fix: full review pass (main process, voice pipeline, Hermes client, HUD) and e2e-validated local voice (v2.1.0.1)

Electron main
- utility-process channel: sherpa output copied into V8 buffers (Electron rejects external
  buffers), audio exchanged as base64; worker restart on timeout, identity-safe exit handling,
  deterministic native unload by restart
- webhook: loopback by default without secret, UTF-8-safe body assembly, clean restart, port validation
- files IPC: realpath-based root check, openPath allow-list (reveal-only outside document folders),
  Windows reserved names; store: debounced async atomic writes with backup of corrupt files;
  logger: streaming writes with rotation; log message size cap
- HTTP stream proxy: socket released on idle timeout / renderer destroyed, id validation
- model manager: inactivity timeout, retrying rm/rename (Windows locks), engine stopped before
  replacing a model; WAV decoder handles float/24-bit; IPC payload validation
- window: opaque rounded window on Windows, navigation lock-down, visibility events;
  Ctrl+Alt+Escape instead of the Task Manager shortcut; single-instance guard; EVEFLOW_USER_DATA

Voice pipeline
- abort semantics (SendHandle.aborted), abort before the stream opens, session id prefixes per transport
- hands-free re-arm after replies without speech, start/stop race, no chime on auto re-arm,
  no silence shipped to STT (400 ms pre-roll), no transcription of empty manual stops
- TTS: bounded prefetch, cancellable segments, non-interrupting notices, volume applied at play time
- SSE CRLF split, usage in chat completions, finish_reason length, phonetic regex hoisted

HUD
- core renderer: no canvas shadows, cached colours, reusable spectrum buffer, theme read on change,
  30 fps idle, stops when the window is hidden; ping flashes on send / tool / speech
- deltas coalesced per animation frame; stable auto-scroll; narrow selectors everywhere
- bundled fonts (offline), reduce-motion fix, error toasts, ops drawer below 1180 px,
  interim transcript and first-token latency in the core caption, Ctrl+K, dialog semantics,
  switch/aria roles, compact widget cleanup, Whisper small recommended for French
- docs/ROADMAP.md: audit results, e2e results and the plan towards a real JARVIS

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017Wn5VX9HNbJ7N54hR24u9Y

src/services/conversation.ts

@@ -1,7 +1,7 @@
 /**
  * Conversation orchestrator: sends user input to Hermes through the resolved transport,
- * streams the answer into the chat store, drives the HUD state, speaks the reply and
- * surfaces approvals / clarifications as pending requests.
+ * streams the answer into the chat store (coalesced per animation frame), drives the HUD
+ * state, speaks the reply and surfaces approvals / clarifications as pending requests.
  */
 import type { HermesPushEvent } from '../../shared/ipc';
 import { Log } from '../lib/log';
@@ -18,6 +18,34 @@ let activeMessageId: string | null = null;
 
 const toolLabel = (name: string) => name.replace(/_/g, ' ');
 
+// ── delta coalescing: one store update per animation frame ─────────────────
+const pendingDelta = { id: '', content: '', reasoning: '', raf: 0 };
+
+function flushDelta(): void {
+  if (pendingDelta.raf) cancelAnimationFrame(pendingDelta.raf);
+  const { id, content, reasoning } = pendingDelta;
+  pendingDelta.id = '';
+  pendingDelta.content = '';
+  pendingDelta.reasoning = '';
+  pendingDelta.raf = 0;
+  if (!id) return;
+  const chat = useChat.getState();
+  if (content) chat.appendToMessage(id, content);
+  if (reasoning) chat.appendToMessage(id, reasoning, 'reasoning');
+}
+
+function queueDelta(id: string, text: string, field: 'content' | 'reasoning'): void {
+  if (pendingDelta.id && pendingDelta.id !== id) flushDelta();
+  pendingDelta.id = id;
+  pendingDelta[field] += text;
+  if (!pendingDelta.raf) pendingDelta.raf = requestAnimationFrame(flushDelta);
+}
+
+/** Bumps a counter the core reads to flash briefly (send, tool start, speech start). */
+export function pingCore(): void {
+  useChat.getState().ping();
+}
+
 function localToolContext() {
   const chat = useChat.getState();
   const settings = useSettings.getState().settings;
@@ -49,7 +77,7 @@ function localToolContext() {
   };
 }
 
