feat(voice): realtime ElevenLabs voice agent alongside classic path

app/src/components/settings/panels/VoicePanel.tsx

@@ -17,9 +17,6 @@ import {
   type VoiceProviderView,
   type VoiceSettings,
 } from '../../../services/api/voiceSettingsApi';
-import { useAppDispatch, useAppSelector } from '../../../store/hooks';
-import { selectVoiceMode, setVoiceMode } from '../../../store/mascotSlice';
-import { VOICE_MODE_FLAG_ENABLED } from '../../../utils/config';
 import {
   openhumanGetVoiceServerSettings,
   openhumanUpdateVoiceServerSettings,
@@ -102,8 +99,6 @@ interface VoicePanelProps {
 
 const VoicePanel = ({ embedded = false }: VoicePanelProps = {}) => {
   const { t } = useT();
-  const dispatch = useAppDispatch();
-  const voiceMode = useAppSelector(selectVoiceMode);
   const { navigateBack, navigateToSettings } = useSettingsNavigation();
   const [settings, setSettings] = useState<VoiceServerSettings | null>(null);
   const [savedSettings, setSavedSettings] = useState<VoiceServerSettings | null>(null);
@@ -582,25 +577,8 @@ const VoicePanel = ({ embedded = false }: VoicePanelProps = {}) => {
           />
         </SettingsSection>
 
-        {/* ─── Realtime voice mode (beta, flag-gated) ──────────────────── */}
-        {VOICE_MODE_FLAG_ENABLED && (
-          <SettingsSection title={t('voice.mode.title')} description={t('voice.mode.desc')}>
-            <SettingsRow
-              htmlFor="voice-mode-realtime"
-              label={t('voice.mode.realtime')}
-              description={t('voice.mode.realtimeDesc')}
-              control={
-                <SettingsSwitch
-                  id="voice-mode-realtime"
-                  data-testid="voice-mode-realtime-toggle"
-                  checked={voiceMode === 'realtime'}
-                  onCheckedChange={next => dispatch(setVoiceMode(next ? 'realtime' : 'classic'))}
-                  aria-label={t('voice.mode.realtime')}
-                />
-              }
-            />
-          </SettingsSection>
-        )}
+        {/* Realtime voice is always on now — its controls live on the Human tab,
+            so the former flag-gated toggle here was removed. */}
 
         {/* ─── Section 1: Voice Provider Chips ─────────────────────────── */}
         {/* Provider chips are intentional bespoke UI — kept as-is. */}

app/src/features/human/HumanPage.tsx

@@ -8,10 +8,8 @@ import {
   selectCustomSecondaryColor,
   selectMascotColor,
   selectSpeakReplies,
-  selectVoiceMode,
   setSpeakReplies,
 } from '../../store/mascotSlice';
-import { VOICE_MODE_FLAG_ENABLED } from '../../utils/config';
 import Conversations from '../conversations/Conversations';
 import {
   CustomGifMascot,
@@ -35,8 +33,8 @@ const HumanPage = () => {
   const speakReplies = useAppSelector(selectSpeakReplies);
 
   const { face, visemeCode } = useHumanMascot({ speakReplies });
-  const voiceMode = useAppSelector(selectVoiceMode);
-  const realtimeEnabled = VOICE_MODE_FLAG_ENABLED && voiceMode === 'realtime';
+  // Realtime voice controls are always shown — no settings/flag gate.
+  const realtimeEnabled = true;
   const mascotColor = useAppSelector(selectMascotColor);
   const customPrimary = useAppSelector(selectCustomPrimaryColor);
   const customSecondary = useAppSelector(selectCustomSecondaryColor);

app/src/features/human/voice/useRealtimeVoiceSession.ts

@@ -3,10 +3,21 @@ import createDebug from 'debug';
 import { useCallback, useEffect, useRef, useState } from 'react';
 
 import { fetchVoiceAgentSignedUrl } from '../../../services/api/voiceAgentApi';
+import { socketService } from '../../../services/socketService';
 import { MASCOT_VOICE_ID } from '../../../utils/config';
 
 const log = createDebug('app:human:realtime-voice');
 
