omnius 1.0.187 → 1.0.189

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,
@@ -259208,6 +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 = [];
@@ -259511,9 +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) {
@@ -260321,6 +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");
@@ -260363,33 +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) {
@@ -260408,7 +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) });
@@ -260447,7 +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,
@@ -260671,6 +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",
@@ -260683,7 +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)
         });
@@ -261735,6 +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 = [];
@@ -261891,9 +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}
+${result.llmContent}` : result.llmContent
   };
 }
 var DEFAULT_SOUND_MODEL, DEFAULT_MUSIC_MODEL, DIFFUSERS_AUDIO_PACKAGES, TRANSFORMERS_AUDIO_PACKAGES, AUDIOCRAFT_PACKAGES, STABLE_AUDIO_PACKAGES, TANGOFLUX_PACKAGES, AUDIO_GENERATION_MODEL_PRESETS, SOUND_GENERATION_QUALITY_LADDER, MUSIC_GENERATION_QUALITY_LADDER, DIFFUSERS_AUDIO_RUNNER, AUDIOCRAFT_RUNNER, TRANSFORMERS_AUDIO_RUNNER, TANGOFLUX_RUNNER, AudioGenerateTool;
@@ -262730,6 +263021,48 @@ if __name__ == "__main__":
         this.lastProgressAt = now;
         this.progressHandler(event);
       }
+      async acquireTransientLoadLease(args) {
+        if (!args.candidate)
+          return null;
+        const host = audioCandidateHost(args.candidate);
+        if (!host)
+          return null;
+        const broker = getModelBroker();
+        const decision2 = await broker.acquireTransientModelLoad({
+          name: args.candidate.model,
+          domain: args.kind,
+          host,
+          owner: `audio-generate-tool/${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;
@@ -262776,6 +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",
@@ -262792,7 +263136,7 @@ if __name__ == "__main__":
         ], {
           cwd: this.cwd,
           timeoutMs: 18e5,
-          env: python.env,
+          env: runnerEnv,
           progressLabel: `Downloading/loading ${args.model}`,
           onProgress: (event) => this.emitProgress(event)
         });
@@ -262872,33 +263216,6 @@ if __name__ == "__main__":
         const candidates = audioGenerationFallbackCandidates(kind, requestedModel, requestedBackend, generationFallbackEnabled2(args));
         const seed = optionalNumberArg2(args["seed"]);
         const playback = playbackRequested(args);
-        const broker = getModelBroker();
-        const firstCandidate = candidates[0];
-        let brokerGpuIndex = null;
-        if (firstCandidate) {
-          const decision2 = await broker.ensureModelLoadable({
-            name: firstCandidate.model,
-            domain: kind === "music" ? "music" : "sound",
-            host: firstCandidate.backend === "audiocraft" ? "audiocraft" : firstCandidate.backend === "tangoflux" ? "tangoflux" : firstCandidate.backend === "transformers" ? "diffusers-py" : "diffusers-py",
-            owner: `audio-generate-tool/${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) {
@@ -262918,15 +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) });
@@ -262954,19 +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) });
@@ -263306,6 +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;
@@ -263871,9 +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) {
@@ -265240,6 +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");
@@ -265295,35 +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: ${decision2.reason}`,
-              durationMs: performance.now() - start2
-            };
-          }
-        }
-        this._brokerGpuIndex = brokerGpuIndex;
         if (candidates.length === 0) {
           return {
             success: false,
@@ -265373,7 +265760,30 @@ if __name__ == "__main__":
             failed.push({ candidate, reason: "ComfyUI backend not yet implemented." });
             continue;
           }
-          const result = await this.prewarmDiffusers({ candidate, start: args.start, python: args.args["python"] });
