done.")],
+            ],
+            [],
+        ),
+    )
+
+    generation = generation_service.Generation(chat_id=chat.id, message_id=message_id)
+    await generation_service._run(generation)
+
+    db.expire_all()
+    assert db.get(Message, waiting.id).queued is False
+    assert generation.injected_ids == [waiting.id]
+
+
+async def test_the_reply_sorts_before_the_prompt_it_took_in(db, user_id, machine, monkeypatch):
+    """Otherwise the next request reads "answer, then the question it answered",
+    and a small model dutifully answers it a second time."""
+    from lembas.services import chat as chat_service
+
+    chat, _profile = _setup(db, user_id, machine, mode=policy.MODE_AUTO)
+    message_id = _pending_reply(db, chat)
+    waiting = chat_service.create_message(db, chat, "user", "and this", queued=True)
+
+    monkeypatch.setattr(
+        generation_service,
+        "stream_chat",
+        _stub_stream([[_chunk("file_list", '{"path": "."}')], [_text("Done.")]], []),
+    )
+
+    generation = generation_service.Generation(chat_id=chat.id, message_id=message_id)
+    await generation_service._run(generation)
+
+    db.expire_all()
+    reply = db.get(Message, message_id)
+    assert reply.created_at > db.get(Message, waiting.id).created_at
+
+
 # --- Plan mode's artifact ---------------------------------------------------------
 def test_plan_submit_is_offered_only_in_plan_mode(db, user_id, machine):
     """It ends the reply. A model in Auto mode that proposed a plan instead of

tests/test_effort.py

@@ -17,6 +17,8 @@ from sqlalchemy import select
 from lembas.db.models import Chat, Connection, Model, User
 from lembas.services import chat as chat_service
 
+from .conftest import control_named
+
 
 def _model(db, **capabilities) -> Model:
     connection = Connection(name="c", base_url="http://127.0.0.1:1", api_key_encrypted="")
@@ -162,3 +164,69 @@ def test_a_model_with_no_defaults_starts_a_plain_chat(client: TestClient, db, re
     client.post("/api/chats/start", data={"content": "hello", "model_id": "m"})
 
     assert db.scalars(select(Chat)).one().params_json == {}
+
+
+# --- Choosable before the first prompt ---------------------------------------
+def test_the_effort_select_carries_its_own_verb(client: TestClient, db, registered):
+    """The same invariant the mode select needs, for the same reason.
+
+    Both were built on `form="…"` pointing at an empty sibling form holding the
+    `hx-patch`, and both therefore wrote nothing at all: `form=` scopes the
+    values a request carries, it does not route the event that starts one.
+    """
+    chat = _chat(db)
+
+    select = control_named(client.get(f"/chat/{chat.id}").text, "reasoning_effort")
+    assert select["hx-patch"] == f"/api/chats/{chat.id}"
+    assert select["form"] == "chat-params-form"
+
+
+def test_the_effort_is_offered_before_there_is_a_chat(client: TestClient, db, registered):
+    """Otherwise it is a setting you can only reach once it is too late to use.
+
+    On the new-chat screen there is nothing to PATCH, so it is an ordinary field
+    of the composer's form and carries no verb -- `_new_chat` reads it.
+    """
+    _model(db)
+
+    select = control_named(client.get("/chat").text, "reasoning_effort")
+    assert "hx-patch" not in select
+    assert "form" not in select
+
+
+def test_starting_a_chat_with_an_effort_stores_it(client: TestClient, db, registered):
+    _model(db)
+
+    client.post(
+        "/api/chats/start",
+        data={"content": "hello", "model_id": "m", "reasoning_effort": "low"},
+    )
+
+    assert db.scalars(select(Chat)).one().params_json["reasoning_effort"] == "low"
+
+
+def test_an_explicit_effort_beats_the_models_default(client: TestClient, db, registered):
+    """An inherited value is a starting point, not a ceiling."""
