@hamp10/agentforge 0.2.9 → 0.2.11

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hamp10/agentforge",
-  "version": "0.2.9",
+  "version": "0.2.11",
   "description": "AgentForge worker — connect your machine to agentforge.ai",
   "type": "module",
   "bin": {

package/src/OllamaAgent.js

@@ -104,6 +104,32 @@ const TOOLS = [
   }
 ];
 
+/**
+ * Detect text-based tool calls from model content.
+ * Some models (qwen3-vl:8b) output tool calls as JSON text lines in content
+ * instead of using the OpenAI tool_calls format.
+ * Returns array of {name, arguments} if ALL non-empty lines are valid tool calls, else null.
+ */
+function _parseTextToolCalls(content) {
+  if (!content) return null;
+  const lines = content.trim().split('\n').map(l => l.trim()).filter(Boolean);
+  if (lines.length === 0) return null;
+  const calls = [];
+  for (const line of lines) {
+    try {
+      const obj = JSON.parse(line);
+      if (typeof obj.name === 'string' && obj.arguments && typeof obj.arguments === 'object') {
+        calls.push({ name: obj.name, arguments: obj.arguments });
+      } else {
+        return null; // Valid JSON but not a tool call — treat whole content as text
+      }
+    } catch {
+      return null; // Non-JSON line — treat whole content as text
+    }
+  }
+  return calls.length > 0 ? calls : null;
+}
+
 /**
  * LocalModelAgent — drop-in replacement for OpenClawCLI.
  * Runs an agentic tool-use loop against ANY OpenAI-compatible local model server.
@@ -336,10 +362,7 @@ export class OllamaAgent extends EventEmitter {
               thinkBuffer = inThinkBlock ? thinkBuffer.slice(thinkBuffer.lastIndexOf('<think>')) : '';
 
               streamContent += out;
-              allOutput += out;
-              if (out) {
-                this.emit('agent_output', { agentId, output: out });
-              }
+              // Don't emit per-token — we check for JSON tool calls after the full turn
             }
           }
         }
@@ -355,8 +378,30 @@ export class OllamaAgent extends EventEmitter {
           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 });
+            // Don't emit here — detection block below handles it
+          }
+        }
+
+        // ── Detect text-based tool calls or emit text content ─────────────────
+        // qwen3-vl:8b outputs tool calls as one JSON object per line in content.
+        // If detected, convert to streamToolCalls and suppress the raw JSON output.
+        // Otherwise, emit the text content to the dashboard.
+        if (Object.keys(streamToolCalls).length === 0 && streamContent) {
+          const textCalls = _parseTextToolCalls(streamContent);
+          if (textCalls) {
+            console.log(`   [${agentId}] 🔍 ${textCalls.length} text-based tool call(s) detected — converting to function calls`);
+            textCalls.forEach((tc, i) => {
+              streamToolCalls[i] = {
+                id: `text-${i}`,
+                type: 'function',
+                function: { name: tc.name, arguments: JSON.stringify(tc.arguments) }
+              };
+            });
+            streamContent = ''; // Don't display raw JSON to user
+          } else {
+            // Regular text response — emit to dashboard
+            allOutput += streamContent;
+            if (streamContent.trim()) this.emit('agent_output', { agentId, output: streamContent });
           }
         }
 
@@ -405,7 +450,11 @@ export class OllamaAgent extends EventEmitter {
 
             // If the tool returned an image (base64), push it as a vision message
             // so the model can actually see what was captured.
+            // Also forward to dashboard so the user sees the screenshot in chat.
             const isImageResult = typeof result === 'string' && result.startsWith('data:image/');
+            if (isImageResult) {
+              this.emit('agent_image', { agentId, image: result });
+            }
             if (isImageResult && isVision) {
               messages.push({
                 role: 'tool',