new feature: External guardrails/judges

This commit is contained in:
2026-08-24 11:14:10 -03:00
parent 05373deff2
commit fd37138c4f
384 changed files with 40593 additions and 3621 deletions

View File

@@ -0,0 +1,230 @@
from __future__ import annotations
from pathlib import Path
from types import SimpleNamespace
import pytest
from agent_framework.observability.code_mapper import ObservabilityCodeMapper
from agent_framework.observability.observer import AgentObserver
from agent_framework.observability.telemetry import Telemetry
from app.observability.telemetry_observer import TelemetryBackedAgentObserver
class CaptureAnalytics:
def __init__(self):
self.calls = []
async def publish(self, event_type, payload):
self.calls.append((event_type, payload))
class CaptureBus:
def __init__(self):
self.calls = []
async def publish(self, event_type, payload, **kwargs):
self.calls.append((event_type, payload, kwargs))
def test_mapper_passthrough_and_contract_fields(tmp_path: Path):
mapper = ObservabilityCodeMapper({"GRL.TOXOUT": "GRL.004"})
code, payload, metadata = mapper.normalize_payload("GRL.TOXOUT", {"x": 1}, {})
assert code == "GRL.004"
assert payload["event_code_internal"] == "GRL.TOXOUT"
assert metadata["event_code_internal"] == "GRL.TOXOUT"
assert metadata["event_code_mapped"] == "GRL.004"
assert mapper.map("GRL.NEW") == "GRL.NEW"
@pytest.mark.asyncio
async def test_agent_observer_maps_before_analytics_and_event_bus():
analytics = CaptureAnalytics()
bus = CaptureBus()
observer = AgentObserver(
analytics=analytics,
event_bus=bus,
code_mapper=ObservabilityCodeMapper({"GRL.TOXOUT": "GRL.004"}),
)
event = await observer.emit("GRL.TOXOUT", {"reason": "tox"})
assert event["eventType"] == "GRL.004"
assert event["payload"]["tag"] == "GRL.004"
assert event["payload"]["event_code_internal"] == "GRL.TOXOUT"
assert analytics.calls[0][0] == "GRL.004"
assert bus.calls[0][0] == "GRL.004"
@pytest.mark.asyncio
async def test_telemetry_maps_direct_events(tmp_path: Path):
mapping = tmp_path / "mapping.yaml"
mapping.write_text("mappings:\n GRL.TOXOUT: GRL.004\n", encoding="utf-8")
settings = SimpleNamespace(
OBSERVABILITY_CODE_MAPPING_ENABLED=True,
OBSERVABILITY_CODE_MAPPING_PATH=str(mapping),
ENABLE_LANGFUSE=False,
ENABLE_OCI_STREAMING=False,
ENABLE_OTEL=False,
LANGFUSE_TRACE_MODE="verbose",
)
telemetry = Telemetry(settings)
capture = []
async def subscriber(evt):
capture.append(evt)
telemetry.event_bus.subscribe(subscriber)
await telemetry.event("GRL.TOXOUT", {"a": 1}, kind="grl")
assert capture
assert capture[0].name == "GRL.004"
assert capture[0].payload["event_code_internal"] == "GRL.TOXOUT"
@pytest.mark.asyncio
async def test_contas_telemetry_observer_returns_mapped_envelope(tmp_path: Path):
mapping = tmp_path / "mapping.yaml"
mapping.write_text("mappings:\n GRL.TOXOUT: GRL.004\n", encoding="utf-8")
settings = SimpleNamespace(
OBSERVABILITY_CODE_MAPPING_ENABLED=True,
OBSERVABILITY_CODE_MAPPING_PATH=str(mapping),
ENABLE_LANGFUSE=False,
ENABLE_OCI_STREAMING=False,
ENABLE_OTEL=False,
LANGFUSE_TRACE_MODE="verbose",
)
telemetry = Telemetry(settings)
observer = TelemetryBackedAgentObserver(telemetry)
event = await observer.emit("GRL.TOXOUT", {"rail": "TOXOUT"})
assert event["eventType"] == "GRL.004"
assert event["body"]["tag"] == "GRL.004"
assert event["metadata"]["event_code_internal"] == "GRL.TOXOUT"
class _FakeObservation:
def __init__(self):
self.updates = []
def update(self, **kwargs):
self.updates.append(kwargs)
class _FakeObservationCM:
def __init__(self, observation):
self.observation = observation
def __enter__(self):
return self.observation
def __exit__(self, exc_type, exc, tb):
return False
