ios: share prompt field across chat and agent views

Integrate the shared Agent room, durable outbox, cursor replay and read state on iPhone and iPad. Reuse message rendering with native messaging layout and expandable tool details.

Validated with 58 Xcode package tests, six Simulator UI scenarios, read-only production connectivity and independent review.

dist/default.conf

@@ -1,3 +1,8 @@
+map $http_upgrade $connection_upgrade {
+  default upgrade;
+  "" close;
+}
+
 server {
   listen 80;
   server_name _;
@@ -19,7 +24,8 @@ server {
     proxy_set_header Host $host;
     proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
     proxy_set_header X-Forwarded-Proto $scheme;
-    proxy_set_header Connection "";
+    proxy_set_header Upgrade $http_upgrade;
+    proxy_set_header Connection $connection_upgrade;
   }
 
   location = /sw.js {

docs/api/rest.md

@@ -31,7 +31,7 @@ Chat upload limits:
 ```json
 {
   "providers": {
-    "openai": { "models": ["gpt-4.1-mini"], "loadedAt": "2026-02-14T00:00:00.000Z", "error": null },
+    "openai": { "models": ["chat-latest", "gpt-4.1-mini"], "loadedAt": "2026-02-14T00:00:00.000Z", "error": null },
     "anthropic": { "models": ["claude-3-5-sonnet-latest"], "loadedAt": null, "error": null },
     "xai": { "models": ["grok-3-mini"], "loadedAt": null, "error": null },
     "gemini": { "models": ["gemini-3.5-flash"], "loadedAt": null, "error": null },
@@ -39,7 +39,7 @@ Chat upload limits:
   }
 }
 ```
-- OpenAI model lists are filtered to models that are expected to work with the backend's Responses API implementation.
+- OpenAI model lists are filtered to models that are expected to work with the backend's Responses API implementation, including the exact `chat-latest` alias when returned by OpenAI. Audio, image-generation, embedding, moderation, and other specialized model ids remain excluded. See [OpenAI's Chat Latest documentation](https://developers.openai.com/api/docs/models/chat-latest) for supported capabilities.
 - Gemini model lists are loaded from Google's native Models API and filtered to Gemini `generateContent` model ids.
 - `hermes-agent` is included only when `HERMES_AGENT_API_KEY` is configured. Set it to Hermes `API_SERVER_KEY`, or any non-empty value if that local server does not require auth. `HERMES_AGENT_API_BASE_URL` defaults to `http://127.0.0.1:8642/v1`; set `HERMES_AGENT_MODEL` only when you need an additional fallback/override model id.
 - The backend loads provider model lists at startup and refreshes them about once every 24 hours. If a later provider refresh fails, the response keeps the last loaded model list for that provider and sets `error` to the latest failure message.
@@ -81,6 +81,8 @@ Behavior notes:
 
 ### `GET /v1/workspace-items`
 - Response: `{ "items": WorkspaceItem[] }`
+- Optional query: `q` (trimmed string, up to 200 characters). Omit or leave blank for the recent-items list.
+- With nonblank `q`, response: `{ "items": WorkspaceItem[], "hasMore": boolean }`; each item also has `searchSnippet: string | null` (plain text, no HTML).
 - `WorkspaceItem` is a discriminated union sorted by `updatedAt` descending:
 ```json
 {
@@ -123,6 +125,15 @@ Behavior notes:
 - Root chats have `parentChatId: null`. Every fork points directly to its single root chat, including a fork created from another fork, so clients can group rows without traversing a fork chain.
 - `starred`/`starredAt` are backed by membership in a reserved `Project` with id `starred`; future project folders can reuse the same project item model.
 
+History search (`q`) notes:
+- Searches all saved history, including items outside the recent-items window. Chats match their title or the text of a user/assistant message; saved web searches match their title, query, or generated answer. System prompts, tool logs, attachments, and raw web-result payloads are excluded.
+- Case- and diacritic-insensitive Unicode word-prefix matching: `plan` matches `planning`. All query words must match within the title, a single message, or one saved web search (across its indexed fields). Punctuation separates words; quotes, `*`, and operators have no special query syntax. Punctuation-only queries return no matches. This is word-prefix search, not arbitrary substring matching.
+- Returns at most 100 distinct matching items, newest `updatedAt` first, with ties ordered by `type` then `id`. `hasMore` indicates additional matches; refine `q` to narrow them. Unmatched fork roots are not added in search mode. Selecting a fork result opens that fork directly. Search-mode fork results carry their root's `starred`/`starredAt`; clients should star/unstar the `parentChatId`, as in the grouped sidebar.
+- `searchSnippet` is a bounded excerpt from a matching message or saved web search; title-only chat matches have `null`. Clients should render it as plain text.
+- The web and iOS sidebar search fields use this endpoint across saved history. Both display matching excerpts; web title-only chat results display “Title match”.
+- Invalid/non-string queries or queries exceeding 200 characters return `400`. Authentication is unchanged.
+- SQLite FTS5 external-content indexes and prefix indexes serve these queries without scanning or transferring every transcript. The migration backfills existing content; transactional triggers maintain the indexes on inserts, updates, deletes, and cascades. Recent chat listing also has an `updatedAt` index.
+
 ## Chats
 
 ### `GET /v1/chats`
@@ -551,3 +562,73 @@ Behavior notes:
 ```
 
 For streaming contracts, see `docs/api/streaming-chat.md`.
+
+## Sybil Agent Room
+
+The Sybil Agent tab connects to one persistent root conversation. The agent owns its messages,
+background work, and history; these endpoints do not create ordinary Sybil chats or
+copy the transcript into the Sybil message database. All endpoints use normal Sybil
+REST authentication. The server connects privately to `SYBIL_AGENT_BASE_URL` (default
+`http://sybil-agent:8788`) using `SYBIL_AGENT_API_KEY`. The agent key never reaches browsers.
+
+### `GET /v1/agent`
+
+Returns the Sybil Agent room descriptor: `id`, `title`, `type: "persistent-agent"`, `model`,
+`protocol: "pi-room.v1"`, `activity`, and `websocket: "/v1/agent/ws"`. The socket path
+is relative to the API base, so web deployments use `/api/v1/agent/ws`. Missing agent
+configuration returns 503; an unreachable agent server or malformed upstream JSON
+returns 502.
+
+### `GET /v1/agent/messages`
+
+Query parameters: optional `after` or `before` (mutually exclusive nonnegative
+numeric cursor strings), and `limit` (1–200). Without a cursor, returns recent
+history. `after` catches up forward; `before` loads earlier history.
+
+```json
+{
+  "conversationId": "1",
+  "messages": [{ "id": "outbound:43", "createdAt": "2026-10-14T18:00:00.000Z", "role": "assistant", "content": "The task finished.", "name": "Sybil Agent", "metadata": { "kind": "message", "entryId": 43, "status": "complete" } }],
+  "cursor": "43",
+  "beforeCursor": "43",
+  "hasMore": false
+}
+```
+
+Messages use `user`, `assistant`, and `tool` roles. Tools use
+`metadata.kind: "tool_call"` and stable IDs; merge by message `id`, since later
+pages can update initiated tools to completed or failed. Internal worker reports,
+event prompts, and model reasoning are excluded. Cursors refer to underlying
+durable entries and can advance without visible messages. Use returned cursors;
+do not derive them from message IDs. `limit` bounds source entries, not displayed
+messages.
+
+### `POST /v1/agent/messages`
+
+Body: `{ "content": "Check the oven", "requestId": "client-generated-uuid" }`.
+Content must contain non-whitespace text and is limited to 40,000 characters;
+`requestId` is required (1–256 characters). Preserve the ID when retrying uncertain
+network results. The proxy fixes `channel: "sybil"`, `whenBusy: "followUp"`, and
+`wait: false`.
+
+Returns the agent's `202 Accepted` receipt, including `id` (submission ID), `status`, and
+`messageId` such as `inbound:sybil:client-generated-uuid`. This is admission, not
+completion: watch the room for answers, including unsolicited later replies.
+Multiple requests can be accepted while the agent works. Attachments are not yet supported.
+The proxy verifies successful receipts contain a positive submission ID, a known
+status (`queued`, `placed`, `done`, or `unanswered`), and the expected `messageId`.
+Invalid receipts return 502; delivery remains uncertain, so retry using the original
+`requestId` to avoid creating a duplicate message.
+
+### `GET /v1/agent/read-state`
+### `PUT /v1/agent/read-state`
+
+GET returns `{ "cursor": "43", "updatedAt": "2026-10-14T18:00:00.000Z" }`.
+Initially: `{ "cursor": "0", "updatedAt": null }`. PUT takes
+`{ "cursor": "43" }` and returns the same shape. The cursor is shared across this
+personal installation and persisted in SQLite. Atomic, monotonic updates prevent
+a stale browser from lowering a cursor saved by another browser. Advance it only
+after showing the corresponding history; poll it to reconcile unread indicators
+across clients. Agent REST responses use `Cache-Control: no-store`.
+
+See the persistent Agent WebSocket section in `streaming-chat.md` for live updates.

docs/api/streaming-chat.md

@@ -1,21 +1,22 @@
 # Streaming Chat API Contract
 
-This document defines the server-sent events (SSE) contract for chat completions.
+This document defines the server-sent events (SSE) contract for chat completions and Quick Questions, and the WebSocket contract for the persistent Agent room.
 
 Endpoint:
 - `POST /v1/chat-completions/stream`
 - `POST /v1/chats/:chatId/stream/attach`
+- `POST /v1/quick-questions/stream`
 
 Transport:
 - HTTP response uses `Content-Type: text/event-stream; charset=utf-8`
 - Events are emitted in SSE format (`event: ...`, `data: ...`)
 - Request body is JSON
-- Request body supports the same inline attachment schema and limits documented in `docs/api/rest.md`.
+- Chat completion request bodies support the same inline attachment schema and limits documented in `docs/api/rest.md`. Quick Questions accept text only.
 
 Authentication:
 - Same as REST endpoints (`Authorization: Bearer <token>` when token mode is enabled)
 
-## Request Body
+## Chat Completion Request Body
 
 ```json
 {
@@ -90,6 +91,32 @@ Persisted chat streams with a `chatId` are backend-owned active runs:
 
 This endpoint is intended for clients that restored an active `chatId` from `GET /v1/active-runs`, especially after browser refresh. Replayed `delta` events may include text that was originally emitted before the client attached.
 
+## Quick Question Endpoint
+
+`POST /v1/quick-questions/stream`
+
+Request body:
+```json
+{
+  "provider": "openai|anthropic|xai|gemini|hermes-agent",
+  "model": "string",
+  "question": "What is the capital of France?",
+  "enabledTools": ["web_search", "fetch_url"],
+  "userLocation": "optional city, region, country",
+  "temperature": 0.2,
+  "maxTokens": 256
+}
+```
+
+Behavior notes:
+- `provider` and `model` are required. Web and native clients default Quick Question to `provider: "openai"` and `model: "chat-latest"` when no Quick Question selection has been saved. Saved Quick Question choices take precedence over this default and are separate from regular chat preferences.
+- `question` is required, trimmed by the server, and must not be empty.
+- The server prepends a Quick Question system prompt that asks for a succinct, direct, self-contained answer without follow-up questions. Clients do not send or maintain this prompt.
+- Quick Questions are always non-persistent. The endpoint does not create a chat or store messages, tool-call logs, assistant output, or `LlmCall` metadata.
+- The response uses the same `meta`, `tool_call`, `delta`, `done`, and `error` SSE events as chat completion streams. The `meta` event has `chatId: null` and `callId: null`.
+- `enabledTools`, `temperature`, and `maxTokens` are optional and behave as they do for chat completion streams. When `enabledTools` is omitted, all available Sybil-managed tools are enabled by default.
+- User location is inferred from the same request headers as chat completion streams when `userLocation` is omitted.
+
 ## Event Stream Contract
 
 Event order:
@@ -248,3 +275,63 @@ event: done
 data: {"type":"done","text":"Hello world"}
 
 ```
+
+## Sybil Agent WebSocket
+
+`GET /v1/agent/ws` upgrades to a continuous WebSocket room subscription. Behind the
+web proxy use `/api/v1/agent/ws`, with `ws:` or `wss:` matching the page. It remains
+connected after each assistant turn and receives later background results. Sending
+messages uses `POST /v1/agent/messages`.
+
+The browser must send this first frame within five seconds:
+
+```json
+{ "type": "auth", "token": "Sybil token", "after": "43" }
+```
+
+In open mode, use `token: null`. Omit `after` for recent history; on reconnect supply
+the last successfully merged cursor. Invalid credentials close with 4401, malformed
+initial frames with 4400. Malformed upgrade URLs and query parameters are rejected
+with HTTP 400; unknown upgrade paths return HTTP 404. Never put tokens in URLs.
+Sybil adds its private Sybil Agent credential only to the upstream Authorization header;
+browser credentials are not forwarded to the agent.
+
+Native `pi-room.v1` frames are forwarded after authentication. This protocol
+identifier is retained for compatibility with existing clients:
+
+```json
+{ "type": "ready", "protocol": "pi-room.v1", "conversationId": "1", "model": "..." }
+```
+
+```json
+{
+  "type": "messages",
+  "conversationId": "1",
+  "messages": [{ "id": "outbound:44", "role": "assistant", "content": "The result is ready.", "createdAt": "2026-10-14T18:00:00.000Z", "metadata": { "kind": "message", "entryId": 44, "status": "complete" } }],
+  "cursor": "44",
+  "beforeCursor": "44",
+  "hasMore": false
+}
+```
+
+```json
+{ "type": "activity", "conversationId": "1", "activity": { "state": "thinking", "isThinking": true, "queued": 0, "backgroundWorkers": 0, "tools": [], "queuedMessages": [] } }
+```
+
+Activity is forwarded from Sybil Agent and also describes tools, queued messages, durable
+waiting, errors, and workers. Use `isThinking` for the thinking indicator and
+`backgroundWorkers` for ongoing delegated work; no model reasoning is exposed. Tools arrive as `role: "tool"` messages
+with `metadata.kind: "tool_call"`, matching Sybil's tool renderer. REST pages and
+socket messages are upserts keyed by stable `id`. Save each returned cursor after
+applying its entire frame, even when it contains no visible messages.
+
+Clients may send `{ "type": "ping" }` (receiving `{ "type": "pong" }`) or
+`{ "type": "subscribe", "after": "44" }` to change the replay cursor. Transport
+ping/pong heartbeats detect dead connections. Other writes belong on REST.
+Errors use `{ "type": "error", "error": "..." }`. Upstream loss closes with 1011
+or 1012; slow consumers may receive 1013. Reconnect with backoff and the last merged
+cursor. Missing agent configuration closes with 4503.
+
+The proxy closes upstream sockets when browsers leave or Sybil shuts down.
+History remains durable in Sybil Agent. Shared read position is separate from delivery:
+update `GET`/`PUT /v1/agent/read-state` only when messages have actually been viewed.

ios/Apps/Sybil/Info.plist

@@ -2,6 +2,9 @@
 <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
 <plist version="1.0">
 <dict>
+	<!-- macOS-only: keep the global Quick Question shortcut available with no windows. -->
+	<key>NSSupportsAutomaticTermination</key>
+	<false/>
 	<key>UIApplicationShortcutItems</key>
 	<array>
 		<dict>

ios/Apps/Sybil/Sources/SybilApp.swift

@@ -8,7 +8,7 @@ struct SybilApp: App
     @UIApplicationDelegateAdaptor(SybilAppDelegate.self) private var appDelegate
 
     var body: some Scene {
-        WindowGroup {
+        WindowGroup(id: SybilCommands.workspaceWindowID) {
             SplitView()
         }
         .commands {
@@ -24,9 +24,18 @@ final class SybilAppDelegate: NSObject, UIApplicationDelegate {
         didFinishLaunchingWithOptions launchOptions: [UIApplication.LaunchOptionsKey: Any]? = nil
     ) -> Bool {
         SybilHomeScreenQuickActionHandler.configureQuickActions()
+        #if targetEnvironment(macCatalyst)
+        SybilMacQuickQuestionController.shared.start()
+        #endif
         return true
     }
 
+    #if targetEnvironment(macCatalyst)
+    func applicationWillTerminate(_ application: UIApplication) {
+        SybilMacQuickQuestionController.shared.stop()
+    }
+    #endif
+
     func application(
         _ application: UIApplication,
         configurationForConnecting connectingSceneSession: UISceneSession,
@@ -56,11 +65,22 @@ final class SybilSceneDelegate: NSObject, UIWindowSceneDelegate {
         willConnectTo session: UISceneSession,
         options connectionOptions: UIScene.ConnectionOptions
     ) {
+        #if targetEnvironment(macCatalyst)
+        if scene is UIWindowScene, session.role == .windowApplication {
+            SybilMacWindowCommands.shared.windowConnected(sessionID: session.persistentIdentifier)
+        }
+        #endif
         if let shortcutItem = connectionOptions.shortcutItem {
             _ = SybilHomeScreenQuickActionHandler.handle(shortcutItem)
         }
     }
 
+    #if targetEnvironment(macCatalyst)
+    func sceneDidDisconnect(_ scene: UIScene) {
+        SybilMacWindowCommands.shared.windowDisconnected(sessionID: scene.session.persistentIdentifier)
+    }
+    #endif
+
     func windowScene(
         _ windowScene: UIWindowScene,
         performActionFor shortcutItem: UIApplicationShortcutItem,

ios/Apps/Sybil/project.yml

@@ -11,6 +11,13 @@ targets:
     dependencies:
       - package: Sybil
         product: Sybil
+      - target: SybilMacQuickQuestion
+        destinationFilters: [macCatalyst]
+        embed: true
+        link: false
+        codeSign: true
+        copy:
+          destination: plugins
     settings:
       base:
         PRODUCT_BUNDLE_IDENTIFIER: net.buzzert.sybil2
@@ -29,6 +36,7 @@ targets:
         INFOPLIST_KEY_CFBundleDisplayName: Sybil
         INFOPLIST_KEY_ITSAppUsesNonExemptEncryption: NO
         INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents: YES
+        "INFOPLIST_KEY_UIApplicationSceneManifest_Generation[sdk=macosx*]": YES
         INFOPLIST_KEY_UILaunchScreen_Generation: YES
         INFOPLIST_KEY_UISupportedInterfaceOrientations_iPhone: UIInterfaceOrientationPortrait
         INFOPLIST_KEY_UISupportedInterfaceOrientations_iPad: UIInterfaceOrientationPortrait UIInterfaceOrientationPortraitUpsideDown UIInterfaceOrientationLandscapeLeft UIInterfaceOrientationLandscapeRight

ios/Apps/SybilMacQuickQuestion/Sources/MacQuickQuestionPlugin.swift

@@ -0,0 +1,229 @@
+import AppKit
+import Carbon
+
+// AppKit is unavailable to Catalyst targets. This native bundle uses only
+// public AppKit/Carbon APIs and communicates with the client through Foundation.
+@MainActor
+@objc(SybilMacQuickQuestionPlugin)
+final class MacQuickQuestionPlugin: NSObject, SybilQuickQuestionPanelBridging, NSWindowDelegate {
+    private weak var delegate: (any SybilQuickQuestionPanelDelegate)?
+    private var panel: QuickQuestionPanel?
+    private var panelView: QuickQuestionPanelView?
+    private var hotKey: EventHotKeyRef?
+    private var eventHandler: EventHandlerRef?
+    private var outsideClickMonitor: Any?
+    private var hotKeyState = SybilQuickQuestionHotKeyState()
+    private var isDismissing = false
+    private var shortcutAvailable = false
+    private var windowlessActivity: NSObjectProtocol?
+    private var latestState = SybilQuickQuestionPanelState()
+    private var previousApplication: NSRunningApplication?
+
+    required override init() {
+        super.init()
+    }
+
+    var isVisible: Bool { panel?.isVisible == true }
+
+    func start(delegate: any SybilQuickQuestionPanelDelegate) -> Int32 {
+        self.delegate = delegate
+        if windowlessActivity == nil {
+            // Keep the hotkey service alive with no windows. This option does
+            // not prevent sleep/App Nap or intercept an explicit Quit command.
+            windowlessActivity = ProcessInfo.processInfo.beginActivity(
+                options: .automaticTerminationDisabled,
+                reason: "Quick Question global shortcut"
+            )
+        }
+        guard eventHandler == nil else { return noErr }
+        var eventTypes = [
+            EventTypeSpec(eventClass: OSType(kEventClassKeyboard), eventKind: UInt32(kEventHotKeyPressed)),
+            EventTypeSpec(eventClass: OSType(kEventClassKeyboard), eventKind: UInt32(kEventHotKeyReleased))
+        ]
+        let context = Unmanaged.passUnretained(self).toOpaque()
+        let handlerStatus = InstallEventHandler(
+            GetApplicationEventTarget(),
+            { _, event, context in
+                guard let event, let context else { return OSStatus(eventNotHandledErr) }
+                var identifier = EventHotKeyID()
+                let status = GetEventParameter(
+                    event, EventParamName(kEventParamDirectObject), EventParamType(typeEventHotKeyID),
+                    nil, MemoryLayout<EventHotKeyID>.size, nil, &identifier
+                )
+                guard status == noErr, identifier.signature == 0x53595151, identifier.id == 1 else {
+                    return OSStatus(eventNotHandledErr)
+                }
+                let isKeyDown = GetEventKind(event) == UInt32(kEventHotKeyPressed)
+                let plugin = Unmanaged<MacQuickQuestionPlugin>.fromOpaque(context).takeUnretainedValue()
+                // Application event handlers run on the main event loop.
+                MainActor.assumeIsolated {
+                    if plugin.hotKeyState.shouldToggle(isKeyDown: isKeyDown) {
+                        plugin.delegate?.quickQuestionToggleRequested()
+                    }
+                }
+                return noErr
+            },
+            eventTypes.count, &eventTypes, context, &eventHandler
+        )
+        guard handlerStatus == noErr else { return handlerStatus }
+        let status = RegisterEventHotKey(
+            UInt32(kVK_Space), UInt32(optionKey),
+            EventHotKeyID(signature: 0x53595151, id: 1),
+            GetApplicationEventTarget(), 0, &hotKey
+        )
+        shortcutAvailable = status == noErr
+        return status
+    }
+
+    func updateState(_ data: Data) {
+        guard let state = try? JSONDecoder().decode(SybilQuickQuestionPanelState.self, from: data) else { return }
+        // Updating a background request must not create a non-hosting window.
+        // The controller creates/orders the panel only after UIKit is foreground.
+        latestState = state
+        panelView?.update(state, shortcutAvailable: shortcutAvailable)
+        if isVisible { resizePanel() }
+    }
+
+    func show() {
+        createPanelIfNeeded()
+        guard let panel, let panelView else { return }
+        panelView.update(latestState, shortcutAvailable: shortcutAvailable)
+        let screen = NSScreen.screens.first { $0.frame.contains(NSEvent.mouseLocation) } ?? NSScreen.main
+        let visibleFrame = screen?.visibleFrame ?? NSRect(x: 0, y: 0, width: 1280, height: 800)
+        let width = min(680, visibleFrame.width - 48)
+        panel.setContentSize(NSSize(width: width, height: 240))
+        panelView.layoutSubtreeIfNeeded()
+        let height = min(panelView.preferredHeight, visibleFrame.height - 64)
+        let top = visibleFrame.maxY - min(visibleFrame.height * 0.18, 160)
+        let y = max(visibleFrame.minY + 24, top - height)
+        panel.setFrame(NSRect(x: visibleFrame.midX - width / 2, y: y, width: width, height: height), display: true)
+        // The controller has verified UIKit foreground readiness. A
+        // nonactivating panel can then take focus without raising a workspace.
+        panel.makeKeyAndOrderFront(nil)
+        panel.makeFirstResponder(panelView.promptView)
+        panelView.preparePromptForPresentation()
+        installOutsideClickMonitor()
+    }
+
+    func hide() {
+        dismissPanel(restoringPreviousApplication: true)
+    }
+
+    func activateForQuickQuestion() {
+        if let frontmost = NSWorkspace.shared.frontmostApplication,
+           frontmost != NSRunningApplication.current {
+            previousApplication = frontmost
+        }
+        // Do not order a window here. Catalyst must finish its own foreground
+        // transition before the controller calls show().
+        NSApp.activate()
+    }
+
+    func activateMainWindow() {
+        // An explicit Open in Chat/Open Sybil action keeps focus in Sybil.
+        previousApplication = nil
+        NSApp.activate()
+        NSApp.windows.first { $0 !== panel && $0.canBecomeMain }?.makeKeyAndOrderFront(nil)
+    }
+
+    private func dismissPanel(restoringPreviousApplication: Bool) {
+        guard !isDismissing else { return }
+        let wasVisible = isVisible
+        isDismissing = true
+        if wasVisible { panel?.orderOut(nil) }
+        removeOutsideClickMonitor()
+        isDismissing = false
+        let applicationToRestore = previousApplication
+        previousApplication = nil
+        if restoringPreviousApplication, NSApp.isActive,
+           let applicationToRestore, !applicationToRestore.isTerminated {
+            NSApp.yieldActivation(to: applicationToRestore)
+            applicationToRestore.activate(options: [])
+        }
+        if wasVisible { delegate?.quickQuestionPanelDismissed() }
+    }
+
+    func stop() {
+        previousApplication = nil
+        hide()
+        removeOutsideClickMonitor()
+        if let hotKey { UnregisterEventHotKey(hotKey) }
+        if let eventHandler { RemoveEventHandler(eventHandler) }
+        hotKey = nil
+        eventHandler = nil
+        hotKeyState = SybilQuickQuestionHotKeyState()
+        if let windowlessActivity {
+            ProcessInfo.processInfo.endActivity(windowlessActivity)
+        }
+        windowlessActivity = nil
+        delegate = nil
+    }
+
+    func windowDidResignKey(_ notification: Notification) {
+        // Focus is already moving to a window/app the user chose.
+        dismissPanel(restoringPreviousApplication: false)
+    }
+
+    private func createPanelIfNeeded() {
+        guard panel == nil else { return }
+        let panel = QuickQuestionPanel(
+            contentRect: NSRect(x: 0, y: 0, width: 680, height: 240),
+            styleMask: [.borderless, .nonactivatingPanel], backing: .buffered, defer: false
+        )
+        panel.title = "Quick Question"
+        panel.level = .floating
+        panel.collectionBehavior = [.canJoinAllSpaces, .fullScreenAuxiliary, .transient, .ignoresCycle]
+        panel.isFloatingPanel = true
+        panel.hidesOnDeactivate = false
+        panel.becomesKeyOnlyIfNeeded = false
+        panel.worksWhenModal = true
+        panel.isReleasedWhenClosed = false
+        panel.isOpaque = false
+        panel.backgroundColor = .clear
+        panel.hasShadow = true
+        panel.isMovableByWindowBackground = true
+        panel.appearance = NSAppearance(named: .darkAqua)
+        panel.delegate = self
+        panel.onDismiss = { [weak self] in self?.hide() }
+        let view = QuickQuestionPanelView()
+        view.delegate = delegate
+        view.onDismiss = { [weak self] in self?.hide() }
+        panel.contentView = view
+        panelView = view
+        self.panel = panel
+    }
+
+    private func resizePanel() {
+        guard let panel, let panelView else { return }
+        panelView.layoutSubtreeIfNeeded()
+        let visibleFrame = panel.screen?.visibleFrame ?? panel.frame
+        let height = min(panelView.preferredHeight, visibleFrame.height - 64)
+        guard abs(height - panel.frame.height) > 1 else { return }
+        var frame = panel.frame
+        frame.origin.y = max(visibleFrame.minY + 24, frame.maxY - height)
+        frame.size.height = height
+        panel.setFrame(frame, display: true)
+    }
+
+    private func installOutsideClickMonitor() {
+        guard outsideClickMonitor == nil else { return }
+        // Mouse-only monitoring does not require Accessibility or Input
+        // Monitoring permission. No global keyboard event tap is installed.
+        outsideClickMonitor = NSEvent.addGlobalMonitorForEvents(matching: [.leftMouseDown, .rightMouseDown]) { [weak self] _ in
+            Task { @MainActor [weak self] in self?.hide() }
+        }
+    }
+
+    private func removeOutsideClickMonitor() {
+        if let outsideClickMonitor { NSEvent.removeMonitor(outsideClickMonitor) }
+        outsideClickMonitor = nil
+    }
+}
+
+@MainActor
+private final class QuickQuestionPanel: NSPanel {
+    var onDismiss: (() -> Void)?
+    override var canBecomeKey: Bool { true }
+    override var canBecomeMain: Bool { false }
+    override func cancelOperation(_ sender: Any?) { onDismiss?() }
+}

ios/Apps/SybilMacQuickQuestion/Sources/QuickQuestionPanelView.swift

@@ -0,0 +1,300 @@
+import AppKit
+
+@MainActor
+final class QuickQuestionPanelView: NSVisualEffectView, NSTextViewDelegate {
+    weak var delegate: (any SybilQuickQuestionPanelDelegate)?
+    var onDismiss: (() -> Void)?
+
+    let promptView = QuickQuestionPromptView()
+    private let stack = NSStackView()
+    private let answerView = NSTextView()
+    private let answerScroll = NSScrollView()
+    private let statusLabel = NSTextField(wrappingLabelWithString: "")
+    private let providerPopup = NSPopUpButton()
+    private let modelPopup = NSPopUpButton()
+    private let convertButton = NSButton(title: "Open in Chat", target: nil, action: nil)
+    private let openAppButton = NSButton(title: "Open Sybil", target: nil, action: nil)
+    private let sendButton = NSButton()
+    private let shortcutLabel = NSTextField(labelWithString: "⌥ Space")
+    private var panelState = SybilQuickQuestionPanelState()
+
+    var preferredHeight: CGFloat { stack.fittingSize.height + 40 }
+
+    override init(frame frameRect: NSRect) {
+        super.init(frame: frameRect)
+        material = .hudWindow
+        blendingMode = .behindWindow
+        state = .active
+        wantsLayer = true
+        layer?.cornerRadius = 18
+        layer?.masksToBounds = true
+        layer?.borderWidth = 1
+        layer?.borderColor = NSColor.white.withAlphaComponent(0.14).cgColor
+
+        stack.orientation = .vertical
+        stack.alignment = .leading
+        stack.spacing = 12
+        stack.detachesHiddenViews = true
+        stack.translatesAutoresizingMaskIntoConstraints = false
+        addSubview(stack)
+        NSLayoutConstraint.activate([
+            stack.leadingAnchor.constraint(equalTo: leadingAnchor, constant: 20),
+            stack.trailingAnchor.constraint(equalTo: trailingAnchor, constant: -20),
+            stack.topAnchor.constraint(equalTo: topAnchor, constant: 20),
+            stack.bottomAnchor.constraint(equalTo: bottomAnchor, constant: -20)
+        ])
+
+        addRow(makeHeader())
+        configureTextView(promptView, fontSize: 22)
+        promptView.delegate = self
+        promptView.isRichText = false
+        promptView.allowsUndo = true
+        promptView.setAccessibilityLabel("Quick question prompt")
+        promptView.onSubmit = { [weak self] in self?.submit() }
+        promptView.onToggle = { [weak self] in self?.delegate?.quickQuestionToggleRequested() }
+        let promptScroll = makeScrollView(documentView: promptView)
+        promptScroll.heightAnchor.constraint(equalToConstant: 64).isActive = true
+        addRow(promptScroll)
+
+        configureTextView(answerView, fontSize: 15)
+        answerView.isEditable = false
+        answerView.isSelectable = true
+        answerView.setAccessibilityLabel("Quick question answer")
+        answerView.linkTextAttributes = [.foregroundColor: NSColor.systemPurple, .underlineStyle: NSUnderlineStyle.single.rawValue]
+        answerScroll.drawsBackground = false
+        answerScroll.hasVerticalScroller = true
+        answerScroll.autohidesScrollers = true
+        answerScroll.documentView = answerView
+        answerScroll.heightAnchor.constraint(equalToConstant: 280).isActive = true
+        addRow(answerScroll)
+
+        statusLabel.font = .systemFont(ofSize: 12)
+        statusLabel.textColor = .secondaryLabelColor
+        statusLabel.maximumNumberOfLines = 3
+        addRow(statusLabel)
+        addRow(makeControls())
+        let hint = NSTextField(labelWithString: "Return to ask   ·   Shift+Return for a new line   ·   Esc to close")
+        hint.font = .systemFont(ofSize: 11)
+        hint.textColor = .secondaryLabelColor
+        addRow(hint)
+        update(SybilQuickQuestionPanelState(), shortcutAvailable: true)
+    }
+
+    required init?(coder: NSCoder) { fatalError("init(coder:) has not been implemented") }
+
+    func update(_ next: SybilQuickQuestionPanelState, shortcutAvailable: Bool) {
+        let previous = panelState
+        panelState = next
+        if promptView.string != next.prompt, !promptView.hasMarkedText() {
+            promptView.string = next.prompt
+            promptView.setSelectedRange(NSRange(location: (next.prompt as NSString).length, length: 0))
+        }
+        promptView.isEditable = !next.isBusy
+        promptView.needsDisplay = true
+        if previous.answer != next.answer {
+            let wasAtBottom = answerScroll.contentView.bounds.maxY >= answerView.bounds.maxY - 36
+            answerView.textStorage?.setAttributedString(Self.renderAnswer(next.answer))
+            if wasAtBottom { answerView.scrollToEndOfDocument(nil) }
+        }
+        answerScroll.isHidden = next.answer.isEmpty
+
+        if providerPopup.numberOfItems == 0 || previous.providers != next.providers {
+            providerPopup.removeAllItems()
+            for provider in next.providers {
+                providerPopup.addItem(withTitle: provider.title)
+                providerPopup.lastItem?.representedObject = provider.id
+            }
+        }
+        if let index = next.providers.firstIndex(where: { $0.id == next.provider }) {
+            providerPopup.selectItem(at: index)
+        }
+        if modelPopup.numberOfItems == 0 || previous.models != next.models || previous.model != next.model {
+            modelPopup.removeAllItems()
+            var models = next.models
+            if !next.model.isEmpty, !models.contains(next.model) { models.insert(next.model, at: 0) }
+            modelPopup.addItems(withTitles: models)
+            modelPopup.selectItem(withTitle: next.model)
+        }
+        providerPopup.isEnabled = next.isAuthenticated && !next.isBusy
+        modelPopup.isEnabled = next.isAuthenticated && !next.isBusy
+        convertButton.isEnabled = next.canConvert
+        sendButton.isEnabled = next.canSend
+        openAppButton.isHidden = next.isAuthenticated
+        convertButton.isHidden = !next.isAuthenticated
+        shortcutLabel.stringValue = shortcutAvailable ? "⌥ Space" : "Shortcut unavailable"
+        shortcutLabel.toolTip = shortcutAvailable ? "Option+Space toggles Quick Question" : "Another app may be using Option+Space. Open Quick Question from Sybil’s File menu."
+
+        let status: String
+        if let error = next.error {
+            status = error
+            statusLabel.textColor = .systemRed
+        } else {
+            statusLabel.textColor = .secondaryLabelColor
+            if next.isCheckingSession {
+                status = "Connecting to Sybil…"
+            } else if !next.isAuthenticated {
+                status = "Connect in Sybil before asking a question."
+            } else if next.isConverting {
+                status = "Opening in chat…"
+            } else if next.isSending {
+                status = next.toolSummaries.last ?? "Thinking…"
+            } else {
+                status = ""
+            }
+        }
+        statusLabel.stringValue = status
+        statusLabel.isHidden = status.isEmpty
+    }
+
+    func preparePromptForPresentation() {
+        // Select only when invoked, never on streamed updates or normal edits.
+        promptView.selectAll(nil)
+    }
+
+    func textDidChange(_ notification: Notification) {
+        guard !panelState.isBusy else { return }
+        promptView.needsDisplay = true
+        delegate?.quickQuestionPromptChanged(promptView.string)
+    }
+
+    private func addRow(_ view: NSView) {
+        stack.addArrangedSubview(view)
+        view.widthAnchor.constraint(equalTo: stack.widthAnchor).isActive = true
+    }
+
+    private func makeHeader() -> NSView {
+        let icon = NSImageView(image: NSImage(systemSymbolName: "sparkles", accessibilityDescription: nil)!)
+        icon.contentTintColor = .systemPurple
+        icon.widthAnchor.constraint(equalToConstant: 18).isActive = true
+        let title = NSTextField(labelWithString: "Quick Question")
+        title.font = .systemFont(ofSize: 14, weight: .semibold)
+        shortcutLabel.font = .monospacedSystemFont(ofSize: 11, weight: .medium)
+        shortcutLabel.textColor = .secondaryLabelColor
+        let close = NSButton(image: NSImage(systemSymbolName: "xmark", accessibilityDescription: "Close Quick Question")!, target: self, action: #selector(closePanel))
+        close.isBordered = false
+        close.toolTip = "Close (Esc)"
+        close.setAccessibilityLabel("Close Quick Question")
+        close.widthAnchor.constraint(equalToConstant: 24).isActive = true
+        let row = NSStackView(views: [icon, title, NSView(), shortcutLabel, close])
+        row.spacing = 8
+        row.alignment = .centerY
+        row.heightAnchor.constraint(equalToConstant: 24).isActive = true
+        return row
+    }
+
+    private func makeControls() -> NSView {
+        providerPopup.target = self
+        providerPopup.action = #selector(changeProvider)
+        providerPopup.setAccessibilityLabel("Quick question provider")
+        providerPopup.widthAnchor.constraint(equalToConstant: 120).isActive = true
+        modelPopup.target = self
+        modelPopup.action = #selector(changeModel)
+        modelPopup.setAccessibilityLabel("Quick question model")
+        modelPopup.widthAnchor.constraint(lessThanOrEqualToConstant: 230).isActive = true
+        modelPopup.setContentCompressionResistancePriority(.defaultLow, for: .horizontal)
+        convertButton.target = self
+        convertButton.action = #selector(convertToChat)
+        convertButton.bezelStyle = .rounded
+        openAppButton.target = self
+        openAppButton.action = #selector(openApp)
+        openAppButton.bezelStyle = .rounded
+        sendButton.image = NSImage(systemSymbolName: "arrow.up", accessibilityDescription: "Ask Quick Question")
+        sendButton.imagePosition = .imageOnly
+        sendButton.bezelStyle = .circular
+        sendButton.contentTintColor = .systemPurple
+        sendButton.target = self
+        sendButton.action = #selector(submit)
+        sendButton.setAccessibilityLabel("Ask Quick Question")
+        sendButton.toolTip = "Ask (Return)"
+        sendButton.widthAnchor.constraint(equalToConstant: 32).isActive = true
+        let row = NSStackView(views: [providerPopup, modelPopup, NSView(), openAppButton, convertButton, sendButton])
+        row.spacing = 8
+        row.alignment = .centerY
+        row.detachesHiddenViews = true
+        row.heightAnchor.constraint(equalToConstant: 32).isActive = true
+        return row
+    }
+
+    private func configureTextView(_ view: NSTextView, fontSize: CGFloat) {
+        view.drawsBackground = false
+        view.textColor = .labelColor
+        view.insertionPointColor = .labelColor
+        view.font = .systemFont(ofSize: fontSize)
+        view.textContainerInset = NSSize(width: 0, height: 4)
+        view.isHorizontallyResizable = false
+        view.isVerticallyResizable = true
+        view.autoresizingMask = [.width]
+        view.textContainer?.widthTracksTextView = true
+        view.textContainer?.lineFragmentPadding = 0
+    }
+
+    private func makeScrollView(documentView: NSView) -> NSScrollView {
+        let scroll = NSScrollView()
+        scroll.drawsBackground = false
+        scroll.hasVerticalScroller = true
+        scroll.autohidesScrollers = true
+        scroll.documentView = documentView
+        return scroll
+    }
+
+    private static func renderAnswer(_ text: String) -> NSAttributedString {
+        let parsed = (try? AttributedString(markdown: text, options: .init(interpretedSyntax: .inlineOnlyPreservingWhitespace))) ?? AttributedString(text)
+        let rendered = NSMutableAttributedString(attributedString: NSAttributedString(parsed))
+        let paragraph = NSMutableParagraphStyle()
+        paragraph.lineSpacing = 4
+        rendered.addAttributes([.foregroundColor: NSColor.labelColor, .paragraphStyle: paragraph], range: NSRange(location: 0, length: rendered.length))
+        for run in parsed.runs {
+            let intent = run.inlinePresentationIntent ?? []
+            var font = intent.contains(.code) ? NSFont.monospacedSystemFont(ofSize: 14, weight: .regular) : NSFont.systemFont(ofSize: 15)
+            if intent.contains(.stronglyEmphasized) { font = NSFontManager.shared.convert(font, toHaveTrait: .boldFontMask) }
+            if intent.contains(.emphasized) { font = NSFontManager.shared.convert(font, toHaveTrait: .italicFontMask) }
+            rendered.addAttribute(.font, value: font, range: NSRange(run.range, in: parsed))
+        }
+        return rendered
+    }
+
+    @objc private func closePanel() { onDismiss?() }
+    @objc private func changeProvider() {
+        guard let provider = providerPopup.selectedItem?.representedObject as? String else { return }
+        delegate?.quickQuestionProviderChanged(provider)
+    }
+    @objc private func changeModel() {
+        guard let model = modelPopup.titleOfSelectedItem else { return }
+        delegate?.quickQuestionModelChanged(model)
+    }
+    @objc private func convertToChat() { delegate?.quickQuestionConvertRequested() }
+    @objc private func openApp() { delegate?.quickQuestionOpenAppRequested() }
+    @objc private func submit() {
+        guard panelState.canSend, !promptView.hasMarkedText() else { return }
+        delegate?.quickQuestionSubmitRequested()
+    }
+}
+
+@MainActor
+final class QuickQuestionPromptView: NSTextView {
+    var onSubmit: (() -> Void)?
+    var onToggle: (() -> Void)?
+
+    override func draw(_ dirtyRect: NSRect) {
+        super.draw(dirtyRect)
+        guard string.isEmpty else { return }
+        ("Ask anything…" as NSString).draw(
+            at: NSPoint(x: textContainerInset.width, y: textContainerInset.height),
+            withAttributes: [.font: font ?? NSFont.systemFont(ofSize: 22), .foregroundColor: NSColor.placeholderTextColor]
+        )
+    }
+
+    override func keyDown(with event: NSEvent) {
+        let modifiers = event.modifierFlags.intersection([.command, .control, .option, .shift])
+        if event.keyCode == 49, modifiers == .option {
+            if !event.isARepeat { onToggle?() }
+            return
+        }
+        if (event.keyCode == 36 || event.keyCode == 76), modifiers.isEmpty, !hasMarkedText() {
+            if !event.isARepeat { onSubmit?() }
+            return
+        }
+        // Let the text input system handle Shift+Return and IME confirmation.
+        super.keyDown(with: event)
+    }
+}

ios/Apps/SybilMacQuickQuestion/project.yml

@@ -0,0 +1,25 @@
+targets:
+  SybilMacQuickQuestion:
+    type: bundle
+    platform: macOS
+    deploymentTarget: "15.0"
+    sources:
+      - Sources
+      - ../../Packages/Sybil/Sources/Sybil/SybilMacQuickQuestionBridge.swift
+    dependencies:
+      - sdk: AppKit.framework
+      - sdk: Carbon.framework
+    settings:
+      base:
+        PRODUCT_BUNDLE_IDENTIFIER: net.buzzert.sybil2.quick-question-panel
+        PRODUCT_NAME: SybilMacQuickQuestion
+        PRODUCT_MODULE_NAME: SybilMacQuickQuestion
+        DEVELOPMENT_TEAM: DQQH5H6GBD
+        CODE_SIGN_STYLE: Automatic
+        SWIFT_VERSION: 6.0
+        SUPPORTS_MACCATALYST: NO
+        GENERATE_INFOPLIST_FILE: YES
+        INFOPLIST_KEY_CFBundleExecutable: $(EXECUTABLE_NAME)
+        INFOPLIST_KEY_NSPrincipalClass: SybilMacQuickQuestionPlugin
+        MACH_O_TYPE: mh_bundle
+        SKIP_INSTALL: YES

ios/Packages/Sybil/Package.swift

@@ -24,7 +24,8 @@ let package = Package(
             name: "Sybil",
             dependencies: [
                 .product(name: "MarkdownUI", package: "swift-markdown-ui")
-            ]),
+            ],
+            resources: [.process("Resources")]),
         .testTarget(
             name: "SybilTests",
             dependencies: ["Sybil"]

ios/Packages/Sybil/Sources/Sybil/Resources/Agent.xcassets/Contents.json

@@ -0,0 +1,6 @@
+{
+  "info": {
+    "author": "xcode",
+    "version": 1
+  }
+}

ios/Packages/Sybil/Sources/Sybil/Resources/Agent.xcassets/character-still.imageset/Contents.json

@@ -0,0 +1,12 @@
+{
+  "images": [
+    {
+      "filename": "character-still.png",
+      "idiom": "universal"
+    }
+  ],
+  "info": {
+    "author": "xcode",
+    "version": 1
+  }
+}

ios/Packages/Sybil/Sources/Sybil/SplitView.swift

@@ -78,6 +78,7 @@ public struct SplitView: View {
         .font(.sybil(.body))
         .preferredColorScheme(.dark)
         .focusedSceneValue(\.sybilKeyboardActions, keyboardActions)
+        #if !targetEnvironment(macCatalyst)
         .sheet(isPresented: $isQuickQuestionPresented, onDismiss: handleQuickQuestionDismissed) {
             SybilQuickQuestionView(
                 viewModel: viewModel,
@@ -85,7 +86,11 @@ public struct SplitView: View {
             )
             .presentationDragIndicator(.visible)
         }
+        #endif
         .task {
+            #if targetEnvironment(macCatalyst)
+            SybilMacQuickQuestionController.shared.attach(viewModel)
+            #endif
             await viewModel.bootstrap()
             presentPendingQuickQuestionIfPossible()
         }
@@ -101,12 +106,18 @@ public struct SplitView: View {
         .onChange(of: viewModel.isAuthenticated) { _, _ in
             presentPendingQuickQuestionIfPossible()
         }
+        .onChange(of: isQuickQuestionPresented) { _, presented in
+            viewModel.agent.isModalPresented = presented
+        }
         .onChange(of: scenePhase) { _, nextPhase in
             switch nextPhase {
             case .background:
                 shouldRefreshOnForeground = true
                 viewModel.markAppInactiveForNetwork()
             case .active:
+                #if targetEnvironment(macCatalyst)
+                SybilMacQuickQuestionController.shared.attach(viewModel)
+                #endif
                 viewModel.markAppActiveForNetwork()
                 guard shouldRefreshOnForeground, horizontalSizeClass != .compact else {
                     return
@@ -151,8 +162,13 @@ public struct SplitView: View {
         }
 
         hasPendingQuickQuestionPresentation = false
+        #if targetEnvironment(macCatalyst)
+        SybilMacQuickQuestionController.shared.attach(viewModel)
+        SybilMacQuickQuestionController.shared.show()
+        #else
         quickQuestionFocusRequest += 1
         isQuickQuestionPresented = true
+        #endif
     }
 
     private func handleQuickQuestionDismissed() {
@@ -161,23 +177,57 @@ public struct SplitView: View {
 }
 
 public struct SybilCommands: Commands {
+    public static let workspaceWindowID = "sybil-workspace"
+
     @FocusedValue(\.sybilKeyboardActions) private var keyboardActions
+    #if targetEnvironment(macCatalyst)
+    @Environment(\.openWindow) private var openWindow
+    @ObservedObject private var windowCommands = SybilMacWindowCommands.shared
+    #endif
 
     public init() {}
 
     public var body: some Commands {
         CommandGroup(replacing: .newItem) {
             Button("New Chat") {
+                #if targetEnvironment(macCatalyst)
+                windowCommands.newChatOrWindow(
+                    newChat: keyboardActions?.newChat,
+                    openWindow: openNewWindow
+                )
+                #else
                 keyboardActions?.newChat()
+                #endif
             }
             .keyboardShortcut("n", modifiers: .command)
+            #if targetEnvironment(macCatalyst)
+            .disabled(!windowCommands.canStartNewChat(hasFocusedActions: keyboardActions != nil))
+            #else
             .disabled(keyboardActions == nil)
+            #endif
+
+            #if targetEnvironment(macCatalyst)
+            Button("New Window", action: openNewWindow)
+                .keyboardShortcut("n", modifiers: [.command, .shift])
+            #endif
 
             Button("New Search") {
                 keyboardActions?.newSearch()
             }
+            #if targetEnvironment(macCatalyst)
+            .keyboardShortcut("n", modifiers: [.command, .option])
+            #else
             .keyboardShortcut("n", modifiers: [.command, .shift])
+            #endif
             .disabled(keyboardActions == nil)
+
+            #if targetEnvironment(macCatalyst)
+            Divider()
+            Button("Quick Question") {
+                SybilMacQuickQuestionController.shared.toggle()
+            }
+            .keyboardShortcut(.space, modifiers: .option)
+            #endif
         }
 
         CommandMenu("Conversation") {
@@ -194,6 +244,13 @@ public struct SybilCommands: Commands {
             .disabled(keyboardActions == nil)
         }
     }
+    #if targetEnvironment(macCatalyst)
+    private func openNewWindow() {
+        // Targeting a WindowGroup without a value creates a fresh scene every
+        // time, even when another workspace window is already open.
+        openWindow(id: Self.workspaceWindowID)
+    }
+    #endif
 }
 
 private struct SybilKeyboardActions {

ios/Packages/Sybil/Sources/Sybil/SybilAPIClient.swift

@@ -17,7 +17,7 @@ struct AnyEncodable: Encodable {
     }
 }
 
-actor SybilAPIClient: SybilAPIClienting {
+actor SybilAPIClient: SybilAPIClienting, AgentAPIClienting {
     private let configuration: APIConfiguration
     private let session: URLSession
 
@@ -40,6 +40,48 @@ actor SybilAPIClient: SybilAPIClienting {
         self.session = session
     }
 
+    func agentDescriptor() async throws -> AgentDescriptor {
+        let value = try await request("/v1/agent", method: "GET", responseType: AgentDescriptor.self)
+        guard value.type == "persistent-agent", value.protocol == "pi-room.v1" else {
+            throw AgentFailure(message: "Unsupported Agent room protocol.")
+        }
+        return value
+    }
+
+    func agentMessages(after: String?, before: String?) async throws -> AgentPage {
+        guard after == nil || before == nil else { throw AgentFailure(message: "Use one history cursor.") }
+        var query = [URLQueryItem(name: "limit", value: "200")]
+        if let after { query.append(URLQueryItem(name: "after", value: try AgentCursor.normalize(after))) }
+        if let before { query.append(URLQueryItem(name: "before", value: try AgentCursor.normalize(before))) }
+        let page = try await request("/v1/agent/messages", method: "GET", queryItems: query, responseType: AgentPage.self)
+        try page.validate()
+        return page
+    }
+
+    func agentReadState() async throws -> AgentReadState {
+        try await request("/v1/agent/read-state", method: "GET", responseType: AgentReadState.self)
+    }
+
+    func agentMarkRead(cursor: String) async throws -> AgentReadState {
+        try await request("/v1/agent/read-state", method: "PUT", body: AnyEncodable(["cursor": try AgentCursor.normalize(cursor)]), responseType: AgentReadState.self)
+    }
+
+    func agentSubmit(content: String, requestID: String) async throws -> AgentReceipt {
+        let receipt = try await request("/v1/agent/messages", method: "POST", body: AnyEncodable(["content": content, "requestId": requestID]), responseType: AgentReceipt.self)
+        try receipt.validate(requestID: requestID)
+        return receipt
+    }
+
+    func agentSocket(after: String) async throws -> any AgentSocket {
+        let task = session.webSocketTask(with: try configuration.agentWebSocketURL())
+        task.maximumMessageSize = 8 * 1024 * 1024
+        task.resume()
+        let socket = AgentURLSocket(task: task)
+        do { try await socket.send(configuration.agentAuthFrame(after: after)) }
+        catch { await socket.close(); throw error }
+        return socket
+    }
+
     func verifySession() async throws -> AuthSession {
         try await request("/v1/auth/session", method: "GET", responseType: AuthSession.self)
     }
@@ -49,6 +91,15 @@ actor SybilAPIClient: SybilAPIClienting {
         return response.items
     }
 
+    func searchWorkspaceItems(query: String) async throws -> WorkspaceSearchResponse {
+        try await request(
+            "/v1/workspace-items",
+            method: "GET",
+            queryItems: [URLQueryItem(name: "q", value: query)],
+            responseType: WorkspaceSearchResponse.self
+        )
+    }
+
     func listChats() async throws -> [ChatSummary] {
         let response = try await request("/v1/chats", method: "GET", responseType: ChatListResponse.self)
         return response.chats
@@ -74,6 +125,16 @@ actor SybilAPIClient: SybilAPIClienting {
         return response.chat
     }
 
+    func forkChat(chatID: String, messageID: String? = nil) async throws -> ChatSummary {
+        let response = try await request(
+            "/v1/chats/\(chatID)/fork",
+            method: "POST",
+            body: AnyEncodable(ChatForkBody(messageId: messageID)),
+            responseType: ChatCreateResponse.self
+        )
+        return response.chat
+    }
+
     func updateChatTitle(chatID: String, title: String) async throws -> ChatSummary {
         let response = try await request(
             "/v1/chats/\(chatID)",
@@ -183,6 +244,29 @@ actor SybilAPIClient: SybilAPIClienting {
         SybilLog.info(SybilLog.network, "Chat stream completed")
     }
 
+    func runQuickQuestionStream(
+        body: QuickQuestionStreamRequest,
+        onEvent: @escaping @Sendable (CompletionStreamEvent) async -> Void
+    ) async throws {
+        let request = try makeRequest(
+            path: "/v1/quick-questions/stream",
+            method: "POST",
+            body: AnyEncodable(body),
+            acceptsSSE: true
+        )
+
+        SybilLog.info(
+            SybilLog.network,
+            "Starting quick question stream POST \(request.url?.absoluteString ?? "<unknown>")"
+        )
+
+        try await stream(request: request) { eventName, dataText in
+            try await Self.handleCompletionStreamEvent(eventName: eventName, dataText: dataText, onEvent: onEvent)
+        }
+
+        SybilLog.info(SybilLog.network, "Quick question stream completed")
+    }
+
     func attachCompletionStream(
         chatID: String,
         onEvent: @escaping @Sendable (CompletionStreamEvent) async -> Void
@@ -257,9 +341,10 @@ actor SybilAPIClient: SybilAPIClienting {
         _ path: String,
         method: String,
         body: AnyEncodable? = nil,
+        queryItems: [URLQueryItem] = [],
         responseType: Response.Type
     ) async throws -> Response {
-        let request = try makeRequest(path: path, method: method, body: body, acceptsSSE: false)
+        let request = try makeRequest(path: path, method: method, body: body, acceptsSSE: false, queryItems: queryItems)
 
         SybilLog.debug(
             SybilLog.network,
@@ -311,12 +396,14 @@ actor SybilAPIClient: SybilAPIClienting {
         path: String,
         method: String,
         body: AnyEncodable?,
-        acceptsSSE: Bool
+        acceptsSSE: Bool,
+        queryItems: [URLQueryItem] = []
     ) throws -> URLRequest {
-        let url = try buildURL(path: path)
+        let url = try buildURL(path: path, queryItems: queryItems)
         var request = URLRequest(url: url)
         request.httpMethod = method
-        request.timeoutInterval = 120
+        if path.hasPrefix("/v1/agent") { request.cachePolicy = .reloadIgnoringLocalCacheData }
+        request.timeoutInterval = path.hasPrefix("/v1/agent") ? 30 : 120
 
         if acceptsSSE {
             request.setValue("text/event-stream", forHTTPHeaderField: "Accept")
@@ -336,7 +423,7 @@ actor SybilAPIClient: SybilAPIClienting {
         return request
     }
 
-    private func buildURL(path: String) throws -> URL {
+    private func buildURL(path: String, queryItems: [URLQueryItem]) throws -> URL {
         guard var components = URLComponents(url: configuration.baseURL, resolvingAgainstBaseURL: false) else {
             throw APIError.invalidBaseURL
         }
@@ -347,6 +434,9 @@ actor SybilAPIClient: SybilAPIClienting {
             basePath.removeLast()
         }
         components.path = "\(basePath)/\(trimmedPath)"
+        if !queryItems.isEmpty {
+            components.queryItems = (components.queryItems ?? []) + queryItems
+        }
 
         guard let url = components.url else {
             throw APIError.invalidBaseURL
@@ -664,6 +754,12 @@ struct CompletionStreamRequest: Codable, Sendable {
     var userLocation: String? = nil
 }
 
+struct QuickQuestionStreamRequest: Codable, Sendable {
+    var provider: Provider
+    var model: String
+    var question: String
+}
+
 private struct ChatCreateBody: Encodable {
     var title: String?
     var provider: Provider?
@@ -675,6 +771,10 @@ private struct ChatTitleUpdateBody: Encodable {
     var title: String
 }
 
