Feature: Multi-device screen wall on the Go server

Bring the Go web backend up to parity with the MFC server's screen wall.
A browser tab can now stream several devices over one WebSocket: the
single `watching` field becomes a per-connection watch set guarded by the
hub mutex, capped at 32 devices per socket. The challenge advertises
`caps.multiWatch` and `caps.keyframeRequest`, binary frames stay keyed by
the low 32 bits of the device ID, and connect/disconnect replies echo the
device id and request id so the wall can route concurrent responses.

Input commands (mouse, key, clipboard, rdp_reset) resolve their target
from the watch set: an explicit id must be a current subscription, and an
absent id is accepted only while exactly one device is watched. Legacy
frontends that do not send `multiWatch` keep replace-all semantics, so a
repeated device switch cannot grow the set without bound. Offline
devices are removed from every watch set, and teardown is computed from a
snapshot taken outside the hub lock to avoid re-entrant locking.

Add the `keyframe` command: it replays the cached IDR, then asks the
device to emit a fresh one, rate-limited to one request per device per
second. This exposed a Go-only corruption bug: the client answers the
request with TOKEN_KEYFRAME (134), which the dispatcher treated as a
no-op. That token only appears on forced IDRs because the client's GOP is
effectively infinite, so the dropped IDR left the browser decoding
P-frames against a reference chain it never received, producing
persistent garbling. Relay it as a keyframe instead. Also drop the
cached IDR when the resolution changes, mirroring the C++ server, so a
late joiner cannot receive an old-geometry keyframe.

Add tests for the watch-set transitions and for relaying TOKEN_KEYFRAME.

Co-Authored-By: deepseek-flash

server/go/cmd/main.go

@@ -138,11 +138,16 @@ func (h *MyHandler) OnReceive(ctx *connection.Context, data []byte) {
 	case protocol.TokenClipboardText:
 		h.handleClipboardText(ctx, data)
 	case protocol.TokenKeyframe:
-		// Sent by the client only when frameID % m_GOP == 0; the client's
+		// The client switches to this token when it emits an IDR on request
-		// DEFAULT_GOP is 0x7FFFFFFF (effectively infinite), so this token
+		// (COMMAND_SCREEN_KEYFRAME_REQ) or on a GOP boundary. The header
-		// is essentially unused in practice. Treat as a no-op for now —
+		// layout is identical to TOKEN_NEXTSCREEN, but the payload is always
-		// IDRs always arrive in-band via TOKEN_NEXTSCREEN and we catch
+		// an IDR, so relay it as frameType=1 without sniffing. Dropping it
-		// them via the H264 NAL scan in handleScreenFrame.
+		// (the old behaviour, harmless while the client's DEFAULT_GOP of
+		// 0x7FFFFFFF meant this token never appeared) breaks keyframe
+		// recovery: the browser keeps decoding P-frames against a reference
+		// chain it never received, so the picture stays corrupt until the
+		// screen session restarts.
+		h.handleScreenFrame(ctx, data, true)
 	case protocol.CmdCursorImage:
 		// Custom cursor bitmaps — relayed in Phase 5+ when the web cursor
 		// overlay learns to render arbitrary BGRA images. Drop silently for