-function handleEvent(event: HermesStreamEvent, messageId: string, runIdRef: { id: string | null }): void {
+function handleEvent(event: HermesStreamEvent, messageId: string, runIdRef: { id: string | null }, timing: { firstToken: number | null; startedAt: number }): void {
   const chat = useChat.getState();
   switch (event.kind) {
     case 'run.started':
@@ -57,15 +85,20 @@ function handleEvent(event: HermesStreamEvent, messageId: string, runIdRef: { id
       chat.setSending(true, event.runId);
       break;
     case 'delta':
-      chat.appendToMessage(messageId, event.text);
+      if (timing.firstToken === null) {
+        timing.firstToken = Date.now() - timing.startedAt;
+        chat.setLatency(timing.firstToken);
+      }
+      queueDelta(messageId, event.text, 'content');
       speech.pushStream(event.text);
       break;
     case 'reasoning':
-      chat.appendToMessage(messageId, event.text, 'reasoning');
+      queueDelta(messageId, event.text, 'reasoning');
       break;
     case 'tool.start':
       chat.pushActivity({ id: event.id, kind: 'tool', name: event.name, status: 'running', detail: event.args ? previewText(event.args, 160) : undefined });
       chat.setHud('thinking');
+      pingCore();
       break;
     case 'tool.progress':
       if (!chat.activity.some((a) => (event.id && a.id === event.id) || (a.name === event.name && a.status === 'running'))) {
@@ -113,6 +146,7 @@ function handleEvent(event: HermesStreamEvent, messageId: string, runIdRef: { id
       break;
     case 'completed':
       if (event.usage) chat.updateMessage(messageId, { usage: event.usage });
+      if (event.status === 'length') chat.updateMessage(messageId, { status: 'cancelled' });
       break;
     case 'error':
       chat.setError(event.message);
@@ -127,7 +161,8 @@ function addPending(request: Omit<PendingRequest, 'id' | 'createdAt'>): void {
   const chat = useChat.getState();
   chat.addPending(request);
   chat.setHud('alert');
-  speech.say(request.kind === 'approval' ? 'Autorisation requise.' : request.kind === 'clarify' ? request.description : 'Saisie requise.');
+  // Spoken without interrupting the streamed answer.
+  speech.say(request.kind === 'approval' ? 'Autorisation requise.' : request.kind === 'clarify' ? request.description : 'Saisie requise.', { interrupt: false });
 }
 
 export async function sendMessage(text: string, images: string[] = [], source = 'eveflow'): Promise<void> {
@@ -139,6 +174,7 @@ export async function sendMessage(text: string, images: string[] = [], source =
 
   speech.stop();
   chat.setError(null);
+  chat.setLatency(null);
   chat.addMessage({ role: 'user', content: trimmed || '(image)', images: images.length ? images : undefined, source, status: 'done' });
   const history = chat.messages
     .filter((m) => m.role !== 'system' && m.status !== 'error')
@@ -148,7 +184,9 @@ export async function sendMessage(text: string, images: string[] = [], source =
   activeMessageId = messageId;
   chat.setSending(true);
   chat.setHud('thinking');
+  pingCore();
   const startedAt = Date.now();
+  const timing = { firstToken: null as number | null, startedAt };
   const runIdRef = { id: null as string | null };
 
   const client = hermes.client();
@@ -158,7 +196,7 @@ export async function sendMessage(text: string, images: string[] = [], source =
       images,
       history,
       sessionId: settings.hermesSessionId,
-      onEvent: (event) => handleEvent(event, messageId, runIdRef),
+      onEvent: (event) => handleEvent(event, messageId, runIdRef, timing),
       localToolDefinitions: LOCAL_TOOL_DEFINITIONS,
       localToolExecutor: (name, args) => executeLocalTool(name, args, localToolContext())
     },
@@ -169,20 +207,26 @@ export async function sendMessage(text: string, images: string[] = [], source =
 
   try {
     const reply = await handle.result;
-    const current = useChat.getState().messages.find((m) => m.id === messageId);
-    const finalText = reply || current?.content || '';
-    useChat.getState().updateMessage(messageId, { content: finalText, status: 'done' });
-    // Anything not streamed (non-delta transports) is spoken now.
-    if (settings.speech.autoSpeak) {
-      if (!current?.content && finalText) speech.say(finalText);
+    flushDelta();
+    const state = useChat.getState();
+    const current = state.messages.find((m) => m.id === messageId);
+    const streamedText = current?.content ?? '';
+    const finalText = handle.aborted ? streamedText : reply || streamedText;
+    const truncated = current?.status === 'cancelled';
+    state.updateMessage(messageId, { content: finalText || (handle.aborted ? '(interrompu)' : ''), status: handle.aborted || truncated ? 'cancelled' : 'done' });
+    if (handle.aborted) {
+      speech.discardStream();
+    } else if (settings.speech.autoSpeak) {
+      if (!streamedText && finalText) speech.say(finalText);
       else speech.endStream();
     } else {
       speech.discardStream();
     }
-    useChat.getState().setHud(speech.isSpeaking() ? 'speaking' : 'idle');
-    if (useHermes.getState().link === 'offline') useHermes.getState().connect().catch(() => undefined);
-    Log.info('hermes', `reply in ${Date.now() - startedAt} ms via ${handle.transport}`, { chars: finalText.length });
+    state.setHud(speech.isSpeaking() ? 'speaking' : 'idle');
+    if (!handle.aborted && useHermes.getState().link === 'offline') useHermes.getState().connect().catch(() => undefined);
+    Log.info('hermes', `${handle.aborted ? 'aborted' : 'reply'} in ${Date.now() - startedAt} ms via ${handle.transport}`, { chars: finalText.length, firstToken: timing.firstToken });
   } catch (err) {
+    flushDelta();
     const message = (err as Error).message || String(err);
     const partial = useChat.getState().messages.find((m) => m.id === messageId)?.content ?? '';
     useChat.getState().updateMessage(messageId, {
@@ -200,10 +244,14 @@ export async function sendMessage(text: string, images: string[] = [], source =
   } finally {
     active = null;
     activeMessageId = null;
-    useChat.getState().setSending(false);
-    for (const a of useChat.getState().activity) {
-      if (a.status === 'running') useChat.getState().finishActivity({ id: a.id }, { status: 'done' });
+    const state = useChat.getState();
+    state.setSending(false);
+    state.setLatency(null);
+    for (const a of state.activity) {
+      if (a.status === 'running') state.finishActivity({ id: a.id }, { status: 'done' });
     }
+    // Requests bound to this run cannot be answered once it is over.
+    for (const p of state.pending) if (!runIdRef.id || p.runId === runIdRef.id) state.removePending(p.id);
   }
 }
 
