omnius 1.0.169 → 1.0.171

package/dist/index.js

@@ -546736,14 +546736,30 @@ function estimateTokens2(text) {
 function normalizeSignalContent(content) {
   return content.replace(/\r\n/g, "\n").split("\n").map((line) => line.trimEnd()).join("\n").trim();
 }
+function urlSpanAt(text, index) {
+  URL_RE.lastIndex = 0;
+  for (const match of text.matchAll(URL_RE)) {
+    const start2 = match.index ?? 0;
+    const end = start2 + match[0].length;
+    if (start2 < index && index < end)
+      return { start: start2, end };
+  }
+  return null;
+}
 function truncateText2(text, maxChars) {
   if (text.length <= maxChars)
     return { text, truncated: false };
   if (maxChars <= 40)
     return { text: text.slice(0, maxChars), truncated: true };
+  const suffix = CONTEXT_FABRIC_TRUNCATION_SUFFIX;
+  let cutIndex = maxChars - suffix.length;
+  const urlSpan = urlSpanAt(text, cutIndex);
+  if (urlSpan) {
+    cutIndex = urlSpan.start;
+  }
+  const prefix = text.slice(0, Math.max(0, cutIndex)).trimEnd();
   return {
-    text: `${text.slice(0, maxChars - 36).trimEnd()}
-... [truncated by context fabric]`,
+    text: `${prefix}${suffix}`,
     truncated: true
   };
 }
@@ -546812,7 +546828,7 @@ function signalFromBlock(kind, source, content, options2 = {}) {
     metadata: options2.metadata
   };
 }
-var KIND_ORDER, KIND_LABELS, DEFAULT_KIND_CHAR_BUDGET, ContextLedger, ContextFrameBuilder;
+var KIND_ORDER, KIND_LABELS, DEFAULT_KIND_CHAR_BUDGET, URL_RE, CONTEXT_FABRIC_TRUNCATION_SUFFIX, ContextLedger, ContextFrameBuilder;
 var init_context_fabric = __esm({
   "packages/orchestrator/dist/context-fabric.js"() {
     "use strict";
@@ -546864,6 +546880,8 @@ var init_context_fabric = __esm({
       environment: 900,
       compaction_summary: 1200
     };
+    URL_RE = /\bhttps?:\/\/[^\s<>"'()[\]{}]+/gi;
+    CONTEXT_FABRIC_TRUNCATION_SUFFIX = "\n... [truncated by context fabric]";
     ContextLedger = class {
       signals = /* @__PURE__ */ new Map();
       sequence = 0;
@@ -592291,7 +592309,7 @@ function extractMediaReferences(text) {
     const offset = match[0].indexOf("MEDIA:");
     addCandidate(value2, match[0].slice(offset), match.index + offset, match.index + match[0].length);
   }
-  for (const match of matchAll(URL_RE, text)) {
+  for (const match of matchAll(URL_RE2, text)) {
     const original = trimTrailingPunctuation(match[0]);
     if (original.length !== match[0].length) {
       addCandidate(original, original, match.index, match.index + original.length);
@@ -592415,7 +592433,7 @@ function decodeUri(value2) {
 function trimTrailingPunctuation(value2) {
   return value2.replace(/[),.;:!?]+$/, "");
 }
-var IMAGE_EXT, AUDIO_EXT, VIDEO_EXT, DOC_EXT, MEDIA_DIRECTIVE_RE, AUDIO_AS_VOICE_RE, MARKDOWN_IMAGE_RE, HTML_IMG_RE, URL_RE, LOCAL_PATH_RE;
+var IMAGE_EXT, AUDIO_EXT, VIDEO_EXT, DOC_EXT, MEDIA_DIRECTIVE_RE, AUDIO_AS_VOICE_RE, MARKDOWN_IMAGE_RE, HTML_IMG_RE, URL_RE2, LOCAL_PATH_RE;
 var init_media_routing = __esm({
   "packages/cli/src/tui/media-routing.ts"() {
     "use strict";
@@ -592447,7 +592465,7 @@ var init_media_routing = __esm({
     AUDIO_AS_VOICE_RE = /\[\[audio_as_voice\]\]/gi;
     MARKDOWN_IMAGE_RE = /!\[[^\]]*]\(([^)\s]+)(?:\s+"[^"]*")?\)/g;
     HTML_IMG_RE = /<img\b[^>]*\bsrc\s*=\s*["']([^"']+)["'][^>]*>/gi;
-    URL_RE = /\bhttps?:\/\/[^\s<>"')]+/gi;
+    URL_RE2 = /\bhttps?:\/\/[^\s<>"')]+/gi;
