omnius 1.0.640 → 1.0.642

package/dist/index.js

@@ -5859,7 +5859,7 @@ 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:${MODEL_SHA2562}` || manifest.backend !== "onnxruntime-cpu" || manifest.onnxRuntimeVersion !== ONNXRUNTIME_VERSION || manifest.numpyVersion !== NUMPY_VERSION || manifest.preprocessing?.implementation !== FBANK_IMPLEMENTATION || manifest.preprocessing?.validationSignature !== FBANK_VALIDATION_SIGNATURE || !hasExpectedSpeakerRuntimePackages(manifest.packages) || usesSystemSitePackages() || !existsSync11(modelPath2()) || !existsSync11(managedPython())) {
+    if (manifest.runtimeVersion !== RUNTIME_VERSION3 || manifest.model !== MODEL_NAME2 || manifest.modelRevision !== MODEL_REVISION2 || manifest.modelDigest !== `sha256:${MODEL_SHA2562}` || manifest.backend !== "onnxruntime-cpu" || manifest.onnxRuntimeVersion !== ONNXRUNTIME_VERSION || manifest.numpyVersion !== NUMPY_VERSION || manifest.preprocessing?.implementation !== FBANK_IMPLEMENTATION || manifest.preprocessing?.validation !== FBANK_VALIDATION || !hasExpectedSpeakerRuntimePackages(manifest.packages) || usesSystemSitePackages() || !existsSync11(modelPath2()) || !existsSync11(managedPython())) {
       return null;
     }
     return sha256File4(modelPath2()) === MODEL_SHA2562 ? manifest : null;
@@ -5914,8 +5914,8 @@ async function preprocessingProbe(python2, workerScript) {
   try {
     const output2 = await execFileText2(python2, [workerScript, "--preprocessing-probe"], { timeout: 3e4, env: managedEnv3() });
     const event = JSON.parse(output2.trim().split(/\r?\n/).pop() || "{}");
-    if (event.type !== "preprocessing_probe" || event.implementation !== FBANK_IMPLEMENTATION || event.validation_signature !== FBANK_VALIDATION_SIGNATURE) {
-      return { verified: false, error: "managed NumPy Kaldi fbank calibration did not match the pinned signature" };
+    if (event.type !== "preprocessing_probe" || event.implementation !== FBANK_IMPLEMENTATION || event.validation?.validation !== FBANK_VALIDATION) {
+      return { verified: false, error: "managed NumPy Kaldi fbank did not satisfy the pinned versioned calibration invariants" };
     }
     return { verified: true };
   } catch (error) {
@@ -6089,7 +6089,7 @@ async function getSpeakerEmbeddingReadiness() {
       packages: manifest?.packages,
       preprocessing: {
         implementation: FBANK_IMPLEMENTATION,
-        validation_signature: FBANK_VALIDATION_SIGNATURE,
+        validation: FBANK_VALIDATION,
         verified: runtime2.preprocessingVerified
       },
       providers: runtime2.providers,
@@ -6120,7 +6120,7 @@ async function ensureSpeakerEmbeddingSetup() {
     await installPinnedSpeakerRuntimeDependencies(python2);
     const runtime2 = await probeRuntime(python2, findWorkerScript4());
     if (!runtime2.available || runtime2.onnxruntimeVersion !== ONNXRUNTIME_VERSION || runtime2.numpyVersion !== NUMPY_VERSION || !runtime2.preprocessingVerified || !runtime2.providers.includes("CPUExecutionProvider") || usesSystemSitePackages()) {
-      throw new Error(`Managed CPU NumPy Kaldi-fbank validation failed after speaker setup${runtime2.error ? `: ${runtime2.error}` : ""}. The private speaker venv must prove pinned NumPy ${NUMPY_VERSION}, ONNX Runtime ${ONNXRUNTIME_VERSION}, the bundled preprocessing signature, and CPUExecutionProvider without importing Torch or Torchaudio.`);
+      throw new Error(`Managed CPU NumPy Kaldi-fbank validation failed after speaker setup${runtime2.error ? `: ${runtime2.error}` : ""}. The private speaker venv must prove pinned NumPy ${NUMPY_VERSION}, ONNX Runtime ${ONNXRUNTIME_VERSION}, the bundled versioned preprocessing invariants, and CPUExecutionProvider without importing Torch or Torchaudio.`);
     }
     await downloadChecked2(MODEL_URL2, modelPath2(), MODEL_SHA2562);
     const manifest = {
@@ -6134,7 +6134,7 @@ async function ensureSpeakerEmbeddingSetup() {
       packages: speakerRuntimeManifest(),
       preprocessing: {
         implementation: FBANK_IMPLEMENTATION,
-        validationSignature: FBANK_VALIDATION_SIGNATURE
+        validation: FBANK_VALIDATION
       },
       pythonVersion: runtime2.pythonVersion ?? bootstrapRuntime.pythonVersion,
       installedAt: (/* @__PURE__ */ new Date()).toISOString()
@@ -6234,7 +6234,7 @@ async function activateSpeakerEmbedding() {
         }
       });
     });
-    if (ready.pid !== worker2.pid || ready.backend !== "onnxruntime-cpu" || ready.provider !== "CPUExecutionProvider" || ready.model !== MODEL_NAME2 || ready.model_digest !== manifest.modelDigest || ready.dimension !== 512 || !ready.warmed || ready.preprocessing !== FBANK_IMPLEMENTATION || ready.preprocessing_signature !== FBANK_VALIDATION_SIGNATURE) {
+    if (ready.pid !== worker2.pid || ready.backend !== "onnxruntime-cpu" || ready.provider !== "CPUExecutionProvider" || ready.model !== MODEL_NAME2 || ready.model_digest !== manifest.modelDigest || ready.dimension !== 512 || !ready.warmed || ready.preprocessing !== FBANK_IMPLEMENTATION || ready.preprocessing_validation?.validation !== FBANK_VALIDATION) {
       throw new Error("Speaker embedding worker did not report the pinned CPU CAM++ runtime");
     }
     state3.ready = ready;
@@ -6372,7 +6372,7 @@ 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, NUMPY_VERSION, NUMPY_WHEEL, NUMPY_WHEEL_SHA256, FBANK_IMPLEMENTATION, FBANK_VALIDATION_SIGNATURE, SETUP_TIMEOUT_MS3, START_TIMEOUT_MS3, EMBED_TIMEOUT_MS2, WESPEAKER_CAMPP_ONNX_SHA256, WESPEAKER_CAMPP_MODEL_REVISION, WESPEAKER_CAMPP_ONNXRUNTIME_VERSION, WESPEAKER_CAMPP_NUMPY_VERSION, WESPEAKER_CAMPP_FBANK_IMPLEMENTATION, WESPEAKER_CAMPP_EMBEDDING_SET_SLUG, SPEAKER_RUNTIME_WHEELS, state3;
