perf(frontend): stop re-deriving message content on every stream chunk (#4441)

frontend/src/core/messages/utils.ts

@@ -364,7 +364,12 @@ export function extractTextFromMessage(message: Message) {
 const THINK_OPEN_TAG = "<think>";
 const THINK_TAG_RE = /<think>\s*([\s\S]*?)\s*<\/think>/g;
 
-function splitInlineReasoning(content: string) {
+interface InlineReasoningSplit {
+  content: string;
+  reasoning: string | null;
+}
+
+function splitInlineReasoning(content: string): InlineReasoningSplit {
   const reasoningParts: string[] = [];
 
   // First pass: strip every fully closed `<think>...</think>` pair and
@@ -401,11 +406,29 @@ function splitInlineReasoning(content: string) {
   };
 }
 
+// The split is re-derived on every render: `hasContent`, `hasReasoning`,
+// `extractContentFromMessage` and `extractReasoningContentFromMessage` all run
+// over the whole message list on each stream chunk, so an unmemoized scan costs
+// O(total content) per chunk — quadratic across a long run. Cache per message
+// object, keyed by the exact content string it was derived from so a message
+// whose `content` is reassigned recomputes instead of serving a stale split.
+const inlineReasoningCache = new WeakMap<
+  object,
+  { content: string; split: InlineReasoningSplit }
+>();
+
 function splitInlineReasoningFromAIMessage(message: Message) {
   if (message.type !== "ai" || typeof message.content !== "string") {
     return null;
   }
-  return splitInlineReasoning(message.content);
+  const content = message.content;
+  const cached = inlineReasoningCache.get(message);
+  if (cached?.content === content) {
+    return cached.split;
+  }
+  const split = splitInlineReasoning(content);
+  inlineReasoningCache.set(message, { content, split });
+  return split;
 }
 
 export function extractContentFromMessage(message: Message) {
@@ -454,7 +477,7 @@ export function extractReasoningContentFromMessage(message: Message) {
     }
   }
   if (typeof message.content === "string") {
-    return splitInlineReasoning(message.content).reasoning;
+    return splitInlineReasoningFromAIMessage(message)?.reasoning ?? null;
   }
   return null;
 }
@@ -507,7 +530,9 @@ export function hasReasoning(message: Message) {
     return (part as unknown as { type: "thinking" })?.type === "thinking";
   }
   if (typeof message.content === "string") {
-    return splitInlineReasoning(message.content).reasoning !== null;
+    return (
+      (splitInlineReasoningFromAIMessage(message)?.reasoning ?? null) !== null
+    );
   }
   return false;
 }
@@ -567,14 +592,26 @@ export function findToolCallResult(toolCallId: string, messages: Message[]) {
 }
 
 export function isHiddenFromUIMessage(message: Message) {
+  if (message.additional_kwargs?.hide_from_ui === true) {
+    return true;
+  }
+  if (
+    typeof message.name === "string" &&
+    HIDDEN_CONTROL_MESSAGE_NAMES.has(message.name)
+  ) {
+    return true;
+  }
+  // Only the human branch consults the text. Extracting it up front made every
+  // caller pay a full content scan for every AI message it was about to
+  // discard, and this predicate runs over the whole message list on each
+  // stream chunk (grouping, dedup, human-input state).
+  if (message.type !== "human") {
+    return false;
+  }
   const content = extractTextFromMessage(message);
   return (
-    message.additional_kwargs?.hide_from_ui === true ||
-    (typeof message.name === "string" &&
-      HIDDEN_CONTROL_MESSAGE_NAMES.has(message.name)) ||
-    (message.type === "human" &&
-      content.includes("<slash_skill_activation>") &&
-      stripUploadedFilesTag(content).length === 0)
+    content.includes("<slash_skill_activation>") &&
+    stripUploadedFilesTag(content).length === 0
   );
 }
 

frontend/tests/unit/core/messages/utils.test.ts

@@ -14,6 +14,7 @@ import {
   hasContent,
   hasReasoning,
   isAssistantMessageGroupStreaming,
+  isHiddenFromUIMessage,
   parseUploadedFiles,
   stripInternalMarkers,
   stripUploadedFilesTag,
@@ -295,6 +296,96 @@ describe("inline <think> tag splitting", () => {
     expect(extractContentFromMessage(message)).toBe("Documentation: `<think>");
     expect(extractReasoningContentFromMessage(message)).toBeNull();
   });
+
+  test("re-splits when the same message object gets new content", () => {
+    // Streaming replaces `content` on the accumulating message, so a split
+    // cached against the message object must be keyed by the content it was
+    // derived from.
+    const message = aiMessage("<think>first</think>one");
+
+    expect(extractContentFromMessage(message)).toBe("one");
+    expect(extractReasoningContentFromMessage(message)).toBe("first");
+
+    (message as { content: string }).content = "<think>second</think>two";
+
+    expect(extractContentFromMessage(message)).toBe("two");
+    expect(extractReasoningContentFromMessage(message)).toBe("second");
+    expect(hasReasoning(message)).toBe(true);
+  });
+});
+
+describe("isHiddenFromUIMessage", () => {
+  function contentReadCounter(
+    message: Omit<Message, "content">,
+    content: string,
+  ) {
+    const counter = { reads: 0 };
+    const probe = {
+      ...message,
+      get content() {
+        counter.reads++;
+        return content;
+      },
+    } as unknown as Message;
+    return { probe, counter };
+  }
+
+  test("does not read AI content to decide visibility", () => {
+    // Only the human branch consults the text, but this predicate runs over
+    // every message on every stream chunk (grouping, dedup, human-input
+    // state), so reading AI content here costs a full content scan per
+    // message per chunk.
+    const { probe, counter } = contentReadCounter(
+      { id: "ai-visible", type: "ai" } as Message,
+      "<think>long reasoning</think>a long streamed answer",
+    );
+
+    expect(isHiddenFromUIMessage(probe)).toBe(false);
+    expect(counter.reads).toBe(0);
+  });
+
+  test("does not read content when a control message name already hides it", () => {
+    const { probe, counter } = contentReadCounter(
+      { id: "summary-1", type: "ai", name: "summary" } as Message,
+      "compressed context",
+    );
+
+    expect(isHiddenFromUIMessage(probe)).toBe(true);
+    expect(counter.reads).toBe(0);
+  });
+
+  test("still hides a human message that is only slash skill activation", () => {
+    const message = {
+      id: "slash-activation",
+      type: "human",
+      content:
+        "<slash_skill_activation>\n<skill_content># SKILL.md</skill_content>\n</slash_skill_activation>",
+    } as Message;
+
+    expect(isHiddenFromUIMessage(message)).toBe(true);
+  });
+
+  test("keeps a human message that carries real text alongside the activation", () => {
+    const message = {
+      id: "slash-activation-with-task",
+      type: "human",
+      content:
+        "<slash_skill_activation>\n<skill_content># SKILL.md</skill_content>\n</slash_skill_activation>\nreal user task",
+    } as Message;
+
+    expect(isHiddenFromUIMessage(message)).toBe(false);
+  });
+
+  test("hides any message flagged with hide_from_ui", () => {
+    const message = {
+      id: "hidden-1",
+      type: "human",
+      content: "internal reply",
+      additional_kwargs: { hide_from_ui: true },
+    } as Message;
+
+    expect(isHiddenFromUIMessage(message)).toBe(true);
+  });
 });
 
 describe("human message internal context stripping", () => {