From 8ec802ca3826be8d60e3d0bb5ab57eb69dd33b4d Mon Sep 17 00:00:00 2001
From: daniel wang <danielwang@supermicro.com>
Date: Tue, 8 Sep 2026 01:14:53 +0000
Subject: [PATCH] fix job web

---
 Makefile                                      |  26 ++-
 apps/web/src/App.tsx                          | 103 +++++++---
 apps/web/src/computer.css                     |  17 +-
 apps/web/src/handoff.ts                       |  57 ++++++
 apps/web/src/locales/en.ts                    |   1 +
 apps/web/src/locales/zh-TW.ts                 |   2 +-
 apps/web/vnc.html                             |  32 +++-
 crates/api/src/computer.rs                    |  13 +-
 crates/control/src/a11y.rs                    |   6 +-
 crates/control/src/cdp.rs                     |   6 +-
 crates/control/src/cua/browser.rs             |  21 +-
 crates/control/src/cua/client.rs              | 174 +++++++++++++++--
 crates/control/src/cua/mod.rs                 | 115 ++++++++---
 crates/control/src/cua/native.rs              |  44 ++++-
 crates/control/src/cua/translate.rs           | 102 +++++++---
 crates/control/src/lease.rs                   |  10 -
 crates/control/src/legacy.rs                  |  14 +-
 crates/control/src/observe.rs                 |  41 ++++
 crates/control/src/x11.rs                     |  21 --
 crates/sandbox/src/docker.rs                  |  43 ++++-
 docs/cua-compatibility.md                     |  67 ++++++-
 docs/cua-schemas/0.23.2/schema-drag.txt       | 147 ++++++++++++++
 .../cua-schemas/0.23.2/schema-end_session.txt |  17 ++
 .../0.23.2/schema-start_session.txt           |  50 +++++
 docs/development.md                           |  24 +++
 image/computer/Dockerfile                     |  60 +++---
 image/computer/fluxbox/menu                   |   4 -
 image/computer/fluxbox/style                  |  92 ---------
 image/supervisor/Dockerfile                   |   4 +-
 scripts/build-computer-image.sh               |   6 -
 scripts/build-image.sh                        | 136 +++++++++++++
 scripts/cua-smoke-test.sh                     |  12 ++
 tests/frontend.test.mjs                       | 179 +++++++++++++++++-
 33 files changed, 1323 insertions(+), 323 deletions(-)
 create mode 100644 apps/web/src/handoff.ts
 create mode 100644 docs/cua-schemas/0.23.2/schema-drag.txt
 create mode 100644 docs/cua-schemas/0.23.2/schema-end_session.txt
 create mode 100644 docs/cua-schemas/0.23.2/schema-start_session.txt
 delete mode 100644 image/computer/fluxbox/menu
 delete mode 100644 image/computer/fluxbox/style
 delete mode 100755 scripts/build-computer-image.sh
 create mode 100755 scripts/build-image.sh

diff --git a/Makefile b/Makefile
index dd19507..43ec085 100644
--- a/Makefile
+++ b/Makefile
@@ -8,10 +8,14 @@ COMPOSE        ?= docker compose
 COMPUTER_IMAGE ?= lazyboy/computer:local
 WEB_DIR        ?= apps/web
 DATA_DIR       ?= ./data
+BUILDX_BUILDER  ?= lazyboy
+# PUSH=1 publishes the manifest list instead of writing an OCI archive.
+MULTI_FLAGS ?=
+$(if $(filter 1,$(PUSH)),$(eval MULTI_FLAGS := --push))
 
 .PHONY: help env env-force \
         up logs ps health down purge \
-        computer postgres postgres-down pg-collation \
+        computer computer-multi images-multi postgres postgres-down pg-collation \
         cua-smoke \
         build build-api build-supervisor build-controld \
         fmt fmt-check clippy lint audit test clean \
@@ -32,6 +36,8 @@ help: ## Show this help
 	@echo ""
 	@echo "  Individual pieces:"
 	@echo "    make computer       Build the heavy Debian desktop image (lazyboy/computer:local)"
+	@echo "    make computer-multi Cross-build the desktop image for amd64 + arm64 (PUSH=1 to publish)"
+	@echo "    make images-multi   Cross-build desktop + api + supervisor for amd64 + arm64"
 	@echo "    make cua-smoke      Run Cua Driver smoke test in a disposable desktop container"
 	@echo "    make postgres       Start only postgres (127.0.0.1:5434) and wait for ready"
 	@echo "    make postgres-down  Stop postgres"
@@ -93,7 +99,23 @@ purge: ## Stop containers and delete the postgres data volume
 # --- Individual pieces -----------------------------------------------------
 
 computer: ## Build the Debian desktop image used to spawn bot computers
-	docker build -f image/computer/Dockerfile -t $(COMPUTER_IMAGE) .
+	./scripts/build-image.sh --tag $(COMPUTER_IMAGE)
+
+# Cross-builds every supported CPU architecture into one manifest list. Needs a
+# docker-container builder + QEMU binfmt; both are bootstrapped by the script.
+# PUSH=1 publishes to a registry, otherwise an OCI archive is written.
+computer-multi: ## Cross-build the desktop image for all CPU architectures
+	./scripts/build-image.sh --file image/computer/Dockerfile --multi $(MULTI_FLAGS)
+
+# The api and supervisor images carry per-architecture binaries as well (ONNX
+# Runtime, node, the distroless libc), so they get the same treatment as the
+# desktop image instead of only ever existing for the build host.
+images-multi: ## Cross-build every shipped image for all CPU architectures
+	@for dockerfile in image/computer/Dockerfile image/supervisor/Dockerfile \
+		image/api/Dockerfile; do \
+	  echo "== $$dockerfile"; \
+	  ./scripts/build-image.sh --file "$$dockerfile" --multi $(MULTI_FLAGS) || exit 1; \
+	done
 
 cua-smoke: computer ## Run the Cua Driver smoke test inside a disposable desktop container
 	./scripts/cua-smoke-test.sh --docker --image $(COMPUTER_IMAGE)
diff --git a/apps/web/src/App.tsx b/apps/web/src/App.tsx
index 3447b9d..e53754b 100644
--- a/apps/web/src/App.tsx
+++ b/apps/web/src/App.tsx
@@ -16,6 +16,7 @@ import visibility2 from "react-useanimations/lib/visibility2";
 import searchToX from "react-useanimations/lib/searchToX";
 import { api, ApiError } from "./api";
 import { createCoalescer, subscribeToSession } from "./live";
+import { HANDOFF_MS, VEIL_FADE_MS, handoffRemaining, keepScreenUrl, nextVeil, viewOnlyFor, viewerPath, type Veil } from "./handoff";
 import { Avatar, AvatarLookProvider, AvatarStack, BLOBATAR_BACKGROUNDS, BLOBATAR_EXPRESSIONS, BLOBATAR_SHAPES, DEFAULT_LOOK, persistBlobatarShape, readAvatarLooks, resolveBlobatarShape, writeAvatarLook, type AvatarBackground, type AvatarExpression, type AvatarLook } from "./avatar";
 import { dateLocale, getLocale, listJoin, setLocale, t, useLocale, type MessageKey } from "./i18n";
 import type { AvatarShape, Bot, ComputerMode, ComputerStatus, FileSkill, McpCatalogEntry, McpServer, McpTransport, MemoryItem, Message, MessageFile, ModelProviderId, Playbook, PlaybookInput, PlaybookStep, Room, RoomMember, Session, TaughtSkill, VoiceSettings, WorkspaceSettings } from "./types";
@@ -86,6 +87,7 @@ function localizeError(message:string){
   if(message==="示範進行中：先按「完成示範」或「取消」，再送訊息。")return t("teachInProgress");
   if(message==="AI 回應逾時（150 秒）")return t("aiTimeout");
   if(message==="run is not retryable")return t("errorRetryFailed");
+  if(message==="Stop the task first")return t("takeoverBusy");
   return message;
 }
 function hudLabel(computer:ComputerStatus,connecting:boolean,handingOff:boolean){
@@ -160,6 +162,11 @@ export function App(){
   const desktopFrameRef=useRef<HTMLIFrameElement>(null);
   const holderRef=useRef(computer.controlHolder); const paneBotRef=useRef<string|null>(null); const skipHandoffRef=useRef(true);
   const [desktopReady,setDesktopReady]=useState(false); const [handingOff,setHandingOff]=useState(false);
+  // The holder the server has not agreed to yet, when the current handoff
+  // started, and a guard so a fast double click cannot fight itself over the
+  // mouse. All three exist so the click, not the round trip, owns the answer.
+  const expectedHolderRef=useRef<ComputerStatus["controlHolder"]|null>(null); const handoffAtRef=useRef(0); const controlBusyRef=useRef(false); const [controlBusy,setControlBusy]=useState(false);
+  const [veil,setVeil]=useState<Veil>({label:null,leaving:false});
   const [pendingFiles,setPendingFiles]=useState<PendingFile[]>([]);
   const messageEndRef=useRef<HTMLDivElement|null>(null);
   const sentHistoryRef=useRef<string[]>([]); const historyIndexRef=useRef<number|null>(null); const historyDraftRef=useRef("");
@@ -213,7 +220,18 @@ export function App(){
     const status=await api<ComputerStatus>(`/api/computer/${computerBot}/status`);if(refreshSeq!==refreshSeqRef.current)return;
     const [nextMessages,screen,nextSkills]=await Promise.all([messagesJob,status.state==="running"?api<{url:string|null}>(`/api/computer/${computerBot}/screen`).catch(()=>({url:null})):Promise.resolve({url:null}),skillsJob]);
     if(refreshSeq!==refreshSeqRef.current)return;
-    setBusyMembers(nextBusy);setComputer(status);setMessages(nextMessages);setSkills(nextSkills);setScreenUrl(status.botId===computerBot?screen.url:null)
+    setBusyMembers(nextBusy);setComputer(status);setMessages(nextMessages);setSkills(nextSkills);
+    // A missing url is not a dead desktop. While it boots or wakes the mounted
+    // frame keeps its VNC session instead of blacking out and reconnecting.
+    setScreenUrl(current=>status.botId===computerBot?keepScreenUrl(current,screen.url,status.state):null);
+    // Agreement with the server is what ends a handoff. The beat has a floor so
+    // the swap is still visible on a LAN, and the ceiling timer still covers a
+    // reply that never lands.
+    if(expectedHolderRef.current&&status.controlHolder===expectedHolderRef.current){
+      expectedHolderRef.current=null;
+      const wait=handoffRemaining(handoffAtRef.current,Date.now());
+      if(wait>0)window.setTimeout(()=>setHandingOff(false),wait);else setHandingOff(false);
+    }
   },[activeId,activeRoomId,activeSessionId]);
   useEffect(()=>{loadBots().catch(e=>{if(e instanceof ApiError&&e.status===401)setAuthRequired(true);else setError(localizeError(e.message))})},[loadBots]);
   useEffect(()=>{if(!voiceSettings?.enabled)setCallOpen(false)},[voiceSettings?.enabled]);
@@ -245,9 +263,9 @@ export function App(){
   useEffect(()=>{const listener=(event:MessageEvent)=>{if(event.origin!==location.origin||event.source!==desktopFrameRef.current?.contentWindow||!event.data)return;if(event.data.type==="lazyboy-desktop-clipboard"){const text=String(event.data.text||"");setDesktopClipboard(text);navigator.clipboard?.writeText(text).then(()=>setClipboardStatus(t("clipboardSynced"))).catch(()=>setClipboardStatus(t("clipboardSyncBlocked")))}if(event.data.type==="lazyboy-copy-request"&&computer.controlHolder==="user")void copySelection();if(event.data.type==="lazyboy-paste-text"&&typeof event.data.text==="string")pasteText(event.data.text);if(event.data.type==="lazyboy-mobile-key"&&typeof event.data.key==="string"&&/^(?:(?:ctrl|alt)\+)?(?:Return|BackSpace|Tab|Escape|Left|Right|Up|Down|[acvz])$/.test(event.data.key))queueDesktopInput({kind:"key",key:event.data.key});if(event.data.type==="lazyboy-paste-request"&&computer.controlHolder==="user")setClipboardOpen(true);if(event.data.type==="lazyboy-desktop-ready")setDesktopReady(true);if(event.data.type==="lazyboy-desktop-lost")setDesktopReady(false)};window.addEventListener("message",listener);return()=>window.removeEventListener("message",listener)});
   useEffect(()=>{setDesktopReady(false)},[screenUrl,paneBotId]);
   useEffect(()=>{if(skipHandoffRef.current){skipHandoffRef.current=false;holderRef.current=computer.controlHolder;paneBotRef.current=paneBotId;return}if((holderRef.current!==computer.controlHolder||paneBotRef.current!==paneBotId)&&computer.state==="running")setHandingOff(true);holderRef.current=computer.controlHolder;paneBotRef.current=paneBotId},[computer.controlHolder,paneBotId,computer.state]);
-  useEffect(()=>{if(!handingOff)return;const timer=setTimeout(()=>setHandingOff(false),1600);return()=>clearTimeout(timer)},[handingOff]);
-  useEffect(()=>{const frame=desktopFrameRef.current;if(!frame?.contentWindow||!screenUrl)return;frame.contentWindow.postMessage({type:"lazyboy-view-only",viewOnly:computer.controlHolder!=="user"},location.origin)},[computer.controlHolder,screenUrl,desktopReady]);
-  useEffect(()=>{const listener=(event:MessageEvent)=>{if(event.origin!==location.origin||event.source!==desktopFrameRef.current?.contentWindow||event.data?.type!=="lazyboy-request-control"||!paneBotId)return;void action(()=>api(`/api/computer/${paneBotId}/takeover`,{method:"POST",body:"{}"}))};window.addEventListener("message",listener);return()=>window.removeEventListener("message",listener)},[paneBotId]);
+  useEffect(()=>{if(!handingOff)return;handoffAtRef.current=Date.now();const timer=setTimeout(()=>{expectedHolderRef.current=null;setHandingOff(false)},HANDOFF_MS);return()=>clearTimeout(timer)},[handingOff]);
+  useEffect(()=>{const frame=desktopFrameRef.current;if(!frame?.contentWindow||!screenUrl)return;frame.contentWindow.postMessage({type:"lazyboy-view-only",viewOnly:viewOnlyFor(computer.controlHolder)},location.origin)},[computer.controlHolder,screenUrl,desktopReady]);
+  useEffect(()=>{const listener=(event:MessageEvent)=>{if(event.origin!==location.origin||event.source!==desktopFrameRef.current?.contentWindow||event.data?.type!=="lazyboy-request-control"||!paneBotId)return;void setControl("user")};window.addEventListener("message",listener);return()=>window.removeEventListener("message",listener)},[paneBotId]);
   useEffect(()=>{const close=(event:MouseEvent)=>{const target=event.target;if(target instanceof Element&&target.closest(".create-menu-wrap,.account-wrap,.session-picker,.context-menu,.plus-menu-wrap"))return;setContext(null);setRoomContext(null);setSessionMenuOpen(false);setAccountOpen(false);setCreateMenuOpen(false);setPlusOpen(false)};window.addEventListener("click",close);return()=>window.removeEventListener("click",close)},[]);
   useEffect(()=>{const onKey=(event:KeyboardEvent)=>{if(event.key!=="Escape")return;setAccountOpen(false);setAccountDialog(null);setCreateMenuOpen(false);setPlusOpen(false);setTeachOpen(false);setEditingSkillId(null);setSessionMenuOpen(false);setContext(null);setRoomContext(null)};window.addEventListener("keydown",onKey);return()=>window.removeEventListener("keydown",onKey)},[]);
   useEffect(()=>{setWorkspaceName(name=>isDefaultWorkspaceName(name)?t("localWorkspace"):name)},[locale]);
@@ -260,10 +278,10 @@ export function App(){
   async function openLoginScreen(){
     const id=paneBot?.id||active?.id;if(!id)return;
     setRightPart("computer");setRightCollapsed(false);setComputerOpen(true);
-    await action(async()=>{
-      try{await api(`/api/computer/${id}/boot`,{method:"POST",body:"{}"})}catch{/* already up */}
-      await api(`/api/computer/${id}/takeover`,{method:"POST",body:"{}"}).catch(()=>{});
-    });
+    await action(async()=>{try{await api(`/api/computer/${id}/boot`,{method:"POST",body:"{}"})}catch{/* already up */}});
+    // The mouse is the point of the login screen, so it comes over on the
+    // optimistic handoff rather than a second round trip behind a spinner.
+    await setControl("user",id);
   }
   async function reloadSchedules(){if(!paneBotId)return;setSchedules(await api<ScheduleItem[]>(`/api/bots/${paneBotId}/schedules`))}
   async function saveScheduleDraft(){
@@ -336,18 +354,52 @@ export function App(){
   async function logout(){setAccountOpen(false);await api("/api/session",{method:"DELETE",body:"{}"}).catch(()=>{});setBots([]);setRooms([]);setMcpServers([]);setActiveId(null);setActiveRoomId(null);setAuthRequired(true)}
   async function changeComputer(operation:"boot"|"restart"|"stop"){
     const botId=paneBotId;if(!botId)return;
-    const previous=computer;setScreenUrl(null);setDesktopReady(false);
+    const previous=computer;
+    setDesktopReady(false);
+    // Stopping drops the pixels; starting keeps them. The viewer url never
+    // changes, so the frame mounts straight away and noVNC dials in the moment
+    // X answers, instead of booting only after the reply has already arrived.
+    if(operation==="stop")setScreenUrl(null);else setScreenUrl(current=>current||viewerPath(botId));
     setComputer(current=>({...current,state:operation==="stop"?current.state:operation==="boot"&&current.state==="suspended"?"suspended":"booting"}));
-    try{const status=await api<ComputerStatus>(`/api/computer/${botId}/${operation}`,{method:"POST",body:"{}",signal:AbortSignal.timeout(120_000)});if(currentPaneRef.current===botId)setComputer(status)}
+    try{
+      const status=await api<ComputerStatus>(`/api/computer/${botId}/${operation}`,{method:"POST",body:"{}",signal:AbortSignal.timeout(120_000)});
+      if(currentPaneRef.current!==botId)return;
+      setComputer(status);
+      // Anything but running and the warm frame is a lie: drop it so the panel
+      // shows what is really there instead of a desktop that will never return.
+      if(status.state!=="running")setScreenUrl(null);
+    }
     catch(error){
       const status=await api<ComputerStatus>(`/api/computer/${botId}/status`,{signal:AbortSignal.timeout(5_000)}).catch(()=>previous);
-      if(currentPaneRef.current===botId)setComputer(status);
+      if(currentPaneRef.current===botId){setComputer(status);if(status.state!=="running")setScreenUrl(null)}
       throw error;
     }
   }
   const startBoot=()=>changeComputer("boot");
   const stopComputer=()=>changeComputer("stop");
   const restartComputer=()=>changeComputer("restart");
+  /** Tell the viewer, this instant, whether human input counts. The lease is
+   * advisory inside the container, so this is the gate that actually moves the
+   * mouse, and nobody should wait on a round trip to be able to click. */
+  function pushViewOnly(viewOnly:boolean){desktopFrameRef.current?.contentWindow?.postMessage({type:"lazyboy-view-only",viewOnly},location.origin)}
+  /** Take or release the screen. The button, the badge, the veil and the mouse
+   * all move on the local copy; the server only has to agree with what is
+   * already on screen, and if it refuses the read-back puts the mouse back. */
+  async function setControl(holder:ComputerStatus["controlHolder"],botId:string|null=paneBotId){
+    if(!botId||controlBusyRef.current)return;
+    controlBusyRef.current=true;setControlBusy(true);
+    expectedHolderRef.current=holder;
+    setComputer(current=>({...current,controlHolder:holder,takeoverRequested:holder==="user"?false:current.takeoverRequested}));
+    setHandingOff(true);
+    pushViewOnly(viewOnlyFor(holder));
+    try{await api(`/api/computer/${botId}/${holder==="user"?"takeover":"release"}`,{method:"POST",body:"{}"})}
+    catch(error){setError(localizeError(error instanceof Error?error.message:t("operationFailed")))}
+    finally{
+      // Read the truth back once: a refused takeover has to give the mouse up
+      // again, and the status answer is who holds it now.
+      await refresh().catch(()=>{});expectedHolderRef.current=null;controlBusyRef.current=false;setControlBusy(false);
+    }
+  }
   useEffect(() => {
     const viewport = window.visualViewport;
     const update = () => {
@@ -360,8 +412,13 @@ export function App(){
   }, []);
   const connecting=computer.state==="running"&&Boolean(screenUrl)&&!desktopReady;
   const overlayLabel=hudLabel(computer,connecting,handingOff);
+  // The veil fades in with its label and fades back out through the same
+  // mascot, so a fast handoff reads as a finished gesture rather than a frame
+  // that was dropped. Status blips never get to strobe it.
+  useEffect(()=>{setVeil(current=>nextVeil(overlayLabel,current))},[overlayLabel]);
+  useEffect(()=>{if(!veil.leaving)return;const timer=setTimeout(()=>setVeil({label:null,leaving:false}),VEIL_FADE_MS);return()=>clearTimeout(timer)},[veil.leaving]);
   const frame=screenUrl?<iframe ref={desktopFrameRef} key={screenUrl} className="desktop-frame" src={screenUrl} title={t("agentComputer")} allow="fullscreen; clipboard-read; clipboard-write"/>:<EmptyComputer state={computer.state}/>;
-  const hud=paneBot&&overlayLabel?<ComputerHud bot={paneBot} label={overlayLabel}/>:null;
+  const hud=paneBot&&veil.label?<ComputerHud bot={paneBot} label={veil.label} leaving={veil.leaving}/>:null;
   const statusMembers=workingMembers;
   const topTools=<nav className="top-tools" aria-label={t("workTools")}>
     {active&&!activeRoomId?<span className="call-entry" tabIndex={!voiceSettings?.enabled?0:undefined} title={!voiceSettings?.enabled?t("voiceDisabledHint"):undefined} aria-label={!voiceSettings?.enabled?t("voiceDisabledHint"):undefined}><button type="button" className={`top-tool-button ${callOpen?"call-active":""}`} title={!voiceSettings?.enabled?t("voiceDisabledHint"):voiceSettings.ready?t("call"):t("setUpVoiceToCall")} aria-label={t("call")} disabled={!activeSessionId||!voiceSettings?.enabled} onClick={()=>{if(!voiceSettings?.enabled)return;if(!voiceSettings.ready){setAccountDialog("voice");return}setCallOpen(true)}}><PhoneIcon/></button></span>:null}
@@ -411,8 +468,8 @@ export function App(){
 