+private struct ChatForkBody: Encodable {
+    var messageId: String?
+}
+
 private struct StarUpdateBody: Encodable {
     var starred: Bool
 }

ios/Packages/Sybil/Sources/Sybil/SybilAPIClienting.swift

@@ -3,6 +3,7 @@ import Foundation
 protocol SybilAPIClienting: Sendable {
     func verifySession() async throws -> AuthSession
     func listWorkspaceItems() async throws -> [WorkspaceItem]
+    func searchWorkspaceItems(query: String) async throws -> WorkspaceSearchResponse
     func listChats() async throws -> [ChatSummary]
     func createChat(
         title: String?,
@@ -11,6 +12,7 @@ protocol SybilAPIClienting: Sendable {
         messages: [CompletionRequestMessage]?
     ) async throws -> ChatSummary
     func getChat(chatID: String) async throws -> ChatDetail
+    func forkChat(chatID: String, messageID: String?) async throws -> ChatSummary
     func updateChatTitle(chatID: String, title: String) async throws -> ChatSummary
     func updateChatStar(chatID: String, starred: Bool) async throws -> ChatSummary
     func deleteChat(chatID: String) async throws
@@ -27,6 +29,10 @@ protocol SybilAPIClienting: Sendable {
         body: CompletionStreamRequest,
         onEvent: @escaping @Sendable (CompletionStreamEvent) async -> Void
     ) async throws
+    func runQuickQuestionStream(
+        body: QuickQuestionStreamRequest,
+        onEvent: @escaping @Sendable (CompletionStreamEvent) async -> Void
+    ) async throws
     func attachCompletionStream(
         chatID: String,
         onEvent: @escaping @Sendable (CompletionStreamEvent) async -> Void

ios/Packages/Sybil/Sources/Sybil/SybilAgentModels.swift

@@ -0,0 +1,241 @@
+import Foundation
+
+struct AgentFailure: LocalizedError, Sendable {
+    let message: String
+    var errorDescription: String? { message }
+}
+
+// Compare decimal strings without losing precision through Double or overflowing Int.
+enum AgentCursor {
+    static func normalize(_ value: String) throws -> String {
+        guard !value.isEmpty, value.utf8.allSatisfy({ (48...57).contains($0) }) else {
+            throw AgentFailure(message: "Invalid Agent cursor.")
+        }
+        let trimmed = value.drop(while: { $0 == "0" })
+        return trimmed.isEmpty ? "0" : String(trimmed)
+    }
+    static func less(_ a: String, _ b: String) -> Bool {
+        a.count == b.count ? a < b : a.count < b.count
+    }
+    static func advance(_ current: String, _ incoming: String) throws -> String {
+        let value = try normalize(incoming)
+        return less(current, value) ? value : current
+    }
+}
+
+struct AgentActivity: Codable, Equatable, Sendable {
+    struct Tool: Codable, Equatable, Sendable {
+        var id: String
+        var name: String
+        var status: String
+        var output: String?
+    }
+    struct Wait: Codable, Equatable, Sendable {
+        var due: Double
+        var reason: String?
+    }
+    var state: String = "idle"
+    var isThinking = false
+    var queued = 0
+    var backgroundWorkers = 0
+    var tools: [Tool] = []
+    var queuedMessages: [Message] = []
+    var wait: Wait?
+    var error: String?
+    var isBusy: Bool { isThinking || backgroundWorkers > 0 }
+}
+
+struct AgentDescriptor: Decodable, Sendable {
+    var id: String
+    var title: String
+    var type: String
+    var model: String
+    var `protocol`: String
+    var activity: AgentActivity
+    var websocket: String
+}
+struct AgentPage: Codable, Sendable {
+    var conversationId: String
+    var messages: [Message]
+    var cursor: String?
+    var beforeCursor: String?
+    var hasMore: Bool
+
+    func validate() throws {
+        guard !conversationId.isEmpty, messages.allSatisfy({ !$0.id.isEmpty }) else {
+            throw AgentFailure(message: "Invalid Agent history.")
+        }
+        if let cursor { _ = try AgentCursor.normalize(cursor) }
+        if let beforeCursor { _ = try AgentCursor.normalize(beforeCursor) }
+    }
+}
+struct AgentReadState: Codable, Sendable {
+    var cursor: String
+    var updatedAt: String?
+}
+struct AgentReceipt: Codable, Sendable {
+    var id: Int64
+    var status: String
+    var messageId: String
+    func validate(requestID: String) throws {
+        guard id > 0, id <= 9_007_199_254_740_991,
+              ["queued", "placed", "done", "unanswered"].contains(status),
+              messageId == "inbound:sybil:\(requestID)" else {
+            throw AgentFailure(message: "Invalid Agent receipt. Retry to confirm delivery.")
+        }
+    }
+}
+
+enum AgentEvent: Sendable {
+    case ready(String?)
+    case messages(AgentPage)
+    case activity(String, AgentActivity)
+    case heartbeat
+    case error(String)
+
+    static func decode(_ data: Data) throws -> AgentEvent? {
+        struct Envelope: Decodable {
+            var type: String
+            var conversationId: String?
+            var activity: AgentActivity?
+            var error: String?
+        }
+        let envelope = try AgentCoding.decoder().decode(Envelope.self, from: data)
+        switch envelope.type {
+        case "ready": return .ready(envelope.conversationId)
+        case "messages":
+            let page = try AgentCoding.decoder().decode(AgentPage.self, from: data)
+            try page.validate()
+            return .messages(page)
+        case "activity":
+            guard let id = envelope.conversationId, let activity = envelope.activity else {
+                throw AgentFailure(message: "Invalid Agent activity frame.")
+            }
+            return .activity(id, activity)
+        case "pong", "heartbeat": return .heartbeat
+        case "error": return .error(envelope.error ?? "Agent connection error.")
+        default: return nil
+        }
+    }
+}
+
+enum AgentCoding {
+    static func decoder() -> JSONDecoder {
+        let decoder = JSONDecoder()
+        decoder.dateDecodingStrategy = .custom { decoder in
+            let container = try decoder.singleValueContainer()
+            let text = try container.decode(String.self)
+            let formatter = ISO8601DateFormatter()
+            formatter.formatOptions = [.withInternetDateTime, .withFractionalSeconds]
+            if let date = formatter.date(from: text) { return date }
+            formatter.formatOptions = [.withInternetDateTime]
+            guard let date = formatter.date(from: text) else {
+                throw DecodingError.dataCorruptedError(in: container, debugDescription: "Invalid message date")
+            }
+            return date
+        }
+        return decoder
+    }
+}
+
+extension Message {
+    var agentEntryCursor: String? {
+        let value = metadata?.objectValue?["entryId"]
+        if let text = value?.stringValue { return try? AgentCursor.normalize(text) }
+        if let number = value?.numberValue, number >= 0, number <= 9_007_199_254_740_991, number.rounded() == number {
+            return String(Int64(number))
+        }
+        return nil
+    }
+    var isAgentPublicReply: Bool {
+        role == .assistant && metadata?.objectValue?["kind"]?.stringValue == "message"
+    }
+    var agentToolID: String { toolCallMetadata?.toolCallId ?? id }
+}
+
+struct AgentRoomState: Equatable, Sendable {
+    var conversationID: String?
+    var messages: [Message] = []
+    var cursor = "0"
+    var readCursor = "0"
+    var activity = AgentActivity()
+    var unreadCount: Int {
+        messages.filter { $0.isAgentPublicReply && $0.agentEntryCursor.map { AgentCursor.less(readCursor, $0) } == true }.count
+    }
+    mutating func checkConversation(_ id: String) throws {
+        guard conversationID == nil || conversationID == id else {
+            throw AgentFailure(message: "The Agent room changed. Reconnect in Settings to reload it.")
+        }
+        conversationID = id
+    }
+    mutating func apply(_ page: AgentPage) throws {
+        try page.validate()
+        var next = self
+        try next.checkConversation(page.conversationId)
+        next.cursor = try page.cursor.map { try AgentCursor.advance(cursor, $0) } ?? cursor
+        next.messages = Self.merge(messages, page.messages)
+        self = next
+    }
+    static func merge(_ previous: [Message], _ incoming: [Message]) -> [Message] {
+        var byID = Dictionary(previous.map { ($0.id, $0) }, uniquingKeysWith: { _, last in last })
+        for var message in incoming {
+            if let old = byID[message.id] {
+                let prior = old.metadata?.objectValue ?? [:]
+                let next = message.metadata?.objectValue ?? [:]
+                if prior["status"] == .string("complete") && next["status"] == .string("queued") { continue }
+                if old.toolCallMetadata != nil, message.toolCallMetadata != nil {
+                    let stale = ["completed", "failed"].contains(old.toolCallMetadata?.status ?? "") && message.toolCallMetadata?.status == "initiated"
+                    var merged = prior.merging(next, uniquingKeysWith: { a, b in stale ? a : b })
+                    if next["startedAt"] == nil || next["startedAt"] == .null { merged["startedAt"] = prior["startedAt"] }
+                    if let args = prior["args"], next["args"]?.objectValue?.isEmpty != false { merged["args"] = args }
+                    message.metadata = .object(merged)
+                    message.createdAt = min(old.createdAt, message.createdAt)
+                    if stale { message.content = old.content; message.name = old.name }
+                }
+            }
+            byID[message.id] = message
+        }
+        return byID.values.sorted {
+            if $0.createdAt != $1.createdAt { return $0.createdAt < $1.createdAt }
+            if let a = $0.agentEntryCursor, let b = $1.agentEntryCursor, a != b { return AgentCursor.less(a, b) }
+            return $0.id < $1.id
+        }
+    }
+}
+
+// The transcript latch is independent of the truthful sidebar/worker activity.
+struct AgentPresentation {
+    private(set) var suppressed = false
+    private(set) var busySince: Date?
+    private var seenTools: Set<String> = []
+    static let quietTools: Set<String> = ["user_message", "notify_parent", "clock_wait", "event_decision"]
+
+    mutating func receive(_ event: AgentEvent, before: AgentRoomState, after: AgentRoomState, replay: Bool, now: Date) {
+        switch event {
+        case let .messages(page):
+            let existing = Set(before.messages.map(\.id))
+            let fresh = page.messages.filter { !existing.contains($0.id) && $0.agentEntryCursor.map { AgentCursor.less(before.cursor, $0) } == true }
+                .sorted { AgentCursor.less($0.agentEntryCursor ?? "0", $1.agentEntryCursor ?? "0") }
+            if !replay {
+                for message in fresh {
+                    if message.isAgentPublicReply && !message.content.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty { suppressed = true }
+                    else if message.role == .user { suppressed = false }
+                    else if message.toolCallMetadata?.status == "initiated",
+                            !Self.quietTools.contains(message.toolCallMetadata?.toolName ?? ""), !seenTools.contains(message.agentToolID) { suppressed = false }
+                }
+            }
+            for message in page.messages where message.role == .tool { seenTools.insert(message.agentToolID) }
+        case let .activity(_, activity):
+            if !activity.isThinking || activity.tools.contains(where: { !seenTools.contains($0.id) && !Self.quietTools.contains($0.name) }) { suppressed = false }
+            seenTools.formUnion(activity.tools.map(\.id))
+        default: break
+        }
+        reconcile(activity: after.activity, now: now)
+    }
+    mutating func sent(now: Date, activity: AgentActivity) { suppressed = false; reconcile(activity: activity, now: now) }
+    private mutating func reconcile(activity: AgentActivity, now: Date) {
+        if suppressed || !activity.isBusy { busySince = nil }
+        else if busySince == nil { busySince = now }
+    }
+    func isTyping(now: Date) -> Bool { busySince.map { now.timeIntervalSince($0) >= 0.25 } ?? false }
+}

ios/Packages/Sybil/Sources/Sybil/SybilAgentOutbox.swift

@@ -0,0 +1,41 @@
+import CryptoKit
+import Foundation
+
+struct AgentOutboxRow: Codable, Equatable, Sendable {
+    var requestID: String
+    var content: String
+    var createdAt: Date
+    var status: String = "sending"
+    var seenOnServer = false
+    var confirmed = false
+    var error: String?
+    var id: String { "inbound:sybil:\(requestID)" }
+    var message: Message {
+        var info: [String: JSONValue] = ["kind": .string("message"), "status": .string(status), "requestId": .string(requestID)]
+        if let error { info["error"] = .string(error) }
+        return Message(id: id, createdAt: createdAt, role: .user, content: content, metadata: .object(info))
+    }
+}
+
+struct AgentOutboxStorage: Sendable {
+    var read: @Sendable () throws -> [AgentOutboxRow]
+    var write: @Sendable ([AgentOutboxRow]) throws -> Void
+
+    static func file(baseURL: URL) -> Self {
+        // Separate installations, but never persist tokens in a filename or payload.
+        let key = SHA256.hash(data: Data(baseURL.absoluteString.utf8)).map { String(format: "%02x", $0) }.joined()
+        let directory = URL.applicationSupportDirectory.appendingPathComponent("AgentOutbox", isDirectory: true)
+        let url = directory.appendingPathComponent("\(key).json")
+        return Self(read: {
+            guard FileManager.default.fileExists(atPath: url.path) else { return [] }
+            return try JSONDecoder().decode([AgentOutboxRow].self, from: Data(contentsOf: url))
+        }, write: { rows in
+            let data = try JSONEncoder().encode(rows)
+            guard rows.count <= 100, data.count <= 2 * 1024 * 1024 else {
+                throw AgentFailure(message: "Retry pending messages before sending more.")
+            }
+            try FileManager.default.createDirectory(at: directory, withIntermediateDirectories: true)
+            try data.write(to: url, options: [.atomic, .completeFileProtectionUntilFirstUserAuthentication])
+        })
+    }
+}

ios/Packages/Sybil/Sources/Sybil/SybilAgentRoom.swift

@@ -0,0 +1,371 @@
+import Foundation
+import Network
+import Observation
+
+@MainActor
+@Observable
+final class SybilAgentRoom {
+    struct Snapshot: Equatable {
+        var room = AgentRoomState()
+        var isTyping = false
+        var entranceIDs: Set<String> = []
+        var revision = 0
+    }
+    private(set) var snapshot = Snapshot()
+    private(set) var connection = "Offline"
+    private(set) var error: String?
+    private(set) var isLoading = false
+    private(set) var outbox: [AgentOutboxRow] = []
+    var composer = ""
+    var isObscured = false
+    var isModalPresented = false
+    var isSelected = false
+    private(set) var isActive = true
+
+    @ObservationIgnored private var client: (any AgentAPIClienting)?
+    @ObservationIgnored private var storage: AgentOutboxStorage?
+    @ObservationIgnored private var storageReadable = true
+    @ObservationIgnored private var pending = AgentRoomState()
+    @ObservationIgnored private var presenter = AgentPresentation()
+    @ObservationIgnored private var pendingEntrances: Set<String> = []
+    @ObservationIgnored private var generation = UUID()
+    @ObservationIgnored private var connectionTask: Task<Void, Never>?
+    @ObservationIgnored private var maintenanceTask: Task<Void, Never>?
+    @ObservationIgnored private var batchTask: Task<Void, Never>?
+    @ObservationIgnored private var typingTask: Task<Void, Never>?
+    @ObservationIgnored private var readTask: Task<Void, Never>?
+    @ObservationIgnored private var sends: [String: Task<Void, Never>] = [:]
+    @ObservationIgnored private var socket: (any AgentSocket)?
+    @ObservationIgnored private var monitor: NWPathMonitor?
+    @ObservationIgnored private var lastPacket = Date()
+    @ObservationIgnored private var socketStarted = Date()
+    @ObservationIgnored private var desiredRead = "0"
+    @ObservationIgnored private var acknowledgedRead = "0"
+    @ObservationIgnored private var hasLoaded = false
+    @ObservationIgnored private var syncing = false
+    @ObservationIgnored private var wasNetworkOffline = false
+
+    var isBusy: Bool { snapshot.room.activity.isBusy }
+    var unreadCount: Int { snapshot.room.unreadCount }
+    var messages: [Message] {
+        let committed = Set(snapshot.room.messages.map(\.id))
+        let queued = snapshot.room.activity.queuedMessages.filter { !committed.contains($0.id) }
+        var messages = AgentRoomState.merge(snapshot.room.messages, queued)
+        for row in outbox {
+            if let index = messages.firstIndex(where: { $0.id == row.id }) {
+                messages[index].createdAt = row.createdAt
+                if !row.confirmed {
+                    var info = messages[index].metadata?.objectValue ?? [:]
+                    info["status"] = .string(row.status)
+                    if let error = row.error { info["error"] = .string(error) }
+                    messages[index].metadata = .object(info)
+                }
+            } else { messages.append(row.message) }
+        }
+        return AgentRoomState.merge([], messages)
+    }
+
+    func configure(client: any AgentAPIClienting, storage: AgentOutboxStorage) {
+        stop()
+        self.client = client
+        self.storage = storage
+        pending = AgentRoomState(); snapshot = Snapshot(); presenter = AgentPresentation()
+        desiredRead = "0"; acknowledgedRead = "0"; hasLoaded = false; composer = ""
+        error = nil; storageReadable = true
+        do {
+            outbox = try storage.read().map { row in
+                var restored = row
+                if !restored.seenOnServer {
+                    restored.status = "failed"
+                    restored.error = "Delivery was interrupted. Retry to confirm."
+                }
+                return restored
+            }
+        } catch {
+            outbox = []; storageReadable = false
+            self.error = "Could not restore pending messages: \(error.localizedDescription)"
+        }
+        start()
+    }
+
+    func disconnect() {
+        stop()
+        client = nil
+    }
+
+    func stop() {
+        generation = UUID()
+        connectionTask?.cancel(); connectionTask = nil
+        maintenanceTask?.cancel(); maintenanceTask = nil
+        batchTask?.cancel(); batchTask = nil
+        typingTask?.cancel(); typingTask = nil
+        readTask?.cancel(); readTask = nil
+        sends.values.forEach { $0.cancel() }; sends = [:]
+        for index in outbox.indices where !outbox[index].seenOnServer && outbox[index].status == "sending" {
+            outbox[index].status = "failed"
+            outbox[index].error = "Delivery was interrupted. Retry to confirm."
+        }
+        if !outbox.isEmpty { try? persist() }
+        monitor?.cancel(); monitor = nil
+        let old = socket; socket = nil
+        Task { await old?.close() }
+        syncing = false
+        pendingEntrances = []
+        connection = "Offline"
+    }
+
+    func setActive(_ active: Bool) {
+        guard isActive != active else { return }
+        isActive = active
+        if active { start() }
+        else {
+            // Publish successfully merged input before cancelling; replay resumes here.
+            flush()
+            stop()
+        }
+    }
+
+    func retryConnection() { flush(); stop(); if isActive { start() } }
+
+    private func start() {
+        guard isActive, let client, connectionTask == nil else { return }
+        let run = generation
+        isLoading = !hasLoaded
+        connection = "Connecting"
+        let monitor = NWPathMonitor()
+        monitor.pathUpdateHandler = { [weak self] path in
+            let satisfied = path.status == .satisfied
+            Task { @MainActor [weak self] in
+                guard let self, self.generation == run else { return }
+                let recovered = self.wasNetworkOffline && satisfied
+                self.wasNetworkOffline = !satisfied
+                if recovered { self.retryConnection() }
+            }
+        }
+        monitor.start(queue: DispatchQueue(label: "net.buzzert.sybil.agent-network"))
+        self.monitor = monitor
+        connectionTask = Task { [weak self] in
+            guard let self else { return }
+            var delay = 1
+            while !Task.isCancelled, self.generation == run {
+                do {
+                    let descriptor = try await client.agentDescriptor()
+                    guard self.generation == run else { return }
+                    try self.receive(.activity(descriptor.id, descriptor.activity), replay: !self.hasLoaded)
+                    try await self.refreshRead(client: client, run: run)
+                    try await self.catchUp(client: client, run: run, replay: !self.hasLoaded)
+                    guard self.generation == run, !Task.isCancelled else { return }
+                    self.hasLoaded = true; self.isLoading = false; self.flush()
+                    let socket = try await client.agentSocket(after: self.pending.cursor)
+                    guard self.generation == run, !Task.isCancelled else { await socket.close(); return }
+                    self.socket = socket
+                    self.socketStarted = Date(); self.lastPacket = Date()
+                    while !Task.isCancelled, self.generation == run {
+                        let data = try await socket.receive()
+                        guard self.generation == run, !Task.isCancelled else { return }
+                        self.lastPacket = Date()
+                        guard let event = try AgentEvent.decode(data) else { continue }
+                        if case .ready = event { self.connection = "Connected"; self.error = nil; delay = 1 }
+                        try self.receive(event)
+                    }
+                } catch {
+                    guard self.generation == run, !Task.isCancelled else { return }
+                    let failedSocket = self.socket
+                    let code = await failedSocket?.closeCode()
+                    await failedSocket?.close()
+                    guard self.generation == run, !Task.isCancelled else { return }
+                    self.socket = nil
+                    self.isLoading = false
+                    self.error = code == 4401 ? "Agent authentication failed. Check your Sybil token in Settings." : error.localizedDescription
+                    self.connection = code == 4401 ? "Offline" : "Reconnecting"
+                    if code == 4401 { return }
+                }
+                try? await Task.sleep(for: .seconds(delay))
+                delay = min(delay * 2, 30)
+            }
+        }
+        maintenanceTask = Task { [weak self] in
+            while !Task.isCancelled {
+                try? await Task.sleep(for: .seconds(15))
+                guard let self, self.generation == run, !Task.isCancelled else { return }
+                do {
+                    try await self.refreshRead(client: client, run: run)
+                    guard self.generation == run, !Task.isCancelled else { return }
+                    self.flushRead()
+                    if self.connection != "Connected" {
+                        try await self.catchUp(client: client, run: run, replay: !self.hasLoaded)
+                    }
+                    guard self.generation == run, !Task.isCancelled else { return }
+                    if let socket = self.socket {
+                        if Date().timeIntervalSince(self.lastPacket) > 60 || (self.connection != "Connected" && Date().timeIntervalSince(self.socketStarted) > 12) {
+                            await socket.close()
+                        } else { try await socket.send("{\"type\":\"ping\"}") }
+                    }
+                } catch {
+                    if self.generation == run { self.error = error.localizedDescription }
+                }
+            }
+        }
+    }
+
+    func catchUp(client: any AgentAPIClienting, run: UUID? = nil, replay: Bool = false) async throws {
+        let run = run ?? generation
+        guard generation == run, !Task.isCancelled, !syncing else { return }
+        syncing = true
+        defer { if generation == run { syncing = false } }
+        var cursor = pending.cursor
+        while !Task.isCancelled, generation == run {
+            let page = try await client.agentMessages(after: cursor, before: nil)
+            guard generation == run, !Task.isCancelled else { return }
+            try page.validate()
+            let next = try AgentCursor.normalize(page.cursor ?? cursor)
+            guard !AgentCursor.less(next, cursor), !page.hasMore || next != cursor else {
+                throw AgentFailure(message: "Agent history did not advance its cursor.")
+            }
+            try receive(.messages(page), replay: replay)
+            cursor = next
+            if !page.hasMore { return }
+        }
+    }
+
+    func receive(_ event: AgentEvent, replay: Bool = false) throws {
+        let before = pending
+        var next = pending
+        switch event {
+        case let .messages(page): try next.apply(page)
+        case let .activity(id, activity):
+            try next.checkConversation(id)
+            next.activity = activity
+        case let .ready(id): if let id { try next.checkConversation(id) }
+        case let .error(message): throw AgentFailure(message: message)
+        case .heartbeat: return
+        }
+        presenter.receive(event, before: before, after: next, replay: replay, now: Date())
+        pending = next
+        if !replay {
+            let oldIDs = Set(before.messages.map(\.id)).union(outbox.map(\.id))
+            pendingEntrances.formUnion(next.messages.filter { !oldIDs.contains($0.id) && $0.role != .tool }.map(\.id))
+        }
+        reconcileOutbox()
+        schedulePublish()
+    }
+
+    private func schedulePublish() {
+        guard batchTask == nil else { return }
+        batchTask = Task { [weak self] in
+            try? await Task.sleep(for: .milliseconds(100))
+            guard !Task.isCancelled else { return }
+            self?.flush()
+        }
+    }
+
+    func flush() {
+        batchTask?.cancel(); batchTask = nil
+        typingTask?.cancel(); typingTask = nil
+        snapshot = Snapshot(room: pending, isTyping: presenter.isTyping(now: Date()), entranceIDs: pendingEntrances, revision: snapshot.revision + 1)
+        pendingEntrances = []
+        if let start = presenter.busySince, !snapshot.isTyping {
+            let delay = max(0.01, 0.25 - Date().timeIntervalSince(start))
+            typingTask = Task { [weak self] in
+                try? await Task.sleep(for: .seconds(delay))
+                guard !Task.isCancelled else { return }
+                self?.flush()
+            }
+        }
+    }
+
+    private func refreshRead(client: any AgentAPIClienting, run: UUID) async throws {
+        let read = try await client.agentReadState()
+        guard generation == run else { return }
+        try applyRead(read)
+    }
+    func applyRead(_ read: AgentReadState) throws {
+        acknowledgedRead = try AgentCursor.advance(acknowledgedRead, read.cursor)
+        pending.readCursor = acknowledgedRead
+        schedulePublish()
+    }
+    func markRead(renderedCursor: String) {
+        guard hasLoaded, isActive, isSelected, !isObscured, !isModalPresented else { return }
+        do { desiredRead = try AgentCursor.advance(desiredRead, renderedCursor); flushRead() }
+        catch { self.error = error.localizedDescription }
+    }
+    private func flushRead() {
+        guard let client, readTask == nil, AgentCursor.less(acknowledgedRead, desiredRead) else { return }
+        let run = generation
+        readTask = Task { [weak self] in
+            guard let self else { return }
+            defer { if self.generation == run { self.readTask = nil } }
+            do {
+                while self.generation == run, AgentCursor.less(self.acknowledgedRead, self.desiredRead), !Task.isCancelled {
+                    let captured = self.desiredRead
+                    let read = try await client.agentMarkRead(cursor: captured)
+                    guard self.generation == run, !Task.isCancelled else { return }
+                    guard !AgentCursor.less(try AgentCursor.normalize(read.cursor), captured) else {
+                        throw AgentFailure(message: "Agent read state was not acknowledged.")
+                    }
+                    try self.applyRead(read)
+                }
+            } catch { if self.generation == run { self.error = error.localizedDescription } }
+        }
+    }
+
+    @discardableResult
+    func send(_ content: String) -> Bool {
+        guard !content.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty, content.utf16.count <= 40_000 else {
+            error = "Enter a message of at most 40,000 characters."; return false
+        }
+        guard storageReadable, client != nil else { error = "Pending messages could not be restored. Reconnect before sending."; return false }
+        let row = AgentOutboxRow(requestID: UUID().uuidString, content: content, createdAt: Date())
+        outbox.append(row)
+        pendingEntrances.insert(row.id)
+        presenter.sent(now: Date(), activity: pending.activity)
+        flush()
+        dispatch(row.id)
+        return true
+    }
+    func retry(_ id: String) { dispatch(id) }
+    private func persist() throws {
+        guard let storage, storageReadable else { throw AgentFailure(message: "Message storage is unavailable.") }
+        try storage.write(outbox.filter { !$0.confirmed })
+    }
+    private func dispatch(_ id: String) {
+        guard let client, sends[id] == nil, let index = outbox.firstIndex(where: { $0.id == id }), !outbox[index].seenOnServer else { return }
+        outbox[index].status = "sending"; outbox[index].error = nil
+        do { try persist() }
+        catch { outbox[index].status = "failed"; outbox[index].error = "Could not save message: \(error.localizedDescription)"; return }
+        let row = outbox[index], run = generation
+        sends[id] = Task { [weak self] in
+            guard let self else { return }
+            defer { if self.generation == run { self.sends[id] = nil } }
+            do {
+                let receipt = try await client.agentSubmit(content: row.content, requestID: row.requestID)
+                try receipt.validate(requestID: row.requestID)
+                guard self.generation == run, let index = self.outbox.firstIndex(where: { $0.id == id }), !self.outbox[index].confirmed else { return }
+                self.outbox[index].seenOnServer = true
+                self.outbox[index].status = receipt.status == "queued" ? "queued" : "complete"
+                self.outbox[index].error = nil
+                try? self.persist()
+            } catch {
+                guard self.generation == run, let index = self.outbox.firstIndex(where: { $0.id == id }), !self.outbox[index].seenOnServer else { return }
+                self.outbox[index].status = "failed"; self.outbox[index].error = error.localizedDescription
+                try? self.persist()
+            }
+        }
+    }
+    private func reconcileOutbox() {
+        let committed = Set(pending.messages.map(\.id))
+        let queued = Set(pending.activity.queuedMessages.map(\.id))
+        var changed = false
+        for index in outbox.indices {
+            if committed.contains(outbox[index].id) {
+                if !outbox[index].confirmed { changed = true }
+                outbox[index].confirmed = true; outbox[index].seenOnServer = true
+                outbox[index].status = "complete"; outbox[index].error = nil
+            } else if queued.contains(outbox[index].id), !outbox[index].confirmed {
+                changed = true
+                outbox[index].seenOnServer = true; outbox[index].status = "queued"; outbox[index].error = nil
+            }
+        }
+        if changed { try? persist() }
+    }
+}

ios/Packages/Sybil/Sources/Sybil/SybilAgentTransport.swift

@@ -0,0 +1,56 @@
+import Foundation
+
+protocol AgentSocket: Sendable {
+    func receive() async throws -> Data
+    func send(_ text: String) async throws
+    func close() async
+    func closeCode() async -> Int
+}
+
+protocol AgentAPIClienting: Sendable {
+    func agentDescriptor() async throws -> AgentDescriptor
+    func agentMessages(after: String?, before: String?) async throws -> AgentPage
+    func agentReadState() async throws -> AgentReadState
+    func agentMarkRead(cursor: String) async throws -> AgentReadState
+    func agentSubmit(content: String, requestID: String) async throws -> AgentReceipt
+    func agentSocket(after: String) async throws -> any AgentSocket
+}
+
+actor AgentURLSocket: AgentSocket {
+    private let task: URLSessionWebSocketTask
+    init(task: URLSessionWebSocketTask) { self.task = task }
+    func receive() async throws -> Data {
+        switch try await task.receive() {
+        case let .string(text): return Data(text.utf8)
+        case .data: throw AgentFailure(message: "Agent sent a non-JSON text frame.")
+        @unknown default: throw AgentFailure(message: "Unsupported Agent frame.")
+        }
+    }
+    func send(_ text: String) async throws { try await task.send(.string(text)) }
+    func close() { task.cancel(with: .goingAway, reason: nil) }
+    func closeCode() -> Int { task.closeCode.rawValue }
+}
+
+extension APIConfiguration {
+    func agentWebSocketURL() throws -> URL {
+        guard var parts = URLComponents(url: baseURL, resolvingAgainstBaseURL: false),
+              ["http", "https"].contains(parts.scheme), parts.host != nil,
+              parts.user == nil, parts.password == nil else { throw APIError.invalidBaseURL }
+        parts.scheme = parts.scheme == "https" ? "wss" : "ws"
+        parts.path = parts.path.trimmingCharacters(in: CharacterSet(charactersIn: "/")).isEmpty
+            ? "/v1/agent/ws" : "/" + parts.path.trimmingCharacters(in: CharacterSet(charactersIn: "/")) + "/v1/agent/ws"
+        parts.query = nil
+        parts.fragment = nil
+        guard let url = parts.url else { throw APIError.invalidBaseURL }
+        return url
+    }
+    func agentAuthFrame(after: String) throws -> String {
+        let token = authToken?.trimmingCharacters(in: .whitespacesAndNewlines)
+        let value: JSONValue = .object([
+            "type": .string("auth"),
+            "token": token.flatMap { $0.isEmpty ? nil : JSONValue.string($0) } ?? .null,
+            "after": .string(try AgentCursor.normalize(after))
+        ])
+        return String(decoding: try JSONEncoder().encode(value), as: UTF8.self)
+    }
+}

ios/Packages/Sybil/Sources/Sybil/SybilAgentView.swift

@@ -0,0 +1,360 @@
+import SwiftUI
+
+struct SybilAgentView: View {
+    @Bindable var room: SybilAgentRoom
+    var topInset: CGFloat = 0
+    @Environment(\.scenePhase) private var scenePhase
+    @Environment(\.accessibilityReduceMotion) private var reduceMotion
+    @FocusState private var composerFocused: Bool
+    @State private var atLatest = true
+    @State private var isScrolling = false
+    @State private var historyLimit = 100
+    @State private var historyStartID: String?
+    @State private var scrollPosition = ScrollPosition()
+    @State private var scrollMetrics = AgentScrollMetrics()
+    @State private var rowFrames: [String: CGRect] = [:]
+    @State private var restoreAnchor: AgentReadingAnchor?
+    @State private var appeared = false
+    private let bottomID = "agent-bottom"
+
+    private var readKey: String {
+        "\(room.snapshot.room.cursor):\(atLatest):\(scrollMetrics.atBottom):\(isScrolling):\(room.isObscured):\(room.isModalPresented):\(scenePhase == .active):\(room.isLoading):\(room.snapshot.revision)"
+    }
+    var body: some View {
+        let rendered = room.snapshot
+        let allMessages = room.messages
+        let startIndex = historyStartID.flatMap { id in allMessages.firstIndex(where: { $0.id == id }) } ?? max(0, allMessages.count - historyLimit)
+        let displayed = Array(allMessages.dropFirst(startIndex))
+        VStack(spacing: 0) {
+            ScrollViewReader { proxy in
+                ScrollView {
+                    VStack(alignment: .leading, spacing: 15) {
+                        if allMessages.count > displayed.count {
+                            Button("Load earlier messages") {
+                                if let anchor = rowFrames.filter({ $0.value.maxY > 0 }).min(by: { $0.value.minY < $1.value.minY }) {
+                                    restoreAnchor = AgentReadingAnchor(id: anchor.key, offset: anchor.value.minY, expectedCount: min(allMessages.count, displayed.count + 100))
+                                }
+                                atLatest = false
+                                historyLimit += 100
+                                historyStartID = allMessages[max(0, startIndex - 100)].id
+                            }
+                            .frame(maxWidth: .infinity)
+                            .padding(.vertical, 10)
+                        }
+                        if room.isLoading { ProgressView("Loading messages…").frame(maxWidth: .infinity) }
+                        ForEach(Array(displayed.enumerated()), id: \.element.id) { index, message in
+                            VStack(spacing: 14) {
+                                if index == 0 || !Calendar.current.isDate(message.createdAt, inSameDayAs: displayed[index - 1].createdAt) {
+                                    Text(message.createdAt.formatted(date: .abbreviated, time: .omitted))
+                                        .font(.sybil(.caption2))
+                                        .foregroundStyle(SybilTheme.textMuted)
+                                        .frame(maxWidth: .infinity)
+                                        .padding(.vertical, 6)
+                                }
+                                Group {
+                                    if message.isToolCallLog {
+                                        AgentToolCard(message: message)
+                                    } else {
+                                        AgentMessageRow(message: message, retry: { room.retry(message.id) })
+                                    }
+                                }
+                                .background(GeometryReader { geometry in
+                                    Color.clear.preference(key: AgentRowFrames.self, value: [message.id: geometry.frame(in: .named("agent-scroll"))])
+                                })
+                            }
+                            .modifier(AgentEntrance(animate: appeared && rendered.entranceIDs.contains(message.id), reduceMotion: reduceMotion))
+                            .id(message.id)
+                        }
+                        if rendered.isTyping {
+                            HStack(spacing: 9) {
+                                AgentAvatar()
+                                ProgressView().controlSize(.small)
+                                Text("Sybil is thinking…").font(.sybil(.caption))
+                            }
+                            .foregroundStyle(SybilTheme.textMuted)
+                            .accessibilityIdentifier("agent-typing")
+                        }
+                        Color.clear.frame(height: 1).id(bottomID)
+                    }
+                    .scrollTargetLayout()
+                    .padding(.horizontal, 16)
+                    .padding(.vertical, 16)
+                    .frame(maxWidth: SybilLayout.webContentMaxWidth)
+                    .frame(maxWidth: .infinity)
+                }
+                .coordinateSpace(name: "agent-scroll")
+                .scrollPosition($scrollPosition)
+                .defaultScrollAnchor(.bottom, for: .initialOffset)
+                .defaultScrollAnchor(.bottom, for: .alignment)
+                .scrollDismissesKeyboard(.interactively)
+                .onScrollPhaseChange { _, phase in
+                    isScrolling = phase.isScrolling
+                    if phase == .interacting || phase == .tracking { restoreAnchor = nil }
+                }
+                .onScrollGeometryChange(for: AgentScrollMetrics.self) { geometry in
+                    AgentScrollMetrics(offset: geometry.contentOffset.y,
+                        height: geometry.containerSize.height,
+                        contentHeight: geometry.contentSize.height,
+                        bottomInset: geometry.contentInsets.bottom, topInset: geometry.contentInsets.top)
+                } action: { previous, next in
+                    scrollMetrics = next
+                    if isScrolling || next.contentHeight <= next.height { atLatest = next.atBottom }
+                    if atLatest, !isScrolling, restoreAnchor == nil,
+                       previous.height != next.height || previous.contentHeight != next.contentHeight || previous.bottomInset != next.bottomInset {
+                        proxy.scrollTo(bottomID, anchor: .bottom)
+                    }
+                }
+                .onPreferenceChange(AgentRowFrames.self) { frames in
+                    let previousAnchor = rowFrames.filter { $0.value.maxY > 0 && $0.value.minY < scrollMetrics.height }
+                        .min { $0.value.minY < $1.value.minY }
+                        .map { AgentReadingAnchor(id: $0.key, offset: $0.value.minY) }
+                    let anchor = restoreAnchor ?? (!atLatest && !isScrolling ? previousAnchor : nil)
+                    if let anchor, let frame = frames[anchor.id], abs(frame.minY - anchor.offset) > 1 {
+                        // Restore the exact pixel offset after insertion, tool
+                        // expansion or text reflow. Normalize the saved frames to
+                        // the requested offset so the next layout cannot undo it.
+                        let requested = scrollMetrics.offset + frame.minY - anchor.offset
+                        let target = min(max(requested, -scrollMetrics.topInset), max(-scrollMetrics.topInset, scrollMetrics.contentHeight - scrollMetrics.height + scrollMetrics.bottomInset))
+                        let appliedShift = target - scrollMetrics.offset
+                        rowFrames = frames.mapValues { $0.offsetBy(dx: 0, dy: -appliedShift) }
+                        if abs(appliedShift) > 1 { scrollPosition.scrollTo(y: target) }
+                        restoreAnchor = nil
+                    } else {
+                        rowFrames = frames
+                        if let expected = restoreAnchor?.expectedCount, frames.count >= expected { restoreAnchor = nil }
+                    }
+                }
+                .overlay(alignment: .bottomTrailing) {
+                    if !atLatest {
+                        Button {
+                            atLatest = true
+                            withAnimation(reduceMotion ? nil : .easeOut(duration: 0.24)) { proxy.scrollTo(bottomID, anchor: .bottom) }
+                        } label: {
+                            Label(room.unreadCount > 0 ? "\(room.unreadCount) unread" : "Latest", systemImage: "arrow.down")
+                                .font(.sybil(.caption, weight: .semibold))
+                                .padding(12)
+                                .background(SybilTheme.surfaceStrong, in: Capsule())
+                                .overlay(Capsule().stroke(SybilTheme.border))
+                        }
+                        .padding(14)
+                        .accessibilityIdentifier("agent-jump-latest")
+                    }
+                }
+                .onAppear {
+                    appeared = true
+                    if !room.isLoading { historyStartID = displayed.first?.id }
+                }
+                .onChange(of: room.isLoading) { _, loading in
+                    if !loading && historyStartID == nil { historyStartID = displayed.first?.id }
+                }
+                .onChange(of: allMessages.count) { _, _ in
+                    if atLatest { proxy.scrollTo(bottomID, anchor: .bottom) }
+                }
+                .onChange(of: composerFocused) { _, _ in
+                    if atLatest { proxy.scrollTo(bottomID, anchor: .bottom) }
+                }
+            }
+            statusBar
+            composer
+        }
+        .padding(.top, topInset)
+        .onAppear { room.isSelected = true }
+        .onDisappear { room.isSelected = false }
+        .task(id: readKey) {
+            // This cursor belongs to the rendered snapshot. New network input cancels
+            // the task, and cannot be acknowledged while still entering the layout.
+            let cursor = rendered.room.cursor
+            guard atLatest, scrollMetrics.atBottom, !isScrolling, !room.isObscured, !room.isModalPresented, scenePhase == .active, !room.isLoading else { return }
+            try? await Task.sleep(for: .milliseconds(reduceMotion ? 80 : 360))
+            guard !Task.isCancelled, atLatest, scrollMetrics.atBottom, !isScrolling, !room.isObscured, !room.isModalPresented, scenePhase == .active else { return }
+            room.markRead(renderedCursor: cursor)
+        }
+    }
+
+    @ViewBuilder private var statusBar: some View {
+        let activity = room.snapshot.room.activity
+        if let error = room.error ?? activity.error {
+            HStack(alignment: .top) {
+                Text(error).font(.sybil(.caption)).foregroundStyle(SybilTheme.danger)
+                Spacer(minLength: 4)
+                Button("Reconnect") { room.retryConnection() }.font(.sybil(.caption, weight: .semibold))
+            }.padding(.horizontal, 16).padding(.vertical, 8)
+        } else if activity.queued > 0 || activity.backgroundWorkers > 0 || activity.wait != nil {
+            HStack(spacing: 10) {
+                if activity.queued > 0 { Label("\(activity.queued) queued", systemImage: "clock") }
+                if activity.backgroundWorkers > 0 { Label("\(activity.backgroundWorkers) working", systemImage: "gearshape.2") }
+                if let wait = activity.wait {
+                    Text(wait.reason ?? "Waiting")
+                    Text(Date(timeIntervalSince1970: wait.due / 1000), style: .time)
+                }
+                Spacer(minLength: 0)
+            }
+            .font(.sybil(.caption2)).foregroundStyle(SybilTheme.textMuted)
+            .padding(.horizontal, 16).padding(.vertical, 6)
+        }
+    }
+
+    private var canSendComposer: Bool {
+        !room.composer.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty && room.composer.utf16.count <= 40_000
+    }
+
+    private var composer: some View {
+        HStack(alignment: .bottom, spacing: 10) {
+            SybilPromptField(
+                text: $room.composer,
+                focused: $composerFocused,
+                submit: submitComposer
+            )
+            .accessibilityIdentifier("agent-composer")
+
+            SybilPromptSendButton(canSend: canSendComposer, submit: submitComposer)
+            .accessibilityIdentifier("agent-send")
+        }
+        .padding(.horizontal, 14)
+        .padding(.bottom, 12)
+        .padding(.top, 6)
+    }
+
+    private func submitComposer() {
+        guard canSendComposer, room.send(room.composer) else { return }
+        room.composer = ""
+        atLatest = true
+    }
+}
+
+struct AgentAvatar: View {
+    var body: some View {
+        Image("character-still", bundle: .module).resizable().scaledToFill()
+            .frame(width: 27, height: 27).clipShape(RoundedRectangle(cornerRadius: 9))
+            .overlay(RoundedRectangle(cornerRadius: 9).stroke(SybilTheme.primary.opacity(0.3)))
+            .accessibilityHidden(true)
+    }
+}
+
+private struct AgentMessageRow: View {
+    var message: Message
+    var retry: () -> Void
+    private var own: Bool { message.role == .user }
+    private var status: String { message.metadata?.objectValue?["status"]?.stringValue ?? "complete" }
+    var body: some View {
+        HStack(alignment: .bottom, spacing: 8) {
+            if !own { AgentAvatar().padding(.bottom, 23) }
+            VStack(alignment: own ? .trailing : .leading, spacing: 5) {
+                MessageBubble(message: message, isSending: false, streamingAssistantPresentation: nil, isForking: false, onFork: nil, agentStyle: true)
+                HStack(spacing: 6) {
+                    Text(own ? "You" : "Sybil")
+                    Text(message.createdAt, style: .time)
+                    if own && status == "sending" { Text("Sending…") }
+                    if own && status == "queued" { Label("Queued", systemImage: "clock") }
+                    if own && status == "failed" {
+                        Text("Not confirmed").foregroundStyle(SybilTheme.danger)
+                        Button("Retry", action: retry).accessibilityIdentifier("agent-retry")
+                    }
+                }
+                .font(.sybil(.caption2)).foregroundStyle(SybilTheme.textMuted)
+                if own && status == "failed", let error = message.metadata?.objectValue?["error"]?.stringValue {
+                    Text(error).font(.sybil(.caption2)).foregroundStyle(SybilTheme.danger)
+                }
+            }
+        }
+    }
+}
+
+struct AgentToolCard: View {
+    var message: Message
+    @State private var expanded = false
+    var body: some View {
+        if let metadata = message.toolCallMetadata {
+            VStack(alignment: .leading, spacing: 0) {
+                Button {
+                    expanded.toggle()
+                } label: {
+                    HStack(spacing: 8) {
+                        ToolCallActivityChip(metadata: metadata, fallbackContent: message.content, createdAt: message.createdAt, compactLayout: true)
+                            .allowsHitTesting(false)
+                        Spacer(minLength: 0)
+                        Image(systemName: expanded ? "chevron.down" : "chevron.right")
+                            .font(.system(size: 13, weight: .semibold))
+                    }
+                    .contentShape(Rectangle())
+                }
+                .buttonStyle(.plain)
+                .accessibilityIdentifier("agent-tool-\(metadata.toolName ?? "tool")")
+                .accessibilityValue(expanded ? "Expanded" : "Collapsed")
+                if expanded {
+                    VStack(alignment: .leading, spacing: 12) {
+                        if let args = metadata.args { detail("Input", text: pretty(args)) }
+                        if let error = metadata.error { detail("Error", text: error) }
+                        if let result = metadata.result { detail("Result", text: pretty(result)) }
+                        else if let preview = metadata.resultPreview { detail("Result", text: preview) }
+                        if metadata.args == nil && metadata.result == nil && metadata.resultPreview == nil && metadata.error == nil {
+                            Text(message.content).font(.sybil(.caption))
+                        }
+                    }.padding(.top, 8).frame(maxWidth: .infinity, alignment: .leading)
+                }
+            }
+            .foregroundStyle(SybilTheme.textMuted)
+            .padding(8)
+            .background(SybilTheme.surface.opacity(0.8), in: RoundedRectangle(cornerRadius: 14))
+        }
+    }
+    private func detail(_ title: String, text: String) -> some View {
+        VStack(alignment: .leading, spacing: 4) {
+            Text(title).font(.sybil(.caption, weight: .semibold)).accessibilityIdentifier("agent-tool-\(title.lowercased())")
+            Text(text).font(.system(size: 12, design: .monospaced)).textSelection(.enabled)
+        }
+    }
+    private func pretty(_ value: JSONValue) -> String {
+        if let text = value.stringValue { return text }
+        let encoder = JSONEncoder(); encoder.outputFormatting = [.prettyPrinted, .sortedKeys]
+        return (try? encoder.encode(value)).map { String(decoding: $0, as: UTF8.self) } ?? ""
+    }
+}
+
+// Animate both occupied height and content, so incoming rows push older history up.
+private struct AgentGrowingLayout: Layout {
+    var progress: CGFloat
+    var animatableData: CGFloat { get { progress } set { progress = newValue } }
+    func sizeThatFits(proposal: ProposedViewSize, subviews: Subviews, cache: inout ()) -> CGSize {
+        let size = subviews.first?.sizeThatFits(ProposedViewSize(width: proposal.width, height: nil)) ?? .zero
+        return CGSize(width: proposal.width ?? size.width, height: size.height * progress)
+    }
+    func placeSubviews(in bounds: CGRect, proposal: ProposedViewSize, subviews: Subviews, cache: inout ()) {
+        subviews.first?.place(at: bounds.origin, proposal: ProposedViewSize(width: bounds.width, height: nil))
+    }
+}
+private struct AgentEntrance: ViewModifier {
+    var animate: Bool
+    var reduceMotion: Bool
+    @State private var entered = false
+    func body(content: Content) -> some View {
+        let preparing = animate && !entered && !reduceMotion
+        AgentGrowingLayout(progress: preparing ? 0 : 1) {
+            content.opacity(preparing ? 0 : 1).offset(y: preparing ? 16 : 0)
+        }
+        .clipped()
+        .onAppear {
+            withAnimation(animate && !reduceMotion ? .easeOut(duration: 0.24) : nil) { entered = true }
+        }
+    }
+}
+
+private struct AgentScrollMetrics: Equatable {
+    var offset: CGFloat = 0
+    var height: CGFloat = 0
+    var contentHeight: CGFloat = 0
+    var bottomInset: CGFloat = 0
+    var topInset: CGFloat = 0
+    var atBottom: Bool { height > 0 && contentHeight - offset - height + bottomInset <= 24 }
+}
+private struct AgentReadingAnchor {
+    var id: String
+    var offset: CGFloat
+    var expectedCount: Int? = nil
+}
+private struct AgentRowFrames: PreferenceKey {
+    static let defaultValue: [String: CGRect] = [:]
+    static func reduce(value: inout [String: CGRect], nextValue: () -> [String: CGRect]) {
+        value.merge(nextValue(), uniquingKeysWith: { _, next in next })
+    }
+}

ios/Packages/Sybil/Sources/Sybil/SybilChatTranscriptView.swift

@@ -1,13 +1,17 @@
 import MarkdownUI
 import SwiftUI
+import UIKit
 
 struct SybilChatTranscriptView: View {
     var messages: [Message]
     var isLoading: Bool
     var isSending: Bool
+    var streamingAssistantPresentation: StreamingAssistantPresentation?
     var topContentInset: CGFloat = 0
     var bottomContentInset: CGFloat = 0
     var bottomPinRequestID: Int = 0
+    var isForking: Bool = false
+    var onForkResponse: ((String) -> Void)? = nil
 
     @State private var hasTrackedToolCallMessages = false
     @State private var knownToolCallMessageIDs: Set<String> = []
@@ -41,7 +45,15 @@ struct SybilChatTranscriptView: View {
                     ForEach(renderItems) { item in
                         switch item {
                         case let .message(message):
-                            MessageBubble(message: message, isSending: isSending)
+                            MessageBubble(
+                                message: message,
+                                isSending: isSending,
+                                streamingAssistantPresentation: streamingAssistantPresentation,
+                                isForking: isForking,
+                                onFork: {
+                                    onForkResponse?(message.id)
+                                }
+                            )
                                 .frame(maxWidth: .infinity)
                         case let .toolGroup(id, messages):
                             ToolCallStackView(
@@ -137,9 +149,13 @@ func buildTranscriptRenderItems(from messages: [Message]) -> [TranscriptRenderIt
     return items
 }
 
-private struct MessageBubble: View {
+struct MessageBubble: View {
     var message: Message
     var isSending: Bool
+    var streamingAssistantPresentation: StreamingAssistantPresentation?
+    var isForking: Bool
+    var onFork: (() -> Void)?
+    var agentStyle = false
 
     private var toolCallMetadata: ToolCallMetadata? {
         message.toolCallMetadata
@@ -149,13 +165,46 @@ private struct MessageBubble: View {
         message.role == .user
     }
 
-    private var isPendingAssistant: Bool {
-        message.id.hasPrefix("temp-assistant-") &&
-            isSending &&
-            message.content.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+    private var isStreamingAssistant: Bool {
+        message.role == .assistant && message.id.hasPrefix("temp-assistant-")
+    }
+
+    private var matchingStreamingAssistantPresentation: StreamingAssistantPresentation? {
+        guard streamingAssistantPresentation?.messageID == message.id else { return nil }
+        return streamingAssistantPresentation
+    }
+
+    private var canFork: Bool {
+        message.role == .assistant &&
+            !message.id.hasPrefix("temp-") &&
+            onFork != nil
     }
 
     var body: some View {
+        Group {
+            if canFork {
+                bubbleContent
+                    .contextMenu {
+                        Button {
+                            UIPasteboard.general.string = message.content
+                        } label: {
+                            Label("Copy", systemImage: "doc.on.doc")
+                        }
+
+                        Button {
+                            onFork?()
+                        } label: {
+                            Label("Fork", systemImage: "arrow.triangle.branch")
+                        }
+                        .disabled(isForking)
+                    }
+            } else {
+                bubbleContent
+            }
+        }
+    }
+
+    private var bubbleContent: some View {
         HStack(alignment: .top, spacing: 0) {
             leadingSpacer
 
@@ -174,25 +223,26 @@ private struct MessageBubble: View {
                         )
                     }
 
-                    if isPendingAssistant {
-                        HStack(spacing: 8) {
-                            ProgressView()
-                                .controlSize(.small)
-                                .tint(SybilTheme.primary)
-                            Text("Thinking…")
-                                .font(.sybil(.footnote))
-                                .foregroundStyle(SybilTheme.textMuted)
-                        }
-                        .padding(.vertical, 2)
+                    if isStreamingAssistant {
+                        StreamingAssistantContentView(
+                            content: message.content,
+                            isSending: isSending,
+                            presentation: matchingStreamingAssistantPresentation
+                        )
                     } else if !message.content.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty {
-                        Markdown(message.content)
-                            .tint(SybilTheme.primary)
-                            .foregroundStyle(isUser ? SybilTheme.text : SybilTheme.text.opacity(0.95))
-                            .markdownTheme(.sybilReadable)
+                        if isUser {
+                            Text(message.content)
+                                .font(.sybil(.body))
+                        } else {
+                            Markdown(message.content)
+                                .tint(SybilTheme.primary)
+                                .foregroundStyle(SybilTheme.text.opacity(0.95))
+                                .markdownTheme(.sybilReadable)
+                        }
                     }
                 }
-                .padding(.horizontal, isUser ? 14 : 2)
-                .padding(.vertical, isUser ? 13 : 2)
+                .padding(.horizontal, isUser || agentStyle ? 14 : 2)
+                .padding(.vertical, isUser || agentStyle ? 13 : 2)
                 .textSelection(.enabled)
                 .background(
                     Group {
@@ -203,6 +253,10 @@ private struct MessageBubble: View {
                                     RoundedRectangle(cornerRadius: 16)
                                         .stroke(SybilTheme.primary.opacity(0.45), lineWidth: 1)
                                 )
+                        } else if agentStyle {
+                            RoundedRectangle(cornerRadius: 16)
+                                .fill(Color(red: 0.12, green: 0.075, blue: 0.21))
+                                .overlay(RoundedRectangle(cornerRadius: 16).stroke(SybilTheme.primary.opacity(0.18), lineWidth: 1))
                         } else {
                             RoundedRectangle(cornerRadius: 0)
                                 .fill(Color.clear)
@@ -232,6 +286,123 @@ private struct MessageBubble: View {
     }
 }
 
+private struct StreamingTraceHeightPreferenceKey: PreferenceKey {
+    static let defaultValue: CGFloat = 0
+
+    static func reduce(value: inout CGFloat, nextValue: () -> CGFloat) {
+        value = max(value, nextValue())
+    }
+}
+
+struct StreamingAssistantContentView: View {
+    private enum Phase: Hashable {
+        case pending
+        case trace(Int)
+        case answer
+    }
+
+    var content: String
+    var isSending: Bool
+    var presentation: StreamingAssistantPresentation?
+
+    @Environment(\.accessibilityReduceMotion) private var reduceMotion
+    @State private var retainedTraceHeight: CGFloat = 0
+
+    private var traceText: String? {
+        guard let trace = presentation?.ephemeralTrace?.trimmingCharacters(in: .whitespacesAndNewlines),
+              !trace.isEmpty
+        else {
+            return nil
+        }
+        return trace
+    }
+
+    private var hasAnswer: Bool {
+        !content.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+    }
+
+    private var phase: Phase {
+        if hasAnswer {
+            return .answer
+        }
+        if traceText != nil {
+            return .trace(presentation?.traceRevision ?? 0)
+        }
+        return .pending
+    }
+
+    private var traceTransition: AnyTransition {
+        guard !reduceMotion else { return .identity }
+        return .asymmetric(
+            insertion: .opacity.combined(with: .offset(y: 5)),
+            removal: .opacity.combined(with: .offset(y: -5))
+        )
+    }
+
+    private var answerTransition: AnyTransition {
+        guard !reduceMotion else { return .identity }
+        return .opacity.combined(with: .offset(y: 3))
+    }
+
+    var body: some View {
+        ZStack(alignment: .topLeading) {
+            if let traceText {
+                Text(traceText)
+                    .font(.sybil(size: 15))
+                    .italic()
+                    .foregroundStyle(SybilTheme.textMuted.opacity(0.88))
+                    .lineSpacing(5)
+                    .fixedSize(horizontal: false, vertical: true)
+                    .opacity(0.78)
+                    .textSelection(.disabled)
+                    .allowsHitTesting(false)
+                    .accessibilityLabel("Assistant is working")
+                    .background {
+                        GeometryReader { geometry in
+                            Color.clear.preference(
+                                key: StreamingTraceHeightPreferenceKey.self,
+                                value: geometry.size.height
+                            )
+                        }
+                    }
+                    .id("thinking-trace-\(presentation?.traceRevision ?? 0)")
+                    .transition(traceTransition)
+                    .zIndex(1)
+            }
+
+            if hasAnswer {
+                Markdown(content)
+                    .tint(SybilTheme.primary)
+                    .foregroundStyle(SybilTheme.text.opacity(0.95))
+                    .markdownTheme(.sybilReadable)
+                    .textSelection(.enabled)
+                    .id("streaming-assistant-answer")
+                    .transition(answerTransition)
+                    .zIndex(2)
+            }
+
+            if isSending && !hasAnswer && traceText == nil {
+                HStack(spacing: 8) {
+                    ProgressView()
+                        .controlSize(.small)
+                        .tint(SybilTheme.primary)
+                    Text("Thinking…")
+                        .font(.sybil(.footnote))
+                        .foregroundStyle(SybilTheme.textMuted)
+                }
+                .padding(.vertical, 2)
+                .transition(reduceMotion ? .identity : .opacity)
+            }
+        }
+        .frame(maxWidth: .infinity, minHeight: max(24, retainedTraceHeight), alignment: .topLeading)
+        .onPreferenceChange(StreamingTraceHeightPreferenceKey.self) { measuredHeight in
+            guard measuredHeight.isFinite, measuredHeight > retainedTraceHeight else { return }
+            retainedTraceHeight = measuredHeight
+        }
+        .animation(reduceMotion ? nil : .easeOut(duration: 0.20), value: phase)
+    }
+}
+
 private struct ToolCallStackView: View {
     private struct CardLayout {
         var x: CGFloat
@@ -440,7 +611,7 @@ private struct ToolCallStackCard: View {
     }
 }
 
