@datalayer/agent-runtimes 1.0.29 → 1.3.10

package/lib/loop/plugins/chat/ChatView.js

@@ -0,0 +1,1670 @@
+import { jsx as _jsx, jsxs as _jsxs, Fragment as _Fragment } from "react/jsx-runtime";
+import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+import { Box, IconButton, SegmentedControl, Text, Truncate, } from '@primer/react';
+import { useColorPalette } from '@datalayer/primer-addons';
+import { ScreenFullIcon, ScreenNormalIcon } from '@primer/octicons-react';
+import { signal } from '@datalayer/reactor';
+import { useContributions, useSignalValue, useGate, useReactorPlatform, ReactorLazy, ReactorSlot, useSlotComponents, } from '@datalayer/reactor/react';
+import { ChatBase } from '../../../chat/base/ChatBase';
+import { SUGGESTION_CHIP_WIDTH } from '../../../chat/display/EmptyState';
+import { AnonymousKeyExpired } from '@datalayer/core/lib/components/anonymous/AnonymousKeyExpired';
+import { useAnonymousSessionStore } from '../../../runtimes/browser/anonymousToken';
+import { browserProtocolConfig } from '../../../runtimes/browser';
+import { useBrowserInference } from '../../../hooks/useBrowserInference';
+import { getAgentspecs } from '../../../specs/agents';
+import { targetRunsAgentInPage, } from '../agents/switchable';
+import { agentIcon } from '../agents/agentIcons';
+import { specSubagents } from '../agents/specSubagents';
+import { subagentsFor } from '../agents/team';
+import { useOptionalTeamSelection } from '../agents/useTeamSelection';
+/**
+ * Stands in for a workspace with no team.
+ *
+ * `useSignalValue` needs a signal on every render, and a hook cannot be
+ * skipped for the ordinary single-agent case.
+ */
+const EMPTY_SELECTION = signal('');
+/* A signal to read when no example contributed chat extras, so the hook
+   below always has one — a hook cannot be called conditionally. */
+const EMPTY_CHAT_EXTRAS = signal({});
+import { useWorkspaceFullScreen } from '../../shell/useWorkspaceFullScreen';
+import { CHAT_PLUGIN_NAME } from './index';
+import { INPUT_PROMPT_PLUGIN_NAME, } from '../input-prompt';
+import { agentRuntimeStore, useAgentRuntimeContextSnapshot, } from '../../../stores';
+import { useIAMStore } from '../../../state';
+import { useConfig } from '../../../hooks/useConfig';
+import { useSkills, useSkillActions } from '../../../hooks/useSkills';
+import { AI_MODEL_CATALOGUE } from '../../../specs/models';
+import { LoopAgentBlueprint, LoopAgentGate, LoopChatComposer, LoopChatExtras, LoopChatHeader, LoopChatLayout, LoopChatSuggestion, LoopChatSurface, LoopChatTurn, LoopFrontendTool, LoopSlots, canOpenView, onPromptFocusRequest, onSurfaceRequest, useLoopPromptStore, } from '../../core';
+import { turnWritersOf } from './turnState';
+import { isInactiveSurfaceContribution, orderToolContributions, toolsForChatView, } from './chatViewTools';
+import { useAgentCommandTools } from '../../../tools/adapters/commands';
+import { NONE_EDITOR, seedEditorChoice, setEditorOptions, } from '../shell/editorChoice';
+/** No editor: the conversation on its own. */
+const NO_SURFACE = '';
+/**
+ * How fast the full-screen control breathes once it starts asking.
+ *
+ * Slow enough to read as breathing rather than blinking — a fast pulse on a
+ * small control reads as an error state, and this one is an invitation.
+ */
+const FULLSCREEN_HINT_PERIOD_MS = 1400;
+export default function ChatView({ workspace }) {
+    /* The active theme's own colours. Read from the store rather than a
+       provider, so a workspace mounted without the addons theme still gets the
+       default palette instead of throwing. */
+    const palette = useColorPalette();
+    const controlsRef = useRef(null);
+    const { setViewControls } = workspace;
+    const [transient, setTransient] = useState(null);
+    // Read from the reactor rather than taken as a prop: the placeholder and the
+    // default editor are the chat plugin's own configuration, so a host sets them
+    // where it sets anything else about this plugin.
+    const reactor = useReactorPlatform();
+    const config = reactor.getConfig(CHAT_PLUGIN_NAME);
+    const placeholder = config?.placeholder ?? 'Ask anything, type / for commands or @ for mention';
+    const defaultSurface = config?.defaultSurface ?? 'notebook';
+    /*
+     * The input-prompt plugin's own configuration, read from here rather than
+     * from that plugin's component.
+     *
+     * `firstPromptHook` has to fire for *every* way a message reaches
+     * `handleSend` below — typed and sent through the composer, a suggestion
+     * chip clicked, or a host's own `suggestLoopPrompt(text, {submit: true})`
+     * arriving through the prompt store — and this view is the only thing all
+     * three of those already go through. The composer that renders is
+     * swappable (a host may contribute a different one to `LoopChatComposer`
+     * entirely) and does not see the other two paths at all, so watching for
+     * the first send from inside it would miss most of what "first prompt"
+     * needs to mean.
+     */
+    const inputPromptConfig = reactor.getConfig(INPUT_PROMPT_PLUGIN_NAME);
+    /*
+     * The default editor is the opening selection, not something switched on
+     * later.
+     *
+     * It used to be applied by an effect that waited for the surface to become
+     * openable, which meant the picker read `None` until a sandbox existed —
+     * a workspace configured for the notebook opened saying it was configured
+     * for nothing. What the host configured is true from the first render; what
+     * the surface can do about it yet is the surface's business.
+     */
+    /*
+     * Nothing open yet, whatever the host asked for.
+     *
+     * Opening the default on the first render puts an editor on screen before
+     * its surface exists — plugins arrive as they activate — and before the
+     * sandbox it needs is running. What the reader saw then was not the notebook
+     * but the notebook's "needs a running sandbox" placeholder, which is a worse
+     * answer than an empty chat that fills in a moment later.
+     *
+     * The effect below opens it the moment it can actually be opened.
+     */
+    const [surfaceId, setSurfaceId] = useState(NO_SURFACE);
+    /**
+     * Whether the reader has chosen a surface themselves.
+     *
+     * Once they have, the default is spent. Without this it re-applies every
+     * time a new surface lands — so closing the notebook would reopen it as soon
+     * as any other plugin activated, which reads as the close button not working.
+     */
+    const surfaceChosen = useRef(false);
+    const surfaces = useContributions(LoopChatSurface);
+    /* The composer and the title bar arrive as plugins: this view assembles
+       their props, whoever contributed renders them. First contribution wins —
+       these are single-occupant surfaces, not lists. */
+    const composerEntries = useContributions(LoopChatComposer);
+    const headerEntries = useContributions(LoopChatHeader);
+    /* Live per-example chat extras — an error banner, the codemode toggle, the
+       footer's MCP/codemode status. Carried as a signal so the example can
+       update it without the plugin being rebuilt; first contribution wins. */
+    /* How the parts are arranged. No contribution: the split below. A layout
+       plugin — the page layout — takes the same wired parts and arranges them
+       its own way; first contribution wins. */
+    const layoutEntries = useContributions(LoopChatLayout);
+    const layout = layoutEntries[0]?.value;
+    /* The current turn, kept for whoever shows the conversation without the
+       transcript. The chat plugin contributed the feed at build; this view is
+       its writer — begin on send, the reply as it arrives, end when it stops. */
+    const turnEntries = useContributions(LoopChatTurn);
+    const turnFeed = turnWritersOf(turnEntries[0]?.value);
+    /* Reached through a ref from the send and loading handlers, which are
+       memoised with empty deps and must not go stale on it. */
+    const turnFeedRef = useRef(turnFeed);
+    turnFeedRef.current = turnFeed;
+    /* Who is working, for the activity line — read through a ref by a
+       handler memoised with empty deps. Set once the spec is resolved below. */
+    const agentNameRef = useRef('');
+    const extrasEntries = useContributions(LoopChatExtras);
+    const chatExtras = useSignalValue(extrasEntries[0]?.value.extras ?? EMPTY_CHAT_EXTRAS);
+    /* Which editors have ever been mounted; they are never unmounted after —
+       see the render below. A ref, not state: adding to it during render is
+       idempotent and must not schedule another one. */
+    const everMounted = useRef(new Set());
+    // Asked, not assumed: the chat cannot know whether anything is listening,
+    // and whichever plugin does answers the gate. Re-asked on every render with
+    // the live workspace, so switching the sandbox switches the prompt.
+    const usable = useGate(LoopAgentGate, workspace);
+    /* The sandbox's answer. It is not the only thing that can close the chat —
+       an anonymous visitor's trial key runs out too — so the two are combined
+       further down, once the key is known. */
+    const gateBlocked = !usable.allowed;
+    const gateReason = gateBlocked
+        ? (usable.reason ?? 'Chat is unavailable for this sandbox')
+        : undefined;
+    // Stable: ChatBase calls these in effects, and a new identity every render
+    // would make it re-publish on every render.
+    /*
+     * Whether the agent is working, held here rather than read back from the
+     * shell.
+     *
+     * It used to live only in `workspace.viewControls`, written by two
+     * callbacks that each replaced the whole object — and `onSendReady` fires
+     * again during a send, because ChatBase rebuilds its `handleSend` when the
+     * send starts. So the sequence was: loading true, then a fresh
+     * `{ stop }` with no `busy` at all. The prompt went live again one tick
+     * after the agent started, which is exactly when it must not be.
+     *
+     * Local state, and the shell is told from one place below.