server/go/cmd/screen_keyframe_test.go

@@ -0,0 +1,72 @@
+package main
+
+import (
+	"testing"
+
+	"github.com/yuanyuanxiang/SimpleRemoter/server/go/connection"
+	"github.com/yuanyuanxiang/SimpleRemoter/server/go/hub"
+	"github.com/yuanyuanxiang/SimpleRemoter/server/go/protocol"
+)
+
+type frameRecorder struct {
+	frames []recordedFrame
+}
+
+type recordedFrame struct {
+	deviceID string
+	packet   []byte
+	isKey    bool
+}
+
+func (r *frameRecorder) OnScreenFrame(deviceID string, packet []byte, isKeyframe bool) {
+	r.frames = append(r.frames, recordedFrame{deviceID, packet, isKeyframe})
+}
+func (r *frameRecorder) OnDeviceOnline(hub.DeviceInfo)       {}
+func (r *frameRecorder) OnDeviceOffline(string)              {}
+func (r *frameRecorder) OnDeviceUpdate(string, int, string)  {}
+func (r *frameRecorder) OnResolutionChange(string, int, int) {}
+func (r *frameRecorder) OnCursorChange(string, byte)         {}
+func (r *frameRecorder) OnClipboardText(string, string)      {}
+func (r *frameRecorder) OnTerminalReady(string, bool)        {}
+func (r *frameRecorder) OnTerminalData(string, []byte)       {}
+func (r *frameRecorder) OnTerminalClosed(string, string)     {}
+
+// TestOnReceiveRelaysKeyframeToken guards the forced-IDR recovery path: when
+// the browser asks for a keyframe, the client replies with TOKEN_KEYFRAME.
+// Dropping that token (the old no-op) leaves the browser decoding P-frames
+// against a reference chain it never received, i.e. persistent corruption.
+func TestOnReceiveRelaysKeyframeToken(t *testing.T) {
+	h := hub.New()
+	rec := &frameRecorder{}
+	h.Subscribe(rec)
+
+	h.Register(&hub.Device{ID: "dev", PeerIP: "10.0.0.1"},
+		&connection.Context{RemoteAddr: "10.0.0.1:1000"})
+	screenCtx := &connection.Context{RemoteAddr: "10.0.0.1:1001"}
+	if !h.BindScreenConn("dev", screenCtx) {
+		t.Fatal("screen bind failed")
+	}
+
+	// [TOKEN_KEYFRAME][algo=H264][cursorPos:8][cursorIdx:1][IDR NAL]
+	data := []byte{
+		protocol.TokenKeyframe, 2,
+		0, 0, 0, 0, 0, 0, 0, 0,
+		0,
+		0x00, 0x00, 0x00, 0x01, 0x65, 0x88, 0x84,
+	}
+	(&MyHandler{hub: h}).OnReceive(screenCtx, data)
+
+	if len(rec.frames) != 1 {
+		t.Fatalf("frames relayed = %d, want 1 (TOKEN_KEYFRAME must not be dropped)", len(rec.frames))
+	}
+	f := rec.frames[0]
+	if f.deviceID != "dev" {
+		t.Fatalf("deviceID = %q, want dev", f.deviceID)
+	}
+	if !f.isKey || f.packet[4] != 1 {
+		t.Fatalf("isKeyframe=%v frameType=%d, want true/1", f.isKey, f.packet[4])
+	}
+	if want := 9 + len(data) - (1 + protocol.ScreenFrameHeaderLen); len(f.packet) != want {
+		t.Fatalf("packet len = %d, want %d", len(f.packet), want)
+	}
+}

server/go/hub/hub.go

@@ -581,6 +581,13 @@ func (h *Hub) PublishResolution(deviceID string, width, height int) {
 	if d, ok := h.devices[deviceID]; ok {
 		d.screenWidth = width
 		d.screenHeight = height
+		// The cached IDR belongs to the previous capture setup. The client
+		// re-sends TOKEN_BITMAPINFO on the same sub-conn after a window
+		// resize / capture restart, so this is not only the
+		// new-sub-conn path (BindScreenConn already clears there). Serving
+		// the old-resolution IDR alongside the new dimensions leaves the
+		// decoder misconfigured and garbles the picture until the next IDR.
+		d.lastKeyframe = nil
 	}
 	h.mu.Unlock()
 	for _, s := range h.snapshotSubscribers() {

server/go/protocol/commands.go

@@ -112,6 +112,7 @@ const (
 	CommandShell              byte = 40  // COMMAND_SHELL - ask device to open a shell sub-connection
 	CommandTerminalRsize      byte = 81  // CMD_TERMINAL_RESIZE - [cmd:1][cols:2 LE][rows:2 LE]
 	CmdRestoreConsole         byte = 82  // CMD_RESTORE_CONSOLE - RDP session "归位": switch back to the console session and restart capture
+	CommandScreenKeyframeReq  byte = 169 // COMMAND_SCREEN_KEYFRAME_REQ - force the client to emit an IDR immediately
 	CommandBye                byte = 204 // COMMAND_BYE - disconnect
 	CommandHeartbeat          byte = 216 // CMD_HEARTBEAT_ACK
 

