diff --git a/src/agent/async-agent.test.ts b/src/agent/async-agent.test.ts
index 3d359d3d..3ebe01fd 100644
--- a/src/agent/async-agent.test.ts
+++ b/src/agent/async-agent.test.ts
@@ -7,6 +7,7 @@ const internalRunState = { value: false };
 const runMock = vi.fn(async () => ({ text: "", thinking: undefined, error: undefined }));
 const runInternalMock = vi.fn(async () => ({ text: "", thinking: undefined, error: undefined }));
 const flushSessionMock = vi.fn(async () => {});
+const persistAssistantSummaryMock = vi.fn();
 const subscribeAllMock = vi.fn((fn: (event: any) => void) => {
   subscribeCallbacks.push(fn);
   return () => {};
@@ -19,6 +20,7 @@ vi.mock("./runner.js", () => ({
     run = runMock;
     runInternal = runInternalMock;
     flushSession = flushSessionMock;
+    persistAssistantSummary = persistAssistantSummaryMock;
     get isInternalRun() {
       return internalRunState.value;
     }
@@ -84,6 +86,7 @@ describe("AsyncAgent internal flow", () => {
     runMock.mockReset();
     runInternalMock.mockReset();
     flushSessionMock.mockReset();
+    persistAssistantSummaryMock.mockReset();
     subscribeAllMock.mockClear();
     runMock.mockResolvedValue({ text: "", thinking: undefined, error: undefined });
     runInternalMock.mockResolvedValue({ text: "", thinking: undefined, error: undefined });
@@ -207,4 +210,55 @@ describe("AsyncAgent internal flow", () => {
     internalRunState.value = false;
     agent.close();
   });
+
+  it("persists assistant summary when persistResponse is true and result has text", async () => {
+    runInternalMock.mockResolvedValueOnce({ text: "Summary of findings", thinking: undefined, error: undefined });
+    const agent = new AsyncAgent();
+
+    agent.writeInternal("announce findings", { forwardAssistant: true, persistResponse: true });
+    await agent.waitForIdle();
+
+    expect(persistAssistantSummaryMock).toHaveBeenCalledOnce();
+    expect(persistAssistantSummaryMock).toHaveBeenCalledWith("Summary of findings");
+    // flushSession called twice: once after runInternal, once after persistAssistantSummary
+    expect(flushSessionMock).toHaveBeenCalledTimes(2);
+
+    agent.close();
+  });
+
+  it("does not persist assistant summary when result text is NO_REPLY", async () => {
+    runInternalMock.mockResolvedValueOnce({ text: "NO_REPLY", thinking: undefined, error: undefined });
+    const agent = new AsyncAgent();
+
+    agent.writeInternal("announce findings", { forwardAssistant: true, persistResponse: true });
+    await agent.waitForIdle();
+
+    expect(persistAssistantSummaryMock).not.toHaveBeenCalled();
+
+    agent.close();
+  });
+
+  it("does not persist assistant summary when result text is empty", async () => {
+    runInternalMock.mockResolvedValueOnce({ text: "  ", thinking: undefined, error: undefined });
+    const agent = new AsyncAgent();
+
+    agent.writeInternal("announce findings", { forwardAssistant: true, persistResponse: true });
+    await agent.waitForIdle();
+
+    expect(persistAssistantSummaryMock).not.toHaveBeenCalled();
+
+    agent.close();
+  });
+
+  it("does not persist assistant summary when persistResponse is not set", async () => {
+    runInternalMock.mockResolvedValueOnce({ text: "Summary of findings", thinking: undefined, error: undefined });
+    const agent = new AsyncAgent();
+
+    agent.writeInternal("announce findings", { forwardAssistant: true });
+    await agent.waitForIdle();
+
+    expect(persistAssistantSummaryMock).not.toHaveBeenCalled();
+
+    agent.close();
+  });
 });
diff --git a/src/agent/async-agent.ts b/src/agent/async-agent.ts
index a284871b..b3b76ec1 100644
--- a/src/agent/async-agent.ts
+++ b/src/agent/async-agent.ts
@@ -13,6 +13,8 @@ export type ChannelItem = Message | AgentEvent | MulticaEvent;
 export interface WriteInternalOptions {
   /** Forward assistant message_end events to realtime stream during internal runs */
   forwardAssistant?: boolean | undefined;
+  /** After internal run completes, persist the LLM's summary as a non-internal assistant message */
+  persistResponse?: boolean | undefined;
 }
 