+          const leaseDecision = await this.acquireTransientLoadLease({
+            candidate,
+            reason: "video-prewarm-needs-room",
+            start: args.start
+          });
+          if (leaseDecision && "success" in leaseDecision) {
+            failed.push({ candidate, reason: summarizeToolResult3(leaseDecision) });
+            if (index < args.candidates.length - 1) {
+              this.emitProgress({
+                stage: "setup",
+                message: `${candidate.model} did not fit current resources; trying ${args.candidates[index + 1].model}`
+              });
+            }
+            continue;
+          }
+          const lease = leaseDecision?.lease;
+          this._brokerGpuIndex = leaseDecision?.gpuIndex ?? null;
+          let result;
+          try {
+            result = await this.prewarmDiffusers({ candidate, start: args.start, python: args.args["python"] });
+          } finally {
+            await lease?.release();
+            this._brokerGpuIndex = null;
+          }
           if (result.success)
             return annotateVideoFallbackSuccess(result, failed, candidate);
           failed.push({ candidate, reason: summarizeToolResult3(result) });
@@ -265459,26 +265869,48 @@ if __name__ == "__main__":
               start: args.start
             });
           } else {
-            result = await this.generateWithDiffusers({
-              prompt: promptForCandidate,
-              model: candidate.model,
-              preset,
-              kind: args.kind,
-              imageArg: args.imageArg,
-              audioInput: args.audioInput,
-              width,
-              height,
-              numFrames,
-              fps,
-              steps,
-              guidance,
-              negativePrompt,
-              seed: args.seed,
-              hfToken: hfTokenOverride,
-              autoAcceptLicense,
-              start: args.start,
-              python: args.args["python"]
+            const leaseDecision = await this.acquireTransientLoadLease({
+              candidate,
+              reason: "video-gen-needs-room",
+              start: args.start
             });
+            if (leaseDecision && "success" in leaseDecision) {
+              failed.push({ candidate, reason: summarizeToolResult3(leaseDecision) });
+              if (index < args.candidates.length - 1) {
+                this.emitProgress({
+                  stage: "setup",
+                  message: `${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;
@@ -265670,6 +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",
@@ -265685,7 +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)
         });
@@ -568924,6 +569367,78 @@ var init_spinner = __esm({
   }
 });
 
+// packages/cli/src/tui/generative-progress.ts
+function generationKindForToolName(toolName) {
+  if (toolName === "generate_image") return "image";
+  if (toolName === "generate_audio") return "audio";
+  if (toolName === "generate_video") return "video";
+  if (toolName === "generate_tts" || toolName === "create_audio_file") return "tts";
+  return null;
+}
+function formatGenerativeProgress(kind, event, options2 = {}) {
+  const width = Math.max(8, Math.min(32, options2.width ?? (options2.surface === "telegram" ? 12 : 20)));
+  const label = kindLabel(kind);
+  const stage = stageLabel(event.stage);
+  const pct = finitePercent(event.percent);
+  const bytes = formatProgressBytes(event);
+  const elapsed = formatElapsed(event.elapsedMs);
+  const message2 = compactProgressMessage(event.message);
+  if (typeof pct === "number") {
+    const filled = Math.max(0, Math.min(width, Math.round(pct / 100 * width)));
+    const bar = `${"#".repeat(filled)}${"-".repeat(width - filled)}`;
+    return `${label} ${stage}: [${bar}] ${pct}% ${message2}${bytes}${elapsed}`;
+  }
+  return `${label} ${stage}: ${message2}${bytes}${elapsed}`;
+}
+function kindLabel(kind) {
+  if (kind === "tts") return "TTS";
+  return kind.slice(0, 1).toUpperCase() + kind.slice(1);
+}
+function stageLabel(stage) {
+  const normalized = String(stage || "process").trim().toLowerCase();
+  if (normalized === "setup") return "setup";