+var RUNTIME_VERSION3, MODEL_NAME2, MODEL_REVISION2, MODEL_FILE, MODEL_URL2, MODEL_SHA2562, ONNXRUNTIME_VERSION, ONNXRUNTIME_WHEEL, ONNXRUNTIME_WHEEL_SHA256, NUMPY_VERSION, NUMPY_WHEEL, NUMPY_WHEEL_SHA256, FBANK_IMPLEMENTATION, FBANK_VALIDATION, SETUP_TIMEOUT_MS3, START_TIMEOUT_MS3, EMBED_TIMEOUT_MS2, WESPEAKER_CAMPP_ONNX_SHA256, WESPEAKER_CAMPP_MODEL_REVISION, WESPEAKER_CAMPP_ONNXRUNTIME_VERSION, WESPEAKER_CAMPP_NUMPY_VERSION, WESPEAKER_CAMPP_FBANK_IMPLEMENTATION, WESPEAKER_CAMPP_FBANK_VALIDATION, WESPEAKER_CAMPP_EMBEDDING_SET_SLUG, SPEAKER_RUNTIME_WHEELS, state3;
 var init_speaker_embedding_runtime = __esm({
   "packages/execution/dist/speaker-embedding-runtime.js"() {
     "use strict";
@@ -6380,7 +6380,7 @@ var init_speaker_embedding_runtime = __esm({
     init_process_async();
     init_model_store();
     init_venv_paths();
-    RUNTIME_VERSION3 = 4;
+    RUNTIME_VERSION3 = 5;
     MODEL_NAME2 = "wespeaker-voxceleb-campplus";
     MODEL_REVISION2 = "acf623ad8ca746e50baa432255cf8fc57c669c45";
     MODEL_FILE = "voxceleb_CAM++.onnx";
@@ -6393,7 +6393,7 @@ var init_speaker_embedding_runtime = __esm({
     NUMPY_WHEEL = "https://files.pythonhosted.org/packages/fc/a5/4beee6488160798683eed5bdb7eead455892c3b4e1f78d79d8d3f3b084ac/numpy-1.26.4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl";
     NUMPY_WHEEL_SHA256 = "d209d8969599b27ad20994c8e41936ee0964e6da07478d6c35016bc386b66ad4";
     FBANK_IMPLEMENTATION = "kaldi-fbank-numpy-v1";
-    FBANK_VALIDATION_SIGNATURE = "sha256:19bb5552b8068d5cd2ccdbb07ac7a7ed1137400d3bdc3c54fb9c21d4f06e16fa";
+    FBANK_VALIDATION = "kaldi-fbank-invariants-v2";
     SETUP_TIMEOUT_MS3 = 10 * 6e4;
     START_TIMEOUT_MS3 = 6e4;
     EMBED_TIMEOUT_MS2 = 45e3;
@@ -6402,6 +6402,7 @@ var init_speaker_embedding_runtime = __esm({
     WESPEAKER_CAMPP_ONNXRUNTIME_VERSION = ONNXRUNTIME_VERSION;
     WESPEAKER_CAMPP_NUMPY_VERSION = NUMPY_VERSION;
     WESPEAKER_CAMPP_FBANK_IMPLEMENTATION = FBANK_IMPLEMENTATION;
+    WESPEAKER_CAMPP_FBANK_VALIDATION = FBANK_VALIDATION;
     WESPEAKER_CAMPP_EMBEDDING_SET_SLUG = "speaker-identity-wespeaker-voxceleb-campplus-acf623ad8ca746e50baa432255cf8fc57c669c45-fbank80-cmn-fullclip-l2";
     SPEAKER_RUNTIME_WHEELS = [
       {
@@ -345916,6 +345917,55 @@ function ocrPythonEnv(extra = {}) {
     ...extra
   };
 }
+function assessAdvancedOcrEvidence(input) {
+  const text3 = String(input.text ?? "").trim();
+  const chars = text3.length;
+  const confidence2 = Number.isFinite(input.confidence) ? input.confidence : 0;
+  const lines = Number.isFinite(input.lines) ? input.lines : 0;
+  if (chars === 0) {
+    return { state: "low_information", accepted: false, reason: "no_readable_text", chars, confidence: confidence2, lines };
+  }
+  const alnum = [...text3].filter((character) => /[\p{L}\p{N}]/u.test(character)).length;
+  const alnumRatio2 = alnum / Math.max(1, chars);
+  if (chars < MIN_SUBSTANTIVE_OCR_CHARS && confidence2 < MIN_ACCEPTED_OCR_CONFIDENCE) {
+    return {
+      state: "low_information",
+      accepted: false,
+      reason: "insufficient_low_confidence_text",
+      chars,
+      confidence: confidence2,
+      lines,
+      alnum_ratio: Number(alnumRatio2.toFixed(3))
+    };
+  }
+  if (chars >= HIGH_VOLUME_GARBAGE_CHARS && confidence2 < HIGH_VOLUME_GARBAGE_CONFIDENCE || confidence2 < MIN_ACCEPTED_OCR_CONFIDENCE || alnumRatio2 < 0.45) {
+    return {
+      state: "rejected",
+      accepted: false,
+      reason: chars >= HIGH_VOLUME_GARBAGE_CHARS && confidence2 < HIGH_VOLUME_GARBAGE_CONFIDENCE ? "high_volume_low_confidence_text" : "low_confidence_or_symbol_heavy_text",
+      chars,
+      confidence: confidence2,
+      lines,
+      alnum_ratio: Number(alnumRatio2.toFixed(3))
+    };
+  }
+  return {
+    state: "accepted",
+    accepted: true,
+    reason: "confidence_and_text_quality_met",
+    chars,
+    confidence: confidence2,
+    lines,
+    alnum_ratio: Number(alnumRatio2.toFixed(3))
+  };
+}
+function isTerminalSmallCropOcrRejection(evidence) {
+  if (evidence.state !== "rejected")
+    return false;
+  if (evidence.reason === "high_volume_low_confidence_text")
+    return true;
+  return evidence.chars >= TERMINAL_SYMBOL_GARBAGE_CHARS && (evidence.alnum_ratio ?? 1) < TERMINAL_SYMBOL_GARBAGE_ALNUM_RATIO;
+}
 function ocrDiagnostic(code8, message2, extra = {}) {
   return {
     schema: "omnius.ocr-diagnostic.v1",
@@ -345934,7 +345984,7 @@ function parsePipelineDiagnostic(error) {
     return null;
   }
 }
-var OCR_PIPELINE_TIMEOUT_MS, OCR_PROCESS_DRAIN_MS, OcrImageAdvancedTool;
+var OCR_PIPELINE_TIMEOUT_MS, OCR_PROCESS_DRAIN_MS, MIN_ACCEPTED_OCR_CONFIDENCE, MIN_SUBSTANTIVE_OCR_CHARS, HIGH_VOLUME_GARBAGE_CHARS, HIGH_VOLUME_GARBAGE_CONFIDENCE, TERMINAL_SYMBOL_GARBAGE_CHARS, TERMINAL_SYMBOL_GARBAGE_ALNUM_RATIO, OcrImageAdvancedTool;
 var init_ocr_image_advanced = __esm({
   "packages/execution/dist/tools/ocr-image-advanced.js"() {
     "use strict";
@@ -345942,6 +345992,12 @@ var init_ocr_image_advanced = __esm({
     init_ocr_advanced_runtime();
     OCR_PIPELINE_TIMEOUT_MS = 8e4;
     OCR_PROCESS_DRAIN_MS = 3e3;
+    MIN_ACCEPTED_OCR_CONFIDENCE = 50;
+    MIN_SUBSTANTIVE_OCR_CHARS = 12;
+    HIGH_VOLUME_GARBAGE_CHARS = 64;
+    HIGH_VOLUME_GARBAGE_CONFIDENCE = 35;
+    TERMINAL_SYMBOL_GARBAGE_CHARS = 24;
+    TERMINAL_SYMBOL_GARBAGE_ALNUM_RATIO = 0.2;
     OcrImageAdvancedTool = class {
       workingDir;
       name = "ocr_image_advanced";
@@ -346136,7 +346192,10 @@ var init_ocr_image_advanced = __esm({
                 stage: result.diagnostic?.stage ?? "pipeline",
                 deadline_ms: result.diagnostic?.deadline_ms ?? OCR_PIPELINE_TIMEOUT_MS,
