V1: Experiment manager (#11388)

Co-authored-by: openhands <openhands@all-hands.dev>

enterprise/experiments/experiment_manager.py

@@ -1,18 +1,47 @@
+from uuid import UUID
+
 from experiments.constants import (
     ENABLE_EXPERIMENT_MANAGER,
+    EXPERIMENT_SYSTEM_PROMPT_EXPERIMENT,
 )
 from experiments.experiment_versions import (
     handle_condenser_max_step_experiment,
     handle_system_prompt_experiment,
 )
+from experiments.experiment_versions._004_condenser_max_step_experiment import (
+    handle_condenser_max_step_experiment__v1,
+)
 
 from openhands.core.config.openhands_config import OpenHandsConfig
 from openhands.core.logger import openhands_logger as logger
 from openhands.experiments.experiment_manager import ExperimentManager
+from openhands.sdk import Agent
 from openhands.server.session.conversation_init_data import ConversationInitData
 
 
 class SaaSExperimentManager(ExperimentManager):
+    @staticmethod
+    def run_agent_variant_tests__v1(
+        user_id: str | None, conversation_id: UUID, agent: Agent
+    ) -> Agent:
+        if not ENABLE_EXPERIMENT_MANAGER:
+            logger.info(
+                'experiment_manager:run_conversation_variant_test:skipped',
+                extra={'reason': 'experiment_manager_disabled'},
+            )
+            return agent
+
+        agent = handle_condenser_max_step_experiment__v1(
+            user_id, conversation_id, agent
+        )
+
+        if EXPERIMENT_SYSTEM_PROMPT_EXPERIMENT:
+            agent = agent.model_copy(
+                update={'system_prompt_filename': 'system_prompt_long_horizon.j2'}
+            )
+
+        return agent
+
     @staticmethod
     def run_conversation_variant_test(
         user_id, conversation_id, conversation_settings

enterprise/experiments/experiment_versions/_004_condenser_max_step_experiment.py

@@ -5,12 +5,18 @@ This module contains the handler for the condenser max step experiment that test
 different max_size values for the condenser configuration.
 """
 
+from uuid import UUID
+
 import posthog
 from experiments.constants import EXPERIMENT_CONDENSER_MAX_STEP
 from server.constants import IS_FEATURE_ENV
 from storage.experiment_assignment_store import ExperimentAssignmentStore
 
 from openhands.core.logger import openhands_logger as logger
+from openhands.sdk import Agent
+from openhands.sdk.context.condenser import (
+    LLMSummarizingCondenser,
+)
 from openhands.server.session.conversation_init_data import ConversationInitData
 
 
@@ -190,3 +196,37 @@ def handle_condenser_max_step_experiment(
         return conversation_settings
 
     return conversation_settings
+
+
+def handle_condenser_max_step_experiment__v1(
+    user_id: str | None,
+    conversation_id: UUID,
+    agent: Agent,
+) -> Agent:
+    enabled_variant = _get_condenser_max_step_variant(user_id, str(conversation_id))
+
+    if enabled_variant is None:
+        return agent
+
+    if enabled_variant == 'control':
+        condenser_max_size = 120
+    elif enabled_variant == 'treatment':
+        condenser_max_size = 80
+    else:
+        logger.error(
+            'condenser_max_step_experiment:unknown_variant',
+            extra={
+                'user_id': user_id,
+                'convo_id': conversation_id,
+                'variant': enabled_variant,
+                'reason': 'unknown variant; returning original conversation settings',
+            },
+        )
+        return agent
+
+    condenser_llm = agent.llm.model_copy(update={'usage_id': 'condenser'})
+    condenser = LLMSummarizingCondenser(
+        llm=condenser_llm, max_size=condenser_max_size, keep_first=4
+    )
+
+    return agent.model_copy(update={'condenser': condenser})

enterprise/poetry.lock

@@ -5737,7 +5737,7 @@ llama = ["llama-index (>=0.12.29,<0.13.0)", "llama-index-core (>=0.12.29,<0.13.0
 
