From 395c1ea9e3f593848c53b11dd896b60a3b8279b1 Mon Sep 17 00:00:00 2001
From: Engel Nyst <enyst@users.noreply.github.com>
Date: Mon, 3 Mar 2025 00:19:25 +0100
Subject: [PATCH] [Refactor] split runtime initialization (create, connect,
 init) in cli scripts (#7036)

---
 evaluation/benchmarks/EDA/run_infer.py        |  2 +
 .../benchmarks/agent_bench/run_infer.py       |  2 +
 .../benchmarks/aider_bench/run_infer.py       |  3 +-
 evaluation/benchmarks/biocoder/run_infer.py   |  2 +
 evaluation/benchmarks/bird/run_infer.py       |  2 +
 .../browsing_delegation/run_infer.py          |  2 +
 .../benchmarks/commit0_bench/run_infer.py     |  2 +
 .../benchmarks/discoverybench/run_infer.py    |  2 +
 evaluation/benchmarks/gaia/run_infer.py       |  2 +
 evaluation/benchmarks/gorilla/run_infer.py    |  2 +
 evaluation/benchmarks/gpqa/run_infer.py       |  2 +
 .../benchmarks/humanevalfix/run_infer.py      |  2 +
 .../benchmarks/logic_reasoning/run_infer.py   |  2 +
 evaluation/benchmarks/miniwob/run_infer.py    |  2 +
 evaluation/benchmarks/mint/run_infer.py       |  2 +
 evaluation/benchmarks/ml_bench/run_infer.py   |  2 +
 .../benchmarks/scienceagentbench/run_infer.py |  2 +
 evaluation/benchmarks/swe_bench/run_infer.py  |  2 +
 .../benchmarks/the_agent_company/run_infer.py |  3 +-
 evaluation/benchmarks/toolqa/run_infer.py     |  2 +
 .../benchmarks/visualwebarena/run_infer.py    |  3 ++
 evaluation/benchmarks/webarena/run_infer.py   |  2 +
 evaluation/integration_tests/run_infer.py     |  2 +
 openhands/core/cli.py                         | 16 +++++-
 openhands/core/main.py                        | 13 ++++-
 openhands/core/setup.py                       | 50 +++++++++++++------
 26 files changed, 107 insertions(+), 21 deletions(-)

diff --git a/evaluation/benchmarks/EDA/run_infer.py b/evaluation/benchmarks/EDA/run_infer.py
index 636a52e2bd..f216a86ff8 100644
--- a/evaluation/benchmarks/EDA/run_infer.py
+++ b/evaluation/benchmarks/EDA/run_infer.py
@@ -24,6 +24,7 @@ from openhands.core.config import (
 from openhands.core.logger import openhands_logger as logger
 from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import MessageAction
+from openhands.utils.async_utils import call_async_from_sync
 
 game = None
 
@@ -121,6 +122,7 @@ def process_instance(
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
 
     state: State | None = asyncio.run(
         run_controller(
diff --git a/evaluation/benchmarks/agent_bench/run_infer.py b/evaluation/benchmarks/agent_bench/run_infer.py
index 68cf2ff793..a78e402395 100644
--- a/evaluation/benchmarks/agent_bench/run_infer.py
+++ b/evaluation/benchmarks/agent_bench/run_infer.py
@@ -34,6 +34,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import AgentFinishAction, CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 
 def get_config(
@@ -210,6 +211,7 @@ def process_instance(
     # =============================================
 
     runtime: Runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
 
     initialize_runtime(runtime, instance=instance)
 
diff --git a/evaluation/benchmarks/aider_bench/run_infer.py b/evaluation/benchmarks/aider_bench/run_infer.py
index 9c848f67b1..0d97496acd 100644
--- a/evaluation/benchmarks/aider_bench/run_infer.py
+++ b/evaluation/benchmarks/aider_bench/run_infer.py
@@ -34,6 +34,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 # Configure visibility of unit tests to the Agent.
 USE_UNIT_TESTS = os.environ.get('USE_UNIT_TESTS', 'false').lower() == 'true'
@@ -203,7 +204,7 @@ def process_instance(
     # =============================================
 
