omnius 1.0.232 → 1.0.233

package/dist/index.js

@@ -536919,37 +536919,35 @@ function findOrphans() {
       const isHot = parseFloat(cpu) > 10 ? " ⚠ HIGH CPU" : "";
       lines.push(`  ${String(pid).padStart(7)} ${String(ppid).padStart(7)} ${String(cpu).padStart(6)} ${String(mem).padStart(5)} ${String(elapsed).padStart(9)}  ${shortCmd}${isHot}`);
     }
-    lines.push(`
-To kill a process: process_health(action='kill', pid=PID)`);
-    lines.push(`Or from the prompt: /destroy processes --global`);
+    lines.push("\nThese are untracked diagnostics only. They are not killed automatically.");
+    lines.push("To stop a process safely, it must have an active verified Omnius process lease.");
+    lines.push("Use /destroy processes --global to sweep registered stale leases and report untracked matches.");
     return lines.join("\n");
   } catch {
     return "Could not scan for orphan processes.";
   }
 }
-function killProcess2(pid) {
-  try {
-    try {
-      process.kill(-pid, "SIGTERM");
-    } catch {
-    }
-    process.kill(pid, "SIGTERM");
-    try {
-      execSync28(`sleep 1 && kill -0 ${pid} 2>/dev/null && kill -9 ${pid} 2>/dev/null`, {
-        timeout: 3e3,
-        stdio: "pipe"
-      });
-    } catch {
-    }
-    return `Killed PID ${pid} (SIGTERM → SIGKILL)`;
-  } catch (err) {
-    return `Failed to kill PID ${pid}: ${err instanceof Error ? err.message : String(err)}`;
+async function stopRegisteredLeaseByPid(pid) {
+  const lease = listProcessLeases({ includeInactive: false }).find((candidate) => candidate.pid === pid && candidate.status === "active");
+  if (!lease) {
+    return `Refusing to kill PID ${pid}: no active Omnius process lease verifies ownership. Treat it as untracked_omnius_like diagnostics only and stop it manually if needed.`;
+  }
+  const action = await stopProcessLease(lease.leaseId, {
+    reason: "process_health explicit stop"
+  });
+  if (action.action === "killed") {
+    return `Stopped registered Omnius lease ${lease.leaseId} for PID ${pid}: ${action.reason}`;
   }
+  if (action.action === "suspect_unverified") {
+    return `Refusing to kill PID ${pid}: lease ${lease.leaseId} could not verify current PID identity.`;
+  }
+  return `PID ${pid} was not killed; lease ${lease.leaseId} action=${action.action}: ${action.reason}`;
 }
 var OMNIUS_PATTERNS, ProcessHealthTool;
 var init_process_health = __esm({
   "packages/execution/dist/tools/process-health.js"() {
     "use strict";
+    init_process_lifecycle();
     OMNIUS_PATTERNS = [
       "omnius",
       "nexus-daemon",
@@ -536966,7 +536964,7 @@ var init_process_health = __esm({
     ].join("|");
     ProcessHealthTool = class {
       name = "process_health";
-      description = "Check system CPU/memory load and find orphaned processes. Actions: 'status' (CPU + memory + top processes), 'orphans' (find stale Omnius-related processes), 'kill' (kill a specific PID). Use this when the system feels slow or when a shell command seems stuck.";
+      description = "Check system CPU/memory load and find orphaned processes. Actions: 'status' (CPU + memory + top processes), 'orphans' (find stale Omnius-related processes), 'kill' (stop a registered Omnius process lease by PID). Use this when the system feels slow or when a shell command seems stuck.";
       parameters = {
         type: "object",
         properties: {
@@ -536977,7 +536975,7 @@ var init_process_health = __esm({
           },
           pid: {
             type: "number",
-            description: "Process ID to kill (for 'kill' action)"