+     */
+    const [busy, setBusy] = useState(false);
+    const [sendReady, setSendReady] = useState(false);
+    const handleSendReady = useCallback((controls) => {
+        controlsRef.current = controls;
+        setSendReady(!!controls);
+    }, []);
+    /*
+     * What the agent is doing, for the turn feed.
+     *
+     * The newest tool call that has not returned is the activity: "Analyst is
+     * adding a cell…". Named by what the tool does in the page rather than by
+     * the tool — a reader watching the notebook is told what to look for.
+     * Cleared when nothing is running.
+     */
+    const handleDisplayItemsChange = useCallback((items) => {
+        const feed = turnFeedRef.current;
+        if (!feed) {
+            return;
+        }
+        const SETTLED = new Set([
+            'complete',
+            'completed',
+            'done',
+            'error',
+            'failed',
+            'cancelled',
+            'canceled',
+            'rejected',
+            'denied',
+            'stopped',
+        ]);
+        const VERBS = {
+            insertcell: 'adding a cell',
+            updatecell: 'editing a cell',
+            runcell: 'running a cell',
+            deletecells: 'removing cells',
+            readallcells: 'reading the notebook',
+            readcell: 'reading a cell',
+            executecode: 'running code',
+            insertblock: 'adding a block',
+            updateblock: 'editing a block',
+            runblock: 'running a block',
+            readdocument: 'reading the document',
+        };
+        let running;
+        for (let index = items.length - 1; index >= 0; index -= 1) {
+            const item = items[index];
+            if (typeof item.toolName === 'string' && item.toolCallId) {
+                if (!SETTLED.has(String(item.status).toLowerCase())) {
+                    running = item;
+                }
+                break;
+            }
+        }
+        if (!running) {
+            feed.activity(undefined);
+            return;
+        }
+        const key = running.toolName
+            .toLowerCase()
+            .replace(/^datalayer_/, '')
+            .replace(/_lexical$/, '')
+            .replace(/[^a-z]/g, '');
+        const verb = VERBS[key] ?? `using ${running.toolName}`;
+        feed.activity(`${agentNameRef.current || 'The agent'} is ${verb}…`);
+    }, []);
+    /* Read through a ref by the snapshot handler, memoised with empty deps. */
+    const chatExtrasRef = useRef(chatExtras);
+    chatExtrasRef.current = chatExtras;
+    // The window, for the prompt band, the turn feed, and a host counting it.
+    const handleContextSnapshot = useCallback((snapshot) => {
+        setChatUsage(snapshot);
+        turnFeedRef.current?.usage(snapshot);
+        chatExtrasRef.current.onUsage?.(snapshot);
+    }, []);
+    const handleLoadingChange = useCallback((next) => {
+        setBusy(next);
+        if (!next) {
+            turnFeedRef.current?.end('done');
+        }
+    }, []);
+    /*
+     * The reply as it arrives, for the turn feed.
+     *
+     * `onMessagesChange` reports the conversation on every change of it,
+     * streaming included; the last assistant message is the one the current
+     * turn is about. Text only — a tool call is shown where it acted, in the
+     * editor, and the panel is for what the agent *said*.
+     */
+    const handleMessagesChange = useCallback((messages) => {
+        const feed = turnFeedRef.current;
+        if (!feed) {
+            return;
+        }
+        /*
+         * The whole reply since the person last spoke, not the latest segment.
+         *
+         * An agent that calls a tool speaks in more than one message per turn —
+         * "I'll start…", the tool, "Good, the frame shows…" — and a reader
+         * watching the panel wants all of it, in order. Text only: the tool's
+         * work shows where it acted, in the editor.
+         */
+        let lastUser = -1;
+        for (let index = messages.length - 1; index >= 0; index -= 1) {
+            if (messages[index].role === 'user') {
+                lastUser = index;
+                break;
+            }
+        }
+        const textOf = (message) => {
+            const content = message.content;
+            return typeof content === 'string'
+                ? content
+                : content
+                    .map(part => typeof part === 'string'
+                    ? part
+                    : (part.text ?? ''))
+                    .join('');
+        };
+        const reply = messages
+            .slice(lastUser + 1)
+            .filter(message => message.role === 'assistant')
+            .map(textOf)
+            .filter(text => text.trim().length > 0)
+            .join('\n\n');
+        feed.assistant(reply);
+    }, []);
+    // One writer, so neither fact can erase the other.
+    useEffect(() => {
+        setViewControls(sendReady
+            ? {
+                busy,
+                stop: controlsRef.current?.stop,
+                newChat: controlsRef.current?.newChat,
+            }
+            : null);
+    }, [busy, sendReady, setViewControls]);
+    // Stop reporting when the view goes away, or the shell would keep a stop
+    // button for something that is no longer there.
+    useEffect(() => () => setViewControls(null), [setViewControls]);
+    useEffect(() => workspace.prompts.subscribe(message => {
+        // A prompt typed before the adapter is ready is dropped rather than
+        // queued: silently sending it minutes later, out of context, is worse
+        // than saying nothing happened.
+        controlsRef.current?.send(message);
+    }), [workspace.prompts]);
+    /*
+     * Fired once, on the first real send, for the life of this mounted chat —
+     * a ref rather than a signal, because nothing here needs to re-render when
+     * it flips; the host's hook is the only thing that reacts.
+     */
+    const firstPromptFired = useRef(false);
+    const handleSend = useCallback(async (message) => {
+        /*
+         * Every way a message reaches this view runs through here — typed and
+         * sent through the composer, a suggestion chip clicked, or a host's own
+         * `suggestLoopPrompt(text, {submit: true})` arriving through the prompt
+         * store (see the `suggestion.submit` effect below) — so this is the one
+         * place "the first prompt" means the same thing regardless of which one
+         * a person or a host used.
+         */
+        if (!firstPromptFired.current) {
+            firstPromptFired.current = true;
+            inputPromptConfig?.firstPromptHook?.();
+        }
+        setTransient(null);
+        // A new turn: whatever the panel showed is gone, this message is it.
+        turnFeedRef.current?.begin(message);
+        const outcome = await workspace.submit(message);
+        if (!outcome.handled) {
+            /*
+             * Nothing took it, so the words stay where they were typed.
+             *
+             * The reason appears under the prompt; emptying the box as well would
+             * make a failed send look like a successful one and cost the person
+             * the sentence they wrote.
+             */
+            setTransient(outcome.reason ?? null);
+            return;
+        }
+        /*
+         * Sent, so the composer is empty again.
+         *
+         * `InputPromptBase` clears itself only when it owns its own text. This
+         * view hands it a controlled value — a page embedding the workspace can
+         * offer a prompt, which needs somewhere outside the editor to put it —
+         * and a controlled value is the owner's to clear. It was not being
+         * cleared at all: the sentence stayed in the box after the agent had
+         * answered it, so the next thing typed was appended to the last thing
+         * asked, and the placeholder that suggests what to ask never came back.
+         */
+        setDraft('');
+        if (outcome.result?.content) {
+            setTransient(outcome.result.content);
+        }
+    }, [workspace, inputPromptConfig?.firstPromptHook]);
+    const active = useMemo(() => surfaces.find(entry => entry.value.surfaceId === surfaceId)?.value, [surfaces, surfaceId]);
+    /**
+     * The reader picking a surface, including picking none.
+     *
+     * It records the choice as well as making it: from here on the host's
+     * default is spent, so a surface closed stays closed.
+     */
+    const chooseSurface = useCallback((next) => {
+        surfaceChosen.current = true;
+        setSurfaceId(next);
+    }, []);
+    /*
+     * A command asking for a surface — `/notebook`, `/document`, or their
+     * shortcuts. Treated exactly like a reader picking one from the strip, so
+     * the choice sticks the same way and the host's default is spent.
+     */
+    useEffect(() => onSurfaceRequest(chooseSurface), [chooseSurface]);
+    /*
+     * The editor choice store — what the shell's selector, the `/editor`
+     * command and the reactor shell's cycle read — learns two things only the
+     * chat knows: which editors exist, and which one is on screen. Nobody fed
+     * it the options before, so "cycle the editors" cycled an empty list and
+     * always landed on none; and a surface opened by a tool or a plugin, past
+     * the store, left it pointing at the wrong one. Seeded, not chosen: the
+     * chat already shows it, there is nothing to announce.
+     */
+    useEffect(() => {
+        setEditorOptions(surfaces.map(entry => entry.value.surfaceId));
+    }, [surfaces]);
+    useEffect(() => {
+        seedEditorChoice(surfaceId && surfaceId !== NO_SURFACE ? surfaceId : NONE_EDITOR);
+    }, [surfaceId]);
+    /** The surface the host asked to open, once its plugin has contributed it. */
+    const wanted = useMemo(() => defaultSurface === NO_SURFACE || defaultSurface === 'none'
+        ? undefined
+        : surfaces.find(entry => entry.value.surfaceId === defaultSurface), [surfaces, defaultSurface]);
+    /*
+     * Open the default the moment it can be opened.
+     *
+     * Both conditions are the point. A surface arrives when its plugin
+     * activates, so it may not exist on the first render; and these two need a
+     * running sandbox, which takes as long as a Pyodide kernel takes to start.
+     * Waiting for both is what turns "configured for the notebook" into a
+     * notebook rather than a placeholder.
+     */
+    useEffect(() => {
+        if (surfaceChosen.current || !wanted) {
+            return;
+        }
+        if (canOpenView(wanted.value, workspace)) {
+            setSurfaceId(defaultSurface);
+        }
+    }, [wanted, workspace, defaultSurface]);
+    /*
+     * An editor that stops being openable — its sandbox went away — should not
+     * stay on screen claiming otherwise.