+    model = _model(db)
+    model.params_json = {"reasoning_effort": "high"}
+    db.commit()
+
+    client.post(
+        "/api/chats/start",
+        data={"content": "hello", "model_id": "m", "reasoning_effort": "low"},
+    )
+
+    assert db.scalars(select(Chat)).one().params_json["reasoning_effort"] == "low"
+
+
+def test_a_nonsense_effort_at_the_start_falls_back(client: TestClient, db, registered):
+    model = _model(db)
+    model.params_json = {"reasoning_effort": "high"}
+    db.commit()
+
+    client.post(
+        "/api/chats/start",
+        data={"content": "hello", "model_id": "m", "reasoning_effort": "extreme"},
+    )
+
+    assert db.scalars(select(Chat)).one().params_json["reasoning_effort"] == "high"

tests/test_mentions.py

@@ -152,6 +152,143 @@ def test_a_plain_chat_gets_a_picker_with_no_file_half(client: TestClient, db, re
     assert "In the project" not in response.text
 
 
+# --- What `@` offers where there is no machine at all ------------------------
+def _library(db):
+    """A note, a skill and a base belonging to the registered reader."""
+    from lembas.db.models import KnowledgeBase, Note, Skill
+
+    owner = db.scalars(select(User)).first()
+    note = Note(owner_id=owner.id, title="Mallorn notes", body="Golden leaves.")
+    skill = Skill(
+        owner_id=owner.id, name="bake-lembas", description="How to bake it", body="Steps."
+    )
+    base = KnowledgeBase(owner_id=owner.id, name="Contracts")
+    db.add_all([note, skill, base])
+    db.commit()
+    return note, skill, base
+
+
+def test_notes_and_skills_are_offered_in_a_plain_chat(client: TestClient, db, registered):
+    """A chat with no SSH connection has no project files, which is exactly why
+    the rest of the library has to be reachable there."""
+    _library(db)
+
+    body = client.get("/api/files/mention-picker", params={"q": "mallorn"}).text
+    assert "Mallorn notes" in body
+
+    body = client.get("/api/files/mention-picker", params={"q": "lembas"}).text
+    assert "bake-lembas" in body
+
+
+def test_a_knowledge_base_is_only_offered_inside_a_chat(
+    client: TestClient, db, registered, make_chat
+):
+    """There is nothing to attach it to before a chat exists -- the same reason
+    project files are absent on the new-chat screen."""
+    _library(db)
+    chat_id = make_chat()
+
+    assert "Contracts" not in client.get("/api/files/mention-picker").text
+    assert "Contracts" in client.get(
+        "/api/files/mention-picker", params={"chat_id": chat_id}
+    ).text
+
+
+def test_a_url_is_offered_as_a_page_to_read(client: TestClient, db, registered):
+    body = client.get(
+        "/api/files/mention-picker", params={"q": "https://tolkien.test/mallorn"}
+    ).text
+
+    assert "Fetch this page" in body
+    assert "https://tolkien.test/mallorn" in body
+
+
+def test_a_note_arrives_with_its_text_and_its_name(client: TestClient, db, registered, make_chat):
+    note, _skill, _base = _library(db)
+    chat_id = make_chat()
+
+    response = client.post(
+        "/api/files/from-note", data={"note_id": note.id, "chat_id": chat_id}
+    )
+
+    assert response.status_code == 200
+    attachment = db.scalars(select(Attachment)).one()
+    assert "Golden leaves." in attachment.extracted_text
+    # Provenance, for the reason a project file carries it: a model handed four
+    # documents cannot name one back when asked to work on it.
+    assert attachment.source_label == "Note"
+    assert attachment.source_path == "Mallorn notes"
+
+
+def test_a_skill_can_be_handed_over_directly(client: TestClient, db, registered, make_chat):
+    """The index is in the harness and `skill_get` fetches on demand -- but only
+    if the model decides to. `@` is the reader saying "use this one"."""
+    _note, skill, _base = _library(db)
+    chat_id = make_chat()
+
+    client.post("/api/files/from-skill", data={"skill_id": skill.id, "chat_id": chat_id})
+
+    attachment = db.scalars(select(Attachment)).one()
+    assert attachment.source_label == "Skill"
+    assert "Steps." in attachment.extracted_text
+
+
+def test_a_base_is_attached_as_a_reference_not_a_copy(
+    client: TestClient, db, registered, make_chat
+):
+    """Scoping, not copying. A folder of contracts in the window would cost the
+    context on every request forever to answer one question."""
+    from lembas.db.models import Chat
+
+    _note, _skill, base = _library(db)
+    chat_id = make_chat()
+
+    response = client.post(f"/api/chats/{chat_id}/bases", data={"base_id": base.id})
+
+    assert response.status_code == 200
+    db.expire_all()
+    assert [b.id for b in db.get(Chat, chat_id).knowledge_bases] == [base.id]
+    # Nothing was copied into the message.