 [[package]]
 name = "openhands-agent-server"
-version = "1.0.0a2"
+version = "1.0.0a3"
 description = "OpenHands Agent Server - REST/WebSocket interface for OpenHands AI Agent"
 optional = false
 python-versions = ">=3.12"
@@ -5759,8 +5759,8 @@ wsproto = ">=1.2.0"
 [package.source]
 type = "git"
 url = "https://github.com/All-Hands-AI/agent-sdk.git"
-reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
-resolved_reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+resolved_reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
 subdirectory = "openhands-agent-server"
 
 [[package]]
@@ -5805,9 +5805,9 @@ memory-profiler = "^0.61.0"
 numpy = "*"
 openai = "1.99.9"
 openhands-aci = "0.3.2"
-openhands-agent-server = {git = "https://github.com/All-Hands-AI/agent-sdk.git", rev = "512399d896521aee3131eea4bb59087fb9dfa243", subdirectory = "openhands-agent-server"}
+openhands-agent-server = {git = "https://github.com/All-Hands-AI/agent-sdk.git", rev = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e", subdirectory = "openhands-agent-server"}
-openhands-sdk = {git = "https://github.com/All-Hands-AI/agent-sdk.git", rev = "512399d896521aee3131eea4bb59087fb9dfa243", subdirectory = "openhands-sdk"}
+openhands-sdk = {git = "https://github.com/All-Hands-AI/agent-sdk.git", rev = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e", subdirectory = "openhands-sdk"}
-openhands-tools = {git = "https://github.com/All-Hands-AI/agent-sdk.git", rev = "512399d896521aee3131eea4bb59087fb9dfa243", subdirectory = "openhands-tools"}
+openhands-tools = {git = "https://github.com/All-Hands-AI/agent-sdk.git", rev = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e", subdirectory = "openhands-tools"}
 opentelemetry-api = "^1.33.1"
 opentelemetry-exporter-otlp-proto-grpc = "^1.33.1"
 pathspec = "^0.12.1"
@@ -5863,7 +5863,7 @@ url = ".."
 
 [[package]]
 name = "openhands-sdk"
-version = "1.0.0a2"
+version = "1.0.0a3"
 description = "OpenHands SDK - Core functionality for building AI agents"
 optional = false
 python-versions = ">=3.12"
@@ -5887,13 +5887,13 @@ boto3 = ["boto3 (>=1.35.0)"]
 [package.source]
 type = "git"
 url = "https://github.com/All-Hands-AI/agent-sdk.git"
-reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
-resolved_reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+resolved_reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
 subdirectory = "openhands-sdk"
 
 [[package]]
 name = "openhands-tools"
-version = "1.0.0a2"
+version = "1.0.0a3"
 description = "OpenHands Tools - Runtime tools for AI agents"
 optional = false
 python-versions = ">=3.12"
@@ -5914,8 +5914,8 @@ pydantic = ">=2.11.7"
 [package.source]
 type = "git"
 url = "https://github.com/All-Hands-AI/agent-sdk.git"
-reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
-resolved_reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+resolved_reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
 subdirectory = "openhands-tools"
 
 [[package]]

enterprise/tests/unit/experiments/__init__.py

@@ -0,0 +1 @@
+"""Unit tests for experiments module."""