 export class AsyncAgent {
@@ -74,6 +76,7 @@ export class AsyncAgent {
   writeInternal(content: string, options?: WriteInternalOptions): void {
     if (this._closed) throw new Error("Agent is closed");
     const forwardAssistant = options?.forwardAssistant === true;
+    const persistResponse = options?.persistResponse === true;
 
     this.queue = this.queue
       .then(async () => {
@@ -87,6 +90,11 @@ export class AsyncAgent {
             // Internal run errors are for diagnostics only; do not leak to user stream.
             console.error(`[AsyncAgent] Internal run error: ${result.error}`);
           }
+          // Persist the LLM summary so it remains in parent context for future turns
+          if (persistResponse && result.text?.trim() && result.text.trim() !== "NO_REPLY") {
+            this.agent.persistAssistantSummary(result.text.trim());
+            await this.agent.flushSession();
+          }
         } finally {
           this.forwardInternalAssistant = prevForward;
         }
diff --git a/src/agent/runner.ts b/src/agent/runner.ts
index 8955b04c..41c0f7a2 100644
--- a/src/agent/runner.ts
+++ b/src/agent/runner.ts
@@ -563,6 +563,35 @@ export class Agent {
     return this._internalRun;
   }
 
+  /**
+   * Persist a synthetic assistant message into both in-memory state and session JSONL.
+   * Used after an internal run to keep the LLM summary visible in future turns
+   * while the internal prompt stays hidden.
+   */
+  persistAssistantSummary(text: string): void {
+    const model = this.agent.state.model;
+    const message = {
+      role: "assistant" as const,
+      content: [{ type: "text" as const, text }],
+      api: model?.api ?? "openai-completions",
+      provider: model?.provider ?? "internal",
+      model: model?.id ?? "unknown",
+      usage: {
+        input: 0,
+        output: 0,
+        cacheRead: 0,
+        cacheWrite: 0,
+        totalTokens: 0,
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+      },
+      stopReason: "stop" as const,
+      timestamp: Date.now(),
+    };
+
+    this.agent.appendMessage(message);
+    this.session.saveMessage(message);
+  }
+
   /** Ensure session messages are loaded from disk (idempotent) */
   async ensureInitialized(): Promise<void> {
     if (this.initialized) return;
diff --git a/src/agent/subagent/announce.ts b/src/agent/subagent/announce.ts
index 5f9b5438..54c92e75 100644
--- a/src/agent/subagent/announce.ts
+++ b/src/agent/subagent/announce.ts
@@ -293,7 +293,7 @@ export function runCoalescedAnnounceFlow(
       return false;
     }
 
-    parentAgent.writeInternal(message, { forwardAssistant: true });
+    parentAgent.writeInternal(message, { forwardAssistant: true, persistResponse: true });
     return true;
   } catch (err) {
     console.error(`[SubagentAnnounce] Failed to coalesced-announce to parent:`, err);
@@ -341,7 +341,7 @@ export function runSubagentAnnounceFlow(params: SubagentAnnounceParams): boolean
       return false;
     }
 
-    parentAgent.writeInternal(message, { forwardAssistant: true });
+    parentAgent.writeInternal(message, { forwardAssistant: true, persistResponse: true });
     return true;
   } catch (err) {
     console.error(`[SubagentAnnounce] Failed to announce to parent:`, err);