+    assert db.scalars(select(Attachment)).all() == []
+
+
+def test_attaching_the_same_base_twice_is_not_an_error(
+    client: TestClient, db, registered, make_chat
+):
+    from lembas.db.models import Chat
+
+    _note, _skill, base = _library(db)
+    chat_id = make_chat()
+
+    client.post(f"/api/chats/{chat_id}/bases", data={"base_id": base.id})
+    client.post(f"/api/chats/{chat_id}/bases", data={"base_id": base.id})
+
+    db.expire_all()
+    assert len(db.get(Chat, chat_id).knowledge_bases) == 1
+
+
+def test_a_library_document_now_carries_its_provenance(client: TestClient, db, registered):
+    """This was the one attach path that dropped it, while a project file beside
+    it carried path and machine."""
+    from lembas.services.fetch import Fetched
+    from lembas.services.library import documents as documents_service
+
+    owner = db.scalars(select(User)).first()
+    base = documents_service.default_base(db, owner)
+    document = documents_service.store_page(
+        db,
+        owner=owner,
+        base=base,
+        page=Fetched(url="https://tolkien.test/c", title="The Contract", text="Terms."),
+    )
+
+    client.post("/api/files/from-knowledge", data={"document_id": document.id})
+
+    attachment = db.scalars(select(Attachment)).one()
+    assert attachment.source_path == "The Contract"
+    assert attachment.source_label == base.name
+
+
 # --- Attaching ---------------------------------------------------------------
 def test_a_mentioned_file_arrives_with_its_contents(client: TestClient, db, registered, box):
     profile = _profile(db, box)

tests/test_queue.py

@@ -0,0 +1,393 @@
+"""Typing while a reply is being written.
+
+Before this existed, a second message during a stream was simply accepted: it
+wrote a second assistant placeholder, started a second `Generation`, and left
+two replies answering the same chat from two different prefixes of it -- with
+Stop pointing at whichever bubble came first in the document.
+
+Now it queues. The queue is not an object: it is "the rows in this chat with
+`queued` set, oldest first". That is the whole reason it survives a restart and
+the reason `_prune` cannot take it away.
+"""
+
+from __future__ import annotations
+
+import pytest
+from fastapi.testclient import TestClient
+from sqlalchemy import select
+
+from lembas.api.chats import MAX_QUEUED
+from lembas.db.models import ROLE_ASSISTANT, ROLE_USER, Chat, Connection, Message, Model
+from lembas.services import chat as chat_service
+from lembas.services import generation as generation_service
+from lembas.services.crypto import encrypt
+
+
+@pytest.fixture(autouse=True)
+def empty_registry():
+    yield
+    generation_service._RUNNING.clear()
+    generation_service._TASKS.clear()
+
+
+@pytest.fixture
+def no_upstream(monkeypatch):
+    """Replace the producer, so the routes can be driven without a server."""
+    started: list = []
+
+    async def _fake_run(generation):
+        started.append(generation)
+
+    monkeypatch.setattr(generation_service, "_run", _fake_run)
+    return started
+
+
+def _connection(db) -> Connection:
+    connection = Connection(
+        name="Test", base_url="http://127.0.0.1:1", api_key_encrypted=encrypt("")
+    )
+    db.add(connection)
+    db.commit()
+    db.add(Model(connection_id=connection.id, model_id="test-model"))
+    db.commit()
+    return connection
+
+
+def _mid_reply(db, chat_id: str) -> Message:
+    """A chat with a reply still being written, which is the whole precondition."""
+    db.add(Message(chat_id=chat_id, role=ROLE_USER, content="what is lembas?"))
+    reply = Message(chat_id=chat_id, role=ROLE_ASSISTANT, content="Way", complete=False)
+    db.add(reply)
+    db.commit()
+    return reply
+
+
+def _messages(db, chat_id: str) -> list[Message]:
+    return list(
+        db.scalars(
+            select(Message).where(Message.chat_id == chat_id).order_by(Message.created_at)
+        )
+    )
+
+
+# --- The column --------------------------------------------------------------
+def test_an_ordinary_message_is_not_queued(db, registered, make_chat):
+    """Every row that predates the column reads the same way, because
+    `sync_schema` adds a NOT NULL boolean with a literal default of 0."""
