feat: correx-managed model lifecycle slice 5 — Go TUI model swap + telemetry

Go TUI decodes model.changed / model.list / resource.status and renders a
status-bar VRAM/GPU%/RAM gauge plus a models overlay (m key / palette 'models'):
lists configured models with the resident one marked, enter swaps (SwapModel),
c clears the pin (ClearModelPin). All three new server messages are
non-event-bearing control/gauge frames, applied immediately like
provider.status_changed.

Server addition required for the picker: ServerMessage.ModelList (model.list,
NonEventMessage) sent once in the initial snapshot from
ManagedInferenceRouter.availableModelIds()/currentModelId() — the TUI otherwise
cannot enumerate swap targets.

Completes the 5-slice model-lifecycle feature. ./gradlew check green; go build/vet clean.

Plan: docs/plans/2026-05-31-model-lifecycle-management.md (slice 5 of 5).

apps/server/src/main/kotlin/com/correx/apps/server/protocol/ServerMessage.kt

@@ -283,6 +283,19 @@ sealed interface ServerMessage {
         override val sessionSequence: Long? = null,
     ) : ServerMessage, NonEventMessage
 
+    /**
+     * The configured local models and which one is currently resident. Sent once in the initial
+     * snapshot when correx manages the model process, so clients can offer a swap picker.
+     */
+    @Serializable
+    @SerialName("model.list")
+    data class ModelList(
+        val models: List<String>,
+        val current: String?,
+        override val sequence: Long? = null,
+        override val sessionSequence: Long? = null,
+    ) : ServerMessage, NonEventMessage
+
     /**
      * Live resource gauge pushed periodically on the global stream (not event-derived). All fields
      * are nullable: GPU fields are null on a non-NVIDIA host, `processRssMb` is null when no model

apps/server/src/main/kotlin/com/correx/apps/server/ws/GlobalStreamHandler.kt

@@ -149,6 +149,16 @@ class GlobalStreamHandler(private val module: ServerModule) {
             ServerMessage.WorkflowList(workflows = workflows),
         )))
 
+        // Configured models + current, so the client can offer a swap picker (managed path only).
+        module.modelSwapper?.let { swapper ->
+            session.send(Frame.Text(ProtocolSerializer.encodeServerMessage(
+                ServerMessage.ModelList(
+                    models = swapper.availableModelIds(),
+                    current = swapper.currentModelId(),
+                ),
+            )))
+        }
+
         // Initial resource gauge so a freshly-connected client renders VRAM/RAM immediately.
         val snapshot = withContext(Dispatchers.IO) { module.resourceProbe.probe() }
         session.send(Frame.Text(ProtocolSerializer.encodeServerMessage(snapshot.toResourceStatus())))

apps/tui-go/internal/app/model.go

@@ -45,6 +45,7 @@ const (
 	OverlayEventInspector
 	OverlayDiff
 	OverlayToolPalette
+	OverlayModels
 )
 
 // RouterEntry is one line in a session's conversation transcript.
@@ -158,6 +159,14 @@ type Model struct {
 	currentModel string
 	providerType string // LOCAL | REMOTE
 
+	// managed-model swap (nil resource fields = unavailable)
+	availableModels []string
+	modelsIndex     int
+	gpuUsedMB       *int64
+	gpuTotalMB      *int64
+	gpuUtil         *int
+	ramMB           *int64
+
 	// diff / overlay
 	overlay          OverlayKind
 	overlayEventIdx  int

apps/tui-go/internal/app/overlays.go

@@ -37,6 +37,8 @@ func (m Model) renderOverlay(base string) string {
 		return m.center(m.diffModal())
 	case OverlayToolPalette:
 		return m.center(m.toolPaletteModal())
+	case OverlayModels:
+		return m.center(m.modelsModal())
 	}
 	return base
 }
@@ -243,6 +245,38 @@ func (m Model) toolPaletteModal() string {
 	return m.center(modal)
 }
 