     runtime: Runtime = create_runtime(config)
-
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime, instance=instance)
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
diff --git a/evaluation/benchmarks/biocoder/run_infer.py b/evaluation/benchmarks/biocoder/run_infer.py
index b0a06a6ece..f1c98ed066 100644
--- a/evaluation/benchmarks/biocoder/run_infer.py
+++ b/evaluation/benchmarks/biocoder/run_infer.py
@@ -31,6 +31,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
     'CodeActAgent': functools.partial(
@@ -274,6 +275,7 @@ def process_instance(
     instruction += AGENT_CLS_TO_INST_SUFFIX[metadata.agent_class]
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime, instance)
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
diff --git a/evaluation/benchmarks/bird/run_infer.py b/evaluation/benchmarks/bird/run_infer.py
index 4cdd29862f..1c56deb967 100644
--- a/evaluation/benchmarks/bird/run_infer.py
+++ b/evaluation/benchmarks/bird/run_infer.py
@@ -34,6 +34,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 
 def codeact_user_response(state: State) -> str:
@@ -399,6 +400,7 @@ def process_instance(
     instruction += AGENT_CLS_TO_INST_SUFFIX[metadata.agent_class]
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime, instance)
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
diff --git a/evaluation/benchmarks/browsing_delegation/run_infer.py b/evaluation/benchmarks/browsing_delegation/run_infer.py
index 5f3ee99d74..0ef080dbca 100644
--- a/evaluation/benchmarks/browsing_delegation/run_infer.py
+++ b/evaluation/benchmarks/browsing_delegation/run_infer.py
@@ -25,6 +25,7 @@ from openhands.core.config import (
 from openhands.core.logger import openhands_logger as logger
 from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import MessageAction
+from openhands.utils.async_utils import call_async_from_sync
 
 # Only CodeActAgent can delegate to BrowsingAgent
 SUPPORTED_AGENT_CLS = {'CodeActAgent'}
@@ -74,6 +75,7 @@ def process_instance(
     )
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
 
     state: State | None = asyncio.run(
         run_controller(
diff --git a/evaluation/benchmarks/commit0_bench/run_infer.py b/evaluation/benchmarks/commit0_bench/run_infer.py
index cf6148975b..63d394a029 100644
--- a/evaluation/benchmarks/commit0_bench/run_infer.py
+++ b/evaluation/benchmarks/commit0_bench/run_infer.py
@@ -35,6 +35,7 @@ from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation, ErrorObservation
 from openhands.events.serialization.event import event_to_dict
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 from openhands.utils.shutdown_listener import sleep_if_should_continue
 
 USE_HINT_TEXT = os.environ.get('USE_HINT_TEXT', 'false').lower() == 'true'
@@ -394,6 +395,7 @@ def process_instance(
         logger.info(f'Starting evaluation for instance {instance.instance_id}.')
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     try:
         initialize_runtime(runtime, instance)
 
diff --git a/evaluation/benchmarks/discoverybench/run_infer.py b/evaluation/benchmarks/discoverybench/run_infer.py
index f3fdadab8e..d91d01194d 100644
--- a/evaluation/benchmarks/discoverybench/run_infer.py
+++ b/evaluation/benchmarks/discoverybench/run_infer.py
@@ -34,6 +34,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import AgentFinishAction, CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 EVALUATION_LLM = 'gpt-4-1106-preview'
 
@@ -281,6 +282,7 @@ def process_instance(
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime, instance.data_files)
 
     state: State | None = asyncio.run(
diff --git a/evaluation/benchmarks/gaia/run_infer.py b/evaluation/benchmarks/gaia/run_infer.py
index e0e5ed0363..e63026e813 100644
--- a/evaluation/benchmarks/gaia/run_infer.py
+++ b/evaluation/benchmarks/gaia/run_infer.py
@@ -31,6 +31,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import AgentFinishAction, CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 DATASET_CACHE_DIR = os.path.join(os.path.dirname(__file__), 'data')
 