@@ -235,14 +283,13 @@ export async function resolvePending(request: PendingRequest, decision: string):
   chat.removePending(request.id);
   const client = useHermes.getState().client();
   try {
+    if (!request.runId) throw new Error('run inconnu');
     if (request.kind === 'approval') {
       const choice = (['once', 'session', 'always', 'deny'].includes(decision) ? decision : decision === 'approve' ? 'once' : 'deny') as 'once' | 'session' | 'always' | 'deny';
-      if (!request.runId) throw new Error('run inconnu');
       await client.approveRun(request.runId, choice, request.requestId || undefined);
       chat.addMessage({ role: 'system', content: `Autorisation ${choice === 'deny' ? 'refusée' : 'accordée'}${request.tool ? ` (${toolLabel(request.tool)})` : ''}.`, status: 'done' });
     } else {
       // Clarifications and secrets are answered by steering the run with the reply.
-      if (!request.runId) throw new Error('run inconnu');
       await client.steerRun(request.runId, decision);
       chat.addMessage({ role: 'system', content: request.secret ? 'Saisie transmise à Hermes.' : `Réponse transmise : ${decision}`, status: 'done' });
     }
@@ -278,6 +325,7 @@ export function handlePush(event: HermesPushEvent): void {
     speech.say(event.jobName ? `Résultat de ${event.jobName}. ${event.text}` : event.text);
   }
   chat.setHud(event.status?.includes('fail') ? 'alert' : 'success');
+  pingCore();
   setTimeout(() => {
     const s = useChat.getState();
     if (s.hud === 'success' || s.hud === 'alert') s.setHud(speech.isSpeaking() ? 'speaking' : 'idle');

src/services/hermes/client.ts

@@ -261,6 +261,7 @@ export class HermesClient {
       for (const event of normalizeLifecycleEvent(eventName, data)) onEvent(event);
     });
     const errorChunks: string[] = [];
+    let handleReady = false;
     const handle = await httpStream(
       {
         url: `${this.base}${path}`,
@@ -271,7 +272,6 @@ export class HermesClient {
       },
       { onChunk: (text) => (handleReady ? parser.feed(text) : errorChunks.push(text)) }
     );
-    let handleReady = handle.start.ok;
     if (!handle.start.ok) {
       await handle.done.catch(() => undefined);
       throw new HttpError(handle.start.status, errorMessage(handle.start.status, errorChunks.join('')));
@@ -285,25 +285,45 @@ export class HermesClient {
 
   send(options: SendOptions, transport: ResolvedTransport): SendHandle {
     const effective: ResolvedTransport = options.images?.length && transport === 'runs' ? 'completions' : transport;
-    let abort: () => void = () => undefined;
+    let aborted = false;
+    let abortImpl: () => void = () => undefined;
+    const setAbort = (fn: () => void) => {
+      abortImpl = fn;
+      if (aborted) fn();
+    };
+    const isAborted = () => aborted;
     const result =
       effective === 'runs'
-        ? this.sendViaRuns(options, (fn) => (abort = fn))
+        ? this.sendViaRuns(options, setAbort, isAborted)
         : effective === 'sessions'
-          ? this.sendViaSessions(options, (fn) => (abort = fn))
-          : this.sendViaCompletions(options, (fn) => (abort = fn));
-    return { result, abort: () => abort(), transport: effective };
+          ? this.sendViaSessions(options, setAbort, isAborted)
+          : this.sendViaCompletions(options, setAbort, isAborted);
+    return {
+      result,
+      transport: effective,
+      get aborted() {
+        return aborted;
+      },
+      abort: () => {
+        aborted = true;
+        abortImpl();
+      }
+    };
   }
 