+            description: "Process ID to stop, only when it has a verified active Omnius process lease"
           }
         },
         required: []
@@ -536998,7 +536996,7 @@ var init_process_health = __esm({
               if (pid === process.pid || pid === process.ppid) {
                 return { success: false, output: "", error: "Cannot kill self or parent process", durationMs: performance.now() - start2 };
               }
-              return { success: true, output: killProcess2(pid), durationMs: performance.now() - start2 };
+              return { success: true, output: await stopRegisteredLeaseByPid(pid), durationMs: performance.now() - start2 };
             }
             default:
               return { success: false, output: "", error: `Unknown action: ${action}`, durationMs: performance.now() - start2 };
@@ -621021,8 +621019,8 @@ sleep 1
           });
           items.push({
             key: "__kill__",
-            label: "Kill Omnius schedulers + active runs",
-            detail: "Stop scheduler/nexus processes and terminate active Omnius runs"
+            label: "Stop registered schedulers + active runs",
+            detail: "Disable scheduler sources and stop registry-owned Omnius runs"
           });
           const result = await tuiSelect({
             items,
@@ -621386,7 +621384,7 @@ sleep 1
             body: JSON.stringify({}),
             headers: { "Content-Type": "application/json" }
           });
-          renderInfo("Kill signal sent to Omnius scheduler processes.");
+          renderInfo("Scheduler sources disabled and registered scheduler leases stopped where verified.");
         } catch (e2) {
           renderError(e2?.message || String(e2));
         }
@@ -645279,6 +645277,31 @@ function adaptTool5(tool, todoSessionId, progress) {
     }
   });
 }
+function normalizeTelegramImageAnalyzeDetail(value2) {
+  const raw = String(value2 ?? "auto").trim().toLowerCase();
+  return raw === "text" || raw === "visual" || raw === "full" ? raw : "auto";
+}
+function stripTelegramImagePayloadMarkers(value2) {
+  return value2.replace(/\n?\[IMAGE_BASE64:[^\]]+\]/g, "\n[IMAGE_BASE64 omitted from display]");
+}
+function truncateTelegramStageOutput(value2, maxChars = 3500) {
+  const clean5 = stripTelegramImagePayloadMarkers(value2).replace(/\n{3,}/g, "\n\n").trim();
+  if (clean5.length <= maxChars) return clean5;
+  return `${clean5.slice(0, maxChars).trimEnd()}
+... (${clean5.length - maxChars} more chars omitted)`;
+}
+function telegramTextExtractionSignal(text) {
+  const lines = text.split(/\r?\n/).map((line) => line.trim()).filter(Boolean);
+  const chars = text.replace(/\s+/g, "").length;
+  const longestLine = lines.reduce((max, line) => Math.max(max, line.length), 0);
+  return {
+    chars,
+    lines: lines.length,
+    longestLine,
+    dense: chars >= 700 || lines.length >= 8 || longestLine >= 140,
+    present: chars >= 8
+  };
+}
 function telegramBotAccessSettingsFromApi(settings) {
   return {
     isAccessRestricted: Boolean(settings.is_access_restricted),
@@ -645414,7 +645437,7 @@ function renderTelegramSubAgentError(username, error) {
   const preview = error.length > 80 ? error.slice(0, 77) + "..." : error;
   renderTelegramCoalescedRow(`${c3.red("✘")} ${c3.bold(`@${username}`)}: ${c3.dim(preview)}`);
 }