server/go/web/watch_set_test.go

@@ -0,0 +1,96 @@
+package web
+
+import (
+	"fmt"
+	"reflect"
+	"sort"
+	"testing"
+)
+
+func sortedWatchIDs(c *wsClient) []string {
+	out := make([]string, 0, len(c.watchIDs))
+	for id := range c.watchIDs {
+		out = append(out, id)
+	}
+	sort.Strings(out)
+	return out
+}
+
+func TestApplyConnectWatchMulti(t *testing.T) {
+	c := &wsClient{watchIDs: make(map[string]struct{})}
+
+	if stopped, tooMany := applyConnectWatch(c, "a", true); len(stopped) != 0 || tooMany {
+		t.Fatalf("first connect = (%v, %v), want ([], false)", stopped, tooMany)
+	}
+	if stopped, tooMany := applyConnectWatch(c, "b", true); len(stopped) != 0 || tooMany {
+		t.Fatalf("second connect = (%v, %v), want ([], false)", stopped, tooMany)
+	}
+	if got, want := sortedWatchIDs(c), []string{"a", "b"}; !reflect.DeepEqual(got, want) {
+		t.Fatalf("watch set = %v, want %v", got, want)
+	}
+
+	// Repeated connect (focus switch) must be a no-op.
+	if stopped, tooMany := applyConnectWatch(c, "a", true); len(stopped) != 0 || tooMany {
+		t.Fatalf("repeat connect = (%v, %v), want ([], false)", stopped, tooMany)
+	}
+	if got, want := sortedWatchIDs(c), []string{"a", "b"}; !reflect.DeepEqual(got, want) {
+		t.Fatalf("watch set after repeat = %v, want %v", got, want)
+	}
+}
+
+func TestApplyConnectWatchLegacyReplaces(t *testing.T) {
+	c := &wsClient{watchIDs: map[string]struct{}{"a": {}, "b": {}}}
+
+	stopped, tooMany := applyConnectWatch(c, "c", false)
+	if tooMany {
+		t.Fatal("legacy connect hit the cap, want not")
+	}
+	sort.Strings(stopped)
+	if want := []string{"a", "b"}; !reflect.DeepEqual(stopped, want) {
+		t.Fatalf("stopped = %v, want %v", stopped, want)
+	}
+	if got, want := sortedWatchIDs(c), []string{"c"}; !reflect.DeepEqual(got, want) {
+		t.Fatalf("watch set = %v, want %v", got, want)
+	}
+}
+
+func TestApplyConnectWatchCap(t *testing.T) {
+	c := &wsClient{watchIDs: make(map[string]struct{})}
+	for i := 0; i < maxWatchPerClient; i++ {
+		applyConnectWatch(c, fmt.Sprintf("d%02d", i), true)
+	}
+	if _, tooMany := applyConnectWatch(c, "overflow", true); !tooMany {
+		t.Fatal("expected the cap to reject a new device")
+	}
+	if len(c.watchIDs) != maxWatchPerClient {
+		t.Fatalf("watch set size = %d, want %d", len(c.watchIDs), maxWatchPerClient)
+	}
+	// An already watched device must still be accepted at the cap.
+	if _, tooMany := applyConnectWatch(c, "d00", true); tooMany {
+		t.Fatal("repeat connect at the cap was rejected")
+	}
+}
+
+func TestResolveWatchDevice(t *testing.T) {
+	h := &wsHub{}
+	c := &wsClient{watchIDs: map[string]struct{}{"a": {}}}
+
+	if got := h.resolveWatchDevice(c, []byte(`{"id":"a"}`)); got != "a" {
+		t.Fatalf("explicit watched id = %q, want a", got)
+	}
+	if got := h.resolveWatchDevice(c, []byte(`{"id":"nope"}`)); got != "" {
+		t.Fatalf("explicit unwatched id = %q, want empty", got)
+	}
+	// Old frontend: no id, exactly one watched device.
+	if got := h.resolveWatchDevice(c, []byte(`{}`)); got != "a" {
+		t.Fatalf("no id, single device = %q, want a", got)
+	}
+
+	c.watchIDs["b"] = struct{}{}
+	if got := h.resolveWatchDevice(c, []byte(`{}`)); got != "" {
+		t.Fatalf("no id, two devices = %q, want empty", got)
+	}
+	if got := h.resolveWatchDevice(c, []byte(`{"id":"a"}`)); got != "a" {
+		t.Fatalf("explicit id, two devices = %q, want a", got)
+	}
+}

server/go/web/ws.go

@@ -52,8 +52,16 @@ type wsClient struct {
 	token        string // set once authenticated
 	role         string // mirrors session role after login
 	addr         string // client address for logs
-	watching     string // device ID this browser is currently streaming, "" when on the list
 	termWatching string // device ID for an open web terminal session, "" otherwise
+
+	// watchIDs is the set of devices this browser is streaming (the screen
+	// wall watches several at once). Empty while on the device list.
+	// Guarded by wsHub.mu.
+	watchIDs map[string]struct{}
+
+	// keyframeAt throttles keyframe requests to one per device per second.
+	// Unix seconds, guarded by wsHub.mu.
+	keyframeAt map[string]int64
 }
 
