omnius 1.0.636 → 1.0.638

package/dist/index.js

@@ -5824,6 +5824,17 @@ function managedEnv3(extra = {}) {
 function sha256File4(path16) {
   return createHash6("sha256").update(readFileSync10(path16)).digest("hex");
 }
+function speakerSupportManifest() {
+  return SPEAKER_SUPPORT_WHEELS.map((wheel) => ({
+    distribution: wheel.distribution,
+    version: wheel.version,
+    digest: `sha256:${wheel.digest}`
+  }));
+}
+function hasExpectedSpeakerSupportPackages(packages) {
+  const expected = speakerSupportManifest();
+  return Array.isArray(packages) && packages.length === expected.length && expected.every((item) => packages.some((candidate) => candidate.distribution === item.distribution && candidate.version === item.version && candidate.digest === item.digest));
+}
 function usesSystemSitePackages() {
   try {
     return /^include-system-site-packages\s*=\s*true\s*$/im.test(readFileSync10(join11(venvDir2(), "pyvenv.cfg"), "utf8"));
@@ -5834,7 +5845,7 @@ function usesSystemSitePackages() {
 function loadManifest3() {
   try {
     const manifest = JSON.parse(readFileSync10(manifestPath3(), "utf8"));
-    if (manifest.runtimeVersion !== RUNTIME_VERSION3 || manifest.model !== MODEL_NAME2 || manifest.modelRevision !== MODEL_REVISION2 || manifest.modelDigest !== `sha256:${MODEL_SHA2562}` || manifest.backend !== "onnxruntime-cpu" || manifest.onnxRuntimeVersion !== ONNXRUNTIME_VERSION || manifest.torchaudioVersion !== TORCHAUDIO_VERSION || !manifest.torchaudioWheel || !manifest.torchVersion || !manifest.torchSitePackages || !existsSync11(modelPath2()) || !existsSync11(managedPython())) {
+    if (manifest.runtimeVersion !== RUNTIME_VERSION3 || manifest.model !== MODEL_NAME2 || manifest.modelRevision !== MODEL_REVISION2 || manifest.modelDigest !== `sha256:${MODEL_SHA2562}` || manifest.backend !== "onnxruntime-cpu" || manifest.onnxRuntimeVersion !== ONNXRUNTIME_VERSION || manifest.torchaudioVersion !== TORCHAUDIO_VERSION || !hasExpectedSpeakerSupportPackages(manifest.supportPackages) || !manifest.torchaudioWheel || !manifest.torchVersion || !manifest.torchSitePackages || !existsSync11(modelPath2()) || !existsSync11(managedPython())) {
       return null;
     }
     return sha256File4(modelPath2()) === MODEL_SHA2562 ? manifest : null;
@@ -5864,6 +5875,25 @@ function requiredJetson2() {
 function bootstrapPython() {
   return process.env["OMNIUS_AUDIO_PYTHON"]?.trim() || "python3";
 }
+function processDiagnostic(error) {
+  if (error instanceof AsyncProcessError) {
+    const detail = `${error.stderr.toString("utf8")}
+${error.stdout.toString("utf8")}`.trim();
+    if (detail)
+      return detail.slice(-4e3);
+  }
+  return (error instanceof Error ? error.message : String(error)).slice(-4e3);
+}
+function describeSpeakerRuntimeProbeError(detail) {
+  const match = detail.match(/ModuleNotFoundError:\s*No module named ['"]([^'"]+)['"]/i);
+  if (match?.[1]) {
+    return `Managed speaker runtime is missing Python module '${match[1]}'. It is intentionally isolated from project/user site packages.`;
+  }
+  return detail.slice(-4e3);
+}
+function runtimeProbeError(error) {
+  return describeSpeakerRuntimeProbeError(processDiagnostic(error));
+}
 async function probeRuntime(python2) {
   if (!python2 || python2.includes("/") && !existsSync11(python2)) {
     return { available: false, providers: [] };
@@ -5903,8 +5933,8 @@ async function probeRuntime(python2) {
       torchaudioVersion: parsed.torchaudio,
       providers
     };
-  } catch {
-    return { available: false, providers: [] };
+  } catch (error) {
+    return { available: false, providers: [], error: runtimeProbeError(error) };
   }
 }
 function isExpectedJetPackTorchVersion(version5) {
@@ -5915,26 +5945,24 @@ async function inspectBootstrapPython(python2) {
     throw new Error("OMNIUS_AUDIO_PYTHON does not name a usable JetPack Python interpreter");
   }
   const script = [
-    "import json,os,sys,sysconfig,torch",
+    "import json,os,sys,sysconfig",
     "assert sys.version_info[:2] == (3, 10), 'CPython 3.10 is required by the pinned ONNX Runtime wheel'",
     "site = sysconfig.get_paths()['purelib']",
     "assert os.path.isdir(site), 'JetPack Python has no usable site-packages directory'",
-    "print(json.dumps({'python':sys.version.split()[0], 'torch':torch.__version__, 'torch_cuda':torch.version.cuda, 'site':site}))"
+    "print(json.dumps({'python':sys.version.split()[0], 'site':site}))"
   ].join("\n");
   try {
     const output2 = await execFileText2(python2, ["-c", script], { timeout: 3e4, env: managedEnv3() });
     const parsed = JSON.parse(output2.trim().split(/\r?\n/).pop() || "{}");
-    if (!parsed.python || !parsed.torch || !isExpectedJetPackTorchVersion(parsed.torch) || !String(parsed.torch_cuda ?? "").startsWith(JETPACK_CUDA_VERSION) || !parsed.site || !isAbsolute3(parsed.site) || !existsSync11(parsed.site)) {
-      throw new Error(`Expected JetPack Torch ${JETPACK_TORCH_ABI}.x built for CUDA ${JETPACK_CUDA_VERSION}; found torch=${parsed.torch ?? "missing"}, cuda=${parsed.torch_cuda ?? "missing"}`);
+    if (!parsed.python || !parsed.site || !isAbsolute3(parsed.site) || !existsSync11(parsed.site)) {
+      throw new Error("OMNIUS_AUDIO_PYTHON did not expose a usable CPython 3.10 site-packages directory");
     }
     return {
       pythonVersion: parsed.python,
-      torchVersion: parsed.torch,
-      torchCudaVersion: parsed.torch_cuda ?? null,
       torchSitePackages: resolvePath(parsed.site)
     };
   } catch (error) {
-    throw new Error(`JetPack Torch ${JETPACK_TORCH_ABI}.x/CUDA ${JETPACK_CUDA_VERSION} is required in OMNIUS_AUDIO_PYTHON. Omnius will not install or replace Torch: ` + (error instanceof Error ? error.message : String(error)));
+    throw new Error(`OMNIUS_AUDIO_PYTHON bootstrap inspection failed: ${processDiagnostic(error)}`);
   }
 }
 async function managedSitePackages(python2) {
@@ -5999,6 +6027,38 @@ async function materializeTorchaudioWheel(wheel) {
   }
   return source;
 }
+async function installPinnedSpeakerSupportDependencies(python2) {
+  for (const wheel of SPEAKER_SUPPORT_WHEELS) {
+    const filename = basename(new URL(wheel.source).pathname);
+    if (!filename.endsWith(".whl")) {
+      throw new Error(`Pinned speaker support wheel for ${wheel.distribution} has an invalid filename`);
+    }
+    const localWheel = join11(speakerEmbeddingRuntimeDir(), "artifacts", wheel.digest, filename);
+    try {
+      await downloadChecked2(wheel.source, localWheel, wheel.digest);
+      await execFileText2(python2, [
+        "-m",
+        "pip",
+        "install",
+        "--disable-pip-version-check",
+        "--no-deps",
+        "--no-index",
+        "--force-reinstall",
+        localWheel
+      ], { timeout: SETUP_TIMEOUT_MS3, env: managedEnv3() });
+      await execFileText2(python2, [
+        "-c",
+        [
+          "import importlib.metadata",
+          `import ${wheel.importName}`,
+          `assert importlib.metadata.version('${wheel.distribution}') == '${wheel.version}'`
+        ].join("\n")
+      ], { timeout: 3e4, env: managedEnv3() });