+     *
+     * Closed, but remembered. This used to set the surface to none and stop
+     * there, which spent the reader's choice on an interruption: a sandbox
+     * blinking during a target switch closed the notebook, `surfaceChosen`
+     * meant the default could not reopen it, and the control settled on None
+     * for a workspace nobody had asked to empty. What was on screen and what
+     * the control said agreed with each other and with neither the reader nor
+     * the configuration.
+     */
+    /* State, not a ref: the layout below keeps this surface's column while it
+       is away, and a ref changing re-renders nothing. */
+    const [suspendedId, setSuspendedId] = useState(null);
+    /*
+     * Full screen: the whole view, on the whole screen.
+     *
+     * Everything the workspace is showing goes — the editor as well as the
+     * conversation. An earlier version hid the editor, on the theory that a long
+     * transcript wants the room; but somebody working on a notebook who asks for
+     * more room is asking for more room *for the work*, and taking the notebook
+     * away to give it is answering a question they did not ask.
+     *
+     * Done with the Fullscreen API rather than by drawing a big box, because a
+     * component cannot know what it is inside of. `position: fixed` escapes only
+     * as far as the nearest ancestor with a transform, and a page that animates
+     * its sections in — the landing does — leaves one behind permanently: the
+     * "full screen" chat then filled the card it was already in. The API
+     * promotes the element to the browser's top layer, where no ancestor can
+     * hold it, and unlike a portal it does not move in the DOM — so every
+     * inherited theme variable the editors read still resolves.
+     *
+     * The CSS overlay is kept as the fallback for where the API is refused: an
+     * iframe without `allow="fullscreen"`, mostly. Confined to a card is a worse
+     * answer than full screen and a better one than nothing.
+     */
+    const viewRef = useRef(null);
+    // A host's own fixed header, drawn outside this workspace entirely —
+    // reserved at the top of the overlay below, and forcing full screen onto
+    // the overlay door for good; see `ChatPluginConfig.fullScreenTopOffset`.
+    const fullScreenTopOffset = config?.fullScreenTopOffset ?? 0;
+    /* Shared with the workspace header's icon — same machinery, two doors in.
+       See `useWorkspaceFullScreen` for why it is the browser's API rather than
+       a big box, and for the CSS fallback the styling below paints. */
+    const { fullScreen, usingApi: usingFullscreenApi, toggle: toggleFullScreen, topOffsetPx: fullScreenTopOffsetPx, } = useWorkspaceFullScreen(viewRef, {
+        forceOverlay: Boolean(fullScreenTopOffset),
+        topOffset: fullScreenTopOffset,
+    });
+    /*
+     * The full-screen control, pointed at from the first run onwards.
+     *
+     * This view is usually a panel inside somebody else's page — a few hundred
+     * pixels of a landing page or an example. The moment it becomes worth more
+     * room is the moment it starts answering, and that is also the moment a
+     * reader's attention is on the transcript rather than on a small grey icon
+     * in a corner. So from the first answer the icon takes the theme's own
+     * colour and breathes.
+     *
+     * It does not stop on a timer. A reader who is deep in the first answer has
+     * not yet had the thought "this is too small", and a hint that has already
+     * expired by the time they have it is a hint that was never given. What ends
+     * it is the reader: taking the offer, or reaching the thing it was offering.
+     *
+     * `hinted` is a ref because it must not itself cause a render, and because
+     * it has to survive `busy` going true again for the second message — this
+     * starts once, on the first run, not on every turn.
+     */
+    const [hintFullScreen, setHintFullScreen] = useState(false);
+    const hinted = useRef(false);
+    useEffect(() => {
+        if (!busy || hinted.current || fullScreen) {
+            return;
+        }
+        hinted.current = true;
+        setHintFullScreen(true);
+    }, [busy, fullScreen]);
+    // Arriving is the end of it, however they got there — the button, a
+    // keyboard shortcut, the host promoting the workspace. There is nothing
+    // left to point at, and it never comes back.
+    useEffect(() => {
+        if (fullScreen) {
+            setHintFullScreen(false);
+        }
+    }, [fullScreen]);
+    useEffect(() => {
+        if (active && !canOpenView(active, workspace)) {
+            setSuspendedId(active.surfaceId);
+            setSurfaceId(NO_SURFACE);
+        }
+    }, [active, workspace]);
+    /**
+     * The surface that is closed but coming back.
+     *
+     * Held so the layout can keep its column: what is missing is the editor,
+     * not the half of the workspace it lives in.
+     */
+    const waiting = useMemo(() => {
+        if (suspendedId) {
+            return surfaces.find(item => item.value.surfaceId === suspendedId)?.value;
+        }
+        /*
+         * The host's default, before it has managed to open.
+         *
+         * It needs a sandbox, and a sandbox takes as long as a kernel takes to
+         * start — so for the first seconds of every workspace there is a
+         * configured notebook that is not open yet. Counting it as "waiting" is
+         * what lets the control say Notebook from the outset and the column hold
+         * its half of the split, instead of both saying None and then changing
+         * their minds.
+         *
+         * Only until somebody chooses for themselves. After that the default is
+         * spent, and a reader who picked None must not be shown Notebook.
+         */
+        if (!surfaceChosen.current && surfaceId === NO_SURFACE) {
+            return wanted?.value;
+        }
+        return undefined;
+    }, [surfaces, suspendedId, surfaceId, wanted]);
+    // And opened again when it can be. The reader asked for it once; an
+    // interruption is not them changing their mind.
+    useEffect(() => {
+        if (!suspendedId || surfaceId !== NO_SURFACE) {
+            return;
+        }
+        const entry = surfaces.find(item => item.value.surfaceId === suspendedId);
+        if (entry && canOpenView(entry.value, workspace)) {
+            setSuspendedId(null);
+            setSurfaceId(suspendedId);
+        }
+    }, [surfaces, surfaceId, suspendedId, workspace]);
+    // The AG-UI mount is per agent, and `default` is the agent a server runs when
+    // it was not told otherwise, so a workspace opened without one still talks to
+    // something.
+    const agentId = workspace.agentId?.trim() || 'default';
+    /*
+     * Which protocol, decided by where the sandbox is.
+     *
+     * On the browser target there is no server to mount an agent on, so the loop
+     * turns in this page with the Vercel AI SDK and the chat calls it rather
+     * than addressing it. Everywhere else the agent lives on the runtime and the
+     * chat speaks AG-UI to it, exactly as before.
+     *
+     * The spec's own `harness` does not decide this. It says where the agent
+     * normally runs, and a page cannot turn a pydantic-ai loop whatever it asks
+     * for — so the location wins, which is the same rule the examples follow.
+     */
+    const inPage = targetRunsAgentInPage(workspace.sandbox.target ?? 'local');
+    /*
+     * Who is being addressed, when this workspace runs a team.
+     *
+     * The team's own answer, not the workspace's: `workspace.agentId` is the
+     * session's agent, and a team has several behind one front door. Picking a
+     * member changes which spec the next prompt is answered by, and nothing
+     * else — the sandbox, the notebook and the transcript stay where they are.
+     */
+    const team = useOptionalTeamSelection();
+    const selectedMemberId = useSignalValue(team?.selected ?? EMPTY_SELECTION);
+    const member = team?.members.find(entry => entry.id === selectedMemberId);
+    /*
+     * Looked up through `getAgentspecs`, not by subscripting the record.
+     *
+     * Agent ids reach this component versioned as often as not —
+     * `jupyter-data-analyst:0.0.1` — and the record is keyed on the bare id, so
+     * a raw subscript on a versioned id silently misses. What follows from a
+     * miss is not an error but a quiet wrong answer: no model, no openers, no
+     * name. `getAgentspecs` strips the version and retries, which is the whole
+     * reason the specs package exports it.
+     */
+    /*
+     * The spec, found two ways.
+     *
+     * By the agent's id first — a team member's spec id, or a host that names
+     * the spec as the agent, as the Loop Shell does. Failing that, by the
+     * blueprint a capacity plugin contributed: an agent it created on a server
+     * carries an instance name (`subagents-example-agent-…`), not a spec id, and
+     * the same agent turned in this page had no prompt, no subagents and no
+     * icon for want of this lookup — it used the document tools and called
+     * nobody.
+     */
+    const blueprintSpecId = useContributions(LoopAgentBlueprint)[0]?.value.specId;
+    const spec = getAgentspecs(member?.specId ?? agentId) ??
+        (blueprintSpecId ? getAgentspecs(blueprintSpecId) : undefined);
+    /*
+     * The team's name for a member first, then the spec's own.
+     *
+     * A spec is named for the catalogue — "Jupyter Data Analyst" — and a team
+     * names the same agent for the person working with it — "Analyst", which is
+     * what the picker in the header says. The header, the empty state and the
+     * openers' heading say the same word as the picker, or a person reads two
+     * names for one agent on one screen.
+     */
+    agentNameRef.current = member?.name ?? spec?.name ?? agentId;
+    /*
+     * The openers: a capacity plugin's first, then the team's with the
+     * addressed member's own beneath, then — with no team — the agent's own.
+     *
+     * A contribution to `LoopChatSuggestion` is a deliberate statement about
+     * what this workspace is worth asking, so it replaces the spec's generic
+     * list rather than joining it.
+     *
+     * With a team, its openers are always shown: they are what the *team* can
+     * be asked, whoever is addressed, and a person who has just opened the
+     * workspace should see the whole before the part. Under them, in a block
+     * with its name on it, come the addressed member's own openers — picking
+     * the Reviewer used to change nothing on screen, and the reverse (the
+     * member's only) hid the team. An opener the team already lists is not
+     * repeated.
