brain-of-reese/tests/unit/test_agent_short_lived_sessions.py

"""Unit: short-lived DB sessions in the agent loop (SEC-14-04).

Verifies that ``run_agent`` uses a session factory to create a new
short-lived session for each DB operation (tool call), closes it after
the tool result is produced, and completes the agent loop correctly.
"""
from __future__ import annotations

import asyncio
import uuid
from collections.abc import AsyncIterator, Callable
from copy import deepcopy
from datetime import UTC, datetime
from typing import Any, cast
from unittest.mock import MagicMock, patch

from sqlalchemy.orm import Session

from app.config import Settings
from app.models import Document
from app.rag import agent
from app.rag.agent import AgentHolder, AgentPiece, run_agent
from app.rag.llm import LLMClient, StreamPiece, ToolCallPiece

#: Fixture document creation date (phase 106, D5).
_FIXTURE_CREATED_AT = datetime(2024, 6, 15, 12, 0, 0, tzinfo=UTC)


def _settings(**kwargs: Any) -> Settings:
    kwargs.setdefault("_env_file", None)
    return Settings(**kwargs)  # pyright: ignore[reportCallIssue]


def _doc(source: str, path: str, title: str = "Title", content: str = "CONTENT") -> Document:
    return Document(
        id=uuid.uuid4(),
        source=source,
        path=path,
        full_path=f"/tmp/{path}",
        title=title,
        content=content,
        content_hash="0" * 64,
        created_at=_FIXTURE_CREATED_AT,
    )


class TrackingSession:
    """A mock Session that tracks ``close()`` calls and is usable as a
    context manager."""

    def __init__(self, close_count: list[int] | None = None) -> None:
        self.close_count = close_count if close_count is not None else [0]
        self.scalar = MagicMock(return_value=None)
        # scalars() returns a ScalarResult-like object with .all()
        self._scalar_result = MagicMock()
        self._scalar_result.all = MagicMock(return_value=[])
        self.scalars = MagicMock(return_value=self._scalar_result)
        self.execute = MagicMock(return_value=MagicMock(scalars=MagicMock(return_value=[])))
        self.add = MagicMock()

    def __enter__(self) -> TrackingSession:
        return self

    def __exit__(self, *args: Any) -> None:
        self.close_count[0] += 1


class TrackingFactory:
    """A session factory that creates a ``TrackingSession`` each time
    it is called, so the tests can verify that a new session is created
    for each tool call and that it is closed afterwards."""

    def __init__(self) -> None:
        self.sessions: list[TrackingSession] = []

    def __call__(self) -> TrackingSession:
        session = TrackingSession()
        self.sessions.append(session)
        return session


# Type alias for pyright: TrackingFactory is callable that returns Session
TrackingFactoryCallable: type[TrackingFactory] = TrackingFactory  # noqa: N816


class ScriptedLLM:
    """Canned stream sequences for agent-loop tests."""

    def __init__(self, *streams: list[StreamPiece | ToolCallPiece]) -> None:
        self.streams: list[list[StreamPiece | ToolCallPiece]] = list(streams)
        self.requests: list[tuple[list[dict[str, Any]], list[dict[str, Any]] | None]] = []

    async def chat_stream(
        self,
        messages: list[dict[str, str]],
        tools: list[dict[str, Any]] | None = None,
        scaffolding: Any = None,
    ) -> AsyncIterator[StreamPiece | ToolCallPiece]:
        self.requests.append((deepcopy(messages), tools))
        if not self.streams:
            raise AssertionError("ScriptedLLM ran out of canned streams")
        pieces = self.streams.pop(0)
        for piece in pieces:
            yield piece


class ScriptedToolLLM(ScriptedLLM):
    """A scripted LLM that returns exactly one tool call, then an
    empty clean answer on the next round."""

    def __init__(
        self,
        tool_call: ToolCallPiece,
        answer: str = "ANSWER",
    ) -> None:
        super().__init__(
            [tool_call],  # round 1: tool call
            [StreamPiece("content", answer)],  # round 2: answer
        )


async def _consume(
    llm: LLMClient,
    db_factory: Callable[[], Session],
    holder: AgentHolder,
    settings: Settings,
    seed_docs: list[Document] | None = None,
) -> list[AgentPiece]:
    """Consume one ``run_agent`` turn."""
    out: list[AgentPiece] = []
    async for piece in run_agent(
        cast("LLMClient", llm),
        db_factory,
        system_prompt="SYSTEM_PROMPT",
        user_message="QUESTION",
        seed_docs=seed_docs or [],
        settings=settings,
        holder=holder,
    ):
        out.append(piece)
    return out


# ---------- db_factory is called per DB operation ----------


def test_db_factory_called_once_for_answer_no_tools() -> None:
    """When the model answers without calling any tools, the agent
    loop makes no DB calls — but ``run_agent`` still accepts the
    factory (it is simply not invoked)."""
    llm = ScriptedLLM([StreamPiece("content", "DIRECT ANSWER")])
    factory = TrackingFactory()
    holder = AgentHolder()
    settings = _settings(agent_max_rounds=10)

    asyncio.run(
        _consume(cast("LLMClient", llm), cast("Callable[[], Session]", factory), holder, settings)
    )

    # No tool calls means no DB operations — factory never invoked
    assert len(factory.sessions) == 0
    assert holder.tool_calls == 0


