fix(chat): simplify tool phase detection and fix stale cache reply leak

- Revert hasCompletedToolPhase to simple check (segmentHasTools only).
  The lastAssistantHasNoTools guard was too restrictive: during reply
  streaming the last assistant in history still has tool_use (reply only
  exists in streamingMessage). The intermediate-narration edge case is
  already handled by stripProcessMessagePrefix producing empty
  trimmedReplyText, causing graceful fallback to buildSteps(false).

- Fix stale graph cache: filter out stream-generated message steps
  (id prefix 'stream-message') instead of brittle exact-match. These
  steps contain accumulated narration+reply text from streaming phase
  that should not persist after completion.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/pages/Chat/index.tsx

@@ -278,19 +278,15 @@ export function Chat() {
     // Three signals indicate "tools finished, now streaming the reply":
     //   1. `pendingFinal`        — set by tool-result final events
     //   2. `allToolsCompleted`   — all entries in streamingTools are completed
-    //   3. `hasCompletedToolPhase` — the Gateway executed tools server-side
+    //   3. `hasCompletedToolPhase` — historical messages (loaded by the poll)
-    //      (no streaming tool events) and the tool phase is over.  Detected
+    //      contain tool_use blocks, meaning the Gateway executed tools
-    //      by finding tool_use in historical messages AND the last assistant
+    //      server-side without sending streaming tool events to the client.
-    //      message in the segment having no tool_use blocks (meaning the
+    //      During intermediate narration (before reply), stripProcessMessagePrefix
-    //      model has moved past tool calls into the reply phase).
+    //      will produce an empty trimmedReplyText, so the graph stays active.
     const allToolsCompleted = streamingTools.length > 0 && !hasRunningStreamToolStatus;
-    const lastAssistantInSegment = [...segmentMessages].reverse().find((m) => m.role === 'assistant');
+    const hasCompletedToolPhase = segmentMessages.some((msg) =>
-    const segmentHasTools = segmentMessages.some((msg) =>
       msg.role === 'assistant' && extractToolUse(msg).length > 0,
     );
-    const lastAssistantHasNoTools = lastAssistantInSegment != null
-      && extractToolUse(lastAssistantInSegment).length === 0;
-    const hasCompletedToolPhase = segmentHasTools && lastAssistantHasNoTools;
     const rawStreamingReplyCandidate = isLatestOpenRun
       && (pendingFinal || allToolsCompleted || hasCompletedToolPhase)
       && (hasStreamText || hasStreamImages)
@@ -329,16 +325,13 @@ export function Chat() {
       }
       const cached = graphStepCache[runKey];
       if (!cached) return [];
-      // When using cached steps for a completed run, filter out message
+      // The cache was captured during streaming and may contain stream-
-      // steps whose text matches the final reply.  The cache was captured
+      // generated message steps that include accumulated narration + reply
-      // during streaming when all text was narration; now that the run is
+      // text.  Strip these out — historical message steps (from messages[])
-      // complete the reply should not appear inside the graph.
+      // will be properly recomputed on the next render with fresh data.
-      const cachedReplyIdx = cached.replyIndex;
+      const cleanedSteps = cached.steps.filter(
-      const replyMsg = cachedReplyIdx != null ? messages[cachedReplyIdx] : null;
+        (s) => !(s.kind === 'message' && s.id.startsWith('stream-message')),
-      const replyText = replyMsg ? extractText(replyMsg).trim() : '';
+      );
-      const cleanedSteps = replyText
-        ? cached.steps.filter((s) => !(s.kind === 'message' && s.detail?.trim() === replyText))
-        : cached.steps;
       return [{
         triggerIndex: idx,
         replyIndex: cached.replyIndex,