-  private async sendViaRuns(options: SendOptions, setAbort: (fn: () => void) => void): Promise<string> {
+  private async sendViaRuns(options: SendOptions, setAbort: (fn: () => void) => void, isAborted: () => boolean): Promise<string> {
     const { onEvent } = options;
     const run = await this.startRun({
       input: options.text,
-      session_id: options.sessionId || undefined,
+      session_id: plainSession(options.sessionId) || undefined,
       instructions: this.config.instructions || undefined,
       model: this.config.model || undefined
     });
     const runId = run.run_id;
+    if (isAborted()) {
+      this.stopRun(runId).catch(() => undefined);
+      return '';
+    }
     onEvent({ kind: 'run.started', runId });
     Log.info('hermes', `run ${runId} started`);
 
@@ -329,7 +349,7 @@ export class HermesClient {
     });
 
     await handle.done;
-    if (stopped) return streamedText;
+    if (stopped || isAborted()) return streamedText;
     if (failure) throw new Error(failure);
 
     if (!completed) {
@@ -349,7 +369,7 @@ export class HermesClient {
     return finalText || streamedText;
   }
 
-  private async sendViaSessions(options: SendOptions, setAbort: (fn: () => void) => void): Promise<string> {
+  private async sendViaSessions(options: SendOptions, setAbort: (fn: () => void) => void, isAborted: () => boolean): Promise<string> {
     const { onEvent } = options;
     let sessionId = options.sessionId;
     if (!sessionId || !sessionId.startsWith('hs:')) {
@@ -357,6 +377,7 @@ export class HermesClient {
       sessionId = `hs:${session.id}`;
       onEvent({ kind: 'session', sessionId });
     }
+    if (isAborted()) return '';
     const realId = sessionId.slice(3);
     const body: Rec = { input: options.text };
     if (this.config.instructions) body.instructions = this.config.instructions;
@@ -374,11 +395,12 @@ export class HermesClient {
     });
     setAbort(() => handle.abort());
     await handle.done;
+    if (isAborted()) return streamed;
     if (failure) throw new Error(failure);
     return finalText || streamed;
   }
 
-  private async sendViaCompletions(options: SendOptions, setAbort: (fn: () => void) => void): Promise<string> {
+  private async sendViaCompletions(options: SendOptions, setAbort: (fn: () => void) => void, isAborted: () => boolean): Promise<string> {
     const { onEvent } = options;
     const useTools = this.config.localTools && !!options.localToolExecutor && !!options.localToolDefinitions?.length;
     const messages: HermesChatMessage[] = [];
@@ -389,16 +411,13 @@ export class HermesClient {
       : options.text;
     messages.push({ role: 'user', content: userContent });
 
-    let aborted = false;
     let currentHandle: StreamHandle | null = null;
-    setAbort(() => {
-      aborted = true;
-      currentHandle?.abort();
-    });
+    setAbort(() => currentHandle?.abort());
+    const aborted = () => isAborted();
 
     let fullText = '';
     let toolsAllowed = useTools;
-    for (let iteration = 0; iteration < 6 && !aborted; iteration++) {
+    for (let iteration = 0; iteration < 6 && !aborted(); iteration++) {
       const payload: Rec = { model: this.config.model || 'hermes-agent', messages, stream: true };
       if (toolsAllowed) {
         payload.tools = options.localToolDefinitions;
@@ -428,7 +447,8 @@ export class HermesClient {
       });
 
       const headers = this.headers({ Accept: 'text/event-stream' });
-      if (options.sessionId) headers['X-Hermes-Session-Id'] = options.sessionId;
+      const sessionId = plainSession(options.sessionId);
+      if (sessionId) headers['X-Hermes-Session-Id'] = sessionId;
       const handle = await httpStream(
         { url: `${this.base}/v1/chat/completions`, method: 'POST', headers, body: JSON.stringify(payload), timeoutMs: 10 * 60_000 },
         { onChunk: (text) => (ready ? parser.feed(text) : errorChunks.push(text)) }
@@ -454,7 +474,7 @@ export class HermesClient {
       parser.end();
       fullText += iterationText;
 
-      if (finishReason === 'tool_calls' && toolsAllowed && toolCalls.size > 0 && !aborted) {
+      if (finishReason === 'tool_calls' && toolsAllowed && toolCalls.size > 0 && !aborted()) {
         const calls = [...toolCalls.values()];
         messages.push({
           role: 'assistant',
@@ -471,11 +491,16 @@ export class HermesClient {
       }
       break;
     }
-    onEvent({ kind: 'completed', status: aborted ? 'cancelled' : 'completed', text: fullText });
+    onEvent({ kind: 'completed', status: aborted() ? 'cancelled' : 'completed', text: fullText });
     return fullText;
   }
 }
 