     <main className="chat-panel">
       <header className="topbar"><button type="button" className="icon-button mobile-menu" aria-label={t("chooseBot")} aria-expanded={mobileNav} aria-controls="conversation-sidebar" onClick={()=>setMobileNav(v=>!v)}><UseAnimations animation={menu} size={18} strokeColor="var(--ink)"/></button>{activeRoom?<><AvatarStack members={activeRoom.members} size={32} online thinkingIds={busyMembers.map(member=>member.id)}/><strong>{activeRoom.name}</strong><SessionMenu sessions={sessions} activeSessionId={activeSessionId} open={sessionMenuOpen} setOpen={setSessionMenuOpen} busy={busy} onSelect={selectSession} onCreate={createSession} onDelete={id=>{setSessionMenuOpen(false);setSessionToDelete(id)}} onClear={()=>{setSessionMenuOpen(false);setClearOpen(true)}}/><span className="grow"/>{topTools}</>:active?<><Avatar lookId={active.id} name={active.name} color={active.avatarColor} shape={active.avatarShape} active online/><strong>{active.name}</strong><SessionMenu sessions={sessions} activeSessionId={activeSessionId} open={sessionMenuOpen} setOpen={setSessionMenuOpen} busy={busy} onSelect={selectSession} onCreate={createSession} onDelete={id=>{setSessionMenuOpen(false);setSessionToDelete(id)}} onClear={()=>{setSessionMenuOpen(false);setClearOpen(true)}}/><span className="grow"/>{topTools}</>:<><strong>{t("chooseBot")}</strong><span className="grow"/>{topTools}</>}</header>
-      {callOpen&&voiceSettings?.enabled&&active&&activeSessionId&&!activeRoomId?<CallOverlay bot={active} sessionId={activeSessionId} takeover={computer.takeoverRequested||computer.controlHolder==="user"} onHangUp={()=>setCallOpen(false)} onTakeOver={()=>{setRightPart("computer");setRightCollapsed(false);if(active)void action(()=>api(`/api/computer/${active.id}/takeover`,{method:"POST",body:"{}"}))}}/>:null}
-      <div className="messages">{(activeRoom||active)&&messages.length===0?<div className="welcome">{activeRoom?<AvatarStack members={activeRoom.members} size={56} online/>:<Avatar lookId={active!.id} name={active!.name} color={active!.avatarColor} shape={active!.avatarShape} active online size={64}/>}<h1>{activeRoom?t("startRoomDiscussion",{name:activeRoom.name}):t("startBotWork",{name:active!.name})}</h1><p>{activeRoom?t("roomWillReply",{names:listJoin(activeRoom.members.map(member=>member.name))}):active!.description||t("botWelcome")}</p></div>:messages.map(message=>{const spoken=message.role!=="user"&&Boolean(activeRoom);const speakerName=message.speakerName||(spoken?paneBot?.name:undefined);const speakerShape=(message.speakerShape||paneBot?.avatarShape||"blob") as AvatarShape;const files=messageFiles(message.blocks);const chips=chipBlocks(message.blocks);const hideBody=isAutoAttachCaption(message.body,files)||isAutoScheduleCaption(message.body,chips);return <div key={message.id} className={`message ${message.role} ${spoken?"spoken":""} ${files.length?"with-files":""}`}>{spoken&&<span className="msg-avatar"><Avatar lookId={message.speakerBotId||paneBot?.id||undefined} name={speakerName||"agent"} color={message.speakerColor||undefined} shape={speakerShape} size={22}/></span>}{spoken&&<b className="speaker" style={{color:message.speakerColor||undefined}}>{speakerName}</b>}{files.length>0&&<div className="msg-attachments">{files.map(file=><FileCard key={file.name} file={file}/>)}</div>}{chips.map((chip,index)=>chip.kind==="error"?<div className="login-chip error-chip" key={`${message.id}-error-${index}`}><div className="login-label">{errorTitle(chip.code)}</div><div className="error-actions">{errorActions(chip.code).map(next=>next==="retry"?<button type="button" className="primary" key="retry" title={t("errorRetryHint")} disabled={busy||!chip.runId||Boolean(retriedRuns[chip.runId])} onClick={()=>void retryRun(chip.runId)}>{retriedRuns[chip.runId||""]?t("errorRetried"):t("errorRetry")}</button>:next==="screen"?<button type="button" key="screen" onClick={()=>void openLoginScreen()}>{t("loginOpenScreen")}</button>:<button type="button" key="settings" onClick={()=>openAccount("model")}>{t("errorOpenSettings")}</button>)}<RunProbe runId={chip.runId||null} align="end" label={t("errorDetails")}><span className="error-probe" title={t("errorDetails")}><Info/></span></RunProbe></div></div>:chip.kind==="resume"?<div className="login-chip resume-chip" key={`${message.id}-resume-${index}`}><div className="login-label">{resumeTitle(chip.reason)}</div>{(chip.limit||0)>0?<div className="login-why">{t("resumeProgress",{turns:chip.turns||0,limit:chip.limit||0})}</div>:null}{resumedChips[message.id]||(message.seq??0)<lastAssistantSeq?null:<div className="resume-actions"><button type="button" className="primary" disabled={busy} onClick={()=>void continueRun(message.id)}>{t("resumeContinue")}</button><button type="button" className="outline" disabled={busy} onClick={()=>setResumedChips(current=>({...current,[message.id]:true}))}>{t("resumeStop")}</button></div>}</div>:chip.kind==="login"?<div className="login-chip" key={`${message.id}-login-${index}`}><div className="login-label">{t("loginNeedsYou")}</div><div className="login-site">{chip.site||message.body}</div>{chip.why?<div className="login-why">{t("loginWhy",{why:chip.why})}</div>:null}<button type="button" className="primary" onClick={()=>void openLoginScreen()}>{t("loginOpenScreen")}</button></div>:chip.kind==="schedule"?<div className="sched-chip" key={`${message.id}-sched-${index}`}><div className="sched-label">{t("scheduleChip")}</div><strong>{chip.name}</strong><small>{scheduleWhen(chip.cron,chip.human)}</small></div>:<div className="sched-chip" key={`${message.id}-run-${index}`}><div className="sched-label">{t("scheduleRunChip")}</div><strong>{chip.name}</strong><small>{scheduleWhen(chip.cron,chip.human)}</small></div>)}{!hideBody&&(message.role==="assistant"?<div className="message-stack"><div className="message-body md"><ChatMarkdown>{message.body}</ChatMarkdown></div>{message.body.trim()?<CopyMessageButton text={message.body}/>:null}</div>:<span className="message-body">{message.body}</span>)}{!hideBody&&message.body.trim()&&<button type="button" className={`remember-msg ${remembered[message.id]?"saved":""}`} title={remembered[message.id]?t("remembered"):t("remember")} disabled={!!remembered[message.id]} onClick={()=>void rememberMessage(message)}><UseAnimations animation={bookmark} size={14} strokeColor="var(--muted)"/></button>}</div>})}{pausedForUser&&paneBot&&<div className="pause-banner" role="status"><span>{computer.controlHolder==="user"?t("pausedUserControl",{name:paneBot.name}):t("pausedNeedsUser",{name:paneBot.name})}{(computer.queuedRuns||0)>0&&<> {t("queuedMessages",{count:computer.queuedRuns||0})}</>}</span>{computer.controlHolder==="user"?<button type="button" className="primary" disabled={busy} onClick={()=>void action(()=>api(`/api/computer/${paneBot.id}/release`,{method:"POST",body:"{}"}))}>{t("releaseAndContinue")}</button>:<button type="button" className="primary" disabled={busy} onClick={()=>void action(()=>api(`/api/computer/${paneBot.id}/takeover`,{method:"POST",body:"{}"}))}>{t("takeOverNow")}</button>}</div>}{teaching&&active&&<div className="teach-banner recording" role="status"><span><i className="record-dot live"/>{t("teachingLive",{goal:teaching.goal})}<small>{t("teachingHint")}{teaching.eventCount>0&&<> · {t("teachingCaptured",{count:teaching.eventCount})}</>}</small></span><button type="button" className="primary" disabled={busy} onClick={()=>void stopTeaching()}>{t("finishDemo")}</button><button type="button" className="outline" disabled={busy} onClick={()=>void cancelTeaching()}>{t("cancel")}</button></div>}{drafting&&<div className="teach-banner" role="status"><UseAnimations animation={loading} size={18} wrapperStyle={{display:"inline-block",verticalAlign:"middle"}}/><span>{t("distilling",{goal:drafting.goal})}</span></div>}{skillDraft&&active&&<SkillDraftCard key={skillDraft.id} skill={skillDraft} busy={busy} onSave={(name,playbook)=>void saveSkill(skillDraft,name,playbook)} onTest={(name,playbook)=>void testSkill(skillDraft,name,playbook)} onDiscard={()=>void discardSkill(skillDraft)} onEdit={()=>setEditingSkillId(skillDraft.id)} onExport={(name,playbook)=>downloadSkill(name,skillDraft.goal,playbook)}/>}<div ref={messageEndRef} aria-hidden="true"/></div>
+      {callOpen&&voiceSettings?.enabled&&active&&activeSessionId&&!activeRoomId?<CallOverlay bot={active} sessionId={activeSessionId} takeover={computer.takeoverRequested||computer.controlHolder==="user"} onHangUp={()=>setCallOpen(false)} onTakeOver={()=>{setRightPart("computer");setRightCollapsed(false);if(active)void setControl("user",active.id)}}/>:null}
+      <div className="messages">{(activeRoom||active)&&messages.length===0?<div className="welcome">{activeRoom?<AvatarStack members={activeRoom.members} size={56} online/>:<Avatar lookId={active!.id} name={active!.name} color={active!.avatarColor} shape={active!.avatarShape} active online size={64}/>}<h1>{activeRoom?t("startRoomDiscussion",{name:activeRoom.name}):t("startBotWork",{name:active!.name})}</h1><p>{activeRoom?t("roomWillReply",{names:listJoin(activeRoom.members.map(member=>member.name))}):active!.description||t("botWelcome")}</p></div>:messages.map(message=>{const spoken=message.role!=="user"&&Boolean(activeRoom);const speakerName=message.speakerName||(spoken?paneBot?.name:undefined);const speakerShape=(message.speakerShape||paneBot?.avatarShape||"blob") as AvatarShape;const files=messageFiles(message.blocks);const chips=chipBlocks(message.blocks);const hideBody=isAutoAttachCaption(message.body,files)||isAutoScheduleCaption(message.body,chips);return <div key={message.id} className={`message ${message.role} ${spoken?"spoken":""} ${files.length?"with-files":""}`}>{spoken&&<span className="msg-avatar"><Avatar lookId={message.speakerBotId||paneBot?.id||undefined} name={speakerName||"agent"} color={message.speakerColor||undefined} shape={speakerShape} size={22}/></span>}{spoken&&<b className="speaker" style={{color:message.speakerColor||undefined}}>{speakerName}</b>}{files.length>0&&<div className="msg-attachments">{files.map(file=><FileCard key={file.name} file={file}/>)}</div>}{chips.map((chip,index)=>chip.kind==="error"?<div className="login-chip error-chip" key={`${message.id}-error-${index}`}><div className="login-label">{errorTitle(chip.code)}</div><div className="error-actions">{errorActions(chip.code).map(next=>next==="retry"?<button type="button" className="primary" key="retry" title={t("errorRetryHint")} disabled={busy||!chip.runId||Boolean(retriedRuns[chip.runId])} onClick={()=>void retryRun(chip.runId)}>{retriedRuns[chip.runId||""]?t("errorRetried"):t("errorRetry")}</button>:next==="screen"?<button type="button" key="screen" onClick={()=>void openLoginScreen()}>{t("loginOpenScreen")}</button>:<button type="button" key="settings" onClick={()=>openAccount("model")}>{t("errorOpenSettings")}</button>)}<RunProbe runId={chip.runId||null} align="end" label={t("errorDetails")}><span className="error-probe" title={t("errorDetails")}><Info/></span></RunProbe></div></div>:chip.kind==="resume"?<div className="login-chip resume-chip" key={`${message.id}-resume-${index}`}><div className="login-label">{resumeTitle(chip.reason)}</div>{(chip.limit||0)>0?<div className="login-why">{t("resumeProgress",{turns:chip.turns||0,limit:chip.limit||0})}</div>:null}{resumedChips[message.id]||(message.seq??0)<lastAssistantSeq?null:<div className="resume-actions"><button type="button" className="primary" disabled={busy} onClick={()=>void continueRun(message.id)}>{t("resumeContinue")}</button><button type="button" className="outline" disabled={busy} onClick={()=>setResumedChips(current=>({...current,[message.id]:true}))}>{t("resumeStop")}</button></div>}</div>:chip.kind==="login"?<div className="login-chip" key={`${message.id}-login-${index}`}><div className="login-label">{t("loginNeedsYou")}</div><div className="login-site">{chip.site||message.body}</div>{chip.why?<div className="login-why">{t("loginWhy",{why:chip.why})}</div>:null}<button type="button" className="primary" onClick={()=>void openLoginScreen()}>{t("loginOpenScreen")}</button></div>:chip.kind==="schedule"?<div className="sched-chip" key={`${message.id}-sched-${index}`}><div className="sched-label">{t("scheduleChip")}</div><strong>{chip.name}</strong><small>{scheduleWhen(chip.cron,chip.human)}</small></div>:<div className="sched-chip" key={`${message.id}-run-${index}`}><div className="sched-label">{t("scheduleRunChip")}</div><strong>{chip.name}</strong><small>{scheduleWhen(chip.cron,chip.human)}</small></div>)}{!hideBody&&(message.role==="assistant"?<div className="message-stack"><div className="message-body md"><ChatMarkdown>{message.body}</ChatMarkdown></div>{message.body.trim()?<CopyMessageButton text={message.body}/>:null}</div>:<span className="message-body">{message.body}</span>)}{!hideBody&&message.body.trim()&&<button type="button" className={`remember-msg ${remembered[message.id]?"saved":""}`} title={remembered[message.id]?t("remembered"):t("remember")} disabled={!!remembered[message.id]} onClick={()=>void rememberMessage(message)}><UseAnimations animation={bookmark} size={14} strokeColor="var(--muted)"/></button>}</div>})}{pausedForUser&&paneBot&&<div className="pause-banner" role="status"><span>{computer.controlHolder==="user"?t("pausedUserControl",{name:paneBot.name}):t("pausedNeedsUser",{name:paneBot.name})}{(computer.queuedRuns||0)>0&&<> {t("queuedMessages",{count:computer.queuedRuns||0})}</>}</span>{computer.controlHolder==="user"?<button type="button" className="primary" disabled={busy||controlBusy} onClick={()=>void setControl("none",paneBot.id)}>{t("releaseAndContinue")}</button>:<button type="button" className="primary" disabled={busy||controlBusy} onClick={()=>void setControl("user",paneBot.id)}>{t("takeOverNow")}</button>}</div>}{teaching&&active&&<div className="teach-banner recording" role="status"><span><i className="record-dot live"/>{t("teachingLive",{goal:teaching.goal})}<small>{t("teachingHint")}{teaching.eventCount>0&&<> · {t("teachingCaptured",{count:teaching.eventCount})}</>}</small></span><button type="button" className="primary" disabled={busy} onClick={()=>void stopTeaching()}>{t("finishDemo")}</button><button type="button" className="outline" disabled={busy} onClick={()=>void cancelTeaching()}>{t("cancel")}</button></div>}{drafting&&<div className="teach-banner" role="status"><UseAnimations animation={loading} size={18} wrapperStyle={{display:"inline-block",verticalAlign:"middle"}}/><span>{t("distilling",{goal:drafting.goal})}</span></div>}{skillDraft&&active&&<SkillDraftCard key={skillDraft.id} skill={skillDraft} busy={busy} onSave={(name,playbook)=>void saveSkill(skillDraft,name,playbook)} onTest={(name,playbook)=>void testSkill(skillDraft,name,playbook)} onDiscard={()=>void discardSkill(skillDraft)} onEdit={()=>setEditingSkillId(skillDraft.id)} onExport={(name,playbook)=>downloadSkill(name,skillDraft.goal,playbook)}/>}<div ref={messageEndRef} aria-hidden="true"/></div>
       <div className={`composer-dock ${statusMembers.length?"has-status":""}`}>
       {error&&<div className="error-banner"><span>{error}</span><button onClick={()=>setError(null)}><X/></button></div>}
       {otherSessionBusy&&<div className="queue-hint">{t("anotherConversationQueued")}</div>}
@@ -431,8 +488,8 @@ export function App(){
         <div className="side-card-body">
           <div className={`side-part computer-part ${rightPart==="computer"?"":"hidden-part"}`}>
             <div className="computer-status-row">{paneBot?<span>{t("botComputer",{name:paneBot.name})}</span>:<span>{t("computer")}</span>}{computer.state==="booting"?<UseAnimations animation={loading} size={17} wrapperStyle={{display:"inline-block",verticalAlign:"middle"}}/>:<i className={`state-dot ${computer.state}`}/>}<small>{stateLabel(computer.state)}</small></div>
-            {!computerOpen&&hud}<div className="preview">{computerOpen?<EmptyComputer state={computer.state}/>:frame}</div>
-            {paneBot&&<><div className="computer-caption"><span>{t("dedicatedScreen")}</span><button className="outline" onClick={()=>setComputerOpen(true)}>{t("enlarge")}</button></div>{teaching?<div className="control-bar"><div className="computer-actions"><button className="primary" disabled={busy} onClick={()=>void stopTeaching()}>{t("finishDemo")}</button><button className="outline" disabled={busy} onClick={()=>void cancelTeaching()}>{t("cancel")}</button></div></div>:<ControlBar active={paneBot} computer={computer} busy={busy} action={action} paste={pasteClipboard} copy={copyClipboard} sessionId={activeSessionId} onBoot={startBoot} onRestart={restartComputer} onStop={stopComputer}/>}<p className="computer-login-hint">{t("computerLoginHint")}</p><p className="clipboard-status" role="status">{clipboardStatus}</p>{scheduleDraft?<ScheduleEditor draft={scheduleDraft} timezone={scheduleDraft.timezone||Intl.DateTimeFormat().resolvedOptions().timeZone||"Asia/Taipei"} saving={scheduleSaving} error={scheduleError} onChange={next=>setScheduleDraft(current=>current?{...current,...next}:next)} onBack={()=>setScheduleDraft(null)} onSave={()=>void saveScheduleDraft()} onDelete={scheduleDraft.id?()=>void action(async()=>{await api(`/api/schedules/${scheduleDraft.id}`,{method:"DELETE"});setScheduleDraft(null);await reloadSchedules()}):undefined}/>:<ScheduleList items={schedules} runningId={runningScheduleId} onCreate={()=>setScheduleDraft({name:"",instructions:"",enabled:true,preset:defaultCronPreset()})} onOpen={item=>setScheduleDraft({id:item.id,timezone:item.timezone,threadId:item.threadId,name:item.name,instructions:item.instructions,enabled:item.enabled,preset:presetFromCron(item.cron)})} onRun={item=>void action(async()=>{setRunningScheduleId(item.id);try{await api(`/api/schedules/${item.id}/run`,{method:"POST",body:"{}"});await loadSessions()}finally{setRunningScheduleId(null)}})}/>}</>}
+            <div className="preview">{computerOpen?<EmptyComputer state={computer.state}/>:frame}{!computerOpen&&hud}</div>
+            {paneBot&&<><div className="computer-caption"><span>{t("dedicatedScreen")}</span><button className="outline" onClick={()=>setComputerOpen(true)}>{t("enlarge")}</button></div>{teaching?<div className="control-bar"><div className="computer-actions"><button className="primary" disabled={busy} onClick={()=>void stopTeaching()}>{t("finishDemo")}</button><button className="outline" disabled={busy} onClick={()=>void cancelTeaching()}>{t("cancel")}</button></div></div>:<ControlBar active={paneBot} computer={computer} busy={busy} pending={controlBusy} action={action} paste={pasteClipboard} copy={copyClipboard} sessionId={activeSessionId} onBoot={startBoot} onRestart={restartComputer} onStop={stopComputer} onTakeOver={()=>void setControl("user")} onRelease={()=>void setControl("none")}/>}<p className="computer-login-hint">{t("computerLoginHint")}</p><p className="clipboard-status" role="status">{clipboardStatus}</p>{scheduleDraft?<ScheduleEditor draft={scheduleDraft} timezone={scheduleDraft.timezone||Intl.DateTimeFormat().resolvedOptions().timeZone||"Asia/Taipei"} saving={scheduleSaving} error={scheduleError} onChange={next=>setScheduleDraft(current=>current?{...current,...next}:next)} onBack={()=>setScheduleDraft(null)} onSave={()=>void saveScheduleDraft()} onDelete={scheduleDraft.id?()=>void action(async()=>{await api(`/api/schedules/${scheduleDraft.id}`,{method:"DELETE"});setScheduleDraft(null);await reloadSchedules()}):undefined}/>:<ScheduleList items={schedules} runningId={runningScheduleId} onCreate={()=>setScheduleDraft({name:"",instructions:"",enabled:true,preset:defaultCronPreset()})} onOpen={item=>setScheduleDraft({id:item.id,timezone:item.timezone,threadId:item.threadId,name:item.name,instructions:item.instructions,enabled:item.enabled,preset:presetFromCron(item.cron)})} onRun={item=>void action(async()=>{setRunningScheduleId(item.id);try{await api(`/api/schedules/${item.id}/run`,{method:"POST",body:"{}"});await loadSessions()}finally{setRunningScheduleId(null)}})}/>}</>}
           </div>
           {rightPart==="accounts"&&paneBot&&<VaultPane bot={paneBot}/>}
           {rightPart==="memory"&&paneBot&&<MemoryPane bot={paneBot} changed={loadBots}/>}
@@ -442,7 +499,7 @@ export function App(){
       </>
     </aside>}
 