+  if (normalized === "download") return "download";
+  if (normalized === "load") return "load";
+  if (normalized === "generate") return "infer";
+  if (normalized === "save") return "save";
+  if (normalized === "thumbnail") return "thumbnail";
+  if (normalized === "hf_token_required") return "auth";
+  return "process";
+}
+function finitePercent(value2) {
+  if (typeof value2 !== "number" || !Number.isFinite(value2)) return void 0;
+  return Math.max(0, Math.min(100, Math.round(value2)));
+}
+function formatProgressBytes(event) {
+  if (typeof event.totalBytes !== "number" || !Number.isFinite(event.totalBytes) || event.totalBytes <= 0) {
+    return "";
+  }
+  const downloaded = typeof event.downloadedBytes === "number" && Number.isFinite(event.downloadedBytes) ? Math.max(0, event.downloadedBytes) : 0;
+  return ` (${formatBytes3(downloaded)} / ${formatBytes3(event.totalBytes)})`;
+}
+function formatElapsed(elapsedMs2) {
+  if (typeof elapsedMs2 !== "number" || !Number.isFinite(elapsedMs2) || elapsedMs2 <= 1500) return "";
+  return ` ${Math.round(elapsedMs2 / 1e3)}s`;
+}
+function compactProgressMessage(message2) {
+  return String(message2 || "working").replace(/\s+/g, " ").trim().slice(0, 220);
+}
+function formatBytes3(value2) {
+  const units = ["B", "KB", "MB", "GB", "TB"];
+  let amount = Math.max(0, value2);
+  let idx = 0;
+  while (amount >= 1024 && idx < units.length - 1) {
+    amount /= 1024;
+    idx++;
+  }
+  return idx === 0 ? `${Math.round(amount)}B` : `${amount.toFixed(1)}${units[idx]}`;
+}
+var init_generative_progress = __esm({
+  "packages/cli/src/tui/generative-progress.ts"() {
+    "use strict";
+  }
+});
+
 // packages/cli/src/api/py-embed.ts
 var py_embed_exports = {};
 __export(py_embed_exports, {
@@ -575515,7 +576030,7 @@ async function fetchOllamaModels(baseUrl) {
     const family = m2.details?.family;
     return {
       name: m2.name,
-      size: formatBytes3(m2.size),
+      size: formatBytes4(m2.size),
       sizeBytes: m2.size,
       modified: formatRelativeTime(m2.modified_at),
       parameterSize: m2.details?.parameter_size,
@@ -575961,7 +576476,7 @@ async function queryModelCapabilities(baseUrl, modelName) {
     return caps;
   }
 }
-function formatBytes3(bytes) {
+function formatBytes4(bytes) {
   if (bytes < 1024) return `${bytes} B`;
   const units = ["KB", "MB", "GB", "TB"];
   let size = bytes;
@@ -595166,7 +595681,7 @@ function formatWorkspaceExplorer(result) {
   const width = Math.max(12, ...result.entries.map((entry) => entry.path.length));
   for (const entry of result.entries) {
     lines.push(
-      `  ${entry.path.padEnd(Math.min(width, 70)).slice(0, 70)}  ${entry.kind.padEnd(6)}  ${formatBytes4(entry.sizeBytes).padStart(8)}`
+      `  ${entry.path.padEnd(Math.min(width, 70)).slice(0, 70)}  ${entry.kind.padEnd(6)}  ${formatBytes5(entry.sizeBytes).padStart(8)}`
     );
   }
   lines.push("");
@@ -595188,7 +595703,7 @@ function previewWorkspaceFile(root, relPath, options2 = {}) {
     return [
       "",
       `  File Preview: ${relPath}`,
-      `  Size: ${formatBytes4(st.size)} (too large for inline preview)`,
+      `  Size: ${formatBytes5(st.size)} (too large for inline preview)`,
       ""
     ].join("\n");
   }
@@ -595199,7 +595714,7 @@ function previewWorkspaceFile(root, relPath, options2 = {}) {
   return [
     "",
     `  File Preview: ${relPath}`,
-    `  Size: ${formatBytes4(st.size)}    Lines: ${rawLines.length}${rawLines.length > maxLines ? " (truncated)" : ""}`,
+    `  Size: ${formatBytes5(st.size)}    Lines: ${rawLines.length}${rawLines.length > maxLines ? " (truncated)" : ""}`,
     "",
     ...visible.map((line, idx) => `  ${String(idx + 1).padStart(gutter)} | ${line}`),
     ""
@@ -595230,7 +595745,7 @@ function scoreWorkspaceFile(entry, query) {
   }
   return score;
 }
-function formatBytes4(bytes) {
+function formatBytes5(bytes) {
   if (bytes < 1024) return `${bytes} B`;
   if (bytes < 1024 * 1024) return `${(bytes / 1024).toFixed(1)} KB`;
   return `${(bytes / (1024 * 1024)).toFixed(1)} MB`;
@@ -600895,7 +601410,7 @@ function describeTaskComplete(summary, completed, personality = 2, _stark = fals
