omnius 1.0.21 → 1.0.23

package/dist/index.js

@@ -1474,7 +1474,7 @@ var init_security_classifier = __esm({
       // ── Network reads (safe)
       { match: /^(web_search|web_fetch)$/, info: NETWORK_READ },
       // ── Network outbound (mutating or remote inference)
-      { match: /^(image_generate|generate_image|vision|video_understand)$/, info: NETWORK_OUTBOUND },
+      { match: /^(image_generate|generate_image|generate_audio|generate_tts|create_audio_file|vision|video_understand|telegram_send_file)$/, info: NETWORK_OUTBOUND },
       { match: /^(transcribe_file|transcribe_url|youtube_download)$/, info: NETWORK_OUTBOUND },
       { match: /^(fortemi_bridge)$/, info: NETWORK_OUTBOUND },
       // ── Memory tools
@@ -1491,7 +1491,7 @@ var init_security_classifier = __esm({
       { match: /^(file_read|file_explore|list_directory|grep_search|glob_find|find_files)$/, info: LOCAL_READ },
       { match: /^(image_read|ocr|ocr_pdf|ocr_image_advanced|pdf_to_text|structured_read|read_structured_file)$/, info: LOCAL_READ },
       { match: /^(symbol_search|impact_analysis|code_neighbors|repo_map|codebase_map|semantic_map|import_graph)$/, info: LOCAL_READ },
-      { match: /^(diagnostic|git_info|environment_snapshot|process_health|todo_read|explore_tools)$/, info: LOCAL_READ },
+      { match: /^(diagnostic|git_info|environment_snapshot|process_health|todo_read|explore_tools|telegram_media_recent)$/, info: LOCAL_READ },
       { match: /^(log_explore|log_packet|change_log|phase_recall|code_graph)$/, info: LOCAL_READ },
       { match: /^skill_(list|execute|read)$/, info: LOCAL_READ },
       // ── Task completion (neutral signal)
@@ -5733,13 +5733,20 @@ var init_explore_tools = __esm({
       diagnostic: "Run project diagnostics (build, test, lint)",
       image_read: "Read and describe image contents",
       screenshot: "Capture a screenshot of the desktop",
+      ocr: "Extract text from images via OCR",
       ocr_image: "Extract text from images via OCR",
+      ocr_image_advanced: "Advanced OCR for images with layout-aware extraction",
       ocr_pdf: "Extract text from PDF pages via OCR",
       pdf_to_text: "Convert PDF to plain text",
       vision: "Describe what's on screen using Moondream",
+      video_understand: "Analyze a video file with transcription and keyframe understanding",
+      audio_analyze: "Classify sounds, detect speech, inspect spectrum, or analyze audio files",
       desktop_click: "Click at coordinates on the desktop",
       desktop_describe: "Describe a region of the desktop",
       transcribe_file: "Transcribe audio/video files to text",
+      telegram_media_recent: "List recent Telegram media available in the current chat scope",
+      generate_audio: "Generate sound effects or music with local model backends",
+      generate_tts: "Generate speech from text with configured voice/TTS backends",
       create_tool: "Create a new custom tool from a workflow",
       manage_tools: "List, inspect, or remove custom tools",
       skill_list: "List available AIWG skills",
@@ -84452,7 +84459,7 @@ var require_mime_types = __commonJS({
   "../node_modules/mime-types/index.js"(exports) {
     "use strict";
     var db = require_mime_db();
-    var extname16 = __require("path").extname;
+    var extname17 = __require("path").extname;
     var EXTRACT_TYPE_REGEXP = /^\s*([^;\s]*)(?:;|\s|$)/;
     var TEXT_TYPE_REGEXP = /^text\//i;
     exports.charset = charset;
@@ -84506,7 +84513,7 @@ var require_mime_types = __commonJS({
       if (!path11 || typeof path11 !== "string") {
         return false;
       }
-      var extension4 = extname16("x." + path11).toLowerCase().substr(1);
+      var extension4 = extname17("x." + path11).toLowerCase().substr(1);
       if (!extension4) {
         return false;
       }
@@ -250335,6 +250342,12 @@ import { spawn as spawn9 } from "node:child_process";
 import { existsSync as existsSync23, statSync as statSync8 } from "node:fs";
 import { chmod as chmod3, mkdir as mkdir11, writeFile as writeFile16 } from "node:fs/promises";
 import { join as join36, resolve as resolve18 } from "node:path";
+function officialBflModel(model) {
+  return `${OFFICIAL_BFL_ORG}/${model}`;
+}
+function normalizeImageGenerationModel(model) {
+  return IMAGE_GENERATION_MODEL_REPLACEMENTS.get(model) ?? model;
+}
 function parsePercent(text) {
   const match = text.match(/\b(\d{1,3})%\b/);
   if (!match)
@@ -250397,35 +250410,44 @@ function isBackend(value2) {
 function getImageGenerationPreset(model) {
   if (!model)
     return void 0;
-  return IMAGE_GENERATION_MODEL_PRESETS.find((preset) => preset.id === model);
+  const resolved = normalizeImageGenerationModel(model);
+  return IMAGE_GENERATION_MODEL_PRESETS.find((preset) => preset.id === resolved);
 }
 function imageGenerationQualityLadder() {
   return IMAGE_GENERATION_QUALITY_LADDER.map((id) => getImageGenerationPreset(id)).filter((preset) => Boolean(preset));
 }
+function imageGenerationFallbackAlternates(model) {
+  if (!model)
+    return [];
+  const resolved = normalizeImageGenerationModel(model);
+  return IMAGE_GENERATION_MODEL_PRESETS.filter((preset) => preset.fallbackFor?.includes(resolved));
+}
 function inferImageGenerationBackend(model, requested) {
   if (requested && isBackend(requested))
     return requested;
   if (!model || model === "auto")
     return "auto";
-  const preset = getImageGenerationPreset(model);
+  const resolved = normalizeImageGenerationModel(model);
+  const preset = getImageGenerationPreset(resolved);
   if (preset)
     return preset.backend;
-  if (model.startsWith("x/"))
+  if (resolved.startsWith("x/"))
     return "ollama";
-  if (model === "sdcpp" || model.startsWith("sdcpp:"))
+  if (resolved === "sdcpp" || resolved.startsWith("sdcpp:"))
     return "sdcpp";
-  if (model.startsWith("/") || model.endsWith(".gguf") || model.endsWith(".safetensors") || model.endsWith(".ckpt"))
+  if (resolved.startsWith("/") || resolved.endsWith(".gguf") || resolved.endsWith(".safetensors") || resolved.endsWith(".ckpt"))
     return "sdcpp";
   return "diffusers";
 }
 function imageCandidateFor(model, requestedBackend) {
-  let backend = inferImageGenerationBackend(model, requestedBackend);
+  const resolved = normalizeImageGenerationModel(model);
+  let backend = inferImageGenerationBackend(resolved, requestedBackend);
   if (backend === "auto")
     backend = "diffusers";
   return {
-    model,
+    model: resolved,
     backend,
-    preset: getImageGenerationPreset(model)
+    preset: getImageGenerationPreset(resolved)
   };
 }
 function imageGenerationFallbackCandidates(requestedModel, requestedBackend, allowFallback = true) {
@@ -250438,6 +250460,8 @@ function imageGenerationFallbackCandidates(requestedModel, requestedBackend, all
   };
   if (requestedModel) {
     add2(imageCandidateFor(requestedModel, requestedBackend));
+    for (const alternate of imageGenerationFallbackAlternates(requestedModel))
+      add2(imageCandidateFor(alternate.id));
   } else if (requestedBackend && requestedBackend !== "auto") {
     const firstForBackend = ladder.find((preset) => preset.backend === requestedBackend);
     add2(imageCandidateFor(firstForBackend?.id ?? (requestedBackend === "ollama" ? DEFAULT_OLLAMA_IMAGE_MODEL : DEFAULT_DIFFUSERS_IMAGE_MODEL), requestedBackend));
@@ -250446,10 +250470,14 @@ function imageGenerationFallbackCandidates(requestedModel, requestedBackend, all
   }
   if (!allowFallback)
     return candidates.length ? candidates : [imageCandidateFor(DEFAULT_DIFFUSERS_IMAGE_MODEL, requestedBackend)];
-  const primaryIndex = requestedModel ? ladder.findIndex((preset) => preset.id === requestedModel) : requestedBackend && requestedBackend !== "auto" ? ladder.findIndex((preset) => preset.backend === requestedBackend) : 0;
+  const normalizedRequestedModel = requestedModel ? normalizeImageGenerationModel(requestedModel) : void 0;
+  const primaryIndex = normalizedRequestedModel ? ladder.findIndex((preset) => preset.id === normalizedRequestedModel) : requestedBackend && requestedBackend !== "auto" ? ladder.findIndex((preset) => preset.backend === requestedBackend) : 0;
   const fallbackTail = primaryIndex >= 0 ? ladder.slice(primaryIndex) : ladder;
-  for (const preset of fallbackTail)
+  for (const preset of fallbackTail) {
     add2(imageCandidateFor(preset.id));
+    for (const alternate of imageGenerationFallbackAlternates(preset.id))
+      add2(imageCandidateFor(alternate.id));
+  }
   return candidates;
 }
 function imageGenerationDir(repoRoot = ".") {
@@ -250745,13 +250773,34 @@ function parseRunnerJson(stdout) {
   }
   return null;
 }
-var DEFAULT_DIFFUSERS_IMAGE_MODEL, DEFAULT_OLLAMA_IMAGE_MODEL, DIFFUSERS_PYTHON_PACKAGES, SDCPP_PYTHON_PACKAGES, IMAGE_GENERATION_MODEL_PRESETS, IMAGE_GENERATION_QUALITY_LADDER, OLLAMA_IMAGE_MODELS, DIFFUSERS_RUNNER, SDCPP_RUNNER, ImageGenerateTool;
+var DEFAULT_DIFFUSERS_IMAGE_MODEL, DEFAULT_OLLAMA_IMAGE_MODEL, SECONDARY_FLUX_DEV_MODEL, SECONDARY_FLUX_DEV_MIRROR_MODEL, SECONDARY_FLUX_DEV_COMFY_MODEL, SECONDARY_FLUX_FILL_MODEL, SECONDARY_FLUX_FILL_FP8_MODEL, SECONDARY_FLUX2_MODEL, OFFICIAL_BFL_ORG, IMAGE_GENERATION_MODEL_REPLACEMENTS, DIFFUSERS_PYTHON_PACKAGES, SDCPP_PYTHON_PACKAGES, IMAGE_GENERATION_MODEL_PRESETS, IMAGE_GENERATION_QUALITY_LADDER, OLLAMA_IMAGE_MODELS, DIFFUSERS_RUNNER, SDCPP_RUNNER, ImageGenerateTool;
 var init_image_generate = __esm({
   "packages/execution/dist/tools/image-generate.js"() {
     "use strict";
     init_venv_paths();
     DEFAULT_DIFFUSERS_IMAGE_MODEL = "stabilityai/sdxl-turbo";
     DEFAULT_OLLAMA_IMAGE_MODEL = "x/z-image-turbo";
+    SECONDARY_FLUX_DEV_MODEL = "lllyasviel/flux1-dev-bnb-nf4";
+    SECONDARY_FLUX_DEV_MIRROR_MODEL = "ChuckMcSneed/FLUX.1-dev";
+    SECONDARY_FLUX_DEV_COMFY_MODEL = "Comfy-Org/flux1-dev";
+    SECONDARY_FLUX_FILL_MODEL = "diffusers/FLUX.1-Fill-dev-nf4";
+    SECONDARY_FLUX_FILL_FP8_MODEL = "boricuapab/flux1-fill-dev-fp8";
+    SECONDARY_FLUX2_MODEL = "x/flux2-klein";
+    OFFICIAL_BFL_ORG = "black-forest-labs";
+    IMAGE_GENERATION_MODEL_REPLACEMENTS = /* @__PURE__ */ new Map([
+      [officialBflModel("FLUX.1-dev"), SECONDARY_FLUX_DEV_MODEL],
+      [officialBflModel("FLUX.1-dev-FP8"), SECONDARY_FLUX_DEV_MODEL],
+      [officialBflModel("FLUX.1-Krea-dev"), SECONDARY_FLUX_DEV_COMFY_MODEL],
+      [officialBflModel("FLUX.1-schnell"), SECONDARY_FLUX_DEV_MODEL],
+      [officialBflModel("FLUX.1-Fill-dev"), SECONDARY_FLUX_FILL_MODEL],
+      [officialBflModel("FLUX.1-Kontext-dev"), SECONDARY_FLUX_DEV_MODEL],
+      [officialBflModel("FLUX.1-Kontext-dev-NVFP4"), SECONDARY_FLUX_DEV_MODEL],
+      [officialBflModel("FLUX.2-klein-4B"), SECONDARY_FLUX2_MODEL],
+      [officialBflModel("FLUX.2-klein-4b-fp8"), SECONDARY_FLUX2_MODEL],
+      [officialBflModel("FLUX.2-klein-4b-nvfp4"), SECONDARY_FLUX2_MODEL],
+      [officialBflModel("FLUX.2-klein-base-4b-fp8"), SECONDARY_FLUX2_MODEL],
+      [officialBflModel("FLUX.2-klein-base-4b-nvfp4"), SECONDARY_FLUX2_MODEL]
+    ]);
     DIFFUSERS_PYTHON_PACKAGES = [
       "torch",
       "torchvision",
@@ -250785,7 +250834,7 @@ var init_image_generate = __esm({
         note: "Modern fast image model; good default when available in Ollama."