-    {computerOpen&&paneBot&&<div className="computer-overlay"><header><div><RunProbe runId={computer.busyRunId||computer.waitingRunId}><Avatar lookId={paneBot.id} name={paneBot.name} color={paneBot.avatarColor} shape={paneBot.avatarShape} active online/></RunProbe><strong>{modeLabel(computer.mode)}</strong><span className={`control-badge ${teaching?"teaching":""}`}>{teaching?t("teachingBadge"):computer.controlHolder==="user"?t("userControlling"):computer.usingComputer?t("aiReadOnly"):t("readOnly")}</span></div><div>{teaching?<div className="computer-actions"><button className="primary" disabled={busy} onClick={()=>void stopTeaching()}>{t("finishDemo")}</button><button className="outline" disabled={busy} onClick={()=>void cancelTeaching()}>{t("cancel")}</button></div>:<ControlButtons computer={computer} busy={busy} action={action} active={paneBot} sessionId={activeSessionId} onBoot={startBoot} onRestart={restartComputer} onStop={stopComputer}/>}<button className="icon-button" onClick={pasteClipboard} disabled={computer.controlHolder!=="user"} title={t("pasteClipboard")}><ClipboardPaste/></button><button className="icon-button" onClick={copyClipboard} disabled={computer.controlHolder!=="user"||!desktopClipboard} title={t("copyDesktopClipboard")}><UseAnimations animation={copy} size={18} strokeColor="var(--ink)"/></button><button className="icon-button" onClick={()=>setComputerOpen(false)}><X/></button></div></header>{hud}<div className="overlay-screen"><div className="overlay-desktop">{frame}</div></div>{error&&<div className="overlay-error">{error}</div>}</div>}
+    {computerOpen&&paneBot&&<div className="computer-overlay"><header><div><RunProbe runId={computer.busyRunId||computer.waitingRunId}><Avatar lookId={paneBot.id} name={paneBot.name} color={paneBot.avatarColor} shape={paneBot.avatarShape} active online/></RunProbe><strong>{modeLabel(computer.mode)}</strong><span className={`control-badge ${teaching?"teaching":""}`}>{teaching?t("teachingBadge"):computer.controlHolder==="user"?t("userControlling"):computer.usingComputer?t("aiReadOnly"):t("readOnly")}</span></div><div>{teaching?<div className="computer-actions"><button className="primary" disabled={busy} onClick={()=>void stopTeaching()}>{t("finishDemo")}</button><button className="outline" disabled={busy} onClick={()=>void cancelTeaching()}>{t("cancel")}</button></div>:<ControlButtons computer={computer} busy={busy} pending={controlBusy} action={action} active={paneBot} sessionId={activeSessionId} onBoot={startBoot} onRestart={restartComputer} onStop={stopComputer} onTakeOver={()=>void setControl("user")} onRelease={()=>void setControl("none")}/>}<button className="icon-button" onClick={pasteClipboard} disabled={computer.controlHolder!=="user"} title={t("pasteClipboard")}><ClipboardPaste/></button><button className="icon-button" onClick={copyClipboard} disabled={computer.controlHolder!=="user"||!desktopClipboard} title={t("copyDesktopClipboard")}><UseAnimations animation={copy} size={18} strokeColor="var(--ink)"/></button><button className="icon-button" onClick={()=>setComputerOpen(false)}><X/></button></div></header><div className="overlay-screen"><div className="overlay-desktop">{frame}{hud}</div></div>{error&&<div className="overlay-error">{error}</div>}</div>}
     {teachOpen&&active&&<TeachDialog bot={active} busy={busy} close={()=>setTeachOpen(false)} start={goal=>void startTeaching(goal)}/>}
     {editingSkill&&<SkillEditDialog key={editingSkill.id} skill={editingSkill} busy={busy} close={()=>setEditingSkillId(null)} save={(name,playbook)=>void updateSkill(editingSkill,name,playbook)} test={(name,playbook)=>void testSkill(editingSkill,name,playbook)} remove={()=>void deleteSkill(editingSkill)} exportFile={(name,playbook)=>downloadSkill(name,editingSkill.goal,playbook)}/>}
     {clipboardOpen&&<ClipboardDialog close={()=>setClipboardOpen(false)} paste={text=>{pasteText(text);setClipboardOpen(false)}}/>}
@@ -678,12 +735,12 @@ function McpCustomForm({busy,error,onBack,onSubmit}:{busy:boolean;error:string;o
   </form>
 }
 
-function ComputerHud({bot,label}:{bot:{id:string;name:string;avatarColor?:string;avatarShape?:AvatarShape};label:string}){
+function ComputerHud({bot,label,leaving}:{bot:{id:string;name:string;avatarColor?:string;avatarShape?:AvatarShape};label:string;leaving?:boolean}){
   const sep=getLocale()==="en"?": ":"：";
-  return <div className="computer-hud" role="status" aria-label={`${bot.name}${sep}${label}`}><span className="computer-signal" aria-hidden="true"><span className="computer-signal-face"><i/><i/></span></span><span className="computer-hud-label">{label}</span></div>
+  return <div className={`computer-hud ${leaving?"is-leaving":""}`} role="status" aria-label={`${bot.name}${sep}${label}`}><span className="computer-signal" aria-hidden="true"><span className="computer-signal-face"><i/><i/></span></span><span className="computer-hud-label">{label}</span></div>
 }
 function EmptyComputer({state}:{state:ComputerStatus["state"]}){if(state==="booting"||state==="suspended")return <div className="empty-computer is-waiting" aria-hidden="true"/>;return <div className="empty-computer"><Computer/><strong>{stateLabel(state)}</strong><span>{t("computerPreviewHint")}</span></div>}
-function ControlButtons({computer,busy,action,active,sessionId,onBoot,onRestart,onStop}:{computer:ComputerStatus;busy:boolean;action:(w:()=>Promise<unknown>)=>Promise<void>;active:Bot;sessionId?:string|null;onBoot?:()=>Promise<void>;onRestart?:()=>Promise<void>;onStop?:()=>Promise<void>}){
+function ControlButtons({computer,busy,pending,action,active,sessionId,onBoot,onRestart,onStop,onTakeOver,onRelease}:{computer:ComputerStatus;busy:boolean;action:(w:()=>Promise<unknown>)=>Promise<void>;active:Bot;sessionId?:string|null;onBoot?:()=>Promise<void>;onRestart?:()=>Promise<void>;onStop?:()=>Promise<void>;pending?:boolean;onTakeOver:()=>void;onRelease:()=>void}){
   const [menuOpen,setMenuOpen]=useState(false);
   const menuRef=useRef<HTMLDivElement>(null);
   useEffect(()=>{
@@ -700,8 +757,8 @@ function ControlButtons({computer,busy,action,active,sessionId,onBoot,onRestart,
   const run=(operation:"boot"|"restart"|"stop",handler?:()=>Promise<void>)=>{setMenuOpen(false);void action(handler||(()=>api(`/api/computer/${active.id}/${operation}`,{method:"POST",body:"{}"})))};
   return <div className="computer-actions">
     {!running?<button className="primary" disabled={busy||booting} onClick={()=>run("boot",onBoot)}>{booting?t("bootingProgress"):t("openComputer")}</button>
-      :computer.controlHolder==="user"?<button className="primary" disabled={busy} onClick={()=>void action(()=>api(`/api/computer/${active.id}/release`,{method:"POST",body:"{}"}))}>{t("releaseControl")}</button>
-      :<button className="primary" disabled={busy} onClick={()=>void action(()=>api(`/api/computer/${active.id}/takeover`,{method:"POST",body:"{}"}))}>{working?t("takeOverNow"):t("takeControl")}</button>}
+      :computer.controlHolder==="user"?<button className="primary" disabled={busy||pending} onClick={onRelease}>{t("releaseControl")}</button>
+      :<button className="primary" disabled={busy||pending} onClick={onTakeOver}>{working?t("takeOverNow"):t("takeControl")}</button>}
     {running&&working&&<button className="outline" disabled={busy} onClick={()=>void action(()=>api(sessionId?`/api/sessions/${sessionId}/stop`:`/api/bots/${active.id}/stop`,{method:"POST",body:"{}"}))}><Square/>{t("stopTask")}</button>}
     {computer.state!=="stopped"&&<div className="computer-power" ref={menuRef}>
       <button type="button" className="outline computer-power-trigger" disabled={busy||booting} aria-expanded={menuOpen} aria-label={t("computerPower")} title={t("computerPower")} onClick={()=>setMenuOpen(open=>!open)}><Ellipsis/></button>
@@ -713,7 +770,7 @@ function ControlButtons({computer,busy,action,active,sessionId,onBoot,onRestart,
     </div>}
   </div>;
 }
-function ControlBar(props:{active:Bot;computer:ComputerStatus;busy:boolean;action:(w:()=>Promise<unknown>)=>Promise<void>;paste:()=>void;copy:()=>void;sessionId?:string|null;onBoot?:()=>Promise<void>;onRestart?:()=>Promise<void>;onStop?:()=>Promise<void>}){const interactive=props.computer.controlHolder==="user";return <div className="control-bar"><ControlButtons {...props}/><button className="icon-button" disabled={!interactive} title={t("pasteClipboard")} onClick={props.paste}><ClipboardPaste/></button><button className="icon-button" disabled={!interactive} title={t("copyDesktopClipboard")} onClick={props.copy}><UseAnimations animation={copy} size={18} strokeColor="var(--ink)"/></button></div>}
+function ControlBar(props:{active:Bot;computer:ComputerStatus;busy:boolean;action:(w:()=>Promise<unknown>)=>Promise<void>;paste:()=>void;copy:()=>void;sessionId?:string|null;onBoot?:()=>Promise<void>;onRestart?:()=>Promise<void>;onStop?:()=>Promise<void>;pending?:boolean;onTakeOver:()=>void;onRelease:()=>void}){const interactive=props.computer.controlHolder==="user";return <div className="control-bar"><ControlButtons {...props}/><button className="icon-button" disabled={!interactive} title={t("pasteClipboard")} onClick={props.paste}><ClipboardPaste/></button><button className="icon-button" disabled={!interactive} title={t("copyDesktopClipboard")} onClick={props.copy}><UseAnimations animation={copy} size={18} strokeColor="var(--ink)"/></button></div>}
 function TeachDialog({bot,busy,close,start}:{bot:Bot;busy:boolean;close:()=>void;start:(goal:string)=>void}){
   const [goal,setGoal]=useState("");
   return <div className="modal-backdrop" onClick={close}><form className="dialog teach-dialog" onClick={e=>e.stopPropagation()} onSubmit={e=>{e.preventDefault();if(goal.trim())start(goal.trim())}}>
diff --git a/apps/web/src/computer.css b/apps/web/src/computer.css
index 3968c80..175cce1 100644
--- a/apps/web/src/computer.css
+++ b/apps/web/src/computer.css
@@ -28,9 +28,13 @@
 .computer-part .preview .desktop-frame{display:block;aspect-ratio:16 / 10}
 .overlay-screen .overlay-desktop{position:relative;width:min(100%,1440px);height:100%}
 .overlay-screen .overlay-desktop>.desktop-frame,.overlay-screen .overlay-desktop>.empty-computer{width:100%;height:100%;border:1px solid var(--border);border-radius:14px;overflow:hidden;background:var(--main)}
-.computer-hud{position:absolute;top:10px;left:10px;z-index:3;display:grid;justify-items:center;gap:6px;padding:10px 12px 8px;border-radius:16px;background:rgba(10,10,12,.72);backdrop-filter:blur(8px);pointer-events:none;box-shadow:0 10px 30px rgba(0,0,0,.35)}
-.computer-hud .avatar.blobatar.thinking{animation:computer-hud-bob 1.6s ease-in-out infinite,avatar-rainbow-glow 1.8s linear infinite}
-.computer-hud-label{font-size:12px;letter-spacing:.02em;line-height:1.3;white-space:nowrap;background:linear-gradient(90deg,#7a8088 0%,#7a8088 28%,#fff 50%,#7a8088 72%,#7a8088 100%);background-size:220% 100%;-webkit-background-clip:text;background-clip:text;color:transparent;animation:working-shimmer 1.35s linear infinite}
+/* The waiting veil sits over the remote pixels, not beside them: while the
+   desktop is booting, waking, or changing hands, the signal mascot breathes
+   under its amber orbit and the label shimmers. It never takes a click, so the
+   mouse is live the instant the veil lifts, and it fades rather than blinks. */
+.computer-hud{position:absolute;inset:0;z-index:3;display:grid;align-content:center;justify-items:center;gap:14px;padding:10px 12px 8px;border-radius:inherit;background:radial-gradient(ellipse at center,#142722e8,#101012ed);backdrop-filter:blur(8px);pointer-events:none;opacity:1;transition:opacity .22s ease}
+.computer-hud.is-leaving{opacity:0}
+.computer-hud-label{font-size:12px;letter-spacing:.02em;line-height:1.5;white-space:normal;text-align:center;max-width:90%;color:#c9ddd5}
 .empty-computer.is-waiting{min-height:100%;background:var(--inset)}
 .computer-actions{display:flex;align-items:center;justify-content:flex-end;gap:8px;min-width:0;flex-wrap:wrap}
 .computer-actions .restart-computer{font-size:12px;padding-inline:10px;color:var(--muted)}
@@ -47,12 +51,10 @@
 .control-bar{flex-wrap:wrap}
 .control-bar>.computer-actions{flex:1 1 180px}
 /* Independent monitor mascot, with mint halo and amber orbit. */
-.computer-hud{inset:0;align-content:center;border-radius:inherit;gap:14px;background:radial-gradient(ellipse at center,#142722e8,#101012ed);box-shadow:none}
 .computer-signal{position:relative;display:grid;place-items:center;width:64px;height:64px;border:1px solid #65dcb34d;border-radius:24px;animation:monitor-breathe 2.8s ease-in-out infinite}
 .computer-signal::before{content:"";position:absolute;inset:-8px;border:1px solid #65dcb326;border-top-color:#efbe72;border-radius:50%;animation:monitor-orbit 4s linear infinite}
 .computer-signal-face{display:flex;align-items:center;justify-content:center;gap:12px;width:46px;height:38px;border-radius:15px;background:#85dfbb;color:#143429;transform:rotate(-6deg)}
 .computer-signal-face i{width:5px;height:11px;border-radius:5px;background:currentColor;animation:monitor-blink 4.6s ease-in-out infinite}
-.computer-hud-label{color:#c9ddd5;background:none;animation:none;white-space:normal;text-align:center;max-width:90%;line-height:1.5}
 
 .clipboard-status{margin:0;min-height:1.4em;font-size:12px;line-height:1.4;color:var(--muted)}
 
@@ -69,10 +71,9 @@
 .computer-power-menu small{display:block;padding:8px 10px 6px;border-top:1px solid var(--line);color:var(--muted);font-size:11px;line-height:1.6}
 .computer-overlay .computer-power-menu{top:calc(100% + 8px);bottom:auto}
 
-/* HUD belongs outside the remote pixels, including during connection/handoff. */
+/* The overlay tracks the visible viewport so the veil and the controls stay
+   above the keyboard; the veil itself belongs to the desktop box below. */
 .computer-overlay{top:var(--visible-top,0px);bottom:auto;height:var(--visible-height,100dvh)}
 .computer-overlay>header{flex-shrink:0}
-.computer-hud{position:static;inset:auto;display:block;padding:7px 12px;border-radius:0;background:var(--surface);flex:none}
-.computer-hud .computer-signal{display:none}
 .overlay-error{position:static;transform:none;flex:none;margin:0 8px 8px}
 @media(max-width:700px){.overlay-screen{padding:4px}.computer-overlay>header{min-height:44px;gap:4px}.computer-overlay>header>div{gap:5px}.computer-overlay>header strong{font-size:12px}}
diff --git a/apps/web/src/handoff.ts b/apps/web/src/handoff.ts
new file mode 100644
index 0000000..06bcd13
--- /dev/null
+++ b/apps/web/src/handoff.ts
@@ -0,0 +1,57 @@
+// Everything that decides when the screen veil shows, and when the viewer may
+// be thrown away. Both jobs used to be a fixed timer plus "null the url", which
+// is exactly what made starting and handing over feel like a stall: the mascot
+// popped, the desktop went black, and noVNC booted from scratch.
+
+import type { ComputerState } from "./types";
+
+/** The viewer url is the proxy path for a bot and never carries a nonce, so the
+ * frame can mount while the desktop is still coming up and let noVNC dial into
+ * it on retry instead of waiting for a status round trip. */
+export function viewerPath(botId: string): string {
+  return `/view/${encodeURIComponent(botId)}/vnc.html`;
+}
+
+/** Only a computer that is really gone may drop the pixels. While it boots,
+ * wakes, or a poll blinks, the mounted frame keeps its VNC session; remounting
+ * costs a page load, a handshake, and a black flash for no information. */
+export function keepScreenUrl(current: string | null, next: string | null, state: ComputerState): string | null {
+  if (next) return next;
+  if (state === "stopped" || state === "error") return null;
+  return current;
+}
+
+/** A handoff should read as one deliberate beat, not as a wait. This is the
+ * ceiling, used only when the server never confirms; agreement on the new
+ * holder usually releases the veil far sooner. */
+export const HANDOFF_MS = 900;
+/** The floor. Fast as the reply may land, the swap is worth one visible beat,
+ * otherwise the mascot strobes instead of gesturing. */
+export const HANDOFF_MIN_MS = 320;
+/** Kept in step with the veil's CSS fade so it never vanishes mid-animation. */
+export const VEIL_FADE_MS = 220;
+
+/** How much longer the veil has to stay after the server agreed on a holder:
+ * zero once the beat has been served, and never more than the ceiling timer,
+ * which is what a lost reply falls back on. */
+export function handoffRemaining(startedAt: number, now: number): number {
+  const held = now - startedAt;
+  if (held >= HANDOFF_MS) return 0;
+  return Math.max(HANDOFF_MIN_MS - held, 0);
+}
+
+export interface Veil { label: string | null; leaving: boolean }
+
+/** Fade in with a label, fade out once the label clears, and stay out when
+ * there was never anything to show. Pure, so a status blip can be tested. */
+export function nextVeil(label: string | null, current: Veil): Veil {
+  if (label) return current.label === label && !current.leaving ? current : { label, leaving: false };
+  if (!current.label) return current;
+  return current.leaving ? current : { label: current.label, leaving: true };
+}
+
+/** The lease is advisory inside the container, so the browser gate is what
+ * actually moves the mouse: whoever holds control may send input right now. */
+export function viewOnlyFor(holder: "none" | "bot" | "user"): boolean {
+  return holder !== "user";
+}
diff --git a/apps/web/src/locales/en.ts b/apps/web/src/locales/en.ts
index 078827d..b1124c6 100644
--- a/apps/web/src/locales/en.ts
+++ b/apps/web/src/locales/en.ts
@@ -253,6 +253,7 @@ export const en: { [K in keyof typeof zhTW]: string } = {
   computerPower: "Computer menu",
   sharedPowerHint: "This is a shared computer. The action affects every agent using it.",
   stopAndTakeOver: "Stop and take over",
+  takeoverBusy: "Stop the task first, then take the mouse.",
   hudBooting: "Computer starting…",
   hudWaking: "Waking…",
   hudConnecting: "Connecting…",
diff --git a/apps/web/src/locales/zh-TW.ts b/apps/web/src/locales/zh-TW.ts
index e6e3eb6..aeab04d 100644
--- a/apps/web/src/locales/zh-TW.ts
+++ b/apps/web/src/locales/zh-TW.ts
@@ -64,7 +64,7 @@ export const zhTW = {
   mcpNoResults: "找不到符合的 MCP。換個關鍵字，或改用自訂接入。", mcpCustom: "自訂接入", mcpBackToList: "回到列表",
   mcpConnectNamed: "接入 {name}", mcpKeyHint: "這個 MCP 需要憑證才能連。",
   noMcp: "還沒有 MCP。點上面的「選擇 MCP」從市集接入。", toolsCount: "{count} 個工具", disabled: "已關閉", disconnected: "未連線", noTools: "沒有可用工具", reconnect: "重新連線", disable: "停用", enable: "啟用",
-  preparingDesktop: "正在準備 Agent 的獨立桌面…", computerPreviewHint: "開啟電腦後，畫面會顯示在這裡。", bootingProgress: "啟動中…", restartComputer: "重啟", shutDownComputer: "關閉電腦", computerPower: "電腦選單", sharedPowerHint: "此為共用電腦，操作會影響使用它的所有 Agent。", stopAndTakeOver: "停止並接管",
+  preparingDesktop: "正在準備 Agent 的獨立桌面…", computerPreviewHint: "開啟電腦後，畫面會顯示在這裡。", bootingProgress: "啟動中…", restartComputer: "重啟", shutDownComputer: "關閉電腦", computerPower: "電腦選單", sharedPowerHint: "此為共用電腦，操作會影響使用它的所有 Agent。", stopAndTakeOver: "停止並接管", takeoverBusy: "請先停止任務，再接手滑鼠。",
   hudBooting: "電腦啟動中…", hudWaking: "喚醒中…", hudConnecting: "連線中…", hudHandoff: "換手中…",
   pasteToRemoteComputer: "貼到遠端電腦", pasteRemoteHelp: "把外面的文字貼在這裡，再送進 VNC。這個方式在區網 HTTP 也能使用。", pasteTextPlaceholder: "在此貼上文字…", pasteIntoVnc: "貼入 VNC",
   botNamePlaceholder: "例如：研究助理", sharedComputerHint: "與其他機器人共用環境", privateComputerHint: "全新的獨立 Docker", create: "建立",
diff --git a/apps/web/vnc.html b/apps/web/vnc.html
index a5ad5ba..55ec508 100644
--- a/apps/web/vnc.html
+++ b/apps/web/vnc.html
@@ -85,6 +85,20 @@
 
