Compare commits

...
Author SHA1 Message Date
b375d7cf09 fix(server): wire WS event bridge to the bus channels actually publish to (#673)
* fix(server): wire WS event bridge to the bus channels actually publish to

include_all_routes() built the /v1/agents/events WebSocket router from
get_event_bus() — a global singleton nothing in `jarvis serve` publishes
to. The real bus lives at app.state.bus (set in server/app.py and handed
to every channel/agent). Events published on it silently never reached
any connected browser client, since the WS router was subscribed to a
different, disconnected EventBus instance entirely.

* fix(server): keep managed agent events on app bus

---------

Co-authored-by: Ari <ari.silva@paipe.co>
Co-authored-by: Elliot Slusky <elliot@slusky.com>
2026-08-14 18:18:11 -07:00
4 changed files with 160 additions and 23 deletions
+31 -21
View File
@@ -34,6 +34,15 @@ _MEMORY_BACKEND_LOCK_SETUP = threading.Lock()
_MCP_LOCK_SETUP = threading.Lock()
def _get_runtime_event_bus(runtime: Any = None) -> Any:
"""Return the server-owned event bus, falling back outside app runtimes."""
from openjarvis.core.events import get_event_bus
bus = getattr(runtime, "bus", None)
return bus if bus is not None else get_event_bus()
def _start_managed_worker(app_state: Any, target: Any, *, name: str) -> Any:
"""Start and track a managed-agent worker for orderly app shutdown."""
@@ -301,14 +310,13 @@ def _make_lightweight_system(
# Wrap with InstrumentedEngine so agent ticks are recorded
# in telemetry (FLOPs, energy, cost savings).
try:
from openjarvis.core.events import get_event_bus
from openjarvis.telemetry.instrumented_engine import (
InstrumentedEngine,
)
plain_engine = InstrumentedEngine(
plain_engine,
get_event_bus(),
_get_runtime_event_bus(runtime),
)
except Exception:
pass # telemetry is optional
@@ -1644,26 +1652,27 @@ def create_agent_manager_router(
# Re-use the server's engine + model so we don't pick a
# random model from Ollama's list.
server_engine = getattr(request.app.state, "engine", None)
server_model = getattr(request.app.state, "model", "")
server_config = getattr(request.app.state, "config", None)
app_state = request.app.state
server_engine = getattr(app_state, "engine", None)
server_model = getattr(app_state, "model", "")
server_config = getattr(app_state, "config", None)
server_bus = _get_runtime_event_bus(app_state)
def _run_tick():
try:
from openjarvis.agents.executor import AgentExecutor
from openjarvis.core.events import get_event_bus
_ts = getattr(request.app.state, "trace_store", None)
_ts = getattr(app_state, "trace_store", None)
executor = AgentExecutor(
manager=manager,
event_bus=get_event_bus(),
event_bus=server_bus,
trace_store=_ts,
)
system = _make_lightweight_system(
server_engine,
server_model,
server_config,
request.app.state,
app_state,
)
executor.set_system(system)
# The route handler above already called start_tick() to
@@ -1690,7 +1699,7 @@ def create_agent_manager_router(
try:
_start_managed_worker(
request.app.state,
app_state,
_run_tick,
name=f"managed-agent-run-{agent_id}",
)
@@ -1997,11 +2006,12 @@ def create_agent_manager_router(
import time as _time
from openjarvis.agents.executor import AgentExecutor
from openjarvis.core.events import get_event_bus
_srv_engine = getattr(request.app.state, "engine", None)
_srv_model = getattr(request.app.state, "model", "")
_srv_config = getattr(request.app.state, "config", None)
_app_state = request.app.state
_srv_engine = getattr(_app_state, "engine", None)
_srv_model = getattr(_app_state, "model", "")
_srv_config = getattr(_app_state, "config", None)
_srv_bus = _get_runtime_event_bus(_app_state)
def _immediate_tick():
_start = _time.time()
@@ -2011,17 +2021,17 @@ def create_agent_manager_router(
_srv_model,
)
try:
_ts2 = getattr(request.app.state, "trace_store", None)
_ts2 = getattr(_app_state, "trace_store", None)
executor = AgentExecutor(
manager=manager,
event_bus=get_event_bus(),
event_bus=_srv_bus,
trace_store=_ts2,
)
system = _make_lightweight_system(
_srv_engine,
_srv_model,
_srv_config,
request.app.state,
_app_state,
)
executor.set_system(system)
logger.info(
@@ -2055,7 +2065,7 @@ def create_agent_manager_router(
try:
_start_managed_worker(
request.app.state,
_app_state,
_immediate_tick,
name=f"managed-agent-immediate-{agent_id}",
)
@@ -2106,12 +2116,12 @@ def create_agent_manager_router(
return {"learning_log": manager.list_learning_log(agent_id)}
@agents_router.post("/{agent_id}/learning/run")
def trigger_learning(agent_id: str):
def trigger_learning(agent_id: str, request: Request):
if not manager.get_agent(agent_id):
raise HTTPException(status_code=404, detail="Agent not found")
from openjarvis.core.events import EventType, get_event_bus
from openjarvis.core.events import EventType
bus = get_event_bus()
bus = _get_runtime_event_bus(request.app.state)
bus.publish(EventType.AGENT_LEARNING_STARTED, {"agent_id": agent_id})