       },
       {
-        id: "x/flux2-klein",
+        id: SECONDARY_FLUX2_MODEL,
         label: "FLUX.2 Klein",
         backend: "ollama",
         install: "ollama pull x/flux2-klein",
@@ -250801,21 +250850,92 @@ var init_image_generate = __esm({
         note: "Compact FLUX-family Ollama path for interactive local generation."
       },
       {
-        id: "black-forest-labs/FLUX.1-dev",
-        label: "FLUX.1 dev",
+        id: SECONDARY_FLUX_DEV_MODEL,
+        label: "FLUX.1 dev BNB NF4",
         backend: "diffusers",
-        install: 'python3 .omnius/image-gen/diffusers_text2image.py --model black-forest-labs/FLUX.1-dev --steps 28 --guidance 3.5 --width 1024 --height 1024 --prompt "..." --output .omnius/images/out.png',
+        install: 'python3 .omnius/image-gen/diffusers_text2image.py --model lllyasviel/flux1-dev-bnb-nf4 --steps 28 --guidance 3.5 --width 1024 --height 1024 --prompt "..." --output .omnius/images/out.png',
         category: "Primary hyper-realistic baseline",
-        sizeClass: "12B rectified-flow transformer",
-        quality: "Top-tier open-weight photorealism, prompt adherence, texture detail, composition, and typography.",
+        sizeClass: "12B FLUX.1 dev NF4",
+        quality: "Traceable lower-memory FLUX.1 dev secondary-source route with strong photorealism and prompt adherence; avoids gated official BFL downloads.",
+        minVramGB: 12,
+        recommendedVramGB: 16,
+        deployment: "Best with BNB-aware Diffusers/Forge-style runtimes. Falls through cleanly if the current runner cannot load it.",
+        steps: 28,
+        guidance: 3.5,
+        width: 1024,
+        height: 1024,
+        note: "Primary serious-generation FLUX route using the secondary-source NF4 weights from the research package."
+      },
+      {
+        id: SECONDARY_FLUX_DEV_MIRROR_MODEL,
+        label: "FLUX.1 dev mirror",
+        backend: "diffusers",
+        install: 'python3 .omnius/image-gen/diffusers_text2image.py --model ChuckMcSneed/FLUX.1-dev --steps 28 --guidance 3.5 --width 1024 --height 1024 --prompt "..." --output .omnius/images/out.png',
+        category: "Traceable FLUX fallback",
+        sizeClass: "12B FLUX.1 dev mirror",
+        quality: "Full FLUX.1 dev mirror fallback from the secondary-source inventory; higher trust than anonymous mirrors, lower priority than quantized routes.",
         minVramGB: 24,
         recommendedVramGB: 48,
-        deployment: "Heavy. Best with Diffusers CPU offload, FP8/quantized variants, ComfyUI, multi-GPU, or cloud GPU workers.",
+        deployment: "Use after the NF4 route if full-weight mirror provenance and license terms are acceptable.",
+        steps: 28,
+        guidance: 3.5,
+        width: 1024,
+        height: 1024,
+        fallbackFor: [SECONDARY_FLUX_DEV_MODEL],
+        note: "Traceable mirror fallback for FLUX.1 dev from the research package."
+      },
+      {
+        id: SECONDARY_FLUX_DEV_COMFY_MODEL,
+        label: "FLUX.1 dev Comfy-Org",
+        backend: "diffusers",
+        install: 'python3 .omnius/image-gen/diffusers_text2image.py --model Comfy-Org/flux1-dev --steps 28 --guidance 3.5 --width 1024 --height 1024 --prompt "..." --output .omnius/images/out.png',
+        category: "Traceable FLUX fallback",
+        sizeClass: "12B FLUX.1 dev single-file/FP8 family",
+        quality: "ComfyUI-oriented secondary-source FLUX.1 dev route; useful when the quantized/default loader path is unavailable.",
+        minVramGB: 16,
+        recommendedVramGB: 24,
+        deployment: "Best in ComfyUI/single-file workflows; the generic Diffusers runner will try it and fall through if the repo layout is incompatible.",
+        steps: 28,
+        guidance: 3.5,
+        width: 1024,
+        height: 1024,
+        fallbackFor: [SECONDARY_FLUX_DEV_MODEL, SECONDARY_FLUX_DEV_MIRROR_MODEL],
+        note: "Traceable Comfy-Org FLUX.1 fallback from the research package."
+      },
+      {
+        id: SECONDARY_FLUX_FILL_MODEL,
+        label: "FLUX.1 Fill dev NF4",
+        backend: "diffusers",
+        install: 'python3 .omnius/image-gen/diffusers_text2image.py --model diffusers/FLUX.1-Fill-dev-nf4 --steps 28 --guidance 3.5 --width 1024 --height 1024 --prompt "..." --output .omnius/images/out.png',
+        category: "Traceable FLUX fallback",
+        sizeClass: "12B FLUX.1 Fill NF4",
+        quality: "Lower-memory fill/inpainting secondary-source route; listed for FLUX-family coverage rather than generic text-to-image priority.",
+        minVramGB: 12,
+        recommendedVramGB: 16,
+        deployment: "Best for fill/inpainting workflows. The generic text-to-image runner may not be the ideal loader, so it is not in the default ladder.",
         steps: 28,
         guidance: 3.5,
         width: 1024,
         height: 1024,
-        note: "Primary serious-generation baseline for maximum photorealism."
+        note: "Traceable NF4 fallback for FLUX.1 Fill dev from the research package."
+      },
+      {
+        id: SECONDARY_FLUX_FILL_FP8_MODEL,
+        label: "FLUX.1 Fill dev FP8",
+        backend: "diffusers",
+        install: 'python3 .omnius/image-gen/diffusers_text2image.py --model boricuapab/flux1-fill-dev-fp8 --steps 28 --guidance 3.5 --width 1024 --height 1024 --prompt "..." --output .omnius/images/out.png',
+        category: "Traceable FLUX fallback",
+        sizeClass: "12B FLUX.1 Fill FP8",
+        quality: "Traceable FP8 fill/inpainting conversion from the secondary-source inventory; gated status can vary.",
+        minVramGB: 12,
+        recommendedVramGB: 16,
+        deployment: "Best for fill/inpainting workflows when accepted access terms allow the repo.",
+        steps: 28,
+        guidance: 3.5,
+        width: 1024,
+        height: 1024,
+        fallbackFor: [SECONDARY_FLUX_FILL_MODEL],
+        note: "Traceable FP8 fallback for FLUX.1 Fill dev from the research package."
       },
       {
         id: "stabilityai/stable-diffusion-3.5-large",
@@ -250834,23 +250954,6 @@ var init_image_generate = __esm({
         height: 1024,
         note: "Primary serious-generation baseline for the Stable Diffusion ecosystem."
       },
-      {
-        id: "black-forest-labs/FLUX.1-schnell",
-        label: "FLUX.1 schnell",
-        backend: "diffusers",
-        install: 'python3 .omnius/image-gen/diffusers_text2image.py --model black-forest-labs/FLUX.1-schnell --steps 4 --guidance 0 --width 1024 --height 1024 --prompt "..." --output .omnius/images/out.png',
-        category: "Fast large-model iteration",
-        sizeClass: "12B rectified-flow transformer",
-        quality: "FLUX-style output with fewer steps; better for rapid iteration than absolute peak quality.",
-        minVramGB: 16,
-        recommendedVramGB: 24,
-        deployment: "Use for fast prompt iteration; verify current license terms before commercial use.",
-        steps: 4,
-        guidance: 0,
-        width: 1024,
-        height: 1024,
-        note: "Fast FLUX-family iteration path."
-      },
       {
         id: "stabilityai/stable-diffusion-3.5-large-turbo",
         label: "SD3.5 Large Turbo",
@@ -250901,22 +251004,6 @@ var init_image_generate = __esm({
         height: 1024,
         note: "Efficient modern large image model."
       },
-      {
-        id: "black-forest-labs/FLUX.2-klein-4B",
-        label: "FLUX.2 Klein 4B",
-        backend: "diffusers",
-        install: 'python3 .omnius/image-gen/diffusers_text2image.py --model black-forest-labs/FLUX.2-klein-4B --steps 8 --width 1024 --height 1024 --prompt "..." --output .omnius/images/out.png',
-        category: "Modern deployable",
-        sizeClass: "4B compact FLUX-family",
-        quality: "Bridge between practical deployment and modern FLUX-family visual quality.",
-        minVramGB: 12,
-        recommendedVramGB: 16,
-        deployment: "Better fit for consumer GPU experimentation than 8B-12B baselines.",
-        steps: 8,
-        width: 1024,
-        height: 1024,
-        note: "More deployable compact FLUX-family model."