       let rfb = null;
       let reconnectTimer = null;
+      // A desktop that has never answered on this page is still coming up, and
+      // x11vnc opens its port a heartbeat after the last attempt failed. Holding
+      // every retry at the calm cadence leaves up to half a second of veil on top
+      // of a desktop that is already there, so the first window of attempts comes
+      // back eager. Once a session has been seen the calm cadence returns: a
+      // genuinely dropped connection should not be hammered.
+      const CALM_RETRY_MS = 1500;
+      const EAGER_RETRY_MS = 350;
+      const EAGER_ATTEMPTS = 40;
+      let everConnected = false;
+      let eagerAttempts = 0;
+      const retryDelay = () => (!everConnected && eagerAttempts < EAGER_ATTEMPTS)
+        ? (eagerAttempts += 1, EAGER_RETRY_MS)
+        : CALM_RETRY_MS;
       const parentOrigin = window.location.origin;
       const keyboard = document.getElementById("mobile-keyboard");
       const sentinel = "\u200b";
@@ -347,12 +361,21 @@
         }
       }
 
+      // The host flips this gate whenever control changes hands, including
+      // while the frame is still dialling or between retries. Keep the last
+      // thing it asked for and hand it to every new RFB, so a handoff never
+      // depends on which side of a reconnect the message happened to land.
+      let wantedViewOnly = null;
       function applyViewOnly(value) {
+        wantedViewOnly = Boolean(value);
         if (!rfb) return;
-        if (value) { releaseDrag(); gesture = null; modifier = null; if (shortcuts) shortcuts.hidden = true; }
-        rfb.viewOnly = Boolean(value);
+        if (wantedViewOnly) { releaseDrag(); gesture = null; modifier = null; if (shortcuts) shortcuts.hidden = true; }
+        rfb.viewOnly = wantedViewOnly;
         updatePointerControls();
         if (rfb.viewOnly) { keyboard.blur(); resetKeyboard(); }
+        // Taking over has to be usable, not just painted: focus so the very
+        // first keystroke after the veil lands on the desktop.
+        else { try { rfb.focus(); } catch (_) {} }
       }
 
       function connect() {
@@ -368,10 +391,13 @@
         pinTaskbar();
         updatePointerControls();
         rfb.addEventListener("connect", () => {
+          everConnected = true;
+          eagerAttempts = 0;
           setStatus("");
           pinTaskbar();
           updatePointerControls();
           try { rfb.focus(); } catch (_) {}
+          if (wantedViewOnly !== null) applyViewOnly(wantedViewOnly);
           window.parent.postMessage({ type: "lazyboy-desktop-ready" }, parentOrigin);
         });
         rfb.addEventListener("disconnect", (event) => {
@@ -382,7 +408,7 @@
           const clean = event && event.detail && event.detail.clean;
           setStatus(clean ? "Disconnected — retrying" : "Desktop connection lost — retrying");
           if (reconnectTimer) clearTimeout(reconnectTimer);
-          reconnectTimer = setTimeout(connect, 1500);
+          reconnectTimer = setTimeout(connect, retryDelay());
         });
         rfb.addEventListener("clipboard", (event) => {
           const text = event && event.detail ? event.detail.text : "";
diff --git a/crates/api/src/computer.rs b/crates/api/src/computer.rs
index 047e576..d8cf3bc 100644
--- a/crates/api/src/computer.rs
+++ b/crates/api/src/computer.rs
@@ -3,8 +3,8 @@ use std::time::Duration;
 
 use chrono::{TimeDelta, Utc};
 use lazyboy_contracts::{
-    BrowserProfileMode, ComputerCapabilities, ComputerMode, ComputerState, ComputerStatus,
-    ControlHolder, DEFAULT_SCREEN_HEIGHT, DEFAULT_SCREEN_WIDTH,
+    BrowserProfileMode, ComputerCapabilities, ComputerMode, ComputerStatus, ControlHolder,
+    DEFAULT_SCREEN_HEIGHT, DEFAULT_SCREEN_WIDTH,
 };
 use lazyboy_control::{
     AdapterContext, CommandRequest, EnsureScreenRequest, ProvisionRequest, admit_gui,
@@ -998,11 +998,6 @@ pub async fn heartbeat(state: &AppState, actor: &Actor, bot_id: &str) -> Result<
     Ok(())
 }
 
-#[allow(dead_code)]
-pub fn user_has_control(computer: &ComputerRow, bot_id: &str) -> bool {
-    user_has_screen_control(computer, None, bot_id)
-}
-
 pub fn user_has_screen_control(
     computer: &ComputerRow,
     screen: Option<&ScreenRow>,
@@ -1232,7 +1227,3 @@ struct ActiveRunRow {
     thread_id: String,
     step: Option<String>,
 }
-
-pub fn _keep_state(state: ComputerState, holder: ControlHolder, mode: BrowserProfileMode) {
-    let _ = (state, holder, mode);
-}
diff --git a/crates/control/src/a11y.rs b/crates/control/src/a11y.rs
index f5de9c6..dc3a1eb 100644
--- a/crates/control/src/a11y.rs
+++ b/crates/control/src/a11y.rs
@@ -1,8 +1,8 @@
 use lazyboy_contracts::UiElement;
 use serde_json::{Value, json};
 
+use crate::normalize_display;
 use crate::x11::{is_browser_title, parse_ui_elements};
-use crate::{PRIMARY_DISPLAY, normalize_display};
 
 const A11Y_PY: &str = include_str!("a11y.py");
 
@@ -30,10 +30,6 @@ pub fn a11y_command_on(display: &str, request: &Value) -> Vec<String> {
     ]
 }
 
-pub fn a11y_command(request: &Value) -> Vec<String> {
-    a11y_command_on(PRIMARY_DISPLAY, request)
-}
-
 pub fn parse_a11y_page(raw: &str) -> A11yPage {
     let value: Value = serde_json::from_str(raw.trim()).unwrap_or(Value::Null);
     let ok = value.get("ok").and_then(Value::as_bool) == Some(true);
diff --git a/crates/control/src/cdp.rs b/crates/control/src/cdp.rs
index e4f7dd8..b68ccf7 100644
--- a/crates/control/src/cdp.rs
+++ b/crates/control/src/cdp.rs
@@ -2,8 +2,8 @@ use lazyboy_contracts::UiElement;
 use serde::{Deserialize, Serialize};
 use serde_json::{Value, json};
 
+use crate::normalize_display;
 use crate::x11::parse_ui_elements;
-use crate::{PRIMARY_DISPLAY, normalize_display};
 
 const CDP_PY: &str = include_str!("cdp.py");
 
@@ -39,10 +39,6 @@ pub fn cdp_command_on(display: &str, profile: Option<&str>, request: &Value) ->
     ]
 }
 
-pub fn cdp_command(request: &Value) -> Vec<String> {
-    cdp_command_on(PRIMARY_DISPLAY, None, request)
-}
-
 /// Same as `cdp_command_on` but the JSON body is meant to arrive on stdin
 /// so secrets never appear on the process argv.
 pub fn cdp_stdin_command_on(display: &str, profile: Option<&str>) -> Vec<String> {
diff --git a/crates/control/src/cua/browser.rs b/crates/control/src/cua/browser.rs
index fcb9c9f..2a1090f 100644
--- a/crates/control/src/cua/browser.rs
+++ b/crates/control/src/cua/browser.rs
@@ -8,8 +8,6 @@ use crate::controller::ControlError;
 use crate::process::spawn_detached;
 use crate::{BrowserRequest, CdpPage, launch_argv_on};
 
-pub const SESSION: &str = "lazyboy";
-
 #[derive(Debug, Clone)]
 pub struct BrowserBind {
     pub pid: u64,
@@ -253,12 +251,7 @@ async fn attach(
     window: &ListedWindow,
 ) -> Result<BrowserBind, ControlError> {
     client
-        .call(
-            display,
-            "start_session",
-            &json!({ "session": SESSION }),
-            &[],
-        )
+        .call(display, "start_session", &json!({}), &[])
         .await?;
     let pid = window.pid;
     let window_id = window.id;
@@ -269,7 +262,6 @@ async fn attach(
             &json!({
                 "pid": pid,
                 "window_id": window_id,
-                "session": SESSION,
                 "strategy": { "kind": "existing_profile" },
                 "allow_launch": false,
             }),
@@ -289,7 +281,6 @@ async fn attach(
             &json!({
                 "pid": pid,
                 "window_id": window_id,
-                "session": SESSION,
                 "include_screenshot": false,
             }),
             &[],
@@ -317,7 +308,6 @@ pub async fn snapshot(
             &json!({
                 "target_id": bind.target_id,
                 "tab_id": bind.tab_id,
-                "session": SESSION,
                 "snapshot_format": "semantic_v2",
                 "include_screenshot": false,
             }),
@@ -376,7 +366,6 @@ pub async fn run(
                     &json!({
                         "target_id": attached.target_id,
                         "tab_id": attached.tab_id,
-                        "session": SESSION,
                         "url": url,
                     }),
                     &[],
@@ -397,7 +386,6 @@ pub async fn run(
                         "key": key,
                         "pid": attached.pid,
                         "window_id": attached.window_id,
-                        "session": SESSION,
                     }),
                     &[],
                 )
@@ -422,7 +410,7 @@ async fn click(
         .as_deref()
         .ok_or_else(|| ControlError::InvalidAction("browser click needs a selector".into()))?;
     let page = bind.page.as_ref().ok_or(ControlError::StaleReference)?;
-    let Some(r#ref) = find_ref(&page, selector) else {
+    let Some(r#ref) = find_ref(page, selector) else {
         return Ok(CdpPage {
                 ok: false,
                 error: Some(
@@ -444,7 +432,6 @@ async fn click(
             &json!({
                 "target_id": bind.target_id,
                 "tab_id": bind.tab_id,
-                "session": SESSION,
                 "ref": r#ref,
                 "input_route": "dom_event",
             }),
@@ -465,7 +452,7 @@ async fn type_into(
     tracing::info!(backend = "cua", tool = "browser_type", length = text.len());
     if let Some(selector) = request.selector.as_deref() {
         let page = bind.page.as_ref().ok_or(ControlError::StaleReference)?;
-        let Some(r#ref) = find_ref(&page, selector) else {
+        let Some(r#ref) = find_ref(page, selector) else {
             return Ok(CdpPage {
                 ok: false,
                 error: Some("target field is unavailable; no text inserted".into()),
@@ -484,7 +471,6 @@ async fn type_into(
                 &json!({
                     "target_id": bind.target_id,
                     "tab_id": bind.tab_id,
-                    "session": SESSION,
                     "ref": r#ref,
                     "text": text,
                     "replace": true,
@@ -501,7 +487,6 @@ async fn type_into(
                     "text": text,
                     "pid": bind.pid,
                     "window_id": bind.window_id,
-                    "session": SESSION,
                 }),
                 &[],
             )
diff --git a/crates/control/src/cua/client.rs b/crates/control/src/cua/client.rs
index 1970500..d5519c0 100644
--- a/crates/control/src/cua/client.rs
+++ b/crates/control/src/cua/client.rs
@@ -2,7 +2,7 @@ use std::path::PathBuf;
 use std::process::Stdio;
 use std::time::{Duration, Instant};
 
-use serde_json::Value;
+use serde_json::{Value, json};
 use tokio::io::AsyncWriteExt;
 use tokio::process::Command;
 
@@ -36,6 +36,16 @@ impl CuaClient {
         }
     }
 
+    /// Every call runs in its own CLI process, so the driver would otherwise
+    /// give each one an implicit session that dies with the process. Trajectory
+    /// recording, snapshots, and browser binds only line up under one label.
+    pub fn session_for_display(display: &str) -> String {
+        format!(
+            "lazyboy-{}",
+            normalize_display(display).trim_start_matches(':')
+        )
+    }
+
     pub fn dbus_file(display: &str) -> PathBuf {
         let number = normalize_display(display)
             .trim_start_matches(':')
@@ -79,6 +89,33 @@ impl CuaClient {
         payload: &Value,
         extra: &[&str],
     ) -> Result<Value, ControlError> {
+        let mut body = with_session_label(screen, payload);
+        let mut escalated = false;
+        loop {
+            let outcome = self.attempt(screen, tool, &body, extra).await?;
+            if !escalated
+                && outcome.error.is_some()
+                && let Some(mode) = recommended_delivery(&outcome.value)
+                && let Some(map) = body.as_object_mut()
+            {
+                map.insert("delivery_mode".to_string(), json!(mode));
+                escalated = true;
+                continue;
+            }
+            return match outcome.error {
+                Some(error) => Err(error),
+                None => Ok(outcome.value),
+            };
+        }
+    }
+
+    async fn attempt(
+        &self,
+        screen: &str,
+        tool: &str,
+        payload: &Value,
+        extra: &[&str],
+    ) -> Result<Outcome, ControlError> {
         let socket = Self::socket_for_display(screen);
         if !socket.exists() {
             return Err(ControlError::DriverUnavailable);
@@ -102,14 +139,12 @@ impl CuaClient {
         let output = bounded_input_output(&mut command, payload).await?;
         let stdout = String::from_utf8_lossy(&output.stdout);
         let stderr = String::from_utf8_lossy(&output.stderr);
-        let combined = format!("{stdout}\n{stderr}");
-        let value =
-            parse_jsonish(&stdout).unwrap_or_else(|| Value::String(stdout.trim().to_string()));
-        let error = if !output.status.success() || stdout.trim_start().starts_with('❌') {
-            Some(classify_cua_failure(&combined))
-        } else {
-            response_error(&value)
-        };
+        let (value, error) = decode_stdout(
+            &stdout,
+            &stderr,
+            output.status.success(),
+            classify_cua_failure,
+        );
         tracing::info!(
             backend = "cua",
             tool,
@@ -117,14 +152,56 @@ impl CuaClient {
             duration_ms = started.elapsed().as_millis() as u64,
             success = error.is_none()
         );
-        if let Some(error) = error {
-            return Err(error);
-        }
-
-        Ok(value)
+        Ok(Outcome { value, error })
     }
 }
 
+struct Outcome {
+    value: Value,
+    error: Option<ControlError>,
+}
+
+/// Refusals and prose diagnostics can arrive with exit code 0, so only a JSON
+/// object counts as proof that the driver ran the tool.
+fn decode_stdout(
+    stdout: &str,
+    stderr: &str,
+    success: bool,
+    classify: impl Fn(&str) -> ControlError,
+) -> (Value, Option<ControlError>) {
+    let combined = format!("{stdout}\n{stderr}");
+    let trimmed = stdout.trim();
+    let empty = Value::Null;
+    let decoded = (!trimmed.is_empty())
+        .then(|| parse_jsonish(trimmed))
+        .flatten();
+    let error = if !success || trimmed.starts_with('\u{274c}') || decoded.is_none() {
+        Some(classify(&combined))
+    } else {
+        response_error(decoded.as_ref().unwrap_or(&empty))
+    };
+    (decoded.unwrap_or(empty), error)
+}
+
+/// The documented contract for `background_unavailable` is one retry with the
+/// delivery mode named in the structured escalation.
+fn recommended_delivery(value: &Value) -> Option<String> {
+    ["/escalation/recommended", "/error/escalation/recommended"]
+        .into_iter()
+        .filter_map(|pointer| value.pointer(pointer).and_then(Value::as_str))
+        .find(|mode| matches!(*mode, "foreground" | "background"))
+        .map(str::to_string)
+}
+
+fn with_session_label(display: &str, payload: &Value) -> Value {
+    let mut body = payload.clone();
+    if let Some(map) = body.as_object_mut() {
+        map.entry("session")
+            .or_insert_with(|| json!(CuaClient::session_for_display(display)));
+    }
+    body
+}
+
 fn response_error(value: &Value) -> Option<ControlError> {
     if value
         .get("code")
@@ -166,7 +243,15 @@ async fn bounded_input_output(
             Ok::<(), std::io::Error>(())
         };
         let (written, output) = tokio::join!(write, child.wait_with_output());
-        written.map_err(|_| ControlError::DriverUnhealthy)?;
+        // A driver that exits before reading stdin (unknown tool, rejected
+        // arguments) closes the pipe, so a broken pipe is expected and the exit
+        // status plus stderr hold the real reason. Losing them here would
+        // downgrade every fast refusal to a generic driver failure.
+        if let Err(error) = written
+            && error.kind() != std::io::ErrorKind::BrokenPipe
+        {
+            return Err(ControlError::DriverUnhealthy);
+        }
         output.map_err(|_| ControlError::DriverUnhealthy)
     })
     .await
@@ -291,6 +376,44 @@ mod tests {
         assert!(response_error(&serde_json::json!({"outline": "❌ payment declined"})).is_none());
     }
 
+    #[test]
+    fn only_a_json_object_proves_the_tool_ran() {
+        let classify = |text: &str| classify_cua_failure(text);
+        let (value, error) = decode_stdout(r#"{"width":1280}"#, "", true, classify);
+        assert!(error.is_none());
+        assert_eq!(value["width"], 1280);
+
+        // Prose with a zero exit code used to be reported as success.
+        let (value, error) = decode_stdout("no window matched", "", true, classify);
+        assert!(matches!(error, Some(ControlError::DriverUnhealthy)));
+        assert!(value.is_null());
+
+        let (_, error) = decode_stdout("\u{274c} unsupported tool", "", true, classify);
+        assert!(matches!(error, Some(ControlError::Unsupported)));
+
+        let (_, error) = decode_stdout(r#"{"code":"background_unavailable"}"#, "", true, classify);
+        assert!(error.is_some());
+
+        // A crash with empty stdout must never look like an empty success.
+        let (_, error) = decode_stdout("", "signal: 11", false, classify);
+        assert!(error.is_some());
+    }
+
+    #[test]
+    fn escalation_is_read_from_the_documented_pointers() {
+        assert_eq!(
+            recommended_delivery(&json!({"escalation": {"recommended": "foreground"}})).as_deref(),
+            Some("foreground")
+        );
+        assert_eq!(
+            recommended_delivery(&json!({"error": {"escalation": {"recommended": "background"}}}))
+                .as_deref(),
+            Some("background")
+        );
+        assert!(recommended_delivery(&json!({"escalation": {"recommended": "reboot"}})).is_none());
+        assert!(recommended_delivery(&json!({"ok": true})).is_none());
+    }
+
     #[tokio::test]
     async fn piped_json_reaches_eof_without_argv_exposure() {
         let mut command = Command::new("sh");
@@ -319,6 +442,27 @@ mod tests {
         ));
     }
 