+func (m Model) modelsModal() string {
+	t := m.theme
+	w := m.modalWidth()
+
+	var b strings.Builder
+	b.WriteString(m.titleLine("models") + "\n\n")
+
+	if g := m.gaugeText(); g != "" {
+		b.WriteString(mbg(t, g, t.P.Faint) + "\n\n")
+	}
+
+	if len(m.availableModels) == 0 {
+		b.WriteString(mbg(t, "  model management not enabled", t.P.Faint) + "\n")
+	}
+	for i, id := range m.availableModels {
+		marker := mbg(t, "  ", t.P.BgPanel)
+		nameFg := t.P.Fg
+		if i == m.modelsIndex {
+			marker = lipgloss.NewStyle().Foreground(t.P.Accent).Background(t.P.BgPanel).Render("▌ ")
+			nameFg = t.P.FgStrong
+		}
+		line := marker + lipgloss.NewStyle().Foreground(nameFg).Background(t.P.BgPanel).Render(padRaw(id, 28))
+		if id == m.currentModel {
+			line += lipgloss.NewStyle().Foreground(t.P.Accent2).Background(t.P.BgPanel).Render("  ◂ resident")
+		}
+		b.WriteString(line + "\n")
+	}
+	b.WriteString("\n" + modalHints(t, [][2]string{{"↑↓", "select"}, {"enter", "swap"}, {"c", "clear pin"}, {"esc", "close"}}))
+	modal := t.Overlay.Width(w).Render(b.String())
+	return m.center(modal)
+}
+
 func tierColor(t Theme, tier int) lipgloss.Color {
 	switch {
 	case tier >= 3:

apps/tui-go/internal/app/server.go

@@ -190,6 +190,22 @@ func (m *Model) applyServer(msg protocol.ServerMessage) {
 		for _, w := range msg.Workflows {
 			m.workflows = append(m.workflows, Workflow{ID: w.WorkflowID, Description: w.Description})
 		}
+	case protocol.TypeModelChanged:
+		if msg.Loaded {
+			m.currentModel = msg.ModelID
+			m.providerType = "LOCAL"
+		}
+	case protocol.TypeModelList:
+		m.availableModels = append(m.availableModels[:0], msg.Models...)
+		if msg.Current != "" {
+			m.currentModel = msg.Current
+			m.providerType = "LOCAL"
+		}
+	case protocol.TypeResourceStatus:
+		m.gpuUsedMB = msg.GpuMemoryUsedMb
+		m.gpuTotalMB = msg.GpuMemoryTotalMb
+		m.gpuUtil = msg.GpuUtilizationPct
+		m.ramMB = msg.ProcessRssMb
 	}
 
 	// Background-update badge for non-selected sessions.
@@ -202,7 +218,8 @@ func sessionIDOf(msg protocol.ServerMessage) string {
 	switch msg.Type {
 	case protocol.TypeStageManifest, protocol.TypeSnapshotComplete,
 		protocol.TypeProtocolError, protocol.TypeProviderStatus,
-		protocol.TypeWorkflowList, protocol.TypeRouterResponse:
+		protocol.TypeWorkflowList, protocol.TypeRouterResponse,
+		protocol.TypeModelChanged, protocol.TypeModelList, protocol.TypeResourceStatus:
 		return ""
 	default:
 		return msg.SessionID

