fix: 速度胶囊每轮只显示一次 + 技能调用按 run 合并统计

速度胶囊：原本每条 AI 消息都渲染，agentic 多轮 LLM 调用会出现多个胶囊。改为只在每轮 turn-final 消息（下一条 human 之前的最后一条 AI 消息）显示。

技能统计：tool_call_metrics 改为按 (run_id, skill_name) upsert 合并——同一 run 内同一技能的多次调用/重试合并为一行；任一成功即记成功，全部失败则记一行 error 并累积所有报错信息；非技能调用不受影响。

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

frontend-web/src/components/workspace/messages/message-token-usage.tsx

@@ -88,7 +88,13 @@ export function MessageTokenUsage({
   // otherwise wait until usage arrives so the bar doesn't render on its own.
   const usage = getUsageMetadata(message);
   const showUsage = enabled && !isLoading && Boolean(usage);
-  const showSpeed = speedEnabled && resolvedMetric !== undefined && resolvedMetric.durationMs > 0;
+  // The speed pill is shown once per turn — only on the turn-final AI message
+  // — so a turn with several LLM calls no longer renders a pill per call.
+  const showSpeed =
+    speedEnabled &&
+    resolvedMetric !== undefined &&
+    resolvedMetric.durationMs > 0 &&
+    resolvedMetric.isTurnFinal;
 
   if (!showUsage && !showSpeed) {
     return null;

frontend-web/src/components/workspace/messages/use-message-metrics.ts

@@ -36,6 +36,13 @@ export interface MessageMetric {
    * only to add a "~" prefix when the value is approximate.
    */
   fromUsage: boolean;
+  /**
+   * True iff this is the turn-final AI message — the last AI message before
+   * the next human message. An agentic turn makes several LLM calls, each its
+   * own AI message; the speed pill is shown only on the turn-final one so a
+   * turn renders a single pill instead of one per call.
+   */
+  isTurnFinal: boolean;
 }
 
 type Sample = { at: number; chars: number };
@@ -161,6 +168,25 @@ export function useMessageMetrics(
     return () => window.clearInterval(interval);
   }, [isLoading]);
 
+  // Identify the turn-final AI message of each turn — the last AI message
+  // (with a recorded metric) before the next human message. Only these show
+  // the speed pill, collapsing a multi-LLM-call turn to a single pill.
+  const turnFinalIds = new Set<string>();
+  let pendingTurnFinal: string | null = null;
+  for (let i = 0; i < messages.length; i++) {
+    const message = messages[i];
+    if (!message) continue;
+    if (message.type === "human") {
+      if (pendingTurnFinal) turnFinalIds.add(pendingTurnFinal);
+      pendingTurnFinal = null;
+      continue;
+    }
+    if (isAiMessage(message) && timingRef.current.has(getMessageId(message, i))) {
+      pendingTurnFinal = getMessageId(message, i);
+    }
+  }
+  if (pendingTurnFinal) turnFinalIds.add(pendingTurnFinal);
+
   // Build the snapshot map. We compute on every render so consumers see the
   // freshest duration; the underlying timing storage is the source of truth.
   const out = new Map<string, MessageMetric>();
@@ -212,6 +238,7 @@ export function useMessageMetrics(
       durationMs,
       isActive,
       fromUsage,
+      isTurnFinal: turnFinalIds.has(id),
     });
   }