+/** Session ids created by the sessions transport carry an `hs:` prefix; other endpoints get the bare id. */
+export function plainSession(id: string): string {
+  return id.startsWith('hs:') ? id.slice(3) : id;
+}
+
 export function normalizeJob(job: HermesJob): HermesJob {
   const raw = job as Rec;
   const id = String(raw.id ?? raw.job_id ?? raw.name ?? '');

src/services/hermes/events.ts

@@ -212,6 +212,7 @@ export function normalizeCompletionChunk(
 ): { events: HermesStreamEvent[]; finishReason: string | null } {
   const events: HermesStreamEvent[] = [];
   if (!isRec(chunk)) return { events, finishReason: null };
+  if (isRec(chunk.usage)) events.push({ kind: 'completed', status: 'usage', usage: chunk.usage as HermesUsage });
   const choice = Array.isArray(chunk.choices) && isRec(chunk.choices[0]) ? (chunk.choices[0] as Rec) : null;
   if (!choice) return { events, finishReason: null };
   const delta = isRec(choice.delta) ? (choice.delta as Rec) : isRec(choice.message) ? (choice.message as Rec) : null;

src/services/hermes/types.ts

@@ -167,8 +167,10 @@ export interface SendOptions {
 }
 
 export interface SendHandle {
-  /** Resolves with the final assistant text. */
+  /** Resolves with the final assistant text (partial text when aborted). */
   result: Promise<string>;
   abort: () => void;
+  /** True once abort() was called; the caller must not treat the result as a completed reply. */
+  readonly aborted: boolean;
   transport: Exclude<HermesTransport, 'auto'>;
 }

src/services/logService.ts

@@ -1,70 +0,0 @@
-/**
- * LogService — Centralized structured logger for EveFlow.
- *
- * In Electron (production/dev): sends log entries via window.electronAPI.log()
- * which the main process writes to %APPDATA%/EveFlow/eveflow.log
- *
- * Fallback (browser only): console only.
- */
-
-export type LogLevel = 'DEBUG' | 'INFO' | 'WARN' | 'ERROR';
-
-export interface LogEntry {
-  ts: string;        // ISO timestamp
-  level: LogLevel;
-  tag: string;       // Component / service tag (e.g. "HermesAgent", "Minimax")
-  message: string;
-  data?: unknown;    // Any extra serializable payload
-}
-
-export class LogService {
-  private static readonly MAX_CONSOLE_DATA_LEN = 2000;
-
-  // Core write method
-  public static log(level: LogLevel, tag: string, message: string, data?: unknown): void {
-    const entry: LogEntry = {
-      ts: new Date().toISOString(),
-      level,
-      tag,
-      message,
-      data
-    };
-
-    // Always write to console for DevTools visibility
-    const consoleMsg = `[${entry.ts}] [${level}] [${tag}] ${message}`;
-    const dataStr = data !== undefined ? JSON.stringify(data, null, 2) : '';
-
-    switch (level) {
-      case 'ERROR': console.error(consoleMsg, dataStr ? dataStr.slice(0, this.MAX_CONSOLE_DATA_LEN) : ''); break;
-      case 'WARN':  console.warn(consoleMsg, dataStr ? dataStr.slice(0, this.MAX_CONSOLE_DATA_LEN) : ''); break;
-      case 'DEBUG': console.debug(consoleMsg, dataStr ? dataStr.slice(0, this.MAX_CONSOLE_DATA_LEN) : ''); break;
-      default:      console.info(consoleMsg, dataStr ? dataStr.slice(0, this.MAX_CONSOLE_DATA_LEN) : ''); break;
-    }
-
-    // Send to Electron main process for file persistence
-    try {
-      const electronAPI = (window as any).electronAPI;
-      if (electronAPI?.writeLog) {
-        electronAPI.writeLog(entry);
-      }
-    } catch {
-      // Silently ignore: browser context has no electronAPI
-    }
-  }
-
-  public static debug(tag: string, message: string, data?: unknown): void {
-    this.log('DEBUG', tag, message, data);
-  }
-
-  public static info(tag: string, message: string, data?: unknown): void {
-    this.log('INFO', tag, message, data);
-  }
-
-  public static warn(tag: string, message: string, data?: unknown): void {
-    this.log('WARN', tag, message, data);
-  }
-
-  public static error(tag: string, message: string, data?: unknown): void {
-    this.log('ERROR', tag, message, data);
-  }
-}

src/services/voice/audioBus.ts

@@ -71,9 +71,9 @@ class AudioBus {
   }
 
   /** Normalised output spectrum (0..1) with `bins` values, low to high frequency. */
-  spectrum(bins: number, source: 'output' | 'input' = 'output'): Float32Array {
+  spectrum(bins: number, source: 'output' | 'input' = 'output', out: Float32Array = new Float32Array(bins)): Float32Array {
     const analyser = source === 'output' ? this.outputAnalyser : this.inputAnalyser;
-    const out = new Float32Array(bins);
+    out.fill(0);
     if (!analyser) {
       if (this.synthetic > 0) {
         const t = performance.now() / 1000;

src/services/voice/capture.ts

@@ -88,6 +88,7 @@ export class MicCapture {
   private active = false;
   private options: CaptureOptions | null = null;
   private sampleRate = 16_000;
+  private hadSpeech = false;
 
   get isActive(): boolean {
     return this.active;
@@ -98,6 +99,7 @@ export class MicCapture {
     this.options = options;
     this.chunks = [];
     this.totalSamples = 0;
+    this.hadSpeech = false;
     this.vad = options.mode === 'auto' ? new EnergyVad(options.vad ?? DEFAULT_VAD) : null;
 