enterprise/tests/unit/experiments/test_saas_experiment_manager.py

@@ -0,0 +1,137 @@
+# tests/test_condenser_max_step_experiment_v1.py
+
+from unittest.mock import patch
+from uuid import uuid4
+
+from experiments.experiment_manager import SaaSExperimentManager
+
+# SUT imports (update the module path if needed)
+from experiments.experiment_versions._004_condenser_max_step_experiment import (
+    handle_condenser_max_step_experiment__v1,
+)
+from pydantic import SecretStr
+
+from openhands.sdk import LLM, Agent
+from openhands.sdk.context.condenser import LLMSummarizingCondenser
+
+
+def make_agent() -> Agent:
+    """Build a minimal valid Agent."""
+    llm = LLM(
+        usage_id='primary-llm',
+        model='provider/model',
+        api_key=SecretStr('sk-test'),
+    )
+    return Agent(llm=llm)
+
+
+def _patch_variant(monkeypatch, return_value):
+    """Patch the internal variant getter to return a specific value."""
+    monkeypatch.setattr(
+        'experiments.experiment_versions._004_condenser_max_step_experiment._get_condenser_max_step_variant',
+        lambda user_id, conv_id: return_value,
+        raising=True,
+    )
+
+
+def test_control_variant_sets_condenser_with_max_size_120(monkeypatch):
+    _patch_variant(monkeypatch, 'control')
+    agent = make_agent()
+    conv_id = uuid4()
+
+    result = handle_condenser_max_step_experiment__v1('user-1', conv_id, agent)
+
+    # Should be a new Agent instance with a condenser installed
+    assert result is not agent
+    assert isinstance(result.condenser, LLMSummarizingCondenser)
+
+    # The condenser should have its own LLM (usage_id overridden to "condenser")
+    assert result.condenser.llm.usage_id == 'condenser'
+    # The original agent LLM remains unchanged
+    assert agent.llm.usage_id == 'primary-llm'
+
+    # Control: max_size = 120, keep_first = 4
+    assert result.condenser.max_size == 120
+    assert result.condenser.keep_first == 4
+
+
+def test_treatment_variant_sets_condenser_with_max_size_80(monkeypatch):
+    _patch_variant(monkeypatch, 'treatment')
+    agent = make_agent()
+    conv_id = uuid4()
+
+    result = handle_condenser_max_step_experiment__v1('user-2', conv_id, agent)
+
+    assert result is not agent
+    assert isinstance(result.condenser, LLMSummarizingCondenser)
+    assert result.condenser.llm.usage_id == 'condenser'
+    assert result.condenser.max_size == 80
+    assert result.condenser.keep_first == 4
+
+
+def test_none_variant_returns_original_agent_without_changes(monkeypatch):
+    _patch_variant(monkeypatch, None)
+    agent = make_agent()
+    conv_id = uuid4()
+
+    result = handle_condenser_max_step_experiment__v1('user-3', conv_id, agent)
+
+    # No changes—same instance and no condenser attribute added
+    assert result is agent
+    assert getattr(result, 'condenser', None) is None
+
+
+def test_unknown_variant_returns_original_agent_without_changes(monkeypatch):
+    _patch_variant(monkeypatch, 'weird-variant')
+    agent = make_agent()
+    conv_id = uuid4()
+
+    result = handle_condenser_max_step_experiment__v1('user-4', conv_id, agent)
+
+    assert result is agent
+    assert getattr(result, 'condenser', None) is None
+
+
+@patch('experiments.experiment_manager.handle_condenser_max_step_experiment__v1')
+@patch('experiments.experiment_manager.ENABLE_EXPERIMENT_MANAGER', False)
+def test_run_agent_variant_tests_v1_noop_when_manager_disabled(
+    mock_handle_condenser,
+):
+    """If ENABLE_EXPERIMENT_MANAGER is False, the method returns the exact same agent and does not call the handler."""
+    agent = make_agent()
+    conv_id = uuid4()
+
+    result = SaaSExperimentManager.run_agent_variant_tests__v1(
+        user_id='user-123',
+        conversation_id=conv_id,
+        agent=agent,
+    )
+
+    # Same object returned (no copy)
+    assert result is agent
+    # Handler should not have been called
+    mock_handle_condenser.assert_not_called()
+
+
+@patch('experiments.experiment_manager.ENABLE_EXPERIMENT_MANAGER', True)
+@patch('experiments.experiment_manager.EXPERIMENT_SYSTEM_PROMPT_EXPERIMENT', True)
+def test_run_agent_variant_tests_v1_calls_handler_and_sets_system_prompt(monkeypatch):
+    """When enabled, it should call the condenser experiment handler and set the long-horizon system prompt."""
+    agent = make_agent()
+    conv_id = uuid4()
+
+    _patch_variant(monkeypatch, 'treatment')
+
+    result: Agent = SaaSExperimentManager.run_agent_variant_tests__v1(
+        user_id='user-abc',
+        conversation_id=conv_id,
+        agent=agent,
+    )
+
+    # Should be a different instance than the original (copied after handler runs)
+    assert result is not agent
+    assert result.system_prompt_filename == 'system_prompt_long_horizon.j2'
+
+    # The condenser returned by the handler must be preserved after the system-prompt override copy
+    assert isinstance(result.condenser, LLMSummarizingCondenser)
+    assert result.condenser.max_size == 80