-var TELEGRAM_TOOL_ACTION_GROUPS, TELEGRAM_TOOL_ACTION_GROUP, TELEGRAM_TOOL_MUTATING_GROUPS, DEFAULT_TELEGRAM_TOOL_GROUP_POLICY, TELEGRAM_TOOL_BUTTON_LABELS, TELEGRAM_SAFETY_PROMPT, ADMIN_DM_PROMPT, ADMIN_GROUP_PROMPT, TELEGRAM_PUBLIC_SOUL_PROFILE, TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT, TELEGRAM_PUBLIC_MEMORY_SCOPE_CONTRACT, TELEGRAM_PUBLIC_VISION_STACK_CONTRACT, GROUP_REPLY_DISCRETION_PROMPT, TELEGRAM_CHAT_MODE_PROMPT, ADMIN_CHAT_PROFILE_PROMPT, TELEGRAM_ACTION_RESPONSE_CONTRACT, TELEGRAM_EXTERNAL_ACQUISITION_CONTRACT, TELEGRAM_LINK_INTEGRITY_CONTRACT, TELEGRAM_INTERACTION_DECISION_RESPONSE_FORMAT, TELEGRAM_INTERACTION_DECISION_MINIMAL_SCHEMA, TELEGRAM_INTERACTION_DECISION_REPAIR_SCHEMA, TELEGRAM_CHAT_REPLY_RESPONSE_FORMAT, TELEGRAM_SPACED_URL_RE, TELEGRAM_HTTP_URL_RE, TELEGRAM_STUCK_SELF_TALK_PREFIXES, 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_ASSOCIATIVE_FACT_LIMIT, TELEGRAM_ASSOCIATIVE_USER_FACT_LIMIT, TELEGRAM_ASSOCIATIVE_ACTION_LIMIT, TELEGRAM_ASSOCIATIVE_RELATION_LIMIT, TELEGRAM_MEMORY_STOPWORDS, TELEGRAM_MEMORY_GENERIC_QUERY_TOKENS, TELEGRAM_SUB_AGENT_BOUNDED_OPTIONS, TELEGRAM_PUBLIC_FAST_OPTIONS, TELEGRAM_ADMIN_EVIDENCE_OPTIONS, TELEGRAM_SUB_AGENT_DEFAULT_LIMIT, TELEGRAM_SUB_AGENT_MAX_LIMIT, TELEGRAM_SUB_AGENT_BURST_CONTEXT_LIMIT, TELEGRAM_ADMIN_LIVE_PANEL_PAGES, TELEGRAM_ADMIN_LIVE_MUTATION_TOOLS, TELEGRAM_PUBLIC_HELP_COMMANDS2, TELEGRAM_REMINDER_SLASH_COMMANDS, TELEGRAM_REFLECTION_SLASH_COMMANDS, TELEGRAM_PUBLIC_BOT_COMMAND_NAMES, TELEGRAM_IMAGE_EXTENSIONS, MEDIA_CACHE_TTL_MS, TELEGRAM_CHANNEL_DMN_SWEEP_MS, TELEGRAM_CHANNEL_DMN_IDLE_AFTER_MS, TELEGRAM_CHANNEL_DMN_MIN_INTERVAL_MS, TELEGRAM_CHANNEL_DMN_MIN_MESSAGES, TELEGRAM_ALLOWED_UPDATES, TELEGRAM_DEFAULT_LONG_POLL_TIMEOUT_SECONDS, TELEGRAM_ROUTER_AUTO_MIN_PARAMETERS_B, TELEGRAM_PUBLIC_TOOL_QUOTAS, TelegramBridge;
+var TELEGRAM_TOOL_ACTION_GROUPS, TELEGRAM_TOOL_ACTION_GROUP, TELEGRAM_TOOL_MUTATING_GROUPS, DEFAULT_TELEGRAM_TOOL_GROUP_POLICY, TELEGRAM_TOOL_BUTTON_LABELS, TELEGRAM_SAFETY_PROMPT, ADMIN_DM_PROMPT, ADMIN_GROUP_PROMPT, TELEGRAM_PUBLIC_SOUL_PROFILE, TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT, TELEGRAM_PUBLIC_MEMORY_SCOPE_CONTRACT, TELEGRAM_PUBLIC_VISION_STACK_CONTRACT, TELEGRAM_EVIDENCE_SUFFICIENCY_CONTRACT, GROUP_REPLY_DISCRETION_PROMPT, TELEGRAM_CHAT_MODE_PROMPT, ADMIN_CHAT_PROFILE_PROMPT, TELEGRAM_ACTION_RESPONSE_CONTRACT, TELEGRAM_EXTERNAL_ACQUISITION_CONTRACT, TELEGRAM_LINK_INTEGRITY_CONTRACT, TELEGRAM_INTERACTION_DECISION_RESPONSE_FORMAT, TELEGRAM_INTERACTION_DECISION_MINIMAL_SCHEMA, TELEGRAM_INTERACTION_DECISION_REPAIR_SCHEMA, TELEGRAM_CHAT_REPLY_RESPONSE_FORMAT, TELEGRAM_SPACED_URL_RE, TELEGRAM_HTTP_URL_RE, TELEGRAM_STUCK_SELF_TALK_PREFIXES, 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_ASSOCIATIVE_FACT_LIMIT, TELEGRAM_ASSOCIATIVE_USER_FACT_LIMIT, TELEGRAM_ASSOCIATIVE_ACTION_LIMIT, TELEGRAM_ASSOCIATIVE_RELATION_LIMIT, TELEGRAM_MEMORY_STOPWORDS, TELEGRAM_MEMORY_GENERIC_QUERY_TOKENS, TELEGRAM_SUB_AGENT_BOUNDED_OPTIONS, TELEGRAM_PUBLIC_FAST_OPTIONS, TELEGRAM_ADMIN_EVIDENCE_OPTIONS, TELEGRAM_SUB_AGENT_DEFAULT_LIMIT, TELEGRAM_SUB_AGENT_MAX_LIMIT, TELEGRAM_SUB_AGENT_BURST_CONTEXT_LIMIT, TELEGRAM_ADMIN_LIVE_PANEL_PAGES, TELEGRAM_ADMIN_LIVE_MUTATION_TOOLS, TELEGRAM_PUBLIC_HELP_COMMANDS2, TELEGRAM_REMINDER_SLASH_COMMANDS, TELEGRAM_REFLECTION_SLASH_COMMANDS, TELEGRAM_PUBLIC_BOT_COMMAND_NAMES, TELEGRAM_IMAGE_EXTENSIONS, MEDIA_CACHE_TTL_MS, TELEGRAM_CHANNEL_DMN_SWEEP_MS, TELEGRAM_CHANNEL_DMN_IDLE_AFTER_MS, TELEGRAM_CHANNEL_DMN_MIN_INTERVAL_MS, TELEGRAM_CHANNEL_DMN_MIN_MESSAGES, TELEGRAM_ALLOWED_UPDATES, TELEGRAM_DEFAULT_LONG_POLL_TIMEOUT_SECONDS, TELEGRAM_ROUTER_AUTO_MIN_PARAMETERS_B, TELEGRAM_PUBLIC_TOOL_QUOTAS, TelegramBridge;
 var init_telegram_bridge = __esm({
   "packages/cli/src/tui/telegram-bridge.ts"() {
     "use strict";
@@ -645630,12 +645653,21 @@ Public Telegram vision and media stack
 