     if (!navigator.mediaDevices?.getUserMedia) {
@@ -169,6 +171,12 @@ export class MicCapture {
     if (!this.active || !this.options) return;
     this.chunks.push(samples);
     this.totalSamples += samples.length;
+    // Before speech starts only a short pre-roll (~400 ms) is kept, so silence is never shipped to STT.
+    if (this.vad && !this.hadSpeech) {
+      while (this.chunks.length > 1 && this.totalSamples - this.chunks[0].length > this.sampleRate * 0.4) {
+        this.totalSamples -= this.chunks.shift()!.length;
+      }
+    }
     const level = rms(samples);
     this.options.callbacks.onLevel?.(Math.min(1, level * 6));
 
@@ -184,6 +192,7 @@ export class MicCapture {
   private handleVad(signal: VadSignal): void {
     switch (signal) {
       case 'speech-start':
+        this.hadSpeech = true;
         this.options?.callbacks.onSpeechStart?.();
         break;
       case 'speech-end':
@@ -222,10 +231,12 @@ export class MicCapture {
   /** Stop capture. In manual mode this finalises the utterance; otherwise reports the reason. */
   stop(reason: 'stopped' | 'no-speech' | 'too-short' = 'stopped'): void {
     if (!this.active) return;
-    if (reason === 'stopped') {
+    // Manual stop: transcribe only if something was actually said (or no VAD is running).
+    if (reason === 'stopped' && (!this.vad || this.hadSpeech)) {
       this.finishUtterance('stopped');
       return;
     }
+    if (reason === 'stopped') reason = 'no-speech';
     const opts = this.options;
     this.teardown();
     opts?.callbacks.onEnd(reason);

src/services/voice/speech.ts

@@ -35,9 +35,10 @@ class SpeechFacade {
     return this.engine?.isActive ?? false;
   }
 
-  say(text: string): void {
+  say(text: string, options: { interrupt?: boolean } = {}): void {
     if (useSettings.getState().settings.speech.provider === 'off') return;
-    this.tts.speak(text);
+    // While an answer streams, spoken notices are inserted without discarding the rest.
+    this.tts.speak(text, { interrupt: options.interrupt ?? !this.streaming });
   }
 
   pushStream(delta: string): void {

src/services/voice/tts.ts

@@ -1,7 +1,7 @@
 /**
- * Text-to-speech engine with a sentence queue, prefetching and Web Audio playback so the HUD
- * reacts to the actual waveform. Providers: OpenAI-compatible /v1/audio/speech, system voices,
- * and the legacy Google Translate endpoint (no key, online only).
+ * Text-to-speech engine with a sentence queue, bounded prefetching and Web Audio playback so
+ * the HUD reacts to the actual waveform. Providers: in-app sherpa-onnx (Kokoro / Piper),
+ * OpenAI-compatible /v1/audio/speech, system voices, and the legacy Google Translate endpoint.
  */
 import { Log } from '../../lib/log';
 import { bridge } from '../../lib/bridge';
@@ -35,6 +35,8 @@ interface QueueItem {
   audio?: Promise<AudioBuffer | null>;
 }
 
+const LOOKAHEAD = 2;
+
 export function ttsEndpoint(apiUrl: string): string {
   const base = apiUrl.trim().replace(/\/+$/, '');
   if (base.endsWith('/speech')) return base;
@@ -46,8 +48,7 @@ export function ttsEndpoint(apiUrl: string): string {
 export class TtsEngine {
   private queue: QueueItem[] = [];
   private playing = false;
-  private currentSource: AudioBufferSourceNode | null = null;
-  private currentUtterance: SpeechSynthesisUtterance | null = null;
+  private cancelCurrent: (() => void) | null = null;
   private streamBuffer = '';
   private generation = 0;
   private config: TtsConfig;
@@ -60,7 +61,6 @@ export class TtsEngine {
 
   updateConfig(config: TtsConfig): void {
     this.config = config;
-    audioBus.setVolume(config.volume);
   }
 
   onState(listener: (state: TtsState) => void): () => void {
@@ -78,10 +78,18 @@ export class TtsEngine {
     for (const l of this.listeners) l(state);
   }
 
-  /** Speak a finished text (clears any pending speech first). */
-  speak(text: string): void {
-    this.stop();
-    for (const chunk of chunkForSpeech(text)) this.enqueue(chunk);
+  /**
+   * Speak a finished text. With `interrupt` (default) pending speech is discarded first;
+   * without it the text is inserted ahead of the queue and streamed speech resumes afterwards.
+   */
+  speak(text: string, options: { interrupt?: boolean } = {}): void {
+    const interrupt = options.interrupt ?? true;
+    if (interrupt) this.stop();
+    const items = chunkForSpeech(text).map((chunk) => this.makeItem(chunk)).filter((i): i is QueueItem => !!i);
+    if (interrupt) this.queue.push(...items);
+    else this.queue.unshift(...items);
+    this.fillPrefetch();
+    void this.drain();
   }
 
