astrololo/services/logic/tests/test_llm.py

"""Dostawcy LLM (LOG-31) — bez wołania jakiegokolwiek prawdziwego modelu.

Transport podstawiamy przez httpx.MockTransport, więc testy są szybkie,
deterministyczne i nic nie wychodzi na zewnątrz.
"""
import json

import httpx
import pytest

from app.llm import factory
from app.llm.base import Completion, LLMError
from app.llm.providers import AnthropicProvider, ChatCompletionsProvider


def _mock_client(handler):
    """Podmienia httpx.Client na wersję z transportem testowym."""
    class _C(httpx.Client):
        def __init__(self, *a, **kw):
            kw["transport"] = httpx.MockTransport(handler)
            super().__init__(*a, **kw)
    return _C


@pytest.fixture
def chat_ok(monkeypatch):
    def handler(request):
        assert request.url.path.endswith("/chat/completions")
        return httpx.Response(200, json={
            "model": "test-model",
            "choices": [{"message": {"content": "Horoskop testowy."}}],
            "usage": {"prompt_tokens": 100, "completion_tokens": 50},
        })
    monkeypatch.setattr(httpx, "Client", _mock_client(handler))


# ------------------------------------------------------- protokół chat/completions

def test_local_provider_generates(chat_ok):
    p = ChatCompletionsProvider("local", "http://localhost:11434/v1", "m", leaves_lan=False)
    out = p.generate("prompt", 500)
    assert isinstance(out, Completion)
    assert out.text == "Horoskop testowy."
    assert out.usage["completion_tokens"] == 50


def test_local_provider_does_not_leave_lan(chat_ok):
    p = ChatCompletionsProvider("local", "http://localhost:11434/v1", "m", leaves_lan=False)
    assert p.generate("prompt", 100).leaves_lan is False


def test_cloud_provider_marks_leaving_lan(chat_ok):
    p = ChatCompletionsProvider("openai", "https://api.openai.com/v1", "m", "klucz",
                                leaves_lan=True)
    assert p.generate("prompt", 100).leaves_lan is True


def test_api_key_sent_only_when_set(monkeypatch):
    seen = {}

    def handler(request):
        seen["auth"] = request.headers.get("authorization")
        return httpx.Response(200, json={"choices": [{"message": {"content": "x"}}]})

    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
    ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 10)
    assert seen["auth"] is None
    ChatCompletionsProvider("openai", "http://x/v1", "m", "tajny").generate("p", 10)
    assert seen["auth"] == "Bearer tajny"


def test_http_error_becomes_readable_message(monkeypatch):
    monkeypatch.setattr(httpx, "Client",
                        _mock_client(lambda r: httpx.Response(400, text="zly model")))
    with pytest.raises(LLMError, match="400"):
        ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 10)


def test_malformed_response_reported(monkeypatch):
    monkeypatch.setattr(httpx, "Client",
                        _mock_client(lambda r: httpx.Response(200, json={"nonsens": 1})))
    with pytest.raises(LLMError, match="kształt"):
        ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 10)


def test_retries_then_succeeds(monkeypatch):
    calls = {"n": 0}

    def handler(request):
        calls["n"] += 1
        if calls["n"] < 3:
            return httpx.Response(429, text="za duzo")
        return httpx.Response(200, json={"choices": [{"message": {"content": "ok"}}]})

    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
    monkeypatch.setattr("app.llm.providers.time.sleep", lambda s: None)   # bez czekania
    assert ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 10).text == "ok"
    assert calls["n"] == 3


# ------------------------------------------------------------------- Anthropic

def test_anthropic_generates(monkeypatch):
    def handler(request):
        assert request.url.path.endswith("/v1/messages")
        assert request.headers.get("x-api-key") == "klucz"
        assert request.headers.get("anthropic-version")
        return httpx.Response(200, json={
            "model": "claude-x",
            "content": [{"type": "text", "text": "Prognoza."}],
            "usage": {"input_tokens": 10, "output_tokens": 5},
        })

    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
    out = AnthropicProvider("https://api.anthropic.com", "claude-x", "klucz").generate("p", 100)
    assert out.text == "Prognoza." and out.leaves_lan is True


def test_anthropic_requires_key():
    with pytest.raises(LLMError, match="ANTHROPIC_API_KEY"):
        AnthropicProvider("https://api.anthropic.com", "m", "").generate("p", 10)