+     */
+    const suggestionEntries = useContributions(LoopChatSuggestion);
+    const hiddenOpeners = chatExtras.hiddenOpeners;
+    const suggestions = useMemo(() => {
+        // What the host took off this page, by the chip's words — the same key
+        // the two-group merge below uses to tell openers apart.
+        const hidden = new Set(hiddenOpeners ?? []);
+        const offered = (items) => hidden.size === 0 ? items : items.filter(item => !hidden.has(item.text));
+        const contributed = suggestionEntries.flatMap(entry => entry.value.suggestions);
+        if (contributed.length > 0) {
+            return offered(contributed);
+        }
+        const own = offered(spec?.suggestions ?? []);
+        const teams = offered(team?.team.suggestions ?? []);
+        if (teams.length === 0) {
+            return own;
+        }
+        // Two groups, named for the two levels the empty state draws: the
+        // team's openers under the team's name, the member's own under its.
+        const listed = new Set(teams.map(item => item.text));
+        const teamName = team?.team.name ?? 'Team';
+        const memberName = member?.name ?? spec?.name ?? agentId;
+        return [
+            ...teams.map(item => ({ ...item, group: teamName })),
+            ...own
+                .filter(item => !listed.has(item.text))
+                .map(item => ({ ...item, group: memberName })),
+        ];
+    }, [suggestionEntries, team, spec, member, agentId, hiddenOpeners]);
+    /*
+     * The same openers in the shape the chat's empty state asks for.
+     *
+     * `Suggestion` is a title and a message — what a chip shows and what it
+     * sends. A spec's opener with a `summary` shows that and sends its text;
+     * one without shows the whole request; a contributed opener may split the
+     * two its own way.
+     */
+    const chatSuggestions = useMemo(() => suggestions.map(item => ({
+        title: item.summary ?? item.text,
+        message: item.message ?? item.text,
+        group: item.group,
+    })), [suggestions]);
+    /* The icon its spec asked for, at the size the empty state draws. */
+    const BrandIcon = agentIcon(spec?.icon);
+    /* And the team's, for the level above it. */
+    const TeamIcon = agentIcon(team?.team.icon);
+    /*
+     * How an in-page agent reaches the notebook.
+     *
+     * On every other target the agent runs on a server with a sandbox of its
+     * own, and "run this cell" happens there — in the same kernel the notebook
+     * is bound to, which is why it appears to work by magic. In the browser
+     * there is no server and no second kernel: the only way to touch the
+     * notebook is a frontend tool, executed here.
+     *
+     * Without these the browser agent could hold a conversation and do nothing,
+     * which is exactly how it behaved — it answered, and the cell never ran.
+     *
+     * From the chat's own extension point rather than a hard-wired notebook
+     * import: the notebook plugin contributes its cell tools, the document its
+     * lexical ones, and a chat mounted without either simply has fewer tools.
+     * Each factory runs against the live workspace, so tools addressed by
+     * `surfaceId` can never point at a different notebook than the one on
+     * screen.
+     */
+    const toolContributions = useContributions(LoopFrontendTool);
+    const notebookTools = useMemo(() => {
+        /*
+         * One tool per name, first contribution wins — so the order is decided
+         * here rather than by which lazy import landed first. The notebook and
+         * the document each ship a kernel tool — `executeCodeInNotebook` runs in the
+         * sandbox and streams onto the conversation, the document's runs on the
+         * document's own kernel and returns silently. The editor on screen owns
+         * the shared names; with none on screen the notebook does, since its
+         * outputs are what the conversation shows. The document's block tools
+         * (runBlock, insertBlock…) keep their own names and are untouched.
+         */
+        const merged = [];
+        const owners = new Map();
+        // Tools a contribution vouched for in the chat view; see `chatView`.
+        const keep = new Set();
+        // See `isInactiveSurfaceContribution`: with an editor on screen, only
+        // its own `${surfaceId}-tools` contribution survives this filter.
+        const surfaceIds = new Set(surfaces.map(entry => entry.value.surfaceId));
+        for (const entry of orderToolContributions(toolContributions, active?.surfaceId)) {
+            if (isInactiveSurfaceContribution(entry.value.id, surfaceIds, active?.surfaceId)) {
+                continue;
+            }
+            let tools;
+            try {
+                tools = entry.value.tools(workspace);
+            }
+            catch (error) {
+                // One broken contributor must not cost the agent every tool.
+                console.warn(`[loop] frontend tools from '${entry.plugin}' failed:`, error);
+                continue;
+            }
+            for (const tool of tools) {
+                const owner = owners.get(tool.name);
+                if (owner) {
+                    console.warn(`[loop] frontend tool '${tool.name}' from '${entry.plugin}' is shadowed by the one from '${owner}'.`);
+                    continue;
+                }
+                owners.set(tool.name, entry.plugin);
+                merged.push(tool);
+                if (entry.value.chatView) {
+                    keep.add(tool.name);
+                }
+            }
+        }
+        // A host example's own frontend tools, folded in last (first name wins),
+        // so a bespoke client tool like A2UI's run_jupyter_output_demo joins the
+        // agent's toolset without a plugin of its own.
+        for (const tool of chatExtras.frontendTools ?? []) {
+            if (owners.has(tool.name)) {
+                continue;
+            }
+            owners.set(tool.name, 'chat-extras');
+            merged.push(tool);
+        }
+        /*
+         * The view decides the toolset. With an editor on screen the agent has
+         * that editor's own tools (`belongsToInactiveSurface` above, plus
+         * whatever isn't tied to a surface at all); with none — the chat view —
+         * it keeps only what leaves the editors alone: the `executeCodeIn…`
+         * tools, whose outputs stream onto the conversation as surfaces, and
+         * the read tools. The editors stay mounted out of sight either way, so
+         * what it can still call, works.
+         */
+        return active ? merged : toolsForChatView(merged, keep);
+        // The workspace object is rebuilt when its facts change; surfaceId is
+        // the one the factories address by.
+        // eslint-disable-next-line react-hooks/exhaustive-deps
+    }, [
+        toolContributions,
+        workspace.surfaceId,
+        workspace.agentId,
+        chatExtras.frontendTools,
+        active?.surfaceId,
+        surfaces,
+    ]);
+    /*
+     * The commands the mounted plugins declared for an agent — the reactor's
+     * `AgentTools` point — executed on this workspace's own reactor. Folded in
+     * after the contributions, which keep their names, so a plugin with a
+     * richer in-page implementation of the same tool (the decks plugin's) is
+     * what the agent gets, and the bundle's command is the fallback. The agent
+     * spec names none of these: they follow the plugin.
+     */
+    const commandTools = useAgentCommandTools();
+    const agentTools = useMemo(() => {
+        if (commandTools.length === 0) {
+            return notebookTools;
+        }
+        const taken = new Set(notebookTools.map(tool => tool.name));
+        return [
+            ...notebookTools,
+            ...commandTools.filter(tool => !taken.has(tool.name)),
+        ];
+    }, [notebookTools, commandTools]);
+    /*
+     * Where this workspace's agent actually lives.
+     *
+     * A Datalayer runtime brings its own agent-runtimes server: the agent is
+     * created on the pod from an agentspec, not on the server the workspace was
+     * opened against. Addressing the latter would be talking to a machine that
+     * has never heard of it. The same value for every other target, where the
+     * two are the same server.
+     */
+    const agentServerUrl = workspace.sandbox.agentBaseUrl || workspace.serverUrl;
+    /*
+     * Who a single request may be addressed to.
+     *
+     * The same set the harness gets as tools, and deliberately so: a menu that
+     * offered a name the model could not reach would be a menu that produces
+     * silent no-ops. Derived here rather than in the plugin because it changes
+     * with the selected member.
+     */
+    const mentionable = useMemo(() => {
+        if (!team || !member) {
+            // No team: the spec's own subagents, which the in-page loop can reach
+            // and a server-side one already could.
+            return specSubagents(spec).map(subagent => ({
+                name: subagent.name,
+                description: subagent.description,
+                icon: agentIcon(subagent.icon),
+            }));
+        }
+        /*
+         * The team, plus whatever specialists the selected member brings.
+         *
+         * It used to be `subagentsFor(member)` alone, which made the list read as
+         * arbitrary: a team of two offered one name, and the name was whichever
+         * member you were *not* talking to. Everyone the team contains is listed,
+         * because "who is on this team" is the question a person types `@` to
+         * ask; the one already being addressed is shown greyed rather than
+         * dropped, so the list is stable as the selection moves.
+         */
+        const roster = team.members.map(entry => ({
+            name: entry.name,
+            description: entry.description,
+            emoji: entry.emoji,
+            icon: agentIcon(entry.icon),
+            disabled: entry.id === member.id,
+            disabledReason: entry.id === member.id
+                ? `You are already talking to ${entry.name}`
+                : undefined,
+        }));
+        const specialists = subagentsFor(team.team, member.id)
+            .filter(subagent => !team.members.some(m => m.name === subagent.name))
+            .map(subagent => ({
+            name: subagent.name,
+            description: subagent.description,
+            icon: agentIcon(subagent.icon),
+        }));
+        return [...roster, ...specialists];
+    }, [team, member, spec]);
+    /*
+     * Where the model is reached, and on whose key.
+     *
+     * Told whether an in-page agent is going to ask at all: a workspace on a
+     * server-backed agent authenticates itself, and minting a visitor's one
+     * trial key for it would start a clock against a conversation that never
+     * spends it.