                 attempts_completed: result.diagnostic?.attempts_completed ?? null,
-                attempts_planned: result.diagnostic?.attempts_planned ?? null
+                attempts_planned: result.diagnostic?.attempts_planned ?? null,
+                evidence: result.diagnostic?.evidence ?? null,
+                terminal_small_crop_rejection: result.diagnostic?.terminal_small_crop_rejection ?? null,
+                attempt_cap_seconds: result.diagnostic?.attempt_cap_seconds ?? null
               })
             };
           }
@@ -346160,22 +346219,62 @@ var init_ocr_image_advanced = __esm({
               data: result
             };
           }
+          const boundaryEvidence = assessAdvancedOcrEvidence({
+            text: result.text,
+            confidence: result.confidence,
+            lines: result.lines
+          });
+          if (boundaryEvidence.state === "rejected") {
+            return {
+              success: false,
+              output: "",
+              error: "OCR evidence rejected: low-confidence text was suppressed and is not presented as extracted evidence.",
+              durationMs: performance.now() - start2,
+              data: ocrDiagnostic("ocr_evidence_rejected", "OCR output did not meet evidence-quality requirements.", {
+                evidence: boundaryEvidence,
+                terminal_small_crop_rejection: isTerminalSmallCropOcrRejection(boundaryEvidence)
+              })
+            };
+          }
+          const report2 = {
+            ...result,
+            ...boundaryEvidence.state === "low_information" ? { text: "", chars: 0, lines: 0, score: 0 } : {},
+            quality: {
+              ...result.quality ?? {},
+              ...boundaryEvidence,
+              schema: "omnius.ocr-evidence.v1"
+            },
+            ...boundaryEvidence.state === "low_information" && !result.diagnostic ? {
+              diagnostic: {
+                schema: "omnius.ocr-diagnostic.v1",
+                code: "ocr_low_information",
+                message: "OCR produced no accepted readable text; the result is low-information rather than evidence.",
+                stage: "evidence_quality"
+              }
+            } : {}
+          };
           const parts = [];
-          parts.push(`OCR extracted from ${basename11(imagePath)} (${result.image_size})`);
-          parts.push(`Best variant: ${result.variant} (confidence: ${result.confidence}%, ${result.chars} chars, ${result.lines} lines, score: ${result.score})`);
-          parts.push(`Variants tested: ${result.variants_tested}`);
-          if (result.output_files) {
-            const files = result.output_files;
+          parts.push(`OCR extracted from ${basename11(imagePath)} (${report2.image_size})`);
+          if (boundaryEvidence.state === "low_information") {
+            parts.push(`OCR found no accepted readable text (${boundaryEvidence.reason}); low-information result.`);
+          } else {
+            parts.push(`Best variant: ${report2.variant} (confidence: ${report2.confidence}%, ${report2.chars} chars, ${report2.lines} lines, score: ${report2.score})`);
+          }
+          parts.push(`Variants tested: ${report2.variants_tested}`);
+          if (report2.output_files) {
+            const files = report2.output_files;
             const saved = [files.txt, files.csv, files.pdf].filter(Boolean);
             parts.push(`Output files: ${saved.join(", ")}`);
           }
-          parts.push("");
-          parts.push("--- Extracted Text ---");
-          parts.push(result.text);
-          if (result.regions) {
+          if (boundaryEvidence.state === "accepted") {
+            parts.push("");
+            parts.push("--- Extracted Text ---");
+            parts.push(report2.text);
+          }
+          if (report2.regions) {
             parts.push("");
             parts.push("--- Region Extraction ---");
-            for (const [rname, rtext] of Object.entries(result.regions)) {
+            for (const [rname, rtext] of Object.entries(report2.regions)) {
               if (rtext) {
                 parts.push(`
 [${rname.toUpperCase()}]`);
@@ -346183,8 +346282,8 @@ var init_ocr_image_advanced = __esm({
               }
             }
           }
-          if (result.all_variants) {
-            const sorted = Object.entries(result.all_variants).filter(([_, v]) => v.chars > 0).sort((a2, b) => b[1].score - a2[1].score).slice(0, 5);
+          if (report2.all_variants) {
+            const sorted = Object.entries(report2.all_variants).filter(([_, v]) => v.chars > 0).sort((a2, b) => b[1].score - a2[1].score).slice(0, 5);
             if (sorted.length > 1) {
               parts.push("");
               parts.push("--- Top Variants ---");
@@ -346201,7 +346300,7 @@ var init_ocr_image_advanced = __esm({
             success: true,
             output: parts.join("\n"),
             durationMs: performance.now() - start2,
-            data: result
+            data: report2
           };
         } catch (err) {
           const pipelineResult = parsePipelineDiagnostic(err);
@@ -346215,7 +346314,8 @@ var init_ocr_image_advanced = __esm({
                 stage: pipelineResult.diagnostic?.stage ?? "pipeline",
                 deadline_ms: pipelineResult.diagnostic?.deadline_ms ?? OCR_PIPELINE_TIMEOUT_MS,
                 attempts_completed: pipelineResult.diagnostic?.attempts_completed ?? null,
-                attempts_planned: pipelineResult.diagnostic?.attempts_planned ?? null
+                attempts_planned: pipelineResult.diagnostic?.attempts_planned ?? null,
+                evidence: pipelineResult.diagnostic?.evidence ?? null
               })
