feat(admin): Agent 并发 - 后台可调同时对话上限并展示 Agent 服务状态

新增运行策略 agentMaxSessions（默认 30，范围 1-64），保存后同时下发到内嵌进程闸门与独立 yingce-agent 的 /v1/limit，降低上限不会中断已开始的对话。系统性能页展示 Agent 服务健康、进行中/排队数量，并可直接修改上限。

backend/internal/agent/runtime/runtime.go

@@ -222,27 +222,83 @@ func RuntimeDir() (string, error) {
 	return "", errors.New("Agent runtime files are missing; set CANVAS_PI_RUNTIME_DIR")
 }
 
-var (
-	agentSlotsOnce sync.Once
-	agentSlots     chan struct{}
+const (
+	defaultAgentProcessLimit = 30
+	maxAgentProcessLimit     = 64
 )
 
-func acquireAgentProcess(ctx context.Context) (func(), error) {
-	agentSlotsOnce.Do(func() {
-		limit := 8
-		if raw := strings.TrimSpace(os.Getenv("CANVAS_AGENT_MAX_PROCESSES")); raw != "" {
-			if parsed, err := strconv.Atoi(raw); err == nil && parsed > 0 && parsed <= 1000 {
-				limit = parsed
-			}
+type agentProcessGate struct {
+	mu     sync.Mutex
+	limit  int
+	active int
+	wake   chan struct{}
+}
+
+var processGate = newAgentProcessGate()
+
+func newAgentProcessGate() *agentProcessGate {
+	limit := defaultAgentProcessLimit
+	if raw := strings.TrimSpace(os.Getenv("CANVAS_AGENT_MAX_PROCESSES")); raw != "" {
+		if parsed, err := strconv.Atoi(raw); err == nil {
+			limit = clampAgentProcessLimit(parsed)
 		}
-		agentSlots = make(chan struct{}, limit)
-	})
-	select {
-	case agentSlots <- struct{}{}:
-		return func() { <-agentSlots }, nil
-	case <-ctx.Done():
-		return nil, ctx.Err()
 	}
+	return &agentProcessGate{limit: limit, wake: make(chan struct{})}
+}
+
+func clampAgentProcessLimit(value int) int {
+	if value < 1 {
+		return defaultAgentProcessLimit
+	}
+	if value > maxAgentProcessLimit {
+		return maxAgentProcessLimit
+	}
+	return value
+}
+
+// SetProcessLimit changes how many embedded Agent processes may run at once.
+// Lowering the limit does not stop a process that has already started.
+func SetProcessLimit(value int) (active int, limit int) {
+	processGate.mu.Lock()
+	defer processGate.mu.Unlock()
+	processGate.limit = clampAgentProcessLimit(value)
+	close(processGate.wake)
+	processGate.wake = make(chan struct{})
+	return processGate.active, processGate.limit
+}
+
+func ProcessUsage() (active int, limit int) {
+	processGate.mu.Lock()
+	defer processGate.mu.Unlock()
+	return processGate.active, processGate.limit
+}
+
+func acquireAgentProcess(ctx context.Context) (func(), error) {
+	for {
+		processGate.mu.Lock()
+		if processGate.active < processGate.limit {
+			processGate.active++
+			processGate.mu.Unlock()
+			return releaseAgentProcess, nil
+		}
+		wake := processGate.wake
+		processGate.mu.Unlock()
+		select {
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		case <-wake:
+		}
+	}
+}
+
+func releaseAgentProcess() {
+	processGate.mu.Lock()
+	if processGate.active > 0 {
+		processGate.active--
+	}
+	close(processGate.wake)
+	processGate.wake = make(chan struct{})
+	processGate.mu.Unlock()
 }
 
 func nodeMemoryMB() string {

backend/internal/agent/runtime/runtime_test.go

@@ -24,6 +24,25 @@ func fakeRuntime(t *testing.T, script string) {
 	t.Setenv("CANVAS_PI_RUNTIME_DIR", dir)
 }
 
+func TestSetProcessLimitDoesNotStopActiveWork(t *testing.T) {
+	t.Cleanup(func() { SetProcessLimit(defaultAgentProcessLimit) })
+	release, err := acquireAgentProcess(context.Background())
+	if err != nil {
+		t.Fatal(err)
+	}
+	active, limit := SetProcessLimit(1)
+	if active != 1 || limit != 1 {
+		t.Fatalf("active=%d limit=%d, want active 1 and limit 1", active, limit)
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Millisecond)
+	defer cancel()
+	if _, err := acquireAgentProcess(ctx); err == nil {
+		t.Fatal("lowered limit still admitted another process")
+	}
+	release()
+	SetProcessLimit(defaultAgentProcessLimit)
+}
+
 func noopBridge() Bridge {
 	handler := func(context.Context, map[string]json.RawMessage) (any, error) { return map[string]any{"ok": true}, nil }
 	return Bridge{Model: handler, Tool: handler, Event: handler}