     LOCAL_PATH_RE = /(?:^|[\s([])((?:~|\/)[^\s<>"')\]]+\.(?:png|jpe?g|webp|gif|bmp|tiff?|svg|wav|mp3|ogg|oga|opus|m4a|flac|aac|mp4|mov|mkv|webm|avi|m4v|pdf|txt|md|markdown|csv|tsv|jsonl?|xlsx?|docx?|pptx?|html?|xml|zip|tar|gz))(?:$|[\s),.\]])/gim;
   }
 });
@@ -615859,9 +615877,10 @@ function buildRealtimeSystemPrompt(opts) {
       `- Produce one natural spoken turn, normally ${maxReplyWords} words or fewer.`,
       "- Use one sentence when possible; two short sentences only when repair or confirmation needs it.",
       "- Lead with the answer. Do not preface with status, analysis, summaries, or implementation narration.",
-      "- No markdown, bullets, tables, headings, citations, code blocks, JSON, or labels like 'Assistant:'.",
+      "- No markdown, bullets, tables, headings, citations, inline code, code blocks, JSON, or labels like 'Assistant:'.",
       "- Sound like a person on a live call: brief acknowledgment, direct answer, one focused follow-up only if needed.",
       "- If the ASR text is garbled or underspecified, ask a single compact repair question.",
+      "- Do not invent app modes, method names, settings, or implementation details when the caller has not supplied them.",
       "- Do not mention ASR, TTS, prompts, realtime mode, hidden reasoning, tools, or policy unless the caller explicitly asks.",
       "- If a request needs work outside this text-only exchange, say the next handoff in one short sentence."
     ].join("\n"),
@@ -615952,7 +615971,7 @@ function wordParts(text) {
 }
 function finalizeRealtimeReply(text, opts = {}) {
   const maxWords = clampInt2(opts.maxReplyWords, DEFAULT_REALTIME_MAX_REPLY_WORDS, 8, 80);
-  let clean5 = stripHiddenThinking(String(text ?? "")).replace(/```[\s\S]*?```/g, "").split("\n").map((line) => line.replace(/^\s*(?:[-*]+|\d+[.)])\s+/, "").trim()).filter(Boolean).join(" ").replace(/^(?:assistant|omnius|agent)\s*:\s*/i, "").replace(/\s+/g, " ").trim();
+  let clean5 = stripHiddenThinking(String(text ?? "")).replace(/```[\s\S]*?```/g, "").split("\n").map((line) => line.replace(/^\s*(?:[-*]+|\d+[.)])\s+/, "").trim()).filter(Boolean).join(" ").replace(/^(?:assistant|omnius|agent)\s*:\s*/i, "").replace(/`([^`]+)`/g, "$1").replace(/\s+/g, " ").trim();
   if (!clean5) return "I didn't catch that. Can you say it again?";
   const sentences = clean5.match(/[^.!?]+[.!?]+(?=\s|$)|[^.!?]+$/g) ?? [clean5];
   const selected = [];
@@ -627419,7 +627438,7 @@ function parseTelegramReflectionFollowupDecision(text) {
   }
 }
 function convertMarkdownToTelegramHTML(md) {
-  let html = md;
+  let html = normalizeTelegramOutboundLinks(md);
   html = html.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
   html = html.replace(
     /```[\w]*\n([\s\S]*?)```/g,
@@ -627439,6 +627458,13 @@ function convertMarkdownToTelegramHTML(md) {
 function escapeTelegramHTML(text) {
   return text.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
 }
+function normalizeTelegramOutboundLinks(text) {
+  if (!text) return text;
+  return text.replace(
+    TELEGRAM_SPACED_URL_RE,
+    (match) => match.replace(/^(https?:\/\/)\s*/i, "$1").replace(/\s*\.\s*/g, ".").replace(/\s*:\s*(\d+)/g, ":$1")
+  );
+}
 function splitTelegramMessageText(text, maxLength = 3900) {
   const trimmed = text.trim();
   if (!trimmed) return [];
@@ -627544,7 +627570,7 @@ function stripTelegramStuckSelfTalk(text) {
   return kept.join("\n\n");
 }
 function cleanTelegramVisibleReply(text, options2 = {}) {