openhands/app_server/app_conversation/live_status_app_conversation_service.py

@@ -5,7 +5,7 @@ from dataclasses import dataclass
 from datetime import datetime, timedelta
 from time import time
 from typing import AsyncGenerator, Sequence
-from uuid import UUID
+from uuid import UUID, uuid4
 
 import httpx
 from fastapi import Request
@@ -52,6 +52,7 @@ from openhands.app_server.services.injector import InjectorState
 from openhands.app_server.services.jwt_service import JwtService
 from openhands.app_server.user.user_context import UserContext
 from openhands.app_server.utils.async_remote_workspace import AsyncRemoteWorkspace
+from openhands.experiments.experiment_manager import ExperimentManagerImpl
 from openhands.integrations.provider import ProviderType
 from openhands.sdk import LocalWorkspace
 from openhands.sdk.conversation.secret_source import LookupSecret, StaticSecret
@@ -458,10 +459,17 @@ class LiveStatusAppConversationService(GitAppConversationService):
             model=user.llm_model,
             base_url=user.llm_base_url,
             api_key=user.llm_api_key,
-            service_id='agent',
+            usage_id='agent',
         )
         agent = get_default_agent(llm=llm)
+
+        conversation_id = uuid4()
+        agent = ExperimentManagerImpl.run_agent_variant_tests__v1(
+            user.id, conversation_id, agent
+        )
+
         start_conversation_request = StartConversationRequest(
+            conversation_id=conversation_id,
             agent=agent,
             workspace=workspace,
             confirmation_policy=AlwaysConfirm()

openhands/experiments/experiment_manager.py

@@ -1,9 +1,11 @@
 import os
+from uuid import UUID
 
 from pydantic import BaseModel
 
 from openhands.core.config.openhands_config import OpenHandsConfig
 from openhands.core.logger import openhands_logger as logger
+from openhands.sdk import Agent
 from openhands.server.session.conversation_init_data import ConversationInitData
 from openhands.server.shared import file_store
 from openhands.storage.locations import get_experiment_config_filename
@@ -29,6 +31,12 @@ def load_experiment_config(conversation_id: str) -> ExperimentConfig | None:
 
 
 class ExperimentManager:
+    @staticmethod
+    def run_agent_variant_tests__v1(
+        user_id: str | None, conversation_id: UUID, agent: Agent
+    ) -> Agent:
+        return agent
+
     @staticmethod
     def run_conversation_variant_test(
         user_id: str | None,

poetry.lock

@@ -5711,8 +5711,11 @@ files = [
     {file = "lxml-5.4.0-cp36-cp36m-win_amd64.whl", hash = "sha256:7ce1a171ec325192c6a636b64c94418e71a1964f56d002cc28122fceff0b6121"},
     {file = "lxml-5.4.0-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:795f61bcaf8770e1b37eec24edf9771b307df3af74d1d6f27d812e15a9ff3872"},
     {file = "lxml-5.4.0-cp37-cp37m-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:29f451a4b614a7b5b6c2e043d7b64a15bd8304d7e767055e8ab68387a8cacf4e"},
+    {file = "lxml-5.4.0-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:891f7f991a68d20c75cb13c5c9142b2a3f9eb161f1f12a9489c82172d1f133c0"},
     {file = "lxml-5.4.0-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:4aa412a82e460571fad592d0f93ce9935a20090029ba08eca05c614f99b0cc92"},
+    {file = "lxml-5.4.0-cp37-cp37m-manylinux_2_28_aarch64.whl", hash = "sha256:ac7ba71f9561cd7d7b55e1ea5511543c0282e2b6450f122672a2694621d63b7e"},
     {file = "lxml-5.4.0-cp37-cp37m-manylinux_2_28_x86_64.whl", hash = "sha256:c5d32f5284012deaccd37da1e2cd42f081feaa76981f0eaa474351b68df813c5"},
+    {file = "lxml-5.4.0-cp37-cp37m-musllinux_1_2_aarch64.whl", hash = "sha256:ce31158630a6ac85bddd6b830cffd46085ff90498b397bd0a259f59d27a12188"},
     {file = "lxml-5.4.0-cp37-cp37m-musllinux_1_2_x86_64.whl", hash = "sha256:31e63621e073e04697c1b2d23fcb89991790eef370ec37ce4d5d469f40924ed6"},
     {file = "lxml-5.4.0-cp37-cp37m-win32.whl", hash = "sha256:be2ba4c3c5b7900246a8f866580700ef0d538f2ca32535e991027bdaba944063"},
     {file = "lxml-5.4.0-cp37-cp37m-win_amd64.whl", hash = "sha256:09846782b1ef650b321484ad429217f5154da4d6e786636c38e434fa32e94e49"},
@@ -7272,7 +7275,7 @@ llama = ["llama-index (>=0.12.29,<0.13.0)", "llama-index-core (>=0.12.29,<0.13.0
 
