dschlueter
7459690bf5
Im Wurzelverzeichnis liegt nur noch TextToModel.py (Ablauf und CLI). Jede Stufe liegt mit Code und Prompt in einem eigenen Ordner: core, understanding, modeling, checks, solving, reporting, data, benchmark; Tests unter tests/, Aufgaben unter tasks/. understand_task, text_to_model/revise_model und review_assumptions/needs_work/select_best sind aus TextToModel.py zu ihren Prompts gewandert, inhaltlich unveraendert. Prompt-Kennung des Benchmarks bleibt 93744636095a. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
116 lines
4.9 KiB
Python
116 lines
4.9 KiB
Python
import threading
|
||
|
||
import pytest
|
||
|
||
from core import llm
|
||
from core.llm import is_fatal_error, is_transient_error, model_for_step, provider_for
|
||
from core.model_schema import LLMUsage
|
||
|
||
|
||
class FakeCompletions:
|
||
def __init__(self): self.kwargs = None
|
||
|
||
def create_with_completion(self, **kwargs):
|
||
self.kwargs = kwargs
|
||
return "ok", object()
|
||
|
||
|
||
def fake_session(model="deepseek/deepseek-v4.1-flash", step_models=None):
|
||
"""LLMSession ohne echten Client – der Konstruktor würde einen API-Key verlangen."""
|
||
session = llm.LLMSession.__new__(llm.LLMSession)
|
||
session.model, session.provider = model, provider_for(model)
|
||
session.step_models = step_models or {}
|
||
session.usage = LLMUsage(model=model, provider=session.provider)
|
||
session._lock = threading.Lock()
|
||
completions = {}
|
||
session._clients = {}
|
||
for provider in {session.provider, *(provider_for(m) for m in session.step_models.values())}:
|
||
completions[provider] = FakeCompletions()
|
||
session._clients[provider] = type("C", (), {"chat": type("Chat", (), {"completions": completions[provider]})()})()
|
||
session.client = session._clients[session.provider]
|
||
return session, completions
|
||
|
||
|
||
@pytest.mark.parametrize("message", [
|
||
"429 RESOURCE_EXHAUSTED. {'error': {'code': 429, 'message': 'You exceeded your current quota'}}", # Gemini-Kontingent
|
||
"Error code: 402 - {'error': {'message': 'Insufficient credits'}}", # OpenRouter-Guthaben
|
||
"Error code: 401 - {'error': {'message': 'No auth credentials found'}}",
|
||
])
|
||
def test_fatal_errors_are_not_retried(message):
|
||
assert is_fatal_error(message)
|
||
assert not is_transient_error(message)
|
||
|
||
|
||
@pytest.mark.parametrize("message", [
|
||
"503 UNAVAILABLE. {'error': {'code': 503, 'message': 'model is currently experiencing high demand'}}",
|
||
"Error code: 429 - {'error': {'message': 'Rate limit exceeded'}}",
|
||
"Error code: 502 - Bad Gateway",
|
||
"Connection error.",
|
||
])
|
||
def test_transient_errors_are_retried(message):
|
||
assert is_transient_error(message)
|
||
|
||
|
||
def test_validation_error_is_neither():
|
||
message = "1 validation error for OptimizationModelSpecification"
|
||
assert not is_fatal_error(message) and not is_transient_error(message)
|
||
|
||
|
||
def test_provider_detection():
|
||
assert provider_for("deepseek/deepseek-v4-flash-0731") == "openrouter"
|
||
assert provider_for("gemini-2.5-flash") == "gemini"
|
||
|
||
|
||
def test_reasoning_effort_is_sent_only_when_configured(monkeypatch):
|
||
def make_session(effort):
|
||
monkeypatch.setattr(llm, "REASONING_EFFORT", effort)
|
||
session, completions = fake_session()
|
||
session.create(str, [{"role": "user", "content": "x"}])
|
||
return completions["openrouter"].kwargs["extra_body"]
|
||
|
||
assert "reasoning" not in make_session("")
|
||
assert make_session("low")["reasoning"] == {"effort": "low"}
|
||
|
||
|
||
def test_truncated_answer_is_retried():
|
||
# beobachtet: ein Anbieter schnitt die Modellspezifikation ab (max_tokens-Vorgabe des Anbieters)
|
||
message = "IncompleteOutputException: The output is incomplete due to a max_tokens length limit."
|
||
assert is_transient_error(message) and not is_fatal_error(message)
|
||
|
||
|
||
def test_max_tokens_is_sent_to_openrouter():
|
||
session, completions = fake_session()
|
||
session.create(str, [{"role": "user", "content": "x"}])
|
||
assert completions["openrouter"].kwargs["max_tokens"] == llm.LLM_MAX_TOKENS >= 16000
|
||
|
||
|
||
# --- Modell je Pipeline-Stufe (LLM_MODEL_EXTRACTION, LLM_MODEL_SOLVER, ...) ---
|
||
|
||
def test_step_model_falls_back_to_session_model(monkeypatch):
|
||
monkeypatch.delenv("LLM_MODEL_SOLVER", raising=False)
|
||
assert model_for_step("Solver-Code", "a/b") == "a/b"
|
||
assert model_for_step("unbenannt", "a/b") == "a/b" # Stufe ohne Umgebungsvariable
|
||
|
||
|
||
def test_step_model_from_environment(monkeypatch):
|
||
monkeypatch.setenv("LLM_MODEL_SOLVER", "schnell/modell")
|
||
assert model_for_step("Solver-Code", "a/b") == "schnell/modell"
|
||
assert model_for_step("Solver-Reparatur", "a/b") == "schnell/modell" # dieselbe Variable
|
||
assert model_for_step("Extraktion", "a/b") == "a/b"
|
||
|
||
|
||
def test_create_uses_step_model_and_records_it():
|
||
session, completions = fake_session(step_models={"Extraktion": "stark/modell", "Solver-Code": "gemini-2.5-flash"})
|
||
session.create(str, [{"role": "user", "content": "x"}], step="Extraktion")
|
||
assert completions["openrouter"].kwargs["model"] == "stark/modell"
|
||
|
||
# anderer Anbieter je Stufe: Gemini bekommt keine OpenRouter-Parameter
|
||
session.create(str, [{"role": "user", "content": "x"}], step="Solver-Code")
|
||
assert completions["gemini"].kwargs["model"] == "gemini-2.5-flash"
|
||
assert "extra_body" not in completions["gemini"].kwargs
|
||
assert [r.model for r in session.usage.call_records] == ["stark/modell", "gemini-2.5-flash"]
|
||
|
||
|
||
def test_overrides_lists_only_deviating_steps():
|
||
session, _ = fake_session(model="a/b", step_models={"Extraktion": "a/b", "Solver-Code": "schnell/modell"})
|
||
assert session.overrides() == {"Solver-Code": "schnell/modell"}
|