-private struct ToolCallActivityChip: View {
+struct ToolCallActivityChip: View {
     enum VisualState {
         case initiated
         case completed

ios/Packages/Sybil/Sources/Sybil/SybilMacQuickQuestionBridge.swift

@@ -0,0 +1,71 @@
+import Foundation
+
+// This Foundation-only contract is compiled into both the Catalyst client and the
+// native macOS bundle. Only Objective-C-compatible values cross the bundle boundary.
+@MainActor
+@objc(SybilQuickQuestionPanelBridging)
+protocol SybilQuickQuestionPanelBridging: NSObjectProtocol {
+    init()
+    var isVisible: Bool { get }
+    func start(delegate: any SybilQuickQuestionPanelDelegate) -> Int32
+    func updateState(_ data: Data)
+    func show()
+    func hide()
+    func activateForQuickQuestion()
+    func activateMainWindow()
+    func stop()
+}
+
+@MainActor
+@objc(SybilQuickQuestionPanelDelegate)
+protocol SybilQuickQuestionPanelDelegate: NSObjectProtocol {
+    func quickQuestionToggleRequested()
+    func quickQuestionPromptChanged(_ prompt: String)
+    func quickQuestionSubmitRequested()
+    func quickQuestionProviderChanged(_ provider: String)
+    func quickQuestionModelChanged(_ model: String)
+    func quickQuestionConvertRequested()
+    func quickQuestionOpenAppRequested()
+    func quickQuestionPanelDismissed()
+}
+
+struct SybilQuickQuestionPanelState: Codable, Equatable, Sendable {
+    struct ProviderOption: Codable, Equatable, Sendable {
+        var id: String
+        var title: String
+    }
+
+    var prompt = ""
+    var answer = ""
+    var toolSummaries: [String] = []
+    var provider = ""
+    var providers: [ProviderOption] = []
+    var model = ""
+    var models: [String] = []
+    var isSending = false
+    var isConverting = false
+    var canSend = false
+    var canConvert = false
+    var isAuthenticated = false
+    var isCheckingSession = true
+    var error: String?
+
+    var isBusy: Bool { isSending || isConverting }
+    var hasResponse: Bool { !answer.isEmpty || !toolSummaries.isEmpty || isSending || error != nil }
+}
+
+// Carbon can deliver multiple presses while a key is held. Consume one toggle
+// per press/release pair, without suppressing a second deliberate press.
+struct SybilQuickQuestionHotKeyState {
+    private var isPressed = false
+
+    mutating func shouldToggle(isKeyDown: Bool) -> Bool {
+        if !isKeyDown {
+            isPressed = false
+            return false
+        }
+        guard !isPressed else { return false }
+        isPressed = true
+        return true
+    }
+}

ios/Packages/Sybil/Sources/Sybil/SybilMacQuickQuestionController.swift

@@ -0,0 +1,257 @@
+#if targetEnvironment(macCatalyst)
+import Combine
+import Foundation
+import Observation
+import UIKit
+
+@MainActor
+public final class SybilMacQuickQuestionController: NSObject, SybilQuickQuestionPanelDelegate {
+    public static let shared = SybilMacQuickQuestionController()
+
+    private var pluginBundle: Bundle?
+    private var bridge: (any SybilQuickQuestionPanelBridging)?
+    private var viewModel: SybilViewModel?
+    private var observationGeneration = 0
+    private var isPresentationRequested = false
+    private var activationRequested = false
+    private var presentationTask: Task<Void, Never>?
+    private var foregroundObservation: AnyCancellable?
+    private var notificationCenter = NotificationCenter.default
+    private var isApplicationInForeground: @MainActor () -> Bool = {
+        UIApplication.shared.applicationState != .background
+    }
+
+    private override init() {
+        super.init()
+    }
+
+    // Inject foreground readiness as well as the panel so tests can exercise
+    // windowless activation without activating apps or showing real windows.
+    init(
+        bridge: any SybilQuickQuestionPanelBridging,
+        isApplicationInForeground: @escaping @MainActor () -> Bool = { true },
+        notificationCenter: NotificationCenter = NotificationCenter()
+    ) {
+        self.bridge = bridge
+        self.isApplicationInForeground = isApplicationInForeground
+        self.notificationCenter = notificationCenter
+        super.init()
+        _ = bridge.start(delegate: self)
+        observeForegroundTransitionsIfNeeded()
+    }
+
+    public func start() {
+        observeForegroundTransitionsIfNeeded()
+        guard bridge == nil else { return }
+        guard let url = Bundle.main.builtInPlugInsURL?.appendingPathComponent("SybilMacQuickQuestion.bundle"),
+              let bundle = Bundle(url: url)
+        else {
+            SybilLog.error(SybilLog.app, "The macOS Quick Question bundle is missing.")
+            return
+        }
+
+        do {
+            try bundle.loadAndReturnError()
+            guard let pluginClass = bundle.principalClass as? any SybilQuickQuestionPanelBridging.Type else {
+                SybilLog.error(SybilLog.app, "The macOS Quick Question bundle has an invalid principal class.")
+                return
+            }
+            let plugin = pluginClass.init()
+            pluginBundle = bundle
+            bridge = plugin
+            let status = plugin.start(delegate: self)
+            if status != 0 {
+                SybilLog.warning(SybilLog.app, "Option+Space could not be registered (status \(status)). Quick Question remains available in the app menu.")
+            }
+        } catch {
+            SybilLog.error(SybilLog.app, "Could not load the macOS Quick Question panel", error: error)
+        }
+    }
+
+    func attach(_ viewModel: SybilViewModel) {
+        guard self.viewModel !== viewModel else { return }
+        // A window becoming active must not replace an in-flight quick question.
+        if self.viewModel != nil,
+           isPresentationRequested || bridge?.isVisible == true || self.viewModel?.isQuickQuestionSending == true || self.viewModel?.isConvertingQuickQuestion == true {
+            return
+        }
+        self.viewModel = viewModel
+        if bridge?.isVisible == true { updateAndObserve() }
+    }
+
+    public func toggle() {
+        start()
+        if isPresentationRequested || bridge?.isVisible == true {
+            cancelPresentation()
+            bridge?.hide()
+        } else {
+            show()
+        }
+    }
+
+    public func show() {
+        start()
+        isPresentationRequested = true
+        schedulePresentationIfNeeded()
+    }
+
+    public func stop() {
+        cancelPresentation()
+        foregroundObservation = nil
+        viewModel?.cancelQuickQuestion()
+        bridge?.stop()
+        bridge = nil
+        // Keep the loaded bundle alive for the lifetime of its code.
+    }
+
+    private func observeForegroundTransitionsIfNeeded() {
+        guard foregroundObservation == nil else { return }
+        foregroundObservation = notificationCenter.publisher(for: UIApplication.willEnterForegroundNotification)
+            .merge(with: notificationCenter.publisher(for: UIApplication.didBecomeActiveNotification))
+            .sink { [weak self] _ in
+                Task { @MainActor [weak self] in
+                    self?.schedulePresentationIfNeeded()
+                }
+            }
+    }
+
+    private func schedulePresentationIfNeeded() {
+        guard isPresentationRequested, bridge?.isVisible != true, presentationTask == nil else { return }
+        // Leave the Carbon/NSApplication event callback before touching windows.
+        // AppKit activation is asynchronous; an arbitrary delay or NSApp.isActive
+        // is not a substitute for checking UIKit's foreground lifecycle state.
+        presentationTask = Task { @MainActor [weak self] in
+            guard let self, !Task.isCancelled else { return }
+            self.presentationTask = nil
+            guard self.isPresentationRequested else { return }
+            guard self.isApplicationInForeground() else {
+                guard !self.activationRequested else { return }
+                self.activationRequested = true
+                SybilLog.debug(SybilLog.ui, "Waiting for Catalyst foreground before showing Quick Question")
+                self.bridge?.activateForQuickQuestion()
+                // Also handle an activation that completes synchronously. If it
+                // has not completed, subsequent UIKit notifications resume us.
+                self.schedulePresentationIfNeeded()
+                return
+            }
+            self.activationRequested = false
+            self.updateAndObserve()
+            self.bridge?.show()
+        }
+    }
+
+    private func cancelPresentation() {
+        isPresentationRequested = false
+        activationRequested = false
+        presentationTask?.cancel()
+        presentationTask = nil
+        observationGeneration += 1
+    }
+
+    private func updateAndObserve() {
+        guard let bridge else { return }
+        observationGeneration += 1
+        let generation = observationGeneration
+        let state = withObservationTracking {
+            viewModel?.macQuickQuestionPanelState ?? SybilQuickQuestionPanelState()
+        } onChange: { [weak self] in
+            Task { @MainActor [weak self] in
+                guard let self, self.observationGeneration == generation, self.bridge?.isVisible == true else { return }
+                self.updateAndObserve()
+            }
+        }
+        do {
+            bridge.updateState(try JSONEncoder().encode(state))
+        } catch {
+            SybilLog.error(SybilLog.ui, "Could not update the macOS Quick Question panel", error: error)
+        }
+    }
+
+    func quickQuestionToggleRequested() {
+        toggle()
+    }
+
+    func quickQuestionPromptChanged(_ prompt: String) {
+        guard let viewModel, !viewModel.isQuickQuestionSending, !viewModel.isConvertingQuickQuestion else { return }
+        viewModel.updateQuickQuestionPrompt(prompt)
+        updateAndObserve()
+    }
+
+    func quickQuestionSubmitRequested() {
+        guard let viewModel, viewModel.isAuthenticated, !viewModel.isCheckingSession else { return }
+        viewModel.sendQuickQuestion()
+        updateAndObserve()
+    }
+
+    func quickQuestionProviderChanged(_ provider: String) {
+        guard let viewModel, let provider = Provider(rawValue: provider),
+              !viewModel.isQuickQuestionSending, !viewModel.isConvertingQuickQuestion else { return }
+        viewModel.setQuickQuestionProvider(provider)
+        updateAndObserve()
+    }
+
+    func quickQuestionModelChanged(_ model: String) {
+        guard let viewModel, !viewModel.isQuickQuestionSending, !viewModel.isConvertingQuickQuestion else { return }
+        viewModel.setQuickQuestionModel(model)
+        updateAndObserve()
+    }
+
+    func quickQuestionConvertRequested() {
+        guard let viewModel, viewModel.canConvertQuickQuestion else { return }
+        Task { @MainActor [weak self] in
+            guard let self else { return }
+            if await viewModel.convertQuickQuestionToChat() {
+                self.cancelPresentation()
+                self.openMainWindow()
+                self.bridge?.hide()
+            } else {
+                self.updateAndObserve()
+            }
+        }
+    }
+
+    func quickQuestionOpenAppRequested() {
+        cancelPresentation()
+        openMainWindow()
+        bridge?.hide()
+    }
+
+    private func openMainWindow() {
+        let session = UIApplication.shared.connectedScenes
+            .compactMap { $0 as? UIWindowScene }
+            .first?.session
+        UIApplication.shared.requestSceneSessionActivation(session, userActivity: nil, options: nil)
+        bridge?.activateMainWindow()
+    }
+
+    func quickQuestionPanelDismissed() {
+        // Toggling out preserves the draft and lets an answer finish. Reopening
+        // reads the latest state, without starting a new request.
+        cancelPresentation()
+    }
+}
+
+extension SybilViewModel {
+    var macQuickQuestionPanelState: SybilQuickQuestionPanelState {
+        SybilQuickQuestionPanelState(
+            prompt: quickQuestionPrompt,
+            answer: quickQuestionAnswerText,
+            toolSummaries: quickQuestionMessages.compactMap { message in
+                guard let metadata = message.toolCallMetadata else { return nil }
+                return metadata.summary ?? message.content
+            },
+            provider: quickQuestionProvider.rawValue,
+            providers: providerOptions.map { .init(id: $0.rawValue, title: $0.displayName) },
+            model: quickQuestionModel,
+            models: quickQuestionProviderModelOptions,
+            isSending: isQuickQuestionSending,
+            isConverting: isConvertingQuickQuestion,
+            canSend: isAuthenticated && !isCheckingSession && canSendQuickQuestion,
+            canConvert: canConvertQuickQuestion,
+            isAuthenticated: isAuthenticated,
+            isCheckingSession: isCheckingSession,
+            error: quickQuestionError
+        )
+    }
+}
+#endif

ios/Packages/Sybil/Sources/Sybil/SybilMacWindowCommands.swift

@@ -0,0 +1,37 @@
+#if targetEnvironment(macCatalyst)
+import Combine
+
+// Scene lifetime is independent of focused commands: a connected window may
+// still be authenticating, inactive, or minimized. The native Quick Question
+// panel is not a workspace scene and does not count as an open window here.
+@MainActor
+public final class SybilMacWindowCommands: ObservableObject {
+    public static let shared = SybilMacWindowCommands()
+
+    @Published private var connectedWindowIDs: Set<String> = []
+
+    init() {}
+
+    var hasOpenWindows: Bool { !connectedWindowIDs.isEmpty }
+
+    public func windowConnected(sessionID: String) {
+        connectedWindowIDs.insert(sessionID)
+    }
+
+    public func windowDisconnected(sessionID: String) {
+        connectedWindowIDs.remove(sessionID)
+    }
+
+    func canStartNewChat(hasFocusedActions: Bool) -> Bool {
+        !hasOpenWindows || hasFocusedActions
+    }
+
+    func newChatOrWindow(newChat: (() -> Void)?, openWindow: () -> Void) {
+        if hasOpenWindows {
+            newChat?()
+        } else {
+            openWindow()
+        }
+    }
+}
+#endif

ios/Packages/Sybil/Sources/Sybil/SybilModels.swift

@@ -219,6 +219,7 @@ public struct WorkspaceItem: Codable, Identifiable, Hashable, Sendable {
     public var id: String
     public var title: String?
     public var query: String?
+    public var searchSnippet: String? = nil
     public var parentChatId: String? = nil
     public var titleGenerationPending = false
     public var createdAt: Date
@@ -302,6 +303,7 @@ extension WorkspaceItem {
         case id
         case title
         case query
+        case searchSnippet
         case parentChatId
         case titleGenerationPending
         case createdAt
@@ -320,6 +322,7 @@ extension WorkspaceItem {
         id = try container.decode(String.self, forKey: .id)
         title = try container.decodeIfPresent(String.self, forKey: .title)
         query = try container.decodeIfPresent(String.self, forKey: .query)
+        searchSnippet = try container.decodeIfPresent(String.self, forKey: .searchSnippet)
         parentChatId = try container.decodeIfPresent(String.self, forKey: .parentChatId)
         titleGenerationPending = try container.decodeIfPresent(Bool.self, forKey: .titleGenerationPending) ?? false
         createdAt = try container.decode(Date.self, forKey: .createdAt)
@@ -356,9 +359,14 @@ public struct Message: Codable, Identifiable, Hashable, Sendable {
             toolName: object["toolName"]?.stringValue ?? name,
             status: object["status"]?.stringValue,
             summary: object["summary"]?.stringValue,
-            durationMs: object["durationMs"]?.numberValue.map(Int.init),
+            durationMs: object["durationMs"]?.numberValue.flatMap { value in
+                value >= 0 && value < Double(Int.max) ? Int(value) : nil
+            },
             error: object["error"]?.stringValue,
-            resultPreview: object["resultPreview"]?.stringValue
+            resultPreview: object["resultPreview"]?.stringValue,
+            args: object["args"],
+            result: object["result"],
+            startedAt: object["startedAt"]?.stringValue
         )
     }
 
@@ -375,6 +383,9 @@ public struct ToolCallMetadata: Hashable, Sendable {
     public var durationMs: Int?
     public var error: String?
     public var resultPreview: String?
+    public var args: JSONValue? = nil
+    public var result: JSONValue? = nil
+    public var startedAt: String? = nil
 }
 
 public enum JSONValue: Codable, Hashable, Sendable {
@@ -730,6 +741,11 @@ struct SearchListResponse: Codable {
     var searches: [SearchSummary]
 }
 
+struct WorkspaceSearchResponse: Codable, Sendable {
+    var items: [WorkspaceItem]
+    var hasMore: Bool
+}
+
 struct WorkspaceListResponse: Codable {
     var items: [WorkspaceItem]
 }

ios/Packages/Sybil/Sources/Sybil/SybilPhoneShellView.swift

@@ -2,6 +2,7 @@ import Observation
 import SwiftUI
 
 enum PhoneRoute: Hashable {
+    case agent
     case chat(String)
     case search(String)
     case draftChat
@@ -14,6 +15,7 @@ enum PhoneRoute: Hashable {
             return .chat(chatID)
         case let .search(searchID):
             return .search(searchID)
+        case .agent: return .agent
         case .settings:
             return .settings
         }
@@ -64,6 +66,7 @@ struct SybilPhoneShellView: View {
             return .chat(chatID)
         case let .search(searchID):
             return .search(searchID)
+        case .agent: return .agent
         case .draftChat, .draftSearch, .settings:
             return nil
         }
@@ -87,6 +90,10 @@ struct SybilPhoneShellView: View {
         .tint(SybilTheme.primary)
         .animation(.easeOut(duration: 0.22), value: route)
         .animation(.easeOut(duration: 0.18), value: isSidebarOverlayPresented)
+        .onChange(of: shouldRenderSidebarOverlay, initial: true) { _, covered in
+            viewModel.agent.isObscured = covered
+        }
+        .onDisappear { viewModel.agent.isObscured = false }
         .onChange(of: scenePhase) { _, nextPhase in
             switch nextPhase {
             case .background:
@@ -130,7 +137,10 @@ struct SybilPhoneShellView: View {
             route: route,
             onRequestBack: { _ in showSidebarOverlay() },
             onRequestNewChat: sidebarWorkspaceNewChatAction,
-            onShowSidebar: showSidebarOverlay
+            onShowSidebar: showSidebarOverlay,
+            onOpenForkedChat: { chatID in
+                showRoute(.chat(chatID))
+            }
         )
         .background(SybilTheme.background)
         .blur(radius: SidebarOverlaySwipeMetrics.workspaceBlurRadius(for: sidebarOverlayProgress))
@@ -239,6 +249,7 @@ struct SybilPhoneShellView: View {
     }
 
     private func showRoute(_ nextRoute: PhoneRoute) {
+        dismissSidebarKeyboard()
         let update = {
             route = nextRoute
         }
@@ -268,12 +279,19 @@ struct SybilPhoneShellView: View {
     }
 
     private func hideSidebarOverlay() {
+        dismissSidebarKeyboard()
         withAnimation(.easeOut(duration: 0.18)) {
             isSidebarOverlayPresented = false
         }
         resetSidebarSwipe(animated: false)
     }
 
+    private func dismissSidebarKeyboard() {
+        if isSidebarOverlayPresented {
+            UIApplication.shared.sendAction(#selector(UIResponder.resignFirstResponder), to: nil, from: nil, for: nil)
+        }
+    }
+
     private func openSidebarSelection(_ selection: SidebarSelection) {
         if openingSelectionRequestID != nil, sidebarHighlightSelection == selection {
             return
@@ -511,7 +529,12 @@ private struct SybilPhoneSidebarRoot: View {
                 },
                 onSelect: { item in
                     onSelect(item.selection)
-                }
+                },
+                onOpenForkedChat: { chatID in
+                    onSelect(.chat(chatID))
+                },
+                onSelectAgent: { onSelect(.agent) },
+                isAgentSelected: highlightedSelection == .agent
             )
         }
         .background(SybilTheme.panelGradient)
@@ -585,6 +608,7 @@ private struct SybilPhoneDestinationView: View {
     let onRequestBack: (_ animateNavigation: Bool) -> Void
     let onRequestNewChat: (() -> Void)?
     let onShowSidebar: () -> Void
+    let onOpenForkedChat: (String) -> Void
 
     var body: some View {
         SybilWorkspaceView(
@@ -593,7 +617,8 @@ private struct SybilPhoneDestinationView: View {
             navigationLeadingControl: .showSidebar,
             onShowSidebar: onShowSidebar,
             onRequestBack: onRequestBack,
-            onRequestNewChat: onRequestNewChat
+            onRequestNewChat: onRequestNewChat,
+            onOpenForkedChat: onOpenForkedChat
         )
         .frame(maxWidth: .infinity, maxHeight: .infinity, alignment: .topLeading)
         .task(id: route) {
@@ -613,6 +638,8 @@ private struct SybilPhoneDestinationView: View {
                 return
             }
             viewModel.select(.search(searchID))
+        case .agent:
+            viewModel.select(.agent)
         case .draftChat:
             viewModel.startNewChat()
         case .draftSearch:

ios/Packages/Sybil/Sources/Sybil/SybilPromptField.swift

@@ -0,0 +1,62 @@
+import SwiftUI
+
+struct SybilPromptField: View {
+    var placeholder: String = "Enter Prompt"
+    @Binding var text: String
+    @FocusState.Binding var focused: Bool
+    var selection: Binding<TextSelection?>? = nil
+    var axis: Axis = .vertical
+    var submit: () -> Void
+
+    var body: some View {
+        Group {
+            if let selection {
+                TextField(placeholder, text: $text, selection: selection, axis: axis)
+            } else {
+                TextField(placeholder, text: $text, axis: axis)
+            }
+        }
+        .focused($focused)
+        .font(.sybil(.body))
+        .textInputAutocapitalization(.sentences)
+        .autocorrectionDisabled(false)
+        .lineLimit(1 ... 6)
+        .submitLabel(.send)
+        .onSubmit(submit)
+        .submitOnMacReturn(submit)
+        .padding(.horizontal, 12)
+        .padding(.vertical, 10)
+        .background(
+            RoundedRectangle(cornerRadius: 12)
+                .fill(SybilTheme.composerGradient)
+                .opacity(0.98)
+        )
+        .foregroundStyle(SybilTheme.text)
+    }
+}
+
+struct SybilPromptSendButton: View {
+    var systemImage: String = "arrow.up"
+    var canSend: Bool
+    var submit: () -> Void
+
+    var body: some View {
+        Button(action: submit) {
+            Image(systemName: systemImage)
+                .font(.system(size: 17, weight: .semibold))
+                .frame(width: 40, height: 40)
+                .background(
+                    Circle()
+                        .fill(
+                            canSend
+                                ? AnyShapeStyle(SybilTheme.primaryGradient)
+                                : AnyShapeStyle(SybilTheme.surface)
+                        )
+                )
+                .foregroundStyle(canSend ? SybilTheme.text : SybilTheme.textMuted)
+        }
+        .buttonStyle(.plain)
+        .disabled(!canSend)
+        .accessibilityLabel("Send message")
+    }
+}

ios/Packages/Sybil/Sources/Sybil/SybilQuickQuestionView.swift

@@ -8,6 +8,7 @@ struct SybilQuickQuestionView: View {
 
     @Environment(\.dismiss) private var dismiss
     @FocusState private var promptFocused: Bool
+    @State private var promptSelection: TextSelection?
 
     private var hasAnswerContent: Bool {
         !viewModel.quickQuestionMessages.isEmpty || viewModel.quickQuestionError != nil
@@ -36,6 +37,8 @@ struct SybilQuickQuestionView: View {
                 return
             }
             promptFocused = true
+            let prompt = viewModel.quickQuestionPrompt
+            promptSelection = TextSelection(range: prompt.startIndex..<prompt.endIndex)
         }
     }
 
@@ -59,7 +62,11 @@ struct SybilQuickQuestionView: View {
             VStack(alignment: .leading, spacing: 12) {
                 if hasAnswerContent {
                     ForEach(viewModel.quickQuestionMessages) { message in
-                        QuickQuestionMessageView(message: message, isSending: viewModel.isQuickQuestionSending)
+                        QuickQuestionMessageView(
+                            message: message,
+                            isSending: viewModel.isQuickQuestionSending,
+                            streamingAssistantPresentation: viewModel.quickQuestionStreamingAssistantPresentation
+                        )
                     }
 
                     if let error = viewModel.quickQuestionError {
@@ -88,46 +95,19 @@ struct SybilQuickQuestionView: View {
     private var composer: some View {
         VStack(alignment: .leading, spacing: 10) {
             HStack(alignment: .bottom, spacing: 10) {
-                TextField(
-                    "Ask anything...",
+                SybilPromptField(
+                    placeholder: "Ask anything...",
                     text: Binding(
                         get: { viewModel.quickQuestionPrompt },
                         set: { viewModel.updateQuickQuestionPrompt($0) }
                     ),
-                    axis: .vertical
+                    focused: $promptFocused,
+                    selection: $promptSelection,
+                    axis: .horizontal,
+                    submit: submitQuestion
                 )
-                .focused($promptFocused)
-                .font(.body)
-                .textInputAutocapitalization(.sentences)
-                .autocorrectionDisabled(false)
-                .lineLimit(1 ... 6)
-                .submitLabel(.send)
-                .onSubmit(submitQuestion)
-                .padding(.horizontal, 12)
-                .padding(.vertical, 10)
-                .background(
-                    RoundedRectangle(cornerRadius: 12)
-                        .fill(SybilTheme.composerGradient)
-                        .opacity(0.98)
-                )
-                .foregroundStyle(SybilTheme.text)
 
-                Button(action: submitQuestion) {
-                    Image(systemName: "arrow.up")
-                        .font(.body.weight(.semibold))
-                        .frame(width: 40, height: 40)
-                        .background(
-                            Circle()
-                                .fill(
-                                    viewModel.canSendQuickQuestion
-                                        ? AnyShapeStyle(SybilTheme.primaryGradient)
-                                        : AnyShapeStyle(SybilTheme.surfaceStrong.opacity(0.92))
-                                )
-                        )
-                        .foregroundStyle(viewModel.canSendQuickQuestion ? SybilTheme.text : SybilTheme.textMuted)
-                }
-                .buttonStyle(.plain)
-                .disabled(!viewModel.canSendQuickQuestion)
+                SybilPromptSendButton(canSend: viewModel.canSendQuickQuestion, submit: submitQuestion)
                 .accessibilityLabel("Ask quick question")
             }
 
@@ -259,11 +239,11 @@ private struct QuickQuestionPickerPill: View {
 private struct QuickQuestionMessageView: View {
     var message: Message
     var isSending: Bool
+    var streamingAssistantPresentation: StreamingAssistantPresentation?
 
-    private var isPendingAssistant: Bool {
-        message.id.hasPrefix("temp-assistant-quick-") &&
-            isSending &&
-            message.content.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+    private var matchingStreamingAssistantPresentation: StreamingAssistantPresentation? {
+        guard streamingAssistantPresentation?.messageID == message.id else { return nil }
+        return streamingAssistantPresentation
     }
 
     var body: some View {
@@ -272,15 +252,12 @@ private struct QuickQuestionMessageView: View {
                 .font(.caption)
                 .foregroundStyle(SybilTheme.textMuted)
                 .fixedSize(horizontal: false, vertical: true)
-        } else if isPendingAssistant {
-            HStack(spacing: 8) {
-                ProgressView()
-                    .controlSize(.small)
-                    .tint(SybilTheme.primary)
-                Text("Thinking...")
-                    .font(.caption)
-                    .foregroundStyle(SybilTheme.textMuted)
-            }
+        } else if message.id.hasPrefix("temp-assistant-quick-") {
+            StreamingAssistantContentView(
+                content: message.content,
+                isSending: isSending,
+                presentation: matchingStreamingAssistantPresentation
+            )
         } else if !message.content.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty {
             Markdown(message.content)
                 .font(.body)

ios/Packages/Sybil/Sources/Sybil/SybilReturnKeySubmission.swift

@@ -0,0 +1,35 @@
+import SwiftUI
+
+// A vertical TextField treats Return as editing, not form submission. Handle
+// unmodified Return explicitly on Mac, leaving Shift+Return and other modified
+// keys to the native text field. iOS keeps its software-keyboard behavior.
+@MainActor
+enum SybilReturnKeySubmission {
+    static func handle(
+        phase: KeyPress.Phases,
+        modifiers: EventModifiers,
+        submit: () -> Void
+    ) -> KeyPress.Result {
+        guard modifiers.intersection([.shift, .command, .control, .option]).isEmpty else {
+            return .ignored
+        }
+        if phase == .down {
+            submit()
+        }
+        // Consume autorepeat without submitting or inserting a newline.
+        return .handled
+    }
+}
+
+extension View {
+    @ViewBuilder
+    func submitOnMacReturn(_ submit: @escaping () -> Void) -> some View {
+        #if targetEnvironment(macCatalyst)
+        onKeyPress(keys: [.return, KeyEquivalent("\u{3}")], phases: [.down, .repeat]) { press in
+            SybilReturnKeySubmission.handle(phase: press.phase, modifiers: press.modifiers, submit: submit)
+        }
+        #else
+        self
+        #endif
+    }
+}

ios/Packages/Sybil/Sources/Sybil/SybilSettingsStore.swift

@@ -52,9 +52,9 @@ final class SybilSettingsStore {
         self.preferredModelByProvider = preferredModels
 
         self.quickQuestionPreferredProvider =
-            defaults.string(forKey: Keys.quickQuestionPreferredProvider).flatMap(Provider.init(rawValue:)) ?? provider
+            defaults.string(forKey: Keys.quickQuestionPreferredProvider).flatMap(Provider.init(rawValue:)) ?? .openai
         self.quickQuestionPreferredModelByProvider = [
-            .openai: defaults.string(forKey: Keys.quickQuestionPreferredOpenAIModel) ?? preferredModels[.openai] ?? "gpt-4.1-mini",
+            .openai: defaults.string(forKey: Keys.quickQuestionPreferredOpenAIModel) ?? "chat-latest",
             .anthropic: defaults.string(forKey: Keys.quickQuestionPreferredAnthropicModel) ?? preferredModels[.anthropic] ?? "claude-3-5-sonnet-latest",
             .xai: defaults.string(forKey: Keys.quickQuestionPreferredXAIModel) ?? preferredModels[.xai] ?? "grok-3-mini",
             .gemini: defaults.string(forKey: Keys.quickQuestionPreferredGeminiModel) ?? preferredModels[.gemini] ?? "gemini-3.5-flash",

ios/Packages/Sybil/Sources/Sybil/SybilSidebarView.swift

@@ -1,5 +1,6 @@
 import Observation
 import SwiftUI
+import UIKit
 
 struct SybilSidebarView: View {
     @Bindable var viewModel: SybilViewModel
@@ -111,8 +112,12 @@ struct SybilSidebarItemList: View {
     @Bindable var viewModel: SybilViewModel
     var isSelected: (SidebarItem) -> Bool
     var onSelect: (SidebarItem) -> Void
+    var onOpenForkedChat: ((String) -> Void)? = nil
+    var onSelectAgent: (() -> Void)? = nil
+    var isAgentSelected: Bool? = nil
     @State private var renameTarget: SidebarItem?
     @State private var renameTitle = ""
+    @FocusState private var isSearchFocused: Bool
 
     private var isRenameAlertPresented: Binding<Bool> {
         Binding {
@@ -126,37 +131,122 @@ struct SybilSidebarItemList: View {
     }
 
     var body: some View {
+        VStack(spacing: 0) {
+            Button {
+                if let onSelectAgent { onSelectAgent() } else { viewModel.select(.agent) }
+            } label: {
+                HStack(spacing: 10) {
+                    AgentAvatar()
+                    Text("Sybil Agent").font(.sybil(.subheadline, weight: .semibold))
+                    Spacer()
+                    if viewModel.agent.isBusy { ProgressView().controlSize(.small).accessibilityLabel("Agent is working") }
+                    if viewModel.agent.unreadCount > 0 {
+                        Text(String(viewModel.agent.unreadCount))
+                            .font(.sybil(.caption, weight: .semibold))
+                            .padding(.horizontal, 7).padding(.vertical, 3)
+                            .background(SybilTheme.primary, in: Capsule())
+                            .accessibilityLabel("\(viewModel.agent.unreadCount) unread replies")
+                    }
+                }
+                .foregroundStyle(SybilTheme.text)
+                .padding(12)
+                .background((isAgentSelected ?? (viewModel.selectedItem == .agent && viewModel.draftKind == nil)) ? SybilTheme.primary.opacity(0.22) : SybilTheme.surface.opacity(0.7), in: RoundedRectangle(cornerRadius: 12))
+                .overlay(RoundedRectangle(cornerRadius: 12).stroke(SybilTheme.border))
+            }
+            .buttonStyle(.plain)
+            .accessibilityIdentifier("agent-sidebar")
+            .padding(.horizontal, 10).padding(.top, 10)
+            HStack(spacing: 10) {
+                Image(systemName: "magnifyingglass")
+                    .foregroundStyle(SybilTheme.textMuted)
+                TextField("Search chats and web searches", text: Binding(
+                    get: { viewModel.sidebarSearchQuery },
+                    set: { viewModel.setSidebarSearchQuery($0) }
+                ))
+                .font(.sybil(.subheadline))
+                .foregroundStyle(SybilTheme.text)
+                .textInputAutocapitalization(.never)
+                .autocorrectionDisabled()
+                .submitLabel(.search)
+                .focused($isSearchFocused)
+                .accessibilityLabel("Search conversation history and saved web searches")
+                .onSubmit {
+                    isSearchFocused = false
+                    Task { await viewModel.refreshSidebarSearch() }
+                }
+
+                if !viewModel.sidebarSearchQuery.isEmpty {
+                    Button {
+                        viewModel.setSidebarSearchQuery("")
+                    } label: {
+                        Image(systemName: "xmark.circle.fill")
+                            .foregroundStyle(SybilTheme.textMuted)
+                    }
+                    .buttonStyle(.plain)
+                    .accessibilityLabel("Clear history search")
+                }
+            }
+            .padding(12)
+            .background(SybilTheme.surface, in: RoundedRectangle(cornerRadius: 12))
+            .overlay(RoundedRectangle(cornerRadius: 12).stroke(SybilTheme.border, lineWidth: 1))
+            .padding(10)
+
+            itemList
+        }
+    }
+
+    private var itemList: some View {
         Group {
-            if viewModel.isLoadingCollections && viewModel.sidebarItems.isEmpty {
+            if viewModel.isSearchingSidebar ||
+                (!viewModel.isSidebarSearchActive && viewModel.isLoadingCollections && viewModel.sidebarItems.isEmpty) {
                 VStack(alignment: .leading, spacing: 8) {
                     ProgressView()
                         .tint(SybilTheme.primary)
-                    Text("Loading conversations…")
+                    Text(viewModel.isSidebarSearchActive ? "Searching history…" : "Loading conversations…")
                         .font(.sybil(.footnote))
                         .foregroundStyle(SybilTheme.textMuted)
                 }
                 .frame(maxWidth: .infinity, maxHeight: .infinity, alignment: .topLeading)
                 .padding(16)
-            } else if viewModel.sidebarItems.isEmpty {
+            } else if let error = viewModel.sidebarSearchError, viewModel.isSidebarSearchActive {
+                VStack(spacing: 12) {
+                    Text(error)
+                        .font(.sybil(.footnote))
+                        .foregroundStyle(SybilTheme.danger)
+                        .multilineTextAlignment(.center)
+                    Button("Try again") {
+                        Task { await viewModel.refreshSidebarSearch() }
+                    }
+                }
+                .frame(maxWidth: .infinity, maxHeight: .infinity)
+                .padding(16)
+            } else if viewModel.visibleSidebarItems.isEmpty {
                 VStack(spacing: 10) {
-                    Image(systemName: "message.badge")
+                    Image(systemName: viewModel.isSidebarSearchActive ? "magnifyingglass" : "message.badge")
                         .font(.system(size: 20, weight: .medium))
                         .foregroundStyle(SybilTheme.textMuted)
-                    Text("Start a chat or run your first search.")
+                    Text(viewModel.isSidebarSearchActive ? "No matching chats or web searches." : "Start a chat or run your first search.")
                         .font(.sybil(.footnote))
                         .multilineTextAlignment(.center)
                         .foregroundStyle(SybilTheme.textMuted)
+                    if viewModel.isSidebarSearchActive {
+                        Text("Try a title, a query, or words from a conversation.")
+                            .font(.sybil(.caption))
+                            .multilineTextAlignment(.center)
+                            .foregroundStyle(SybilTheme.textMuted)
+                    }
                 }
                 .frame(maxWidth: .infinity, maxHeight: .infinity)
                 .padding(16)
             } else {
                 ScrollView {
                     LazyVStack(alignment: .leading, spacing: 8) {
-                        ForEach(viewModel.sidebarItems) { item in
+                        ForEach(viewModel.visibleSidebarItems) { item in
                             Button {
+                                isSearchFocused = false
                                 onSelect(item)
                             } label: {
-                                SybilSidebarRow(item: item, isSelected: isSelected(item))
+                                SybilSidebarRow(item: item, isSelected: isSelected(item), searchQuery: viewModel.sidebarSearchQuery)
                             }
                             .buttonStyle(.plain)
                             .contextMenu {
@@ -168,13 +258,26 @@ struct SybilSidebarItemList: View {
                                     Label(item.starred ? "Unstar" : "Star", systemImage: item.starred ? "star.slash" : "star")
                                 }
 
-                                if item.kind == .chat {
+                                if case let .chat(chatID) = item.selection {
+                                    Button {
+                                        forkChat(item)
+                                    } label: {
+                                        Label("Fork", systemImage: "arrow.triangle.branch")
+                                    }
+                                    .disabled(viewModel.isForkingChat || item.isRunning)
+
                                     Button {
                                         renameTarget = item
                                         renameTitle = item.title
                                     } label: {
                                         Label("Rename", systemImage: "pencil")
                                     }
+
+                                    Button {
+                                        UIPasteboard.general.string = chatID
+                                    } label: {
+                                        Label("Copy Session ID", systemImage: "doc.on.doc")
+                                    }
                                 }
 
                                 Button(role: .destructive) {
@@ -186,9 +289,16 @@ struct SybilSidebarItemList: View {
                                 }
                             }
                         }
+                        if viewModel.isSidebarSearchActive && viewModel.sidebarSearchHasMore {
+                            Text("Showing the 100 most recent matches. Refine your search to find more.")
+                                .font(.sybil(.caption))
+                                .foregroundStyle(SybilTheme.textMuted)
+                                .padding(12)
+                        }
                     }
                     .padding(10)
                 }
+                .scrollDismissesKeyboard(.interactively)
                 .refreshable {
                     await viewModel.refreshSidebarCollectionsFromPullToRefresh()
                 }
@@ -215,11 +325,52 @@ struct SybilSidebarItemList: View {
             .disabled(renameTitle.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty)
         }
     }
+
+    private func forkChat(_ item: SidebarItem) {
+        guard case let .chat(chatID) = item.selection else {
+            return
+        }
+
+        let selectionAtStart = viewModel.selectedItem
+        let draftKindAtStart = viewModel.draftKind
+        Task {
+            guard let forkedChat = await viewModel.forkChat(chatID: chatID),
+                  viewModel.selectedItem == selectionAtStart,
+                  viewModel.draftKind == draftKindAtStart else {
+                return
+            }
+
+            if let onOpenForkedChat {
+                onOpenForkedChat(forkedChat.id)
+            } else {
+                viewModel.select(.chat(forkedChat.id))
+            }
+        }
+    }
 }
 
 struct SybilSidebarRow: View {
     var item: SidebarItem
     var isSelected: Bool
+    var searchQuery = ""
+
+    private func highlightedSnippet(_ snippet: String) -> AttributedString {
+        var highlighted = AttributedString(snippet)
+        guard let words = try? NSRegularExpression(pattern: #"[\p{L}\p{N}\p{M}\p{Co}]+"#) else { return highlighted }
+        let terms = words.matches(in: searchQuery, range: NSRange(searchQuery.startIndex..., in: searchQuery)).compactMap { match -> String? in
+            guard let range = Range(match.range, in: searchQuery) else { return nil }
+            return String(searchQuery[range]).folding(options: [.diacriticInsensitive, .caseInsensitive], locale: Locale(identifier: "en_US_POSIX"))
+        }.filter { !$0.isEmpty }
+        for word in words.matches(in: snippet, range: NSRange(snippet.startIndex..., in: snippet)) {
+            guard let range = Range(word.range, in: snippet) else { continue }
+            let normalized = String(snippet[range]).folding(options: [.diacriticInsensitive, .caseInsensitive], locale: Locale(identifier: "en_US_POSIX"))
+            guard terms.contains(where: { normalized.hasPrefix($0) }),
+                  let start = AttributedString.Index(range.lowerBound, within: highlighted),
+                  let end = AttributedString.Index(range.upperBound, within: highlighted) else { continue }
+            highlighted[start..<end].font = .sybil(.caption, weight: .semibold)
+        }
+        return highlighted
+    }
 
     private var isHighlighted: Bool {
         isSelected
@@ -281,6 +432,15 @@ struct SybilSidebarRow: View {
                         .frame(maxWidth: .infinity, alignment: .trailing)
                 }
             }
+
+            if let snippet = item.searchSnippet, !snippet.isEmpty {
+                Text(highlightedSnippet(snippet.replacingOccurrences(of: #"\s+"#, with: " ", options: .regularExpression)))
+                    .font(.sybil(.caption))
+                    .foregroundStyle(SybilTheme.textMuted)
+                    .lineLimit(1)
+                    .truncationMode(.middle)
+                    .multilineTextAlignment(.leading)
+            }
         }
         .foregroundStyle(SybilTheme.text)
         .padding(.horizontal, 12)

ios/Packages/Sybil/Sources/Sybil/SybilViewModel.swift

@@ -7,6 +7,7 @@ enum DraftKind {
 }
 
 enum SidebarSelection: Hashable {
+    case agent
     case chat(String)
     case search(String)
     case settings
@@ -17,6 +18,8 @@ enum SidebarSelection: Hashable {
             return "chat:\(chatID)"
         case let .search(searchID):
             return "search:\(searchID)"
+        case .agent:
+            return "agent"
         case .settings:
             return "settings"
         }
@@ -38,11 +41,75 @@ struct SidebarItem: Identifiable, Hashable {
     var starredAt: Date?
     var initiatedLabel: String?
     var isRunning: Bool
+    var searchSnippet: String? = nil
+}
+
+enum StreamingAssistantContentUpdate: Equatable, Sendable {
+    case unchanged
+    case replace(String)
+}
+
+struct StreamingAssistantState: Equatable, Sendable {
+    private(set) var segmentText = ""
+    private(set) var initiatedToolCallIDs: Set<String> = []
+    private(set) var ephemeralTrace: String?
+    private(set) var traceRevision = 0
+
+    mutating func receiveDelta(_ text: String) -> StreamingAssistantContentUpdate {
+        guard !text.isEmpty else { return .unchanged }
+        segmentText += text
+        ephemeralTrace = nil
+        return .replace(segmentText)
+    }
+
+    mutating func receiveToolCall(id: String, status: String) -> StreamingAssistantContentUpdate {
+        guard status.lowercased() == "initiated", initiatedToolCallIDs.insert(id).inserted else {
+            return .unchanged
+        }
+
+        let trace = segmentText.trimmingCharacters(in: .whitespacesAndNewlines)
+        segmentText = ""
+        if !trace.isEmpty {
+            ephemeralTrace = trace
+            traceRevision += 1
+        }
+        return .replace("")
+    }
+
+    mutating func receiveDone(_ text: String) -> StreamingAssistantContentUpdate {
+        segmentText = ""
+        ephemeralTrace = nil
+        return .replace(text)
+    }
+
+    mutating func receiveError() -> StreamingAssistantContentUpdate {
+        segmentText = ""
+        ephemeralTrace = nil
+        return .unchanged
+    }
+}
+
+struct StreamingAssistantPresentation: Equatable, Sendable {
+    var messageID: String
+    var ephemeralTrace: String?
+    var traceRevision: Int
 }
 
 private struct PendingChatState {
     var chatID: String?
     var messages: [Message]
+    var streamingAssistantState = StreamingAssistantState()
+
+    var streamingAssistantPresentation: StreamingAssistantPresentation? {
+        guard let messageID = messages.last(where: { $0.id.hasPrefix("temp-assistant-") })?.id else {
+            return nil
+        }
+        return StreamingAssistantPresentation(
+            messageID: messageID,
+            ephemeralTrace: streamingAssistantState.ephemeralTrace,
+            traceRevision: streamingAssistantState.traceRevision
+        )
+    }
 }
 
 private enum ActiveSendContext: Hashable {
@@ -89,6 +156,7 @@ private actor SearchStreamStatus {
 @Observable
 final class SybilViewModel {
     let settings: SybilSettingsStore
+    let agent = SybilAgentRoom()
 
     var isCheckingSession = true
     var isAuthenticated = false
@@ -98,6 +166,15 @@ final class SybilViewModel {
     var chats: [ChatSummary] = []
     var searches: [SearchSummary] = []
     var workspaceItems: [WorkspaceItem] = []
+    private(set) var sidebarSearchQuery = ""
+    private(set) var sidebarSearchResults: [WorkspaceItem] = []
+    private(set) var isSearchingSidebar = false
+    private(set) var sidebarSearchError: String?
+    private(set) var sidebarSearchHasMore = false
+    @ObservationIgnored
+    private var sidebarSearchTask: Task<Void, Never>?
+    @ObservationIgnored
+    private var sidebarSearchRequestID = UUID()
 
     var selectedItem: SidebarSelection?
     var selectedChat: ChatDetail?
@@ -107,6 +184,7 @@ final class SybilViewModel {
     var isLoadingCollections = false
     var isLoadingSelection = false
     var isCreatingSearchChat = false
+    var isForkingChat = false
     var chatBottomPinRequestID = 0
     var errorMessage: String?
 
@@ -125,6 +203,7 @@ final class SybilViewModel {
     var quickQuestionSubmittedModel: String?
     var isQuickQuestionSending = false
     var isConvertingQuickQuestion = false
+    private var quickQuestionStreamingAssistantState = StreamingAssistantState()
 
     @ObservationIgnored
     private var hasBootstrapped = false
@@ -157,7 +236,7 @@ final class SybilViewModel {
     private let clientFactory: (APIConfiguration) -> any SybilAPIClienting
 
     private let fallbackModels: [Provider: [String]] = [
-        .openai: ["gpt-4.1-mini"],
+        .openai: ["gpt-4.1-mini", "chat-latest"],
         .anthropic: ["claude-3-5-sonnet-latest"],
         .xai: ["grok-3-mini"],
         .gemini: ["gemini-3.5-flash", "gemini-flash-latest"],
@@ -264,6 +343,8 @@ final class SybilViewModel {
             }
             return "Search"
 
+        case .agent:
+            return "Sybil Agent"
         case .settings:
             return "Settings"
         }
@@ -303,6 +384,7 @@ final class SybilViewModel {
     }
 
     var showsComposer: Bool {
+        if selectedItem == .agent { return false }
         if case .settings = selectedItem {
             return false
         }
@@ -360,6 +442,30 @@ final class SybilViewModel {
         return canonical
     }
 
+    var displayedStreamingAssistantPresentation: StreamingAssistantPresentation? {
+        if case let .chat(chatID) = selectedItem,
+           let pending = pendingChatStates[chatID] {
+            return pending.streamingAssistantPresentation
+        }
+
+        if draftKind == .chat, let pending = pendingDraftChatState {
+            return pending.streamingAssistantPresentation
+        }
+
+        return nil
+    }
+
+    var quickQuestionStreamingAssistantPresentation: StreamingAssistantPresentation? {
+        guard let messageID = quickQuestionMessages.last(where: { $0.id.hasPrefix("temp-assistant-quick-") })?.id else {
+            return nil
+        }
+        return StreamingAssistantPresentation(
+            messageID: messageID,
+            ephemeralTrace: quickQuestionStreamingAssistantState.ephemeralTrace,
+            traceRevision: quickQuestionStreamingAssistantState.traceRevision
+        )
+    }
+
     var displayedSearch: SearchDetail? {
         if case let .search(searchID) = selectedItem,
            let activeSearch = activeSearchDetails[searchID] {
@@ -393,7 +499,19 @@ final class SybilViewModel {
     }
 
     var sidebarItems: [SidebarItem] {
-        workspaceItems.map { item in
+        makeSidebarItems(workspaceItems)
+    }
+
+    var isSidebarSearchActive: Bool {
+        !sidebarSearchQuery.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+    }
+
+    var visibleSidebarItems: [SidebarItem] {
+        isSidebarSearchActive ? makeSidebarItems(sidebarSearchResults) : sidebarItems
+    }
+
+    private func makeSidebarItems(_ items: [WorkspaceItem]) -> [SidebarItem] {
+        items.map { item in
             switch item.type {
             case .chat:
                 let initiatedLabel: String?
@@ -407,7 +525,8 @@ final class SybilViewModel {
                     initiatedLabel = nil
                 }
                 let starOwner = item.parentChatId.flatMap { parentChatID in
-                    workspaceItems.first(where: { $0.type == .chat && $0.id == parentChatID })
+                    workspaceItems.first(where: { $0.type == .chat && $0.id == parentChatID }) ??
+                        sidebarSearchResults.first(where: { $0.type == .chat && $0.id == parentChatID })
                 } ?? item
 
                 return SidebarItem(
@@ -418,7 +537,8 @@ final class SybilViewModel {
                     starred: starOwner.starred,
                     starredAt: starOwner.starredAt,
                     initiatedLabel: initiatedLabel,
-                    isRunning: isChatRowRunning(item.id)
+                    isRunning: isChatRowRunning(item.id),
+                    searchSnippet: item.searchSnippet
                 )
 
             case .search:
@@ -430,20 +550,70 @@ final class SybilViewModel {
                     starred: item.starred,
                     starredAt: item.starredAt,
                     initiatedLabel: "exa",
-                    isRunning: isSearchRowRunning(item.id)
+                    isRunning: isSearchRowRunning(item.id),
+                    searchSnippet: item.searchSnippet
                 )
             }
         }
     }
 
+    func setSidebarSearchQuery(_ query: String) {
+        let previous = sidebarSearchQuery.trimmingCharacters(in: .whitespacesAndNewlines)
+        sidebarSearchQuery = query
+        guard query.trimmingCharacters(in: .whitespacesAndNewlines) != previous else { return }
+        scheduleSidebarSearch()
+    }
+
+    func refreshSidebarSearch() async {
+        guard isSidebarSearchActive else { return }
+        scheduleSidebarSearch(debounce: false)
+        await sidebarSearchTask?.value
+    }
+
+    private func scheduleSidebarSearch(debounce: Bool = true) {
+        sidebarSearchTask?.cancel()
+        let requestID = UUID()
+        sidebarSearchRequestID = requestID
+        sidebarSearchResults = []
+        sidebarSearchError = nil
+        sidebarSearchHasMore = false
+        isSearchingSidebar = false
+        let query = sidebarSearchQuery.trimmingCharacters(in: .whitespacesAndNewlines)
+        guard !query.isEmpty, isAuthenticated else { return }
+        guard query.utf16.count <= 200 else {
+            sidebarSearchError = "Use 200 characters or fewer."
+            return
+        }
+        isSearchingSidebar = true
+        sidebarSearchTask = Task { [weak self] in
+            do {
+                if debounce { try await Task.sleep(for: .milliseconds(300)) }
+                guard let self, !Task.isCancelled else { return }
+                let response = try await self.client().searchWorkspaceItems(query: query)
+                guard !Task.isCancelled, self.sidebarSearchRequestID == requestID else { return }
+                self.sidebarSearchResults = response.items
+                self.sidebarSearchHasMore = response.hasMore
+                self.isSearchingSidebar = false
+            } catch {
+                guard let self, !Task.isCancelled, self.sidebarSearchRequestID == requestID else { return }
+                self.isSearchingSidebar = false
+                if !self.isCancellation(error) {
+                    self.sidebarSearchError = self.normalizeAPIError(error)
+                }
+            }
+        }
+    }
+
     var selectedChatSummary: ChatSummary? {
         guard case let .chat(chatID) = selectedItem else { return nil }
-        return chats.first(where: { $0.id == chatID })
+        return chats.first(where: { $0.id == chatID }) ??
+            sidebarSearchResults.first(where: { $0.type == .chat && $0.id == chatID })?.chatSummary
     }
 
     var selectedSearchSummary: SearchSummary? {
         guard case let .search(searchID) = selectedItem else { return nil }
-        return searches.first(where: { $0.id == searchID })
+        return searches.first(where: { $0.id == searchID }) ??
+            sidebarSearchResults.first(where: { $0.type == .search && $0.id == searchID })?.searchSummary
     }
 
     var hasRefreshableSelection: Bool {
@@ -454,7 +624,7 @@ final class SybilViewModel {
         switch selectedItem {
         case .chat, .search:
             return true
-        case .settings:
+        case .settings, .agent:
             return false
         }
     }
@@ -470,6 +640,8 @@ final class SybilViewModel {
     }
 
     func reconnect() async {
+        agent.disconnect()
+        setSidebarSearchQuery("")
         isCheckingSession = true
         authError = nil
         errorMessage = nil
@@ -507,6 +679,10 @@ final class SybilViewModel {
             )
 
             await loadInitialData(using: client)
+            if isAuthenticated, let agentClient = client as? any AgentAPIClienting, let baseURL = settings.normalizedAPIBaseURL {
+                agent.configure(client: agentClient, storage: .file(baseURL: baseURL))
+                agent.setActive(isAppActive)
+            }
             startActiveRunPolling()
         } catch {
             stopActiveRunPolling()
@@ -620,6 +796,7 @@ final class SybilViewModel {
         quickQuestionTask = nil
         quickQuestionRunID = nil
         isQuickQuestionSending = false
+        quickQuestionStreamingAssistantState = StreamingAssistantState()
     }
 
     @discardableResult
@@ -636,9 +813,12 @@ final class SybilViewModel {
         }
 
         cancelQuickQuestion()
+        // Lock submission before the task starts, including repeated Return
+        // events delivered in the same main-actor turn.
+        isQuickQuestionSending = true
         let selectedProvider = quickQuestionProvider
         let task = Task { [weak self] in
-            guard let self else {
+            guard let self, !Task.isCancelled else {
                 return
             }
             await self.runQuickQuestion(prompt: content, provider: selectedProvider, model: selectedModel)
@@ -772,6 +952,15 @@ final class SybilViewModel {
     private func beginSelecting(_ selection: SidebarSelection) -> Task<Void, Never>? {
         SybilLog.debug(SybilLog.ui, "Selecting \(selection.id)")
 
+        // Search can open history outside the recent collection. Retain its
+        // summary for provider settings, starring, and subsequent refreshes.
+        if let item = sidebarSearchResults.first(where: {
+            ($0.type == .chat ? SidebarSelection.chat($0.id) : .search($0.id)) == selection
+        }) {
+            if let chat = item.chatSummary { applyChatSummary(chat, moveToFront: false) }
+            if let search = item.searchSummary { applySearchSummary(search, moveToFront: false) }
+        }
+
         if draftKind == nil, selectedItem == selection {
             errorMessage = nil
             if case .search = selection {
@@ -804,7 +993,7 @@ final class SybilViewModel {
             selectedChat = nil
             composerAttachments = []
 
-        case .settings:
+        case .settings, .agent:
             selectedChat = nil
             selectedSearch = nil
             return nil
@@ -822,7 +1011,7 @@ final class SybilViewModel {
     }
 
     private func selectAdjacentSidebarItem(offset: Int) {
-        let items = sidebarItems
+        let items = visibleSidebarItems
         guard !items.isEmpty else {
             return
         }
@@ -842,6 +1031,7 @@ final class SybilViewModel {
     }
 
     func deleteItem(_ selection: SidebarSelection) async {
+        guard selection != .agent else { return }
         guard isAuthenticated else {
             return
         }
@@ -854,10 +1044,11 @@ final class SybilViewModel {
                     try await client.deleteChat(chatID: chatID)
                 case let .search(searchID):
                     try await client.deleteSearch(searchID: searchID)
-                case .settings:
+                case .settings, .agent:
                     break
                 }
                 await refreshCollections(preferredSelection: nil)
+                await refreshSidebarSearch()
             } catch {
                 errorMessage = normalizeAPIError(error)
                 SybilLog.error(SybilLog.ui, "Delete failed", error: error)
@@ -866,6 +1057,36 @@ final class SybilViewModel {
         }
     }
 
+    @discardableResult
+    func forkChat(chatID: String, messageID: String? = nil) async -> ChatSummary? {
+        guard isAuthenticated, !isForkingChat else {
+            return nil
+        }
+
+        guard messageID != nil || !isChatRowRunning(chatID) else {
+            errorMessage = "Wait for the response to finish before forking this chat."
+            return nil
+        }
+
+        let forkDescription = messageID.map { "chat \(chatID) from response \($0)" } ?? "chat \(chatID)"
+        SybilLog.info(SybilLog.ui, "Forking \(forkDescription)")
+        errorMessage = nil
+        isForkingChat = true
+        defer {
+            isForkingChat = false
+        }
+
+        do {
+            let forkedChat = try await client().forkChat(chatID: chatID, messageID: messageID)
+            applyChatSummary(forkedChat, moveToFront: true)
+            return forkedChat
+        } catch {
+            errorMessage = normalizeAPIError(error)
+            SybilLog.error(SybilLog.ui, "Fork failed", error: error)
+            return nil
+        }
+    }
+
     func renameChat(chatID: String, title: String) async {
         guard isAuthenticated else {
             return
@@ -883,6 +1104,7 @@ final class SybilViewModel {
         do {
             let updated = try await client().updateChatTitle(chatID: chatID, title: trimmedTitle)
             applyChatSummary(updated, moveToFront: true)
+            await refreshSidebarSearch()
         } catch {
             errorMessage = normalizeAPIError(error)
             SybilLog.error(SybilLog.ui, "Rename failed", error: error)
@@ -890,6 +1112,7 @@ final class SybilViewModel {
     }
 
     func setItemStarred(_ selection: SidebarSelection, starred: Bool) async {
+        guard selection != .agent else { return }
         guard isAuthenticated else {
             return
         }
@@ -907,7 +1130,7 @@ final class SybilViewModel {
                 case let .search(searchID):
                     let updated = try await client.updateSearchStar(searchID: searchID, starred: starred)
                     applySearchSummary(updated, moveToFront: false)
-                case .settings:
+                case .settings, .agent:
                     break
                 }
             } catch {
@@ -930,12 +1153,14 @@ final class SybilViewModel {
         }
 
         isAppActive = false
+        agent.setActive(false)
         appLifecycleGeneration += 1
         SybilLog.debug(SybilLog.app, "App became inactive for network lifecycle generation \(appLifecycleGeneration)")
     }
 
     func markAppActiveForNetwork() {
         isAppActive = true
+        agent.setActive(true)
     }
 
     func refreshAfterAppBecameActive(refreshCollections shouldRefreshCollections: Bool, refreshSelection shouldRefreshSelection: Bool) async {
@@ -982,6 +1207,10 @@ final class SybilViewModel {
     }
 
     func refreshSidebarCollectionsFromPullToRefresh() async {
+        if isSidebarSearchActive {
+            await refreshSidebarSearch()
+            return
+        }
         guard isAuthenticated, !isCheckingSession else {
             return
         }
@@ -1137,6 +1366,7 @@ final class SybilViewModel {
         quickQuestionSubmittedPrompt = prompt
         quickQuestionSubmittedProvider = provider
         quickQuestionSubmittedModel = model
+        quickQuestionStreamingAssistantState = StreamingAssistantState()
         quickQuestionMessages = [
             Message(
                 id: "temp-assistant-quick-\(UUID().uuidString)",
@@ -1159,13 +1389,11 @@ final class SybilViewModel {
         let streamStatus = CompletionStreamStatus()
 
         do {
-            try await client().runCompletionStream(
-                body: CompletionStreamRequest(
-                    chatId: nil,
-                    persist: false,
+            try await client().runQuickQuestionStream(
+                body: QuickQuestionStreamRequest(
                     provider: provider,
                     model: model,
-                    messages: [CompletionRequestMessage(role: .user, content: prompt)]
+                    question: prompt
                 )
             ) { [weak self] event in
                 guard let self else { return }
@@ -1183,6 +1411,8 @@ final class SybilViewModel {
                 return
             }
 
+            let update = quickQuestionStreamingAssistantState.receiveError()
+            applyQuickQuestionAssistantContentUpdate(update)
             quickQuestionError = normalizeAPIError(error)
             SybilLog.error(SybilLog.ui, "Quick question failed", error: error)
         }
@@ -1197,17 +1427,16 @@ final class SybilViewModel {
             upsertQuickQuestionToolCallMessage(payload)
 
         case let .delta(payload):
-            guard !payload.text.isEmpty else { return }
-            mutateQuickQuestionAssistantMessage { existing in
-                existing + payload.text
-            }
+            let update = quickQuestionStreamingAssistantState.receiveDelta(payload.text)
+            applyQuickQuestionAssistantContentUpdate(update)
 
         case let .done(payload):
-            mutateQuickQuestionAssistantMessage { _ in
-                payload.text
-            }
+            let update = quickQuestionStreamingAssistantState.receiveDone(payload.text)
+            applyQuickQuestionAssistantContentUpdate(update)
 
         case let .error(payload):
+            let update = quickQuestionStreamingAssistantState.receiveError()
+            applyQuickQuestionAssistantContentUpdate(update)
             await streamStatus.setError(payload.message)
 
         case .ignored:
@@ -1259,7 +1488,7 @@ final class SybilViewModel {
 
             if let nextSelection {
                 switch nextSelection {
-                case .settings:
+                case .settings, .agent:
                     selectedChat = nil
                     selectedSearch = nil
                 case .chat, .search:
@@ -1322,7 +1551,15 @@ final class SybilViewModel {
             async let activeRunsValue = client.getActiveRuns()
             let (nextWorkspaceItems, nextActiveRuns) = try await (workspaceItemsValue, activeRunsValue)
 
-            applyWorkspaceItems(nextWorkspaceItems)
+            var items = nextWorkspaceItems
+            // A selected historical search result may be outside the latest 100.
+            if let preferredSelection,
+               let selectedSummary = workspaceItems.first(where: {
+                   ($0.type == .chat ? SidebarSelection.chat($0.id) : .search($0.id)) == preferredSelection
+               }), !items.contains(where: { $0.type == selectedSummary.type && $0.id == selectedSummary.id }) {
+                items.append(selectedSummary)
+            }
+            applyWorkspaceItems(items)
             applyActiveRuns(nextActiveRuns)
 
             SybilLog.info(
@@ -1500,6 +1737,11 @@ final class SybilViewModel {
     }
 
     private func upsertWorkspaceItem(_ item: WorkspaceItem, moveToFront: Bool) {
+        if let index = sidebarSearchResults.firstIndex(where: { $0.type == item.type && $0.id == item.id }) {
+            var updated = item
+            updated.searchSnippet = sidebarSearchResults[index].searchSnippet
+            sidebarSearchResults[index] = updated
+        }
         if let existingIndex = workspaceItems.firstIndex(where: { $0.type == item.type && $0.id == item.id }) {
             workspaceItems.remove(at: existingIndex)
             if moveToFront {
@@ -1523,6 +1765,9 @@ final class SybilViewModel {
         if let parentChatID = workspaceItems.first(where: { $0.type == .chat && $0.id == chatID })?.parentChatId {
             return parentChatID
         }
+        if let parentChatID = sidebarSearchResults.first(where: { $0.type == .chat && $0.id == chatID })?.parentChatId {
+            return parentChatID
+        }
         return chatID
     }
 
@@ -1554,7 +1799,7 @@ final class SybilViewModel {
                 await self?.attachToActiveSearchStream(searchID: searchID)
             }
 
-        case .settings, nil:
+        case .settings, .agent, nil:
             return
         }
     }
@@ -1606,21 +1851,28 @@ final class SybilViewModel {
             }
 
             serverActiveChatIDs.remove(chatID)
-            pendingChatStates[chatID] = nil
             await refreshCollections(preferredSelection: selectedItem, refreshSelection: false, attachVisibleActiveRun: false)
 
             if selectedItem == selection, draftKind == nil {
-                selectedChat = try await client.getChat(chatID: chatID)
-                selectedSearch = nil
+                let refreshedChat = try await client.getChat(chatID: chatID)
+                if selectedItem == selection, draftKind == nil {
+                    selectedChat = refreshedChat
+                    selectedSearch = nil
+                }
             }
+            pendingChatStates[chatID] = nil
         } catch {
             serverActiveChatIDs.remove(chatID)
-            pendingChatStates[chatID] = nil
+            defer {
+                pendingChatStates[chatID] = nil
+            }
 
             if isCancellation(error) {
                 return
             }
 
+            applyPendingAssistantError(chatID: chatID)
+
             if isActiveStreamNotFound(error) {
                 SybilLog.info(SybilLog.app, "Active chat stream \(chatID) no longer exists")
             } else if shouldSuppressInactiveTransportError(error) {
@@ -1634,8 +1886,11 @@ final class SybilViewModel {
 
             if selectedItem == selection, draftKind == nil {
                 do {
-                    selectedChat = try await client().getChat(chatID: chatID)
-                    selectedSearch = nil
+                    let refreshedChat = try await client().getChat(chatID: chatID)
+                    if selectedItem == selection, draftKind == nil {
+                        selectedChat = refreshedChat
+                        selectedSearch = nil
+                    }
                 } catch {
                     SybilLog.warning(SybilLog.app, "Chat refresh after attach failure failed: \(SybilLog.describe(error))")
                 }
@@ -1753,12 +2008,13 @@ final class SybilViewModel {
             return selectedChat?.id != chatID
         case let .search(searchID):
             return selectedSearch?.id != searchID
-        case .settings:
+        case .settings, .agent:
             return false
         }
     }
 
     private func refreshSelectionIfNeeded() async {
+        if selectedItem == .agent { isLoadingSelection = false; return }
         guard let target = selectedItem else {
             selectedChat = nil
             selectedSearch = nil
@@ -1800,7 +2056,7 @@ final class SybilViewModel {
                     selectedSearch = search
                     selectedChat = nil
 