-      },
       {
         id: "deepseek-ai/Janus-Pro-7B",
         label: "Janus-Pro-7B",
@@ -251070,13 +251157,12 @@ var init_image_generate = __esm({
       }
     ];
     IMAGE_GENERATION_QUALITY_LADDER = [
-      "black-forest-labs/FLUX.1-dev",
-      "stabilityai/stable-diffusion-3.5-large",
+      SECONDARY_FLUX_DEV_MODEL,
+      SECONDARY_FLUX_DEV_MIRROR_MODEL,
+      SECONDARY_FLUX_DEV_COMFY_MODEL,
       DEFAULT_OLLAMA_IMAGE_MODEL,
-      "black-forest-labs/FLUX.1-schnell",
-      "stabilityai/stable-diffusion-3.5-large-turbo",
       "Tongyi-MAI/Z-Image-Turbo",
-      "black-forest-labs/FLUX.2-klein-4B",
+      SECONDARY_FLUX2_MODEL,
       DEFAULT_DIFFUSERS_IMAGE_MODEL,
       "Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers",
       "SimianLuo/LCM_Dreamshaper_v7",
@@ -251265,7 +251351,7 @@ if __name__ == "__main__":
 `;
     ImageGenerateTool = class {
       name = "generate_image";
-      description = "Generate an image from a text prompt using a local image-generation backend. Supports Ollama image models (x/z-image-turbo, x/flux2-klein), Python Diffusers models (SDXL Turbo default, FLUX.1 dev, SD3.5 Large, Tiny-SD, LCM, Sana Sprint), and stable-diffusion.cpp local checkpoints/GGUF. When fallback is enabled, auto generation tries ranked high-quality candidates first and falls back to smaller models if setup, download, or generation fails. Saves a PNG under .omnius/images and returns the file path.";
+      description = "Generate an image from a text prompt using a local image-generation backend. Supports Ollama image models (x/z-image-turbo, x/flux2-klein), Python Diffusers models (SDXL Turbo default, FLUX.1 dev, SD3.5 Large, Tiny-SD, LCM, Sana Sprint), and stable-diffusion.cpp local checkpoints/GGUF. When fallback is enabled, auto generation tries ranked high-quality candidates first, including official/traceable FLUX fallbacks for Black Forest Labs models, and then falls back to smaller models if setup, download, or generation fails. Saves a PNG under .omnius/images and returns the file path.";
       parameters = {
         type: "object",
         properties: {
@@ -251929,7 +252015,7 @@ ${errText.slice(0, 800)}`,
 });
 
 // packages/execution/dist/tools/audio-generate.js
-import { spawn as spawn10 } from "node:child_process";
+import { execFileSync as execFileSync2, spawn as spawn10 } from "node:child_process";
 import { existsSync as existsSync24, readdirSync as readdirSync10, statSync as statSync9 } from "node:fs";
 import { chmod as chmod4, mkdir as mkdir12, writeFile as writeFile17 } from "node:fs/promises";
 import { join as join37 } from "node:path";
@@ -251953,6 +252039,56 @@ function backendPackages(backend) {
     return TANGOFLUX_PACKAGES;
   return DIFFUSERS_AUDIO_PACKAGES;
 }
+function detectLegacyCudaComputeCapability() {
+  try {
+    const out = execFileSync2("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 {
+    return null;
+  }
+}
+function isLegacyCudaCapability(major, minor) {
+  return major < 7 || major === 7 && minor < 5;
+}
+function torchInstallPlan(forceLegacyCuda = false) {
+  if (process.env["OMNIUS_AUDIO_TORCH_INDEX_URL"]) {
+    return {
+      args: ["torch", "torchaudio", "--index-url", process.env["OMNIUS_AUDIO_TORCH_INDEX_URL"]],
+      description: `env override ${process.env["OMNIUS_AUDIO_TORCH_INDEX_URL"]}`
+    };
+  }
+  if (forceLegacyCuda) {
+    return {
+      args: ["torch==2.3.1", "torchaudio==2.3.1", "--index-url", "https://download.pytorch.org/whl/cu118"],
+      description: "runtime-detected legacy CUDA GPU; using PyTorch 2.3.1 cu118 to avoid cuDNN 9 incompatibility"
+    };
+  }
+  if (process.platform === "linux" && process.arch === "x64") {
+    const gpu = detectLegacyCudaComputeCapability();
+    if (gpu && isLegacyCudaCapability(gpu.major, gpu.minor)) {
+      return {
+        args: ["torch==2.3.1", "torchaudio==2.3.1", "--index-url", "https://download.pytorch.org/whl/cu118"],
+        description: `CUDA legacy GPU SM ${gpu.major}.${gpu.minor}${gpu.name ? ` ${gpu.name}` : ""}; using PyTorch 2.3.1 cu118 to avoid cuDNN 9 incompatibility`
+      };
+    }
+  }
+  return { args: ["torch", "torchaudio"], description: "default PyTorch wheel selection" };
+}
+function withoutTorchPackages(packages) {
+  return packages.filter((pkg) => pkg !== "torch" && pkg !== "torchaudio");
+}
 function backendImportCheck(backend) {
   if (backend === "transformers")
     return "import torch, torchaudio, transformers, scipy\nfrom transformers import AutoProcessor, MusicgenForConditionalGeneration\n";
@@ -252151,6 +252287,69 @@ async function pythonCanImport2(command, code8, repoRoot, env2) {
 async function pythonImportResult(command, code8, repoRoot, env2) {
   return await runProcess3(command, ["-c", code8], { cwd: repoRoot, timeoutMs: 6e4, env: env2 });
 }
+async function torchRuntimeCompatibilityResult(command, repoRoot, env2) {
+  const code8 = [
+    "import json, sys",
+    "import torch",
+    "payload={'torch': getattr(torch, '__version__', '?'), 'cuda_available': bool(torch.cuda.is_available())}",
+    "if torch.cuda.is_available():",
+    "    cap=torch.cuda.get_device_capability(0)",
+    "    cudnn=torch.backends.cudnn.version() or 0",
+    "    payload.update({'capability': list(cap), 'cudnn': int(cudnn), 'device': torch.cuda.get_device_name(0)})",
+    "    if int(cudnn) >= 90000 and tuple(cap) < (7, 5):",
+    "        print(json.dumps(payload))",
+    "        raise SystemExit(42)",
+    "print(json.dumps(payload))"
+  ].join("\n");
+  return await runProcess3(command, ["-c", code8], { cwd: repoRoot, timeoutMs: 6e4, env: env2 });
+}
+async function repairTorchRuntime(command, repoRoot, env2, forceLegacyCuda = false, onProgress) {
+  const plan = torchInstallPlan(forceLegacyCuda);
+  onProgress?.({ stage: "setup", message: `Installing PyTorch runtime: ${plan.description}` });
+  const result = await runProcess3(command, [
+    "-m",
+    "pip",
+    "install",
+    "--progress-bar",
+    "on",
+    "--prefer-binary",
+    "--force-reinstall",
+    ...plan.args
+  ], {
+    cwd: repoRoot,
+    timeoutMs: 18e5,
+    env: env2,
+    progressLabel: `Installing PyTorch runtime (${plan.description})`,
+    onProgress
+  });
+  if (result.code !== 0) {
+    throw new Error(`Failed to install compatible PyTorch runtime (${plan.description}).
+${trimProcessText2(result.stderr || result.stdout)}`);
+  }
+}
+async function ensureCompatibleTorchRuntime(command, repoRoot, env2, onProgress) {
+  const existing = await torchRuntimeCompatibilityResult(command, repoRoot, env2);
+  if (existing.code === 0)
+    return;
+  if (existing.code === 42) {
+    await repairTorchRuntime(command, repoRoot, env2, true, onProgress);
+  } else {
+    await repairTorchRuntime(command, repoRoot, env2, false, onProgress);
+  }
+  const installed = await torchRuntimeCompatibilityResult(command, repoRoot, env2);
+  if (installed.code === 0)
+    return;
+  if (installed.code === 42) {
+    await repairTorchRuntime(command, repoRoot, env2, true, onProgress);
+    const repaired = await torchRuntimeCompatibilityResult(command, repoRoot, env2);
+    if (repaired.code === 0)
+      return;
+    throw new Error(`Audio-generation PyTorch runtime remains incompatible after cu118 repair.
+${trimProcessText2(repaired.stderr || repaired.stdout)}`);
+  }
+  throw new Error(`Audio-generation PyTorch runtime could not be prepared.
+${trimProcessText2(installed.stderr || installed.stdout)}`);
+}
 function formatAudioSetupFailure(backend, text) {
   const body = trimProcessText2(text);
   const lowered = text.toLowerCase();
@@ -252161,6 +252360,9 @@ 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.");
+  }
   return [body, ...notes2.map((note) => `
 ${note}`)].filter(Boolean).join("");
 }
@@ -252189,9 +252391,13 @@ ${trimProcessText2(created.stderr || created.stdout)}`);
     }
   }
   if (await pythonCanImport2(command, backendImportCheck(backend), repoRoot, pythonEnv)) {
-    return { command, env: pythonEnv };
+    await ensureCompatibleTorchRuntime(command, repoRoot, pythonEnv, onProgress);
+    if (await pythonCanImport2(command, backendImportCheck(backend), repoRoot, pythonEnv)) {
+      return { command, env: pythonEnv };
+    }
   }
   const packages = backendPackages(backend);
+  await ensureCompatibleTorchRuntime(command, repoRoot, pythonEnv, onProgress);
   onProgress?.({ stage: "setup", message: `Installing ${backend} audio-generation Python packages` });
   const pipArgs = [
     "-m",
@@ -252203,7 +252409,7 @@ ${trimProcessText2(created.stderr || created.stdout)}`);
     ...backend === "audiocraft" ? ["--only-binary", "av"] : [],
     "-U",
     "pip",
-    ...packages
+    ...withoutTorchPackages(packages)
   ];
   const pip = await runProcess3(command, pipArgs, {
     cwd: repoRoot,
@@ -252220,6 +252426,12 @@ ${formatAudioSetupFailure(backend, pip.stderr || pip.stdout)}`);
   if (importCheck.code !== 0) {
     throw new Error(`Audio-generation Python environment at ${venvDir} was created, but required ${backend} imports still fail.
 ${formatAudioSetupFailure(backend, importCheck.stderr || importCheck.stdout)}`);
+  }
+  await ensureCompatibleTorchRuntime(command, repoRoot, pythonEnv, onProgress);
+  if (!await pythonCanImport2(command, backendImportCheck(backend), repoRoot, pythonEnv)) {
+    const retry = await pythonImportResult(command, backendImportCheck(backend), repoRoot, pythonEnv);
+    throw new Error(`Audio-generation Python environment at ${venvDir} lost required ${backend} imports after PyTorch repair.
