@hamp10/agentforge 0.2.15 → 0.2.16

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hamp10/agentforge",
-  "version": "0.2.15",
+  "version": "0.2.16",
   "description": "AgentForge worker — connect your machine to agentforge.ai",
   "type": "module",
   "bin": {

package/src/OllamaAgent.js

@@ -8,112 +8,63 @@ import { fileURLToPath } from 'url';
 const execAsync = promisify(exec);
 const __dirname = path.dirname(fileURLToPath(import.meta.url));
 
-// Tool definitions — used both for embedded system prompt (qwen3 format)
-// and kept as structured objects for _toolDesc lookups.
+// Minimal tool definitions — one compact JSON per line, embedded in system prompt.
+// Ollama's `tools` API param is broken for qwen3 (github.com/ollama/ollama/issues/14601).
+// Descriptions kept short to fit within qwen3-vl:8b's 4096 token context.
 const TOOL_DEFS = [
   {
     type: 'function',
     function: {
       name: 'bash',
-      description: 'Execute a shell command in the working directory. Returns stdout and stderr.',
-      parameters: {
-        type: 'object',
-        properties: {
-          command: { type: 'string', description: 'The shell command to run' }
-        },
-        required: ['command']
-      }
+      description: 'Run a shell command. Returns stdout/stderr.',
+      parameters: { type: 'object', properties: { command: { type: 'string' } }, required: ['command'] }
     }
   },
   {
     type: 'function',
     function: {
       name: 'read_file',
-      description: 'Read the full contents of a file.',
-      parameters: {
-        type: 'object',
-        properties: {
-          path: { type: 'string', description: 'Path to the file (absolute or relative to workdir)' }
-        },
-        required: ['path']
-      }
+      description: 'Read a file.',
+      parameters: { type: 'object', properties: { path: { type: 'string' } }, required: ['path'] }
     }
   },
   {
     type: 'function',
     function: {
       name: 'write_file',
-      description: 'Write content to a file, creating it and any missing parent directories.',
-      parameters: {
-        type: 'object',
-        properties: {
-          path: { type: 'string', description: 'Path to write (absolute or relative to workdir)' },
-          content: { type: 'string', description: 'File content to write' }
-        },
-        required: ['path', 'content']
-      }
+      description: 'Write a file.',
+      parameters: { type: 'object', properties: { path: { type: 'string' }, content: { type: 'string' } }, required: ['path', 'content'] }
     }
   },
   {
     type: 'function',
     function: {
       name: 'list_directory',
-      description: 'List files and subdirectories at a path.',
-      parameters: {
-        type: 'object',
-        properties: {
-          path: { type: 'string', description: 'Directory path (absolute or relative to workdir)' }
-        },
-        required: ['path']
-      }
+      description: 'List files in a directory.',
+      parameters: { type: 'object', properties: { path: { type: 'string' } }, required: ['path'] }
     }
   },
   {
     type: 'function',
     function: {
       name: 'web_fetch',
-      description: 'Fetch the text content of a URL (first 4000 chars).',
-      parameters: {
-        type: 'object',
-        properties: {
-          url: { type: 'string', description: 'URL to fetch' }
-        },
-        required: ['url']
-      }
+      description: 'Fetch text content from a URL.',
+      parameters: { type: 'object', properties: { url: { type: 'string' } }, required: ['url'] }
     }
   },
   {
     type: 'function',
     function: {
       name: 'take_screenshot',
-      description: 'Take a screenshot of the current screen or the agent browser (port 9223). Returns base64 image data you can analyze visually. Use this to check what a webpage looks like, verify a build result, or monitor a running process. Set send_to_user=true ONLY when the user explicitly asked to see a screenshot.',
-      parameters: {
-        type: 'object',
-        properties: {
-          target: {
-            type: 'string',
-            enum: ['screen', 'browser'],
-            description: 'screen = full screen capture. browser = screenshot of the agent browser (port 9223).'
-          },
-          url: {
-            type: 'string',
-            description: 'Optional: navigate the browser to this URL before taking the screenshot.'
-          },
-          send_to_user: {
-            type: 'boolean',
-            description: 'If true, send the screenshot to the user\'s chat. Only set this when the user explicitly asked to see a screenshot or visual output.'
-          }
-        },
-        required: ['target']
-      }
+      description: 'Screenshot the screen. Set send_to_user=true only if user asked to see it.',
+      parameters: { type: 'object', properties: { target: { type: 'string', enum: ['screen', 'browser'] }, send_to_user: { type: 'boolean' } }, required: ['target'] }
     }
   }
 ];
 