+    chat_id = make_chat()
+    chat = db.get(Chat, chat_id)
+
+    assert chat_service.create_message(db, chat, ROLE_USER, "hello").queued is False
+
+
+def test_a_queued_message_is_left_out_of_the_request(db, registered, make_chat):
+    """It is in the transcript and it is not in the request. That distinction is
+    the entire feature."""
+    chat_id = make_chat()
+    chat = db.get(Chat, chat_id)
+    chat_service.create_message(db, chat, ROLE_USER, "first")
+    chat_service.create_message(db, chat, ROLE_ASSISTANT, "an answer")
+    chat_service.create_message(db, chat, ROLE_USER, "typed while it worked", queued=True)
+
+    sent = chat_service.build_messages(db, chat)
+
+    assert [m["content"] for m in sent] == ["first", "an answer"]
+
+
+def test_a_delivered_message_is_sent_on_the_next_turn(db, registered, make_chat):
+    chat_id = make_chat()
+    chat = db.get(Chat, chat_id)
+    chat_service.create_message(db, chat, ROLE_USER, "first")
+    waiting = chat_service.create_message(db, chat, ROLE_USER, "second", queued=True)
+
+    waiting.queued = False
+    db.commit()
+
+    assert [m["content"] for m in chat_service.build_messages(db, chat)] == ["first", "second"]
+
+
+# --- Queueing ----------------------------------------------------------------
+def test_a_message_sent_during_a_reply_is_queued(
+    client: TestClient, db, registered, make_chat, no_upstream
+):
+    """The bug this replaces: a second POST used to start a second generation."""
+    _connection(db)
+    chat_id = make_chat()
+    _mid_reply(db, chat_id)
+
+    response = client.post(f"/api/chats/{chat_id}/messages", data={"content": "and also this"})
+    assert response.status_code == 200
+
+    rows = _messages(db, chat_id)
+    assert rows[-1].content == "and also this"
+    assert rows[-1].queued is True
+    # Exactly one reply in flight, which is the point.
+    assert len([m for m in rows if m.role == ROLE_ASSISTANT and not m.complete]) == 1
+
+
+def test_a_queued_bubble_never_carries_a_streaming_shell(
+    client: TestClient, db, registered, make_chat, no_upstream
+):
+    """`sse-connect` is the only thing that starts a generation, so a queued
+    turn carrying one would be the second concurrent reply all over again.
+
+    Asserted on the body rather than on a row, unusually and deliberately: the
+    attribute *is* the behaviour here.
+    """
+    _connection(db)
+    chat_id = make_chat()
+    _mid_reply(db, chat_id)
+
+    body = client.post(f"/api/chats/{chat_id}/messages", data={"content": "later"}).text
+
+    assert "sse-connect" not in body
+    assert "Waiting to be sent" in body
+
+
+def test_a_message_with_nothing_in_flight_is_sent_as_before(
+    client: TestClient, db, registered, make_chat, no_upstream
+):
+    _connection(db)
+    chat_id = make_chat()
+
+    response = client.post(f"/api/chats/{chat_id}/messages", data={"content": "hello"})
+
+    assert "sse-connect" in response.text
+    rows = _messages(db, chat_id)
+    assert rows[0].queued is False
+    assert rows[1].role == ROLE_ASSISTANT and rows[1].complete is False
+
+
+def test_the_queue_is_bounded(client: TestClient, db, registered, make_chat, no_upstream):
+    """A `for` loop in the terminal panel with Auto send on can produce commands
+    far faster than any model answers them."""