-                case .settings:
+                case .settings, .agent:
                     break
                 }
                 errorMessage = nil
@@ -1855,6 +2111,13 @@ final class SybilViewModel {
         } else {
             pendingDraftChatState = PendingChatState(chatID: nil, messages: optimisticMessages)
         }
+        defer {
+            if let resolvedChatID = chatID {
+                pendingChatStates[resolvedChatID] = nil
+            } else {
+                clearPendingChatState(for: sendContext)
+            }
+        }
         requestChatBottomPin()
 
         if chatID == nil {
@@ -2044,17 +2307,13 @@ final class SybilViewModel {
             upsertPendingToolCallMessage(payload, chatID: chatID)
 
         case let .delta(payload):
-            guard !payload.text.isEmpty else { return }
-            mutatePendingAssistantMessage(chatID: chatID) { existing in
-                existing + payload.text
-            }
+            applyPendingAssistantDelta(payload.text, chatID: chatID)
 
         case let .done(payload):
-            mutatePendingAssistantMessage(chatID: chatID) { _ in
-                payload.text
-            }
+            finalizePendingAssistant(payload.text, chatID: chatID)
 
         case let .error(payload):
+            applyPendingAssistantError(chatID: chatID)
             await streamStatus.setError(payload.message)
 
         case .ignored:
@@ -2232,29 +2491,68 @@ final class SybilViewModel {
         }
     }
 
-    private func mutatePendingAssistantMessage(chatID: String, _ transform: (String) -> String) {
+    private func applyPendingAssistantDelta(_ text: String, chatID: String) {
         guard var pending = pendingChatStates[chatID], !pending.messages.isEmpty else {
             return
         }
 
-        let index = pending.messages.indices.last { pending.messages[$0].id.hasPrefix("temp-assistant-") }
-        guard let index else {
-            return
-        }
-
-        var message = pending.messages[index]
-        message.content = transform(message.content)
-        pending.messages[index] = message
+        let update = pending.streamingAssistantState.receiveDelta(text)
+        applyStreamingAssistantContentUpdate(
+            update,
+            to: &pending.messages,
+            assistantMessagePrefix: "temp-assistant-"
+        )
         pendingChatStates[chatID] = pending
     }
 
-    private func mutateQuickQuestionAssistantMessage(_ transform: (String) -> String) {
-        let index = quickQuestionMessages.indices.last { quickQuestionMessages[$0].id.hasPrefix("temp-assistant-quick-") }
-        guard let index else {
+    private func finalizePendingAssistant(_ text: String, chatID: String) {
+        guard var pending = pendingChatStates[chatID], !pending.messages.isEmpty else {
             return
         }
 
-        quickQuestionMessages[index].content = transform(quickQuestionMessages[index].content)
+        let update = pending.streamingAssistantState.receiveDone(text)
+        applyStreamingAssistantContentUpdate(
+            update,
+            to: &pending.messages,
+            assistantMessagePrefix: "temp-assistant-"
+        )
+        pendingChatStates[chatID] = pending
+    }
+
+    private func applyPendingAssistantError(chatID: String) {
+        guard var pending = pendingChatStates[chatID] else {
+            return
+        }
+
+        let update = pending.streamingAssistantState.receiveError()
+        applyStreamingAssistantContentUpdate(
+            update,
+            to: &pending.messages,
+            assistantMessagePrefix: "temp-assistant-"
+        )
+        pendingChatStates[chatID] = pending
+    }
+
+    private func applyQuickQuestionAssistantContentUpdate(_ update: StreamingAssistantContentUpdate) {
+        applyStreamingAssistantContentUpdate(
+            update,
+            to: &quickQuestionMessages,
+            assistantMessagePrefix: "temp-assistant-quick-"
+        )
+    }
+
+    private func applyStreamingAssistantContentUpdate(
+        _ update: StreamingAssistantContentUpdate,
+        to messages: inout [Message],
+        assistantMessagePrefix: String
+    ) {
+        guard case let .replace(content) = update,
+              let index = messages.indices.last(where: { messages[$0].id.hasPrefix(assistantMessagePrefix) })
+        else {
+            return
+        }
+
+        messages[index].content = content
     }
 
     private func upsertPendingToolCallMessage(_ payload: CompletionStreamToolCall, chatID: String) {
@@ -2262,6 +2560,16 @@ final class SybilViewModel {
             return
         }
 
+        let update = pending.streamingAssistantState.receiveToolCall(
+            id: payload.toolCallId,
+            status: payload.status
+        )
+        applyStreamingAssistantContentUpdate(
+            update,
+            to: &pending.messages,
+            assistantMessagePrefix: "temp-assistant-"
+        )
+
         if let existingIndex = pending.messages.firstIndex(where: { $0.toolCallMetadata?.toolCallId == payload.toolCallId || $0.id == "temp-tool-\(payload.toolCallId)" }) {
             pending.messages[existingIndex] = toolCallMessage(for: payload, id: pending.messages[existingIndex].id)
             pendingChatStates[chatID] = pending
@@ -2280,6 +2588,12 @@ final class SybilViewModel {
     }
 
     private func upsertQuickQuestionToolCallMessage(_ payload: CompletionStreamToolCall) {
+        let update = quickQuestionStreamingAssistantState.receiveToolCall(
+            id: payload.toolCallId,
+            status: payload.status
+        )
+        applyQuickQuestionAssistantContentUpdate(update)
+
         if let existingIndex = quickQuestionMessages.firstIndex(where: { $0.toolCallMetadata?.toolCallId == payload.toolCallId || $0.id == "temp-tool-\(payload.toolCallId)" }) {
             quickQuestionMessages[existingIndex] = toolCallMessage(for: payload, id: quickQuestionMessages[existingIndex].id)
             return
@@ -2429,7 +2743,8 @@ final class SybilViewModel {
 
     private func clearPendingChatState(for context: ActiveSendContext) {
         switch context {
-        case .draftChat:
+        case let .draftChat(identity):
+            guard draftKind == .chat, draftIdentity == identity else { return }
             pendingDraftChatState = nil
         case let .chat(chatID):
             pendingChatStates[chatID] = nil
@@ -2457,7 +2772,7 @@ final class SybilViewModel {
             return chats.contains(where: { $0.id == chatID })
         case let .search(searchID):
             return searches.contains(where: { $0.id == searchID })
-        case .settings:
+        case .settings, .agent:
             return true
         }
     }

ios/Packages/Sybil/Sources/Sybil/SybilWorkspaceView.swift

@@ -1,4 +1,3 @@
-import ImageIO
 import Observation
 import PhotosUI
 import SwiftUI
@@ -19,6 +18,7 @@ struct SybilWorkspaceView: View {
     var onShowSidebar: (() -> Void)? = nil
     var onRequestBack: ((_ animateNavigation: Bool) -> Void)? = nil
     var onRequestNewChat: (() -> Void)? = nil
+    var onOpenForkedChat: ((String) -> Void)? = nil
     @FocusState private var composerFocused: Bool
     @Environment(\.dismiss) private var dismiss
     @State private var isShowingAttachmentOptions = false
@@ -155,8 +155,6 @@ struct SybilWorkspaceView: View {
             workspaceContentStack
 
             if showsCustomWorkspaceNavigation {
-                SybilWorkspaceCharacterBackdrop(isBusy: viewModel.isActiveSelectionSending)
-                    .allowsHitTesting(false)
                 customWorkspaceNavigationBar
             }
         }
@@ -175,6 +173,8 @@ struct SybilWorkspaceView: View {
                 if isSettingsSelected {
                     SybilSettingsView(viewModel: viewModel)
                         .padding(.top, showsCustomWorkspaceNavigation ? customWorkspaceNavigationContentInset : 0)
+                } else if viewModel.selectedItem == .agent {
+                    SybilAgentView(room: viewModel.agent, topInset: showsCustomWorkspaceNavigation ? customWorkspaceNavigationContentInset : 0)
                 } else if viewModel.isSearchMode {
                     SybilSearchResultsView(
                         search: viewModel.displayedSearch,
@@ -193,9 +193,12 @@ struct SybilWorkspaceView: View {
                         messages: viewModel.displayedMessages,
                         isLoading: viewModel.isLoadingSelection,
                         isSending: viewModel.isSendingVisibleChat,
+                        streamingAssistantPresentation: viewModel.displayedStreamingAssistantPresentation,
                         topContentInset: showsCustomWorkspaceNavigation ? customWorkspaceNavigationContentInset : 0,
                         bottomContentInset: viewModel.showsComposer ? composerOverlayContentInset : 0,
-                        bottomPinRequestID: viewModel.chatBottomPinRequestID
+                        bottomPinRequestID: viewModel.chatBottomPinRequestID,
+                        isForking: viewModel.isForkingChat,
+                        onForkResponse: forkResponse
                     )
                     .id(transcriptScrollContextID)
                 }
@@ -229,6 +232,28 @@ struct SybilWorkspaceView: View {
         }
     }
 
+    private func forkResponse(messageID: String) {
+        guard viewModel.draftKind == nil,
+              case let .chat(chatID) = viewModel.selectedItem else {
+            return
+        }
+
+        let selectionAtStart = viewModel.selectedItem
+        Task {
+            guard let forkedChat = await viewModel.forkChat(chatID: chatID, messageID: messageID),
+                  viewModel.selectedItem == selectionAtStart,
+                  viewModel.draftKind == nil else {
+                return
+            }
+
+            if let onOpenForkedChat {
+                onOpenForkedChat(forkedChat.id)
+            } else {
+                viewModel.select(.chat(forkedChat.id))
+            }
+        }
+    }
+
     private var customWorkspaceNavigationBar: some View {
         HStack(spacing: 14) {
             workspaceNavigationLeadingControl
@@ -262,7 +287,7 @@ struct SybilWorkspaceView: View {
                 .lineLimit(1)
                 .minimumScaleFactor(0.78)
 
-            Text(selectedProviderModelSubtitle)
+            Text(viewModel.selectedItem == .agent ? viewModel.agent.connection : selectedProviderModelSubtitle)
                 .font(.sybil(size: 10, weight: .medium))
                 .foregroundStyle(SybilTheme.textMuted)
                 .lineLimit(1)
@@ -484,7 +509,7 @@ struct SybilWorkspaceView: View {
 
     @ViewBuilder
     private var workspaceNavigationTrailingControl: some View {
-        if isSettingsSelected {
+        if isSettingsSelected || viewModel.selectedItem == .agent {
             EmptyView()
         } else if viewModel.isSearchMode {
             searchModeNavigationLabel
@@ -588,46 +613,19 @@ struct SybilWorkspaceView: View {
                     .accessibilityLabel("Attach file")
                 }
 
-                TextField(
-                    viewModel.isSearchMode ? "Search the web" : "Enter Prompt",
+                SybilPromptField(
+                    placeholder: viewModel.isSearchMode ? "Search the web" : "Enter Prompt",
                     text: $viewModel.composer,
-                    axis: .vertical
+                    focused: $composerFocused,
+                    submit: submitComposer
                 )
-                .focused($composerFocused)
-                .textInputAutocapitalization(.sentences)
-                .autocorrectionDisabled(false)
-                .lineLimit(1 ... 6)
-                .submitLabel(.send)
-                .onSubmit {
-                    submitComposer()
-                }
-                .padding(.horizontal, 12)
-                .padding(.vertical, 10)
-                .background(
-                    RoundedRectangle(cornerRadius: 12)
-                        .fill(SybilTheme.composerGradient)
-                        .opacity(0.98)
-                )
-                .foregroundStyle(SybilTheme.text)
 
-                Button {
-                    submitComposer()
-                } label: {
-                    Image(systemName: viewModel.isSearchMode ? "magnifyingglass" : "arrow.up")
-                        .font(.system(size: 17, weight: .semibold))
-                        .frame(width: 40, height: 40)
-                        .background(
-                            Circle()
-                                .fill(
-                                    viewModel.canSendComposer
-                                        ? AnyShapeStyle(SybilTheme.primaryGradient)
-                                        : AnyShapeStyle(SybilTheme.surface)
-                                )
-                        )
-                        .foregroundStyle(viewModel.canSendComposer ? SybilTheme.text : SybilTheme.textMuted)
-                }
-                .buttonStyle(.plain)
-                .disabled(!viewModel.canSendComposer)
+                SybilPromptSendButton(
+                    systemImage: viewModel.isSearchMode ? "magnifyingglass" : "arrow.up",
+                    canSend: viewModel.canSendComposer,
+                    submit: submitComposer
+                )
+                .accessibilityLabel(viewModel.isSearchMode ? "Search" : "Send message")
             }
         }
         .padding(.horizontal, 14)
@@ -1348,173 +1346,6 @@ private struct SybilNavigationFadeBackground: View {
     }
 }
 
-private struct SybilWorkspaceCharacterBackdrop: View {
-    var isBusy: Bool
-
-    var body: some View {
-        ZStack(alignment: .topTrailing) {
-            RadialGradient(
-                colors: [
-                    SybilTheme.primary.opacity(0.36),
-                    SybilTheme.primary.opacity(0.13),
-                    Color.clear
-                ],
-                center: .center,
-                startRadius: 10,
-                endRadius: 150
-            )
-            .frame(width: 136, height: 118)
-            .blur(radius: 7)
-            .offset(x: 28, y: -24)
-
-            SybilAnimatedGIFView(resourceName: isBusy ? "character-busy" : "character-idle")
-                .frame(width: 172, height: 172)
-                .opacity(0.92)
-                .mask {
-                    LinearGradient(
-                        colors: [
-                            Color.black,
-                            Color.black,
-                            Color.black.opacity(0)
-                        ],
-                        startPoint: .top,
-                        endPoint: .bottom
-                    )
-                }
-                .mask {
-                    LinearGradient(
-                        colors: [
-                            Color.clear,
-                            Color.black.opacity(0.98),
-                            Color.black
-                        ],
-                        startPoint: .leading,
-                        endPoint: .trailing
-                    )
-                }
-                .offset(x: 8, y: 4)
-        }
-        .frame(maxWidth: .infinity, minHeight: 156, maxHeight: 156, alignment: .topTrailing)
-        .clipped()
-        .ignoresSafeArea(edges: .top)
-        .accessibilityHidden(true)
-    }
-}
-
-private struct SybilAnimatedGIFView: UIViewRepresentable {
-    var resourceName: String
-
-    func makeCoordinator() -> Coordinator {
-        Coordinator()
-    }
-
-    func makeUIView(context: Context) -> UIImageView {
-        let imageView = SybilAnimatedUIImageView()
-        imageView.backgroundColor = .clear
-        imageView.contentMode = .scaleAspectFill
-        imageView.clipsToBounds = false
-        imageView.isOpaque = false
-        imageView.setContentHuggingPriority(.defaultLow, for: .horizontal)
-        imageView.setContentHuggingPriority(.defaultLow, for: .vertical)
-        imageView.setContentCompressionResistancePriority(.defaultLow, for: .horizontal)
-        imageView.setContentCompressionResistancePriority(.defaultLow, for: .vertical)
-        return imageView
-    }
-
-    func updateUIView(_ imageView: UIImageView, context: Context) {
-        guard context.coordinator.resourceName != resourceName else {
-            if !imageView.isAnimating {
-                imageView.startAnimating()
-            }
-            return
-        }
-
-        context.coordinator.resourceName = resourceName
-        imageView.image = SybilAnimatedGIFCache.image(named: resourceName)
-        imageView.startAnimating()
-    }
-
-    final class Coordinator {
-        var resourceName: String?
-    }
-}
-
-private final class SybilAnimatedUIImageView: UIImageView {
-    override var intrinsicContentSize: CGSize {
-        CGSize(width: UIView.noIntrinsicMetric, height: UIView.noIntrinsicMetric)
-    }
-}
-
-@MainActor
-private enum SybilAnimatedGIFCache {
-    private static var images: [String: UIImage] = [:]
-
-    static func image(named name: String) -> UIImage? {
-        if let cached = images[name] {
-            return cached
-        }
-
-        guard let image = loadImage(named: name) else {
-            return nil
-        }
-
-        images[name] = image
-
-        return image
-    }
-
-    private static func loadImage(named name: String) -> UIImage? {
-        let url = Bundle.main.url(forResource: name, withExtension: "gif", subdirectory: "Character") ??
-            Bundle.main.url(forResource: name, withExtension: "gif")
-
-        guard let url,
-              let data = try? Data(contentsOf: url),
-              let source = CGImageSourceCreateWithData(data as CFData, nil)
-        else {
-            return nil
-        }
-
-        let frameCount = CGImageSourceGetCount(source)
-        guard frameCount > 0 else {
-            return nil
-        }
-
-        var frames: [UIImage] = []
-        frames.reserveCapacity(frameCount)
-        var duration: TimeInterval = 0
-
-        for index in 0 ..< frameCount {
-            guard let cgImage = CGImageSourceCreateImageAtIndex(source, index, nil) else {
-                continue
-            }
-            frames.append(UIImage(cgImage: cgImage))
-            duration += frameDuration(at: index, source: source)
-        }
-
-        guard !frames.isEmpty else {
-            return nil
-        }
-
-        if frames.count == 1 {
-            return frames[0]
-        }
-
-        return UIImage.animatedImage(with: frames, duration: max(duration, 0.1))
-    }
-
-    private static func frameDuration(at index: Int, source: CGImageSource) -> TimeInterval {
-        guard let properties = CGImageSourceCopyPropertiesAtIndex(source, index, nil) as? [CFString: Any],
-              let gifProperties = properties[kCGImagePropertyGIFDictionary] as? [CFString: Any]
-        else {
-            return 0.08
-        }
-
-        let unclampedDelay = gifProperties[kCGImagePropertyGIFUnclampedDelayTime] as? TimeInterval
-        let delay = unclampedDelay ?? gifProperties[kCGImagePropertyGIFDelayTime] as? TimeInterval ?? 0.08
-        return max(delay, 0.03)
-    }
-}
-
 private struct NewChatSwipeBackdrop: View {
     var progress: CGFloat
     var hasLatched: Bool

ios/Packages/Sybil/Tests/SybilTests/AgentTests.swift

@@ -0,0 +1,327 @@
+import Foundation
+import Testing
+@testable import Sybil
+
+private func agentMessage(_ id: String, entry: String, role: MessageRole = .assistant, info: [String: JSONValue] = [:]) -> Message {
+    Message(id: id, createdAt: Date(timeIntervalSince1970: 1_700_000_000), role: role, content: "Message \(id)", metadata: .object(["kind": .string(role == .tool ? "tool_call" : "message"), "entryId": .string(entry)].merging(info, uniquingKeysWith: { _, b in b })))
+}
+private func agentPage(_ messages: [Message], cursor: String, more: Bool = false) -> AgentPage {
+    AgentPage(conversationId: "root", messages: messages, cursor: cursor, beforeCursor: nil, hasMore: more)
+}
+
+private final class MemoryOutbox: @unchecked Sendable {
+    private let lock = NSLock()
+    private var rows: [AgentOutboxRow] = []
+    var storage: AgentOutboxStorage {
+        AgentOutboxStorage(read: { self.read() }, write: { self.save($0) })
+    }
+    func read() -> [AgentOutboxRow] { lock.withLock { rows } }
+    func save(_ value: [AgentOutboxRow]) { lock.withLock { rows = value } }
+}
+private actor ControlledAgentSocket: AgentSocket {
+    let stream: AsyncThrowingStream<Data, any Error>
+    let continuation: AsyncThrowingStream<Data, any Error>.Continuation
+    private(set) var closed = false
+    init() { (stream, continuation) = AsyncThrowingStream.makeStream() }
+    func receive() async throws -> Data {
+        var iterator = stream.makeAsyncIterator()
+        guard let data = try await iterator.next() else { throw CancellationError() }
+        return data
+    }
+    func send(_ text: String) {}
+    func close() { closed = true; continuation.finish(throwing: URLError(.networkConnectionLost)) }
+    func closeCode() -> Int { 0 }
+    func push(_ text: String) { continuation.yield(Data(text.utf8)) }
+}
+private actor TestAgentClient: AgentAPIClienting {
+    var pages: [String: AgentPage]
+    var requests: [String] = []
+    var submissions: [(String, String)] = []
+    var read = "0"
+    var failRead = false
+    var failSubmit = false
+    var malformedReceipt = false
+    var socketAfters: [String] = []
+    var sockets: [ControlledAgentSocket] = []
+    var holdRead = false
+    var heldRead: CheckedContinuation<AgentReadState, Never>?
+    var beforeSubmit: (@Sendable () -> Bool)?
+    init(pages: [String: AgentPage] = [:]) { self.pages = pages }
+    func agentDescriptor() -> AgentDescriptor {
+        AgentDescriptor(id: "root", title: "Sybil Agent", type: "persistent-agent", model: "test", protocol: "pi-room.v1", activity: AgentActivity(), websocket: "/v1/agent/ws")
+    }
+    func agentMessages(after: String?, before: String?) -> AgentPage {
+        let cursor = after ?? "0"; requests.append(cursor)
+        return pages[cursor] ?? agentPage([], cursor: cursor)
+    }
+    func agentReadState() async -> AgentReadState {
+        if holdRead { return await withCheckedContinuation { heldRead = $0 } }
+        return AgentReadState(cursor: read)
+    }
+    func pauseRead() { holdRead = true }
+    func releaseRead() { holdRead = false; heldRead?.resume(returning: AgentReadState(cursor: read)); heldRead = nil }
+    func hasHeldRead() -> Bool { heldRead != nil }
+    func agentMarkRead(cursor: String) throws -> AgentReadState {
+        if failRead { throw AgentFailure(message: "Offline") }
+        read = try AgentCursor.advance(read, cursor)
+        return AgentReadState(cursor: read)
+    }
+    func agentSubmit(content: String, requestID: String) async throws -> AgentReceipt {
+        #expect(beforeSubmit?() ?? true)
+        submissions.append((content, requestID))
+        try await Task.sleep(for: .milliseconds(80))
+        if failSubmit { throw AgentFailure(message: "Lost acknowledgement") }
+        return AgentReceipt(id: 1, status: "queued", messageId: malformedReceipt ? "wrong" : "inbound:sybil:\(requestID)")
+    }
+    func agentSocket(after: String) async -> any AgentSocket {
+        socketAfters.append(after)
+        let socket = ControlledAgentSocket(); sockets.append(socket)
+        await socket.push(#"{"type":"ready","conversationId":"root"}"#)
+        return socket
+    }
+    func connectedAfters() -> [String] { socketAfters }
+    func latestSocket() -> ControlledAgentSocket? { sockets.last }
+    func setFailures(submit: Bool = false, read: Bool = false, malformed: Bool = false) { failSubmit = submit; failRead = read; malformedReceipt = malformed }
+    func requirePersisted(_ check: @escaping @Sendable () -> Bool) { beforeSubmit = check }
+    func submittedIDs() -> [String] { submissions.map(\.1) }
+    func requestedCursors() -> [String] { requests }
+}
+
+@MainActor
+private func eventually(_ predicate: () -> Bool) async -> Bool {
+    for _ in 0..<200 {
+        if predicate() { return true }
+        try? await Task.sleep(for: .milliseconds(10))
+    }
+    return false
+}
+
+struct AgentTests {
+    @Test func cursorAndMalformedFrames() throws {
+        #expect(try AgentCursor.normalize("00012") == "12")
+        #expect(try AgentCursor.advance("9", "10") == "10")
+        #expect(try AgentCursor.advance("9007199254740993", "9007199254740992") == "9007199254740993")
+        #expect(throws: (any Error).self) { try AgentCursor.normalize("-1") }
+        var state = AgentRoomState()
+        try state.apply(agentPage([agentMessage("reply", entry: "10")], cursor: "10"))
+        let original = state
+        #expect(throws: (any Error).self) { try state.apply(agentPage([], cursor: "bad")) }
+        #expect(state == original)
+        #expect(throws: (any Error).self) { try AgentEvent.decode(Data(#"{"type":"messages","conversationId":"root","cursor":"99","beforeCursor":null,"hasMore":false,"messages":[{"id":"broken"}]}"#.utf8)) }
+        #expect(state.cursor == "10")
+        let oversizedDuration = agentMessage("tool", entry: "11", role: .tool, info: ["durationMs": .number(1e100)])
+        #expect(oversizedDuration.toolCallMetadata?.durationMs == nil)
+    }
+
+    @Test func terminalToolAndRestSocketRace() throws {
+        let start = agentMessage("tool:1", entry: "5", role: .tool, info: ["status": .string("initiated"), "toolName": .string("search"), "args": .object(["query": .string("fixture")]), "startedAt": .string("start")])
+        var end = agentMessage("tool:1", entry: "8", role: .tool, info: ["status": .string("failed"), "error": .string("fixture failure"), "resultPreview": .string("detail"), "args": .object([:])])
+        end.createdAt.addTimeInterval(5)
+        var state = AgentRoomState()
+        try state.apply(agentPage([start], cursor: "5"))
+        try state.apply(agentPage([end, agentMessage("reply", entry: "9")], cursor: "9"))
+        try state.apply(agentPage([start], cursor: "5"))
+        #expect(state.cursor == "9")
+        #expect(state.messages.count == 2)
+        let tool = try #require(state.messages.first?.toolCallMetadata)
+        #expect(tool.status == "failed")
+        #expect(tool.error == "fixture failure")
+        #expect(tool.args == .object(["query": .string("fixture")]))
+        #expect(tool.startedAt == "start")
+        #expect(state.messages.first?.createdAt == start.createdAt)
+    }
+
+    @Test func unreadExcludesToolsAndUsers() throws {
+        var state = AgentRoomState()
+        try state.apply(agentPage([agentMessage("a", entry: "9"), agentMessage("b", entry: "10"), agentMessage("u", entry: "11", role: .user), agentMessage("t", entry: "12", role: .tool)], cursor: "12"))
+        state.readCursor = "9"
+        #expect(state.unreadCount == 1)
+        state.readCursor = try AgentCursor.advance(state.readCursor, "12")
+        state.readCursor = try AgentCursor.advance(state.readCursor, "2")
+        #expect(state.unreadCount == 0)
+    }
+
+    @Test func typingReplyAndStaleActivityAreAtomic() throws {
+        var presentation = AgentPresentation(), state = AgentRoomState()
+        let now = Date(timeIntervalSince1970: 100)
+        var busy = AgentActivity(); busy.isThinking = true; busy.backgroundWorkers = 1
+        var after = state; after.activity = busy
+        presentation.receive(.activity("root", busy), before: state, after: after, replay: false, now: now)
+        state = after
+        #expect(!presentation.isTyping(now: now.addingTimeInterval(0.249)))
+        #expect(presentation.isTyping(now: now.addingTimeInterval(0.251)))
+        let reply = agentPage([agentMessage("reply", entry: "1")], cursor: "1")
+        try after.apply(reply)
+        presentation.receive(.messages(reply), before: state, after: after, replay: false, now: now.addingTimeInterval(1))
+        state = after
+        #expect(!presentation.isTyping(now: now.addingTimeInterval(2)))
+        #expect(state.activity.isBusy)
+        presentation.receive(.activity("root", busy), before: state, after: state, replay: false, now: now.addingTimeInterval(3))
+        #expect(!presentation.isTyping(now: now.addingTimeInterval(4)))
+        for name in AgentPresentation.quietTools {
+            busy.tools.append(.init(id: name, name: name, status: "initiated"))
+            presentation.receive(.activity("root", busy), before: state, after: state, replay: false, now: now.addingTimeInterval(5))
+            #expect(!presentation.isTyping(now: now.addingTimeInterval(6)))
+        }
+        busy.tools.append(.init(id: "real", name: "search", status: "initiated"))
+        presentation.receive(.activity("root", busy), before: state, after: state, replay: false, now: now.addingTimeInterval(7))
+        #expect(presentation.isTyping(now: now.addingTimeInterval(7.3)))
+    }
+
+    @Test func socketURLAndAuthentication() throws {
+        let config = APIConfiguration(baseURL: URL(string: "https://sybil.bajor.cloud/api/?accidental=query")!, authToken: " token ")
+        #expect(try config.agentWebSocketURL().absoluteString == "wss://sybil.bajor.cloud/api/v1/agent/ws")
+        let frame = try JSONDecoder().decode(JSONValue.self, from: Data(config.agentAuthFrame(after: "0009").utf8))
+        #expect(frame.objectValue?["token"] == .string("token"))
+        #expect(frame.objectValue?["after"] == .string("9"))
+        let open = APIConfiguration(baseURL: URL(string: "http://localhost:8787")!, authToken: nil)
+        let openFrame = try JSONDecoder().decode(JSONValue.self, from: Data(open.agentAuthFrame(after: "0").utf8))
+        #expect(openFrame.objectValue?["token"] == .null)
+    }
+
+    @Test @MainActor func emptyPagesAndInitialHistoryDoNotAnimate() async throws {
+        let client = TestAgentClient(pages: ["0": agentPage([], cursor: "3", more: true), "3": agentPage([], cursor: "7", more: true), "7": agentPage([agentMessage("reply", entry: "8")], cursor: "8")])
+        let room = SybilAgentRoom(); room.configure(client: client, storage: MemoryOutbox().storage)
+        defer { room.stop() }
+        #expect(await eventually { !room.isLoading })
+        #expect(await client.requestedCursors() == ["0", "3", "7"])
+        #expect(room.snapshot.room.cursor == "8")
+        #expect(room.unreadCount == 1)
+        #expect(room.snapshot.entranceIDs.isEmpty)
+        try room.receive(.messages(agentPage([agentMessage("late", entry: "9")], cursor: "9")))
+        room.flush()
+        #expect(room.snapshot.entranceIDs == ["late"])
+        try room.receive(.messages(agentPage([agentMessage("late", entry: "9")], cursor: "9")))
+        room.flush()
+        #expect(room.snapshot.entranceIDs.isEmpty)
+        #expect(room.messages.count == 2)
+    }
+
+    @Test @MainActor func outboxPersistsBeforePostAndEchoWinsFailure() async throws {
+        let memory = MemoryOutbox(), client = TestAgentClient()
+        await client.setFailures(submit: true)
+        await client.requirePersisted { !memory.read().isEmpty }
+        let room = SybilAgentRoom(); room.configure(client: client, storage: memory.storage)
+        defer { room.stop() }
+        #expect(await eventually { !room.isLoading })
+        #expect(room.send("identical")); #expect(room.send("identical"))
+        #expect(room.messages.count == 2)
+        let first = room.outbox[0], second = room.outbox[1]
+        #expect(first.requestID != second.requestID)
+        #expect(memory.read().count == 2)
+        try room.receive(.messages(agentPage([agentMessage(first.id, entry: "1", role: .user)], cursor: "1")))
+        room.flush()
+        #expect(await eventually { room.outbox[1].status == "failed" })
+        #expect(room.outbox[0].status == "complete")
+        #expect(room.messages.first(where: { $0.id == first.id })?.createdAt == first.createdAt)
+        room.retry(second.id)
+        #expect(await eventually { room.outbox[1].status == "failed" })
+        let ids = await client.submittedIDs()
+        #expect(ids.filter { $0 == second.requestID }.count == 2)
+        let restored = SybilAgentRoom(); restored.configure(client: client, storage: memory.storage)
+        defer { restored.stop() }
+        #expect(restored.outbox.count == 1)
+        #expect(restored.outbox[0].requestID == second.requestID)
+        #expect(restored.outbox[0].createdAt == second.createdAt)
+    }
+
+    @Test @MainActor func malformedReceiptAndStorageFailureRemainRetryable() async {
+        let client = TestAgentClient(); await client.setFailures(malformed: true)
+        let room = SybilAgentRoom(); room.configure(client: client, storage: MemoryOutbox().storage)
+        defer { room.stop() }
+        #expect(await eventually { !room.isLoading })
+        room.send("fixture")
+        #expect(await eventually { room.outbox.first?.status == "failed" })
+        #expect(room.outbox.first?.seenOnServer == false)
+        let failed = SybilAgentRoom()
+        failed.configure(client: client, storage: AgentOutboxStorage(read: { [] }, write: { _ in throw AgentFailure(message: "Disk full") }))
+        defer { failed.stop() }
+        failed.send("must not submit")
+        #expect(failed.outbox.first?.status == "failed")
+        #expect(await client.submittedIDs().count == 1)
+    }
+
+    @Test @MainActor func readOnlyAfterAcknowledgementAndCapturedSnapshot() async throws {
+        let client = TestAgentClient(pages: ["0": agentPage([agentMessage("a", entry: "1")], cursor: "1")])
+        let room = SybilAgentRoom(); room.configure(client: client, storage: MemoryOutbox().storage)
+        defer { room.stop() }
+        #expect(await eventually { !room.isLoading })
+        room.isSelected = true
+        await client.setFailures(read: true)
+        room.markRead(renderedCursor: "1")
+        #expect(await eventually { room.error != nil })
+        #expect(room.snapshot.room.readCursor == "0")
+        await client.setFailures()
+        try room.receive(.messages(agentPage([agentMessage("b", entry: "2")], cursor: "2")))
+        room.markRead(renderedCursor: "1")
+        #expect(await eventually { room.snapshot.room.readCursor == "1" })
+        #expect(room.unreadCount == 1)
+        room.isObscured = true; room.markRead(renderedCursor: "2")
+        try? await Task.sleep(for: .milliseconds(100))
+        #expect(room.snapshot.room.readCursor == "1")
+        try room.applyRead(AgentReadState(cursor: "2")); room.flush()
+        #expect(room.unreadCount == 0)
+        room.setActive(false); room.setActive(true)
+        #expect(await eventually { room.connection == "Connecting" })
+        #expect(room.messages.count == 2)
+    }
+    @Test @MainActor func interruptedAdmissionForegroundKeepsRetryIdentity() async throws {
+        let client = TestAgentClient(), memory = MemoryOutbox()
+        let room = SybilAgentRoom(); room.configure(client: client, storage: memory.storage)
+        defer { room.stop() }
+        #expect(await eventually { room.connection == "Connected" })
+        room.send("interrupted fixture")
+        let original = try #require(room.outbox.first)
+        room.setActive(false)
+        #expect(room.outbox.first?.status == "failed")
+        #expect(memory.read().first?.requestID == original.requestID)
+        room.setActive(true)
+        #expect(await eventually { room.connection == "Connected" })
+        room.retry(original.id)
+        #expect(await eventually { room.outbox.first?.seenOnServer == true })
+        #expect(room.outbox.first?.createdAt == original.createdAt)
+        #expect(await client.submittedIDs().allSatisfy { $0 == original.requestID })
+    }
+
+    @Test @MainActor func disconnectCannotRestartOldClientAndStaleReadCannotPoisonNewRun() async throws {
+        let old = TestAgentClient(); await old.pauseRead()
+        let room = SybilAgentRoom(); room.configure(client: old, storage: MemoryOutbox().storage)
+        for _ in 0..<100 {
+            if await old.hasHeldRead() { break }
+            try? await Task.sleep(for: .milliseconds(10))
+        }
+        room.disconnect()
+        room.setActive(false); room.setActive(true)
+        #expect(await old.connectedAfters().isEmpty)
+        let next = TestAgentClient(pages: ["0": agentPage([agentMessage("new", entry: "2")], cursor: "2")])
+        room.configure(client: next, storage: MemoryOutbox().storage)
+        defer { room.stop() }
+        #expect(await eventually { room.connection == "Connected" })
+        await old.releaseRead()
+        try? await Task.sleep(for: .milliseconds(50))
+        try await room.catchUp(client: next)
+        #expect(room.messages.map(\.id) == ["new"])
+        #expect(await next.requestedCursors() == ["0", "2"])
+        #expect(await old.connectedAfters().isEmpty)
+    }
+
+    @Test @MainActor func socketReplayAfterForegroundDeduplicatesAndRetainsMergedCursor() async throws {
+        let client = TestAgentClient()
+        let room = SybilAgentRoom(); room.configure(client: client, storage: MemoryOutbox().storage)
+        defer { room.stop() }
+        #expect(await eventually { room.connection == "Connected" })
+        let socket = try #require(await client.latestSocket())
+        let frame = #"{"type":"messages","conversationId":"root","messages":[{"id":"late","createdAt":"2026-10-14T18:00:00Z","role":"assistant","content":"Late reply","metadata":{"kind":"message","entryId":7}}],"cursor":"7","beforeCursor":null,"hasMore":false}"#
+        await socket.push(frame)
+        #expect(await eventually { room.snapshot.room.cursor == "7" })
+        room.setActive(false); room.setActive(true)
+        #expect(await eventually { room.connection == "Connected" })
+        let next = try #require(await client.latestSocket())
+        await next.push(frame)
+        try? await Task.sleep(for: .milliseconds(150))
+        #expect(await client.connectedAfters() == ["0", "7"])
+        #expect(room.messages.count == 1)
+        #expect(await socket.closed)
+    }
+
+}

ios/README.md

@@ -0,0 +1,149 @@
+# Sybil for iOS and Mac Catalyst
+
+Run `just run` for the iPhone simulator, or `just run-mac` for Mac Catalyst.
+`just build-mac` builds without launching the Mac app. Both regenerate the Xcode
+project from `project.yml`. The Mac recipes use ad-hoc signing for local development;
+release/archive signing settings are unchanged.
+
+The conversation sidebar includes **Search chats and web searches** on iPhone,
+iPad, and Mac. It searches all saved history by title, conversation text, and saved
+web-search query/answer, with single-line excerpts that bold matching words and
+truncate in the middle. Search waits briefly while typing,
+cancels superseded requests, and leaves the current conversation open. Clear the
+field to return to recent items. Searches use word prefixes (for example, `plan`
+finds `planning`); if more than 100 items match, refine the query.
+
+## Sybil Agent
+
+The permanent **Sybil Agent** sidebar item opens the persistent room on both
+phone and iPad. It uses the configured Sybil API URL (including an explicit
+`/api` path) and token. The room stays subscribed while other chats are selected;
+foreground and network recovery replay from the last merged cursor. Shared unread
+counts include only public assistant replies and are acknowledged after the
+visible transcript reaches the bottom and its entrance animation finishes.
+
+Multiple messages can be queued while the Agent works. The local outbox is saved
+atomically under Application Support before submission and is scoped to the API
+base URL. An interrupted send stays on its original bubble with **Retry**, retaining
+its request ID and timestamp across relaunch. A canonical server echo confirms it
+even if the HTTP acknowledgement is lost. Tool cards expand to show input, result,
+and error details. Attachments and model selection belong to ordinary chats.
+
+Agent contracts remain in `../docs/api/rest.md` and
+`../docs/api/streaming-chat.md`; this client requires no backend deployment.
+
+### Agent regression and Simulator checks
+
+Run these commands from this worktree's `ios/` directory:
+
+```sh
+just build
+just test
+just run
+just screenshot build/agent-evidence/simulator.png
+```
+
+`just test` runs reducer, cursor, outbox, presentation, shared-read and lifecycle
+regressions along with the existing app tests through Xcode on iPhone 16e.
+For interactive UI coverage, start the local fixture (Python with `aiohttp`):
+
+```sh
+python3 Tests/Fixtures/agent_server.py
+```
+
+In a second terminal, run:
+
+```sh
+just generate
+xcodebuild test -scheme SybilAgentUI \
+  -destination 'platform=iOS Simulator,name=iPhone 16e,OS=latest' \
+  -parallel-testing-enabled NO \
+  -resultBundlePath build/agent-evidence/phone-ui.xcresult
+```
+
+The fixture binds only to `127.0.0.1:8799`, uses `fixture-token`, and never contacts
+external services. UI tests exercise the real REST/WebSocket client and retain
+screenshots in the result bundle. They cover queued submissions, uncertain
+admission/retry, tools, asynchronous replies, inactive sidebar unread counts,
+shared reads, reconnect/malformed replay, foreground recovery, long-history scroll
+anchors and multiline keyboard layout. The launch arguments select the fixture
+through ordinary API settings; no fixture mode is compiled into the application.
+Use a fresh result-bundle path for each run. Check the same UI on iPad and with
+Simulator Reduce Motion enabled before release.
+
+## Quick Question on macOS
+
+While the Catalyst app is running, **Option+Space** toggles a floating Quick
+Question panel, including over other applications. The prompt is focused as soon
+as it opens, with any existing prompt selected so typing replaces it. The same
+select-on-open behavior applies to Quick Question on iPhone and iPad.
+**Return** submits; **Shift+Return** inserts a new line. **Escape**, a
+second **Option+Space**, the close button, or clicking outside dismisses it.
+Hiding the panel preserves the draft and answer and lets an in-flight answer finish.
+**Open in Chat** saves the question and answer in the regular workspace.
+
+Quick Question defaults to OpenAI `chat-latest` on iPhone, iPad, and Mac. Its saved
+provider/model selection takes precedence and is separate from regular chat preferences.
+
+The regular Mac chat/search composer uses the same Return/Shift+Return rules.
+Holding Return does not repeatedly submit. On iOS, Return submits from Quick
+Question, while the normal chat/search composer continues to use Return for a
+new line.
+
+Closing the last Mac window leaves Sybil running, so Option+Space still works.
+Use **Sybil → Quit Sybil** or **Command+Q** to quit the application. The Catalyst
+build opts out of automatic termination, and the native shortcut service holds a
+matching automatic-termination activity until it stops. No hidden keep-alive
+window or sleep-preventing activity is used. Mac scene support also allows the
+workspace to reopen after its last window is closed.
+
+When Catalyst is backgrounded, opening Quick Question first requests app
+activation and waits for UIKit to enter the foreground before creating or ordering
+the native panel. Presentation is deferred out of the hotkey event callback.
+Pressing Option+Space again cancels a pending presentation. Dismissing a panel
+that required activation returns focus to the previous app; explicitly opening
+the workspace keeps focus in Sybil.
+
+On Mac, **Command+N** starts a new chat in the focused workspace, or opens a new
+workspace window when none are open. **Command+Shift+N** always opens a new window,
+including while signed out. **Command+Option+N** starts a new search; iOS retains
+its existing Command+Shift+N search shortcut.
+
+The shortcut is also shown under **File → Quick Question**. If another application
+already owns Option+Space, the panel reports that the shortcut is unavailable;
+the menu item can still be clicked. No Accessibility or Input Monitoring grant is
+needed for the registered hotkey.
+
+The native AppKit panel and Carbon hotkey live in `Apps/SybilMacQuickQuestion`,
+a macOS bundle embedded only for the Mac Catalyst destination. Catalyst loads it
+through the Foundation-only `SybilMacQuickQuestionBridge` contract. Requests,
+provider/model preferences, streaming, and conversion to chat remain in the
+existing `SybilViewModel`; there is no separate backend or API contract. The panel
+uses AppKit controls rather than trying to load the macOS SwiftUI runtime into a
+Catalyst process.
+
+The iPhone/iPad app keeps its existing Quick Question sheet and Home Screen quick
+action, and does not build or embed the macOS helper.
+
+## Verification
+
+- `just test`: existing Swift package tests on the iPhone simulator.
+- From `Packages/Sybil`, run `xcodebuild test -scheme Sybil -destination 'platform=macOS,variant=Mac Catalyst' -parallel-testing-enabled NO` for package tests including the Catalyst panel controller.
+- For a manual Mac check, open the panel from another app, type immediately,
+  submit, hide/reopen during streaming, and verify Escape/outside-click dismissal
+  returns input to the previous app. Check a second display and a full-screen Space
+  when available.
+- Close every Mac window, invoke Option+Space, and verify the saved question and
+  answer remain available. Reopen the workspace from the panel, then explicitly
+  quit with Command+Q. Rebuild and relaunch after changing lifecycle plist keys.
+- In both Mac and iOS Quick Question, reopen an existing prompt and type to replace
+  it. On iOS, verify Return submits Quick Question while Return in the normal
+  prompt adds a newline. In the Mac workspace, verify Return submits and
+  Shift+Return adds a newline.
+- Check Command+N both with a workspace open and after closing every workspace.
+  Command+Shift+N must open another window even when one already exists, and
+  Command+Option+N must start a search without opening a window.
+- With all workspace windows closed, switch to another app and repeatedly open
+  and dismiss Quick Question, including two quick Option+Space presses while it
+  is opening. Check that no late panel appears after cancellation and that no
+  `UINSAppLifecycleStateRunningNoOpenWindows` assertion occurs.

ios/Tests/Fixtures/agent_server.py

@@ -0,0 +1,177 @@
+#!/usr/bin/env python3
+"""Local-only deterministic Agent contract fixture. Never contacts external services.
+Run with Python + aiohttp, then xcodebuild test -scheme SybilAgentUI.
+"""
+import asyncio
+import datetime
+import json
+from aiohttp import web
+
+sockets = set()
+messages = []
+requests = {}
+read_cursor = 0
+cursor = 0
+busy = False
+attempts = {}
+log = []
+reply_tasks = set()
+
+def stamp():
+    return datetime.datetime.now(datetime.timezone.utc).isoformat(timespec='milliseconds').replace('+00:00', 'Z')
+
+def activity():
+    return dict(state='thinking' if busy else 'idle', isThinking=busy, queued=0,
+                backgroundWorkers=0, tools=[], queuedMessages=[])
+
+def message(content, role='assistant', id=None, info=None):
+    global cursor
+    cursor += 1
+    row = dict(id=id or f'outbound:{cursor}', role=role, content=content, createdAt=stamp(),
+               metadata=dict(kind='message', entryId=cursor, status='complete', **(info or {})))
+    messages.append(row)
+    return row
+
+def page(rows, position=None, more=False):
+    return dict(conversationId='fixture-root', messages=rows, cursor=str(cursor if position is None else position), beforeCursor='1', hasMore=more)
+
+async def broadcast(frame):
+    for socket in list(sockets):
+        try:
+            await socket.send_json(frame)
+        except ConnectionError:
+            sockets.discard(socket)
+
+async def emit(content):
+    row = message(content)
+    await broadcast(dict(type='messages', **page([row])))
+    return row
+
+async def reply(content):
+    global busy, cursor
+    await asyncio.sleep(1.2)
+    tool = message('Inspecting the fixture', 'tool', id=f'tool:{cursor + 1}')
+    tool['metadata'].update(kind='tool_call', toolCallId=tool['id'], toolName='fixture_search', status='initiated', args={'query': content}, startedAt=stamp())
+    await broadcast(dict(type='messages', **page([tool])))
+    await asyncio.sleep(0.8)
+    cursor += 1
+    tool = dict(tool, metadata=dict(tool['metadata'], entryId=cursor, status='completed', resultPreview='Fixture result: 3 matches'))
+    messages[messages.index(next(m for m in messages if m['id'] == tool['id']))] = tool
+    await broadcast(dict(type='messages', **page([tool])))
+    failed = message('Fixture unavailable', 'tool', id=f'tool:{cursor + 1}')
+    failed['metadata'].update(kind='tool_call', toolName='fixture_error', status='failed', args={'source': 'local only'}, error='Simulated tool failure')
+    await broadcast(dict(type='messages', **page([failed])))
+    await emit(f'**Finished:** {content}\n\nThe local fixture returned three matches.')
+    # Deliberately stale busy frame after the public reply.
+    await broadcast(dict(type='activity', conversationId='fixture-root', activity=activity()))
+    await asyncio.sleep(0.5)
+    busy = False
+    await broadcast(dict(type='activity', conversationId='fixture-root', activity=activity()))
+
+async def route(request):
+    global read_cursor, cursor, busy
+    path = request.path
+    log.append(dict(method=request.method, path=path, query=dict(request.query)))
+    if path == '/control':
+        body = await request.json()
+        action = body.get('action')
+        if action == 'reset':
+            for task in reply_tasks:
+                task.cancel()
+            reply_tasks.clear()
+            messages.clear(); requests.clear(); attempts.clear(); cursor = 0; read_cursor = 0; busy = False
+            for i in range(body.get('count', 3)):
+                message(f'History message {i + 1}: local simulator fixture.')
+        elif action == 'reply':
+            await emit(body.get('content', 'Late asynchronous fixture reply'))
+        elif action == 'long-tool':
+            row = message('Local tool output', 'tool')
+            row['metadata'].update(kind='tool_call', toolName='fixture_long', status='completed',
+                                   args={'source': 'local only'}, result='\n'.join(f'Result line {i}' for i in range(80)))
+            await broadcast(dict(type='messages', **page([row])))
+        elif action == 'busy':
+            busy = body.get('value', True)
+            await broadcast(dict(type='activity', conversationId='fixture-root', activity=activity()))
+        elif action == 'read':
+            read_cursor = cursor
+        elif action == 'disconnect':
+            for socket in list(sockets):
+                await socket.close(code=1012)
+        elif action == 'malformed':
+            await broadcast(dict(type='messages', conversationId='fixture-root', messages=[{'id': 'broken'}], cursor='999999', beforeCursor=None, hasMore=False))
+        return web.json_response(dict(cursor=cursor, readCursor=read_cursor, messages=messages, requests=requests, log=log))
+    if path == '/api/v1/auth/session':
+        return web.json_response(dict(authenticated=True, mode='token'))
+    if path == '/api/v1/workspace-items':
+        return web.json_response(dict(items=[], hasMore=False))
+    if path == '/api/v1/chats':
+        return web.json_response(dict(chats=[]))
+    if path == '/api/v1/searches':
+        return web.json_response(dict(searches=[]))
+    if path == '/api/v1/models':
+        return web.json_response(dict(providers={}))
+    if path == '/api/v1/active-runs':
+        return web.json_response(dict(chats=[], searches=[]))
+    if path.startswith('/api/v1/agent') and path != '/api/v1/agent/ws':
+        if request.headers.get('Authorization') != 'Bearer fixture-token':
+            return web.json_response(dict(error='Fixture token missing'), status=401)
+    if path == '/api/v1/agent':
+        return web.json_response(dict(id='fixture-root', title='Sybil Agent', type='persistent-agent', model='fixture', protocol='pi-room.v1', websocket='/v1/agent/ws', activity=activity()))
+    if path == '/api/v1/agent/read-state':
+        if request.method == 'PUT':
+            read_cursor = max(read_cursor, int((await request.json())['cursor']))
+        return web.json_response(dict(cursor=str(read_cursor), updatedAt=stamp()))
+    if path == '/api/v1/agent/messages':
+        if request.method == 'POST':
+            body = await request.json()
+            key, content = body['requestId'], body['content']
+            attempts[key] = attempts.get(key, 0) + 1
+            if content == 'retry fixture' and attempts[key] == 1:
+                return web.json_response(dict(error='Simulated uncertain admission; retry'), status=502)
+            receipt = dict(id=len(requests) + 1, status='queued', messageId=f'inbound:sybil:{key}')
+            if key not in requests:
+                requests[key] = receipt
+                row = message(content, 'user', id=receipt['messageId'])
+                busy = True
+                await broadcast(dict(type='messages', **page([row])))
+                await broadcast(dict(type='activity', conversationId='fixture-root', activity=activity()))
+                task = asyncio.create_task(reply(content))
+                reply_tasks.add(task)
+                task.add_done_callback(reply_tasks.discard)
+            if content == 'echo wins fixture':
+                await asyncio.sleep(0.3)
+                return web.json_response(dict(error='Lost HTTP acknowledgement'), status=502)
+            return web.json_response(requests[key], status=202)
+        after = int(request.query.get('after', 0))
+        rows = [m for m in messages if m['metadata']['entryId'] > after]
+        limit = min(int(request.query.get('limit', 200)), 200)
+        selected = rows[:limit]
+        return web.json_response(page(selected, selected[-1]['metadata']['entryId'] if len(rows) > limit else cursor, len(rows) > limit))
+    if path == '/api/v1/agent/ws':
+        if request.query:
+            return web.Response(status=400)
+        socket = web.WebSocketResponse(); await socket.prepare(request)
+        auth = await socket.receive_json(timeout=5)
+        if auth.get('token') != 'fixture-token' or auth.get('type') != 'auth':
+            await socket.close(code=4401); return socket
+        log.append(dict(authAfter=auth.get('after')))
+        sockets.add(socket)
+        await socket.send_json(dict(type='ready', protocol='pi-room.v1', conversationId='fixture-root'))
+        await socket.send_json(dict(type='messages', **page([m for m in messages if m['metadata']['entryId'] > int(auth.get('after', 0))])))
+        await socket.send_json(dict(type='activity', conversationId='fixture-root', activity=activity()))
+        try:
+            async for frame in socket:
+                if frame.type == web.WSMsgType.TEXT:
+                    data = json.loads(frame.data)
+                    if data.get('type') == 'ping':
+                        await socket.send_json(dict(type='pong'))
+        finally:
+            sockets.discard(socket)
+        return socket
+    return web.json_response(dict(error='No fixture for this endpoint'), status=404)
+
+for i in range(3):
+    message(f'History message {i + 1}: local simulator fixture.')
+app = web.Application()
+app.router.add_route('*', '/{path:.*}', route)
+web.run_app(app, host='127.0.0.1', port=8799)

ios/Tests/SybilUITests/AgentUITests.swift

@@ -0,0 +1,175 @@
+import XCTest
+import UIKit
+
+@MainActor
+final class AgentUITests: XCTestCase {
+    let app = XCUIApplication()
+    override func setUp() async throws {
+        continueAfterFailure = false
+        try await control(["action": "reset", "count": 3])
+        app.launchArguments = ["-sybil.ios.apiBaseURL", "http://127.0.0.1:8799/api", "-sybil.ios.adminToken", "fixture-token"]
+        app.launch()
+        openAgent()
+    }
+    private func openAgent() {
+        if !app.buttons["agent-sidebar"].waitForExistence(timeout: 2) || !app.buttons["agent-sidebar"].isHittable {
+            XCTAssertTrue(app.buttons["Show sidebar"].waitForExistence(timeout: 15))
+            app.buttons["Show sidebar"].tap()
+        }
+        XCTAssertTrue(app.buttons["agent-sidebar"].waitForExistence(timeout: 5))
+        app.buttons["agent-sidebar"].tap()
+        XCTAssertTrue(app.textFields["agent-composer"].waitForExistence(timeout: 5) || app.textViews["agent-composer"].exists)
+    }
+    private var composer: XCUIElement {
+        app.textFields["agent-composer"].exists ? app.textFields["agent-composer"] : app.textViews["agent-composer"]
+    }
+    private func send(_ text: String) {
+        composer.tap(); composer.typeText(text); app.buttons["agent-send"].tap()
+    }
+    private func capture(_ name: String) {
+        let attachment = XCTAttachment(screenshot: XCUIScreen.main.screenshot()); attachment.name = name; attachment.lifetime = .keepAlways
+        add(attachment)
+    }
+    @discardableResult private func control(_ value: [String: Any]) async throws -> [String: Any] {
+        var request = URLRequest(url: URL(string: "http://127.0.0.1:8799/control")!)
+        request.httpMethod = "POST"; request.httpBody = try JSONSerialization.data(withJSONObject: value)
+        request.setValue("application/json", forHTTPHeaderField: "Content-Type")
+        let (data, _) = try await URLSession.shared.data(for: request)
+        return try XCTUnwrap(JSONSerialization.jsonObject(with: data) as? [String: Any])
+    }
+    func testSendQueueToolsAndRetry() async throws {
+        capture("agent-history")
+        send("first fixture")
+        send("second fixture")
+        XCTAssertTrue(app.staticTexts["second fixture"].waitForExistence(timeout: 5))
+        XCTAssertTrue(app.staticTexts.containing(NSPredicate(format: "label CONTAINS %@", "Finished: second fixture")).firstMatch.waitForExistence(timeout: 12))
+        XCTAssertFalse(app.otherElements["agent-typing"].exists)
+        capture("agent-replies-keyboard")
+        app.swipeDown()
+        send("retry fixture")
+        XCTAssertTrue(app.buttons["agent-retry"].waitForExistence(timeout: 5))
+        capture("agent-retry")
+        app.buttons["agent-retry"].tap()
+        XCTAssertTrue(app.staticTexts.containing(NSPredicate(format: "label CONTAINS %@", "Finished: retry fixture")).firstMatch.waitForExistence(timeout: 12))
+        send("echo wins fixture")
+        XCTAssertTrue(app.staticTexts.containing(NSPredicate(format: "label CONTAINS %@", "Finished: echo wins fixture")).firstMatch.waitForExistence(timeout: 12))
+        XCTAssertFalse(app.buttons["agent-retry"].exists)
+        composer.tap(); app.swipeDown()
+        let failures = app.buttons.matching(identifier: "agent-tool-fixture_error")
+        failures.element(boundBy: failures.count - 1).tap()
+        XCTAssertTrue(app.staticTexts["agent-tool-input"].waitForExistence(timeout: 3))
+        XCTAssertTrue(app.staticTexts["agent-tool-error"].exists)
+        capture("agent-tool-error")
+        let state = try await control(["action": "state"])
+        XCTAssertEqual((state["requests"] as? [String: Any])?.count, 4)
+    }
+    func testReduceMotionSend() async throws {
+        try XCTSkipUnless(UIAccessibility.isReduceMotionEnabled, "Run with Simulator Reduce Motion enabled")
+        send("reduce motion fixture")
+        XCTAssertTrue(app.staticTexts.containing(NSPredicate(format: "label CONTAINS %@", "Finished: reduce motion fixture")).firstMatch.waitForExistence(timeout: 12))
+        XCTAssertFalse(app.otherElements["agent-typing"].exists)
+        capture("agent-reduce-motion")
+    }
+
+    func testExpandedToolCollapseToShortHistory() async throws {
+        try await control(["action": "long-tool"])
+        let tool = app.buttons["agent-tool-fixture_long"]
+        XCTAssertTrue(tool.waitForExistence(timeout: 5))
+        tool.tap()
+        XCTAssertTrue(app.staticTexts["agent-tool-result"].waitForExistence(timeout: 3))
+        let scroll = app.scrollViews.firstMatch
+        for _ in 0..<12 {
+            if tool.isHittable { break }
+            scroll.swipeDown()
+        }
+        XCTAssertTrue(tool.isHittable)
+        tool.tap()
+        XCTAssertEqual(tool.value as? String, "Collapsed")
+        try await control(["action": "reply", "content": "Reply after collapsing a tall tool"])
+        XCTAssertTrue(app.staticTexts["Reply after collapsing a tall tool"].waitForExistence(timeout: 5))
+        XCTAssertTrue(app.staticTexts["Reply after collapsing a tall tool"].isHittable)
+        XCTAssertFalse(app.buttons["agent-jump-latest"].exists)
+        capture("agent-collapsed-tool-short-history")
+    }
+
+    func testTabletSidebarAndInactiveChat() async throws {
+        try XCTSkipUnless(UIDevice.current.userInterfaceIdiom == .pad, "iPad layout check")
+        XCUIDevice.shared.orientation = .landscapeLeft
+        XCTAssertTrue(app.buttons["agent-sidebar"].waitForExistence(timeout: 5))
+        XCTAssertTrue(app.buttons["agent-sidebar"].isHittable)
+        app.buttons["New chat"].tap()
+        try await control(["action": "busy", "value": true])
+        try await control(["action": "reply", "content": "iPad inactive chat reply"])
+        let unread = XCTNSPredicateExpectation(predicate: NSPredicate(format: "label CONTAINS %@", "unread replies"), object: app.buttons["agent-sidebar"])
+        await fulfillment(of: [unread], timeout: 5)
+        capture("agent-ipad-sidebar")
+        app.buttons["agent-sidebar"].tap()
+        XCTAssertTrue(app.staticTexts["iPad inactive chat reply"].waitForExistence(timeout: 5))
+        capture("agent-ipad-room-before-send")
+        send("iPad fixture")
+        XCTAssertTrue(app.staticTexts.containing(NSPredicate(format: "label CONTAINS %@", "Finished: iPad fixture")).firstMatch.waitForExistence(timeout: 12))
+        capture("agent-ipad-room")
+        XCUIDevice.shared.orientation = .portrait
+    }
+
+    func testScrollPreservationAndKeyboard() async throws {
+        app.terminate()
+        try await control(["action": "reset", "count": 160])
+        app.launch(); openAgent()
+        let scroll = app.scrollViews.firstMatch
+        XCTAssertTrue(app.staticTexts["History message 160: local simulator fixture."].isHittable)
+        scroll.swipeDown()
+        scroll.swipeDown()
+        XCTAssertTrue(app.buttons["agent-jump-latest"].waitForExistence(timeout: 3))
+        let anchor = try XCTUnwrap(app.staticTexts.matching(NSPredicate(format: "label BEGINSWITH %@", "History message")).allElementsBoundByIndex.first { $0.isHittable })
+        let before = anchor.frame.minY
+        try await control(["action": "reply", "content": "Reply while reading earlier history"])
+        try await Task.sleep(for: .milliseconds(500))
+        XCTAssertEqual(anchor.frame.minY, before, accuracy: 2)
+        capture("agent-reading-anchor")
+        for _ in 0..<25 {
+            if app.buttons["Load earlier messages"].isHittable { break }
+            scroll.swipeDown(velocity: .fast)
+        }
+        XCTAssertTrue(app.buttons["Load earlier messages"].isHittable)
+        let first = app.staticTexts["History message 61: local simulator fixture."]
+        let beforeLoad = first.frame.minY
+        app.buttons["Load earlier messages"].tap()
+        XCTAssertEqual(first.frame.minY, beforeLoad, accuracy: 3)
+        capture("agent-older-history-anchor")
+        app.buttons["agent-jump-latest"].tap()
+        XCTAssertTrue(app.staticTexts["Reply while reading earlier history"].isHittable)
+        composer.tap(); composer.typeText("A multiline draft\nwith a second line\nand a third line")
+        XCTAssertTrue(app.staticTexts["Reply while reading earlier history"].isHittable)
+        capture("agent-multiline-keyboard")
+    }
+
+    func testBackgroundUnreadReadReconciliationAndReplay() async throws {
+        app.buttons["Show sidebar"].tap()
+        try await control(["action": "busy", "value": true])
+        try await control(["action": "reply", "content": "Late asynchronous fixture reply"])
+        let unread = XCTNSPredicateExpectation(predicate: NSPredicate(format: "label CONTAINS %@", "1 unread replies"), object: app.buttons["agent-sidebar"])
+        await fulfillment(of: [unread], timeout: 5)
+        capture("agent-unread-sidebar")
+        let covered = try await control(["action": "state"])
+        XCTAssertLessThan(covered["readCursor"] as? Int ?? 0, covered["cursor"] as? Int ?? 0)
+        try await control(["action": "read"])
+        let gone = NSPredicate(format: "NOT (label CONTAINS %@)", "unread replies")
+        let cleared = XCTNSPredicateExpectation(predicate: gone, object: app.buttons["agent-sidebar"])
+        await fulfillment(of: [cleared], timeout: 20)
+        app.buttons["agent-sidebar"].tap()
+        XCTAssertTrue(app.staticTexts["Late asynchronous fixture reply"].waitForExistence(timeout: 5))
+        try await control(["action": "disconnect"])
+        try await control(["action": "reply", "content": "Reconnect fixture reply"])
+        XCTAssertTrue(app.staticTexts["Reconnect fixture reply"].waitForExistence(timeout: 10))
+        try await control(["action": "malformed"])
+        try await control(["action": "reply", "content": "Recovered malformed frame"])
+        XCTAssertTrue(app.staticTexts["Recovered malformed frame"].waitForExistence(timeout: 12))
+        XCUIDevice.shared.press(.home)
+        try await control(["action": "reply", "content": "Foreground fixture reply"])
+        app.activate()
+        XCTAssertTrue(app.staticTexts["Foreground fixture reply"].waitForExistence(timeout: 12))
+        XCTAssertEqual(app.staticTexts.matching(NSPredicate(format: "label == %@", "Foreground fixture reply")).count, 1)
+        capture("agent-foreground-replay")
+    }
+}

ios/justfile

@@ -9,7 +9,7 @@ generate:
   xcodegen --spec project.yml
 
 build: generate
-  if command -v xcbeautify >/dev/null 2>&1; then \
+  set -o pipefail; if command -v xcbeautify >/dev/null 2>&1; then \
     xcodebuild -scheme Sybil -destination '{{simulator}}' | xcbeautify; \
   else \
     xcodebuild -scheme Sybil -destination '{{simulator}}'; \
@@ -24,6 +24,12 @@ run: generate
   xcrun simctl install booted '{{derived_data}}/Build/Products/Debug-iphonesimulator/Sybil.app'
   xcrun simctl launch booted net.buzzert.sybil2
 
+build-mac: generate
+  xcodebuild -scheme Sybil -destination 'platform=macOS,variant=Mac Catalyst' -derivedDataPath '{{derived_data}}' CODE_SIGN_IDENTITY=- CODE_SIGNING_REQUIRED=NO
+
+run-mac: build-mac
+  open '{{derived_data}}/Build/Products/Debug-maccatalyst/Sybil.app'
+
 beta:
   fastlane ios beta
 

ios/project.yml

@@ -7,3 +7,31 @@ packages:
     path: Packages/Sybil
 include:
   - Apps/Sybil/project.yml
+  - Apps/SybilMacQuickQuestion/project.yml
+
+# Dedicated local-fixture UI suite; ordinary release archives only build SybilApp.
+targets:
+  SybilAgentUITests:
+    type: bundle.ui-testing
+    platform: iOS
+    deploymentTarget: "18.0"
+    sources:
+      - Tests/SybilUITests
+    dependencies:
+      - target: SybilApp
+    settings:
+      base:
+        PRODUCT_BUNDLE_IDENTIFIER: net.buzzert.sybil2.agent-ui-tests
+        GENERATE_INFOPLIST_FILE: YES
+        TEST_TARGET_NAME: SybilApp
+        SWIFT_VERSION: 6.0
+schemes:
+  SybilAgentUI:
+    build:
+      targets:
+        SybilApp: all
+        SybilAgentUITests: [test]
+    test:
+      config: Debug
+      targets:
+        - SybilAgentUITests

scripts/devbox-start.sh

@@ -0,0 +1,90 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+SESSION="${SYBIL_TMUX_SESSION:-sybil-dev}"
+LOG_DIR="$ROOT/.devbox/logs"
+PID_DIR="$ROOT/.devbox/pids"
+
+require_env() {
+  local name="$1"
+  if [[ -z "${!name:-}" ]]; then
+    echo "Missing required environment variable: $name" >&2
+    exit 1
+  fi
+}
+
+require_env OPENAI_API_KEY
+require_env EXA_API_KEY
+
+mkdir -p "$LOG_DIR" "$PID_DIR"
+
+export HOST="${HOST:-0.0.0.0}"
+export PORT="${PORT:-8787}"
+export DATABASE_URL="${DATABASE_URL:-file:$ROOT/server/dev.db}"
+
+echo "Installing server dependencies..."
+npm ci --prefix "$ROOT/server" --no-audit --no-fund
+
+echo "Installing web dependencies..."
+npm ci --prefix "$ROOT/web" --no-audit --no-fund
+
+if command -v tmux >/dev/null 2>&1; then
+  tmux kill-session -t "$SESSION" >/dev/null 2>&1 || true
+  tmux new-session -d -s "$SESSION" -n server -c "$ROOT/server" "sleep infinity"
+  tmux set-environment -t "$SESSION" OPENAI_API_KEY "$OPENAI_API_KEY"
+  tmux set-environment -t "$SESSION" EXA_API_KEY "$EXA_API_KEY"
+  tmux set-environment -t "$SESSION" HOST "$HOST"
+  tmux set-environment -t "$SESSION" PORT "$PORT"
+  tmux set-environment -t "$SESSION" DATABASE_URL "$DATABASE_URL"
+  if [[ -n "${BRIX_CLUSTER:-}" ]]; then
+    tmux set-environment -t "$SESSION" BRIX_CLUSTER "$BRIX_CLUSTER"
+  fi
+  for name in \
+    CHAT_CODEX_TOOL_ENABLED \
+    CHAT_CODEX_REMOTE_HOST \
+    CHAT_CODEX_REMOTE_USER \
+    CHAT_CODEX_REMOTE_PORT \
+    CHAT_CODEX_REMOTE_WORKDIR \
+    CHAT_CODEX_SSH_KEY_PATH \
+    CHAT_CODEX_SSH_PRIVATE_KEY_B64 \
+    CHAT_CODEX_EXEC_TIMEOUT_MS \
+    CHAT_SHELL_TOOL_ENABLED \
+    CHAT_SHELL_EXEC_TIMEOUT_MS; do
+    if [[ -n "${!name:-}" ]]; then
+      tmux set-environment -t "$SESSION" "$name" "${!name}"
+    fi
+  done
+  tmux respawn-pane -k -t "$SESSION:server.0" -c "$ROOT/server" "npm run dev"
+  tmux new-window -t "$SESSION" -n web -c "$ROOT/web" \
+    "npm run dev -- --host 0.0.0.0 --port 5173"
+  echo "Started Sybil in tmux session '$SESSION'."
+  echo "Server: http://127.0.0.1:8787"
+  echo "Web:    http://127.0.0.1:5173"
+  exit 0
+fi
+
+echo "tmux not found; starting background processes with nohup."
+
+if [[ -f "$PID_DIR/server.pid" ]]; then
+  kill "$(cat "$PID_DIR/server.pid")" >/dev/null 2>&1 || true
+fi
+if [[ -f "$PID_DIR/web.pid" ]]; then
+  kill "$(cat "$PID_DIR/web.pid")" >/dev/null 2>&1 || true
+fi
+
+(
+  cd "$ROOT/server"
+  exec env HOST="$HOST" PORT="$PORT" DATABASE_URL="$DATABASE_URL" npm run dev
+) >"$LOG_DIR/server.log" 2>&1 &
+echo "$!" >"$PID_DIR/server.pid"
+
+(
+  cd "$ROOT/web"
+  exec npm run dev -- --host 0.0.0.0 --port 5173
+) >"$LOG_DIR/web.log" 2>&1 &
+echo "$!" >"$PID_DIR/web.pid"
+
+echo "Started Sybil background processes."
+echo "Server log: $LOG_DIR/server.log"
+echo "Web log:    $LOG_DIR/web.log"

scripts/devbox-update.sh

@@ -0,0 +1,357 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+REPO="${REPO:-Sybil-2}"
+WORKLOAD="${WORKLOAD:-sybil-dev}"
+CLUSTER="${CLUSTER:-viper}"
+NAMESPACE="${NAMESPACE:-buzzert}"
+QUOTA="${QUOTA:-research-api}"
+SECRET="${SECRET:-sybil-dev-env}"
+REMOTE_ROOT="${REMOTE_ROOT:-/root/code/$REPO}"
+INGRESS_PORT="${INGRESS_PORT:-5173}"
+CODEX_IMAGE="${CODEX_IMAGE:-openai.azurecr.io/rcall:crow-1b540b5e15f9-2026-01-22-2404-312}"
+CODEX_SERVICE="${CODEX_SERVICE:-sybil-codex-devbox}"
+CODEX_CONTAINER="${CODEX_CONTAINER:-codex-devbox}"
+CODEX_SSH_SECRET="${CODEX_SSH_SECRET:-sybil-codex-ssh}"
+CODEX_SSH_PORT="${CODEX_SSH_PORT:-2222}"
+CODEX_WORKDIR="${CODEX_WORKDIR:-/workspace/sybil-codex}"
+CODEX_SSH_KEY_PATH="${CODEX_SSH_KEY_PATH:-/run/secrets/sybil-codex-ssh/id_ed25519}"
+
+echo "Syncing $REPO to $WORKLOAD on $CLUSTER/$NAMESPACE..."
+brix create \
+  --project-root "$(dirname "$ROOT")" \
+  --path "$REPO=$ROOT" \
+  --repositories "$REPO" \
+  --include-untracked \
+  --reconcile-policy continuous \
+  -- \
+  "cluster=$CLUSTER" \
+  "namespace=$NAMESPACE" \
+  "name=$WORKLOAD" \
+  "pool.spec.quota=$QUOTA" \
+  "pool.spec.size.resource=cpu" \
+  "pool.spec.size.quantity=1" \
+  "pool.spec.priority=low"
+
+if [[ "$REPO" != "openai" ]]; then
+  if kubectl --context "$CLUSTER" -n "$NAMESPACE" get git "$WORKLOAD" -o json 2>/dev/null | jq -e '.spec.repositories.openai?' >/dev/null; then
+    kubectl --context "$CLUSTER" -n "$NAMESPACE" patch git "$WORKLOAD" --type=json -p='[{"op":"remove","path":"/spec/repositories/openai"}]'
+  fi
+fi
+
+if [[ -n "${OPENAI_API_KEY:-}" && -n "${EXA_API_KEY:-}" ]]; then
+  kubectl --context "$CLUSTER" -n "$NAMESPACE" create secret generic "$SECRET" \
+    --from-literal=OPENAI_API_KEY="$OPENAI_API_KEY" \
+    --from-literal=EXA_API_KEY="$EXA_API_KEY" \
+    --dry-run=client -o yaml | kubectl --context "$CLUSTER" -n "$NAMESPACE" apply -f -
+elif [[ -n "${OPENAI_API_KEY:-}" || -n "${EXA_API_KEY:-}" ]]; then
+  echo "Only one API key env var is set; keeping existing Kubernetes secret $SECRET." >&2
+fi
+
+if [[ "$(kubectl --context "$CLUSTER" -n "$NAMESPACE" get secret "$SECRET" -o name 2>/dev/null || true)" == "" ]]; then
+  echo "Secret $SECRET does not exist. Re-run with OPENAI_API_KEY and EXA_API_KEY set." >&2
+  exit 1
+fi
+
+if [[ "$(kubectl --context "$CLUSTER" -n "$NAMESPACE" get secret "$CODEX_SSH_SECRET" -o name 2>/dev/null || true)" == "" ]]; then
+  tmpdir="$(mktemp -d)"
+  trap 'rm -rf "$tmpdir"' EXIT
+  ssh-keygen -t ed25519 -N "" -C "sybil-codex@$WORKLOAD" -f "$tmpdir/id_ed25519" >/dev/null
+  kubectl --context "$CLUSTER" -n "$NAMESPACE" create secret generic "$CODEX_SSH_SECRET" \
+    --from-file=id_ed25519="$tmpdir/id_ed25519" \
+    --from-file=id_ed25519.pub="$tmpdir/id_ed25519.pub" \
+    --dry-run=client -o yaml | kubectl --context "$CLUSTER" -n "$NAMESPACE" apply -f -
+fi
+
+kubectl --context "$CLUSTER" -n "$NAMESPACE" patch secret "$SECRET" --type=merge -p "$(jq -n \
+  --arg host "$CODEX_SERVICE" \
+  --arg port "$CODEX_SSH_PORT" \
+  --arg workdir "$CODEX_WORKDIR" \
+  --arg keyPath "$CODEX_SSH_KEY_PATH" \
+  '{
+    stringData: {
+      CHAT_CODEX_TOOL_ENABLED: "true",
+      CHAT_CODEX_REMOTE_HOST: $host,
+      CHAT_CODEX_REMOTE_USER: "root",
+      CHAT_CODEX_REMOTE_PORT: $port,
+      CHAT_CODEX_REMOTE_WORKDIR: $workdir,
+      CHAT_CODEX_SSH_KEY_PATH: $keyPath,
+      CHAT_CODEX_EXEC_TIMEOUT_MS: "600000",
+      CHAT_SHELL_TOOL_ENABLED: "true"
+    }
+  }')"
+
+ensure_codex_deployment() {
+  kubectl --context "$CLUSTER" -n "$NAMESPACE" apply -f - <<KUBE_MANIFEST
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: $CODEX_SERVICE
+  labels:
+    app.kubernetes.io/name: $CODEX_SERVICE
+    app.kubernetes.io/part-of: sybil-dev
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: $CODEX_SERVICE
+  template:
+    metadata:
+      labels:
+        app.kubernetes.io/name: $CODEX_SERVICE
+        app.kubernetes.io/part-of: sybil-dev
+    spec:
+      tolerations:
+        - key: openai.com/team
+          operator: Exists
+          effect: NoSchedule
+      containers:
+        - name: $CODEX_CONTAINER
+          image: $CODEX_IMAGE
+          imagePullPolicy: IfNotPresent
+          command: ["/bin/bash", "-lc"]
+          args:
+            - |
+              set -euo pipefail
+              export DEBIAN_FRONTEND=noninteractive
+              if ! command -v sshd >/dev/null 2>&1; then
+                apt-get update
+                apt-get install -y --no-install-recommends openssh-server ca-certificates git curl
+                rm -rf /var/lib/apt/lists/*
+              fi
+              if ! command -v codex >/dev/null 2>&1; then
+                npm install -g @openai/codex
+              fi
+              mkdir -p /run/sshd /root/.ssh "$CODEX_WORKDIR"
+              cp /run/secrets/sybil-codex-ssh/id_ed25519.pub /root/.ssh/authorized_keys
+              chmod 700 /root/.ssh
+              chmod 600 /root/.ssh/authorized_keys
+              {
+                if [[ -n "\${OPENAI_API_KEY:-}" ]]; then
+                  printf 'export OPENAI_API_KEY=%q\n' "\$OPENAI_API_KEY"
+                fi
+                if [[ -n "\${EXA_API_KEY:-}" ]]; then
+                  printf 'export EXA_API_KEY=%q\n' "\$EXA_API_KEY"
+                fi
+              } >/root/.sybil-codex-env
+              chmod 600 /root/.sybil-codex-env
+              cat >/usr/local/bin/sybil-codex-ssh-command <<'SSH_COMMAND'
+              #!/usr/bin/env bash
+              set -euo pipefail
+              if [[ -f /root/.sybil-codex-env ]]; then
+                source /root/.sybil-codex-env
+              fi
+              if [[ -n "\${SSH_ORIGINAL_COMMAND:-}" ]]; then
+                exec /bin/bash -lc "\$SSH_ORIGINAL_COMMAND"
+              fi
+              exec /bin/bash -l
+              SSH_COMMAND
+              chmod 700 /usr/local/bin/sybil-codex-ssh-command
+              ssh-keygen -A
+              cat >/tmp/sybil-codex-sshd_config <<SSHD_CONFIG
+              Port $CODEX_SSH_PORT
+              ListenAddress 0.0.0.0
+              PermitRootLogin prohibit-password
+              PasswordAuthentication no
+              PubkeyAuthentication yes
+              AcceptEnv OPENAI_API_KEY EXA_API_KEY
+              AuthorizedKeysFile .ssh/authorized_keys
+              PermitUserEnvironment no
+              ForceCommand /usr/local/bin/sybil-codex-ssh-command
+              AllowTcpForwarding no
+              X11Forwarding no
+              Subsystem sftp internal-sftp
+              PidFile /run/sshd/sybil-codex.pid
+              SSHD_CONFIG
+              exec \$(command -v sshd) -D -e -f /tmp/sybil-codex-sshd_config
+          ports:
+            - name: ssh
+              containerPort: $CODEX_SSH_PORT
+          envFrom:
+            - secretRef:
+                name: $SECRET
+          resources:
+            requests:
+              cpu: "1"
+              memory: 4Gi
+            limits:
+              cpu: "1"
+              memory: 4Gi
+          volumeMounts:
+            - name: sybil-codex-ssh
+              mountPath: /run/secrets/sybil-codex-ssh
+              readOnly: true
+            - name: sybil-codex-workspace
+              mountPath: /workspace
+      volumes:
+        - name: sybil-codex-ssh
+          secret:
+            secretName: $CODEX_SSH_SECRET
+            defaultMode: 384
+        - name: sybil-codex-workspace
+          emptyDir: {}
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: $CODEX_SERVICE
+  labels:
+    app.kubernetes.io/name: $CODEX_SERVICE
+    app.kubernetes.io/part-of: sybil-dev
+spec:
+  type: ClusterIP
+  selector:
+    app.kubernetes.io/name: $CODEX_SERVICE
+  ports:
+    - name: ssh
+      port: $CODEX_SSH_PORT
+      targetPort: ssh
+KUBE_MANIFEST
+
+  kubectl --context "$CLUSTER" -n "$NAMESPACE" rollout status "deployment/$CODEX_SERVICE" --timeout=300s
+}
+
+ensure_codex_deployment
+
+ensure_env_from() {
+  local resource="$1"
+  local name="$2"
+  local jsonpath="$3"
+  local patch_path="$4"
+  local existing_env_from
+
+  existing_env_from="$(kubectl --context "$CLUSTER" -n "$NAMESPACE" get "$resource" "$name" \
+    -o "jsonpath=$jsonpath" 2>/dev/null || true)"
+
+  if [[ "$existing_env_from" == *"$SECRET"* ]]; then
+    return
+  fi
+
+  local patch
+  if [[ -z "$existing_env_from" ]]; then
+    patch="[{\"op\":\"add\",\"path\":\"$patch_path\",\"value\":[{\"secretRef\":{\"name\":\"$SECRET\"}}]}]"
+  else
+    patch="[{\"op\":\"add\",\"path\":\"$patch_path/-\",\"value\":{\"secretRef\":{\"name\":\"$SECRET\"}}}]"
+  fi
+  kubectl --context "$CLUSTER" -n "$NAMESPACE" patch "$resource" "$name" --type=json -p "$patch"
+}
+
+ensure_env_from \
+  workload \
+  "$WORKLOAD" \
+  "{.spec.pools[0].template.template.spec.workers.template.spec.containers[0].envFrom}" \
+  "/spec/pools/0/template/template/spec/workers/template/spec/containers/0/envFrom"
+
+ensure_env_from \
+  pool \
+  "$WORKLOAD" \
+  "{.spec.workers.template.spec.containers[0].envFrom}" \
+  "/spec/workers/template/spec/containers/0/envFrom"
+
+codex_ssh_volume_json="$(jq -n --arg secret "$CODEX_SSH_SECRET" '{
+  name: "sybil-codex-ssh",
+  secret: {
+    secretName: $secret,
+    defaultMode: 384
+  }
+}')"
+codex_main_mount_json="$(jq -n '{name: "sybil-codex-ssh", mountPath: "/run/secrets/sybil-codex-ssh", readOnly: true}')"
+
+patch_pool_codex_mount() {
+  local pool_json containers volumes patch
+  pool_json="$(kubectl --context "$CLUSTER" -n "$NAMESPACE" get pool "$WORKLOAD" -o json)"
+  containers="$(jq \
+    --arg sidecarName "$CODEX_CONTAINER" \
+    --argjson mainMount "$codex_main_mount_json" \
+    '
+    def ensure_mount($mount):
+      .volumeMounts = (((.volumeMounts // []) | map(select(.name != $mount.name))) + [$mount]);
+    (.spec.workers.template.spec.containers // [])
+    | map(if .name == "main" then ensure_mount($mainMount) else . end)
+    | map(select(.name != $sidecarName))
+    ' <<<"$pool_json")"
+  volumes="$(jq \
+    --argjson sshVolume "$codex_ssh_volume_json" \
+    '
+    (.spec.workers.template.spec.volumes // [])
+    | map(select(.name != $sshVolume.name and .name != "sybil-codex-workspace"))
+    + [$sshVolume]
+    ' <<<"$pool_json")"
+  patch="$(jq -n --argjson containers "$containers" --argjson volumes "$volumes" '{
+    spec: {
+      workers: {
+        template: {
+          spec: {
+            containers: $containers,
+            volumes: $volumes
+          }
+        }
+      }
+    }
+  }')"
+  kubectl --context "$CLUSTER" -n "$NAMESPACE" patch pool "$WORKLOAD" --type=merge -p "$patch"
+}
+
+patch_workload_codex_mount() {
+  local workload_json pools patch
+  workload_json="$(kubectl --context "$CLUSTER" -n "$NAMESPACE" get workload "$WORKLOAD" -o json)"
+  pools="$(jq \
+    --arg groupName "$WORKLOAD" \
+    --arg sidecarName "$CODEX_CONTAINER" \
+    --argjson mainMount "$codex_main_mount_json" \
+    --argjson sshVolume "$codex_ssh_volume_json" \
+    '
+    def ensure_mount($mount):
+      .volumeMounts = (((.volumeMounts // []) | map(select(.name != $mount.name))) + [$mount]);
+    .spec.pools
+    | map(
+        if .groupName == $groupName then
+          .template.template.spec.workers.template.spec.containers =
+            ((.template.template.spec.workers.template.spec.containers // [])
+              | map(if .name == "main" then ensure_mount($mainMount) else . end)
+              | map(select(.name != $sidecarName)))
+          | .template.template.spec.workers.template.spec.volumes =
+            ((.template.template.spec.workers.template.spec.volumes // [])
+              | map(select(.name != $sshVolume.name and .name != "sybil-codex-workspace"))
+              + [$sshVolume])
+        else
+          .
+        end
+      )
+    ' <<<"$workload_json")"
+  patch="$(jq -n --argjson pools "$pools" '{spec: {pools: $pools}}')"
+  kubectl --context "$CLUSTER" -n "$NAMESPACE" patch workload "$WORKLOAD" --type=merge -p "$patch"
+}
+
+patch_workload_codex_mount
+patch_pool_codex_mount
+
+echo "Waiting for $WORKLOAD-0 to run the current pool revision..."
+deadline=$((SECONDS + 300))
+while true; do
+  pool_revision="$(kubectl --context "$CLUSTER" -n "$NAMESPACE" get pool "$WORKLOAD" -o jsonpath='{.status.revision}' 2>/dev/null || true)"
+  pod_revision="$(kubectl --context "$CLUSTER" -n "$NAMESPACE" get pod "$WORKLOAD-0" -o jsonpath='{.metadata.labels.brix\.openai\.com/revision}' 2>/dev/null || true)"
+  pod_ready="$(kubectl --context "$CLUSTER" -n "$NAMESPACE" get pod "$WORKLOAD-0" -o jsonpath='{.status.conditions[?(@.type=="Ready")].status}' 2>/dev/null || true)"
+  pod_deletion_timestamp="$(kubectl --context "$CLUSTER" -n "$NAMESPACE" get pod "$WORKLOAD-0" -o jsonpath='{.metadata.deletionTimestamp}' 2>/dev/null || true)"
+
+  if [[ -n "$pool_revision" && "$pod_revision" == "$pool_revision" && "$pod_ready" == "True" && -z "$pod_deletion_timestamp" ]]; then
+    break
+  fi
+
+  if (( SECONDS >= deadline )); then
+    echo "Timed out waiting for current ready pod. pool_revision=$pool_revision pod_revision=$pod_revision pod_ready=$pod_ready deleting=$pod_deletion_timestamp" >&2
+    exit 1
+  fi
+
+  sleep 5
+done
+
+brix run --clusters "$CLUSTER" --namespace "$NAMESPACE" --pools "$WORKLOAD" --dir "$REMOTE_ROOT" -- ./scripts/devbox-start.sh
+
+ssh_check="ssh -n -i '$CODEX_SSH_KEY_PATH' -o BatchMode=yes -o StrictHostKeyChecking=accept-new -o UserKnownHostsFile=/tmp/sybil-codex-known-hosts -p '$CODEX_SSH_PORT' root@'$CODEX_SERVICE' true"
+brix run --clusters "$CLUSTER" --namespace "$NAMESPACE" --pools "$WORKLOAD" --dir "$REMOTE_ROOT" -- /bin/bash -lc "$ssh_check"
+
+url="$(brix ingress create --cluster "$CLUSTER" --namespace "$NAMESPACE" --pod "$WORKLOAD-0" --ports "web:$INGRESS_PORT" --root web --timeout 300 "$WORKLOAD" | tail -1)"
+
+echo "Devbox update complete: $url"

server/README.md

@@ -106,3 +106,27 @@ When `chatId` is provided to completion endpoints, you can send full conversatio
 - User accounts + per-device API keys
 - Postgres support + migrations for prod
 - Attachments + embeddings + semantic search
+
+## Sybil Agent
+
+Configure the backend with:
+
+```dotenv
+SYBIL_AGENT_BASE_URL=http://sybil-agent:8788
+SYBIL_AGENT_API_KEY=<the Sybil Agent service API key>
+```
+
+Use the same `SYBIL_AGENT_API_KEY` value configured on the agent service.
+The base URL defaults to `http://sybil-agent:8788` and has no `/v1` suffix. In Docker,
+attach Sybil's **server** container and Sybil Agent to the shared `agents` network so that
+`sybil-agent` resolves privately. Supply the key through a private server environment
+file; it is never a `VITE_*` variable or browser setting. Keep the existing Sybil
+`ADMIN_TOKEN` authentication for clients.
+
+The Sybil Agent tab uses `/v1/agent`, `/v1/agent/messages`, and `/v1/agent/ws` to proxy the
+same persistent root conversation across devices. Sybil Agent remains the source of
+truth for messages and background replies. Sybil stores only the shared,
+monotonically increasing read cursor in SQLite through `/v1/agent/read-state`;
+its migration is applied automatically at startup. The bundled nginx configuration
+forwards WebSocket upgrades. See `docs/api/rest.md` and
+`docs/api/streaming-chat.md` at the repository root for the client contracts.

server/package-lock.json

@@ -22,10 +22,12 @@
         "openai": "^6.16.0",
         "pino-pretty": "^13.1.3",
         "prisma": "^6.6.0",
+        "ws": "8.22.0",
         "zod": "^4.3.6"
       },
       "devDependencies": {
         "@types/node": "^25.0.10",
+        "@types/ws": "8.18.1",
         "tsx": "^4.21.0",
         "typescript": "^5.9.3"
       }
@@ -876,6 +878,16 @@
         "undici-types": "~7.16.0"
       }
     },
+    "node_modules/@types/ws": {
+      "version": "8.18.1",
+      "resolved": "https://registry.npmjs.org/@types/ws/-/ws-8.18.1.tgz",
+      "integrity": "sha512-ThVF6DCVhA8kUGy+aazFQ4kXQ7E1Ty7A3ypFOe0IcJV8O/M511G99AW24irKrW56Wt44yG9+ij8FaqoBGkuBXg==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@types/node": "*"
+      }
+    },
     "node_modules/abstract-logging": {
       "version": "2.0.1",
       "resolved": "https://registry.npmjs.org/abstract-logging/-/abstract-logging-2.0.1.tgz",
@@ -2241,6 +2253,27 @@
       "integrity": "sha512-l4Sp/DRseor9wL6EvV2+TuQn63dMkPjZ/sp9XkghTEbV9KlPS1xUsZ3u7/IQO4wxtcFB4bgpQPRcR3QCvezPcQ==",
       "license": "ISC"
     },
+    "node_modules/ws": {
+      "version": "8.22.0",
+      "resolved": "https://registry.npmjs.org/ws/-/ws-8.22.0.tgz",
+      "integrity": "sha512-Ydggc987+RO0AnWtZ/7Wq9FtNvcrL1b/RO0ud9mWjUPgDrsAAwQSF51sm2hm1XofbU/4jkpGEsLFsZZxU+1DOg==",
+      "license": "MIT",
+      "engines": {
+        "node": ">=10.0.0"
+      },
+      "peerDependencies": {
+        "bufferutil": "^4.0.1",
+        "utf-8-validate": ">=5.0.2"
+      },
+      "peerDependenciesMeta": {
+        "bufferutil": {
+          "optional": true
+        },
+        "utf-8-validate": {
+          "optional": true
+        }
+      }
+    },
     "node_modules/yaml": {
       "version": "2.8.2",
       "resolved": "https://registry.npmjs.org/yaml/-/yaml-2.8.2.tgz",

server/package.json

@@ -31,10 +31,12 @@
     "openai": "^6.16.0",
     "pino-pretty": "^13.1.3",
     "prisma": "^6.6.0",
+    "ws": "8.22.0",
     "zod": "^4.3.6"
   },
   "devDependencies": {
     "@types/node": "^25.0.10",
+    "@types/ws": "8.18.1",
     "tsx": "^4.21.0",
     "typescript": "^5.9.3"
   }

server/prisma/migrations/20261007000000_add_workspace_full_text_search/migration.sql

@@ -0,0 +1,70 @@
+-- External-content indexes avoid storing a second copy of conversation text.
+-- Keep these virtual tables and triggers in SQL: Prisma does not model FTS5.
+CREATE VIRTUAL TABLE "ChatSearch" USING fts5(
+  title, content='Chat', content_rowid='rowid',
+  tokenize='unicode61 remove_diacritics 2', prefix='1 2 3'
+);
+
+-- Hidden system prompts and tool payloads are not searchable conversation text.
+CREATE VIEW "MessageSearchContent" AS
+  SELECT rowid, content FROM "Message" WHERE role IN ('user', 'assistant');
+CREATE VIRTUAL TABLE "MessageSearch" USING fts5(
+  content, content='MessageSearchContent', content_rowid='rowid',
+  tokenize='unicode61 remove_diacritics 2', prefix='1 2 3'
+);
+
+CREATE VIRTUAL TABLE "SavedSearchSearch" USING fts5(
+  title, query, answerText, content='Search', content_rowid='rowid',
+  tokenize='unicode61 remove_diacritics 2', prefix='1 2 3'
+);
+
+CREATE TRIGGER "ChatSearch_insert" AFTER INSERT ON "Chat" BEGIN
+  INSERT INTO "ChatSearch" (rowid, title) VALUES (new.rowid, new.title);
+END;
+CREATE TRIGGER "ChatSearch_delete" AFTER DELETE ON "Chat" BEGIN
+  INSERT INTO "ChatSearch" ("ChatSearch", rowid, title) VALUES ('delete', old.rowid, old.title);
+END;
+CREATE TRIGGER "ChatSearch_update" AFTER UPDATE OF title ON "Chat"
+WHEN old.title IS NOT new.title BEGIN
+  INSERT INTO "ChatSearch" ("ChatSearch", rowid, title) VALUES ('delete', old.rowid, old.title);
+  INSERT INTO "ChatSearch" (rowid, title) VALUES (new.rowid, new.title);
+END;
+
+CREATE TRIGGER "MessageSearch_insert" AFTER INSERT ON "Message"
+WHEN new.role IN ('user', 'assistant') BEGIN
+  INSERT INTO "MessageSearch" (rowid, content) VALUES (new.rowid, new.content);
+END;
+CREATE TRIGGER "MessageSearch_delete" AFTER DELETE ON "Message"
+WHEN old.role IN ('user', 'assistant') BEGIN
+  INSERT INTO "MessageSearch" ("MessageSearch", rowid, content) VALUES ('delete', old.rowid, old.content);
+END;
+CREATE TRIGGER "MessageSearch_update" AFTER UPDATE OF content, role ON "Message"
+WHEN old.content IS NOT new.content OR old.role IS NOT new.role BEGIN
+  INSERT INTO "MessageSearch" ("MessageSearch", rowid, content)
+    SELECT 'delete', old.rowid, old.content WHERE old.role IN ('user', 'assistant');
+  INSERT INTO "MessageSearch" (rowid, content)
+    SELECT new.rowid, new.content WHERE new.role IN ('user', 'assistant');
+END;
+
+CREATE TRIGGER "SavedSearchSearch_insert" AFTER INSERT ON "Search" BEGIN
+  INSERT INTO "SavedSearchSearch" (rowid, title, query, answerText)
+    VALUES (new.rowid, new.title, new.query, new.answerText);
+END;
+CREATE TRIGGER "SavedSearchSearch_delete" AFTER DELETE ON "Search" BEGIN
+  INSERT INTO "SavedSearchSearch" ("SavedSearchSearch", rowid, title, query, answerText)
+    VALUES ('delete', old.rowid, old.title, old.query, old.answerText);
+END;
+CREATE TRIGGER "SavedSearchSearch_update" AFTER UPDATE OF title, query, answerText ON "Search"
+WHEN old.title IS NOT new.title OR old.query IS NOT new.query OR old.answerText IS NOT new.answerText BEGIN
+  INSERT INTO "SavedSearchSearch" ("SavedSearchSearch", rowid, title, query, answerText)
+    VALUES ('delete', old.rowid, old.title, old.query, old.answerText);
+  INSERT INTO "SavedSearchSearch" (rowid, title, query, answerText)
+    VALUES (new.rowid, new.title, new.query, new.answerText);
+END;
+
+-- Index existing history during migration, then maintain it transactionally.
+INSERT INTO "ChatSearch" ("ChatSearch") VALUES ('rebuild');
+INSERT INTO "MessageSearch" ("MessageSearch") VALUES ('rebuild');
+INSERT INTO "SavedSearchSearch" ("SavedSearchSearch") VALUES ('rebuild');
+
+CREATE INDEX "Chat_updatedAt_idx" ON "Chat"("updatedAt");

server/prisma/migrations/20261014000000_add_agent_read_state/migration.sql

@@ -0,0 +1,5 @@
+CREATE TABLE "AgentReadState" (
+  "id" TEXT NOT NULL PRIMARY KEY,
+  "cursor" TEXT NOT NULL DEFAULT '0',
+  "updatedAt" DATETIME NOT NULL
+);

server/prisma/schema.prisma

@@ -76,6 +76,7 @@ model Chat {
 
   @@index([userId])
   @@index([parentChatId])
+  @@index([updatedAt])
 }
 
 model Message {
@@ -214,3 +215,11 @@ model ProjectItem {
   @@index([chatId])
   @@index([searchId])
 }
+
+// Shared personal read position for the external persistent Sybil Agent room.
+// Messages remain in the agent's durable transcript, not duplicated in Sybil Chat.
+model AgentReadState {
+  id        String @id
+  cursor    String @default("0")
+  updatedAt DateTime @updatedAt
+}

server/src/agent.ts

@@ -0,0 +1,171 @@
+import type { IncomingMessage } from 'node:http';
+import type { Duplex } from 'node:stream';
+import type { FastifyInstance } from 'fastify';
+import { WebSocket, WebSocketServer } from 'ws';
+import { z } from 'zod';
+import { requireAdmin } from './auth.js';
+import { prisma } from './db.js';
+import { env } from './env.js';
+
+const Cursor = z.string().regex(/^\d{1,16}$/).refine(value => Number.isSafeInteger(Number(value)), 'Invalid cursor')
+  .transform(value => String(Number(value)));
+const PageQuery = z.object({ after: Cursor.optional(), before: Cursor.optional(), limit: z.coerce.number().int().min(1).max(200).optional() })
+  .refine(value => !(value.after && value.before), 'Use after or before, not both');
+const MessageBody = z.object({ content: z.string().min(1).max(40_000).refine(value => value.trim().length > 0, 'content is required'), requestId: z.string().min(1).max(256) });
+const ReadBody = z.object({ cursor: Cursor });
+const AuthFrame = z.object({ type: z.literal('auth'), token: z.string().nullable().optional(), after: Cursor.optional() });
+const SubscribeFrame = z.object({ type: z.literal('subscribe'), after: Cursor.optional() });
+const SubmissionReceipt = z.object({ id: z.number().int().positive().max(Number.MAX_SAFE_INTEGER),
+  status: z.enum(['queued', 'placed', 'done', 'unanswered']), messageId: z.string() });
+
+type AgentOptions = { baseUrl?: string; apiKey?: string };
+const parse = <T>(schema: z.ZodType<T>, value: unknown): T => {
+  const result = schema.safeParse(value);
+  if (!result.success) throw Object.assign(new Error(result.error.issues.map(issue => issue.message).join('; ')), { statusCode: 400 });
+  return result.data;
+};
+
+async function readState() {
+  const value = await prisma.agentReadState.findUnique({ where: { id: 'root' } });
+  return { cursor: value?.cursor || '0', updatedAt: value?.updatedAt.toISOString() || null };
+}
+
+/** Sybil Agent owns the conversation; the chat server stores only its shared read cursor. */
+export async function registerAgentRoutes(app: FastifyInstance, options: AgentOptions = {}) {
+  const baseUrl = options.baseUrl || env.SYBIL_AGENT_BASE_URL;
+  const apiKey = options.apiKey ?? env.SYBIL_AGENT_API_KEY;
+  const proxyJson = async (path: string, method = 'GET', body?: unknown) => {
+    if (!apiKey) throw app.httpErrors.serviceUnavailable('Sybil Agent is not configured');
+    let response: Response;
+    try {
+      response = await fetch(new URL(path, baseUrl), { method,
+        headers: { Authorization: `Bearer ${apiKey}`, ...(body === undefined ? {} : { 'Content-Type': 'application/json' }) },
+        ...(body === undefined ? {} : { body: JSON.stringify(body) }), signal: AbortSignal.timeout(15_000),
+      });
+    } catch { throw app.httpErrors.badGateway('Sybil Agent is unavailable'); }
+    let data: unknown;
+    try { data = await response.json(); }
+    catch { throw app.httpErrors.badGateway('Sybil Agent returned invalid JSON'); }
+    return { status: response.status, data };
+  };
+
+  app.get('/v1/agent', async (req, reply) => {
+    reply.header('Cache-Control', 'no-store');
+    requireAdmin(req);
+    const result = await proxyJson('/api/conversations/root');
+    reply.code(result.status);
+    if (result.status !== 200) return result.data;
+    return { ...result.data as object, websocket: '/v1/agent/ws' };
+  });
+  app.get('/v1/agent/messages', async (req, reply) => {
+    reply.header('Cache-Control', 'no-store');
+    requireAdmin(req);
+    const query = parse(PageQuery, req.query);
+    const params = new URLSearchParams();
+    for (const [name, value] of Object.entries(query)) if (value !== undefined) params.set(name, String(value));
+    const result = await proxyJson(`/api/conversations/root/messages?${params}`);
+    reply.code(result.status);
+    return result.data;
+  });
+  app.post('/v1/agent/messages', { bodyLimit: 256 * 1024 }, async (req, reply) => {
+    reply.header('Cache-Control', 'no-store');
+    requireAdmin(req);
+    const body = parse(MessageBody, req.body);
+    const result = await proxyJson('/api/conversations/root/messages', 'POST', { ...body, channel: 'sybil', whenBusy: 'followUp', wait: false });
+    if (result.status >= 200 && result.status < 300) {
+      const receipt = SubmissionReceipt.safeParse(result.data);
+      if (!receipt.success || receipt.data.messageId !== `inbound:sybil:${body.requestId}`) {
+        throw app.httpErrors.badGateway('Sybil Agent returned an invalid message receipt');
+      }
+    }
+    reply.code(result.status);
+    return result.data;
+  });
+  app.get('/v1/agent/read-state', async (req, reply) => {
+    reply.header('Cache-Control', 'no-store');
+    requireAdmin(req);
+    return readState();
+  });
+  app.put('/v1/agent/read-state', async (req, reply) => {
+    reply.header('Cache-Control', 'no-store');
+    requireAdmin(req);
+    const { cursor } = parse(ReadBody, req.body);
+    // One SQLite statement makes concurrent tabs monotonic. Never write an older
+    // cursor over one saved by another connection between a read and an update.
+    await prisma.$executeRaw`INSERT INTO "AgentReadState" ("id", "cursor", "updatedAt") VALUES ('root', ${cursor}, CURRENT_TIMESTAMP)
+      ON CONFLICT("id") DO UPDATE SET
+        "cursor" = CASE WHEN CAST(excluded."cursor" AS INTEGER) > CAST("AgentReadState"."cursor" AS INTEGER) THEN excluded."cursor" ELSE "AgentReadState"."cursor" END,
+        "updatedAt" = CASE WHEN CAST(excluded."cursor" AS INTEGER) > CAST("AgentReadState"."cursor" AS INTEGER) THEN CURRENT_TIMESTAMP ELSE "AgentReadState"."updatedAt" END`;
+    return readState();
+  });
+
+  attachAgentSocket(app, baseUrl, apiKey);
+}
+
+function attachAgentSocket(app: FastifyInstance, baseUrl: string, apiKey?: string) {
+  const sockets = new WebSocketServer({ noServer: true, maxPayload: 16 * 1024 });
+  const cleanup = new Set<() => void>();
+  const upgrade = (request: IncomingMessage, socket: Duplex, head: Buffer) => {
+    let url: URL;
+    try { url = new URL(request.url || '/', 'http://localhost'); }
+    catch { socket.end('HTTP/1.1 400 Bad Request\r\nConnection: close\r\n\r\n'); return; }
+    if (url.pathname !== '/v1/agent/ws') { socket.end('HTTP/1.1 404 Not Found\r\nConnection: close\r\n\r\n'); return; }
+    // Credentials must never enter URLs and access logs.
+    if (url.search) { socket.end('HTTP/1.1 400 Bad Request\r\nConnection: close\r\n\r\n'); return; }
+    sockets.handleUpgrade(request, socket, head, browser => {
+      let upstream: WebSocket | undefined;
+      let authenticated = false, disposed = false, alive = true;
+      const send = (value: unknown) => {
+        if (browser.readyState !== WebSocket.OPEN) return;
+        if (browser.bufferedAmount > 2 * 1024 * 1024) { browser.close(1013, 'Slow consumer; reconnect with cursor'); return; }
+        browser.send(typeof value === 'string' ? value : JSON.stringify(value));
+      };
+      const authTimeout = setTimeout(() => browser.close(4401, 'Authentication required'), 5000);
+      const heartbeat = setInterval(() => { if (!alive) browser.terminate(); else { alive = false; browser.ping(); } }, 25_000);
+      browser.on('pong', () => { alive = true; });
+      const dispose = () => {
+        if (disposed) return;
+        disposed = true; clearTimeout(authTimeout); clearInterval(heartbeat); cleanup.delete(dispose);
+        upstream?.terminate(); browser.terminate();
+      };
+      cleanup.add(dispose);
+      browser.on('close', dispose); browser.on('error', dispose);
+      browser.on('message', (raw, isBinary) => {
+        try {
+          if (isBinary) throw new Error('JSON text frames are required');
+          const frame = JSON.parse(raw.toString());
+          if (!authenticated) {
+            const auth = parse(AuthFrame, frame);
+            if (env.ADMIN_TOKEN && auth.token !== env.ADMIN_TOKEN) { browser.close(4401, 'Invalid token'); return; }
+            if (!apiKey) { browser.close(4503, 'Sybil Agent is not configured'); return; }
+            authenticated = true; clearTimeout(authTimeout);
+            const target = new URL('/api/ws', baseUrl); target.protocol = target.protocol === 'https:' ? 'wss:' : 'ws:';
+            upstream = new WebSocket(target, { headers: { Authorization: `Bearer ${apiKey}` }, handshakeTimeout: 10_000, maxPayload: 16 * 1024 * 1024 });
+            upstream.on('open', () => upstream!.send(JSON.stringify({ type: 'subscribe', ...(auth.after === undefined ? {} : { after: auth.after }) })));
+            upstream.on('message', (data, binary) => {
+              if (binary) { browser.close(1011, 'Invalid Sybil Agent frame'); return; }
+              send(data.toString());
+            });
+            upstream.on('close', () => { if (!disposed) browser.close(1012, 'Sybil Agent disconnected; reconnect with cursor'); });
+            upstream.on('error', () => { send({ type: 'error', error: 'Sybil Agent is unavailable' }); browser.close(1011, 'Sybil Agent is unavailable'); });
+            return;
+          }
+          if (frame.type === 'ping') { send({ type: 'pong' }); return; }
+          const subscribe = parse(SubscribeFrame, frame);
+          if (upstream?.readyState !== WebSocket.OPEN) { send({ type: 'error', error: 'Sybil Agent is still connecting' }); return; }
+          upstream.send(JSON.stringify(subscribe));
+        } catch (error) {
+          if (!authenticated) browser.close(4400, 'Invalid authentication frame');
+          else send({ type: 'error', error: error instanceof Error ? error.message : 'Invalid frame' });
+        }
+      });
+    });
+  };
+  app.server.on('upgrade', upgrade);
+  app.addHook('preClose', async () => {
+    app.server.off('upgrade', upgrade);
+    for (const stop of cleanup) stop();
+    for (const client of sockets.clients) client.terminate();
+    sockets.close();
+  });
+}

server/src/env.ts

@@ -70,6 +70,8 @@ const EnvSchema = z.object({
   HERMES_AGENT_API_BASE_URL: HermesAgentApiBaseUrlSchema,
   HERMES_AGENT_API_KEY: OptionalTrimmedStringSchema,
   HERMES_AGENT_MODEL: OptionalTrimmedStringSchema,
+  SYBIL_AGENT_BASE_URL: z.string().url().default("http://sybil-agent:8788"),
+  SYBIL_AGENT_API_KEY: OptionalTrimmedStringSchema,
   EXA_API_KEY: z.string().optional(),
   BRAVE_SEARCH_API_KEY: OptionalTrimmedStringSchema,
 

server/src/index.ts

@@ -6,6 +6,7 @@ import sensible from "@fastify/sensible";
 import { env } from "./env.js";
 import { ensureDatabaseReady } from "./db-init.js";
 import { startModelCatalogRefreshLoop, warmModelCatalog } from "./llm/model-catalog.js";
+import { registerAgentRoutes } from "./agent.js";
 import { registerRoutes } from "./routes.js";
 
 const app = Fastify({
@@ -26,7 +27,7 @@ const stopModelCatalogRefreshLoop = startModelCatalogRefreshLoop(app.log);
 await app.register(cors, {
   origin: true,
   credentials: true,
-  methods: ["GET", "HEAD", "POST", "DELETE", "OPTIONS"],
+  methods: ["GET", "HEAD", "POST", "PUT", "DELETE", "OPTIONS"],
 });
 
 await app.register(swagger, {
@@ -86,6 +87,7 @@ app.addHook("onClose", async () => {
 });
 
 await registerRoutes(app);
+await registerAgentRoutes(app);
 
 await app.listen({ port: env.PORT, host: env.HOST });
 app.log.info(`listening on http://${env.HOST}:${env.PORT}`);

server/src/llm/provider-adapters.ts

@@ -95,7 +95,7 @@ function isLikelyResponsesApiModel(model: string) {
   if (id.includes("audio") || id.includes("realtime") || id.includes("transcribe") || id.includes("tts")) return false;
   if (id.includes("image") || id.includes("dall-e") || id.includes("sora")) return false;
   if (id.includes("search") || id.includes("computer-use")) return false;
-  return /^(gpt-|o\d|chatgpt-)/.test(id);
+  return id === "chat-latest" || /^(gpt-|o\d|chatgpt-)/.test(id);
 }
 
 function isLikelyGeminiChatModel(model: string) {

server/src/llm/quick-question.ts

@@ -0,0 +1,30 @@
+import type { MultiplexRequest, Provider } from "./types.js";
+
+export const QUICK_QUESTION_SYSTEM_PROMPT =
+  "You are answering a quick question in a one-shot experience intended to replace a quick Google search. " +
+  "Give a succinct, direct, self-contained answer. Do not ask follow-up questions, invite the user to continue, " +
+  "or offer additional help. If the question is ambiguous, make the most reasonable assumption and state it briefly only when needed.";
+
+export type QuickQuestionRequest = {
+  provider: Provider;
+  model: string;
+  question: string;
+  enabledTools?: string[];
+  userLocation?: string;
+  temperature?: number;
+  maxTokens?: number;
+};
+
+export function buildQuickQuestionMultiplexRequest({
+  question,
+  ...request
+}: QuickQuestionRequest): MultiplexRequest {
+  return {
+    ...request,
+    persist: false,
+    messages: [
+      { role: "system", content: QUICK_QUESTION_SYSTEM_PROMPT },
+      { role: "user", content: question },
+    ],
+  };
+}

server/src/routes.ts

@@ -11,10 +11,12 @@ import { runMultiplex } from "./llm/multiplexer.js";
 import { runMultiplexStream, type StreamEvent } from "./llm/streaming.js";
 import { getAvailableChatTools, normalizeEnabledChatTools } from "./llm/chat-tools.js";
 import { getModelCatalogSnapshot } from "./llm/model-catalog.js";
+import { buildQuickQuestionMultiplexRequest } from "./llm/quick-question.js";
 import { openaiClient } from "./llm/providers.js";
 import { serializeProviderFields, toPrismaProvider } from "./llm/provider-ids.js";
 import { exaClient } from "./search/exa.js";
 import { isFreshSearchCacheHit, normalizeSearchQuery } from "./search-cache.js";
+import { searchWorkspace } from "./workspace-search.js";
 import type { ChatAttachment } from "./llm/types.js";
 
 const ProviderSchema = z.enum(["openai", "anthropic", "xai", "gemini", "hermes-agent"]);
@@ -205,6 +207,16 @@ const CompletionStreamBody = z
     }
   });
 
+const QuickQuestionStreamBody = z.object({
+  provider: ProviderSchema,
+  model: z.string().min(1),
+  question: z.string().trim().min(1),
+  enabledTools: EnabledToolsSchema.optional(),
+  userLocation: z.string().trim().min(1).max(200).optional(),
+  temperature: z.number().min(0).max(2).optional(),
+  maxTokens: z.number().int().positive().optional(),
+});
+
 function mergeAttachmentsIntoMetadata(metadata: unknown, attachments?: ChatAttachment[]) {
   if (!attachments?.length) return metadata as any;
   if (!metadata || typeof metadata !== "object" || Array.isArray(metadata)) {
@@ -916,6 +928,37 @@ export async function registerRoutes(app: FastifyInstance) {
 
   app.get("/v1/workspace-items", async (req) => {
     requireAdmin(req);
+    const parsed = z.object({ q: z.string().trim().max(200).optional() }).safeParse(req.query);
+    if (!parsed.success) throw app.httpErrors.badRequest("q must be a string of at most 200 characters");
+    if (parsed.data.q) {
+      const { hits, hasMore } = await searchWorkspace(parsed.data.q);
+      const [chats, searches] = await Promise.all([
+        prisma.chat.findMany({
+          where: { id: { in: hits.filter((hit) => hit.type === "chat").map((hit) => hit.id) } },
+          select: {
+            ...chatSummarySelect,
+            parentChat: { select: { projectItems: starredProjectItemsSelect } },
+          },
+        }),
+        prisma.search.findMany({
+          where: { id: { in: hits.filter((hit) => hit.type === "search").map((hit) => hit.id) } },
+          select: searchSummarySelect,
+        }),
+      ]);
+      const summaries = new Map<string, Record<string, unknown>>([
+        ...chats.map(({ parentChat, ...chat }) => [
+          `chat:${chat.id}`,
+          // Search omits unmatched roots, but sidebar fork stars belong to the root.
+          { ...serializeChatLike(chat), ...serializeStarFields(parentChat ?? chat) },
+        ] as const),
+        ...searches.map((search) => [`search:${search.id}`, serializeSearchLike(search)] as const),
+      ]);
+      const items = hits.flatMap((hit) => {
+        const summary = summaries.get(`${hit.type}:${hit.id}`);
+        return summary ? [{ ...summary, ...hit }] : [];
+      });
+      return { items, hasMore };
+    }
     return { items: await listWorkspaceItems() };
   });
 
@@ -1564,7 +1607,28 @@ export async function registerRoutes(app: FastifyInstance) {
     };
   });
 
-  // Streaming SSE endpoint.
+  // One-shot, non-persistent Quick Question SSE endpoint.
+  app.post("/v1/quick-questions/stream", async (req, reply) => {
+    requireAdmin(req);
+
+    const parsed = QuickQuestionStreamBody.safeParse(req.body);
+    if (!parsed.success) return app.httpErrors.badRequest(parsed.error.message);
+    const body = withRequestUserLocation(parsed.data, req);
+
+    reply.raw.writeHead(200, buildSseHeaders(typeof req.headers.origin === "string" ? req.headers.origin : undefined));
+    reply.raw.flushHeaders();
+
+    for await (const ev of runMultiplexStream(buildQuickQuestionMultiplexRequest(body))) {
+      writeSseEvent(reply, mapChatStreamEvent(ev));
+    }
+
+    if (!reply.raw.destroyed && !reply.raw.writableEnded) {
+      reply.raw.end();
+    }
+    return reply;
+  });
+
+  // General chat completion SSE endpoint.
   app.post("/v1/chat-completions/stream", async (req, reply) => {
     requireAdmin(req);
 

server/src/workspace-search.ts

@@ -0,0 +1,74 @@
+import { Prisma } from "@prisma/client";
+import { prisma } from "./db.js";
+
+export const WORKSPACE_SEARCH_LIMIT = 100;
+
+// Treat input as literal Unicode words, never FTS operators/column expressions.
+// Prefix matching supports type-ahead, including short terms such as "AI".
+export function workspaceSearchExpression(query: string): string | null {
+  const words = query.match(/[\p{L}\p{N}\p{M}\p{Co}]+/gu);
+  return words?.map((word) => `"${word}"*`).join(" AND ") ?? null;
+}
+
+export type WorkspaceSearchHit = {
+  type: "chat" | "search";
+  id: string;
+  searchSnippet: string | null;
+};
+
+export function workspaceSearchSQL(expression: string) {
+  // Batch excerpts after limiting matches. Unary + prevents SQLite from using
+  // the IN clause as repeated rowid lookups (and restarting FTS for every row).
+  // FTS still uses MATCH and the rowid bounds; IN filters that single scan.
+  return Prisma.sql`
+    WITH chat_hits AS (
+      SELECT c.id, NULL AS message_rowid
+      FROM "ChatSearch" JOIN "Chat" c ON c.rowid = "ChatSearch".rowid
+      WHERE "ChatSearch" MATCH ${expression}
+      UNION ALL
+      SELECT m.chatId AS id, MIN(m.rowid) AS message_rowid
+      FROM "MessageSearch" JOIN "Message" m ON m.rowid = "MessageSearch".rowid
+      WHERE "MessageSearch" MATCH ${expression}
+      GROUP BY m.chatId
+    ), matches AS (
+      SELECT 'chat' AS type, c.id, c.updatedAt, MIN(h.message_rowid) AS message_rowid,
+        NULL AS search_rowid
+      FROM chat_hits h JOIN "Chat" c ON c.id = h.id
+      GROUP BY c.id
+      UNION ALL
+      SELECT 'search' AS type, s.id, s.updatedAt, NULL AS message_rowid, s.rowid AS search_rowid
+      FROM "SavedSearchSearch" JOIN "Search" s ON s.rowid = "SavedSearchSearch".rowid
+      WHERE "SavedSearchSearch" MATCH ${expression}
+    ), limited AS MATERIALIZED (
+      SELECT * FROM matches ORDER BY updatedAt DESC, type, id LIMIT ${WORKSPACE_SEARCH_LIMIT + 1}
+    ), message_excerpts AS MATERIALIZED (
+      SELECT rowid, substr(snippet("MessageSearch", 0, '', '', ' … ', 24), 1, 480) AS snippet
+      FROM "MessageSearch"
+      WHERE "MessageSearch" MATCH ${expression}
+        AND rowid >= (SELECT MIN(message_rowid) FROM limited)
+        AND rowid <= (SELECT MAX(message_rowid) FROM limited)
+        AND +rowid IN (SELECT message_rowid FROM limited)
+    ), search_excerpts AS MATERIALIZED (
+      SELECT rowid, substr(snippet("SavedSearchSearch", -1, '', '', ' … ', 24), 1, 480) AS snippet
+      FROM "SavedSearchSearch"
+      WHERE "SavedSearchSearch" MATCH ${expression}
+        AND rowid >= (SELECT MIN(search_rowid) FROM limited)
+        AND rowid <= (SELECT MAX(search_rowid) FROM limited)
+        AND +rowid IN (SELECT search_rowid FROM limited)
+    )
+    SELECT type, id, COALESCE(m.snippet, s.snippet) AS searchSnippet
+    FROM limited
+    LEFT JOIN message_excerpts m ON m.rowid = limited.message_rowid
+    LEFT JOIN search_excerpts s ON s.rowid = limited.search_rowid
+    ORDER BY updatedAt DESC, type, id
+  `;
+}
+
+export async function searchWorkspace(query: string) {
+  const expression = workspaceSearchExpression(query);
+  if (!expression) return { hits: [], hasMore: false };
+  // Match via the inverted indexes first. Hydrate summaries and generate excerpts
+  // only for the bounded result set, never fetch full transcripts into JS.
+  const hits = await prisma.$queryRaw<WorkspaceSearchHit[]>(workspaceSearchSQL(expression));
+  return { hits: hits.slice(0, WORKSPACE_SEARCH_LIMIT), hasMore: hits.length > WORKSPACE_SEARCH_LIMIT };
+}

server/tests/agent.test.ts

@@ -0,0 +1,184 @@
+import assert from 'node:assert/strict';
+import { execFileSync } from 'node:child_process';
+import { once } from 'node:events';
+import { mkdtempSync, rmSync } from 'node:fs';
+import { createServer } from 'node:http';
+import { createConnection } from 'node:net';
+import { tmpdir } from 'node:os';
+import { dirname, join, resolve } from 'node:path';
+import test from 'node:test';
+import { fileURLToPath } from 'node:url';
+import Fastify from 'fastify';
+import sensible from '@fastify/sensible';
+import { WebSocket, WebSocketServer } from 'ws';
+
+const root = resolve(dirname(fileURLToPath(import.meta.url)), '..');
+const directory = mkdtempSync(join(tmpdir(), 'sybil-agent-'));
+process.env.DATABASE_URL = `file:${join(directory, 'agent.db')}`;
+process.env.ADMIN_TOKEN = 'sybil-test-token-0123456789';
+execFileSync(process.execPath, [join(root, 'node_modules/prisma/build/index.js'), 'migrate', 'deploy'], {
+  cwd: root, env: { ...process.env, PRISMA_HIDE_UPDATE_MESSAGE: '1' }, stdio: 'pipe',
+});
+const [{ registerAgentRoutes }, { prisma }, { env }] = await Promise.all([
+  import('../src/agent.js'), import('../src/db.js'), import('../src/env.js'),
+]);
+const auth = { authorization: `Bearer ${process.env.ADMIN_TOKEN}` };
+const privateKey = 'sybil-agent-private-key-for-test';
+const requests: { method?: string; url?: string; body: any }[] = [];
+const accepted = new Map<string, number>();
+let upstreamReply: { status: number; body: string } | undefined;
+let connections = 0;
+const subscriptions: any[] = [];
+const upstream = createServer(async (req, res) => {
+  assert.equal(req.headers.authorization, `Bearer ${privateKey}`);
+  const chunks = [];
+  for await (const chunk of req) chunks.push(chunk);
+  const body = chunks.length ? JSON.parse(Buffer.concat(chunks).toString()) : null;
+  requests.push({ method: req.method, url: req.url, body });
+  res.setHeader('Content-Type', 'application/json');
+  if (upstreamReply) { res.statusCode = upstreamReply.status; res.end(upstreamReply.body); return; }
+  if (req.method === 'POST') {
+    const id = accepted.get(body.requestId) || accepted.size + 1; accepted.set(body.requestId, id);
+    res.statusCode = 202; res.end(JSON.stringify({ id, status: 'queued', messageId: `inbound:sybil:${body.requestId}` }));
+  } else if (req.url?.includes('/messages')) {
+    res.end(JSON.stringify({ conversationId: '1', messages: [{ id: 'outbound:43', role: 'assistant', content: 'Later result' }], cursor: '43', beforeCursor: '43', hasMore: false }));
+  } else res.end(JSON.stringify({ id: '1', type: 'persistent-agent', title: 'Sybil Agent', websocket: '/api/ws', protocol: 'pi-room.v1', activity: { thinking: true } }));
+});
+const sourceSockets = new WebSocketServer({ noServer: true });
+upstream.on('upgrade', (req, socket, head) => {
+  assert.equal(req.headers.authorization, `Bearer ${privateKey}`);
+  assert.equal(req.url, '/api/ws');
+  connections++;
+  sourceSockets.handleUpgrade(req, socket, head, ws => {
+    ws.on('message', raw => {
+      const frame = JSON.parse(raw.toString()); subscriptions.push(frame);
+      ws.send(JSON.stringify({ type: 'ready', protocol: 'pi-room.v1', conversationId: '1' }));
+      ws.send(JSON.stringify({ type: 'messages', conversationId: '1', messages: [{ id: 'outbound:43', role: 'assistant', content: 'Later result' }], cursor: '43', hasMore: false }));
+      ws.send(JSON.stringify({ type: 'activity', conversationId: '1', activity: { thinking: true, responding: true } }));
+    });
+  });
+});
+upstream.listen(0, '127.0.0.1'); await once(upstream, 'listening');
+const upstreamAddress = upstream.address() as { port: number };
+const app = Fastify({ logger: false }); await app.register(sensible);
+await registerAgentRoutes(app, { baseUrl: `http://127.0.0.1:${upstreamAddress.port}`, apiKey: privateKey });
+await app.listen({ port: 0, host: '127.0.0.1' });
+const address = app.server.address() as { port: number };
+const wsUrl = `ws://127.0.0.1:${address.port}/v1/agent/ws`;
+
+test.after(async () => {
+  await app.close();
+  for (const socket of sourceSockets.clients) socket.terminate();
+  sourceSockets.close();
+  await new Promise<void>(resolve => upstream.close(() => resolve()));
+  await prisma.$disconnect();
+  rmSync(directory, { recursive: true, force: true });
+});
+
+async function eventually(check: () => boolean) {
+  for (let i = 0; i < 100; i++) { if (check()) return; await new Promise(resolve => setTimeout(resolve, 10)); }
+  assert.fail('Expected condition was not reached');
+}
+
+test('REST authenticates, proxies the shared room, and retains stable message IDs without revealing Sybil Agent credentials', async () => {
+  assert.equal((await app.inject({ url: '/v1/agent' })).statusCode, 401);
+  const descriptor = await app.inject({ url: '/v1/agent', headers: auth });
+  assert.equal(descriptor.statusCode, 200);
+  assert.equal(descriptor.json().websocket, '/v1/agent/ws');
+  assert.ok(!descriptor.body.includes(privateKey));
+  const page = await app.inject({ url: '/v1/agent/messages?after=42&limit=20', headers: auth });
+  assert.equal(page.json().messages[0].role, 'assistant');
+  assert.match(requests.at(-1)!.url!, /after=42&limit=20/);
+  const input = { content: '  Keep watching\n', requestId: 'same-client-message' };
+  const first = await app.inject({ method: 'POST', url: '/v1/agent/messages', headers: auth, payload: input });
+  const second = await app.inject({ method: 'POST', url: '/v1/agent/messages', headers: auth, payload: input });
+  assert.equal(first.statusCode, 202); assert.deepEqual(first.json(), second.json());
+  assert.deepEqual(requests.at(-1)!.body, { ...input, channel: 'sybil', whenBusy: 'followUp', wait: false });
+  assert.equal((await prisma.chat.count()), 0, 'Sybil Agent messages are not copied into ordinary Sybil chats');
+  assert.equal((await app.inject({ url: '/v1/agent/messages?after=2&before=3', headers: auth })).statusCode, 400);
+  assert.equal((await app.inject({ method: 'POST', url: '/v1/agent/messages', headers: auth, payload: { content: 'Hello' } })).statusCode, 400);
+});
+
+test('the read cursor stays monotonic under concurrent tabs and survives reconnecting to SQLite', async () => {
+  const initial = await app.inject({ url: '/v1/agent/read-state', headers: auth });
+  assert.equal(initial.json().cursor, '0');
+  assert.equal(initial.headers['cache-control'], 'no-store');
+  await Promise.all(['900', '4', '100', '12'].map(cursor => app.inject({ method: 'PUT', url: '/v1/agent/read-state', headers: auth, payload: { cursor } })));
+  assert.equal((await app.inject({ url: '/v1/agent/read-state', headers: auth })).json().cursor, '900');
+  await prisma.$disconnect();
+  assert.equal((await app.inject({ method: 'PUT', url: '/v1/agent/read-state', headers: auth, payload: { cursor: '5' } })).json().cursor, '900');
+  assert.equal((await app.inject({ method: 'PUT', url: '/v1/agent/read-state', headers: auth, payload: { cursor: '-1' } })).statusCode, 400);
+});
+
+test('malformed upstream JSON and unconfirmed message receipts return 502, preserving safe retries', async () => {
+  const payload = { content: 'One durable message', requestId: 'invalid-receipt-retry' };
+  try {
+    upstreamReply = { status: 200, body: '{truncated' };
+    for (const url of ['/v1/agent', '/v1/agent/messages']) {
+      const response = await app.inject({ url, headers: auth });
+      assert.equal(response.statusCode, 502);
+      assert.equal(response.headers['cache-control'], 'no-store');
+    }
+    for (const body of ['{truncated', '{}', 'null', JSON.stringify({ id: 1, status: 'queued', messageId: 'wrong-message' }),
+      JSON.stringify({ id: 1, status: 'unknown', messageId: `inbound:sybil:${payload.requestId}` })]) {
+      upstreamReply = { status: 202, body };
+      const response = await app.inject({ method: 'POST', url: '/v1/agent/messages', headers: auth, payload });
+      assert.equal(response.statusCode, 502, body);
+      assert.equal(response.headers['cache-control'], 'no-store');
+    }
+  } finally { upstreamReply = undefined; }
+  const retry = await app.inject({ method: 'POST', url: '/v1/agent/messages', headers: auth, payload });
+  assert.equal(retry.statusCode, 202);
+  assert.equal(retry.json().messageId, `inbound:sybil:${payload.requestId}`);
+});
+
+test('malformed and unmatched upgrade targets close without crashing or opening Sybil Agent connections', async () => {
+  const before = connections;
+  const requestUpgrade = (target: string): Promise<string> => new Promise((resolve, reject) => {
+    let response = '';
+    const socket = createConnection({ host: '127.0.0.1', port: address.port }, () => {
+      socket.write(`GET ${target} HTTP/1.1\r\nHost: localhost\r\nConnection: Upgrade\r\nUpgrade: websocket\r\nSec-WebSocket-Version: 13\r\nSec-WebSocket-Key: dGhlIHNhbXBsZSBub25jZQ==\r\n\r\n`);
+    });
+    socket.setTimeout(1000, () => socket.destroy(new Error('Upgrade rejection did not close the socket')));
+    socket.on('data', chunk => { response += chunk.toString(); });
+    socket.on('error', reject);
+    socket.on('end', () => { socket.destroy(); resolve(response); });
+  });
+  assert.match(await requestUpgrade('http://['), /^HTTP\/1\.1 400 /);
+  assert.match(await requestUpgrade('/v1/not-an-agent/ws'), /^HTTP\/1\.1 404 /);
+  assert.match(await requestUpgrade('/v1/agent/ws?token=forbidden'), /^HTTP\/1\.1 400 /);
+  assert.equal(connections, before);
+  assert.equal((await app.inject({ url: '/v1/agent', headers: auth })).statusCode, 200);
+});
+
+test('WebSocket first-frame authentication gates the private connection and forwards resumable native frames', async () => {
+  const before = connections;
+  const denied = new WebSocket(wsUrl); await once(denied, 'open');
+  denied.send(JSON.stringify({ type: 'auth', token: 'wrong' }));
+  const [code] = await once(denied, 'close');
+  assert.equal(code, 4401); assert.equal(connections, before);
+  const browser = new WebSocket(wsUrl);
+  const frames: any[] = []; browser.on('message', raw => frames.push(JSON.parse(raw.toString())));
+  await once(browser, 'open'); browser.send(JSON.stringify({ type: 'auth', token: env.ADMIN_TOKEN, after: '42' }));
+  await eventually(() => frames.some(frame => frame.type === 'activity'));
+  assert.deepEqual(subscriptions.at(-1), { type: 'subscribe', after: '42' });
+  assert.deepEqual(frames.map(frame => frame.type), ['ready', 'messages', 'activity']);
+  assert.equal(frames[1].messages[0].content, 'Later result');
+  assert.equal(frames[2].activity.thinking, true);
+  assert.ok(!JSON.stringify(frames).includes(privateKey));
+  browser.send(JSON.stringify({ type: 'subscribe', after: '43' }));
+  await eventually(() => subscriptions.at(-1)?.after === '43');
+  browser.close(); await once(browser, 'close');
+  await eventually(() => sourceSockets.clients.size === 0);
+});
+
+test('open-mode WebSocket authentication accepts null and still keeps the Sybil Agent token server-side', async () => {
+  const saved = env.ADMIN_TOKEN; env.ADMIN_TOKEN = undefined;
+  try {
+    const browser = new WebSocket(wsUrl); const frames: any[] = [];
+    browser.on('message', raw => frames.push(JSON.parse(raw.toString())));
+    await once(browser, 'open'); browser.send(JSON.stringify({ type: 'auth', token: null, after: '0' }));
+    await eventually(() => frames.some(frame => frame.type === 'ready'));
+    browser.close(); await once(browser, 'close');
+  } finally { env.ADMIN_TOKEN = saved; }
+});

server/tests/chat-forks.test.ts

@@ -1,6 +1,6 @@
 import assert from "node:assert/strict";
 import { execFileSync } from "node:child_process";
-import { mkdtempSync, rmSync } from "node:fs";
+import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
 import { tmpdir } from "node:os";
 import { dirname, join, resolve } from "node:path";
 import test from "node:test";
@@ -11,6 +11,7 @@ import sensible from "@fastify/sensible";
 const serverRoot = resolve(dirname(fileURLToPath(import.meta.url)), "..");
 const databaseDir = mkdtempSync(join(tmpdir(), "sybil-chat-forks-"));
 process.env.DATABASE_URL = `file:${join(databaseDir, "test.db")}`;
+writeFileSync(join(databaseDir, "test.db"), "");
 process.env.OPENAI_API_KEY = "";
 delete process.env.ADMIN_TOKEN;
 

server/tests/provider-adapters.test.ts

@@ -1,6 +1,50 @@
 import assert from "node:assert/strict";
 import test from "node:test";
-import { describeProviderChatBackend } from "../src/llm/provider-adapters.js";
+import { env } from "../src/env.js";
+import { describeProviderChatBackend, fetchProviderCatalogModels } from "../src/llm/provider-adapters.js";
+
+for (const { name, listedModels, expectedModels } of [
+  {
+    name: "OpenAI catalog includes chat-latest while excluding non-chat models",
+    listedModels: [
+      "gpt-4.1-mini",
+      "chat-latest",
+      "o3",
+      "chatgpt-4o-latest",
+      "chat-latest",
+      "text-embedding-3-small",
+      "omni-moderation-latest",
+      "gpt-4o-audio-preview",
+      "gpt-realtime",
+      "gpt-4o-transcribe",
+      "gpt-4o-mini-tts",
+      "gpt-image-1",
+      "chatgpt-image-latest",
+      "gpt-4o-search-preview",
+      "computer-use-preview",
+      "sora-2",
+    ],
+    expectedModels: ["chat-latest", "chatgpt-4o-latest", "gpt-4.1-mini", "o3"],
+  },
+  {
+    name: "OpenAI catalog does not invent models absent from the provider response",
+    listedModels: ["gpt-4.1-mini"],
+    expectedModels: ["gpt-4.1-mini"],
+  },
+]) {
+  test(name, async (t) => {
+    const originalKey = env.OPENAI_API_KEY;
+    env.OPENAI_API_KEY = "test-openai-key";
+    t.after(() => { env.OPENAI_API_KEY = originalKey; });
+    t.mock.method(globalThis, "fetch", async (input: RequestInfo | URL) => {
+      const url = input instanceof Request ? input.url : String(input);
+      assert.equal(url, "https://api.openai.com/v1/models");
+      return Response.json({ data: listedModels.map((id) => ({ id })) });
+    });
+
+    assert.deepEqual(await fetchProviderCatalogModels("openai"), expectedModels);
+  });
+}
 
 test("provider backend registry selects chat protocol and managed-tool mode", () => {
   assert.deepEqual(describeProviderChatBackend("openai", []), {

server/tests/quick-question.test.ts

@@ -0,0 +1,27 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import {
+  buildQuickQuestionMultiplexRequest,
+  QUICK_QUESTION_SYSTEM_PROMPT,
+} from "../src/llm/quick-question.js";
+
+test("quick question requests receive the server-owned one-shot system prompt", () => {
+  const request = buildQuickQuestionMultiplexRequest({
+    provider: "openai",
+    model: "gpt-4.1-mini",
+    question: "How do I reset my password?",
+    enabledTools: ["web_search"],
+    userLocation: "Los Angeles, CA",
+  });
+
+  assert.equal(request.persist, false);
+  assert.equal(request.chatId, undefined);
+  assert.deepEqual(request.messages, [
+    { role: "system", content: QUICK_QUESTION_SYSTEM_PROMPT },
+    { role: "user", content: "How do I reset my password?" },
+  ]);
+  assert.deepEqual(request.enabledTools, ["web_search"]);
+  assert.equal(request.userLocation, "Los Angeles, CA");
+  assert.match(QUICK_QUESTION_SYSTEM_PROMPT, /succinct, direct, self-contained answer/i);
+  assert.match(QUICK_QUESTION_SYSTEM_PROMPT, /do not ask follow-up questions/i);
+});

server/tests/workspace-search.test.ts

@@ -0,0 +1,188 @@
+import assert from "node:assert/strict";
+import { execFileSync } from "node:child_process";
+import { cpSync, mkdtempSync, readdirSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { dirname, join, resolve } from "node:path";
+import test from "node:test";
+import { fileURLToPath } from "node:url";
+import Fastify from "fastify";
+import sensible from "@fastify/sensible";
+import { Prisma } from "@prisma/client";
+
+const serverRoot = resolve(dirname(fileURLToPath(import.meta.url)), "..");
+const databaseDir = mkdtempSync(join(tmpdir(), "sybil-workspace-search-"));
+const migration = "20261007000000_add_workspace_full_text_search";
+process.env.DATABASE_URL = `file:${join(databaseDir, "test.db")}`;
+process.env.OPENAI_API_KEY = "";
+delete process.env.ADMIN_TOKEN;
+writeFileSync(join(databaseDir, "test.db"), "");
+cpSync(join(serverRoot, "prisma/schema.prisma"), join(databaseDir, "schema.prisma"));
+for (const name of readdirSync(join(serverRoot, "prisma/migrations"))) {
+  if (name !== migration) cpSync(join(serverRoot, "prisma/migrations", name), join(databaseDir, "migrations", name), { recursive: true });
+}
+function migrate() {
+  execFileSync(process.execPath, [join(serverRoot, "node_modules/prisma/build/index.js"), "migrate", "deploy", "--schema", join(databaseDir, "schema.prisma")], {
+    cwd: serverRoot, env: process.env, stdio: "pipe",
+  });
+}
+migrate();
+const { prisma } = await import("../src/db.js");
+const oldChat = await prisma.chat.create({ data: {
+  title: "Old constellation notes",
+  messages: { create: [
+    { role: "user", content: "We photographed a nebula before the migration." },
+    { role: "system", content: "secretmigrationprompt" },
+  ] },
+} });
+const oldSearch = await prisma.search.create({ data: { query: "constellation map", answerText: "The nebula is visible." } });
+cpSync(join(serverRoot, "prisma/migrations", migration), join(databaseDir, "migrations", migration), { recursive: true });
+migrate();
+
+const { registerRoutes } = await import("../src/routes.js");
+const { workspaceSearchExpression, workspaceSearchSQL } = await import("../src/workspace-search.js");
+const app = Fastify({ logger: false });
+await app.register(sensible);
+await registerRoutes(app);
+await app.ready();
+
+async function search(query: string) {
+  const response = await app.inject({ method: "GET", url: `/v1/workspace-items?q=${encodeURIComponent(query)}` });
+  assert.equal(response.statusCode, 200, response.body);
+  return response.json();
+}
+
+test.afterEach(async () => {
+  await prisma.chat.deleteMany();
+  await prisma.search.deleteMany();
+});
+test.after(async () => {
+  await app.close();
+  await prisma.$disconnect();
+  rmSync(databaseDir, { recursive: true, force: true });
+});
+
+test("migration indexes existing titles, messages, queries, and answers", async () => {
+  assert.deepEqual(new Set((await search("nebula")).items.map((item: any) => item.id)), new Set([oldChat.id, oldSearch.id]));
+  assert.equal((await search("constellation")).items.length, 2);
+  assert.equal((await search("secretmigrationprompt")).items.length, 0);
+});
+
+test("matches full history, deduplicates messages, and excludes hidden content", async () => {
+  const chat = await prisma.chat.create({ data: {
+    title: "Travel notes", updatedAt: new Date("2000-01-01"),
+    messages: { create: [
+      { role: "user", content: "Let's discuss café planning in Zürich" },
+      { role: "assistant", content: "CAFÉ PLANNING includes booking a table" },
+      { role: "system", content: "hiddensystemvalue" },
+      { role: "tool", content: "hiddentoolvalue" },
+    ] },
+  } });
+  await prisma.chat.createMany({ data: Array.from({ length: 105 }, (_, i) => ({ title: `Recent ${i}` })) });
+  const recent = (await app.inject({ method: "GET", url: "/v1/workspace-items" })).json();
+  assert.equal(recent.items.some((item: any) => item.id === chat.id), false);
+  const result = await search("CAFE plan");
+  assert.equal(result.hasMore, false);
+  assert.deepEqual(result.items.map((item: any) => item.id), [chat.id]);
+  assert.match(result.items[0].searchSnippet, /café planning/i);
+  assert.equal(result.items[0].messages, undefined);
+  assert.equal((await search("zuri")).items.length, 1);
+  assert.equal((await search("hiddensystemvalue")).items.length, 0);
+  assert.equal((await search("hiddentoolvalue")).items.length, 0);
+});
+
+test("indexes edits, role changes, forks, rollbacks, and cascading deletes", async () => {
+  const chat = await prisma.chat.create({ data: {
+    title: "originaltitle", messages: { create: { role: "assistant", content: "originalbody" } },
+  }, include: { messages: true } });
+  const messageID = chat.messages[0].id;
+  await prisma.chat.update({ where: { id: chat.id }, data: { title: "replacementtitle" } });
+  await prisma.message.update({ where: { id: messageID }, data: { content: "replacementbody" } });
+  assert.equal((await search("original")).items.length, 0);
+  assert.equal((await search("replacementbody")).items.length, 1);
+  await prisma.message.update({ where: { id: messageID }, data: { role: "tool" } });
+  assert.equal((await search("replacementbody")).items.length, 0);
+  await prisma.message.update({ where: { id: messageID }, data: { role: "user" } });
+  assert.equal((await search("replacementbody")).items.length, 1);
+  await assert.rejects(prisma.$transaction(async (tx) => {
+    await tx.message.update({ where: { id: messageID }, data: { content: "rolledbackvalue" } });
+    throw new Error("rollback");
+  }));
+  assert.equal((await search("rolledbackvalue")).items.length, 0);
+  assert.equal((await search("replacementbody")).items.length, 1);
+  const fork = await app.inject({ method: "POST", url: `/v1/chats/${chat.id}/fork`, payload: {} });
+  assert.equal(fork.statusCode, 200, fork.body);
+  assert.equal((await search("replacementbody")).items.length, 2);
+  await prisma.message.delete({ where: { id: messageID } });
+  const onlyFork = (await search("replacementbody")).items;
+  assert.deepEqual(onlyFork.map((item: any) => item.id), [fork.json().chat.id]);
+  const starResponse = await app.inject({ method: "PATCH", url: `/v1/chats/${chat.id}/star`, payload: { starred: true } });
+  assert.equal(starResponse.statusCode, 200, starResponse.body);
+  const starredFork = (await search("replacementbody")).items[0];
+  assert.equal(starredFork.starred, true);
+  assert.ok(starredFork.starredAt);
+  assert.equal(starredFork.parentChat, undefined);
+  await prisma.chat.delete({ where: { id: chat.id } });
+  assert.equal((await search("replacement")).items.length, 0);
+  for (const table of ["ChatSearch", "MessageSearch", "SavedSearchSearch"]) {
+    await prisma.$executeRawUnsafe(`INSERT INTO "${table}" ("${table}", rank) VALUES ('integrity-check', 1)`);
+  }
+});
+
+test("saved searches match query and answer, and update/delete their index", async () => {
+  const saved = await prisma.search.create({ data: { title: "Bookmarks", query: "AI design", answerText: "Use a nebula illustration" } });
+  assert.equal((await search("ai des")).items[0].type, "search");
+  assert.match((await search("nebula")).items[0].searchSnippet, /nebula/);
+  await prisma.search.update({ where: { id: saved.id }, data: { title: "New title", query: "newquery", answerText: "newanswer" } });
+  assert.equal((await search("nebula")).items.length, 0);
+  assert.equal((await search("design")).items.length, 0);
+  assert.equal((await search("Bookmarks")).items.length, 0);
+  assert.equal((await search("newanswer")).items.length, 1);
+  await prisma.search.delete({ where: { id: saved.id } });
+  assert.equal((await search("new")).items.length, 0);
+});
+
+test("literal queries, validation, blank queries, and bounded snippets", async () => {
+  await prisma.chat.create({ data: { title: "AI OR safety", messages: { create: { role: "user", content: "longword".repeat(1000) } } } });
+  assert.equal((await search('"AI" OR safety*')).items.length, 1);
+  assert.equal((await search("AI OR other")).items.length, 0);
+  assert.equal((await search("AI")).items.length, 1);
+  assert.equal((await search("*** : () --")).items.length, 0);
+  assert.equal((await search("' UNION SELECT * FROM Chat; --")).items.length, 0);
+  assert.equal((await search("longword")).items[0].searchSnippet.length, 480);
+  const recent = (await app.inject({ method: "GET", url: "/v1/workspace-items" })).json();
+  assert.deepEqual(await search("  "), recent);
+  for (const query of ["q=" + "x".repeat(201), "q=one&q=two"]) {
+    const response = await app.inject({ method: "GET", url: `/v1/workspace-items?${query}` });
+    assert.equal(response.statusCode, 400, response.body);
+  }
+});
+
+test("caps combined results at 100 with deterministic newest-first ordering", async () => {
+  const date = new Date("2026-01-01");
+  await prisma.chat.createMany({ data: Array.from({ length: 100 }, (_, i) => ({ id: `chat-${String(i).padStart(3, "0")}`, title: "commonword", updatedAt: date })) });
+  await prisma.search.create({ data: { id: "saved", query: "commonword", updatedAt: new Date("2026-02-01") } });
+  const result = await search("common");
+  assert.equal(result.items.length, 100);
+  assert.equal(result.hasMore, true);
+  assert.equal(result.items[0].id, "saved");
+  assert.equal(result.items[1].id, "chat-000");
+  assert.equal(result.items.at(-1).id, "chat-098");
+});
+
+test("query plan uses all three FTS indexes", async () => {
+  const plan = await prisma.$queryRaw<Array<{ detail: string }>>(Prisma.sql`EXPLAIN QUERY PLAN ${workspaceSearchSQL(workspaceSearchExpression("needle")!)}`);
+  for (const table of ["ChatSearch", "MessageSearch", "SavedSearchSearch"]) {
+    assert.ok(plan.some((row) => row.detail.includes(table) && /VIRTUAL TABLE INDEX.*M\d/.test(row.detail)), plan.map((row) => row.detail).join("\n"));
+  }
+  assert.equal(plan.some((row) => row.detail.includes("CORRELATED SCALAR SUBQUERY")), false, "Excerpts must not restart FTS for each result");
+});
+
+test("vacuum preserves external-content index row identities", async () => {
+  const removed = await prisma.chat.create({ data: { title: "removed" } });
+  const kept = await prisma.chat.create({ data: { title: "kepttitle", messages: { create: { role: "user", content: "keptbody" } } } });
+  await prisma.chat.delete({ where: { id: removed.id } });
+  await prisma.$executeRaw`VACUUM`;
+  assert.equal((await search("kepttitle")).items[0].id, kept.id);
+  assert.equal((await search("kepttitle")).items[0].searchSnippet, null);
+  assert.equal((await search("keptbody")).items[0].searchSnippet, "keptbody");
+});

web/README.md

@@ -34,6 +34,13 @@ Default dev URL: `http://localhost:5173`
 ## UI
 
 - Left panel: mixed list of chat conversations and Exa searches.
+- Sidebar search uses `GET /v1/workspace-items?q=…` to search all saved titles,
+  conversation messages, and web-search queries/answers. It waits 300 ms while
+  typing and cancels superseded requests. Results include a single-line matching excerpt
+  with matching words in bold and middle truncation
+  (or “Title match” for title-only chat matches), with up to 100 newest matches.
+  Clear the field or press Escape to return to recent items. Matching uses word
+  prefixes, so `plan` also finds `planning`.
 - Right panel:
   - Chat mode: transcript + provider/model controls.
   - Search mode: top AI answer block + Google-style Exa results view.

web/src/components/agent/agent-conversation.css

@@ -0,0 +1,538 @@
+.agent-room {
+  height:100%;
+  min-height:0;
+  min-width:0;
+  display:flex;
+  flex-direction:column;
+  background:radial-gradient(ellipse at 80% 0%,hsl(266 80% 30%/.12),transparent 42%),hsl(237 43% 6%/.72);
+  color:hsl(258 36% 96%);
+}
+
+.agent-room__body {
+  position:relative;
+  display:flex;
+  flex:1;
+  min-height:0;
+  min-width:0;
+}
+
+.agent-room__scroll {
+  width:100%;
+  overflow:auto;
+  overscroll-behavior:contain;
+  scrollbar-gutter:stable;
+  outline:none;
+  overflow-anchor:none;
+}
+
+.agent-room__scroll:focus-visible {
+  outline:1px solid hsl(264 80% 74%/.48);
+  outline-offset:-2px;
+}
+
+.agent-room__timeline {
+  display:flex;
+  flex-direction:column;
+  justify-content:flex-end;
+  gap:15px;
+  max-width:848px;
+  min-height:100%;
+  margin:0 auto;
+  padding:24px 30px 22px;
+  box-sizing:border-box;
+}
+
+.agent-room__entry {
+  flex-shrink:0;
+  min-width:0;
+}
+
+.agent-room__day {
+  display:flex;
+  align-items:center;
+  justify-content:center;
+  gap:18px;
+  margin:17px 0 24px;
+  color:hsl(255 19% 64%);
+  font-size:10px;
+  letter-spacing:.05em;
+}
+
+.agent-room__day:before,.agent-room__day:after {
+  content:"";
+  height:1px;
+  width:48px;
+  background:linear-gradient(90deg,transparent,hsl(260 48% 65%/.19));
+}
+
+.agent-room__day:after {
+  transform:rotate(180deg);
+}
+
+.agent-room__message {
+  display:flex;
+  align-items:flex-end;
+  gap:9px;
+  min-width:0;
+  max-width:89%;
+}
+
+.agent-room__message[data-animation="enter"] {
+  animation:agent-room-message-enter 240ms cubic-bezier(.2,.7,.25,1) both;
+}
+
+.agent-room__message.is-own {
+  margin-left:auto;
+  justify-content:flex-end;
+  max-width:78%;
+}
+
+.agent-room__message-avatar {
+  display:grid;
+  place-items:center;
+  flex:none;
+  align-self:flex-end;
+  width:26px;
+  height:26px;
+  margin-bottom:23px;
+  border:1px solid hsl(267 75% 73%/.2);
+  border-radius:9px;
+  background:hsl(261 52% 20%/.56);
+  color:hsl(269 72% 80%);
+  overflow:hidden;
+}
+
+.agent-room__message-avatar img {
+  display:block;
+  width:100%;
+  height:100%;
+  object-fit:cover;
+}
+
+.agent-room__message-content {
+  min-width:0;
+}
+
+.agent-room__plain-message {
+  margin:0;
+  white-space:pre-wrap;
+}
+
+.agent-room__message-meta {
+  display:flex;
+  align-items:center;
+  gap:7px;
+  margin:5px 3px 0;
+  color:hsl(253 18% 55%);
+  font-size:9px;
+  line-height:1.4;
+}
+
+.agent-room__message.is-own .agent-room__message-meta {
+  justify-content:flex-end;
+}
+
+.agent-room__message-meta>time:before {
+  content:"·";
+  margin-right:7px;
+  opacity:.6;
+}
+
+.agent-room__delivery {
+  display:inline-flex;
+  align-items:center;
+  gap:3px;
+  color:hsl(272 52% 77%);
+  white-space:nowrap;
+}
+
+.agent-room__delivery.is-error {
+  color:hsl(349 70% 78%);
+}
+
+.agent-room__delivery button {
+  border:0;
+  padding:0 0 0 3px;
+  background:transparent;
+  color:inherit;
+  font:inherit;
+  text-decoration:underline;
+  cursor:pointer;
+}
+
+.agent-room .md-content p {
+  margin-top:.55em;
+  margin-bottom:.55em;
+}
+
+.agent-room .md-content>:first-child {
+  margin-top:0;
+}
+
+.agent-room .md-content>:last-child {
+  margin-bottom:0;
+}
+
+.agent-room .md-content pre {
+  max-width:100%;
+  font-size:12px;
+}
+
+.agent-room .md-content h1 {
+  font-size:1.25em;
+}
+
+.agent-room .md-content h2 {
+  font-size:1.15em;
+}
+
+.agent-room .md-content h3 {
+  font-size:1.05em;
+}
+
+.agent-room .md-content a {
+  color:hsl(189 79% 75%);
+}
+
+.agent-room__tool {
+  margin:0 0 0 35px;
+  max-width:calc(89% - 35px);
+  border:1px solid hsl(256 35% 47%/.23);
+  border-radius:11px;
+  background:hsl(246 32% 12%/.56);
+  color:hsl(255 26% 75%);
+  font-size:11px;
+}
+
+.agent-room__tool summary {
+  display:flex;
+  align-items:center;
+  gap:9px;
+  min-height:37px;
+  padding:8px 11px;
+  cursor:pointer;
+  list-style:none;
+}
+
+.agent-room__tool summary::-webkit-details-marker {
+  display:none;
+}
+
+.agent-room__tool-icon {
+  display:grid;
+  place-items:center;
+  color:hsl(267 69% 76%);
+}
+
+.agent-room__tool-title {
+  min-width:0;
+  flex:1;
+  overflow:hidden;
+  text-overflow:ellipsis;
+  white-space:nowrap;
+  text-transform:capitalize;
+}
+
+.agent-room__tool-state {
+  color:hsl(254 18% 51%);
+  font-size:9px;
+}
+
+.agent-room__tool.is-running .agent-room__tool-state {
+  color:hsl(184 54% 72%);
+}
+
+.agent-room__tool.is-failed .agent-room__tool-state {
+  color:hsl(351 71% 76%);
+}
+
+.agent-room__tool-chevron {
+  color:hsl(255 20% 52%);
+  transition:transform .15s;
+}
+
+.agent-room__tool[open] .agent-room__tool-chevron {
+  transform:rotate(180deg);
+}
+
+.agent-room__tool-body {
+  padding:0 11px 11px;
+}
+
+.agent-room__tool-body pre {
+  max-height:210px;
+  overflow:auto;
+  white-space:pre-wrap;
+  overflow-wrap:anywhere;
+  font:10px/1.6 ui-monospace,monospace;
+  margin:8px 0 0;
+  padding:9px;
+  border-radius:6px;
+  background:hsl(239 42% 6%/.7);
+}
+
+.agent-room__typing,.agent-room__waiting {
+  display:flex;
+  align-items:center;
+  gap:9px;
+  flex:none;
+  min-height:20px;
+  margin-left:36px;
+  color:hsl(254 22% 63%);
+  font-size:11px;
+}
+
+.agent-room__typing-dots {
+  display:flex;
+  align-items:center;
+  gap:3px;
+  padding:6px 8px;
+  border-radius:9px;
+  background:hsl(258 43% 22%/.42);
+}
+
+.agent-room__typing-dots i {
+  width:3px;
+  height:3px;
+  border-radius:50%;
+  background:hsl(268 72% 78%);
+  animation:agent-room-dot 1.4s ease-in-out infinite;
+}
+
+.agent-room__typing-dots i:nth-child(2) {
+  animation-delay:.16s;
+}
+
+.agent-room__typing-dots i:nth-child(3) {
+  animation-delay:.32s;
+}
+
+.agent-room__worker-count {
+  padding:2px 5px;
+  border-radius:5px;
+  background:hsl(265 58% 27%/.3);
+  color:hsl(268 58% 76%);
+  font-size:9px;
+}
+
+.agent-room__jump {
+  position:absolute;
+  z-index:2;
+  left:50%;
+  bottom:14px;
+  transform:translateX(-50%);
+  display:flex;
+  align-items:center;
+  gap:7px;
+  white-space:nowrap;
+  padding:9px 14px;
+  border:1px solid hsl(266 76% 72%/.43);
+  border-radius:99px;
+  background:hsl(256 46% 20%/.96);
+  color:hsl(263 60% 93%);
+  box-shadow:0 5px 24px hsl(240 70% 2%/.5);
+  font-size:11px;
+  cursor:pointer;
+  backdrop-filter:blur(14px);
+}
+
+.agent-room__jump:hover {
+  background:hsl(261 48% 27%);
+}
+
+.agent-room__composer-area {
+  flex:none;
+  border-top:1px solid hsl(253 40% 40%/.14);
+  padding:16px 30px max(13px,var(--safe-area-bottom));
+  background:linear-gradient(180deg,hsl(240 41% 7%/.72),hsl(240 43% 7%/.97));
+}
+
+.agent-room__composer-inner {
+  max-width:788px;
+  margin:0 auto;
+}
+
+.agent-room__composer {
+  display:flex;
+  align-items:flex-end;
+  gap:10px;
+  padding:7px 8px 7px 5px;
+  border:1px solid hsl(263 51% 58%/.3);
+  border-radius:17px;
+  background:hsl(247 35% 13%/.9);
+  box-shadow:inset 0 1px 0 hsl(266 80% 92%/.04),0 6px 24px hsl(239 80% 2%/.12);
+  transition:border-color .15s,box-shadow .15s;
+}
+
+.agent-room__composer:focus-within {
+  border-color:hsl(265 74% 69%/.65);
+  box-shadow:0 0 0 3px hsl(266 69% 55%/.08);
+}
+
+.agent-room__composer textarea {
+  display:block;
+  flex:1;
+  min-width:0;
+  min-height:46px;
+  max-height:156px;
+  margin:0;
+  resize:none;
+  overflow:auto;
+  border:0;
+  outline:0;
+  background:transparent;
+  padding:13px 11px;
+  font-size:14px;
+  line-height:20px;
+  color:hsl(259 44% 96%);
+  box-shadow:none;
+}
+
+.agent-room__composer textarea::placeholder {
+  color:hsl(255 23% 53%);
+}
+
+.agent-room__send {
+  display:grid;
+  place-items:center;
+  flex:none;
+  width:38px;
+  height:38px;
+  margin-bottom:4px;
+  border:1px solid hsl(265 86% 79%/.32);
+  border-radius:11px;
+  background:linear-gradient(140deg,hsl(264 80% 59%),hsl(270 75% 49%));
+  color:hsl(265 65% 98%);
+  box-shadow:0 3px 12px hsl(264 75% 45%/.22);
+  cursor:pointer;
+  transition:filter .15s,opacity .15s;
+}
+
+.agent-room__send:hover:not(:disabled) {
+  filter:brightness(1.17);
+}
+
+.agent-room__send:disabled {
+  opacity:.34;
+  box-shadow:none;
+  cursor:default;
+}
+
+.agent-room__error {
+  display:flex;
+  align-items:center;
+  gap:9px;
+  margin-bottom:11px;
+  padding:10px 12px;
+  border:1px solid hsl(350 62% 62%/.2);
+  border-radius:11px;
+  background:hsl(340 45% 22%/.22);
+  color:hsl(349 70% 85%);
+  font-size:11px;
+}
+
+.agent-room__error>span {
+  min-width:0;
+  flex:1;
+}
+
+.agent-room__error>svg {
+  flex:none;
+}
+
+.agent-room__error button {
+  flex:none;
+  border:1px solid hsl(350 57% 66%/.28);
+  border-radius:6px;
+  padding:4px 8px;
+  color:hsl(349 66% 90%);
+  font-size:10px;
+  cursor:pointer;
+}
+
+.agent-room__empty {
+  flex:none;
+  margin:auto;
+  max-width:360px;
+  padding:48px 15px;
+  text-align:center;
+  color:hsl(254 25% 65%);
+}
+
+.agent-room button:focus-visible,.agent-room summary:focus-visible {
+  outline:2px solid hsl(188 82% 71%);
+  outline-offset:3px;
+}
+
+@keyframes agent-room-dot {
+  0%,70%,100% {
+    opacity:.35;
+    transform:translateY(0);
+  }
+  35% {
+    opacity:1;
+    transform:translateY(-2px);
+  }
+}
+
+@keyframes agent-room-message-enter {
+  from {
+    opacity:0;
+    transform:translateY(10px);
+  }
+  to {
+    opacity:1;
+    transform:translateY(0);
+  }
+}
+
+@media(max-width:767px) {
+  .agent-room__timeline {
+    padding:17px 14px 20px;
+    gap:14px;
+  }
+  .agent-room__message {
+    max-width:96%;
+    gap:7px;
+  }
+  .agent-room__message.is-own {
+    max-width:87%;
+  }
+  .agent-room__message-avatar {
+    width:23px;
+    height:23px;
+    border-radius:8px;
+  }
+  .agent-room__tool {
+    margin-left:30px;
+    max-width:calc(96% - 30px);
+  }
+  .agent-room__typing,.agent-room__waiting {
+    margin-left:30px;
+    font-size:10px;
+  }
+  .agent-room__composer-area {
+    padding:12px 13px max(11px,var(--safe-area-bottom));
+  }
+  .agent-room__composer textarea {
+    font-size:16px;
+    line-height:22px;
+  }
+  .agent-room__empty {
+    max-width:300px;
+  }
+}
+
+@media(prefers-reduced-motion:reduce) {
+  .agent-room__message[data-animation="enter"] {
+    animation:none;
+  }
+  .agent-room__typing-dots i {
+    animation:none;
+    opacity:.7;
+  }
+  .agent-room * {
+    scroll-behavior:auto!important;
+    transition:none!important;
+  }
+}

web/src/components/agent/agent-conversation.tsx

@@ -0,0 +1,350 @@
+import { useEffect, useLayoutEffect, useMemo, useRef, useState } from "preact/hooks";
+import { ArrowDown, ChevronDown, Clock3, LoaderCircle, Menu, RefreshCw, SendHorizontal, WifiOff, Wrench } from "lucide-preact";
+import type { Message } from "@/lib/api";
+import type { AgentRoomHookResult } from "@/hooks/use-agent-room";
+import { MarkdownContent } from "@/components/markdown/markdown-content";
+import { Button } from "@/components/ui/button";
+import "./agent-conversation.css";
+
+type Props = {
+  room: AgentRoomHookResult;
+  onOpenSidebar: () => void;
+  /** False while a mobile navigation overlay covers this pane. */
+  isVisible?: boolean;
+};
+
+type ReadingAnchor = { id: string; offset: number };
+const DRAFT_KEY = "sybil.pi-root.draft";
+const timeFormat = new Intl.DateTimeFormat(undefined, { hour: "numeric", minute: "2-digit" });
+const dayFormat = new Intl.DateTimeFormat(undefined, { weekday: "short", month: "short", day: "numeric" });
+
+function savedDraft() {
+  try { return sessionStorage.getItem(DRAFT_KEY) || ""; } catch { return ""; }
+}
+function saveDraft(value: string) {
+  try { if (value) sessionStorage.setItem(DRAFT_KEY, value); else sessionStorage.removeItem(DRAFT_KEY); } catch { /* A blocked browser store does not prevent chatting. */ }
+}
+function metadata(message: Message): Record<string, unknown> {
+  return message.metadata && typeof message.metadata === "object" && !Array.isArray(message.metadata)
+    ? message.metadata as Record<string, unknown> : {};
+}
+function dateOf(value: string) {
+  const date = new Date(value);
+  return Number.isNaN(date.getTime()) || date.getFullYear() < 2000 ? null : date;
+}
+function dayKey(date: Date | null) {
+  return date ? `${date.getFullYear()}-${date.getMonth()}-${date.getDate()}` : "older";
+}
+function dayLabel(date: Date | null) {
+  if (!date) return "Earlier";
+  const today = new Date();
+  if (dayKey(date) === dayKey(today)) return "Today";
+  today.setDate(today.getDate() - 1);
+  return dayKey(date) === dayKey(today) ? "Yesterday" : dayFormat.format(date);
+}
+function toolName(value: unknown) {
+  return typeof value === "string" ? value.replace(/_/g, " ") : "Tool";
+}
+
+function AgentToolCard({ message }: { message: Message }) {
+  const info = metadata(message);
+  const active = info.status === "initiated";
+  const failed = info.status === "failed";
+  const title = toolName(info.toolName || message.name);
+  const preview = typeof info.resultPreview === "string" ? info.resultPreview : typeof info.error === "string" ? info.error : "";
+  return (
+    <details className={`agent-room__tool${active ? " is-running" : ""}${failed ? " is-failed" : ""}`} data-agent-message-id={message.id}>
+      <summary>
+        <span className="agent-room__tool-icon"><Wrench size={14} aria-hidden="true" /></span>
+        <span className="agent-room__tool-title">{title}</span>
+        <span className="agent-room__tool-state">{active ? "Running" : failed ? "Failed" : "Done"}</span>
+        <ChevronDown className="agent-room__tool-chevron" size={14} aria-hidden="true" />
+      </summary>
+      <div className="agent-room__tool-body">
+        {info.args && typeof info.args === "object" ? <pre aria-label="Tool input">{JSON.stringify(info.args, null, 2)}</pre> : null}
+        {preview ? <pre aria-label={failed ? "Tool error" : "Tool result"}>{preview}</pre> : null}
+        {!info.args && !preview ? <p>{active ? "This tool is still running." : "No additional details."}</p> : null}
+      </div>
+    </details>
+  );
+}
+
+function AgentBubble({ message, animate, onRetry }: { message: Message; animate: boolean; onRetry: (id: string) => void }) {
+  const own = message.role === "user";
+  const info = metadata(message);
+  const status = typeof info.status === "string" ? info.status : "complete";
+  const failed = status === "failed" || status === "unconfirmed";
+  const date = dateOf(message.createdAt);
+  return (
+    <article className={`agent-room__message ${own ? "is-own" : "is-agent"}`} data-agent-message-id={message.id} data-message-status={status} data-animation={animate ? "enter" : undefined} aria-label={own ? "Your message" : "Sybil's message"}>
+      {!own ? <span className="agent-room__message-avatar" aria-hidden="true"><img src="/character-still.png" alt="" draggable={false} /></span> : null}
+      <div className="agent-room__message-content">
+        <div className={`message-bubble${own ? " message-bubble--user" : ""}`}>
+          {own ? <p className="agent-room__plain-message">{message.content}</p> : <MarkdownContent markdown={message.content} openLinksInNewTab />}
+        </div>
+        <div className="agent-room__message-meta">
+          <span>{own ? "You" : "Sybil"}</span>
+          {date ? <time dateTime={message.createdAt} title={date.toLocaleString()}>{timeFormat.format(date)}</time> : null}
+          {own && status === "sending" ? <span className="agent-room__delivery" role="status">Sending…</span> : null}
+          {own && status === "queued" ? <span className="agent-room__delivery" role="status"><Clock3 size={11} aria-hidden="true" /> Queued</span> : null}
+          {own && failed ? <span className="agent-room__delivery is-error" title={typeof info.error === "string" ? info.error : undefined} role="status">
+            {status === "unconfirmed" ? "Unconfirmed" : "Not sent"}
+            <button type="button" onClick={() => onRetry(message.id)} aria-label="Retry message">Retry</button>
+          </span> : null}
+        </div>
+      </div>
+    </article>
+  );
+}
+
+export function AgentConversation({ room, onOpenSidebar, isVisible = true }: Props) {
+  const [draft, setDraft] = useState(savedDraft);
+  const draftRef = useRef(draft);
+  const seenMessageIdsRef = useRef(new Set(room.messages.map(message => message.id)));
+  const enteringMessageIdsRef = useRef(new Set<string>());
+  const animatedEntryIdsRef = useRef(new Set<string>());
+  const entryAnimationsRef = useRef(new Set<Animation>());
+  const typingHeightRef = useRef(0);
+  const historyReadyRef = useRef(!room.isLoading);
+  const [atBottom, setAtBottom] = useState(true);
+  const scrollerRef = useRef<HTMLDivElement>(null);
+  const timelineRef = useRef<HTMLDivElement>(null);
+  const textareaRef = useRef<HTMLTextAreaElement>(null);
+  const pinnedRef = useRef(true);
+  const restoredScrollTopRef = useRef<number | null>(null);
+  const anchorRef = useRef<ReadingAnchor | null>(null);
+  const markReadRef = useRef(room.markRead);
+  const visibleRef = useRef(isVisible);
+  const loadingRef = useRef(room.isLoading);
+  const readFrameRef = useRef<number | null>(null);
+  const mountedRef = useRef(true);
+  markReadRef.current = room.markRead;
+  visibleRef.current = isVisible;
+  loadingRef.current = room.isLoading;
+
+  const distanceFromBottom = () => {
+    const element = scrollerRef.current;
+    return element ? element.scrollHeight - element.scrollTop - element.clientHeight : Infinity;
+  };
+  const scheduleRead = () => {
+    if (readFrameRef.current !== null) cancelAnimationFrame(readFrameRef.current);
+    readFrameRef.current = requestAnimationFrame(() => {
+      readFrameRef.current = null;
+      if (mountedRef.current && visibleRef.current && !loadingRef.current && entryAnimationsRef.current.size === 0 && document.visibilityState === "visible" && document.hasFocus() && distanceFromBottom() <= 24) {
+        markReadRef.current();
+      }
+    });
+  };
+  const rememberAnchor = () => {
+    const element = scrollerRef.current;
+    if (!element || pinnedRef.current) { anchorRef.current = null; return; }
+    const top = element.getBoundingClientRect().top;
+    // Anchor to layout, independently of the bubble's entrance transform.
+    const node = Array.from(element.querySelectorAll<HTMLElement>("[data-agent-entry-id]")).find(item => item.getBoundingClientRect().bottom > top + 8);
+    anchorRef.current = node ? { id: node.dataset.agentEntryId!, offset: node.getBoundingClientRect().top - top } : null;
+  };
+  const restoreViewport = () => {
+    const element = scrollerRef.current;
+    if (!element) return;
+    if (pinnedRef.current) element.scrollTop = element.scrollHeight;
+    else if (anchorRef.current) {
+      const anchor = anchorRef.current;
+      const node = Array.from(element.querySelectorAll<HTMLElement>("[data-agent-entry-id]")).find(item => item.dataset.agentEntryId === anchor.id);
+      if (node) element.scrollTop += node.getBoundingClientRect().top - element.getBoundingClientRect().top - anchor.offset;
+    }
+    restoredScrollTopRef.current = element.scrollTop;
+    rememberAnchor();
+    scheduleRead();
+  };
+  const jumpToLatest = () => {
+    pinnedRef.current = true;
+    anchorRef.current = null;
+    setAtBottom(true);
+    if (scrollerRef.current) {
+      scrollerRef.current.scrollTop = scrollerRef.current.scrollHeight;
+      restoredScrollTopRef.current = scrollerRef.current.scrollTop;
+    }
+    scheduleRead();
+  };
+
+  useLayoutEffect(() => {
+    const timeline = timelineRef.current;
+    if (!timeline) return;
+    const gap = parseFloat(getComputedStyle(timeline).rowGap) || 0;
+    const typing = timeline.querySelector<HTMLElement>(".agent-room__typing");
+    // Transfer the departing indicator's space to the reply before it grows.
+    // Removing that space first would briefly move the transcript downward.
+    let replacedTypingHeight = typing ? 0 : typingHeightRef.current;
+    typingHeightRef.current = typing ? typing.offsetHeight + (typing.previousElementSibling ? gap : 0) : 0;
+    const reduceMotion = window.matchMedia("(prefers-reduced-motion: reduce)").matches;
+    for (const entry of timeline.querySelectorAll<HTMLElement>("[data-agent-entry-id]")) {
+      const id = entry.dataset.agentEntryId!;
+      if (animatedEntryIdsRef.current.has(id) || !enteringMessageIdsRef.current.has(id)) continue;
+      animatedEntryIdsRef.current.add(id);
+      if (reduceMotion) continue;
+      const height = entry.getBoundingClientRect().height;
+      const spacing = entry.previousElementSibling ? gap : 0;
+      const isReply = Boolean(entry.querySelector(".agent-room__message.is-agent"));
+      const initialHeight = isReply ? Math.min(replacedTypingHeight, height + spacing) : 0;
+      if (isReply) replacedTypingHeight = 0;
+      // Include the inter-message gap in the reveal. Clipping keeps the bubble's
+      // positive-Y transform from prematurely extending the scrollable area.
+      entry.style.overflow = "clip";
+      const animation = entry.animate([
+        { height: `${initialHeight}px`, marginTop: `${-spacing}px` },
+        { height: `${height}px`, marginTop: "0px" },
+      ], { duration: 240, easing: "cubic-bezier(.2,.7,.25,1)" });
+      entryAnimationsRef.current.add(animation);
+      animation.onfinish = () => {
+        entry.style.overflow = "";
+        entryAnimationsRef.current.delete(animation);
+        restoreViewport();
+      };
+    }
+  }, [room.messages, room.isTyping]);
+
+  useLayoutEffect(() => {
+    restoreViewport();
+  }, [room.messages, room.isTyping, room.activity?.state, room.isLoading, isVisible]);
+
+  useEffect(() => {
+    mountedRef.current = true;
+    const observer = new ResizeObserver(restoreViewport);
+    if (scrollerRef.current) observer.observe(scrollerRef.current);
+    if (timelineRef.current) observer.observe(timelineRef.current);
+    const onVisibility = () => scheduleRead();
+    window.addEventListener("focus", onVisibility);
+    document.addEventListener("visibilitychange", onVisibility);
+    return () => {
+      mountedRef.current = false;
+      observer.disconnect();
+      for (const animation of entryAnimationsRef.current) animation.cancel();
+      entryAnimationsRef.current.clear();
+      window.removeEventListener("focus", onVisibility);
+      document.removeEventListener("visibilitychange", onVisibility);
+      if (readFrameRef.current !== null) cancelAnimationFrame(readFrameRef.current);
+    };
+  }, []);
+
+  useLayoutEffect(() => {
+    const textarea = textareaRef.current;
+    if (!textarea) return;
+    textarea.style.height = "0px";
+    textarea.style.height = `${Math.min(156, Math.max(46, textarea.scrollHeight))}px`;
+  }, [draft]);
+
+  const updateDraft = (value: string) => {
+    draftRef.current = value;
+    saveDraft(value);
+    setDraft(value);
+  };
+
+  const retryMessage = (id: string) => {
+    // Admission errors remain attached to their own bubble.
+    void room.retryMessage(id).catch(() => {});
+  };
+  const submit = () => {
+    const content = draftRef.current.trim();
+    if (!content) return;
+    const delivery = room.sendMessage(content);
+    updateDraft("");
+    jumpToLatest();
+    void delivery.catch(() => {});
+    textareaRef.current?.focus();
+  };
+
+  const messages = useMemo(() => {
+    let previousDay = "";
+    for (const message of room.messages) {
+      if (!seenMessageIdsRef.current.has(message.id) && historyReadyRef.current && !room.isLoading && message.role !== "tool") {
+        enteringMessageIdsRef.current.add(message.id);
+      }
+      seenMessageIdsRef.current.add(message.id);
+    }
+    if (!room.isLoading) historyReadyRef.current = true;
+    return room.messages.filter(message => message.role !== "system").map(message => {
+      const date = dateOf(message.createdAt), day = dayKey(date), separator = previousDay !== day;
+      previousDay = day;
+      return (
+        <div className="agent-room__entry" key={message.id} data-agent-entry-id={message.id}>
+          {separator ? <div className="agent-room__day"><span>{dayLabel(date)}</span></div> : null}
+          {message.role === "tool" ? <AgentToolCard message={message} /> : <AgentBubble message={message} animate={enteringMessageIdsRef.current.has(message.id)} onRetry={retryMessage} />}
+        </div>
+      );
+    });
+  }, [room.messages, room.isLoading, room.retryMessage]);
+
+  const activity = room.activity;
+  const running = room.isTyping;
+  const waiting = activity?.state === "waiting";
+  const connected = room.connection === "connected";
+  const connectionLabel = connected ? "Connected" : room.connection === "offline" ? "Offline" : room.connection === "reconnecting" ? "Reconnecting" : "Connecting";
+  const activeTool = activity?.tools?.find(tool => tool.status !== "done");
+  const workingLabel = activeTool ? `Using ${toolName(activeTool.name)}`
+    : activity?.isThinking ? "Sybil is thinking" : (activity?.backgroundWorkers || 0) > 0 ? "Working in the background" : "";
+
+  return (
+    <section className="agent-room" aria-label="Sybil Agent conversation">
+      <header className="agent-room__header flex shrink-0 items-center justify-between gap-2 border-b border-violet-300/12 bg-[linear-gradient(180deg,hsl(243_48%_10%_/_0.86),hsl(236_48%_6%_/_0.66))] px-4 py-3 md:gap-3 md:px-7">
+        <div className="flex min-w-0 items-center gap-2">
+          <Button type="button" size="icon" variant="ghost" className="workspace-sidebar-trigger -ml-1 h-8 w-8 md:hidden" onClick={onOpenSidebar} aria-label="Open sidebar">
+            <Menu className="h-4 w-4" />
+          </Button>
+          <h1 className="truncate text-sm font-semibold text-violet-50 md:text-base">Sybil Agent</h1>
+        </div>
+        <div className="flex h-10 shrink-0 items-center justify-end gap-2">
+          <span className="sr-only" role="status">{connectionLabel}</span>
+          {!connected ? <Button type="button" size="icon" variant="ghost" onClick={room.retry} aria-label="Reconnect to Sybil Agent" title={connectionLabel}><RefreshCw className="h-4 w-4" /></Button> : null}
+        </div>
+      </header>
+
+      <div className="agent-room__body">
+        <div
+          ref={scrollerRef}
+          className="agent-room__scroll"
+          tabIndex={0}
+          aria-label="Conversation messages"
+          onScroll={() => {
+            // A scroll event from our last resize correction can arrive after
+            // the next animation frame has already grown a tall reply. That
+            // temporary bottom distance does not mean the reader scrolled up.
+            if (restoredScrollTopRef.current !== null && scrollerRef.current
+              && Math.abs(scrollerRef.current.scrollTop - restoredScrollTopRef.current) < 1) return;
+            const bottom = distanceFromBottom() <= 48;
+            pinnedRef.current = bottom;
+            setAtBottom(current => current === bottom ? current : bottom);
+            rememberAnchor();
+            if (bottom) scheduleRead();
+          }}
+        >
+          <div ref={timelineRef} className="agent-room__timeline">
+            {room.isLoading && !messages.length ? <div className="agent-room__empty" role="status"><LoaderCircle className="h-5 w-5 animate-spin" aria-hidden="true" /><span className="sr-only">Loading conversation</span></div> : null}
+            {messages}
+            {running ? <div className="agent-room__typing" role="status" aria-live="polite"><span className="agent-room__typing-dots" aria-hidden="true"><i /><i /><i /></span><span>{workingLabel}</span>{(activity?.backgroundWorkers || 0) > 1 ? <span className="agent-room__worker-count">{activity!.backgroundWorkers} workers</span> : null}</div> : null}
+            {!running && waiting && activity?.wait ? <div className="agent-room__waiting" role="status"><Clock3 size={13} /><span>Checking back at {timeFormat.format(new Date(activity.wait.due))}</span></div> : null}
+          </div>
+        </div>
+        {!atBottom ? <button className="agent-room__jump" onClick={jumpToLatest}><ArrowDown size={15} /><span>{room.unreadCount ? `${room.unreadCount} new message${room.unreadCount === 1 ? "" : "s"}` : "Jump to latest"}</span></button> : null}
+      </div>
+
+      <div className="agent-room__composer-area">
+        <div className="agent-room__composer-inner">
+          {room.error ? <div className="agent-room__error" role="alert"><WifiOff size={15} /><span>{room.error}</span><button onClick={room.retry}>Retry</button></div> : null}
+          <form className="agent-room__composer" onSubmit={event => { event.preventDefault(); submit(); }}>
+            <textarea
+              ref={textareaRef}
+              value={draft}
+              rows={1}
+              onInput={event => updateDraft(event.currentTarget.value)}
+              onKeyDown={event => { if (event.key === "Enter" && !event.shiftKey && !event.isComposing) { event.preventDefault(); submit(); } }}
+              placeholder="Message Sybil Agent…"
+              aria-label="Message Sybil Agent"
+              enterkeyhint="send"
+            />
+            <button type="submit" className="agent-room__send" disabled={!draft.trim()} aria-label="Send message"><SendHorizontal size={19} /></button>
+          </form>
+        </div>
+      </div>
+    </section>
+  );
+}

web/src/components/chat/chat-attachment-list.tsx

@@ -16,7 +16,7 @@ function getTextPreview(value: string) {
 
 function getSurfaceClasses(tone: Props["tone"]) {
   if (tone === "user") {
-    return "border-white/12 bg-black/16 text-fuchsia-50";
+    return "border-white/12 bg-black/16 text-inherit";
   }
   if (tone === "assistant") {
     return "border-violet-300/16 bg-violet-400/8 text-violet-50";

web/src/components/chat/chat-messages-panel.tsx

@@ -1,8 +1,9 @@
-import { useEffect, useMemo, useRef, useState } from "preact/hooks";
+import { useEffect, useLayoutEffect, useMemo, useRef, useState } from "preact/hooks";
 import type { ComponentChildren, JSX } from "preact";
 import { cn } from "@/lib/utils";
 import { ChatAttachmentList } from "@/components/chat/chat-attachment-list";
 import { getMessageAttachments, type Message } from "@/lib/api";
+import { asStreamingAssistantMetadata } from "@/lib/chat-stream-presentation";
 import { MarkdownContent } from "@/components/markdown/markdown-content";
 import { ChevronDown, ChevronUp, Globe2, Link2, Wrench } from "lucide-preact";
 