+    } catch (error) {
+      throw new Error(`Managed speaker venv could not install required support module '${wheel.importName}' (${wheel.distribution}==${wheel.version}): ${processDiagnostic(error)}`);
+    }
+  }
+}
 async function installPinnedTorchaudio(python2, wheel) {
   const localWheel = await materializeTorchaudioWheel(wheel);
   await execFileText2(python2, [
@@ -6088,6 +6148,7 @@ async function getSpeakerEmbeddingReadiness() {
       torch_version: runtime2.torchVersion,
       torch_cuda_version: runtime2.torchCudaVersion,
       torchaudio_version: runtime2.torchaudioVersion,
+      support_packages: manifest?.supportPackages,
       ...manifest ? {
         torchaudio_wheel: {
           source: manifest.torchaudioWheel.source,
@@ -6097,7 +6158,8 @@ async function getSpeakerEmbeddingReadiness() {
         torch_site_packages: manifest.torchSitePackages
       } : {},
       providers: runtime2.providers,
-      imports_verified: importsVerified
+      imports_verified: importsVerified,
+      ...runtime2.error ? { probe_error: runtime2.error } : {}
     },
     worker: {
       ready: Boolean(state3.ready && state3.worker?.exitCode === null && state3.worker.signalCode === null),
@@ -6120,12 +6182,13 @@ async function ensureSpeakerEmbeddingSetup() {
     const bootstrapRuntime = await inspectBootstrapPython(bootstrap2);
     await createManagedVenv(bootstrap2, bootstrapRuntime.torchSitePackages);
     const python2 = managedPython();
+    await installPinnedSpeakerSupportDependencies(python2);
     const torchaudioWheel = configuredTorchaudioWheel();
     await installPinnedTorchaudio(python2, torchaudioWheel);
     await installPinnedOnnxRuntime(python2);
     const runtime2 = await probeRuntime(python2);
     if (!runtime2.available || runtime2.torchaudioVersion !== TORCHAUDIO_VERSION || !runtime2.torchVersion || !isExpectedJetPackTorchVersion(runtime2.torchVersion) || !String(runtime2.torchCudaVersion ?? "").startsWith(JETPACK_CUDA_VERSION) || !runtime2.torchPath || !resolvePath(runtime2.torchPath).startsWith(bootstrapRuntime.torchSitePackages) || !runtime2.providers.includes("CPUExecutionProvider")) {
-      throw new Error("Managed Torchaudio/Torch ABI validation failed after speaker setup. The exact CPU kaldi.fbank probe must use the operator-selected JetPack Torch 2.2/CUDA 12.2 and managed torchaudio 2.2.0.");
+      throw new Error(`Managed Torchaudio/Torch ABI validation failed after speaker setup${runtime2.error ? `: ${runtime2.error}` : ""}. The exact CPU kaldi.fbank probe must use the operator-selected JetPack Torch 2.2/CUDA 12.2 and managed torchaudio 2.2.0.`);
     }
     await downloadChecked2(MODEL_URL2, modelPath2(), MODEL_SHA2562);
     const manifest = {
@@ -6136,6 +6199,7 @@ async function ensureSpeakerEmbeddingSetup() {
       backend: "onnxruntime-cpu",
       onnxRuntimeVersion: ONNXRUNTIME_VERSION,
       torchaudioVersion: runtime2.torchaudioVersion,
+      supportPackages: speakerSupportManifest(),
       torchaudioWheel,
       pythonVersion: runtime2.pythonVersion ?? bootstrapRuntime.pythonVersion,
       torchVersion: runtime2.torchVersion,
@@ -6376,7 +6440,7 @@ async function bootstrapSpeakerEmbeddingRuntime() {
   await ensureSpeakerEmbeddingSetup();
   return activateSpeakerEmbedding();
 }
-var RUNTIME_VERSION3, MODEL_NAME2, MODEL_REVISION2, MODEL_FILE, MODEL_URL2, MODEL_SHA2562, ONNXRUNTIME_VERSION, ONNXRUNTIME_WHEEL, ONNXRUNTIME_WHEEL_SHA256, TORCHAUDIO_VERSION, TORCHAUDIO_WHEEL, TORCHAUDIO_WHEEL_SHA256, JETPACK_TORCH_ABI, JETPACK_CUDA_VERSION, SETUP_TIMEOUT_MS3, START_TIMEOUT_MS3, EMBED_TIMEOUT_MS2, WESPEAKER_CAMPP_ONNX_SHA256, WESPEAKER_CAMPP_MODEL_REVISION, WESPEAKER_CAMPP_ONNXRUNTIME_VERSION, WESPEAKER_CAMPP_EMBEDDING_SET_SLUG, state3;
+var RUNTIME_VERSION3, MODEL_NAME2, MODEL_REVISION2, MODEL_FILE, MODEL_URL2, MODEL_SHA2562, ONNXRUNTIME_VERSION, ONNXRUNTIME_WHEEL, ONNXRUNTIME_WHEEL_SHA256, TORCHAUDIO_VERSION, TORCHAUDIO_WHEEL, TORCHAUDIO_WHEEL_SHA256, TYPING_EXTENSIONS_VERSION, TYPING_EXTENSIONS_WHEEL, TYPING_EXTENSIONS_WHEEL_SHA256, JETPACK_TORCH_ABI, JETPACK_CUDA_VERSION, SETUP_TIMEOUT_MS3, START_TIMEOUT_MS3, EMBED_TIMEOUT_MS2, WESPEAKER_CAMPP_ONNX_SHA256, WESPEAKER_CAMPP_MODEL_REVISION, WESPEAKER_CAMPP_ONNXRUNTIME_VERSION, WESPEAKER_CAMPP_EMBEDDING_SET_SLUG, SPEAKER_SUPPORT_WHEELS, state3;
 var init_speaker_embedding_runtime = __esm({
   "packages/execution/dist/speaker-embedding-runtime.js"() {
     "use strict";
@@ -6384,7 +6448,7 @@ var init_speaker_embedding_runtime = __esm({
     init_process_async();
     init_model_store();
     init_venv_paths();
-    RUNTIME_VERSION3 = 2;
+    RUNTIME_VERSION3 = 3;
     MODEL_NAME2 = "wespeaker-voxceleb-campplus";
     MODEL_REVISION2 = "acf623ad8ca746e50baa432255cf8fc57c669c45";
     MODEL_FILE = "voxceleb_CAM++.onnx";
@@ -6396,6 +6460,9 @@ var init_speaker_embedding_runtime = __esm({
     TORCHAUDIO_VERSION = "2.2.0";
     TORCHAUDIO_WHEEL = "https://files.pythonhosted.org/packages/45/a5/74d8a03fdf47cf89e9a2f6c58a65ffe4b392e8cfa503f148baec43377f24/torchaudio-2.2.0-cp310-cp310-manylinux2014_aarch64.whl";
     TORCHAUDIO_WHEEL_SHA256 = "d4ea094b8721a361982db062ee993f2a6f71dfe16f62a84f8900b2364f33a2e4";
+    TYPING_EXTENSIONS_VERSION = "4.10.0";
+    TYPING_EXTENSIONS_WHEEL = "https://files.pythonhosted.org/packages/f9/de/dc04a3ea60b22624b51c703a84bbe0184abcd1d0b9bc8074b5d6b7ab90bb/typing_extensions-4.10.0-py3-none-any.whl";
+    TYPING_EXTENSIONS_WHEEL_SHA256 = "69b1a937c3a517342112fb4c6df7e72fc39a38e7891a5730ed4985b5214b5475";
     JETPACK_TORCH_ABI = "2.2";
     JETPACK_CUDA_VERSION = "12.2";
     SETUP_TIMEOUT_MS3 = 10 * 6e4;
@@ -6405,6 +6472,15 @@ var init_speaker_embedding_runtime = __esm({
     WESPEAKER_CAMPP_MODEL_REVISION = MODEL_REVISION2;
     WESPEAKER_CAMPP_ONNXRUNTIME_VERSION = ONNXRUNTIME_VERSION;
     WESPEAKER_CAMPP_EMBEDDING_SET_SLUG = "speaker-identity-wespeaker-voxceleb-campplus-acf623ad8ca746e50baa432255cf8fc57c669c45-fbank80-cmn-fullclip-l2";
+    SPEAKER_SUPPORT_WHEELS = [
+      {
+        distribution: "typing_extensions",
+        importName: "typing_extensions",
+        version: TYPING_EXTENSIONS_VERSION,
+        source: TYPING_EXTENSIONS_WHEEL,
+        digest: TYPING_EXTENSIONS_WHEEL_SHA256
+      }
+    ];
     state3 = {
       setup: null,
       start: null,
@@ -688266,9 +688342,12 @@ __export(listen_exports, {
   getListenEngine: () => getListenEngine,
   getListenLiveState: () => getListenLiveState,
   isAudioPath: () => isAudioPath,
+  isEggReSpeakerPulseSource: () => isEggReSpeakerPulseSource,
+  isKnownVirtualPulseRemap: () => isKnownVirtualPulseRemap,
   isTranscribablePath: () => isTranscribablePath,
   isVideoPath: () => isVideoPath,
   resolveAsrConsensusModel: () => resolveAsrConsensusModel,
+  selectPreferredPulseMicSource: () => selectPreferredPulseMicSource,
   transcribeFileViaWhisper: () => transcribeFileViaWhisper,
   waitForTranscribeCli: () => waitForTranscribeCli
 });
@@ -688351,10 +688430,30 @@ function renderMicSpectrum(samples) {
   }
   return bands.join("");
 }
+function isEggReSpeakerPulseSource(source) {
