diff --git a/cmd/olliesrv/internal/agent/cancellation_test.go b/cmd/olliesrv/internal/agent/cancellation_test.go
index 8c77b5f..e93e562 100644
--- a/cmd/olliesrv/internal/agent/cancellation_test.go
+++ b/cmd/olliesrv/internal/agent/cancellation_test.go
@@ -61,7 +61,7 @@ func TestStripColdCancellation(t *testing.T) {
 	done := make(chan struct{})
 
 	go func() {
-		h.stripCold(ctx, b)
+		_, _ = h.stripCold(ctx, b)
 		close(done)
 	}()
 
@@ -99,7 +99,7 @@ func TestPreRunCompactionCancellation(t *testing.T) {
 			Backend: b,
 		},
 		history: newHistory("test"),
-		log: noOpLogger,
+		log:     noOpLogger,
 	}
 
 	// Add enough messages to exceed the compact threshold.
diff --git a/cmd/olliesrv/internal/agent/history.go b/cmd/olliesrv/internal/agent/history.go
index 6f1653c..b7afdc4 100644
--- a/cmd/olliesrv/internal/agent/history.go
+++ b/cmd/olliesrv/internal/agent/history.go
@@ -381,10 +381,10 @@ func (s *History) estimateTokens() int {
 // stripCold summarizes large tool-result messages outside the hot tail
 // using a single batched LLM call. Messages in the last hotTailSize slots
 // are left verbatim.
-func (s *History) stripCold(ctx context.Context, b backend.Backend) {
+func (s *History) stripCold(ctx context.Context, b backend.Backend) (backend.Usage, bool) {
 	hot := len(s.messages) - hotTailSize
 	if hot < 0 {
-		return
+		return backend.Usage{}, false
 	}
 
 	// Collect cold-zone tool results exceeding 200 chars.
@@ -404,7 +404,7 @@ func (s *History) stripCold(ctx context.Context, b backend.Backend) {
 		}
 	}
 	if len(items) == 0 {
-		return
+		return backend.Usage{}, false
 	}
 
 	// Build a single prompt listing each item.
@@ -416,17 +416,27 @@ func (s *History) stripCold(ctx context.Context, b backend.Backend) {
 
 	ch, err := b.ChatStream(ctx, []backend.Message{{Role: "user", Content: sb.String()}}, nil, backend.GenerationParams{})
 	if err != nil {
-		return
+		return backend.Usage{}, false
 	}
 	var resp strings.Builder
+	var usage backend.Usage
+	var reported bool
 	for ev := range ch {
 		if ev.Content != "" {
 			resp.WriteString(ev.Content)
 		}
 		if ev.Done {
+			usage = ev.Usage
+			reported = true
 			break
 		}
 	}
+	if !reported {
+		usage.InputTokens = len(sb.String()) / 4
+		usage.OutputTokens = resp.Len() / 4
+	}
+	estimated := !reported
+
 	summaries := make(map[string]string)
 	json.Unmarshal([]byte(resp.String()), &summaries)
 
@@ -437,6 +447,7 @@ func (s *History) stripCold(ctx context.Context, b backend.Backend) {
 			s.messages[it.idx].Content = summary
 		}
 	}
+	return usage, estimated
 }
 
 // resolveCompactionModel returns the model to use for compaction.
diff --git a/cmd/olliesrv/internal/agent/turn.go b/cmd/olliesrv/internal/agent/turn.go
index d2097c0..a002d85 100644
--- a/cmd/olliesrv/internal/agent/turn.go
+++ b/cmd/olliesrv/internal/agent/turn.go
@@ -204,7 +204,11 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 
 	// Strip cold-zone tool results (one batched LLM call per turn).
 	if ag.history != nil {
-		ag.history.stripCold(actCtx, ag.runtime.Backend)
+		usage, estimated := ag.history.stripCold(actCtx, ag.runtime.Backend)
+		if usage.InputTokens > 0 || usage.OutputTokens > 0 {
+			ag.history.addUsage(usage, estimated)
+			ag.notifyChange()
+		}
 	}
 
 	// Inject the preamble system message into history before running.
@@ -257,6 +261,9 @@ func (ag *Agent) executeTurn(ctx context.Context, input string) string {
 	ag.flush()
 
 	if err != nil {
+		if ag.history != nil {
+			ag.history.recordTurnCost(ag.runtime.Backend.Model())
+		}
 		// Keep completed work — only remove cancelled tool results.
 		if ag.history != nil {
 			ag.history.removeCancelledToolResults()