return {"status": "triggered"}
+6 -2
View File
@@ -1087,12 +1087,16 @@ def include_all_routes(app) -> None:
except ImportError:
pass
# WebSocket bridge for real-time agent events
# WebSocket bridge for real-time agent events. Must subscribe on the
# same EventBus instance channels/agents actually publish to
# (app.state.bus, set in server/app.py) — the get_event_bus() global
# singleton is a *different* bus that nothing in `jarvis serve` ever
# publishes to, so events silently never reached this endpoint.
try:
from openjarvis.core.events import get_event_bus
from openjarvis.server.ws_bridge import create_ws_router
ws_router = create_ws_router(get_event_bus())
ws_router = create_ws_router(getattr(app.state, "bus", None) or get_event_bus())
app.include_router(ws_router)
except Exception:
logger.debug("WebSocket bridge not available", exc_info=True)
+33
View File
@@ -605,6 +605,39 @@ class TestLightweightSystemEngineResolution:
)
assert captured["key"] == "llamacpp"
def test_instrumented_engine_uses_runtime_event_bus(self, monkeypatch):
pytest.importorskip("fastapi")
from openjarvis.core.events import EventBus
from openjarvis.server import agent_manager_routes as amr
from openjarvis.telemetry import instrumented_engine
resolved_engine = MagicMock()
wrapped_engine = MagicMock()
runtime_bus = EventBus()
runtime = SimpleNamespace(
bus=runtime_bus,
memory_backend=object(),
channel_backend=None,
channel_bridge=None,
knowledge_db_path=None,
)
monkeypatch.setattr(
"openjarvis.engine._discovery.get_engine",
MagicMock(return_value=("resolved", resolved_engine)),
)
instrumented = MagicMock(return_value=wrapped_engine)
monkeypatch.setattr(instrumented_engine, "InstrumentedEngine", instrumented)
system = amr._make_lightweight_system(
engine=MagicMock(),
model="m",
config=self._cfg("vllm", "ollama"),
runtime=runtime,
)
instrumented.assert_called_once_with(resolved_engine, runtime_bus)
assert system.engine is wrapped_engine
def test_caches_tool_memory_backend_when_prompt_context_is_disabled(
self,
monkeypatch,
+90
View File
@@ -3,8 +3,10 @@
from __future__ import annotations
import asyncio
import threading
import time
from types import SimpleNamespace
from unittest.mock import MagicMock, patch
import pytest
@@ -161,3 +163,91 @@ class TestWSBridge:
]
asyncio.run(exercise())
class TestIncludeAllRoutesBusWiring:
"""Regression: the WS endpoint must subscribe on the same EventBus that
channels/agents actually publish to (app.state.bus), not the unrelated
get_event_bus() global singleton — publishing on the latter used to
silently never reach any connected browser client."""
def test_uses_app_state_bus_not_global_singleton(self):
from openjarvis.core.events import reset_event_bus
from openjarvis.server.api_routes import include_all_routes
reset_event_bus() # isolate from other tests' global singleton state
app = FastAPI()
real_bus = EventBus()
app.state.bus = real_bus
include_all_routes(app)
client = TestClient(app)
with client.websocket_connect("/v1/agents/events") as ws:
real_bus.publish(EventType.AGENT_TICK_START, {"agent_id": "test-123"})
time.sleep(0.05)
data = ws.receive_json()
assert data["data"]["agent_id"] == "test-123"
@pytest.mark.parametrize(
("path", "payload"),
[
("/v1/managed-agents/test-123/run", None),
(
"/v1/managed-agents/test-123/messages",
{"content": "run now", "mode": "immediate", "stream": False},
),
],
)
def test_managed_agent_run_paths_publish_to_app_bus(self, path, payload):
from openjarvis.agents.executor import AgentExecutor
from openjarvis.core.events import reset_event_bus
from openjarvis.server.api_routes import include_all_routes
reset_event_bus()
app_bus = EventBus()
manager = MagicMock()
manager.get_agent.return_value = {
"id": "test-123",
"name": "test",
"status": "idle",
"config": {},
}
manager.send_message.return_value = {
"id": "message-123",
"agent_id": "test-123",
"content": "run now",
"mode": "immediate",
}
app = FastAPI()
app.state.bus = app_bus
app.state.agent_manager = manager
include_all_routes(app)
executed = threading.Event()
observed_buses = []
def publish_tick(executor, agent_id, **_kwargs):
observed_buses.append(executor._bus)
executor._bus.publish(EventType.AGENT_TICK_START, {"agent_id": agent_id})
executed.set()
client = TestClient(app)
with (
patch.object(AgentExecutor, "execute_tick", publish_tick),
patch(
"openjarvis.server.agent_manager_routes._make_lightweight_system",
return_value=MagicMock(),
) as make_system,
client.websocket_connect("/v1/agents/events?agent_id=test-123") as ws,
):
request_kwargs = {"json": payload} if payload is not None else {}
response = client.post(path, **request_kwargs)
assert response.status_code == 200
assert executed.wait(timeout=1)
assert observed_buses == [app_bus]
data = ws.receive_json()
assert data["type"] == "agent_tick_start"
assert data["data"]["agent_id"] == "test-123"
make_system.assert_called_once()