             };
           }
@@ -616978,6 +617078,7 @@ __export(dist_exports2, {
   VisualTriggerTool: () => VisualTriggerTool,
   WESPEAKER_CAMPP_EMBEDDING_SET_SLUG: () => WESPEAKER_CAMPP_EMBEDDING_SET_SLUG,
   WESPEAKER_CAMPP_FBANK_IMPLEMENTATION: () => WESPEAKER_CAMPP_FBANK_IMPLEMENTATION,
+  WESPEAKER_CAMPP_FBANK_VALIDATION: () => WESPEAKER_CAMPP_FBANK_VALIDATION,
   WESPEAKER_CAMPP_MODEL_REVISION: () => WESPEAKER_CAMPP_MODEL_REVISION,
   WESPEAKER_CAMPP_NUMPY_VERSION: () => WESPEAKER_CAMPP_NUMPY_VERSION,
   WESPEAKER_CAMPP_ONNXRUNTIME_VERSION: () => WESPEAKER_CAMPP_ONNXRUNTIME_VERSION,
@@ -687469,8 +687570,11 @@ import { existsSync as existsSync128, mkdirSync as mkdirSync79, readFileSync as
 import { dirname as dirname51, join as join140 } from "node:path";
 import { homedir as homedir41 } from "node:os";
 import { fileURLToPath as fileURLToPath25 } from "node:url";
+function managedAsrRuntimeRoot() {
+  return process.env["OMNIUS_ASR_RUNTIME_ROOT"]?.trim() || MANAGED_ASR_ROOT;
+}
 function managedAsrStateFile(engineId) {
-  return join140(MANAGED_ASR_ROOT, engineId, "models.json");
+  return join140(managedAsrRuntimeRoot(), engineId, "models.json");
 }
 function managedAsrModelDir(engineId) {
   return join140(MANAGED_ASR_MODEL_ROOT, engineId);
@@ -687483,16 +687587,18 @@ function readManagedAsrState(engineId) {
   }
 }
 function persistManagedAsrReadiness(readiness) {
-  const root = join140(MANAGED_ASR_ROOT, readiness.engineId);
+  const root = join140(managedAsrRuntimeRoot(), readiness.engineId);
   mkdirSync79(root, { recursive: true });
   const state5 = readManagedAsrState(readiness.engineId);
   state5.models ??= {};
   state5.models[readiness.modelId] = {
     installed: readiness.installed,
     weightsReady: readiness.weightsReady,
+    active: readiness.active,
     device: readiness.device,
     lastError: readiness.lastError,
-    pulledAt: readiness.pulledAt
+    pulledAt: readiness.pulledAt,
+    activeAt: readiness.activeAt
   };
   writeFileSync70(managedAsrStateFile(readiness.engineId), `${JSON.stringify(state5, null, 2)}
 `, "utf8");
@@ -687885,17 +687991,34 @@ ${log22.slice(-2e3)}` };
 function getManagedAsrReadiness(engineId, modelId) {
   assertManagedAsrModel(engineId, modelId);
   const saved = readManagedAsrState(engineId).models?.[modelId];
+  const active = saved?.active === true && saved.weightsReady === true && typeof saved.activeAt === "string" && /^cuda(?::\d+)?$/i.test(saved.device ?? "");
   return {
     engineId,
     modelId,
     installed: Boolean(saved?.installed && existsSync128(getVenvPython())),
     weightsReady: Boolean(saved?.weightsReady),
-    active: false,
+    // Managed direct ASR has no resident process. `active` therefore means
+    // this exact model last completed a CUDA inference successfully, not that
+    // a Python child happens to still be alive after a request returned.
+    active,
     device: saved?.device,
     lastError: saved?.lastError,
-    pulledAt: saved?.pulledAt
+    pulledAt: saved?.pulledAt,
+    activeAt: saved?.activeAt
   };
 }
+function recordManagedAsrInference(input, outcome) {
+  const prior = getManagedAsrReadiness(input.engineId, input.modelId);
+  const cuda = typeof input.device === "string" && /^cuda(?::\d+)?$/i.test(input.device);
+  const activeAt = outcome.ok && cuda ? (/* @__PURE__ */ new Date()).toISOString() : void 0;
+  persistManagedAsrReadiness({
+    ...prior,
+    device: input.device || prior.device,
+    active: Boolean(activeAt),
+    activeAt,
+    lastError: outcome.ok ? void 0 : outcome.error
+  });
+}
 function ensureManagedAsrModel(input) {
   const engineId = input.engineId.trim().toLowerCase();
   const modelId = input.modelId.trim().toLowerCase();
@@ -688058,12 +688181,16 @@ function transcribeManagedNemotronFile(input) {
   const ready = events.find((event) => event["type"] === "ready");
   const error = events.find((event) => event["type"] === "error");
   if (result.status !== 0 || !transcript) {
-    throw new Error(String(error?.["message"] ?? (`${result.stderr || ""}`.trim() || result.error || "Nemotron file transcription failed")));
+    const message2 = String(error?.["message"] ?? (`${result.stderr || ""}`.trim() || result.error || "Nemotron file transcription failed"));
+    recordManagedAsrInference({ engineId: "nemotron-streaming", modelId, device: existing.device }, { ok: false, error: message2 });
+    throw new Error(message2);
   }
+  const device2 = typeof ready?.["device"] === "string" ? ready["device"] : existing.device;
+  recordManagedAsrInference({ engineId: "nemotron-streaming", modelId, device: device2 }, { ok: true });
   return {
     text: String(transcript["text"] ?? ""),
     duration: Number.isFinite(Number(transcript["audioSeconds"])) ? Number(transcript["audioSeconds"]) : null,
-    device: typeof ready?.["device"] === "string" ? ready["device"] : existing.device,
+    device: device2,
     segments: [],
     warnings: []
   };
@@ -688102,6 +688229,7 @@ function transcribeManagedWhisperFile(input) {
   const payload = [...parseWorkerEvents(`${result.stdout || ""}`)].reverse().find((event) => event["ok"] === true);
   if (result.status !== 0 || !payload) {
     const detail = `${result.stderr || ""}`.trim() || String(result.error ?? "Whisper file transcription failed");
+    recordManagedAsrInference({ engineId: "openai-whisper", modelId, device: existing.device }, { ok: false, error: detail });
     throw new Error(detail);
   }
   const segments = Array.isArray(payload["segments"]) ? payload["segments"].flatMap((segment) => {
@@ -688113,10 +688241,12 @@ function transcribeManagedWhisperFile(input) {
       text: String(value2["text"] ?? "")