# --------------------------------------------------------------------- fabryka

def test_default_provider_is_local(monkeypatch):
    monkeypatch.delenv("LLM_PROVIDER", raising=False)
    monkeypatch.delenv("LLM_MODEL", raising=False)
    monkeypatch.delenv("LLM_BASE_URL", raising=False)
    p = factory.build_provider()
    assert p.name == "local"
    # domyślnie NIC nie opuszcza sieci — prompt niesie opisy z baz (LOG-32)
    assert p.leaves_lan is False


def test_openai_requires_key(monkeypatch):
    monkeypatch.delenv("LLM_API_KEY", raising=False)
    monkeypatch.delenv("OPENAI_API_KEY", raising=False)
    # komunikat wskazuje ZMIENNĄ DO USTAWIENIA dla tego dostawcy, nie ogólne LLM_API_KEY
    with pytest.raises(LLMError, match="OPENAI_API_KEY"):
        factory.build_provider("openai")


def test_unknown_provider_rejected():
    with pytest.raises(LLMError, match="Nieznany dostawca"):
        factory.build_provider("bzdura")


def test_env_overrides_model_and_url(monkeypatch):
    monkeypatch.setenv("LLM_MODEL", "moj-model")
    monkeypatch.setenv("LLM_BASE_URL", "http://serwer:8000/v1")
    p = factory.build_provider("local")
    assert p.model == "moj-model" and p.base_url == "http://serwer:8000/v1"


# ---------------------------------------- konfiguracja per dostawca (regresja LOG-31)
# UI pozwala przelaczac dostawce przy kazdym zadaniu, wiec ustawienia JEDNEGO nie moga
# przeciekac na pozostalych. Wczesniej wspolne LLM_BASE_URL/LLM_MODEL kierowaly zadania
# do OpenAI na adres lokalnej Ollamy i prosily Anthropic o model llama.

def _clear(monkeypatch):
    for v in ("LLM_PROVIDER", "LLM_MODEL", "LLM_BASE_URL", "LLM_API_KEY",
              "LOCAL_MODEL", "LOCAL_BASE_URL", "LOCAL_API_KEY",
              "OPENAI_MODEL", "OPENAI_BASE_URL", "OPENAI_API_KEY",
              "ANTHROPIC_MODEL", "ANTHROPIC_BASE_URL", "ANTHROPIC_API_KEY"):
        monkeypatch.delenv(v, raising=False)


def test_local_config_does_not_leak_to_cloud(monkeypatch):
    """Sedno bledu: skonfigurowany model lokalny przejmowal zadania do chmury."""
    _clear(monkeypatch)
    monkeypatch.setenv("LLM_PROVIDER", "local")
    monkeypatch.setenv("LLM_BASE_URL", "http://ollama:11434/v1")   # konfiguracja lokalnego
    monkeypatch.setenv("LLM_MODEL", "llama3.1:8b")
    monkeypatch.setenv("OPENAI_API_KEY", "sk-test")

    local = factory.build_provider("local")
    assert local.base_url == "http://ollama:11434/v1" and local.model == "llama3.1:8b"

    openai = factory.build_provider("openai")
    assert openai.base_url == "https://api.openai.com/v1", "zadanie do OpenAI poszloby do Ollamy"
    assert openai.model == "gpt-4o-mini", "OpenAI dostalby nazwe modelu llama"


def test_provider_specific_settings_win(monkeypatch):
    _clear(monkeypatch)
    monkeypatch.setenv("LLM_PROVIDER", "local")
    monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-ant")
    monkeypatch.setenv("ANTHROPIC_MODEL", "claude-opus-4-8")
    p = factory.build_provider("anthropic")
    assert p.model == "claude-opus-4-8" and p.api_key == "sk-ant"


def test_generic_vars_apply_only_to_default_provider(monkeypatch):
    """Zgodnosc wstecz: wspolne LLM_* konfiguruja dostawce domyslnego i tylko jego."""
    _clear(monkeypatch)
    monkeypatch.setenv("LLM_PROVIDER", "openai")
    monkeypatch.setenv("LLM_API_KEY", "sk-generic")
    monkeypatch.setenv("LLM_MODEL", "gpt-4o")
    assert factory.build_provider("openai").model == "gpt-4o"
    assert factory.build_provider("local").model == "llama3.1:8b"   # nie dziedziczy


def test_cloud_without_key_is_rejected_clearly(monkeypatch):
    _clear(monkeypatch)
    monkeypatch.setenv("LLM_PROVIDER", "local")
    for name in ("openai", "anthropic"):
        with pytest.raises(LLMError, match=f"{name.upper()}_API_KEY"):
            factory.build_provider(name)


def test_local_needs_no_key(monkeypatch):
    _clear(monkeypatch)
    assert factory.build_provider("local").api_key == ""