+  return source.channels >= 6 && /(?:^|:)alsa_input\.usb-SEEED_ReSpeaker_4_Mic_Array__UAC1\.0_-00\.multichannel-input$/i.test(source.name);
+}
+function isKnownVirtualPulseRemap(sourceName) {
+  const name10 = String(sourceName ?? "").replace(/^pulse:/i, "");
+  return /^nx_remapped_out(?:[._-].*)?$/i.test(name10);
+}
+function selectPreferredPulseMicSource(sources) {
+  const candidates = sources.filter((source) => !/\.monitor$/i.test(source.name));
+  const rank = (source) => {
+    let score = 0;
+    if (isEggReSpeakerPulseSource(source)) score -= 1e3;
+    if (isKnownVirtualPulseRemap(source.name)) score += 100;
+    if (source.state === "RUNNING") score -= 20;
+    if (source.state === "SUSPENDED") score += 5;
+    if (/usb|mic|array|respeaker|seeed/i.test(source.name)) score -= 10;
+    return score;
+  };
+  return [...candidates].sort((a2, b) => rank(a2) - rank(b))[0];
+}
 async function resolveMicSource() {
   const configured = String(process.env["OMNIUS_MIC_DEVICE"] ?? "").trim();
   if (configured.startsWith("hw:") || configured.startsWith("plughw:")) {
-    return { alsaDevice: configured, label: configured };
+    return { alsaDevice: configured, sourceKind: "configured-alsa", label: configured };
   }
   const configuredPulse = configured.startsWith("pulse:") ? configured.slice("pulse:".length) : configured || void 0;
   let sources = [];
@@ -688380,21 +688479,19 @@ async function resolveMicSource() {
     return {
       pulseSource: configuredPulse,
       channels: match?.channels,
+      sourceKind: "configured-pulse",
       label: `pulse:${configuredPulse}`
     };
   }
-  const candidates = sources.filter((s2) => !/\.monitor$/i.test(s2.name));
-  if (candidates.length === 0) return null;
-  const rank = (s2) => {
-    let score = 0;
-    if (s2.state === "RUNNING") score -= 20;
-    if (s2.state === "SUSPENDED") score += 5;
-    if (/usb|mic|array|respeaker|seeed/i.test(s2.name)) score -= 10;
-    return score;
+  const best = selectPreferredPulseMicSource(sources);
+  if (!best) return null;
+  const isReSpeaker = isEggReSpeakerPulseSource(best);
+  return {
+    pulseSource: best.name,
+    channels: best.channels,
+    sourceKind: isReSpeaker ? "respeaker-usb" : "pulse-auto",
+    label: isReSpeaker ? `pulse:${best.name} (ReSpeaker physical 6ch; channel 0)` : `pulse:${best.name}`
   };
-  candidates.sort((a2, b) => rank(a2) - rank(b));
-  const best = candidates[0];
-  return { pulseSource: best.name, channels: best.channels, label: `pulse:${best.name}` };
 }
 async function ensureSourceCaptureVolume(source) {
   try {
@@ -688431,7 +688528,10 @@ async function findMicCaptureCommand() {
       return {
         cmd: "arecord",
         args: ["-D", resolved.alsaDevice, "-f", "S16_LE", "-r", "16000", "-c", "1", "-t", "raw", "-q", "-"],
-        device: resolved.alsaDevice
+        device: resolved.alsaDevice,
+        sourceKind: resolved.sourceKind,
+        sourceChannels: null,
+        selectedChannel: null
       };
     }
     if (resolved?.pulseSource && await commandExists2("ffmpeg")) {
@@ -688455,14 +688555,20 @@ async function findMicCaptureCommand() {
           "quiet",
           "pipe:1"
         ],
-        device: resolved.label
+        device: resolved.label,
+        sourceKind: resolved.sourceKind,
+        sourceChannels: resolved.channels ?? null,
+        selectedChannel: (resolved.channels ?? 1) > 2 ? 0 : null
       };
     }
     if (await commandExists2("arecord")) {
       return {
         cmd: "arecord",
         args: ["-f", "S16_LE", "-r", "16000", "-c", "1", "-t", "raw", "-q", "-"],
-        device: "default (arecord)"
+        device: "default (arecord)",
+        sourceKind: "fallback",
+        sourceChannels: null,
+        selectedChannel: null
       };
     }
   }
@@ -688484,7 +688590,10 @@ async function findMicCaptureCommand() {
           "signed-integer",
           "-"
         ],
-        device: "default (sox)"
+        device: "default (sox)",
+        sourceKind: "fallback",
+        sourceChannels: null,
+        selectedChannel: null
       };
     }
   }
@@ -688507,7 +688616,10 @@ async function findMicCaptureCommand() {
           "quiet",
           "pipe:1"
         ],
-        device: "pulse:default"
+        device: "pulse:default",
+        sourceKind: "fallback",
+        sourceChannels: null,
+        selectedChannel: null
       };
     } else if (platform13 === "darwin") {
       return {
@@ -688527,7 +688639,10 @@ async function findMicCaptureCommand() {
           "quiet",
           "pipe:1"
         ],
-        device: "avfoundation:0"
+        device: "avfoundation:0",
+        sourceKind: "fallback",
+        sourceChannels: null,
+        selectedChannel: null
       };
     }
   }
@@ -688851,6 +688966,9 @@ var init_listen = __esm({
       backend: "",
       model: "",
       micDevice: "",
+      micSourceKind: "unknown",
+      micSourceChannels: null,
+      micSelectedChannel: null,
       micLevelDb: null,
       noiseFloorDb: null,
       speechActive: false,
@@ -689284,7 +689402,12 @@ ${text3}`.slice(-2e3);
           stdio: ["pipe", "pipe", "pipe"],
           env: { ...process.env }
         });
-        updateListenLiveState({ micDevice: micCmd.device });
+        updateListenLiveState({
+          micDevice: micCmd.device,
+          micSourceKind: micCmd.sourceKind,
+          micSourceChannels: micCmd.sourceChannels,
+          micSelectedChannel: micCmd.selectedChannel
+        });
         this.micProcess.stdout?.on("data", (chunk) => {
           this.meterMicChunk(chunk);
           if (this.active && !this.paused && this.liveTranscriber) {
@@ -692573,14 +692696,28 @@ function isAudioOutputMonitorDevice(device2) {
   if (!device2) return false;
   return isAudioOutputMonitorId(device2.id) || isAudioOutputMonitorId(device2.label) || isAudioOutputMonitorId(device2.detail);
 }
+function isEggReSpeakerLiveInput(device2) {
+  if (!device2?.id.startsWith("pulse:")) return false;
+  return isEggReSpeakerPulseSource({
+    name: device2.id.slice("pulse:".length),
+    channels: liveDeviceChannels(device2) ?? 0
+  });
+}
+function shouldReplaceStaleVirtualAudioSelection(configured, usable) {
+  return Boolean(configured && isKnownVirtualPulseRemap(configured) && usable.some(isEggReSpeakerLiveInput));
+}
 function preferredLiveAudioInputId(devices, configured) {
   const usable = devices.filter((device2) => device2.id && !isAudioOutputMonitorDevice(device2));
-  if (configured && usable.some((device2) => device2.id === configured)) return configured;
+  if (configured && usable.some((device2) => device2.id === configured) && !shouldReplaceStaleVirtualAudioSelection(configured, usable)) {
+    return configured;
+  }
   return rankLiveAudioInputDevices(devices, configured)[0]?.id ?? firstDeviceId(devices);
 }
 function liveAudioDeviceRank(device2, configured) {
   let rank = 0;
   if (device2.id === configured) rank -= 40;
+  if (isEggReSpeakerLiveInput(device2)) rank -= 1e3;
+  if (isKnownVirtualPulseRemap(device2.id)) rank += 100;