 [[package]]
 name = "openhands-agent-server"
-version = "1.0.0a2"
+version = "1.0.0a3"
 description = "OpenHands Agent Server - REST/WebSocket interface for OpenHands AI Agent"
 optional = false
 python-versions = ">=3.12"
@@ -7294,13 +7297,13 @@ wsproto = ">=1.2.0"
 [package.source]
 type = "git"
 url = "https://github.com/All-Hands-AI/agent-sdk.git"
-reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
-resolved_reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+resolved_reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
 subdirectory = "openhands-agent-server"
 
 [[package]]
 name = "openhands-sdk"
-version = "1.0.0a2"
+version = "1.0.0a3"
 description = "OpenHands SDK - Core functionality for building AI agents"
 optional = false
 python-versions = ">=3.12"
@@ -7324,13 +7327,13 @@ boto3 = ["boto3 (>=1.35.0)"]
 [package.source]
 type = "git"
 url = "https://github.com/All-Hands-AI/agent-sdk.git"
-reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
-resolved_reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+resolved_reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
 subdirectory = "openhands-sdk"
 
 [[package]]
 name = "openhands-tools"
-version = "1.0.0a2"
+version = "1.0.0a3"
 description = "OpenHands Tools - Runtime tools for AI agents"
 optional = false
 python-versions = ">=3.12"
@@ -7351,8 +7354,8 @@ pydantic = ">=2.11.7"
 [package.source]
 type = "git"
 url = "https://github.com/All-Hands-AI/agent-sdk.git"
-reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
-resolved_reference = "512399d896521aee3131eea4bb59087fb9dfa243"
+resolved_reference = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e"
 subdirectory = "openhands-tools"
 
 [[package]]
@@ -16521,4 +16524,4 @@ third-party-runtimes = ["daytona", "e2b-code-interpreter", "modal", "runloop-api
 [metadata]
 lock-version = "2.1"
 python-versions = "^3.12,<3.14"
-content-hash = "03639ad9782d05163b25c507e7232d797572902ee57408bf999b72c21e3adf5e"
+content-hash = "fd68ed845befeb646ee910db46f1ef9c5a1fd2e6d1ac6189c04864e0665f66ed"

pyproject.toml

@@ -113,10 +113,10 @@ e2b-code-interpreter = { version = "^2.0.0", optional = true }
 pybase62 = "^1.0.0"
 
 # V1 dependencies
-openhands-agent-server = { git = "https://github.com/All-Hands-AI/agent-sdk.git", subdirectory = "openhands-agent-server", rev = "512399d896521aee3131eea4bb59087fb9dfa243" }
+openhands-agent-server = { git = "https://github.com/All-Hands-AI/agent-sdk.git", subdirectory = "openhands-agent-server", rev = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e" }
-openhands-sdk = { git = "https://github.com/All-Hands-AI/agent-sdk.git", subdirectory = "openhands-sdk", rev = "512399d896521aee3131eea4bb59087fb9dfa243" }
+openhands-sdk = { git = "https://github.com/All-Hands-AI/agent-sdk.git", subdirectory = "openhands-sdk", rev = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e" }
 # This refuses to install
-openhands-tools = { git = "https://github.com/All-Hands-AI/agent-sdk.git", subdirectory = "openhands-tools", rev = "512399d896521aee3131eea4bb59087fb9dfa243" }
+openhands-tools = { git = "https://github.com/All-Hands-AI/agent-sdk.git", subdirectory = "openhands-tools", rev = "8d8134ca5a87cc3e90e3ff968327a7f4c961e22e" }
 python-jose = { version = ">=3.3", extras = [ "cryptography" ] }
 sqlalchemy = { extras = [ "asyncio" ], version = "^2.0.40" }
 pg8000 = "^1.31.5"

tests/unit/experiments/test_experiment_manager.py

@@ -0,0 +1,215 @@
+"""Unit tests for ExperimentManager class, focusing on the v1 agent method."""
+
+from types import SimpleNamespace
+from unittest.mock import Mock, patch
+from uuid import UUID, uuid4
+
+import pytest
+
+from openhands.app_server.app_conversation.live_status_app_conversation_service import (
+    LiveStatusAppConversationService,
+)
+from openhands.experiments.experiment_manager import ExperimentManager
+from openhands.sdk import Agent
+from openhands.sdk.llm import LLM
+
+
+class TestExperimentManager:
+    """Test cases for ExperimentManager class."""
+
+    def setup_method(self):
+        """Set up test fixtures."""
+        self.user_id = 'test_user_123'
+        self.conversation_id = uuid4()
+
+        # Create a mock LLM
+        self.mock_llm = Mock(spec=LLM)
+        self.mock_llm.model = 'gpt-4'
+        self.mock_llm.usage_id = 'agent'
+
+        # Create a mock Agent
+        self.mock_agent = Mock(spec=Agent)
+        self.mock_agent.llm = self.mock_llm
+        self.mock_agent.system_prompt_filename = 'default_system_prompt.j2'
+        self.mock_agent.model_copy = Mock(return_value=self.mock_agent)
+
+    def test_run_agent_variant_tests__v1_returns_agent_unchanged(self):
+        """Test that the base ExperimentManager returns the agent unchanged."""
+        result = ExperimentManager.run_agent_variant_tests__v1(
+            self.user_id, self.conversation_id, self.mock_agent
+        )
+
+        assert result is self.mock_agent
+        assert result == self.mock_agent
+
+    def test_run_agent_variant_tests__v1_with_none_user_id(self):
+        """Test that the method works with None user_id."""
+        # Act
+        result = ExperimentManager.run_agent_variant_tests__v1(
+            None, self.conversation_id, self.mock_agent
+        )
+
+        # Assert
+        assert result is self.mock_agent
+
+    def test_run_agent_variant_tests__v1_with_different_conversation_ids(self):
+        """Test that the method works with different conversation IDs."""
+        conversation_id_1 = uuid4()
+        conversation_id_2 = uuid4()
+
+        # Act
+        result_1 = ExperimentManager.run_agent_variant_tests__v1(
+            self.user_id, conversation_id_1, self.mock_agent
+        )
+        result_2 = ExperimentManager.run_agent_variant_tests__v1(
+            self.user_id, conversation_id_2, self.mock_agent
+        )
+
+        # Assert
+        assert result_1 is self.mock_agent
+        assert result_2 is self.mock_agent
+
+
+class TestExperimentManagerIntegration:
+    """Integration tests for ExperimentManager with start_app_conversation."""
+
+    def setup_method(self):
+        """Set up test fixtures."""
+        self.user_id = 'test_user_123'
+        self.conversation_id = uuid4()
+
+        # Create a mock LLM
+        self.mock_llm = Mock(spec=LLM)
+        self.mock_llm.model = 'gpt-4'
+        self.mock_llm.usage_id = 'agent'
+
+        # Create a mock Agent
+        self.mock_agent = Mock(spec=Agent)
+        self.mock_agent.llm = self.mock_llm
+        self.mock_agent.system_prompt_filename = 'default_system_prompt.j2'
+        self.mock_agent.model_copy = Mock(return_value=self.mock_agent)
+
+    @patch('openhands.experiments.experiment_manager.ExperimentManagerImpl')
+    def test_start_app_conversation_calls_experiment_manager_v1(
+        self, mock_experiment_manager_impl
+    ):
+        """Test that start_app_conversation calls the experiment manager v1 method with correct parameters."""
+        # Arrange
+        mock_experiment_manager_impl.run_agent_variant_tests__v1.return_value = (
+            self.mock_agent
+        )
+
+        # Create a mock service instance
+        mock_service = Mock(spec=LiveStatusAppConversationService)
+
+        # Mock the _build_start_conversation_request_for_user method to simulate the call
+        with patch.object(mock_service, '_build_start_conversation_request_for_user'):
+            # Simulate the part of the code that calls the experiment manager
+            from uuid import uuid4
+
+            conversation_id = uuid4()
+
+            # This simulates the call that happens in the actual service
+            result_agent = mock_experiment_manager_impl.run_agent_variant_tests__v1(
