Trademind/tests/test_llm.py

"""Ollama-Anbindung: Antwortverarbeitung, Fehlerdiagnose, Prompt-Aufbau."""

from __future__ import annotations

import asyncio

import pytest
from aiohttp import web

from trademind.config import OllamaConfig
from trademind.llm import SYSTEM_PROMPT, OllamaClient, build_status_prompt

STATUS = {
    "mode": "paper",
    "exchange": "binance",
    "symbols": ["BTC/USDT", "ETH/USDT"],
    "timeframe": "5m",
    "quote_currency": "USDT",
    "portfolio": {"equity": 10123.45, "total_return_pct": 1.23, "trades": 7, "win_rate": 0.42,
                  "profit_factor": 1.1, "max_drawdown_pct": 2.5, "open_positions": 1},
    "strategy": {"candidates_seen": 40, "candidates_accepted": 9,
                 "learner": {"samples_seen": 900, "trade_samples": 7, "online_accuracy": 0.53}},
    "risk": {"halted": False, "halt_reason": "", "max_open_positions": 3, "max_position_pct": 0.2},
    "trading": {"active": True, "simulated": True},
    "feature_weights": {"ema_spread": 0.39, "trend_dist": -0.29, "rsi_norm": 0.01},
    "positions": [{"symbol": "BTC/USDT", "entry_price": 70000.0, "mark_price": 70500.0,
                   "unrealized_pct": 0.71, "bars_held": 4, "confidence": 0.62}],
    "recent_trades": [{"symbol": "ETH/USDT", "pnl_pct": -0.004, "exit_reason": "stop_loss",
                       "bars_held": 9}],
}


# ------------------------------------------------------------------- Prompt


def test_prompt_contains_the_real_numbers():
    prompt = build_status_prompt(STATUS)
    for needle in ("paper", "binance", "BTC/USDT", "10123.45", "ema_spread", "stop_loss"):
        assert needle in prompt, f"{needle} fehlt im Prompt"


def test_prompt_ranks_weights_by_magnitude():
    prompt = build_status_prompt(STATUS, top_weights=2)
    assert "ema_spread" in prompt and "trend_dist" in prompt
    assert "rsi_norm" not in prompt, "das schwächste Gewicht sollte wegfallen"


def test_prompt_survives_a_bare_status():
    prompt = build_status_prompt({})
    assert "PORTFOLIO" in prompt and "LERNMODELL" in prompt


def test_system_prompt_forbids_advice():
    for needle in ("keine Anlageempfehlung", "keine Kursprognose", "Erfinde keine Zahlen"):
        assert needle in SYSTEM_PROMPT


# ------------------------------------------------------------ Falsches Ollama


def fake_ollama(handler):
    app = web.Application()
    app.router.add_post("/api/generate", handler)
    app.router.add_get("/api/tags", lambda _: web.json_response({"models": [{"name": "testmodell"}]}))
    return app


@pytest.fixture
async def client_for(aiohttp_server):
    async def _make(handler, **overrides) -> OllamaClient:
        server = await aiohttp_server(fake_ollama(handler))
        options = {"model": "testmodell", "timeout_seconds": 5, **overrides}
        config = OllamaConfig(
            enabled=True, base_url=str(server.make_url("/")).rstrip("/"), **options
        )
        return OllamaClient(config)

    return _make


async def test_plain_answer_is_returned(client_for):
    async def handler(request):
        assert (await request.json())["stream"] is False
        return web.json_response({"response": "Alles ruhig.", "done_reason": "stop"})

    client = await client_for(handler)
    result = await client.generate("frage")
    assert result.ok and result.text == "Alles ruhig."
    assert result.model == "testmodell"
    await client.close()


async def test_thinking_block_is_stripped(client_for):
    async def handler(_):
        return web.json_response(
            {"response": "<think>erst überlegen</think>Das Ergebnis.", "done_reason": "stop"}
        )

    client = await client_for(handler)
    result = await client.generate("frage")
    assert result.ok and result.text == "Das Ergebnis."
    await client.close()


async def test_reasoning_model_without_answer_is_diagnosed(client_for):
    """Der reale Fall: qwen3 verbraucht das Token-Budget für 'thinking'."""

    async def handler(_):
        return web.json_response({"response": "", "thinking": "x" * 1800, "done_reason": "length"})

    client = await client_for(handler)
    result = await client.generate("frage")
    assert not result.ok
    assert "Denkschritte" in result.error
    assert "llm.think" in result.error, "die Meldung muss den Ausweg nennen"
    await client.close()


async def test_truncated_answer_is_diagnosed(client_for):
    async def handler(_):
        return web.json_response({"response": "", "done_reason": "length"})

    client = await client_for(handler)
    result = await client.generate("frage")
    assert not result.ok and "max_tokens" in result.error
    await client.close()


async def test_think_flag_is_sent(client_for):
    seen = {}

    async def handler(request):
        seen.update(await request.json())
        return web.json_response({"response": "ok", "done_reason": "stop"})

    client = await client_for(handler)
    await client.generate("frage")
    assert seen["think"] is False, "Reasoning ist standardmäßig aus"
    await client.close()


async def test_old_ollama_without_think_field_still_works(client_for):
    """Ältere Versionen lehnen das Feld ab – dann ohne es erneut versuchen."""
    calls = []

    async def handler(request):
        body = await request.json()
        calls.append("think" in body)
        if "think" in body:
            return web.json_response({"error": "unknown field think"}, status=400)
        return web.json_response({"response": "Klappt doch.", "done_reason": "stop"})

    client = await client_for(handler)
    result = await client.generate("frage")
    assert result.ok and result.text == "Klappt doch."
    assert calls == [True, False]
    await client.close()


async def test_http_error_is_reported(client_for):
    async def handler(_):
        return web.json_response({"error": "model not found"}, status=404)

    client = await client_for(handler)
    result = await client.generate("frage")
    assert not result.ok and "404" in result.error
    await client.close()


async def test_timeout_is_reported_with_a_hint(client_for):
    async def handler(_):
        await asyncio.sleep(2)
        return web.json_response({"response": "zu spät"})

    client = await client_for(handler, timeout_seconds=0.2)
    result = await client.generate("frage")
    assert not result.ok
    assert "Zeitüberschreitung" in result.error and "timeout_seconds" in result.error
    await client.close()


async def test_unreachable_server_is_reported(aiohttp_server):
    """Server starten, wieder beenden, dann anfragen – die Verbindung wird abgelehnt."""

    async def handler(_):
        return web.json_response({"response": "nie erreicht"})

    server = await aiohttp_server(fake_ollama(handler))
    url = str(server.make_url("/")).rstrip("/")
    await server.close()

    client = OllamaClient(OllamaConfig(enabled=True, base_url=url, timeout_seconds=5))
    result = await client.generate("frage")
    assert not result.ok
    assert "nicht erreichbar" in result.error and url in result.error
    await client.close()


async def test_answer_is_capped(client_for):
    async def handler(_):
        return web.json_response({"response": "y" * 5000, "done_reason": "stop"})

    client = await client_for(handler, max_answer_chars=100)
    result = await client.generate("frage")
    assert result.ok and len(result.text) == 100
    await client.close()


async def test_model_listing(client_for):
    async def handler(_):
        return web.json_response({"response": "ok"})

    client = await client_for(handler)
    assert await client.available_models() == ["testmodell"]
    await client.close()


def test_llm_is_off_by_default():
    from trademind.config import Config

    assert Config().llm.enabled is False