# ------------------------------------------- pusta odpowiedz modelu (cicha awaria)
# Regresja: model potrafi oddac pusta tresc (prompt zjadl caly kontekst ->
# finish_reason=length, completion_tokens=0). Wczesniej generate() zwracalo pusty
# tekst BEZ bledu, widok nic nie renderowal i uzytkownik dostawal pusta strone
# bez zadnego wyjasnienia. Pusta odpowiedz MUSI byc bledem.

def test_empty_completion_raises_instead_of_silent_blank(monkeypatch):
    monkeypatch.setattr(httpx, "Client", _mock_client(lambda r: httpx.Response(200, json={
        "model": "llama3.1:8b",
        "choices": [{"message": {"content": ""}, "finish_reason": "length"}],
        "usage": {"prompt_tokens": 8000, "completion_tokens": 0},
    })))
    with pytest.raises(LLMError, match="nie zwrócił żadnej treści"):
        ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 2000)


def test_empty_completion_explains_context_window(monkeypatch):
    """Komunikat ma prowadzic do przyczyny, a nie tylko stwierdzac fakt."""
    monkeypatch.setattr(httpx, "Client", _mock_client(lambda r: httpx.Response(200, json={
        "choices": [{"message": {"content": "   "}, "finish_reason": "length"}],
        "usage": {"prompt_tokens": 8000, "completion_tokens": 0},
    })))
    with pytest.raises(LLMError) as ei:
        ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 2000)
    msg = str(ei.value)
    assert "kontekstu" in msg and "budżet" in msg
    assert "powód zakończenia: length" in msg      # diagnostyka w tresci bledu


def test_whitespace_only_is_treated_as_empty(monkeypatch):
    monkeypatch.setattr(httpx, "Client", _mock_client(lambda r: httpx.Response(200, json={
        "choices": [{"message": {"content": "\n\n  \t "}}],
    })))
    with pytest.raises(LLMError, match="nie zwrócił żadnej treści"):
        ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 100)


def test_anthropic_empty_completion_raises(monkeypatch):
    monkeypatch.setattr(httpx, "Client", _mock_client(lambda r: httpx.Response(200, json={
        "model": "claude-x", "content": [], "stop_reason": "max_tokens",
        "usage": {"input_tokens": 9000, "output_tokens": 0},
    })))
    with pytest.raises(LLMError, match="nie zwrócił żadnej treści"):
        AnthropicProvider("https://api.anthropic.com", "claude-x", "klucz").generate("p", 100)


def test_normal_response_still_passes(monkeypatch):
    """Straznik nie moze psuc poprawnej odpowiedzi."""
    monkeypatch.setattr(httpx, "Client", _mock_client(lambda r: httpx.Response(200, json={
        "choices": [{"message": {"content": "Horoskop."}, "finish_reason": "stop"}],
        "usage": {"prompt_tokens": 100, "completion_tokens": 20},
    })))
    assert ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 100).text == "Horoskop."


# ------------------------------------- kontynuacja: horoskop MA powstac zawsze
# Sedno wymagania: niezaleznie od objetosci promptu i limitu wyjscia, pelna tresc
# ma wrocic do uzytkownika. Model urwany na max_tokens jest proszony o dokonczenie
# w ramach tej samej rozmowy, a kawalki sa sklejane.

def _scripted(responses):
    """Transport oddajacy kolejne odpowiedzi z listy (po jednej na ture)."""
    seq = list(responses)
    seen = []

    def handler(request):
        seen.append(request)
        return httpx.Response(200, json=seq.pop(0) if seq else seq_last)

    seq_last = responses[-1]
    return handler, seen


def _chat(text, finish):
    return {"choices": [{"message": {"content": text}, "finish_reason": finish}],
            "usage": {"completion_tokens": 10}}


def test_truncated_answer_is_continued_and_joined(monkeypatch):
    handler, seen = _scripted([
        _chat("Czesc pierwsza.", "length"),
        _chat("Czesc druga. KONIEC", "stop"),
    ])
    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
    out = ChatCompletionsProvider("local", "http://x/v1", "m").generate("prompt", 40000)
    assert "Czesc pierwsza." in out.text and "Czesc druga." in out.text
    assert "KONIEC" not in out.text          # znacznik nie trafia do horoskopu
    assert out.usage["turns"] == 2


def test_continuation_asks_in_same_conversation(monkeypatch):
    """Kontynuacja musi isc jako kolejna tura rozmowy, a ostatnia wiadomosc MUSI
    byc od uzytkownika — Claude odrzuca prefill w turze asystenta (400)."""
    handler, seen = _scripted([_chat("Poczatek", "length"), _chat("Reszta", "stop")])
    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
    ChatCompletionsProvider("local", "http://x/v1", "m").generate("prompt", 40000)
    msgs = json.loads(seen[1].content)["messages"]
    assert msgs[-1]["role"] == "user", "ostatnia wiadomosc nie moze byc prefillem asystenta"
    assert msgs[1]["role"] == "assistant" and "Poczatek" in msgs[1]["content"]


