Add tests/test_api.py
Browse files- tests/test_api.py +62 -0
tests/test_api.py
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import pytest
|
| 2 |
+
from fastapi.testclient import TestClient
|
| 3 |
+
from app import app
|
| 4 |
+
from codereview_env.models import TaskId, ActionType, Category, Severity, Verdict
|
| 5 |
+
|
| 6 |
+
def test_api_health():
|
| 7 |
+
client = TestClient(app)
|
| 8 |
+
response = client.get("/health")
|
| 9 |
+
assert response.status_code == 200
|
| 10 |
+
assert response.json()["status"] == "ok"
|
| 11 |
+
assert response.json()["env_ready"] is True
|
| 12 |
+
|
| 13 |
+
def test_api_workflow():
|
| 14 |
+
client = TestClient(app)
|
| 15 |
+
|
| 16 |
+
# 1. Reset
|
| 17 |
+
reset_resp = client.post("/reset", json={"task_id": "bug_detection", "seed": 1})
|
| 18 |
+
assert reset_resp.status_code == 200
|
| 19 |
+
data = reset_resp.json()
|
| 20 |
+
episode_id = data["episode_id"]
|
| 21 |
+
assert "observation" in data["result"]
|
| 22 |
+
|
| 23 |
+
# 2. Step
|
| 24 |
+
action = {
|
| 25 |
+
"action_type": "comment",
|
| 26 |
+
"body": "Starting review",
|
| 27 |
+
}
|
| 28 |
+
step_resp = client.post(f"/step/{episode_id}", json=action)
|
| 29 |
+
assert step_resp.status_code == 200
|
| 30 |
+
assert "observation" in step_resp.json()
|
| 31 |
+
|
| 32 |
+
# 3. Get Result
|
| 33 |
+
result_resp = client.get(f"/result/{episode_id}")
|
| 34 |
+
assert result_resp.status_code == 200
|
| 35 |
+
assert result_resp.json()["final_score"] >= 0
|
| 36 |
+
|
| 37 |
+
def test_api_leaderboard():
|
| 38 |
+
client = TestClient(app)
|
| 39 |
+
# Submit a score
|
| 40 |
+
sub = {
|
| 41 |
+
"agent_name": "test_agent",
|
| 42 |
+
"task_id": "bug_detection",
|
| 43 |
+
"score": 0.95,
|
| 44 |
+
"seed": 42
|
| 45 |
+
}
|
| 46 |
+
resp = client.post("/submit", json=sub)
|
| 47 |
+
assert resp.status_code == 200
|
| 48 |
+
assert resp.json()["status"] == "submitted"
|
| 49 |
+
|
| 50 |
+
# Check leaderboard
|
| 51 |
+
lb_resp = client.get("/leaderboard")
|
| 52 |
+
assert lb_resp.status_code == 200
|
| 53 |
+
assert len(lb_resp.json()["bug_detection"]) > 0
|
| 54 |
+
assert lb_resp.json()["bug_detection"][0]["agent_name"] == "test_agent"
|
| 55 |
+
|
| 56 |
+
def test_api_invalid_episode():
|
| 57 |
+
client = TestClient(app)
|
| 58 |
+
response = client.post("/step/nonexistent-id", json={
|
| 59 |
+
"action_type": "comment",
|
| 60 |
+
"body": "hello"
|
| 61 |
+
})
|
| 62 |
+
assert response.status_code == 404
|