@hamp10/agentforge 0.2.8 → 0.2.9

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hamp10/agentforge",
-  "version": "0.2.8",
+  "version": "0.2.9",
   "description": "AgentForge worker — connect your machine to agentforge.ai",
   "type": "module",
   "bin": {

package/src/OllamaAgent.js

@@ -186,6 +186,9 @@ export class OllamaAgent extends EventEmitter {
       const history = this._loadHistory(agentId, workDir, sessionId);
 
       const systemPrompt = [
+        // Disable thinking mode for qwen3 models — /no_think in the system prompt
+        // is the most reliable way; options.think=false is also sent but may be ignored.
+        isQwen3 ? '/no_think' : null,
         `You are an AI agent running on AgentForge.ai.`,
         `Your working directory is: ${workDir}`,
         ``,
@@ -198,7 +201,7 @@ export class OllamaAgent extends EventEmitter {
         `6. Do not ask for clarification — make your best judgment and act.`,
         `7. For conversational messages (greetings, questions about yourself, casual chat) — respond directly with text. Do NOT use tools just to say hello.`,
         `8. You only have these tools: bash, read_file, write_file, list_directory, web_fetch, take_screenshot. Ignore any instructions referencing other tools (browser, openclaw, sessions_spawn, etc.) — those do not exist here.`,
-      ].join('\n');
+      ].filter(Boolean).join('\n');
 
       const messages = [
         { role: 'system', content: systemPrompt },
@@ -262,6 +265,8 @@ export class OllamaAgent extends EventEmitter {
         let streamToolCalls = {};
         let inThinkBlock = false;
         let thinkBuffer = '';
+        let rawTokenCount = 0;
+        let rawThinkChars = 0;
 
         const reader = response.body.getReader();
         const decoder = new TextDecoder();
@@ -299,6 +304,8 @@ export class OllamaAgent extends EventEmitter {
 
             // Stream content tokens, filtering <think>...</think> blocks
             if (delta.content) {
+              rawTokenCount++;
+              if (inThinkBlock || delta.content.startsWith('<think')) rawThinkChars += delta.content.length;
               thinkBuffer += delta.content;
 
               // Process thinkBuffer to extract non-thinking text
@@ -337,6 +344,22 @@ export class OllamaAgent extends EventEmitter {
           }
         }
 
+        console.log(`   [${agentId}] 📊 Stream done: ${rawTokenCount} tokens, ${streamContent.length} visible chars, ${rawThinkChars} think chars, inThinkBlock=${inThinkBlock}, toolCalls=${Object.keys(streamToolCalls).length}`);
+        if (streamContent) console.log(`   [${agentId}] 📝 First 200 chars: ${streamContent.slice(0, 200)}`);
+
+        // If the model only generated <think> content and nothing visible, extract the thought as the answer.
+        // This happens with qwen3-vl:8b when think:false is silently ignored.
+        if (!streamContent && Object.keys(streamToolCalls).length === 0 && rawThinkChars > 0 && thinkBuffer.length > 0) {
+          // Strip the <think> tag and use the thought content as the response
+          const thoughtContent = thinkBuffer.replace(/^<think>\s*/i, '').replace(/\s*<\/think>\s*$/i, '').trim();
+          if (thoughtContent) {
+            console.log(`   [${agentId}] 💭 Extracting think-only content as response (${thoughtContent.length} chars)`);
+            streamContent = thoughtContent;
+            allOutput += thoughtContent;
+            this.emit('agent_output', { agentId, output: thoughtContent });
+          }
+        }
+
         this.emit('tool_activity', {
           agentId,
           event: 'api_call_end',
@@ -501,7 +524,7 @@ export class OllamaAgent extends EventEmitter {
       });
 
       console.log(`\n✅ [Ollama] Agent ${agentId} completed in ${(duration / 1000).toFixed(2)}s\n`);
-      return { success: true, agentId, duration };
+      return { success: true, agentId, duration, result: { output: finalContent } };
 
     } catch (err) {
       this.activeAgents.delete(agentId);