+    #[tokio::test]
+    async fn driver_that_never_reads_stdin_still_reports_its_exit() {
+        // Bigger than the pipe buffer: the child never reads, so the write can
+        // only fail with EPIPE and must not swallow the driver's own error.
+        let payload = json!({ "blob": "x".repeat(1 << 20) });
+        let mut command = Command::new("sh");
+        command.args(["-c", "echo invalid_action_target >&2; exit 1"]);
+        let output = tokio::time::timeout(
+            Duration::from_secs(10),
+            bounded_input_output(&mut command, &payload),
+        )
+        .await
+        .expect("bounded")
+        .expect("exit status survives a broken stdin pipe");
+        assert!(!output.status.success());
+        assert!(
+            String::from_utf8_lossy(&output.stderr).contains("invalid_action_target"),
+            "driver stderr must reach the classifier"
+        );
+    }
+
     #[test]
     fn unknown_driver_failure_does_not_echo_secret() {
         let error = classify_cua_failure("failed typing secret-password");
diff --git a/crates/control/src/cua/mod.rs b/crates/control/src/cua/mod.rs
index e619ca1..b039740 100644
--- a/crates/control/src/cua/mod.rs
+++ b/crates/control/src/cua/mod.rs
@@ -7,6 +7,7 @@ mod translate;
 use std::collections::HashMap;
 use std::path::PathBuf;
 use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
 use std::time::{SystemTime, UNIX_EPOCH};
 
 use async_trait::async_trait;
@@ -22,12 +23,29 @@ use crate::controller::{
 use crate::process::spawn_detached;
 use crate::{
     ActionRequest, ActionResult, BrowserRequest, CdpPage, RecordingRequest, RecordingResult,
-    RecordingSession, action_pause_ms, normalize_display, observation_from_png,
+    RecordingSession, action_pause_ms, image_dimensions, normalize_display, observation_from_png,
     observation_with_elements, teach_trajectory_dir,
 };
 use client::first_array_of_objects;
 
 pub use client::CuaClient;
+
+/// Last resort when neither the screenshot nor the driver reports a mode.
+const FALLBACK_SCREEN: (u32, u32) = (1280, 800);
+
+/// `image/computer/Dockerfile` pins `CUA_DRIVER_RS_VERSION`; only that tool
+/// surface is guaranteed. Patch releases stay compatible, a new minor does not.
+const PINNED_DRIVER: (u32, u32) = (0, 23);
+
+/// `cua-driver --version` prints `cua-driver 0.23.2`, and some builds append a
+/// target suffix (`cua-driver 0.23.2 (x86_64-linux)`), so only the leading
+/// numeric version is trusted.
+fn driver_release(version: &str) -> Option<(u32, u32)> {
+    let digits = version.find(|character: char| character.is_ascii_digit())?;
+    let mut parts = version[digits..].split('.');
+    Some((parts.next()?.parse().ok()?, parts.next()?.parse().ok()?))
+}
+
 pub use translate::{TranslatedAction, translate_action};
 
 #[derive(Debug, Default)]
@@ -36,6 +54,7 @@ pub struct CuaController {
     screens: tokio::sync::Mutex<HashMap<String, Arc<tokio::sync::Mutex<()>>>>,
     native: tokio::sync::Mutex<HashMap<String, HashMap<String, native::NativeTarget>>>,
     browser: tokio::sync::Mutex<HashMap<String, browser::BrowserBind>>,
+    snapshots: AtomicU64,
 }
 
 #[async_trait]
@@ -52,7 +71,9 @@ impl ComputerController for CuaController {
             .await;
         match report {
             Ok(report) => {
-                let compatible = version.as_deref() == Some("cua-driver 0.23.2");
+                let compatible = version
+                    .as_deref()
+                    .is_some_and(|text| driver_release(text) == Some(PINNED_DRIVER));
                 let healthy =
                     compatible && report.get("overall").and_then(Value::as_str) == Some("ok");
                 let mut details: Vec<String> = report
@@ -64,7 +85,12 @@ impl ComputerController for CuaController {
                     .filter_map(|check| check["message"].as_str().map(str::to_string))
                     .collect();
                 if !compatible {
-                    details.push("expected pinned Cua Driver 0.23.2".into());
+                    details.push(format!(
+                        "Cua Driver {} is not the pinned {}.{} series",
+                        version.as_deref().unwrap_or("unknown"),
+                        PINNED_DRIVER.0,
+                        PINNED_DRIVER.1
+                    ));
                 }
                 Ok(ControllerHealth {
                     backend: "cua".into(),
@@ -292,15 +318,6 @@ impl CuaController {
                 completed += 5;
                 continue;
             }
-            if matches!(
-                action,
-                ComputerAction::Pointer {
-                    pointer_type: PointerType::Down | PointerType::Up,
-                    ..
-                }
-            ) {
-                return Err(ControlError::Unsupported);
-            }
             if let ComputerAction::Ref {
                 verb,
                 target,
@@ -362,7 +379,10 @@ impl CuaController {
         if png.is_empty() {
             return Err(ControlError::Internal("screenshot failed".into()));
         }
-        let (width, height) = png_dimensions(&png);
+        let (width, height) = match image_dimensions(&png) {
+            Some(dimensions) => dimensions,
+            None => self.screen_size(display).await,
+        };
         let cursor = self.cursor(display).await;
         let windows = self.windows(display).await.unwrap_or_default();
         let active = windows
@@ -372,8 +392,15 @@ impl CuaController {
                 id: window.id.to_string(),
                 title: Some(window.title.clone()).filter(|title| !title.is_empty()),
             });
-        let (mut elements, targets) =
-            native::observe(&self.client, display, &windows, &png_path.to_string_lossy()).await?;
+        // Selectors only need to be short and unique: a process-local counter
+        // keeps stale handles from resolving without pasting a temp path into
+        // every identifier the Agent echoes back.
+        let snapshot = self.snapshots.fetch_add(1, Ordering::Relaxed);
+        let native::NativeObservation {
+            mut elements,
+            targets,
+            complete,
+        } = native::observe(&self.client, display, &windows, &format!("s{snapshot}")).await;
         elements.extend(
             windows
                 .into_iter()
@@ -402,10 +429,38 @@ impl CuaController {
             observation_from_png(png, width, height, cursor, active),
             elements,
         );
-        observation.native_observation_complete = true;
+        observation.native_observation_complete = complete;
         Ok(observation)
     }
 
+    /// `scroll` on the desktop plane is aimed at a point and the action DSL
+    /// does not carry one, so aim at the pointer; the screen centre is the
+    /// next best guess when the pointer cannot be read.
+    async fn scroll_point(&self, display: &str) -> (u32, u32) {
+        if let Some(cursor) = self.cursor(display).await
+            && cursor.x >= 0
+            && cursor.y >= 0
+        {
+            return (cursor.x as u32, cursor.y as u32);
+        }
+        let (width, height) = self.screen_size(display).await;
+        (width / 2, height / 2)
+    }
+
+    async fn screen_size(&self, display: &str) -> (u32, u32) {
+        self.client
+            .call(display, "get_screen_size", &json!({}), &[])
+            .await
+            .ok()
+            .and_then(|value| {
+                Some((
+                    value.get("width").and_then(Value::as_u64)? as u32,
+                    value.get("height").and_then(Value::as_u64)? as u32,
+                ))
+            })
+            .unwrap_or(FALLBACK_SCREEN)
+    }
+
     async fn cursor(&self, display: &str) -> Option<CursorPosition> {
         let value = self
             .client
@@ -449,6 +504,11 @@ impl CuaController {
             }
             TranslatedAction::FocusTitle { title } => self.focus_title(display, &title).await,
             TranslatedAction::Cua { tool, mut payload } => {
+                if tool == "scroll" && payload.get("x").is_none() {
+                    let (x, y) = self.scroll_point(display).await;
+                    payload["x"] = json!(x);
+                    payload["y"] = json!(y);
+                }
                 if tool == "drag" {
                     let x = payload["from_x"].as_f64().unwrap_or(0.0);
                     let y = payload["from_y"].as_f64().unwrap_or(0.0);
@@ -662,12 +722,6 @@ fn number(value: &Value, key: &str) -> Option<u64> {
     })
 }
 
-fn png_dimensions(bytes: &[u8]) -> (u32, u32) {
-    image::load_from_memory(bytes)
-        .map(|image| (image.width(), image.height()))
-        .unwrap_or((1280, 800))
-}
-
 fn observe_png_path(display: &str) -> PathBuf {
     let nanos = SystemTime::now()
         .duration_since(UNIX_EPOCH)
@@ -683,6 +737,23 @@ fn observe_png_path(display: &str) -> PathBuf {
 mod tests {
     use super::*;
 
+    #[test]
+    fn driver_release_reads_the_pinned_minor_series() {
+        assert_eq!(driver_release("cua-driver 0.23.2"), Some((0, 23)));
+        assert_eq!(
+            driver_release("cua-driver 0.23.2 (x86_64-linux)"),
+            Some((0, 23))
+        );
+        assert_eq!(driver_release("cua-driver"), None);
+        assert_eq!(driver_release(""), None);
+    }
+
+    #[test]
+    fn a_new_minor_series_is_not_compatible() {
+        assert_ne!(driver_release("cua-driver 0.24.0"), Some(PINNED_DRIVER));
+        assert_eq!(driver_release("cua-driver 0.23.9"), Some(PINNED_DRIVER));
+    }
+
     #[test]
     fn normalized_drag_uses_one_driver_gesture() {
         let actions = crate::parse_computer_actions(
diff --git a/crates/control/src/cua/native.rs b/crates/control/src/cua/native.rs
index df04286..3eca0a3 100644
--- a/crates/control/src/cua/native.rs
+++ b/crates/control/src/cua/native.rs
@@ -13,19 +13,30 @@ pub(super) struct NativeTarget {
     token: String,
 }
 
+#[derive(Debug, Default)]
+pub(super) struct NativeObservation {
+    pub(super) elements: Vec<UiElement>,
+    pub(super) targets: HashMap<String, NativeTarget>,
+    /// False when any window answered with a degraded tree, which means the
+    /// element list is partial and a second opinion is still worth taking.
+    pub(super) complete: bool,
+}
+
 pub(super) async fn observe(
     client: &CuaClient,
     display: &str,
     windows: &[ListedWindow],
-    generation: &str,
-) -> Result<(Vec<UiElement>, HashMap<String, NativeTarget>), ControlError> {
-    let mut elements = Vec::new();
-    let mut targets = HashMap::new();
+    snapshot_id: &str,
+) -> NativeObservation {
+    let mut observed = NativeObservation {
+        complete: true,
+        ..NativeObservation::default()
+    };
     for window in windows
         .iter()
         .filter(|window| !window.app_name.to_ascii_lowercase().contains("chrom"))
     {
-        let state = client
+        let state = match client
             .call(
                 display,
                 "get_window_state",
@@ -35,7 +46,20 @@ pub(super) async fn observe(
                 }),
                 &[],
             )
-            .await?;
+            .await
+        {
+            Ok(state) => state,
+            Err(_) => {
+                observed.complete = false;
+                continue;
+            }
+        };
+        // A degraded reply is window metadata plus a root node: usable for
+        // discovery, not as an element tree, and never as proof of coverage.
+        if state.get("degraded").and_then(Value::as_bool) == Some(true) {
+            observed.complete = false;
+            continue;
+        }
         for item in state
             .get("elements")
             .and_then(Value::as_array)
@@ -48,7 +72,7 @@ pub(super) async fn observe(
             if item.get("enabled").and_then(Value::as_bool) == Some(false) {
                 continue;
             }
-            let selector = format!("cua:{generation}:{}:{token}", window.id);
+            let selector = format!("cua:{snapshot_id}:{}:{token}", window.id);
             let title = item.get("label").and_then(Value::as_str).unwrap_or("");
             let role = item.get("role").and_then(Value::as_str).unwrap_or("");
             let frame = item.get("frame").unwrap_or(&Value::Null);
@@ -59,7 +83,7 @@ pub(super) async fn observe(
                     .unwrap_or(0.0)
                     .clamp(0.0, u32::MAX as f64) as u32
             };
-            elements.push(UiElement {
+            observed.elements.push(UiElement {
                 id: 0,
                 title: if title.is_empty() {
                     role.into()
@@ -74,7 +98,7 @@ pub(super) async fn observe(
                 kind: Some("a11y".into()),
                 role: Some(role.into()),
             });
-            targets.insert(
+            observed.targets.insert(
                 selector,
                 NativeTarget {
                     pid: window.pid,
@@ -84,7 +108,7 @@ pub(super) async fn observe(
             );
         }
     }
-    Ok((elements, targets))
+    observed
 }
 
 pub(super) async fn act(
diff --git a/crates/control/src/cua/translate.rs b/crates/control/src/cua/translate.rs
index bdbd491..7b330c4 100644
--- a/crates/control/src/cua/translate.rs
+++ b/crates/control/src/cua/translate.rs
@@ -36,7 +36,7 @@ pub fn translate_action(
             y,
             pointer_type,
             button,
-        } => Ok(translate_pointer(*x, *y, pointer_type, *button)),
+        } => translate_pointer(*x, *y, pointer_type, *button),
         ComputerAction::Clipboard { text } => {
             tracing::info!(backend = "cua", tool = "type_text", length = text.len());
             Ok(TranslatedAction::Cua {
@@ -44,7 +44,6 @@ pub fn translate_action(
                 payload: json!({
                     "text": text,
                     "scope": "desktop",
-                    "target": { "kind": "desktop", "display_id": "primary" },
                 }),
             })
         }
@@ -56,10 +55,9 @@ pub fn translate_action(
                     ScrollDirection::Up => "up",
                     ScrollDirection::Down => "down",
                 },
-                "amount": amount.unwrap_or(12),
+                "amount": amount.unwrap_or(12).clamp(1, 50),
                 "by": "line",
                 "scope": "desktop",
-                "target": { "kind": "desktop", "display_id": "primary" },
             }),
         }),
     }
@@ -70,40 +68,33 @@ fn translate_pointer(
     y: u32,
     pointer_type: &PointerType,
     button: Option<PointerButton>,
-) -> TranslatedAction {
+) -> Result<TranslatedAction, ControlError> {
     let button = match button.unwrap_or(PointerButton::Left) {
         PointerButton::Left => "left",
         PointerButton::Middle => "middle",
         PointerButton::Right => "right",
     };
     match pointer_type {
-        PointerType::Move => TranslatedAction::Cua {
+        PointerType::Move => Ok(TranslatedAction::Cua {
             tool: "move_cursor",
             payload: json!({
                 "x": x,
                 "y": y,
                 "scope": "desktop",
-                "target": { "kind": "desktop", "display_id": "primary" },
             }),
-        },
-        PointerType::Click => TranslatedAction::Cua {
+        }),
+        PointerType::Click => Ok(TranslatedAction::Cua {
             tool: "click",
             payload: json!({
                 "x": x,
                 "y": y,
                 "button": button,
                 "scope": "desktop",
-                "target": { "kind": "desktop", "display_id": "primary" },
             }),
-        },
-        PointerType::Down => TranslatedAction::Cua {
-            tool: "mouse_button_down",
-            payload: json!({ "x": x, "y": y, "button": button }),
-        },
-        PointerType::Up => TranslatedAction::Cua {
-            tool: "mouse_button_up",
-            payload: json!({ "x": x, "y": y, "button": button }),
-        },
+        }),
+        // A held-button gesture collapses into one `drag` call before it gets
+        // here; the CLI has no lease that keeps a button down between calls.
+        PointerType::Down | PointerType::Up => Err(ControlError::Unsupported),
     }
 }
 
@@ -118,7 +109,6 @@ fn translate_key(key: &str, modifiers: Option<&[String]>) -> TranslatedAction {
                 payload: json!({
                     "keys": keys,
                     "scope": "desktop",
-                    "target": { "kind": "desktop", "display_id": "primary" },
                 }),
             }
         }
@@ -127,7 +117,6 @@ fn translate_key(key: &str, modifiers: Option<&[String]>) -> TranslatedAction {
             payload: json!({
                 "key": key,
                 "scope": "desktop",
-                "target": { "kind": "desktop", "display_id": "primary" },
             }),
         },
     }
@@ -166,7 +155,76 @@ mod tests {
         assert_eq!(payload["x"], 40);
         assert_eq!(payload["y"], 80);
         assert_eq!(payload["scope"], "desktop");
-        assert_eq!(payload["target"]["display_id"], "primary");
+        // 0.23.2 rejects an explicit desktop target with invalid_action_target.
+        assert!(payload.get("target").is_none());
+    }
+
+    #[test]
+    fn desktop_actions_omit_the_rejected_target() {
+        for action in [
+            ComputerAction::Clipboard { text: "hi".into() },
+            ComputerAction::Key {
+                key: "return".into(),
+                modifiers: None,
+            },
+            ComputerAction::Key {
+                key: "c".into(),
+                modifiers: Some(vec!["ctrl".into()]),
+            },
+            ComputerAction::Pointer {
+                x: 1,
+                y: 2,
+                pointer_type: PointerType::Move,
+                button: None,
+            },
+            ComputerAction::Scroll {
+                direction: ScrollDirection::Down,
+                amount: None,
+            },
+        ] {
+            let TranslatedAction::Cua { payload, .. } =
+                translate_action(&action, ":1", None).unwrap()
+            else {
+                panic!("every desktop action translates to a Cua tool");
+            };
+            assert!(payload.get("target").is_none(), "{action:?}");
+        }
+    }
+
+    #[test]
+    fn scroll_amount_stays_inside_the_driver_window() {
+        let scroll = |amount| match translate_action(
+            &ComputerAction::Scroll {
+                direction: ScrollDirection::Down,
+                amount: Some(amount),
+            },
+            ":1",
+            None,
+        )
+        .unwrap()
+        {
+            TranslatedAction::Cua { payload, .. } => payload["amount"].as_i64().unwrap(),
+            other => panic!("expected scroll, got {other:?}"),
+        };
+        assert_eq!(scroll(0), 1);
+        assert_eq!(scroll(12), 12);
+        assert_eq!(scroll(200), 50);
+    }
+
+    #[test]
+    fn held_buttons_are_not_translated_one_by_one() {
+        for pointer_type in [PointerType::Down, PointerType::Up] {
+            let action = ComputerAction::Pointer {
+                x: 1,
+                y: 2,
+                pointer_type,
+                button: Some(PointerButton::Left),
+            };
+            assert!(matches!(
+                translate_action(&action, ":1", None),
+                Err(ControlError::Unsupported)
+            ));
+        }
     }
 
     #[test]
diff --git a/crates/control/src/lease.rs b/crates/control/src/lease.rs
index 497e475..9856c7e 100644
--- a/crates/control/src/lease.rs
+++ b/crates/control/src/lease.rs
@@ -1,15 +1,9 @@
-use thiserror::Error;
-
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ScreenLease {
     pub owner_id: String,
     pub fence: u32,
 }
 
-#[derive(Debug, Error, Clone, PartialEq, Eq)]
-#[error("computer is busy")]
-pub struct ComputerBusyError;
-
 pub fn screen_lease_id(run_id: &str, fence: u32) -> String {
     format!("{run_id}:{fence}")
 }
@@ -57,10 +51,6 @@ pub fn can_release_screen_lease(existing: Option<&str>, incoming: Option<&str>)
     }
 }
 
-pub fn next_fence(current: u32) -> u32 {
-    current.saturating_add(1)
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;
diff --git a/crates/control/src/legacy.rs b/crates/control/src/legacy.rs
index b74d20e..067f636 100644
--- a/crates/control/src/legacy.rs
+++ b/crates/control/src/legacy.rs
@@ -1,5 +1,7 @@
 use async_trait::async_trait;
-use lazyboy_contracts::{ComputerAction, ComputerObservation, RefVerb};
+use lazyboy_contracts::{
+    ComputerAction, ComputerObservation, DEFAULT_SCREEN_HEIGHT, DEFAULT_SCREEN_WIDTH, RefVerb,
+};
 use tokio::time::{Duration, sleep};
 
 use crate::controller::{
@@ -11,7 +13,7 @@ use crate::process::{
 use crate::{
     ActionRequest, ActionResult, BrowserRequest, CdpPage, RecordingRequest, RecordingResult,
     RecordingSession, a11y_command_on, action_pause_ms, cdp_command_on, cdp_record_command_on,
-    cdp_record_stop_command, devtools_port, launch_argv_on, observation_from_png,
+    cdp_record_stop_command, devtools_port, image_dimensions, launch_argv_on, observation_from_png,
     observation_with_elements, open_argv_on, parse_a11y_page, parse_cdp_page, parse_pointer_state,
     parse_ui_elements, pointer_state_command_on, screenshot_command_on, teach_recorder_output,
     teach_trajectory_dir, window_list_command_on, xdotool_argv_on,
@@ -153,13 +155,17 @@ async fn legacy_browser(
 }
 
 pub async fn observe_display(display: &str) -> Result<ComputerObservation, ControlError> {
-    let png = capture_stdout(&screenshot_command_on(display))
+    let frame = capture_stdout(&screenshot_command_on(display))
         .await
         .map_err(|_| ControlError::Internal("screenshot failed".into()))?;
+    // The capture is a JPEG of the root window, so the frame carries the real
+    // geometry; the constants are only a fallback for an undecodable frame.
+    let (width, height) =
+        image_dimensions(&frame).unwrap_or((DEFAULT_SCREEN_WIDTH, DEFAULT_SCREEN_HEIGHT));
     let ((cursor, window), elements) =
         tokio::join!(run_pointer_state(display), run_window_list(display));
     Ok(observation_with_elements(
-        observation_from_png(png, 1280, 800, cursor, window),
+        observation_from_png(frame, width, height, cursor, window),
         elements,
     ))
 }
diff --git a/crates/control/src/observe.rs b/crates/control/src/observe.rs
index 57cb3f5..4f21771 100644
--- a/crates/control/src/observe.rs
+++ b/crates/control/src/observe.rs
@@ -25,6 +25,20 @@ pub fn observation_from_png(
     }
 }
 
