fix(voice): return agent reply items for remote run_input

handleRunInput routed run_input through the registered text-input
callback, which is fire-and-forget (interrupt + generateReply) and
never surfaces the resulting chat items, so the remote driver always
received empty responses even though the agent replied. Drive the
reply through session.run() directly and capture result.events,
mirroring the Python SessionHost. Remove the now-unused
SessionHost.registerTextInput; participant text input is handled by
RoomIO itself.

Author: u9g

agents/src/voice/agent_session.ts

@@ -81,12 +81,7 @@ import {
 import { AgentInput, AgentOutput } from './io.js';
 import { RecorderIO } from './recorder_io/index.js';
 import { RoomSessionTransport, SessionHost } from './remote_session.js';
-import {
-  DEFAULT_TEXT_INPUT_CALLBACK,
-  RoomIO,
-  type RoomInputOptions,
-  type RoomOutputOptions,
-} from './room_io/index.js';
+import { RoomIO, type RoomInputOptions, type RoomOutputOptions } from './room_io/index.js';
 import type { UnknownUserData } from './run_context.js';
 import type { SpeechHandle } from './speech_handle.js';
 import { RunResult } from './testing/run_result.js';
@@ -537,11 +532,6 @@ export class AgentSession<
       const transport = new RoomSessionTransport(room, this._roomIO);
       this.sessionHost = new SessionHost(transport);
       this.sessionHost.registerSession(this);
-      if (inputOptions?.textEnabled !== false) {
-        this.sessionHost.registerTextInput(
-          inputOptions?.textInputCallback ?? DEFAULT_TEXT_INPUT_CALLBACK,
-        );
-      }
     }
 
     const ctx = getJobContext(false);

agents/src/voice/remote_session.ts

@@ -626,7 +626,6 @@ export class SessionHost {
   private eventsRegistered = false;
   private recvTask: Task<void> | undefined;
   private readonly tasks = new Set<Task<void>>();
-  private textInputCb: TextInputCallback | undefined;
 
   constructor(
     transport: SessionTransport,
@@ -654,10 +653,6 @@ export class SessionHost {
     }
   }
 
-  registerTextInput(textInputCb: TextInputCallback): void {
-    this.textInputCb = textInputCb;
-  }
-
   async start(): Promise<void> {
     if (this.started) return;
     this.started = true;
@@ -991,26 +986,31 @@ export class SessionHost {
     let items: pb.ChatContext_ChatItem[] = [];
     let error: string | undefined;
 
-    if (text) {
-      if (this.textInputCb) {
-        const cbResult = this.textInputCb(this.session!, { text });
-        if (cbResult instanceof Promise) {
-          await cbResult;
-        }
-      } else {
-        try {
-          await this.session!.interrupt({ force: true }).await;
-        } catch {
-          // ignore
-        }
+    if (!text) {
+      error = 'empty run_input text';
+    } else {
+      // Drive the reply through session.run() directly and capture its events,
+      // ignoring any registered text-input callback. The room's default text
+      // input callback is fire-and-forget (interrupt + generateReply) and never
+      // surfaces the resulting chat items, so routing run_input through it would
+      // always return empty responses to the remote driver. This mirrors the
+      // Python SessionHost behavior.
+      try {
+        await this.session!.interrupt({ force: true }).await;
+      } catch {
+        // ignore
+      }
 
-        const result = this.session!.run({ userInput: text });
-        try {
-          await result.wait();
-        } catch (e) {
-          error = e instanceof Error ? e.message : String(e);
-        }
-        items = chatItemsToProto(result.events.map((ev) => ev.item));
+      const result = this.session!.run({ userInput: text });
+      try {
+        await result.wait();
+      } catch (e) {
+        error = e instanceof Error ? e.message : String(e);
+      }
+      items = chatItemsToProto(result.events.map((ev) => ev.item));
+
+      if (items.length === 0 && !error) {
+        error = 'agent produced no response items';
       }
     }