apps/tui-go/internal/app/update.go

@@ -144,6 +144,8 @@ func (m Model) handleNormalKey(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 		m.overlayEventIdx = 0
 	case "t":
 		m.overlay = OverlayToolPalette
+	case "m":
+		m.openModelsOverlay()
 	case "w":
 		if ds == StateIdle {
 			m.wfVisible = !m.wfVisible
@@ -333,10 +335,43 @@ func (m Model) handleOverlayKey(k tea.KeyMsg) (tea.Model, tea.Cmd) {
 		if runeIs(k, "t") {
 			m.overlay = OverlayNone
 		}
+	case OverlayModels:
+		switch {
+		case k.Type == tea.KeyUp || runeIs(k, "k"):
+			if m.modelsIndex > 0 {
+				m.modelsIndex--
+			}
+		case k.Type == tea.KeyDown || runeIs(k, "j"):
+			if m.modelsIndex < len(m.availableModels)-1 {
+				m.modelsIndex++
+			}
+		case k.Type == tea.KeyEnter:
+			if m.modelsIndex >= 0 && m.modelsIndex < len(m.availableModels) {
+				m.client.Send(protocol.SwapModel(m.availableModels[m.modelsIndex]))
+				m.overlay = OverlayNone
+			}
+		case runeIs(k, "c"):
+			m.client.Send(protocol.ClearModelPin())
+			m.overlay = OverlayNone
+		case runeIs(k, "m"):
+			m.overlay = OverlayNone
+		}
 	}
 	return m, nil
 }
 
+// openModelsOverlay opens the model picker, pre-selecting the resident model.
+func (m *Model) openModelsOverlay() {
+	m.overlay = OverlayModels
+	m.modelsIndex = 0
+	for i, id := range m.availableModels {
+		if id == m.currentModel {
+			m.modelsIndex = i
+			break
+		}
+	}
+}
+
 func runeIs(k tea.KeyMsg, s string) bool {
 	return k.Type == tea.KeyRunes && string(k.Runes) == s
 }
@@ -374,6 +409,7 @@ func paletteCommands() []paletteCmd {
 	return []paletteCmd{
 		{"workflows", "start workflow", "open the workflow picker"},
 		{"tools", "tool palette", "tools for the current stage"},
+		{"models", "swap model", "pick / pin the local model"},
 		{"events", "event inspector", "browse the event stream"},
 		{"mode", "toggle mode", "switch chat / steering"},
 		{"cancel", "cancel session", "stop the selected session"},
@@ -404,6 +440,8 @@ func (m Model) execPalette(id string) (tea.Model, tea.Cmd) {
 		m.wfIndex = 0
 	case "tools":
 		m.overlay = OverlayToolPalette
+	case "models":
+		m.openModelsOverlay()
 	case "events":
 		m.overlay = OverlayEventInspector
 		m.overlayEventIdx = 0

apps/tui-go/internal/app/view.go

@@ -81,6 +81,9 @@ func (m Model) renderStatus() string {
 	left := strings.Join(parts, sep)
 
 	var right []string
+	if g := m.gaugeText(); g != "" {
+		right = append(right, span(g, t.P.Faint))
+	}
 	if s := m.session(m.selectedID); s != nil && s.Active {
 		right = append(right, lipgloss.NewStyle().Foreground(t.P.Accent).Background(bg).Render(m.spinner())+span(" active", t.P.Accent2))
 	}
@@ -93,6 +96,21 @@ func (m Model) renderStatus() string {
 	return m.justify(left, strings.Join(right, span("  ·  ", t.P.Faint)), m.width, bg)
 }
 
+// gaugeText renders the live VRAM/GPU/RAM gauge, omitting any unavailable field.
+func (m Model) gaugeText() string {
+	var parts []string
+	if m.gpuUsedMB != nil && m.gpuTotalMB != nil {
+		parts = append(parts, "VRAM "+itoa(int(*m.gpuUsedMB))+"/"+itoa(int(*m.gpuTotalMB))+"M")
+	}
+	if m.gpuUtil != nil {
+		parts = append(parts, "GPU "+itoa(*m.gpuUtil)+"%")
+	}
+	if m.ramMB != nil {
+		parts = append(parts, "RAM "+itoa(int(*m.ramMB))+"M")
+	}
+	return strings.Join(parts, " ")
+}
+
 var spinnerFrames = []string{"⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"}
 