   /** Feed streamed tokens; complete sentences are spoken as soon as they are available. */
@@ -100,32 +108,32 @@ export class TtsEngine {
   }
 
   enqueue(text: string): void {
-    if (this.config.provider === 'off') return;
-    const clean = cleanForSpeech(text);
-    if (!clean || !/[\p{L}\p{N}]/u.test(clean)) return;
-    const item: QueueItem = { text: clean };
-    if (this.config.provider !== 'system') item.audio = this.prefetch(clean, this.generation);
+    const item = this.makeItem(text);
+    if (!item) return;
     this.queue.push(item);
+    this.fillPrefetch();
     void this.drain();
   }
 
+  private makeItem(text: string): QueueItem | null {
+    if (this.config.provider === 'off') return null;
+    const clean = cleanForSpeech(text);
+    if (!clean || !/[\p{L}\p{N}]/u.test(clean)) return null;
+    return { text: clean };
+  }
+
+  /** Keep at most LOOKAHEAD synthesis requests in flight ahead of playback. */
+  private fillPrefetch(): void {
+    if (this.config.provider === 'system') return;
+    for (const item of this.queue.slice(0, LOOKAHEAD)) item.audio ??= this.prefetch(item.text, this.generation);
+  }
+
   stop(): void {
     this.generation++;
     this.queue = [];
     this.streamBuffer = '';
-    if (this.currentSource) {
-      try {
-        this.currentSource.onended = null;
-        this.currentSource.stop();
-      } catch {
-        /* already stopped */
-      }
-      this.currentSource = null;
-    }
-    if (this.currentUtterance) {
-      this.currentUtterance.onend = null;
-      this.currentUtterance = null;
-    }
+    this.cancelCurrent?.();
+    this.cancelCurrent = null;
     if (typeof speechSynthesis !== 'undefined') speechSynthesis.cancel();
     audioBus.setSyntheticLevel(0);
     this.playing = false;
@@ -138,15 +146,17 @@ export class TtsEngine {
     const gen = this.generation;
     while (this.queue.length > 0 && gen === this.generation) {
       const item = this.queue.shift()!;
+      this.fillPrefetch();
       try {
-        if (item.audio) {
+        if (this.config.provider !== 'system') {
           this.setState('loading');
-          const buffer = await item.audio;
+          const buffer = await (item.audio ?? this.prefetch(item.text, gen));
           if (gen !== this.generation) break;
           if (buffer) {
             this.setState('speaking');
             await this.playBuffer(buffer, gen);
           } else {
+            this.setState('speaking');
             await this.speakSystem(item.text, gen);
           }
         } else {
@@ -174,7 +184,7 @@ export class TtsEngine {
         return audioBus.context.decodeAudioData(copy);
       })
       .catch((err) => {
-        Log.warn('tts', `synthesis failed, falling back to system voice: ${(err as Error).message}`);
+        if (gen === this.generation) Log.warn('tts', `synthesis failed, falling back to system voice: ${(err as Error).message}`);
         return null;
       });
   }
@@ -238,15 +248,26 @@ export class TtsEngine {
     return new Promise((resolve) => {
       if (gen !== this.generation) return resolve();
       const ctx = audioBus.context;
+      audioBus.setVolume(this.config.volume);
       const source = ctx.createBufferSource();
       source.buffer = buffer;
       if (this.config.provider === 'google-free') source.playbackRate.value = Math.max(0.5, Math.min(2, this.config.speed || 1));
       source.connect(audioBus.output);
-      source.onended = () => {
-        if (this.currentSource === source) this.currentSource = null;
+      const finish = () => {
+        if (this.cancelCurrent === cancel) this.cancelCurrent = null;
         resolve();
       };
-      this.currentSource = source;
+      const cancel = () => {
+        source.onended = null;
+        try {
+          source.stop();
+        } catch {
+          /* already stopped */
+        }
+        finish();
+      };
+      source.onended = finish;
+      this.cancelCurrent = cancel;
       source.start();
     });
   }
@@ -268,12 +289,17 @@ export class TtsEngine {
         if (pulse) clearInterval(pulse);
         pulse = null;
         audioBus.setSyntheticLevel(0);
-        if (this.currentUtterance === utterance) this.currentUtterance = null;
+        if (this.cancelCurrent === cancel) this.cancelCurrent = null;
         resolve();
       };
+      const cancel = () => {
+        utterance.onend = null;
+        utterance.onerror = null;
+        finish();
+      };
       utterance.onend = finish;
       utterance.onerror = finish;
-      this.currentUtterance = utterance;
+      this.cancelCurrent = cancel;
       speechSynthesis.speak(utterance);
     });
   }

src/services/voice/voiceController.ts

@@ -51,6 +51,7 @@ class VoiceController {
   private browser = new BrowserRecognizer();
   private handsFreeTimer: ReturnType<typeof setTimeout> | null = null;
   private stopping = false;
+  private startSeq = 0;
   private unsubscribeVoice: (() => void) | null = null;
 