   if (isAudioOutputMonitorDevice(device2)) rank += 500;
   if (device2.source === "pulse" || device2.source === "pipewire") rank -= 20;
   if (device2.source === "alsa") rank -= 10;
@@ -693202,7 +693339,10 @@ var init_live_sensors = __esm({
         };
         if (!this.config.selectedCamera) this.config.selectedCamera = firstDeviceId(video);
         this.config.cameraStreams = this.reconcileCameraStreams(video);
-        if (!this.config.selectedAudioInput) this.config.selectedAudioInput = preferredLiveAudioInputId(inputs.devices);
+        const preferredAudioInput = preferredLiveAudioInputId(inputs.devices, this.config.selectedAudioInput);
+        if (!this.config.selectedAudioInput || shouldReplaceStaleVirtualAudioSelection(this.config.selectedAudioInput, inputs.devices)) {
+          this.config.selectedAudioInput = preferredAudioInput;
+        }
         if (!this.config.selectedAudioOutput) this.config.selectedAudioOutput = firstDeviceId(outputs.devices);
         this.persist();
         return this.devices;
@@ -792531,6 +792671,10 @@ function getRuntimeStatus() {
     asrBackend: asrEngineId === "vibevoice-transformers" ? "vibevoice-transformers" : listenState.backend,
     asrPhase: asrEngineId === "vibevoice-transformers" ? vibePhase : listenState.phase,
     asrReady: asrEngineId === "vibevoice-transformers" ? vibe.active : Boolean(_listenEngine?.isActive),
+    micDevice: listenState.micDevice || null,
+    micSourceKind: listenState.micSourceKind,
+    micSourceChannels: listenState.micSourceChannels,
+    micSelectedChannel: listenState.micSelectedChannel,
     clientCount: _clients2.size,
     loadedAt: _loadedAt,
     lastError: _lastError
@@ -828308,7 +828452,7 @@ function getOpenApiSpec() {
       "/v1/agents": { get: { summary: "List agent types", tags: ["Tools"], responses: { 200: { description: "Agent type registry" } } } },
       "/v1/engines": { get: { summary: "List long-running engines", tags: ["Engines"], responses: { 200: { description: "Engine status + state files" } } } },
       // ───── Voice / Audio surface (live) ─────
-      "/v1/voice/state": { get: { summary: "Voice runtime status (engine load state, active model, current clone ref, client count)", tags: ["Voice"], responses: { 200: { description: "Runtime status snapshot" } } } },
+      "/v1/voice/state": { get: { summary: "Voice runtime status (engine/model state plus actual selected microphone source and channel provenance)", tags: ["Voice"], responses: { 200: { description: "Runtime status snapshot" } } } },
       "/v1/voice/start": { post: { summary: "Enable and warm the daemon voice runtime; optional body {modelId|model|voice}", tags: ["Voice"], responses: { 200: { description: "Voice enabled and ready" }, 400: { description: "Unknown model" }, 409: { description: "Model disabled by configuration" }, 500: { description: "Warmup failed" } } } },
       "/v1/voice/stop": { post: { summary: "Pause daemon voice input immediately (TTS models remain warm)", tags: ["Voice"], responses: { 200: { description: "Voice input paused" }, 500: { description: "Stop failed" } } } },
       "/v1/voice/models": { get: { summary: "List TTS voice models with backend metadata and managed readiness", tags: ["Voice"], responses: { 200: { description: "Voice model list, active selection, and Voxtral pull/deploy state" } } } },
@@ -828436,7 +828580,7 @@ function getOpenApiSpec() {
         post: {
           summary: "Provision and activate one role-typed audio embedding runtime",
           tags: ["Audio"],
-          description: "Admin-only. Requires kind=acoustic|speaker|semantic in the query (canonical) or JSON body. acoustic reuses the pinned JetPack YAMNet/TensorRT setup; speaker installs the isolated CPU-only WeSpeaker CAM++ ONNX Runtime and pinned model, then installs only the checksummed CPython 3.10/aarch64 Torchaudio 2.2.0 wheel into ~/.omnius/runtimes/audio/speaker/venv with --no-deps/--no-index. It retains --system-site-packages and links the selected OMNIUS_AUDIO_PYTHON JetPack Torch 2.2/CUDA 12.2 provider without writing to it; setup proves the Torch/Torchaudio ABI and CPU kaldi.fbank before readiness. An operator may replace that wheel only by supplying both OMNIUS_SPEAKER_TORCHAUDIO_WHEEL and OMNIUS_SPEAKER_TORCHAUDIO_WHEEL_SHA256. semantic installs the isolated JetPack CUDA CLAP dependencies and pinned model. This is the only REST operation allowed to provision. It activates and warms only the requested role worker; no role is substituted for another.",
+          description: "Admin-only. Requires kind=acoustic|speaker|semantic in the query (canonical) or JSON body. acoustic reuses the pinned JetPack YAMNet/TensorRT setup; speaker installs the isolated CPU-only WeSpeaker CAM++ ONNX Runtime and pinned model, then installs only checksummed wheels into ~/.omnius/runtimes/audio/speaker/venv: CPython 3.10/aarch64 Torchaudio 2.2.0 and the self-contained typing_extensions support dependency, each with --no-deps/--no-index. It retains --system-site-packages and links the selected OMNIUS_AUDIO_PYTHON JetPack Torch 2.2/CUDA 12.2 provider without writing to it; setup proves the Torch/Torchaudio ABI and CPU kaldi.fbank before readiness. An operator may replace the Torchaudio wheel only by supplying both OMNIUS_SPEAKER_TORCHAUDIO_WHEEL and OMNIUS_SPEAKER_TORCHAUDIO_WHEEL_SHA256. semantic installs the isolated JetPack CUDA CLAP dependencies and pinned model. This is the only REST operation allowed to provision. It activates and warms only the requested role worker; no role is substituted for another.",
           parameters: [{ name: "kind", in: "query", required: false, schema: { type: "string", enum: ["acoustic", "speaker", "semantic"] }, description: "Canonical required role selector; JSON body.kind is accepted for compatibility." }],
           requestBody: { required: false, content: { "application/json": { schema: { type: "object", properties: { kind: { type: "string", enum: ["acoustic", "speaker", "semantic"] } } } } } },
           responses: { 200: { description: "Requested role runtime provisioned and warm." }, 202: { description: "Provisioning finished but requested role did not prove warm; poll its health URL." }, 400: { description: "kind is missing or invalid." }, 403: { description: "Admin scope required." }, 500: { description: "Role-specific setup/preflight failed." } }
@@ -830030,6 +830174,71 @@ function getBackendTimeoutMs(perRequestSeconds) {
   }
   return BACKEND_TIMEOUT_DEFAULT_MS;
 }
+function getAgentLoopToolTimeoutMs() {
+  const configured = Number(process.env["OMNIUS_AGENT_LOOP_TOOL_TIMEOUT_MS"]);
+  if (!Number.isFinite(configured) || configured <= 0) {
+    return AGENT_LOOP_TOOL_TIMEOUT_DEFAULT_MS;
+  }
+  return Math.max(1e3, Math.min(Math.floor(configured), AGENT_LOOP_TOOL_TIMEOUT_MAX_MS));
+}
+function cancelAgentLoopTool(tool) {
+  if (!tool || typeof tool.cancel !== "function" || cancelledAgentLoopTools.has(tool)) {
+    return;
+  }
+  cancelledAgentLoopTools.add(tool);
+  try {
+    const cancelled = tool.cancel();
+    if (cancelled && typeof cancelled.then === "function") {
