tests/test_orchestrator.py

from __future__ import annotations

import pytest

from app.llm.base import ChatResult, ProviderError, ToolCall
from app.models import ChatMessage, ChatSession
from app.orchestrator import run_turn
from app.tools.registry import ToolContext


class FakeRouter:
    def __init__(self, results: list[ChatResult]):
        self._results = list(results)
        self.calls = 0

    async def chat(self, messages, tools, tier="large", temperature=None):
        self.calls += 1
        return self._results.pop(0)


class FakeShopify:
    async def execute(self, query, variables=None):
        return {"orders": {"edges": []}}


@pytest.fixture(autouse=True)
def no_real_search(monkeypatch):
    async def fake_search(session, query, k=4, *, tenant_id=None):
        return []

    from app.rag import index

    monkeypatch.setattr(index, "search", fake_search)


async def _session(db):
    s = ChatSession(shop="x")
    db.add(s)
    await db.flush()
    return s


async def test_info_turn_uses_knowledge_tool_then_answers(db_session):
    router = FakeRouter(
        [
            ChatResult(
                content=None,
                tool_calls=[ToolCall(id="t1", name="search_knowledge", arguments={"query": "envíos"})],
                finish_reason="tool_calls",
            ),
            ChatResult(content="Hacemos envíos gratis desde 199€.", tool_calls=[], finish_reason="stop"),
        ]
    )
    ctx = ToolContext(db=db_session, session=await _session(db_session))
    resp = await run_turn(router, ctx, "¿hacéis envíos gratis?", brand_name="Tienda")

    assert resp.reply == "Hacemos envíos gratis desde 199€."
    assert resp.used_tools == ["search_knowledge"]
    assert router.calls == 2

    # persisted user + assistant
    from sqlalchemy import select

    msgs = (await db_session.execute(select(ChatMessage))).scalars().all()
    roles = [m.role for m in msgs]
    assert "user" in roles and "assistant" in roles


async def test_order_turn_unverified_asks_for_credentials(db_session):
    router = FakeRouter(
        [
            ChatResult(
                content=None,
                tool_calls=[ToolCall(id="o1", name="lookup_order", arguments={})],
                finish_reason="tool_calls",
            ),
            ChatResult(
                content="Para localizar tu pedido necesito tu email y número de pedido.",
                tool_calls=[],
                finish_reason="stop",
            ),
        ]
    )
    ctx = ToolContext(db=db_session, session=await _session(db_session), shopify=FakeShopify())
    resp = await run_turn(router, ctx, "¿dónde está mi pedido?")
    assert "email" in resp.reply.lower()
    assert resp.used_tools == ["lookup_order"]


class _ToolFailRouter:
    """Fails (ProviderError) whenever tools are sent; answers when no tools."""

    def __init__(self, reply="Respuesta sin herramientas."):
        self.reply = reply
        self.calls_with_tools = 0

    async def chat(self, messages, tools, tier="large", temperature=None):
        if tools:
            self.calls_with_tools += 1
            raise ProviderError("tool_use_failed")
        return ChatResult(content=self.reply, tool_calls=[], finish_reason="stop")


class _AlwaysFailRouter:
    async def chat(self, messages, tools, tier="large", temperature=None):
        raise ProviderError("down")


async def test_degrades_to_no_tools_when_tool_calling_fails(db_session):
    router = _ToolFailRouter()
    ctx = ToolContext(db=db_session, session=await _session(db_session))
    resp = await run_turn(router, ctx, "¿qué vendéis?")
    assert resp.reply == "Respuesta sin herramientas."
    assert router.calls_with_tools >= 1  # it tried tools first, then degraded


async def test_total_provider_failure_returns_fallback_not_500(db_session):
    from app.orchestrator import FALLBACK_REPLY

    ctx = ToolContext(db=db_session, session=await _session(db_session))
    resp = await run_turn(_AlwaysFailRouter(), ctx, "hola")
    assert resp.reply == FALLBACK_REPLY


async def test_tool_budget_exhausted_forces_final_answer(db_session):
    # Always returns tool_calls; orchestrator must stop and force a final answer.
    loop_result = ChatResult(
        content=None,
        tool_calls=[ToolCall(id="t", name="search_knowledge", arguments={"query": "x"})],
        finish_reason="tool_calls",
    )
    router = FakeRouter(
        [loop_result, loop_result, loop_result, loop_result,
         ChatResult(content="Respuesta final.", tool_calls=[], finish_reason="stop")]
    )
    ctx = ToolContext(db=db_session, session=await _session(db_session))
    resp = await run_turn(router, ctx, "bucle")
    assert resp.reply == "Respuesta final."
    assert router.calls == 5  # 4 tool iters + 1 forced final