+    _connection(db)
+    chat_id = make_chat()
+    _mid_reply(db, chat_id)
+
+    for n in range(MAX_QUEUED):
+        assert client.post(
+            f"/api/chats/{chat_id}/messages", data={"content": f"line {n}"}
+        ).status_code == 200
+
+    refused = client.post(f"/api/chats/{chat_id}/messages", data={"content": "one too many"})
+
+    assert refused.status_code == 409
+    assert len([m for m in _messages(db, chat_id) if m.queued]) == MAX_QUEUED
+
+
+# --- Delivery ----------------------------------------------------------------
+async def test_a_finished_reply_delivers_the_next_prompt(
+    db, registered, make_chat, no_upstream
+):
+    _connection(db)
+    chat_id = make_chat()
+    reply = _mid_reply(db, chat_id)
+    chat = db.get(Chat, chat_id)
+    waiting = chat_service.create_message(db, chat, ROLE_USER, "next please", queued=True)
+
+    generation = generation_service.Generation(chat_id=chat_id, message_id=reply.id)
+    generation_service._drain(generation)
+
+    db.refresh(waiting)
+    assert waiting.queued is False
+    assert generation.drained is True
+    assert len([m for m in _messages(db, chat_id) if not m.complete]) == 2
+
+
+async def test_only_one_prompt_is_delivered_at_a_time(db, registered, make_chat, no_upstream):
+    """Draining the lot would put two consecutive user turns in the next
+    request, which several local chat templates refuse outright."""
+    _connection(db)
+    chat_id = make_chat()
+    reply = _mid_reply(db, chat_id)
+    chat = db.get(Chat, chat_id)
+    first = chat_service.create_message(db, chat, ROLE_USER, "one", queued=True)
+    second = chat_service.create_message(db, chat, ROLE_USER, "two", queued=True)
+
+    generation_service._drain(
+        generation_service.Generation(chat_id=chat_id, message_id=reply.id)
+    )
+
+    db.refresh(first)
+    db.refresh(second)
+    assert first.queued is False
+    assert second.queued is True
+
+
+async def test_a_stopped_reply_leaves_the_queue_alone(db, registered, make_chat, no_upstream):
+    """Stop means stop. This is the decision the whole feature was shaped
+    around, and it must never be relaxed into "stop this one and start the
+    next"."""
+    _connection(db)
+    chat_id = make_chat()
+    reply = _mid_reply(db, chat_id)
+    chat = db.get(Chat, chat_id)
+    waiting = chat_service.create_message(db, chat, ROLE_USER, "next please", queued=True)
+
+    generation = generation_service.Generation(chat_id=chat_id, message_id=reply.id)
+    generation.stopped = True
+    generation_service._drain(generation)
+
+    db.refresh(waiting)
+    assert waiting.queued is True
+    assert generation.drained is False
+
+
+async def test_an_errored_reply_leaves_the_queue_alone(db, registered, make_chat, no_upstream):
+    """The endpoint has just failed; sending the next prompt into it spends
+    somebody's words to produce a second failure."""
+    _connection(db)
+    chat_id = make_chat()
+    reply = _mid_reply(db, chat_id)
+    chat = db.get(Chat, chat_id)
+    waiting = chat_service.create_message(db, chat, ROLE_USER, "next please", queued=True)
+
+    generation = generation_service.Generation(chat_id=chat_id, message_id=reply.id)
+    generation.error = "The endpoint refused."
+    generation_service._drain(generation)
+
+    db.refresh(waiting)
+    assert waiting.queued is True
+
+
+async def test_a_superseded_generation_does_not_drain(db, registered, make_chat, no_upstream):
+    """A regeneration cancels its predecessor, whose `finally:` still runs --
+    the same reason `_persist` refuses. Without this, regenerating would drain
+    the queue as a side effect."""
+    _connection(db)
+    chat_id = make_chat()
+    reply = _mid_reply(db, chat_id)
+    chat = db.get(Chat, chat_id)
+    waiting = chat_service.create_message(db, chat, ROLE_USER, "next please", queued=True)
+
+    abandoned = generation_service.Generation(chat_id=chat_id, message_id=reply.id)
+    generation_service._RUNNING[reply.id] = generation_service.Generation(
+        chat_id=chat_id, message_id=reply.id
+    )
+    generation_service._drain(abandoned)
+
+    db.refresh(waiting)
+    assert waiting.queued is True
+
+
+# --- Send now and Discard ----------------------------------------------------
+def test_discarding_removes_the_row(client: TestClient, db, registered, make_chat):
+    _connection(db)
+    chat_id = make_chat()
+    chat = db.get(Chat, chat_id)
+    waiting = chat_service.create_message(db, chat, ROLE_USER, "never mind", queued=True)
+    waiting_id = waiting.id
+
+    response = client.post(f"/api/chats/{chat_id}/messages/{waiting_id}/discard")
+
+    assert response.status_code == 200
+    # Queried rather than `db.get`: the route committed in a session of its own,
+    # and this one still holds the instance.