+/// Pixel size of a captured frame, or `None` when the bytes are not decodable.
+/// Each driver captures with a different codec (the legacy pipeline emits JPEG
+/// from `xwd | convert`, the Cua driver writes PNG), so the dimensions have to
+/// come from the frame itself rather than from a configured constant.
+pub fn image_dimensions(bytes: &[u8]) -> Option<(u32, u32)> {
+    use image::ImageDecoder;
+
+    let reader = image::ImageReader::new(std::io::Cursor::new(bytes))
+        .with_guessed_format()
+        .ok()?;
+    let decoder = reader.into_decoder().ok()?;
+    Some(decoder.dimensions())
+}
+
 pub fn observation_to_control_json(observation: &ComputerObservation) -> Value {
     let mut body = json!({
         "png_base64": base64::engine::general_purpose::STANDARD.encode(&observation.image),
@@ -109,6 +123,33 @@ pub fn signatures_similar(a: &[u8], b: &[u8]) -> bool {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use image::codecs::jpeg::JpegEncoder;
+    use image::codecs::png::PngEncoder;
+    use image::{ExtendedColorType, ImageEncoder, Rgb, RgbImage};
+    use std::io::Cursor;
+
+    #[test]
+    fn dimensions_come_from_the_frame_for_both_codecs() {
+        let frame = RgbImage::from_pixel(96, 48, Rgb([10, 20, 30]));
+
+        let mut png = Cursor::new(Vec::new());
+        PngEncoder::new(&mut png)
+            .write_image(frame.as_raw(), 96, 48, ExtendedColorType::Rgb8)
+            .unwrap();
+        assert_eq!(image_dimensions(&png.into_inner()), Some((96, 48)));
+
+        // The legacy driver captures JPEG, so a hardcoded size would drift the
+        // moment the Xvfb geometry changes.
+        let mut jpeg = Cursor::new(Vec::new());
+        JpegEncoder::new_with_quality(&mut jpeg, 60)
+            .encode(frame.as_raw(), 96, 48, ExtendedColorType::Rgb8)
+            .unwrap();
+        assert_eq!(image_dimensions(&jpeg.into_inner()), Some((96, 48)));
+
+        // A truncated capture must not invent a size.
+        assert_eq!(image_dimensions(&[0xFF, 0xD8, 0xFF]), None);
+        assert_eq!(image_dimensions(&[]), None);
+    }
 
     #[test]
     fn identical_bytes_share_a_frame_id() {
diff --git a/crates/control/src/x11.rs b/crates/control/src/x11.rs
index 65d6980..562ee44 100644
--- a/crates/control/src/x11.rs
+++ b/crates/control/src/x11.rs
@@ -7,7 +7,6 @@ pub fn is_browser_title(title: &str) -> bool {
 
 use crate::screen::{PRIMARY_DISPLAY, normalize_display};
 
-pub const DISPLAY: &str = PRIMARY_DISPLAY;
 pub const HOME: &str = "/home/lazyboy";
 
 fn display_env(display: &str) -> String {
@@ -170,10 +169,6 @@ pub fn action_pause_ms(action: &ComputerAction) -> u64 {
     }
 }
 
-pub fn pointer_state_command() -> Vec<String> {
-    pointer_state_command_on(PRIMARY_DISPLAY)
-}
-
 pub fn pointer_state_command_on(display: &str) -> Vec<String> {
     vec![
         "env".into(),
@@ -200,10 +195,6 @@ print(json.dumps({"x": int(vals.get("X") or 0), "y": int(vals.get("Y") or 0), "i
     ]
 }
 
-pub fn window_list_command() -> Vec<String> {
-    window_list_command_on(PRIMARY_DISPLAY)
-}
-
 pub fn window_list_command_on(display: &str) -> Vec<String> {
     vec![
         "env".into(),
@@ -309,10 +300,6 @@ pub fn parse_pointer_state(
     (cursor, window)
 }
 
-pub fn open_argv(path: &str) -> Vec<String> {
-    open_argv_on(PRIMARY_DISPLAY, None, path)
-}
-
 pub fn open_argv_on(display: &str, profile: Option<&str>, path: &str) -> Vec<String> {
     if path.starts_with("http://") || path.starts_with("https://") {
         return browser_argv(display, profile, Some(path));
@@ -325,10 +312,6 @@ pub fn open_argv_on(display: &str, profile: Option<&str>, path: &str) -> Vec<Str
     ]
 }
 
-pub fn launch_argv(application: &str, uri: Option<&str>) -> Option<Vec<String>> {
-    launch_argv_on(PRIMARY_DISPLAY, None, application, uri)
-}
-
 pub fn launch_argv_on(
     display: &str,
     profile: Option<&str>,
@@ -371,10 +354,6 @@ fn browser_argv(display: &str, profile: Option<&str>, uri: Option<&str>) -> Vec<
     argv
 }
 
-pub fn screenshot_command() -> Vec<String> {
-    screenshot_command_on(PRIMARY_DISPLAY)
-}
-
 pub fn screenshot_command_on(display: &str) -> Vec<String> {
     vec![
         "bash".into(),
diff --git a/crates/sandbox/src/docker.rs b/crates/sandbox/src/docker.rs
index 893cdec..630d61d 100644
--- a/crates/sandbox/src/docker.rs
+++ b/crates/sandbox/src/docker.rs
@@ -1,13 +1,16 @@
 use async_trait::async_trait;
 use base64::Engine;
 use lazyboy_contracts::ComputerCapabilities;
-use lazyboy_contracts::{ActiveWindow, ComputerObservation, CursorPosition, SandboxKind};
+use lazyboy_contracts::{
+    ActiveWindow, ComputerObservation, CursorPosition, DEFAULT_SCREEN_HEIGHT, DEFAULT_SCREEN_WIDTH,
+    SandboxKind,
+};
 use lazyboy_control::{
     ActionRequest, ActionResult, AdapterContext, BrowserRequest, CdpPage, CommandRequest,
     CommandResult, ComputerRef, EnsureScreenRequest, EnsureScreenResult, FileEntry,
     ProvisionRequest, RecordingRequest, RecordingResult, RecordingSession, SandboxError,
-    SandboxProvider, ScreenSession, observation_from_png, observation_with_elements,
-    parse_ui_elements,
+    SandboxProvider, ScreenSession, image_dimensions, observation_from_png,
+    observation_with_elements, parse_ui_elements,
 };
 use reqwest::Client;
 use serde_json::Value;
@@ -546,8 +549,13 @@ fn decode_observation(body: &Value) -> Result<ComputerObservation, SandboxError>
         .get("elements")
         .map(|value| parse_ui_elements(&value.to_string()))
         .unwrap_or_default();
+    // The control endpoint does not put the mode on the wire, so the frame
+    // itself decides: a desktop that is not 1280x800 must not be reported to
+    // the model as 1280x800.
+    let (width, height) =
+        image_dimensions(&png).unwrap_or((DEFAULT_SCREEN_WIDTH, DEFAULT_SCREEN_HEIGHT));
     let mut observation = observation_with_elements(
-        observation_from_png(png, 1280, 800, cursor, window),
+        observation_from_png(png, width, height, cursor, window),
         elements,
     );
     observation.native_observation_complete = body
@@ -556,3 +564,30 @@ fn decode_observation(body: &Value) -> Result<ComputerObservation, SandboxError>
         .unwrap_or(false);
     Ok(observation)
 }
+
+#[cfg(test)]
+mod tests {
+    use super::decode_observation;
+    use lazyboy_contracts::{DEFAULT_SCREEN_HEIGHT, DEFAULT_SCREEN_WIDTH};
+    use serde_json::json;
+
+    /// A 3x2 solid-red PNG: small enough to inline, real enough to decode.
+    const FRAME_3X2: &str = "iVBORw0KGgoAAAANSUhEUgAAAAMAAAACCAIAAAASFvFNAAAAEElEQVR4nGP4z8AAQQxwFgBB\
+0gX7h/C5SAAAAABJRU5ErkJggg==";
+
+    #[test]
+    fn the_mode_is_read_from_the_frame() {
+        let observation = decode_observation(&json!({ "png_base64": FRAME_3X2 }))
+            .expect("a decodable frame makes an observation");
+        assert_eq!(observation.width, 3);
+        assert_eq!(observation.height, 2);
+    }
+
+    #[test]
+    fn an_undecodable_frame_falls_back_to_the_default_mode() {
+        let observation =
+            decode_observation(&json!({ "png_base64": "AAAAAAAAAAA=" })).expect("base64 decodes");
+        assert_eq!(observation.width, DEFAULT_SCREEN_WIDTH);
+        assert_eq!(observation.height, DEFAULT_SCREEN_HEIGHT);
+    }
+}
diff --git a/docs/cua-compatibility.md b/docs/cua-compatibility.md
index 71448b2..cbf6406 100644
--- a/docs/cua-compatibility.md
+++ b/docs/cua-compatibility.md
@@ -56,11 +56,59 @@ Element refs are snapshot-scoped (`p1:1`, `p4:1`, … `p28:1`). Reusing an old r
 
 ## Relevant tools (0.23.2 `list-tools`)
 
-Observation / native input: `get_desktop_state`, `list_windows`, `get_accessibility_tree`, `get_window_state`, `click`, `type_text`, `press_key`, `hotkey`, `scroll`, `drag`, `get_cursor_position`, `get_screen_size`.
+Observation / native input used by `controld`: `get_desktop_state`, `list_windows`,
+`get_window_state`, `click`, `type_text`, `press_key`, `hotkey`, `scroll`, `drag`,
+`move_cursor`, `set_value`, `bring_to_front`, `get_cursor_position`, `get_screen_size`.
 
-Browser (attach only): `browser_prepare` (`strategy.kind=existing_profile`, `allow_launch=false`), `get_browser_state` (`semantic_v2`), `browser_navigate` (http/https/about only), `browser_click`, `browser_type`.
+Browser (attach only): `browser_prepare` (`strategy.kind=existing_profile`,
+`allow_launch=false`), `get_browser_state` (`semantic_v2`), `browser_navigate`
+(http/https/about only), `browser_click`, `browser_type`.
 
-Not used here: recording, isolated `launch_app` browsers, Wayland helpers. These names must not be exposed to the LLM.
+Lifecycle / diagnostics: `start_session`, `end_session`, `health_report`. Skill
+teaching also uses `start_recording` / `stop_recording`.
+
+Not used: `mouse_button_down`, `mouse_button_up`, `mouse_drag` (held-button
+background X11 tools). LazyBoy's action DSL has no partial-pointer state, so
+`Pointer{Down}` / `Pointer{Up}` translate to `ControlError::Unsupported` instead
+of a half-pressed button nobody releases. Also unused: isolated `launch_app`
+browsers, Wayland helpers, and the deprecated `get_session_state` /
+`escalate_session` aliases. None of these names may be exposed to the LLM.
+
+## Driver contract rules (enforced in `crates/control/src/cua`)
+
+Each of these was confirmed against a real 0.23.2 daemon, and each one fails
+silently (exit 0) if violated:
+
+1. **Repeat the `session` label on every call.** `CuaClient::call` injects
+   `lazyboy-<display>` unless the caller already set one. Without it each CLI
+   process gets an ephemeral `cli-<uuid>` session, so trajectory turns,
+   snapshots, and browser binds never line up, and every call also emits a bogus
+   `end_session` turn.
+2. **Never send `target: {kind: "desktop", display_id: "primary"}`.** The Linux
+   driver rejects it with `invalid_action_target` (exit 0). Omit `target` to use
+   the global input route.
+3. **Desktop `scroll` needs a point.** With `scope: "desktop"`, `x`/`y` are
+   required (`missing field x`); `dispatch` aims at the pointer and falls back to
+   the screen centre. `amount` is clamped to the schema range `1..=50`.
+4. **Only a JSON object proves the tool ran.** A refusal or prose banner that
+   arrives with exit 0 is a failure (`decode_stdout`), never an empty success.
+5. **One escalation retry.** `background_unavailable` carries
+   `escalation.recommended`; the client retries once with that `delivery_mode`
+   and never loops.
+6. **`get_window_state` can be degraded.** AT-SPI intermittently answers with
+   `degraded: true` and a root-only tree. Such windows are skipped and the
+   observation reports `native_observation_complete: false` rather than failing
+   the whole `observe`.
+7. **Nothing is validated for you.** The Linux schemas declare
+   `additionalProperties: false` and numeric bounds, but 0.23.2 accepts unknown
+   keys and out-of-range values anyway (a bogus key on `list_windows` and
+   `scroll amount: 0` both return exit 0 with `effect: unverifiable`), so the
+   bounds in `docs/cua-schemas/0.23.2/` are enforced here, by the client.
+   `session` is accepted by every tool, including the ones whose own schema
+   omits it (`list_windows`, `bring_to_front`, `health_report`,
+   `start_recording`), which is what lets rule 1 be applied uniformly.
+8. **Socket peer uid.** `/tmp/lazyboy/cua*.sock` rejects uid 0; `controld` runs as
+   the desktop user (uid 1000).
 
 ## Integration notes for the next PR
 
@@ -94,3 +142,16 @@ for the build architecture (`cua-driver 0.23.2` on linux/arm64 in this run).
 `make cua-smoke` is the acceptance entry: raw Driver smoke, adapter E2E,
 isolation, and pause/restart persistence. Production defaults remain `legacy`.
 See [the migration audit](cua-review.md).
+
+One upstream caveat about that persistence claim: Chromium writes its cookie
+database on a ~30 s timer and does not flush on `SIGTERM`. A profile survives
+`docker pause` / `docker restart` once that write has landed; stopping a desktop
+seconds after a login can still lose the cookie, and no LazyBoy code controls
+the timer. `scripts/cua-smoke-test.sh` waits for the fixture cookie to reach the
+profile before it restarts, so the check measures profile persistence instead of
+the flush timer.
+
+Image architectures are capped at `linux/amd64` and `linux/arm64` by upstream
+binaries: the Cua Driver ships only `linux-x86_64` / `linux-arm64` and ONNX
+Runtime only `linux-x64` / `linux-aarch64`. See
+[development.md](development.md#映像與-cpu-架構).
diff --git a/docs/cua-schemas/0.23.2/schema-drag.txt b/docs/cua-schemas/0.23.2/schema-drag.txt
new file mode 100644
index 0000000..80ec380
--- /dev/null
+++ b/docs/cua-schemas/0.23.2/schema-drag.txt
@@ -0,0 +1,147 @@
+name: drag
+
+description:
+Press-drag-release gesture from (from_x, from_y) to (to_x, to_y) in window-local screenshot pixels via XSendEvent (ButtonPress + MotionNotify × steps + ButtonRelease). duration_ms (default 500), steps (default 20). No focus steal.
+
+input_schema:
+{
+  "additionalProperties": false,
+  "properties": {
+    "button": {
+      "description": "Mouse button. Default \"left\".",
+      "enum": [
+        "left",
+        "right",
+        "middle"
+      ],
+      "type": "string"
+    },
+    "cursor_id": {
+      "description": "Optional multi-cursor instance id. Default: 'default'.",
+      "type": "string"
+    },
+    "delivery_mode": {
+      "default": "background",
+      "description": "Input delivery mode. 'background' (default) never activates or raises the target window. On X11 it injects via XTEST / the XInput2 master pointer (no focus steal). On Wayland it goes through libei + xdg-desktop-portal, which injects to the compositor's input focus — Wayland's security model has no per-window background targeting, so a specific non-focused window cannot be aimed at; when no libei backend is available the tool returns a structured background_unavailable error. 'foreground' is the explicit escalation: activate the target (X11 _NET_ACTIVE_WINDOW; Wayland compositor activate), inject, then restore the prior active window — a brief focus swap unless the target was already active. Matches the macOS / Windows delivery_mode surface.",
+      "enum": [
+        "background",
+        "foreground"
+      ],
+      "type": "string"
+    },
+    "duration_ms": {
+      "description": "Total drag duration. Default: 500.",
+      "maximum": 10000,
+      "minimum": 0,
+      "type": "integer"
+    },
+    "from_x": {
+      "type": "number"
+    },
+    "from_y": {
+      "type": "number"
+    },
+    "from_zoom": {
+      "type": "boolean"
+    },
+    "modifier": {
+      "description": "Modifier keys held during the action: cmd, shift, option/alt, ctrl.",
+      "items": {
+        "type": "string"
+      },
+      "type": "array"
+    },
+    "pid": {
+      "type": "integer"
+    },
+    "scope": {
+      "default": "window",
+      "enum": [
+        "window",
+        "desktop"
+      ],
+      "type": "string"
+    },
+    "session": {
+      "description": "For multi-call work, prefer a short public session label and repeat it on every call that accepts it. Omit it to use the authenticated transport's implicit lifecycle session.",
+      "type": "string"
+    },
+    "steps": {
+      "description": "Intermediate MotionNotify events. Default: 20.",
+      "maximum": 200,
+      "minimum": 1,
+      "type": "integer"
+    },
+    "target": {
+      "anyOf": [
+        {
+          "description": "Exact capture/input target selected independently for each action.\n\n`display_id=\"primary\"` is the portable desktop target in this release.\nPlatforms that cannot address another display reject it explicitly rather\nthan silently changing coordinate spaces.",
+          "oneOf": [
+            {
+              "additionalProperties": true,
+              "properties": {
+                "kind": {
+                  "const": "window",
+                  "type": "string"
+                },
+                "pid": {
+                  "format": "uint32",
+                  "minimum": 0,
+                  "type": "integer"
+                },
+                "window_id": {
+                  "format": "uint64",
+                  "minimum": 0,
+                  "type": "integer"
+                }
+              },
+              "required": [
+                "kind",
+                "pid",
+                "window_id"
+              ],
+              "type": "object"
+            },
+            {
+              "additionalProperties": true,
+              "properties": {
+                "display_id": {
+                  "type": "string"
+                },
+                "kind": {
+                  "const": "desktop",
+                  "type": "string"
+                }
+              },
+              "required": [
+                "kind",
+                "display_id"
+              ],
+              "type": "object"
+            }
+          ]
+        },
+        {
+          "type": "null"
+        }
+      ]
+    },
+    "to_x": {
+      "type": "number"
+    },
+    "to_y": {
+      "type": "number"
+    },
+    "window_id": {
+      "description": "Target window XID. Required.",
+      "type": "integer"
+    }
+  },
+  "required": [
+    "from_x",
+    "from_y",
+    "to_x",
+    "to_y"
+  ],
+  "type": "object"
+}
diff --git a/docs/cua-schemas/0.23.2/schema-end_session.txt b/docs/cua-schemas/0.23.2/schema-end_session.txt
new file mode 100644
index 0000000..e2f4efb
--- /dev/null
+++ b/docs/cua-schemas/0.23.2/schema-end_session.txt
@@ -0,0 +1,17 @@
+name: end_session
+
+description:
+End one visible lifecycle session and run its cursor, recording, configuration, and other cleanup hooks exactly once. Omit `session` to end the authenticated transport's implicit session. Idempotent.
+
+input_schema:
+{
+  "additionalProperties": true,
+  "properties": {
+    "session": {
+      "description": "Optional public label to end. When omitted, end the caller's attached\nimplicit session.",
+      "type": "string"
+    }
+  },
+  "required": [],
+  "type": "object"
+}
diff --git a/docs/cua-schemas/0.23.2/schema-start_session.txt b/docs/cua-schemas/0.23.2/schema-start_session.txt
new file mode 100644
index 0000000..6a7c5a1
--- /dev/null
+++ b/docs/cua-schemas/0.23.2/schema-start_session.txt
@@ -0,0 +1,50 @@
+name: start_session
+
+description:
+Optionally create or return a lifecycle session before acting. For multi-call work, prefer a short public `session` label and repeat it on every call that accepts it; an omitted value uses the authenticated transport lease's implicit session instead. This tool is optional because an ordinary action can create or reuse a named run directly. Use it to set the initial cursor theme before acting or to revive a public name after it has ended; ordinary actions never revive ended names. `capture_scope` is deprecated compatibility input; new callers select window or desktop modality per action. Idempotent.
+
+input_schema:
+{
+  "additionalProperties": true,
+  "properties": {
+    "capture_scope": {
+      "description": "Deprecated compatibility policy. New callers select window or desktop\nmodality on each action instead of storing it on the session.",
+      "enum": [
+        "auto",
+        "window",
+        "desktop"
+      ],
+      "type": "string"
+    },
+    "cursor_theme": {
+      "description": "Optional initial cursor theme. The host applies it before the cursor is\nfirst made visible, avoiding a flash of the default theme.",
+      "properties": {
+        "reduced_motion": {
+          "default": "auto",
+          "enum": [
+            "auto",
+            "on",
+            "off"
+          ],
+          "type": "string"
+        },
+        "theme_id": {
+          "type": "string"
+        }
+      },
+      "required": [
+        "theme_id"
+      ],
+      "type": [
+        "object",
+        "null"
+      ]
+    },
+    "session": {
+      "description": "Optional stable public label for this run (e.g. \"research-run-1\").\nWhen omitted, the authenticated transport lease's implicit session is\ncreated or returned.",
+      "type": "string"
+    }
+  },
+  "required": [],
+  "type": "object"
+}
diff --git a/docs/development.md b/docs/development.md
index d3145f7..f94a4aa 100644
--- a/docs/development.md
+++ b/docs/development.md
@@ -87,6 +87,30 @@ Lint 政策集中在三處，新增 crate 時只要補 `[lints] workspace = true
 依賴來源。它不是內建工具，第一次要先 `cargo install --locked cargo-deny`。
 已知無法升級的項目會寫進 `deny.toml` 的 `ignore` 並附原因與重檢時機。
 