+      void Promise.resolve(cancelled).catch(() => void 0);
+    }
+  } catch {
+  }
+}
+async function executeAgentLoopToolBounded(tool, args, timeoutMs, clientDisconnect) {
+  const startedAt2 = performance.now();
+  const execution = Promise.resolve().then(() => tool.execute(args)).then(
+    (result) => ({ kind: "result", result }),
+    (error) => ({ kind: "error", error })
+  );
+  let timer;
+  try {
+    const outcome = await Promise.race([
+      execution,
+      new Promise((resolveTimeout) => {
+        timer = setTimeout(() => resolveTimeout({ kind: "timeout" }), timeoutMs);
+        timer.unref();
+      }),
+      clientDisconnect.then(() => ({ kind: "client_disconnected" }))
+    ]);
+    if (outcome.kind === "result") {
+      return { result: outcome.result, clientDisconnected: false };
+    }
+    if (outcome.kind === "error") throw outcome.error;
+    cancelAgentLoopTool(tool);
+    if (outcome.kind === "client_disconnected") {
+      return {
+        result: {
+          success: false,
+          output: "",
+          error: "agent_loop_client_disconnected",
+          durationMs: performance.now() - startedAt2
+        },
+        clientDisconnected: true
+      };
+    }
+    return {
+      result: {
+        success: false,
+        output: "",
+        error: `agent_loop_tool_timeout: daemon tool exceeded ${timeoutMs}ms`,
+        durationMs: performance.now() - startedAt2
+      },
+      clientDisconnected: false
+    };
+  } finally {
+    if (timer) clearTimeout(timer);
+  }
+}
 function getModelListTimeoutMs() {
   const envS = process.env["OMNIUS_MODEL_LIST_TIMEOUT_S"];
   if (envS) {
@@ -832379,361 +832588,447 @@ async function handleApiTags(res) {
 }
 async function runAgentLoopChatCompletions(opts) {
   const { req: req3, res, ollamaUrl, requestBody, perReqTimeoutS } = opts;
-  const startMs = Date.now();
-  const reqTimeoutMs = getBackendTimeoutMs(perReqTimeoutS);
-  const totalDeadline = startMs + (perReqTimeoutS && perReqTimeoutS > 0 ? Math.min(perReqTimeoutS * 1e3, BACKEND_TIMEOUT_MAX_MS) : 30 * 60 * 1e3);
-  const model = requestBody["model"] || "unknown";
-  const route = resolveModelEndpoint(model);
-  const targetUrl = route?.endpoint.url ?? ollamaUrl;
-  const targetType = route?.endpoint.type ?? loadConfig().backendType ?? "ollama";
-  const originalModel = route?.originalId ?? model;
-  const maxTurns = typeof requestBody["max_turns"] === "number" ? Math.max(1, Math.min(64, requestBody["max_turns"])) : 8;
-  const remoteIp = (req3.socket?.remoteAddress || "").replace(/^::ffff:/, "");
-  const origin = /^(127\.\d+\.\d+\.\d+|::1|localhost)$/.test(remoteIp) ? "loopback" : "remote";
-  const reqAuth = req3;
-  const scope = reqAuth._authScope ?? (origin === "loopback" ? "admin" : "read");
-  const profileName = typeof requestBody["profile"] === "string" ? requestBody["profile"] : typeof req3.headers["x-tool-profile"] === "string" ? req3.headers["x-tool-profile"] : req3._authProfile;
-  const activeProfileMeta = profileName ? loadProfileWithMeta(
-    profileName,
-    typeof req3.headers["x-profile-password"] === "string" ? req3.headers["x-profile-password"] : void 0,
-    process.cwd()
-  ) : null;
-  if (profileName && !activeProfileMeta) {
-    jsonResponse(res, 400, {
-      error: "Profile not found or wrong password",
-      profile: profileName,
-      search_order: [
-        "preset",
-        `${process.cwd()}/.omnius/profiles/${profileName}.json`,
-        "~/.omnius/profiles"
-      ]
-    });
-    return;
-  }
-  const execMod = await Promise.resolve().then(() => (init_dist5(), dist_exports2)).catch(() => null);
-  if (!execMod) {
-    jsonResponse(res, 500, { error: "Execution module unavailable" });
-    return;
-  }
-  const classify2 = execMod.classifyTool;
-  const canInvoke = execMod.canInvokeTool;
-  const daemonTools = /* @__PURE__ */ new Map();
-  for (const [classKey, value2] of Object.entries(execMod)) {
-    if (typeof value2 !== "function") continue;
-    const proto = value2.prototype;
-    if (!proto || typeof proto.execute !== "function") continue;
-    let probe = null;
-    try {
-      probe = new value2();
-    } catch {
+  let activeTool = null;
+  let clientDisconnected = false;
+  let resolveClientDisconnect;
+  const clientDisconnect = new Promise((resolveDisconnect) => {
+    resolveClientDisconnect = resolveDisconnect;
+  });
+  const cancelActiveTool = () => cancelAgentLoopTool(activeTool);
+  const markClientDisconnected = () => {
+    if (clientDisconnected) return;
+    clientDisconnected = true;
+    cancelActiveTool();
+    resolveClientDisconnect();
+  };
+  const isClientDisconnected = () => {
+    if (!clientDisconnected && (req3.aborted || res.destroyed && !res.writableEnded)) {
+      markClientDisconnected();
+    }
+    return clientDisconnected;
+  };
+  const onRequestAborted = () => markClientDisconnected();
+  const onRequestClose = () => {
+    if (!req3.complete && !res.writableEnded) markClientDisconnected();
+  };
+  const onResponseClose = () => {
+    if (!res.writableEnded) markClientDisconnected();
+  };
+  const onRequestError = () => markClientDisconnected();
+  const onResponseError = () => markClientDisconnected();
+  req3.once("aborted", onRequestAborted);
+  req3.once("close", onRequestClose);
+  req3.once("error", onRequestError);
+  res.once("close", onResponseClose);
+  res.once("error", onResponseError);
+  try {
+    const startMs = Date.now();
+    const reqTimeoutMs = getBackendTimeoutMs(perReqTimeoutS);
+    const totalDeadline = startMs + AGENT_LOOP_TOTAL_MAX_MS;
+    const model = requestBody["model"] || "unknown";
+    const route = resolveModelEndpoint(model);
+    const targetUrl = route?.endpoint.url ?? ollamaUrl;
+    const targetType = route?.endpoint.type ?? loadConfig().backendType ?? "ollama";
+    const originalModel = route?.originalId ?? model;
+    const maxTurns = typeof requestBody["max_turns"] === "number" ? Math.max(1, Math.min(64, requestBody["max_turns"])) : 8;
+    const remoteIp = (req3.socket?.remoteAddress || "").replace(/^::ffff:/, "");
+    const origin = /^(127\.\d+\.\d+\.\d+|::1|localhost)$/.test(remoteIp) ? "loopback" : "remote";
+    const reqAuth = req3;
+    const scope = reqAuth._authScope ?? (origin === "loopback" ? "admin" : "read");
+    const profileName = typeof requestBody["profile"] === "string" ? requestBody["profile"] : typeof req3.headers["x-tool-profile"] === "string" ? req3.headers["x-tool-profile"] : req3._authProfile;
+    const activeProfileMeta = profileName ? loadProfileWithMeta(
+      profileName,
+      typeof req3.headers["x-profile-password"] === "string" ? req3.headers["x-profile-password"] : void 0,
+      process.cwd()
+    ) : null;
+    if (profileName && !activeProfileMeta) {
+      jsonResponse(res, 400, {
+        error: "Profile not found or wrong password",
+        profile: profileName,
+        search_order: [
+          "preset",
+          `${process.cwd()}/.omnius/profiles/${profileName}.json`,
+          "~/.omnius/profiles"
+        ]