@@ -148,6 +149,7 @@ def process_instance(
     logger.info(f'Instruction:\n{instruction}', extra={'msg_type': 'OBSERVATION'})
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime, instance)
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
diff --git a/evaluation/benchmarks/gorilla/run_infer.py b/evaluation/benchmarks/gorilla/run_infer.py
index 22b42a8545..e856fa267c 100644
--- a/evaluation/benchmarks/gorilla/run_infer.py
+++ b/evaluation/benchmarks/gorilla/run_infer.py
@@ -26,6 +26,7 @@ from openhands.core.config import (
 from openhands.core.logger import openhands_logger as logger
 from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import MessageAction
+from openhands.utils.async_utils import call_async_from_sync
 
 AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
     'CodeActAgent': codeact_user_response,
@@ -82,6 +83,7 @@ def process_instance(
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     state: State | None = asyncio.run(
         run_controller(
             config=config,
diff --git a/evaluation/benchmarks/gpqa/run_infer.py b/evaluation/benchmarks/gpqa/run_infer.py
index 0f19755c34..e297e3fb9e 100644
--- a/evaluation/benchmarks/gpqa/run_infer.py
+++ b/evaluation/benchmarks/gpqa/run_infer.py
@@ -49,6 +49,7 @@ from openhands.events.action import (
     MessageAction,
 )
 from openhands.events.observation import Observation
+from openhands.utils.async_utils import call_async_from_sync
 
 ACTION_FORMAT = """
 <<FINAL_ANSWER||
@@ -214,6 +215,7 @@ Ok now its time to start solving the question. Good luck!
 """
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     state: State | None = asyncio.run(
         run_controller(
             config=config,
diff --git a/evaluation/benchmarks/humanevalfix/run_infer.py b/evaluation/benchmarks/humanevalfix/run_infer.py
index 82e89f06e9..fbf88859b6 100644
--- a/evaluation/benchmarks/humanevalfix/run_infer.py
+++ b/evaluation/benchmarks/humanevalfix/run_infer.py
@@ -39,6 +39,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 IMPORT_HELPER = {
     'python': [
@@ -232,6 +233,7 @@ def process_instance(
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime, instance)
     state: State | None = asyncio.run(
         run_controller(
diff --git a/evaluation/benchmarks/logic_reasoning/run_infer.py b/evaluation/benchmarks/logic_reasoning/run_infer.py
index 543416deea..fac82f29f5 100644
--- a/evaluation/benchmarks/logic_reasoning/run_infer.py
+++ b/evaluation/benchmarks/logic_reasoning/run_infer.py
@@ -31,6 +31,7 @@ from openhands.events.action import (
 )
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
     'CodeActAgent': codeact_user_response,
@@ -206,6 +207,7 @@ def process_instance(
     instruction += AGENT_CLS_TO_INST_SUFFIX[metadata.agent_class]
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime, instance)
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
diff --git a/evaluation/benchmarks/miniwob/run_infer.py b/evaluation/benchmarks/miniwob/run_infer.py
index a6b354716a..55e510818a 100644
--- a/evaluation/benchmarks/miniwob/run_infer.py
+++ b/evaluation/benchmarks/miniwob/run_infer.py
@@ -41,6 +41,7 @@ from openhands.runtime.browser.browser_env import (
     BROWSER_EVAL_GET_GOAL_ACTION,
     BROWSER_EVAL_GET_REWARDS_ACTION,
 )
+from openhands.utils.async_utils import call_async_from_sync
 
 SUPPORTED_AGENT_CLS = {'BrowsingAgent', 'CodeActAgent'}
 
@@ -145,6 +146,7 @@ def process_instance(
         logger.info(f'Starting evaluation for instance {env_id}.')
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     task_str, obs = initialize_runtime(runtime)
 
     task_str += (
diff --git a/evaluation/benchmarks/mint/run_infer.py b/evaluation/benchmarks/mint/run_infer.py
index 320c9caf47..bd1a394332 100644
--- a/evaluation/benchmarks/mint/run_infer.py
+++ b/evaluation/benchmarks/mint/run_infer.py
@@ -35,6 +35,7 @@ from openhands.events.action import (
 )
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 
 def codeact_user_response_mint(state: State, task: Task, task_config: dict[str, int]):
@@ -184,6 +185,7 @@ def process_instance(
     )
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime)
 
     state: State | None = asyncio.run(
diff --git a/evaluation/benchmarks/ml_bench/run_infer.py b/evaluation/benchmarks/ml_bench/run_infer.py
index 371314a74d..5eff173b46 100644
--- a/evaluation/benchmarks/ml_bench/run_infer.py
+++ b/evaluation/benchmarks/ml_bench/run_infer.py
@@ -43,6 +43,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 config = load_app_config()
 