 Public Telegram runs have the full scoped media-analysis stack for media posted in this chat:
 - Use telegram_media_recent to find recent scoped media, then use path/media aliases 'reply' and 'latest' instead of exposing local paths to users.
-- Use ocr_image_advanced for complex textual imagery: screenshots, dense documents, forms, receipts, scans, diagrams with labels, low-contrast photos, or uneven lighting.
-- Use ocr for quick image text extraction, image_read for image metadata + OCR + multimodal image payload, and vision for captioning, visual QA, object detection, or pointing.
+- For image questions, prefer telegram_image_analyze first. It resolves omitted/reply/latest media, starts with low-fidelity image intake, uses basic OCR as the text extraction probe, escalates to advanced OCR when text is dense or under-extracted, and escalates to Moondream vision when visual QA/captioning is needed.
+- Use ocr for quick image text extraction, ocr_image_advanced when basic OCR shows dense or degraded text, image_read for image metadata + multimodal image payload, and vision for direct Moondream captioning, visual QA, object detection, or pointing.
 - Use pdf_to_text for embedded-text PDFs and ocr_pdf for scanned PDFs.
 - Use video_understand and transcribe_file for video/audio media posted in this chat.
 - Use identity_memory for explicit user-provided identity assertions, staged next-image names, and "who is this?" recall from scoped media. Do not guess real identities from images.
 - These tools are current-chat scoped. Never inspect arbitrary local files, reveal local paths, or claim access to media outside this Telegram chat scope.
+`.trim();
+    TELEGRAM_EVIDENCE_SUFFICIENCY_CONTRACT = `
+Evidence sufficiency contract
+
+- Before making a factual claim, identify whether the answer is already supported by the supplied Telegram context, scoped memory, attached media extraction, or fresh tool/web evidence.
+- Use action/tools when the answer depends on current or recent external facts, volatile public events, named people or organizations in news-like claims, media extraction, source verification, or the user's wording asks you to find/check/look up something.
+- Quick-chat may answer only when the reply is conversational, based on supplied context, or based on stable common knowledge that is unlikely to have changed.
+- If a media tool call fails or lacks an argument, recover through telegram_image_analyze, telegram_media_recent, or the reply/latest aliases before telling the user you cannot inspect the media.
+- Do not fill evidence gaps with plausible memory. If evidence is missing and tools are unavailable, say what is missing instead of fabricating.
 `.trim();
     GROUP_REPLY_DISCRETION_PROMPT = `
 Reply discretion: you are in a group chat. The live router selected this turn
@@ -649477,6 +649509,31 @@ ${mediaContext}` : ""
         }
         return true;
       }
+      telegramScopedMediaParameters(parameters, fields, description) {
+        const base3 = parameters ?? {};
+        const baseRecord = base3;
+        const rawProperties = baseRecord["properties"];
+        const properties = rawProperties && typeof rawProperties === "object" && !Array.isArray(rawProperties) ? { ...rawProperties } : {};
+        for (const field of fields) {
+          const property = properties[field];
+          if (property && typeof property === "object" && !Array.isArray(property)) {
+            const currentDescription = typeof property["description"] === "string" ? String(property["description"]) : "";
+            properties[field] = {
+              ...property,
+              description: [currentDescription, description].filter(Boolean).join(" ")
+            };
+          }
+        }
+        const required = Array.isArray(baseRecord["required"]) ? baseRecord["required"].filter((item) => typeof item === "string" && !fields.includes(item)) : void 0;
+        const rest = { ...baseRecord };
+        delete rest["properties"];
+        delete rest["required"];
+        return {
+          ...rest,
+          properties,
+          ...required && required.length > 0 ? { required } : {}
+        };
+      }
       resolveTelegramScopedMediaPath(rawValue, chatId, currentMsg, kind) {
         const raw = String(rawValue ?? "").trim();
         const repoRoot = this.repoRoot || ".";
@@ -650671,6 +650728,7 @@ ${lines.join("\n")}`);
           toolContext,
           authorityContext: [
             baseContract,
+            TELEGRAM_EVIDENCE_SUFFICIENCY_CONTRACT,
             TELEGRAM_PUBLIC_MEMORY_SCOPE_CONTRACT,
             TELEGRAM_PUBLIC_VISION_STACK_CONTRACT
           ].join("\n\n"),
@@ -651844,6 +651902,8 @@ ${stimulationProbe.context}`,
           `- chat: a short conversational answer can be produced without tools.`,
           `- action: tools, workspace context, media processing, web lookup, delegation, or a multi-step agent loop may be needed.`,
           `Route discipline: infer whether the current request can be completed as immediate conversation or needs an external capability. Do not route from keyword categories.`,
+          `Evidence sufficiency: choose chat only when the answer is supported by supplied Telegram context, scoped memory, attached media context, or stable common knowledge. Choose action when answering would require web lookup, media extraction, current/recent public facts, named person/organization/news verification, or tool recovery.`,
+          `Do not answer volatile external claims from model memory on the fast path. Route to action so web/media tools can gather evidence.`,
           ``,
           `Reply discretion: make a human-like attention decision from the full social context. Observe the message, relationship stream, reply graph, conversation momentum, prior bot involvement, speaker intent, and notification-like signals, then decide whether a visible reply would be natural.`,
           `No hard triggers: direct address, @mentions, name/identity references, private-chat delivery, replies, active threads, and stimulation score are evidence only. They may raise or lower salience, but none guarantees should_reply=true or should_reply=false.`,
@@ -652359,6 +652419,7 @@ Profile: ${profile}
 Tool context: ${toolContext}
 ${chatLabel}`,
           TELEGRAM_ACTION_RESPONSE_CONTRACT,
+          TELEGRAM_EVIDENCE_SUFFICIENCY_CONTRACT,
           TELEGRAM_LINK_INTEGRITY_CONTRACT
         ];
         sections.push(conversationStream);
@@ -652392,6 +652453,8 @@ ${TELEGRAM_PUBLIC_SOUL_PROFILE}
 
