fix: restore controls after task completion

agent/agent.py

@@ -631,7 +631,8 @@ async def handle_job(ctx: JobContext) -> None:
             voice_enabled = mode == "voice"
             logger.info("Switching conversation mode to %s", mode)
             if not voice_enabled:
-                asyncio.create_task(session.interrupt())
+                # interrupt() returns an already-scheduled Future, not a coroutine.
+                session.interrupt()
             session.input.set_audio_enabled(voice_enabled)
             session.output.set_audio_enabled(voice_enabled)
             return

agent/task_worker.py

@@ -102,11 +102,11 @@ class TaskWorker:
                 if not text:
                     reg.fail(self.task_id, "LLM returned empty response")
                     return
-                if self._is_deferred_answer(text):
+                if self._is_deferred_answer(text) or self._is_underfilled_news_answer(text):
                     reg.add_step(
                         self.task_id,
                         "thinking",
-                        "Draft deferred the work; continuing automatically.",
+                        "Draft was incomplete; continuing automatically.",
                     )
                     if iteration == max_iterations - 1:
                         reg.fail(self.task_id, "Could not produce a complete answer")
@@ -117,8 +117,9 @@ class TaskWorker:
                             {
                                 "role": "user",
                                 "content": (
-                                    "Do not ask me a follow-up or offer to continue later. "
-                                    "Continue the task now, run any more targeted tool calls "
+                                    "The previous draft did not fully answer the original request. "
+                                    "Do not ask a follow-up or offer to continue later. Continue "
+                                    "the task now, run any more targeted tool calls "
                                     "you need, make reasonable assumptions, and only finish "
                                     "when you have directly answered the original request."
                                 ),
@@ -153,6 +154,23 @@ class TaskWorker:
         )
         return "?" in text or any(phrase in lowered for phrase in deferral_phrases)
 
+    def _is_underfilled_news_answer(self, text: str) -> bool:
+        """Reject generic news summaries that do not name actual events."""
+        description = self.description.lower()
+        if "news" not in description and "headline" not in description:
+            return False
+        lowered = text.lower()
+        generic_phrases = (
+            "reports also cover global developments",
+            "other sources indicate",
+            "various international perspectives",
+            "ongoing coverage of world events",
+            "recommend checking major",
+            "unable to retrieve",
+            "could not retrieve",
+        )
+        return any(phrase in lowered for phrase in generic_phrases)
+
     def _system_prompt(self) -> str:
         today = datetime.now(timezone.utc).strftime("%B %d, %Y")
         return (
@@ -162,8 +180,10 @@ class TaskWorker:
             "to finish later; make reasonable assumptions and continue autonomously. If a web "
             "search is broad or inconclusive, immediately run a more targeted query and scrape "
             "the most promising result pages, including today's date and specific reputable "
-            "sources when freshness matters. For headline "
-            "requests, return concrete named events rather than links to news homepages, and "
+            "sources when freshness matters. For plural news or headline requests, provide three "
+            "distinct concrete events unless the user requests a different count. For three "
+            "headlines, write exactly three sentences with one event per sentence and no extras. Never pad the "
+            "answer with generic phrases about broad coverage or international perspectives, and "
             "respect any requested item count exactly. Only "
             "produce a final answer once it fulfills the original request. The final answer must "
             "be 1-3 concise sentences suitable for speaking aloud, with plain text only and no "

tests/test_background_tasks.py

@@ -56,6 +56,14 @@ def test_deferred_task_answers_are_rejected():
         "Reuters reports that the agreement was signed today, while the BBC says voting begins Monday."
     )
 
+    news_worker = TaskWorker("test", "Find the latest international news headlines")
+    assert news_worker._is_underfilled_news_answer(
+        "One flood was reported. Other sources indicate ongoing coverage of world events."
+    )
+    assert not news_worker._is_underfilled_news_answer(
+        "Nepal began flood recovery, Venezuela signed an oil agreement, and Iran resumed talks."
+    )
+
 
 def test_task_event_directories_are_isolated(tmp_path):
     first = TaskRegistry(tmp_path / "room-one")

web/app.js

@@ -166,9 +166,16 @@ function updateModeUI() {
 
 async function setConversationMode(mode) {
     if (mode !== "voice" && mode !== "text") return;
+    const previousMode = conversationMode;
     conversationMode = mode;
     localStorage.setItem(MODE_KEY, mode);
     updateModeUI();
+    if (mode === "voice" && previousMode === "text") {
+        chatInput.blur();
+        window.setTimeout(() => {
+            document.querySelector(".workspace").scrollIntoView({ block: "start", behavior: "smooth" });
+        }, 0);
+    }
 
     if (!room || !room.localParticipant) return;
     sendModeMessage(mode);
@@ -314,9 +321,10 @@ function inferState() {
     userSpeaking = micLevel > 0.035;
     if (agentAudioLevel > 0.02) return "speaking";
     if (userSpeaking) return "listening";
-    // Not speaking, agent silent: stay in thinking until audio arrives,
-    // otherwise settle back to listening.
-    if (uiState === "thinking" || uiState === "speaking") return "thinking";
+    // Once agent audio ends, return to listening. The previous implementation
+    // converted speaking → thinking forever, leaving the UI apparently frozen.
+    if (uiState === "speaking") return "listening";
+    if (uiState === "thinking") return "thinking";
     return "listening";
 }
 

web/style.css

@@ -511,6 +511,7 @@ button {
 .controls {
     position: relative;
     z-index: 2;
+    flex-shrink: 0;
     display: flex;
     align-items: center;
     justify-content: center;
@@ -656,6 +657,7 @@ button {
 
 .connection-strip {
     width: min(340px, 100%);
+    flex-shrink: 0;
     margin-top: 1rem;
 }
 
@@ -685,6 +687,7 @@ button {
 
 .privacy-note {
     display: flex;
+    flex-shrink: 0;
     align-items: center;
     gap: 0.4rem;
     margin-top: 1.15rem;
@@ -1223,6 +1226,39 @@ body[data-mode="text"] .mic-meter {
 .task-error { margin-top: 0.5rem; color: #ffc2ba; background: rgba(255, 128, 111, 0.1); }
 
 /* ── Responsive ─────────────────────────────────────────────────────────── */
+@media (max-height: 760px) and (min-width: 881px) {
+    .workspace {
+        min-height: 540px;
+    }
+
+    .voice-console {
+        padding-top: 1.5rem;
+        padding-bottom: 1rem;
+    }
+
+    .console-intro h2 {
+        font-size: 1.65rem;
+    }
+
+    .console-intro p {
+        margin-top: 0.35rem;
+    }
+
+    .orb-stage {
+        min-height: 220px;
+        margin: -1rem 0;
+    }
+
+    .orb {
+        transform: scale(0.78);
+    }
+
+    .connection-strip,
+    .privacy-note {
+        margin-top: 0.65rem;
+    }
+}
+
 @media (max-width: 880px) {
     body {
         padding: calc(0.85rem + env(safe-area-inset-top))