-  const clean5 = stripTelegramHiddenThinking(text).trim();
+  const clean5 = normalizeTelegramOutboundLinks(stripTelegramHiddenThinking(text)).trim();
   if (!clean5) return "";
   if (options2.suppressPotentialNoReplyPrefix && isTelegramPotentialNoReplyPrefix(clean5)) return "";
   if (isTelegramInternalStatusText(clean5)) return "";
@@ -628621,7 +628647,7 @@ function renderTelegramSubAgentError(username, error) {
   process.stdout.write(`    ${c3.dim("│")} ${c3.magenta("✘")} @${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_INTERACTION_DECISION_RESPONSE_FORMAT, TELEGRAM_INTERACTION_DECISION_MINIMAL_SCHEMA, TELEGRAM_INTERACTION_DECISION_REPAIR_SCHEMA, TELEGRAM_CHAT_REPLY_RESPONSE_FORMAT, 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_SUB_AGENT_DEFAULT_LIMIT, TELEGRAM_SUB_AGENT_MAX_LIMIT, TELEGRAM_SUB_AGENT_BURST_CONTEXT_LIMIT, 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_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, 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_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_SUB_AGENT_DEFAULT_LIMIT, TELEGRAM_SUB_AGENT_MAX_LIMIT, TELEGRAM_SUB_AGENT_BURST_CONTEXT_LIMIT, 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_PUBLIC_TOOL_QUOTAS, TelegramBridge;
 var init_telegram_bridge = __esm({
   "packages/cli/src/tui/telegram-bridge.ts"() {
     "use strict";
@@ -628889,6 +628915,13 @@ External acquisition contract:
 - Treat repeated invalid downloads (saved HTML/JSON/login pages instead of the expected file type) as the same blocker; do not retry past a small bounded number of attempts.
 - The browser tool is for interactive web work; it does not save arbitrary rendered files to disk. For file acquisition, use the download/file tool and verify content-type + size before reporting success.
 - Report the exact blocker concisely and offer lawful alternatives (library hold, official store, summary/discussion if the user supplies an authorized copy).
+`.trim();
+    TELEGRAM_LINK_INTEGRITY_CONTRACT = `
+Telegram link integrity contract:
+- When sending public http(s) URLs, copy the URL exactly from source/tool output.
+- Never defang, space out, redact, or rewrite dots, slashes, host labels, or TLDs in normal public URLs.
+- Prefer Markdown links when helpful, but keep the href URL contiguous and parseable.
+- If you are uncertain a URL is real, say so or omit it instead of reconstructing a broken URL.
 `.trim();
     TELEGRAM_INTERACTION_DECISION_RESPONSE_FORMAT = {
       type: "json_object"
@@ -628913,6 +628946,7 @@ External acquisition contract:
         }
       }
     };
+    TELEGRAM_SPACED_URL_RE = /\bhttps?:\/\/\s*[A-Za-z0-9-]+(?:\s*\.\s*[A-Za-z0-9-]+)+(?:\s*:\s*\d+)?(?:[/?#][^\s<>"'()[\]{}]*)?/gi;
     TELEGRAM_STUCK_SELF_TALK_PREFIXES = [
       /^i'?ve been stuck for\b/i,
       /^i am (still |currently )?stuck\b/i,
@@ -634268,7 +634302,8 @@ Todo/session id: ${sessionId}
 Profile: ${profile}
 Tool context: ${toolContext}
 ${chatLabel}`,
-          TELEGRAM_ACTION_RESPONSE_CONTRACT
+          TELEGRAM_ACTION_RESPONSE_CONTRACT,
+          TELEGRAM_LINK_INTEGRITY_CONTRACT
         ];
         sections.push(conversationStream);
         const sendTargetContext = this.buildTelegramSendTargetContext(msg, isAdminDM);
@@ -634637,10 +634672,11 @@ ${TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT}`);
        * streamed content. Returns the message_id for subsequent edits.
