portfolio-optimiser/tests/test_mcp_run_loadbearing.py

"""Load-bearing: configured MCP servers become tools the AGENTS can call during a run, they are
opened and closed around the debate, and nothing is contacted that was not announced (Trekk B2/B3).

Krav 3 is only met if the external service is reachable *while the run works*. Three detach points,
each RED on its own:

* drop the MCP tools from ``debate_tools`` -> the agents never get the tool (and on the bundle path
  they get NO tools at all, which is what that path had before);
* skip the ``AsyncExitStack`` entry -> the tools are constructed but never connected, so they are
  present and useless — the failure mode that looks like success;
* let a dry run enter them -> ``--live-dry-run`` would contact a third party while claiming to stop
  before the first call.

The control asserts an un-configured run's tool list is unchanged, so no assertion above can pass
on something the run does anyway.
"""

from __future__ import annotations

from pathlib import Path
from typing import Any

import pytest

from portfolio_optimiser import run as run_module
from portfolio_optimiser.mcp_tools import McpServerConfig
from portfolio_optimiser.run import run_project
from portfolio_optimiser.simulation import ScriptedChatClient
from portfolio_optimiser.verdicts import VerdictStore

BUNDLE_DIR = Path(__file__).resolve().parents[1] / "shared" / "examples" / "bygg-energi-mikro"
_VERDICT_INPUT = {"decision": "approved", "rationale": "expert reviewed (sim)"}
_REPLY = (
    '{"measure":"LED-retrofit av kontorbelysning","affected_items":'
    '[{"code":"ENERGI-TOTAL-EL","quantity":300000,"unit_cost":1.0}],"claimed_saving_nok":30000}'
)

_SERVER = McpServerConfig(
    name="prisregister",
    transport="http",
    url="https://intern.example/mcp",
    allowed_tools=("lookup_unit_price",),
    timeout_seconds=15,
)


class _FakeMcpTool:
    """Stands in for a MAF ``MCPTool``: an async context manager that RECORDS its lifecycle.

    A real one would open a connection, which is exactly what a test must not do — and the thing
    worth asserting is the lifecycle itself, not the protocol MAF already owns.
    """

    def __init__(self, name: str) -> None:
        self.name = name
        self.entered = 0
        self.exited = 0
        self.entered_before_debate: bool | None = None

    async def __aenter__(self) -> _FakeMcpTool:
        self.entered += 1
        return self

    async def __aexit__(self, *exc: object) -> None:
        self.exited += 1


@pytest.fixture()
def fake_tools(monkeypatch: pytest.MonkeyPatch) -> list[_FakeMcpTool]:
    """Replace the MAF client construction with recording doubles — the config still travels the
    real path, only the socket-opening object is swapped."""
    built: list[_FakeMcpTool] = []

    def _build(configs: tuple[McpServerConfig, ...]) -> list[Any]:
        # Return ONLY this call's tools. Accumulating across calls handed the second project both
        # projects' tools under one name, and MAF refused it ("Duplicate tool name") — a defect in
        # the double, but a useful reminder that each run builds its own clients.
        fresh = [_FakeMcpTool(c.name) for c in configs]
        built.extend(fresh)
        return list(fresh)

    monkeypatch.setattr(run_module, "build_mcp_tools", _build)
    return built


@pytest.fixture()
def captured_tools(monkeypatch: pytest.MonkeyPatch) -> list[list[Any]]:
    """Capture what actually reaches the debate's ``tools=`` — the agents' real surface."""
    seen: list[list[Any]] = []
    original = run_module.fresh_workflow

    def _spy(*args: Any, **kwargs: Any) -> Any:
        seen.append(list(kwargs.get("tools") or []))
        return original(*args, **kwargs)

    monkeypatch.setattr(run_module, "fresh_workflow", _spy)
    return seen


def _factory(_role: str) -> ScriptedChatClient:
    return ScriptedChatClient(reply=_REPLY)


async def _run(**kwargs: Any):
    return await run_project(
        "BYGG-KONTOR-NORD",
        "local",
        docs_dir=str(BUNDLE_DIR),
        bundle_dir=str(BUNDLE_DIR),
        verdict_input=_VERDICT_INPUT,
        store=VerdictStore(verdicts=[]),
        client_factory=_factory,
        **kwargs,
    )


async def test_configured_server_becomes_a_tool_the_agents_have(fake_tools, captured_tools) -> None:
    """On the bundle path the agents had NO tools at all; a configured server is the first one."""
    await _run(mcp_servers=(_SERVER,))
    assert captured_tools, "the debate was never built"
    assert any(isinstance(t, _FakeMcpTool) for t in captured_tools[0])


async def test_tools_are_opened_and_closed_around_the_debate(fake_tools) -> None:
    """Constructed is not connected. An ``MCPTool`` must be entered to expose anything, and exited
    or the process hangs — so both halves are asserted, not just the first."""
    await _run(mcp_servers=(_SERVER,))
    assert fake_tools, "no MCP tool was built"
    assert fake_tools[0].entered == 1
    assert fake_tools[0].exited == 1


async def test_dry_run_never_contacts_a_configured_server(fake_tools) -> None:
    """``--live-dry-run`` stops before the first model call, and that promise has to cover egress
    too: a dry run that opened a connection to a third party would be lying by omission."""
    await _run(mcp_servers=(_SERVER,), live_dry_run=True)
    assert all(t.entered == 0 for t in fake_tools)


async def test_run_without_mcp_servers_keeps_the_tool_list_unchanged(captured_tools) -> None:
    """CONTROL: with nothing configured the bundle path still hands the agents no tools — the
    pre-Trekk-B behaviour, so every assertion above rests on the configuration and not on the run."""
    await _run()
    assert captured_tools[0] == []


async def test_portfolio_mode_gives_every_project_the_configured_tools(
    fake_tools, captured_tools
) -> None:
    """A configured server reaches EVERY project in a portfolio pass. Without the threading the
    flag would be accepted and silently dropped in one of the two modes — the defect class this
    repo refuses by name ('refused, never ignored')."""
    result = await run_module.run_portfolio(
        ["FV42-GSV-E1", "RV13-RAS-TP"],
        "local",
        store=VerdictStore(verdicts=[]),
        client_factory=_factory,
        mcp_servers=(_SERVER,),
    )
    assert len(result.runs) == 2
    assert len(captured_tools) == 2
    assert all(any(isinstance(t, _FakeMcpTool) for t in tools) for tools in captured_tools)