fix(backend): stabilize ACP session resume and workspace git poll under contention (#1242)

apps/backend/internal/agentctl/server/adapter/transport/acp/adapter.go

@@ -47,6 +47,14 @@ const (
 // allow rather than a tight RPC deadline.
 const wakeupPromptTimeout = 30 * time.Minute
 
+// notifQueueCapacity sizes the buffered channel that feeds the update
+// worker. 4096 covers any realistic session/load replay burst (the failure
+// case that motivated this was ~5-10k notifications spread over several
+// seconds, well within the worker's sub-microsecond drain rate). Set big
+// enough that the SDK's 1024-slot upstream queue stays near-empty under
+// burst, small enough that worst-case memory is bounded.
+const notifQueueCapacity = 4096
+
 // AgentInfo contains information about the connected agent.
 type AgentInfo struct {
 	Name    string `json:"name"`
@@ -83,6 +91,26 @@ type Adapter struct {
 	// Update channel
 	updatesCh chan AgentEvent
 
+	// notifQueue decouples the ACP SDK's notification goroutine from the
+	// per-notification processing we do in handleACPUpdate. The SDK's inbound
+	// channel is a 1024-slot non-blocking send — if our handler ever stalls
+	// (slow disk, GC, downstream backpressure on updatesCh), the SDK closes
+	// the connection. By front-loading a buffered channel + dedicated worker,
+	// the SDK only ever sees nanosecond-scale enqueues. workerWg lets Close
+	// wait for in-flight notifications to drain before tearing down updatesCh.
+	//
+	// Items are notifWork so the queue can also carry barrier-sync requests
+	// from syncNotifQueue — the worker closes the barrier's channel in FIFO
+	// order, which lets sendPrompt wait for queued text chunks before it
+	// emits EventTypeComplete.
+	//
+	// Drained-by-cancel, not by close: Close cancels lifetimeCtx (which the
+	// worker selects on) but never closes notifQueue, so any items queued
+	// after Close are dropped on the floor rather than re-delivered. Close
+	// is terminal — fine for shutdown but worth knowing when reading the loop.
+	notifQueue chan notifWork
+	workerWg   sync.WaitGroup
+
 	// Permission handler
 	permissionHandler PermissionHandler
 
@@ -196,6 +224,7 @@ func NewAdapter(cfg *shared.Config, log *logger.Logger) *Adapter {
 		agentID:         cfg.AgentID,
 		normalizer:      NewNormalizer(cfg.AgentID),
 		updatesCh:       make(chan AgentEvent, 100),
+		notifQueue:      make(chan notifWork, notifQueueCapacity),
 		activeToolCalls: make(map[string]*streams.NormalizedPayload),
 		activeMonitors:  make(map[string]map[string]string),
 		pendingWakeups:  make(map[string]*pendingWakeup),
@@ -206,6 +235,12 @@ func NewAdapter(cfg *shared.Config, log *logger.Logger) *Adapter {
 		lifetimeCancel:  cancel,
 	}
 	a.wakeup = newWakeupScheduler(l, a.fireWakeup)
+	// Start the update worker before returning so any caller that connects
+	// the SDK (Initialize, or any future direct wiring) always has a live
+	// consumer for notifQueue. Moving this out of Initialize closes a latent
+	// footgun where a future caller bypassing Initialize would silently fill
+	// the queue and lose notifications.
+	a.startUpdateWorker()
 	return a
 }
 
@@ -241,11 +276,11 @@ func (a *Adapter) Initialize(ctx context.Context) error {
 	a.logger.Info("initializing ACP adapter",
 		zap.String("workdir", a.cfg.WorkDir))
 
-	// Create ACP client with update handler that converts to AgentEvent
+	// Create ACP client with update handler that enqueues for the worker.
 	a.acpClient = acpclient.NewClient(
 		acpclient.WithLogger(a.logger.Zap()),
 		acpclient.WithWorkspaceRoot(a.cfg.WorkDir),
-		acpclient.WithUpdateHandler(a.handleACPUpdate),
+		acpclient.WithUpdateHandler(a.enqueueACPUpdate),
 		acpclient.WithPermissionHandler(a.handlePermissionRequest),
 	)
 
@@ -367,17 +402,19 @@ func (a *Adapter) sendUpdate(event AgentEvent) {
 // Close releases resources held by the adapter.
 func (a *Adapter) Close() error {
 	a.mu.Lock()
-	defer a.mu.Unlock()
-
 	if a.closed {
+		a.mu.Unlock()
 		return nil
 	}
 	a.closed = true
+	a.mu.Unlock()
 
 	a.logger.Info("closing ACP adapter")
 
 	// Stop any pending ScheduleWakeup timer so it doesn't fire after close,
 	// and cancel the lifetime context so any in-flight wakeup prompt aborts.
+	// Cancelling lifetimeCtx also unblocks enqueueACPUpdate senders and
+	// signals the update worker to exit.
 	if a.wakeup != nil {
 		a.wakeup.cancel()
 	}
@@ -385,6 +422,11 @@ func (a *Adapter) Close() error {
 		a.lifetimeCancel()
 	}
 
+	// Wait for the update worker to exit before closing updatesCh.
+	// handleACPUpdate may call sendUpdate, so updatesCh must remain open
+	// until the worker is gone.
+	a.workerWg.Wait()
+
 	// Clean up any saved attachments
 	a.attachMgr.Cleanup()
 

apps/backend/internal/agentctl/server/adapter/transport/acp/adapter_prompt.go

@@ -150,6 +150,21 @@ func (a *Adapter) sendPrompt(ctx context.Context, message string, attachments []
 		return err
 	}
 
+	// Drain queued ACP notifications before running the post-prompt sweeps and
+	// the complete event. The worker (now async) may still hold final frames
+	// the agent emitted right before its prompt response — the final text
+	// chunk, the terminal monitor_end tool_call_update, the registration
+	// frame for a Monitor whose events haven't yet been routed. Without this
+	// barrier:
+	//   - cancelActiveToolCalls / sweepMonitorsOnPromptEnd race the worker
+	//     for activeToolCalls and activeMonitors. If the worker hasn't added
+	//     a Monitor yet when the sweep takes the map, subsequent monitor_event
+	//     text frames find no tracking and drop their events on the floor.
+	//   - The complete event emitted to updatesCh outruns the final text chunk,
+	//     so the downstream buffer flush yields empty and the turn persists as
+	//     had_output=false even when the agent did produce text.
+	a.syncNotifQueue()
+
 	// Cancel any tool calls still in-flight (e.g. a denied permission leaves the
 	// tool_call without a terminal status update from the agent).
 	a.cancelActiveToolCalls(sessionID)

apps/backend/internal/agentctl/server/adapter/transport/acp/adapter_updates.go

@@ -10,20 +10,107 @@ import (
 	"go.uber.org/zap"
 )
 
-// handleACPUpdate converts ACP SessionNotification to protocol-agnostic AgentEvent.
-func (a *Adapter) handleACPUpdate(n acp.SessionNotification) {
-	// Marshal once for both debug logging and tracing
-	rawData, _ := json.Marshal(n)
+// notifWork is the item type carried on notifQueue. Exactly one of the two
+// fields is populated: notif for a real SDK notification (the common case),
+// sync for a barrier posted by syncNotifQueue. The worker closes sync when
+// it pops the barrier off the queue, releasing the waiter once everything
+// queued ahead of it has been processed.
+type notifWork struct {
+	notif acp.SessionNotification
+	sync  chan struct{}
+}
 
-	// Log raw event for debugging
-	if len(rawData) > 0 {
-		shared.LogRawEvent(shared.ProtocolACP, a.agentID, string(n.SessionId), "session_notification", rawData)
+// enqueueACPUpdate is the SDK-facing notification handler. It pushes the
+// notification onto notifQueue and returns immediately, so the ACP SDK's
+// receive loop is never blocked by our per-item processing (json.Marshal,
+// debug log I/O, monitor capture, downstream sendUpdate). The actual work
+// runs on runUpdateWorker.
+//
+// The select honours lifetimeCtx so a Close in flight unblocks any sender
+// that would otherwise stall on a full queue. Sending blocks (rather than
+// dropping) when the queue is full: with the handleACPUpdate fast path the
+// worker drains in nanoseconds-to-microseconds, so a full 4096-slot queue
+// means we're in a catastrophic stall and slowing the SDK is preferable to
+// silently losing notifications.
+func (a *Adapter) enqueueACPUpdate(n acp.SessionNotification) {
+	select {
+	case <-a.lifetimeCtx.Done():
+		return
+	case a.notifQueue <- notifWork{notif: n}:
 	}
+}
 
-	// During session/load, suppress history replay notifications.
-	// ACP agents stream the entire conversation history during load, which should not
-	// be emitted as new message events to avoid duplicating messages in the UI.
-	// We suppress: message chunks, thinking chunks, tool calls, and tool updates.
+// syncNotifQueue blocks until every notifWork enqueued before this call has
+// been processed by the worker. It works by posting a barrier item onto the
+// same FIFO queue and waiting for the worker to close it.
+//
+// Motivation: the worker is async, so after conn.Prompt() returns the final
+// text chunks emitted by the agent right before the prompt response may
+// still be sitting in notifQueue. If sendPrompt emits EventTypeComplete
+// immediately, the downstream "turn complete" handler flushes the message
+// buffer before those chunks land, the agent's text is dropped, and the
+// turn persists as had_output=false. Calling this before the complete emit
+// guarantees the chunks are delivered to updatesCh first.
+//
+// No caller-context honored: returning before the barrier closes would
+// reintroduce the very race this primitive exists to prevent (sweeps and
+// EventTypeComplete running while the worker still has queued frames).
+// Only adapter shutdown via lifetimeCtx can release the wait early.
+func (a *Adapter) syncNotifQueue() {
+	ch := make(chan struct{})
+	select {
+	case <-a.lifetimeCtx.Done():
+		return
+	case a.notifQueue <- notifWork{sync: ch}:
+	}
+	select {
+	case <-a.lifetimeCtx.Done():
+	case <-ch:
+	}
+}
+
+// startUpdateWorker spawns the goroutine that drains notifQueue and calls
+// handleACPUpdate for each notification. Called exactly once from
+// NewAdapter so the queue always has a consumer for the adapter's
+// lifetime — do not call again (a second invocation would Add(1) to
+// workerWg and spawn a second reader, breaking the single-worker FIFO
+// guarantee). Close waits for the worker to exit via workerWg before
+// closing updatesCh.
+func (a *Adapter) startUpdateWorker() {
+	a.workerWg.Add(1)
+	go a.runUpdateWorker()
+}
+
+// runUpdateWorker is the worker loop. It exits when lifetimeCtx is cancelled
+// (by Close). A single worker preserves FIFO ordering across notifications,
+// matching the SDK's own serial delivery contract.
+func (a *Adapter) runUpdateWorker() {
+	defer a.workerWg.Done()
+	for {
+		select {
+		case <-a.lifetimeCtx.Done():
+			return
+		case item := <-a.notifQueue:
+			if item.sync != nil {
+				close(item.sync)
+				continue
+			}
+			a.handleACPUpdate(item.notif)
+		}
+	}
+}
+
+// handleACPUpdate converts ACP SessionNotification to protocol-agnostic AgentEvent.
+// Runs synchronously on the update worker goroutine; do not call from the SDK's
+// notification path (use enqueueACPUpdate instead). Unit tests invoke this
+// directly to exercise the conversion logic without spinning up the worker.
+func (a *Adapter) handleACPUpdate(n acp.SessionNotification) {
+	// Fast path during session/load: history-replay notifications can arrive as
+	// a burst large enough to overflow the ACP SDK's 1024-deep notification
+	// queue if the per-item handler is slow. Check the loading flag first and
+	// skip json.Marshal + LogRawEvent for replay notifications we'd suppress
+	// anyway. We still capture the Plan + Monitor state needed to reconcile
+	// after replay.
 	a.mu.RLock()
 	isLoading := a.isLoadingSession
 	a.mu.RUnlock()
@@ -51,12 +138,16 @@ func (a *Adapter) handleACPUpdate(n acp.SessionNotification) {
 		if u.AgentMessageChunk != nil || u.UserMessageChunk != nil || u.AgentThoughtChunk != nil ||
 			u.ToolCall != nil || u.ToolCallUpdate != nil ||
 			u.Plan != nil || u.CurrentModeUpdate != nil || u.ConfigOptionUpdate != nil {
-			a.logger.Debug("suppressing history replay notification during session load",
-				zap.String("session_id", string(n.SessionId)))
 			return
 		}
 	}
 
+	// Marshal once for both debug logging and tracing.
+	rawData, _ := json.Marshal(n)
+	if len(rawData) > 0 {
+		shared.LogRawEvent(shared.ProtocolACP, a.agentID, string(n.SessionId), "session_notification", rawData)
+	}
+
 	sessionID := string(n.SessionId)
 
 	event := a.convertNotification(n)

apps/backend/internal/agentctl/server/adapter/transport/acp/load_burst_test.go

@@ -0,0 +1,381 @@
+package acp
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"testing"
+	"time"
+
+	acp "github.com/coder/acp-go-sdk"
+	"github.com/kandev/kandev/internal/agentctl/types/streams"
+)
+
+// burstAgent is a minimal acp.Agent stub used by the load-replay regression
+// test. None of these methods are exercised — we only need a peer that holds
+// the connection open while we push notifications agent → client.
+type burstAgent struct{}
+
+func (burstAgent) Initialize(context.Context, acp.InitializeRequest) (acp.InitializeResponse, error) {
+	return acp.InitializeResponse{}, nil
+}
+
+func (burstAgent) Authenticate(context.Context, acp.AuthenticateRequest) (acp.AuthenticateResponse, error) {
+	return acp.AuthenticateResponse{}, nil
+}
+func (burstAgent) Cancel(context.Context, acp.CancelNotification) error { return nil }
+func (burstAgent) CloseSession(context.Context, acp.CloseSessionRequest) (acp.CloseSessionResponse, error) {
+	return acp.CloseSessionResponse{}, nil
+}
+
+func (burstAgent) ListSessions(context.Context, acp.ListSessionsRequest) (acp.ListSessionsResponse, error) {
+	return acp.ListSessionsResponse{}, nil
+}
+
+func (burstAgent) NewSession(context.Context, acp.NewSessionRequest) (acp.NewSessionResponse, error) {
+	return acp.NewSessionResponse{}, nil
+}
+func (burstAgent) Prompt(context.Context, acp.PromptRequest) (acp.PromptResponse, error) {
+	return acp.PromptResponse{}, nil
+}
+
+func (burstAgent) ResumeSession(context.Context, acp.ResumeSessionRequest) (acp.ResumeSessionResponse, error) {
+	return acp.ResumeSessionResponse{}, nil
+}
+
+func (burstAgent) SetSessionConfigOption(context.Context, acp.SetSessionConfigOptionRequest) (acp.SetSessionConfigOptionResponse, error) {
+	return acp.SetSessionConfigOptionResponse{}, nil
+}
+
+func (burstAgent) SetSessionMode(context.Context, acp.SetSessionModeRequest) (acp.SetSessionModeResponse, error) {
+	return acp.SetSessionModeResponse{}, nil
+}
+
+// burstClient is a minimal acp.Client stub that forwards SessionUpdate
+// notifications to the adapter's real handler.
+type burstClient struct {
+	onUpdate func(acp.SessionNotification)
+}
+
+func (b *burstClient) SessionUpdate(_ context.Context, n acp.SessionNotification) error {
+	b.onUpdate(n)
+	return nil
+}
+
+func (*burstClient) ReadTextFile(context.Context, acp.ReadTextFileRequest) (acp.ReadTextFileResponse, error) {
+	return acp.ReadTextFileResponse{}, errors.New("not implemented")
+}
+
+func (*burstClient) WriteTextFile(context.Context, acp.WriteTextFileRequest) (acp.WriteTextFileResponse, error) {
+	return acp.WriteTextFileResponse{}, errors.New("not implemented")
+}
+
+func (*burstClient) RequestPermission(context.Context, acp.RequestPermissionRequest) (acp.RequestPermissionResponse, error) {
+	return acp.RequestPermissionResponse{}, errors.New("not implemented")
+}
+
+func (*burstClient) CreateTerminal(context.Context, acp.CreateTerminalRequest) (acp.CreateTerminalResponse, error) {
+	return acp.CreateTerminalResponse{}, errors.New("not implemented")
+}
+
+func (*burstClient) KillTerminal(context.Context, acp.KillTerminalRequest) (acp.KillTerminalResponse, error) {
+	return acp.KillTerminalResponse{}, errors.New("not implemented")
+}
+
+func (*burstClient) TerminalOutput(context.Context, acp.TerminalOutputRequest) (acp.TerminalOutputResponse, error) {
+	return acp.TerminalOutputResponse{}, errors.New("not implemented")
+}
+
+func (*burstClient) ReleaseTerminal(context.Context, acp.ReleaseTerminalRequest) (acp.ReleaseTerminalResponse, error) {
+	return acp.ReleaseTerminalResponse{}, errors.New("not implemented")
+}
+
+func (*burstClient) WaitForTerminalExit(context.Context, acp.WaitForTerminalExitRequest) (acp.WaitForTerminalExitResponse, error) {
+	return acp.WaitForTerminalExitResponse{}, errors.New("not implemented")
+}
+
+// newBurstPair wires a real acp.ClientSideConnection / AgentSideConnection
+// pair over in-memory pipes. The client's SessionUpdate forwards to the
+// supplied handler. Cleanup is registered on the test.
+func newBurstPair(t *testing.T, onUpdate func(acp.SessionNotification)) (*acp.ClientSideConnection, *acp.AgentSideConnection) {
+	t.Helper()
+
+	c2aR, c2aW := io.Pipe()
+	a2cR, a2cW := io.Pipe()
+
+	clientConn := acp.NewClientSideConnection(&burstClient{onUpdate: onUpdate}, c2aW, a2cR)
+	agentConn := acp.NewAgentSideConnection(burstAgent{}, a2cW, c2aR)
+
+	t.Cleanup(func() {
+		_ = c2aR.Close()
+		_ = c2aW.Close()
+		_ = a2cR.Close()
+		_ = a2cW.Close()
+	})
+
+	return clientConn, agentConn
+}
+
+// TestLoadReplayBurst_HandlesLargeReplay is a regression test for the
+// "notification queue overflow" failure we hit on a 304-exchange auggie
+// session load. The acp-go-sdk exposes a hardcoded 1024-slot inbound channel
+// (defaultMaxQueuedNotifications) and shuts down the connection when the
+// single-goroutine consumer falls behind. Before the adapter_updates.go
+// fast-path, json.Marshal + LogRawEvent on every replayed notification was
+// slow enough that the session/load burst would back the queue up.
+//
+// This test wires a real acp.ClientSideConnection / AgentSideConnection pair
+// over io.Pipes, sets isLoadingSession=true on the adapter, and pushes a
+// large burst of replay notifications agent → client. It then sends an
+// AvailableCommandsUpdate (which is intentionally NOT suppressed during
+// load) as a sentinel and asserts:
+//  1. the connection is still alive after the burst (overflow would have
+//     closed it),
+//  2. the sentinel flows through the un-suppressed path,
+//  3. the last Plan from the burst is captured for post-load re-emit.
+//
+// Note: in-memory io.Pipe is synchronous, so the producer is naturally
+// gated by the consumer drain rate; an overflow here would mean the
+// handler is hanging, not just slow. The throughput-sensitive scenario
+// is exercised by BenchmarkLoadReplayHandler below.
+func TestLoadReplayBurst_HandlesLargeReplay(t *testing.T) {
+	const notificationBurstSize = 20000
+	const sentinelCmd = "burst-sentinel"
+
+	a := newTestAdapter()
+	a.mu.Lock()
+	a.isLoadingSession = true
+	a.mu.Unlock()
+
+	// Exercise the production path: SDK → enqueueACPUpdate → notifQueue →
+	// runUpdateWorker → handleACPUpdate. NewAdapter starts the worker; we
+	// just rely on Close to drain it.
+	t.Cleanup(func() { _ = a.Close() })
+
+	clientConn, agentConn := newBurstPair(t, a.enqueueACPUpdate)
+
+	const sessionID = "burst-session"
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	// Push the burst of replay notifications. These would all be suppressed
+	// by the adapter (AgentMessageChunk + ToolCall + Plan are on the suppress
+	// list) — the point is whether the SDK's 1024-deep queue stays drained.
+	for i := 0; i < notificationBurstSize; i++ {
+		if err := agentConn.SessionUpdate(ctx, makeReplayNotification(sessionID, i)); err != nil {
+			t.Fatalf("SessionUpdate at i=%d failed: %v", i, err)
+		}
+	}
+
+	// Sentinel: AvailableCommandsUpdate passes through during load, so it
+	// will land on updatesCh once the consumer drains the queue.
+	if err := agentConn.SessionUpdate(ctx, acp.SessionNotification{
+		SessionId: acp.SessionId(sessionID),
+		Update: acp.SessionUpdate{
+			AvailableCommandsUpdate: &acp.SessionAvailableCommandsUpdate{
+				AvailableCommands: []acp.AvailableCommand{
+					{Name: sentinelCmd, Description: "burst sentinel"},
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatalf("sentinel SessionUpdate failed: %v", err)
+	}
+
+	// The sentinel-wait loop below is the authoritative overflow check: the
+	// SDK tears down both sides if its inbound queue ever stalls, so a closed
+	// Done() observed there proves overflow. A non-blocking peek here would
+	// race the async teardown and could produce a false PASS.
+
+	// Wait for the sentinel to flow through to the adapter's updates channel.
+	deadline := time.After(15 * time.Second)
+	var got *AgentEvent
+	for got == nil {
+		select {
+		case ev := <-a.updatesCh:
+			if ev.Type == streams.EventTypeAvailableCommands {
+				got = &ev
+			}
+		case <-clientConn.Done():
+			t.Fatal("client connection closed while waiting for sentinel — overflow")
+		case <-deadline:
+			t.Fatal("timeout waiting for sentinel AvailableCommands event after replay burst")
+		}
+	}
+
+	if len(got.AvailableCommands) != 1 || got.AvailableCommands[0].Name != sentinelCmd {
+		t.Fatalf("unexpected sentinel payload: %+v", got)
+	}
+
+	// Plan was sent during the burst and should have been captured (last
+	// Plan seen during load is stashed in loadReplayPlan for re-emit).
+	a.mu.RLock()
+	captured := a.loadReplayPlan
+	a.mu.RUnlock()
+	if captured == nil {
+		t.Fatal("expected loadReplayPlan to be captured during replay burst")
+	}
+}
+
+// makeReplayNotification builds a representative replay notification of one
+// of the suppressed kinds (AgentMessageChunk / ToolCall / Plan). Shared by
+// the burst test above and the benchmarks below.
+func makeReplayNotification(sessionID string, i int) acp.SessionNotification {
+	var update acp.SessionUpdate
+	switch i % 3 {
+	case 0:
+		update = acp.SessionUpdate{
+			AgentMessageChunk: &acp.SessionUpdateAgentMessageChunk{
+				Content: acp.TextBlock(fmt.Sprintf("replay chunk %d", i)),
+			},
+		}
+	case 1:
+		update = acp.SessionUpdate{
+			ToolCall: &acp.SessionUpdateToolCall{
+				ToolCallId: acp.ToolCallId(fmt.Sprintf("tc-%d", i)),
+				Title:      "replay tool call",
+			},
+		}
+	case 2:
+		update = acp.SessionUpdate{
+			Plan: &acp.SessionUpdatePlan{
+				Entries: []acp.PlanEntry{
+					{Content: fmt.Sprintf("plan entry %d", i), Status: "in_progress"},
+				},
+			},
+		}
+	}
+	return acp.SessionNotification{
+		SessionId: acp.SessionId(sessionID),
+		Update:    update,
+	}
+}
+
+// BenchmarkHandleACPUpdate_LoadSuppressed measures the per-notification cost
+// of the suppressed-during-load fast path. This is the hot loop during
+// session/load replay, and the path that previously did json.Marshal +
+// LogRawEvent on every notification before being short-circuited.
+func BenchmarkHandleACPUpdate_LoadSuppressed(b *testing.B) {
+	a := newTestAdapter()
+	b.Cleanup(func() { _ = a.Close() })
+	a.mu.Lock()
+	a.isLoadingSession = true
+	a.mu.Unlock()
+
+	notes := make([]acp.SessionNotification, 256)
+	for i := range notes {
+		notes[i] = makeReplayNotification("bench-session", i)
+	}
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		a.handleACPUpdate(notes[i%len(notes)])
+	}
+}
+
+// BenchmarkHandleACPUpdate_NormalPath measures the per-notification cost of
+// the non-loading path (json.Marshal + LogRawEvent + convertNotification +
+// updatesCh send). Drain updatesCh in a background goroutine so we don't
+// stall on the unbuffered/buffered send.
+func BenchmarkHandleACPUpdate_NormalPath(b *testing.B) {
+	a := newTestAdapter()
+	b.Cleanup(func() { _ = a.Close() })
+
+	done := make(chan struct{})
+	go func() {
+		for {
+			select {
+			case <-a.updatesCh:
+			case <-done:
+				return
+			}
+		}
+	}()
+	defer close(done)
+
+	notes := make([]acp.SessionNotification, 256)
+	for i := range notes {
+		notes[i] = makeReplayNotification("bench-session", i)
+	}
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		a.handleACPUpdate(notes[i%len(notes)])
+	}
+}
+
+// TestUpdateWorker_FIFOAndDrainOnClose verifies the SDK-decoupling worker:
+//  1. enqueueACPUpdate hands notifications to a single worker that processes
+//     them in FIFO order (preserving the SDK's serial-delivery contract);
+//  2. Close cancels the worker via lifetimeCtx, workerWg.Wait returns, and
+//     updatesCh is then safely closed — no goroutine leak, no panic.
+//
+// We don't poll the queue length directly; the contract we care about is
+// "the events arrive on updatesCh in order, and Close completes".
+func TestUpdateWorker_FIFOAndDrainOnClose(t *testing.T) {
+	const burst = 32
+
+	a := newTestAdapter()
+
+	// Push a burst of AvailableCommands notifications (one-shot events that
+	// land on updatesCh and carry an identifier we can assert order on).
+	for i := 0; i < burst; i++ {
+		a.enqueueACPUpdate(acp.SessionNotification{
+			SessionId: acp.SessionId("s1"),
+			Update: acp.SessionUpdate{
+				AvailableCommandsUpdate: &acp.SessionAvailableCommandsUpdate{
+					AvailableCommands: []acp.AvailableCommand{
+						{Name: fmt.Sprintf("cmd-%d", i)},
+					},
+				},
+			},
+		})
+	}
+
+	// Drain updatesCh until we've seen the full ordered burst.
+	deadline := time.After(5 * time.Second)
+	for i := 0; i < burst; i++ {
+		select {
+		case ev := <-a.updatesCh:
+			if ev.Type != streams.EventTypeAvailableCommands {
+				t.Fatalf("event %d: unexpected type %q", i, ev.Type)
+			}
+			if len(ev.AvailableCommands) != 1 {
+				t.Fatalf("event %d: expected 1 command, got %d", i, len(ev.AvailableCommands))
+			}
+			want := fmt.Sprintf("cmd-%d", i)
+			if got := ev.AvailableCommands[0].Name; got != want {
+				t.Fatalf("event %d: out-of-order: got %q want %q (worker is supposed to be FIFO)", i, got, want)
+			}
+		case <-deadline:
+			t.Fatalf("timeout after %d/%d events — worker stalled", i, burst)
+		}
+	}
+
+	// Close must return quickly and the worker goroutine must exit.
+	closed := make(chan error, 1)
+	go func() { closed <- a.Close() }()
+	select {
+	case err := <-closed:
+		if err != nil {
+			t.Fatalf("Close returned error: %v", err)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("Close did not return — worker likely leaked")
+	}
+
+	// Post-Close enqueues are no-ops (lifetimeCtx is cancelled). Must not
+	// block, must not panic.
+	done := make(chan struct{})
+	go func() {
+		a.enqueueACPUpdate(acp.SessionNotification{SessionId: "s1"})
+		close(done)
+	}()
+	select {
+	case <-done:
+	case <-time.After(time.Second):
+		t.Fatal("enqueueACPUpdate blocked after Close — lifetimeCtx guard missing")
+	}
+}

apps/backend/internal/agentctl/server/adapter/transport/acp/sync_barrier_test.go

@@ -0,0 +1,49 @@
+package acp
+
+import (
+	"fmt"
+	"testing"
+
+	"github.com/coder/acp-go-sdk"
+	"github.com/kandev/kandev/internal/agentctl/types/streams"
+)
+
+// TestSyncNotifQueue_BlocksUntilQueueDrained verifies the barrier-sync
+// primitive used by sendPrompt before emitting EventTypeComplete: every
+// notification enqueued before syncNotifQueue is processed (and visible on
+// updatesCh) by the time the call returns.
+func TestSyncNotifQueue_BlocksUntilQueueDrained(t *testing.T) {
+	const burst = 50
+
+	a := newTestAdapter()
+	t.Cleanup(func() { _ = a.Close() })
+
+	for i := 0; i < burst; i++ {
+		a.enqueueACPUpdate(acp.SessionNotification{
+			SessionId: acp.SessionId("s1"),
+			Update: acp.SessionUpdate{
+				AvailableCommandsUpdate: &acp.SessionAvailableCommandsUpdate{
+					AvailableCommands: []acp.AvailableCommand{
+						{Name: fmt.Sprintf("cmd-%d", i)},
+					},
+				},
+			},
+		})
+	}
+
+	a.syncNotifQueue()
+
+	events := drainEvents(a)
+	if len(events) != burst {
+		t.Fatalf("expected %d events on updatesCh after sync, got %d", burst, len(events))
+	}
+	for i, ev := range events {
+		if ev.Type != streams.EventTypeAvailableCommands {
+			t.Fatalf("event %d: unexpected type %q", i, ev.Type)
+		}
+		want := fmt.Sprintf("cmd-%d", i)
+		if len(ev.AvailableCommands) != 1 || ev.AvailableCommands[0].Name != want {
+			t.Fatalf("event %d: got %+v, want command %q", i, ev.AvailableCommands, want)
+		}
+	}
+}

apps/backend/internal/agentctl/server/process/workspace_git_cmd.go

@@ -4,6 +4,8 @@ import (
 	"context"
 	"os"
 	"os/exec"
+
+	"github.com/kandev/kandev/internal/common/subproc"
 )
 
 // gitOptionalLocksOff is the env var git reads to skip "optional" locks, i.e.
@@ -26,3 +28,44 @@ func (wt *WorkspaceTracker) pollingGitCommand(ctx context.Context, args ...strin
 	cmd.Env = append(os.Environ(), gitOptionalLocksOff)
 	return cmd
 }
+
+// gitCmdContext builds a per-command timeout-bounded ctx and an exec.Cmd
+// for a git invocation. When polling is true, the cmd is built via
+// pollingGitCommand (which sets GIT_OPTIONAL_LOCKS=0) so the background
+// poll loop doesn't contend with user-initiated git operations on the
+// index lock. The returned cancel must be deferred by the caller to
+// release the timer once the subprocess exits.
+func (wt *WorkspaceTracker) gitCmdContext(ctx context.Context, polling bool, args ...string) (context.Context, context.CancelFunc, *exec.Cmd) {
+	cctx, cancel := context.WithTimeout(ctx, gitCommandTimeout)
+	if polling {
+		return cctx, cancel, wt.pollingGitCommand(cctx, args...)
+	}
+	cmd := exec.CommandContext(cctx, "git", args...)
+	cmd.Dir = wt.workDir
+	return cctx, cancel, cmd
+}
+
+// runGitOutput runs a git command with a per-command timeout and returns its
+// stdout. The derived ctx ensures cancellation SIGKILLs the subprocess and
+// releases its throttle slot even when the outer poll ctx is long-lived.
+func (wt *WorkspaceTracker) runGitOutput(ctx context.Context, args ...string) ([]byte, error) {
+	cctx, cancel, cmd := wt.gitCmdContext(ctx, false, args...)
+	defer cancel()
+	return subproc.RunGitOutput(cctx, cmd)
+}
+
+// runGit is runGitOutput's no-stdout sibling for verify-style probes where
+// only the exit code matters.
+func (wt *WorkspaceTracker) runGit(ctx context.Context, args ...string) error {
+	cctx, cancel, cmd := wt.gitCmdContext(ctx, false, args...)
+	defer cancel()
+	return subproc.RunGit(cctx, cmd)
+}
+
+// runPollingGitOutput is runGitOutput's polling sibling — see
+// gitCmdContext for the GIT_OPTIONAL_LOCKS rationale.
+func (wt *WorkspaceTracker) runPollingGitOutput(ctx context.Context, args ...string) ([]byte, error) {
+	cctx, cancel, cmd := wt.gitCmdContext(ctx, true, args...)
+	defer cancel()
+	return subproc.RunGitOutput(cctx, cmd)
+}

apps/backend/internal/agentctl/server/process/workspace_git_cmd_test.go

@@ -0,0 +1,127 @@
+//go:build !windows
+
+package process
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/kandev/kandev/internal/common/subproc"
+)
+
+// TestRunGitOutput_ReleasesThrottleOnTimeout is the regression test for the
+// "empty Changes panel after freeze" symptom (see workspace_monitor.go
+// gitCommandTimeout). It proves that when a git invocation wedges past
+// gitCommandTimeout, runGitOutput SIGKILLs the subprocess and releases its
+// gitThrottle slot so queued callers can still make progress.
+//
+// Mechanism: install a 'git' shim on $PATH that sleeps for an hour, shrink
+// gitCommandTimeout to 200ms and the throttle cap to 2, then spawn 6
+// goroutines all calling runGitOutput. Without the per-command timeout the
+// 4 queued goroutines would block on Acquire forever (the test would hang
+// past its bound). With the fix each batch of 2 times out, releases, and
+// the next pair Acquires — total ≈ 200ms × 3 batches = 600ms.
+func TestRunGitOutput_ReleasesThrottleOnTimeout(t *testing.T) {
+	prev := gitCommandTimeout
+	gitCommandTimeout = 200 * time.Millisecond
+	t.Cleanup(func() { gitCommandTimeout = prev })
+
+	restoreCap := subproc.Git().SetCapForTest(2)
+	t.Cleanup(restoreCap)
+
+	shimDir := installSleepGitShim(t)
+	t.Setenv("PATH", shimDir+string(os.PathListSeparator)+os.Getenv("PATH"))
+
+	_, wt := setupTestDir(t)
+
+	const N = 6
+	var wg sync.WaitGroup
+	wg.Add(N)
+	done := make(chan struct{})
+	start := time.Now()
+	for i := 0; i < N; i++ {
+		go func() {
+			defer wg.Done()
+			_, _ = wt.runGitOutput(context.Background(), "status")
+		}()
+	}
+	go func() { wg.Wait(); close(done) }()
+
+	select {
+	case <-done:
+		elapsed := time.Since(start)
+		// 6 calls through cap=2 with ~200ms each ⇒ ~600ms. 5s is a generous
+		// upper bound that still catches a leaked slot (which would hang
+		// indefinitely and only be observed via the select timeout below).
+		if elapsed > 5*time.Second {
+			t.Fatalf("runGitOutput took %v for %d calls through cap=2; throttle slot may have leaked", elapsed, N)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("runGitOutput hung past 5s — throttle slot leaked on per-command timeout")
+	}
+}
+
+// TestRunGitOutput_PollingVariantReleasesThrottleOnTimeout mirrors the
+// regression above for runPollingGitOutput, which is the entry point used
+// by the actual workspace poll loop. Same setup, same expectations.
+func TestRunGitOutput_PollingVariantReleasesThrottleOnTimeout(t *testing.T) {
+	prev := gitCommandTimeout
+	gitCommandTimeout = 200 * time.Millisecond
+	t.Cleanup(func() { gitCommandTimeout = prev })
+
+	restoreCap := subproc.Git().SetCapForTest(2)
+	t.Cleanup(restoreCap)
+
+	shimDir := installSleepGitShim(t)
+	t.Setenv("PATH", shimDir+string(os.PathListSeparator)+os.Getenv("PATH"))
+
+	_, wt := setupTestDir(t)
+
+	const N = 6
+	var wg sync.WaitGroup
+	wg.Add(N)
+	done := make(chan struct{})
+	start := time.Now()
+	for i := 0; i < N; i++ {
+		go func() {
+			defer wg.Done()
+			_, _ = wt.runPollingGitOutput(context.Background(), "status")
+		}()
+	}
+	go func() { wg.Wait(); close(done) }()
+
+	select {
+	case <-done:
+		if elapsed := time.Since(start); elapsed > 5*time.Second {
+			t.Fatalf("runPollingGitOutput took %v for %d calls through cap=2; throttle slot may have leaked", elapsed, N)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("runPollingGitOutput hung past 5s — throttle slot leaked on per-command timeout")
+	}
+}
+
+// installSleepGitShim writes an executable shell script named 'git' into a
+// fresh tmpdir that blocks for an hour, and returns the dir. Callers must
+// prepend it to $PATH so exec.Command("git", ...) picks it up before the
+// real binary. The long sleep guarantees that any successful completion
+// must come from the per-command timeout SIGKILLing the subprocess.
+//
+// `exec sleep 3600` (instead of `sleep 3600`) makes sleep replace sh as
+// the process image, so SIGKILL on the exec.Cmd's pid lands directly on
+// sleep. Otherwise sh would die, leave sleep orphaned still holding the
+// stdout/stderr pipes, and cmd.Output() would block waiting for pipe
+// closure for the full hour — making the test flaky.
+func installSleepGitShim(t *testing.T) string {
+	t.Helper()
+	dir := t.TempDir()
+	shim := filepath.Join(dir, "git")
+	const script = "#!/bin/sh\nexec sleep 3600\n"
+	if err := os.WriteFile(shim, []byte(script), 0o755); err != nil {
+		t.Fatalf("write shim: %v", err)
+	}
+	return dir
+}

apps/backend/internal/agentctl/server/process/workspace_git_diff.go

@@ -37,31 +37,43 @@ const (
 	diffSkipReasonBudgetExceeded = "budget_exceeded"
 )
 
-// enrichWithDiffData adds diff information (additions, deletions, diff content) to file info
-func (wt *WorkspaceTracker) enrichWithDiffData(ctx context.Context, update *types.GitStatusUpdate) {
+// enrichWithDiffData adds diff information (additions, deletions, diff content)
+// to file info. The prior status is threaded through so per-file diff data can
+// be carried forward when a git command times out — without it a single bad
+// poll would wipe the diffs panel for files that genuinely still have changes.
+//
+// Carry-forward runs once at the end, after every enrichment phase has had a
+// chance to populate fresh data. Running it inside enrichWithUnstagedDiff would
+// pre-fill fi.Diff for staged-only files, and enrichWithStagedDiff's
+// `if fileInfo.Diff == ""` guard would then skip the fresh staged diff fetch.
+func (wt *WorkspaceTracker) enrichWithDiffData(ctx context.Context, update *types.GitStatusUpdate, prior types.GitStatusUpdate) {
 	// Always diff against HEAD for unstaged/staged content so that files committed
 	// locally (but not yet pushed) show only their uncommitted changes rather than
 	// the entire file as new. The remote branch is only relevant for ahead/behind counts.
-	wt.enrichWithUnstagedDiff(ctx, update, "HEAD")
-	wt.enrichWithStagedDiff(ctx, update, "HEAD")
+	wt.enrichWithUnstagedDiff(ctx, update, "HEAD", prior)
+	wt.enrichWithStagedDiff(ctx, update, "HEAD", prior)
 	wt.enrichUntrackedFileDiffs(ctx, update)
+	carryForwardFileDiffs(update, prior)
 }
 
 // enrichWithBranchDiff computes the total additions/deletions for the entire branch
 // vs the merge-base, covering committed + staged + unstaged changes in one pass.
 // Untracked file line counts (already computed) are added on top.
 // The result is stored in BranchAdditions/BranchDeletions for the sidebar display.
-func (wt *WorkspaceTracker) enrichWithBranchDiff(ctx context.Context, update *types.GitStatusUpdate) {
+//
+// On numstat failure the totals are carried forward from prior when HEAD
+// hasn't moved, mirroring the per-command carry-forward used elsewhere in
+// getGitStatus to avoid a transient git timeout clearing the sidebar count.
+func (wt *WorkspaceTracker) enrichWithBranchDiff(ctx context.Context, update *types.GitStatusUpdate, prior types.GitStatusUpdate) {
 	if update.BaseCommit == "" {
 		return
 	}
 
 	// git diff --numstat <merge-base> covers committed + staged + unstaged changes.
-	numstatCmd := exec.CommandContext(ctx, "git", "diff", "--numstat", update.BaseCommit)
-	numstatCmd.Dir = wt.workDir
-	numstatOut, err := subproc.RunGitOutput(ctx, numstatCmd)
+	numstatOut, err := wt.runGitOutput(ctx, "diff", "--numstat", update.BaseCommit)
 	if err != nil {
-		wt.logger.Debug("enrichWithBranchDiff: numstat failed", zap.Error(err))
+		wt.logger.Debug("enrichWithBranchDiff: numstat failed, carrying forward", zap.Error(err))
+		carryBranchDiff(update, prior)
 		return
 	}
 
@@ -101,6 +113,83 @@ func totalDiffBytes(update *types.GitStatusUpdate) int64 {
 	return total
 }
 
+// carryBranchDiff copies prior.BranchAdditions/BranchDeletions onto update when
+// neither HEAD nor BaseCommit has moved. A new HEAD or a re-pointed merge-base
+// invalidates the cached totals — both are inputs to `git diff --numstat
+// <merge-base>`, so a change in either makes the prior count stale.
+func carryBranchDiff(update *types.GitStatusUpdate, prior types.GitStatusUpdate) {
+	if prior.HeadCommit == "" || prior.HeadCommit != update.HeadCommit {
+		return
+	}
+	if prior.BaseCommit != update.BaseCommit {
+		return
+	}
+	update.BranchAdditions = prior.BranchAdditions
+	update.BranchDeletions = prior.BranchDeletions
+}
+
+// carryForwardFileDiffs copies per-file diff content, additions, deletions, and
+// skip reason from prior.Files into update.Files for files still present in
+// update.Files. Used when a `git diff --numstat` call (the gate that drives
+// per-file diff enrichment) fails: the porcelain status already reported which
+// files changed; we just can't compute fresh diffs this poll, so we keep the
+// last known diff data visible until the next successful poll. HEAD is
+// required to be unchanged so we don't show stale diffs after a reset/commit.
+func carryForwardFileDiffs(update *types.GitStatusUpdate, prior types.GitStatusUpdate) {
+	if prior.HeadCommit == "" || prior.HeadCommit != update.HeadCommit {
+		return
+	}
+	for path, fi := range update.Files {
+		// A non-empty Diff means an earlier enrichment pass already populated
+		// this entry (e.g. unstaged numstat succeeded, then staged numstat
+		// failed). Leave the freshly-computed entry alone — only fill in
+		// completely-missing diff data.
+		if fi.Diff != "" {
+			continue
+		}
+		// A skip reason set this poll (budget_exceeded, too_large, binary)
+		// is an intentional decision to not surface diff content; restoring
+		// a prior diff would defeat that skip and could show stale content
+		// for a file that has since grown past the size or budget cap.
+		if fi.DiffSkipReason != "" {
+			continue
+		}
+		priorFi, ok := prior.Files[path]
+		if !ok || priorFi.Diff == "" {
+			continue
+		}
+		fi.Diff = priorFi.Diff
+		fi.DiffSkipReason = priorFi.DiffSkipReason
+		if fi.Additions == 0 {
+			fi.Additions = priorFi.Additions
+		}
+		if fi.Deletions == 0 {
+			fi.Deletions = priorFi.Deletions
+		}
+		update.Files[path] = fi
+	}
+}
+
+// carryForwardFileDiff fills fi.Diff/DiffSkipReason from prior for a single
+// file when HEAD matches and prior had diff content. Used when numstat
+// succeeded (so Additions/Deletions are fresh) but the per-file capDiffOutput
+// call returned empty — typically a 10s gitCommandTimeout on a large diff,
+// throttle starvation, or a broken pipe. Without this, a single slow file
+// would blank the diff content in the UI even though we have a usable cached
+// copy. Returns the (possibly updated) FileInfo so callers can re-store it.
+func carryForwardFileDiff(fi types.FileInfo, filePath string, update *types.GitStatusUpdate, prior types.GitStatusUpdate) types.FileInfo {
+	if prior.HeadCommit == "" || prior.HeadCommit != update.HeadCommit {
+		return fi
+	}
+	priorFi, ok := prior.Files[filePath]
+	if !ok || priorFi.Diff == "" {
+		return fi
+	}
+	fi.Diff = priorFi.Diff
+	fi.DiffSkipReason = priorFi.DiffSkipReason
+	return fi
+}
+
 // capDiffOutput runs a git diff command and returns at most maxDiffOutputSize bytes.
 // Returns the output string and whether it was truncated.
 //
@@ -110,14 +199,16 @@ func totalDiffBytes(update *types.GitStatusUpdate) int64 {
 // pipe-based reads. Slot is acquired before Start; if Start fails we
 // release immediately, else release runs after Wait.
 func capDiffOutput(ctx context.Context, workDir string, args ...string) (string, bool) {
-	cmd := exec.CommandContext(ctx, "git", args...)
+	cctx, cancel := context.WithTimeout(ctx, gitCommandTimeout)
+	defer cancel()
+	cmd := exec.CommandContext(cctx, "git", args...)
 	cmd.Dir = workDir
 
 	stdout, err := cmd.StdoutPipe()
 	if err != nil {
 		return "", false
 	}
-	release, err := subproc.Git().Acquire(ctx)
+	release, err := subproc.Git().Acquire(cctx)
 	if err != nil {
 		_ = stdout.Close()
 		return "", false
@@ -169,13 +260,17 @@ func resolveNumstatPath(numstatPath string) string {
 }
 