        */
       async sendLiveMessage(chatId, initialText, replyToMessageId) {
+        const normalizedInitialText = normalizeTelegramOutboundLinks(initialText);
         try {
           const body = {
             chat_id: chatId,
-            text: initialText,
+            text: normalizedInitialText,
             parse_mode: "HTML"
           };
           const replyParameters = telegramReplyParameters(replyToMessageId);
@@ -634657,7 +634693,7 @@ ${TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT}`);
           try {
             const body = {
               chat_id: chatId,
-              text: initialText
+              text: normalizedInitialText
             };
             const replyParameters = telegramReplyParameters(replyToMessageId);
             if (replyParameters) body["reply_parameters"] = replyParameters;
@@ -634692,7 +634728,8 @@ ${TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT}`);
        * Handles rate limits and Telegram's "message is not modified" error silently.
        */
       async editLiveMessage(chatId, messageId, html) {
-        const truncated = html.length > 4e3 ? html.slice(0, 3950) + "\n\n<i>... (streaming)</i>" : html;
+        const normalizedHtml = normalizeTelegramOutboundLinks(html);
+        const truncated = normalizedHtml.length > 4e3 ? normalizedHtml.slice(0, 3950) + "\n\n<i>... (streaming)</i>" : normalizedHtml;
         try {
           const result = await this.apiCall("editMessageText", {
             chat_id: chatId,
@@ -634705,7 +634742,7 @@ ${TELEGRAM_PUBLIC_ORCHESTRATOR_CONTRACT}`);
           return true;
         } catch {
           try {
-            const plain = html.replace(/<[^>]+>/g, "");
+            const plain = normalizedHtml.replace(/<[^>]+>/g, "");
             const truncPlain = plain.length > 4e3 ? plain.slice(0, 3950) + "\n\n... (streaming)" : plain;
             const result = await this.apiCall("editMessageText", {
               chat_id: chatId,
@@ -635404,6 +635441,8 @@ ${TELEGRAM_PUBLIC_VISION_STACK_CONTRACT}`;
             role: "system",
             content: `${TELEGRAM_CHAT_MODE_PROMPT}
 
+${TELEGRAM_LINK_INTEGRITY_CONTRACT}
+
 ## Runtime Context
 
 ${runtime}
@@ -635729,6 +635768,8 @@ ${GROUP_REPLY_DISCRETION_PROMPT}` : "";
 
 ${TELEGRAM_ACTION_RESPONSE_CONTRACT}
 
+${TELEGRAM_LINK_INTEGRITY_CONTRACT}
+
 ${TELEGRAM_EXTERNAL_ACQUISITION_CONTRACT}
 
 ${reminderToolContract}
@@ -635741,6 +635782,7 @@ ${currentTelegramPrompt}`;
             "You have access to isolated per-chat memory (memory_write, memory_read, memory_search) scoped to this conversation.",
             "memory_search may use scope=group/current_chat for this group or scope=user with user_id/username for a participant in this same group. Other groups, admin chats, and private DMs are not accessible here.",
             "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'.",
             formatIdentityMemoryContext(chatLabel || "Telegram private chat"),
@@ -638321,9 +638363,10 @@ ${text}`.trim());
         return result.messageId ?? null;
       }
       async sendMessageHTMLDetailed(chatId, html, replyToMessageId, options2 = {}) {
-        const extracted = extractMediaReferences(html);
+        const normalizedHtml = normalizeTelegramOutboundLinks(html);
+        const extracted = extractMediaReferences(normalizedHtml);
         const mediaRefs = this.filterTelegramMediaReferences(extracted.media, options2);
-        const sendHtml = extracted.text || (extracted.media.length > 0 ? "" : html);
+        const sendHtml = extracted.text || (extracted.media.length > 0 ? "" : normalizedHtml);
         let sentId = null;
         if (!sendHtml.trim()) {
           for (let idx = 0; idx < mediaRefs.length; idx++) {
@@ -638790,7 +638833,8 @@ Content-Type: ${mimeForPath(pathOrFileId, field === "photo" ? "image" : "video")
        * article carrying the message text Telegram should send to the guest chat.