+     */
+    const { inference, anonymous } = useBrowserInference(inPage);
+    /* Read from the store rather than passed in: the sign-in form inside the
+       expiry panel writes here, so the same render that gains a member gains
+       the wording that goes with one. */
+    const signedInUser = useIAMStore(state => state.user);
+    /* The trial is over, and this chat is the thing that stops working. Only
+       for an in-page agent: everywhere else the runtime holds its own
+       credentials and never saw the visitor's key. */
+    const keyExpired = inPage && anonymous.status === 'expired';
+    /*
+     * Whose key ran out, which decides what the chat is allowed to call it.
+     *
+     * The anonymous store only ever holds the trial key, so reaching this
+     * through `anonymous.status` means the reader never signed in. A member
+     * whose own session expires arrives by a different route and must not be
+     * told a temporary key ran out — they never had one, and would go looking
+     * for something that was never theirs.
+     */
+    const expiredKeyIsTemporary = !signedInUser;
+    const chatDisabled = gateBlocked || keyExpired;
+    const disabledReason = keyExpired
+        ? expiredKeyIsTemporary
+            ? 'This demo runs on a shared key, and its time is up. Sign in to keep going.'
+            : 'Your key has expired. Sign in to keep going.'
+        : gateReason;
+    /*
+     * The prompt's text, held here rather than inside `InputPrompt`.
+     *
+     * A page embedding the workspace can offer a request — "try asking it
+     * this" — and the only honest place to put such an offer is the box the
+     * visitor would have typed it into, where they can read it, change their
+     * mind, or send it. That needs a controlled prompt, which needs the text to
+     * live somewhere both the suggestion and the composer can reach.
+     */
+    const [draft, setDraft] = useState('');
+    /*
+     * The model this conversation is on.
+     *
+     * Three answers, and the order matters more than it looks. The reader's own
+     * pick wins, then the *agent's spec*, then whatever the host mounted the
+     * workspace with.
+     *
+     * The spec used to come last, behind `workspace.model` — which is a
+     * workspace-wide default, set once for whichever agent happens to be
+     * addressed. So launching an agent whose spec names one model showed a
+     * different one ticked in the menu, and the footer disagreed with the thing
+     * actually answering. An agent's spec is the definition of that agent; a
+     * host default is a fallback for when there is no definition, and a fallback
+     * that overrides the definition is not a fallback.
+     *
+     * Held here rather than in the workspace because the choice belongs to the
+     * conversation: it is what the next message is sent with, and it resets when
+     * the member being addressed changes.
+     */
+    const [pickedModel, setPickedModel] = useState();
+    const activeModel = pickedModel ?? spec?.model ?? workspace.model ?? '';
+    /*
+     * The team, in the shape the footer asks for.
+     *
+     * The footer knows nothing about teamspecs, and should not: it renders a
+     * list of names and reports which one was chosen. Everything that makes a
+     * member a member stays on this side of the boundary.
+     */
+    /* What is left of the context window, as the runtime last reported it. The
+       footer shows nothing at all until there is a real number, so a workspace
+       whose agent has not answered yet simply has no bar. */
+    /*
+     * The window, and only for the agent being addressed now.
+     *
+     * The store keeps one snapshot rather than one per agent, so switching
+     * member left the previous one's numbers on screen under the new one's
+     * name. Ignored until the store hands back a different object, which is the
+     * only evidence available that the new agent has reported.
+     */
+    const storedUsage = useAgentRuntimeContextSnapshot();
+    const usageAtSwitch = useRef(storedUsage);
+    const addressedAgent = useRef(selectedMemberId);
+    const storedUsageRef = useRef(storedUsage);
+    storedUsageRef.current = storedUsage;
+    useEffect(() => {
+        if (addressedAgent.current === selectedMemberId) {
+            return;
+        }
+        addressedAgent.current = selectedMemberId;
+        usageAtSwitch.current = storedUsageRef.current;
+        // And the model goes back to whatever the new agent's spec asks for. A
+        // pick made for one member is not a statement about the next.
+        setPickedModel(undefined);
+    }, [selectedMemberId]);
+    const fromStore = storedUsage && storedUsage !== usageAtSwitch.current ? storedUsage : null;
+    /*
+     * What the chat itself accounts for, when nothing else does.
+     *
+     * A server-side runtime pushes a snapshot over the socket and that is what
+     * `fromStore` holds. An in-page agent has no server to push one, and its
+     * only account of the window is the totals its harness reports — which
+     * `ChatBase` assembles and hands out. Without this the browser agent's
+     * footer had no numbers at all, and the space kept for them showed as a
+     * white stripe under the prompt.
+     */
+    const [chatUsage, setChatUsage] = useState();
+    const contextUsage = fromStore ?? chatUsage ?? null;
+    // Read to decide whether the strip exists at all, not to render it.
+    const inpromptMenu = useSlotComponents(LoopSlots.inpromptMenu);
+    // Buttons for the prompt's footer bar, beside the session menus. Asked
+    // whether anyone contributed before anything is drawn, for the same reason
+    // as the in-prompt menu: an empty slot is still an element.
+    const promptActions = useSlotComponents(LoopSlots.promptAction);
+    // Same question for the chat's own title bar, which plugins may add to.
+    const chatHeaderItems = useSlotComponents(LoopSlots.chatHeader);
+    const footerAgents = useMemo(() => team?.members.map(entry => ({
+        id: entry.id,
+        name: entry.name,
+        description: entry.description,
+        icon: agentIcon(entry.icon),
+    })) ?? [], [team]);
+    const suggestion = useLoopPromptStore(state => state.pending);
+    /*
+     * `/prompt`, or its keystroke, asking for the caret.
+     *
+     * A command runs outside the component tree, so the ask arrives on the
+     * focus channel; each one bumps a nonce that joins the suggestion's in
+     * `focusTrigger` below. The chat answers wherever its composer is —
+     * docked, in-column, or floating.
+     */
+    const [focusAsked, setFocusAsked] = useState(0);
+    useEffect(() => onPromptFocusRequest(() => setFocusAsked(nonce => nonce + 1)), []);
+    useEffect(() => {
+        if (!suggestion) {
+            return;
+        }
+        setDraft(suggestion.text);
+        // Taken, so a second click on the same button offers it again rather than
+        // finding the store already holding it.
+        useLoopPromptStore.getState().consume();
+        if (suggestion.submit) {
+            void handleSend(suggestion.text);
+        }
+        // `handleSend` is stable for the life of a connection; re-running this on
+        // a new identity would resend the last suggestion.
+        // eslint-disable-next-line react-hooks/exhaustive-deps
+    }, [suggestion]);
+    const protocol = useMemo(() => inPage
+        ? browserProtocolConfig({
+            agentId: member?.specId ?? agentId,
+            instructions: spec?.systemPrompt,
+            model: activeModel || undefined,
+            inference,
+            // Who this member may hand work to, and what they are told. Both
+            // come from the teamspec: the supervisor routes to the others, a
+            // member reaches only its own specialists, and `context.sharing`
+            // decides whether the child sees this conversation.
+            // The agent runs them itself, so they go to the harness rather
+            // than to the chat — handing them to both would run each tool
+            // twice.
+            frontendTools: agentTools,
+            // A team's members and specialists when there is a team; the
+            // spec's own subagents otherwise — a researcher and a writer are
+            // as much the agent's as its prompt is.
+            subagents: team && member
+                ? subagentsFor(team.team, member.id)
+                : specSubagents(spec),
+            sharing: team?.sharing,
+            // What a delegated run does lands in the same store the
+            // server's monitoring stream fills, so the side panel and the
+            // composer's pulse draw an in-page delegation like any other.
+            onSubagentEvent: event => agentRuntimeStore.getState().appendSubagentEvent(event),
+        })
+        : {
+            type: 'ag-ui',
+            endpoint: `${agentServerUrl}/api/v1/ag-ui/${agentId}/`,
+            agentId,
+            // `/api/v1/configure`, not `/api/v1/configure/config`: the hooks
+            // strip one trailing `config`/`configure` segment to find the API
+            // base, so the longer form doubles it.
+            configEndpoint: `${agentServerUrl}/api/v1/configure`,
+            enableConfigQuery: true,
+        }, [
+        agentId,
+        inPage,
+        inference,
+        member,
+        agentTools,
+        team,
+        spec,
+        activeModel,
+        agentServerUrl,
+    ]);
+    /*
+     * The model catalogue, and the builtin tools and skills the agent reports.
+     *
+     * Read here rather than left to `ChatBase`, because this view draws its own
+     * prompt: the toolbar under it can only offer what this component hands it.
+     *
+     * `useSkills` reads the runtime's codemode status out of the store, so it
+     * needs no endpoint and works for an in-page agent too. `useConfig` does
+     * need one, and a browser agent has none — hence the menus below being
+     * gated on the data rather than switched on unconditionally.
+     */
+    const configQuery = useConfig(!inPage, protocol.type === 'ag-ui' ? protocol.configEndpoint : undefined, undefined, agentId);
+    const skillsQuery = useSkills(true);
+    const skillActions = useSkillActions(agentId);
+    const [catalogModels, setCatalogModels] = useState([]);
+    useEffect(() => {
+        /*
+         * Only where there is a server to ask.
+         *
+         * An in-page agent has none, and `agentServerUrl` then holds whatever the
+         * host defaulted to — on a public page that was a developer's localhost,
+         * so every visitor's console carried a connection-refused error for a
+         * catalogue that was never going to arrive. The agentspecs catalogue below
+         * is the answer for this case and needs no request at all.
+         */
+        if (inPage) {
+            return undefined;
+        }
+        let cancelled = false;
+        void fetch(`${agentServerUrl}/api/v1/configure/models`)
+            .then(response => (response.ok ? response.json() : { models: [] }))
+            .then(payload => {
+            if (!cancelled) {
+                setCatalogModels(payload.models ?? []);
+            }
+        })
+            // No catalogue is not an error worth a banner: the menu simply has
+            // nothing to offer, and the agent answers on whatever it was given.
