fix(chat): refine tool phase detection and clean stale reply from graph cache

- hasCompletedToolPhase now checks that the last assistant message in the
  segment has no tool_use blocks, preventing false positives during
  intermediate tool rounds that would suppress the trailing thinking indicator
- Filter reply text from cached graph steps when a completed run falls
  back to the step cache, preventing the final response from appearing
  inside the graph when expanding after completion
- Remove debug logging

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

src/pages/Chat/index.tsx

@@ -278,13 +278,19 @@ 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` — historical messages (loaded by the poll)
+    //   3. `hasCompletedToolPhase` — the Gateway executed tools server-side
-    //      contain tool_use blocks, meaning the Gateway executed tools
+    //      (no streaming tool events) and the tool phase is over.  Detected
-    //      server-side without sending streaming tool events to the client
+    //      by finding tool_use in historical messages AND the last assistant
+    //      message in the segment having no tool_use blocks (meaning the
+    //      model has moved past tool calls into the reply phase).
     const allToolsCompleted = streamingTools.length > 0 && !hasRunningStreamToolStatus;
-    const hasCompletedToolPhase = segmentMessages.some((msg) =>
+    const lastAssistantInSegment = [...segmentMessages].reverse().find((m) => m.role === 'assistant');
+    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)
@@ -323,6 +329,16 @@ export function Chat() {
       }
       const cached = graphStepCache[runKey];
       if (!cached) return [];
+      // When using cached steps for a completed run, filter out message
+      // steps whose text matches the final reply.  The cache was captured
+      // during streaming when all text was narration; now that the run is
+      // complete the reply should not appear inside the graph.
+      const cachedReplyIdx = cached.replyIndex;
+      const replyMsg = cachedReplyIdx != null ? messages[cachedReplyIdx] : null;
+      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,
@@ -330,8 +346,8 @@ export function Chat() {
         agentLabel: cached.agentLabel,
         sessionLabel: cached.sessionLabel,
         segmentEnd: nextUserIndex === -1 ? messages.length - 1 : nextUserIndex - 1,
-        steps: cached.steps,
+        steps: cleanedSteps,
-        messageStepTexts: getPrimaryMessageStepTexts(cached.steps),
+        messageStepTexts: getPrimaryMessageStepTexts(cleanedSteps),
         streamingReplyText: null,
       }];
     }