        */
       async answerGuestQuery(guestQueryId, text, options2 = {}) {
-        const truncated = text.length > 4e3 ? text.slice(0, 3950) + "\n\n... (truncated)" : text;
+        const normalizedText = normalizeTelegramOutboundLinks(text);
+        const truncated = normalizedText.length > 4e3 ? normalizedText.slice(0, 3950) + "\n\n... (truncated)" : normalizedText;
         const inputMessageContent = {
           message_text: truncated
         };
@@ -657279,6 +657323,49 @@ function bodyString(body, keys) {
   }
   return "";
 }
+function realtimeFallbackCacheKey(ollamaUrl, missingModel) {
+  return `${ollamaUrl}
+${missingModel}`;
+}
+function isOllamaMissingModelError(body) {
+  return /model ['\"]?[^'\"]+['\"]? not found/i.test(body);
+}
+async function resolveRealtimeOllamaFallbackModel(ollamaUrl, timeoutMs, missingModel) {
+  try {
+    const cacheKey = realtimeFallbackCacheKey(ollamaUrl, missingModel);
+    const cached = realtimeOllamaFallbackCache.get(cacheKey);
+    if (cached) return cached;
+    const result = await ollamaRequest(ollamaUrl, "/api/tags", "GET", void 0, Math.min(timeoutMs, 1e4));
+    if (result.status >= 400) return null;
+    const parsed = JSON.parse(result.body);
+    const names = (parsed.models ?? []).map((entry) => typeof entry.name === "string" ? entry.name : typeof entry.model === "string" ? entry.model : "").filter(Boolean);
+    if (!names.length) return null;
+    const remember = (name10) => {
+      realtimeOllamaFallbackCache.set(cacheKey, name10);
+      return name10;
+    };
+    const exactLatest = `${missingModel}:latest`;
+    if (names.includes(exactLatest)) return remember(exactLatest);
+    const preferred = [
+      "qwen3.5-9b-r10:q4km",
+      "open-agents-qwen35-9b-r10-q4km:latest",
+      "open-agents-qwen35-9b-r10-parsed-q4km:latest",
+      "open-agents-qwen35-9b-r9-q4km:latest",
+      "qwen3:8b",
+      "open-agents-qwen3-8b:latest",
+      "omnius-qwen36-35b:latest",
+      "open-agents-qwen36:latest",
+      "qwen3.6:35b"
+    ];
+    for (const name10 of preferred) {
+      if (names.includes(name10)) return remember(name10);
+    }
+    const fallback = names.find((name10) => /qwen/i.test(name10) && !/embed|vision/i.test(name10)) ?? names.find((name10) => !/embed|vision|moondream/i.test(name10)) ?? null;
+    return fallback ? remember(fallback) : null;
+  } catch {
+    return null;
+  }
+}
 function realtimeEndpointMessages(body) {
   const messages2 = [];
   const suppliedSoul = bodyString(body, ["soul_md", "soul", "soulMd"]);
@@ -657305,13 +657392,14 @@ ${suppliedContext}` });
 }
 async function completeRealtimeTextOnly(opts) {
   const cfg = loadConfig();
-  const model = bodyString(opts.body, ["model"]) || cfg.model;
+  const requestedModel = bodyString(opts.body, ["model"]);
+  const model = requestedModel || opts.defaultModel || cfg.model;
   const route = resolveModelEndpoint(model);
   const limitErr = route?.endpoint ? checkEndpointRateLimit(route.endpoint) : null;
   if (limitErr) throw new Error(limitErr);
   const targetUrl = route?.endpoint.url ?? opts.ollamaUrl;
-  const targetType = route?.endpoint.type ?? cfg.backendType ?? "ollama";
-  const originalModel = route?.originalId ?? model.replace(/^[a-z]+\//, "");
+  const targetType = route?.endpoint.type ?? opts.defaultBackendType ?? cfg.backendType ?? "ollama";
