diff --git a/openhands-agent-server/openhands/agent_server/event_service.py b/openhands-agent-server/openhands/agent_server/event_service.py index 6388c65de43..93e3f22d49d 100644 --- a/openhands-agent-server/openhands/agent_server/event_service.py +++ b/openhands-agent-server/openhands/agent_server/event_service.py @@ -81,7 +81,10 @@ StreamingDeltaEvent, ) from openhands.sdk.event.conversation_error import ConversationErrorEvent -from openhands.sdk.event.conversation_state import ConversationStateUpdateEvent +from openhands.sdk.event.conversation_state import ( + FULL_STATE_KEY, + ConversationStateUpdateEvent, +) from openhands.sdk.event.error_classification import ErrorClassification, FailureKind from openhands.sdk.event.llm_completion_log import LLMCompletionLogEvent from openhands.sdk.git.exceptions import GitCommandError, GitRepositoryError @@ -111,6 +114,23 @@ _GIT_INIT_LOCK = threading.Lock() +def _without_skill_bodies( + event: ConversationStateUpdateEvent, +) -> ConversationStateUpdateEvent: + """Leave skill bodies, most of the agent's size, out of the agent that + sockets and webhooks receive, as conversation responses do by default.""" + if event.key == FULL_STATE_KEY: + agent = event.value.get("agent") or {} + elif event.key == "agent": + agent = event.value.model_dump(mode="json") + event = event.model_copy(update={"value": agent}) + else: + return event + if agent.get("agent_context"): + agent["agent_context"]["skills"] = [] + return event + + class ConversationRunLimitExceeded(RateLimited, RuntimeError): """The server has no capacity to start another conversation run.""" @@ -847,7 +867,8 @@ def _create_state_update_event_sync(self) -> ConversationStateUpdateEvent: raise ValueError("inactive_service") state = self._conversation._state with state: - return ConversationStateUpdateEvent.from_conversation_state(state) + event = ConversationStateUpdateEvent.from_conversation_state(state) + return _without_skill_bodies(event) def _create_initial_state_sync( self, @@ -1325,7 +1346,7 @@ async def start(self): # Create and store callback wrapper to allow flushing pending events self._callback_wrapper = AsyncCallbackWrapper( - self._pub_sub, loop=asyncio.get_running_loop() + self._publish_to_subscribers, loop=asyncio.get_running_loop() ) # Token streaming is wired only for agents that can actually emit token @@ -2492,6 +2513,11 @@ async def get_state(self) -> ConversationState: raise ValueError("inactive_service") return self._conversation._state + async def _publish_to_subscribers(self, event: Event) -> None: + if isinstance(event, ConversationStateUpdateEvent): + event = _without_skill_bodies(event) + await self._pub_sub(event) + async def _publish_state_update(self): """Publish a ConversationStateUpdateEvent with the current state.""" if not self._conversation: diff --git a/tests/agent_server/test_event_service.py b/tests/agent_server/test_event_service.py index dccebd6ef9b..92100cb3ea4 100644 --- a/tests/agent_server/test_event_service.py +++ b/tests/agent_server/test_event_service.py @@ -39,6 +39,7 @@ from openhands.agent_server.pub_sub import Subscriber from openhands.sdk import LLM, Agent, AgentBase, Conversation, Message from openhands.sdk.agent import ACPAgent +from openhands.sdk.context import AgentContext from openhands.sdk.context.condenser import NoOpCondenser from openhands.sdk.conversation.event_store import EventLog from openhands.sdk.conversation.exceptions import ConversationRunError @@ -70,6 +71,7 @@ from openhands.sdk.mcp.config import coerce_mcp_config from openhands.sdk.secret import StaticSecret from openhands.sdk.security.confirmation_policy import NeverConfirm +from openhands.sdk.skills import Skill from openhands.sdk.subagent.schema import AgentDefinition from openhands.sdk.utils.cipher import FERNET_TOKEN_PREFIX, Cipher from openhands.sdk.utils.pydantic_secrets import REDACTED_SECRET_VALUE @@ -2808,6 +2810,64 @@ def mock_conversation_with_real_lock(self): conversation._state = state return conversation + @pytest.mark.asyncio + async def test_initial_state_leaves_out_skill_bodies( + self, event_service, mock_conversation_with_real_lock + ): + state = mock_conversation_with_real_lock._state + state.model_dump.return_value = { + "execution_status": "idle", + "agent": { + "llm": {"model": "gpt-4o-mini"}, + "agent_context": { + "skills": [{"name": "big", "content": "x" * 100_000}], + "system_message_suffix": "kept", + }, + }, + } + event_service._conversation = mock_conversation_with_real_lock + received: list[Event] = [] + + class Recorder(Subscriber[Event]): + async def __call__(self, event: Event): + received.append(event) + + await event_service.subscribe_to_events(Recorder()) + + snapshot = cast(ConversationStateUpdateEvent, received[0]) + assert snapshot.key == "full_state" + assert snapshot.value["agent"]["agent_context"] == { + "skills": [], + "system_message_suffix": "kept", + } + assert snapshot.value["agent"]["llm"] == {"model": "gpt-4o-mini"} + assert snapshot.value["execution_status"] == "idle" + + @pytest.mark.asyncio + async def test_agent_update_leaves_out_skill_bodies(self, event_service): + agent = _sample_agent().model_copy( + update={ + "agent_context": AgentContext( + skills=[Skill(name="big", content="x" * 100_000)], + system_message_suffix="kept", + ) + } + ) + received: list[Event] = [] + + class Recorder(Subscriber[Event]): + async def __call__(self, event: Event): + received.append(event) + + event_service._pub_sub.subscribe(Recorder()) + update = ConversationStateUpdateEvent(key="agent", value=agent) + await event_service._publish_to_subscribers(update) + + sent = cast(ConversationStateUpdateEvent, received[0]) + assert sent.value["agent_context"]["skills"] == [] + assert sent.value["agent_context"]["system_message_suffix"] == "kept" + assert len(update.value.agent_context.skills) == 1 + @pytest.mark.asyncio async def test_concurrent_subscriptions_no_deadlock( self, event_service, mock_conversation_with_real_lock diff --git a/tests/cross/test_remote_conversation_live_server.py b/tests/cross/test_remote_conversation_live_server.py index 972fc1d765d..cb5cc072413 100644 --- a/tests/cross/test_remote_conversation_live_server.py +++ b/tests/cross/test_remote_conversation_live_server.py @@ -86,16 +86,15 @@ def _assert_secret(value: "str | SecretStr", expected: str) -> None: assert value == expected -@pytest.mark.parametrize("skill_mib", [2, 17]) -def test_attach_receives_large_initial_state(tmp_path: Path, monkeypatch, skill_mib): +@pytest.mark.parametrize("context_mib", [2, 17]) +def test_attach_receives_large_initial_state(tmp_path: Path, monkeypatch, context_mib): monkeypatch.setenv("OPENHANDS_REMOTE_WS_READY_TIMEOUT", "5") + # Skill bodies are left out of the snapshot, so grow a field that is sent. agent = Agent( llm=LLM(model="gpt-4o-mini", api_key=SecretStr("test")), tools=[], agent_context=AgentContext( - skills=[ - Skill(name="large-context", content="x" * (skill_mib * 1024 * 1024)) - ] + system_message_suffix="x" * (context_mib * 1024 * 1024) ), ) received = [] @@ -122,7 +121,7 @@ def on_event(event): response.raise_for_status() conversation_id = UUID(response.json()["id"]) try: - if skill_mib == 17: + if context_mib == 17: with pytest.raises( WebSocketConnectionError, match="message too big.*1009" ):