+${formatAudioSetupFailure(backend, retry.stderr || retry.stdout)}`);
   }
   return { command, env: pythonEnv };
 }
@@ -252911,6 +253123,10 @@ def _snapshot_model(repo_id):
 def _device():
     import torch
     if torch.cuda.is_available():
+        cap = torch.cuda.get_device_capability(0)
+        cudnn = torch.backends.cudnn.version() or 0
+        if int(cudnn) >= 90000 and tuple(cap) < (7, 5):
+            raise RuntimeError(f"PyTorch cuDNN {cudnn} is incompatible with CUDA device {torch.cuda.get_device_name(0)} SM {cap[0]}.{cap[1]}; recreate the audio venv or let Omnius repair it with a cu118-compatible Torch wheel")
         return "cuda"
     if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
         return "mps"
@@ -253103,6 +253319,10 @@ def _snapshot_model(repo_id):
 def _device():
     import torch
     if torch.cuda.is_available():
+        cap = torch.cuda.get_device_capability(0)
+        cudnn = torch.backends.cudnn.version() or 0
+        if int(cudnn) >= 90000 and tuple(cap) < (7, 5):
+            raise RuntimeError(f"PyTorch cuDNN {cudnn} is incompatible with CUDA device {torch.cuda.get_device_name(0)} SM {cap[0]}.{cap[1]}; recreate the audio venv or let Omnius repair it with a cu118-compatible Torch wheel")
         return "cuda"
     if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
         return "mps"
@@ -477268,7 +477488,7 @@ var require_path_browserify = __commonJS({
           return path11.slice(start2, end);
         }
       },
-      extname: function extname16(path11) {
+      extname: function extname17(path11) {
         assertPath(path11);
         var startDot = -1;
         var startPart = 0;
@@ -507429,22 +507649,22 @@ Saved to: ${tempFile}`,
 });
 
 // packages/execution/dist/tools/audio-playback.js
-import { execFileSync as execFileSync2, execSync as execSync29, spawn as spawn16 } from "node:child_process";
+import { execFileSync as execFileSync3, execSync as execSync29, spawn as spawn16 } from "node:child_process";
 import { copyFileSync as copyFileSync2, existsSync as existsSync40, statSync as statSync18, writeFileSync as writeFileSync16, mkdirSync as mkdirSync16, readdirSync as readdirSync14 } from "node:fs";
 import { basename as basename12, extname as extname10, isAbsolute, join as join58 } from "node:path";
 import { homedir as homedir14, tmpdir as tmpdir11 } from "node:os";
 function hasCommand3(command) {
   try {
     if (process.platform === "win32") {
-      execFileSync2("where", [command], { stdio: "ignore", timeout: 2e3 });
+      execFileSync3("where", [command], { stdio: "ignore", timeout: 2e3 });
     } else {
-      execFileSync2("command", ["-v", command], { stdio: "ignore", timeout: 2e3 });
+      execFileSync3("command", ["-v", command], { stdio: "ignore", timeout: 2e3 });
     }
     return true;
   } catch {
     if (process.platform !== "win32") {
       try {
-        execFileSync2("which", [command], { stdio: "ignore", timeout: 2e3 });
+        execFileSync3("which", [command], { stdio: "ignore", timeout: 2e3 });
         return true;
       } catch {
         return false;
@@ -507499,7 +507719,7 @@ function playSoundFile(file, opts = {}) {
     };
   }
   try {
-    execFileSync2(command.command, command.args, { timeout: opts.timeoutMs ?? 3e5, stdio: "pipe" });
+    execFileSync3(command.command, command.args, { timeout: opts.timeoutMs ?? 3e5, stdio: "pipe" });
     return { ok: true, player: command.label };
   } catch (err) {
     return { ok: false, error: `Playback via ${command.label} failed: ${err instanceof Error ? err.message.slice(0, 300) : String(err).slice(0, 300)}` };
@@ -507646,13 +507866,13 @@ function ensureSupertonicInstalled() {
     const py = findPython32();
     if (!py)
       throw new Error("python3 is required to set up Supertonic TTS.");
-    execFileSync2(py, ["-m", "venv", join58(voiceDir(), "supertonic3-venv")], { stdio: "pipe", timeout: 18e4 });
+    execFileSync3(py, ["-m", "venv", join58(voiceDir(), "supertonic3-venv")], { stdio: "pipe", timeout: 18e4 });
   }
   try {
-    execFileSync2(venvPy, ["-c", "import supertonic"], { stdio: "pipe", timeout: 1e4 });
+    execFileSync3(venvPy, ["-c", "import supertonic"], { stdio: "pipe", timeout: 1e4 });
   } catch {
-    execFileSync2(venvPy, ["-m", "pip", "install", "--quiet", "--upgrade", "pip"], { stdio: "pipe", timeout: 12e4 });
-    execFileSync2(venvPy, ["-m", "pip", "install", "--quiet", "supertonic"], { stdio: "pipe", timeout: 6e5 });
+    execFileSync3(venvPy, ["-m", "pip", "install", "--quiet", "--upgrade", "pip"], { stdio: "pipe", timeout: 12e4 });
+    execFileSync3(venvPy, ["-m", "pip", "install", "--quiet", "supertonic"], { stdio: "pipe", timeout: 6e5 });
   }
   mkdirSync16(voiceDir(), { recursive: true });
   writeFileSync16(supertonicInferScript(), SUPERTONIC_INFER_PY, "utf-8");
@@ -507667,19 +507887,19 @@ function ensureMlxInstalled() {
     const py = findPython32();
     if (!py)
       throw new Error("python3 is required to set up MLX Audio.");
-    execFileSync2(py, ["-m", "venv", join58(voiceDir(), "mlx-venv")], { stdio: "pipe", timeout: 18e4 });
+    execFileSync3(py, ["-m", "venv", join58(voiceDir(), "mlx-venv")], { stdio: "pipe", timeout: 18e4 });
   }
   try {
-    execFileSync2(venvPy, ["-c", "import mlx_audio"], { stdio: "pipe", timeout: 1e4 });
+    execFileSync3(venvPy, ["-c", "import mlx_audio"], { stdio: "pipe", timeout: 1e4 });
   } catch {
-    execFileSync2(venvPy, ["-m", "pip", "install", "--quiet", "--upgrade", "pip"], { stdio: "pipe", timeout: 12e4 });
-    execFileSync2(venvPy, ["-m", "pip", "install", "--quiet", "mlx-audio"], { stdio: "pipe", timeout: 6e5 });
+    execFileSync3(venvPy, ["-m", "pip", "install", "--quiet", "--upgrade", "pip"], { stdio: "pipe", timeout: 12e4 });
+    execFileSync3(venvPy, ["-m", "pip", "install", "--quiet", "mlx-audio"], { stdio: "pipe", timeout: 6e5 });
   }
   return venvPy;
 }
 function pythonCanImportLuxTts(venvPy) {
   try {
-    execFileSync2(venvPy, [
+    execFileSync3(venvPy, [
       "-c",
       "import sys, os; sys.path.insert(0, os.environ['LUXTTS_REPO_PATH']); from zipvoice.luxvoice import LuxTTS; print('ok')"
     ], {
@@ -507693,7 +507913,7 @@ function pythonCanImportLuxTts(venvPy) {
   }
 }
 function pipInstall(venvPy, packages, timeout2 = 9e5) {
-  execFileSync2(venvPy, ["-m", "pip", "install", "--prefer-binary", ...packages], {
+  execFileSync3(venvPy, ["-m", "pip", "install", "--prefer-binary", ...packages], {
     stdio: "pipe",
     timeout: timeout2,
     env: process.env
@@ -507711,9 +507931,9 @@ function ensureLuxttsInstalled() {
   if (!py)
     throw new Error("python3 is required to set up LuxTTS voice cloning.");
   if (!existsSync40(venvPy)) {
-    execFileSync2(py, ["-m", "venv", luxttsVenvDir()], { stdio: "pipe", timeout: 18e4 });
+    execFileSync3(py, ["-m", "venv", luxttsVenvDir()], { stdio: "pipe", timeout: 18e4 });
   }
-  execFileSync2(venvPy, ["-m", "pip", "install", "--upgrade", "pip", "wheel", "setuptools<81"], {
+  execFileSync3(venvPy, ["-m", "pip", "install", "--upgrade", "pip", "wheel", "setuptools<81"], {
     stdio: "pipe",
     timeout: 3e5
   });
@@ -507721,7 +507941,7 @@ function ensureLuxttsInstalled() {
   if (!existsSync40(join58(repoDir, "zipvoice", "luxvoice.py"))) {
     if (!hasCommand3("git"))
       throw new Error("git is required to set up LuxTTS voice cloning.");
-    execFileSync2("git", ["clone", "--depth", "1", "https://github.com/ysharma3501/LuxTTS.git", repoDir], {
+    execFileSync3("git", ["clone", "--depth", "1", "https://github.com/ysharma3501/LuxTTS.git", repoDir], {
       stdio: "pipe",
       timeout: 3e5
     });
@@ -507761,10 +507981,10 @@ function ensurePiperInstalled() {
     if (!py)
       throw new Error("python3 is required to set up Piper TTS.");
     mkdirSync16(voiceDir(), { recursive: true });
-    execFileSync2(py, ["-m", "venv", piperVenvDir()], { stdio: "pipe", timeout: 18e4 });
+    execFileSync3(py, ["-m", "venv", piperVenvDir()], { stdio: "pipe", timeout: 18e4 });
     const venvPy = process.platform === "win32" ? join58(piperVenvDir(), "Scripts", "python.exe") : join58(piperVenvDir(), "bin", "python3");
-    execFileSync2(venvPy, ["-m", "pip", "install", "--quiet", "--upgrade", "pip"], { stdio: "pipe", timeout: 12e4 });
-    execFileSync2(venvPy, ["-m", "pip", "install", "--quiet", "piper-tts"], { stdio: "pipe", timeout: 6e5 });
+    execFileSync3(venvPy, ["-m", "pip", "install", "--quiet", "--upgrade", "pip"], { stdio: "pipe", timeout: 12e4 });
+    execFileSync3(venvPy, ["-m", "pip", "install", "--quiet", "piper-tts"], { stdio: "pipe", timeout: 6e5 });
   }
   if (!existsSync40(bin)) {
     throw new Error("Piper TTS installed but the piper executable was not found in the managed venv.");
@@ -508294,7 +508514,7 @@ ${tried.map((line) => `- ${line}`).join("\n")}`,
           "d=(np.clip(wav.cpu().numpy().squeeze(), -1, 1)*32767).astype(np.int16)",
           "f=wave.open(args['output'], 'wb'); f.setnchannels(1); f.setsampwidth(2); f.setframerate(48000); f.writeframes(d.tobytes()); f.close()"
         ].join("; ");
-        execFileSync2(venvPy, ["-c", pyScript, JSON.stringify({ text, output: outputPath2, clone_ref: cloneRef, repo: repoDir, speed })], {
+        execFileSync3(venvPy, ["-c", pyScript, JSON.stringify({ text, output: outputPath2, clone_ref: cloneRef, repo: repoDir, speed })], {
           stdio: "pipe",
           timeout: 12e4,
           env: { ...process.env, LUXTTS_REPO_PATH: repoDir }
@@ -508307,7 +508527,7 @@ ${tried.map((line) => `- ${line}`).join("\n")}`,
         const lang = typeof args["lang"] === "string" ? args["lang"] : "en";
         const speed = numberArg3(args["speed"], 1.05);
         const totalStep = Math.round(numberArg3(args["total_step"], 8));
-        const stdout = execFileSync2(venvPy, [supertonicInferScript()], {
+        const stdout = execFileSync3(venvPy, [supertonicInferScript()], {
           input: JSON.stringify({ text, output_path: outputPath2, voice_name: voice, lang, speed, total_step: totalStep }),
           encoding: "utf8",
           stdio: ["pipe", "pipe", "pipe"],
@@ -508330,7 +508550,7 @@ ${tried.map((line) => `- ${line}`).join("\n")}`,
           "args=json.loads(sys.argv[1])",
           "tts_gen.main(['--model', args['model'], '--text', args['text'], '--voice', args['voice'], '--lang_code', args['lang'], '--audio_path', args['output']])"