 // queue writes a JSON text frame onto the send buffer. Drops silently if the
@@ -85,6 +93,24 @@ func (c *wsClient) close() {
 	})
 }
 
+// challengeCaps advertises optional protocol features. It is appended to every
+// challenge, including the rotation after a failed login, because the browser
+// overwrites its capability state on each challenge. Old frontends ignore it.
+const challengeCaps = `,"caps":{"multiWatch":true,"keyframeRequest":true}`
+
+// maxWatchPerClient bounds the number of devices one WebSocket may stream at
+// once. The wall tops out at 25 tiles; the slack covers a focus switch racing
+// a teardown. This keeps the per-frame broadcast loop cheap for a misbehaving
+// or buggy client.
+const maxWatchPerClient = 32
+
+// isWatching reports whether this connection subscribed to deviceID.
+// Callers must hold wsHub.mu (read or write).
+func (c *wsClient) isWatching(deviceID string) bool {
+	_, ok := c.watchIDs[deviceID]
+	return ok
+}
+
 // ----- ws hub: registry of all connected browsers -------------------------
 
 type wsHub struct {
@@ -253,10 +279,19 @@ func (h *wsHub) OnDeviceOffline(id string) {
 		"cmd": "device_offline",
 		"id":  id,
 	})
-	h.mu.RLock()
+	// Drop the id from every watch set as well: the device is gone, so this
-	defer h.mu.RUnlock()
+	// connection no longer holds a subscription for it. The browser does the
+	// same with its local refcount. Without this, a stale entry would keep
+	// countWatchers non-zero and later block the screen teardown.
+	h.mu.Lock()
+	defer h.mu.Unlock()
 	for c := range h.clients {
-		if c.watching == id && c.token != "" {
+		if !c.isWatching(id) {
+			continue
+		}
+		delete(c.watchIDs, id)
+		delete(c.keyframeAt, id)
+		if c.token != "" {
 			c.queue(msg)
 		}
 	}
@@ -269,12 +304,13 @@ func (h *wsHub) OnDeviceOffline(id string) {
 func (h *wsHub) OnCursorChange(deviceID string, index byte) {
 	msg := mustJSON(map[string]any{
 		"cmd":   "cursor",
+		"id":    deviceID,
 		"index": index,
 	})
 	h.mu.RLock()
 	defer h.mu.RUnlock()
 	for c := range h.clients {
-		if c.watching == deviceID && c.token != "" {
+		if c.isWatching(deviceID) && c.token != "" {
 			c.queue(msg)
 		}
 	}
@@ -294,7 +330,7 @@ func (h *wsHub) OnResolutionChange(deviceID string, width, height int) {
 	h.mu.RLock()
 	defer h.mu.RUnlock()
 	for c := range h.clients {
-		if c.watching == deviceID && c.token != "" {
+		if c.isWatching(deviceID) && c.token != "" {
 			c.queue(msg)
 		}
 	}
@@ -311,7 +347,7 @@ func (h *wsHub) OnClipboardText(deviceID string, text string) {
 	h.mu.RLock()
 	defer h.mu.RUnlock()
 	for c := range h.clients {
-		if c.watching == deviceID && c.token != "" {
+		if c.isWatching(deviceID) && c.token != "" {
 			c.queue(msg)
 		}
 	}
@@ -325,7 +361,7 @@ func (h *wsHub) OnScreenFrame(deviceID string, packet []byte, _ bool) {
 	h.mu.RLock()
 	defer h.mu.RUnlock()
 	for c := range h.clients {
-		if c.watching == deviceID && c.token != "" {
+		if c.isWatching(deviceID) && c.token != "" {
 			c.queueBinary(packet)
 		}
 	}
@@ -424,12 +460,22 @@ func (h *wsHub) register(c *wsClient) {
 func (h *wsHub) unregister(c *wsClient) {
 	h.mu.Lock()
 	delete(h.clients, c)
+	// Snapshot the watch set and clear it under the same lock. The teardown
+	// below must run OUTSIDE the lock: countWatchers takes h.mu.RLock and
+	// deferredCloseScreen takes h.mu.Lock.
+	watched := make([]string, 0, len(c.watchIDs))
+	for id := range c.watchIDs {
+		watched = append(watched, id)
+	}
+	c.watchIDs = nil
+	c.keyframeAt = nil
 	h.mu.Unlock()
 	// If this client was the last viewer of a device, tear down the screen
-	// session so the device stops encoding. Done OUTSIDE the lock so the
+	// session so the device stops encoding.
-	// hub's mutators can take their own locks without risk of recursion.
+	for _, id := range watched {
-	if c.watching != "" && h.countWatchers(c.watching) == 0 {
+		if h.countWatchers(id) == 0 {
-		h.deferredCloseScreen(c.watching)
+			h.deferredCloseScreen(id)
+		}
 	}
 	// Terminal sessions are single-viewer by design, so any open session
 	// belongs to this client. Tear it down so the next viewer doesn't
@@ -497,6 +543,8 @@ func (h *wsHub) serve(w http.ResponseWriter, r *http.Request) {
 		closed:     make(chan struct{}),
 		nonce:      nonce,
 		addr:       clientIP(r, h.trustForwardedFor),
+		watchIDs:   make(map[string]struct{}),
+		keyframeAt: make(map[string]int64),
 	}
 	h.register(client)
 	defer h.unregister(client)
@@ -504,7 +552,7 @@ func (h *wsHub) serve(w http.ResponseWriter, r *http.Request) {
 	go h.writeLoop(client)
 