 // enrichWithUnstagedDiff populates additions/deletions and diff content for files
-// with unstaged changes by comparing the worktree against baseRef.
-func (wt *WorkspaceTracker) enrichWithUnstagedDiff(ctx context.Context, update *types.GitStatusUpdate, baseRef string) {
-	numstatCmd := exec.CommandContext(ctx, "git", "diff", "--numstat", baseRef)
-	numstatCmd.Dir = wt.workDir
-	numstatOut, err := subproc.RunGitOutput(ctx, numstatCmd)
+// with unstaged changes by comparing the worktree against baseRef. On numstat
+// failure the function returns early and leaves the carry-forward to
+// enrichWithDiffData — preventing a single git timeout from blanking the diffs
+// panel for files whose porcelain status still shows them as changed, while
+// keeping the staged phase free to populate fresh data for staged-only files.
+func (wt *WorkspaceTracker) enrichWithUnstagedDiff(ctx context.Context, update *types.GitStatusUpdate, baseRef string, prior types.GitStatusUpdate) {
+	numstatOut, err := wt.runGitOutput(ctx, "diff", "--numstat", baseRef)
 	if err != nil {
-		wt.logger.Debug("failed to get numstat", zap.Error(err))
+		// Carry-forward happens once at the end of enrichWithDiffData so the
+		// staged phase can still populate fresh diffs for staged-only files.
+		wt.logger.Debug("enrichWithUnstagedDiff: numstat failed", zap.Error(err))
 		return
 	}
 
@@ -221,6 +316,8 @@ func (wt *WorkspaceTracker) enrichWithUnstagedDiff(ctx context.Context, update *
 			if truncated {
 				fileInfo.DiffSkipReason = diffSkipReasonTruncated
 			}
+		} else {
+			fileInfo = carryForwardFileDiff(fileInfo, filePath, update, prior)
 		}
 
 		update.Files[filePath] = fileInfo
@@ -228,15 +325,18 @@ func (wt *WorkspaceTracker) enrichWithUnstagedDiff(ctx context.Context, update *
 }
 
 // enrichWithStagedDiff populates additions/deletions and diff content for staged files
-// that have no additional unstaged changes, using git diff --cached.
-func (wt *WorkspaceTracker) enrichWithStagedDiff(ctx context.Context, update *types.GitStatusUpdate, baseRef string) {
+// that have no additional unstaged changes, using git diff --cached. On --cached
+// numstat failure the function returns early; carry-forward happens once at the
+// end of enrichWithDiffData (same rationale as enrichWithUnstagedDiff).
+func (wt *WorkspaceTracker) enrichWithStagedDiff(ctx context.Context, update *types.GitStatusUpdate, baseRef string, prior types.GitStatusUpdate) {
 	// For staged files that don't have unstaged changes, we need to get the diff from the index.
 	// The first diff (git diff baseRef) shows worktree vs baseRef, but if a file is staged
 	// and has no additional unstaged changes, its diff won't appear there.
-	stagedCmd := exec.CommandContext(ctx, "git", "diff", "--cached", "--numstat", baseRef)
-	stagedCmd.Dir = wt.workDir
-	stagedOut, err := subproc.RunGitOutput(ctx, stagedCmd)
+	stagedOut, err := wt.runGitOutput(ctx, "diff", "--cached", "--numstat", baseRef)
 	if err != nil {
+		// Carry-forward happens at the end of enrichWithDiffData; running it
+		// here would mask the unstaged phase's fresh data.
+		wt.logger.Debug("enrichWithStagedDiff: --cached numstat failed", zap.Error(err))
 		return
 	}
 
@@ -283,6 +383,8 @@ func (wt *WorkspaceTracker) enrichWithStagedDiff(ctx context.Context, update *ty
 				if truncated {
 					fileInfo.DiffSkipReason = diffSkipReasonTruncated
 				}
+			} else {
+				fileInfo = carryForwardFileDiff(fileInfo, filePath, update, prior)
 			}
 		}
 		update.Files[filePath] = fileInfo

apps/backend/internal/agentctl/server/process/workspace_git_diff_test.go

@@ -273,7 +273,7 @@ func TestEnrichStagedDiff_RenamedFile(t *testing.T) {
 		},
 	}
 
-	wt.enrichWithStagedDiff(context.Background(), update, "HEAD")
+	wt.enrichWithStagedDiff(context.Background(), update, "HEAD", streams.GitStatusUpdate{})
 
 	fi := update.Files["renamed.txt"]
 	if fi.Diff == "" {
@@ -283,3 +283,271 @@ func TestEnrichStagedDiff_RenamedFile(t *testing.T) {
 		t.Error("expected non-zero additions or deletions for renamed+modified file")
 	}
 }
+
+// TestCarryBranchDiff covers carry-forward of branch additions/deletions when
+// the `git diff --numstat <merge-base>` call fails — without it, a transient
+// timeout would clear the sidebar's branch-totals count.
+func TestCarryBranchDiff(t *testing.T) {
+	head := "abc123"
+	base := "base000"
+	tests := []struct {
+		name          string
+		prior         streams.GitStatusUpdate
+		updateHead    string
+		updateBase    string
+		wantAdditions int
+		wantDeletions int
+	}{
+		{
+			name:          "same head and base preserves totals",
+			prior:         streams.GitStatusUpdate{HeadCommit: head, BaseCommit: base, BranchAdditions: 42, BranchDeletions: 7},
+			updateHead:    head,
+			updateBase:    base,
+			wantAdditions: 42,
+			wantDeletions: 7,
+		},
+		{
+			name:          "different head drops totals",
+			prior:         streams.GitStatusUpdate{HeadCommit: head, BaseCommit: base, BranchAdditions: 42, BranchDeletions: 7},
+			updateHead:    "def456",
+			updateBase:    base,
+			wantAdditions: 0,
+			wantDeletions: 0,
+		},
+		{
+			name:          "different base drops totals",
+			prior:         streams.GitStatusUpdate{HeadCommit: head, BaseCommit: base, BranchAdditions: 42, BranchDeletions: 7},
+			updateHead:    head,
+			updateBase:    "newbase",
+			wantAdditions: 0,
+			wantDeletions: 0,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			update := &streams.GitStatusUpdate{HeadCommit: tt.updateHead, BaseCommit: tt.updateBase}
+			carryBranchDiff(update, tt.prior)
+			if update.BranchAdditions != tt.wantAdditions || update.BranchDeletions != tt.wantDeletions {
+				t.Errorf("additions/deletions = %d/%d, want %d/%d",
+					update.BranchAdditions, update.BranchDeletions, tt.wantAdditions, tt.wantDeletions)
+			}
+		})
+	}
+}
+
+// TestCarryForwardFileDiffs covers per-file diff carry-forward. The numstat
+// command is the gate that drives per-file enrichment; when it fails we keep
+// the prior diff content visible (as long as HEAD hasn't moved) instead of
+// clearing the diff panel on a transient timeout.
+func TestCarryForwardFileDiffs(t *testing.T) {
+	head := "abc123"
+	priorDiff := "diff --git a/foo b/foo\n+new line\n"
+
+	t.Run("preserves diff additions deletions when head matches", func(t *testing.T) {
+		prior := streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go", Diff: priorDiff, Additions: 5, Deletions: 2, DiffSkipReason: ""},
+			},
+		}
+		update := &streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go"},
+			},
+		}
+		carryForwardFileDiffs(update, prior)
+		got := update.Files["foo.go"]
+		if got.Diff != priorDiff {
+			t.Errorf("Diff = %q, want %q", got.Diff, priorDiff)
+		}
+		if got.Additions != 5 || got.Deletions != 2 {
+			t.Errorf("Additions/Deletions = %d/%d, want 5/2", got.Additions, got.Deletions)
+		}
+	})
+
+	t.Run("skips when head moved", func(t *testing.T) {
+		prior := streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go", Diff: priorDiff, Additions: 5, Deletions: 2},
+			},
+		}
+		update := &streams.GitStatusUpdate{
+			HeadCommit: "different",
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go"},
+			},
+		}
+		carryForwardFileDiffs(update, prior)
+		got := update.Files["foo.go"]
+		if got.Diff != "" || got.Additions != 0 || got.Deletions != 0 {
+			t.Errorf("expected zeroed FileInfo on head mismatch, got %+v", got)
+		}
+	})
+
+	t.Run("skips files not in prior", func(t *testing.T) {
+		prior := streams.GitStatusUpdate{HeadCommit: head, Files: map[string]streams.FileInfo{}}
+		update := &streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files:      map[string]streams.FileInfo{"new.go": {Path: "new.go"}},
+		}
+		carryForwardFileDiffs(update, prior)
+		got := update.Files["new.go"]
+		if got.Diff != "" {
+			t.Errorf("expected no carry-forward for new file, got Diff=%q", got.Diff)
+		}
+	})
+
+	t.Run("does not overwrite already-populated update entry", func(t *testing.T) {
+		newDiff := "diff --git a/foo b/foo\n+brand new\n"
+		prior := streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go", Diff: priorDiff, Additions: 5, Deletions: 2},
+			},
+		}
+		update := &streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go", Diff: newDiff, Additions: 1, Deletions: 0},
+			},
+		}
+		carryForwardFileDiffs(update, prior)
+		got := update.Files["foo.go"]
+		if got.Diff != newDiff {
+			t.Errorf("Diff overwritten; got %q, want %q", got.Diff, newDiff)
+		}
+		if got.Additions != 1 || got.Deletions != 0 {
+			t.Errorf("counts overwritten; got %d/%d, want 1/0", got.Additions, got.Deletions)
+		}
+	})
+
+	t.Run("respects skip reason set this poll", func(t *testing.T) {
+		skipReasons := []string{
+			diffSkipReasonBudgetExceeded,
+			diffSkipReasonTooLarge,
+			diffSkipReasonBinary,
+		}
+		for _, reason := range skipReasons {
+			t.Run(reason, func(t *testing.T) {
+				prior := streams.GitStatusUpdate{
+					HeadCommit: head,
+					Files: map[string]streams.FileInfo{
+						"foo.go": {Path: "foo.go", Diff: priorDiff, Additions: 5, Deletions: 2},
+					},
+				}
+				update := &streams.GitStatusUpdate{
+					HeadCommit: head,
+					Files: map[string]streams.FileInfo{
+						"foo.go": {Path: "foo.go", DiffSkipReason: reason},
+					},
+				}
+				carryForwardFileDiffs(update, prior)
+				got := update.Files["foo.go"]
+				if got.Diff != "" {
+					t.Errorf("Diff carried forward despite skip reason %q; got %q", reason, got.Diff)
+				}
+				if got.DiffSkipReason != reason {
+					t.Errorf("DiffSkipReason = %q, want %q", got.DiffSkipReason, reason)
+				}
+			})
+		}
+	})
+}
+
+func TestCarryForwardFileDiff(t *testing.T) {
+	head := "abc123"
+	priorDiff := "diff --git a/foo b/foo\n+kept line\n"
+
+	t.Run("fills empty Diff from prior when head matches", func(t *testing.T) {
+		prior := streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go", Diff: priorDiff, DiffSkipReason: "truncated"},
+			},
+		}
+		update := &streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files: map[string]streams.FileInfo{
+				// numstat already populated additions/deletions; capDiffOutput then returned "".
+				"foo.go": {Path: "foo.go", Additions: 7, Deletions: 3},
+			},
+		}
+		fi := update.Files["foo.go"]
+		fi = carryForwardFileDiff(fi, "foo.go", update, prior)
+		if fi.Diff != priorDiff {
+			t.Errorf("Diff = %q, want %q", fi.Diff, priorDiff)
+		}
+		if fi.DiffSkipReason != "truncated" {
+			t.Errorf("DiffSkipReason = %q, want %q", fi.DiffSkipReason, "truncated")
+		}
+		if fi.Additions != 7 || fi.Deletions != 3 {
+			t.Errorf("counts should be untouched, got %d/%d, want 7/3", fi.Additions, fi.Deletions)
+		}
+	})
+
+	t.Run("no-op when head moved", func(t *testing.T) {
+		prior := streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go", Diff: priorDiff},
+			},
+		}
+		update := &streams.GitStatusUpdate{
+			HeadCommit: "different",
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go"},
+			},
+		}
+		fi := carryForwardFileDiff(update.Files["foo.go"], "foo.go", update, prior)
+		if fi.Diff != "" {
+			t.Errorf("expected no carry-forward on head mismatch, got Diff=%q", fi.Diff)
+		}
+	})
+
+	t.Run("no-op when file absent from prior", func(t *testing.T) {
+		prior := streams.GitStatusUpdate{HeadCommit: head, Files: map[string]streams.FileInfo{}}
+		update := &streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files:      map[string]streams.FileInfo{"new.go": {Path: "new.go"}},
+		}
+		fi := carryForwardFileDiff(update.Files["new.go"], "new.go", update, prior)
+		if fi.Diff != "" {
+			t.Errorf("expected no carry-forward for new file, got Diff=%q", fi.Diff)
+		}
+	})
+
+	t.Run("no-op when prior diff empty", func(t *testing.T) {
+		prior := streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go", Additions: 1, Deletions: 1},
+			},
+		}
+		update := &streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files:      map[string]streams.FileInfo{"foo.go": {Path: "foo.go"}},
+		}
+		fi := carryForwardFileDiff(update.Files["foo.go"], "foo.go", update, prior)
+		if fi.Diff != "" || fi.DiffSkipReason != "" {
+			t.Errorf("expected no carry-forward when prior Diff empty, got %+v", fi)
+		}
+	})
+
+	t.Run("no-op when prior head empty", func(t *testing.T) {
+		prior := streams.GitStatusUpdate{
+			Files: map[string]streams.FileInfo{
+				"foo.go": {Path: "foo.go", Diff: priorDiff},
+			},
+		}
+		update := &streams.GitStatusUpdate{
+			HeadCommit: head,
+			Files:      map[string]streams.FileInfo{"foo.go": {Path: "foo.go"}},
+		}
+		fi := carryForwardFileDiff(update.Files["foo.go"], "foo.go", update, prior)
+		if fi.Diff != "" {
+			t.Errorf("expected no carry-forward on empty prior head, got Diff=%q", fi.Diff)
+		}
+	})
+}

