Parameterize local CUA template tasks and centralize the computer tool shape.

This makes local smoke tests easier to target while keeping the OpenAI computer tool definition consistent across the Python and TypeScript templates.

Made-with: Cursor

Author: rgarcia

pkg/templates/python/openai-computer-use/agent/agent.py

@@ -117,6 +117,9 @@
     "strict": False,
 }
 
+# Keep this shape aligned with CUA and current OpenAI Responses API.
+OPENAI_COMPUTER_TOOL = {"type": "computer"}
+
 
 class Agent:
     """An agent that uses OpenAI CUA with Kernel's native computer control API."""
@@ -140,9 +143,7 @@ def __init__(
 
         if computer:
             self.tools += [
-                {
-                    "type": "computer",
-                },
+                dict(OPENAI_COMPUTER_TOOL),
                 BATCH_TOOL,
                 EXTRA_TOOL,
             ]

pkg/templates/python/openai-computer-use/run_local.py

@@ -3,7 +3,7 @@
 No Kernel app deployment needed.
 
 Usage:
-    KERNEL_API_KEY=... OPENAI_API_KEY=... uv run run_local.py --output text
+    KERNEL_API_KEY=... OPENAI_API_KEY=... uv run run_local.py --task "go to example.com and summarize it"
 """
 
 import argparse
@@ -26,6 +26,8 @@
 )
 from computers.kernel_computer import KernelComputer
 
+DEFAULT_TASK = "go to example.com and summarize what the page says"
+
 
 def parse_args():
     parser = argparse.ArgumentParser(description="Run OpenAI CUA local test")
@@ -40,6 +42,11 @@ def parse_args():
         action="store_true",
         help="Enable verbose debug payload logging",
     )
+    parser.add_argument(
+        "--task",
+        default=DEFAULT_TASK,
+        help="User task prompt to run in the browser session",
+    )
     return parser.parse_args()
 
 
@@ -74,7 +81,7 @@ def main():
             },
             {
                 "role": "user",
-                "content": "go to ebay.com and look up oberheim ob-x prices and give me a report",
+                "content": args.task,
             },
         ]
 

pkg/templates/typescript/openai-computer-use/lib/agent.ts

@@ -17,6 +17,8 @@ import type { CuaAction, KernelComputer } from './kernel-computer';
 
 const BATCH_FUNC_NAME = 'batch_computer_actions';
 const EXTRA_FUNC_NAME = 'computer_use_extra';
+// Keep this shape aligned with CUA and current OpenAI Responses API.
+const OPENAI_COMPUTER_TOOL = { type: 'computer' } as unknown as Tool;
 
 export class Agent {
   private model: string;
@@ -40,9 +42,7 @@ export class Agent {
     this.ackCb = opts.acknowledge_safety_check_callback ?? ((): boolean => true);
 
     this.tools = [
-      {
-        type: 'computer',
-      } as unknown as Tool,
+      OPENAI_COMPUTER_TOOL,
       batchComputerTool,
       computerUseExtraTool,
       ...(opts.tools ?? []),

pkg/templates/typescript/openai-computer-use/run_local.ts

@@ -21,15 +21,18 @@ dotenv.config({ override: true, quiet: true });
  * No Kernel app deployment needed.
  *
  * Usage:
- *   KERNEL_API_KEY=... OPENAI_API_KEY=... npx tsx run_local.ts
+ *   KERNEL_API_KEY=... OPENAI_API_KEY=... npx tsx run_local.ts --task "go to example.com and summarize it"
  */
 
+const DEFAULT_TASK = 'go to example.com and summarize what the page says';
+
 export async function runLocalTest(args: string[] = process.argv.slice(2)): Promise<void> {
   if (!process.env.KERNEL_API_KEY) throw new Error('KERNEL_API_KEY is not set');
   if (!process.env.OPENAI_API_KEY) throw new Error('OPENAI_API_KEY is not set');
 
   const client = new Kernel({ apiKey: process.env.KERNEL_API_KEY });
   const outputMode = parseOutputMode(args);
+  const task = parseTask(args);
   const debug = args.includes('--debug');
   const onEvent = createEventLogger({ output: outputMode, verbose: debug });
 
@@ -69,7 +72,7 @@ export async function runLocalTest(args: string[] = process.argv.slice(2)): Prom
           content: [
             {
               type: 'input_text',
-              text: 'go to ebay.com and look up oberheim ob-x prices and give me a report',
+                  text: task,
             },
           ],
         },
@@ -100,6 +103,15 @@ function parseOutputMode(args: string[]): OutputMode {
   return output === 'jsonl' ? 'jsonl' : 'text';
 }
 
+function parseTask(args: string[]): string {
+  const taskFromEquals = args.find((arg) => arg.startsWith('--task='))?.slice('--task='.length).trim();
+  const taskFlagIndex = args.findIndex((arg) => arg === '--task');
+  const nextArg = taskFlagIndex >= 0 ? args[taskFlagIndex + 1] : undefined;
+  const taskFromNext = nextArg && !nextArg.startsWith('--') ? nextArg.trim() : undefined;
+  const task = taskFromEquals || taskFromNext;
+  return task && task.length > 0 ? task : DEFAULT_TASK;
+}
+
 function isDirectRun(): boolean {
   const entry = process.argv[1];
   if (!entry) return false;