+### 映像與 CPU 架構
+
+三個映像（`image/computer`、`image/api`、`image/supervisor`）一律用 `scripts/build-image.sh`
+建置。支援的架構是 **linux/amd64 與 linux/arm64**，這個上限由上游二進位發行決定，不是實作取捨：
+
+- Cua Driver `cua-driver-rs-v0.23.2` 的 Linux 發行檔只有 `linux-x86_64` 與 `linux-arm64`。
+- ONNX Runtime `v1.24.1` 的 Linux 發行檔只有 `linux-x64` 與 `linux-aarch64`。
+
+`image/computer/Dockerfile` 依 `TARGETARCH` 選檔並驗證 SHA256，`scripts/fetch-onnxruntime.sh`
+同樣只認這兩種架構，其他架構直接建置失敗；`build-image.sh` 在註冊 QEMU 與開始跨修編
+之前就先擋掉不支援的 `--platforms`，不會燒掉數 GB 才報錯。
+
+```bash
+make computer            # 本機架構，結果直接進本地 image 庫
+make computer-multi      # amd64 + arm64 -> dist/lazyboy-computer-multi.oci.tar
+make images-multi        # 三個映像都跨修 amd64 + arm64
+make images-multi PUSH=1 # 直接發布 manifest list 到 registry
+```
+
+多架構需要 `docker-container` builder 與 QEMU binfmt，`build-image.sh` 會自行建立 `lazyboy`
+builder 並註冊 binfmt；主機本身的架構會跳過註冊，否則在特權容器內會誤報失敗。
+不推送 registry 時輸出 OCI archive（`nerdctl load -i <檔>` 可載入），因為 Docker 的
+`docker` driver 無法匯出 manifest list。
+
 ### 專案結構
 
 ```text
diff --git a/image/computer/Dockerfile b/image/computer/Dockerfile
index c352593..d90aacb 100644
--- a/image/computer/Dockerfile
+++ b/image/computer/Dockerfile
@@ -8,6 +8,7 @@
 FROM rust:1-bookworm AS controld
 WORKDIR /src
 COPY Cargo.toml Cargo.lock ./
+COPY .cargo .cargo
 COPY crates crates
 COPY migrations migrations
 RUN --mount=type=cache,target=/usr/local/cargo/registry,sharing=locked \
@@ -16,8 +17,14 @@ RUN --mount=type=cache,target=/usr/local/cargo/registry,sharing=locked \
 
 FROM debian:bookworm-slim
 
-# Keep Chromium, XFCE, CJK, AT-SPI, git. Skip Debian novnc (pulls nodejs),
-# fonts-noto-core (Latin is DejaVu/Liberation/huninn), and unused xterm.
+# Keep Chromium, XFCE, CJK, AT-SPI, git. Skip Debian novnc (pulls nodejs) and
+# fonts-noto-core (Latin is DejaVu/Liberation/huninn).
+#
+# x11-apps + imagemagick look removable and are not: the legacy driver's
+# screenshot path is `xwd -root -silent | convert xwd:- … jpeg:-`
+# (control/src/x11.rs screenshot_command_on), reached by legacy observe_display
+# and by the supervisor's exec fallback. Nothing else in the image provides
+# /usr/bin/xwd or /usr/bin/convert, so removing either breaks the default driver.
 RUN printf '%s\n' \
       'path-include=/usr/share/locale/zh_TW/*' \
       'path-include=/usr/share/locale/zh/*' \
@@ -33,7 +40,6 @@ RUN printf '%s\n' \
     fonts-noto-color-emoji \
     git \
     htop \
-    imagemagick \
     locales \
     procps \
     python3 \
@@ -49,6 +55,7 @@ RUN printf '%s\n' \
     x11vnc \
     xclip \
     xdg-utils \
+    imagemagick \
     xfce4-panel \
     xfce4-settings \
     xfce4-terminal \
@@ -73,7 +80,10 @@ RUN printf '%s\n' \
   && echo "en_US.UTF-8 UTF-8" >> /etc/locale.gen \
   && locale-gen \
   && mkdir -p /usr/share/novnc \
-  && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/* /usr/share/doc/* /usr/share/man/*
+  && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/* /usr/share/doc/* /usr/share/man/* \
+       /usr/share/i18n /usr/share/info /usr/share/gtk-doc /usr/share/backgrounds \
+  && find /usr -name __pycache__ -type d -prune -exec rm -rf '{}' + \
+  && find /usr -name '*.py[co]' -delete
 
 # jf open 粉圓 (system UI) + MesloLGS NF (Powerlevel10k glyphs, CJK via fontconfig).
 RUN mkdir -p /usr/share/fonts/truetype/huninn /usr/share/fonts/truetype/meslo \
@@ -102,6 +112,27 @@ RUN useradd --create-home --uid 1000 --shell /bin/zsh lazyboy \
     && mkdir -p /home/lazyboy /tmp/lazyboy /usr/share/lazyboy/skel /usr/share/lazyboy/xfce-skel /etc/gtk-3.0 /etc/fonts/conf.d \
     && chown -R 1000:1000 /home/lazyboy /tmp/lazyboy
 
+# Pin Cua Driver outside the persisted /home/lazyboy bind-mount.
+# Checksums are from the matching official release; select the target architecture.
+ARG CUA_DRIVER_RS_VERSION=0.23.2
+ARG TARGETARCH
+ARG CUA_DRIVER_RS_SHA256_AMD64=01bf8339ec129cc00f4b4b2c6056ef1a7c5b52df39ff83ad17c9b16818aec500
+ARG CUA_DRIVER_RS_SHA256_ARM64=be22768a207796a4bc1de50c52f32f9ef680b5e86e58c059e02eec2caba2e7bb
+RUN case "$TARGETARCH" in \
+      amd64) cua_arch=x86_64; cua_sha="$CUA_DRIVER_RS_SHA256_AMD64" ;; \
+      arm64) cua_arch=arm64; cua_sha="$CUA_DRIVER_RS_SHA256_ARM64" ;; \
+      *) echo "Unsupported Cua architecture: $TARGETARCH" >&2; exit 1 ;; \
+    esac \
+  && curl -fsSL -o /tmp/cua-driver.tar.gz \
+      "https://github.com/trycua/cua/releases/download/cua-driver-rs-v${CUA_DRIVER_RS_VERSION}/cua-driver-rs-${CUA_DRIVER_RS_VERSION}-linux-${cua_arch}-binary.tar.gz" \
+  && echo "${cua_sha}  /tmp/cua-driver.tar.gz" | sha256sum -c \
+  && mkdir -p /usr/local/lib/cua-driver \
+  && tar -xzf /tmp/cua-driver.tar.gz -C /usr/local/lib/cua-driver \
+  && chmod 755 /usr/local/lib/cua-driver/cua-driver \
+  && ln -sf /usr/local/lib/cua-driver/cua-driver /usr/local/bin/cua-driver \
+  && rm -f /tmp/cua-driver.tar.gz \
+  && cua-driver --version
+
 COPY --from=controld --chmod=755 /lazyboy-controld /usr/local/bin/lazyboy-controld
 COPY --chmod=755 image/computer/rotate-logs.py /usr/local/bin/lazyboy-rotate-logs
 COPY --chmod=755 image/computer/lazyboy-screen /usr/local/bin/lazyboy-screen
@@ -140,27 +171,6 @@ COPY --chmod=644 image/computer/xfce/helper-browser.desktop /usr/share/xfce4/hel
 COPY --chmod=755 image/computer/chromium /usr/local/bin/chromium
 RUN sed -i 's|^Exec=/usr/bin/chromium|Exec=/usr/local/bin/lazyboy-browser|' /usr/share/applications/chromium.desktop || true
 
-# Pin Cua Driver outside the persisted /home/lazyboy bind-mount.
-# Checksums are from the matching official release; select the target architecture.
-ARG CUA_DRIVER_RS_VERSION=0.23.2
-ARG TARGETARCH
-ARG CUA_DRIVER_RS_SHA256_AMD64=01bf8339ec129cc00f4b4b2c6056ef1a7c5b52df39ff83ad17c9b16818aec500
-ARG CUA_DRIVER_RS_SHA256_ARM64=be22768a207796a4bc1de50c52f32f9ef680b5e86e58c059e02eec2caba2e7bb
-RUN case "$TARGETARCH" in \
-      amd64) cua_arch=x86_64; cua_sha="$CUA_DRIVER_RS_SHA256_AMD64" ;; \
-      arm64) cua_arch=arm64; cua_sha="$CUA_DRIVER_RS_SHA256_ARM64" ;; \
-      *) echo "Unsupported Cua architecture: $TARGETARCH" >&2; exit 1 ;; \
-    esac \
-  && curl -fsSL -o /tmp/cua-driver.tar.gz \
-      "https://github.com/trycua/cua/releases/download/cua-driver-rs-v${CUA_DRIVER_RS_VERSION}/cua-driver-rs-${CUA_DRIVER_RS_VERSION}-linux-${cua_arch}-binary.tar.gz" \
-  && echo "${cua_sha}  /tmp/cua-driver.tar.gz" | sha256sum -c \
-  && mkdir -p /usr/local/lib/cua-driver \
-  && tar -xzf /tmp/cua-driver.tar.gz -C /usr/local/lib/cua-driver \
-  && chmod 755 /usr/local/lib/cua-driver/cua-driver \
-  && ln -sf /usr/local/lib/cua-driver/cua-driver /usr/local/bin/cua-driver \
-  && rm -f /tmp/cua-driver.tar.gz \
-  && cua-driver --version
-
 COPY --chmod=755 image/computer/start.sh /usr/local/bin/lazyboy-computer
 COPY --chmod=755 image/computer/entrypoint.sh /usr/local/bin/lazyboy-entrypoint
 COPY --chmod=644 image/computer/cua-smoke.html /usr/share/lazyboy/cua-smoke.html
diff --git a/image/computer/fluxbox/menu b/image/computer/fluxbox/menu
deleted file mode 100644
index 5881cfb..0000000
--- a/image/computer/fluxbox/menu
+++ /dev/null
@@ -1,4 +0,0 @@
-[begin] (桌面)
-  [exec] (終端機) {/usr/local/bin/lazyboy-terminal}
-  [exec] (瀏覽器) {/usr/local/bin/lazyboy-browser https://duckduckgo.com}
-[end]
diff --git a/image/computer/fluxbox/style b/image/computer/fluxbox/style
deleted file mode 100644
index e77019f..0000000
--- a/image/computer/fluxbox/style
+++ /dev/null
@@ -1,92 +0,0 @@
-! LazyBoy Fluxbox style — flat dark, Traditional Chinese UI.
-! Debian ships Fluxbox 1.3.5 (last stable). This is a custom skin, not a fork.
-toolbar:                             flat solid
-toolbar.color:                       #0b1220
-toolbar.borderColor:                 #1e293b
-toolbar.borderWidth:                 0
-toolbar.height:                      36
-toolbar.justify:                     center
-toolbar.clock:                       flat solid
-toolbar.clock.color:                 #0b1220
-toolbar.clock.textColor:             #cbd5e1
-toolbar.clock.font:                  jf open 粉圓 2.1-11
-toolbar.workspace:                   flat solid
-toolbar.workspace.color:             #0b1220
-toolbar.workspace.textColor:         #7dd3fc
-toolbar.workspace.font:              jf open 粉圓 2.1-11
-toolbar.iconbar.empty:               flat solid
-toolbar.iconbar.empty.color:         #0b1220
-toolbar.iconbar.focused:             flat solid
-toolbar.iconbar.focused.color:       #1d4ed8
-toolbar.iconbar.focused.textColor:   #f8fafc
-toolbar.iconbar.focused.font:        jf open 粉圓 2.1-11
-toolbar.iconbar.unfocused:           flat solid
-toolbar.iconbar.unfocused.color:     #162032
-toolbar.iconbar.unfocused.textColor: #94a3b8
-toolbar.iconbar.unfocused.font:      jf open 粉圓 2.1-11
-toolbar.button:                      flat solid
-toolbar.button.color:                #0b1220
-toolbar.button.picColor:             #7dd3fc
-toolbar.button.pressed.color:        #1e3a8a
-
-menu:                                flat solid
-menu.borderColor:                    #1e293b
-menu.borderWidth:                    1
-menu.bullet:                         triangle
-menu.bullet.position:                right
-menu.title:                          flat solid
-menu.title.color:                    #1e3a8a
-menu.title.textColor:                #f8fafc
-menu.title.font:                     jf open 粉圓 2.1-12
-menu.title.justify:                  center
-menu.frame:                          flat solid
-menu.frame.color:                    #0f172a
-menu.frame.textColor:                #e2e8f0
-menu.frame.disableColor:             #64748b
-menu.frame.font:                     jf open 粉圓 2.1-12
-menu.hilight:                        flat solid
-menu.hilight.color:                  #2563eb
-menu.hilight.textColor:              #ffffff
-
-window.title.height:                 22
-window.justify:                      left
-window.borderWidth:                  1
-window.borderColor:                  #334155
-window.handleWidth:                  3
-window.bevelWidth:                   0
-window.roundCorners:                 none
-window.font:                         jf open 粉圓 2.1-11
-window.title.focus:                  flat solid
-window.title.focus.color:            #1e293b
-window.title.unfocus:                flat solid
-window.title.unfocus.color:          #0f172a
-window.label.focus:                  flat solid
-window.label.focus.color:            #1e293b
-window.label.focus.textColor:        #f8fafc
-window.label.unfocus:                flat solid
-window.label.unfocus.color:          #0f172a
-window.label.unfocus.textColor:      #64748b
-window.handle.focus:                 flat solid
-window.handle.focus.color:           #1e293b
-window.handle.unfocus:               flat solid
-window.handle.unfocus.color:         #0f172a
-window.grip.focus:                   flat solid
-window.grip.focus.color:             #2563eb
-window.grip.unfocus:                 flat solid
-window.grip.unfocus.color:           #1e293b
-window.button.focus:                 flat solid
-window.button.focus.color:           #1e293b
-window.button.focus.picColor:        #e2e8f0
-window.button.unfocus:               flat solid
-window.button.unfocus.color:         #0f172a
-window.button.unfocus.picColor:      #64748b
-window.button.pressed.color:         #2563eb
-window.close.pressed.color:          #dc2626
-
-*font:                               jf open 粉圓 2.1-11
-borderWidth:                         1
-bevelWidth:                          0
-handleWidth:                         3
-borderColor:                         #334155
-background:                          solid
-background.color:                    #0f172a
diff --git a/image/supervisor/Dockerfile b/image/supervisor/Dockerfile
index 51e1a44..6e91537 100644
--- a/image/supervisor/Dockerfile
+++ b/image/supervisor/Dockerfile
@@ -9,7 +9,9 @@ RUN --mount=type=cache,target=/usr/local/cargo/registry,sharing=locked \
     --mount=type=cache,target=/src/target,id=lazyboy-supervisor-release,sharing=locked \
     cargo build --locked --release -p lazyboy-supervisor && cp /src/target/release/lazyboy-supervisor /lazyboy-supervisor
 
-FROM gcr.io/distroless/cc-debian12:latest
+# Pin the multi-arch (amd64 + arm64) index by digest: `latest` is mutable and a
+# floating base silently changes the supervisor's libc on every rebuild.
+FROM gcr.io/distroless/cc-debian12@sha256:e5d81ddde149641e2a9ba55be4545bc125c67de07508b03ba4c22e6eb0ded5aa
 COPY --from=build /lazyboy-supervisor /usr/local/bin/lazyboy-supervisor
 EXPOSE 7091
 CMD ["/usr/local/bin/lazyboy-supervisor"]
diff --git a/scripts/build-computer-image.sh b/scripts/build-computer-image.sh
deleted file mode 100755
index 66cfd8b..0000000
--- a/scripts/build-computer-image.sh
+++ /dev/null
@@ -1,6 +0,0 @@
-#!/usr/bin/env bash
-set -euo pipefail
-root="$(cd "$(dirname "$0")/.." && pwd)"
-cd "$root"
-docker build -f image/computer/Dockerfile -t lazyboy/computer:local .
-echo "built lazyboy/computer:local"
diff --git a/scripts/build-image.sh b/scripts/build-image.sh
new file mode 100755
index 0000000..c9d6305
--- /dev/null
+++ b/scripts/build-image.sh
@@ -0,0 +1,136 @@
+#!/usr/bin/env bash
+# Build one LazyBoy image. Single entry point for everything in image/.
+#
+#   scripts/build-image.sh                                   # desktop image, host arch, into docker
+#   scripts/build-image.sh --file image/api/Dockerfile       # any image, host arch
+#   scripts/build-image.sh --platforms linux/arm64           # one foreign arch, needs QEMU
+#   scripts/build-image.sh --multi                           # amd64 + arm64 -> OCI archive
+#   scripts/build-image.sh --multi --push                    # amd64 + arm64 -> registry
+#
+# Supported arches: linux/amd64 and linux/arm64 -- see the ceiling note below.
+# --multi/--platforms cannot use the default "docker" builder: that driver
+# cannot export a manifest list. A docker-container builder plus QEMU binfmt is
+# bootstrapped here.
+set -euo pipefail
+root="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$root"
+
+file="image/computer/Dockerfile"
+platforms=""
+multi=0
+push=0
+target=""
+oci_name=""
+builder="${BUILDX_BUILDER:-lazyboy}"
+tag=""
+
+while [ $# -gt 0 ]; do
+  case "$1" in
+    --file) file="$2"; shift 2 ;;
+    --tag) tag="$2"; shift 2 ;;
+    --target) target="$2"; shift 2 ;;
+    --platforms) platforms="$2"; shift 2 ;;
+    --oci) oci_name="$2"; shift 2 ;;
+    --builder) builder="$2"; shift 2 ;;
+    --multi) multi=1; shift ;;
+    --push) push=1; shift ;;
+    -h|--help) sed -n '2,13p' "$0"; exit 0 ;;
+    *) echo "unknown option: $1" >&2; exit 2 ;;
+  esac
+done
+
+# Defaults follow docker-compose.yml: the desktop image is published under a
+# lazyboy/ namespace, the services under the compose project name.
+if [ -z "$tag" ]; then
+  case "$file" in
+    image/computer/Dockerfile) tag="lazyboy/computer:local" ;;
+    image/supervisor/Dockerfile) tag="lazyboy-supervisor:latest" ;;
+    image/api/Dockerfile) tag="lazyboy-api:latest" ;;
+    *) echo "--tag is required for $file" >&2; exit 2 ;;
+  esac
+fi
+if [ "$multi" = 1 ] && [ -z "$platforms" ]; then
+  platforms="linux/amd64,linux/arm64"
+fi
+if [ -z "$oci_name" ]; then
+  case "$file" in
+    image/computer/Dockerfile) oci_name="dist/lazyboy-computer-multi.oci.tar" ;;
+    *) oci_name="dist/$(basename "$(dirname "$file")")-multi.oci.tar" ;;
+  esac
+fi
+
+if [ "$push" = 1 ] && [ -z "$platforms" ]; then
+  echo "--push only makes sense with --multi or --platforms" >&2
+  exit 2
+fi
+
+# Architecture ceiling, set by two upstream binary distributions: the Cua Driver
+# (cua-driver-rs-v0.23.2) ships Linux builds for x86_64 and arm64 only, and
+# ONNX Runtime (v1.24.1) for x64 and aarch64 only. image/computer/Dockerfile and
+# scripts/fetch-onnxruntime.sh fail fast on anything else; checking here rejects
+# a bad --platforms before binfmt + GBs of cross-compilation are spent.
+supported_arches="amd64 arm64"
+for platform in ${platforms//,/ }; do
+  os="${platform%%/*}"
+  arch="${platform#*/}"
+  arch="${arch%%/*}"
+  if [ "$os" != "linux" ]; then
+    echo "--platforms: only Linux images are built (got $platform)" >&2
+    exit 2
+  fi
+  case " $supported_arches " in
+    *" $arch "*) ;;
+    *)
+      echo "--platforms: unsupported architecture '$arch' in '$platform'." >&2
+      echo "  LazyBoy images exist for linux/amd64 and linux/arm64: the Cua" >&2
+      echo "  Driver ships only linux-x86_64/linux-arm64 and ONNX Runtime only" >&2
+      echo "  linux-x64/linux-aarch64, so no other Linux arch can run them." >&2
+      exit 2
+      ;;
+  esac
+done
+
+if [ -z "$platforms" ]; then
+  # Native build: the docker driver, so the result lands in the local store and
+  # `docker compose up` picks it up without a registry.
+  args=(docker build -f "$file" -t "$tag")
+  [ -n "$target" ] && args+=(--target "$target")
+  "${args[@]}" .
+  echo "built $tag"
+  exit 0
+fi
+
+docker buildx inspect "$builder" >/dev/null 2>&1 \
+  || docker buildx create --name "$builder" --driver docker-container --bootstrap
+# QEMU is only needed for architectures the host cannot run natively; asking
+# binfmt to register the host's own handler fails inside a privileged container
+# and would look like a real failure, so the native arch is skipped.
+native="$(dpkg --print-architecture 2>/dev/null || true)"
+if [ -z "$native" ]; then
+  case "$(uname -m)" in
