omnius 1.0.354 → 1.0.355

package/dist/index.js

@@ -570895,25 +570895,24 @@ Your hypotheses MUST address this specific error, not generic causes.
         return best && best.count >= 3 ? best : null;
       }
       /**
-       * Backend adapter for AUXILIARY inference (adversary critiques, branch
-       * extraction) — tool-less, think-off, JSON-shaped calls. The main backend's
-       * chatCompletion routes to Ollama's /v1/chat/completions, where qwen3-family
-       * models IGNORE think:false and /no_think and (with no tools to anchor
-       * output) emit a reasoning-only response that gets stripped to EMPTY. The
-       * native /api/chat path honors think:false. This adapter prefers it and sets
-       * a responseFormat so the native path enforces JSON mode. Falls back to
-       * chatCompletion for non-Ollama backends.
+       * Backend adapter for AUXILIARY inference (adversary critiques, resolution
+       * gate, branch extraction) — tool-less, think-off, JSON-shaped calls. Uses the
+       * SAME inference backend the main agent loop uses — whatever provider the user
+       * selected — and only ensures a JSON responseFormat for these structured
+       * calls. Provider-specific quirks (e.g. Ollama /v1 + qwen3 returning an empty
+       * reasoning-only response for tool-less think-off calls) are recovered INSIDE
+       * the backend's own chatCompletion (empty-recovery → native /api/chat
+       * fallback), so no caller here is Ollama-aware.
        */
       _auxInferenceBackend() {
         const b = this.backend;
-        const useNative = typeof b.nativeOllamaChatCompletion === "function";
         return {
           chatCompletion: (req3) => {
             const r2 = {
               ...req3,
               responseFormat: req3.responseFormat ?? { type: "json_object" }
             };
-            return useNative ? b.nativeOllamaChatCompletion(r2) : b.chatCompletion(r2);
+            return b.chatCompletion(r2);
           }
         };
       }
@@ -585983,6 +585982,15 @@ ${description}`
           const justSuppressed = this._thinkSuppressed && this._thinkFailStreak === _OllamaAgenticBackend._thinkFailThreshold;
           const shouldRetryThinkGuard = outcome !== null && effectiveThink === true && (justSuppressed || outcome === "empty_after_strip" || outcome === "unclosed_think");
           if (shouldRetryThinkGuard || shouldRecoverFromEmpty) {
+            if (shouldRecoverFromEmpty) {
+              try {
+                const _native = await this.nativeOllamaChatCompletion(request);
+                const _nText = String(_native.choices?.[0]?.message?.content ?? "");
+                if (_nText.trim().length >= 2)
+                  return _native;
+              } catch {
+              }
+            }
             const retryMessages = injectNoThinkDirective(requestMessages);
             const retryBody = {
               model: this.model,

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.354",
+  "version": "1.0.355",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.354",
+      "version": "1.0.355",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.354",
+  "version": "1.0.355",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",