     }];
   }) : [];
+  const device2 = typeof payload["device"] === "string" ? payload["device"] : existing.device;
+  recordManagedAsrInference({ engineId: "openai-whisper", modelId, device: device2 }, { ok: true });
   return {
     text: String(payload["text"] ?? ""),
     duration: Number.isFinite(Number(payload["duration"])) ? Number(payload["duration"]) : null,
-    device: typeof payload["device"] === "string" ? payload["device"] : existing.device,
+    device: device2,
     segments,
     warnings: []
   };
@@ -688152,17 +688282,21 @@ function transcribeManagedAsrFile(input) {
     const payload = [...events].reverse().find((event) => event["type"] === "transcript");
     if (result.status !== 0 || !payload) {
       const workerError = [...events].reverse().find((event) => event["type"] === "error");
-      throw new Error(String(workerError?.["message"] ?? result.stderr ?? result.error ?? "Voxtral file transcription failed"));
+      const message2 = String(workerError?.["message"] ?? result.stderr ?? result.error ?? "Voxtral file transcription failed");
+      recordManagedAsrInference({ engineId, modelId, device: existing.device }, { ok: false, error: message2 });
+      throw new Error(message2);
     }
     const segments = Array.isArray(payload["segments"]) ? payload["segments"].flatMap((segment) => {
       if (!segment || typeof segment !== "object") return [];
       const value2 = segment;
       return [{ start: Number(value2["start"] ?? 0), end: Number(value2["end"] ?? 0), text: String(value2["text"] ?? "") }];
     }) : [];
+    const device2 = typeof payload["device"] === "string" ? payload["device"] : existing.device;
+    recordManagedAsrInference({ engineId, modelId, device: device2 }, { ok: true });
     return {
       text: String(payload["text"] ?? ""),
       duration: Number.isFinite(Number(payload["duration"])) ? Number(payload["duration"]) : null,
-      device: existing.device,
+      device: device2,
       segments,
       warnings: []
     };
@@ -688625,6 +688759,11 @@ function updateListenLiveState(patch) {
 function getListenLiveState() {
   return listenLiveState;
 }
+function childProcessIsRunning(process4) {
+  return Boolean(
+    process4 && process4.exitCode === null && process4.signalCode === null && !process4.killed
+  );
+}
 function clamp0112(value2) {
   if (!Number.isFinite(value2)) return 0;
   return Math.max(0, Math.min(1, value2));
@@ -689286,7 +689425,7 @@ var init_listen = __esm({
       lastVisibleVadAt = 0;
       stderrTail = "";
       get ready() {
-        return this._ready;
+        return this._ready && childProcessIsRunning(this.process);
       }
       async start() {
         let pyPath = "python3";
@@ -689425,17 +689564,32 @@ ${text3}`.slice(-2e3);
           });
           onChildError(this.process, (err) => {
             clearTimeout(timeout2);
-            reject(err);
+            const wasReady = this._ready;
+            this._ready = false;
+            this.process = null;
+            if (wasReady) {
+              updateListenLiveState({ phase: "error", lastStatus: err.message.slice(0, 160) });
+              this.emit("error", err);
+            } else {
+              reject(err);
+            }
           });
           onChildClose(this.process, (code8) => {
             this.unregisterBrokerModel();
-            if (!this._ready) {
+            const wasReady = this._ready;
+            this._ready = false;
+            this.process = null;
+            if (!wasReady) {
               clearTimeout(timeout2);
               reject(
                 new Error(
                   `${this.engineId} worker exited with code ${code8} before ready` + (this.stderrTail ? `: ${this.stderrTail.trim()}` : "")
                 )
               );
+            } else {
+              const message2 = `${this.engineId} worker exited with code ${code8}` + (this.stderrTail ? `: ${this.stderrTail.trim()}` : "");
+              updateListenLiveState({ phase: "error", lastStatus: message2.slice(0, 160) });
+              this.emit("error", new Error(message2));
             }
           });
         });
@@ -689500,6 +689654,10 @@ ${text3}`.slice(-2e3);
       liveTranscriber = null;
       // TranscribeLive from transcribe-cli or WhisperFallbackTranscriber
       active = false;
+      /** Set only after the selected worker has emitted/finished its ready signal. */
+      liveWorkerReady = false;
+      /** Identity captured when that worker became ready; never infer it from a later selection. */
+      runningSelection = null;
       paused = false;
       // Pause state for voicechat (stops mic but keeps transcriber)
       silenceTimer = null;
@@ -689542,6 +689700,27 @@ ${text3}`.slice(-2e3);
       get currentMode() {
         return this.config.mode;
       }
+      /**
+       * Runtime truth for a resident live ASR pipeline.  This does not claim that
+       * a selected-but-unstarted model is ready, and it verifies fallback worker
+       * process liveness rather than trusting an old `ready` event.
+       */
+      get liveRuntimeReadiness() {
+        const fallbackHealthy = !(this.liveTranscriber instanceof WhisperFallbackTranscriber) || this.liveTranscriber.ready === true;
+        const workerReady = Boolean(
+          this.active && this.liveWorkerReady && this.runningSelection && this.liveTranscriber && fallbackHealthy
+        );
+        return {
+          // Keep the captured identity while reporting a failed worker so the
+          // daemon can surface its error against the exact selection rather than
+          // silently falling back to persisted weights state.
