From 835f12ef47ee9e545b5db9f782e1bf4f41b31d67 Mon Sep 17 00:00:00 2001
From: Vincent Koc <vincentkoc@ieee.org>
Date: Sat, 1 Aug 2026 14:42:36 +0800
Subject: [PATCH] fix(talk): release owner after startup failure

---
 .../chat/realtime-talk-lifecycle.test.ts      | 76 +++++++++++++++++++
 ui/src/pages/chat/realtime-talk.ts            | 60 +++++++++++----
 2 files changed, 121 insertions(+), 15 deletions(-)

diff --git a/ui/src/pages/chat/realtime-talk-lifecycle.test.ts b/ui/src/pages/chat/realtime-talk-lifecycle.test.ts
index ab231487ed82..353563915458 100644
--- a/ui/src/pages/chat/realtime-talk-lifecycle.test.ts
+++ b/ui/src/pages/chat/realtime-talk-lifecycle.test.ts
@@ -167,6 +167,82 @@ describe("RealtimeTalkSession lifecycle", () => {
     secondReplacement.stop();
   });
 
+  it("restores the existing owner when same-session transport startup fails", async () => {
+    const transcriptEntryIds: string[] = [];
+    const request = vi.fn(async (method: string, params?: { entryId?: string }) => {
+      if (method === "talk.client.create") {
+        return {
+          provider: "openai",
+          transport: "webrtc",
+          voiceSessionId: "voice-existing",
+          clientSecret: "secret",
+        };
+      }
+      if (method === "talk.client.transcript") {
+        transcriptEntryIds.push(String(params?.entryId));
+      }
+      return { ok: true };
+    });
+    const client = { request } as never;
+    const session = new RealtimeTalkSession(client, "agent:main:main");
+
+    await session.start();
+    const existingContext = transcriptContext(transportMock.webRtcContexts);
+    transportMock.start.mockRejectedValueOnce(new Error("replacement startup failed"));
+
+    await expect(session.start()).rejects.toThrow("replacement startup failed");
+    expect(request.mock.calls.some(([method]) => method === "talk.client.close")).toBe(false);
+
+    existingContext.callbacks.onTranscript?.({ role: "user", text: "still active", final: true });
+    await existingContext.flushTranscriptWrites?.();
+    expect(transcriptEntryIds).toEqual(["1"]);
+
+    const concurrent = new RealtimeTalkSession(client, "agent:main:main");
+    await concurrent.start();
+    concurrent.stop();
+    session.stop();
+  });
+
+  it("releases newly allocated owners after transport startup failures", async () => {
+    let createCount = 0;
+    const request = vi.fn(async (method: string) => {
+      if (method === "talk.client.create") {
+        createCount += 1;
+        return {
+          provider: "openai",
+          transport: "webrtc",
+          voiceSessionId: `voice-start-${createCount}`,
+          clientSecret: "secret",
+        };
+      }
+      return { ok: true };
+    });
+    const client = { request } as never;
+    transportMock.start
+      .mockRejectedValueOnce(new Error("first startup failed"))
+      .mockRejectedValueOnce(new Error("second startup failed"));
+
+    const first = new RealtimeTalkSession(client, "agent:main:main");
+    await expect(first.start()).rejects.toThrow("first startup failed");
+    await vi.waitFor(() =>
+      expect(request.mock.calls.filter(([method]) => method === "talk.client.close")).toHaveLength(
+        1,
+      ),
+    );
+
+    const second = new RealtimeTalkSession(client, "agent:main:main");
+    await expect(second.start()).rejects.toThrow("second startup failed");
+    await vi.waitFor(() =>
+      expect(request.mock.calls.filter(([method]) => method === "talk.client.close")).toHaveLength(
+        2,
+      ),
+    );
+
+    const recovered = new RealtimeTalkSession(client, "agent:main:main");
+    await recovered.start();
+    recovered.stop();
+  });
+
   it("surfaces transcript failure after three attempts", async () => {
     vi.useFakeTimers();
     const warn = vi.spyOn(console, "warn").mockImplementation(() => undefined);
diff --git a/ui/src/pages/chat/realtime-talk.ts b/ui/src/pages/chat/realtime-talk.ts
index 9bff063598f4..ffb96303e188 100644
--- a/ui/src/pages/chat/realtime-talk.ts
+++ b/ui/src/pages/chat/realtime-talk.ts
@@ -240,8 +240,20 @@ export class RealtimeTalkSession {
     const lifecycleGeneration = ++this.lifecycleGeneration;
     this.closed = false;
     this.callbacks.onStatus?.("connecting");
+    const existingTransport = this.transport;
+    const existingTransportGeneration = this.transportGeneration;
     const existingVoiceSessionId = this.voiceSessionId;
+    const existingAcceptingTranscripts = this.acceptingTranscripts;
+    const existingServerOwnedVoiceSession = this.serverOwnedVoiceSession;
     const existingOwner = this.clientVoiceSessionOwner;
+    const restoreExistingSession = () => {
+      this.transport = existingTransport;
+      this.transportGeneration = existingTransportGeneration;
+      this.voiceSessionId = existingVoiceSessionId;
+      this.acceptingTranscripts = existingAcceptingTranscripts;
+      this.serverOwnedVoiceSession = existingServerOwnedVoiceSession;
+      this.clientVoiceSessionOwner = existingOwner;
+    };
     const owner = reserveClientVoiceSessionOwner(this.client, this.sessionKey);
     let ownerTransferred = false;
     try {
@@ -307,21 +319,39 @@ export class RealtimeTalkSession {
               adoptedOwner.signal,
             );
       const transcriptQueue = this.transcriptQueue;
-      this.transport = createTransport(session, {
-        client: this.client,
-        sessionKey: this.sessionKey,
-        voiceSessionId,
-        flushTranscriptWrites: async () => await transcriptQueue.flush(),
-        callbacks,
-        inputDeviceId: this.localOptions.inputDeviceId,
-        videoDeviceId: this.localOptions.videoDeviceId,
-        consultThinkingLevel: session.consultThinkingLevel,
-        consultFastMode: session.consultFastMode,
-      });
-      this.callbacks.onVideoCapability?.(
-        providerVideoCapable && typeof this.transport.setVideoEnabled === "function",
-      );
-      await this.transport.start();
+      let nextTransport: RealtimeTalkTransport | null = null;
+      try {
+        nextTransport = createTransport(session, {
+          client: this.client,
+          sessionKey: this.sessionKey,
+          voiceSessionId,
+          flushTranscriptWrites: async () => await transcriptQueue.flush(),
+          callbacks,
+          inputDeviceId: this.localOptions.inputDeviceId,
+          videoDeviceId: this.localOptions.videoDeviceId,
+          consultThinkingLevel: session.consultThinkingLevel,
+          consultFastMode: session.consultFastMode,
+        });
+        this.transport = nextTransport;
+        this.callbacks.onVideoCapability?.(
+          providerVideoCapable && typeof nextTransport.setVideoEnabled === "function",
+        );
+        await nextTransport.start();
+      } catch (error) {
+        nextTransport?.stop();
+        if (existingOwner && adoptedOwner === existingOwner) {
+          // A same-session replacement borrows the existing owner and queue.
+          // Restore the live transport instead of closing their shared allocation.
+          restoreExistingSession();
+        } else {
+          const detached = this.detachVoiceSession();
+          restoreExistingSession();
+          if (detached) {
+            this.closeLogicalVoiceSession(detached);
+          }
+        }
+        throw error;
+      }
     } finally {
       if (!ownerTransferred) {
         owner.release();