113 lines
4.2 KiB
Python
113 lines
4.2 KiB
Python
import json
|
|
|
|
import pytest
|
|
|
|
from duck_core.events.store import EventStore
|
|
from duck_core.model_client import ModelResponse
|
|
from duck_core.approvals.service import ApprovalService
|
|
from duck_core.runtime_loop import RuntimeLoop
|
|
from duck_core.tasks.store import TaskStore
|
|
|
|
|
|
class FakeToolModelClient:
|
|
async def chat(self, role, messages):
|
|
if role == "action":
|
|
return ModelResponse(
|
|
role=role,
|
|
model="local-main",
|
|
content=json.dumps(
|
|
{
|
|
"kind": "action_directive",
|
|
"intent": "read requested file",
|
|
"risk_level": "low",
|
|
"actions": [
|
|
{
|
|
"tool": "file_read",
|
|
"args": {"path": "note.txt"},
|
|
"reason": "User asked for file contents",
|
|
}
|
|
],
|
|
}
|
|
),
|
|
reasoning_content=None,
|
|
raw={},
|
|
latency_ms=5.0,
|
|
)
|
|
assert role == "thinker"
|
|
assert any("tool_observations" in message["content"] for message in messages)
|
|
return ModelResponse(
|
|
role=role,
|
|
model="local-main",
|
|
content="The file says: hello from tool",
|
|
reasoning_content="used file_read",
|
|
raw={},
|
|
latency_ms=12.0,
|
|
)
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_runtime_executes_action_directive_tool_and_finishes_with_observation(tmp_path):
|
|
(tmp_path / "note.txt").write_text("hello from tool")
|
|
db_path = str(tmp_path / "duck.sqlite3")
|
|
task_store = TaskStore(db_path)
|
|
event_store = EventStore(db_path)
|
|
loop = RuntimeLoop(task_store, event_store, FakeToolModelClient())
|
|
|
|
result = await loop.run_chat("read note.txt", str(tmp_path), debug=True)
|
|
events = await event_store.list_events(result.task_id)
|
|
event_types = [event.event_type for event in events]
|
|
tool_finished = next(event for event in events if event.event_type == "tool_call_finished")
|
|
|
|
assert result.status == "completed"
|
|
assert result.final_response == "The file says: hello from tool"
|
|
assert "action_directive" in event_types
|
|
assert "tool_call_started" in event_types
|
|
assert tool_finished.payload["tool"] == "file_read"
|
|
assert tool_finished.payload["result"]["ok"] is True
|
|
assert tool_finished.payload["result"]["output"] == "hello from tool"
|
|
|
|
|
|
class FakeApprovalModelClient:
|
|
async def chat(self, role, messages):
|
|
if role == "action":
|
|
return ModelResponse(
|
|
role=role,
|
|
model="local-main",
|
|
content=json.dumps(
|
|
{
|
|
"kind": "action_directive",
|
|
"intent": "run command",
|
|
"risk_level": "medium",
|
|
"actions": [
|
|
{
|
|
"tool": "shell_exec_safe",
|
|
"args": {"command": "uname -a"},
|
|
"reason": "User requested system information",
|
|
}
|
|
],
|
|
}
|
|
),
|
|
reasoning_content=None,
|
|
raw={},
|
|
latency_ms=5.0,
|
|
)
|
|
raise AssertionError("thinker must not be called while approval is pending")
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_runtime_creates_pending_approval_when_tool_requires_it(tmp_path):
|
|
db_path = str(tmp_path / "duck.sqlite3")
|
|
task_store = TaskStore(db_path)
|
|
event_store = EventStore(db_path)
|
|
approvals = ApprovalService(db_path)
|
|
loop = RuntimeLoop(task_store, event_store, FakeApprovalModelClient(), approval_service=approvals)
|
|
|
|
result = await loop.run_chat("run uname", str(tmp_path), debug=True)
|
|
pending = await approvals.pending()
|
|
events = await event_store.list_events(result.task_id)
|
|
|
|
assert result.status == "waiting_for_approval"
|
|
assert pending[0].task_id == result.task_id
|
|
assert pending[0].normalized_action["tool"] == "shell_exec_safe"
|
|
assert any(event.event_type == "tool_approval_requested" for event in events)
|