   }
   return `Task completed, but no summary was generated to describe the outcome`;
 }
-function formatBytes5(bytes) {
+function formatBytes6(bytes) {
   if (bytes < 1024) return `${bytes}B`;
   if (bytes < 1024 * 1024) return `${(bytes / 1024).toFixed(0)}KB`;
   return `${(bytes / (1024 * 1024)).toFixed(1)}MB`;
@@ -603605,7 +604120,7 @@ Error: ${err instanceof Error ? err.message : String(err)}`
               const pct = Math.round(received / contentLength * 100);
               if (pct === 25 || pct === 50 || pct === 75 || pct === 100) {
                 renderInfo(
-                  `  ${pct}% (${formatBytes5(received)} / ${formatBytes5(contentLength)})`
+                  `  ${pct}% (${formatBytes6(received)} / ${formatBytes6(contentLength)})`
                 );
               }
             }
@@ -603613,7 +604128,7 @@ Error: ${err instanceof Error ? err.message : String(err)}`
           const fullBuffer = Buffer.concat(chunks);
           writeFileSync54(onnxPath, fullBuffer);
           renderInfo(
-            `${model.label} model downloaded (${formatBytes5(fullBuffer.length)}).`
+            `${model.label} model downloaded (${formatBytes6(fullBuffer.length)}).`
           );
         }
       }
@@ -611742,15 +612257,7 @@ async function handleImageCommand(ctx3, arg, hasLocal) {
   return "handled";
 }
 function formatImageGenerationProgress(event) {
-  const pct = event.percent;
-  const elapsed = event.elapsedMs && event.elapsedMs > 1500 ? ` ${Math.round(event.elapsedMs / 1e3)}s` : "";
-  if (typeof pct === "number") {
-    const width = 20;
-    const filled = Math.max(0, Math.min(width, Math.round(pct / 100 * width)));
-    const bar = `${"#".repeat(filled)}${"-".repeat(width - filled)}`;
-    return `Image ${event.stage}: [${bar}] ${pct}% ${event.message}${elapsed}`;
-  }
-  return `Image ${event.stage}: ${event.message}${elapsed}`;
+  return formatGenerativeProgress("image", event);
 }
 function rateVideoPresetForHardware(preset, specs) {
   const min = preset.minVramGB;
@@ -612030,15 +612537,7 @@ async function handleVideoCommand(ctx3, arg, hasLocal) {
   return "handled";
 }
 function formatVideoGenerationProgress(event) {
-  const pct = event.percent;
-  const elapsed = event.elapsedMs && event.elapsedMs > 1500 ? ` ${Math.round(event.elapsedMs / 1e3)}s` : "";
-  if (typeof pct === "number") {
-    const width = 20;
-    const filled = Math.max(0, Math.min(width, Math.round(pct / 100 * width)));
-    const bar = `${"#".repeat(filled)}${"-".repeat(width - filled)}`;
-    return `Video ${event.stage}: [${bar}] ${pct}% ${event.message}${elapsed}`;
-  }
-  return `Video ${event.stage}: ${event.message}${elapsed}`;
+  return formatGenerativeProgress("video", event);
 }
 function activeAudioModel(settings, kind) {
   return kind === "music" ? settings.musicModel : settings.soundModel;
@@ -612346,16 +612845,7 @@ async function handleAudioGenerationCommand(ctx3, arg, hasLocal, kind) {
   return "handled";
 }
 function formatAudioGenerationProgress(event) {
-  const pct = event.percent;
-  const elapsed = event.elapsedMs && event.elapsedMs > 1500 ? ` ${Math.round(event.elapsedMs / 1e3)}s` : "";
-  const bytes = typeof event.totalBytes === "number" && event.totalBytes > 0 ? ` (${formatFileSize(event.downloadedBytes ?? 0)} / ${formatFileSize(event.totalBytes)})` : "";
-  if (typeof pct === "number") {
-    const width = 20;
-    const filled = Math.max(0, Math.min(width, Math.round(pct / 100 * width)));
-    const bar = `${"#".repeat(filled)}${"-".repeat(width - filled)}`;
-    return `Audio ${event.stage}: [${bar}] ${pct}% ${event.message}${bytes}${elapsed}`;
-  }
-  return `Audio ${event.stage}: ${event.message}${bytes}${elapsed}`;
+  return formatGenerativeProgress("audio", event);
 }
 async function showHelpMenu(ctx3) {
   const slashCommands = getSlashHelpEntries();
@@ -617504,6 +617994,7 @@ var init_commands = __esm({
     "use strict";
     init_model_picker();
     init_render();
+    init_generative_progress();
     init_command_registry();
     init_hf_token_prompt();
     init_dist5();
@@ -626249,6 +626740,10 @@ function scopedTool(base3, root, mode) {