+/**
+ * Instruction prefix for "speak-back". A slow voice turn (e.g. an email summary)
+ * is acknowledged aloud, finishes in the background, and its result is delivered
+ * to chat AND pushed here as a `voice_speak` event. We send it back into the live
+ * ElevenLabs session as a user message wrapped with this prefix so the agent reads
+ * it verbatim. MUST match `VOICE_READBACK_PREFIX` in `voice/realtime_harness.rs`,
+ * which uses it to avoid re-arming speak-back on the read-back turn (loop guard).
+ */
+const READBACK_PREFIX = 'Please read the following to me, word for word, and say nothing else:';
+
 /**
  * Lifecycle of a realtime ElevenLabs Agents voice session (#5399).
  * `idle → connecting → active → idle`, or `→ error`.
@@ -125,6 +136,25 @@ export function useRealtimeVoiceSession(opts?: { voiceId?: string }): RealtimeVo
     []
   );
 
+  // Speak-back: a slow voice turn (email/calendar summary) is acknowledged aloud,
+  // finishes in the background, and the core emits its result as a `voice_speak`
+  // event. While the call is still open, read it aloud by sending it back into the
+  // live ElevenLabs session wrapped in the verbatim prefix (a fast read-back turn).
+  // The result also lands in chat regardless (delivered core-side) — this is the
+  // spoken copy. Refs keep the subscription set up once while always seeing the
+  // live conversation and liveness.
+  useEffect(() => {
+    const handler = (payload: unknown) => {
+      if (!liveRef.current) return; // call already ended — the chat copy stands alone
+      const text = (payload as { full_response?: string } | undefined)?.full_response?.trim();
+      if (!text) return;
+      log('speak-back: reading deferred result aloud (%d chars)', text.length);
+      conversationRef.current.sendUserMessage(`${READBACK_PREFIX}\n\n${text}`);
+    };
+    socketService.on('voice_speak', handler);
+    return () => socketService.off('voice_speak', handler);
+  }, []);
+
   return {
     state,
     isSpeaking: conversation.isSpeaking,

src/openhuman/agent/harness/session/runtime.rs

@@ -178,6 +178,16 @@ impl Agent {
         &self.model_name
     }
 
+    /// Override the base model this session runs its top-level turns on. Set
+    /// once before running: per-turn classification is disabled (the main agent
+    /// is pinned to its configured model for KV-cache stability — see the model
+    /// pin in `turn/core.rs`), so this sticks for the session and is not flipped
+    /// mid-conversation. The realtime voice harness uses it to pin a fast,
+    /// non-thinking model within the provider's response-time ceiling.
+    pub fn set_model_name(&mut self, model_name: impl Into<String>) {
+        self.model_name = model_name.into();
+    }
+
     /// The agent's currently-configured temperature.
     pub fn temperature(&self) -> f64 {
         self.temperature

src/openhuman/agent/harness/session/turn/core.rs

@@ -923,6 +923,16 @@ impl Agent {
 
         let mut agent_context_prepared_sources: Vec<harness::AgentContextPreparedSource> =
             Vec::new();
+        // Triggered memory-agent recall runs on EVERY channel, voice included:
+        // dropping it on voice would strip the user's remembered context
+        // (preferences, people, prior facts) from spoken answers — a real quality
+        // loss the transcript alone can't replace. Recall adds a few seconds of
+        // embedding + retrieval before the first model token, but on realtime
+        // voice that latency is already covered end-to-end: the backend relay
+        // streams an audible keepalive filler from t=0 so the cloud session never
+        // sees a silent stall, and the desktop's ~8s ack-defer closes the spoken
+        // turn and finishes in the background if the work runs long. So the recall
+        // path is byte-for-byte identical across voice and chat.
         let (enriched, memory_agent_context_injected) = self
             .inject_triggered_memory_agent_context(user_message, enriched, &parent_context)
             .await;

src/openhuman/memory/diff/mod.rs

@@ -76,4 +76,5 @@ pub use tinycortex::memory::diff::types::{
     ChangeKind, Checkpoint, CrossSourceDiff, DiffResult, DiffSummary, ItemChange, Snapshot,
     SnapshotTrigger,
 };
+#[cfg(feature = "memory-git")]
 pub use tools::MemoryDiffTool;

src/openhuman/memory/diff/stub.rs

@@ -26,7 +26,7 @@
 use crate::openhuman::config::Config;
 use crate::openhuman::memory::sources::types::MemorySourceEntry;
 
-use super::types::{Checkpoint, CrossSourceDiff, Snapshot};
+use tinycortex::memory::diff::types::{Checkpoint, CrossSourceDiff, Snapshot};
 
 /// The message every disabled entry point returns.
 ///