 	// Greet with a challenge nonce so the browser can compute the login response.
-	client.queue([]byte(`{"cmd":"challenge","nonce":"` + nonce + `"}`))
+	client.queue([]byte(`{"cmd":"challenge","nonce":"` + nonce + `"` + challengeCaps + `}`))
 
 	h.readLoop(client)
 }

server/go/web/ws_handlers.go

@@ -24,7 +24,8 @@ func (h *wsHub) rotateChallenge(c *wsClient) {
 		return
 	}
 	c.nonce = n
-	c.queue([]byte(`{"cmd":"challenge","nonce":"` + n + `"}`))
+	// Must carry caps too: the browser replaces serverCaps on every challenge.
+	c.queue([]byte(`{"cmd":"challenge","nonce":"` + n + `"` + challengeCaps + `}`))
 }
 
 // dispatch routes one inbound message to its handler. The `raw` payload is
@@ -50,6 +51,8 @@ func (h *wsHub) dispatch(c *wsClient, cmd string, raw []byte) {
 
 	case "connect":
 		h.handleConnect(c, raw)
+	case "keyframe":
+		h.handleKeyframe(c, raw)
 	case "rdp_reset":
 		h.handleRdpReset(c, raw)
 	case "audio_toggle":
@@ -240,66 +243,196 @@ func (h *wsHub) resetLoginThrottle(c *wsClient, username string) {
 // screen sub-connection. Closing the TCP sub-conn is the signal the C++
 // device firmware uses to stop screen capture, so this is how we ask the
 // device to free its encoder.
-func (h *wsHub) handleDisconnect(c *wsClient, _ []byte) {
+func (h *wsHub) handleDisconnect(c *wsClient, raw []byte) {
-	// Mirror handleConnect: take h.mu so event-handler readers
+	var in struct {
-	// (OnResolutionChange/OnScreenFrame) get a consistent view of c.watching.
+		ID         string `json:"id"`
+		MultiWatch bool   `json:"multiWatch"`
+		Req        string `json:"req"`
+	}
+	_ = json.Unmarshal(raw, &in)
+
+	// Take h.mu so event-handler readers (OnResolutionChange/OnScreenFrame)
+	// get a consistent view of the watch set.
+	var closed []string
 	h.mu.Lock()
-	prev := c.watching
+	if in.MultiWatch && in.ID != "" {
-	c.watching = ""
+		// New frontend: drop only this device; the other wall tiles keep
+		// streaming on the same connection.
+		if c.isWatching(in.ID) {
+			delete(c.watchIDs, in.ID)
+			delete(c.keyframeAt, in.ID)
+			closed = append(closed, in.ID)
+		}
+	} else {
+		// Old frontend, or a disconnect without an id: it only ever watched
+		// one device, so clear everything.
+		for id := range c.watchIDs {
+			closed = append(closed, id)
+		}
+		c.watchIDs = make(map[string]struct{})
+		c.keyframeAt = make(map[string]int64)
+	}
 	h.mu.Unlock()
-	c.queue([]byte(`{"cmd":"disconnect_result","ok":true}`))
+
-	if prev != "" && h.countWatchers(prev) == 0 {
+	res := map[string]any{"cmd": "disconnect_result", "ok": true}
-		h.deferredCloseScreen(prev)
+	if in.ID != "" {
+		res["id"] = in.ID
+	}
+	if in.Req != "" {
+		res["req"] = in.Req
+	}
+	c.queue(mustJSON(res))
+
+	// deferredCloseScreen takes h.mu, so it must run outside the lock above.
+	for _, id := range closed {
+		if h.countWatchers(id) == 0 {
+			h.deferredCloseScreen(id)
+		}
 	}
 }
 