class _FakeLangfuse:
def __init__(self):
self.started = []
def start_as_current_observation(self, **kwargs):
self.started.append(dict(kwargs))
return _FakeObservationCM(_FakeObservation())
@pytest.mark.asyncio
async def test_generation_observation_name_is_mapped_before_langfuse_and_event_bus(tmp_path: Path):
mapping = tmp_path / "mapping.yaml"
mapping.write_text(
"mappings:\n guardrail.dlex_in: GRL.004\n guardrail.tox: GRL.005\n",
encoding="utf-8",
)
settings = SimpleNamespace(
OBSERVABILITY_CODE_MAPPING_ENABLED=True,
OBSERVABILITY_CODE_MAPPING_PATH=str(mapping),
ENABLE_LANGFUSE=False,
ENABLE_OCI_STREAMING=False,
ENABLE_OTEL=False,
LANGFUSE_TRACE_MODE="verbose",
)
telemetry = Telemetry(settings)
fake = _FakeLangfuse()
telemetry.langfuse = fake
telemetry.enabled = True
capture = []
async def subscriber(evt):
capture.append(evt)
telemetry.event_bus.subscribe(subscriber)
async with telemetry.generation_span(
"guardrail.dlex_in",
"model-x",
"input",
metadata={"component": "guardrail.dlex_in"},
) as generation:
generation.set_output("ok")
assert fake.started[0]["name"] == "GRL.004"
assert fake.started[0]["metadata"]["observability_name_internal"] == "guardrail.dlex_in"
assert fake.started[0]["metadata"]["observability_name_mapped"] == "GRL.004"
assert capture[0].name == "GRL.004"
assert capture[0].payload["observability_name_internal"] == "guardrail.dlex_in"
@pytest.mark.asyncio
async def test_span_name_is_mapped_before_otel_event_bus_and_langfuse(tmp_path: Path):
mapping = tmp_path / "mapping.yaml"
mapping.write_text("mappings:\n guardrail.tox: GRL.005\n", encoding="utf-8")
settings = SimpleNamespace(
OBSERVABILITY_CODE_MAPPING_ENABLED=True,
OBSERVABILITY_CODE_MAPPING_PATH=str(mapping),
ENABLE_LANGFUSE=False,
ENABLE_OCI_STREAMING=False,
ENABLE_OTEL=False,
LANGFUSE_TRACE_MODE="verbose",
)
telemetry = Telemetry(settings)
fake = _FakeLangfuse()
telemetry.langfuse = fake
telemetry.enabled = True
capture = []
async def subscriber(evt):
capture.append(evt)
telemetry.event_bus.subscribe(subscriber)
async with telemetry.span("guardrail.tox", input={"text": "x"}):
pass
assert fake.started[0]["name"] == "GRL.005"
assert fake.started[0]["metadata"]["observability_name_internal"] == "guardrail.tox"
assert capture[0].name == "GRL.005.started"
assert capture[-1].name == "GRL.005.completed"
@pytest.mark.asyncio
async def test_langfuse_analytics_direct_sdk_path_also_maps_observation_name(tmp_path: Path):
from agent_framework.analytics.providers.langfuse import LangfuseAnalyticsPublisher
mapping = tmp_path / "mapping.yaml"
mapping.write_text("mappings:\n guardrail.dlex_in: GRL.004\n", encoding="utf-8")
settings = SimpleNamespace(
OBSERVABILITY_CODE_MAPPING_ENABLED=True,
OBSERVABILITY_CODE_MAPPING_PATH=str(mapping),
LANGFUSE_PUBLIC_KEY="pk",
LANGFUSE_SECRET_KEY="sk",
LANGFUSE_HOST="http://localhost",
)
fake = _FakeLangfuse()
publisher = LangfuseAnalyticsPublisher(settings=settings, langfuse=fake)
# Give the technical event correlation so the provider does not intentionally skip it.
from agent_framework.observability.context import set_observability_context, clear_observability_context
clear_observability_context()
set_observability_context(request_id="req-123", trace_id="req-123")
try:
await publisher.publish("guardrail.dlex_in", {"metadata": {"request_id": "req-123"}})
finally:
clear_observability_context()
assert fake.started, "Langfuse analytics provider did not create an observation"
assert fake.started[0]["name"] == "GRL.004"
assert fake.started[0]["metadata"]["observability_name_internal"] == "guardrail.dlex_in"
assert fake.started[0]["metadata"]["observability_name_mapped"] == "GRL.004"