         ].join("; ");
-        execFileSync2(py, ["-c", pyScript, JSON.stringify({ text, model, voice, lang, output: outputPath2 })], {
+        execFileSync3(py, ["-c", pyScript, JSON.stringify({ text, model, voice, lang, output: outputPath2 })], {
           stdio: "pipe",
           timeout: 18e4,
           cwd: tmpdir11()
@@ -508351,7 +508571,7 @@ ${tried.map((line) => `- ${line}`).join("\n")}`,
         } else {
           throw new Error(`${requireModel ? "Raw ONNX" : "Piper"} TTS requires model=<path.onnx> or voice=<path.onnx>.`);
         }
-        execFileSync2(piper, argv, { input: text, stdio: ["pipe", "pipe", "pipe"], timeout: 12e4 });
+        execFileSync3(piper, argv, { input: text, stdio: ["pipe", "pipe", "pipe"], timeout: 12e4 });
         return summary;
       }
       synthesizeEspeak(text, outputPath2, args) {
@@ -508359,7 +508579,7 @@ ${tried.map((line) => `- ${line}`).join("\n")}`,
           throw new Error("Local fallback TTS command not found.");
         const voice = typeof args["voice"] === "string" ? args["voice"] : "en";
         const speed = Math.round(numberArg3(args["speed"], 160));
-        execFileSync2("espeak-ng", ["-v", voice, "-s", String(speed), "-w", outputPath2, text], {
+        execFileSync3("espeak-ng", ["-v", voice, "-s", String(speed), "-w", outputPath2, text], {
           stdio: "pipe",
           timeout: 6e4
         });
@@ -575505,7 +575725,7 @@ __export(image_ascii_preview_exports, {
   extractSavedImagePath: () => extractSavedImagePath,
   formatImageAsciiContext: () => formatImageAsciiContext
 });
-import { execFileSync as execFileSync3 } from "node:child_process";
+import { execFileSync as execFileSync4 } from "node:child_process";
 import { createRequire as createRequire5 } from "node:module";
 import { existsSync as existsSync94, readFileSync as readFileSync75, statSync as statSync32 } from "node:fs";
 import { resolve as resolve37 } from "node:path";
@@ -575642,7 +575862,7 @@ function convertWithFfmpeg(imagePath, width, height, timeoutMs) {
       `scale=${width}:${height}`,
       "format=gray"
     ].join(",");
-    const raw = execFileSync3(
+    const raw = execFileSync4(
       "ffmpeg",
       [
         "-hide_banner",
@@ -596827,6 +597047,17 @@ var init_tool_policy = __esm({
       "todo_write",
       "web_search",
       "web_fetch",
+      "image_read",
+      "ocr",
+      "ocr_image_advanced",
+      "ocr_pdf",
+      "pdf_to_text",
+      "vision",
+      "transcribe_file",
+      "video_understand",
+      "audio_analyze",
+      "explore_tools",
+      "telegram_media_recent",
       "generate_image",
       "generate_audio",
       "generate_tts",
@@ -596843,6 +597074,17 @@ var init_tool_policy = __esm({
       "web_search",
       "web_fetch",
       "web_crawl",
+      "image_read",
+      "ocr",
+      "ocr_image_advanced",
+      "ocr_pdf",
+      "pdf_to_text",
+      "vision",
+      "transcribe_file",
+      "video_understand",
+      "audio_analyze",
+      "explore_tools",
+      "telegram_media_recent",
       "generate_image",
       "generate_audio",
       "generate_tts",
@@ -596934,6 +597176,7 @@ function buildTelegramCreativeTools(repoRoot, chatId, backendUrl, imageDefaults
   ensureManifest(root);
   return [
     scopedTool(new FileWriteTool(root), root, "create"),
+    scopedTool(new FileEditTool(root), root, "edit"),
     scopedTool(new StructuredFileTool(root), root, "create"),
     scopedTool(new ImageGenerateTool(root, backendUrl, imageDefaults), root, "generate"),
     scopedTool(new AudioGenerateTool(root, audioDefaults), root, "generate"),
@@ -597000,6 +597243,7 @@ function scopedTool(base3, root, mode) {
         }
       }
       const pathKey = PATH_KEYS.find((key) => typeof next[key] === "string" && String(next[key]).trim());
+      let restoredEditPath = null;
       if (pathKey) {
         const guarded = guardPath(rootAbs, String(next[pathKey]));
         if (!guarded.ok) return denied(guarded.error);
@@ -597010,6 +597254,14 @@ function scopedTool(base3, root, mode) {
         if (mode === "edit" && !manifestHas(rootAbs, guarded.path.rel)) {
           return denied(`Public Telegram creative tools can only edit files created in this chat workspace: ${guarded.path.rel}`);
         }
+        if (mode === "edit" && !existsSync104(guarded.path.abs)) {
+          const materialized = materializeTelegramCreativeArtifactForSend(rootAbs, guarded.path.rel);
+          if (!materialized.ok) return denied(materialized.error);
+          mkdirSync59(dirname33(guarded.path.abs), { recursive: true });
+          writeFileSync55(guarded.path.abs, readFileSync85(materialized.path));
+          materialized.cleanup?.();
+          restoredEditPath = guarded.path.abs;
+        }
         if (mode === "create" && existsSync104(guarded.path.abs) && !manifestHas(rootAbs, guarded.path.rel)) {
           return denied(`Refusing to overwrite a file that is not owned by this chat workspace manifest: ${guarded.path.rel}`);
         }
@@ -597018,9 +597270,14 @@ function scopedTool(base3, root, mode) {
         return denied(`${base3.name} requires a path inside the public creative workspace.`);
       }
       const result = await base3.execute(next);
-      if (result.success && mode === "create" && pathKey) {
+      if (result.success && (mode === "create" || mode === "edit") && pathKey) {
         const guarded = guardPath(rootAbs, String(args[pathKey]));
         if (guarded.ok) rememberCreated(rootAbs, guarded.path.abs);
+      } else if (restoredEditPath) {
+        try {
+          unlinkSync19(restoredEditPath);
+        } catch {
+        }
       }
       if (result.success) {
         for (const path11 of collectGeneratedArtifactPathsFromText(result.output, rootAbs)) {
@@ -597417,12 +597674,12 @@ __export(vision_ingress_exports, {
   queryVisionModel: () => queryVisionModel,
   runVisionIngress: () => runVisionIngress
 });
-import { execFileSync as execFileSync4 } from "node:child_process";
+import { execFileSync as execFileSync5 } from "node:child_process";
 import { existsSync as existsSync105, readFileSync as readFileSync86, unlinkSync as unlinkSync20 } from "node:fs";
 import { join as join120 } from "node:path";
 function isTesseractAvailable() {
   try {
-    execFileSync4("tesseract", ["--version"], { stdio: "ignore", timeout: 3e3 });
+    execFileSync5("tesseract", ["--version"], { stdio: "ignore", timeout: 3e3 });
     return true;
   } catch {
     return false;
@@ -597463,7 +597720,7 @@ function advancedOcr(imagePath) {
   for (const psm of psmModes) {
     const outFile = `${tmpBase}_psm${psm}`;
     try {
-      execFileSync4("tesseract", [
+      execFileSync5("tesseract", [
         imagePath,
         outFile,
         "--psm",
@@ -597562,7 +597819,7 @@ var init_vision_ingress = __esm({
 
 // packages/cli/src/tui/telegram-bridge.ts
 import { mkdirSync as mkdirSync60, existsSync as existsSync106, unlinkSync as unlinkSync21, readdirSync as readdirSync36, statSync as statSync36, readFileSync as readFileSync87, writeFileSync as writeFileSync57 } from "node:fs";
-import { join as join121, resolve as resolve39, basename as basename23, relative as relative13, isAbsolute as isAbsolute7 } from "node:path";
+import { join as join121, resolve as resolve39, basename as basename23, relative as relative13, isAbsolute as isAbsolute7, extname as extname15 } from "node:path";
 import { writeFile as writeFileAsync } from "node:fs/promises";
 import { createHash as createHash19, randomInt } from "node:crypto";
 function parseTelegramInteractionDecision(text, forcedRoute, options2 = {}) {
@@ -597760,6 +598017,19 @@ function summarizeTelegramMessageAttachments(msg) {
       parts.push(`caption: ${truncateTelegramContextLine(msg.media.caption, 180)}`);
     }
   }
+  if (msg.replyToMedia) {
+    const details = [
+      msg.replyToMedia.type,
+      msg.replyToMedia.mimeType,
+      msg.replyToMedia.fileName,
+      msg.replyToMedia.duration ? `${msg.replyToMedia.duration}s` : "",
+      msg.replyToMedia.fileSize ? `${msg.replyToMedia.fileSize} bytes` : ""
+    ].filter(Boolean).join(", ");
+    parts.push(`replied-to media: ${details}`);
+    if (msg.replyToMedia.caption) {
+      parts.push(`replied-to caption: ${truncateTelegramContextLine(msg.replyToMedia.caption, 180)}`);
+    }
+  }
   if (msg.poll) {
     parts.push(`poll: ${truncateTelegramContextLine(msg.poll.question, 180)}`);
   }
@@ -598133,6 +598403,25 @@ function telegramImageMime(media) {
   if (ext === ".tif" || ext === ".tiff") return "image/tiff";
   return "image/jpeg";
 }
+function telegramCachedMediaIsImage(entry) {
+  if (entry.mediaType === "photo") return true;
+  if (entry.mimeType?.toLowerCase().startsWith("image/")) return true;
+  return TELEGRAM_IMAGE_EXTENSIONS.has(extname15(entry.localPath).toLowerCase());
+}
+function telegramCachedMediaIsPdf(entry) {
+  if (entry.mimeType?.toLowerCase() === "application/pdf") return true;
+  return extname15(entry.localPath).toLowerCase() === ".pdf";
+}
+function telegramCachedMediaIsAudio(entry) {
+  if (entry.mediaType === "audio" || entry.mediaType === "voice") return true;
+  if (entry.mimeType?.toLowerCase().startsWith("audio/")) return true;
+  return [".wav", ".mp3", ".flac", ".aac", ".m4a", ".ogg", ".opus"].includes(extname15(entry.localPath).toLowerCase());
+}
+function telegramCachedMediaIsVideo(entry) {
+  if (entry.mediaType === "video" || entry.mediaType === "video_note" || entry.mediaType === "live_photo") return true;
+  if (entry.mimeType?.toLowerCase().startsWith("video/")) return true;
+  return [".mp4", ".mkv", ".avi", ".mov", ".webm"].includes(extname15(entry.localPath).toLowerCase());
+}
 function isPathInside(root, path11) {
   const rel = relative13(resolve39(root), resolve39(path11));
   return rel === "" || Boolean(rel) && !rel.startsWith("..") && !isAbsolute7(rel);
@@ -598166,6 +598455,10 @@ function normalizeTelegramUpdate(update2) {
   const username = message2.from?.username ?? message2.sender_chat?.username ?? "";
   const chatType = message2.chat?.type ?? "private";
   const media = normalizeTelegramMedia(message2);
+  const replyTo = message2.reply_to_message && typeof message2.reply_to_message === "object" ? message2.reply_to_message : void 0;
+  const replyToMedia = replyTo ? normalizeTelegramMedia(replyTo) : void 0;
+  const replyToPoll = replyTo ? normalizeTelegramPoll(replyTo.poll) : void 0;
