omnius 1.0.173 → 1.0.174

package/dist/index.js

@@ -562106,6 +562106,7 @@ ${description}`
         if (effectiveThink === true && (effectiveMaxTokens ?? 0) < 4096) {
           effectiveMaxTokens = 4096;
         }
+        const responseFormat = request.responseFormat ?? request.response_format;
         const body = {
           model: this.model,
           messages: cleanedMessages,
@@ -562116,6 +562117,16 @@ ${description}`
           stream_options: { include_usage: true },
           think: effectiveThink
         };
+        if (responseFormat !== void 0) {
+          body["response_format"] = responseFormat;
+        }
+        const reqNumCtx = request.numCtx;
+        if (Number.isFinite(reqNumCtx) && (reqNumCtx ?? 0) > 0) {
+          const opts = body["options"] ?? {};
+          opts["num_ctx"] = reqNumCtx;
+          body["options"] = opts;
+          body["num_ctx"] = reqNumCtx;
+        }
         let poolSlot = shouldUseOllamaPoolForBaseUrl(this.baseUrl) ? await getOllamaPool({ baseInstanceUrl: this.baseUrl }).acquire({
           model: this.model
         }) : null;
@@ -633170,7 +633181,8 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
             backend,
             { ...suppressed, responseFormat: TELEGRAM_INTERACTION_DECISION_RESPONSE_FORMAT },
             inferenceKind,
-            sessionKey
+            sessionKey,
+            { stream: false, reason: "router-json" }
           );
           const visible = jsonModeResult.choices.some(
             (choice) => stripTelegramHiddenThinking(choice.message.content ?? "").trim().length > 0
@@ -633192,7 +633204,8 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
             backend,
             suppressed,
             inferenceKind,
-            sessionKey
+            sessionKey,
+            { stream: false, reason: "router-plain-retry" }
           );
           if (diagnostics) {
             const plainVisible = plainResult.choices.some(
@@ -633231,7 +633244,7 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
        *      content tokens emitted, the registry shows it, and the
        *      hard-deadline retire path becomes diagnosable instead of opaque
        */
-      async telegramObservableInference(backend, request, kind, sessionKey) {
+      async telegramObservableInference(backend, request, kind, sessionKey, options2 = {}) {
         const model = this.agentConfig?.model ?? "?";
         const promptTokens = estimatePromptTokensFromRequest(request);
         const broker = getModelBroker();
@@ -633259,7 +633272,17 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
         let completionTokens = 0;
         try {
           let result;
-          if (typeof streamFn !== "function") {
+          const streamAllowed = options2.stream !== false;
+          if (!streamAllowed) {
+            result = await backend.chatCompletion(requestWithCtx);
+            this.updateTelegramInferenceFinal(id, result);
+            if (!streamAllowed && process.env["OMNIUS_BROKER_TRACE"] === "1") {
+              this.tuiWrite(() => renderTelegramSubAgentEvent(
+                sessionKey,
+                `inference ${id}: non-stream direct (${options2.reason ?? "requested"}) ${this.telegramInferenceRequestDiagnostic(requestWithCtx)}`
+              ));
+            }
+          } else if (typeof streamFn !== "function") {
             result = await backend.chatCompletion(requestWithCtx);
             this.updateTelegramInferenceFinal(id, result);
           } else {
@@ -633289,6 +633312,28 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
           this.deregisterTelegramInference(id);
         }
       }
+      telegramBackendDiagnostic() {
+        const config = this.agentConfig;
+        if (!config) return "backend=unconfigured model=?";
+        return `backend=${config.backendType} url=${config.backendUrl} model=${config.model}`;
+      }
+      telegramInferenceRequestDiagnostic(request) {
+        const responseFormat = request.responseFormat ?? request.response_format;
+        const responseFormatType = responseFormat && typeof responseFormat["type"] === "string" ? responseFormat["type"] : responseFormat ? "present" : "none";
+        const numCtx = request.numCtx;
+        const think = request.think;
+        const tools = Array.isArray(request.tools) ? request.tools.length : 0;
+        return `${this.telegramBackendDiagnostic()} response_format=${responseFormatType} num_ctx=${Number.isFinite(numCtx) ? numCtx : "unset"} think=${think === void 0 ? "default" : String(think)} tools=${tools} timeoutMs=${Number.isFinite(request.timeoutMs) ? request.timeoutMs : "unset"}`;
+      }
+      telegramStreamInactivityDiagnostic(request, inferenceId, inactivityMs, contentChars, thinkingChars) {
+        const entry = this.telegramActiveInferences.get(inferenceId);
+        const now = performance.now();
+        const elapsed = entry ? `${((now - entry.startTs) / 1e3).toFixed(1)}s` : "unknown";
+        const idle = entry ? `${((now - entry.lastTokenAt) / 1e3).toFixed(1)}s` : "unknown";
+        const ttfb = entry?.firstChunkAt !== void 0 ? `${((entry.firstChunkAt - entry.startTs) / 1e3).toFixed(1)}s` : "never";
+        const phase = entry?.firstChunkAt === void 0 ? "before-first-chunk" : "mid-stream";
+        return `stream-inactivity: no chunks for ${(inactivityMs / 1e3).toFixed(0)}s (phase=${phase}; elapsed=${elapsed}; idle=${idle}; ttfb=${ttfb}; content=${contentChars}c thinking=${thinkingChars}c; ${this.telegramInferenceRequestDiagnostic(request)}; stream_endpoint=no-sse-chunk)`;
+      }
       /**
        * Drive a chatCompletionStream to exhaustion, accumulating tokens into a
        * chatCompletion-shaped result. Live-emits content + thinking tokens
@@ -633325,7 +633370,13 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`,
             const inactivityPromise = new Promise((_, reject) => {
               timeoutHandle = setTimeout(
                 () => reject(new Error(
-                  `stream-inactivity: no chunks for ${(inactivityMs / 1e3).toFixed(0)}s (content=${contentBuf.length}c thinking=${thinkingBuf.length}c so far) — Ollama likely cold-loading the model or wedged; falling back to non-stream`
+                  this.telegramStreamInactivityDiagnostic(
+                    request,
+                    inferenceId,
+                    inactivityMs,
+                    contentBuf.length,
+                    thinkingBuf.length
+                  )
                 )),
                 inactivityMs
               );
@@ -634180,6 +634231,7 @@ ${this.quoteTelegramContextBlock(msg.text, 1200)}`
       composeTelegramRouterDiagnosticNote(invalidRouterPreview, failureNarrative, headline) {
         const segments = [];
         segments.push(headline);
+        segments.push(this.telegramBackendDiagnostic());
         if (failureNarrative.summary) segments.push(failureNarrative.summary);
         if (invalidRouterPreview) segments.push(`invalid router output preview: ${invalidRouterPreview}`);
         if (failureNarrative.detail) segments.push(`router-failure trace: ${failureNarrative.detail}`);

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.173",
+  "version": "1.0.174",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.173",
+      "version": "1.0.174",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.173",
+  "version": "1.0.174",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",