@@ -377,13 +378,13 @@ function ToolCallStack({
           );
         })}
         {!expanded && hiddenCount ? (
-          <span className="absolute bottom-1 right-10 z-20 rounded-full border border-cyan-300/30 bg-slate-950/86 px-2 py-0.5 text-[10px] font-semibold leading-none text-cyan-100 shadow-sm">
+          <span className="absolute bottom-1 right-10 z-20 rounded-none border border-cyan-300/30 bg-slate-950/86 px-2 py-0.5 text-[10px] font-semibold leading-none text-cyan-100 shadow-sm">
             +{hiddenCount}
           </span>
         ) : null}
         <button
           type="button"
-          className="tool-call-stack-toggle absolute right-0 top-2 z-20 flex h-8 w-8 items-center justify-center rounded-full"
+          className="tool-call-stack-toggle absolute right-0 top-2 z-20 flex h-8 w-8 items-center justify-center rounded-none"
           aria-expanded={expanded ? "true" : "false"}
           aria-label={`${expanded ? "Collapse" : "Expand"} ${countLabel}`}
           title={`${expanded ? "Collapse" : "Expand"} ${countLabel}`}
@@ -396,8 +397,144 @@ function ToolCallStack({
   );
 }
 
+type RenderedThinkingTrace = {
+  text: string;
+  revision: number;
+};
+
+const THINKING_TRACE_EXIT_MS = 220;
+
+export function StreamingAssistantPresentation({ message, isSending }: { message: Message; isSending: boolean }) {
+  const metadata = asStreamingAssistantMetadata(message.metadata);
+  const traceText = metadata?.ephemeralTrace?.trim() || null;
+  const traceRevision = metadata?.traceRevision ?? 0;
+  const [activeTrace, setActiveTrace] = useState<RenderedThinkingTrace | null>(() =>
+    traceText ? { text: traceText, revision: traceRevision } : null
+  );
+  const [outgoingTrace, setOutgoingTrace] = useState<RenderedThinkingTrace | null>(null);
+  const [reservedTraceHeight, setReservedTraceHeight] = useState(0);
+  const activeTraceRef = useRef<HTMLDivElement | null>(null);
+  const outgoingTraceRef = useRef<HTMLDivElement | null>(null);
+  const exitTimerRef = useRef<number | null>(null);
+
+  useLayoutEffect(() => {
+    const isSameTrace =
+      (!activeTrace && !traceText) ||
+      (activeTrace?.text === traceText && activeTrace?.revision === traceRevision && Boolean(activeTrace) === Boolean(traceText));
+    if (isSameTrace) return;
+
+    if (exitTimerRef.current !== null) {
+      window.clearTimeout(exitTimerRef.current);
+      exitTimerRef.current = null;
+    }
+
+    if (activeTrace) {
+      setOutgoingTrace(activeTrace);
+      exitTimerRef.current = window.setTimeout(() => {
+        setOutgoingTrace(null);
+        exitTimerRef.current = null;
+      }, THINKING_TRACE_EXIT_MS);
+    } else {
+      setOutgoingTrace(null);
+    }
+
+    setActiveTrace(traceText ? { text: traceText, revision: traceRevision } : null);
+  }, [traceRevision, traceText]);
+
+  useEffect(() => {
+    return () => {
+      if (exitTimerRef.current !== null) window.clearTimeout(exitTimerRef.current);
+    };
+  }, []);
+
+  useLayoutEffect(() => {
+    const traceElements = [activeTraceRef.current, outgoingTraceRef.current].filter(
+      (element): element is HTMLDivElement => element !== null
+    );
+    if (!traceElements.length) return;
+
+    const reserveMeasuredHeight = () => {
+      const measuredHeight = Math.max(...traceElements.map((element) => element.getBoundingClientRect().height));
+      if (measuredHeight > 0) {
+        setReservedTraceHeight((current) => Math.max(current, Math.ceil(measuredHeight)));
+      }
+    };
+
+    reserveMeasuredHeight();
+    if (typeof ResizeObserver === "undefined") return;
+
+    const observer = new ResizeObserver(reserveMeasuredHeight);
+    for (const element of traceElements) observer.observe(element);
+    return () => observer.disconnect();
+  }, [activeTrace, outgoingTrace]);
+
+  const hasAnswer = message.content.trim().length > 0;
+  const showTyping = isSending && !hasAnswer && !traceText && !activeTrace && !outgoingTrace;
+  const phase = hasAnswer ? "answer" : traceText || activeTrace || outgoingTrace ? "trace" : "pending";
+
+  return (
+    <div
+      className="streaming-assistant-slot"
+      data-streaming-assistant="true"
+      data-streaming-phase={phase}
+      style={reservedTraceHeight ? { minHeight: `${reservedTraceHeight}px` } : undefined}
+    >
+      {outgoingTrace ? (
+        <div
+          ref={outgoingTraceRef}
+          key={`outgoing-trace-${outgoingTrace.revision}`}
+          className="streaming-assistant-layer ephemeral-thinking-trace ephemeral-thinking-trace-exit"
+          data-thinking-trace="true"
+          data-thinking-trace-state="outgoing"
+          aria-hidden="true"
+        >
+          <p className="whitespace-pre-wrap">{outgoingTrace.text}</p>
+        </div>
+      ) : null}
+      {activeTrace ? (
+        <div
+          ref={activeTraceRef}
+          key={`active-trace-${activeTrace.revision}`}
+          className="streaming-assistant-layer ephemeral-thinking-trace ephemeral-thinking-trace-enter"
+          data-thinking-trace="true"
+          data-thinking-trace-state="active"
+          aria-hidden="true"
+        >
+          <p className="whitespace-pre-wrap">{activeTrace.text}</p>
+        </div>
+      ) : null}
+      {activeTrace && !hasAnswer ? (
+        <span className="sr-only" role="status" aria-live="polite">
+          Assistant is working
+        </span>
+      ) : null}
+      {hasAnswer ? (
+        <div className="streaming-assistant-layer streaming-assistant-answer" data-streaming-answer="true">
+          <MarkdownContent
+            markdown={message.content}
+            openLinksInNewTab
+            className="leading-[1.82] text-violet-50 [&_a]:text-inherit [&_a]:underline"
+          />
+        </div>
+      ) : null}
+      {showTyping ? (
+        <span
+          className="streaming-assistant-layer inline-flex items-center gap-1"
+          data-streaming-pending="true"
+          aria-label="Assistant is typing"
+          role="status"
+        >
+          <span className="inline-block h-1.5 w-1.5 animate-bounce rounded-full bg-muted-foreground [animation-delay:0ms]" />
+          <span className="inline-block h-1.5 w-1.5 animate-bounce rounded-full bg-muted-foreground [animation-delay:140ms]" />
+          <span className="inline-block h-1.5 w-1.5 animate-bounce rounded-full bg-muted-foreground [animation-delay:280ms]" />
+        </span>
+      ) : null}
+    </div>
+  );
+}
+
 export function ChatMessagesPanel({ messages, isLoading, isSending, onMessageContextMenu }: Props) {
-  const hasPendingAssistant = messages.some((message) => message.id.startsWith("temp-assistant-") && message.content.trim().length === 0);
+  const hasPendingAssistant = messages.some((message) => message.id.startsWith("temp-assistant-"));
   const renderItems = useMemo(() => buildMessageRenderItems(messages), [messages]);
   const toolCallMessageIDs = useMemo(() => getToolCallMessageIDs(messages), [messages]);
   const seenToolCallMessageIDsRef = useRef<Set<string> | null>(null);
@@ -457,16 +594,17 @@ export function ChatMessagesPanel({ messages, isLoading, isSending, onMessageCon
           }
 
           const isUser = message.role === "user";
+          const streamingAssistantMetadata = asStreamingAssistantMetadata(message.metadata);
           const isPendingAssistant = message.id.startsWith("temp-assistant-") && isSending && message.content.trim().length === 0;
           const attachments = getMessageAttachments(message.metadata);
           return (
             <div key={message.id} className={cn("flex", isUser ? "justify-end" : "justify-start")}>
               <div
                 className={cn(
-                  "max-w-[85%] space-y-3",
+                  "min-w-0 space-y-3",
                   isUser
-                    ? "rounded-xl border border-violet-300/24 bg-[linear-gradient(135deg,hsl(258_86%_48%_/_0.86),hsl(278_72%_29%_/_0.86))] px-4 py-3 text-sm leading-6 text-fuchsia-50 shadow-sm"
-                    : "text-base leading-7 text-violet-50"
+                    ? "message-bubble message-bubble--user max-w-[87%] md:max-w-[78%]"
+                    : "max-w-[85%] text-base leading-7 text-violet-50"
                 )}
                 onContextMenu={
                   message.role === "assistant" && !message.id.startsWith("temp-") && onMessageContextMenu
@@ -475,7 +613,9 @@ export function ChatMessagesPanel({ messages, isLoading, isSending, onMessageCon
                 }
               >
                 {attachments.length ? <ChatAttachmentList attachments={attachments} tone={isUser ? "user" : "assistant"} /> : null}
-                {isPendingAssistant ? (
+                {streamingAssistantMetadata ? (
+                  <StreamingAssistantPresentation message={message} isSending={isSending} />
+                ) : isPendingAssistant ? (
                   <span className="inline-flex items-center gap-1" aria-label="Assistant is typing" role="status">
                     <span className="inline-block h-1.5 w-1.5 animate-bounce rounded-full bg-muted-foreground [animation-delay:0ms]" />
                     <span className="inline-block h-1.5 w-1.5 animate-bounce rounded-full bg-muted-foreground [animation-delay:140ms]" />
@@ -485,7 +625,7 @@ export function ChatMessagesPanel({ messages, isLoading, isSending, onMessageCon
                   <MarkdownContent
                     markdown={message.content}
                     openLinksInNewTab
-                    className={cn("[&_a]:text-inherit [&_a]:underline", isUser ? "leading-[1.78] text-fuchsia-50" : "leading-[1.82] text-violet-50")}
+                    className={cn("[&_a]:text-inherit [&_a]:underline", !isUser && "leading-[1.82] text-violet-50")}
                   />
                 ) : null}
               </div>

web/src/components/search/search-match-label.tsx

@@ -0,0 +1,96 @@
+/// <reference lib="es2022.intl" />
+import { useLayoutEffect, useMemo, useRef, useState } from "preact/hooks";
+
+const words = /[\p{L}\p{N}\p{M}\p{Co}]+/gu;
+const fold = (text: string) => text.normalize("NFD").replace(/\p{M}/gu, "").toLowerCase();
+const graphemes = new Intl.Segmenter(undefined, { granularity: "grapheme" });
+
+export function SearchMatchLabel({ text: source, query }: { text: string; query: string }) {
+  const text = source.replace(/\s+/gu, " ").trim();
+  const containerRef = useRef<HTMLSpanElement>(null);
+  const measureRef = useRef<HTMLSpanElement>(null);
+  const ellipsisRef = useRef<HTMLSpanElement>(null);
+  const [kept, setKept] = useState<number | null>(null);
+  const boundaries = useMemo(() => [0, ...Array.from(graphemes.segment(text), ({ index, segment }) => index + segment.length)], [text]);
+  const runs = useMemo(() => {
+    const terms = (query.match(words) ?? []).map(fold).filter(Boolean);
+    const result: Array<{ start: number; end: number; match: boolean }> = [];
+    let cursor = 0;
+    for (const word of text.matchAll(words)) {
+      if (word.index > cursor) result.push({ start: cursor, end: word.index, match: false });
+      cursor = word.index + word[0].length;
+      result.push({ start: word.index, end: cursor, match: terms.some((term) => fold(word[0]).startsWith(term)) });
+    }
+    if (cursor < text.length) result.push({ start: cursor, end: text.length, match: false });
+    return result;
+  }, [text, query]);
+
+  const renderRange = (start: number, end: number) => runs.flatMap((run) => {
+    const value = text.slice(Math.max(start, run.start), Math.min(end, run.end));
+    if (run.end <= start || run.start >= end) return [];
+    return [run.match
+      ? <strong key={`${start}:${run.start}`} className="font-semibold text-violet-100/95">{value}</strong>
+      : <span key={`${start}:${run.start}`}>{value}</span>];
+  });
+
+  useLayoutEffect(() => {
+    const container = containerRef.current;
+    const measure = measureRef.current;
+    const ellipsis = ellipsisRef.current;
+    if (!container || !measure || !ellipsis) return;
+    let disposed = false;
+    const update = () => {
+      if (disposed) return;
+      const width = container.getBoundingClientRect().width;
+      if (measure.getBoundingClientRect().width <= width) { setKept(null); return; }
+      // Measure the styled DOM, including bold runs, instead of guessing a
+      // character limit. Grapheme boundaries keep accents and emoji intact.
+      const nodes = Array.from(measure.children, (child) => child.firstChild as Text);
+      const range = document.createRange();
+      const measureRange = (start: number, end: number) => {
+        if (start >= end) return 0;
+        let offset = 0;
+        let started = false;
+        for (const node of nodes) {
+          const next = offset + node.length;
+          if (!started && start <= next) { range.setStart(node, start - offset); started = true; }
+          if (end <= next) { range.setEnd(node, end - offset); break; }
+          offset = next;
+        }
+        return range.getBoundingClientRect().width;
+      };
+      const count = boundaries.length - 1;
+      const ellipsisWidth = ellipsis.getBoundingClientRect().width;
+      let low = 0;
+      let high = count - 1;
+      while (low < high) {
+        const middle = Math.ceil((low + high) / 2);
+        const prefix = boundaries[Math.ceil(middle / 2)];
+        const suffix = boundaries[count - Math.floor(middle / 2)];
+        if (measureRange(0, prefix) + ellipsisWidth + measureRange(suffix, text.length) <= width - 1) low = middle;
+        else high = middle - 1;
+      }
+      setKept(low);
+    };
+    update();
+    const observer = new ResizeObserver(update);
+    observer.observe(container);
+    void document.fonts.ready.then(update);
+    document.fonts.addEventListener("loadingdone", update);
+    return () => { disposed = true; observer.disconnect(); document.fonts.removeEventListener("loadingdone", update); };
+  }, [text, runs, boundaries]);
+
+  const count = boundaries.length - 1;
+  const truncated = kept !== null && kept < count;
+  return (
+    <span ref={containerRef} title={text} className="relative col-start-2 block min-w-0 overflow-hidden whitespace-nowrap text-xs leading-5 text-secondary-foreground/75">
+      <span className="sr-only">{text}</span>
+      <span aria-hidden="true">
+        {truncated ? <>{renderRange(0, boundaries[Math.ceil(kept / 2)])}…{renderRange(boundaries[count - Math.floor(kept / 2)], text.length)}</> : renderRange(0, text.length)}
+      </span>
+      <span aria-hidden="true" className="pointer-events-none invisible absolute left-0 top-0 whitespace-pre">
+        <span ref={measureRef}>{renderRange(0, text.length)}</span><span ref={ellipsisRef}>…</span>
+      </span>
+    </span>
+  );
+}

web/src/hooks/use-agent-room.ts

@@ -0,0 +1,229 @@
+import { useCallback, useEffect, useMemo, useRef, useState } from "preact/hooks";
+import { getApiBaseUrl, getAuthToken, type Message } from "@/lib/api";
+import {
+  AGENT_ADMISSION_STORAGE_KEY, AgentRoomClient, advanceAgentCursor, agentAuthFrame, agentDisplayMessages, agentUnreadCount,
+  applyAgentEvent, applyAgentReadState,
+  createAgentReadWriter, createAgentRoomState, parseAgentRoomEvent, syncAgentHistory,
+  type AgentActivity, type AgentRoomState,
+} from "@/lib/agent-room";
+import { AGENT_OUTBOX_STORAGE_KEY, createAgentOptimisticOutbox } from "@/lib/agent-optimistic-outbox";
+import { createAgentRoomPresenter } from "@/lib/agent-room-presentation";
+
+export type AgentRoomConnection = "connecting" | "connected" | "reconnecting" | "offline";
+export type AgentRoomHookResult = {
+  messages: Message[]; activity: AgentActivity | null; unreadCount: number;
+  connection: AgentRoomConnection; error: string | null; isLoading: boolean; isTyping: boolean;
+  sendMessage(content: string): Promise<void>; retryMessage(messageId: string): Promise<void>; markRead(): void; retry(): void;
+};
+const errorText = (error: unknown) => error instanceof Error ? error.message : String(error);
+
+/** Mount once in App; selection controls reading, never the lifetime of this subscription. */
+export function useAgentRoom(enabled: boolean): AgentRoomHookResult {
+  const [view, setView] = useState(() => ({ room: createAgentRoomState(), isTyping: false }));
+  const room = view.room;
+  const current = useRef(room);
+  const [connection, setConnection] = useState<AgentRoomConnection>("offline");
+  const [error, setError] = useState<string | null>(null);
+  const [isLoading, setLoading] = useState(enabled);
+  const [revision, setRevision] = useState(0);
+  const client = useRef<AgentRoomClient | null>(null);
+  const mark = useRef<((cursor: string) => void) | null>(null);
+  const [outboxRevision, setOutboxRevision] = useState(0);
+  const update = useCallback((change: (state: AgentRoomState) => AgentRoomState) => {
+    const next = change(current.current); current.current = next; setView(previous => ({ ...previous, room: next }));
+  }, []);
+  const outbox = useRef<ReturnType<typeof createAgentOptimisticOutbox> | null>(null);
+  if (!outbox.current) {
+    const stored = (key: string) => { try { return JSON.parse(sessionStorage.getItem(key) || "null"); } catch { return null; } };
+    outbox.current = createAgentOptimisticOutbox({
+      submit: (content, requestId) => {
+        if (!client.current) throw new Error("Agent connection is unavailable.");
+        return client.current.submit(content, requestId);
+      },
+      changed: () => setOutboxRevision(value => value + 1),
+      storage: {
+        read: () => stored(AGENT_OUTBOX_STORAGE_KEY),
+        legacyRows: () => stored("sybil.pi-root.outbox"),
+        legacyIds: () => stored(AGENT_ADMISSION_STORAGE_KEY),
+        write: rows => {
+          try {
+            // Keep [] as a migration marker; legacy keys are retired only once
+            // their replacement has been saved successfully.
+            sessionStorage.setItem(AGENT_OUTBOX_STORAGE_KEY, JSON.stringify(rows));
+            sessionStorage.removeItem("sybil.pi-root.outbox");
+            sessionStorage.removeItem(AGENT_ADMISSION_STORAGE_KEY);
+          } catch { throw new Error("Could not save this message for retry. Allow session storage or free browser storage, then retry."); }
+        },
+      },
+    });
+  }
+
+  useEffect(() => {
+    if (!enabled) {
+      client.current = null; mark.current = null;
+      update(() => createAgentRoomState()); setLoading(false); setConnection("offline"); setError(null);
+      return;
+    }
+    let active = true;
+    let socket: WebSocket | undefined;
+    let reconnectTimer: ReturnType<typeof setTimeout> | undefined;
+    let startupTimer: ReturnType<typeof setTimeout> | undefined;
+    let socketTimeout: ReturnType<typeof setTimeout> | undefined;
+    let reconnectAttempt = 0;
+    let started = false;
+    let maintaining = false;
+    let lastPacket = Date.now();
+    let socketAuthenticated = false;
+    let authFailed = false;
+    const controller = new AbortController();
+    const api = new AgentRoomClient({ baseUrl: getApiBaseUrl(), token: getAuthToken });
+    client.current = api;
+    const presenter = createAgentRoomPresenter({ initial: current.current, reduceEvent: applyAgentEvent,
+      publish: (next, isTyping) => {
+        if (!active) return;
+        outbox.current?.reconcile(next);
+        current.current = next;
+        setView({ room: next, isTyping });
+      },
+    });
+    setLoading(current.current.conversationId === null); setConnection("connecting"); setError(null);
+    const reportError = (failure: unknown) => { if (active && !controller.signal.aborted) setError(errorText(failure)); };
+    // Read acknowledgments may be ahead of the batched UI for ~100ms. The
+    // writer must observe that acknowledged cursor immediately to avoid a
+    // repeated PUT loop while the presenter waits to publish its batch.
+    let acknowledgedReadCursor = current.current.readCursor;
+    const readWriter = createAgentReadWriter({
+      current: () => acknowledgedReadCursor,
+      save: cursor => api.markRead(cursor, controller.signal),
+      apply: read => {
+        acknowledgedReadCursor = advanceAgentCursor(acknowledgedReadCursor, read.cursor);
+        if (active) presenter.update(state => applyAgentReadState(state, read));
+      },
+    });
+    mark.current = cursor => { void readWriter.request(cursor).catch(reportError); };
+
+    const refreshRead = async () => {
+      const read = await api.readState(controller.signal);
+      acknowledgedReadCursor = advanceAgentCursor(acknowledgedReadCursor, read.cursor);
+      if (active) presenter.update(state => applyAgentReadState(state, read));
+    };
+    const catchUp = async (replay = false) => syncAgentHistory(api, current.current.cursor, page => {
+      if (active) presenter.push({ type: "messages", ...page }, { replay });
+    }, controller.signal);
+    const refreshActivity = async () => {
+      const descriptor = await api.descriptor(controller.signal);
+      if (active) presenter.push({ type: "activity", conversationId: descriptor.id, activity: descriptor.activity });
+    };
+
+    const connect = () => {
+      if (!active || authFailed) return;
+      clearTimeout(reconnectTimer); clearTimeout(socketTimeout);
+      setConnection(reconnectAttempt ? "reconnecting" : "connecting");
+      socketAuthenticated = false;
+      const next = new WebSocket(api.websocketUrl(window.location.href));
+      socket = next;
+      socketTimeout = setTimeout(() => { if (!socketAuthenticated) next.close(); }, 12_000);
+      next.onopen = () => {
+        if (!active || socket !== next) { next.close(); return; }
+        lastPacket = Date.now();
+        next.send(agentAuthFrame(getAuthToken(), current.current.cursor));
+      };
+      next.onmessage = ({ data }) => {
+        if (!active || socket !== next) return;
+        lastPacket = Date.now();
+        try {
+          const event = parseAgentRoomEvent(String(data));
+          if (!event) return;
+          if (event.type === "error") { reportError(new Error(event.error || event.message || "Agent connection error.")); return; }
+          if (event.type === "ready") {
+            clearTimeout(socketTimeout); socketAuthenticated = true; reconnectAttempt = 0;
+            setConnection("connected"); setError(null); return;
+          }
+          presenter.push(event);
+        } catch (failure) {
+          reportError(failure);
+          // A malformed message must not advance the replay cursor. Reconnect
+          // from the last applied page, rather than silently skipping that input.
+          next.close();
+        }
+      };
+      next.onerror = () => { if (active) setConnection("reconnecting"); };
+      next.onclose = event => {
+        if (!active || socket !== next) return;
+        clearTimeout(socketTimeout); socketAuthenticated = false;
+        if (event.code === 4401) {
+          authFailed = true; setConnection("offline"); setError("Agent authentication failed. Sign in again or retry."); return;
+        }
+        setConnection("reconnecting");
+        const wait = Math.min(30_000, 1000 * 2 ** Math.min(reconnectAttempt++, 5));
+        reconnectTimer = setTimeout(connect, wait);
+      };
+    };
+
+    const start = async () => {
+      try {
+        await Promise.all([refreshRead(), refreshActivity()]);
+        // The first start begins at zero, not the latest page. Unread counts
+        // therefore include older public replies even after a long absence.
+        await catchUp(!started);
+        if (!active) return;
+        presenter.flush();
+        started = true; setLoading(false); connect();
+      } catch (failure) {
+        if (!active) return;
+        reportError(failure); setLoading(false); setConnection("offline");
+        startupTimer = setTimeout(() => { void start(); }, 10_000);
+      }
+    };
+    const maintain = async (foreground = false) => {
+      if (!active || !started || maintaining || authFailed) return;
+      maintaining = true;
+      try {
+        await refreshRead();
+        await readWriter.flush();
+        if (!socketAuthenticated) {
+          await Promise.all([catchUp(), refreshActivity()]);
+        } else if (foreground) await catchUp();
+        if (socket?.readyState === WebSocket.OPEN) {
+          if (Date.now() - lastPacket > 60_000) socket.close();
+          else socket.send(JSON.stringify({ type: "ping" }));
+        }
+      } catch (failure) { reportError(failure); }
+      finally { maintaining = false; }
+    };
+    const focus = () => { if (document.visibilityState === "visible") void maintain(true); };
+    const interval = setInterval(() => { void maintain(); }, 15_000);
+    window.addEventListener("focus", focus);
+    document.addEventListener("visibilitychange", focus);
+    void start();
+    return () => {
+      active = false; controller.abort();
+      presenter.dispose();
+      clearInterval(interval); clearTimeout(reconnectTimer); clearTimeout(startupTimer); clearTimeout(socketTimeout);
+      window.removeEventListener("focus", focus); document.removeEventListener("visibilitychange", focus);
+      socket?.close();
+      if (client.current === api) client.current = null;
+      mark.current = null;
+    };
+  }, [enabled, revision, update]);
+
+  const renderedCursor = room.cursor;
+  const markRead = useCallback(() => {
+    if (!enabled || isLoading) return;
+    // This callback belongs to this rendered snapshot. Reading current.current
+    // here would incorrectly mark a socket message that has not rendered yet.
+    mark.current?.(renderedCursor);
+  }, [enabled, isLoading, renderedCursor]);
+  const sendMessage = useCallback(async (content: string) => {
+    if (!enabled || !outbox.current) throw new Error("Agent connection is unavailable.");
+    await outbox.current.send(content);
+  }, [enabled]);
+  const retryMessage = useCallback(async (messageId: string) => {
+    if (!enabled || !outbox.current) throw new Error("Agent connection is unavailable.");
+    await outbox.current.retry(messageId);
+  }, [enabled]);
+  const retry = useCallback(() => { setRevision(value => value + 1); }, []);
+  const messages = useMemo(() => outbox.current?.display(agentDisplayMessages(room)) || agentDisplayMessages(room), [room.messages, room.activity, outboxRevision]);
+  return { messages: enabled ? messages : [], activity: room.activity, unreadCount: agentUnreadCount(room), connection, error, isLoading,
+    isTyping: enabled && view.isTyping, sendMessage, retryMessage, markRead, retry };
+}

web/src/hooks/use-workspace-search.ts

@@ -0,0 +1,30 @@
+import { useEffect, useState } from "preact/hooks";
+import { searchWorkspaceItems, type WorkspaceItem, type WorkspaceSearchResponse } from "@/lib/api";
+import { scheduleWorkspaceSearch } from "@/lib/workspace-search";
+
+type SearchState = WorkspaceSearchResponse & { query: string; revision: number; error: string | null };
+const EMPTY_ITEMS: WorkspaceItem[] = [];
+
+export function useWorkspaceSearch(query: string, enabled: boolean, revision: number) {
+  const trimmed = query.trim();
+  const [state, setState] = useState<SearchState | null>(null);
+  const validationError = trimmed.length > 200 ? "Use 200 characters or fewer." : null;
+
+  useEffect(() => {
+    setState(null);
+    if (!enabled || !trimmed || validationError) return;
+    return scheduleWorkspaceSearch(trimmed, searchWorkspaceItems,
+      (result) => setState({ ...result, query: trimmed, revision, error: null }),
+      (error) => setState({ items: [], hasMore: false, query: trimmed, revision, error: error instanceof Error ? error.message : String(error) })
+    );
+  }, [trimmed, enabled, revision, validationError]);
+
+  // Hide obsolete results immediately, even before effect cleanup runs.
+  const current = enabled && state?.query === trimmed && state.revision === revision ? state : null;
+  return {
+    items: current?.items ?? EMPTY_ITEMS,
+    hasMore: current?.hasMore ?? false,
+    error: validationError ?? current?.error ?? null,
+    isLoading: enabled && Boolean(trimmed) && !validationError && !current,
+  };
+}

web/src/index.css

@@ -31,7 +31,7 @@
   --secondary-foreground: 254 38% 94%;
   --accent: 188 86% 50%;
   --accent-foreground: 230 45% 8%;
-  --radius: 0.5rem;
+  --radius: 0px;
 }
 
 * {
@@ -72,6 +72,7 @@ select,
 textarea {
   font: inherit;
   color-scheme: dark;
+  border-radius: var(--radius);
 }
 
 ::selection {
@@ -113,13 +114,39 @@ textarea {
     max(0.75rem, var(--safe-area-left));
 }
 
+.chat-settings-header {
+  padding:
+    calc(0.5rem + var(--safe-area-top))
+    max(0.75rem, var(--safe-area-right))
+    0.5rem
+    max(0.75rem, var(--safe-area-left));
+}
+
+.chat-settings-content {
+  padding:
+    1rem
+    max(1rem, var(--safe-area-right))
+    max(1.5rem, var(--safe-area-bottom))
+    max(1rem, var(--safe-area-left));
+  scroll-padding-block: 1rem;
+}
+
+.chat-settings-error {
+  padding:
+    0.75rem
+    max(1rem, var(--safe-area-right))
+    max(0.75rem, var(--safe-area-bottom))
+    max(1rem, var(--safe-area-left));
+}
+
 @media (min-width: 768px) {
-  .app-safe-frame {
-    padding:
-      max(0.5rem, var(--safe-area-top))
-      max(0.5rem, var(--safe-area-right))
-      max(0.5rem, var(--safe-area-bottom))
-      max(0.5rem, var(--safe-area-left));
+  .chat-settings-header,
+  .chat-settings-content {
+    padding: 1.25rem;
+  }
+
+  .chat-settings-error {
+    padding: 0.75rem 1.25rem;
   }
 
   .app-search-safe-pad {
@@ -132,19 +159,11 @@ textarea {
 }
 
 @media (horizontal-viewport-segments: 2) {
-  .app-safe-frame {
-    padding:
-      max(0.5rem, var(--safe-area-top))
-      max(0.5rem, var(--safe-area-right))
-      max(0.5rem, var(--safe-area-bottom))
-      max(0.5rem, var(--safe-area-left));
-  }
-
   .workspace-shell {
     display: grid;
     grid-template-columns:
-      calc(env(viewport-segment-width 0 0) - max(0.5rem, var(--safe-area-left)))
-      calc(env(viewport-segment-width 1 0) - max(0.5rem, var(--safe-area-right)));
+      calc(env(viewport-segment-width 0 0) - var(--safe-area-left))
+      calc(env(viewport-segment-width 1 0) - var(--safe-area-right));
     column-gap: calc(env(viewport-segment-left 1 0) - env(viewport-segment-right 0 0));
   }
 
@@ -153,14 +172,10 @@ textarea {
     width: 100%;
     max-width: none;
     transform: none;
-    border-width: 1px;
-    border-radius: 1rem;
   }
 
   .workspace-content {
     min-width: 0;
-    border-width: 1px;
-    border-radius: 1rem;
     touch-action: auto;
   }
 
@@ -179,6 +194,37 @@ textarea {
     0 14px 36px hsl(240 80% 2% / 0.28);
 }
 
+/* Shared by regular chats and the persistent agent transcript. */
+.message-bubble {
+  min-width: 0;
+  padding: 13px 17px;
+  border: 1px solid hsl(256 37% 57% / .18);
+  border-radius: 18px 18px 18px 5px;
+  background: linear-gradient(150deg, hsl(251 36% 15% / .93), hsl(244 35% 11% / .93));
+  box-shadow: 0 3px 12px hsl(239 55% 2% / .13);
+  font-size: 14px;
+  line-height: 1.68;
+  overflow-wrap: anywhere;
+}
+
+.message-bubble--user {
+  border-color: hsl(266 80% 72% / .3);
+  border-radius: 18px 18px 5px 18px;
+  background: linear-gradient(135deg, hsl(264 59% 33% / .86), hsl(259 58% 26% / .9));
+  color: hsl(265 55% 97%);
+}
+
+.message-bubble .md-content {
+  font-size: inherit;
+  line-height: inherit;
+}
+
+@media (max-width: 767px) {
+  .message-bubble {
+    padding: 11px 14px;
+  }
+}
+
 .tool-call-stack-shell {
   perspective: 900px;
   transform-style: preserve-3d;
@@ -321,6 +367,97 @@ textarea {
   }
 }
 
+.streaming-assistant-slot {
+  display: grid;
+  min-width: 0;
+  min-height: 1.82em;
+  align-items: start;
+}
+
+.streaming-assistant-layer {
+  grid-area: 1 / 1;
+  min-width: 0;
+  transform-origin: top left;
+}
+
+.ephemeral-thinking-trace {
+  position: relative;
+  z-index: 1;
+  color: hsl(252 35% 88% / 0.8);
+  font-style: italic;
+  opacity: 0.78;
+  filter: saturate(0.72);
+  will-change: opacity, transform, filter;
+}
+
+.ephemeral-thinking-trace-enter {
+  animation: ephemeral-thinking-trace-enter 200ms cubic-bezier(0.2, 0.78, 0.24, 1) both;
+}
+
+.ephemeral-thinking-trace-exit {
+  pointer-events: none;
+  animation: ephemeral-thinking-trace-exit 220ms cubic-bezier(0.4, 0, 0.8, 0.28) both;
+}
+
+.streaming-assistant-answer {
+  position: relative;
+  z-index: 2;
+  animation: streaming-assistant-answer-enter 180ms cubic-bezier(0.2, 0.78, 0.24, 1) both;
+}
+
+@keyframes ephemeral-thinking-trace-enter {
+  from {
+    opacity: 0;
+    filter: saturate(0.72) blur(1px);
+    transform: translate3d(0, 0.32rem, 0);
+  }
+
+  to {
+    opacity: 0.78;
+    filter: saturate(0.72) blur(0);
+    transform: translate3d(0, 0, 0);
+  }
+}
+
+@keyframes ephemeral-thinking-trace-exit {
+  from {
+    opacity: 0.78;
+    filter: saturate(0.72) blur(0);
+    transform: translate3d(0, 0, 0);
+  }
+
+  to {
+    opacity: 0;
+    filter: saturate(0.72) blur(1px);
+    transform: translate3d(0, -0.32rem, 0);
+  }
+}
+
+@keyframes streaming-assistant-answer-enter {
+  from {
+    opacity: 0;
+    transform: translate3d(0, 0.18rem, 0);
+  }
+
+  to {
+    opacity: 1;
+    transform: translate3d(0, 0, 0);
+  }
+}
+
+@media (prefers-reduced-motion: reduce) {
+  .ephemeral-thinking-trace-enter,
+  .ephemeral-thinking-trace-exit,
+  .streaming-assistant-answer {
+    animation: none;
+    transform: none;
+  }
+
+  .ephemeral-thinking-trace-exit {
+    opacity: 0;
+  }
+}
+
 .md-content {
   word-break: break-word;
 }
@@ -339,7 +476,7 @@ textarea {
   overflow-x: auto;
   overflow-y: hidden;
   border: 1px solid hsl(var(--border) / 0.86);
-  border-radius: 0.625rem;
+  border-radius: var(--radius);
   background: hsl(246 34% 10% / 0.76);
   box-shadow: inset 0 1px 0 hsl(258 80% 88% / 0.06);
 }
@@ -358,7 +495,7 @@ textarea {
 }
 
 .md-table-scroll::-webkit-scrollbar-thumb {
-  border-radius: 9999px;
+  border-radius: var(--radius);
   background: hsl(263 78% 72% / 0.34);
 }
 
@@ -470,7 +607,7 @@ textarea {
 
 .md-content code {
   background: hsl(249 40% 10% / 0.78);
-  border-radius: 0.3rem;
+  border-radius: var(--radius);
   padding: 0.05rem 0.3rem;
   font-size: 0.86em;
   box-decoration-break: clone;
@@ -480,7 +617,7 @@ textarea {
 .md-content pre {
   overflow-x: auto;
   border: 1px solid hsl(253 31% 29% / 0.72);
-  border-radius: 0.625rem;
+  border-radius: var(--radius);
   background: hsl(249 40% 10% / 0.82);
   padding: 0.75rem;
   box-shadow: inset 0 1px 0 hsl(258 80% 88% / 0.05);
@@ -504,7 +641,7 @@ textarea {
 .md-cite-token {
   display: inline-flex;
   align-items: center;
-  border-radius: 9999px;
+  border-radius: var(--radius);
   border: 1px solid hsl(274 40% 55%);
   background: hsl(277 33% 21%);
   color: hsl(304 42% 96%);

web/src/lib/agent-optimistic-outbox.ts

@@ -0,0 +1,166 @@
+import type { Message } from "./api";
+
+export const AGENT_OUTBOX_STORAGE_KEY = "sybil.agent.optimistic-outbox.v1";
+export type AgentDeliveryStatus = "sending" | "queued" | "complete" | "failed" | "unconfirmed";
+export type AgentOutboxRow = {
+  id: string; requestId: string; content: string; createdAt: string;
+  status: AgentDeliveryStatus; error?: string; seenOnServer: boolean; confirmed: boolean;
+};
+export type AgentOutboxStorage = {
+  read(): unknown; write(rows: AgentOutboxRow[]): void;
+  legacyRows?(): unknown; legacyIds?(): unknown;
+};
+type Receipt = { status: string; messageId?: string };
+type RoomMessages = { messages: readonly Message[]; activity?: { queuedMessages?: readonly Message[] } | null };
+const errorText = (error: unknown) => (error instanceof Error ? error.message : String(error)).slice(0, 240);
+const info = (message: Message) => message.metadata && typeof message.metadata === "object" && !Array.isArray(message.metadata)
+  ? message.metadata as Record<string, unknown> : {};
+const idFor = (requestId: string) => `inbound:sybil:${requestId}`;
+const validRequestId = (value: unknown): value is string => typeof value === "string" && value.length > 0 && value.length <= 256;
+
+/** Local rows use the server's identity from their first render through final echo. */
+export function createAgentOptimisticOutbox(options: {
+  submit(content: string, requestId: string): Promise<Receipt>;
+  storage: AgentOutboxStorage;
+  changed(): void;
+  newId?: () => string;
+  now?: () => number;
+}) {
+  const newId = options.newId || (() => crypto.randomUUID());
+  const now = options.now || Date.now;
+  const rows = new Map<string, AgentOutboxRow>();
+  const inFlight = new Map<string, Promise<void>>();
+  const rejectedStorage = new Set<string>();
+  const confirmedEchoes = new Map<string, Message>();
+  let latestTime = 0;
+  const timestamp = () => {
+    latestTime = Math.max(now(), latestTime + 1);
+    return new Date(latestTime).toISOString();
+  };
+  const restored = options.storage.read();
+  if (Array.isArray(restored)) {
+    for (const value of restored.slice(0, 100)) {
+      if (!value || typeof value.content !== "string" || !value.content.trim() || !validRequestId(value.requestId)) continue;
+      const time = typeof value.createdAt === "string" && Number.isFinite(Date.parse(value.createdAt)) ? value.createdAt : timestamp();
+      latestTime = Math.max(latestTime, Date.parse(time));
+      const status: AgentDeliveryStatus = ["failed", "queued", "complete", "unconfirmed"].includes(value.status) ? value.status : "unconfirmed";
+      const row: AgentOutboxRow = { id: idFor(value.requestId), requestId: value.requestId, content: value.content, createdAt: time,
+        status, seenOnServer: Boolean(value.seenOnServer), confirmed: false,
+        ...(typeof value.error === "string" && value.error ? { error: value.error.slice(0, 240) } : {}) };
+      rows.set(row.id, row);
+    }
+  } else {
+    // Join the former UI-only failed cards to the old sender's receipt keys.
+    // Neither representation is retried automatically during migration.
+    const oldRows = options.storage.legacyRows?.();
+    const oldIds = options.storage.legacyIds?.();
+    const ids = new Map<string, string>();
+    if (Array.isArray(oldIds)) for (const value of oldIds) {
+      if (value && typeof value.content === "string" && validRequestId(value.requestId)) ids.set(value.content, value.requestId);
+    }
+    const contents = new Map<string, { content: string; status?: string; error?: string }>();
+    if (Array.isArray(oldRows)) for (const value of oldRows) {
+      if (value && typeof value.content === "string" && value.content.trim()) contents.set(value.content, value);
+    }
+    for (const content of ids.keys()) if (!contents.has(content)) contents.set(content, { content });
+    for (const value of [...contents.values()].slice(0, 100)) {
+      const requestId = ids.get(value.content) || newId();
+      rows.set(idFor(requestId), { id: idFor(requestId), requestId, content: value.content, createdAt: timestamp(),
+        status: value.status === "failed" ? "failed" : "unconfirmed", seenOnServer: false, confirmed: false,
+        error: typeof value.error === "string" && value.error ? value.error.slice(0, 240)
+          : ids.has(value.content) ? "Delivery was interrupted; retry to confirm." : "Restored draft; check the conversation before retrying." });
+    }
+  }
+  const persist = () => {
+    const pending = [...rows.values()].filter(row => !row.confirmed && !rejectedStorage.has(row.id));
+    if (pending.length > 100 || JSON.stringify(pending).length > 2 * 1024 * 1024) throw new Error("Too many pending agent messages. Retry existing messages before sending more.");
+    options.storage.write(pending.map(row => ({ ...row })));
+  };
+  const saveBestEffort = () => { try { persist(); } catch { /* Existing receipt IDs stay in memory; acknowledged messages remain sent. */ } };
+  const publish = () => options.changed();
+  const dispatch = (row: AgentOutboxRow): Promise<void> => {
+    const running = inFlight.get(row.id);
+    if (running) return running;
+    if (row.confirmed || row.seenOnServer) return Promise.resolve();
+    row.status = "sending"; delete row.error;
+    rejectedStorage.delete(row.id);
+    try { persist(); }
+    catch (error) {
+      // A refused new draft must not prevent older persisted drafts from
+      // draining their outbox. The refused row remains visibly retryable.
+      rejectedStorage.add(row.id);
+      row.status = "failed"; row.error = errorText(error); publish();
+      return Promise.reject(error);
+    }
+    // Both durable local storage and the visible row exist before POST begins.
+    publish();
+    const pending = Promise.resolve().then(() => options.submit(row.content, row.requestId)).then(receipt => {
+      if (receipt.messageId && receipt.messageId !== row.id) throw new Error("Agent returned a different message identity.");
+      if (!row.confirmed) {
+        row.seenOnServer = true;
+        row.status = receipt.status === "queued" ? "queued" : "complete";
+        delete row.error; saveBestEffort(); publish();
+      }
+    }).catch(error => {
+      // A canonical echo wins over a lost HTTP response. The message has already
+      // been admitted, so a transport error must not paint it failed or resend it.
+      if (row.confirmed || row.seenOnServer) return;
+      row.status = "failed"; row.error = errorText(error); saveBestEffort(); publish();
+      throw error;
+    }).finally(() => { inFlight.delete(row.id); });
+    inFlight.set(row.id, pending);
+    return pending;
+  };
+  // Complete migration even when the user leaves restored drafts untouched:
+  // their newly assigned identity/time must survive the next reload too.
+  saveBestEffort();
+  return {
+    send(content: string): Promise<void> {
+      if (!content.trim()) return Promise.reject(new Error("Enter a message."));
+      const requestId = newId();
+      const row: AgentOutboxRow = { id: idFor(requestId), requestId, content, createdAt: timestamp(), status: "sending", seenOnServer: false, confirmed: false };
+      rows.set(row.id, row);
+      return dispatch(row);
+    },
+    retry(messageId: string): Promise<void> {
+      const row = rows.get(messageId);
+      return row ? dispatch(row) : Promise.reject(new Error("This message is no longer pending."));
+    },
+    reconcile(room: RoomMessages): void {
+      const committed = new Map(room.messages.map(message => [message.id, message]));
+      const queued = new Set((room.activity?.queuedMessages || []).map(message => message.id));
+      let changed = false;
+      for (const row of rows.values()) {
+        if (committed.has(row.id)) {
+          confirmedEchoes.set(row.id, committed.get(row.id)!);
+          if (!row.confirmed || row.status !== "complete" || row.error) {
+            row.confirmed = true; row.seenOnServer = true; row.status = "complete"; delete row.error; changed = true;
+          }
+        } else if (!row.confirmed && queued.has(row.id) && (row.status !== "queued" || !row.seenOnServer || row.error)) {
+          row.seenOnServer = true; row.status = "queued"; delete row.error; changed = true;
+        }
+      }
+      if (changed) { saveBestEffort(); publish(); }
+    },
+    display(serverMessages: readonly Message[]): Message[] {
+      const displayed = new Map(serverMessages.map(message => {
+        const row = rows.get(message.id);
+        if (!row) return [message.id, message] as const;
+        const serverInfo = info(message);
+        const status = serverInfo.status === "complete" || row.confirmed ? "complete" : row.status;
+        return [message.id, { ...message, createdAt: row.createdAt,
+          metadata: { ...serverInfo, status, ...(row.error && status !== "complete" ? { error: row.error } : {}) } }] as const;
+      }));
+      for (const row of rows.values()) if (!displayed.has(row.id)) {
+        const echo = confirmedEchoes.get(row.id);
+        if (row.confirmed && !echo) continue;
+        displayed.set(row.id, echo ? { ...echo, createdAt: row.createdAt } : {
+          id: row.id, createdAt: row.createdAt, role: "user", content: row.content, name: "buzzert",
+          metadata: { kind: "message", status: row.status, channel: "sybil", requestId: row.requestId, ...(row.error ? { error: row.error } : {}) },
+        });
+      }
+      return [...displayed.values()].sort((a, b) => a.createdAt.localeCompare(b.createdAt));
+    },
+    snapshot(): AgentOutboxRow[] { return [...rows.values()].map(row => ({ ...row })); },
+  };
+}

web/src/lib/agent-room-presentation.ts

@@ -0,0 +1,120 @@
+import type { AgentRoomEvent, AgentRoomState } from './agent-room';
+import type { Message } from './api';
+
+type Scheduler = {
+  now(): number;
+  schedule(callback: () => void, delay: number): unknown;
+  cancel(handle: unknown): void;
+};
+export type AgentRoomPresenterOptions = {
+  initial: AgentRoomState;
+  reduceEvent(state: AgentRoomState, event: AgentRoomEvent): AgentRoomState;
+  /** Publish room and transcript typing together in one UI state update. */
+  publish(state: AgentRoomState, isTyping: boolean): void;
+  batchMs?: number;
+  typingDelayMs?: number;
+  clock?: Scheduler;
+};
+const quietTools = new Set(['user_message', 'notify_parent', 'clock_wait', 'event_decision']);
+const data = (message: Message): Record<string, unknown> => message.metadata && typeof message.metadata === 'object'
+  && !Array.isArray(message.metadata) ? message.metadata as Record<string, unknown> : {};
+const entryId = (message: Message): bigint | undefined => {
+  const value = data(message).entryId;
+  return (typeof value === 'string' && /^\d+$/.test(value)) || (typeof value === 'number' && Number.isSafeInteger(value) && value >= 0)
+    ? BigInt(value) : undefined;
+};
+const toolId = (message: Message) => typeof data(message).toolCallId === 'string' ? String(data(message).toolCallId) : message.id;
+
+/**
+ * Sybil Agent sends messages and activity as separate frames. This adapter publishes them
+ * as one room snapshot and latches transcript typing off when a new public reply
+ * arrives. An unchanged busy frame cannot restore the old turn's indicator.
+ * Actual activity remains untouched for the sidebar, workers, and queue display.
+ */
+export function createAgentRoomPresenter(options: AgentRoomPresenterOptions) {
+  const clock: Scheduler = options.clock || {
+    now: () => Date.now(),
+    schedule: (callback, delay) => setTimeout(callback, delay),
+    cancel: handle => clearTimeout(handle as ReturnType<typeof setTimeout>),
+  };
+  const batchMs = options.batchMs ?? 100;
+  const typingDelayMs = options.typingDelayMs ?? 250;
+  let state = options.initial;
+  let stopped = false;
+  let batchTimer: unknown;
+  let typingTimer: unknown;
+  let suppressTyping = false;
+  let candidateSince: number | undefined;
+  const seenTools = new Set(state.messages.filter(message => message.role === 'tool').map(toolId));
+  for (const tool of state.activity?.tools || []) seenTools.add(tool.id);
+
+  const reconcile = () => {
+    const work = Boolean(state.activity?.isThinking || (state.activity?.backgroundWorkers || 0) > 0);
+    if (!work || suppressTyping) candidateSince = undefined;
+    else if (candidateSince === undefined) candidateSince = clock.now();
+  };
+  const clearTypingTimer = () => { if (typingTimer !== undefined) clock.cancel(typingTimer); typingTimer = undefined; };
+  const flush = () => {
+    if (stopped) return;
+    if (batchTimer !== undefined) clock.cancel(batchTimer);
+    batchTimer = undefined; clearTypingTimer();
+    reconcile();
+    const isTyping = candidateSince !== undefined && clock.now() - candidateSince >= typingDelayMs;
+    options.publish(state, isTyping);
+    if (candidateSince !== undefined && !isTyping) typingTimer = clock.schedule(flush, Math.max(0, typingDelayMs - (clock.now() - candidateSince)));
+  };
+  const schedule = () => {
+    reconcile();
+    // A fixed maximum batch window avoids starving publication during bursts.
+    if (batchTimer === undefined) batchTimer = clock.schedule(flush, batchMs);
+  };
+  const release = () => { suppressTyping = false; };
+  const latch = () => {
+    suppressTyping = true;
+    candidateSince = undefined;
+    clearTypingTimer();
+  };
+
+  return {
+    /** replay:true is for initial historical loading, not ongoing catch-up. */
+    push(event: AgentRoomEvent, hints: { replay?: boolean } = {}) {
+      if (stopped) return;
+      const before = state;
+      const next = options.reduceEvent(before, event); // Validate/reduce before touching presentation state.
+      state = next;
+      if (event.type === 'messages') {
+        const existing = new Set(before.messages.map(message => message.id));
+        const fresh = event.messages.filter(message => !existing.has(message.id) && (entryId(message) ?? -1n) > BigInt(before.cursor))
+          .sort((left, right) => (entryId(left) || 0n) < (entryId(right) || 0n) ? -1 : 1);
+        if (!hints.replay) for (const message of fresh) {
+          const info = data(message);
+          if (message.role === 'assistant' && info.kind === 'message' && message.content.trim()) latch();
+          else if (suppressTyping && message.role === 'user') release();
+          else if (suppressTyping && message.role === 'tool' && info.status === 'initiated'
+            && !quietTools.has(String(info.toolName || message.name)) && !seenTools.has(toolId(message))) release();
+          if (message.role === 'tool') seenTools.add(toolId(message));
+        }
+        for (const message of event.messages) if (message.role === 'tool') seenTools.add(toolId(message));
+      } else if (event.type === 'activity') {
+        const activity = event.activity;
+        if (suppressTyping && (!activity.isThinking
+          || activity.tools.some(tool => !seenTools.has(tool.id) && !quietTools.has(tool.name)))) release();
+        for (const tool of activity.tools) seenTools.add(tool.id);
+      }
+      schedule();
+    },
+    /** Read-position updates share the pending snapshot; they never lose queued frames. */
+    update(change: (state: AgentRoomState) => AgentRoomState) {
+      if (stopped) return;
+      state = change(state); schedule();
+    },
+    flush,
+    dispose() {
+      stopped = true;
+      if (batchTimer !== undefined) clock.cancel(batchTimer);
+      batchTimer = undefined; clearTypingTimer();
+      // Unpublished input is deliberately discarded; reconnect replays from the
+      // last published cursor, so neither delivery nor read tracking gets ahead.
+    },
+  };
+}

web/src/lib/agent-room.ts

@@ -0,0 +1,216 @@
+import type { Message } from "./api";
+
+export type AgentActivity = {
+  state: "idle" | "thinking" | "tool" | "waiting" | "error";
+  isThinking: boolean;
+  queued: number;
+  backgroundWorkers: number;
+  tools: Array<{ id: string; name: string; status: string; output?: string }>;
+  queuedMessages: Message[];
+  wait?: { due: number; reason?: string };
+  error?: string;
+};
+export type AgentMessagesPage = {
+  conversationId: string; messages: Message[]; cursor: string | null;
+  beforeCursor: string | null; hasMore: boolean;
+};
+export type AgentDescriptor = {
+  id: string; title: string; model: string; activity: AgentActivity;
+  websocket: string; protocol: string;
+};
+export type AgentReadState = { cursor: string; updatedAt: string | null };
+export type AgentRoomState = {
+  conversationId: string | null; messages: Message[]; cursor: string;
+  readCursor: string; activity: AgentActivity | null;
+};
+export type AgentRoomEvent =
+  | ({ type: "messages" } & AgentMessagesPage)
+  | { type: "activity"; conversationId: string; activity: AgentActivity }
+  | { type: "ready"; conversationId?: string }
+  | { type: "pong" | "heartbeat" }
+  | { type: "error"; error?: string; message?: string };
+
+const object = (value: unknown): Record<string, unknown> | undefined => value !== null && typeof value === "object" && !Array.isArray(value)
+  ? value as Record<string, unknown> : undefined;
+const defined = (value: Record<string, unknown>) => Object.fromEntries(Object.entries(value).filter(([, item]) => item !== undefined));
+export function normalizeAgentCursor(value: unknown): string {
+  if (typeof value !== "string" || !/^\d+$/.test(value)) throw new Error("Invalid agent message cursor.");
+  return BigInt(value).toString();
+}
+export function advanceAgentCursor(current: string, incoming: string): string {
+  const a = normalizeAgentCursor(current), b = normalizeAgentCursor(incoming);
+  return BigInt(a) >= BigInt(b) ? a : b;
+}
+function metadata(message: Message) { return object(message.metadata); }
+function messageCursor(message: Message): string | null {
+  const entryId = metadata(message)?.entryId;
+  if (typeof entryId === "number" && Number.isSafeInteger(entryId) && entryId >= 0) return String(entryId);
+  return typeof entryId === "string" && /^\d+$/.test(entryId) ? normalizeAgentCursor(entryId) : null;
+}
+function mergeMessage(previous: Message, incoming: Message): Message {
+  const prior = metadata(previous), next = metadata(incoming);
+  // A delayed activity snapshot must not demote an already placed user message.
+  if (prior?.status === "complete" && next?.status === "queued") return previous;
+  if (prior?.kind !== "tool_call" || next?.kind !== "tool_call") return { ...previous, ...incoming };
+  const terminal = prior.status === "completed" || prior.status === "failed";
+  const staleStart = terminal && next.status === "initiated";
+  const merged = staleStart ? { ...defined(next), ...defined(prior) } : { ...defined(prior), ...defined(next) };
+  if (!next.startedAt && prior.startedAt) merged.startedAt = prior.startedAt;
+  const oldArgs = object(prior.args), newArgs = object(next.args);
+  if (oldArgs && (!newArgs || Object.keys(newArgs).length === 0)) merged.args = oldArgs;
+  return { ...previous, ...incoming, ...(staleStart ? { content: previous.content, name: previous.name } : {}),
+    createdAt: previous.createdAt < incoming.createdAt ? previous.createdAt : incoming.createdAt, metadata: merged };
+}
+
+/** Replays and separate tool-start/result pages update stable rows instead of duplicating them. */
+export function mergeAgentMessages(previous: readonly Message[], incoming: readonly Message[]): Message[] {
+  const messages = new Map(previous.map(message => [message.id, message]));
+  for (const message of incoming) {
+    const prior = messages.get(message.id);
+    messages.set(message.id, prior ? mergeMessage(prior, message) : message);
+  }
+  return [...messages.values()].sort((a, b) => {
+    const byTime = a.createdAt.localeCompare(b.createdAt);
+    if (byTime) return byTime;
+    const left = messageCursor(a), right = messageCursor(b);
+    if (left !== null && right !== null && left !== right) return BigInt(left) < BigInt(right) ? -1 : 1;
+    return a.id.localeCompare(b.id);
+  });
+}
+export const createAgentRoomState = (): AgentRoomState => ({
+  conversationId: null, messages: [], cursor: "0", readCursor: "0", activity: null,
+});
+export function applyAgentPage(state: AgentRoomState, page: AgentMessagesPage): AgentRoomState {
+  if (state.conversationId !== null && state.conversationId !== page.conversationId) throw new Error("Agent conversation changed; reload the page.");
+  // Validate before changing state. Failed delivery/decoding cannot advance replay position.
+  const cursor = page.cursor === null ? state.cursor : advanceAgentCursor(state.cursor, page.cursor);
+  return { ...state, conversationId: page.conversationId, messages: mergeAgentMessages(state.messages, page.messages), cursor };
+}
+export function applyAgentReadState(state: AgentRoomState, read: AgentReadState): AgentRoomState {
+  return { ...state, readCursor: advanceAgentCursor(state.readCursor, read.cursor) };
+}
+export function applyAgentEvent(state: AgentRoomState, event: AgentRoomEvent): AgentRoomState {
+  if (event.type === "messages") return applyAgentPage(state, event);
+  if (event.type !== "activity") return state;
+  if (state.conversationId !== null && state.conversationId !== event.conversationId) throw new Error("Agent activity belongs to a different conversation.");
+  return { ...state, conversationId: event.conversationId, activity: event.activity };
+}
+export function agentUnreadCount(state: AgentRoomState): number {
+  return state.messages.filter(message => {
+    if (message.role !== "assistant" || metadata(message)?.kind !== "message") return false;
+    const cursor = messageCursor(message);
+    return cursor !== null && BigInt(cursor) > BigInt(state.readCursor);
+  }).length;
+}
+export function agentDisplayMessages(state: AgentRoomState): Message[] {
+  const committed = new Set(state.messages.map(message => message.id));
+  return mergeAgentMessages(state.messages, (state.activity?.queuedMessages || []).filter(message => !committed.has(message.id)));
+}
+export function parseAgentRoomEvent(raw: string): AgentRoomEvent | null {
+  const event = object(JSON.parse(raw));
+  if (!event || typeof event.type !== "string") throw new Error("Invalid agent room frame.");
+  if (!["messages", "activity", "ready", "pong", "heartbeat", "error"].includes(event.type)) return null;
+  if (event.type === "messages") {
+    if (typeof event.conversationId !== "string" || !Array.isArray(event.messages) || typeof event.hasMore !== "boolean") throw new Error("Invalid agent messages frame.");
+    if (event.cursor !== null) normalizeAgentCursor(event.cursor);
+    for (const item of event.messages) {
+      const message = object(item);
+      if (!message || typeof message.id !== "string" || typeof message.content !== "string" || typeof message.createdAt !== "string"
+        || !["system", "user", "assistant", "tool"].includes(String(message.role))) throw new Error("Invalid agent message.");
+    }
+  }
+  if (event.type === "activity" && (typeof event.conversationId !== "string" || !object(event.activity))) throw new Error("Invalid agent activity frame.");
+  return event as AgentRoomEvent;
+}
+export function agentAuthFrame(token: string | null, cursor: string): string {
+  return JSON.stringify({ type: "auth", token, after: normalizeAgentCursor(cursor) });
+}
+
+export type AgentSubmission = { id: number; status: "queued" | "placed" | "done" | "unanswered"; messageId: string };
+export type AgentRoomClientOptions = { baseUrl?: string; token?: () => string | null; fetch?: typeof globalThis.fetch };
+export class AgentRoomClient {
+  private readonly baseUrl: string;
+  private readonly fetcher: typeof globalThis.fetch;
+  private readonly token: () => string | null;
+  constructor(options: AgentRoomClientOptions = {}) {
+    this.baseUrl = (options.baseUrl || "/api").replace(/\/+$/, "");
+    this.fetcher = (options.fetch || globalThis.fetch).bind(globalThis);
+    this.token = options.token || (() => null);
+  }
+  private async request<T>(path: string, init: RequestInit = {}): Promise<T> {
+    const headers = new Headers(init.headers);
+    const token = this.token();
+    if (token) headers.set("Authorization", `Bearer ${token}`);
+    if (init.body !== undefined) headers.set("Content-Type", "application/json");
+    const response = await this.fetcher(`${this.baseUrl}/v1/agent${path}`, { ...init, headers });
+    let value: unknown;
+    try { value = await response.json(); }
+    catch { throw new Error(`Agent API returned ${response.status} without a JSON response.`); }
+    if (!response.ok) {
+      const error = object(value);
+      throw new Error(typeof error?.message === "string" ? error.message : typeof error?.error === "string" ? error.error : `Agent API returned ${response.status}.`);
+    }
+    return value as T;
+  }
+  descriptor(signal?: AbortSignal): Promise<AgentDescriptor> { return this.request("", { signal }); }
+  messages(after: string, signal?: AbortSignal): Promise<AgentMessagesPage> {
+    return this.request(`/messages?${new URLSearchParams({ after: normalizeAgentCursor(after), limit: "200" })}`, { signal });
+  }
+  readState(signal?: AbortSignal): Promise<AgentReadState> { return this.request("/read-state", { signal }); }
+  markRead(cursor: string, signal?: AbortSignal): Promise<AgentReadState> {
+    return this.request("/read-state", { method: "PUT", body: JSON.stringify({ cursor: normalizeAgentCursor(cursor) }), signal });
+  }
+  async submit(content: string, requestId: string): Promise<AgentSubmission> {
+    const receipt = object(await this.request<unknown>("/messages", { method: "POST", body: JSON.stringify({ content, requestId }) }));
+    // An HTTP success alone does not acknowledge delivery. Keep the original
+    // outbox receipt key retryable if a proxy or upstream returns malformed data.
+    if (!receipt || typeof receipt.id !== "number" || !Number.isSafeInteger(receipt.id) || receipt.id <= 0
+      || typeof receipt.status !== "string" || !["queued", "placed", "done", "unanswered"].includes(receipt.status)
+      || receipt.messageId !== `inbound:sybil:${requestId}`) throw new Error("Agent API returned an invalid message receipt. Retry to confirm delivery.");
+    return receipt as AgentSubmission;
+  }
+  websocketUrl(pageUrl: string): string {
+    const url = new URL(`${this.baseUrl}/v1/agent/ws`, pageUrl);
+    url.protocol = url.protocol === "https:" ? "wss:" : "ws:";
+    return url.toString();
+  }
+}
+
+/** Read every newer page, including pages containing only invisible internal entries. */
+export async function syncAgentHistory(client: Pick<AgentRoomClient, "messages">, from: string,
+  apply: (page: AgentMessagesPage) => void, signal?: AbortSignal): Promise<void> {
+  let cursor = normalizeAgentCursor(from);
+  while (!signal?.aborted) {
+    const page = await client.messages(cursor, signal);
+    if (signal?.aborted) return;
+    const next = normalizeAgentCursor(page.cursor ?? cursor);
+    if (BigInt(next) < BigInt(cursor) || (page.hasMore && next === cursor)) throw new Error("Agent history did not advance its cursor.");
+    apply(page);
+    cursor = next;
+    if (!page.hasMore) return;
+  }
+}
+
+// Retained only to migrate receipt IDs from the former admission sender.
+export const AGENT_ADMISSION_STORAGE_KEY = "sybil.agent.pending-admissions.v1";
+
+/** Only successfully persisted, explicitly observed cursors become read state. */
+export function createAgentReadWriter(options: {
+  current: () => string; save: (cursor: string) => Promise<AgentReadState>; apply: (read: AgentReadState) => void;
+}) {
+  let desired = "0";
+  let running: Promise<void> | undefined;
+  const flush = (): Promise<void> => {
+    if (running) return running;
+    running = (async () => {
+      while (BigInt(desired) > BigInt(options.current())) {
+        const captured = desired;
+        const saved = await options.save(captured);
+        if (BigInt(normalizeAgentCursor(saved.cursor)) < BigInt(captured)) throw new Error("Agent read state did not acknowledge the seen messages.");
+        options.apply(saved);
+      }
+    })().finally(() => { running = undefined; });
+    return running;
+  };
+  return { request(cursor: string) { desired = advanceAgentCursor(desired, cursor); return flush(); }, flush };
+}

web/src/lib/api.ts

@@ -33,7 +33,9 @@ export type SearchWorkspaceItem = SearchSummary & {
   type: "search";
 };
 