+            .catch(() => undefined);
+        return () => {
+            cancelled = true;
+        };
+    }, [agentServerUrl, inPage]);
+    /*
+     * What to offer when the server has no catalogue of its own.
+     *
+     * `/configure/models` answers for a server-backed agent and is
+     * authoritative. An in-page agent has no server to ask, and a model menu
+     * with nothing in it is a control that opens onto an empty list — so the
+     * agentspecs catalogue stands in, filtered to the models marked available.
+     */
+    const offeredModels = useMemo(() => catalogModels.length > 0
+        ? catalogModels
+        : Object.values(AI_MODEL_CATALOGUE)
+            .filter(model => model.available)
+            .map(model => ({
+            id: model.id,
+            name: model.name,
+            provider: model.provider,
+        })), [catalogModels]);
+    /*
+     * The tools the agent can call, from both sides.
+     *
+     * The server's builtins, and the notebook tools this workspace hands the
+     * harness — which are the only ones an in-page agent has, and the reason
+     * "run this cell" works at all. Listing a tool and running it are different
+     * jobs, so a tool the chat does not execute still belongs in the menu.
+     */
+    const offeredTools = useMemo(() => {
+        const fromServer = configQuery.data?.builtinTools ?? [];
+        const seen = new Set(fromServer.map(tool => tool.name));
+        return [
+            ...fromServer,
+            ...agentTools
+                .filter(tool => !seen.has(tool.name))
+                .map(tool => ({ id: tool.name, name: tool.name })),
+        ];
+    }, [configQuery.data?.builtinTools, agentTools]);
+    /* Which skills are on, as the runtime last reported them. Derived rather
+       than held: the source of truth is the agent, and a local copy would drift
+       the moment a skill was enabled from anywhere else. */
+    const enabledSkills = useMemo(() => new Set((skillsQuery.data?.skills ?? [])
+        // `enabled` and `loaded` both mean the agent has it; `available`
+        // means it could. There is no `disabled` — a skill that is off is
+        // simply one that is merely available.
+        .filter(skill => skill.status === 'enabled' || skill.status === 'loaded')
+        .map(skill => skill.id)), [skillsQuery.data]);
+    /*
+     * Taken here first, then told to the server.
+     *
+     * Locally, because an in-page agent has no server to tell and its model is
+     * whatever this view hands the harness — without recording the choice, the
+     * menu ticked a row for a moment and the next message went to the old model.
+     * And on the server for the agents that live there, where the chip in its
+     * header makes the same call: whichever control a person used, the other has
+     * to agree.
+     */
+    const selectModel = useCallback(async (model) => {
+        setPickedModel(model);
+        if (inPage) {
+            return;
+        }
+        await fetch(`${agentServerUrl}/api/v1/configure/inference/provider`, {
+            method: 'PUT',
+            headers: { 'content-type': 'application/json' },
+            body: JSON.stringify({ model }),
+        }).catch(() => undefined);
+    }, [agentServerUrl, inPage]);
+    /* `bottom-chat` keeps the prompt in the chat column; `bottom`, the default,
+       spans the workspace. */
+    const besideChat = config?.promptPlacement === 'bottom-chat';
+    /* `floating` renders the same composer in a draggable card over the
+       workspace — the prompt below simply gains `draggable`, and the card takes
+       itself out of the layout. */
+    const floatingPrompt = config?.promptPlacement === 'floating';
+    /* `top` puts the composer above the editors as a horizontal bar, with the
+       openers as chips under it — for a host whose editor is the point. */
+    const topPrompt = config?.promptPlacement === 'top';
+    /* No prompt at all, for a workspace where something else owns the typing —
+       by configuration, or live through the extras channel when the host
+       changes its mind while the workspace is running. */
+    const promptHidden = config?.hidePrompt === true || chatExtras.hidePrompt === true;
+    /*
+     * The composer, assembled here and rendered by whoever contributed to the
+     * `LoopChatComposer` point — the input-prompt plugin, normally.
+     *
+     * The wiring is this view's knowledge: the controlled draft (so an outside
+     * suggestion has somewhere to land, with `focusTrigger` putting the caret
+     * back), the lexical variant (the `@` menu needs it), the menus' data with
+     * every control drawn even when its list is empty (a menu that opens onto
+     * "0" is a statement; a missing control is not), the `promptAction` and
+     * `inpromptMenu` slots (asked whether anyone contributed first — an empty
+     * `ReactorSlot` is still an element and would pad out a band of nothing).
+     * The surface is the plugin's: no contribution, no composer, which is the
+     * honest outcome for a workspace mounted without one.
+     */
+    const composerProps = {
+        onSend: () => void handleSend(draft),
+        input: draft,
+        setInput: setDraft,
+        focusTrigger: suggestion || focusAsked
+            ? (suggestion?.nonce ?? 0) + focusAsked
+            : undefined,
+        // In a draggable card over the workspace when the host asked for the
+        // floating placement; the dragging is the prompt's own feature.
+        placement: floatingPrompt ||
+            layout?.prompt === 'floating' ||
+            layout?.prompt === 'floating-top'
+            ? 'floating'
+            : 'docked',
+        // Where a floating card starts: the page layout wants it over the top
+        // of the sheet, like a document's title bar; everyone else, bottom.
+        floatingAnchor: layout?.prompt === 'floating-top' ? 'top' : 'bottom',
+        // The agent is working: `isLoading` disables the editor and turns the
+        // send button into a stop — the person keeps a way out.
+        isLoading: busy,
+        onStop: () => workspace.viewControls.stop?.(),
+        // The reason where the typing would go, read before the person types.
+        placeholder: disabledReason ?? placeholder,
+        typingSuggestions: suggestions.map(item => item.text),
+        // The same openers behind the composer's suggestions control.
+        suggestions: chatSuggestions,
+        disableInputPrompt: chatDisabled,
+        connectionConfirmed: !chatDisabled,
+        // Tighter on top: every pixel of the composer is a pixel of the editor
+        // under it pushed down, and the bar reads as a command line at this size.
+        // Both of the page layout's placements are that bar — one floating over
+        // the page, one docked above it — so both take the tighter padding.
+        padding: topPrompt ||
+            layout?.prompt === 'floating-top' ||
+            layout?.prompt === 'docked-top'
+            ? 2
+            : 3,
+        promptVariant: 'lexical',
+        // True unless the host asked otherwise: the workspace is there to be typed
+        // into. A page that embeds the loop as one section among many turns it off,
+        // because focusing an input on mount scrolls the browser to it.
+        autoFocus: config?.autoFocusPrompt ?? true,
+        mentionableAgents: mentionable,
+        headerContent: inpromptMenu.length > 0 ? (_jsx(ReactorSlot, { slot: LoopSlots.inpromptMenu, props: { workspace } })) : undefined,
+        footerExtras: promptActions.length > 0 ? (_jsx(ReactorSlot, { slot: LoopSlots.promptAction, props: { workspace } })) : undefined,
+        showAgentsMenu: true,
+        showInlineAgentsMenu: false,
+        agents: footerAgents,
+        selectedAgentId: selectedMemberId,
+        onSelectAgent: id => team?.select(id),
+        // On unless the host said otherwise: a public page turns the counters
+        // off, because they answer questions a visitor is not asking.
+        showTokenUsage: chatExtras.showTokenUsage ?? true,
+        showContextRing: true,
+        agentUsage: contextUsage ?? undefined,
+        showModelSelector: true,
+        models: offeredModels,
+        selectedModel: activeModel,
+        onModelSelect: model => void selectModel(model),
+        showToolsMenu: true,
+        availableTools: offeredTools,
+        mcpServers: configQuery.data?.mcpServers ?? [],
+        // Live from the host example, when one feeds it: the footer's MCP status
+        // indicator. (Codemode surfaces through the toggle below, not a blob.)
+        mcpStatusData: chatExtras.mcpStatusData ?? undefined,
+        showSkillsMenu: true,
+        skills: skillsQuery.data?.skills ?? [],
+        skillsLoading: skillsQuery.isLoading,
+        enabledSkills,
+        onToggleSkill: skillId => enabledSkills.has(skillId)
+            ? skillActions.disableSkill(skillId)
+            : skillActions.enableSkill(skillId),
+        onToggleAllSkills: (skillIds, enable) => skillIds.forEach(skillId => enable
+            ? skillActions.enableSkill(skillId)
+            : skillActions.disableSkill(skillId)),
+        codemodeEnabled: chatExtras.codemodeEnabled ?? false,
+        onToggleCodemode: chatExtras.onToggleCodemode,
+        isA2AProtocol: false,
+        hasConfigData: !!configQuery.data,
+        hasSkillsData: !!skillsQuery.data,
+    };
+    const ComposerComponent = composerEntries[0]?.value.Component;
+    const HeaderComponent = headerEntries[0]?.value.Component;
+    const prompt = ComposerComponent ? (_jsx(ComposerComponent, { workspace: workspace, composer: composerProps })) : null;
+    /*
+     * The parts, named, so a layout can arrange them.
+     *
+     * Each is wired here — the surfaces' hidden mounts, the transcript's
+     * stream, the composer's draft — and only *placed* below: by the split
+     * this view has always drawn, or by whatever `LoopChatLayout` contributed.
+     */
+    const pickerEl = surfaces.length > 0 && config?.showSurfaceSelector !== false ? (_jsx(SurfacePicker, { surfaces: surfaces.map(entry => entry.value), 
+        /*
+            What is on screen, not what was asked for.
+            
+            It read `surfaceId` — the request — while the column beside it
+            renders `active`, the surface that request actually resolved to.