+  const replyToText = replyTo ? replyTo.text || replyTo.caption || (replyToPoll ? formatTelegramPollSummary(replyToPoll) : "") : "";
   const poll = normalizeTelegramPoll(message2.poll);
   const livePhoto = normalizeTelegramLivePhoto(message2.live_photo);
   const text = message2.text || message2.caption || (poll ? formatTelegramPollSummary(poll) : "");
@@ -598180,6 +598473,8 @@ function normalizeTelegramUpdate(update2) {
     chatType,
     chatTitle: message2.chat?.title,
     media,
+    replyToMedia,
+    replyToText: replyToText || void 0,
     poll,
     livePhoto,
     guestQueryId: typeof message2.guest_query_id === "string" ? message2.guest_query_id : void 0,
@@ -598188,9 +598483,9 @@ function normalizeTelegramUpdate(update2) {
     isGuestMessage: sourceUpdateType === "guest_message",
     isDirectMessages: Boolean(message2.chat?.is_direct_messages),
     parentChatId: message2.chat?.parent_chat?.id ?? message2.direct_messages_topic?.parent_topic?.id,
-    replyToMessageId: message2.reply_to_message?.message_id,
-    replyToUsername: message2.reply_to_message?.from?.username ?? message2.reply_to_message?.sender_chat?.username,
-    replyToBot: Boolean(message2.reply_to_message?.from?.is_bot),
+    replyToMessageId: replyTo?.message_id,
+    replyToUsername: replyTo?.from?.username ?? replyTo?.sender_chat?.username,
+    replyToBot: Boolean(replyTo?.from?.is_bot),
     mentionedUsernames: extractTelegramMentionedUsernames(message2, text),
     sourceUpdateType
   };
@@ -598337,7 +598632,7 @@ function renderTelegramSubAgentError(username, error) {
   process.stdout.write(`    ${c3.dim("⎿")} ${c3.red("✘")} @${username}: ${c3.dim(preview)}
 `);
 }
-var TELEGRAM_SAFETY_PROMPT, ADMIN_DM_PROMPT, ADMIN_GROUP_PROMPT, TELEGRAM_PUBLIC_SOUL_PROFILE, TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT, TELEGRAM_PUBLIC_MEMORY_SCOPE_CONTRACT, GROUP_REPLY_DISCRETION_PROMPT, TELEGRAM_CHAT_MODE_PROMPT, ADMIN_CHAT_PROFILE_PROMPT, TELEGRAM_ACTION_RESPONSE_CONTRACT, TELEGRAM_CHAT_HISTORY_LIMIT, TELEGRAM_CONTEXT_RECENT_DEFAULT, TELEGRAM_CONTEXT_LINE_LIMIT, TELEGRAM_CONTEXT_SAMPLE_LIMIT, TELEGRAM_MEMORY_CARD_LIMIT, TELEGRAM_MEMORY_NOTE_LIMIT, TELEGRAM_MEMORY_STOPWORDS, TELEGRAM_PUBLIC_HELP_COMMANDS, MEDIA_CACHE_TTL_MS, TelegramBridge;
+var TELEGRAM_SAFETY_PROMPT, ADMIN_DM_PROMPT, ADMIN_GROUP_PROMPT, TELEGRAM_PUBLIC_SOUL_PROFILE, TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT, TELEGRAM_PUBLIC_MEMORY_SCOPE_CONTRACT, GROUP_REPLY_DISCRETION_PROMPT, TELEGRAM_CHAT_MODE_PROMPT, ADMIN_CHAT_PROFILE_PROMPT, TELEGRAM_ACTION_RESPONSE_CONTRACT, TELEGRAM_CHAT_HISTORY_LIMIT, TELEGRAM_CONTEXT_RECENT_DEFAULT, TELEGRAM_CONTEXT_LINE_LIMIT, TELEGRAM_CONTEXT_SAMPLE_LIMIT, TELEGRAM_MEMORY_CARD_LIMIT, TELEGRAM_MEMORY_NOTE_LIMIT, TELEGRAM_MEMORY_STOPWORDS, TELEGRAM_PUBLIC_HELP_COMMANDS, TELEGRAM_IMAGE_EXTENSIONS, MEDIA_CACHE_TTL_MS, TelegramBridge;
 var init_telegram_bridge = __esm({
   "packages/cli/src/tui/telegram-bridge.ts"() {
     "use strict";
@@ -598533,6 +598828,7 @@ Telegram response contract:
       "your"
     ]);
     TELEGRAM_PUBLIC_HELP_COMMANDS = /* @__PURE__ */ new Set(["help", "start", "auth", "call"]);
+    TELEGRAM_IMAGE_EXTENSIONS = /* @__PURE__ */ new Set([".png", ".jpg", ".jpeg", ".gif", ".webp", ".bmp", ".tiff", ".tif", ".svg"]);
     MEDIA_CACHE_TTL_MS = 30 * 60 * 1e3;
     TelegramBridge = class {
       constructor(botToken, onMessage, agentConfig, repoRoot, toolPolicyConfig) {
@@ -598944,6 +599240,80 @@ Telegram response contract:
           }
         }
       }
+      updateLastTelegramUserMessageText(msg, text) {
+        const sessionKey = this.sessionKeyForMessage(msg);
+        const history = this.chatHistory.get(sessionKey);
+        if (!history || !text.trim()) return;
+        for (let i2 = history.length - 1; i2 >= 0; i2--) {
+          const entry = history[i2];
+          if (entry.role !== "user") continue;
+          if (entry.messageId === msg.messageId || !entry.messageId && entry.text === msg.text) {
+            entry.text = text.trim();
+            entry.mediaSummary = summarizeTelegramMessageAttachments(msg) || entry.mediaSummary;
+            this.updateTelegramMemoryCards(sessionKey, entry);
+            this.saveTelegramConversationState(sessionKey);
+            return;
+          }
+        }
+      }
+      recentTelegramMediaEntries(chatId, limit = 12) {
+        const now = Date.now();
+        return [...this.mediaCache.values()].filter((entry) => {
+          if (chatId !== void 0 && String(entry.chatId) !== String(chatId)) return false;
+          return now - entry.cachedAt <= MEDIA_CACHE_TTL_MS;
+        }).sort((a2, b) => b.cachedAt - a2.cachedAt).slice(0, limit);
+      }
+      telegramMediaEntryMatchesKind(entry, kind) {
+        if (kind === "image") return telegramCachedMediaIsImage(entry);
+        if (kind === "pdf") return telegramCachedMediaIsPdf(entry);
+        if (kind === "audio") return telegramCachedMediaIsAudio(entry);
+        if (kind === "video") return telegramCachedMediaIsVideo(entry);
+        if (kind === "transcribable") {
+          return telegramCachedMediaIsAudio(entry) || telegramCachedMediaIsVideo(entry);
+        }
+        return true;
+      }
+      resolveTelegramScopedMediaPath(rawValue, chatId, currentMsg, kind) {
+        const raw = String(rawValue ?? "").trim();
+        const repoRoot = this.repoRoot || ".";
+        const creativeRoot = telegramCreativeWorkspaceRoot(repoRoot, chatId);
+        const mediaEntries = this.recentTelegramMediaEntries(chatId, 60).filter((entry) => this.telegramMediaEntryMatchesKind(entry, kind));
+        const aliases = /* @__PURE__ */ new Set(["", "latest", "last", "current", "this", "that", "it", "reply", "replied", "replied-to", "replied_to"]);
+        if (aliases.has(raw.toLowerCase())) {
+          const replied = currentMsg?.replyToMessageId ? mediaEntries.find((entry2) => entry2.messageId === currentMsg.replyToMessageId) : void 0;
+          const entry = replied ?? mediaEntries[0];
+          if (!entry) {
+            return { ok: false, error: `No recent ${kind} media is available in this Telegram chat scope.` };
+          }
+          return { ok: true, path: entry.localPath };
+        }
+        const matchingEntry = mediaEntries.find((entry) => {
+          if (resolve39(entry.localPath) === resolve39(raw)) return true;
+          if (basename23(entry.localPath) === raw) return true;
+          if (entry.fileUniqueId === raw || entry.fileId === raw) return true;
+          if (entry.messageId && String(entry.messageId) === raw) return true;
+          return false;
+        });
+        if (matchingEntry) return { ok: true, path: matchingEntry.localPath };
+        const creativeCandidate = isAbsolute7(raw) ? resolve39(raw) : resolve39(creativeRoot, raw);
+        if (isPathInside(creativeRoot, creativeCandidate) && existsSync106(creativeCandidate)) {
+          return { ok: true, path: creativeCandidate };
+        }
+        return {
+          ok: false,
+          error: `Path is outside this Telegram chat's media/workspace scope or does not exist: ${raw || "(empty)"}`
+        };
+      }
+      resolveTelegramScopedOutputPath(rawValue, chatId, fallbackName) {
+        const repoRoot = this.repoRoot || ".";
+        const creativeRoot = telegramCreativeWorkspaceRoot(repoRoot, chatId);
+        const raw = String(rawValue || fallbackName).trim() || fallbackName;
+        const outputPath2 = isAbsolute7(raw) ? resolve39(raw) : resolve39(creativeRoot, raw);
+        if (!isPathInside(creativeRoot, outputPath2)) {
+          return { ok: false, error: `Output path must stay inside this Telegram chat's creative workspace: ${raw}` };
+        }
+        return { ok: true, path: outputPath2 };
+      }
       updateTelegramParticipantProfile(sessionKey, msg, text) {
         const participantKey = String(msg.fromUserId || msg.username || msg.firstName || "unknown");
         const participants = this.chatParticipants.get(sessionKey) ?? /* @__PURE__ */ new Map();
@@ -599118,6 +599488,22 @@ ${notes2}`;
           sections.push(`### Zettelkasten Memory Recall
 ${cardLines.join("\n")}`);
         }
+        const recentMedia = this.recentTelegramMediaEntries(msg.chatId, 10);
+        if (recentMedia.length > 0) {
+          const mediaLines = recentMedia.map((entry) => {
+            const kind = telegramCachedMediaIsImage(entry) ? "image" : entry.mediaType;
+            const replyMark = msg.replyToMessageId && entry.messageId === msg.replyToMessageId ? " replied-to" : "";
+            const caption = entry.caption ? ` caption:${truncateTelegramContextLine(entry.caption, 120)}` : "";
+            const extracted = entry.extractedContent ? `
+    ${truncateTelegramContextLine(entry.extractedContent.replace(/\s+/g, " "), 220)}` : "";
+            return `- message_id ${entry.messageId}${replyMark}: ${kind}; path ${entry.localPath}; file ${basename23(entry.localPath)}${caption}${extracted}`;
+          });
+          sections.push([
+            "### Recent Chat Media",
+            "Use these paths only as tool inputs when the user asks about media in this chat. Do not quote local paths in the visible Telegram reply.",
+            mediaLines.join("\n")
+          ].join("\n"));
+        }
         if (olderCount > 0) {
           const older = history.slice(0, olderCount);
           const bySpeaker = /* @__PURE__ */ new Map();
@@ -599814,8 +600200,8 @@ Join: ${newUrl}`);
             }
           }
           let steeringText = msg.text;
-          if (msg.media) {
-            const mediaContext = await this.processMedia(msg);
+          if (msg.media || msg.replyToMedia) {
+            const mediaContext = await this.processMediaContextForMessage(msg);
             if (mediaContext) {
               steeringText += `
 
@@ -599889,8 +600275,8 @@ ${mediaContext}`;
         this.tuiWrite(() => renderTelegramSubAgentStart(msg.username, msg.text, isAdminDM));
         try {
           let mediaContext = "";
-          if (msg.media) {
-            mediaContext = await this.processMedia(msg);
+          if (msg.media || msg.replyToMedia) {
+            mediaContext = await this.processMediaContextForMessage(msg);
           }
           const result = await this.runSubAgent(msg, subAgent, mediaContext);
           if (subAgent.typingInterval) {