+    db.expire_all()
+    assert db.scalar(select(Message).where(Message.id == waiting_id)) is None
+
+
+def test_discarding_a_delivered_message_is_refused(client: TestClient, db, registered, make_chat):
+    """Discard removes a row outright, so it must only ever reach one that was
+    never sent."""
+    _connection(db)
+    chat_id = make_chat()
+    chat = db.get(Chat, chat_id)
+    sent = chat_service.create_message(db, chat, ROLE_USER, "already gone")
+
+    assert client.post(f"/api/chats/{chat_id}/messages/{sent.id}/discard").status_code == 404
+    assert db.get(Message, sent.id) is not None
+
+
+def test_send_now_delivers_and_starts_a_reply(
+    client: TestClient, db, registered, make_chat, no_upstream
+):
+    _connection(db)
+    chat_id = make_chat()
+    chat = db.get(Chat, chat_id)
+    waiting = chat_service.create_message(db, chat, ROLE_USER, "go on then", queued=True)
+
+    response = client.post(f"/api/chats/{chat_id}/messages/{waiting.id}/send-now")
+
+    assert response.status_code == 200
+    db.refresh(waiting)
+    assert waiting.queued is False
+    assert len(no_upstream) == 1
+
+
+def test_send_now_is_refused_while_a_reply_is_running(
+    client: TestClient, db, registered, make_chat, no_upstream
+):
+    """Jumping the queue is starting a second generation, which is the thing
+    this whole mechanism exists to stop."""
+    _connection(db)
+    chat_id = make_chat()
+    _mid_reply(db, chat_id)
+    chat = db.get(Chat, chat_id)
+    waiting = chat_service.create_message(db, chat, ROLE_USER, "me first", queued=True)
+
+    response = client.post(f"/api/chats/{chat_id}/messages/{waiting.id}/send-now")
+
+    assert response.status_code == 409
+    db.refresh(waiting)
+    assert waiting.queued is True
+
+
+def test_neither_route_reaches_another_readers_chat(
+    client: TestClient, db, registered, make_chat
+):
+    _connection(db)
+    chat_id = make_chat()
+    chat = db.get(Chat, chat_id)
+    waiting = chat_service.create_message(db, chat, ROLE_USER, "mine", queued=True)
+
+    client.post("/auth/logout", follow_redirects=False)
+    client.post(
+        "/auth/register",
+        data={"name": "Sam", "email": "sam@shire.test", "password": "potatoes-po-ta-toes"},
+        follow_redirects=False,
+    )
+
+    assert client.post(f"/api/chats/{chat_id}/messages/{waiting.id}/discard").status_code == 404
+    assert client.post(f"/api/chats/{chat_id}/messages/{waiting.id}/send-now").status_code == 404
+    assert db.get(Message, waiting.id) is not None
+
+
+# --- Everything else that touches the thread ---------------------------------
+def test_editing_is_refused_while_a_reply_is_running(
+    client: TestClient, db, registered, make_chat, no_upstream
+):
+    """Editing rewinds and then starts a reply unconditionally. Pressing it
+    mid-stream was a second concurrent generation behind a pencil icon, and was
+    reachable before the queue existed too."""
+    _connection(db)
+    chat_id = make_chat()
+    _mid_reply(db, chat_id)
+    first = _messages(db, chat_id)[0]
+
+    response = client.post(
+        f"/api/chats/{chat_id}/messages/{first.id}/edit", data={"content": "rewritten"}
+    )
+
+    assert response.status_code == 409
+    db.refresh(first)
+    assert first.content == "what is lembas?"
+    assert len([m for m in _messages(db, chat_id) if not m.complete]) == 1
+
+
+def test_compaction_does_not_summarise_a_waiting_prompt(db, registered, make_chat):
+    """It would fold words no model has seen into the record, and then deliver
+    them again afterwards."""
+    from lembas.services import compaction as compaction_service
+
+    _connection(db)
+    chat_id = make_chat()
+    chat = db.get(Chat, chat_id)
+    chat_service.create_message(db, chat, ROLE_USER, "what is lembas?")
+    reply = chat_service.create_message(db, chat, ROLE_ASSISTANT, "Waybread.")
+    chat_service.create_message(db, chat, ROLE_USER, "still waiting", queued=True)
+
+    text = compaction_service.transcript(db, chat, upto=reply)
+
+    assert "still waiting" not in text