Joseph Doherty
104 lines
3.3 KiB
Python
104 lines
3.3 KiB
Python
"""Tests for the event-lifecycle detection service (T52).
|
|
|
|
Per Phase 3, after each narrated turn we ask a classifier whether any
|
|
active events transitioned (started, completed, cancelled). The bias is
|
|
strongly toward an empty result — most turns do NOT resolve or start a
|
|
known event, and the turn-flow caller (T61) only appends an
|
|
event_started/completed/cancelled record when this service yields one.
|
|
|
|
These tests cover:
|
|
|
|
* The classifier returning a single transition is honored end-to-end.
|
|
* An empty ``active_events`` list short-circuits before any LLM call,
|
|
so callers that hold no live events pay zero classifier cost.
|
|
* Three rounds of malformed JSON exhaust ``classify``'s retries and we
|
|
fall back to the empty default — graceful degradation per §3.3.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
|
|
import pytest
|
|
|
|
from chat.llm.mock import MockLLMClient
|
|
from chat.services.event_lifecycle import (
|
|
EventLifecycleDecision,
|
|
detect_event_transitions,
|
|
)
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_detects_one_transition_happy_path():
|
|
canned = json.dumps(
|
|
{
|
|
"transitions": [
|
|
{
|
|
"event_id": "evt_1",
|
|
"new_status": "completed",
|
|
"reason": "they arrived at the park",
|
|
}
|
|
]
|
|
}
|
|
)
|
|
mock = MockLLMClient(canned=[canned])
|
|
result = await detect_event_transitions(
|
|
mock,
|
|
classifier_model="x",
|
|
narrative_text="They walked through the park gate, finally there.",
|
|
active_events=[
|
|
{
|
|
"event_id": "evt_1",
|
|
"kind": "date_at_park",
|
|
"status": "active",
|
|
"props": {},
|
|
}
|
|
],
|
|
)
|
|
assert isinstance(result, EventLifecycleDecision)
|
|
assert len(result.transitions) == 1
|
|
assert result.transitions[0].event_id == "evt_1"
|
|
assert result.transitions[0].new_status == "completed"
|
|
assert result.transitions[0].reason == "they arrived at the park"
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_empty_active_events_short_circuits_without_classifier_call():
|
|
"""No active events -> no classifier call.
|
|
|
|
The mock has an empty canned list; any ``generate`` call would raise
|
|
``IndexError`` from ``list.pop(0)``. The test passing proves the
|
|
short-circuit holds.
|
|
"""
|
|
mock = MockLLMClient(canned=[])
|
|
result = await detect_event_transitions(
|
|
mock,
|
|
classifier_model="x",
|
|
narrative_text="Just a quiet moment between them.",
|
|
active_events=[],
|
|
)
|
|
assert isinstance(result, EventLifecycleDecision)
|
|
assert result.transitions == []
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_classifier_failure_returns_empty_default():
|
|
"""``classify`` retries 3 times; after all fail it returns the empty
|
|
default so the turn flow keeps moving (§3.3 graceful degradation)."""
|
|
mock = MockLLMClient(canned=["bad", "bad", "bad"])
|
|
result = await detect_event_transitions(
|
|
mock,
|
|
classifier_model="x",
|
|
narrative_text="Some text the classifier will choke on.",
|
|
active_events=[
|
|
{
|
|
"event_id": "evt_1",
|
|
"kind": "date_at_park",
|
|
"status": "active",
|
|
"props": {},
|
|
}
|
|
],
|
|
)
|
|
assert isinstance(result, EventLifecycleDecision)
|
|
assert result.transitions == []
|