@@ -234,6 +235,7 @@ def process_instance(instance: Any, metadata: EvalMetadata, reset_logger: bool =
     instruction += AGENT_CLS_TO_INST_SUFFIX[metadata.agent_class]
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime, instance)
 
     # Run the agent
diff --git a/evaluation/benchmarks/scienceagentbench/run_infer.py b/evaluation/benchmarks/scienceagentbench/run_infer.py
index b84b9230e5..fe0cd7ef3a 100644
--- a/evaluation/benchmarks/scienceagentbench/run_infer.py
+++ b/evaluation/benchmarks/scienceagentbench/run_infer.py
@@ -29,6 +29,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
     'CodeActAgent': codeact_user_response,
@@ -195,6 +196,7 @@ If the program uses some packages that are incompatible, please figure out alter
 """
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime, instance)
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
diff --git a/evaluation/benchmarks/swe_bench/run_infer.py b/evaluation/benchmarks/swe_bench/run_infer.py
index 266fc6fa23..df9042969e 100644
--- a/evaluation/benchmarks/swe_bench/run_infer.py
+++ b/evaluation/benchmarks/swe_bench/run_infer.py
@@ -40,6 +40,7 @@ from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation, ErrorObservation
 from openhands.events.serialization.event import event_to_dict
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 from openhands.utils.shutdown_listener import sleep_if_should_continue
 
 USE_HINT_TEXT = os.environ.get('USE_HINT_TEXT', 'false').lower() == 'true'
@@ -464,6 +465,7 @@ def process_instance(
             f'This is the {runtime_failure_count + 1}th attempt for instance {instance.instance_id}, setting resource factor to {config.sandbox.remote_runtime_resource_factor}'
         )
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
 
     try:
         initialize_runtime(runtime, instance)
diff --git a/evaluation/benchmarks/the_agent_company/run_infer.py b/evaluation/benchmarks/the_agent_company/run_infer.py
index fc5dadb27c..732a431aa1 100644
--- a/evaluation/benchmarks/the_agent_company/run_infer.py
+++ b/evaluation/benchmarks/the_agent_company/run_infer.py
@@ -28,6 +28,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import BrowserOutputObservation, CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 
 def get_config(
@@ -275,7 +276,7 @@ if __name__ == '__main__':
         args.task_image_name, task_short_name, temp_dir, agent_llm_config, agent_config
     )
     runtime: Runtime = create_runtime(config)
-
+    call_async_from_sync(runtime.connect)
     init_task_env(runtime, args.server_hostname, env_llm_config)
 
     dependencies = load_dependencies(runtime)
diff --git a/evaluation/benchmarks/toolqa/run_infer.py b/evaluation/benchmarks/toolqa/run_infer.py
index b4b9724dce..2fc670e568 100644
--- a/evaluation/benchmarks/toolqa/run_infer.py
+++ b/evaluation/benchmarks/toolqa/run_infer.py
@@ -27,6 +27,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import CmdRunAction, MessageAction
 from openhands.events.observation import CmdOutputObservation
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
     'CodeActAgent': codeact_user_response,
@@ -104,6 +105,7 @@ def process_instance(instance: Any, metadata: EvalMetadata, reset_logger: bool =
     logger.info(f'Instruction:\n{instruction}', extra={'msg_type': 'OBSERVATION'})
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     initialize_runtime(runtime)
 
     # Here's how you can run the agent (similar to the `main` function) and get the final task state
diff --git a/evaluation/benchmarks/visualwebarena/run_infer.py b/evaluation/benchmarks/visualwebarena/run_infer.py
index 4a1bce5f97..75529155a2 100644
--- a/evaluation/benchmarks/visualwebarena/run_infer.py
+++ b/evaluation/benchmarks/visualwebarena/run_infer.py
@@ -37,6 +37,7 @@ from openhands.runtime.browser.browser_env import (
     BROWSER_EVAL_GET_GOAL_ACTION,
     BROWSER_EVAL_GET_REWARDS_ACTION,
 )
+from openhands.utils.async_utils import call_async_from_sync
 