   if (typeof baseSetExpander === "function") {
     wrapper.setPromptExpander = (expander) => baseSetExpander.call(base3, expander);
   }
+  const baseSetProgress = base3.setProgressCallback;
+  if (typeof baseSetProgress === "function") {
+    wrapper.setProgressCallback = (handler) => baseSetProgress.call(base3, handler);
+  }
   return wrapper;
 }
 function withTelegramAutoAttachmentNotice(result, artifactCount) {
@@ -626557,6 +627052,16 @@ var init_telegram_creative_tools = __esm({
         },
         required: []
       };
+      progressHandler = null;
+      setProgressCallback(handler) {
+        this.progressHandler = handler;
+      }
+      emitProgress(start2, event) {
+        try {
+          this.progressHandler?.({ ...event, elapsedMs: performance.now() - start2 });
+        } catch {
+        }
+      }
       async execute(args) {
         const start2 = performance.now();
         const text = typeof args["text"] === "string" && args["text"].trim() ? args["text"].trim() : typeof args["input"] === "string" && args["input"].trim() ? args["input"].trim() : typeof args["prompt"] === "string" && args["prompt"].trim() ? args["prompt"].trim() : "";
@@ -626590,8 +627095,10 @@ var init_telegram_creative_tools = __esm({
         }
         let result;
         try {
+          this.emitProgress(start2, { stage: "setup", message: "Preparing scoped TTS audio file" });
           await mkdir19(dirname37(guarded.path.abs), { recursive: true });
           const tts = new TtsGenerateTool();
+          this.emitProgress(start2, { stage: "load", message: "Starting TTS backend" });
           result = await tts.execute({
             text,
             output: guarded.path.abs,
@@ -626624,6 +627131,7 @@ ${(result.error || result.output || "").slice(0, 1200)}`,
         }
         rememberCreated(this.root, guarded.path.abs);
         const sizeKB = Math.round(statSync43(guarded.path.abs).size / 1024);
+        this.emitProgress(start2, { stage: "save", message: `Saved scoped audio file (${sizeKB}KB)` });
         return withTelegramAutoAttachmentNotice({
           success: true,
           output: `Created audio file: ${guarded.path.abs} (${sizeKB}KB WAV)
@@ -631246,7 +631754,13 @@ function normalizeTelegramCallbackQuery(update2) {
     data
   };
 }
-function adaptTool5(tool, todoSessionId) {
+function adaptTool5(tool, todoSessionId, progress) {
+  const progressTool = tool;
+  if (generationKindForToolName(tool.name) && typeof progressTool.setProgressCallback === "function") {
+    progressTool.setProgressCallback((event) => {
+      progress?.onProgress(tool.name, event);
+    });
+  }
   return {
     name: tool.name,
     description: tool.description,
@@ -631258,7 +631772,15 @@ function adaptTool5(tool, todoSessionId) {
       }
       try {
         const result = await tool.execute(args);
+        progress?.complete(tool.name, result);
         return { success: result.success, output: result.output, error: result.error, llmContent: result.llmContent };
+      } catch (err) {
+        progress?.complete(tool.name, {
+          success: false,
+          output: "",
+          error: err instanceof Error ? err.message : String(err)
+        });
+        throw err;
       } finally {
         if (todoSessionId && (tool.name === "todo_write" || tool.name === "todo_read")) {
           setTodoSessionId(previousTodoSession);
@@ -631429,6 +631951,7 @@ var init_telegram_bridge = __esm({
     init_scoped_personality();
     init_voice_soul();
     init_telegram_creative_tools();
+    init_generative_progress();
     init_omnius_directory();
     init_stimulation();
     init_pid_controller();
@@ -637910,6 +638433,91 @@ ${TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT}`);
       async sendChatAction(chatId, action) {
         await this.apiCall("sendChatAction", { chat_id: chatId, action });
       }
+      createTelegramGenerativeProgressBridge(chatId, msg) {
+        if (chatId === void 0) return void 0;
+        const states = /* @__PURE__ */ new Map();
+        const stateFor = (toolName) => {
+          let state = states.get(toolName);
+          if (!state) {