 func (m Model) spinner() string { return spinnerFrames[m.frame%len(spinnerFrames)] }
@@ -121,7 +139,7 @@ func (m Model) renderFooter() string {
 	case m.displayState() == StateIdle:
 		hints = []string{hint("i", "name"), hint("/", "filter"), hint("enter", "open"), hint("jk", "move"), hint("w", "workflows"), hint("p", "cmds"), hint("q", "quit")}
 	default: // StateInSession
-		hints = []string{hint("i", "message"), hint("e", "events"), hint("t", "tools"), hint("s", "mode"), hint("l", "back"), hint("p", "cmds"), hint("q", "quit")}
+		hints = []string{hint("i", "message"), hint("e", "events"), hint("t", "tools"), hint("m", "model"), hint("s", "mode"), hint("l", "back"), hint("p", "cmds"), hint("q", "quit")}
 	}
 	left := strings.Join(hints, gap)
 

apps/tui-go/internal/protocol/protocol.go

@@ -38,6 +38,9 @@ const (
 	TypeSnapshotComplete = "snapshot_complete"
 	TypeWorkflowList     = "workflow.list"
 	TypeToolAssessed     = "tool.assessed"
+	TypeModelChanged     = "model.changed"
+	TypeModelList        = "model.list"
+	TypeResourceStatus   = "resource.status"
 )
 
 // ServerMessage is a flat decode of every server->client variant. Field names
@@ -69,6 +72,20 @@ type ServerMessage struct {
 	ElapsedMs       int64 `json:"elapsedMs"`
 	LastSequence    int64 `json:"lastSequence"`
 
+	// model.changed
+	ModelID string `json:"modelId"`
+	Loaded  bool   `json:"loaded"`
+
+	// model.list (current reuses the field below)
+	Models  []string `json:"models"`
+	Current string   `json:"current"`
+
+	// resource.status (nil = unavailable)
+	GpuMemoryUsedMb   *int64 `json:"gpuMemoryUsedMb"`
+	GpuMemoryTotalMb  *int64 `json:"gpuMemoryTotalMb"`
+	GpuUtilizationPct *int   `json:"gpuUtilizationPct"`
+	ProcessRssMb      *int64 `json:"processRssMb"`
+
 	State          *SessionStateDto   `json:"state"`
 	RiskSummary    *RiskSummaryDto    `json:"riskSummary"`
 	Status         *ProviderHealthDto `json:"status"`
@@ -206,6 +223,16 @@ func Ping(ts int64) []byte {
 	return encode("Ping", map[string]any{"timestamp": ts})
 }
 
+// SwapModel asks the server to make modelID resident and pin it.
+func SwapModel(modelID string) []byte {
+	return encode("SwapModel", map[string]any{"modelId": modelID})
+}
+
+// ClearModelPin releases the manual-swap pin so per-stage selection resumes.
+func ClearModelPin() []byte {
+	return encode("ClearModelPin", map[string]any{})
+}
+
 // CreateGrant pre-authorizes a tool/tier so future calls skip the approval gate.
 // scope is "SESSION" or "STAGE"; permittedTiers are tier names like "T3".
 func CreateGrant(sessionID, scope string, permittedTiers []string, reason, toolName string) []byte {

infrastructure/inference/commons/src/main/kotlin/com/correx/infrastructure/inference/commons/ManagedInferenceRouter.kt

@@ -79,6 +79,12 @@ class ManagedInferenceRouter(
     /** The currently pinned model id, or null when no manual swap is active. */
     fun pinnedModel(): String? = pinnedModelId
 
+    /** All configured model ids, in config order — the swap candidates surfaced to clients. */
+    fun availableModelIds(): List<String> = ordered.map { it.modelId }
+
+    /** The currently resident model id, or null when none is loaded. */
+    fun currentModelId(): String? = manager.currentModel()?.modelId
+
     /**
      * Selects the configured model whose declared capabilities cover every required capability,
      * preferring the highest summed score over those required capabilities. Returns null when no