+    x86_64) native=amd64 ;;
+    aarch64) native=arm64 ;;
+    *) native=unknown ;;
+  esac
+fi
+for platform in ${platforms//,/ }; do
+  arch="${platform#*/}"
+  if [ "$arch" != "$native" ]; then
+    docker run --privileged --rm tonistiigi/binfmt --install "$arch" >/dev/null 2>&1 \
+      || echo "WARNING: no QEMU handler for $arch; that build fails with 'exec format error'" >&2
+  fi
+done
+args=(docker buildx build --builder "$builder" -f "$file" -t "$tag" --platform "$platforms")
+[ -n "$target" ] && args+=(--target "$target")
+if [ "$push" = 1 ]; then
+  args+=(--push)
+else
+  mkdir -p "$(dirname "$oci_name")"
+  args+=(--output "type=oci,dest=$oci_name")
+fi
+"${args[@]}" .
+if [ "$push" = 1 ]; then
+  echo "published $tag for ${platforms}"
+else
+  echo "wrote $oci_name for ${platforms} (load with: nerdctl load -i $oci_name)"
+fi
diff --git a/scripts/cua-smoke-test.sh b/scripts/cua-smoke-test.sh
index 3cadce4..16983f5 100755
--- a/scripts/cua-smoke-test.sh
+++ b/scripts/cua-smoke-test.sh
@@ -100,6 +100,18 @@ if [[ "$code" -eq 0 ]]; then
   code=$?
 fi
 if [[ "$code" -eq 0 ]]; then
+  # Chromium writes its cookie database on a ~30 s timer and does not flush on
+  # SIGTERM, so restarting right away races that timer: the check below would
+  # then measure flush timing instead of profile persistence. Wait until the
+  # fixture cookie is on disk before restarting.
+  echo "waiting for the browser to persist its cookie"
+  for _ in $(seq 1 90); do
+    if docker exec -u 1000:1000 "$name" sh -c \
+      'grep -qa cuaSmoke "$HOME"/.browser-profiles/displays/*/Default/Cookies 2>/dev/null'; then
+      break
+    fi
+    sleep 1
+  done
   docker restart "$name" >/dev/null
   code=$?
 fi
diff --git a/tests/frontend.test.mjs b/tests/frontend.test.mjs
index c7a6adb..5cccc62 100644
--- a/tests/frontend.test.mjs
+++ b/tests/frontend.test.mjs
@@ -4,7 +4,13 @@ import fs from 'node:fs';
 import vm from 'node:vm';
 import ts from '../apps/web/node_modules/typescript/lib/typescript.js';
 function loadCatalog(file){
-  const code=ts.transpileModule(fs.readFileSync(file,'utf8'),{compilerOptions:{module:ts.ModuleKind.CommonJS}}).outputText;
+  const source=fs.readFileSync(file,'utf8');
+  // transpileModule prints output even for a catalog that does not parse, so a
+  // missing comma silently loads a shorter catalog and every key check below
+  // still passes. Ask the parser, which actually keeps its errors.
+  const parsed=ts.createSourceFile(file,source,ts.ScriptTarget.ESNext,true,ts.ScriptKind.TS);
+  assert.deepEqual((parsed.parseDiagnostics||[]).map(d=>ts.flattenDiagnosticMessageText(d.messageText,' ')),[],`${file} must parse`);
+  const code=ts.transpileModule(source,{compilerOptions:{module:ts.ModuleKind.CommonJS}}).outputText;
   const sandbox={exports:{},require:()=>({})};
   vm.runInNewContext(code,sandbox);
   return sandbox.exports.zhTW||sandbox.exports.en;
@@ -274,17 +280,15 @@ test('mobile shortcut row sends modifiers and hides with the keyboard',()=>{
  assert.equal(v.keyboard.focused,false);
  assert.equal(v.elements['keyboard-shortcuts'].hidden,true);
 });
-test('status and HUD sit outside the remote pixels; screenshots stay opt-in',()=>{
- const app=fs.readFileSync('apps/web/src/App.tsx','utf8');
- assert.doesNotMatch(app,/RunStatus/);
+test('status sits outside the remote pixels; screenshots stay opt-in',()=>{
+  const app=fs.readFileSync('apps/web/src/App.tsx','utf8');
+  assert.doesNotMatch(app,/RunStatus/);
  assert.match(app,/<RunProbe runId=\{computer\.busyRunId\|\|computer\.waitingRunId\}><Avatar/);
- assert.match(app,/\{!computerOpen&&hud\}<div className="preview">/);
- assert.match(app,/--visible-height/);
+  assert.match(app,/--visible-height/);
  assert.match(app,/<div className=\{`composer-dock \$\{statusMembers\.length\?"has-status":""\}`\}>\s*\{error&&<div className="error-banner"/);
  const css=fs.readFileSync('apps/web/src/computer.css','utf8');
  assert.doesNotMatch(css,/\.run-status/);
- assert.match(css,/\.computer-hud\{position:static/);
- const chat=fs.readFileSync('apps/web/src/chat.css','utf8');
+  const chat=fs.readFileSync('apps/web/src/chat.css','utf8');
  assert.match(chat,/\.composer-dock\{position:relative/);
  assert.doesNotMatch(chat,/\.composer-dock\{position:absolute/);
  assert.match(chat,/\.messages\{min-width:0;padding-bottom:24px/);
@@ -447,3 +451,162 @@ test('chat follows the event stream instead of a fixed two second poll',()=>{
   assert.doesNotMatch(app,/const timer=setInterval\(\(\)=>\{refresh\(\)/,'the 2s transcript poll should be gone');
   assert.match(app,/const heartbeat=window\.setInterval\(\(\)=>\{const beat=roomsRef\.current[\s\S]*\},HEARTBEAT_MS\)/,'the heartbeat keeps its own minute cadence');
 });
+
+// The screen veil and the frame that survives it: both used to be driven by a
+// timer plus "null the url", which is what made booting and handing over feel
+// like a stall instead of a gesture.
+const handoffJs=ts.transpileModule(fs.readFileSync('apps/web/src/handoff.ts','utf8'),{compilerOptions:{module:ts.ModuleKind.CommonJS}}).outputText;
+const handoffBox={exports:{},require:()=>({})};vm.runInNewContext(handoffJs,handoffBox);
+const {viewerPath,keepScreenUrl,handoffRemaining,nextVeil,viewOnlyFor,HANDOFF_MS,HANDOFF_MIN_MS,VEIL_FADE_MS}=handoffBox.exports;
+
+test('the viewer path is stable, so a frame can mount before the desktop answers',()=>{
+  assert.equal(viewerPath('bot 7'),'/view/bot%207/vnc.html');
+  assert.equal(viewerPath('a/b'),'/view/a%2Fb/vnc.html');
+  assert.equal(viewerPath('bot'),viewerPath('bot'),'no nonce, or the frame would remount every poll');
+});
+
+test('the path the frame mounts on is the byte the api hands back',()=>{
+  // The viewer mounts on the client's own path while the desktop is coming up,
+  // then adopts the url from the status round trip. If the two ever differ the
+  // iframe remounts mid-boot, which is the black flash this whole path exists
+  // to avoid, so the invariant is pinned to the server's own format string.
+  const routes=fs.readFileSync('crates/api/src/routes.rs','utf8');
+  const served=(routes.match(/"url": format!\("([^"]+)"\)/)||[])[1];
+  assert.equal(served,'/view/{id}/vnc.html');
+  const id='6f1d3a0e-2b7c-4d5e-8f90-1a2b3c4d5e6f';
+  assert.equal(viewerPath(id),served.replace('{id}',id));
+});
+
+test('pixels are dropped only for a computer that is really gone',()=>{
+  const url='/view/bot/vnc.html';
+  for(const state of ['booting','suspended','running'])assert.equal(keepScreenUrl(url,null,state),url,state+' must keep its VNC session');
+  for(const state of ['stopped','error'])assert.equal(keepScreenUrl(url,null,state),null,state+' has no desktop to keep');
+  assert.equal(keepScreenUrl(null,null,'running'),null,'nothing was ever mounted');
+  assert.equal(keepScreenUrl(url,'/other','booting'),'/other','a fresh url still wins');
+  assert.equal(keepScreenUrl(null,url,'booting'),url,'the first url arrives while booting');
+});
+
+test('a handoff holds one visible beat, never the whole round trip',()=>{
+  assert.ok(HANDOFF_MIN_MS<HANDOFF_MS,'the beat needs a floor and a ceiling');
+  assert.ok(VEIL_FADE_MS<=HANDOFF_MIN_MS,'the fade has to finish inside the beat');
+  assert.ok(HANDOFF_MS<=1_000,'a reply that never lands cannot strand the mascot');
+  assert.equal(handoffRemaining(1_000,1_000),HANDOFF_MIN_MS,'a reply this instant still owes the beat');
+  assert.equal(handoffRemaining(1_000,1_000+HANDOFF_MIN_MS),0,'served');
+  assert.equal(handoffRemaining(1_000,1_000+(HANDOFF_MIN_MS+HANDOFF_MS)/2),0,'past the floor is never a wait');
+  assert.equal(handoffRemaining(1_000,1_000+HANDOFF_MS),0);
+  const midway=handoffRemaining(1_000,1_000+HANDOFF_MIN_MS-10);
+  assert.equal(midway,10,'the remaining beat shrinks with the clock');
+});
+
+test('the veil fades in with its label and back out through the same mascot',()=>{
+  const empty={label:null,leaving:false};
+  assert.equal(nextVeil(null,empty),empty,'with nothing to say there is nothing to show');
+  const shown=nextVeil('換手中',empty);
+  assert.equal(shown.label,'換手中');assert.equal(shown.leaving,false);
+  assert.equal(nextVeil('換手中',shown),shown,'a repeated status must not restart the animation');
+  const leaving=nextVeil(null,shown);
+  assert.equal(leaving.label,'換手中','the label is held while it fades');assert.equal(leaving.leaving,true);
+  assert.equal(nextVeil(null,leaving),leaving,'it leaves exactly once');
+  const replaced=nextVeil('喚醒中',shown);
+  assert.equal(replaced.label,'喚醒中');assert.equal(replaced.leaving,false,'a new status replaces the label');
+  const revived=nextVeil('喚醒中',leaving);
+  assert.equal(revived.label,'喚醒中');assert.equal(revived.leaving,false,'and pulls it back out of the fade');
+});
+
+test('only a human moves the mouse, and the veil never holds it back',()=>{
+  assert.equal(viewOnlyFor('user'),false);
+  for(const holder of ['none','bot'])assert.equal(viewOnlyFor(holder),true,holder);
+});
+
+test('taking the screen flips the mouse on the click, not on the reply',()=>{
+  const app=fs.readFileSync('apps/web/src/App.tsx','utf8');
+  const body=app.slice(app.indexOf('async function setControl('));
+  const flip=body.indexOf('controlHolder:holder');
+  const gate=body.indexOf('pushViewOnly(viewOnlyFor(holder))');
+  const post=body.indexOf('await api(`/api/computer/${botId}/${holder==="user"?"takeover":"release"}`');
+  const readBack=body.indexOf('finally{');
+  assert.ok(flip>0&&gate>0&&post>0&&readBack>0,'optimistic flip, gate, request, and read back all present');
+  assert.ok(flip<post&&gate<post,'badge and input move before the server answers');
+  assert.ok(readBack>post,'the truth is read back after the request');
+  assert.match(body.slice(readBack),/await refresh\(\)/);
+  assert.match(body,/if\(!botId\|\|controlBusyRef\.current\)return/,'a double click cannot fight itself');
+  assert.match(app,/const listener=\(event:MessageEvent\)=>\{[\s\S]*?lazyboy-request-control[\s\S]*?void setControl\("user"\)/);
+  assert.match(app,/if\(expectedHolderRef\.current&&status\.controlHolder===expectedHolderRef\.current\)/,'agreement, not a timer, ends the handoff');
+  assert.match(app,/onClick=\{onTakeOver\}/);
+  assert.match(app,/onClick=\{onRelease\}/);
+  assert.doesNotMatch(app,/action\(\(\)=>api\(`\/api\/computer\/[^`]*takeover/,'takeover no longer rides the global busy path');
+  assert.equal((app.match(/\/api\/computer\/\$\{[^}]*\}\/(takeover|release)/g)||[]).length,0,'every handoff goes through setControl');
+  assert.equal((app.match(/holder==="user"\?"takeover":"release"/g)||[]).length,1,'one request path, one owner of it');
+  assert.match(app,/void setControl\("user",active\.id\)/,'the call overlay hands over the same way');
+  assert.match(app,/await setControl\("user",id\)/,'the login screen boots, then takes the mouse without a second spinner');
+});
+
+test('the waiting veil covers the desktop the way it always did',()=>{
+  const app=fs.readFileSync('apps/web/src/App.tsx','utf8');
+  assert.match(app,/<div className="preview">\{computerOpen\?<EmptyComputer state=\{computer\.state\}\/>:frame\}\{!computerOpen&&hud\}<\/div>/);
+  assert.match(app,/<div className="overlay-desktop">\{frame\}\{hud\}<\/div>/);
+  assert.match(app,/const hud=paneBot&&veil\.label\?<ComputerHud bot=\{paneBot\} label=\{veil\.label\} leaving=\{veil\.leaving\}\/>:null/);
+  const css=fs.readFileSync('apps/web/src/computer.css','utf8');
+  // Both parents are what make inset:0 cover the desktop rather than the page.
+  assert.match(css,/\.computer-part \.preview\{position:relative/);
+  assert.match(css,/\.overlay-screen \.overlay-desktop\{position:relative/);
+  // These values were compared against main in a real browser: rendered pixel
+  // for pixel identical, so drift here means the veil visibly changed.
+  const hud=css.match(/^\.computer-hud\{([^}]*)\}/m)[1];
+  for(const rule of ['position:absolute','inset:0','z-index:3','display:grid','align-content:center','justify-items:center','gap:14px','padding:10px 12px 8px','border-radius:inherit','background:radial-gradient(ellipse at center,#142722e8,#101012ed)','backdrop-filter:blur(8px)','pointer-events:none'])assert.ok(hud.includes(rule),`the veil needs ${rule}`);
+  assert.match(css,/\.computer-hud\.is-leaving\{opacity:0\}/);
+  assert.match(css,/\.computer-hud\{[^}]*transition:opacity \.22s ease\}/,'it fades instead of blinking off');
+  // main drew the label as static mint text; the shimmer belongs to the chat
+  // working label, and the veil keeps the mascot as the only moving part.
+  const label=css.match(/^\.computer-hud-label\{([^}]*)\}/m)[1];
+  assert.ok(!label.includes('animation'),'the veil label does not animate');
+  assert.ok(label.includes('color:#c9ddd5')&&label.includes('font-size:12px')&&label.includes('text-align:center')&&label.includes('max-width:90%'));
+  assert.ok(!/\.computer-hud[^{]*\{[^}]*working-shimmer/.test(css),'nothing shimmers inside the veil');
+  assert.match(css,/\.computer-signal\{[^}]*width:64px[^}]*animation:monitor-breathe/);
+  assert.match(css,/\.computer-signal::before\{[^}]*monitor-orbit/);
+  const responsive=fs.readFileSync('apps/web/src/responsive.css','utf8');
+  assert.match(responsive,/prefers-reduced-motion:reduce\)\{\.computer-signal,\.computer-signal::before,\.computer-signal-face i[^}]*animation:none/,'reduced motion still has to still the mascot');
+});
+
+test('a desktop that is still starting is dialled again fast, a dropped session calmly',()=>{
+  const source=fs.readFileSync('apps/web/vnc.html','utf8').match(/<script type="module">([\s\S]*?)<\/script>/)[1].replace(/import RFB[^;]+;/,'');
+  // Any element the viewer touches answers, so the disconnect path runs the
+  // same way it does in a browser without a DOM to stand in.
+  const element=()=>new Proxy({style:{},classList:{toggle(){}},dataset:{},hidden:false,textContent:''},{get:(node,key)=>key in node?node[key]:()=>undefined,set:(node,key,value)=>{node[key]=value;return true}});
+  const instances=[],scheduled=[];
+  class RFB{constructor(){this._handlers={};instances.push(this._handlers);}
+    addEventListener(name,handler){(this._handlers[name]||(this._handlers[name]=[])).push(handler);} focus(){} blur(){} sendKey(){}}
+  const window={location:{pathname:'/vnc.html',protocol:'http:',host:'localhost',origin:'http://localhost',hash:''},parent:{postMessage(){}},addEventListener(){}};
+  vm.runInNewContext(source,{window,document:{location:{href:'http://localhost/vnc.html?view_only=false'},getElementById:()=>element(),querySelector:()=>null},navigator:{clipboard:{}},RFB,setTimeout:(fn,ms)=>{scheduled.push(ms);return scheduled.length},clearTimeout(){}});
+  const fire=(name,event)=>{for(const handlers of instances)(handlers[name]||[]).forEach(handler=>handler(event||{}));};
+  assert.equal(instances.length,1,'the viewer dials once on load');
+  fire('disconnect',{clean:false});
+  assert.equal(scheduled.at(-1),350,'a desktop that has never answered is retried eagerly');
+  fire('connect');
+  fire('disconnect',{clean:true});
+  assert.equal(scheduled.at(-1),1500,'a session that was really there is retried calmly');
+});
+
+test('a session dialed after a handoff starts with the mouse already handed over',()=>{
+  const vnc=fs.readFileSync('apps/web/vnc.html','utf8');
+  // Run the viewer rather than read it. The host hands the mouse over while the
+  // viewer sits between two sessions: exactly the moment a fresh RFB used to
+  // fall back to its view-only default and leave a taken-over desktop deaf with
+  // the veil already lifted. Only the gate message arriving is real here.
+  const body=vnc.match(/<script type="module">([\s\S]*?)<\/script>/)[1].replace(/import RFB[^;]+;/,'');
+  const element=()=>new Proxy({style:{},classList:{toggle(){}},dataset:{},hidden:false,textContent:''},{get:(node,key)=>key in node?node[key]:()=>undefined,set:(node,key,value)=>{node[key]=value;return true}});
+  const instances=[],scheduled=[],messages=[];
+  class RFB{constructor(){this._handlers={};instances.push(this);}addEventListener(name,handler){(this._handlers[name]||(this._handlers[name]=[])).push(handler);}focus(){this.focused=true;}blur(){}}
+  const parent={postMessage(){}};
+  const window={location:{pathname:'/vnc.html',protocol:'http:',host:'localhost',origin:'http://localhost',hash:''},parent,addEventListener(name,handler){if(name==='message')messages.push(handler);}};
+  vm.runInNewContext(body,{window,document:{location:{href:'http://localhost/vnc.html'},getElementById:()=>element(),querySelector:()=>null},navigator:{clipboard:{}},RFB,setTimeout:(fn)=>{scheduled.push(fn);return scheduled.length},clearTimeout(){}});
+  const fire=(name,event)=>instances.forEach(rfb=>(rfb._handlers[name]||[]).forEach(handler=>handler(event||{})));
+  fire('disconnect',{clean:false});
+  assert.ok(messages.length,'the viewer listens for the host gate');
+  messages.forEach(handler=>handler({origin:'http://localhost',source:parent,data:{type:'lazyboy-view-only',viewOnly:false}}));
+  scheduled.at(-1)();
+  fire('connect');
+  assert.equal(instances.length,2,'the retry dialed a new session');
+  assert.equal(instances.at(-1).viewOnly,false,'a handoff made mid-reconnect still owns the new session');
+  assert.equal(instances.at(-1).focused,true,'taking over lands the first keystroke');
+});