 ${TELEGRAM_PUBLIC_MEMORY_SCOPE_CONTRACT}
 
+${TELEGRAM_EVIDENCE_SUFFICIENCY_CONTRACT}
+
 ${TELEGRAM_PUBLIC_VISION_STACK_CONTRACT}
 
 ${TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT}`);
@@ -652404,6 +652467,8 @@ ${TELEGRAM_PUBLIC_SOUL_PROFILE}
 
 ${TELEGRAM_PUBLIC_MEMORY_SCOPE_CONTRACT}
 
+${TELEGRAM_EVIDENCE_SUFFICIENCY_CONTRACT}
+
 ${TELEGRAM_PUBLIC_VISION_STACK_CONTRACT}
 
 ${TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT}`);
@@ -654120,6 +654185,8 @@ ${TELEGRAM_PUBLIC_VISION_STACK_CONTRACT}`;
             role: "system",
             content: `${TELEGRAM_CHAT_MODE_PROMPT}
 
+${TELEGRAM_EVIDENCE_SUFFICIENCY_CONTRACT}
+
 ${TELEGRAM_LINK_INTEGRITY_CONTRACT}
 
 ## Runtime Context
@@ -654467,6 +654534,8 @@ ${GROUP_REPLY_DISCRETION_PROMPT}` : "";
 
 ${TELEGRAM_ACTION_RESPONSE_CONTRACT}
 
+${TELEGRAM_EVIDENCE_SUFFICIENCY_CONTRACT}
+
 ${TELEGRAM_LINK_INTEGRITY_CONTRACT}
 
 ${TELEGRAM_EXTERNAL_ACQUISITION_CONTRACT}
@@ -654483,7 +654552,8 @@ ${currentTelegramPrompt}`;
             "You can remember facts about users and retrieve them later. Durable associative memory in the prompt includes participant profiles, relationships, scoped facts, and prior actions retained across days, sessions, and Omnius updates. You also have web_search and web_fetch to look up information.",
             TELEGRAM_LINK_INTEGRITY_CONTRACT,
             "If a user explicitly states a durable preference for reply cadence/order, call telegram_preference_set. Do not infer or classify reply-mode preferences from keywords, style, tone, or task type.",
-            "You have the full scoped Telegram media-analysis stack by default: telegram_media_recent, image_read, ocr, ocr_image_advanced, vision, pdf_to_text, ocr_pdf, transcribe_file, video_understand, audio_analyze, and identity_memory. For complex textual imagery, screenshots, forms, scans, or dense labels, prefer ocr_image_advanced after resolving media with path='reply' or path='latest'.",
+            TELEGRAM_EVIDENCE_SUFFICIENCY_CONTRACT,
+            "You have the full scoped Telegram media-analysis stack by default: telegram_image_analyze, telegram_media_recent, image_read, ocr, ocr_image_advanced, vision, pdf_to_text, ocr_pdf, transcribe_file, video_understand, audio_analyze, and identity_memory. For image questions, prefer telegram_image_analyze first; it resolves omitted/reply/latest media, starts with low-fidelity image intake, uses basic OCR as the text extraction probe, escalates to advanced OCR when text is dense or under-extracted, and escalates to Moondream vision when visual QA/captioning is needed.",
             formatIdentityMemoryContext(chatLabel || "Telegram private chat"),
             reminderToolContract,
             "If the user asks you to create an image, audio file, video, 3D/CAD model, or document artifact, create it with the scoped creative tools. Freshly generated artifacts are recorded and automatically attached to this Telegram chat when the turn completes, so do not call telegram_send_file for those same artifacts unless the user asked for a specific caption, existing/unrecorded file, or non-default target.",
@@ -654959,6 +655029,7 @@ ${lines.join("\n")}`
             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.",
+              parameters: this.telegramScopedMediaParameters(tool.parameters, ["path"], "In Telegram scope this is optional; omit it to use the replied-to image first, otherwise the latest chat image."),
               execute: async (args) => {
                 const resolved = this.resolveTelegramScopedMediaPath(args["path"], chatId, currentMsg, "image");
                 if (!resolved.ok) return { success: false, output: "", error: resolved.error };
@@ -654970,6 +655041,7 @@ ${lines.join("\n")}`
             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.",
+              parameters: this.telegramScopedMediaParameters(tool.parameters, ["path"], "In Telegram scope this is optional; omit it to use the replied-to image first, otherwise the latest chat image."),
               execute: async (args) => {
                 const resolved = this.resolveTelegramScopedMediaPath(args["path"], chatId, currentMsg, "image");
                 if (!resolved.ok) return { success: false, output: "", error: resolved.error };
@@ -654981,6 +655053,7 @@ ${lines.join("\n")}`
             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.",
+              parameters: this.telegramScopedMediaParameters(tool.parameters, ["image"], "In Telegram scope this is optional; omit it to use the replied-to image first, otherwise the latest chat image."),
               execute: async (args) => {
                 const resolved = this.resolveTelegramScopedMediaPath(args["image"], chatId, currentMsg, "image");
                 if (!resolved.ok) return { success: false, output: "", error: resolved.error };
@@ -654992,6 +655065,7 @@ ${lines.join("\n")}`
             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.",
+              parameters: this.telegramScopedMediaParameters(tool.parameters, ["image"], "In Telegram scope this is optional; omit it to use the replied-to image first, otherwise the latest chat image."),
               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");
@@ -655010,6 +655084,7 @@ ${lines.join("\n")}`
             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.",
+              parameters: this.telegramScopedMediaParameters(tool.parameters, ["path"], "In Telegram scope this is optional; omit it to use the replied-to audio/video first, otherwise the latest transcribable chat media."),
               execute: async (args) => {
                 const resolved = this.resolveTelegramScopedMediaPath(args["path"], chatId, currentMsg, "transcribable");
                 if (!resolved.ok) return { success: false, output: "", error: resolved.error };
@@ -655021,6 +655096,7 @@ ${lines.join("\n")}`
             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.",
+              parameters: this.telegramScopedMediaParameters(tool.parameters, ["path"], "In Telegram scope this is optional; omit it to use the replied-to PDF first, otherwise the latest chat PDF."),
               execute: async (args) => {
                 const resolved = this.resolveTelegramScopedMediaPath(args["path"], chatId, currentMsg, "pdf");
                 if (!resolved.ok) return { success: false, output: "", error: resolved.error };
@@ -655032,6 +655108,7 @@ ${lines.join("\n")}`
             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.",
+              parameters: this.telegramScopedMediaParameters(tool.parameters, ["input"], "In Telegram scope this is optional; omit it to use the replied-to PDF first, otherwise the latest chat PDF."),
               execute: async (args) => {
                 const input = this.resolveTelegramScopedMediaPath(args["input"], chatId, currentMsg, "pdf");
                 if (!input.ok) return { success: false, output: "", error: input.error };
@@ -655049,6 +655126,7 @@ ${lines.join("\n")}`
             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'.",
+              parameters: this.telegramScopedMediaParameters(tool.parameters, ["path"], "In Telegram scope this is optional; omit it to use the replied-to video first, otherwise the latest chat video."),
               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");
@@ -655061,6 +655139,7 @@ ${lines.join("\n")}`
             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.",
+              parameters: this.telegramScopedMediaParameters(tool.parameters, ["file", "path"], "In Telegram scope file/path is optional; omit it to use the replied-to audio first, otherwise the latest chat audio."),
               execute: async (args) => {
                 if (String(args["action"] || "").toLowerCase() === "listen") {
                   return { success: false, output: "", error: "Continuous microphone listening is not available in Telegram public scope." };
@@ -655139,6 +655218,8 @@ ${lines.join("\n")}`
 