+            The two are the same once everything has settled and differ
+            exactly while it has not: a surface still arriving, or one closed
+            because its sandbox went away. So the control could say None over
+            a notebook, which is the one thing a control must never do.
+            
+            Reporting the rendered surface makes them agree by construction.
+          */
+        active: active?.surfaceId ?? waiting?.surfaceId ?? NO_SURFACE, onChange: chooseSurface, workspace: workspace })) : null;
+    const editorsEl = (_jsxs(_Fragment, { children: [surfaces.map(entry => {
+                const surface = entry.value;
+                const shown = active?.surfaceId === surface.surfaceId;
+                const mount = shown ||
+                    everMounted.current.has(surface.surfaceId) ||
+                    canOpenView(surface, workspace);
+                if (!mount) {
+                    return null;
+                }
+                everMounted.current.add(surface.surfaceId);
+                return (_jsx(Box, { sx: shown
+                        ? { flex: '1 1 0', minWidth: 0, minHeight: 0 }
+                        : {
+                            position: 'absolute',
+                            inset: 0,
+                            visibility: 'hidden',
+                            pointerEvents: 'none',
+                            zIndex: -1,
+                        }, children: _jsx(ReactorLazy, { load: surface.load, props: { surfaceId: surface.surfaceId, workspace }, fallback: _jsxs(Centered, { children: ["Loading ", surface.title, "\u2026"] }), errorFallback: error => (_jsxs(Centered, { children: [surface.title, " failed to load: ", error.message] })) }) }, surface.surfaceId));
+            }), !active && waiting ? (_jsx(Box, { sx: { flex: '1 1 0', minWidth: 0, minHeight: 0 }, children: _jsxs(Centered, { children: ["Starting ", waiting.title, "\u2026"] }) })) : null] }));
+    const transcriptEl = (_jsx(_Fragment, { children: _jsxs(Box, { sx: {
+                /*
+                    Beside an editor, with the prompt on top, the conversation is
+                    the narrower column: the editor is what a `top` host put the
+                    prompt over, and the transcript reads fine at a third of the
+                    width. Every other placement splits the row evenly.
+                  */
+                flex: topPrompt && active ? '0 0 34%' : '1 1 0',
+                maxWidth: topPrompt && active ? 460 : undefined,
+                minWidth: 0,
+                minHeight: 0,
+                display: 'flex',
+                flexDirection: 'column',
+                borderLeft: active ? '1px solid' : undefined,
+                borderColor: 'border.default',
+                // So the expired-key panel below can cover exactly this column and
+                // nothing else: the notebook beside it still works.
+                position: 'relative',
+            }, children: [chatExtras.errorBanner ? (_jsx(Box, { sx: {
+                        flexShrink: 0,
+                        px: 3,
+                        py: 2,
+                        fontSize: 1,
+                        borderBottom: '1px solid',
+                        bg: chatExtras.errorBanner.variant === 'warning'
+                            ? 'attention.subtle'
+                            : 'danger.subtle',
+                        color: chatExtras.errorBanner.variant === 'warning'
+                            ? 'attention.fg'
+                            : 'danger.fg',
+                        borderColor: chatExtras.errorBanner.variant === 'warning'
+                            ? 'attention.muted'
+                            : 'danger.muted',
+                    }, children: chatExtras.errorBanner.message })) : null, _jsx(Box, { sx: { flex: '1 1 auto', minHeight: 0, display: 'flex' }, children: _jsx(ChatBase
+                    // The header says why, beside the title, for the same reason the
+                    // placeholder does: a dead control with no explanation is worse
+                    // than an absent one.
+                    , { 
+                        // The header says why, beside the title, for the same reason the
+                        // placeholder does: a dead control with no explanation is worse
+                        // than an absent one.
+                        disabled: chatDisabled, disableReason: disabledReason, protocol: protocol, 
+                        /*
+                            Who is answering, in the words its spec uses.
+            
+                            `ChatBase` falls back to "Start a conversation with the AI
+                            agent", which is true of every chat ever built and says
+                            nothing about this one. The empty state is the first thing a
+                            person sees and the only place the agent introduces itself.
+                          */
+                        title: member?.name ?? spec?.name ?? agentId, description: spec?.description, 
+                        /*
+                            Two sizes, because it is drawn in two places.
+            
+                            The header wants a mark beside a line of text; the empty state
+                            wants the thing a person's eye lands on first. One `brandIcon`
+                            served both, so a 48px icon meant for the empty state sat in
+                            the header at three times the height of the words next to it.
+                          */
+                        // Big enough to read as the agent's mark rather than as
+                        // punctuation before its name, and still short enough not to
+                        // set the header's height.
+                        brandIcon: _jsx(BrandIcon, { size: 20 }), emptyState: {
+                            icon: _jsx(BrandIcon, { size: 48 }),
+                            // `ChatEmptyState` reads its heading from here and nowhere
+                            // else — the `title` above reaches the header only — so
+                            // without this the agent introduced itself as "Start a
+                            // conversation".
+                            title: member?.name ?? spec?.name ?? agentId,
+                            /*
+                              Two levels when there is a team: the team first — its name,
+                              what it is for, what it can be asked — and under it the
+                              member being addressed, with its own description and its
+                              own openers. The groups match the ones the openers carry.
+                            */
+                            ...(team && member
+                                ? {
+                                    sections: [
+                                        {
+                                            group: team.team.name,
+                                            icon: _jsx(TeamIcon, { size: 48 }),
+                                            title: team.team.name,
+                                            subtitle: team.team.description,
+                                        },
+                                        {
+                                            group: member.name,
+                                            icon: _jsx(BrandIcon, { size: 32 }),
+                                            title: member.name,
+                                            subtitle: spec?.description,
+                                        },
+                                    ],
+                                }
+                                : null),
+                        }, 
+                        /*
+                            And what it can be asked. From the team rather than the
+                            member: the supervisor answers first, and somebody who has
+                            just opened the workspace does not yet know there are two
+                            agents behind it.
+                          */
+                        // With the prompt on top the openers are already chips under
+                        // it; the empty state repeating them was the same three buttons
+                        // twice on one screen.
+                        suggestions: topPrompt || layout ? [] : chatSuggestions, 
+                        /*
+                            The title bar arrives as a plugin: the chat-header plugin
+                            contributes the component, this view hands it the assembled
+                            props through `renderHeader`. No contribution — or a host
+                            that set `hideHeader` — means no bar.
+                          */
+                        showHeader: !config?.hideHeader && !!HeaderComponent, renderHeader: HeaderComponent
+                            ? headerProps => (_jsx(HeaderComponent, { workspace: workspace, header: headerProps }))
+                            : undefined, 
+                        /*
+                            The (i), which opens the agent's details over the transcript.
+                            
+                            Worth having here in particular: this workspace can be moved
+                            between four runtimes and several agents, and the details pane
+                            is the only place that says which one is actually answering
+                            and where it is running.
+                          */
+                        showInformation: true, 
+                        /*
+                            Full screen: this view, on the whole screen, editor and all.
+                            See `toggleFullScreen` for why it is the browser's API rather
+                            than a big box.
+                          */
+                        headerActions: _jsxs(Box, { sx: { display: 'inline-flex', alignItems: 'center', gap: 2 }, children: [workspace.chatHeaderActions, chatHeaderItems.length > 0 ? (_jsx(ReactorSlot, { slot: LoopSlots.chatHeader, props: { workspace } })) : null, _jsx(IconButton, { icon: fullScreen ? ScreenNormalIcon : ScreenFullIcon, "aria-label": fullScreen ? 'Exit full screen' : 'Enter full screen', variant: "invisible", size: "small", onClick: () => {
+                                        // Whether or not it takes, the hint has been answered:
+                                        // they found the control, which is all it was for.
+                                        setHintFullScreen(false);
+                                        toggleFullScreen();
+                                    }, sx: hintFullScreen
+                                        ? {
+                                            /*
+                                                Colour and opacity, never geometry.
+                  
+                                                The colour is the theme's own `primary` — the
+                                                same brand the page around this workspace is
+                                                already using — because a grey icon in a row of
+                                                grey icons cannot be picked out however hard it
+                                                fades. Colour is what makes it findable; the
+                                                fade is what makes it look like it is asking.
+                  
+                                                It never fades to nothing: a control that
+                                                vanishes and returns reads as a rendering fault,
+                                                not an invitation. And nothing here moves or
+                                                resizes, so the header does not reflow and the
+                                                controls beside it stay where a reader last saw
+                                                them.
+                                              */
+                                            '@keyframes dla-fullscreen-hint': {
+                                                '0%, 100%': { opacity: 1 },
+                                                '50%': { opacity: 0.4 },
+                                            },
+                                            /* `&&` doubles the specificity, because Primer
+                                                 ships Button as a CSS module and
+                                                 `prc-Button-*` outranks a single generated
+                                                 class — without it the keyframes register and
+                                                 the button sits there at `animation: none`.
+                                                 The colour needs the same treatment: an
+                                                 invisible IconButton sets `color` itself.
+                  
+                                                 And no `:focus` or `:hover` clause to stop it:
+                                                 anything that matches the button while it is
+                                                 being pointed at would cancel the pointing —
+                                                 which is exactly how the first version of this
+                                                 silently did nothing. It ends when the reader
+                                                 takes the offer, or arrives without it. */
+                                            '&&': {
+                                                /*
+                                                    The brand on the box, not on the glyph.
+                    
+                                                    Painting the icon `palette.primary` was the
+                                                    obvious reading and it is measurably
+                                                    backwards: in the theme this workspace ships
+                                                    on, `primary` is #FFC107, which sits at 1.55
+                                                    contrast against the header — where the
+                                                    icon's ordinary olive sits at 5.07. Recolouring
+                                                    the stroke made the control roughly three
+                                                    times harder to see, and at the bottom of the
+                                                    fade it was 1.2 and simply gone.