+            state = {
+              messageId: null,
+              pump: null,
+              queuedHtml: null,
+              lastRenderedAt: 0
+            };
+            states.set(toolName, state);
+          }
+          return state;
+        };
+        const kindTitle = (toolName) => {
+          const kind = generationKindForToolName(toolName);
+          if (!kind) return "Generation";
+          if (kind === "tts") return "TTS audio";
+          return `${kind.slice(0, 1).toUpperCase()}${kind.slice(1)}`;
+        };
+        const enqueue = (state, html) => {
+          state.queuedHtml = html;
+          if (state.pump) return;
+          state.pump = (async () => {
+            while (state.queuedHtml) {
+              const nextHtml = state.queuedHtml;
+              state.queuedHtml = null;
+              if (state.messageId === null) {
+                const messageId = await this.sendLiveMessage(chatId, nextHtml, msg?.messageId);
+                if (!messageId) return;
+                state.messageId = messageId;
+              } else {
+                await this.editLiveMessage(chatId, state.messageId, nextHtml);
+              }
+            }
+          })().catch(() => {
+          }).finally(() => {
+            state.pump = null;
+            if (state.queuedHtml) enqueue(state, state.queuedHtml);
+          });
+        };
+        const progressHtml = (toolName, event) => {
+          const kind = generationKindForToolName(toolName);
+          if (!kind) return null;
+          const rendered = redactTelegramLocalPaths(formatGenerativeProgress(kind, event, { surface: "telegram" }));
+          return `<b>${escapeTelegramHTML(kindTitle(toolName))} generation</b>
+<code>${escapeTelegramHTML(rendered)}</code>`;
+        };
+        const completeHtml = (toolName, result) => {
+          if (!generationKindForToolName(toolName)) return null;
+          if (result.success) {
+            return `<b>${escapeTelegramHTML(kindTitle(toolName))} generation complete</b>`;
+          }
+          const reason = redactTelegramLocalPaths((result.error || result.output || "Generation failed").slice(0, 900));
+          return `<b>${escapeTelegramHTML(kindTitle(toolName))} generation failed</b>
+<code>${escapeTelegramHTML(reason)}</code>`;
+        };
+        return {
+          onProgress: (toolName, event) => {
+            const html = progressHtml(toolName, event);
+            if (!html) return;
+            const state = stateFor(toolName);
+            const now = Date.now();
+            const stage = String(event.stage || "process");
+            const percent = typeof event.percent === "number" && Number.isFinite(event.percent) ? Math.round(event.percent) : void 0;
+            const terminalStage = stage === "save" || stage === "thumbnail" || stage === "hf_token_required";
+            const shouldRender = state.lastRenderedAt === 0 || stage !== state.lastStage || typeof percent === "number" && (state.lastPercent === void 0 || Math.abs(percent - state.lastPercent) >= 5) || now - state.lastRenderedAt >= 3500 || terminalStage;
+            if (!shouldRender) return;
+            state.lastRenderedAt = now;
+            state.lastStage = stage;
+            state.lastPercent = percent;
+            enqueue(state, html);
+          },
+          complete: (toolName, result) => {
+            if (!generationKindForToolName(toolName)) return;
+            const state = states.get(toolName);
+            if (!state && result.success) return;
+            const html = completeHtml(toolName, result);
+            if (!html) return;
+            enqueue(state ?? stateFor(toolName), html);
+          }
+        };
+      }
       // ── Live message streaming (editMessageText pattern) ─────────────────
       /**
        * Send a placeholder message that will be progressively edited with
@@ -639942,6 +640550,7 @@ Scoped workspace: ${scopedRoot}`,
         const imageDefaults = this.imageGenerationDefaultsForRepo(repoRoot);
         const audioDefaults = this.audioGenerationDefaultsForRepo(repoRoot);
         const videoDefaults = this.videoGenerationDefaultsForRepo(repoRoot);