 ${TELEGRAM_PUBLIC_MEMORY_SCOPE_CONTRACT}
 
+${TELEGRAM_EVIDENCE_SUFFICIENCY_CONTRACT}
+
 ${TELEGRAM_PUBLIC_VISION_STACK_CONTRACT}
 
 ${TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT}
@@ -655246,6 +655327,7 @@ Scoped workspace: ${scopedRoot}`,
           new WebSearchTool(),
           new WebCrawlTool(repoRoot),
           // Vision/OCR tools
+          this.buildTelegramImageAnalyzeTool(repoRoot, chatId, msg),
           new ImageReadTool(repoRoot),
           new OCRTool(repoRoot),
           new VisionTool(repoRoot),
@@ -655304,6 +655386,7 @@ Scoped workspace: ${scopedRoot}`,
           new StructuredReadTool(repoRoot),
           new StructuredFileTool(repoRoot),
           new CodeSandboxTool(repoRoot),
+          this.buildTelegramImageAnalyzeTool(repoRoot, chatId, msg),
           new ImageReadTool(repoRoot),
           new ScreenshotTool(repoRoot),
           new OCRTool(repoRoot),
@@ -655424,7 +655507,7 @@ Scoped workspace: ${scopedRoot}`,
       }
       telegramPublicQuotaKind(toolName) {
         if (toolName === "web_fetch") return "web";
-        if (/^(image_read|ocr|ocr_image_advanced|ocr_pdf|pdf_to_text|vision|transcribe_file|video_understand|audio_analyze)$/.test(toolName)) return "media";
+        if (/^(telegram_image_analyze|image_read|ocr|ocr_image_advanced|ocr_pdf|pdf_to_text|vision|transcribe_file|video_understand|audio_analyze)$/.test(toolName)) return "media";
         if (toolName === "generate_video") return "video-generation";
         if (/^(generate_image|generate_audio|generate_model|generate_tts|create_audio_file)$/.test(toolName)) return "generation";
         if (toolName === "telegram_send_file") return "upload";
@@ -656728,11 +656811,144 @@ Scoped workspace: ${scopedRoot}`,
           }
         };
       }
+      buildTelegramImageAnalyzeTool(repoRoot, chatId, currentMsg) {
+        const bridge = this;
+        return {
+          name: "telegram_image_analyze",
+          description: [
+            "Analyze an image from the current Telegram chat using a staged evidence ladder.",
+            "Omit image, or use image='reply', image='latest', message_id:<id>, or a listed basename.",
+            "The tool starts with low-fidelity image intake, uses basic OCR as the text extraction probe, escalates to advanced OCR when text is dense or under-extracted, and escalates to Moondream vision for visual QA/captioning."
+          ].join(" "),
+          parameters: {
+            type: "object",
+            properties: {
+              image: {
+                type: "string",
+                description: "Optional Telegram media reference. Omit for replied-to image first, otherwise latest image in this chat."
+              },
+              question: {
+                type: "string",
+                description: "Optional visual question to answer about the image."
+              },
+              detail: {
+                type: "string",
+                enum: ["auto", "text", "visual", "full"],
+                description: "auto chooses stages from observed extraction signal; text emphasizes OCR; visual emphasizes Moondream; full runs both."
+              }
+            }
+          },
+          async execute(args) {
+            const start2 = performance.now();
+            const resolved = bridge.resolveTelegramScopedMediaPath(args["image"], chatId, currentMsg, "image");
+            if (!resolved.ok) {
+              return { success: false, output: "", error: resolved.error, durationMs: performance.now() - start2 };
+            }
+            const detail = normalizeTelegramImageAnalyzeDetail(args["detail"]);
+            const question = String(args["question"] ?? args["prompt"] ?? currentMsg?.text ?? "").trim();
+            const stageLines = [
+              `Telegram image analysis: media=${basename35(resolved.path)} detail=${detail}`,
+              "Stage ladder: low_fidelity_image_read -> basic OCR probe -> advanced OCR when text density/under-extraction warrants it -> Moondream vision when visual QA/captioning is warranted."
+            ];
+            const llmParts = [];
+            const runStage = async (name10, reason, fn) => {
+              const result = await fn();
+              const body = result.success ? truncateTelegramStageOutput(result.output || result.llmContent || "") : `ERROR: ${result.error || result.output || "stage failed"}`;
+              stageLines.push(`
+## ${name10}
+Reason: ${reason}
+${body}`);
+              llmParts.push(`## ${name10}
+Reason: ${reason}
+${result.llmContent ?? result.output ?? ""}${result.success ? "" : `
+ERROR: ${result.error || "stage failed"}`}`);
+              return result;
+            };
+            const imageRead = await runStage(
+              "low_fidelity_image_read",
+              "Always start with scoped image metadata and a multimodal image payload for the model.",
+              () => new ImageReadTool(repoRoot).execute({
+                path: resolved.path,
+                ocr: false,
+                max_size_kb: typeof args["max_size_kb"] === "number" ? args["max_size_kb"] : 10240
+              })
+            );
+            const textRequested = detail === "text" || detail === "full";
+            const visualRequested = detail === "visual" || detail === "full";
+            const shouldRunTextProbe = detail !== "visual";
+            let textStage = null;
+            let textSignal = telegramTextExtractionSignal("");
+            if (shouldRunTextProbe) {
+              textStage = await runStage(
+                "ocr",
+                "After low-fidelity intake, basic OCR measures whether text extraction is sufficient before any heavier escalation.",
+                () => new OCRTool(repoRoot).execute({
+                  path: resolved.path,
+                  language: typeof args["language"] === "string" ? args["language"] : "eng"
+                })
+              );
+              textSignal = telegramTextExtractionSignal(textStage.output ?? "");