apps/backend/internal/agentctl/server/process/workspace_git_status.go

@@ -2,13 +2,11 @@ package process
 
 import (
 	"context"
-	"os/exec"
 	"strconv"
 	"strings"
 	"time"
 
 	"github.com/kandev/kandev/internal/agentctl/types"
-	"github.com/kandev/kandev/internal/common/subproc"
 	"go.uber.org/zap"
 )
 
@@ -76,7 +74,12 @@ func (wt *WorkspaceTracker) GetGitStatus(ctx context.Context, fresh bool) (types
 	return status, nil
 }
 
-// getGitStatus retrieves the current git status
+// getGitStatus retrieves the current git status. Secondary commands
+// (ahead/behind counts, diff enrichment) that fail or time out carry their
+// values forward from the prior cached status rather than overwriting them
+// with zero. The two primary commands (branch info and `git status
+// --porcelain`) still propagate errors upward; on those, updateGitStatus
+// skips the cache write entirely so the prior state stays visible to the UI.
 func (wt *WorkspaceTracker) getGitStatus(ctx context.Context) (types.GitStatusUpdate, error) {
 	update := types.GitStatusUpdate{
 		Timestamp:      time.Now(),
@@ -99,47 +102,50 @@ func (wt *WorkspaceTracker) getGitStatus(ctx context.Context) (types.GitStatusUp
 		return update, nil
 	}
 
+	// Snapshot the prior cache once so per-command carry-forward sees a
+	// consistent view even if a concurrent RefreshGitStatus replaces
+	// currentStatus mid-poll.
+	wt.mu.RLock()
+	prior := wt.currentStatus
+	wt.mu.RUnlock()
+
 	if err := wt.getGitBranchInfo(ctx, &update); err != nil {
 		return update, err
 	}
 
-	wt.getAheadBehindCounts(ctx, &update)
+	wt.getAheadBehindCounts(ctx, &update, prior)
 
 	if err := wt.parseGitStatusOutput(ctx, &update); err != nil {
 		return update, err
 	}
 
 	// Enrich file info with diff data (additions, deletions, and actual diff content)
-	wt.enrichWithDiffData(ctx, &update)
+	wt.enrichWithDiffData(ctx, &update, prior)
 
 	// Compute full branch totals vs merge-base (committed + staged + unstaged + untracked)
-	wt.enrichWithBranchDiff(ctx, &update)
+	wt.enrichWithBranchDiff(ctx, &update, prior)
 
 	return update, nil
 }
 
 // getGitBranchInfo populates branch, remote branch, head commit, and base commit fields.
+// Each command runs under a per-command timeout via the runGit* helpers so a
+// single wedged git invocation cannot pin the shared throttle slot.
 func (wt *WorkspaceTracker) getGitBranchInfo(ctx context.Context, update *types.GitStatusUpdate) error {
 	// Get current branch
-	branchCmd := exec.CommandContext(ctx, "git", "rev-parse", "--abbrev-ref", "HEAD")
-	branchCmd.Dir = wt.workDir
-	branchOut, err := subproc.RunGitOutput(ctx, branchCmd)
+	branchOut, err := wt.runGitOutput(ctx, "rev-parse", "--abbrev-ref", "HEAD")
 	if err != nil {
 		return err
 	}
 	update.Branch = strings.TrimSpace(string(branchOut))
 
 	// Get remote branch
-	remoteCmd := exec.CommandContext(ctx, "git", "rev-parse", "--abbrev-ref", "@{upstream}")
-	remoteCmd.Dir = wt.workDir
-	if remoteOut, err := subproc.RunGitOutput(ctx, remoteCmd); err == nil {
+	if remoteOut, err := wt.runGitOutput(ctx, "rev-parse", "--abbrev-ref", "@{upstream}"); err == nil {
 		update.RemoteBranch = strings.TrimSpace(string(remoteOut))
 	}
 
 	// Get current HEAD commit SHA
-	headCmd := exec.CommandContext(ctx, "git", "rev-parse", "HEAD")
-	headCmd.Dir = wt.workDir
-	if headOut, err := subproc.RunGitOutput(ctx, headCmd); err == nil {
+	if headOut, err := wt.runGitOutput(ctx, "rev-parse", "HEAD"); err == nil {
 		update.HeadCommit = strings.TrimSpace(string(headOut))
 	}
 