-// Build the <tools> XML block to embed in the system prompt.
-// Ollama's `tools` API parameter is broken for qwen3 models (malformed JSON in the prompt).
-// The reliable fix is to embed tool definitions directly in the system prompt as XML.
-const TOOLS_XML = `<tools>\n${TOOL_DEFS.map(t => JSON.stringify(t)).join('\n')}\n</tools>`;
+// Minimal <tools> XML for system prompt — one compact JSON per line, no outer array.
+// Per qwen3 Hermes chat template (tokenizer_config.json).
+const TOOLS_XML = `<tools>\n${TOOL_DEFS.map(t => JSON.stringify(t.function)).join('\n')}\n</tools>`;
 
 /**
  * Parse <tool_call>...</tool_call> blocks from streamed content.
@@ -281,24 +232,34 @@ export class OllamaAgent extends EventEmitter {
       // Load conversation history from disk (session persistence)
       const history = this._loadHistory(agentId, workDir, sessionId);
 
-      // For qwen3 models: embed tool definitions in the system prompt.
-      // Ollama's `tools` API param is broken for qwen3 (malformed JSON sent to model).
-      // Embedding as XML matches the model's native Hermes-style chat template.
-      const toolsBlock = isQwen3 ? `\n\n${TOOLS_XML}\n\nFor each tool call, output ONLY a <tool_call> block with no surrounding text:\n<tool_call>\n{"name": "<tool_name>", "arguments": {<args>}}\n</tool_call>` : '';
-
-      const systemPrompt = [
-        isQwen3 ? '/no_think' : null,
-        `You are an AI agent running on AgentForge.ai.`,
-        `Your working directory is: ${workDir}`,
-        ``,
-        `CRITICAL RULES:`,
-        `1. Use tools to act. Do NOT describe steps or write code blocks — call the actual tool.`,
-        `2. bash = run shell commands. write_file = write files. read_file = read files. take_screenshot = screenshot.`,
-        `3. For conversational messages (greetings, casual chat) — respond with plain text. No tools needed.`,
-        `4. Do not ask for clarification — make your best judgment and act immediately.`,
-        `5. After completing work, write a brief summary of what you did.`,
-        toolsBlock,
-      ].filter(Boolean).join('\n');
+      // System prompt uses the exact format from qwen3's Hermes chat template.
+      // Tools are embedded as <tools> XML — never passed via the API `tools` param (broken in Ollama).
+      const systemPrompt = isQwen3
+        ? [
+            '/no_think',
+            `You are a helpful assistant. Working directory: ${workDir}`,
+            ``,
+            `# Tools`,
+            ``,
+            `You may call one or more functions to complete the task.`,
+            ``,
+            `You are provided with function signatures within <tools></tools> XML tags:`,
+            TOOLS_XML,
+            ``,
+            `For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:`,
+            `<tool_call>`,
+            `{"name": <function-name>, "arguments": <args-json-object>}`,
+            `</tool_call>`,
+            ``,
+            `Rules:`,
+            `- Call tools to take actions. Do NOT describe what you would do — just do it.`,
+            `- For simple conversation (greetings, questions) respond with plain text, no tools.`,
+            `- After finishing, write a brief summary.`,
+          ].join('\n')
+        : [
+            `You are a helpful AI agent. Working directory: ${workDir}`,
+            `Use the provided tools to complete tasks. Don't describe — act.`,
+          ].join('\n');
 
       const messages = [
         { role: 'system', content: systemPrompt },
@@ -329,10 +290,13 @@ export class OllamaAgent extends EventEmitter {
             model: effectiveModel,
             messages,
             stream: true,
-            // qwen3: tools embedded in system prompt — do NOT pass tools param (broken in Ollama)
-            // Other models: pass tools normally via API
+            // qwen3: tools embedded in system prompt — do NOT pass tools param (broken in Ollama for qwen3)
+            // Other models: pass tools normally
             ...(!isQwen3 ? { tools: TOOL_DEFS, tool_choice: 'auto' } : {}),
-            ...(isQwen3 ? { options: { think: false } } : {}),
+            options: {
+              num_ctx: 8192,  // explicit context — Ollama defaults to 2048 which is too small
+              ...(isQwen3 ? { think: false } : {}),  // CRITICAL: thinking + tools corrupts template
+            },
           };
 
           response = await fetch(`${this.baseUrl}/v1/chat/completions`, {