-// countWatchers returns how many authenticated clients still have their
+// countWatchers returns how many clients still have deviceID in their watch
-// `watching` field pointing at deviceID. Called from disconnect paths.
+// set. Called from disconnect paths.
 func (h *wsHub) countWatchers(deviceID string) int {
 	h.mu.RLock()
 	defer h.mu.RUnlock()
 	n := 0
 	for c := range h.clients {
-		if c.watching == deviceID {
+		if c.isWatching(deviceID) {
 			n++
 		}
 	}
 	return n
 }
 
+// applyConnectWatch mutates c.watchIDs for a connect request. The caller must
+// hold h.mu.
+//
+// It returns the ids dropped by the legacy replace-all semantics; the caller
+// must stop their screen sessions outside the lock (only if nobody else still
+// watches them). tooMany reports that the per-connection cap blocked the id.
+func applyConnectWatch(c *wsClient, id string, multiWatch bool) (stopped []string, tooMany bool) {
+	if c.watchIDs == nil {
+		c.watchIDs = make(map[string]struct{})
+	}
+	if !multiWatch {
+		// Old frontend: one device per connection. Replace the whole set so
+		// switching devices repeatedly cannot grow it without bound.
+		for old := range c.watchIDs {
+			if old != id {
+				stopped = append(stopped, old)
+			}
+		}
+		c.watchIDs = make(map[string]struct{})
+	}
+	if _, ok := c.watchIDs[id]; ok {
+		return stopped, false // repeated connect (focus switch) is a no-op
+	}
+	if len(c.watchIDs) >= maxWatchPerClient {
+		return stopped, true
+	}
+	c.watchIDs[id] = struct{}{}
+	return stopped, false
+}
+
+// resolveWatchDevice maps an input command to the device it targets. The new
+// frontend sends an explicit id, which must be in this connection's watch set
+// so a stale or forged id can never inject input into a device the browser is
+// not viewing. Old frontends omit id — accepted only while exactly one device
+// is watched. With several devices and no id the command is dropped.
+func (h *wsHub) resolveWatchDevice(c *wsClient, raw []byte) string {
+	var in struct {
+		ID string `json:"id"`
+	}
+	_ = json.Unmarshal(raw, &in)
+	h.mu.RLock()
+	defer h.mu.RUnlock()
+	if in.ID != "" {
+		if c.isWatching(in.ID) {
+			return in.ID
+		}
+		return ""
+	}
+	if len(c.watchIDs) == 1 {
+		for id := range c.watchIDs {
+			return id
+		}
+	}
+	return ""
+}
+
 func (h *wsHub) handleConnect(c *wsClient, raw []byte) {
 	if !h.requireAuth(c, raw, "connect_result") {
 		return
 	}
 	var in struct {
 		ID         string `json:"id"`
+		MultiWatch bool   `json:"multiWatch"`
+		Req        string `json:"req"`
 	}
 	if err := json.Unmarshal(raw, &in); err != nil || in.ID == "" {
 		c.queue(mustJSON(map[string]any{"cmd": "connect_result", "ok": false, "msg": "Bad request"}))
 		return
 	}
 