# ---------- each tool call creates a new session ----------


def test_each_tool_call_creates_a_new_session() -> None:
    """Each tool call the model emits creates its own short-lived
    session via the factory; sessions are closed after the tool
    result is produced."""
    tool_call = ToolCallPiece(id="call_1", name="ls", arguments={})
    llm = ScriptedToolLLM(tool_call)
    factory = TrackingFactory()
    holder = AgentHolder()
    settings = _settings(agent_max_rounds=10)

    asyncio.run(
        _consume(cast("LLMClient", llm), cast("Callable[[], Session]", factory), holder, settings)
    )

    # One tool call → one session created
    assert len(factory.sessions) == 1
    # The session was closed after the tool result
    assert factory.sessions[0].close_count[0] == 1
    # The holder records the executed call
    assert holder.tool_calls == 1


def test_multiple_tool_calls_create_separate_sessions() -> None:
    """When the model makes multiple tool calls across rounds, each
    round creates a new session that is closed after the result."""
    tool_call_1 = ToolCallPiece(id="call_1", name="ls", arguments={})
    tool_call_2 = ToolCallPiece(id="call_2", name="ls", arguments={})
    llm = ScriptedToolLLM(tool_call_1)
    # Override the second round to also emit a tool call
    llm.streams = [
        [tool_call_1],  # round 1: ls
        [tool_call_2],  # round 2: ls (another listing)
        [StreamPiece("content", "FINAL ANSWER")],  # round 3: answer
    ]
    factory = TrackingFactory()
    holder = AgentHolder()
    settings = _settings(agent_max_rounds=10)

    asyncio.run(
        _consume(cast("LLMClient", llm), cast("Callable[[], Session]", factory), holder, settings)
    )

    # Two tool calls → two sessions created and closed
    assert len(factory.sessions) == 2
    for session in factory.sessions:
        assert session.close_count[0] == 1
    assert holder.tool_calls == 2


# ---------- sessions are closed after use (no pinning) ----------


def test_sessions_closed_after_tool_result() -> None:
    """Verify that the session is closed AFTER the tool result is
    produced but BEFORE the next model round — no session is held
    across rounds."""
    tool_call = ToolCallPiece(id="call_1", name="ls", arguments={})
    llm = ScriptedToolLLM(tool_call)
    factory = TrackingFactory()
    holder = AgentHolder()
    settings = _settings(agent_max_rounds=10)

    asyncio.run(
        _consume(cast("LLMClient", llm), cast("Callable[[], Session]", factory), holder, settings)
    )

    # The session was closed (close_count incremented)
    assert factory.sessions[0].close_count[0] == 1
    # Only one session was created (not reused across rounds)
    assert len(factory.sessions) == 1


# ---------- deflected path works without DB factory usage ----------


def test_deflected_path_no_db_factory_calls() -> None:
    """A deflected turn (LOW mode) does not run the agent loop, so
    the session factory is never invoked."""
    llm = ScriptedLLM([StreamPiece("content", "I don't know about that.")])
    factory = TrackingFactory()
    holder = AgentHolder()
    # agent_max_rounds=0 disables tools → single tools=None request
    settings = _settings(agent_max_rounds=0)

    asyncio.run(
        _consume(cast("LLMClient", llm), cast("Callable[[], Session]", factory), holder, settings)
    )

    assert len(factory.sessions) == 0
    assert holder.tool_calls == 0


# ---------- agent loop completes correctly ----------


def test_agent_loop_completes_with_mock_factory() -> None:
    """The agent loop completes correctly with a mock session factory:
    tool calls are executed, the answer is streamed, and the holder
    records the correct state."""
    tool_call = ToolCallPiece(id="call_1", name="grep", arguments={"pattern": "hello"})
    llm = ScriptedToolLLM(tool_call)
    factory = TrackingFactory()
    holder = AgentHolder()
    settings = _settings(agent_max_rounds=10)

    pieces = asyncio.run(
        _consume(cast("LLMClient", llm), cast("Callable[[], Session]", factory), holder, settings)
    )

    # The stream contains the tool call piece and the answer piece
    types = [getattr(p, "kind", "tool") for p in pieces]
    assert "tool" in types
    assert "content" in types

    # The holder records one executed call
    assert holder.tool_calls == 1
    assert len(factory.sessions) == 1


# ---------- monkeypatched DB accessors work with factory ----------


def test_monkeypatched_accessors_with_factory() -> None:
    """DB accessors that are monkeypatched (as in the existing unit
    test suite) work correctly when the agent loop calls them through
    a session factory."""
    # Patch ls_top to return a canned result
    canned_ls = [("TestSource", 5, None)]

    def fake_ls_top(db: Session) -> list[tuple[str, int, str | None]]:  # type: ignore[return-value]
        return list(canned_ls)

    with (
        patch.object(agent, "ls_top", fake_ls_top),
        patch.object(agent, "list_source_names", return_value=["TestSource"]),
    ):
            tool_call = ToolCallPiece(id="call_1", name="ls", arguments={})
            llm = ScriptedToolLLM(tool_call)
            factory = TrackingFactory()
            holder = AgentHolder()
            settings = _settings(agent_max_rounds=10)

            asyncio.run(
        _consume(cast("LLMClient", llm), cast("Callable[[], Session]", factory), holder, settings)
    )

            assert holder.tool_calls == 1
            assert len(factory.sessions) == 1
            # The factory was called exactly once for this tool round
            assert factory.sessions[0].close_count[0] == 1