Release 2.9.3: switch primary to gpt-oss:120b-cloud with mistral vision fallback

- config: primary model → gpt-oss:120b-cloud (faster general chat, 2.72s median)
- server-v2: fix vision regex to exclude gpt-oss (gpt(?!-oss) negative lookahead)
  * gpt-oss was incorrectly matching /gpt/ → image data sent to non-vision model
  * applied at all 3 _supportsVision check sites (lines 5204, 6362, 7832)
- server-v2: vision fallback routing — when primary lacks vision support and user
  message contains images, automatically route to orchestration.secondary
  (mistral-large-3:675b-cloud, 4.27s avg, fastest vision cloud model)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

.smallclaw/config.json

@@ -21,7 +21,7 @@
     "providers": {
       "ollama": {
         "endpoint": "http://localhost:11434",
-        "model": "gemini-3-flash-preview:cloud"
+        "model": "gpt-oss:120b-cloud"
       },
       "lm_studio": {
         "endpoint": "http://host.docker.internal:1234",
@@ -45,11 +45,11 @@
     }
   },
   "models": {
-    "primary": "gemini-3-flash-preview:cloud",
+    "primary": "gpt-oss:120b-cloud",
     "roles": {
-      "manager": "gemini-3-flash-preview:cloud",
-      "executor": "gemini-3-flash-preview:cloud",
-      "verifier": "gemini-3-flash-preview:cloud",
+      "manager": "gpt-oss:120b-cloud",
+      "executor": "gpt-oss:120b-cloud",
+      "verifier": "gpt-oss:120b-cloud",
       "background_task": ""
     }
   },

src/gateway/server-v2.ts

@@ -5201,7 +5201,7 @@ async function handleChat(
         ? (isBrowserTool ? buildBrowserAck(toolName, toolResult) : buildDesktopAck(toolName, toolResult))
         : toolResult.result;
       const _imagingTools3 = new Set(['image_edit', 'image_info', 'image_read', 'python_eval']);
-      const _supportsVision3 = /kimi|openai|claude|gpt|gemini/i.test(String(_activeModelName || primaryProvider || ''));
+      const _supportsVision3 = /kimi|openai|claude|gemini|gpt(?!-oss)/i.test(String(_activeModelName || primaryProvider || ''));
       const _resolvedToolContent3 = _imagingTools3.has(toolName) && typeof toolMessageContent === 'string'
         ? resolveToolImageContent(toolMessageContent, workspacePath, _supportsVision3)
         : toolMessageContent;
@@ -6359,7 +6359,7 @@ RULES:
           ? (isBrowserTool ? buildBrowserAck(toolName, toolResult) : buildDesktopAck(toolName, toolResult))
           : toolResult.result;
         const _imagingTools2 = new Set(['image_edit', 'image_info', 'image_read', 'python_eval']);
-        const _supportsVision2 = /kimi|openai|claude|gpt|gemini/i.test(String(_activeModelName || primaryProvider || ''));
+        const _supportsVision2 = /kimi|openai|claude|gemini|gpt(?!-oss)/i.test(String(_activeModelName || primaryProvider || ''));
         const _resolvedContent2 = _imagingTools2.has(toolName) && typeof toolMessageContent === 'string'
           ? resolveToolImageContent(toolMessageContent, workspacePath, _supportsVision2)
           : toolMessageContent;
@@ -6496,8 +6496,20 @@ RULES:
       );
       const primaryThinkMode: boolean | 'high' | 'medium' | 'low' = (multiAgentActive && !isActiveAutomationOp) ? true : false;
       const needsLongOutput = /pptx|powerpoint|presentation|슬라이드|발표|프레젠테이션|게임|코드|스크립트|함수|클래스|구현해|만들어줘|만들어\s*줘|작성해|짜줘|build.*app|create.*app|write.*code|implement.*class|game.*make|전체.*코드|완성된.*코드/i.test(message);
-      // Model priority: explicit request > skill override > config default
+      // Vision fallback: if primary doesn't support vision but messages contain images, use secondary vision model
+      const _visionFallbackModel: string | undefined = (() => {
+        if (/kimi|openai|claude|gemini|gpt(?!-oss)/i.test(String(_activeModelName || ''))) return undefined;
+        const hasImages = messages.some((m: any) =>
+          Array.isArray(m.content) && m.content.some((p: any) => p.type === 'image_url')
+        );
+        if (!hasImages) return undefined;
+        const sec = rawCfgForPreempt.orchestration?.secondary;
+        const secModel = String(sec?.model || '').trim();
+        return (secModel && sec?.vision === true) ? secModel : undefined;
+      })();
+      // Model priority: explicit request > vision fallback > skill override > config default
       const effectiveModel = String(modelOverride || '').trim()
+        || _visionFallbackModel
         || skillsManager.getModelOverrideForUser(username ? getUserWorkspace(username) : null)
         || undefined;
       // Always use streaming so text tokens appear in real-time. The streaming
@@ -7817,7 +7829,7 @@ RULES:
         ? (isBrowserTool ? buildBrowserAck(toolName, toolResult) : buildDesktopAck(toolName, toolResult))
         : toolResult.result;
       const _imagingTools = new Set(['image_edit', 'image_info', 'image_read', 'python_eval']);
-      const _supportsVision = /kimi|openai|claude|gpt|gemini/i.test(String(_activeModelName || primaryProvider || ''));
+      const _supportsVision = /kimi|openai|claude|gemini|gpt(?!-oss)/i.test(String(_activeModelName || primaryProvider || ''));
       const _resolvedToolContent = _imagingTools.has(toolName) && typeof toolMessageContent === 'string'
         ? resolveToolImageContent(toolMessageContent, workspacePath, _supportsVision)
         : toolMessageContent;