 SUPPORTED_AGENT_CLS = {'VisualBrowsingAgent'}
 AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
@@ -159,6 +160,8 @@ def process_instance(
         logger.info(f'Starting evaluation for instance {env_id}.')
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
+
     task_str, goal_image_urls = initialize_runtime(runtime)
     initial_user_action = MessageAction(content=task_str, image_urls=goal_image_urls)
     state: State | None = asyncio.run(
diff --git a/evaluation/benchmarks/webarena/run_infer.py b/evaluation/benchmarks/webarena/run_infer.py
index 3da1cb4a8d..a9b251b90a 100644
--- a/evaluation/benchmarks/webarena/run_infer.py
+++ b/evaluation/benchmarks/webarena/run_infer.py
@@ -36,6 +36,7 @@ from openhands.runtime.browser.browser_env import (
     BROWSER_EVAL_GET_GOAL_ACTION,
     BROWSER_EVAL_GET_REWARDS_ACTION,
 )
+from openhands.utils.async_utils import call_async_from_sync
 
 SUPPORTED_AGENT_CLS = {'BrowsingAgent'}
 
@@ -144,6 +145,7 @@ def process_instance(
         logger.info(f'Starting evaluation for instance {env_id}.')
 
     runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     task_str = initialize_runtime(runtime)
 
     state: State | None = asyncio.run(
diff --git a/evaluation/integration_tests/run_infer.py b/evaluation/integration_tests/run_infer.py
index 61c7e3bde9..d215b0599b 100644
--- a/evaluation/integration_tests/run_infer.py
+++ b/evaluation/integration_tests/run_infer.py
@@ -30,6 +30,7 @@ from openhands.core.main import create_runtime, run_controller
 from openhands.events.action import MessageAction
 from openhands.events.serialization.event import event_to_dict
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 FAKE_RESPONSES = {
     'CodeActAgent': fake_user_response,
@@ -108,6 +109,7 @@ def process_instance(
     # create sandbox and run the agent
     # =============================================
     runtime: Runtime = create_runtime(config)
+    call_async_from_sync(runtime.connect)
     try:
         test_class.initialize_runtime(runtime)
 
diff --git a/openhands/core/cli.py b/openhands/core/cli.py
index 0e43cfc2d6..bb134803ae 100644
--- a/openhands/core/cli.py
+++ b/openhands/core/cli.py
@@ -14,7 +14,12 @@ from openhands.core.config import (
 from openhands.core.logger import openhands_logger as logger
 from openhands.core.loop import run_agent_until_done
 from openhands.core.schema import AgentState
-from openhands.core.setup import create_agent, create_controller, create_runtime
+from openhands.core.setup import (
+    create_agent,
+    create_controller,
+    create_runtime,
+    initialize_repository_for_runtime,
+)
 from openhands.events import EventSource, EventStreamSubscriber
 from openhands.events.action import (
     Action,
@@ -109,7 +114,6 @@ async def main(loop: asyncio.AbstractEventLoop):
         sid=sid,
         headless_mode=True,
         agent=agent,
-        selected_repository=config.sandbox.selected_repo,
     )
 
     controller, _ = create_controller(agent, runtime, config)
@@ -165,6 +169,14 @@ async def main(loop: asyncio.AbstractEventLoop):
 
     await runtime.connect()
 
+    # Initialize repository if needed
+    if config.sandbox.selected_repo:
+        initialize_repository_for_runtime(
+            runtime,
+            agent=agent,
+            selected_repository=config.sandbox.selected_repo,
+        )
+
     if initial_user_action:
         # If there's an initial user action, enqueue it and do not prompt again
         event_stream.add_event(initial_user_action, EventSource.USER)
diff --git a/openhands/core/main.py b/openhands/core/main.py
index ab642553b3..4b282864f2 100644
--- a/openhands/core/main.py
+++ b/openhands/core/main.py
@@ -20,6 +20,7 @@ from openhands.core.setup import (
     create_controller,
     create_runtime,
     generate_sid,
+    initialize_repository_for_runtime,
 )
 from openhands.events import EventSource, EventStreamSubscriber
 from openhands.events.action import MessageAction, NullAction
@@ -29,6 +30,7 @@ from openhands.events.observation import AgentStateChangedObservation
 from openhands.events.serialization import event_from_dict
 from openhands.io import read_input, read_task
 from openhands.runtime.base import Runtime
+from openhands.utils.async_utils import call_async_from_sync
 