+          engineId: this.runningSelection?.engineId ?? null,
+          modelId: this.runningSelection?.modelId ?? null,
+          workerReady,
+          micActive: workerReady && !this.paused && childProcessIsRunning(this.micProcess),
+          paused: this.paused
+        };
+      }
       get pendingTranscript() {
         return this.pendingText;
       }
@@ -689812,6 +689991,8 @@ ${text3}`.slice(-2e3);
        */
       async start() {
         if (this.active) return "Already listening.";
+        this.liveWorkerReady = false;
+        this.runningSelection = null;
         const unavailableReason = getAsrEngineUnavailableReason(this.config.engineId);
         if (unavailableReason) return unavailableReason;
         if (this.config.engineId === "vibevoice-transformers") {
@@ -689941,6 +690122,8 @@ ${text3}`.slice(-2e3);
                   }
                 );
                 this.liveTranscriber.on("error", (err) => {
+                  this.liveWorkerReady = false;
+                  updateListenLiveState({ phase: "error", lastStatus: err.message.slice(0, 160) });
                   this.emit("error", err);
                 });
                 await new Promise((resolve90, reject) => {
@@ -689999,6 +690182,8 @@ ${text3}`.slice(-2e3);
               if (this.config.mode === "auto") this.resetSilenceTimer();
             });
             fallback.on("error", (err) => {
+              this.liveWorkerReady = false;
+              updateListenLiveState({ phase: "error", lastStatus: err.message.slice(0, 160) });
               this.emit("error", err);
             });
             this.liveTranscriber = fallback;
@@ -690011,6 +690196,11 @@ transcribe-cli error: ${transcribeCliError}` : "";
           }
         }
         this.active = true;
+        this.liveWorkerReady = true;
+        this.runningSelection = {
+          engineId: this.config.engineId,
+          modelId: this.config.modelId
+        };
         this.paused = false;
         this.spawnMicProcess(micCmd);
         this.blinkState = true;
@@ -690037,6 +690227,8 @@ transcribe-cli error: ${transcribeCliError}` : "";
       async stop() {
         if (!this.active) return "Not listening.";
         this.active = false;
+        this.liveWorkerReady = false;
+        this.runningSelection = null;
         this.owners.clear();
         this.blinkState = false;
         this.micWaterfall = [];
@@ -792931,6 +793123,13 @@ function getRuntimeStatus() {
   const persisted = _listenEngine ? null : loadGlobalSettings();
   const asrEngineId = _listenEngine?.currentEngine ?? persisted?.asrEngine ?? "openai-whisper";
   const asrModelId = _listenEngine?.currentModel ?? persisted?.asrModel ?? "medium";
+  const liveAsr = _listenEngine?.liveRuntimeReadiness ?? null;
+  const selectedLivePipeline = Boolean(
+    liveAsr?.engineId === asrEngineId && liveAsr?.modelId === asrModelId
+  );
+  const selectedLiveWorker = selectedLivePipeline && liveAsr?.workerReady === true;
+  const selectedLiveMic = selectedLiveWorker && liveAsr?.micActive === true;
+  const managedAsr = asrEngineId === "openai-whisper" || asrEngineId === "nemotron-streaming" || asrEngineId === "voxtral-transformers" ? getManagedAsrReadiness(asrEngineId, asrModelId) : null;
   const vibePhase = vibe.active ? "ready" : vibe.lastError ? "error" : vibe.weightsReady ? "weights-ready" : vibe.installed ? "runtime-installed" : "not-installed";
   return {
     state: _state3,
@@ -792942,9 +793141,9 @@ function getRuntimeStatus() {
     listenPaused: _listenEngine?.isPaused ?? false,
     asrEngineId,
     asrModelId,
-    asrBackend: asrEngineId === "vibevoice-transformers" ? "vibevoice-transformers" : listenState.backend,
-    asrPhase: asrEngineId === "vibevoice-transformers" ? vibePhase : listenState.phase,
-    asrReady: asrEngineId === "vibevoice-transformers" ? vibe.active : Boolean(_listenEngine?.isActive),
+    asrBackend: asrEngineId === "vibevoice-transformers" ? "vibevoice-transformers" : managedAsr ? asrEngineId : listenState.backend,
+    asrPhase: asrEngineId === "vibevoice-transformers" ? vibePhase : selectedLiveMic ? "listening" : selectedLiveWorker && liveAsr?.paused && listenState.phase === "paused" ? "paused" : selectedLivePipeline && listenState.phase === "error" ? "error" : managedAsr?.active ? "ready" : managedAsr?.lastError ? "error" : managedAsr?.weightsReady ? "weights-ready" : "not-started",
+    asrReady: asrEngineId === "vibevoice-transformers" ? vibe.active : managedAsr ? managedAsr.active || selectedLiveMic : Boolean(_listenEngine?.isActive),
     micDevice: listenState.micDevice || null,
     micSourceKind: listenState.micSourceKind,
     micSourceChannels: listenState.micSourceChannels,
@@ -793173,6 +793372,7 @@ var init_voice_runtime = __esm({
     init_listen();
     init_dist5();
     init_voicechat();
+    init_py_embed();
     _voiceEngine = null;
     _listenEngine = null;
     _voiceChatSession = null;
@@ -828251,8 +828451,10 @@ function getOpenApiSpec() {
             include_daemon_tools: { type: "array", items: { type: "string", enum: ["read", "run", "admin"] }, description: "Scopes from which Omnius may offer its bounded core daemon-tool catalog. Use daemon_tool_names to request other exact tools." },
             daemon_tool_names: { type: "array", items: { type: "string" }, description: "Optional exact allowlist of daemon tool names. Use this to keep local-model tool prompts small." },
             agent_timeout_s: { type: "number", minimum: 5, maximum: 600, default: 45, description: "Total server-side agent-loop deadline. Distinct from per-backend timeout_s." },
+            agent_max_tool_rounds: { type: "integer", minimum: 1, maximum: 8, default: 1, description: "Maximum daemon-tool planning rounds. The default executes one tool round, then removes daemon schemas for lower-latency final synthesis." },
+            agent_prefetch_web_search: { type: "boolean", default: false, description: "Explicitly execute authorized web_search with the latest user text before one backend synthesis. factual-first enables this automatically." },
             max_turns: { type: "integer", description: "Q2 — agent_loop max iterations (default 8, max 64)." },
-            prompt_template: { type: "string", enum: ["factual-first"], description: "Q8 — prepended system policy template. 'factual-first' instructs model to call web_search FIRST for any factual question." }
+            prompt_template: { type: "string", enum: ["factual-first"], description: "Factual-first prefetches authorized web_search from the latest user turn and performs one grounded synthesis generation." }
           } } } } },
           responses: { 200: { description: "OpenAI chat.completion shape, SSE if stream=true. agent_loop responses include _agent_loop:{turns,log,done,reason,elapsed_ms,backend_transport}." }, 504: { description: "Backend round or total agent-loop deadline expired" }, 508: { description: "The model repeated an identical daemon tool call" }, ...ErrorResponses }
         }
@@ -828873,7 +829075,7 @@ 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 creates a private CPython 3.10 CPU-only WeSpeaker CAM++ venv, installs only checksum-pinned NumPy, ONNX Runtime, and its locked CPU support wheels with --no-deps/--no-index, then downloads the pinned model. Its 80-bin Kaldi-compatible NumPy fbank is validated with a deterministic signature before readiness. The speaker path never imports, links, replaces, or otherwise depends on JetPack Torch/Torchaudio, so an incompatible generic Torchaudio wheel cannot affect CUDA-enabled Egg Torch. 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 creates a private CPython 3.10 CPU-only WeSpeaker CAM++ venv, installs only checksum-pinned NumPy, ONNX Runtime, and its locked CPU support wheels with --no-deps/--no-index, then downloads the pinned model. Its 80-bin Kaldi-compatible NumPy fbank is validated with a versioned cross-platform numeric-invariant probe before readiness. The speaker path never imports, links, replaces, or otherwise depends on JetPack Torch/Torchaudio, so an incompatible generic Torchaudio wheel cannot affect CUDA-enabled Egg Torch. 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." } }
@@ -830485,6 +830687,32 @@ function getAgentLoopTotalTimeoutMs(body) {
     AGENT_LOOP_TOTAL_TIMEOUT_MAX_MS
   );
 }
+function getAgentLoopPlannerMaxTokens() {
+  return boundedModelCheckInteger(
+    Number(process.env["OMNIUS_AGENT_LOOP_PLANNER_MAX_TOKENS"]),
+    AGENT_LOOP_PLANNER_DEFAULT_MAX_TOKENS,
+    32,
+    512
+  );
+}
+function getAgentLoopMaxToolRounds(body) {
+  const requested = body["agent_max_tool_rounds"];
+  const configured = Number(process.env["OMNIUS_AGENT_LOOP_MAX_TOOL_ROUNDS"]);
+  return boundedModelCheckInteger(
+    requested ?? configured,
+    AGENT_LOOP_MAX_TOOL_ROUNDS_DEFAULT,
+    1,
+    8
+  );
+}
+function getAgentLoopToolResultMaxChars() {
+  return boundedModelCheckInteger(
+    Number(process.env["OMNIUS_AGENT_LOOP_TOOL_RESULT_MAX_CHARS"]),
+    AGENT_LOOP_TOOL_RESULT_DEFAULT_MAX_CHARS,
+    1e3,
+    32e3
+  );
+}
 function compactAgentLoopToolResult(value2) {
   let encoded;
   try {
@@ -830492,11 +830720,12 @@ function compactAgentLoopToolResult(value2) {
   } catch {
     encoded = JSON.stringify({ success: false, error: "tool result was not JSON serializable" });
   }
-  if (encoded.length <= AGENT_LOOP_TOOL_RESULT_MAX_CHARS) return encoded;
+  const maxChars = getAgentLoopToolResultMaxChars();
+  if (encoded.length <= maxChars) return encoded;
   return JSON.stringify({
     truncated: true,
     original_chars: encoded.length,
-    preview: encoded.slice(0, AGENT_LOOP_TOOL_RESULT_MAX_CHARS)
+    preview: encoded.slice(0, maxChars)
   });
 }
 function parseJsonObject8(value2) {
@@ -832987,6 +833216,7 @@ async function runAgentLoopChatCompletions(opts) {
     const targetTransport = requestProviderTransport(targetUrl, route?.endpoint);
     const useNativeOllamaChat = targetTransport.protocol === "ollama";
     const maxTurns = typeof requestBody["max_turns"] === "number" ? Math.max(1, Math.min(64, requestBody["max_turns"])) : 8;
+    const maxDaemonToolRounds = getAgentLoopMaxToolRounds(requestBody);
     const remoteIp = (req3.socket?.remoteAddress || "").replace(/^::ffff:/, "");
     const origin = /^(127\.\d+\.\d+\.\d+|::1|localhost)$/.test(remoteIp) ? "loopback" : "remote";
     const reqAuth = req3;
@@ -833118,6 +833348,77 @@ ${messages2[firstSystemIdx].content}`
     const chatId = `chatcmpl-${randomBytes30(12).toString("hex")}`;
     const turnsLog = [];
     const seenDaemonCalls = /* @__PURE__ */ new Set();
+    let daemonToolRoundsExecuted = 0;
+    const prefetchWebSearch = (promptTemplate === "factual-first" || requestBody["agent_prefetch_web_search"] === true) && advertisedDaemonNames.has("web_search") && callerTools.length === 0 && !messages2.some((message2) => message2.role === "tool" || (message2.tool_calls?.length ?? 0) > 0);
+    const latestUserQuery = [...messages2].reverse().find((message2) => message2.role === "user" && typeof message2.content === "string")?.content?.trim().slice(0, 1e3);
+    if (prefetchWebSearch && latestUserQuery) {
+      const meta = daemonTools.get("web_search");
+      let tool = null;
+      if (meta) {
+        try {
+          tool = new meta.ToolClass(process.cwd());
+        } catch {
+          try {
+            tool = new meta.ToolClass();
+          } catch {
+            tool = null;
+          }
+        }
+      }
+      if (tool) {
+        const callId = `prefetch_web_search_${randomBytes30(6).toString("hex")}`;
+        const args = { query: latestUserQuery };
+        const fingerprint3 = `web_search:${JSON.stringify(args)}`;
+        messages2.push({
+          role: "assistant",
+          content: "",
+          tool_calls: [{
+            id: callId,
+            type: "function",
+            function: { name: "web_search", arguments: JSON.stringify(args) }
+          }]
+        });
+        let toolResult;
+        try {
+          activeTool = tool;
+          const remainingMs = totalDeadline - Date.now();
+          if (remainingMs <= 0) throw new Error("agent loop deadline expired before factual-first search");
+          const execution = await executeAgentLoopToolBounded(
+            tool,
+            args,
+            Math.min(getAgentLoopToolTimeoutMs(), remainingMs),
+            clientDisconnect
+          );
+          if (execution.clientDisconnected || isClientDisconnected()) return;
+          toolResult = execution.result;
+        } catch (error) {
+          if (isClientDisconnected()) return;
+          toolResult = {
+            success: false,
+            output: "",
+            error: error instanceof Error ? error.message : String(error),
+            durationMs: 0
+          };
+        } finally {
+          activeTool = null;
+        }
+        messages2.push({
+          role: "tool",
+          tool_call_id: callId,
+          name: "web_search",
+          content: compactAgentLoopToolResult(toolResult)
+        });
+        seenDaemonCalls.add(fingerprint3);
+        daemonToolRoundsExecuted = 1;
+        turnsLog.push({
+          turn: 0,