+                self.user_id, conversation_id, self.mock_agent
+            )
+
+            # Assert
+            mock_experiment_manager_impl.run_agent_variant_tests__v1.assert_called_once_with(
+                self.user_id, conversation_id, self.mock_agent
+            )
+            assert result_agent == self.mock_agent
+
+    @pytest.mark.asyncio
+    async def test_experiment_manager_called_with_correct_parameters_in_context__noop_pass_through(
+        self,
+    ):
+        """
+        Use the real LiveStatusAppConversationService to build a StartConversationRequest,
+        and verify ExperimentManagerImpl.run_agent_variant_tests__v1:
+        - is called exactly once with the (user_id, generated conversation_id, agent)
+        - returns the *same* agent instance (no copy/mutation)
+        - does not tweak agent fields (LLM, system prompt, etc.)
+        """
+        # --- Arrange: fixed UUID to assert call parameters deterministically
+        fixed_conversation_id = UUID('00000000-0000-0000-0000-000000000001')
+
+        # Create a stable Agent (and LLM) we can identity-check later
+        mock_llm = Mock(spec=LLM)
+        mock_llm.model = 'gpt-4'
+        mock_llm.usage_id = 'agent'
+
+        mock_agent = Mock(spec=Agent)
+        mock_agent.llm = mock_llm
+        mock_agent.system_prompt_filename = 'default_system_prompt.j2'
+
+        # Minimal, real-ish user context used by the service
+        class DummyUserContext:
+            async def get_user_info(self):
+                # confirmation_mode=False -> NeverConfirm()
+                return SimpleNamespace(
+                    id='test_user_123',
+                    llm_model='gpt-4',
+                    llm_base_url=None,
+                    llm_api_key=None,
+                    confirmation_mode=False,
+                )
+
+            async def get_secrets(self):
+                return {}
+
+            async def get_latest_token(self, provider):
+                return None
+
+            async def get_user_id(self):
+                return 'test_user_123'
+
+        user_context = DummyUserContext()
+
+        # The service requires a lot of deps, but for this test we won't exercise them.
+        app_conversation_info_service = Mock()
+        app_conversation_start_task_service = Mock()
+        sandbox_service = Mock()
+        sandbox_spec_service = Mock()
+        jwt_service = Mock()
+        httpx_client = Mock()
+
+        service = LiveStatusAppConversationService(
+            init_git_in_empty_workspace=False,
+            user_context=user_context,
+            app_conversation_info_service=app_conversation_info_service,
+            app_conversation_start_task_service=app_conversation_start_task_service,
+            sandbox_service=sandbox_service,
+            sandbox_spec_service=sandbox_spec_service,
+            jwt_service=jwt_service,
+            sandbox_startup_timeout=30,
+            sandbox_startup_poll_frequency=1,
+            httpx_client=httpx_client,
+            web_url=None,
+            access_token_hard_timeout=None,
+        )
+
+        # Patch the pieces invoked by the service
+        with (
+            patch(
+                'openhands.app_server.app_conversation.live_status_app_conversation_service.get_default_agent',
+                return_value=mock_agent,
+            ),
+            patch(
+                'openhands.app_server.app_conversation.live_status_app_conversation_service.uuid4',
+                return_value=fixed_conversation_id,
+            ),
+        ):
+            # --- Act: build the start request
+            start_req = await service._build_start_conversation_request_for_user(
+                initial_message=None,
+                git_provider=None,  # Keep secrets path simple
+                working_dir='/tmp/project',  # Arbitrary path
+            )
+
+            # The agent in the StartConversationRequest is the *same* object we provided
+            assert start_req.agent is mock_agent
+
+            # No tweaks to agent fields by the experiment manager (noop)
+            assert start_req.agent.llm is mock_llm
+            assert start_req.agent.system_prompt_filename == 'default_system_prompt.j2'