@@ -599992,8 +600378,8 @@ ${mediaContext}`;
         this.tuiWrite(() => renderTelegramSubAgentEvent(msg.username, `admin chat with full context/tools (${this.interactionMode})`));
         try {
           let mediaContext = "";
-          if (msg.media) {
-            mediaContext = await this.processMedia(msg);
+          if (msg.media || msg.replyToMedia) {
+            mediaContext = await this.processMediaContextForMessage(msg);
           }
           const result = await this.runSubAgent(msg, subAgent, mediaContext, "chat");
           if (subAgent.typingInterval) {
@@ -600076,7 +600462,7 @@ ${mediaContext}`;
         }
         this.tuiWrite(() => renderTelegramSubAgentEvent(msg.username, `live inference: chat reply (${this.interactionMode})`));
         try {
-          const mediaContext = msg.media || msg.livePhoto ? "Attachment received. Quick-chat mode does not inspect media; use action mode for media analysis." : "";
+          const mediaContext = msg.media || msg.replyToMedia || msg.livePhoto ? await this.processMediaContextForMessage(msg) : "";
           const finalText = await this.runTelegramChatCompletion(
             msg,
             toolContext,
@@ -600569,6 +600955,128 @@ ${lines.join("\n\n")}` };
               }
             };
           }
+          if (tool.name === "image_read") {
+            return {
+              ...tool,
+              description: "Read only images from this Telegram chat's media cache or creative workspace. Use path='reply' for the replied-to image or path='latest' for the most recent chat image.",
+              execute: async (args) => {
+                const resolved = this.resolveTelegramScopedMediaPath(args["path"], chatId, currentMsg, "image");
+                if (!resolved.ok) return { success: false, output: "", error: resolved.error };
+                return tool.execute({ ...args, path: resolved.path });
+              }
+            };
+          }
+          if (tool.name === "ocr") {
+            return {
+              ...tool,
+              description: "Extract text only from images in this Telegram chat's media cache or creative workspace. Use path='reply' or path='latest' for chat media references.",
+              execute: async (args) => {
+                const resolved = this.resolveTelegramScopedMediaPath(args["path"], chatId, currentMsg, "image");
+                if (!resolved.ok) return { success: false, output: "", error: resolved.error };
+                return tool.execute({ ...args, path: resolved.path });
+              }
+            };
+          }
+          if (tool.name === "vision") {
+            return {
+              ...tool,
+              description: "Analyze only images from this Telegram chat's media cache or creative workspace. Use image='reply' for the replied-to image or image='latest' for the most recent chat image.",
+              execute: async (args) => {
+                const resolved = this.resolveTelegramScopedMediaPath(args["image"], chatId, currentMsg, "image");
+                if (!resolved.ok) return { success: false, output: "", error: resolved.error };
+                return tool.execute({ ...args, image: resolved.path });
+              }
+            };
+          }
+          if (tool.name === "ocr_image_advanced") {
+            return {
+              ...tool,
+              description: "Advanced OCR only for images in this Telegram chat's media cache or creative workspace. Batch directory mode is disabled in public Telegram scope.",
+              execute: async (args) => {
+                if (args["batch"] === true) return { success: false, output: "", error: "Batch directory OCR is not available in public Telegram scope." };
+                const resolved = this.resolveTelegramScopedMediaPath(args["image"], chatId, currentMsg, "image");
+                if (!resolved.ok) return { success: false, output: "", error: resolved.error };
+                const next = { ...args, image: resolved.path };
+                if (typeof next["output_dir"] === "string" && next["output_dir"].trim()) {
+                  const output = this.resolveTelegramScopedOutputPath(next["output_dir"], chatId, "ocr-output");
+                  if (!output.ok) return { success: false, output: "", error: output.error };
+                  next["output_dir"] = output.path;
+                }
+                return tool.execute(next);
+              }
+            };
+          }
+          if (tool.name === "transcribe_file") {
+            return {
+              ...tool,
+              description: "Transcribe only audio/video files from this Telegram chat's media cache or creative workspace. Use path='reply' or path='latest' for chat media references.",
+              execute: async (args) => {
+                const resolved = this.resolveTelegramScopedMediaPath(args["path"], chatId, currentMsg, "transcribable");
+                if (!resolved.ok) return { success: false, output: "", error: resolved.error };
+                return tool.execute({ ...args, path: resolved.path });
+              }
+            };
+          }
+          if (tool.name === "pdf_to_text") {
+            return {
+              ...tool,
+              description: "Extract text only from PDFs in this Telegram chat's media cache or creative workspace. Use path='reply' or path='latest' for chat document references.",
+              execute: async (args) => {
+                const resolved = this.resolveTelegramScopedMediaPath(args["path"], chatId, currentMsg, "pdf");
+                if (!resolved.ok) return { success: false, output: "", error: resolved.error };
+                return tool.execute({ ...args, path: resolved.path });
+              }
+            };
+          }
+          if (tool.name === "ocr_pdf") {
+            return {
+              ...tool,
+              description: "OCR only PDFs from this Telegram chat's media cache or creative workspace. Output, when requested, is forced into this chat's creative workspace.",
+              execute: async (args) => {
+                const input = this.resolveTelegramScopedMediaPath(args["input"], chatId, currentMsg, "pdf");
+                if (!input.ok) return { success: false, output: "", error: input.error };
+                const next = { ...args, input: input.path };
+                if (typeof next["output"] === "string" && next["output"].trim()) {
+                  const output = this.resolveTelegramScopedOutputPath(next["output"], chatId, `ocr-${Date.now()}.pdf`);
+                  if (!output.ok) return { success: false, output: "", error: output.error };
+                  next["output"] = output.path;
+                }
+                return tool.execute(next);
+              }
+            };
+          }
+          if (tool.name === "video_understand") {
+            return {
+              ...tool,
+              description: "Analyze only video files from this Telegram chat's media cache or creative workspace. URL download is disabled in public Telegram scope; use path='reply' or path='latest'.",
+              execute: async (args) => {
+                if (args["url"]) return { success: false, output: "", error: "URL video analysis is not available in public Telegram scope. Use a video posted in this chat." };
+                const resolved = this.resolveTelegramScopedMediaPath(args["path"], chatId, currentMsg, "video");
+                if (!resolved.ok) return { success: false, output: "", error: resolved.error };
+                return tool.execute({ ...args, path: resolved.path });
+              }
+            };
+          }
+          if (tool.name === "audio_analyze") {
+            return {
+              ...tool,
+              description: "Analyze only audio files from this Telegram chat's media cache or creative workspace. Microphone/listen mode is disabled in public Telegram scope.",
+              execute: async (args) => {
+                if (String(args["action"] || "").toLowerCase() === "listen") {
+                  return { success: false, output: "", error: "Continuous microphone listening is not available in Telegram public scope." };
+                }
+                const resolved = this.resolveTelegramScopedMediaPath(args["file"] ?? args["path"], chatId, currentMsg, "audio");
+                if (!resolved.ok) return { success: false, output: "", error: resolved.error };
+                return tool.execute({ ...args, file: resolved.path, path: resolved.path });
+              }
+            };
+          }
+          if (tool.name === "explore_tools") {
+            return {
+              ...tool,
+              description: "List and explain the tools available in this Telegram public/group scope. Do not invent unavailable tool names."
+            };
+          }
           return tool;
         });
       }
@@ -600732,11 +601240,16 @@ Scoped workspace: ${scopedRoot}`,
           new ImageReadTool(repoRoot),
           new OCRTool(repoRoot),
           new VisionTool(repoRoot),
+          new OcrImageAdvancedTool(repoRoot),
           new OcrPdfTool(repoRoot),
           new PdfToTextTool(repoRoot),
           // Transcription tools
           new TranscribeFileTool(repoRoot),
-          new TranscribeUrlTool(repoRoot)
+          new TranscribeUrlTool(repoRoot),
+          new VideoUnderstandTool(repoRoot),
+          new AudioAnalyzeTool(),
+          new ExploreToolsTool(),
+          this.buildTelegramMediaRecentTool(chatId, msg)
         ];
         const adminTools = [
           new ShellTool(repoRoot),
@@ -600839,6 +601352,55 @@ Scoped workspace: ${scopedRoot}`,
         ]);
         return tools.filter((tool) => !blocked.has(tool.name));
       }
+      buildTelegramMediaRecentTool(chatId, currentMsg) {
+        const bridge = this;
+        return {
+          name: "telegram_media_recent",
+          description: "List recent media files available in this Telegram chat scope, including safe aliases for image_read, ocr, vision, transcribe_file, pdf_to_text, video_understand, and audio_analyze.",
+          parameters: {
+            type: "object",
+            properties: {
+              kind: {
+                type: "string",
+                enum: ["media", "image", "audio", "video", "pdf", "transcribable"],
+                description: "Filter by media kind. Defaults to all recent chat media."