+      });
+      return;
+    }
+    const execMod = await Promise.resolve().then(() => (init_dist5(), dist_exports2)).catch(() => null);
+    if (isClientDisconnected()) return;
+    if (!execMod) {
+      jsonResponse(res, 500, { error: "Execution module unavailable" });
+      return;
+    }
+    const classify2 = execMod.classifyTool;
+    const canInvoke = execMod.canInvokeTool;
+    const daemonTools = /* @__PURE__ */ new Map();
+    for (const [classKey, value2] of Object.entries(execMod)) {
+      if (typeof value2 !== "function") continue;
+      const proto = value2.prototype;
+      if (!proto || typeof proto.execute !== "function") continue;
+      let probe = null;
       try {
-        probe = new value2(process.cwd());
+        probe = new value2();
       } catch {
-        probe = null;
+        try {
+          probe = new value2(process.cwd());
+        } catch {
+          probe = null;
+        }
+      }
+      if (probe?.name && typeof probe.name === "string") {
+        daemonTools.set(probe.name, { ToolClass: value2, classKey });
       }
     }
-    if (probe?.name && typeof probe.name === "string") {
-      daemonTools.set(probe.name, { ToolClass: value2, classKey });
-    }
-  }
-  const includeDaemonTools = Array.isArray(requestBody["include_daemon_tools"]) ? requestBody["include_daemon_tools"].filter(
-    (s2) => typeof s2 === "string"
-  ) : [];
-  const advertisedDaemonNames = /* @__PURE__ */ new Set();
-  if (includeDaemonTools.length > 0) {
-    for (const [toolName, { ToolClass: _T }] of daemonTools.entries()) {
-      void _T;
-      const sec = classify2(toolName);
-      if (!sec) continue;
-      if (activeProfileMeta && !isToolAllowed(activeProfileMeta.profile, toolName))
-        continue;
-      if (!includeDaemonTools.includes(sec.requires_scope)) continue;
-      if (canInvoke({ toolName, origin, scope }) !== null) continue;
-      advertisedDaemonNames.add(toolName);
+    const includeDaemonTools = Array.isArray(requestBody["include_daemon_tools"]) ? requestBody["include_daemon_tools"].filter(
+      (s2) => typeof s2 === "string"
+    ) : [];
+    const advertisedDaemonNames = /* @__PURE__ */ new Set();
+    if (includeDaemonTools.length > 0) {
+      for (const [toolName, { ToolClass: _T }] of daemonTools.entries()) {
+        void _T;
+        const sec = classify2(toolName);
+        if (!sec) continue;
+        if (activeProfileMeta && !isToolAllowed(activeProfileMeta.profile, toolName))
+          continue;
+        if (!includeDaemonTools.includes(sec.requires_scope)) continue;
+        if (canInvoke({ toolName, origin, scope }) !== null) continue;
+        advertisedDaemonNames.add(toolName);
+      }
     }
-  }
-  const callerTools = Array.isArray(requestBody["tools"]) ? requestBody["tools"].filter(
-    (t2) => t2?.type === "function" && typeof t2?.function?.name === "string"
-  ) : [];
-  const callerToolNames = new Set(callerTools.map((t2) => t2.function.name));
-  const daemonToolEntries = [];
-  for (const name10 of advertisedDaemonNames) {
-    if (callerToolNames.has(name10)) continue;
-    const meta = daemonTools.get(name10);
-    if (!meta) continue;
-    let inst = null;
-    try {
-      inst = new meta.ToolClass();
-    } catch {
+    const callerTools = Array.isArray(requestBody["tools"]) ? requestBody["tools"].filter(
+      (t2) => t2?.type === "function" && typeof t2?.function?.name === "string"
+    ) : [];
+    const callerToolNames = new Set(callerTools.map((t2) => t2.function.name));
+    const daemonToolEntries = [];
+    for (const name10 of advertisedDaemonNames) {
+      if (callerToolNames.has(name10)) continue;
+      const meta = daemonTools.get(name10);
+      if (!meta) continue;
+      let inst = null;
       try {
-        inst = new meta.ToolClass(process.cwd());
+        inst = new meta.ToolClass();
       } catch {
-        inst = null;
+        try {
+          inst = new meta.ToolClass(process.cwd());
+        } catch {
+          inst = null;
+        }
       }
+      if (!inst) continue;
+      daemonToolEntries.push({
+        type: "function",
+        function: {
+          name: name10,
+          description: typeof inst.description === "string" ? inst.description : "",
+          parameters: inst.parameters ?? { type: "object", properties: {} }
+        }
+      });
     }
-    if (!inst) continue;
-    daemonToolEntries.push({
-      type: "function",
-      function: {
-        name: name10,
-        description: typeof inst.description === "string" ? inst.description : "",
-        parameters: inst.parameters ?? { type: "object", properties: {} }
-      }
-    });
-  }
-  const mergedTools = [...callerTools, ...daemonToolEntries];
-  const messages2 = Array.isArray(requestBody["messages"]) ? requestBody["messages"].slice() : [];
-  const promptTemplate = typeof requestBody["prompt_template"] === "string" ? requestBody["prompt_template"] : null;
-  if (promptTemplate === "factual-first" && advertisedDaemonNames.has("web_search")) {
-    const SYSTEM_FACTUAL_FIRST = [
-      "[POLICY: factual-first]",
-      "For any user question that asks for a fact, statistic, current event,",
-      "person/place/product detail, or anything you might be tempted to answer",
-      "from training data alone — your FIRST tool_call MUST be `web_search`",
-      "with a query derived from the user's question. Do NOT answer from",
-      "training data. After web_search returns, optionally call `web_fetch`",
-      "on the most relevant URL, then synthesize a grounded answer citing",
-      "the source. Decline to answer factual questions if web_search is",
-      "unavailable."
-    ].join(" ");
-    const firstSystemIdx = messages2.findIndex((m2) => m2.role === "system");
-    if (firstSystemIdx >= 0 && typeof messages2[firstSystemIdx].content === "string") {
-      messages2[firstSystemIdx] = {
-        ...messages2[firstSystemIdx],
-        content: `${SYSTEM_FACTUAL_FIRST}
+    const mergedTools = [...callerTools, ...daemonToolEntries];
+    const forceNoThinkForTools = targetType === "ollama" && mergedTools.length > 0;
+    const messages2 = Array.isArray(requestBody["messages"]) ? requestBody["messages"].slice() : [];
+    const promptTemplate = typeof requestBody["prompt_template"] === "string" ? requestBody["prompt_template"] : null;
+    if (promptTemplate === "factual-first" && advertisedDaemonNames.has("web_search")) {
+      const SYSTEM_FACTUAL_FIRST = [
+        "[POLICY: factual-first]",
+        "For any user question that asks for a fact, statistic, current event,",
+        "person/place/product detail, or anything you might be tempted to answer",
+        "from training data alone — your FIRST tool_call MUST be `web_search`",
+        "with a query derived from the user's question. Do NOT answer from",
+        "training data. After web_search returns, optionally call `web_fetch`",
+        "on the most relevant URL, then synthesize a grounded answer citing",
+        "the source. Decline to answer factual questions if web_search is",
+        "unavailable."