+  let originalModel = route?.originalId ?? model.replace(/^[a-z]+\//, "");
   const realtimeOpts = {
     ...realtimeOptionsFromBody(opts.body, process.cwd(), opts.sessionId),
     surface: "voice_adapter"
@@ -657333,13 +657421,24 @@ async function completeRealtimeTextOnly(opts) {
   }
   const maxTokens = typeof requestBody["max_tokens"] === "number" ? requestBody["max_tokens"] : 120;
   const temperature = typeof requestBody["temperature"] === "number" ? requestBody["temperature"] : 0.6;
-  const result = await ollamaRequest(targetUrl, "/api/chat", "POST", JSON.stringify({
-    model: originalModel,
+  if (!requestedModel) {
+    originalModel = realtimeOllamaFallbackCache.get(realtimeFallbackCacheKey(targetUrl, originalModel)) ?? originalModel;
+  }
+  const makeOllamaChatBody = (modelName) => JSON.stringify({
+    model: modelName,
     messages: requestBody["messages"],
     stream: false,
     think: false,
     options: { temperature, num_predict: maxTokens }
-  }), timeoutMs, route?.endpoint);
+  });
+  let result = await ollamaRequest(targetUrl, "/api/chat", "POST", makeOllamaChatBody(originalModel), timeoutMs, route?.endpoint);
+  if (result.status >= 400 && !requestedModel && isOllamaMissingModelError(result.body)) {
+    const fallbackModel = await resolveRealtimeOllamaFallbackModel(targetUrl, timeoutMs, originalModel);
+    if (fallbackModel && fallbackModel !== originalModel) {
+      originalModel = fallbackModel;
+      result = await ollamaRequest(targetUrl, "/api/chat", "POST", makeOllamaChatBody(originalModel), timeoutMs, route?.endpoint);
+    }
+  }
   if (result.status >= 400) throw new Error(`Backend HTTP ${result.status}: ${result.body.slice(0, 300)}`);
   const parsed = JSON.parse(result.body);
   const rawReply = String(parsed?.message?.content ?? "").trim();
@@ -657354,7 +657453,7 @@ async function completeRealtimeTextOnly(opts) {
     }
   };
 }
-async function handleRealtimeText(req2, res, ollamaUrl) {
+async function handleRealtimeText(req2, res, ollamaUrl, defaults3 = {}) {
   const body = await parseJsonBody(req2);
   if (!body || typeof body !== "object") {
     jsonResponse(res, 400, { error: "invalid_request", message: "Expected a JSON object." });
@@ -657367,7 +657466,14 @@ async function handleRealtimeText(req2, res, ollamaUrl) {
   }
   try {
     const sessionId = typeof body["session_id"] === "string" ? body["session_id"] : void 0;
-    const result = await completeRealtimeTextOnly({ body, messages: messages2, ollamaUrl, sessionId });
+    const result = await completeRealtimeTextOnly({
+      body,
+      messages: messages2,
+      ollamaUrl,
+      defaultModel: defaults3.model,
+      defaultBackendType: defaults3.backendType,
+      sessionId
+    });
     const wantsPlain = String(req2.headers["accept"] ?? "").includes("text/plain") || body["format"] === "text";
     if (wantsPlain) {
       res.writeHead(200, { "Content-Type": "text/plain; charset=utf-8", "Cache-Control": "no-store" });
@@ -660331,7 +660437,7 @@ async function handlePostCommand(res, cmd) {
     });
   }
 }
-async function handleRequest(req2, res, ollamaUrl, verbose) {
+async function handleRequest(req2, res, ollamaUrl, verbose, runtimeDefaults = {}) {
   try {
     const _liveCfg = loadConfig();
     if (_liveCfg.backendUrl) ollamaUrl = _liveCfg.backendUrl;