-	// If a screen session is already live for this device (another browser
+	// Every response echoes id and req: one socket may now serve several
-	// is already watching), reuse it: hand the new viewer the current
+	// devices (screen wall), so the browser routes by id (and req when it
-	// resolution and the most recent IDR keyframe so its decoder can start
+	// needs to pair a reply with a specific request).
-	// rendering immediately, without waiting for the next IDR (~15 s).
+	reply := func(ok bool, msg string) map[string]any {
-	cache := h.devices.ScreenState(in.ID)
+		m := map[string]any{"cmd": "connect_result", "ok": ok, "id": in.ID}
-	if cache.Active {
+		if in.Req != "" {
-		c.queue(mustJSON(map[string]any{
+			m["req"] = in.Req
-			"cmd": "connect_result", "ok": true,
-			"width": cache.Width, "height": cache.Height,
-		}))
-		if len(cache.Keyframe) > 0 {
-			c.queueBinary(cache.Keyframe)
 		}
+		if msg != "" {
+			m["msg"] = msg
+		}
+		return m
+	}
+
+	var stopped []string
+	var tooMany bool
 	h.mu.Lock()
-		// Cancel any pending deferred close so the relay stays alive.
+	stopped, tooMany = applyConnectWatch(c, in.ID, in.MultiWatch)
+	// A pending deferred close means the previous session is still alive for
+	// a few seconds; cancel it so the cached resolution + IDR hot path wins.
+	// Only when the subscription was accepted: a rejected connect must not
+	// cancel the teardown of a device nobody is watching any more, or that
+	// session would keep streaming forever.
+	if !tooMany {
 		if t, ok := h.screenCloseTimers[in.ID]; ok {
 			t.Stop()
 			delete(h.screenCloseTimers, in.ID)
 		}
-		c.watching = in.ID
+	}
 	h.mu.Unlock()
+
+	if tooMany {
+		c.queue(mustJSON(reply(false, "Too many devices on one connection")))
+		return
+	}
+	// Only the legacy replace-all semantics can produce stopped ids. Their
+	// screen sessions are torn down outside the lock; deferredCloseScreen is
+	// a no-op while another browser still watches the device.
+	for _, id := range stopped {
+		if h.countWatchers(id) == 0 {
+			h.deferredCloseScreen(id)
+		}
+	}
+
+	// If a screen session is already live for this device (another browser
+	// is already watching, or this connection reconnected), reuse it: hand
+	// the viewer the cached resolution and the most recent IDR keyframe so
+	// its decoder can start rendering immediately, without waiting for the
+	// next IDR (~15 s).
+	cache := h.devices.ScreenState(in.ID)
+	if cache.Active {
+		m := reply(true, "")
+		m["width"] = cache.Width
+		m["height"] = cache.Height
+		c.queue(mustJSON(m))
+		if len(cache.Keyframe) > 0 {
+			c.queueBinary(cache.Keyframe)
+		}
 		return
 	}
 
@@ -312,41 +445,101 @@ func (h *wsHub) handleConnect(c *wsClient, raw []byte) {
 	cmd[2] = protocol.AlgorithmH264
 	cmd[3] = 1
 
-	// CRITICAL: bind c.watching BEFORE asking the device to start capturing.
+	// CRITICAL: the watch entry is inserted above (applyConnectWatch) BEFORE
-	// On fast reconnects the device's screen sub-conn handshake completes in
+	// asking the device to start capturing. On fast reconnects the device's
-	// <100 ms, so TOKEN_BITMAPINFO and even the first H264 frame can arrive
+	// screen sub-conn handshake completes in <100 ms, so TOKEN_BITMAPINFO and
-	// before this handler finishes — and the resolution_changed / frame
+	// even the first H264 frame can arrive before this handler finishes — and
-	// broadcasts in wsHub filter on c.watching. With the assignment after
+	// the resolution_changed / frame broadcasts filter on the watch set. If
-	// SendToDevice the new viewer silently misses the very first IDR and
+	// the entry were added afterwards the new viewer would silently miss the
-	// resolution_changed, leaving the page stuck on "Waiting for video".
+	// very first IDR and resolution_changed, leaving the page stuck on
-	//
+	// "Waiting for video".
-	// The write needs to share the lock event handlers use to read c.watching
-	// (they iterate h.clients under h.mu.RLock). Without that the write is a
-	// data race; on a fast reconnect the reader goroutine can keep observing
-	// the previous value ("") long enough to drop the first resolution_changed
-	// and the first IDR, which produces the exact "every other quick reconnect
-	// goes black" symptom — the C++ server avoids it because it does the same
-	// state mutation under std::mutex and reaps the memory-barrier as a bonus.
-	h.mu.Lock()
-	c.watching = in.ID
-	h.mu.Unlock()
-
 	if err := h.devices.SendToDevice(in.ID, cmd); err != nil {
-		// Roll back the watching flag if we never managed to kick capture.
+		// Roll back the subscription: no screen session was started.
 		h.mu.Lock()
-		c.watching = ""
+		delete(c.watchIDs, in.ID)
+		delete(c.keyframeAt, in.ID)
 		h.mu.Unlock()
 		msg := "Device offline"
 		if err != hub.ErrDeviceOffline {
 			msg = "Failed to start screen capture"
 			h.log.Error("SendToDevice(%s): %v", in.ID, err)
 		}
-		c.queue(mustJSON(map[string]any{"cmd": "connect_result", "ok": false, "msg": msg}))
+		c.queue(mustJSON(reply(false, msg)))
 		return
 	}
 	h.log.Debug("[timing] COMMAND_SCREEN_SPY sent to device=%s (cold start)", in.ID)
 