-export type WorkspaceItem = ChatWorkspaceItem | SearchWorkspaceItem;
+export type WorkspaceItem = (ChatWorkspaceItem | SearchWorkspaceItem) & { searchSnippet?: string | null };
+
+export type WorkspaceSearchResponse = { items: WorkspaceItem[]; hasMore: boolean };
 
 export type Message = {
   id: string;
@@ -191,6 +193,14 @@ type CompletionStreamHandlers = {
   onError?: (payload: { message: string }) => void;
 };
 
+function dispatchCompletionStreamEvent(handlers: CompletionStreamHandlers, eventName: string, payload: any) {
+  if (eventName === "meta") handlers.onMeta?.(payload);
+  else if (eventName === "tool_call") handlers.onToolCall?.(payload);
+  else if (eventName === "delta") handlers.onDelta?.(payload);
+  else if (eventName === "done") handlers.onDone?.(payload);
+  else if (eventName === "error") handlers.onError?.(payload);
+}
+
 type CreateChatRequest = {
   title?: string;
   provider?: Provider;
@@ -220,6 +230,14 @@ export function getConfiguredToken() {
   return ENV_ADMIN_TOKEN;
 }
 
+export function getAuthToken() {
+  return authToken;
+}
+
+export function getApiBaseUrl() {
+  return API_BASE_URL;
+}
+
 export function setAuthToken(token: string | null) {
   authToken = token?.trim() || null;
 }
@@ -264,6 +282,10 @@ export async function listWorkspaceItems() {
   return data.items;
 }
 
+export function searchWorkspaceItems(query: string, signal?: AbortSignal) {
+  return api<WorkspaceSearchResponse>(`/v1/workspace-items?${new URLSearchParams({ q: query })}`, { signal });
+}
+
 export async function verifySession() {
   return api<{ authenticated: true; mode: "open" | "token" }>("/v1/auth/session");
 }
@@ -627,13 +649,34 @@ export async function runCompletionStream(
     signal: options?.signal,
   });
 
-  await readSseStream(response, (eventName, payload) => {
-    if (eventName === "meta") handlers.onMeta?.(payload);
-    else if (eventName === "tool_call") handlers.onToolCall?.(payload);
-    else if (eventName === "delta") handlers.onDelta?.(payload);
-    else if (eventName === "done") handlers.onDone?.(payload);
-    else if (eventName === "error") handlers.onError?.(payload);
+  await readSseStream(response, (eventName, payload) => dispatchCompletionStreamEvent(handlers, eventName, payload));
+}
+
+export async function runQuickQuestionStream(
+  body: {
+    provider: Provider;
+    model: string;
+    question: string;
+  },
+  handlers: CompletionStreamHandlers,
+  options?: { signal?: AbortSignal }
+) {
+  const headers = new Headers({
+    Accept: "text/event-stream",
+    "Content-Type": "application/json",
   });
+  if (authToken) {
+    headers.set("Authorization", `Bearer ${authToken}`);
+  }
+
+  const response = await fetch(`${API_BASE_URL}/v1/quick-questions/stream`, {
+    method: "POST",
+    headers,
+    body: JSON.stringify(body),
+    signal: options?.signal,
+  });
+
+  await readSseStream(response, (eventName, payload) => dispatchCompletionStreamEvent(handlers, eventName, payload));
 }
 
 export async function attachCompletionStream(chatId: string, handlers: CompletionStreamHandlers, options?: { signal?: AbortSignal }) {
@@ -650,11 +693,5 @@ export async function attachCompletionStream(chatId: string, handlers: Completio
     signal: options?.signal,
   });
 
-  await readSseStream(response, (eventName, payload) => {
-    if (eventName === "meta") handlers.onMeta?.(payload);
-    else if (eventName === "tool_call") handlers.onToolCall?.(payload);
-    else if (eventName === "delta") handlers.onDelta?.(payload);
-    else if (eventName === "done") handlers.onDone?.(payload);
-    else if (eventName === "error") handlers.onError?.(payload);
-  });
+  await readSseStream(response, (eventName, payload) => dispatchCompletionStreamEvent(handlers, eventName, payload));
 }

web/src/lib/chat-stream-presentation.ts

@@ -0,0 +1,139 @@
+import type { Message, ToolCallEvent } from "@/lib/api";
+
+export const STREAMING_ASSISTANT_METADATA_KIND = "streaming_assistant";
+
+export type StreamingAssistantMetadata = {
+  kind: typeof STREAMING_ASSISTANT_METADATA_KIND;
+  ephemeralTrace: string | null;
+  traceRevision: number;
+};
+
+export type StreamingAttemptState = {
+  segmentText: string;
+  initiatedToolCallIds: string[];
+};
+
+export type StreamingToolCallTransition = {
+  attempt: StreamingAttemptState;
+  startedNewToolCall: boolean;
+  traceText: string | null;
+};
+
+export function createStreamingAssistantMetadata(): StreamingAssistantMetadata {
+  return {
+    kind: STREAMING_ASSISTANT_METADATA_KIND,
+    ephemeralTrace: null,
+    traceRevision: 0,
+  };
+}
+
+export function asStreamingAssistantMetadata(value: unknown): StreamingAssistantMetadata | null {
+  if (!value || typeof value !== "object" || Array.isArray(value)) return null;
+  const record = value as Record<string, unknown>;
+  if (record.kind !== STREAMING_ASSISTANT_METADATA_KIND) return null;
+
+  return {
+    kind: STREAMING_ASSISTANT_METADATA_KIND,
+    ephemeralTrace: typeof record.ephemeralTrace === "string" ? record.ephemeralTrace : null,
+    traceRevision:
+      typeof record.traceRevision === "number" && Number.isFinite(record.traceRevision)
+        ? Math.max(0, Math.floor(record.traceRevision))
+        : 0,
+  };
+}
+
+export function createStreamingAttemptState(): StreamingAttemptState {
+  return {
+    segmentText: "",
+    initiatedToolCallIds: [],
+  };
+}
+
+export function appendStreamingDelta(attempt: StreamingAttemptState, delta: string): StreamingAttemptState {
+  if (!delta) return attempt;
+  return {
+    ...attempt,
+    segmentText: attempt.segmentText + delta,
+  };
+}
+
+export function transitionStreamingToolCall(
+  attempt: StreamingAttemptState,
+  event: Pick<ToolCallEvent, "toolCallId" | "status">
+): StreamingToolCallTransition {
+  if (event.status !== "initiated" || attempt.initiatedToolCallIds.includes(event.toolCallId)) {
+    return { attempt, startedNewToolCall: false, traceText: null };
+  }
+
+  const traceText = attempt.segmentText.trim();
+  return {
+    attempt: {
+      segmentText: "",
+      initiatedToolCallIds: attempt.initiatedToolCallIds.concat(event.toolCallId),
+    },
+    startedNewToolCall: true,
+    traceText: traceText || null,
+  };
+}
+
+function updateStreamingAssistant(
+  messages: Message[],
+  assistantMessagePrefix: string,
+  update: (message: Message, metadata: StreamingAssistantMetadata) => Message
+) {
+  let didUpdate = false;
+  const nextMessages = messages.map((message, index, all) => {
+    const isTarget = index === all.length - 1 && message.id.startsWith(assistantMessagePrefix);
+    if (!isTarget) return message;
+    didUpdate = true;
+    return update(message, asStreamingAssistantMetadata(message.metadata) ?? createStreamingAssistantMetadata());
+  });
+  return didUpdate ? nextMessages : messages;
+}
+
+export function setStreamingAssistantSegment(messages: Message[], assistantMessagePrefix: string, segmentText: string) {
+  return updateStreamingAssistant(messages, assistantMessagePrefix, (message, metadata) => ({
+    ...message,
+    content: segmentText,
+    metadata: {
+      ...metadata,
+      ephemeralTrace: null,
+    } satisfies StreamingAssistantMetadata,
+  }));
+}
+
+export function promoteStreamingAssistantTrace(messages: Message[], assistantMessagePrefix: string, traceText: string) {
+  const normalizedTrace = traceText.trim();
+  return updateStreamingAssistant(messages, assistantMessagePrefix, (message, metadata) => ({
+    ...message,
+    content: "",
+    metadata: normalizedTrace
+      ? ({
+          ...metadata,
+          ephemeralTrace: normalizedTrace,
+          traceRevision: metadata.traceRevision + 1,
+        } satisfies StreamingAssistantMetadata)
+      : metadata,
+  }));
+}
+
+export function finalizeStreamingAssistant(messages: Message[], assistantMessagePrefix: string, finalText: string) {
+  return updateStreamingAssistant(messages, assistantMessagePrefix, (message, metadata) => ({
+    ...message,
+    content: finalText,
+    metadata: {
+      ...metadata,
+      ephemeralTrace: null,
+    } satisfies StreamingAssistantMetadata,
+  }));
+}
+
+export function clearStreamingAssistantTrace(messages: Message[], assistantMessagePrefix: string) {
+  return updateStreamingAssistant(messages, assistantMessagePrefix, (message, metadata) => ({
+    ...message,
+    metadata: {
+      ...metadata,
+      ephemeralTrace: null,
+    } satisfies StreamingAssistantMetadata,
+  }));
+}

web/src/lib/quick-question.ts

@@ -0,0 +1,25 @@
+import type { Provider } from "./api";
+
+export const QUICK_QUESTION_DEFAULT_MODEL = "chat-latest";
+
+export function pickQuickQuestionModel(provider: Provider, options: string[], preferred: string | null = null) {
+  if (preferred?.trim()) return preferred.trim();
+  if (provider === "openai") return QUICK_QUESTION_DEFAULT_MODEL;
+  return options[0] ?? "";
+}
+
+export function buildQuickQuestionRequest({
+  provider,
+  model,
+  content,
+}: {
+  provider: Provider;
+  model: string;
+  content: string;
+}) {
+  return {
+    provider,
+    model,
+    question: content,
+  };
+}

web/src/lib/sidebar-selection.ts

@@ -1,6 +1,17 @@
-export type SidebarSelection = { kind: "chat" | "search"; id: string };
+export type HistorySelection = { kind: "chat" | "search"; id: string };
+export type SidebarSelection = HistorySelection | { kind: "agent"; id: "root" };
 
-type WorkspaceSelectionItem = { type: SidebarSelection["kind"]; id: string };
+type WorkspaceSelectionItem = { type: HistorySelection["kind"]; id: string };
+
+export function retainSelectedHistoryItem<T extends WorkspaceSelectionItem>(
+  items: T[], selection: SidebarSelection | null, retained: T | null
+): T[] {
+  if (retained && selection?.kind === retained.type && selection.id === retained.id &&
+      !items.some((item) => item.type === retained.type && item.id === retained.id)) {
+    return [...items, retained];
+  }
+  return items;
+}
 
 type ResolveSidebarSelectionOptions = {
   initialSelection?: SidebarSelection;
@@ -14,6 +25,8 @@ export function resolveSidebarSelectionAfterRefresh(
 ): SidebarSelection | null {
   const hasItem = (candidate: SidebarSelection | null | undefined) => {
     if (!candidate) return false;
+    // The persistent agent is always available, independent of saved history.
+    if (candidate.kind === "agent") return true;
     return workspaceItems.some((item) => item.type === candidate.kind && item.id === candidate.id);
   };
 

web/src/lib/workspace-search.ts

@@ -0,0 +1,24 @@
+import type { WorkspaceSearchResponse } from "./api";
+
+// Cleanup cancels both the debounce and fetch, and also rejects late responses
+// from transports that finish after cancellation.
+export function scheduleWorkspaceSearch(
+  query: string,
+  request: (query: string, signal: AbortSignal) => Promise<WorkspaceSearchResponse>,
+  onResult: (result: WorkspaceSearchResponse) => void,
+  onError: (error: unknown) => void
+) {
+  const controller = new AbortController();
+  const timer = setTimeout(async () => {
+    try {
+      const result = await request(query, controller.signal);
+      if (!controller.signal.aborted) onResult(result);
+    } catch (error) {
+      if (!controller.signal.aborted) onError(error);
+    }
+  }, 300);
+  return () => {
+    clearTimeout(timer);
+    controller.abort();
+  };
+}

web/tailwind.config.ts

@@ -29,9 +29,14 @@ const config: Config = {
         },
       },
       borderRadius: {
+        // Keep surfaces square at every size; rounded-full is reserved for dots.
+        DEFAULT: "var(--radius)",
+        sm: "var(--radius)",
+        md: "var(--radius)",
         lg: "var(--radius)",
-        md: "calc(var(--radius) - 2px)",
-        sm: "calc(var(--radius) - 4px)",
+        xl: "var(--radius)",
+        "2xl": "var(--radius)",
+        "3xl": "var(--radius)",
       },
     },
   },

web/tests/agent-optimistic-outbox.test.mjs

@@ -0,0 +1,139 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { createAgentOptimisticOutbox } from "../src/lib/agent-optimistic-outbox.ts";
+
+const canonical = (id, content = "Hello", time = "2026-10-14T12:01:00.000Z") => ({
+  id, content, createdAt: time, role: "user", name: "buzzert",
+  metadata: { kind: "message", entryId: 123, status: "complete", channel: "sybil" },
+});
+const deferred = () => { let resolve, reject; const promise = new Promise((yes, no) => { resolve = yes; reject = no; }); return { promise, resolve, reject }; };
+function fixture(submit, legacy = {}) {
+  let stored = null;
+  let serial = 0;
+  const changes = [];
+  let box;
+  const storage = { read: () => structuredClone(stored), write: rows => { stored = structuredClone(rows); }, ...legacy };
+  const open = () => {
+    box = createAgentOptimisticOutbox({ submit, storage, newId: () => `request-${++serial}`, now: () => Date.parse("2026-10-14T12:00:00.000Z"), changed: () => changes.push(box?.snapshot()) });
+    return box;
+  };
+  return { open, changes, stored: () => stored, storage };
+}
+
+test("send publishes and persists one canonical row before POST, retaining it through receipt and echo", async () => {
+  const network = deferred(); let called = false;
+  const state = fixture(async (_content, requestId) => { called = true; assert.equal(requestId, "request-1"); return network.promise; });
+  const box = state.open();
+  const done = box.send("Hello");
+  const first = box.display([])[0];
+  assert.equal(called, false, "network starts after synchronous local publication");
+  assert.equal(first.id, "inbound:sybil:request-1");
+  assert.equal(first.metadata.status, "sending");
+  assert.equal(state.stored()[0].id, first.id);
+  assert.equal(state.changes.length, 1);
+  network.resolve({ status: "queued", messageId: first.id }); await done;
+  assert.equal(box.display([])[0].metadata.status, "queued");
+  assert.equal(box.display([])[0].createdAt, first.createdAt);
+  const echo = canonical(first.id);
+  box.reconcile({ messages: [echo] });
+  const displayed = box.display([echo]);
+  assert.equal(displayed.length, 1);
+  assert.equal(displayed[0].id, first.id);
+  assert.equal(displayed[0].createdAt, first.createdAt, "server receipt time cannot reorder the optimistic bubble");
+  assert.equal(displayed[0].metadata.status, "complete");
+  assert.deepEqual(state.stored(), []);
+  assert.equal(box.display([]).length, 1, "even a stale display snapshot cannot create an acknowledgment gap");
+});
+
+test("canonical echo before an HTTP failure wins and never paints a delivered bubble failed", async () => {
+  const network = deferred(); const state = fixture(() => network.promise); const box = state.open();
+  const done = box.send("Hello"); const row = box.display([])[0];
+  const echo = canonical(row.id);
+  box.reconcile({ messages: [echo] });
+  network.reject(new Error("response lost"));
+  await done;
+  assert.equal(box.display([echo])[0].metadata.status, "complete");
+  assert.equal(box.display([echo])[0].metadata.error, undefined);
+  assert.deepEqual(state.stored(), []);
+});
+
+test("queue echo before HTTP acknowledgment preserves one row when the queue snapshot disappears", async () => {
+  const network = deferred(); const state = fixture(() => network.promise); const box = state.open();
+  const done = box.send("Hello"); const row = box.display([])[0];
+  const queued = { ...canonical(row.id), metadata: { kind: "message", status: "queued" } };
+  box.reconcile({ messages: [], activity: { queuedMessages: [queued] } });
+  assert.equal(box.display([queued]).length, 1);
+  box.reconcile({ messages: [], activity: { queuedMessages: [] } });
+  assert.equal(box.display([])[0].id, row.id, "queue-to-history transfer must not hide the bubble");
+  network.reject(new Error("HTTP acknowledgment lost")); await done;
+  assert.equal(box.display([])[0].metadata.status, "queued");
+});
+
+test("failed retry and reload preserve the message ID, timestamp, and original receipt key", async () => {
+  const requests = []; let fail = true;
+  const state = fixture(async (content, requestId) => {
+    requests.push({ content, requestId });
+    if (fail) throw new Error("network down");
+    return { status: "placed", messageId: `inbound:sybil:${requestId}` };
+  });
+  let box = state.open();
+  await assert.rejects(box.send("Hello"), /network down/);
+  const failed = box.display([])[0];
+  assert.equal(failed.metadata.status, "failed");
+  assert.equal(failed.metadata.error, "network down");
+  box = state.open();
+  assert.equal(requests.length, 1, "restoring drafts never sends automatically");
+  assert.equal(box.display([])[0].metadata.status, "failed");
+  assert.equal(box.display([])[0].createdAt, failed.createdAt);
+  fail = false;
+  const retry = box.retry(failed.id);
+  assert.equal(box.display([])[0].metadata.status, "sending");
+  await retry;
+  assert.deepEqual(requests.map(row => row.requestId), ["request-1", "request-1"]);
+  assert.equal(box.display([]).length, 1);
+  assert.equal(box.display([])[0].metadata.status, "complete");
+  assert.equal(box.display([])[0].createdAt, failed.createdAt);
+});
+
+test("in-flight reload becomes unconfirmed and two identical sends remain independent rows", async () => {
+  const first = deferred(), second = deferred(); let calls = 0;
+  const state = fixture(() => ++calls === 1 ? first.promise : second.promise); const box = state.open();
+  const send1 = box.send("Same content"), send2 = box.send("Same content");
+  const rows = box.display([]);
+  assert.equal(rows.length, 2);
+  assert.notEqual(rows[0].id, rows[1].id);
+  assert.ok(rows[0].createdAt < rows[1].createdAt, "same-millisecond sends retain insertion order");
+  const reloaded = state.open();
+  assert.deepEqual(reloaded.display([]).map(row => row.metadata.status), ["unconfirmed", "unconfirmed"]);
+  assert.deepEqual(reloaded.display([]).map(row => row.id), rows.map(row => row.id));
+  first.resolve({ status: "queued" }); second.resolve({ status: "queued" });
+  await Promise.all([send1, send2]);
+});
+
+test("legacy failed cards join their saved receipt keys without automatic delivery", () => {
+  const state = fixture(() => assert.fail("migration must not submit"), {
+    legacyRows: () => [{ content: "old message", status: "failed", error: "old error" }],
+    legacyIds: () => [{ content: "old message", requestId: "old-request" }],
+  });
+  const box = state.open();
+  assert.equal(box.display([])[0].id, "inbound:sybil:old-request");
+  assert.equal(box.display([])[0].metadata.status, "failed");
+  assert.equal(box.display([])[0].metadata.error, "old error");
+  const migrated = state.stored()[0];
+  assert.equal(migrated.id, "inbound:sybil:old-request");
+  assert.equal(state.open().display([])[0].createdAt, migrated.createdAt, "migration is saved before any retry");
+});
+
+test("storage failures still leave a failed inline row and prevent untracked POST", async () => {
+  let calls = 0, visible = 0;
+  const box = createAgentOptimisticOutbox({ submit: async () => { calls++; return { status: "queued" }; },
+    storage: { read: () => null, write: () => { throw new Error("storage unavailable"); } },
+    newId: () => "request", changed: () => visible++,
+  });
+  const sending = box.send("Do not lose this text");
+  assert.equal(box.display([])[0].metadata.status, "failed");
+  assert.equal(box.display([])[0].content, "Do not lose this text");
+  assert.equal(visible, 1);
+  await assert.rejects(sending, /storage unavailable/);
+  assert.equal(calls, 0);
+});

web/tests/agent-room-presentation.test.mjs

@@ -0,0 +1,100 @@
+import assert from 'node:assert/strict';
+import test from 'node:test';
+import { applyAgentEvent, applyAgentReadState, createAgentRoomState } from '../src/lib/agent-room.ts';
+import { createAgentRoomPresenter } from '../src/lib/agent-room-presentation.ts';
+
+const message = (entry, role = 'assistant', extra = {}) => ({ id: `entry:${entry}`, createdAt: '2026-10-14T00:00:00.000Z', role,
+  content: `Message ${entry}`, metadata: { kind: 'message', entryId: entry, status: 'complete' }, ...extra });
+const page = (cursor, messages) => ({ type: 'messages', conversationId: '1', messages, cursor: String(cursor), beforeCursor: '0', hasMore: false });
+const activity = (isThinking, extra = {}) => ({ type: 'activity', conversationId: '1', activity: {
+  state: isThinking ? 'thinking' : 'idle', isThinking, queued: 0, backgroundWorkers: 0, tools: [], queuedMessages: [], ...extra,
+} });
+function fixture(initial = createAgentRoomState()) {
+  let now = 0, sequence = 0;
+  const tasks = new Map();
+  const frames = [];
+  const clock = { now: () => now, schedule(callback, delay) { const id = ++sequence; tasks.set(id, { at: now + delay, callback }); return id; }, cancel(id) { tasks.delete(id); } };
+  const presenter = createAgentRoomPresenter({ initial, reduceEvent: applyAgentEvent, publish: (room, isTyping) => frames.push({ room, isTyping }), clock });
+  const advance = ms => {
+    const end = now + ms;
+    while (true) {
+      const next = [...tasks].sort((a, b) => a[1].at - b[1].at)[0];
+      if (!next || next[1].at > end) break;
+      now = next[1].at; tasks.delete(next[0]); next[1].callback();
+    }
+    now = end;
+  };
+  return { presenter, advance, frames, last: () => frames.at(-1) };
+}
+
+test('a reply and removal of its old thinking indicator publish together even when idle is delayed', () => {
+  const f = fixture();
+  f.presenter.push(activity(true)); f.advance(250); assert.equal(f.last().isTyping, true);
+  f.presenter.push(page(2, [message(2)])); f.advance(100);
+  assert.equal(f.last().room.messages.length, 1); assert.equal(f.last().isTyping, false);
+  assert.equal(f.last().room.activity.isThinking, true, 'raw sidebar activity remains unchanged');
+  f.presenter.push(activity(true)); f.advance(1000); assert.equal(f.last().isTyping, false, 'identical busy snapshots cannot revive stale typing');
+  f.presenter.push(activity(false)); f.advance(100);
+  assert.ok(f.frames.every(frame => !frame.room.messages.length || !frame.isTyping));
+});
+
+test('split messages/activity frames batch without reply-plus-thinking flashes or transient typing starts', () => {
+  const f = fixture();
+  f.presenter.push(activity(true)); f.advance(50);
+  f.presenter.push(page(2, [message(1, 'user'), message(2)])); f.advance(20);
+  f.presenter.push(activity(false)); f.advance(30);
+  assert.equal(f.frames.length, 1); assert.equal(f.last().room.messages.length, 2); assert.equal(f.last().isTyping, false);
+  f.advance(1000); assert.equal(f.frames.length, 1);
+});
+
+test('new user work and new tools can start typing again, but delivery tools cannot', () => {
+  const f = fixture();
+  f.presenter.push(activity(true)); f.advance(250);
+  f.presenter.push(page(2, [message(2)])); f.advance(100);
+  f.presenter.push(activity(true, { tools: [{ id: 'send-reply', name: 'user_message', status: 'running' }] }));
+  f.advance(300); assert.equal(f.last().isTyping, false);
+  f.presenter.push(page(3, [message(3, 'user')])); f.advance(250); assert.equal(f.last().isTyping, true);
+  f.presenter.push(page(4, [message(4)])); f.advance(100); assert.equal(f.last().isTyping, false);
+  f.presenter.push(activity(true, { tools: [{ id: 'new-browser', name: 'browser', status: 'running' }] }));
+  f.advance(250); assert.equal(f.last().isTyping, true);
+});
+
+test('initial history and duplicate replay do not suppress genuine new activity', () => {
+  const f = fixture();
+  f.presenter.push(activity(true));
+  f.presenter.push(page(7, [message(5), message(7)]), { replay: true });
+  f.advance(250); assert.equal(f.last().isTyping, true);
+  f.presenter.push(page(7, [message(5), message(7)])); f.advance(100);
+  assert.equal(f.last().isTyping, true); assert.equal(f.last().room.messages.length, 2);
+  f.presenter.push(page(9, [message(9)])); f.advance(100); assert.equal(f.last().isTyping, false);
+  f.presenter.push(activity(false)); f.advance(100);
+  f.presenter.push(activity(true)); f.advance(250); assert.equal(f.last().isTyping, true);
+  f.presenter.push(page(9, [message(9)])); f.advance(100); assert.equal(f.last().isTyping, true);
+});
+
+test('independent background work remains visible after the root yields', () => {
+  const f = fixture();
+  f.presenter.push(activity(true, { backgroundWorkers: 1 })); f.advance(250);
+  f.presenter.push(page(2, [message(2)])); f.advance(100); assert.equal(f.last().isTyping, false);
+  f.presenter.push(activity(false, { backgroundWorkers: 1 })); f.advance(250);
+  assert.equal(f.last().isTyping, true); assert.equal(f.last().room.activity.isThinking, false);
+});
+
+test('bursts preserve users, tools, queued messages, cursor, and read state in bounded publications', () => {
+  const f = fixture();
+  f.presenter.push(page(1, [message(1, 'user')])); f.advance(50);
+  const tool = message(2, 'tool', { metadata: { kind: 'tool_call', toolCallId: 'browser-1', toolName: 'browser', status: 'initiated', entryId: 2 } });
+  f.presenter.push(page(2, [tool])); f.advance(40);
+  const queued = message(3, 'user', { metadata: { kind: 'message', status: 'queued' } });
+  f.presenter.push(activity(true, { queued: 1, queuedMessages: [queued] }));
+  f.presenter.update(state => applyAgentReadState(state, { cursor: '1', updatedAt: null })); f.advance(10);
+  assert.equal(f.frames.length, 1); assert.equal(f.last().room.cursor, '2'); assert.equal(f.last().room.readCursor, '1');
+  assert.equal(f.last().room.messages.length, 2); assert.deepEqual(f.last().room.activity.queuedMessages, [queued]);
+  f.presenter.push(page(4, [{ ...tool, metadata: { ...tool.metadata, status: 'completed', entryId: 4 } }]));
+  f.presenter.flush(); assert.equal(f.last().room.messages.length, 2); assert.equal(f.last().room.cursor, '4');
+});
+
+test('disposing an unpublished batch leaves delivery and read cursors at the last rendered snapshot', () => {
+  const f = fixture(); f.presenter.push(page(2, [message(2)])); f.presenter.dispose(); f.advance(1000);
+  assert.equal(f.frames.length, 0);
+});

web/tests/agent-room.test.mjs

@@ -0,0 +1,210 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import {
+  AgentRoomClient, advanceAgentCursor, agentAuthFrame, agentDisplayMessages, agentUnreadCount,
+  applyAgentEvent, applyAgentPage, applyAgentReadState, createAgentReadWriter,
+  createAgentRoomState, mergeAgentMessages, normalizeAgentCursor, parseAgentRoomEvent, syncAgentHistory,
+} from "../src/lib/agent-room.ts";
+import { createAgentOptimisticOutbox } from "../src/lib/agent-optimistic-outbox.ts";
+
+const message = (id, role = "assistant", extra = {}) => ({
+  id: `entry:${id}`, createdAt: `2026-10-14T10:00:${String(id % 60).padStart(2, "0")}.000Z`,
+  role, content: `message ${id}`, name: role === "assistant" ? "Sybil Agent" : "buzzert",
+  metadata: { kind: "message", entryId: id, status: "complete" }, ...extra,
+});
+const page = (cursor, messages = [], hasMore = false) => ({
+  conversationId: "1", messages, cursor, beforeCursor: messages.length ? String(messages[0].metadata.entryId) : cursor, hasMore,
+});
+const activity = (queuedMessages = []) => ({ state: "idle", isThinking: false, queued: queuedMessages.length, backgroundWorkers: 0, tools: [], queuedMessages });
+
+test("tool starts and results merge across replay without downgrading completed cards", () => {
+  const start = message(2, "tool", { id: "tool:1:call1", metadata: { kind: "tool_call", entryId: 2, status: "initiated", args: { query: "cookie" }, startedAt: "2026-10-14T10:00:00Z" } });
+  const done = message(3, "tool", { id: start.id, metadata: { kind: "tool_call", entryId: 3, status: "completed", resultPreview: "found" } });
+  const messages = mergeAgentMessages(mergeAgentMessages([], [start]), [done, start]);
+  assert.equal(messages.length, 1);
+  assert.equal(messages[0].metadata.status, "completed");
+  assert.equal(messages[0].metadata.resultPreview, "found");
+  assert.deepEqual(messages[0].metadata.args, { query: "cookie" });
+  assert.equal(messages[0].metadata.entryId, 3);
+  assert.equal(messages[0].createdAt, start.createdAt);
+});
+
+test("initial synchronization reads every page including invisible entries and counts only unread assistant messages", async () => {
+  const requests = [];
+  const responses = [page("2", [message(1), message(2, "user")], true), page("4", [], true), page("7", [message(5), message(6, "tool"), message(7)], false)];
+  let state = createAgentRoomState();
+  await syncAgentHistory({ messages: async after => { requests.push(after); return responses.shift(); } }, "0", update => { state = applyAgentPage(state, update); });
+  assert.deepEqual(requests, ["0", "2", "4"]);
+  assert.equal(state.cursor, "7");
+  assert.equal(agentUnreadCount(state), 3);
+  state = applyAgentReadState(state, { cursor: "5", updatedAt: null });
+  assert.equal(agentUnreadCount(state), 1);
+  state = applyAgentReadState(state, { cursor: "1", updatedAt: null });
+  assert.equal(state.readCursor, "5", "a stale tab response cannot regress read state");
+  assert.equal(agentUnreadCount(state), 1);
+});
+
+test("queued user messages keep stable IDs and disappear if withdrawn or promoted", () => {
+  const pending = message(1, "user", { id: "inbound:sybil:request1", metadata: { kind: "message", status: "queued" } });
+  let state = applyAgentEvent(createAgentRoomState(), { type: "activity", conversationId: "1", activity: activity([pending]) });
+  assert.equal(agentDisplayMessages(state).length, 1);
+  assert.equal(state.cursor, "0", "pending rows never advance replay or read cursors");
+  assert.equal(agentUnreadCount(state), 0);
+  state = applyAgentPage(state, page("4", [message(4, "user", { id: pending.id })]));
+  assert.equal(agentDisplayMessages(state).length, 1);
+  assert.equal(agentDisplayMessages(state)[0].metadata.status, "complete");
+  state = applyAgentEvent(state, { type: "activity", conversationId: "1", activity: activity() });
+  assert.equal(agentDisplayMessages(state).length, 1, "committed rows survive queue removal");
+  const onlyPending = applyAgentEvent(createAgentRoomState(), { type: "activity", conversationId: "1", activity: activity([pending]) });
+  assert.deepEqual(agentDisplayMessages(applyAgentEvent(onlyPending, { type: "activity", conversationId: "1", activity: activity() })), []);
+});
+
+test("failed or stalled page delivery preserves the last applied cursor for replay", async () => {
+  let state = createAgentRoomState();
+  let calls = 0;
+  await assert.rejects(syncAgentHistory({ messages: async () => {
+    if (++calls === 1) return page("10", [message(10)], true);
+    throw new Error("network unavailable");
+  } }, "0", next => { state = applyAgentPage(state, next); }), /network unavailable/);
+  assert.equal(state.cursor, "10");
+  await assert.rejects(syncAgentHistory({ messages: async () => page("10", [], true) }, state.cursor, next => { state = applyAgentPage(state, next); }), /did not advance/);
+  assert.equal(state.cursor, "10");
+  assert.throws(() => applyAgentPage(state, page("bad", [message(11)])), /cursor/);
+  assert.equal(state.messages.length, 1);
+  assert.throws(() => applyAgentPage(state, { ...page("11"), conversationId: "different" }), /conversation changed/);
+  assert.equal(advanceAgentCursor("9007199254740993", "9007199254740992"), "9007199254740993");
+  assert.throws(() => normalizeAgentCursor("-1"), /cursor/);
+});
+
+test("read acknowledgments use only the rendered cursor and failed persistence retains unread replies", async () => {
+  let state = applyAgentPage(createAgentRoomState(), page("5", [message(5)]));
+  const renderedCursor = state.cursor;
+  state = applyAgentPage(state, page("6", [message(6)]));
+  const saved = [];
+  let fail = true;
+  const writer = createAgentReadWriter({
+    current: () => state.readCursor,
+    save: async cursor => { saved.push(cursor); if (fail) throw new Error("database offline"); return { cursor, updatedAt: null }; },
+    apply: read => { state = applyAgentReadState(state, read); },
+  });
+  await assert.rejects(writer.request(renderedCursor), /database offline/);
+  assert.equal(state.readCursor, "0");
+  assert.equal(agentUnreadCount(state), 2);
+  fail = false;
+  await writer.flush();
+  assert.deepEqual(saved, ["5", "5"]);
+  assert.equal(state.readCursor, "5");
+  assert.equal(agentUnreadCount(state), 1, "message6 arrived after render and must remain unread");
+});
+
+test("read acknowledgments coalesce concurrent seen snapshots and reject unacknowledged cursors", async () => {
+  let cursor = "0";
+  let release;
+  const gate = new Promise(resolve => { release = resolve; });
+  const saves = [];
+  const writer = createAgentReadWriter({ current: () => cursor, save: async value => {
+    saves.push(value); if (value === "3") await gate; return { cursor: value, updatedAt: null };
+  }, apply: read => { cursor = read.cursor; } });
+  const first = writer.request("3"), second = writer.request("7");
+  release(); await Promise.all([first, second]);
+  assert.deepEqual(saves, ["3", "7"]);
+  assert.equal(cursor, "7");
+  const broken = createAgentReadWriter({ current: () => "0", save: async () => ({ cursor: "0", updatedAt: null }), apply: () => assert.fail("must not apply a failed acknowledgment") });
+  await assert.rejects(broken.request("2"), /did not acknowledge/);
+});
+
+test("failed persistence after acknowledgment retains the original receipt key across reload", async () => {
+  let stored = [], writes = 0;
+  const storage = { read: () => structuredClone(stored), write: rows => {
+    if (++writes === 3) throw new Error("storage unavailable after acknowledgment");
+    stored = structuredClone(rows);
+  } };
+  const requests = [];
+  const open = () => createAgentOptimisticOutbox({
+    storage, changed: () => {}, newId: () => "stable-request",
+    submit: async (_content, requestId) => {
+      requests.push(requestId);
+      return { status: "placed", messageId: `inbound:sybil:${requestId}` };
+    },
+  });
+  const original = open();
+  await original.send("retry me");
+  const accepted = original.display([])[0];
+  assert.equal(accepted.metadata.status, "complete", "a storage error cannot undo server acceptance");
+  assert.equal(stored[0].status, "sending", "the pre-admission receipt remains durable if its update fails");
+  const restored = open();
+  assert.deepEqual(requests, ["stable-request"], "reload never retries automatically");
+  const uncertain = restored.display([])[0];
+  assert.equal(uncertain.metadata.status, "unconfirmed");
+  assert.equal(uncertain.id, accepted.id);
+  assert.equal(uncertain.createdAt, accepted.createdAt);
+  await restored.retry(uncertain.id);
+  assert.deepEqual(requests, ["stable-request", "stable-request"], "manual retry cannot create a second admission");
+  restored.reconcile({ messages: [message(1, "user", { id: accepted.id, content: "retry me" })] });
+  assert.deepEqual(stored, [], "canonical echo finishes local cleanup");
+});
+
+test("client uses Sybil auth and routes while websocket credentials remain in the first frame", async () => {
+  const calls = [];
+  let token = "first-session-token";
+  const client = new AgentRoomClient({ baseUrl: "https://sybil.example/api/", token: () => token, fetch: async (url, init) => {
+    calls.push({ url, init }); return new Response(JSON.stringify(init.method === "POST"
+      ? { id: 1, status: "queued", messageId: "inbound:sybil:request-1" } : { cursor: "0", updatedAt: null }), { status: 200 });
+  } });
+  await client.messages("0");
+  token = "second-session-token";
+  await client.markRead("42");
+  await client.submit("hi", "request-1");
+  assert.equal(calls[0].url, "https://sybil.example/api/v1/agent/messages?after=0&limit=200");
+  assert.equal(calls[0].init.headers.get("Authorization"), "Bearer first-session-token");
+  assert.equal(calls[1].init.headers.get("Authorization"), "Bearer second-session-token");
+  assert.equal(calls[1].init.method, "PUT");
+  assert.deepEqual(JSON.parse(calls[1].init.body), { cursor: "42" });
+  assert.deepEqual(JSON.parse(calls[2].init.body), { content: "hi", requestId: "request-1" });
+  const url = client.websocketUrl("https://sybil.example/");
+  assert.equal(url, "wss://sybil.example/api/v1/agent/ws");
+  assert.equal(new URL(url).search, "");
+  assert.deepEqual(JSON.parse(agentAuthFrame(null, "42")), { type: "auth", token: null, after: "42" });
+  assert.deepEqual(JSON.parse(agentAuthFrame(token, "42")), { type: "auth", token, after: "42" });
+});
+
+test("malformed successful receipts keep optimistic messages retryable under the original request ID", async () => {
+  const requests = [];
+  let response = {};
+  const client = new AgentRoomClient({ fetch: async (_url, init) => {
+    requests.push(JSON.parse(init.body));
+    return new Response(JSON.stringify(response), { status: 202 });
+  } });
+  const box = createAgentOptimisticOutbox({ submit: (content, requestId) => client.submit(content, requestId),
+    storage: { read: () => [], write: () => {} }, changed: () => {}, newId: () => "retry-id",
+  });
+  await assert.rejects(box.send("Keep this request"), /invalid message receipt/);
+  const row = box.display([])[0];
+  assert.equal(row.metadata.status, "failed");
+  assert.equal(box.snapshot()[0].seenOnServer, false);
+  const valid = { id: 1, status: "queued", messageId: row.id };
+  for (const malformed of [null, { error: "Invalid Sybil Agent response" }, { ...valid, id: 0 }, { ...valid, id: 1.5 },
+    { ...valid, id: Number.MAX_SAFE_INTEGER + 1 }, { ...valid, status: "unknown" }, { ...valid, status: ["queued"] },
+    { ...valid, messageId: "inbound:sybil:other" }]) {
+    response = malformed;
+    await assert.rejects(box.retry(row.id), /invalid message receipt/);
+    assert.equal(box.display([])[0].metadata.status, "failed");
+  }
+  response = valid;
+  await box.retry(row.id);
+  assert.equal(box.display([])[0].metadata.status, "queued");
+  assert.ok(requests.every(request => request.requestId === "retry-id"), "uncertain responses cannot mint a duplicate admission");
+  for (const status of ["placed", "done", "unanswered"]) {
+    response = { ...valid, status };
+    assert.equal((await client.submit("Keep this request", "retry-id")).status, status);
+  }
+});
+
+test("protocol ignores unknown frames and rejects malformed message cursors", () => {
+  assert.equal(parseAgentRoomEvent('{"type":"future-frame"}'), null);
+  assert.deepEqual(parseAgentRoomEvent('{"type":"pong"}'), { type: "pong" });
+  assert.throws(() => parseAgentRoomEvent(JSON.stringify({ type: "messages", ...page("oops") })), /cursor/);
+  const state = applyAgentEvent(createAgentRoomState(), { type: "activity", conversationId: "1", activity: { ...activity(), state: "thinking", isThinking: true } });
+  assert.equal(state.activity.isThinking, true);
+  assert.deepEqual(state.messages, [], "a thinking indicator never becomes a chat message");
+});

web/tests/chat-stream-presentation.test.mjs

@@ -0,0 +1,148 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import {
+  appendStreamingDelta,
+  asStreamingAssistantMetadata,
+  clearStreamingAssistantTrace,
+  createStreamingAssistantMetadata,
+  createStreamingAttemptState,
+  finalizeStreamingAssistant,
+  promoteStreamingAssistantTrace,
+  setStreamingAssistantSegment,
+  transitionStreamingToolCall,
+} from "../src/lib/chat-stream-presentation.ts";
+
+function assistantMessage(content = "") {
+  return {
+    id: "temp-assistant-test",
+    createdAt: "2026-08-30T00:00:00.000Z",
+    role: "assistant",
+    content,
+    name: null,
+    metadata: createStreamingAssistantMetadata(),
+  };
+}
+
+function metadata(messages) {
+  return asStreamingAssistantMetadata(messages.at(-1)?.metadata);
+}
+
+test("a tool boundary promotes the current segment and the next delta starts the answer in its place", () => {
+  let messages = [assistantMessage()];
+  let attempt = appendStreamingDelta(createStreamingAttemptState(), "I'll search for that.");
+  messages = setStreamingAssistantSegment(messages, "temp-assistant-", attempt.segmentText);
+
+  const toolTransition = transitionStreamingToolCall(attempt, {
+    toolCallId: "search-1",
+    status: "initiated",
+  });
+  attempt = toolTransition.attempt;
+  messages = promoteStreamingAssistantTrace(messages, "temp-assistant-", toolTransition.traceText ?? "");
+
+  assert.equal(messages.at(-1).content, "");
+  assert.deepEqual(metadata(messages), {
+    kind: "streaming_assistant",
+    ephemeralTrace: "I'll search for that.",
+    traceRevision: 1,
+  });
+
+  attempt = appendStreamingDelta(attempt, "Here is the answer");
+  messages = setStreamingAssistantSegment(messages, "temp-assistant-", attempt.segmentText);
+
+  assert.equal(messages.at(-1).content, "Here is the answer");
+  assert.equal(metadata(messages)?.ephemeralTrace, null);
+
+  messages = finalizeStreamingAssistant(messages, "temp-assistant-", "Here is the answer.");
+  assert.equal(messages.at(-1).content, "Here is the answer.");
+  assert.equal(metadata(messages)?.ephemeralTrace, null);
+});
+
+test("a later tool round replaces the previous trace and parallel initiated calls do not erase it", () => {
+  let messages = [assistantMessage("First trace")];
+  let attempt = appendStreamingDelta(createStreamingAttemptState(), "First trace");
+  let transition = transitionStreamingToolCall(attempt, { toolCallId: "tool-1", status: "initiated" });
+  attempt = transition.attempt;
+  messages = promoteStreamingAssistantTrace(messages, "temp-assistant-", transition.traceText ?? "");
+
+  attempt = appendStreamingDelta(attempt, "I need one more lookup.");
+  messages = setStreamingAssistantSegment(messages, "temp-assistant-", attempt.segmentText);
+  transition = transitionStreamingToolCall(attempt, { toolCallId: "tool-2", status: "initiated" });
+  attempt = transition.attempt;
+  messages = promoteStreamingAssistantTrace(messages, "temp-assistant-", transition.traceText ?? "");
+
+  assert.equal(metadata(messages)?.ephemeralTrace, "I need one more lookup.");
+  assert.equal(metadata(messages)?.traceRevision, 2);
+
+  transition = transitionStreamingToolCall(attempt, { toolCallId: "tool-3", status: "initiated" });
+  assert.equal(transition.startedNewToolCall, true);
+  assert.equal(transition.traceText, null);
+  messages = promoteStreamingAssistantTrace(messages, "temp-assistant-", transition.traceText ?? "");
+
+  assert.equal(metadata(messages)?.ephemeralTrace, "I need one more lookup.");
+  assert.equal(metadata(messages)?.traceRevision, 2);
+});
+
+test("terminal and duplicate tool events never promote or reset the in-progress segment", () => {
+  const initial = appendStreamingDelta(createStreamingAttemptState(), "candidate");
+  const completed = transitionStreamingToolCall(initial, { toolCallId: "tool-1", status: "completed" });
+
+  assert.equal(completed.startedNewToolCall, false);
+  assert.strictEqual(completed.attempt, initial);
+
+  const initiated = transitionStreamingToolCall(initial, { toolCallId: "tool-1", status: "initiated" });
+  const withNextSegment = appendStreamingDelta(initiated.attempt, "next segment");
+  const duplicate = transitionStreamingToolCall(withNextSegment, { toolCallId: "tool-1", status: "initiated" });
+
+  assert.equal(duplicate.startedNewToolCall, false);
+  assert.strictEqual(duplicate.attempt, withNextSegment);
+  assert.equal(duplicate.attempt.segmentText, "next segment");
+});
+
+test("a fresh per-attempt accumulator can replay and reclassify an already-seen tool round", () => {
+  let messages = [assistantMessage("stale partial answer")];
+  let replayAttempt = appendStreamingDelta(createStreamingAttemptState(), "Replayed trace");
+  messages = setStreamingAssistantSegment(messages, "temp-assistant-", replayAttempt.segmentText);
+
+  const replayedTool = transitionStreamingToolCall(replayAttempt, { toolCallId: "same-tool-id", status: "initiated" });
+  messages = promoteStreamingAssistantTrace(messages, "temp-assistant-", replayedTool.traceText ?? "");
+
+  assert.equal(replayedTool.startedNewToolCall, true);
+  assert.equal(messages.at(-1).content, "");
+  assert.equal(metadata(messages)?.ephemeralTrace, "Replayed trace");
+});
+
+test("message helpers preserve identity when no matching temporary assistant exists", () => {
+  const messages = [
+    {
+      id: "persisted-assistant",
+      createdAt: "2026-08-30T00:00:00.000Z",
+      role: "assistant",
+      content: "Persisted",
+      name: null,
+      metadata: null,
+    },
+  ];
+
+  assert.strictEqual(setStreamingAssistantSegment(messages, "temp-assistant-", "new"), messages);
+  assert.strictEqual(promoteStreamingAssistantTrace(messages, "temp-assistant-", "trace"), messages);
+  assert.strictEqual(finalizeStreamingAssistant(messages, "temp-assistant-", "done"), messages);
+  assert.strictEqual(clearStreamingAssistantTrace(messages, "temp-assistant-"), messages);
+});
+
+test("a terminal error clears the ephemeral trace without discarding a partial answer", () => {
+  let messages = [
+    {
+      ...assistantMessage("Partial answer"),
+      metadata: {
+        ...createStreamingAssistantMetadata(),
+        ephemeralTrace: "Still checking one source.",
+        traceRevision: 1,
+      },
+    },
+  ];
+
+  messages = clearStreamingAssistantTrace(messages, "temp-assistant-");
+
+  assert.equal(messages.at(-1).content, "Partial answer");
+  assert.equal(metadata(messages)?.ephemeralTrace, null);
+});

web/tests/quick-question.test.mjs

@@ -0,0 +1,39 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { buildQuickQuestionRequest, pickQuickQuestionModel } from "../src/lib/quick-question.ts";
+
+test("quick questions default to chat-latest before and after the model catalog loads", () => {
+  for (const options of [[], ["gpt-4.1-mini"], ["gpt-4.1-mini", "chat-latest"]]) {
+    const model = pickQuickQuestionModel("openai", options);
+    assert.equal(model, "chat-latest");
+    assert.deepEqual(buildQuickQuestionRequest({ provider: "openai", model, content: "What is UTC?" }), {
+      provider: "openai",
+      model: "chat-latest",
+      question: "What is UTC?",
+    });
+  }
+});
+
+test("quick questions preserve saved models and use provider-specific choices when switching providers", () => {
+  assert.equal(pickQuickQuestionModel("openai", ["chat-latest"], " gpt-4.1-mini "), "gpt-4.1-mini");
+  assert.equal(pickQuickQuestionModel("openai", ["gpt-4.1-mini"], " "), "chat-latest");
+  assert.equal(pickQuickQuestionModel("anthropic", ["claude-3-5-sonnet-latest"]), "claude-3-5-sonnet-latest");
+  assert.equal(pickQuickQuestionModel("anthropic", ["claude-3-5-sonnet-latest"], "claude-3-haiku"), "claude-3-haiku");
+  assert.equal(pickQuickQuestionModel("anthropic", []), "");
+  assert.equal(pickQuickQuestionModel("openai", ["gpt-4.1-mini", "chat-latest"]), "chat-latest");
+});
+
+test("quick question requests use the dedicated server endpoint shape", () => {
+  const request = buildQuickQuestionRequest({
+    provider: "openai",
+    model: "gpt-4.1-mini",
+    content: "How do I reset my password?",
+  });
+
+  assert.deepEqual(request, {
+    provider: "openai",
+    model: "gpt-4.1-mini",
+    question: "How do I reset my password?",
+  });
+  assert.equal("additionalSystemPrompt" in request, false);
+});

web/tests/sidebar-selection.test.mjs

@@ -8,6 +8,14 @@ const workspaceItems = [
   { type: "search", id: "selected-search" },
 ];
 
+test("persistent agent selection survives history refreshes and empty collections", () => {
+  const agent = { kind: "agent", id: "root" };
+  assert.deepEqual(resolveSidebarSelectionAfterRefresh(agent, workspaceItems, { selectFallback: true }), agent);
+  assert.deepEqual(resolveSidebarSelectionAfterRefresh(agent, []), agent);
+  assert.deepEqual(resolveSidebarSelectionAfterRefresh(null, workspaceItems, { initialSelection: agent, selectFallback: true }), agent);
+  assert.deepEqual(resolveSidebarSelectionAfterRefresh({ kind: "chat", id: "selected-chat" }, workspaceItems, { initialSelection: agent }), { kind: "chat", id: "selected-chat" });
+});
+
 test("a collection refresh preserves the current thread selection", () => {
   assert.deepEqual(
     resolveSidebarSelectionAfterRefresh({ kind: "chat", id: "selected-chat" }, workspaceItems),

web/tests/workspace-search.test.mjs

@@ -0,0 +1,80 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { scheduleWorkspaceSearch } from "../src/lib/workspace-search.ts";
+import { retainSelectedHistoryItem, resolveSidebarSelectionAfterRefresh } from "../src/lib/sidebar-selection.ts";
+
+const flush = async () => { await Promise.resolve(); await Promise.resolve(); };
+
+test("typing is debounced and superseded requests never start", async (t) => {
+  t.mock.timers.enable({ apis: ["setTimeout"] });
+  const queries = [];
+  const results = [];
+  const request = async (query) => { queries.push(query); return { items: [], hasMore: false }; };
+  const reportError = (error) => assert.fail(String(error));
+  const cancelFirst = scheduleWorkspaceSearch("ne", request, (result) => results.push(result), reportError);
+  t.mock.timers.tick(299);
+  assert.deepEqual(queries, []);
+  cancelFirst();
+  const cancelLast = scheduleWorkspaceSearch("nebula", request, (result) => results.push(result), reportError);
+  t.mock.timers.tick(300);
+  await flush();
+  assert.deepEqual(queries, ["nebula"]);
+  assert.deepEqual(results, [{ items: [], hasMore: false }]);
+  cancelLast();
+});
+
+test("cancellation aborts the transport and discards stale results and errors", async (t) => {
+  t.mock.timers.enable({ apis: ["setTimeout"] });
+  let resolveFirst;
+  let rejectSecond;
+  let signal;
+  const results = [];
+  const errors = [];
+  const first = scheduleWorkspaceSearch("first", (_query, value) => {
+    signal = value;
+    return new Promise((resolve) => { resolveFirst = resolve; });
+  }, (result) => results.push(result), (error) => errors.push(error));
+  t.mock.timers.tick(300);
+  first();
+  assert.equal(signal.aborted, true);
+  const second = scheduleWorkspaceSearch("second", () => new Promise((_resolve, reject) => { rejectSecond = reject; }),
+    (result) => results.push(result), (error) => errors.push(error));
+  t.mock.timers.tick(300);
+  second();
+  resolveFirst({ items: [{ id: "stale" }], hasMore: true });
+  rejectSecond(new Error("stale failure"));
+  await flush();
+  assert.deepEqual(results, []);
+  assert.deepEqual(errors, []);
+});
+
+test("search results retain excerpts and truncation info; failures reach the UI", async (t) => {
+  t.mock.timers.enable({ apis: ["setTimeout"] });
+  const expected = { items: [{ type: "chat", id: "old", searchSnippet: "A nebula in the sky" }], hasMore: true };
+  let received;
+  let error;
+  const cancel = scheduleWorkspaceSearch("nebula", async () => expected, (result) => { received = result; }, assert.fail);
+  t.mock.timers.tick(300);
+  await flush();
+  assert.deepEqual(received, expected);
+  cancel();
+  const cancelFailure = scheduleWorkspaceSearch("nebula", async () => { throw new Error("Offline"); }, assert.fail, (result) => { error = result; });
+  t.mock.timers.tick(300);
+  await flush();
+  assert.equal(error.message, "Offline");
+  cancelFailure();
+});
+
+test("opening older chat/search results survives collection refresh without restoring deleted or unrelated items", () => {
+  const recent = [{ type: "chat", id: "recent" }];
+  for (const kind of ["chat", "search"]) {
+    const selection = { kind, id: "old" };
+    const retained = { type: kind, id: "old" };
+    const refreshed = retainSelectedHistoryItem(recent, selection, retained);
+    assert.deepEqual(resolveSidebarSelectionAfterRefresh(selection, refreshed), selection);
+    assert.equal(retainSelectedHistoryItem(refreshed, selection, retained), refreshed);
+    assert.equal(retainSelectedHistoryItem(recent, { kind, id: "different" }, retained), recent);
+    assert.equal(retainSelectedHistoryItem(recent, selection, null), recent);
+    assert.equal(retainSelectedHistoryItem(recent, null, retained), recent);
+  }
+});

web/tsconfig.tsbuildinfo

@@ -1 +1 @@
-{"root":["./src/App.tsx","./src/main.tsx","./src/pwa.ts","./src/root-router.tsx","./src/vite-env.d.ts","./src/components/sybil-character.tsx","./src/components/auth/auth-screen.tsx","./src/components/chat/chat-attachment-list.tsx","./src/components/chat/chat-composer.tsx","./src/components/chat/chat-messages-panel.tsx","./src/components/markdown/markdown-content.tsx","./src/components/search/search-results-panel.tsx","./src/components/ui/button.tsx","./src/components/ui/input.tsx","./src/components/ui/scroll-area.tsx","./src/components/ui/separator.tsx","./src/components/ui/textarea.tsx","./src/hooks/use-session-auth.ts","./src/lib/api.ts","./src/lib/chat-forking.ts","./src/lib/chat-model-selection.ts","./src/lib/sidebar-selection.ts","./src/lib/utils.ts","./src/pages/search-route-page.tsx"],"version":"5.9.3"}
+{"root":["./src/App.tsx","./src/main.tsx","./src/pwa.ts","./src/root-router.tsx","./src/vite-env.d.ts","./src/components/sybil-character.tsx","./src/components/auth/auth-screen.tsx","./src/components/chat/chat-attachment-list.tsx","./src/components/chat/chat-composer.tsx","./src/components/chat/chat-messages-panel.tsx","./src/components/markdown/markdown-content.tsx","./src/components/search/search-results-panel.tsx","./src/components/ui/button.tsx","./src/components/ui/input.tsx","./src/components/ui/scroll-area.tsx","./src/components/ui/separator.tsx","./src/components/ui/textarea.tsx","./src/hooks/use-session-auth.ts","./src/lib/api.ts","./src/lib/chat-forking.ts","./src/lib/chat-model-selection.ts","./src/lib/quick-question.ts","./src/lib/sidebar-selection.ts","./src/lib/utils.ts","./src/pages/search-route-page.tsx"],"version":"5.9.3"}