def test_anthropic_thinking_only_turn_is_continued(monkeypatch):
    """DOKLADNIE zgloszony objaw: cala tura poszla na myslenie, tekst pusty.
    Wczesniej konczylo sie to pusta strona; teraz pytamy o tresc dalej."""
    seq = [
        {"content": [{"type": "thinking", "thinking": ""}], "stop_reason": "max_tokens",
         "usage": {"output_tokens": 2000}},
        {"content": [{"type": "text", "text": "Horoskop urodzeniowy..."}],
         "stop_reason": "end_turn", "usage": {"output_tokens": 500}},
    ]

    def handler(request):
        return httpx.Response(200, json=seq.pop(0) if seq else seq[-1])

    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
    out = AnthropicProvider("https://api.anthropic.com", "claude-opus-4-8", "k").generate("p", 40000)
    assert out.text == "Horoskop urodzeniowy..."
    assert out.usage["turns"] == 2


def test_anthropic_sends_thinking_config(monkeypatch):
    """Bez jawnego `thinking` Sonnet 5 wlacza myslenie sam — konfigurujemy to wprost."""
    seen = []

    def handler(request):
        seen.append(json.loads(request.content))
        return httpx.Response(200, json={"content": [{"type": "text", "text": "ok"}],
                                         "stop_reason": "end_turn"})

    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
    monkeypatch.delenv("ANTHROPIC_THINKING", raising=False)
    AnthropicProvider("https://api.anthropic.com", "claude-opus-4-8", "k").generate("p", 5000)
    assert seen[0]["thinking"] == {"type": "adaptive"}

    seen.clear()
    monkeypatch.setenv("ANTHROPIC_THINKING", "off")
    AnthropicProvider("https://api.anthropic.com", "claude-opus-4-8", "k").generate("p", 5000)
    assert seen[0]["thinking"] == {"type": "disabled"}


def test_complete_answer_does_not_loop(monkeypatch):
    """Straznik nie moze mnozyc zapytan, gdy model skonczyl normalnie."""
    calls = {"n": 0}

    def handler(request):
        calls["n"] += 1
        return httpx.Response(200, json=_chat("Gotowe.", "stop"))

    monkeypatch.setattr(httpx, "Client", _mock_client(handler))
    out = ChatCompletionsProvider("local", "http://x/v1", "m").generate("p", 40000)
    assert out.text == "Gotowe." and calls["n"] == 1


# ------------------------------------------- wybor modelu przez uzytkownika (UI)

def test_model_from_request_wins_over_config(monkeypatch):
    _clear(monkeypatch)
    monkeypatch.setenv("ANTHROPIC_MODEL", "claude-opus-4-8")
    monkeypatch.setenv("ANTHROPIC_API_KEY", "k")
    p = factory.build_provider("anthropic", "claude-fable-5")
    assert p.model == "claude-fable-5", "wybor z UI musi wygrac nad konfiguracja"


def test_blank_model_falls_back_to_configured_default(monkeypatch):
    _clear(monkeypatch)
    monkeypatch.setenv("ANTHROPIC_MODEL", "claude-sonnet-5")
    monkeypatch.setenv("ANTHROPIC_API_KEY", "k")
    assert factory.build_provider("anthropic", "   ").model == "claude-sonnet-5"


def test_resolve_model_needs_no_api_key(monkeypatch):
    """Budzet promptu zalezy od okna kontekstu modelu — nie moze wymagac klucza.

    Wczesniej liczenie budzetu szlo przez build_provider(), ktory bez klucza
    rzuca bledem, wiec „maksymalny kontekst" cicho spadal do wartosci zapasowej.
    """
    _clear(monkeypatch)
    provider, model = factory.resolve_model("anthropic", "claude-haiku-4-5")
    assert (provider, model) == ("anthropic", "claude-haiku-4-5")
    with pytest.raises(LLMError):                 # samo zbudowanie nadal wymaga klucza
        factory.build_provider("anthropic", "claude-haiku-4-5")


def test_max_budget_differs_between_models(monkeypatch):
    """Sedno funkcji: wieksze okno = wiekszy budzet promptu."""
    from app.llm.limits import prompt_token_budget
    _clear(monkeypatch)
    opus = prompt_token_budget(*factory.resolve_model("anthropic", "claude-opus-4-8"))
    haiku = prompt_token_budget(*factory.resolve_model("anthropic", "claude-haiku-4-5"))
    local = prompt_token_budget(*factory.resolve_model("local", "llama3.1:8b"))
    assert opus > haiku > local > 0