Show per-request tokens in the visualizer

openhands/sdk/conversation/visualizer.py

@@ -234,14 +234,13 @@ def _format_metrics_subtitle(self) -> str | None:
             return None
 
         combined_metrics = self._conversation_stats.get_combined_metrics()
-        if not combined_metrics or not combined_metrics.accumulated_token_usage:
+        if not combined_metrics or not combined_metrics.token_usages:
             return None
 
-        usage = combined_metrics.accumulated_token_usage
+        latest_usage = combined_metrics.token_usages[-1]
         cost = combined_metrics.accumulated_cost or 0.0
 
-        # helper: 1234 -> "1.2K", 1200000 -> "1.2M"
-        def abbr(n: int | float) -> str:
+        def abbr(n: int | float) -> str:  # helper: 1234 -> "1.2K", 1200000 -> "1.2M"
             n = int(n or 0)
             if n >= 1_000_000_000:
                 s = f"{n / 1_000_000_000:.2f}B"
@@ -253,26 +252,23 @@ def abbr(n: int | float) -> str:
                 return str(n)
             return s.replace(".0", "")
 
-        input_tokens = abbr(usage.prompt_tokens or 0)
-        output_tokens = abbr(usage.completion_tokens or 0)
+        input_tokens = latest_usage.prompt_tokens or 0
+        output_tokens = latest_usage.completion_tokens or 0
+        cache_read = latest_usage.cache_read_tokens or 0
+        cache_rate = (
+            f"{(cache_read / input_tokens * 100):.2f}%" if input_tokens > 0 else "N/A"
+        )
+        reasoning_tokens = latest_usage.reasoning_tokens or 0
 
-        # Cache hit rate (prompt + cache)
-        prompt = usage.prompt_tokens or 0
-        cache_read = usage.cache_read_tokens or 0
-        cache_rate = f"{(cache_read / prompt * 100):.2f}%" if prompt > 0 else "N/A"
-        reasoning_tokens = usage.reasoning_tokens or 0
+        cost_str = f"{cost:.4f}"
 
-        # Cost
-        cost_str = f"{cost:.4f}" if cost > 0 else "$0.00"
-
-        # Build with fixed color scheme
         parts: list[str] = []
-        parts.append(f"[cyan]↑ input {input_tokens}[/cyan]")
+        parts.append(f"[cyan]↑ input {abbr(input_tokens)}[/cyan]")
         parts.append(f"[magenta]cache hit {cache_rate}[/magenta]")
         if reasoning_tokens > 0:
             parts.append(f"[yellow] reasoning {abbr(reasoning_tokens)}[/yellow]")
-        parts.append(f"[blue]↓ output {output_tokens}[/blue]")
-        parts.append(f"[green]$ {cost_str}[/green]")
+        parts.append(f"[blue]↓ output {abbr(output_tokens)}[/blue]")
+        parts.append(f"[green]$ {cost_str} (total)[/green]")
 
         return "Tokens: " + " • ".join(parts)
 

tests/sdk/conversation/test_visualizer.py

@@ -337,7 +337,47 @@ def test_metrics_formatting():
     assert "500" in subtitle  # Output tokens
     assert "20.00%" in subtitle  # Cache hit rate
     assert "200" in subtitle  # Reasoning tokens
-    assert "0.0234" in subtitle  # Cost
+    assert "$ 0.0234 (total)" in subtitle
+
+
+def test_metrics_formatting_uses_latest_request():
+    """Tokens should reflect the latest request while cost stays cumulative."""
+    from openhands.sdk.conversation.conversation_stats import ConversationStats
+    from openhands.sdk.llm.utils.metrics import Metrics
+
+    conversation_stats = ConversationStats()
+    metrics = Metrics(model_name="test-model")
+    metrics.add_cost(0.1)
+    metrics.add_token_usage(
+        prompt_tokens=120,
+        completion_tokens=40,
+        cache_read_tokens=12,
+        cache_write_tokens=0,
+        reasoning_tokens=5,
+        context_window=8000,
+        response_id="first",
+    )
+    metrics.add_cost(0.05)
+    metrics.add_token_usage(
+        prompt_tokens=200,
+        completion_tokens=75,
+        cache_read_tokens=25,
+        cache_write_tokens=0,
+        reasoning_tokens=0,
+        context_window=8000,
+        response_id="second",
+    )
+    conversation_stats.service_to_metrics["test_service"] = metrics
+
+    visualizer = ConversationVisualizer(conversation_stats=conversation_stats)
+
+    subtitle = visualizer._format_metrics_subtitle()
+    assert subtitle is not None
+    assert "input 200" in subtitle
+    assert "output 75" in subtitle
+    assert "cache hit 10.00%" not in subtitle  # ensure using latest cache values
+    assert "cache hit 12.50%" in subtitle
+    assert "$ 0.1500 (total)" in subtitle
 
 
 def test_event_base_fallback_visualize():