   init(): void {
@@ -61,6 +62,13 @@ class VoiceController {
         this.scheduleHandsFree(350);
       }
     });
+    // Hands-free also re-arms when a reply ends without speech (voice off, empty answer, error).
+    useChat.subscribe((state, prev) => {
+      if (prev.isSending && !state.isSending) {
+        const v = useVoice.getState();
+        if (v.handsFree && v.phase === 'off' && !speech.isSpeaking()) this.scheduleHandsFree(400);
+      }
+    });
   }
 
   dispose(): void {
@@ -83,6 +91,7 @@ class VoiceController {
     if (on && !this.isListening && !speech.isSpeaking()) void this.start();
     if (!on) {
       this.clearHandsFreeTimer();
+      if (this.isListening) this.cancel();
     }
   }
 
@@ -90,7 +99,8 @@ class VoiceController {
     this.clearHandsFreeTimer();
     this.handsFreeTimer = setTimeout(() => {
       const v = useVoice.getState();
-      if (v.handsFree && v.phase === 'off' && !useChat.getState().isSending && !speech.isSpeaking()) void this.start();
+      const bargeIn = useSettings.getState().settings.voice.bargeIn;
+      if (v.handsFree && v.phase === 'off' && !useChat.getState().isSending && (bargeIn || !speech.isSpeaking())) void this.start(true);
     }, delay);
   }
 
@@ -99,14 +109,16 @@ class VoiceController {
     this.handsFreeTimer = null;
   }
 
-  async start(): Promise<void> {
+  /** @param auto true when re-armed by the hands-free loop (no chime, patient timeout). */
+  async start(auto = false): Promise<void> {
     const voice = useVoice.getState();
     if (voice.phase !== 'off') return;
     const settings = useSettings.getState().settings.voice;
+    const seq = ++this.startSeq;
     voice.setError(null);
     voice.setInterim('');
-    if (!settings.bargeIn) speech.stop();
-    else speech.stop();
+    // Barge-in keeps the assistant talking while the mic opens (echo cancellation handles the overlap).
+    if (!(auto && settings.bargeIn)) speech.stop();
     useChat.getState().setHud('listening');
     voice.setPhase('arming');
 
@@ -125,18 +137,29 @@ class VoiceController {
           silenceMs: settings.silenceMs,
           speechRatio: SENSITIVITY_RATIO[sensitivity],
           minRms: SENSITIVITY_MIN_RMS[sensitivity],
-          noSpeechTimeoutMs: useVoice.getState().handsFree ? 20_000 : 9_000
+          noSpeechTimeoutMs: useVoice.getState().handsFree ? Number.POSITIVE_INFINITY : 9_000
         },
         callbacks: {
-          onLevel: (level) => useVoice.getState().setInputLevel(level),
-          onSpeechStart: () => useVoice.getState().setPhase('speech'),
+          onLevel: (level) => {
+            const q = Math.round(level * 20) / 20;
+            if (q !== useVoice.getState().inputLevel) useVoice.getState().setInputLevel(q);
+          },
+          onSpeechStart: () => {
+            useVoice.getState().setPhase('speech');
+            useChat.getState().ping();
+          },
           onUtterance: (wav) => void this.transcribe(wav),
           onEnd: (reason) => this.onCaptureEnd(reason),
           onError: (message) => useVoice.getState().setError(message)
         }
       });
+      if (seq !== this.startSeq || useVoice.getState().phase !== 'arming') {
+        // stop() was called while the microphone was opening.
+        this.capture.cancel();
+        return;
+      }
       useVoice.getState().setPhase('listening');
-      this.playChime(true);
+      if (!auto) this.playChime(true);
     } catch (err) {
       const message = (err as Error).message;
       useVoice.getState().setError(message);
@@ -166,6 +189,7 @@ class VoiceController {
 
   stop(): void {
     this.stopping = true;
+    this.startSeq++;
     this.clearHandsFreeTimer();
     if (useSettings.getState().settings.voice.provider === 'browser') this.browser.stop();
     else if (this.capture.isActive) this.capture.stop('stopped');
@@ -174,6 +198,7 @@ class VoiceController {
   }
 
   cancel(): void {
+    this.startSeq++;
     this.clearHandsFreeTimer();
     this.browser.stop();
     this.capture.cancel();
@@ -224,6 +249,7 @@ class VoiceController {
       this.attentionUntil = 0;
       if (text.trim()) {
         await sendMessage(text, [], 'voice');
+        if (useVoice.getState().handsFree && !speech.isSpeaking()) this.scheduleHandsFree(350);
       } else {
         voice.setError('Transcription vide.');
         useChat.getState().setHud('idle');
@@ -237,6 +263,7 @@ class VoiceController {
       useChat.getState().setHud('error');
       Log.error('voice', `transcription failed: ${message}`);
       setTimeout(() => useChat.getState().setHud('idle'), 3000);
+      if (useVoice.getState().handsFree) this.scheduleHandsFree(3000);
     }
   }