-	c.queue(mustJSON(map[string]any{"cmd": "connect_result", "ok": true}))
+	c.queue(mustJSON(reply(true, "")))
+}
+
+// handleKeyframe asks the device to emit an IDR immediately. The browser calls
+// it when it resumes decoding a wall tile, or right after connecting to reset
+// the reference chain. The server first replays its cached IDR so the decoder
+// has something to paint at once, then sends COMMAND_SCREEN_KEYFRAME_REQ to
+// the client, whose fresh IDR replaces the cache (the cached one cannot
+// guarantee the following P-frames reference it).
+//
+// Mirrors CWebService::HandleKeyframe. Requests are limited to one per device
+// per second, so a retrying frontend cannot spin the encoder.
+func (h *wsHub) handleKeyframe(c *wsClient, raw []byte) {
+	if !h.requireAuth(c, raw, "keyframe_result") {
+		return
+	}
+	var in struct {
+		ID  string `json:"id"`
+		Req string `json:"req"`
+	}
+	_ = json.Unmarshal(raw, &in)
+
+	reply := func(ok bool, msg string) map[string]any {
+		m := map[string]any{"cmd": "keyframe_result", "ok": ok}
+		if in.ID != "" {
+			m["id"] = in.ID
+		}
+		if in.Req != "" {
+			m["req"] = in.Req
+		}
+		if msg != "" {
+			m["msg"] = msg
+		}
+		return m
+	}
+	if in.ID == "" {
+		c.queue(mustJSON(reply(false, "Bad request")))
+		return
+	}
+
+	allowed := false
+	h.mu.Lock()
+	if c.isWatching(in.ID) {
+		now := time.Now().Unix()
+		if last, ok := c.keyframeAt[in.ID]; !ok || now > last {
+			if c.keyframeAt == nil {
+				c.keyframeAt = make(map[string]int64)
+			}
+			c.keyframeAt[in.ID] = now
+			allowed = true
+		}
+	}
+	h.mu.Unlock()
+	if !allowed {
+		c.queue(mustJSON(reply(false, "Not watching device or rate limited")))
+		return
+	}
+
+	cache := h.devices.ScreenState(in.ID)
+	if !cache.Active {
+		c.queue(mustJSON(reply(false, "No active screen session")))
+		return
+	}
+	if len(cache.Keyframe) > 0 {
+		c.queueBinary(cache.Keyframe)
+	}
+	if err := h.devices.SendToScreen(in.ID, []byte{protocol.CommandScreenKeyframeReq}); err != nil {
+		c.queue(mustJSON(reply(false, "Failed to request keyframe")))
+		return
+	}
+	c.queue(mustJSON(reply(true, "")))
 }
 
 // handleRdpReset asks the device to switch its screen capture back to the
@@ -363,7 +556,7 @@ func (h *wsHub) handleRdpReset(c *wsClient, raw []byte) {
 	if !h.requireAuth(c, raw, "rdp_reset_result") {
 		return
 	}
-	deviceID := c.watching
+	deviceID := h.resolveWatchDevice(c, raw)
 	if deviceID == "" {
 		h.log.Warn("rdp_reset: no device watched (addr=%s role=%s)", c.addr, c.role)
 		return
@@ -440,7 +633,7 @@ func (h *wsHub) handleMouse(c *wsClient, raw []byte) {
 	if err := json.Unmarshal(raw, &in); err != nil {
 		return
 	}
-	deviceID := c.watching
+	deviceID := h.resolveWatchDevice(c, raw)
 	if deviceID == "" {
 		return // browser hasn't picked a device yet
 	}
@@ -525,7 +718,7 @@ func (h *wsHub) handleKey(c *wsClient, raw []byte) {
 	if in.KeyCode == protocol.VKLWin || in.KeyCode == protocol.VKRWin {
 		return // Windows keys stay local; matches both C++ server and client
 	}
-	deviceID := c.watching
+	deviceID := h.resolveWatchDevice(c, raw)
 	if deviceID == "" {
 		return
 	}
@@ -583,7 +776,7 @@ func (h *wsHub) handleClipboardSet(c *wsClient, raw []byte) {
 	if in.Text == "" {
 		return
 	}
-	deviceID := c.watching
+	deviceID := h.resolveWatchDevice(c, raw)
 	if deviceID == "" {
 		return
 	}
@@ -607,7 +800,7 @@ func (h *wsHub) handleClipboardGet(c *wsClient, raw []byte) {
 	if !h.requireAuth(c, raw, "clipboard_get_result") {
 		return
 	}
-	deviceID := c.watching
+	deviceID := h.resolveWatchDevice(c, raw)
 	if deviceID == "" {
 		return
 	}