 
 class FakeUserResponseFunc(Protocol):
@@ -97,8 +99,17 @@ async def run_controller(
             sid=sid,
             headless_mode=headless_mode,
             agent=agent,
-            selected_repository=config.sandbox.selected_repo,
         )
+        # Connect to the runtime
+        call_async_from_sync(runtime.connect)
+
+        # Initialize repository if needed
+        if config.sandbox.selected_repo:
+            initialize_repository_for_runtime(
+                runtime,
+                agent=agent,
+                selected_repository=config.sandbox.selected_repo,
+            )
 
     event_stream = runtime.event_stream
 
diff --git a/openhands/core/setup.py b/openhands/core/setup.py
index 9142be6713..709ef923b9 100644
--- a/openhands/core/setup.py
+++ b/openhands/core/setup.py
@@ -21,7 +21,6 @@ from openhands.runtime import get_runtime_cls
 from openhands.runtime.base import Runtime
 from openhands.security import SecurityAnalyzer, options
 from openhands.storage import get_file_store
-from openhands.utils.async_utils import call_async_from_sync
 
 
 def create_runtime(
@@ -29,18 +28,19 @@ def create_runtime(
     sid: str | None = None,
     headless_mode: bool = True,
     agent: Agent | None = None,
-    selected_repository: str | None = None,
-    github_token: SecretStr | None = None,
 ) -> Runtime:
     """Create a runtime for the agent to run on.
 
-    config: The app config.
-    sid: (optional) The session id. IMPORTANT: please don't set this unless you know what you're doing.
-        Set it to incompatible value will cause unexpected behavior on RemoteRuntime.
-    headless_mode: Whether the agent is run in headless mode. `create_runtime` is typically called within evaluation scripts,
-        where we don't want to have the VSCode UI open, so it defaults to True.
-    selected_repository: (optional) The GitHub repository to use.
-    github_token: (optional) The GitHub token to use.
+    Args:
+        config: The app config.
+        sid: (optional) The session id. IMPORTANT: please don't set this unless you know what you're doing.
+            Set it to incompatible value will cause unexpected behavior on RemoteRuntime.
+        headless_mode: Whether the agent is run in headless mode. `create_runtime` is typically called within evaluation scripts,
+            where we don't want to have the VSCode UI open, so it defaults to True.
+        agent: (optional) The agent instance to use for configuring the runtime.
+
+    Returns:
+        The created Runtime instance (not yet connected or initialized).
     """
     # if sid is provided on the command line, use it as the name of the event stream
     # otherwise generate it on the basis of the configured jwt_secret
@@ -74,8 +74,30 @@ def create_runtime(
         headless_mode=headless_mode,
     )
 
-    call_async_from_sync(runtime.connect)
+    logger.debug(
+        f'Runtime created with plugins: {[plugin.name for plugin in runtime.plugins]}'
+    )
 
+    return runtime
+
+
+def initialize_repository_for_runtime(
+    runtime: Runtime,
+    agent: Agent | None = None,
+    selected_repository: str | None = None,
+    github_token: SecretStr | None = None,
+) -> str | None:
+    """Initialize the repository for the runtime.
+
+    Args:
+        runtime: The runtime to initialize the repository for.
+        agent: (optional) The agent to load microagents for.
+        selected_repository: (optional) The GitHub repository to use.
+        github_token: (optional) The GitHub token to use.
+
+    Returns:
+        The repository directory path if a repository was cloned, None otherwise.
+    """
     # clone selected repository if provided
     repo_directory = None
     github_token = (
@@ -98,11 +120,7 @@ def create_runtime(
         agent.prompt_manager.load_microagents(microagents)
         agent.prompt_manager.set_repository_info(selected_repository, repo_directory)
 
-    logger.debug(
-        f'Runtime initialized with plugins: {[plugin.name for plugin in runtime.plugins]}'
-    )
-
-    return runtime
+    return repo_directory
 
 
 def create_agent(config: AppConfig) -> Agent: