omnius 1.0.186 → 1.0.188

package/dist/index.js

@@ -1412,6 +1412,18 @@ var init_tool_executor = __esm({
 import { EventEmitter } from "node:events";
 import { totalmem, freemem } from "node:os";
 import { exec } from "node:child_process";
+function dedupeLoadedModels(models) {
+  const seen = /* @__PURE__ */ new Set();
+  const out = [];
+  for (const model of models) {
+    const key = `${model.host}:${model.name}`;
+    if (seen.has(key))
+      continue;
+    seen.add(key);
+    out.push(model);
+  }
+  return out;
+}
 function ramSnapshotMB() {
   const total = Math.round(totalmem() / (1024 * 1024));
   const free = Math.round(freemem() / (1024 * 1024));
@@ -1690,6 +1702,104 @@ var init_model_broker = __esm({
         this.emit("rejected", spec, reason);
         return { kind: "reject", reason };
       }
+      /**
+       * Acquire a short-lived load lease for media/subprocess generation.
+       *
+       * Media generation often needs to temporarily free VRAM occupied by Ollama
+       * chat models. This helper refreshes Ollama state, asks the broker what must
+       * be evicted, unloads those Ollama models with keep_alive=0, and returns a
+       * lease whose release() clears transient inflight state, unloads any
+       * Ollama-hosted requested model, and warms the evicted Ollama models again.
+       */
+      async acquireTransientModelLoad(spec, options2 = {}) {
+        const reason = options2.reason ?? `${spec.domain}-transient-load`;
+        const evictedModels = [];
+        let gpuIndex = null;
+        let admitted = false;
+        await this.pollOnce().catch(() => {
+        });
+        for (let attempt = 0; attempt < 4; attempt++) {
+          const decision2 = await this.ensureModelLoadable(spec);
+          if (decision2.kind === "wait-for-inflight") {
+            const waited = await decision2.promise.catch((err) => ({
+              kind: "reject",
+              reason: err instanceof Error ? err.message : String(err)
+            }));
+            if (waited.kind === "ok") {
+              gpuIndex = waited.gpuIndex ?? null;
+              admitted = true;
+              break;
+            }
+            if (waited.kind === "evict") {
+              for (const target of waited.evictTargets) {
+                if (await this.evict(target.host, target.name, reason))
+                  evictedModels.push(target);
+              }
+              await this.pollOnce().catch(() => {
+              });
+              continue;
+            }
+            if (waited.kind === "degrade")
+              return waited;
+            if (waited.kind === "reject")
+              return waited;
+            continue;
+          }
+          if (decision2.kind === "ok") {
+            gpuIndex = decision2.gpuIndex ?? null;
+            admitted = true;
+            break;
+          }
+          if (decision2.kind === "evict") {
+            for (const target of decision2.evictTargets) {
+              const before = this._loaded.get(`${target.host}:${target.name}`) ?? target;
+              if (await this.evict(target.host, target.name, reason)) {
+                evictedModels.push(before);
+              }
+            }
+            await this.pollOnce().catch(() => {
+            });
+            continue;
+          }
+          if (decision2.kind === "degrade")
+            return decision2;
+          return decision2;
+        }
+        if (!admitted) {
+          return {
+            kind: "reject",
+            reason: `could not acquire transient load lease for ${spec.host}:${spec.name} after repeated evictions`
+          };
+        }
+        const evictedOllamaModels = dedupeLoadedModels(evictedModels.filter((m2) => m2.host === "ollama"));
+        const broker = this;
+        let released = false;
+        return {
+          kind: "ok",
+          lease: {
+            spec,
+            gpuIndex,
+            evictedModels: dedupeLoadedModels(evictedModels),
+            evictedOllamaModels,
+            async release() {
+              if (released)
+                return;
+              released = true;
+              broker.clearInflight(spec.host, spec.name);
+              if ((options2.unloadRequestedOllama ?? true) && spec.host === "ollama") {
+                await broker.unloadOllamaModel(spec.name, `${reason}-complete`).catch(() => false);
+              }
+              if (options2.restoreOllama !== false && evictedOllamaModels.length > 0) {
+                await broker.restoreOllamaModels(evictedOllamaModels, {
+                  keepAlive: options2.restoreKeepAlive ?? "30m"
+                });
+              }
+              await broker.pollOnce().catch(() => {
+              });
+            }
+          }
+        };
+      }
       /**
        * Register a model that has been successfully loaded.
        * Callers MUST call this after a successful load so the broker can track LRU.
@@ -1743,6 +1853,66 @@ var init_model_broker = __esm({
         this.emit("evicted", m2, reason);
         return actively;
       }
+      /** Force-unload an Ollama model even when it is not currently tracked. */
+      async unloadOllamaModel(modelName, reason = "ollama-unload") {
+        const key = `ollama:${modelName}`;
+        const existing = this._loaded.get(key);
+        const ok3 = await this.ollamaUnload(modelName).catch(() => false);
+        if (existing) {
+          this._loaded.delete(key);
+          this.emit("evicted", existing, reason);
+        }
+        return ok3;
+      }
+      /** Best-effort warm/reload of an Ollama model after temporary eviction. */
+      async warmOllamaModel(modelName, keepAlive = "30m") {
+        const bodies = [
+          { model: modelName, prompt: "", stream: false, keep_alive: keepAlive, options: { num_predict: 0 } },
+          { model: modelName, prompt: "", stream: false, keep_alive: keepAlive, options: { num_predict: 1 } }
+        ];
+        for (const body of bodies) {
+          try {
+            const res = await fetch(`${this._ollamaBaseUrl}/api/generate`, {
+              method: "POST",
+              headers: { "Content-Type": "application/json" },
+              body: JSON.stringify(body),
+              signal: AbortSignal.timeout(12e4)
+            });
+            if (!res.ok)
+              continue;
+            await this.refreshOllamaPs().catch(() => {
+            });
+            return true;
+          } catch {
+          }
+        }
+        try {
+          const res = await fetch(`${this._ollamaBaseUrl}/api/generate`, {
+            method: "POST",
+            headers: { "Content-Type": "application/json" },
+            body: JSON.stringify({
+              model: modelName,
+              stream: false,
+              keep_alive: keepAlive
+            }),
+            signal: AbortSignal.timeout(12e4)
+          });
+          if (!res.ok)
+            return false;
+          await this.refreshOllamaPs().catch(() => {
+          });
+          return true;
+        } catch {
+          return false;
+        }
+      }
+      /** Restore a set of previously evicted Ollama models, oldest first. */
+      async restoreOllamaModels(models, options2 = {}) {
+        const unique2 = dedupeLoadedModels(models.filter((m2) => m2.host === "ollama")).sort((a2, b) => a2.lastUsedAt - b.lastUsedAt);
+        for (const model of unique2) {
+          await this.warmOllamaModel(model.name, options2.keepAlive ?? "30m").catch(() => false);
+        }
+      }
       // ------------------------------------------------------------------
       // Internal — Ollama
       // ------------------------------------------------------------------
@@ -1885,7 +2055,7 @@ var init_model_broker = __esm({
         );
         const idle = (m2) => now - m2.lastUsedAt > this.idleEvictMs;
         const onTargetGpu = (m2) => req2.targetGpu === void 0 || req2.targetGpu === null ? true : m2.gpuIndex === req2.targetGpu;
-        const evictable = [...this._loaded.values()].filter((m2) => m2.priority <= req2.requestingPriority).filter(sameDomainOk).filter(onTargetGpu).sort((a2, b) => {
+        const evictable = [...this._loaded.values()].filter((m2) => m2.priority <= req2.requestingPriority).filter(sameDomainOk).filter(onTargetGpu).filter((m2) => !this.hasActiveSlotForModel(m2)).sort((a2, b) => {
           const aIdle = idle(a2) ? 0 : 1;
           const bIdle = idle(b) ? 0 : 1;
           if (aIdle !== bIdle)
@@ -1931,6 +2101,13 @@ var init_model_broker = __esm({
             n2++;
         return n2;
       }
+      hasActiveSlotForModel(model) {
+        for (const slot of this._activeSlots.values()) {
+          if (slot.model === model.name)
+            return true;
+        }
+        return false;
+      }
       // ------------------------------------------------------------------
       // Internal — fallback resolution
       // ------------------------------------------------------------------
@@ -22761,6 +22938,17 @@ function evictModelsToFreeSpace(args) {
   writeMeta(meta);
   return { evicted, bytesFreed, finalFreeBytes: disk.freeBytes };
 }
+function estimateReclaimableCacheBytes(keepRepos) {
+  const keep = new Set(keepRepos ?? []);
+  let total = 0;
+  for (const entry of readMeta().entries) {
+    if (keep.has(entry.repo))
+      continue;
+    const measured = measureRepoCacheBytes(entry.repo);
+    total += Math.max(0, measured || entry.sizeBytes || 0);
+  }
+  return total;
+}
 function ensureDiskSpaceForDownload(args) {
   ensureUnifiedCacheDirs();
   const safetyMargin = args.safetyMarginBytes ?? 1 * 1024 ** 3;
@@ -22769,6 +22957,10 @@ function ensureDiskSpaceForDownload(args) {
   if (disk.freeBytes >= target) {
     return { ok: true, evicted: [], freeBytes: disk.freeBytes };
   }
+  const reclaimableBytes = estimateReclaimableCacheBytes(args.keepRepos);
+  if (disk.freeBytes + reclaimableBytes < target) {
+    throw new InsufficientDiskSpaceError(args.approxDownloadBytes, disk.freeBytes, disk.totalBytes, []);
+  }
   const evictionResult = evictModelsToFreeSpace({
     neededBytes: args.approxDownloadBytes,
     keepRepos: args.keepRepos,
@@ -25161,11 +25353,154 @@ ${content}`
   }
 });
 
+// packages/execution/dist/tools/cuda-device-filter.js
+import { execFileSync as execFileSync3 } from "node:child_process";
+function cleanEnvValue(value2) {
+  const trimmed = value2?.trim();
+  return trimmed ? trimmed : void 0;
+}
+function envFlagEnabled(value2) {
+  return /^(1|true|yes|on)$/i.test(value2?.trim() ?? "");
+}
+function parseCudaComputeCapability(value2) {
+  const match = value2?.match(/(\d+(?:\.\d+)?)/);
+  if (!match)
+    return null;
+  const parsed = Number(match[1]);
+  return Number.isFinite(parsed) ? parsed : null;
+}
+function modalityEnvPrefix(modality) {
+  return `OMNIUS_${modality.toUpperCase()}`;
+}
+function parseMinCudaComputeCapability(modality, env2) {
+  const prefix = modalityEnvPrefix(modality);
+  const parsed = parseCudaComputeCapability(env2[`${prefix}_MIN_CUDA_CC`]) ?? parseCudaComputeCapability(env2["OMNIUS_MEDIA_MIN_CUDA_CC"]);
+  return parsed && parsed > 0 ? parsed : DEFAULT_MEDIA_MIN_CUDA_COMPUTE_CAPABILITY;
+}
+function splitCudaVisibleDevices(value2) {
+  return (value2 ?? "").split(",").map((part) => part.trim()).filter(Boolean);
+}
+function parseCudaDeviceInfo(text) {
+  const devices = [];
+  for (const line of text.split(/\r?\n/)) {
+    const trimmed = line.trim();
+    if (!trimmed)
+      continue;
+    const parts = trimmed.split(",").map((part) => part.trim());
+    if (parts.length < 4)
+      continue;
+    const index = Number.parseInt(parts.shift() ?? "", 10);
+    const capability = parseCudaComputeCapability(parts.pop());
+    const uuid = cleanEnvValue(parts.shift());
+    const name10 = cleanEnvValue(parts.join(", "));
+    if (!Number.isFinite(index) || index < 0)
+      continue;
+    devices.push({ index, uuid, name: name10, computeCapability: capability });
+  }
+  return devices;
+}
+function detectCudaDevices() {
+  try {
+    const out = execFileSync3("nvidia-smi", ["--query-gpu=index,uuid,name,compute_cap", "--format=csv,noheader,nounits"], {
+      encoding: "utf8",
+      timeout: 5e3,
+      stdio: ["ignore", "pipe", "ignore"]
+    });
+    return parseCudaDeviceInfo(out);
+  } catch {
+    return [];
+  }
+}
+function resolveMediaCudaVisibleDevicesForEnv(args) {
+  const env2 = args.env ?? process.env;
+  const prefix = modalityEnvPrefix(args.modality);
+  const explicit = cleanEnvValue(env2[`${prefix}_CUDA_VISIBLE_DEVICES`]) ?? cleanEnvValue(env2["OMNIUS_MEDIA_CUDA_VISIBLE_DEVICES"]) ?? (args.modality === "audio" ? cleanEnvValue(env2["OMNIUS_AUDIO_GPU"]) : void 0);
+  if (explicit)
+    return explicit;
+  const current = cleanEnvValue(env2["CUDA_VISIBLE_DEVICES"]);
+  if (envFlagEnabled(env2[`${prefix}_DISABLE_CUDA_FILTER`]) || envFlagEnabled(env2["OMNIUS_MEDIA_DISABLE_CUDA_FILTER"])) {
+    return current;
+  }
+  const devices = args.devices ?? detectCudaDevices();
+  const minComputeCapability = args.minComputeCapability ?? parseMinCudaComputeCapability(args.modality, env2);
+  const compatible = devices.filter((device) => device.computeCapability !== null && device.computeCapability >= minComputeCapability);
+  if (compatible.length === 0)
+    return current;
+  const compatibleTokens = /* @__PURE__ */ new Set();
+  for (const device of compatible) {
+    compatibleTokens.add(String(device.index));
+    if (device.uuid)
+      compatibleTokens.add(device.uuid);
+  }
+  if (current) {
+    const requested = splitCudaVisibleDevices(current);
+    const canFilter = requested.length > 0 && requested.every((token) => /^\d+$/.test(token) || token.startsWith("GPU-"));
+    if (!canFilter)
+      return current;
+    const kept = requested.filter((token) => compatibleTokens.has(token));
+    return (kept.length > 0 ? kept : compatible.map((device) => String(device.index))).join(",");
+  }
+  return compatible.map((device) => String(device.index)).join(",");
+}
+function mediaBrokerGpuIndexIsCompatible(gpuIndex, modality, env2 = process.env, devices = detectCudaDevices()) {
+  const prefix = modalityEnvPrefix(modality);
+  if (envFlagEnabled(env2[`${prefix}_DISABLE_CUDA_FILTER`]) || envFlagEnabled(env2["OMNIUS_MEDIA_DISABLE_CUDA_FILTER"])) {
+    return true;
+  }
+  if (devices.length === 0)
+    return true;
+  const minComputeCapability = parseMinCudaComputeCapability(modality, env2);
+  const device = devices.find((candidate) => candidate.index === gpuIndex);
+  if (!device || device.computeCapability === null)
+    return true;
+  return device.computeCapability >= minComputeCapability;
+}
+function applyMediaCudaDeviceFilterToEnv(env2, modality) {
+  const cudaVisibleDevices = resolveMediaCudaVisibleDevicesForEnv({ modality, env: env2 });
+  if (cudaVisibleDevices) {
+    env2["CUDA_VISIBLE_DEVICES"] = cudaVisibleDevices;
+    env2["PYTORCH_NVML_BASED_CUDA_CHECK"] = process.env["PYTORCH_NVML_BASED_CUDA_CHECK"] ?? "1";
+  }
+  return env2;
+}
+var DEFAULT_MEDIA_MIN_CUDA_COMPUTE_CAPABILITY;
+var init_cuda_device_filter = __esm({
+  "packages/execution/dist/tools/cuda-device-filter.js"() {
+    "use strict";
+    DEFAULT_MEDIA_MIN_CUDA_COMPUTE_CAPABILITY = 7.5;
+  }
+});
+
 // packages/execution/dist/tools/transcribe-tool.js
 import { existsSync as existsSync29, mkdirSync as mkdirSync13, writeFileSync as writeFileSync14, readFileSync as readFileSync22, unlinkSync as unlinkSync3, readdirSync as readdirSync13 } from "node:fs";
 import { join as join32, basename as basename6, extname as extname3, resolve as resolve17 } from "node:path";
 import { homedir as homedir10 } from "node:os";
-import { execFileSync as execFileSync3, execSync as execSync15 } from "node:child_process";
+import { execFileSync as execFileSync4, execSync as execSync15 } from "node:child_process";
+function transcriptionPythonEnv(extra = {}) {
+  const env2 = { ...process.env, ...extra };
+  applyMediaCudaDeviceFilterToEnv(env2, "asr");
+  return env2;
+}
+async function withProcessEnv(env2, fn) {
+  const previous = /* @__PURE__ */ new Map();
+  for (const [key, value2] of Object.entries(env2)) {
+    previous.set(key, process.env[key]);
+    if (value2 === void 0)
+      delete process.env[key];
+    else
+      process.env[key] = value2;
+  }
+  try {
+    return await fn();
+  } finally {
+    for (const [key, value2] of previous) {
+      if (value2 === void 0)
+        delete process.env[key];
+      else
+        process.env[key] = value2;
+    }
+  }
+}
 function whisperRamEstimate(model) {
   const m2 = model.toLowerCase();
   if (m2.includes("large"))
@@ -25251,6 +25586,7 @@ var init_transcribe_tool = __esm({
     "use strict";
     init_model_broker();
     init_network_egress_policy();
+    init_cuda_device_filter();
     AUDIO_EXTS = /* @__PURE__ */ new Set([
       ".mp3",
       ".wav",
@@ -25353,13 +25689,13 @@ var init_transcribe_tool = __esm({
           return this.execViaCli(filePath, model, diarize, start2);
         }
         try {
-          const result = await tc.transcribe(filePath, {
+          const result = await withProcessEnv(transcriptionPythonEnv(), () => tc.transcribe(filePath, {
             model,
             format: "json",
             diarize,
             wordTimestamps: true
             // Always get timestamps for structured output
-          });
+          }));
           const transcriptDir = join32(this.workingDir, ".omnius", "transcripts");
           mkdirSync13(transcriptDir, { recursive: true });
           const fileBase = basename6(filePath).replace(/\.[^.]+$/, "");
@@ -25453,7 +25789,8 @@ var init_transcribe_tool = __esm({
             timeout: 3e5,
             // 5 min max
             cwd: this.workingDir,
-            stdio: ["pipe", "pipe", "pipe"]
+            stdio: ["pipe", "pipe", "pipe"],
+            env: transcriptionPythonEnv()
           });
           return {
             success: true,
@@ -25522,7 +25859,7 @@ var init_transcribe_tool = __esm({
             }
             tmpFile = `${tmpBase}.mp3`;
             try {
-              execFileSync3("yt-dlp", ["-x", "--audio-format", "mp3", "--audio-quality", "5", "-o", `${tmpBase}.%(ext)s`, url], { timeout: 3e5, stdio: ["pipe", "pipe", "pipe"] });
+              execFileSync4("yt-dlp", ["-x", "--audio-format", "mp3", "--audio-quality", "5", "-o", `${tmpBase}.%(ext)s`, url], { timeout: 3e5, stdio: ["pipe", "pipe", "pipe"] });
               if (!existsSync29(tmpFile)) {
                 const files = readdirSync13(tmpDir).filter((f2) => f2.startsWith(`download-`) && f2 !== ".gitkeep");
                 const match = files.find((f2) => f2.includes(basename6(tmpBase)));
@@ -25647,13 +25984,13 @@ ${result.output}`,
         try {
           let title = "download";
           try {
-            title = execFileSync3("yt-dlp", ["--get-title", url], { timeout: 15e3, stdio: "pipe" }).toString().trim().replace(/[<>:"/\\|?*]/g, "_").slice(0, 100);
+            title = execFileSync4("yt-dlp", ["--get-title", url], { timeout: 15e3, stdio: "pipe" }).toString().trim().replace(/[<>:"/\\|?*]/g, "_").slice(0, 100);
           } catch {
           }
           if (format3 === "mp4") {
             const outPath = join32(outputDir, `${title}.mp4`);
             const outTemplate = join32(outputDir, `${title}.%(ext)s`);
-            execFileSync3("yt-dlp", [
+            execFileSync4("yt-dlp", [
               "-f",
               "bestvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best",
               "--merge-output-format",
@@ -25673,7 +26010,7 @@ Format: mp4`,
           } else {
             const outPath = join32(outputDir, `${title}.mp3`);
             const outTemplate = join32(outputDir, `${title}.%(ext)s`);
-            execFileSync3("yt-dlp", ["-x", "--audio-format", "mp3", "--audio-quality", "0", "-o", outTemplate, url], { timeout: 6e5, stdio: "pipe", cwd: outputDir });
+            execFileSync4("yt-dlp", ["-x", "--audio-format", "mp3", "--audio-quality", "0", "-o", outTemplate, url], { timeout: 6e5, stdio: "pipe", cwd: outputDir });
             const actualPath = existsSync29(outPath) ? outPath : outTemplate.replace("%(ext)s", "mp3");
             return {
               success: true,
@@ -259063,6 +259400,12 @@ function imageCandidateFor(model, requestedBackend) {
     preset: getImageGenerationPreset(resolved)
   };
 }
+function imageCandidateHost(candidate) {
+  return candidate.backend === "ollama" ? "ollama" : "diffusers-py";
+}
+function imageCandidateEstimatedVramMB(candidate) {
+  return candidate.preset?.minVramGB !== void 0 ? Math.ceil(candidate.preset.minVramGB * 1024) : void 0;
+}
 function imageGenerationFallbackCandidates(requestedModel, requestedBackend, allowFallback = true) {
   const ladder = imageGenerationQualityLadder();
   const candidates = [];
@@ -259246,7 +259589,9 @@ function formatDiffusersFailure(stderrOrStdout) {
 Note: ${note}`)].filter(Boolean).join("");
 }
 function imageGenerationPythonEnv(_repoRoot) {
-  return unifiedPythonEnv();
+  const env2 = unifiedPythonEnv();
+  applyMediaCudaDeviceFilterToEnv(env2, "image");
+  return env2;
 }
 function approxImageDownloadBytes(preset) {
   if (preset?.approxDownloadGB)
@@ -259364,9 +259709,15 @@ function annotateImageFallbackSuccess(result, failed, winner) {
     ...failed.map((attempt, index) => `  ${formatImageAttempt(attempt.candidate, attempt.reason, index)}`),
     ""
   ].join("\n");
+  const llmPrefix = [
+    `Fallback ladder used ${winner.model} [${winner.backend}] after ${failed.length} failed attempt(s).`,
+    ...failed.map((attempt, index) => formatImageAttempt(attempt.candidate, attempt.reason, index))
+  ].join("\n");
   return {
     ...result,
-    output: prefix + result.output
+    output: prefix + result.output,
+    llmContent: result.llmContent ? `${llmPrefix}
+${result.llmContent}` : result.llmContent
   };
 }
 function parseRunnerJson(stdout) {
@@ -259389,6 +259740,7 @@ var init_image_generate = __esm({
     init_venv_paths();
     init_model_store();
     init_hf_media_models();
+    init_cuda_device_filter();
     DEFAULT_DIFFUSERS_IMAGE_MODEL = "Efficient-Large-Model/SANA1.5_1.6B_1024px_diffusers";
     DEFAULT_OLLAMA_IMAGE_MODEL = "x/flux2-klein";
     LEGACY_SDXL_TURBO_MODEL = "stabilityai/sdxl-turbo";
@@ -260173,6 +260525,45 @@ if __name__ == "__main__":
         this.lastProgressAt = now;
         this.progressHandler(event);
       }
+      async acquireTransientLoadLease(args) {
+        if (!args.candidate)
+          return null;
+        const broker = getModelBroker();
+        const decision2 = await broker.acquireTransientModelLoad({
+          name: args.candidate.model,
+          domain: "image-gen",
+          host: imageCandidateHost(args.candidate),
+          owner: "image-generate-tool",
+          estimatedVramMB: imageCandidateEstimatedVramMB(args.candidate)
+        }, {
+          reason: args.reason,
+          restoreOllama: true,
+          unloadRequestedOllama: true
+        });
+        if (decision2.kind === "reject") {
+          return {
+            success: false,
+            output: "",
+            error: `Image generation blocked by resource broker: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.kind === "degrade") {
+          return {
+            success: false,
+            output: "",
+            error: `Image generation needs a broker fallback (${decision2.fallback.name}), but image candidate fallback must be selected by the image ladder: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.lease.evictedOllamaModels.length > 0) {
+          this.emitProgress({
+            stage: "setup",
+            message: `Temporarily unloaded ${decision2.lease.evictedOllamaModels.length} Ollama inference model(s) to free VRAM for image generation`
+          });
+        }
+        return { lease: decision2.lease, gpuIndex: decision2.lease.gpuIndex };
+      }
       async execute(args) {
         const start2 = performance.now();
         const action = String(args["action"] ?? "generate");
@@ -260215,33 +260606,6 @@ if __name__ == "__main__":
         const requestedBackend = args["backend"] ? String(args["backend"]) : this.defaultBackend;
         const seed = optionalNumberArg(args["seed"]);
         const candidates = imageGenerationFallbackCandidates(requestedModel, requestedBackend, generationFallbackEnabled(args));
-        const broker = getModelBroker();
-        const firstCandidate = candidates[0];
-        let brokerGpuIndex = null;
-        if (firstCandidate) {
-          const decision2 = await broker.ensureModelLoadable({
-            name: firstCandidate.model,
-            domain: "image-gen",
-            host: firstCandidate.backend === "ollama" ? "ollama" : "diffusers-py",
-            owner: "image-generate-tool"
-          });
-          if (decision2.kind === "evict") {
-            for (const target of decision2.evictTargets) {
-              await broker.evict(target.host, target.name, "image-gen-needs-room");
-            }
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "ok") {
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "reject") {
-            return {
-              success: false,
-              output: "",
-              error: `Image generation blocked by resource broker: ${decision2.reason}`,
-              durationMs: performance.now() - start2
-            };
-          }
-        }
-        this._brokerGpuIndex = brokerGpuIndex;
         try {
           return await this.generateCandidateLadder({ candidates, prompt, args, seed, start: start2 });
         } catch (err) {
@@ -260260,7 +260624,30 @@ if __name__ == "__main__":
             stage: "setup",
             message: `Preparing image model ${candidate.model} (${candidate.backend}) [${index + 1}/${args.candidates.length}]`
           });
-          const result = candidate.backend === "ollama" ? await this.prewarmOllama({ model: candidate.model, start: args.start }) : candidate.backend === "sdcpp" ? await this.prewarmSdCpp({ model: candidate.model, start: args.start, python: args.args["python"] }) : await this.prewarmDiffusers({ model: candidate.model, start: args.start, python: args.args["python"] });
+          const leaseDecision = await this.acquireTransientLoadLease({
+            candidate,
+            reason: "image-prewarm-needs-room",
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; trying ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = candidate.backend === "ollama" ? await this.prewarmOllama({ model: candidate.model, start: args.start }) : candidate.backend === "sdcpp" ? await this.prewarmSdCpp({ model: candidate.model, start: args.start, python: args.args["python"] }) : await this.prewarmDiffusers({ model: candidate.model, start: args.start, python: args.args["python"] });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success)
             return annotateImageFallbackSuccess(result, failed, candidate);
           failed.push({ candidate, reason: summarizeToolResult(result) });
@@ -260299,7 +260686,30 @@ if __name__ == "__main__":
             message: `Using image model ${candidate.model} (${candidate.backend}) [${index + 1}/${args.candidates.length}]`
           });
           const promptForCandidate = expansionEnabled ? await this.expandPromptForCandidate(args.prompt, candidate, index, args.candidates.length) : args.prompt;
-          const result = candidate.backend === "ollama" ? await this.generateWithOllama({ prompt: promptForCandidate, model: candidate.model, width, height, steps, start: args.start }) : candidate.backend === "sdcpp" ? await this.generateWithSdCpp({ prompt: promptForCandidate, model: candidate.model, width, height, steps, seed: args.seed, start: args.start, python: args.args["python"] }) : await this.generateWithDiffusers({ prompt: promptForCandidate, model: candidate.model, width, height, steps, guidance, seed: args.seed, start: args.start, python: args.args["python"] });
+          const leaseDecision = await this.acquireTransientLoadLease({
+            candidate,
+            reason: "image-gen-needs-room",
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; falling back to ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = candidate.backend === "ollama" ? await this.generateWithOllama({ prompt: promptForCandidate, model: candidate.model, width, height, steps, start: args.start }) : candidate.backend === "sdcpp" ? await this.generateWithSdCpp({ prompt: promptForCandidate, model: candidate.model, width, height, steps, seed: args.seed, start: args.start, python: args.args["python"] }) : await this.generateWithDiffusers({ prompt: promptForCandidate, model: candidate.model, width, height, steps, guidance, seed: args.seed, start: args.start, python: args.args["python"] });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success) {
             await this.writeImageSidecar(result, {
               originalPrompt: args.prompt,
@@ -260523,6 +260933,17 @@ ${errText.slice(0, 1200)}`,
         }
         ensureUnifiedCacheDirs();
         this.emitProgress({ stage: "load", message: `Downloading/loading image model ${args.model}` });
+        const runnerEnv = { ...python.env };
+        if (this._brokerGpuIndex !== null) {
+          if (mediaBrokerGpuIndexIsCompatible(this._brokerGpuIndex, "image", runnerEnv)) {
+            runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          } else {
+            this.emitProgress({
+              stage: "setup",
+              message: `Broker selected CUDA GPU ${this._brokerGpuIndex}, but image CUDA filtering excluded it; using CUDA_VISIBLE_DEVICES=${runnerEnv["CUDA_VISIBLE_DEVICES"] ?? "default"}`
+            });
+          }
+        }
         const result = await runProcess2(python.command, [
           runner,
           "--model",
@@ -260535,7 +260956,7 @@ ${errText.slice(0, 1200)}`,
         ], {
           cwd: this.cwd,
           timeoutMs: 18e5,
-          env: python.env,
+          env: runnerEnv,
           progressLabel: `Downloading/loading ${args.model}`,
           onProgress: (event) => this.emitProgress(event)
         });
@@ -260746,7 +261167,14 @@ ${errText.slice(0, 800)}`,
         this.emitProgress({ stage: "load", message: `Starting image generation with ${args.model}` });
         const runnerEnv = { ...python.env };
         if (this._brokerGpuIndex !== null) {
-          runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          if (mediaBrokerGpuIndexIsCompatible(this._brokerGpuIndex, "image", runnerEnv)) {
+            runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          } else {
+            this.emitProgress({
+              stage: "setup",
+              message: `Broker selected CUDA GPU ${this._brokerGpuIndex}, but image CUDA filtering excluded it; using CUDA_VISIBLE_DEVICES=${runnerEnv["CUDA_VISIBLE_DEVICES"] ?? "default"}`
+            });
+          }
         }
         const result = await runProcess2(python.command, argv, {
           cwd: this.cwd,
@@ -260948,6 +261376,7 @@ __export(audio_generate_exports, {
   AudioGenerateTool: () => AudioGenerateTool,
   DEFAULT_MUSIC_MODEL: () => DEFAULT_MUSIC_MODEL,
   DEFAULT_SOUND_MODEL: () => DEFAULT_SOUND_MODEL,
+  audioBrokerGpuIndexIsCompatible: () => audioBrokerGpuIndexIsCompatible,
   audioGenerationDir: () => audioGenerationDir,
   audioGenerationFallbackCandidates: () => audioGenerationFallbackCandidates,
   audioGenerationModelPresets: () => audioGenerationModelPresets,
@@ -260958,9 +261387,11 @@ __export(audio_generate_exports, {
   findNonGatedAudioFallback: () => findNonGatedAudioFallback,
   getAudioGenerationPreset: () => getAudioGenerationPreset,
   inferAudioGenerationBackend: () => inferAudioGenerationBackend,
-  isAudioPresetGated: () => isAudioPresetGated
+  isAudioPresetGated: () => isAudioPresetGated,
+  parseAudioCudaDeviceInfo: () => parseAudioCudaDeviceInfo,
+  resolveAudioCudaVisibleDevicesForEnv: () => resolveAudioCudaVisibleDevicesForEnv
 });
-import { execFileSync as execFileSync4, spawn as spawn8 } from "node:child_process";
+import { spawn as spawn8 } from "node:child_process";
 import { existsSync as existsSync32, readdirSync as readdirSync14, statSync as statSync13 } from "node:fs";
 import { chmod as chmod4, mkdir as mkdir13, writeFile as writeFile18 } from "node:fs/promises";
 import { join as join42 } from "node:path";
@@ -261025,25 +261456,36 @@ function backendPackages(backend) {
     return TANGOFLUX_PACKAGES;
   return DIFFUSERS_AUDIO_PACKAGES;
 }
+function splitCudaVisibleDevices2(value2) {
+  return (value2 ?? "").split(",").map((part) => part.trim()).filter(Boolean);
+}
+function parseAudioCudaDeviceInfo(text) {
+  return parseCudaDeviceInfo(text);
+}
+function detectAudioCudaDevices() {
+  return detectCudaDevices();
+}
+function resolveAudioCudaVisibleDevicesForEnv(args = {}) {
+  return resolveMediaCudaVisibleDevicesForEnv({
+    modality: "audio",
+    env: args.env,
+    devices: args.devices,
+    minComputeCapability: args.minComputeCapability
+  });
+}
+function audioBrokerGpuIndexIsCompatible(gpuIndex, env2 = process.env, devices = detectAudioCudaDevices()) {
+  return mediaBrokerGpuIndexIsCompatible(gpuIndex, "audio", env2, devices);
+}
 function detectLegacyCudaComputeCapability() {
-  try {
-    const out = execFileSync4("nvidia-smi", ["--query-gpu=compute_cap,name", "--format=csv,noheader,nounits"], {
-      encoding: "utf8",
-      timeout: 5e3,
-      stdio: ["ignore", "pipe", "ignore"]
-    }).trim();
-    const first2 = out.split(/\r?\n/).map((line) => line.trim()).find(Boolean);
-    const match = first2?.match(/^(\d+)\.(\d+)\s*,?\s*(.*)$/);
-    if (!match)
-      return null;
-    const major = Number(match[1]);
-    const minor = Number(match[2]);
-    if (!Number.isFinite(major) || !Number.isFinite(minor))
-      return null;
-    return { major, minor, name: match[3]?.trim() || void 0 };
-  } catch {
+  const devices = detectAudioCudaDevices();
+  const visible = new Set(splitCudaVisibleDevices2(resolveAudioCudaVisibleDevicesForEnv({ devices })));
+  const candidates = visible.size > 0 ? devices.filter((device) => visible.has(String(device.index)) || device.uuid && visible.has(device.uuid)) : devices;
+  const legacy = candidates.find((device) => device.computeCapability !== null && isLegacyCudaCapability(Math.floor(device.computeCapability), Math.round(device.computeCapability % 1 * 10)));
+  if (!legacy || legacy.computeCapability === null)
     return null;
-  }
+  const major = Math.floor(legacy.computeCapability);
+  const minor = Math.round(legacy.computeCapability % 1 * 10);
+  return { major, minor, name: legacy.name };
 }
 function isLegacyCudaCapability(major, minor) {
   return major < 7 || major === 7 && minor < 5;
@@ -261087,7 +261529,9 @@ function backendImportCheck(backend) {
   return "import torch, diffusers, scipy\nfrom diffusers import AudioLDMPipeline\n";
 }
 function audioGenerationPythonEnv(_repoRoot) {
-  return unifiedPythonEnv();
+  const env2 = unifiedPythonEnv();
+  applyMediaCudaDeviceFilterToEnv(env2, "audio");
+  return env2;
 }
 function approxAudioDownloadBytes(preset) {
   if (preset?.approxDownloadGB)
@@ -261341,8 +261785,8 @@ function formatAudioSetupFailure(backend, text) {
   if (lowered.includes("cuda") && lowered.includes("not available")) {
     notes2.push("CUDA was not available to the selected Python environment; install a Torch build matching this machine's CUDA runtime or use CPU-compatible settings.");
   }
-  if (lowered.includes("cudnn version") && lowered.includes("sm < 7.5")) {
-    notes2.push("The installed PyTorch wheel uses cuDNN 9 on a legacy CUDA GPU. Omnius now repairs audio-generation venvs by reinstalling PyTorch 2.3.1 from the cu118 index for SM < 7.5 hardware.");
+  if (lowered.includes("cudnn") && lowered.includes("incompatible") || lowered.includes("sm < 7.5") || lowered.includes("not compatible") && (lowered.includes("sm_") || lowered.includes("compute capability"))) {
+    notes2.push(`The installed PyTorch wheel is touching a legacy CUDA GPU. Audio generation auto-filters CUDA devices below SM 7.5; current resolved CUDA_VISIBLE_DEVICES=${resolveAudioCudaVisibleDevicesForEnv() ?? "unset"}. Override with OMNIUS_AUDIO_CUDA_VISIBLE_DEVICES or disable with OMNIUS_AUDIO_DISABLE_CUDA_FILTER=1.`);
   }
   return [body, ...notes2.map((note) => `
 ${note}`)].filter(Boolean).join("");
@@ -261564,6 +262008,18 @@ function audioCandidateFor(kind, model, requestedBackend) {
     preset: getAudioGenerationPreset(model, kind)
   };
 }
+function audioCandidateHost(candidate) {
+  if (candidate.backend === "project")
+    return null;
+  if (candidate.backend === "audiocraft")
+    return "audiocraft";
+  if (candidate.backend === "tangoflux")
+    return "tangoflux";
+  return "diffusers-py";
+}
+function audioCandidateEstimatedVramMB(candidate) {
+  return candidate.preset ? Math.ceil(candidate.preset.minVramGB * 1024) : void 0;
+}
 function audioGenerationFallbackCandidates(kind, requestedModel, requestedBackend, allowFallback = true) {
   const ladder = audioGenerationQualityLadder(kind);
   const candidates = [];
@@ -261720,9 +262176,15 @@ function annotateAudioFallbackSuccess(result, failed, winner) {
     ...failed.map((attempt, index) => `  ${formatAudioAttempt(attempt.candidate, attempt.reason, index)}`),
     ""
   ].join("\n");
+  const llmPrefix = [
+    `Fallback ladder used ${winner.model} [${winner.backend}] after ${failed.length} failed attempt(s).`,
+    ...failed.map((attempt, index) => formatAudioAttempt(attempt.candidate, attempt.reason, index))
+  ].join("\n");
   return {
     ...result,
-    output: prefix + result.output
+    output: prefix + result.output,
+    llmContent: result.llmContent ? `${llmPrefix}
+${result.llmContent}` : result.llmContent
   };
 }
 var DEFAULT_SOUND_MODEL, DEFAULT_MUSIC_MODEL, DIFFUSERS_AUDIO_PACKAGES, TRANSFORMERS_AUDIO_PACKAGES, AUDIOCRAFT_PACKAGES, STABLE_AUDIO_PACKAGES, TANGOFLUX_PACKAGES, AUDIO_GENERATION_MODEL_PRESETS, SOUND_GENERATION_QUALITY_LADDER, MUSIC_GENERATION_QUALITY_LADDER, DIFFUSERS_AUDIO_RUNNER, AUDIOCRAFT_RUNNER, TRANSFORMERS_AUDIO_RUNNER, TANGOFLUX_RUNNER, AudioGenerateTool;
@@ -261733,6 +262195,7 @@ var init_audio_generate = __esm({
     init_venv_paths();
     init_model_store();
     init_hf_media_models();
+    init_cuda_device_filter();
     DEFAULT_SOUND_MODEL = "cvssp/audioldm-s-full-v2";
     DEFAULT_MUSIC_MODEL = "facebook/musicgen-small";
     DIFFUSERS_AUDIO_PACKAGES = [
@@ -262094,9 +262557,9 @@ var init_audio_generate = __esm({
 import argparse, json, os, sys, time
 from pathlib import Path
 
-# Broker-picked GPU pinning — must run before importing torch.
-_omnius_gpu = os.environ.get("OMNIUS_GPU_INDEX", "").strip()
-if _omnius_gpu and "CUDA_VISIBLE_DEVICES" not in os.environ:
+# Broker/audio GPU pinning — must run before importing torch.
+_omnius_gpu = os.environ.get("OMNIUS_GPU_INDEX", "").strip() or os.environ.get("OMNIUS_AUDIO_GPU", "").strip()
+if _omnius_gpu:
     os.environ["CUDA_VISIBLE_DEVICES"] = _omnius_gpu
 
 def _format_bytes(value):
@@ -262239,9 +262702,14 @@ if __name__ == "__main__":
     main()
 `;
     AUDIOCRAFT_RUNNER = String.raw`#!/usr/bin/env python3
-import argparse, json, sys, time
+import argparse, json, os, sys, time
 from pathlib import Path
 
+# Broker/audio GPU pinning — must run before importing torch/audiocraft.
+_omnius_gpu = os.environ.get("OMNIUS_GPU_INDEX", "").strip() or os.environ.get("OMNIUS_AUDIO_GPU", "").strip()
+if _omnius_gpu:
+    os.environ["CUDA_VISIBLE_DEVICES"] = _omnius_gpu
+
 def _progress(stage, message, percent=None):
     payload = {"omnius_progress": True, "stage": stage, "message": message}
     if percent is not None:
@@ -262295,9 +262763,9 @@ if __name__ == "__main__":
 import argparse, json, os, sys, time
 from pathlib import Path
 
-# Broker-picked GPU pinning — must run before importing torch.
-_omnius_gpu = os.environ.get("OMNIUS_GPU_INDEX", "").strip()
-if _omnius_gpu and "CUDA_VISIBLE_DEVICES" not in os.environ:
+# Broker/audio GPU pinning — must run before importing torch.
+_omnius_gpu = os.environ.get("OMNIUS_GPU_INDEX", "").strip() or os.environ.get("OMNIUS_AUDIO_GPU", "").strip()
+if _omnius_gpu:
     os.environ["CUDA_VISIBLE_DEVICES"] = _omnius_gpu
 
 def _format_bytes(value):
@@ -262411,9 +262879,14 @@ if __name__ == "__main__":
     main()
 `;
     TANGOFLUX_RUNNER = String.raw`#!/usr/bin/env python3
-import argparse, json, sys, time
+import argparse, json, os, sys, time
 from pathlib import Path
 
+# Broker/audio GPU pinning — must run before importing torch/tangoflux.
+_omnius_gpu = os.environ.get("OMNIUS_GPU_INDEX", "").strip() or os.environ.get("OMNIUS_AUDIO_GPU", "").strip()
+if _omnius_gpu:
+    os.environ["CUDA_VISIBLE_DEVICES"] = _omnius_gpu
+
 def _format_bytes(value):
     try:
         n = float(value)
@@ -262548,6 +263021,48 @@ if __name__ == "__main__":
         this.lastProgressAt = now;
         this.progressHandler(event);
       }
+      async acquireTransientLoadLease(args) {
+        if (!args.candidate)
+          return null;
+        const host = audioCandidateHost(args.candidate);
+        if (!host)
+          return null;
+        const broker = getModelBroker();
+        const decision2 = await broker.acquireTransientModelLoad({
+          name: args.candidate.model,
+          domain: args.kind,
+          host,
+          owner: `audio-generate-tool/${args.kind}`,
+          estimatedVramMB: audioCandidateEstimatedVramMB(args.candidate)
+        }, {
+          reason: args.reason,
+          restoreOllama: true,
+          unloadRequestedOllama: false
+        });
+        if (decision2.kind === "reject") {
+          return {
+            success: false,
+            output: "",
+            error: `${args.kind === "music" ? "Music" : "Sound"} generation blocked by resource broker: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.kind === "degrade") {
+          return {
+            success: false,
+            output: "",
+            error: `${args.kind === "music" ? "Music" : "Sound"} generation needs a broker fallback (${decision2.fallback.name}), but audio candidate fallback must be selected by the audio ladder: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.lease.evictedOllamaModels.length > 0) {
+          this.emitProgress({
+            stage: "setup",
+            message: `Temporarily unloaded ${decision2.lease.evictedOllamaModels.length} Ollama inference model(s) to free VRAM for ${args.kind} generation`
+          });
+        }
+        return { lease: decision2.lease, gpuIndex: decision2.lease.gpuIndex };
+      }
       async prewarmPythonBackend(args) {
         const runner = await ensureAudioRunner(this.cwd, args.runnerBackend);
         let python;
@@ -262594,6 +263109,17 @@ if __name__ == "__main__":
         }
         ensureUnifiedCacheDirs();
         this.emitProgress({ stage: "load", message: `Downloading/loading ${args.kind} model ${args.model}` });
+        const runnerEnv = { ...python.env };
+        if (this._brokerGpuIndex !== null) {
+          if (audioBrokerGpuIndexIsCompatible(this._brokerGpuIndex, runnerEnv)) {
+            runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          } else {
+            this.emitProgress({
+              stage: "setup",
+              message: `Broker selected CUDA GPU ${this._brokerGpuIndex}, but audio CUDA filtering excluded it; using CUDA_VISIBLE_DEVICES=${runnerEnv["CUDA_VISIBLE_DEVICES"] ?? "default"}`
+            });
+          }
+        }
         const result = await runProcess3(python.command, [
           runner,
           "--kind",
@@ -262610,7 +263136,7 @@ if __name__ == "__main__":
         ], {
           cwd: this.cwd,
           timeoutMs: 18e5,
-          env: python.env,
+          env: runnerEnv,
           progressLabel: `Downloading/loading ${args.model}`,
           onProgress: (event) => this.emitProgress(event)
         });
@@ -262690,33 +263216,6 @@ if __name__ == "__main__":
         const candidates = audioGenerationFallbackCandidates(kind, requestedModel, requestedBackend, generationFallbackEnabled2(args));
         const seed = optionalNumberArg2(args["seed"]);
         const playback = playbackRequested(args);
-        const broker = getModelBroker();
-        const firstCandidate = candidates[0];
-        let brokerGpuIndex = null;
-        if (firstCandidate) {
-          const decision2 = await broker.ensureModelLoadable({
-            name: firstCandidate.model,
-            domain: kind === "music" ? "music" : "sound",
-            host: firstCandidate.backend === "audiocraft" ? "audiocraft" : firstCandidate.backend === "tangoflux" ? "tangoflux" : firstCandidate.backend === "transformers" ? "diffusers-py" : "diffusers-py",
-            owner: `audio-generate-tool/${kind}`
-          });
-          if (decision2.kind === "evict") {
-            for (const target of decision2.evictTargets) {
-              await broker.evict(target.host, target.name, `${kind}-gen-needs-room`);
-            }
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "ok") {
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "reject") {
-            return {
-              success: false,
-              output: "",
-              error: `${kind === "music" ? "Music" : "Sound"} generation blocked by resource broker: ${decision2.reason}`,
-              durationMs: performance.now() - start2
-            };
-          }
-        }
-        this._brokerGpuIndex = brokerGpuIndex;
         try {
           return await this.generateCandidateLadder({ kind, candidates, prompt, args, seed, playback, start: start2 });
         } catch (err) {
@@ -262736,15 +263235,39 @@ if __name__ == "__main__":
             stage: "setup",
             message: `Preparing ${args.kind} model ${candidate.model} (${candidate.backend}) [${index + 1}/${args.candidates.length}]`
           });
-          const result = candidate.backend === "project" ? this.projectProfileResult(args.kind, candidate, args.start) : await this.prewarmPythonBackend({
+          const leaseDecision = await this.acquireTransientLoadLease({
             kind: args.kind,
-            backend: candidate.backend,
-            runnerBackend: candidate.backend,
-            model: candidate.model,
-            duration,
-            start: args.start,
-            python: args.args["python"]
-          });
+            candidate,
+            reason: `${args.kind}-prewarm-needs-room`,
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult2(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; trying ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = candidate.backend === "project" ? this.projectProfileResult(args.kind, candidate, args.start) : await this.prewarmPythonBackend({
+              kind: args.kind,
+              backend: candidate.backend,
+              runnerBackend: candidate.backend,
+              model: candidate.model,
+              duration,
+              start: args.start,
+              python: args.args["python"]
+            });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success)
             return annotateAudioFallbackSuccess(result, failed, candidate);
           failed.push({ candidate, reason: summarizeToolResult2(result) });
@@ -262772,19 +263295,43 @@ if __name__ == "__main__":
             stage: "setup",
             message: `Using ${args.kind} model ${candidate.model} (${candidate.backend}) [${index + 1}/${args.candidates.length}]`
           });
-          const result = candidate.backend === "project" ? this.projectProfileResult(args.kind, candidate, args.start) : await this.generateWithPythonBackend({
+          const leaseDecision = await this.acquireTransientLoadLease({
             kind: args.kind,
-            backend: candidate.backend,
-            runnerBackend: candidate.backend,
-            prompt: args.prompt,
-            model: candidate.model,
-            duration,
-            steps,
-            seed: args.seed,
-            playback: args.playback,
-            start: args.start,
-            python: args.args["python"]
-          });
+            candidate,
+            reason: `${args.kind}-gen-needs-room`,
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult2(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; falling back to ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = candidate.backend === "project" ? this.projectProfileResult(args.kind, candidate, args.start) : await this.generateWithPythonBackend({
+              kind: args.kind,
+              backend: candidate.backend,
+              runnerBackend: candidate.backend,
+              prompt: args.prompt,
+              model: candidate.model,
+              duration,
+              steps,
+              seed: args.seed,
+              playback: args.playback,
+              start: args.start,
+              python: args.args["python"]
+            });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success)
             return annotateAudioFallbackSuccess(result, failed, candidate);
           failed.push({ candidate, reason: summarizeToolResult2(result) });
@@ -262885,7 +263432,14 @@ if __name__ == "__main__":
         this.emitProgress({ stage: "load", message: `Starting ${args.kind} generation with ${args.model}` });
         const runnerEnv = { ...python.env };
         if (this._brokerGpuIndex !== null) {
-          runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          if (audioBrokerGpuIndexIsCompatible(this._brokerGpuIndex, runnerEnv)) {
+            runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          } else {
+            this.emitProgress({
+              stage: "setup",
+              message: `Broker selected CUDA GPU ${this._brokerGpuIndex}, but audio CUDA filtering excluded it; using CUDA_VISIBLE_DEVICES=${runnerEnv["CUDA_VISIBLE_DEVICES"] ?? "default"}`
+            });
+          }
         }
         const result = await runProcess3(python.command, argv, {
           cwd: this.cwd,
@@ -263117,6 +263671,12 @@ function videoCandidateFor(model, requestedBackend, requestedKind) {
   }
   return { model, backend, preset };
 }
+function videoCandidateHost(candidate) {
+  return candidate.backend === "comfyui" ? "comfyui" : "diffusers-py";
+}
+function videoCandidateEstimatedVramMB(candidate) {
+  return candidate.preset ? Math.ceil(candidate.preset.minVramGB * 1024) : void 0;
+}
 function videoGenerationFallbackCandidates(requestedModel, requestedBackend, requestedKind, allowFallback = true, options2 = {}) {
   const preferAudioVideo = Boolean(options2.preferNativeAudioVideo);
   const baseLadderIds = preferAudioVideo ? [...VIDEO_AUDIO_QUALITY_LADDER, ...VIDEO_GENERATION_QUALITY_LADDER] : VIDEO_GENERATION_QUALITY_LADDER;
@@ -263359,6 +263919,7 @@ function resolveHfToken() {
 }
 function videoGenerationPythonEnv(_repoRoot) {
   const env2 = unifiedPythonEnv();
+  applyMediaCudaDeviceFilterToEnv(env2, "video");
   const token = resolveHfToken();
   if (token) {
     env2["HF_TOKEN"] = token;
@@ -263681,9 +264242,15 @@ function annotateVideoFallbackSuccess(result, failed, winner) {
     ...failed.map((attempt, index) => `  ${formatVideoAttempt(attempt.candidate, attempt.reason, index)}`),
     ""
   ].join("\n");
+  const llmPrefix = [
+    `Fallback ladder used ${winner.model} [${winner.backend}] after ${failed.length} failed attempt(s).`,
+    ...failed.map((attempt, index) => formatVideoAttempt(attempt.candidate, attempt.reason, index))
+  ].join("\n");
   return {
     ...result,
-    output: prefix + result.output
+    output: prefix + result.output,
+    llmContent: result.llmContent ? `${llmPrefix}
+${result.llmContent}` : result.llmContent
   };
 }
 function parseRunnerJson3(stdout) {
@@ -263717,6 +264284,7 @@ var init_video_generate = __esm({
     init_venv_paths();
     init_model_store();
     init_hf_media_models();
+    init_cuda_device_filter();
     DEFAULT_DIFFUSERS_VIDEO_MODEL = "Efficient-Large-Model/SANA-Video_2B_480p";
     SANA_VIDEO_480P_MODEL = "Efficient-Large-Model/SANA-Video_2B_480p";
     SANA_VIDEO_720P_MODEL = "Efficient-Large-Model/SANA-Video_2B_720p";
@@ -265049,6 +265617,45 @@ if __name__ == "__main__":
         this.lastProgressAt = now;
         this.progressHandler(event);
       }
+      async acquireTransientLoadLease(args) {
+        if (!args.candidate)
+          return null;
+        const broker = getModelBroker();
+        const decision2 = await broker.acquireTransientModelLoad({
+          name: args.candidate.model,
+          domain: "video-gen",
+          host: videoCandidateHost(args.candidate),
+          owner: "video-generate-tool",
+          estimatedVramMB: videoCandidateEstimatedVramMB(args.candidate)
+        }, {
+          reason: args.reason,
+          restoreOllama: true,
+          unloadRequestedOllama: false
+        });
+        if (decision2.kind === "reject") {
+          return {
+            success: false,
+            output: "",
+            error: `Video generation blocked by resource broker: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.kind === "degrade") {
+          return {
+            success: false,
+            output: "",
+            error: `Video generation needs a broker fallback (${decision2.fallback.name}), but video candidate fallback must be selected by the video ladder: ${decision2.reason}`,
+            durationMs: performance.now() - args.start
+          };
+        }
+        if (decision2.lease.evictedOllamaModels.length > 0) {
+          this.emitProgress({
+            stage: "setup",
+            message: `Temporarily unloaded ${decision2.lease.evictedOllamaModels.length} Ollama inference model(s) to free VRAM for video generation`
+          });
+        }
+        return { lease: decision2.lease, gpuIndex: decision2.lease.gpuIndex };
+      }
       async execute(args) {
         const start2 = performance.now();
         const action = String(args["action"] ?? "generate");
@@ -265104,35 +265711,6 @@ if __name__ == "__main__":
         const withAudio = booleanArg3(args["with_audio"], false);
         const audioInput = typeof args["audio_input"] === "string" && args["audio_input"].trim() ? String(args["audio_input"]).trim() : void 0;
         const candidates = videoGenerationFallbackCandidates(requestedModel, requestedBackend, inferredKind, generationFallbackEnabled3(args), { preferNativeAudioVideo: withAudio || Boolean(audioInput) });
-        const broker = getModelBroker();
-        const firstCandidate = candidates[0];
-        let brokerGpuIndex = null;
-        if (firstCandidate) {
-          const preset = firstCandidate.preset;
-          const decision2 = await broker.ensureModelLoadable({
-            name: firstCandidate.model,
-            domain: "video-gen",
-            host: firstCandidate.backend === "comfyui" ? "comfyui" : "diffusers-py",
-            owner: "video-generate-tool",
-            estimatedVramMB: preset ? preset.minVramGB * 1024 : void 0
-          });
-          if (decision2.kind === "evict") {
-            for (const target of decision2.evictTargets) {
-              await broker.evict(target.host, target.name, "video-gen-needs-room");
-            }
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "ok") {
-            brokerGpuIndex = decision2.gpuIndex ?? null;
-          } else if (decision2.kind === "reject") {
-            return {
-              success: false,
-              output: "",
-              error: `Video generation blocked by resource broker: ${decision2.reason}`,
-              durationMs: performance.now() - start2
-            };
-          }
-        }
-        this._brokerGpuIndex = brokerGpuIndex;
         if (candidates.length === 0) {
           return {
             success: false,
@@ -265182,7 +265760,30 @@ if __name__ == "__main__":
             failed.push({ candidate, reason: "ComfyUI backend not yet implemented." });
             continue;
           }
-          const result = await this.prewarmDiffusers({ candidate, start: args.start, python: args.args["python"] });
+          const leaseDecision = await this.acquireTransientLoadLease({
+            candidate,
+            reason: "video-prewarm-needs-room",
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult3(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; trying ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = await this.prewarmDiffusers({ candidate, start: args.start, python: args.args["python"] });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success)
             return annotateVideoFallbackSuccess(result, failed, candidate);
           failed.push({ candidate, reason: summarizeToolResult3(result) });
@@ -265268,26 +265869,48 @@ if __name__ == "__main__":
               start: args.start
             });
           } else {
-            result = await this.generateWithDiffusers({
-              prompt: promptForCandidate,
-              model: candidate.model,
-              preset,
-              kind: args.kind,
-              imageArg: args.imageArg,
-              audioInput: args.audioInput,
-              width,
-              height,
-              numFrames,
-              fps,
-              steps,
-              guidance,
-              negativePrompt,
-              seed: args.seed,
-              hfToken: hfTokenOverride,
-              autoAcceptLicense,
-              start: args.start,
-              python: args.args["python"]
+            const leaseDecision = await this.acquireTransientLoadLease({
+              candidate,
+              reason: "video-gen-needs-room",
+              start: args.start
             });
+            if (leaseDecision && "success" in leaseDecision) {
+              failed.push({ candidate, reason: summarizeToolResult3(leaseDecision) });
+              if (index < args.candidates.length - 1) {
+                this.emitProgress({
+                  stage: "setup",
+                  message: `${candidate.model} did not fit current resources; falling back to ${args.candidates[index + 1].model}`
+                });
+              }
+              continue;
+            }
+            const lease = leaseDecision?.lease;
+            this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+            try {
+              result = await this.generateWithDiffusers({
+                prompt: promptForCandidate,
+                model: candidate.model,
+                preset,
+                kind: args.kind,
+                imageArg: args.imageArg,
+                audioInput: args.audioInput,
+                width,
+                height,
+                numFrames,
+                fps,
+                steps,
+                guidance,
+                negativePrompt,
+                seed: args.seed,
+                hfToken: hfTokenOverride,
+                autoAcceptLicense,
+                start: args.start,
+                python: args.args["python"]
+              });
+            } finally {
+              await lease?.release();
+              this._brokerGpuIndex = null;
+            }
           }
           let nativeAudio = preset.nativeAudioVideo === true;
           let audioPath;
@@ -265479,6 +266102,17 @@ ${llmAnnotation}` : result.llmContent;
         }
         ensureUnifiedCacheDirs();
         this.emitProgress({ stage: "load", message: `Downloading/loading video model ${args.candidate.model}` });
+        const runnerEnv = { ...python.env };
+        if (this._brokerGpuIndex !== null) {
+          if (mediaBrokerGpuIndexIsCompatible(this._brokerGpuIndex, "video", runnerEnv)) {
+            runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          } else {
+            this.emitProgress({
+              stage: "setup",
+              message: `Broker selected CUDA GPU ${this._brokerGpuIndex}, but video CUDA filtering excluded it; using CUDA_VISIBLE_DEVICES=${runnerEnv["CUDA_VISIBLE_DEVICES"] ?? "default"}`
+            });
+          }
+        }
         const result = await runProcess4(python.command, [
           runner,
           "--model",
@@ -265494,7 +266128,7 @@ ${llmAnnotation}` : result.llmContent;
         ], {
           cwd: this.cwd,
           timeoutMs: 18e5,
-          env: python.env,
+          env: runnerEnv,
           progressLabel: `Downloading/loading ${args.candidate.model}`,
           onProgress: (event) => this.emitProgress(event)
         });
@@ -265555,7 +266189,14 @@ ${llmAnnotation}` : result.llmContent;
           runnerEnv["HUGGING_FACE_HUB_TOKEN"] = effectiveToken;
         }
         if (this._brokerGpuIndex !== null) {
-          runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          if (mediaBrokerGpuIndexIsCompatible(this._brokerGpuIndex, "video", runnerEnv)) {
+            runnerEnv["OMNIUS_GPU_INDEX"] = String(this._brokerGpuIndex);
+          } else {
+            this.emitProgress({
+              stage: "setup",
+              message: `Broker selected CUDA GPU ${this._brokerGpuIndex}, but video CUDA filtering excluded it; using CUDA_VISIBLE_DEVICES=${runnerEnv["CUDA_VISIBLE_DEVICES"] ?? "default"}`
+            });
+          }
         }
         const argv = [
           runner,
@@ -266682,6 +267323,11 @@ import { readFileSync as readFileSync24, existsSync as existsSync34, statSync as
 import { execSync as execSync16, spawn as spawn10, spawnSync as spawnSync4 } from "node:child_process";
 import { resolve as resolve23, extname as extname6, basename as basename8, dirname as dirname9, join as join44 } from "node:path";
 import { fileURLToPath as fileURLToPath4 } from "node:url";
+function visionPythonEnv(extra = {}) {
+  const env2 = { ...process.env, ...extra };
+  applyMediaCudaDeviceFilterToEnv(env2, "vision");
+  return env2;
+}
 async function probeStation(endpoint) {
   try {
     const healthUrl = endpoint.replace(/\/v1\/?$/, "/health");
@@ -266752,7 +267398,8 @@ async function autoLaunchStation(port = 2020) {
     return false;
   return new Promise((resolvePromise) => {
     const child = spawn10(pythonBin, [launcherScript, "--port", String(port)], {
-      stdio: ["ignore", "pipe", "pipe"]
+      stdio: ["ignore", "pipe", "pipe"],
+      env: visionPythonEnv()
     });
     stationProcess = child;
     const cleanupStation = () => {
@@ -267067,7 +267714,11 @@ function tryHuggingFacePointBackend(options2) {
     hfPointUnavailable = "Python not found";
     return null;
   }
-  const deps = spawnSync4(python, ["-c", "import torch, transformers, PIL"], { stdio: "pipe", timeout: 1e4 });
+  const deps = spawnSync4(python, ["-c", "import torch, transformers, PIL"], {
+    stdio: "pipe",
+    timeout: 1e4,
+    env: visionPythonEnv()
+  });
   if (deps.status !== 0) {
     hfPointUnavailable = bufferishToString3(deps.stderr) || "Python dependencies torch, transformers, and pillow are not importable";
     return null;
@@ -267114,7 +267765,7 @@ print(json.dumps(last_result))
       encoding: "utf8",
       stdio: ["pipe", "pipe", "pipe"],
       timeout: Math.max(options2.timeoutMs ?? 6e4, 3e5),
-      env: { ...process.env }
+      env: visionPythonEnv()
     });
     if (run2.status !== 0) {
       hfPointUnavailable = run2.stderr || run2.stdout || "Hugging Face Moondream point backend failed";
@@ -267167,6 +267818,7 @@ var init_vision = __esm({
   "packages/execution/dist/tools/vision.js"() {
     "use strict";
     init_model_broker();
+    init_cuda_device_filter();
     moondreamClient = null;
     moondreamError = null;
     stationProcess = null;
@@ -520577,6 +521229,11 @@ import { execFileSync as execFileSync5, execSync as execSync30, spawn as spawn15
 import { copyFileSync as copyFileSync3, existsSync as existsSync47, statSync as statSync23, writeFileSync as writeFileSync19, mkdirSync as mkdirSync20, readdirSync as readdirSync18, writeSync } from "node:fs";
 import { basename as basename15, extname as extname10, isAbsolute as isAbsolute2, join as join63 } from "node:path";
 import { homedir as homedir16, tmpdir as tmpdir11 } from "node:os";
+function ttsPythonEnv(extra = {}) {
+  const env2 = { ...process.env, ...extra };
+  applyMediaCudaDeviceFilterToEnv(env2, "tts");
+  return env2;
+}
 function hasCommand3(command) {
   try {
     if (process.platform === "win32") {
@@ -521144,7 +521801,7 @@ function ensureLuxttsDaemon() {
     const daemon = spawn15(venvPy, [inferScript], {
       stdio: ["pipe", "pipe", "pipe"],
       cwd: tmpdir11(),
-      env: { ...process.env, LUXTTS_REPO_PATH: repoDir }
+      env: ttsPythonEnv({ LUXTTS_REPO_PATH: repoDir })
     });
     _luxttsDaemon = daemon;
     _luxttsBuffer = "";
@@ -521224,6 +521881,7 @@ var init_audio_playback = __esm({
   "packages/execution/dist/tools/audio-playback.js"() {
     "use strict";
     init_hf_media_models();
+    init_cuda_device_filter();
     _luxttsDaemon = null;
     _luxttsReady = false;
     _luxttsRequestId = 0;
@@ -521605,7 +522263,7 @@ ${tried.map((line) => `- ${line}`).join("\n")}`,
         execFileSync5(venvPy, ["-c", pyScript, JSON.stringify({ text, output: outputPath3, clone_ref: cloneRef, repo: repoDir, speed })], {
           stdio: "pipe",
           timeout: 12e4,
-          env: { ...process.env, LUXTTS_REPO_PATH: repoDir }
+          env: ttsPythonEnv({ LUXTTS_REPO_PATH: repoDir })
         });
         return `${basename15(cloneRef)} (LuxTTS standalone)`;
       }
@@ -521619,7 +522277,8 @@ ${tried.map((line) => `- ${line}`).join("\n")}`,
           input: JSON.stringify({ text, output_path: outputPath3, voice_name: voice, lang, speed, total_step: totalStep }),
           encoding: "utf8",
           stdio: ["pipe", "pipe", "pipe"],
-          timeout: 18e4
+          timeout: 18e4,
+          env: ttsPythonEnv()
         });
         const line = stdout.trim().split(/\r?\n/).pop() || "";
         const parsed = JSON.parse(line);
@@ -523041,10 +523700,16 @@ import { execSync as execSync36 } from "node:child_process";
 import { existsSync as existsSync50, mkdirSync as mkdirSync22, writeFileSync as writeFileSync20, readFileSync as readFileSync35 } from "node:fs";
 import { join as join65 } from "node:path";
 import { homedir as homedir17, tmpdir as tmpdir13 } from "node:os";
+function audioAnalysisPythonEnv(extra = {}) {
+  const env2 = { ...process.env, ...extra };
+  applyMediaCudaDeviceFilterToEnv(env2, "asr");
+  return env2;
+}
 var VENV_DIR, VENV_PIP, VENV_PYTHON, AudioAnalyzeTool;
 var init_audio_analyze = __esm({
   "packages/execution/dist/tools/audio-analyze.js"() {
     "use strict";
+    init_cuda_device_filter();
     VENV_DIR = join65(homedir17(), ".omnius", "audio-ml-venv");
     VENV_PIP = join65(VENV_DIR, "bin", "pip");
     VENV_PYTHON = join65(VENV_DIR, "bin", "python3");
@@ -523334,15 +523999,15 @@ Context saved to: ${contextFile}`,
       /** Ensure Python venv with required packages */
       async ensureVenv(packages) {
         if (!existsSync50(VENV_PYTHON)) {
-          execSync36(`python3 -m venv ${VENV_DIR}`, { timeout: 3e4, stdio: "pipe" });
+          execSync36(`python3 -m venv ${VENV_DIR}`, { timeout: 3e4, stdio: "pipe", env: audioAnalysisPythonEnv() });
         }
         for (const pkg of packages) {
           const importName = pkg.replace(/[<>=!].*/g, "").replace(/-/g, "_");
           try {
-            execSync36(`${VENV_PYTHON} -c "import ${importName}"`, { timeout: 1e4, stdio: "pipe" });
+            execSync36(`${VENV_PYTHON} -c "import ${importName}"`, { timeout: 1e4, stdio: "pipe", env: audioAnalysisPythonEnv() });
           } catch {
             try {
-              execSync36(`${VENV_PIP} install "${pkg}"`, { timeout: 3e5, stdio: "pipe" });
+              execSync36(`${VENV_PIP} install "${pkg}"`, { timeout: 3e5, stdio: "pipe", env: audioAnalysisPythonEnv() });
             } catch {
             }
           }
@@ -523356,7 +524021,7 @@ Context saved to: ${contextFile}`,
           const output = execSync36(`${VENV_PYTHON} ${scriptFile}`, {
             encoding: "utf8",
             timeout: 3e5,
-            env: { ...process.env, TF_CPP_MIN_LOG_LEVEL: "3", TF_ENABLE_ONEDNN_OPTS: "0", PYTHONUNBUFFERED: "1" }
+            env: audioAnalysisPythonEnv({ TF_CPP_MIN_LOG_LEVEL: "3", TF_ENABLE_ONEDNN_OPTS: "0", PYTHONUNBUFFERED: "1" })
           });
           try {
             const result = JSON.parse(output.trim().split("\n").pop());
@@ -524192,10 +524857,16 @@ import { execSync as execSync38 } from "node:child_process";
 import { existsSync as existsSync52, mkdirSync as mkdirSync24, writeFileSync as writeFileSync22, readFileSync as readFileSync37 } from "node:fs";
 import { join as join67 } from "node:path";
 import { homedir as homedir19, tmpdir as tmpdir15 } from "node:os";
+function visualMemoryPythonEnv(extra = {}) {
+  const env2 = { ...process.env, ...extra };
+  applyMediaCudaDeviceFilterToEnv(env2, "vision");
+  return env2;
+}
 var VMEM_DIR, VENV_DIR2, VENV_PY, VENV_PIP2, VisualMemoryTool;
 var init_visual_memory = __esm({
   "packages/execution/dist/tools/visual-memory.js"() {
     "use strict";
+    init_cuda_device_filter();
     VMEM_DIR = join67(homedir19(), ".omnius", "visual-memory");
     VENV_DIR2 = join67(homedir19(), ".omnius", "vision-ml-venv");
     VENV_PY = join67(VENV_DIR2, "bin", "python3");
@@ -524723,19 +525394,23 @@ ${objects.join("\n") || "  (none taught)"}`,
       async ensureVenv() {
         if (existsSync52(VENV_PY)) {
           try {
-            execSync38(`${VENV_PY} -c "import insightface, transformers, torch"`, { timeout: 15e3, stdio: "pipe" });
+            execSync38(`${VENV_PY} -c "import insightface, transformers, torch"`, {
+              timeout: 15e3,
+              stdio: "pipe",
+              env: visualMemoryPythonEnv()
+            });
             return true;
           } catch {
           }
         }
         try {
           if (!existsSync52(VENV_PY)) {
-            execSync38(`python3 -m venv ${VENV_DIR2}`, { timeout: 3e4, stdio: "pipe" });
+            execSync38(`python3 -m venv ${VENV_DIR2}`, { timeout: 3e4, stdio: "pipe", env: visualMemoryPythonEnv() });
           }
-          execSync38(`${VENV_PIP2} install "setuptools<81" wheel`, { timeout: 6e4, stdio: "pipe" });
-          execSync38(`${VENV_PIP2} install torch torchvision`, { timeout: 6e5, stdio: "pipe" });
-          execSync38(`${VENV_PIP2} install insightface onnxruntime opencv-python-headless`, { timeout: 3e5, stdio: "pipe" });
-          execSync38(`${VENV_PIP2} install transformers pillow`, { timeout: 3e5, stdio: "pipe" });
+          execSync38(`${VENV_PIP2} install "setuptools<81" wheel`, { timeout: 6e4, stdio: "pipe", env: visualMemoryPythonEnv() });
+          execSync38(`${VENV_PIP2} install torch torchvision`, { timeout: 6e5, stdio: "pipe", env: visualMemoryPythonEnv() });
+          execSync38(`${VENV_PIP2} install insightface onnxruntime opencv-python-headless`, { timeout: 3e5, stdio: "pipe", env: visualMemoryPythonEnv() });
+          execSync38(`${VENV_PIP2} install transformers pillow`, { timeout: 3e5, stdio: "pipe", env: visualMemoryPythonEnv() });
           return true;
         } catch {
           return false;
@@ -524748,7 +525423,7 @@ ${objects.join("\n") || "  (none taught)"}`,
           const output = execSync38(`${VENV_PY} ${scriptFile}`, {
             encoding: "utf8",
             timeout: timeoutMs,
-            env: { ...process.env, PYTHONUNBUFFERED: "1" }
+            env: visualMemoryPythonEnv({ PYTHONUNBUFFERED: "1" })
           });
           const lastLine = output.trim().split("\n").pop() || "{}";
           return JSON.parse(lastLine);
@@ -525364,6 +526039,11 @@ import { existsSync as existsSync54, mkdirSync as mkdirSync26, writeFileSync as
 import { dirname as dirname15, join as join69, resolve as resolve35 } from "node:path";
 import { tmpdir as tmpdir17, homedir as homedir21 } from "node:os";
 import { fileURLToPath as fileURLToPath8 } from "node:url";
+function asrPythonEnv(extra = {}) {
+  const env2 = { ...process.env, ...extra };
+  applyMediaCudaDeviceFilterToEnv(env2, "asr");
+  return env2;
+}
 function _findNemotronScript() {
   const candidates = [];
   try {
@@ -525399,6 +526079,7 @@ var init_asr_listen = __esm({
   "packages/execution/dist/tools/asr-listen.js"() {
     "use strict";
     init_hf_media_models();
+    init_cuda_device_filter();
     AsrListenTool = class {
       name = "asr_listen";
       description = "Record from microphone and transcribe speech to text. Backends: 'whisper' (default, battle-tested openai-whisper / faster-whisper), 'nemotron' (nvidia/nemotron-speech-streaming-en-0.6b — faster streaming), or 'parallel' (runs BOTH engines on the same audio and returns a side-by-side comparison with per-engine latency and character counts). Actions: 'listen' to record + transcribe in one step, 'transcribe' to run on an existing file. Use this when you need to HEAR what a human is saying — ask a question via audio_playback speak, then use asr_listen to capture and transcribe their response.";
@@ -525633,7 +526314,7 @@ print(json.dumps({"ok": False, "error": "No whisper backend available"}))
               const output = execSync40(`"${pyPath}" "${scriptFile}"`, {
                 encoding: "utf8",
                 timeout: 12e4,
-                env: { ...process.env, PYTHONUNBUFFERED: "1" }
+                env: asrPythonEnv({ PYTHONUNBUFFERED: "1" })
               }).trim();
               const lines = output.split("\n");
               for (let i2 = lines.length - 1; i2 >= 0; i2--) {
@@ -525682,7 +526363,8 @@ print(json.dumps({"ok": False, "error": "No whisper backend available"}))
           const result = spawnSync7("python3", [script, "--file", audioFile], {
             encoding: "utf8",
             timeout: 6e5,
-            stdio: ["ignore", "pipe", "pipe"]
+            stdio: ["ignore", "pipe", "pipe"],
+            env: asrPythonEnv()
           });
           if (result.error) {
             return {
@@ -530791,6 +531473,7 @@ __export(dist_exports, {
   CustomTool: () => CustomTool,
   DEFAULT_DIFFUSERS_IMAGE_MODEL: () => DEFAULT_DIFFUSERS_IMAGE_MODEL,
   DEFAULT_DIFFUSERS_VIDEO_MODEL: () => DEFAULT_DIFFUSERS_VIDEO_MODEL,
+  DEFAULT_MEDIA_MIN_CUDA_COMPUTE_CAPABILITY: () => DEFAULT_MEDIA_MIN_CUDA_COMPUTE_CAPABILITY,
   DEFAULT_MUSIC_MODEL: () => DEFAULT_MUSIC_MODEL,
   DEFAULT_OLLAMA_IMAGE_MODEL: () => DEFAULT_OLLAMA_IMAGE_MODEL,
   DEFAULT_SOUND_MODEL: () => DEFAULT_SOUND_MODEL,
@@ -530905,6 +531588,7 @@ __export(dist_exports, {
   addProjectConstraint: () => addProjectConstraint,
   addSessionConstraint: () => addSessionConstraint,
   aliasTool: () => aliasTool,
+  applyMediaCudaDeviceFilterToEnv: () => applyMediaCudaDeviceFilterToEnv,
   applyPatch: () => applyPatch,
   applyToolResultTriage: () => applyToolResultTriage,
   artifactManifestFromBytes: () => artifactManifestFromBytes,
@@ -530945,6 +531629,7 @@ __export(dist_exports, {
   defaultExtensionForMime: () => defaultExtensionForMime,
   deleteMediaModelAdapter: () => deleteMediaModelAdapter,
   deleteTodos: () => deleteTodos,
+  detectCudaDevices: () => detectCudaDevices,
   detectElevationMethod: () => detectElevationMethod,
   detectLegacyCaches: () => detectLegacyCaches,
   detectSearchProvider: () => detectSearchProvider,
@@ -531046,6 +531731,7 @@ __export(dist_exports, {
   markSessionValidated: () => markSessionValidated,
   measureRepoCacheBytes: () => measureRepoCacheBytes,
   mediaBackendCompatibleWithModality: () => mediaBackendCompatibleWithModality,
+  mediaBrokerGpuIndexIsCompatible: () => mediaBrokerGpuIndexIsCompatible,
   mediaMimeFromPath: () => mediaMimeFromPath,
   mediaModelCatalogDir: () => mediaModelCatalogDir,
   mediaModelSlug: () => mediaModelSlug,
@@ -531056,6 +531742,8 @@ __export(dist_exports, {
   normalizeSponsorMediaConfig: () => normalizeSponsorMediaConfig,
   omniusHomeDir: () => omniusHomeDir,
   packetPath: () => packetPath,
+  parseCudaComputeCapability: () => parseCudaComputeCapability,
+  parseCudaDeviceInfo: () => parseCudaDeviceInfo,
   parseMcpMarkdown: () => parseMcpMarkdown,
   parseMcpToolName: () => parseMcpToolName,
   parseSponsorMediaCapability: () => parseSponsorMediaCapability,
@@ -531078,6 +531766,7 @@ __export(dist_exports, {
   renderCustomToolDocs: () => renderCustomToolDocs,
   resetDepCache: () => resetDepCache,
   resetMoondreamClient: () => resetMoondreamClient,
+  resolveMediaCudaVisibleDevicesForEnv: () => resolveMediaCudaVisibleDevicesForEnv,
   resolveMediaModel: () => resolveMediaModel,
   resolveSecret: () => resolveSecret,
   revokeSecret: () => revokeSecret,
@@ -531199,6 +531888,7 @@ var init_dist5 = __esm({
     init_embedding_store();
     init_image_generate();
     init_audio_generate();
+    init_cuda_device_filter();
     init_model_store();
     init_video_generate();
     init_sponsor_media();
@@ -589502,6 +590192,7 @@ ${CONTENT_BG_SEQ}`);
 });
 
 // packages/cli/src/tui/tui-select.ts
+import { AsyncLocalStorage } from "node:async_hooks";
 function ansi3(code8, text) {
   return isTTY2 ? `\x1B[${code8}m${text}\x1B[0m` : text;
 }
@@ -589511,6 +590202,48 @@ function fg2563(code8, text) {
 function stripAnsi3(s2) {
   return s2.replace(/\x1B\[[0-9;]*m/g, "");
 }
+function stripTerminalControl(s2) {
+  return s2.replace(/\x1B(?:\[[\d;?]*[ -/]*[@-~]|\][^\x07\x1B]*(?:\x07|\x1B\\)?|[@-Z\\-_])/g, "");
+}
+function isNonInteractiveSelectSurface() {
+  return Boolean(nonInteractiveSelectSurface.getStore());
+}
+function runWithNonInteractiveSelectSurface(fn, opts = {}) {
+  return nonInteractiveSelectSurface.run(opts, fn);
+}
+function renderNonInteractiveSelect(opts, currentTitle, skipSet) {
+  const surface = nonInteractiveSelectSurface.getStore();
+  const maxItems = Math.max(1, surface?.maxItems ?? 30);
+  const lines = [];
+  if (currentTitle) lines.push(stripTerminalControl(stripAnsi3(currentTitle)));
+  if (lines.length) lines.push("");
+  let idx = 1;
+  let shown = 0;
+  let omitted = 0;
+  for (const item of opts.items) {
+    const isSkip = skipSet.has(item.key);
+    const labelPlain = stripTerminalControl(stripAnsi3(item.label)).trim();
+    const detailPlain = item.detail ? stripTerminalControl(stripAnsi3(item.detail)).trim() : "";
+    if (isSkip) {
+      if (labelPlain) lines.push(labelPlain);
+      continue;
+    }
+    if (shown >= maxItems) {
+      omitted++;
+      idx++;
+      continue;
+    }
+    const num = String(idx).padStart(2, " ");
+    const detail = detailPlain ? ` - ${detailPlain}` : "";
+    lines.push(`  ${num}. ${labelPlain}${detail}`);
+    shown++;
+    idx++;
+  }
+  if (omitted > 0) lines.push(`  ... ${omitted} more`);
+  if (opts.customKeyHint) lines.push("", stripTerminalControl(stripAnsi3(opts.customKeyHint)));
+  lines.push("", surface?.hint ?? "(non-interactive: menu shown as text; open the TUI for selection)");
+  process.stdout.write(lines.join("\n").trimEnd() + "\n");
+}
 function defaultRenderRow(item, focused, isActive) {
   const marker = isActive ? selectColors.green("●") : focused ? selectColors.blue("●") : selectColors.dim("○");
   const label = focused ? selectColors.blue(selectColors.bold(item.label)) : isActive ? selectColors.green(item.label) : item.label;
@@ -589535,27 +590268,8 @@ function tuiSelect(opts) {
   if (items.length === 0) {
     return Promise.resolve({ confirmed: false, key: null, index: -1 });
   }
-  if (!process.stdin.isTTY && process.env["OMNIUS_TUI_FORCE_INTERACTIVE"] !== "1") {
-    const lines = [];
-    if (currentTitle) lines.push(currentTitle);
-    if (lines.length) lines.push("");
-    let idx = 1;
-    for (const item of items) {
-      const isSkip = skipSet.has(item.key);
-      const labelPlain = stripAnsi3(item.label);
-      const detailPlain = item.detail ? stripAnsi3(item.detail) : "";
-      if (isSkip) {
-        lines.push(labelPlain);
-      } else {
-        const num = String(idx).padStart(2, " ");
-        const detail = detailPlain ? `  — ${detailPlain}` : "";
-        lines.push(`  ${num}. ${labelPlain}${detail}`);
-        idx++;
-      }
-    }
-    if (opts.customKeyHint) lines.push("", opts.customKeyHint);
-    lines.push("", "(non-interactive: list shown above; pick options by re-running this command from the TUI)");
-    process.stdout.write(lines.join("\n") + "\n");
+  if (isNonInteractiveSelectSurface() || !process.stdin.isTTY && process.env["OMNIUS_TUI_FORCE_INTERACTIVE"] !== "1") {
+    renderNonInteractiveSelect(opts, currentTitle, skipSet);
     return Promise.resolve({ confirmed: false, key: null, index: -1 });
   }
   const isSkippable = (idx) => skipSet.has(items[idx].key);
@@ -590136,7 +590850,7 @@ ${tuiBgSeq()}`);
     }
   });
 }
-var isTTY2, MENU_ACTIVE_GREEN_256, selectColors;
+var isTTY2, MENU_ACTIVE_GREEN_256, selectColors, nonInteractiveSelectSurface;
 var init_tui_select = __esm({
   "packages/cli/src/tui/tui-select.ts"() {
     "use strict";
@@ -590156,6 +590870,7 @@ var init_tui_select = __esm({
       /** Readable grey for non-matching items */
       matchDark: (t2) => fg2563(tuiTextDim(), t2)
     };
+    nonInteractiveSelectSurface = new AsyncLocalStorage();
   }
 });
 
@@ -590182,12 +590897,17 @@ import { join as join114, dirname as dirname32 } from "node:path";
 import { homedir as homedir36 } from "node:os";
 import { execSync as execSync50, spawn as spawn26 } from "node:child_process";
 import { fileURLToPath as fileURLToPath15 } from "node:url";
+function personaplexPythonEnv(extra = {}) {
+  const env2 = { ...process.env, ...extra };
+  applyMediaCudaDeviceFilterToEnv(env2, "voice");
+  return env2;
+}
 function execAsync(cmd, opts = {}) {
   return new Promise((resolve59, reject) => {
     const child = spawn26("bash", ["-c", cmd], {
       stdio: ["ignore", "pipe", "pipe"],
       timeout: opts.timeout ?? 3e5,
-      env: opts.env ?? process.env
+      env: personaplexPythonEnv(opts.env ?? {})
     });
     let stdout = "";
     let stderr = "";
@@ -590261,7 +590981,8 @@ function detectPersonaPlexCapability() {
     try {
       execSync50('python3 -c "import torch; assert torch.cuda.is_available()"', {
         timeout: 1e4,
-        stdio: "pipe"
+        stdio: "pipe",
+        env: personaplexPythonEnv()
       });
     } catch {
       const tier2 = selectWeightTier(vramGB);
@@ -590438,7 +591159,8 @@ async function installPersonaPlex(onInfo, weightTier) {
     const sitePackages = execSync50(`"${python}" -c "import moshi, os; print(os.path.dirname(moshi.__file__))"`, {
       encoding: "utf8",
       timeout: 5e3,
-      stdio: "pipe"
+      stdio: "pipe",
+      env: personaplexPythonEnv()
     }).trim();
     const serverFile = join114(sitePackages, "server.py");
     if (existsSync99(serverFile)) {
@@ -590455,7 +591177,8 @@ async function installPersonaPlex(onInfo, weightTier) {
     const sitePackages = execSync50(`"${python}" -c "import moshi, os; print(os.path.dirname(moshi.__file__))"`, {
       encoding: "utf8",
       timeout: 5e3,
-      stdio: "pipe"
+      stdio: "pipe",
+      env: personaplexPythonEnv()
     }).trim();
     const loadersFile = join114(sitePackages, "models", "loaders.py");
     if (existsSync99(loadersFile)) {
@@ -590559,7 +591282,8 @@ $2if filename.endswith(".safetensors"):`
     const sitePackages2 = execSync50(`"${python}" -c "import moshi, os; print(os.path.dirname(moshi.__file__))"`, {
       encoding: "utf8",
       timeout: 5e3,
-      stdio: "pipe"
+      stdio: "pipe",
+      env: personaplexPythonEnv()
     }).trim();
     const hybridDest = join114(sitePackages2, "hybrid_agent.py");
     const serverDest = join114(sitePackages2, "server.py");
@@ -590693,7 +591417,7 @@ async function startPersonaPlexDaemon(onInfo) {
       try {
         const weightPath = execSync50(
           `"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', '${repoInfo.file}', token=False))"`,
-          { encoding: "utf8", timeout: 6e4, stdio: "pipe" }
+          { encoding: "utf8", timeout: 6e4, stdio: "pipe", env: personaplexPythonEnv() }
         ).trim();
         if (existsSync99(weightPath)) {
           if (!existsSync99(cachedBf16)) {
@@ -590706,7 +591430,7 @@ state = {k: v.to(torch.bfloat16) if v.is_floating_point() else v for k, v in sta
 save_file(state, '${cachedBf16}')
 print('Converted')
 "`,
-              { timeout: 18e4, stdio: "pipe" }
+              { timeout: 18e4, stdio: "pipe", env: personaplexPythonEnv() }
             );
           }
           if (existsSync99(cachedBf16)) {
@@ -590732,13 +591456,13 @@ print('Converted')
       try {
         const weightPath = execSync50(
           `"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', '${repoInfo.file}'${repoInfo.needsToken ? "" : ", token=False"}))"`,
-          { encoding: "utf8", timeout: 3e4, stdio: "pipe" }
+          { encoding: "utf8", timeout: 3e4, stdio: "pipe", env: personaplexPythonEnv() }
         ).trim();
         if (existsSync99(dequantScript) && existsSync99(weightPath)) {
           try {
             execSync50(
               `"${venvPython2}" "${dequantScript}" --input "${weightPath}" --output "${cachedBf16}"`,
-              { timeout: 3e5, stdio: "pipe" }
+              { timeout: 3e5, stdio: "pipe", env: personaplexPythonEnv() }
             );
             if (existsSync99(cachedBf16)) {
               extraArgs.push("--moshi-weight", cachedBf16);
@@ -590751,7 +591475,7 @@ print('Converted')
         try {
           const mimiPath = execSync50(
             `"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', 'tokenizer-e351c8d8-checkpoint125.safetensors', token=False))"`,
-            { encoding: "utf8", timeout: 3e4, stdio: "pipe" }
+            { encoding: "utf8", timeout: 3e4, stdio: "pipe", env: personaplexPythonEnv() }
           ).trim();
           if (existsSync99(mimiPath)) extraArgs.push("--mimi-weight", mimiPath);
         } catch {
@@ -590759,7 +591483,7 @@ print('Converted')
         try {
           const tokPath = execSync50(
             `"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', 'tokenizer_spm_32k_3.model', token=False))"`,
-            { encoding: "utf8", timeout: 3e4, stdio: "pipe" }
+            { encoding: "utf8", timeout: 3e4, stdio: "pipe", env: personaplexPythonEnv() }
           ).trim();
           if (existsSync99(tokPath)) extraArgs.push("--tokenizer", tokPath);
         } catch {
@@ -590814,7 +591538,7 @@ print('Converted')
   ];
   if (hybridEnabled) serverArgs.push("--hybrid");
   if (needsOffload) serverArgs.push("--cpu-offload");
-  const serverEnv = { ...process.env };
+  const serverEnv = personaplexPythonEnv();
   if (hybridEnabled) {
     serverEnv["HYBRID_ENABLED"] = "1";
     serverEnv["HYBRID_LLM_MODEL"] = ollamaModel;
@@ -590959,7 +591683,7 @@ async function clonePersonaPlexVoice(inputWav, voiceName, onInfo) {
       "cuda"
     ], {
       stdio: ["ignore", "pipe", "pipe"],
-      env: { ...process.env },
+      env: personaplexPythonEnv(),
       cwd: PERSONAPLEX_DIR
     });
     let output = "";
@@ -591135,6 +591859,7 @@ var init_personaplex = __esm({
     init_render();
     init_daemon_registry();
     init_typed_node_events();
+    init_dist5();
     WEIGHT_REPOS = {
       original: { repo: "nvidia/personaplex-7b-v1", file: "model.safetensors", sizeGB: 15.6, needsToken: true },
       nf4: { repo: "cudabenchmarktest/personaplex-7b-nf4", file: "model-nf4.safetensors", sizeGB: 4.1, needsToken: false },
@@ -599715,6 +600440,11 @@ import {
   spawn as nodeSpawn
 } from "node:child_process";
 import { createRequire as createRequire6 } from "node:module";
+function voicePythonEnv(extra = {}) {
+  const env2 = { ...process.env, ...extra };
+  applyMediaCudaDeviceFilterToEnv(env2, "tts");
+  return env2;
+}
 function sanitizeForTTS(text) {
   return text.replace(/^#{1,6}\s+/gm, "").replace(/\*{1,3}([^*]+)\*{1,3}/g, "$1").replace(/_{1,3}([^_]+)_{1,3}/g, "$1").replace(/~~([^~]+)~~/g, "$1").replace(/`([^`]+)`/g, "$1").replace(/```[\s\S]*?```/g, "").replace(/\[([^\]]+)\]\([^)]+\)/g, "$1").replace(/!\[([^\]]*)\]\([^)]+\)/g, "$1").replace(/^[\s]*[-*+]\s+/gm, "").replace(/^[\s]*\d+\.\s+/gm, "").replace(/^>\s+/gm, "").replace(/^[-*_]{3,}$/gm, "").replace(/\[[ xX]\]\s*/g, "").replace(/[\u{1F600}-\u{1F64F}]/gu, "").replace(/[\u{1F300}-\u{1F5FF}]/gu, "").replace(/[\u{1F680}-\u{1F6FF}]/gu, "").replace(/[\u{1F1E0}-\u{1F1FF}]/gu, "").replace(/[\u{2600}-\u{26FF}]/gu, "").replace(/[\u{2700}-\u{27BF}]/gu, "").replace(/[\u{FE00}-\u{FE0F}]/gu, "").replace(/[\u{1F900}-\u{1F9FF}]/gu, "").replace(/[\u{1FA00}-\u{1FA6F}]/gu, "").replace(/[\u{1FA70}-\u{1FAFF}]/gu, "").replace(/[\u{200D}]/gu, "").replace(/[\u{20E3}]/gu, "").replace(/[✓✔✗✘✕✖⚠️⏸⏹⏵●○◆◇■□▪▫►▼▲◀⬆⬇⬅➡↑↓←→⇐⇒⇑⇓]/g, "").replace(/[─━│┃┌┐└┘├┤┬┴┼╔╗╚╝╠╣╦╩╬⎿⎾▕▏⏐░▒▓█⠀-⣿]/g, "").replace(/\s{2,}/g, " ").trim();
 }
@@ -600620,6 +601350,7 @@ var init_voice = __esm({
     init_typed_node_events();
     init_render();
     init_daemon_registry();
+    init_dist5();
     VOICE_MODELS = {
       glados: {
         id: "glados",
@@ -602134,7 +602865,8 @@ except Exception as exc:
         return new Promise((resolve59, reject) => {
           const proc = nodeSpawn("sh", ["-c", command], {
             stdio: ["ignore", "pipe", "pipe"],
-            cwd: tmpdir20()
+            cwd: tmpdir20(),
+            env: voicePythonEnv()
           });
           let stdout = "";
           let stderr = "";
@@ -602909,7 +603641,7 @@ if __name__ == '__main__':
         const venvPy = luxttsVenvPy2();
         if (!existsSync109(venvPy)) return false;
         return new Promise((resolve59) => {
-          const env2 = { ...process.env, LUXTTS_REPO_PATH: luxttsRepoDir2() };
+          const env2 = voicePythonEnv({ LUXTTS_REPO_PATH: luxttsRepoDir2() });
           const daemon = nodeSpawn(venvPy, [luxttsInferScript2()], {
             stdio: ["pipe", "pipe", "pipe"],
             cwd: tmpdir20(),
@@ -625604,6 +626336,146 @@ var init_telegram_stats_menu = __esm({
   }
 });
 
+// packages/cli/src/tui/telegram-command-menu.ts
+function isBareTelegramGenerativeCommand(input) {
+  const trimmed = input.trim();
+  if (!trimmed.startsWith("/")) return false;
+  const parts = trimmed.split(/\s+/);
+  const name10 = (parts[0] ?? "").slice(1).split("@")[0]?.toLowerCase() ?? "";
+  return parts.length === 1 && GENERATIVE_COMMANDS.has(name10);
+}
+function buildTelegramCommandMenuItems(scope) {
+  const commands = listCommandRegistry({ includePlanned: false }).filter((cmd) => cmd.implementationStatus === "implemented").filter((cmd) => scope === "admin" || ["help", "start"].includes(cmd.name));
+  const seen = /* @__PURE__ */ new Set();
+  const items = [];
+  for (const cmd of commands) {
+    const signature = cmd.signatures[0]?.signature;
+    if (!signature || seen.has(cmd.name)) continue;
+    seen.add(cmd.name);
+    items.push({
+      label: `/${cmd.name}`,
+      command: `/${cmd.name}`,
+      description: cmd.signatures[0]?.description ?? signature,
+      adminOnly: scope === "admin"
+    });
+  }
+  return items.sort((a2, b) => a2.label.localeCompare(b.label));
+}
+function buildTelegramGenerativeMenuItems(commandName) {
+  const name10 = commandName.replace(/^\//, "").toLowerCase();
+  if (!GENERATIVE_COMMANDS.has(name10)) return [];
+  const title = name10[0].toUpperCase() + name10.slice(1);
+  return [
+    { label: `${title} models`, command: `/${name10} list`, description: `List available ${name10} models and hardware fit.` },
+    { label: `${title} setup`, command: `/${name10} setup`, description: `Show setup commands for the ${name10} backend.` }
+  ];
+}
+function encodeTelegramCommandMenuCallback(action, value2) {
+  const data = `${CALLBACK_PREFIX2}:${action[0]}:${value2}`;
+  return Buffer.byteLength(data, "utf8") <= MAX_CALLBACK_DATA_BYTES ? data : data.slice(0, MAX_CALLBACK_DATA_BYTES);
+}
+function decodeTelegramCommandMenuCallback(data) {
+  const parts = data.split(":");
+  if (parts.length !== 3 || parts[0] !== CALLBACK_PREFIX2) return null;
+  const action = parts[1] === "p" ? "page" : parts[1] === "r" ? "run" : parts[1] === "c" ? "close" : null;
+  if (!action) return null;
+  return { action, value: parts[2] ?? "" };
+}
+function renderTelegramCommandMenu(state) {
+  const totalPages = Math.max(1, Math.ceil(state.items.length / PAGE_SIZE2));
+  const page2 = Math.max(0, Math.min(state.page, totalPages - 1));
+  const start2 = page2 * PAGE_SIZE2;
+  const visible = state.items.slice(start2, start2 + PAGE_SIZE2);
+  const title = state.kind === "generative" ? "Generative command" : "Commands";
+  const scope = state.scope === "admin" ? "admin" : "public";
+  const lines = [
+    `<b>${escapeHTML3(title)}</b>`,
+    `<i>${escapeHTML3(scope)} scope - page ${page2 + 1}/${totalPages}</i>`,
+    "",
+    ...visible.flatMap((item) => [
+      `<code>${escapeHTML3(item.command)}</code>`,
+      escapeHTML3(item.description)
+    ])
+  ];
+  const keyboard = visible.map((item, offset) => [{
+    text: item.label.slice(0, 32),
+    callback_data: encodeTelegramCommandMenuCallback("run", start2 + offset)
+  }]);
+  const nav = [];
+  nav.push({ text: "Close", callback_data: encodeTelegramCommandMenuCallback("close", 0) });
+  if (page2 > 0) nav.push({ text: "Prev", callback_data: encodeTelegramCommandMenuCallback("page", page2 - 1) });
+  nav.push({ text: `${page2 + 1}/${totalPages}`, callback_data: encodeTelegramCommandMenuCallback("page", page2) });
+  if (page2 < totalPages - 1) nav.push({ text: "Next", callback_data: encodeTelegramCommandMenuCallback("page", page2 + 1) });
+  keyboard.push(nav);
+  return { text: lines.join("\n"), reply_markup: { inline_keyboard: keyboard } };
+}
+function handleTelegramCommandMenuCallback(data, state, now = Date.now()) {
+  const decoded = decodeTelegramCommandMenuCallback(data);
+  if (!decoded) return null;
+  if (state.expiresAt <= now) return null;
+  if (decoded.action === "close") return { close: true };
+  if (decoded.action === "page") {
+    const totalPages = Math.max(1, Math.ceil(state.items.length / PAGE_SIZE2));
+    const page2 = Math.max(0, Math.min(Number.parseInt(decoded.value, 10) || 0, totalPages - 1));
+    const newState = { ...state, page: page2 };
+    return { newState, render: renderTelegramCommandMenu(newState) };
+  }
+  const index = Number.parseInt(decoded.value, 10);
+  const item = Number.isFinite(index) ? state.items[index] : void 0;
+  return item ? { command: item.command } : null;
+}
+function escapeHTML3(text) {
+  return text.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
+}
+var CALLBACK_PREFIX2, PAGE_SIZE2, TTL_MS, MAX_CALLBACK_DATA_BYTES, GENERATIVE_COMMANDS, TelegramCommandMenuStateStore;
+var init_telegram_command_menu = __esm({
+  "packages/cli/src/tui/telegram-command-menu.ts"() {
+    "use strict";
+    init_command_registry();
+    CALLBACK_PREFIX2 = "ocm";
+    PAGE_SIZE2 = 8;
+    TTL_MS = 10 * 60 * 1e3;
+    MAX_CALLBACK_DATA_BYTES = 64;
+    GENERATIVE_COMMANDS = /* @__PURE__ */ new Set(["image", "video", "sound", "music"]);
+    TelegramCommandMenuStateStore = class {
+      states = /* @__PURE__ */ new Map();
+      key(chatId, messageId) {
+        return `${chatId}:${messageId}`;
+      }
+      create(input, now = Date.now()) {
+        return {
+          ...input,
+          createdAt: now,
+          expiresAt: now + TTL_MS
+        };
+      }
+      set(state) {
+        this.states.set(this.key(state.chatId, state.messageId), state);
+      }
+      get(chatId, messageId, now = Date.now()) {
+        const state = this.states.get(this.key(chatId, messageId));
+        if (!state) return void 0;
+        if (state.expiresAt <= now) {
+          this.delete(chatId, messageId);
+          return void 0;
+        }
+        return state;
+      }
+      delete(chatId, messageId) {
+        this.states.delete(this.key(chatId, messageId));
+      }
+      prune(now = Date.now()) {
+        for (const [key, state] of this.states) {
+          if (state.expiresAt <= now) this.states.delete(key);
+        }
+      }
+      clear() {
+        this.states.clear();
+      }
+    };
+  }
+});
+
 // packages/cli/src/tui/telegram-creative-tools.ts
 import { createCipheriv as createCipheriv4, createDecipheriv as createDecipheriv4, randomBytes as randomBytes23 } from "node:crypto";
 import {
@@ -630996,6 +631868,7 @@ var init_telegram_bridge = __esm({
     init_command_registry();
     init_telegram_help_menu();
     init_telegram_stats_menu();
+    init_telegram_command_menu();
     init_scoped_personality();
     init_voice_soul();
     init_telegram_creative_tools();
@@ -631606,6 +632479,8 @@ Telegram link integrity contract:
       statsMenuTimers = null;
       /** Prune expired stats menu states every 5 minutes */
       statsMenuPruneTimer = null;
+      /** Telegram-native command and generative command menus */
+      telegramCommandMenuStates = new TelegramCommandMenuStateStore();
       /** Command handler for admin DM slash commands (wired from interactive.ts) */
       commandHandler = null;
       /** Callback fired after a Telegram user completes the TUI-only admin auth challenge */
@@ -631938,6 +632813,10 @@ Telegram link integrity contract:
         const name10 = this.telegramSlashName(input);
         return name10 === "help" || name10 === "h" || name10 === "commands" || name10 === "cmds";
       }
+      isTelegramCommandsMenuCommand(input) {
+        const name10 = this.telegramSlashName(input);
+        return name10 === "commands" || name10 === "cmds";
+      }
       isTelegramStatsCommand(input) {
         const name10 = this.telegramSlashName(input);
         return name10 === "stats" || name10 === "metrics";
@@ -632531,6 +633410,49 @@ ${message2}`)
           this.helpMenuTimers.startTimer(state);
         }
       }
+      async replyWithTelegramCommandMenu(msg, isAdmin, kind, commandName) {
+        const scope = isAdmin ? "admin" : "public";
+        const items = kind === "generative" ? buildTelegramGenerativeMenuItems(commandName ?? "") : buildTelegramCommandMenuItems(scope);
+        if (items.length === 0) {
+          await this.replyToTelegramMessage(msg, "No Telegram command menu entries are available.");
+          return;
+        }
+        if (msg.guestQueryId || !isAdmin) {
+          const lines = items.slice(0, 24).map((item) => `${item.command} - ${item.description}`);
+          const text = ["Available commands:", "", ...lines].join("\n");
+          if (msg.guestQueryId) {
+            await this.answerGuestQuery(msg.guestQueryId, text);
+          } else {
+            await this.replyToTelegramMessage(msg, text);
+          }
+          return;
+        }
+        const previewState = this.telegramCommandMenuStates.create({
+          chatId: msg.chatId,
+          messageId: 0,
+          invokerMessageId: msg.messageId,
+          fromUserId: msg.fromUserId ?? 0,
+          scope,
+          kind,
+          page: 0,
+          items
+        });
+        const menu = renderTelegramCommandMenu(previewState);
+        const sent = await this.apiCall("sendMessage", {
+          chat_id: msg.chatId,
+          text: menu.text,
+          parse_mode: "HTML",
+          reply_markup: JSON.stringify(menu.reply_markup),
+          ...msg.chatType !== "private" ? { reply_to_message_id: msg.messageId } : {}
+        });
+        if (sent.ok && sent.result?.message_id) {
+          this.telegramCommandMenuStates.prune();
+          this.telegramCommandMenuStates.set({
+            ...previewState,
+            messageId: sent.result.message_id
+          });
+        }
+      }
       collectSessionMetricsSnapshot() {
         if (this._metricsProvider) {
           try {
@@ -637339,6 +638261,7 @@ ${TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT}`);
         this.telegramActiveWorkGenerations.clear();
         this.telegramActiveWorkStartedAtMs.clear();
         this.telegramAdminLivePanels.clear();
+        this.telegramCommandMenuStates.clear();
         this.flushTelegramViewWrites();
         this.flushTelegramTuiWrites();
         this.telegramActiveInferences.clear();
@@ -637717,6 +638640,10 @@ ${summary}` : ""
           return;
         }
         const isAdmin = this.isAdminUser(msg);
+        if (msg.text.trim().startsWith("/") && this.isTelegramCommandsMenuCommand(normalizedCommandText)) {
+          await this.replyWithTelegramCommandMenu(msg, isAdmin, "commands");
+          return;
+        }
         if (msg.text.trim().startsWith("/") && this.isTelegramHelpCommand(normalizedCommandText)) {
           await this.replyWithTelegramHelp(msg, isAdmin);
           return;
@@ -637751,6 +638678,10 @@ ${summary}` : ""
         const toolContext = this.resolveToolContext(msg, isAdmin);
         const isAdminDM = toolContext === "telegram-admin-dm";
         const sessionKey = this.sessionKeyForMessage(msg);
+        if (isAdminDM && isBareTelegramGenerativeCommand(normalizedCommandText)) {
+          await this.replyWithTelegramCommandMenu(msg, isAdmin, "generative", telegramSlash);
+          return;
+        }
         if (msg.text.trim().startsWith("/") && TELEGRAM_REMINDER_SLASH_COMMANDS.has(telegramSlash)) {
           await this.handleTelegramReminderSlash(msg, normalizedCommandText, toolContext);
           return;
@@ -640485,6 +641416,90 @@ Scoped workspace: ${scopedRoot}`,
         return Boolean(result.ok);
       }
       async handleTelegramCallbackQuery(callback) {
+        const commandMenuDecoded = decodeTelegramCommandMenuCallback(callback.data);
+        if (commandMenuDecoded) {
+          let answerText2 = "";
+          let alert2 = false;
+          let answered = false;
+          try {
+            const chatId = callback.chatId;
+            const messageId = callback.messageId;
+            if (!chatId || !messageId) {
+              answerText2 = "Cannot identify menu message.";
+              alert2 = true;
+              return;
+            }
+            const menuState = this.telegramCommandMenuStates.get(chatId, messageId);
+            if (!menuState) {
+              answerText2 = "This command menu expired. Send /commands for a fresh one.";
+              alert2 = true;
+              return;
+            }
+            const isAdmin = this.isAdminActor(callback.fromUserId, callback.username);
+            if (callback.fromUserId !== menuState.fromUserId && !isAdmin) {
+              answerText2 = "Only the user who opened this menu can use it.";
+              alert2 = true;
+              return;
+            }
+            if (!isAdmin) {
+              answerText2 = "That command requires Telegram admin authentication.";
+              alert2 = true;
+              return;
+            }
+            const result = handleTelegramCommandMenuCallback(callback.data, menuState);
+            if (!result) {
+              answerText2 = "Unknown or expired command menu action.";
+              alert2 = true;
+              return;
+            }
+            if (result.close) {
+              this.telegramCommandMenuStates.delete(chatId, messageId);
+              await this.apiCall("deleteMessage", { chat_id: chatId, message_id: messageId }).catch(() => {
+              });
+              if (menuState.invokerMessageId) {
+                await this.apiCall("deleteMessage", { chat_id: chatId, message_id: menuState.invokerMessageId }).catch(() => {
+                });
+              }
+              answered = await this.answerCallbackQuery(callback.id).catch(() => false);
+              return;
+            }
+            if (result.render && result.newState) {
+              this.telegramCommandMenuStates.set(result.newState);
+              await this.apiCall("editMessageText", {
+                chat_id: chatId,
+                message_id: messageId,
+                text: result.render.text,
+                parse_mode: "HTML",
+                reply_markup: JSON.stringify(result.render.reply_markup)
+              });
+              return;
+            }
+            if (result.command) {
+              if (!this.commandHandler) {
+                answerText2 = "No command handler is available.";
+                alert2 = true;
+                return;
+              }
+              answered = await this.answerCallbackQuery(callback.id, `Running ${result.command}...`).catch(() => false);
+              const output = await this.commandHandler(result.command);
+              if (output) {
+                await this.sendMessageHTML(chatId, convertMarkdownToTelegramHTML(output));
+              }
+              return;
+            }
+          } catch (err) {
+            answerText2 = err instanceof Error ? err.message : String(err);
+            alert2 = true;
+          } finally {
+            if (answered) {
+            } else if (answerText2) {
+              await this.answerCallbackQuery(callback.id, answerText2.slice(0, 180), alert2).catch(() => false);
+            } else {
+              await this.answerCallbackQuery(callback.id).catch(() => false);
+            }
+          }
+          return;
+        }
         const helpDecoded = decodeHelpCallback(callback.data);
         if (helpDecoded) {
           let answerText2 = "";
@@ -673904,14 +674919,24 @@ Respond concisely and safely. Remember: you are talking to the general public.`;
       telegramBridge.setCommandHandler(async (input) => {
         const captured = [];
         const origWrite = process.stdout.write;
-        process.stdout.write = function(chunk, ..._args) {
+        process.stdout.write = function(chunk, ...args) {
           if (typeof chunk === "string") {
             captured.push(chunk);
+          } else if (Buffer.isBuffer(chunk)) {
+            captured.push(chunk.toString("utf8"));
           }
+          const cb = args.find((arg) => typeof arg === "function");
+          if (cb) cb();
           return true;
         };
         try {
-          const result = await handleSlashCommand(input, commandCtx);
+          const result = await runWithNonInteractiveSelectSurface(
+            () => handleSlashCommand(input, commandCtx),
+            {
+              maxItems: 24,
+              hint: "(Telegram: interactive menu shown as text; use concrete slash arguments or open the TUI to select)"
+            }
+          );
           process.stdout.write = origWrite;
           if (statusBar.isActive) statusBar.handleResize();
           if (result === "exit") {
@@ -673925,7 +674950,7 @@ Respond concisely and safely. Remember: you are talking to the general public.`;
             return `Skill invoked: ${result.name}`;
           }
           const raw = captured.join("");
-          const clean5 = raw.replace(/\x1B\[[0-9;]*[A-Za-z]/g, "").replace(/\x1B\][^\x07]*\x07/g, "").replace(/\x1B[()][A-Z0-9]/g, "").replace(/\x1B\[?\??[0-9;]*[a-zA-Z]/g, "").replace(/\x1B/g, "").replace(/[─━│┃┌┐└┘├┤┬┴┼╔╗╚╝╠╣╦╩╬⎿⎾▕▏⏐]/g, "").replace(/\n{3,}/g, "\n\n").trim();
+          const clean5 = raw.replace(/\x1B(?:\[[\d;?]*[ -/]*[@-~]|\][^\x07\x1B]*(?:\x07|\x1B\\)?|[@-Z\\-_])/g, "").replace(/\x1B/g, "").replace(/[─━│┃┌┐└┘├┤┬┴┼╔╗╚╝╠╣╦╩╬⎿⎾▕▏⏐]/g, "").replace(/\n{3,}/g, "\n\n").trim();
           if (!clean5) return null;
           return clean5.length > 3900 ? clean5.slice(0, 3900) + "\n..." : clean5;
         } catch (err) {