@@ -660606,7 +660712,7 @@ async function handleRequest(req2, res, ollamaUrl, verbose) {
         status = 401;
         return;
       }
-      await handleRealtimeText(req2, res, ollamaUrl);
+      await handleRealtimeText(req2, res, ollamaUrl, runtimeDefaults);
       return;
     }
     if (pathname === "/v1/files" && method === "GET") {
@@ -662614,13 +662720,14 @@ ${historyLines}
       }));
     }
   } finally {
-    recordMetric(method, pathname, status);
+    const finalStatus = res.headersSent ? res.statusCode : status;
+    recordMetric(method, pathname, finalStatus);
     const latencyMs = Math.round(performance.now() - startMs);
     logRequest({
       requestId,
       method,
       path: pathname,
-      status,
+      status: finalStatus,
       latencyMs,
       user: req2._authUser ?? "anonymous",
       scope: req2._authScope ?? "none"
@@ -662630,7 +662737,7 @@ ${historyLines}
       requestId,
       method,
       path: pathname,
-      status,
+      status: finalStatus,
       user: req2._authUser ?? "anonymous",
       scope: req2._authScope ?? "none",
       latencyMs: Math.round(performance.now() - startMs),
@@ -663552,7 +663659,10 @@ function startApiServer(options2 = {}) {
       }
     } catch {
     }
-    handleRequest(req2, res, ollamaUrl, verbose).catch((err) => {
+    handleRequest(req2, res, ollamaUrl, verbose, {
+      model: options2.model ?? config.model,
+      backendType: options2.backendType ?? config.backendType
+    }).catch((err) => {
       metrics.totalErrors++;
       try {
         jsonResponse(res, 500, {
@@ -664374,7 +664484,9 @@ async function apiServeCommand(opts, config) {
     port: opts.port,
     // Let startApiServer() parse OMNIUS_HOST env if no explicit --port
     verbose: opts.verbose,
-    ollamaUrl: config.backendUrl
+    ollamaUrl: config.backendUrl,
+    model: config.model,
+    backendType: config.backendType
   });
   await new Promise((resolve57) => {
     server2.on("close", resolve57);
@@ -664429,7 +664541,7 @@ function setTimerEnabled(name10, enabled2) {
     return false;
   }
 }
-var require4, NEXUS_DIRECTORY_ORIGIN2, NEXUS_SPONSORS_URL2, endpointRegistry, modelRouteMap, endpointUsage, _lastEndpointDiagnostics, BACKEND_TIMEOUT_DEFAULT_MS, BACKEND_TIMEOUT_MAX_MS, MODEL_LIST_TIMEOUT_DEFAULT_MS, metrics, startedAt, runningProcesses, perKeyUsage, CRON_MARKER2;
+var require4, NEXUS_DIRECTORY_ORIGIN2, NEXUS_SPONSORS_URL2, endpointRegistry, modelRouteMap, endpointUsage, _lastEndpointDiagnostics, BACKEND_TIMEOUT_DEFAULT_MS, BACKEND_TIMEOUT_MAX_MS, MODEL_LIST_TIMEOUT_DEFAULT_MS, metrics, startedAt, realtimeOllamaFallbackCache, runningProcesses, perKeyUsage, CRON_MARKER2;
 var init_serve = __esm({
   "packages/cli/src/api/serve.ts"() {
     "use strict";
@@ -664477,6 +664589,7 @@ var init_serve = __esm({
       totalErrors: 0
     };
     startedAt = Date.now();
+    realtimeOllamaFallbackCache = /* @__PURE__ */ new Map();
     runningProcesses = /* @__PURE__ */ new Map();
     perKeyUsage = /* @__PURE__ */ new Map();
     CRON_MARKER2 = "# OMNIUS-SCHEDULED:";

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.169",
+  "version": "1.0.171",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.169",
+      "version": "1.0.171",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.169",
+  "version": "1.0.171",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",