+                    
+                                                    A ring and a wash carry the same colour on a
+                                                    far larger area, and they do it without asking
+                                                    the glyph to be legible in a brand colour that
+                                                    changes with every theme — this palette has
+                                                    six, from a dark teal to this amber, and a
+                                                    rule that works for one fails for another. The
+                                                    glyph keeps the colour it can be read in.
+                                                  */
+                                                bg: `color-mix(in srgb, ${palette.primary} 16%, transparent)`,
+                                                boxShadow: `inset 0 0 0 1.5px ${palette.primary}`,
+                                                animation: `dla-fullscreen-hint ${FULLSCREEN_HINT_PERIOD_MS}ms ease-in-out infinite`,
+                                            },
+                                            // A reader who asked the machine to hold still
+                                            // keeps the colour and loses the knocking: the
+                                            // control is still the findable one, it just
+                                            // holds still.
+                                            '@media (prefers-reduced-motion: reduce)': {
+                                                '&&': { animation: 'none' },
+                                            },
+                                        }
+                                        : undefined })] }), 
+                        // This view owns the prompt, below, so the chat does not draw a
+                        // second one: two input boxes on one screen is one too many.
+                        showInput: false, onSendReady: handleSendReady, 
+                        /*
+                            The notebook tools' results, drawn into the transcript — the
+                            chat's own machinery, switched on only while no editor is on
+                            screen: the hidden notebook is where the agent works, and the
+                            transcript is then the one place the reader can see what a
+                            tool did. With an editor open the change is visible where it
+                            happened, and the default tool row is enough.
+                          */
+                        notebookToolSurfacesId: active ? undefined : workspace.surfaceId, 
+                        // A host example's own tool-result renderer, when it feeds one
+                        // through the extras channel — the A2UI examples draw their
+                        // surface here. Wins over the notebook surfaces in ChatBase.
+                        renderToolResult: chatExtras.renderToolResult, showTurnFooter: chatExtras.showTokenUsage ?? true, onContextSnapshot: handleContextSnapshot, onLoadingChange: handleLoadingChange, onItemsChange: handleMessagesChange, onDisplayItemsChange: handleDisplayItemsChange, enableStreaming: true }) }), besideChat && !promptHidden ? prompt : null] }) }));
+    const chipsEl = chatSuggestions.length > 0 ? (_jsx(Box, { "aria-label": "Suggested prompts", sx: { display: 'flex', flexWrap: 'wrap', gap: 2, px: 2, pb: '6px' }, children: chatSuggestions.map(item => (_jsx(Box, { as: "button", type: "button", 
+            // The whole request, where the chip shows only its summary.
+            title: item.message, disabled: busy || chatDisabled, onClick: () => void handleSend(item.message), sx: {
+                appearance: 'none',
+                font: 'inherit',
+                fontSize: 0,
+                fontWeight: 'semibold',
+                // Room for descenders: the truncated span clips what falls
+                // below its line, and at a line-height of one that is the
+                // tail of every "y".
+                lineHeight: 1.25,
+                px: '10px',
+                py: '5px',
+                width: SUGGESTION_CHIP_WIDTH,
+                flex: `0 0 ${SUGGESTION_CHIP_WIDTH}px`,
+                borderRadius: '999px',
+                border: '1px solid',
+                borderColor: 'border.default',
+                bg: 'canvas.subtle',
+                color: 'fg.default',
+                cursor: 'pointer',
+                whiteSpace: 'nowrap',
+                '&:hover:not(:disabled)': {
+                    borderColor: 'accent.fg',
+                    color: 'accent.fg',
+                },
+                '&:disabled': { opacity: 0.5, cursor: 'default' },
+            }, children: _jsx(Truncate, { title: item.title, maxWidth: "100%", sx: { minWidth: 0 }, children: item.title }) }, item.title))) })) : null;
+    const transientEl = transient ? (_jsx(Box, { sx: {
+            flex: '0 0 auto',
+            px: 3,
+            py: 2,
+            borderTop: '1px solid',
+            borderColor: 'border.default',
+            bg: 'canvas.subtle',
+            fontSize: 1,
+            maxHeight: '40%',
+            overflowY: 'auto',
+        }, children: transient })) : null;
+    return (_jsxs(Box, { ref: viewRef, sx: {
+            height: '100%',
+            minHeight: 0,
+            display: 'flex',
+            flexDirection: 'column',
+            // The positioned root the expiry panel below covers.
+            position: 'relative',
+            /*
+              A background of its own once it is full screen, in both modes.
+    
+              The element is promoted out of whatever was painting behind it — the
+              browser's top layer, or the page — and a workspace that inherited its
+              canvas from an ancestor would arrive transparent over black.
+            */
+            ...(fullScreen ? { bg: 'canvas.default' } : null),
+            // The fallback, for a host where the API was refused, or where
+            // `fullScreenTopOffset` asked for the overlay outright. Covers as
+            // much as the nearest transformed ancestor allows; see the note
+            // above. `top` leaves room for a host's own fixed header rather than
+            // `inset: 0`'s full viewport — 0 (no offset given) is the same rect
+            // either way.
+            ...(fullScreen && !usingFullscreenApi.current
+                ? {
+                    position: 'fixed',
+                    top: fullScreenTopOffsetPx,
+                    right: 0,
+                    bottom: 0,
+                    left: 0,
+                    // Over the `height: '100%'` above, which would otherwise win
+                    // over `top`/`bottom` and push the bottom off the window.
+                    height: 'auto',
+                    zIndex: 1000,
+                }
+                : null),
+        }, children: [layout ? (_jsx(layout.Component, { workspace: workspace, editors: editorsEl, hasEditor: !!active, transcript: transcriptEl, prompt: promptHidden ? null : prompt, chips: promptHidden ? null : chipsEl, picker: pickerEl, transient: transientEl })) : (_jsxs(_Fragment, { children: [topPrompt && !promptHidden ? (_jsxs(Box, { sx: {
+                            flex: '0 0 auto',
+                            borderBottom: '1px solid',
+                            borderColor: 'border.default',
+                        }, children: [prompt, chipsEl] })) : null, pickerEl, _jsxs(Box, { sx: {
+                            flex: '1 1 auto',
+                            minHeight: 0,
+                            display: 'flex',
+                            // The hidden editors position themselves against this row.
+                            position: 'relative',
+                        }, children: [editorsEl, transcriptEl] }), transientEl, besideChat || promptHidden || topPrompt ? null : prompt] })), keyExpired ? (_jsx(Box, { sx: {
+                    position: 'absolute',
+                    inset: 0,
+                    zIndex: 5,
+                    bg: 'canvas.default',
+                    display: 'flex',
+                    flexDirection: 'column',
+                    minHeight: 0,
+                }, children: _jsx(AnonymousKeyExpired, { agentName: member?.name ?? spec?.name, 
+                    // On the browser target the kernel is in this page and owes the
+                    // inference service nothing, so the notebook is genuinely
+                    // unaffected and the panel may say so.
+                    sandboxStillRuns: inPage, temporary: expiredKeyIsTemporary, 
+                    // The trial key is spent once somebody signs in, and a session
+                    // left `expired` would keep this panel over a working chat.
+                    onSignedIn: () => useAnonymousSessionStore.getState().clear() }) })) : null] }));
+}
+/**
+ * Which editor sits beside the conversation.
+ *
+ * Rendered only when a plugin contributed one — a chat on its own has nothing
+ * to pick between, and a control with a single option is noise.
+ */
+function SurfacePicker({ surfaces, active, onChange, workspace, }) {
+    const ordered = [...surfaces].sort((left, right) => (left.order ?? 100) - (right.order ?? 100));
+    return (_jsxs(Box, { sx: {
+            flex: '0 0 auto',
+            display: 'flex',
+            alignItems: 'center',
+            gap: 2,
+            px: 3,
+            py: 2,
+            borderBottom: '1px solid',
+            borderColor: 'border.default',
+        }, children: [_jsx(Text, { sx: { fontSize: 0, color: 'fg.muted' }, children: "Beside the chat" }), _jsxs(SegmentedControl, { "aria-label": "Editor beside the chat", size: "small", children: [_jsx(SegmentedControl.Button, { selected: active === NO_SURFACE, onClick: () => onChange(NO_SURFACE), children: "Chat" }), ordered.map(surface => {
+                        const openable = canOpenView(surface, workspace);
+                        const reason = openable
+                            ? undefined
+                            : (surface.unavailableReason?.(workspace) ??
+                                'Not available right now');
+                        return (_jsx(SegmentedControl.Button, { selected: active === surface.surfaceId, "aria-disabled": !openable, 
+                            // The native attribute rather than Primer's `Tooltip`: that
+                            // component requires its child to *be* the interactive element,
+                            // and `SegmentedControl.Button` renders a list item around its
+                            // button — wrapping it throws before the workspace can paint.
+                            title: reason ?? surface.title, 
+                            // Primer types this as its own icon shape; a contribution may
+                            // bring any component, which is the point of the extension point.
+                            leadingIcon: surface.icon, onClick: () => {
+                                if (openable) {
+                                    onChange(surface.surfaceId);
+                                }
+                            }, children: surface.title }, surface.surfaceId));
+                    })] })] }));
+}
+function Centered({ children }) {
+    return (_jsx(Box, { sx: {
+            height: '100%',
+            display: 'flex',
+            alignItems: 'center',
+            justifyContent: 'center',
+            color: 'fg.muted',
+            fontSize: 1,
+            px: 3,
+            textAlign: 'center',
+        }, children: children }));
+}