+      ].join(" ");
+      const firstSystemIdx = messages2.findIndex((m2) => m2.role === "system");
+      if (firstSystemIdx >= 0 && typeof messages2[firstSystemIdx].content === "string") {
+        messages2[firstSystemIdx] = {
+          ...messages2[firstSystemIdx],
+          content: `${SYSTEM_FACTUAL_FIRST}
 
 ${messages2[firstSystemIdx].content}`
-      };
-    } else {
-      messages2.unshift({ role: "system", content: SYSTEM_FACTUAL_FIRST });
-    }
-  }
-  const chatId = `chatcmpl-${randomBytes30(12).toString("hex")}`;
-  const turnsLog = [];
-  for (let turn = 1; turn <= maxTurns; turn++) {
-    if (Date.now() > totalDeadline) {
-      jsonResponse(res, 504, {
-        error: "agent_loop timed out",
-        turn,
-        elapsed_ms: Date.now() - startMs,
-        timeout_s: perReqTimeoutS ?? null
-      });
-      return;
+        };
+      } else {
+        messages2.unshift({ role: "system", content: SYSTEM_FACTUAL_FIRST });
+      }
     }
-    const turnBody = {
-      ...requestBody,
-      model: originalModel,
-      messages: messages2,
-      tools: mergedTools.length > 0 ? mergedTools : void 0,
-      stream: false,
-      // Strip our own loop-control fields so they don't confuse the backend
-      agent_loop: void 0,
-      include_daemon_tools: void 0,
-      max_turns: void 0,
-      timeout_s: void 0,
-      realtime: void 0,
-      realtime_options: void 0,
-      realtime_max_history_messages: void 0,
-      realtime_max_tokens: void 0
-    };
-    for (const k of Object.keys(turnBody))
-      if (turnBody[k] === void 0) delete turnBody[k];
-    let parsed;
-    try {
-      const path16 = targetType === "vllm" || targetType === "openai" ? "/v1/chat/completions" : "/v1/chat/completions";
-      const result = await ollamaRequest(
-        targetUrl,
-        path16,
-        "POST",
-        JSON.stringify(turnBody),
-        reqTimeoutMs,
-        route?.endpoint
-      );
-      if (result.status !== 200) {
-        jsonResponse(res, result.status, {
-          error: "Backend request failed during agent_loop",
+    const chatId = `chatcmpl-${randomBytes30(12).toString("hex")}`;
+    const turnsLog = [];
+    for (let turn = 1; turn <= maxTurns; turn++) {
+      if (isClientDisconnected()) return;
+      if (Date.now() > totalDeadline) {
+        jsonResponse(res, 504, {
+          error: "agent_loop timed out",
           turn,
-          details: result.body.slice(0, 500)
+          elapsed_ms: Date.now() - startMs,
+          timeout_s: perReqTimeoutS ?? null
         });
         return;
       }
-      parsed = JSON.parse(result.body);
-    } catch (err) {
-      jsonResponse(res, 502, {
-        error: "Backend proxy error during agent_loop",
-        message: err instanceof Error ? err.message : String(err),
-        turn
-      });
-      return;
-    }
-    const choice = parsed?.choices?.[0];
-    const message2 = choice?.message || { role: "assistant", content: "" };
-    const toolCalls = Array.isArray(message2.tool_calls) ? message2.tool_calls : [];
-    if (toolCalls.length === 0) {
-      jsonResponse(res, 200, {
-        ...parsed,
-        id: chatId,
-        _agent_loop: {
-          turns: turn,
-          log: turnsLog,
-          done: true,
-          reason: "model_returned_content"
-        }
-      });
-      return;
-    }
-    const daemonCalls = [];
-    const clientCalls = [];
-    for (const tc of toolCalls) {
-      const name10 = tc?.function?.name ?? tc?.name ?? "";
-      if (advertisedDaemonNames.has(name10)) daemonCalls.push(tc);
-      else clientCalls.push(tc);
-    }
-    if (clientCalls.length > 0) {
-      turnsLog.push({
-        turn,
-        tool_calls: toolCalls.length,
-        daemon_executed: 0,
-        client_yielded: clientCalls.length
-      });
-      jsonResponse(res, 200, {
-        ...parsed,
-        id: chatId,
-        _agent_loop: {
-          turns: turn,
-          log: turnsLog,
-          done: false,
-          reason: "client_tools_pending",
-          pending_tool_calls: clientCalls.length
-        }
-      });
-      return;
-    }
-    messages2.push(message2);
-    let executed = 0;
-    for (const tc of daemonCalls) {
-      const name10 = tc?.function?.name ?? tc?.name ?? "";
-      const argsRaw = tc?.function?.arguments ?? tc?.arguments ?? "{}";
-      let args = {};
-      try {
-        args = typeof argsRaw === "string" ? JSON.parse(argsRaw) : argsRaw || {};
-      } catch {
-        args = {};
+      const remainingMs = totalDeadline - Date.now();
+      if (remainingMs <= 0) {
+        jsonResponse(res, 504, {
+          error: "agent_loop timed out",
+          turn,
+          elapsed_ms: Date.now() - startMs,
+          timeout_s: perReqTimeoutS ?? null
+        });
+        return;
       }
-      const denial = canInvoke({ toolName: name10, origin, scope });
-      if (denial) {
-        messages2.push({
-          role: "tool",
-          tool_call_id: tc.id,
-          name: name10,
-          content: JSON.stringify({
-            error: denial.reason,
-            status: denial.status
-          })
+      const turnBody = {
+        ...requestBody,
+        model: originalModel,
+        messages: forceNoThinkForTools ? messages2.map((message3) => typeof message3.content === "string" ? { ...message3, content: stripNoThinkPromptDirectives(message3.content) } : message3) : messages2,
+        tools: mergedTools.length > 0 ? mergedTools : void 0,
+        stream: false,
+        ...forceNoThinkForTools ? { think: false } : {},
+        // Strip our own loop-control fields so they don't confuse the backend
+        agent_loop: void 0,
+        include_daemon_tools: void 0,
+        max_turns: void 0,
+        timeout_s: void 0,
+        realtime: void 0,
+        realtime_options: void 0,
+        realtime_max_history_messages: void 0,
+        realtime_max_tokens: void 0
+      };
+      for (const k of Object.keys(turnBody))
+        if (turnBody[k] === void 0) delete turnBody[k];
+      let parsed;
+      try {
+        const path16 = targetType === "vllm" || targetType === "openai" ? "/v1/chat/completions" : "/v1/chat/completions";
+        const result = await ollamaRequest(
+          targetUrl,
+          path16,
+          "POST",
+          JSON.stringify(turnBody),
+          Math.min(reqTimeoutMs, remainingMs),
+          route?.endpoint
+        );
+        if (isClientDisconnected()) return;
+        if (result.status !== 200) {
+          jsonResponse(res, result.status, {
+            error: "Backend request failed during agent_loop",
+            turn,
+            details: result.body.slice(0, 500)
+          });
+          return;
+        }
+        parsed = JSON.parse(result.body);
+      } catch (err) {
+        if (isClientDisconnected()) return;
+        jsonResponse(res, 502, {
+          error: "Backend proxy error during agent_loop",
+          message: err instanceof Error ? err.message : String(err),
+          turn
         });
-        continue;
+        return;
       }
-      if (activeProfileMeta && !isToolAllowed(activeProfileMeta.profile, name10)) {
-        messages2.push({
-          role: "tool",
-          tool_call_id: tc.id,
-          name: name10,
-          content: JSON.stringify({
-            success: false,
-            error: `Tool '${name10}' is denied by profile '${activeProfileMeta.profile.name}'.`,
-            active_profile: {
-              name: activeProfileMeta.profile.name,
-              source: activeProfileMeta.source,
-              path: activeProfileMeta.path,
-              allowed_tools: allowedToolsForProfile(
-                activeProfileMeta.profile,
-                daemonTools.keys()
-              )
-            }
-          })
+      const choice = parsed?.choices?.[0];
+      const rawMessage = choice?.message || {};
+      const message2 = {
+        ...rawMessage,
+        role: typeof rawMessage.role === "string" ? rawMessage.role : "assistant",
+        // Tool-only OpenAI/Ollama turns can return null content. Preserve the
+        // tool_calls but normalize content for the next protocol turn.