+          tool_calls: 1,
+          daemon_executed: 1,
+          client_yielded: 0,
+          prefetched: true
+        });
+      }
+    }
     for (let turn = 1; turn <= maxTurns; turn++) {
       if (isClientDisconnected()) return;
       if (Date.now() > totalDeadline) {
@@ -833142,15 +833443,19 @@ ${messages2[firstSystemIdx].content}`
         return;
       }
       const transportMessages = forceNoThinkForTools ? messages2.map((message3) => typeof message3.content === "string" ? { ...message3, content: stripNoThinkPromptDirectives(message3.content) } : message3) : messages2;
+      const daemonToolsForTurn = daemonToolRoundsExecuted < maxDaemonToolRounds ? daemonToolEntries : [];
+      const toolsForTurn = [...callerTools, ...daemonToolsForTurn];
+      const planningTurn = toolsForTurn.length > 0;
       const ollamaOptions = {};
       if (typeof requestBody["temperature"] === "number") ollamaOptions["temperature"] = requestBody["temperature"];
       if (typeof requestBody["top_p"] === "number") ollamaOptions["top_p"] = requestBody["top_p"];
-      if (typeof requestBody["max_tokens"] === "number") ollamaOptions["num_predict"] = requestBody["max_tokens"];
+      if (planningTurn) ollamaOptions["num_predict"] = getAgentLoopPlannerMaxTokens();
+      else if (typeof requestBody["max_tokens"] === "number") ollamaOptions["num_predict"] = requestBody["max_tokens"];
       if (typeof requestBody["seed"] === "number") ollamaOptions["seed"] = requestBody["seed"];
       const turnBody = useNativeOllamaChat ? {
         model: originalModel,
         messages: agentLoopMessagesForOllama(transportMessages),
-        tools: mergedTools.length > 0 ? mergedTools : void 0,
+        tools: toolsForTurn.length > 0 ? toolsForTurn : void 0,
         stream: false,
         think: false,
         options: ollamaOptions
@@ -833158,7 +833463,8 @@ ${messages2[firstSystemIdx].content}`
         ...requestBody,
         model: originalModel,
         messages: transportMessages,
-        tools: mergedTools.length > 0 ? mergedTools : void 0,
+        tools: toolsForTurn.length > 0 ? toolsForTurn : void 0,
+        ...planningTurn ? { max_tokens: getAgentLoopPlannerMaxTokens() } : {},
         stream: false,
         ...forceNoThinkForTools ? { think: false } : {},
         // Strip our own loop-control fields so they don't confuse the backend
@@ -833166,6 +833472,8 @@ ${messages2[firstSystemIdx].content}`
         include_daemon_tools: void 0,
         daemon_tool_names: void 0,
         agent_timeout_s: void 0,
+        agent_max_tool_rounds: void 0,
+        agent_prefetch_web_search: void 0,
         max_turns: void 0,
         timeout_s: void 0,
         realtime: void 0,
@@ -833432,6 +833740,7 @@ ${messages2[firstSystemIdx].content}`
         daemon_executed: executed,
         client_yielded: 0
       });
+      if (executed > 0) daemonToolRoundsExecuted += 1;
     }
     jsonResponse(res, 200, {
       id: chatId,
@@ -837627,7 +837936,10 @@ data: ${JSON.stringify(data)}
         const baseReadiness = engine.id === "vibevoice-transformers" ? vibe : getManagedAsrReadiness(engine.id, selectedModel ?? engine.models[0].id);
         const readiness = {
           ...baseReadiness,
-          active: unavailableReason ? false : baseReadiness.active,
+          // Direct managed ASR is request-scoped rather than a resident child
+          // process. Expose its successful CUDA inference only for the exact
+          // selected engine/model; never promote an unselected model to active.
+          active: unavailableReason ? false : selected && baseReadiness.active,
           disabled: Boolean(unavailableReason),
           ...unavailableReason ? { unavailableReason, lastError: unavailableReason } : {}
         };
@@ -837641,7 +837953,7 @@ data: ${JSON.stringify(data)}
             const modelBaseReadiness = engine.id === "vibevoice-transformers" ? vibe : getManagedAsrReadiness(engine.id, model.id);
             const modelReadiness = {
               ...modelBaseReadiness,
-              active: unavailableReason ? false : modelBaseReadiness.active,
+              active: unavailableReason ? false : selected && model.id === listen.currentModel && modelBaseReadiness.active,
               disabled: Boolean(unavailableReason),
               ...unavailableReason ? { unavailableReason, lastError: unavailableReason } : {}
             };
@@ -837683,9 +837995,11 @@ data: ${JSON.stringify(data)}
         runtimes: listAsrEngines().filter((engine) => engine.id === "openai-whisper" || engine.id === "nemotron-streaming" || engine.id === "voxtral-transformers").flatMap((engine) => engine.models.map((model) => {
           const readiness = getManagedAsrReadiness(engine.id, model.id);
           const unavailableReason = getAsrEngineUnavailableReason(engine.id);
+          const selected = engine.id === listen.currentEngine && model.id === listen.currentModel;
           return {
             ...readiness,
-            active: unavailableReason ? false : readiness.active,
+            active: unavailableReason ? false : selected && readiness.active,
+            selected,
             disabled: Boolean(unavailableReason),
             ...unavailableReason ? { unavailableReason, lastError: unavailableReason } : {}
           };
@@ -841937,7 +842251,7 @@ 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, AGENT_LOOP_TOTAL_TIMEOUT_DEFAULT_MS, AGENT_LOOP_TOTAL_TIMEOUT_MIN_MS, AGENT_LOOP_TOTAL_TIMEOUT_MAX_MS, AGENT_LOOP_TOOL_TIMEOUT_DEFAULT_MS, AGENT_LOOP_TOOL_TIMEOUT_MAX_MS, AGENT_LOOP_TOOL_RESULT_MAX_CHARS, DEFAULT_AGENT_LOOP_DAEMON_TOOL_NAMES, 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 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_TIMEOUT_DEFAULT_MS, AGENT_LOOP_TOTAL_TIMEOUT_MIN_MS, AGENT_LOOP_TOTAL_TIMEOUT_MAX_MS, AGENT_LOOP_TOOL_TIMEOUT_DEFAULT_MS, AGENT_LOOP_TOOL_TIMEOUT_MAX_MS, AGENT_LOOP_TOOL_RESULT_DEFAULT_MAX_CHARS, AGENT_LOOP_PLANNER_DEFAULT_MAX_TOKENS, AGENT_LOOP_MAX_TOOL_ROUNDS_DEFAULT, DEFAULT_AGENT_LOOP_DAEMON_TOOL_NAMES, 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();
@@ -841998,7 +842312,9 @@ var init_serve = __esm({
     AGENT_LOOP_TOTAL_TIMEOUT_MAX_MS = 10 * 60 * 1e3;
     AGENT_LOOP_TOOL_TIMEOUT_DEFAULT_MS = 3e4;
     AGENT_LOOP_TOOL_TIMEOUT_MAX_MS = 12e4;
-    AGENT_LOOP_TOOL_RESULT_MAX_CHARS = 16e3;
+    AGENT_LOOP_TOOL_RESULT_DEFAULT_MAX_CHARS = 6e3;
+    AGENT_LOOP_PLANNER_DEFAULT_MAX_TOKENS = 96;
+    AGENT_LOOP_MAX_TOOL_ROUNDS_DEFAULT = 1;
     DEFAULT_AGENT_LOOP_DAEMON_TOOL_NAMES = /* @__PURE__ */ new Set([
       "web_search",
       "web_fetch",