+              },
+              limit: { type: "number", description: "Maximum entries to return, 1-20. Default: 10." }
+            }
+          },
+          async execute(args) {
+            const start2 = performance.now();
+            const kind = String(args["kind"] || "media").toLowerCase();
+            const limit = typeof args["limit"] === "number" && Number.isFinite(args["limit"]) ? Math.max(1, Math.min(20, Math.floor(args["limit"]))) : 10;
+            const entries = bridge.recentTelegramMediaEntries(chatId, 60).filter((entry) => bridge.telegramMediaEntryMatchesKind(entry, kind)).slice(0, limit);
+            if (entries.length === 0) {
+              return { success: true, output: `No recent ${kind} media is available in this Telegram chat scope.`, durationMs: performance.now() - start2 };
+            }
+            const lines = entries.map((entry, index) => {
+              const parts = [
+                `${index + 1}. message_id ${entry.messageId || "unknown"}`,
+                currentMsg?.replyToMessageId === entry.messageId ? "replied-to" : "",
+                telegramCachedMediaIsImage(entry) ? "image" : telegramCachedMediaIsPdf(entry) ? "pdf" : telegramCachedMediaIsAudio(entry) ? "audio" : telegramCachedMediaIsVideo(entry) ? "video" : entry.mediaType,
+                `file=${basename23(entry.localPath)}`,
+                `path=${entry.localPath}`,
+                entry.caption ? `caption=${truncateTelegramContextLine(entry.caption, 140)}` : ""
+              ].filter(Boolean);
+              const extracted = entry.extractedContent ? `
+   context: ${truncateTelegramContextLine(entry.extractedContent.replace(/\s+/g, " "), 240)}` : "";
+              return `${parts.join("; ")}${extracted}`;
+            });
+            return {
+              success: true,
+              output: [
+                "Recent scoped Telegram media:",
+                "Use path='reply' for replied-to media, path='latest' for the most recent matching item, or one of the listed paths.",
+                lines.join("\n")
+              ].join("\n"),
+              durationMs: performance.now() - start2
+            };
+          }
+        };
+      }
       imageGenerationDefaultsForRepo(repoRoot) {
         const settings = resolveSettings(repoRoot);
         return {
@@ -601056,30 +601618,36 @@ ${knownList}` : "Private-user telegram_send_file target must be this DM or a kno
        * Downloads the file, runs it through the appropriate pipeline,
        * caches it, and returns a text description for the agent.
        */
-      async processMedia(msg) {
-        if (!msg.media) return "";
-        const { type, fileId, fileUniqueId, mimeType, caption } = msg.media;
-        const isImageMedia = telegramMediaIsImage(msg.media);
+      async processMedia(msg, source = "message") {
+        const media = source === "reply" ? msg.replyToMedia : msg.media;
+        if (!media) return "";
+        const { type, fileId, fileUniqueId, mimeType, caption } = media;
+        const isImageMedia = telegramMediaIsImage(media);
+        const sourceMessageId = source === "reply" ? msg.replyToMessageId : msg.messageId;
+        const sourceLabel = source === "reply" ? "replied-to " : "";
         let ext = ".bin";
-        if (isImageMedia) ext = telegramImageExtension(msg.media);
+        if (isImageMedia) ext = telegramImageExtension(media);
         else if (type === "audio" || type === "voice") ext = ".ogg";
         else if (type === "video" || type === "video_note" || type === "live_photo") ext = ".mp4";
-        else if (msg.media.fileName) {
-          const dotIdx = msg.media.fileName.lastIndexOf(".");
-          if (dotIdx >= 0) ext = msg.media.fileName.slice(dotIdx);
+        else if (media.fileName) {
+          const dotIdx = media.fileName.lastIndexOf(".");
+          if (dotIdx >= 0) ext = media.fileName.slice(dotIdx);
         }
         const localPath = await this.downloadTelegramFile(fileId, ext);
         if (!localPath) return `[Media: ${type} — failed to download]`;
         const cacheEntry = {
           localPath,
           fileId,
+          fileUniqueId,
           chatId: msg.chatId,
+          messageId: sourceMessageId ?? 0,
           username: msg.username,
           mediaType: type,
           mimeType,
+          caption,
           cachedAt: Date.now()
         };
-        this.mediaCache.set(fileUniqueId, cacheEntry);
+        this.mediaCache.set(`${String(msg.chatId)}:${String(sourceMessageId ?? 0)}:${fileUniqueId}`, cacheEntry);
         const metadataKey = String(msg.chatId);
         if (!this.mediaMetadata.has(metadataKey)) {
           this.mediaMetadata.set(metadataKey, []);
@@ -601100,7 +601668,7 @@ ${knownList}` : "Private-user telegram_send_file target must be this DM or a kno
               {
                 path: localPath,
                 buffer: readFileSync87(localPath),
-                mime: telegramImageMime(msg.media)
+                mime: telegramImageMime(media)
               },
               this.agentConfig?.model ?? ""
             );
@@ -601109,10 +601677,10 @@ ${knownList}` : "Private-user telegram_send_file target must be this DM or a kno
           } catch {
           }
           if (visionContext) {
-            description = `[Image received: ${localPath}${caption ? ` — caption: "${caption}"` : ""}
+            description = `[${sourceLabel}image received: ${localPath}${caption ? ` — caption: "${caption}"` : ""}
 ${visionContext}]`;
           } else {
-            description = `[Image received and saved to ${localPath}${caption ? ` — caption: "${caption}"` : ""}. You can use image_read or vision tools to analyze it if available.]`;
+            description = `[${sourceLabel}image received and saved to ${localPath}${caption ? ` — caption: "${caption}"` : ""}. You can use image_read, ocr, or vision tools to analyze it.]`;
           }
           try {
             await fetch("http://127.0.0.1:11435/v1/memory/ingest", {
@@ -601136,9 +601704,9 @@ ${visionContext}]`;
           } catch {
           }
           if (transcription) {
-            description = `[Voice message transcribed: "${transcription}"${caption ? ` — caption: "${caption}"` : ""}]`;
+            description = `[${sourceLabel}voice message transcribed: "${transcription}"${caption ? ` — caption: "${caption}"` : ""}]`;
           } else {
-            description = `[Audio/voice message received and saved to ${localPath}${caption ? ` — caption: "${caption}"` : ""}. You can use transcribe_file to transcribe it if available.]`;
+            description = `[${sourceLabel}audio/voice message received and saved to ${localPath}${caption ? ` — caption: "${caption}"` : ""}. You can use transcribe_file to transcribe it.]`;
           }
           try {
             await fetch("http://127.0.0.1:11435/v1/memory/ingest", {
@@ -601151,13 +601719,30 @@ ${visionContext}]`;
           }
         } else if (type === "video" || type === "video_note" || type === "live_photo") {
           const label = type === "live_photo" ? "Live photo" : "Video";
-          description = `[${label} received and saved to ${localPath}${caption ? ` — caption: "${caption}"` : ""}.]`;
+          description = `[${sourceLabel}${label.toLowerCase()} received and saved to ${localPath}${caption ? ` — caption: "${caption}"` : ""}. You can use video_understand or transcribe_file to analyze it.]`;
         } else if (type === "document") {
-          description = `[Document received: ${msg.media.fileName || "unnamed"}${mimeType ? ` (${mimeType})` : ""}, saved to ${localPath}${caption ? ` — caption: "${caption}"` : ""}.]`;
+          description = `[${sourceLabel}document received: ${media.fileName || "unnamed"}${mimeType ? ` (${mimeType})` : ""}, saved to ${localPath}${caption ? ` — caption: "${caption}"` : ""}.]`;
         }
         cacheEntry.extractedContent = description;
         return description;
       }
+      async processMediaContextForMessage(msg) {
+        const parts = [];
+        if (msg.media) {
+          const current = await this.processMedia(msg, "message");
+          if (current) parts.push(current);
+        }
+        if (msg.replyToMedia) {
+          const replied = await this.processMedia(msg, "reply");
+          if (replied) parts.push(replied);
+        }
+        const text = parts.join("\n\n");
+        if (text) this.updateLastTelegramUserMessageText(msg, `${msg.text}
+
+[Media context]
+${text}`.trim());
+        return text;
+      }
       /** Clean up expired media cache entries (older than 30 minutes) */
       cleanupMediaCache() {
         const now = Date.now();
@@ -625743,7 +626328,7 @@ var clipboard_media_exports = {};
 __export(clipboard_media_exports, {
   pasteClipboardImageToFile: () => pasteClipboardImageToFile
 });
-import { execFileSync as execFileSync5, execSync as execSync58 } from "node:child_process";
+import { execFileSync as execFileSync6, execSync as execSync58 } from "node:child_process";
 import { mkdirSync as mkdirSync72, readFileSync as readFileSync99, rmSync as rmSync5, writeFileSync as writeFileSync67 } from "node:fs";
 import { join as join136 } from "node:path";
 function pasteClipboardImageToFile(repoRoot) {
@@ -625760,7 +626345,7 @@ function readClipboardImage() {
     try {
       execSync58("command -v pngpaste", { stdio: "ignore", timeout: 1e3 });
       const tmp = `/tmp/omnius-clipboard-${Date.now()}.png`;
-      execFileSync5("pngpaste", [tmp], { timeout: 3e3 });
+      execFileSync6("pngpaste", [tmp], { timeout: 3e3 });
       const buffer2 = readFileSync99(tmp);
       try {
         rmSync5(tmp);
@@ -625780,7 +626365,7 @@ function readClipboardImage() {
     ];
     for (const attempt of attempts) {
       try {
-        const buffer2 = execFileSync5(attempt.cmd, attempt.args, { timeout: 3e3, maxBuffer: 25 * 1024 * 1024 });
+        const buffer2 = execFileSync6(attempt.cmd, attempt.args, { timeout: 3e3, maxBuffer: 25 * 1024 * 1024 });
         if (buffer2.length > 0) return { buffer: buffer2, mime: attempt.mime, ext: attempt.ext };
       } catch {
         continue;
@@ -625797,7 +626382,7 @@ function readClipboardImage() {
         "$img.Save($ms,[Drawing.Imaging.ImageFormat]::Png);",
         "[Console]::OpenStandardOutput().Write($ms.ToArray(),0,$ms.Length)"
       ].join("");
-      const buffer2 = execFileSync5("powershell.exe", ["-NoProfile", "-Command", ps], {
+      const buffer2 = execFileSync6("powershell.exe", ["-NoProfile", "-Command", ps], {
         timeout: 5e3,
         maxBuffer: 25 * 1024 * 1024
       });
@@ -625816,7 +626401,7 @@ var init_clipboard_media = __esm({
 
 // packages/cli/src/tui/interactive.ts
 import { cwd } from "node:process";
-import { resolve as resolve44, join as join137, dirname as dirname38, extname as extname15, relative as relative14 } from "node:path";
+import { resolve as resolve44, join as join137, dirname as dirname38, extname as extname16, relative as relative14 } from "node:path";
 import { createRequire as createRequire8 } from "node:module";
 import { fileURLToPath as fileURLToPath18 } from "node:url";
 import {
@@ -633118,7 +633703,7 @@ Execute this skill now. Follow the behavioral guidance above.`;
           const imgPath = resolve44(repoRoot, cleanPath);
           const imgBuffer = readFileSync100(imgPath);
           const base642 = imgBuffer.toString("base64");
-          const ext = extname15(cleanPath).toLowerCase();
+          const ext = extname16(cleanPath).toLowerCase();
           const mime = ext === ".png" ? "image/png" : ext === ".gif" ? "image/gif" : ext === ".webp" ? "image/webp" : "image/jpeg";
           const asciiContext = await renderAsciiPreviewForImage(
             imgPath,