+        content: typeof rawMessage.content === "string" ? rawMessage.content : ""
+      };
+      const toolCalls = Array.isArray(message2.tool_calls) ? message2.tool_calls : [];
+      if (toolCalls.length === 0) {
+        jsonResponse(res, 200, {
+          ...parsed,
+          id: chatId,
+          _agent_loop: {
+            turns: turn,
+            log: turnsLog,
+            done: true,
+            reason: "model_returned_content"
+          }
         });
-        continue;
+        return;
       }
-      const meta = daemonTools.get(name10);
-      if (!meta) {
-        messages2.push({
-          role: "tool",
-          tool_call_id: tc.id,
-          name: name10,
-          content: JSON.stringify({ error: `Daemon tool '${name10}' not found` })
+      const daemonCalls = [];
+      const clientCalls = [];
+      for (const tc of toolCalls) {
+        const name10 = tc?.function?.name ?? tc?.name ?? "";
+        if (advertisedDaemonNames.has(name10)) daemonCalls.push(tc);
+        else clientCalls.push(tc);
+      }
+      if (clientCalls.length > 0) {
+        turnsLog.push({
+          turn,
+          tool_calls: toolCalls.length,
+          daemon_executed: 0,
+          client_yielded: clientCalls.length
         });
-        continue;
+        jsonResponse(res, 200, {
+          ...parsed,
+          id: chatId,
+          _agent_loop: {
+            turns: turn,
+            log: turnsLog,
+            done: false,
+            reason: "client_tools_pending",
+            pending_tool_calls: clientCalls.length
+          }
+        });
+        return;
       }
-      let tool = null;
-      try {
-        tool = new meta.ToolClass(process.cwd());
-      } catch {
+      messages2.push(message2);
+      let executed = 0;
+      for (const tc of daemonCalls) {
+        const name10 = tc?.function?.name ?? tc?.name ?? "";
+        const argsRaw = tc?.function?.arguments ?? tc?.arguments ?? "{}";
+        let args = {};
         try {
-          tool = new meta.ToolClass();
+          args = typeof argsRaw === "string" ? JSON.parse(argsRaw) : argsRaw || {};
         } catch {
-          tool = null;
+          args = {};
+        }
+        const denial = canInvoke({ toolName: name10, origin, scope });
+        if (denial) {
+          messages2.push({
+            role: "tool",
+            tool_call_id: tc.id,
+            name: name10,
+            content: JSON.stringify({
+              error: denial.reason,
+              status: denial.status
+            })
+          });
+          continue;
+        }
+        if (activeProfileMeta && !isToolAllowed(activeProfileMeta.profile, name10)) {
+          messages2.push({
+            role: "tool",
+            tool_call_id: tc.id,
+            name: name10,
+            content: JSON.stringify({
+              success: false,
+              error: `Tool '${name10}' is denied by profile '${activeProfileMeta.profile.name}'.`,
+              active_profile: {
+                name: activeProfileMeta.profile.name,
+                source: activeProfileMeta.source,
+                path: activeProfileMeta.path,
+                allowed_tools: allowedToolsForProfile(
+                  activeProfileMeta.profile,
+                  daemonTools.keys()
+                )
+              }
+            })
+          });
+          continue;
+        }
+        const meta = daemonTools.get(name10);
+        if (!meta) {
+          messages2.push({
+            role: "tool",
+            tool_call_id: tc.id,
+            name: name10,
+            content: JSON.stringify({ error: `Daemon tool '${name10}' not found` })
+          });
+          continue;
+        }
+        let tool = null;
+        try {
+          tool = new meta.ToolClass(process.cwd());
+        } catch {
+          try {
+            tool = new meta.ToolClass();
+          } catch {
+            tool = null;
+          }
+        }
+        if (!tool) {
+          messages2.push({
+            role: "tool",
+            tool_call_id: tc.id,
+            name: name10,
+            content: JSON.stringify({
+              error: `Could not instantiate ${meta.classKey}`
+            })
+          });
+          continue;
+        }
+        if (isClientDisconnected()) return;
+        const remainingToolMs = totalDeadline - Date.now();
+        if (remainingToolMs <= 0) {
+          jsonResponse(res, 504, {
+            error: "agent_loop timed out",
+            turn,
+            elapsed_ms: Date.now() - startMs,
+            timeout_s: perReqTimeoutS ?? null
+          });
+          return;
+        }
+        let toolResult;
+        try {
+          activeTool = tool;
+          const execution = await executeAgentLoopToolBounded(
+            tool,
+            args,
+            Math.min(getAgentLoopToolTimeoutMs(), remainingToolMs),
+            clientDisconnect
+          );
+          if (execution.clientDisconnected || isClientDisconnected()) return;
+          toolResult = execution.result;
+        } catch (e2) {
+          if (isClientDisconnected()) return;
+          toolResult = {
+            success: false,
+            output: "",
+            error: e2 instanceof Error ? e2.message : String(e2),
+            durationMs: 0
+          };
+        } finally {
+          activeTool = null;
         }
-      }
-      if (!tool) {
         messages2.push({
           role: "tool",
           tool_call_id: tc.id,
           name: name10,
-          content: JSON.stringify({
-            error: `Could not instantiate ${meta.classKey}`
-          })
+          content: JSON.stringify(toolResult)
         });
-        continue;
+        executed++;
       }
-      let toolResult;
-      try {
-        toolResult = await tool.execute(args);
-      } catch (e2) {
-        toolResult = {
-          success: false,
-          output: "",
-          error: e2 instanceof Error ? e2.message : String(e2),
-          durationMs: 0
-        };
-      }
-      messages2.push({
-        role: "tool",
-        tool_call_id: tc.id,
-        name: name10,
-        content: JSON.stringify(toolResult)
+      turnsLog.push({
+        turn,
+        tool_calls: toolCalls.length,
+        daemon_executed: executed,
+        client_yielded: 0
       });
-      executed++;
     }
-    turnsLog.push({
-      turn,
-      tool_calls: toolCalls.length,
-      daemon_executed: executed,
-      client_yielded: 0
+    jsonResponse(res, 200, {
+      id: chatId,
+      object: "chat.completion",
+      created: Math.floor(Date.now() / 1e3),
+      model,
+      choices: [
+        {
+          index: 0,
+          message: {
+            role: "assistant",
+            content: "[agent_loop reached max_turns without producing a final answer]"
+          },
+          finish_reason: "length"
+        }
+      ],
+      _agent_loop: {
+        turns: maxTurns,
+        log: turnsLog,
+        done: false,
+        reason: "max_turns_exhausted",
+        max_turns: maxTurns
+      }
     });
+  } finally {
+    req3.removeListener("aborted", onRequestAborted);
+    req3.removeListener("close", onRequestClose);
+    req3.removeListener("error", onRequestError);
+    res.removeListener("close", onResponseClose);
+    res.removeListener("error", onResponseError);
   }
-  jsonResponse(res, 200, {
-    id: chatId,
-    object: "chat.completion",
-    created: Math.floor(Date.now() / 1e3),
-    model,
-    choices: [
-      {
-        index: 0,
-        message: {
-          role: "assistant",
-          content: "[agent_loop reached max_turns without producing a final answer]"
-        },
-        finish_reason: "length"
-      }
-    ],
-    _agent_loop: {
-      turns: maxTurns,
-      log: turnsLog,
-      done: false,
-      reason: "max_turns_exhausted",
-      max_turns: maxTurns
-    }
-  });
 }
 async function handleV1ChatCompletions(req3, res, ollamaUrl) {
   const body = await parseJsonBody(req3);
@@ -841203,7 +841498,7 @@ function setTimerEnabled(name10, enabled2) {
     return false;
   }
 }
-var require4, NEXUS_DIRECTORY_ORIGIN3, NEXUS_SPONSORS_URL3, SERVER_BOOT_IDENTITY, voiceTtsTransactionTail, endpointRegistry, modelRouteMap, endpointUsage, _lastEndpointDiagnostics, BACKEND_TIMEOUT_DEFAULT_MS, BACKEND_TIMEOUT_MAX_MS, MODEL_LIST_TIMEOUT_DEFAULT_MS, MODEL_CHECK_TIMEOUT_DEFAULT_MS, MODEL_CHECK_TIMEOUT_MAX_MS, MODEL_CHECK_MAX_TOKENS, metrics, startedAt, FILE_MIME_BY_EXT, realtimeOllamaFallbackCache, runningProcesses, LOCAL_UI_SESSION_COOKIE, LOCAL_UI_SESSION_TOKEN, perKeyUsage, CRON_MARKER2;
+var require4, NEXUS_DIRECTORY_ORIGIN3, NEXUS_SPONSORS_URL3, SERVER_BOOT_IDENTITY, voiceTtsTransactionTail, endpointRegistry, modelRouteMap, endpointUsage, _lastEndpointDiagnostics, BACKEND_TIMEOUT_DEFAULT_MS, BACKEND_TIMEOUT_MAX_MS, AGENT_LOOP_TOTAL_MAX_MS, AGENT_LOOP_TOOL_TIMEOUT_DEFAULT_MS, AGENT_LOOP_TOOL_TIMEOUT_MAX_MS, MODEL_LIST_TIMEOUT_DEFAULT_MS, MODEL_CHECK_TIMEOUT_DEFAULT_MS, MODEL_CHECK_TIMEOUT_MAX_MS, MODEL_CHECK_MAX_TOKENS, cancelledAgentLoopTools, metrics, startedAt, FILE_MIME_BY_EXT, realtimeOllamaFallbackCache, runningProcesses, LOCAL_UI_SESSION_COOKIE, LOCAL_UI_SESSION_TOKEN, perKeyUsage, CRON_MARKER2;
 var init_serve = __esm({
   "packages/cli/src/api/serve.ts"() {
     init_config();
@@ -841259,10 +841554,14 @@ var init_serve = __esm({
     _lastEndpointDiagnostics = [];
     BACKEND_TIMEOUT_DEFAULT_MS = 12e4;
     BACKEND_TIMEOUT_MAX_MS = 36e5;
+    AGENT_LOOP_TOTAL_MAX_MS = 30 * 60 * 1e3;
+    AGENT_LOOP_TOOL_TIMEOUT_DEFAULT_MS = 3e4;
+    AGENT_LOOP_TOOL_TIMEOUT_MAX_MS = 12e4;
     MODEL_LIST_TIMEOUT_DEFAULT_MS = 1500;
     MODEL_CHECK_TIMEOUT_DEFAULT_MS = 1e4;
     MODEL_CHECK_TIMEOUT_MAX_MS = 3e4;
     MODEL_CHECK_MAX_TOKENS = 32;
+    cancelledAgentLoopTools = /* @__PURE__ */ new WeakSet();
     metrics = {
       requests: /* @__PURE__ */ new Map(),
       totalTokensIn: 0,