+        const generativeProgress = this.createTelegramGenerativeProgressBridge(chatId, msg);
         const taskComplete = {
           name: "task_complete",
           description: "Internal completion signal for Telegram runs. Put the actual user-facing reply in assistant text before calling this. Use summary 'no_reply' only to silently skip responding; never write that sentinel as assistant text.",
@@ -640088,7 +640697,7 @@ Scoped workspace: ${scopedRoot}`,
             }
           }
         }
-        const unfilteredAdaptedTools = allTools.map((tool) => adaptTool5(tool, todoSessionId));
+        const unfilteredAdaptedTools = allTools.map((tool) => adaptTool5(tool, todoSessionId, generativeProgress));
         let adaptedTools = unfilteredAdaptedTools;
         adaptedTools = applyToolPolicy(adaptedTools, context2, this.toolPolicyConfig);
         if (context2 === "telegram-admin-dm") {
@@ -640103,9 +640712,9 @@ Scoped workspace: ${scopedRoot}`,
             imageDefaults,
             audioDefaults,
             videoDefaults
-          ).map((tool) => adaptTool5(tool, todoSessionId));
+          ).map((tool) => adaptTool5(tool, todoSessionId, generativeProgress));
           adaptedTools.push(...creativeTools);
-          adaptedTools.push(adaptTool5(this.buildTelegramSendFileTool(context2, repoRoot, chatId, msg), todoSessionId));
+          adaptedTools.push(adaptTool5(this.buildTelegramSendFileTool(context2, repoRoot, chatId, msg), todoSessionId, generativeProgress));
           adaptedTools = this.filterNonAdminTelegramTools(adaptedTools);
           adaptedTools = adaptedTools.map((tool) => this.applyTelegramPublicQuota(tool, context2, chatId, msg));
         }
@@ -642850,7 +643459,7 @@ function buildShellLiveBlockLines(state, width) {
   const w = Math.max(36, width);
   const inner = Math.max(1, w - 4);
   const elapsed = Math.max(0, Date.now() - state.startedAt);
-  const status = state.status === "running" ? `live ${formatElapsed(elapsed)}` : state.status;
+  const status = state.status === "running" ? `live ${formatElapsed2(elapsed)}` : state.status;
   const title = ` Shell ${status} `;
   const top = `╭${fitWithFill(`─${title}`, w - 2, "─")}╮`;
   const bottom = `╰${"─".repeat(w - 2)}╯`;
@@ -642895,7 +643504,7 @@ function fitWithFill(value2, width, fill) {
   if (chars.length > width) return chars.slice(0, width).join("");
   return value2 + fill.repeat(width - chars.length);
 }
-function formatElapsed(ms) {
+function formatElapsed2(ms) {
   const seconds = Math.floor(ms / 1e3);
   if (seconds < 60) return `${seconds}s`;
   const minutes = Math.floor(seconds / 60);
@@ -668394,6 +669003,12 @@ function getVersion4() {
   return "0.0.0";
 }
 function adaptTool6(tool) {
+  const progressTool = tool;
+  if (generationKindForToolName(tool.name) && typeof progressTool.setProgressCallback === "function") {
+    progressTool.setProgressCallback((event) => {
+      _generativeProgressSink?.(tool.name, event);
+    });
+  }
   return {
     name: tool.name,
     description: tool.description,
@@ -668704,6 +669319,17 @@ function audioGenerationDefaultsForRepo(repoRoot) {
 function createConfiguredAudioGenerateTool(repoRoot) {
   return new AudioGenerateTool(repoRoot, audioGenerationDefaultsForRepo(repoRoot));
 }
+function videoGenerationDefaultsForRepo(repoRoot) {
+  const settings = resolveSettings(repoRoot);
+  return {
+    model: typeof settings.videoModel === "string" && settings.videoModel.trim() ? settings.videoModel : void 0,
+    backend: settings.videoBackend,
+    defaultKind: settings.videoKind
+  };
+}
+function createConfiguredVideoGenerateTool(repoRoot) {
+  return new VideoGenerateTool(repoRoot, videoGenerationDefaultsForRepo(repoRoot));
+}
 function buildSubAgentTools(repoRoot, config) {
   return [
     // File + search
@@ -668802,6 +669428,7 @@ function buildSubAgentTools(repoRoot, config) {
     new CameraCaptureTool(),
     createConfiguredImageGenerateTool(repoRoot, config.backendUrl),
     createConfiguredAudioGenerateTool(repoRoot),
+    createConfiguredVideoGenerateTool(repoRoot),
     // Hardware sensors + radios (read-only scans)
