Fix Auto-GPT never timing out

Signed-off-by: Merwane Hamadi <[email protected]>

agbenchmark/agent_interface.py

@@ -42,35 +42,45 @@ def run_agent(
         )
 
         start_time = time.time()
-        timeout = config["cutoff"]
 
-        while True:
-            if process.stdout is None:
-                continue
+        print(
+            f"Running Python function '{config['entry_path']}' with timeout {config['cutoff']}"
+        )
+        command = [sys.executable, "-m", config["entry_path"], str(task)]
+        process = subprocess.Popen(
+            command,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+            universal_newlines=True,
+        )
 
-            while output := process.stdout.readline():
-                print(output.strip())
+        start_time = time.time()
 
-            # Check if process has ended
-            if process.poll() is not None:
-                print("The Python function has finished running.")
-                break
+        while True:
+            output = ""
+            if process.stdout is not None:
+                output = process.stdout.readline()
+                print(output.strip())
 
-            # Check if process has exceeded timeout
-            if time.time() - start_time > timeout:
-                print(
-                    "The Python function has exceeded the time limit and was terminated."
-                )
-                # Terminate the process group
-                process.terminate()
+            # Check if process has ended, has no more output, or exceeded timeout
+            if (
+                process.poll() is not None
+                or output == ""
+                or (time.time() - start_time > config["cutoff"])
+            ):
                 break
 
-            # Optional: sleep for a while
-            time.sleep(0.1)
+        if time.time() - start_time > config["cutoff"]:
+            print("The Python function has exceeded the time limit and was terminated.")
+            process.kill()
+        else:
+            print("The Python function has finished running.")
 
-        # Wait for process to terminate, then get return code
         process.wait()
 
+        if process.returncode != 0:
+            print(f"The agent timed out")
+
 
 def copy_artifacts_into_workspace(
     workspace: str, artifact_folder_name: str, challenge_dir_path: str