diff --git a/pkg/agent/loop.go b/pkg/agent/loop.go
index 59dc93e..f73d882 100644
--- a/pkg/agent/loop.go
+++ b/pkg/agent/loop.go
@@ -47,19 +47,19 @@ type Config struct {
 
 // Iteration represents a single cycle of the agent loop.
 type Iteration struct {
-	Number     int
-	ToolCalls  []llm.ToolCall
-	ToolsUsed  int
-	Duration   time.Duration
+	Number    int
+	ToolCalls []llm.ToolCall
+	ToolsUsed int
+	Duration  time.Duration
 }
 
 // Response is the final output of the agent loop.
 type Response struct {
-	Content     string
-	ToolCalls   []llm.ToolCall
-	Iterations  int
-	Duration    time.Duration
-	TokenUsage  llm.TokenUsage
+	Content    string
+	ToolCalls  []llm.ToolCall
+	Iterations int
+	Duration   time.Duration
+	TokenUsage llm.TokenUsage
 }
 
 // Loop is the main agent loop that orchestrates LLM calls and tool execution.
@@ -82,17 +82,21 @@ func (l *Loop) Run(ctx context.Context, input string, history ...llm.Message) (R
 	start := time.Now()
 
 	messages := l.buildInitialMessages(input, history)
+	// Tool schemas don't change between iterations, so build the JSON once
+	// per Run instead of re-marshaling every tool on every loop pass.
+	toolSchemas := l.getToolSchemas()
 	var finalContent string
 	var allToolCalls []llm.ToolCall
 	var totalUsage llm.TokenUsage
 	iterations := 0
+	completed := false
 
 	for iterations < l.cfg.MaxIters {
 		iterations++
 
 		resp, err := l.cfg.LLM.Generate(ctx, llm.CompletionRequest{
 			Messages:           messages,
-			Tools:              l.getToolSchemas(),
+			Tools:              toolSchemas,
 			ChatTemplateKwargs: l.cfg.ChatTemplateKwargs,
 		})
 		if err != nil {
@@ -105,6 +109,7 @@ func (l *Loop) Run(ctx context.Context, input string, history ...llm.Message) (R
 
 		if len(resp.ToolCalls) == 0 {
 			finalContent = resp.Content
+			completed = true
 			break
 		}
 
@@ -138,16 +143,20 @@ func (l *Loop) Run(ctx context.Context, input string, history ...llm.Message) (R
 	}
 
 	duration := time.Since(start)
-	if iterations >= l.cfg.MaxIters {
+	// Only report the max-iterations failure when the loop actually ran out
+	// of budget without producing a final answer — an answer that arrives
+	// exactly on the last allowed iteration is still a success (the old
+	// `iterations >= MaxIters` check threw that valid response away).
+	if !completed {
 		return Response{}, fmt.Errorf("max iterations (%d) reached", l.cfg.MaxIters)
 	}
 
 	return Response{
-		Content:     finalContent,
-		ToolCalls:   allToolCalls,
-		Iterations:  iterations,
-		Duration:    duration,
-		TokenUsage:  totalUsage,
+		Content:    finalContent,
+		ToolCalls:  allToolCalls,
+		Iterations: iterations,
+		Duration:   duration,
+		TokenUsage: totalUsage,
 	}, nil
 }
 
@@ -157,6 +166,8 @@ func (l *Loop) Run(ctx context.Context, input string, history ...llm.Message) (R
 func (l *Loop) RunStream(ctx context.Context, input string, history ...llm.Message) iter.Seq2[llm.StreamChunk, error] {
 	return func(yield func(llm.StreamChunk, error) bool) {
 		messages := l.buildInitialMessages(input, history)
+		// Same as Run: the schemas are identical on every iteration.
+		toolSchemas := l.getToolSchemas()
 		iterations := 0
 
 		for iterations < l.cfg.MaxIters {
@@ -164,7 +175,7 @@ func (l *Loop) RunStream(ctx context.Context, input string, history ...llm.Messa
 
 			stream := l.cfg.LLM.Stream(ctx, llm.CompletionRequest{
 				Messages:           messages,
-				Tools:              l.getToolSchemas(),
+				Tools:              toolSchemas,
 				ChatTemplateKwargs: l.cfg.ChatTemplateKwargs,
 			})
 
@@ -226,11 +237,23 @@ func (l *Loop) RunStream(ctx context.Context, input string, history ...llm.Messa
 				// real token counts.
 				hasContent := chunk.Delta != "" || chunk.ReasoningDelta != ""
 				hasUsage := chunk.Usage.TotalTokens > 0
-				if !hasToolCalls && (hasContent || hasUsage) {
+				switch {
+				case !hasToolCalls && (hasContent || hasUsage):
 					responseBuilder.WriteString(chunk.Delta)
 					if !yield(chunk, nil) {
 						return
 					}
+				case hasToolCalls && hasUsage:
+					// The content gate above exists to hide raw provider
+					// deltas during a tool-call round, but it also swallowed
+					// that round's token usage — so callers tracking context
+					// occupancy (e.g. a UI's context bar deciding when to
+					// compact) only ever saw the usage of the final,
+					// tool-free round. Forward the usage on its own,
+					// without the content.
+					if !yield(llm.StreamChunk{Usage: chunk.Usage}, nil) {
+						return
+					}
 				}
 			}