     new GpsLocationTool(),
     new WifiControlTool(),
@@ -668889,6 +669516,8 @@ function buildTools(repoRoot, config, contextWindowSize, modelTier) {
     createConfiguredImageGenerateTool(repoRoot, config.backendUrl),
     // Sound/music Generation — local Diffusers/AudioCraft/Stable Audio paths
     createConfiguredAudioGenerateTool(repoRoot),
+    // Video Generation — local Diffusers/ComfyUI video pipelines
+    createConfiguredVideoGenerateTool(repoRoot),
     // Structured file reading (CSV, JSON, Markdown, binary detection)
     new StructuredReadTool(repoRoot),
     // Vision tools (Moondream — desktop awareness + point-and-click)
@@ -669551,16 +670180,6 @@ async function renderAsciiPreviewForImage(imagePath, displayPath, title, writer)
     return "";
   }
 }
-function formatImageGenerationProgress2(event) {
-  const elapsed = event.elapsedMs && event.elapsedMs > 1500 ? ` ${Math.round(event.elapsedMs / 1e3)}s` : "";
-  if (typeof event.percent === "number") {
-    const width = 20;
-    const filled = Math.max(0, Math.min(width, Math.round(event.percent / 100 * width)));
-    const bar = `${"#".repeat(filled)}${"-".repeat(width - filled)}`;
-    return `Image ${event.stage}: [${bar}] ${event.percent}% ${event.message}${elapsed}`;
-  }
-  return `Image ${event.stage}: ${event.message}${elapsed}`;
-}
 async function renderAsciiPreviewForToolResult(toolName, output, repoRoot, writer) {
   if (!output) return;
   try {
@@ -670701,14 +671320,11 @@ ${entry.fullContent}`
       fn();
     }
   };
-  for (const tool of tools) {
-    const maybeProgressTool = tool;
-    if (typeof maybeProgressTool.setProgressCallback === "function") {
-      maybeProgressTool.setProgressCallback((event) => {
-        contentWrite(() => renderInfo(formatImageGenerationProgress2(event)));
-      });
-    }
-  }
+  _generativeProgressSink = (toolName, event) => {
+    const kind = generationKindForToolName(toolName);
+    if (!kind) return;
+    contentWrite(() => renderInfo(formatGenerativeProgress(kind, event, { surface: "tui" })));
+  };
   let liveShellBlock = null;
   const scheduleLiveShellRepaint = () => {
     if (!liveShellBlock || liveShellBlock.repaintTimer || !statusBar?.isActive)
@@ -677484,7 +678100,7 @@ Rules:
     process.exit(1);
   }
 }
-var NEXUS_DIRECTORY_ORIGIN3, NEXUS_AGENT_DIRECTORY_URL, NEXUS_SPONSORS_URL3, _interactiveSessionActive, _interactiveSessionReason, _voiceChatSession2, taskManager, _apiCallbacks, _shellToolRef, _replToolRef, _fullSubAgentToolRef, _agentToolRef, _sendMessageToolRef, _agentLifecycleMgr, _activeRunnerRef, _parentRunnerForArchive, _wireSubAgentCallbacks, _wireAgentToolCallbacks, _wireSubAgentToolCallbacks, _autoUpdatedThisSession, _mcpManager, _pluginManager, _mcpTools, SELF_IMPROVE_INTERVAL, _tasksSinceImprove;
+var NEXUS_DIRECTORY_ORIGIN3, NEXUS_AGENT_DIRECTORY_URL, NEXUS_SPONSORS_URL3, _generativeProgressSink, _interactiveSessionActive, _interactiveSessionReason, _voiceChatSession2, taskManager, _apiCallbacks, _shellToolRef, _replToolRef, _fullSubAgentToolRef, _agentToolRef, _sendMessageToolRef, _agentLifecycleMgr, _activeRunnerRef, _parentRunnerForArchive, _wireSubAgentCallbacks, _wireAgentToolCallbacks, _wireSubAgentToolCallbacks, _autoUpdatedThisSession, _mcpManager, _pluginManager, _mcpTools, SELF_IMPROVE_INTERVAL, _tasksSinceImprove;
 var init_interactive = __esm({
   "packages/cli/src/tui/interactive.ts"() {
     "use strict";
@@ -677493,6 +678109,7 @@ var init_interactive = __esm({
     init_dist8();
     init_dist8();
     init_dist5();
+    init_generative_progress();
     init_dist();
     init_listen();
     init_voice_session();
@@ -677543,6 +678160,7 @@ var init_interactive = __esm({
     NEXUS_DIRECTORY_ORIGIN3 = (process.env["OMNIUS_NEXUS_DIRECTORY_ORIGIN"] || process.env["OMNIUS_NEXUS_SIGNALING_SERVER"] || "https://openagents.nexus").replace(/\/+$/, "");
     NEXUS_AGENT_DIRECTORY_URL = `${NEXUS_DIRECTORY_ORIGIN3}/api/v1/directory`;
     NEXUS_SPONSORS_URL3 = `${NEXUS_DIRECTORY_ORIGIN3}/api/v1/sponsors`;
+    _generativeProgressSink = null;
     _interactiveSessionActive = false;
     _interactiveSessionReason = "";
     _voiceChatSession2 = null;