+              stageLines.push(
+                `
+Text extraction signal: chars=${textSignal.chars}, lines=${textSignal.lines}, longest_line=${textSignal.longestLine}, dense=${textSignal.dense ? "yes" : "no"}.`
+              );
+              llmParts.push(`Text extraction signal: ${JSON.stringify(textSignal)}`);
+              const shouldUseAdvancedOcr = detail === "full" || textSignal.dense || textRequested && !textSignal.present;
+              if (shouldUseAdvancedOcr) {
+                textStage = await runStage(
+                  "ocr_image_advanced",
+                  textSignal.dense ? "Basic OCR observed dense text, so multi-pass OCR is warranted." : "The request emphasized text extraction and basic intake did not provide enough text.",
+                  () => new OcrImageAdvancedTool(repoRoot).execute({
+                    image: resolved.path,
+                    language: typeof args["language"] === "string" ? args["language"] : "eng"
+                  })
+                );
+                const advancedSignal = telegramTextExtractionSignal(textStage.output ?? "");
+                stageLines.push(
+                  `
+Advanced text extraction signal: chars=${advancedSignal.chars}, lines=${advancedSignal.lines}, longest_line=${advancedSignal.longestLine}, dense=${advancedSignal.dense ? "yes" : "no"}.`
+                );
+                llmParts.push(`Advanced text extraction signal: ${JSON.stringify(advancedSignal)}`);
+                if (advancedSignal.present || !textSignal.present) textSignal = advancedSignal;
+              }
+            } else {
+              stageLines.push("\n## OCR skipped\nReason: Visual detail was explicitly requested and text extraction was not needed.");
+              llmParts.push("## OCR skipped\nReason: Visual detail was explicitly requested and text extraction was not needed.");
+            }
+            const shouldRunVision = visualRequested || Boolean(question && detail !== "text") || !textSignal.present && detail !== "text";
+            let visionStage = null;
+            if (shouldRunVision) {
+              visionStage = await runStage(
+                "vision_moondream",
+                question ? "A visual question is present, so Moondream visual QA is warranted after lower-cost intake." : "Text extraction did not provide enough content or visual detail was requested, so Moondream captioning is warranted.",
+                () => new VisionTool(repoRoot).execute({
+                  image: resolved.path,
+                  action: question ? "query" : "caption",
+                  ...question ? { prompt: question } : {},
+                  length: "normal"
+                })
+              );
+            } else {
+              stageLines.push("\n## vision_moondream skipped\nReason: OCR/text extraction produced sufficient evidence for the requested detail.");
+              llmParts.push("## vision_moondream skipped\nReason: OCR/text extraction produced sufficient evidence for the requested detail.");
+            }
+            const anySuccess = imageRead.success || Boolean(textStage?.success) || Boolean(visionStage?.success);
+            const output = stageLines.join("\n");
+            return {
+              success: anySuccess,
+              output,
+              llmContent: llmParts.join("\n\n"),
+              error: anySuccess ? void 0 : "Telegram image analysis could not extract usable image evidence.",
+              durationMs: performance.now() - start2
+            };
+          }
+        };
+      }
       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, ocr_image_advanced, vision, identity_memory, transcribe_file, pdf_to_text, video_understand, and audio_analyze.",
+          description: "List recent media files available in this Telegram chat scope, including safe aliases for telegram_image_analyze, image_read, ocr, ocr_image_advanced, vision, identity_memory, transcribe_file, pdf_to_text, video_understand, and audio_analyze.",
           parameters: {
             type: "object",
             properties: {
@@ -657140,7 +657356,7 @@ ${knownList}` : "Private-user telegram_send_file target must be this DM or a kno
             description = `[${sourceLabel}image received: path_alias=${mediaAlias}${safeCaption}
 ${visionContext}]`;
           } else {
-            description = `[${sourceLabel}image received: path_alias=${mediaAlias}${safeCaption}. Use path='${source === "reply" ? "reply" : "latest"}' or path='${mediaAlias}' with image_read, ocr, ocr_image_advanced, vision, or identity_memory.]`;
+            description = `[${sourceLabel}image received: path_alias=${mediaAlias}${safeCaption}. Use image='${source === "reply" ? "reply" : "latest"}' or image='${mediaAlias}' with telegram_image_analyze, or use path/image aliases with image_read, ocr, ocr_image_advanced, vision, or identity_memory.]`;
           }
           const ingestPayload = this.telegramMemoryIngestPayload(msg, media, localPath, source, cacheEntry.extractedContent);
           let visualIdentityContext = "";

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.232",
+  "version": "1.0.233",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.232",
+      "version": "1.0.233",
       "bundleDependencies": [
         "image-to-ascii"
       ],
@@ -6690,9 +6690,9 @@
       }
     },
     "node_modules/streamx": {
-      "version": "2.26.0",
-      "resolved": "https://registry.npmjs.org/streamx/-/streamx-2.26.0.tgz",
-      "integrity": "sha512-VvNG1K72Po/xwJzxZFnZ++Tbrv4lwSptsbkFuzXCJAYZvCK5nnxsvXU6ajqkv7chyiI1Y0YXq2Jh8Iy8Y7NF/A==",
+      "version": "2.27.0",
+      "resolved": "https://registry.npmjs.org/streamx/-/streamx-2.27.0.tgz",
+      "integrity": "sha512-WZ189TKnHoAokYHvwzaAQMpd55cgUmFIcJFzBSgGcb886jau5DL+XdDhTWV4ps3FLvk+OORp0dLRTPsLZ21CSA==",
       "license": "MIT",
       "optional": true,
       "dependencies": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.232",
+  "version": "1.0.233",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",