@@ -148,9 +154,7 @@ func (wt *WorkspaceTracker) getGitBranchInfo(ctx context.Context, update *types.
 	// so we show all changes this branch introduces compared to the main development line.
 	var baseBranch string
 	for _, candidate := range []string{"origin/main", "origin/master", "main", "master"} {
-		checkCmd := exec.CommandContext(ctx, "git", "rev-parse", "--verify", candidate)
-		checkCmd.Dir = wt.workDir
-		if err := subproc.RunGit(ctx, checkCmd); err == nil {
+		if err := wt.runGit(ctx, "rev-parse", "--verify", candidate); err == nil {
 			baseBranch = candidate
 			break
 		}
@@ -158,9 +162,7 @@ func (wt *WorkspaceTracker) getGitBranchInfo(ctx context.Context, update *types.
 	if baseBranch != "" {
 		// Use merge-base to find common ancestor between current branch and base branch.
 		// This is correct even when the branch has diverged from main.
-		mergeBaseCmd := exec.CommandContext(ctx, "git", "merge-base", baseBranch, "HEAD")
-		mergeBaseCmd.Dir = wt.workDir
-		if mergeBaseOut, err := subproc.RunGitOutput(ctx, mergeBaseCmd); err == nil {
+		if mergeBaseOut, err := wt.runGitOutput(ctx, "merge-base", baseBranch, "HEAD"); err == nil {
 			update.BaseCommit = strings.TrimSpace(string(mergeBaseOut))
 		}
 	}
@@ -168,44 +170,63 @@ func (wt *WorkspaceTracker) getGitBranchInfo(ctx context.Context, update *types.
 	return nil
 }
 
-// getAheadBehindCounts populates the Ahead/Behind fields relative to the base branch
-// (origin/main or origin/master). Always compares against the base branch rather than
-// the remote tracking branch, because after a rebase the tracking branch has stale
-// commit SHAs that produce inflated counts.
-func (wt *WorkspaceTracker) getAheadBehindCounts(ctx context.Context, update *types.GitStatusUpdate) {
+// getAheadBehindCounts populates the Ahead/Behind fields relative to the base
+// branch (origin/main or origin/master). Always compares against the base
+// branch rather than the remote tracking branch, because after a rebase the
+// tracking branch has stale commit SHAs that produce inflated counts.
+//
+// If either the compareRef lookup or the count command fails (e.g. timeout
+// under throttle saturation, transient FS hang), Ahead/Behind are carried
+// forward from prior when HEAD hasn't moved. Otherwise a single bad poll
+// would silently overwrite the cached counts with 0/0 and hide a legitimate
+// "Pull N" / "Push N" indicator in the UI.
+func (wt *WorkspaceTracker) getAheadBehindCounts(ctx context.Context, update *types.GitStatusUpdate, prior types.GitStatusUpdate) {
 	// Always compare against the base branch (origin/main or origin/master).
 	// Using the remote tracking branch (origin/<feature-branch>) gives wrong counts
 	// after rebase because rebased commits have new SHAs.
 	var compareRef string
 	for _, candidate := range []string{"origin/main", "origin/master"} {
-		checkCmd := exec.CommandContext(ctx, "git", "rev-parse", "--verify", candidate)
-		checkCmd.Dir = wt.workDir
-		if err := subproc.RunGit(ctx, checkCmd); err == nil {
+		if err := wt.runGit(ctx, "rev-parse", "--verify", candidate); err == nil {
 			compareRef = candidate
 			break
 		}
 	}
 	if compareRef == "" {
+		carryAheadBehind(update, prior)
 		return
 	}
-	countCmd := exec.CommandContext(ctx, "git", "rev-list", "--left-right", "--count", update.Branch+"..."+compareRef)
-	countCmd.Dir = wt.workDir
-	if countOut, err := subproc.RunGitOutput(ctx, countCmd); err == nil {
-		parts := strings.Fields(string(countOut))
-		if len(parts) == 2 {
-			update.Ahead, _ = strconv.Atoi(parts[0])
-			update.Behind, _ = strconv.Atoi(parts[1])
-		}
+	countOut, err := wt.runGitOutput(ctx, "rev-list", "--left-right", "--count", update.Branch+"..."+compareRef)
+	if err != nil {
+		wt.logger.Debug("getAheadBehindCounts: rev-list failed, carrying forward", zap.Error(err))
+		carryAheadBehind(update, prior)
+		return
 	}
+	parts := strings.Fields(string(countOut))
+	if len(parts) != 2 {
+		carryAheadBehind(update, prior)
+		return
+	}
+	update.Ahead, _ = strconv.Atoi(parts[0])
+	update.Behind, _ = strconv.Atoi(parts[1])
+}
+
+// carryAheadBehind copies prior.Ahead/Behind onto update when HEAD hasn't
+// moved. A new HEAD invalidates prior counts (the user committed, pulled, or
+// reset), so we leave update zeroed in that case.
+func carryAheadBehind(update *types.GitStatusUpdate, prior types.GitStatusUpdate) {
+	if prior.HeadCommit == "" || prior.HeadCommit != update.HeadCommit {
+		return
+	}
+	update.Ahead = prior.Ahead
+	update.Behind = prior.Behind
 }
 
 // parseGitStatusOutput runs git status --porcelain and populates the file lists and map.
 func (wt *WorkspaceTracker) parseGitStatusOutput(ctx context.Context, update *types.GitStatusUpdate) error {
 	// --untracked-files=all shows all files in untracked directories, not just the directory name.
-	// GIT_OPTIONAL_LOCKS=0 (via pollingGitCommand) prevents the background poll loop from taking
-	// .git/index.lock, which would race with concurrent user-initiated git operations.
-	statusCmd := wt.pollingGitCommand(ctx, "status", "--porcelain", "--untracked-files=all")
-	statusOut, err := subproc.RunGitOutput(ctx, statusCmd)
+	// GIT_OPTIONAL_LOCKS=0 (via runPollingGitOutput) prevents the background poll loop from
+	// taking .git/index.lock, which would race with concurrent user-initiated git operations.
+	statusOut, err := wt.runPollingGitOutput(ctx, "status", "--porcelain", "--untracked-files=all")
 	if err != nil {
 		return err
 	}

apps/backend/internal/agentctl/server/process/workspace_git_status_test.go

@@ -5,6 +5,8 @@ import (
 	"os"
 	"path/filepath"
 	"testing"
+
+	"github.com/kandev/kandev/internal/agentctl/types"
 )
 
 func TestUnquoteGitPath(t *testing.T) {
@@ -208,3 +210,49 @@ func mapKeys[V any](m map[string]V) []string {
 	}
 	return keys
 }
+
+// TestCarryAheadBehind covers the carry-forward fallback used when the
+// ahead/behind git command fails (timeout, missing upstream). The contract:
+// preserve prior counts when HEAD is unchanged, leave them zero when HEAD
+// moved (the prior counts are stale by definition) or when prior is empty.
+func TestCarryAheadBehind(t *testing.T) {
+	head := "abc123"
+	tests := []struct {
+		name       string
+		prior      types.GitStatusUpdate
+		updateHead string
+		wantAhead  int
+		wantBehind int
+	}{
+		{
+			name:       "same head preserves counts",
+			prior:      types.GitStatusUpdate{HeadCommit: head, Ahead: 1, Behind: 3},
+			updateHead: head,
+			wantAhead:  1,
+			wantBehind: 3,
+		},
+		{
+			name:       "different head drops counts",
+			prior:      types.GitStatusUpdate{HeadCommit: head, Ahead: 1, Behind: 3},
+			updateHead: "def456",
+			wantAhead:  0,
+			wantBehind: 0,
+		},
+		{
+			name:       "empty prior head no-op",
+			prior:      types.GitStatusUpdate{Ahead: 9, Behind: 9},
+			updateHead: head,
+			wantAhead:  0,
+			wantBehind: 0,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			update := &types.GitStatusUpdate{HeadCommit: tt.updateHead}
+			carryAheadBehind(update, tt.prior)
+			if update.Ahead != tt.wantAhead || update.Behind != tt.wantBehind {
+				t.Errorf("ahead/behind = %d/%d, want %d/%d", update.Ahead, update.Behind, tt.wantAhead, tt.wantBehind)
+			}
+		})
+	}
+}

apps/backend/internal/agentctl/server/process/workspace_monitor.go

@@ -20,9 +20,19 @@ import (
 // causing "too many open files" errors in large workspaces.
 const DefaultFilePollInterval = 2 * time.Second
 
-// gitCommandTimeout is the maximum time to wait for git commands during polling.
-// This prevents the monitor loop from hanging if git is blocked (e.g., index lock).
-const gitCommandTimeout = 10 * time.Second
+// gitCommandTimeout is the maximum time to wait for any single git subprocess
+// the workspace tracker spawns (monitor loop quick-status, full status refresh,
+// diff enrichment). Without a per-command bound, a wedged invocation (FS hang,
+// index lock, EDR scan) would inherit only the tracker's lifetime ctx and pin
+// its gitThrottle slot for the lifetime of the session, blocking subsequent
+// poll cycles and producing the "empty changes panel after freeze" symptom.
+// 10s is generous for normal completion (typical poll commands finish in
+// <100ms) while still freeing the slot in time for the next poll cycle.
+//
+// A var (not const) so the regression test in workspace_git_cmd_test.go can
+// shrink it to milliseconds and prove that timed-out invocations release
+// their gitThrottle slot. Production code must not mutate it.
+var gitCommandTimeout = 10 * time.Second
 
 // maxConsecutiveGitFailures is the number of consecutive git command failures
 // before the polling loop gives up and stops. At 2-second intervals, 5 failures